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
376 if (pos >= f->real_file_size)
377 pos = f->file_offset;
387 static int get_next_block(struct thread_data *td, struct io_u *io_u,
388 enum fio_ddir ddir, int rw_seq,
389 unsigned int *is_random)
391 struct fio_file *f = io_u->file;
395 assert(ddir_rw(ddir));
401 if (should_do_random(td, ddir)) {
402 ret = get_next_rand_block(td, f, ddir, &b);
406 io_u_set(io_u, IO_U_F_BUSY_OK);
407 ret = get_next_seq_offset(td, f, ddir, &offset);
409 ret = get_next_rand_block(td, f, ddir, &b);
413 ret = get_next_seq_offset(td, f, ddir, &offset);
416 io_u_set(io_u, IO_U_F_BUSY_OK);
419 if (td->o.rw_seq == RW_SEQ_SEQ) {
420 ret = get_next_seq_offset(td, f, ddir, &offset);
422 ret = get_next_rand_block(td, f, ddir, &b);
425 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
426 if (f->last_start[ddir] != -1ULL)
427 offset = f->last_start[ddir] - f->file_offset;
432 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
439 io_u->offset = offset;
441 io_u->offset = b * td->o.ba[ddir];
443 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
452 * For random io, generate a random new block and see if it's used. Repeat
453 * until we find a free one. For sequential io, just return the end of
454 * the last io issued.
456 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
457 unsigned int *is_random)
459 struct fio_file *f = io_u->file;
460 enum fio_ddir ddir = io_u->ddir;
463 assert(ddir_rw(ddir));
465 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
467 td->ddir_seq_nr = td->o.ddir_seq_nr;
470 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
473 if (io_u->offset >= f->io_size) {
474 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
475 (unsigned long long) io_u->offset,
476 (unsigned long long) f->io_size);
480 io_u->offset += f->file_offset;
481 if (io_u->offset >= f->real_file_size) {
482 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
483 (unsigned long long) io_u->offset,
484 (unsigned long long) f->real_file_size);
491 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
492 unsigned int *is_random)
494 if (td->flags & TD_F_PROFILE_OPS) {
495 struct prof_io_ops *ops = &td->prof_io_ops;
497 if (ops->fill_io_u_off)
498 return ops->fill_io_u_off(td, io_u, is_random);
501 return __get_next_offset(td, io_u, is_random);
504 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
507 struct fio_file *f = io_u->file;
509 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
512 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
513 unsigned int is_random)
515 int ddir = io_u->ddir;
516 unsigned int buflen = 0;
517 unsigned int minbs, maxbs;
521 assert(ddir_rw(ddir));
523 if (td->o.bs_is_seq_rand)
524 ddir = is_random ? DDIR_WRITE: DDIR_READ;
526 minbs = td->o.min_bs[ddir];
527 maxbs = td->o.max_bs[ddir];
533 * If we can't satisfy the min block size from here, then fail
535 if (!io_u_fits(td, io_u, minbs))
538 frand_max = rand_max(&td->bsrange_state);
540 r = __rand(&td->bsrange_state);
542 if (!td->o.bssplit_nr[ddir]) {
543 buflen = 1 + (unsigned int) ((double) maxbs *
544 (r / (frand_max + 1.0)));
551 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
552 struct bssplit *bsp = &td->o.bssplit[ddir][i];
556 if ((r * 100UL <= frand_max * perc) &&
557 io_u_fits(td, io_u, buflen))
562 if (td->o.verify != VERIFY_NONE)
563 buflen = (buflen + td->o.verify_interval - 1) &
564 ~(td->o.verify_interval - 1);
566 if (!td->o.bs_unaligned && is_power_of_2(minbs))
567 buflen &= ~(minbs - 1);
569 } while (!io_u_fits(td, io_u, buflen));
574 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
575 unsigned int is_random)
577 if (td->flags & TD_F_PROFILE_OPS) {
578 struct prof_io_ops *ops = &td->prof_io_ops;
580 if (ops->fill_io_u_size)
581 return ops->fill_io_u_size(td, io_u, is_random);
584 return __get_next_buflen(td, io_u, is_random);
587 static void set_rwmix_bytes(struct thread_data *td)
592 * we do time or byte based switch. this is needed because
593 * buffered writes may issue a lot quicker than they complete,
594 * whereas reads do not.
596 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
597 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
600 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
604 v = rand32_between(&td->rwmix_state, 1, 100);
606 if (v <= td->o.rwmix[DDIR_READ])
612 int io_u_quiesce(struct thread_data *td)
617 * We are going to sleep, ensure that we flush anything pending as
618 * not to skew our latency numbers.
620 * Changed to only monitor 'in flight' requests here instead of the
621 * td->cur_depth, b/c td->cur_depth does not accurately represent
622 * io's that have been actually submitted to an async engine,
623 * and cur_depth is meaningless for sync engines.
625 if (td->io_u_queued || td->cur_depth) {
628 ret = td_io_commit(td);
631 while (td->io_u_in_flight) {
634 ret = io_u_queued_complete(td, 1);
642 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
644 enum fio_ddir odir = ddir ^ 1;
647 assert(ddir_rw(ddir));
648 now = utime_since_now(&td->start);
651 * if rate_next_io_time is in the past, need to catch up to rate
653 if (td->rate_next_io_time[ddir] <= now)
657 * We are ahead of rate in this direction. See if we
660 if (td_rw(td) && td->o.rwmix[odir]) {
662 * Other direction is behind rate, switch
664 if (td->rate_next_io_time[odir] <= now)
668 * Both directions are ahead of rate. sleep the min
669 * switch if necissary
671 if (td->rate_next_io_time[ddir] <=
672 td->rate_next_io_time[odir]) {
673 usec = td->rate_next_io_time[ddir] - now;
675 usec = td->rate_next_io_time[odir] - now;
679 usec = td->rate_next_io_time[ddir] - now;
681 if (td->o.io_submit_mode == IO_MODE_INLINE)
684 usec = usec_sleep(td, usec);
690 * Return the data direction for the next io_u. If the job is a
691 * mixed read/write workload, check the rwmix cycle and switch if
694 static enum fio_ddir get_rw_ddir(struct thread_data *td)
699 * see if it's time to fsync
701 if (td->o.fsync_blocks &&
702 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
703 td->io_issues[DDIR_WRITE] && should_fsync(td))
707 * see if it's time to fdatasync
709 if (td->o.fdatasync_blocks &&
710 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
711 td->io_issues[DDIR_WRITE] && should_fsync(td))
712 return DDIR_DATASYNC;
715 * see if it's time to sync_file_range
717 if (td->sync_file_range_nr &&
718 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
719 td->io_issues[DDIR_WRITE] && should_fsync(td))
720 return DDIR_SYNC_FILE_RANGE;
724 * Check if it's time to seed a new data direction.
726 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
728 * Put a top limit on how many bytes we do for
729 * one data direction, to avoid overflowing the
732 ddir = get_rand_ddir(td);
734 if (ddir != td->rwmix_ddir)
737 td->rwmix_ddir = ddir;
739 ddir = td->rwmix_ddir;
740 } else if (td_read(td))
742 else if (td_write(td))
747 td->rwmix_ddir = rate_ddir(td, ddir);
748 return td->rwmix_ddir;
751 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
753 enum fio_ddir ddir = get_rw_ddir(td);
755 if (td_trimwrite(td)) {
756 struct fio_file *f = io_u->file;
757 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
763 io_u->ddir = io_u->acct_ddir = ddir;
765 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
766 td->o.barrier_blocks &&
767 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
768 td->io_issues[DDIR_WRITE])
769 io_u_set(io_u, IO_U_F_BARRIER);
772 void put_file_log(struct thread_data *td, struct fio_file *f)
774 unsigned int ret = put_file(td, f);
777 td_verror(td, ret, "file close");
780 void put_io_u(struct thread_data *td, struct io_u *io_u)
787 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
788 put_file_log(td, io_u->file);
791 io_u_set(io_u, IO_U_F_FREE);
793 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
795 assert(!(td->flags & TD_F_CHILD));
797 io_u_qpush(&td->io_u_freelist, io_u);
799 td_io_u_free_notify(td);
802 void clear_io_u(struct thread_data *td, struct io_u *io_u)
804 io_u_clear(io_u, IO_U_F_FLIGHT);
808 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
810 struct io_u *__io_u = *io_u;
811 enum fio_ddir ddir = acct_ddir(__io_u);
813 dprint(FD_IO, "requeue %p\n", __io_u);
820 io_u_set(__io_u, IO_U_F_FREE);
821 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
822 td->io_issues[ddir]--;
824 io_u_clear(__io_u, IO_U_F_FLIGHT);
825 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
827 assert(!(td->flags & TD_F_CHILD));
830 io_u_rpush(&td->io_u_requeues, __io_u);
832 td_io_u_free_notify(td);
836 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
838 unsigned int is_random;
840 if (td->io_ops->flags & FIO_NOIO)
843 set_rw_ddir(td, io_u);
846 * fsync() or fdatasync() or trim etc, we are done
848 if (!ddir_rw(io_u->ddir))
852 * See if it's time to switch to a new zone
854 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
855 struct fio_file *f = io_u->file;
858 f->file_offset += td->o.zone_range + td->o.zone_skip;
861 * Wrap from the beginning, if we exceed the file size
863 if (f->file_offset >= f->real_file_size)
864 f->file_offset = f->real_file_size - f->file_offset;
865 f->last_pos[io_u->ddir] = f->file_offset;
866 td->io_skip_bytes += td->o.zone_skip;
870 * No log, let the seq/rand engine retrieve the next buflen and
873 if (get_next_offset(td, io_u, &is_random)) {
874 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
878 io_u->buflen = get_next_buflen(td, io_u, is_random);
880 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
884 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
885 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
886 dprint(FD_IO, " off=%llu/%lu > %llu\n",
887 (unsigned long long) io_u->offset, io_u->buflen,
888 (unsigned long long) io_u->file->real_file_size);
893 * mark entry before potentially trimming io_u
895 if (td_random(td) && file_randommap(td, io_u->file))
896 mark_random_map(td, io_u);
899 dprint_io_u(io_u, "fill_io_u");
900 td->zone_bytes += io_u->buflen;
904 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
933 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
935 __io_u_mark_map(td->ts.io_u_submit, nr);
936 td->ts.total_submit++;
939 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
941 __io_u_mark_map(td->ts.io_u_complete, nr);
942 td->ts.total_complete++;
945 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
949 switch (td->cur_depth) {
971 td->ts.io_u_map[idx] += nr;
974 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1011 assert(idx < FIO_IO_U_LAT_U_NR);
1012 td->ts.io_u_lat_u[idx]++;
1015 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1056 assert(idx < FIO_IO_U_LAT_M_NR);
1057 td->ts.io_u_lat_m[idx]++;
1060 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1063 io_u_mark_lat_usec(td, usec);
1065 io_u_mark_lat_msec(td, usec / 1000);
1069 * Get next file to service by choosing one at random
1071 static struct fio_file *get_next_file_rand(struct thread_data *td,
1072 enum fio_file_flags goodf,
1073 enum fio_file_flags badf)
1075 uint64_t frand_max = rand_max(&td->next_file_state);
1083 r = __rand(&td->next_file_state);
1084 fno = (unsigned int) ((double) td->o.nr_files
1085 * (r / (frand_max + 1.0)));
1088 if (fio_file_done(f))
1091 if (!fio_file_open(f)) {
1094 if (td->nr_open_files >= td->o.open_files)
1095 return ERR_PTR(-EBUSY);
1097 err = td_io_open_file(td, f);
1103 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1104 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1108 td_io_close_file(td, f);
1113 * Get next file to service by doing round robin between all available ones
1115 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1118 unsigned int old_next_file = td->next_file;
1124 f = td->files[td->next_file];
1127 if (td->next_file >= td->o.nr_files)
1130 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1131 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);
1144 dprint(FD_FILE, "error %d on open of %s\n",
1152 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1154 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1158 td_io_close_file(td, f);
1161 } while (td->next_file != old_next_file);
1163 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1167 static struct fio_file *__get_next_file(struct thread_data *td)
1171 assert(td->o.nr_files <= td->files_index);
1173 if (td->nr_done_files >= td->o.nr_files) {
1174 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1175 " nr_files=%d\n", td->nr_open_files,
1181 f = td->file_service_file;
1182 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1183 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1185 if (td->file_service_left--)
1189 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1190 td->o.file_service_type == FIO_FSERVICE_SEQ)
1191 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1193 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1198 td->file_service_file = f;
1199 td->file_service_left = td->file_service_nr - 1;
1202 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1204 dprint(FD_FILE, "get_next_file: NULL\n");
1208 static struct fio_file *get_next_file(struct thread_data *td)
1210 if (td->flags & TD_F_PROFILE_OPS) {
1211 struct prof_io_ops *ops = &td->prof_io_ops;
1213 if (ops->get_next_file)
1214 return ops->get_next_file(td);
1217 return __get_next_file(td);
1220 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1225 f = get_next_file(td);
1226 if (IS_ERR_OR_NULL(f))
1232 if (!fill_io_u(td, io_u))
1235 put_file_log(td, f);
1236 td_io_close_file(td, f);
1238 fio_file_set_done(f);
1239 td->nr_done_files++;
1240 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1241 td->nr_done_files, td->o.nr_files);
1247 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1248 unsigned long tusec, unsigned long max_usec)
1251 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1252 td_verror(td, ETIMEDOUT, "max latency exceeded");
1253 icd->error = ETIMEDOUT;
1256 static void lat_new_cycle(struct thread_data *td)
1258 fio_gettime(&td->latency_ts, NULL);
1259 td->latency_ios = ddir_rw_sum(td->io_blocks);
1260 td->latency_failed = 0;
1264 * We had an IO outside the latency target. Reduce the queue depth. If we
1265 * are at QD=1, then it's time to give up.
1267 static bool __lat_target_failed(struct thread_data *td)
1269 if (td->latency_qd == 1)
1272 td->latency_qd_high = td->latency_qd;
1274 if (td->latency_qd == td->latency_qd_low)
1275 td->latency_qd_low--;
1277 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1279 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1282 * When we ramp QD down, quiesce existing IO to prevent
1283 * a storm of ramp downs due to pending higher depth.
1290 static bool lat_target_failed(struct thread_data *td)
1292 if (td->o.latency_percentile.u.f == 100.0)
1293 return __lat_target_failed(td);
1295 td->latency_failed++;
1299 void lat_target_init(struct thread_data *td)
1301 td->latency_end_run = 0;
1303 if (td->o.latency_target) {
1304 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1305 fio_gettime(&td->latency_ts, NULL);
1307 td->latency_qd_high = td->o.iodepth;
1308 td->latency_qd_low = 1;
1309 td->latency_ios = ddir_rw_sum(td->io_blocks);
1311 td->latency_qd = td->o.iodepth;
1314 void lat_target_reset(struct thread_data *td)
1316 if (!td->latency_end_run)
1317 lat_target_init(td);
1320 static void lat_target_success(struct thread_data *td)
1322 const unsigned int qd = td->latency_qd;
1323 struct thread_options *o = &td->o;
1325 td->latency_qd_low = td->latency_qd;
1328 * If we haven't failed yet, we double up to a failing value instead
1329 * of bisecting from highest possible queue depth. If we have set
1330 * a limit other than td->o.iodepth, bisect between that.
1332 if (td->latency_qd_high != o->iodepth)
1333 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1335 td->latency_qd *= 2;
1337 if (td->latency_qd > o->iodepth)
1338 td->latency_qd = o->iodepth;
1340 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1343 * Same as last one, we are done. Let it run a latency cycle, so
1344 * we get only the results from the targeted depth.
1346 if (td->latency_qd == qd) {
1347 if (td->latency_end_run) {
1348 dprint(FD_RATE, "We are done\n");
1351 dprint(FD_RATE, "Quiesce and final run\n");
1353 td->latency_end_run = 1;
1354 reset_all_stats(td);
1363 * Check if we can bump the queue depth
1365 void lat_target_check(struct thread_data *td)
1367 uint64_t usec_window;
1371 usec_window = utime_since_now(&td->latency_ts);
1372 if (usec_window < td->o.latency_window)
1375 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1376 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1377 success_ios *= 100.0;
1379 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1381 if (success_ios >= td->o.latency_percentile.u.f)
1382 lat_target_success(td);
1384 __lat_target_failed(td);
1388 * If latency target is enabled, we might be ramping up or down and not
1389 * using the full queue depth available.
1391 bool queue_full(const struct thread_data *td)
1393 const int qempty = io_u_qempty(&td->io_u_freelist);
1397 if (!td->o.latency_target)
1400 return td->cur_depth >= td->latency_qd;
1403 struct io_u *__get_io_u(struct thread_data *td)
1405 struct io_u *io_u = NULL;
1413 if (!io_u_rempty(&td->io_u_requeues))
1414 io_u = io_u_rpop(&td->io_u_requeues);
1415 else if (!queue_full(td)) {
1416 io_u = io_u_qpop(&td->io_u_freelist);
1421 io_u->end_io = NULL;
1425 assert(io_u->flags & IO_U_F_FREE);
1426 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1427 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1431 io_u->acct_ddir = -1;
1433 assert(!(td->flags & TD_F_CHILD));
1434 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1436 } else if (td_async_processing(td)) {
1438 * We ran out, wait for async verify threads to finish and
1441 assert(!(td->flags & TD_F_CHILD));
1442 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1450 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1452 if (!(td->flags & TD_F_TRIM_BACKLOG))
1455 if (td->trim_entries) {
1458 if (td->trim_batch) {
1461 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1462 td->last_ddir != DDIR_READ) {
1463 td->trim_batch = td->o.trim_batch;
1464 if (!td->trim_batch)
1465 td->trim_batch = td->o.trim_backlog;
1469 if (get_trim && !get_next_trim(td, io_u))
1476 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1478 if (!(td->flags & TD_F_VER_BACKLOG))
1481 if (td->io_hist_len) {
1484 if (td->verify_batch)
1486 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1487 td->last_ddir != DDIR_READ) {
1488 td->verify_batch = td->o.verify_batch;
1489 if (!td->verify_batch)
1490 td->verify_batch = td->o.verify_backlog;
1494 if (get_verify && !get_next_verify(td, io_u)) {
1504 * Fill offset and start time into the buffer content, to prevent too
1505 * easy compressible data for simple de-dupe attempts. Do this for every
1506 * 512b block in the range, since that should be the smallest block size
1507 * we can expect from a device.
1509 static void small_content_scramble(struct io_u *io_u)
1511 unsigned int i, nr_blocks = io_u->buflen / 512;
1513 unsigned int offset;
1520 boffset = io_u->offset;
1521 io_u->buf_filled_len = 0;
1523 for (i = 0; i < nr_blocks; i++) {
1525 * Fill the byte offset into a "random" start offset of
1526 * the buffer, given by the product of the usec time
1527 * and the actual offset.
1529 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1530 offset &= ~(sizeof(uint64_t) - 1);
1531 if (offset >= 512 - sizeof(uint64_t))
1532 offset -= sizeof(uint64_t);
1533 memcpy(p + offset, &boffset, sizeof(boffset));
1535 end = p + 512 - sizeof(io_u->start_time);
1536 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1543 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1544 * etc. The returned io_u is fully ready to be prepped and submitted.
1546 struct io_u *get_io_u(struct thread_data *td)
1550 int do_scramble = 0;
1553 io_u = __get_io_u(td);
1555 dprint(FD_IO, "__get_io_u failed\n");
1559 if (check_get_verify(td, io_u))
1561 if (check_get_trim(td, io_u))
1565 * from a requeue, io_u already setup
1571 * If using an iolog, grab next piece if any available.
1573 if (td->flags & TD_F_READ_IOLOG) {
1574 if (read_iolog_get(td, io_u))
1576 } else if (set_io_u_file(td, io_u)) {
1578 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1584 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1588 assert(fio_file_open(f));
1590 if (ddir_rw(io_u->ddir)) {
1591 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1592 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1596 f->last_start[io_u->ddir] = io_u->offset;
1597 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1599 if (io_u->ddir == DDIR_WRITE) {
1600 if (td->flags & TD_F_REFILL_BUFFERS) {
1601 io_u_fill_buffer(td, io_u,
1602 td->o.min_bs[DDIR_WRITE],
1604 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1605 !(td->flags & TD_F_COMPRESS))
1607 if (td->flags & TD_F_VER_NONE) {
1608 populate_verify_io_u(td, io_u);
1611 } else if (io_u->ddir == DDIR_READ) {
1613 * Reset the buf_filled parameters so next time if the
1614 * buffer is used for writes it is refilled.
1616 io_u->buf_filled_len = 0;
1621 * Set io data pointers.
1623 io_u->xfer_buf = io_u->buf;
1624 io_u->xfer_buflen = io_u->buflen;
1628 if (!td_io_prep(td, io_u)) {
1629 if (!td->o.disable_lat)
1630 fio_gettime(&io_u->start_time, NULL);
1632 small_content_scramble(io_u);
1636 dprint(FD_IO, "get_io_u failed\n");
1638 return ERR_PTR(ret);
1641 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1643 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1645 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1648 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1649 io_u->file ? " on file " : "",
1650 io_u->file ? io_u->file->file_name : "",
1651 strerror(io_u->error),
1652 io_ddir_name(io_u->ddir),
1653 io_u->offset, io_u->xfer_buflen);
1655 if (td->io_ops->errdetails) {
1656 char *err = td->io_ops->errdetails(io_u);
1658 log_err("fio: %s\n", err);
1663 td_verror(td, io_u->error, "io_u error");
1666 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1668 __io_u_log_error(td, io_u);
1670 __io_u_log_error(td->parent, io_u);
1673 static inline bool gtod_reduce(struct thread_data *td)
1675 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1676 || td->o.gtod_reduce;
1679 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1680 struct io_completion_data *icd,
1681 const enum fio_ddir idx, unsigned int bytes)
1683 const int no_reduce = !gtod_reduce(td);
1684 unsigned long lusec = 0;
1690 lusec = utime_since(&io_u->issue_time, &icd->time);
1692 if (!td->o.disable_lat) {
1693 unsigned long tusec;
1695 tusec = utime_since(&io_u->start_time, &icd->time);
1696 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1698 if (td->flags & TD_F_PROFILE_OPS) {
1699 struct prof_io_ops *ops = &td->prof_io_ops;
1702 icd->error = ops->io_u_lat(td, tusec);
1705 if (td->o.max_latency && tusec > td->o.max_latency)
1706 lat_fatal(td, icd, tusec, td->o.max_latency);
1707 if (td->o.latency_target && tusec > td->o.latency_target) {
1708 if (lat_target_failed(td))
1709 lat_fatal(td, icd, tusec, td->o.latency_target);
1713 if (!td->o.disable_clat) {
1714 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1715 io_u_mark_latency(td, lusec);
1718 if (!td->o.disable_bw)
1719 add_bw_sample(td, idx, bytes, &icd->time);
1722 add_iops_sample(td, idx, bytes, &icd->time);
1724 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1725 uint32_t *info = io_u_block_info(td, io_u);
1726 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1727 if (io_u->ddir == DDIR_TRIM) {
1728 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1729 BLOCK_INFO_TRIMS(*info) + 1);
1730 } else if (io_u->ddir == DDIR_WRITE) {
1731 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1738 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1739 struct io_completion_data *icd)
1741 struct io_u *io_u = *io_u_ptr;
1742 enum fio_ddir ddir = io_u->ddir;
1743 struct fio_file *f = io_u->file;
1745 dprint_io_u(io_u, "io complete");
1747 assert(io_u->flags & IO_U_F_FLIGHT);
1748 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1751 * Mark IO ok to verify
1755 * Remove errored entry from the verification list
1758 unlog_io_piece(td, io_u);
1760 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1765 if (ddir_sync(ddir)) {
1766 td->last_was_sync = 1;
1768 f->first_write = -1ULL;
1769 f->last_write = -1ULL;
1774 td->last_was_sync = 0;
1775 td->last_ddir = ddir;
1777 if (!io_u->error && ddir_rw(ddir)) {
1778 unsigned int bytes = io_u->buflen - io_u->resid;
1781 td->io_blocks[ddir]++;
1782 td->this_io_blocks[ddir]++;
1783 td->io_bytes[ddir] += bytes;
1785 if (!(io_u->flags & IO_U_F_VER_LIST))
1786 td->this_io_bytes[ddir] += bytes;
1788 if (ddir == DDIR_WRITE) {
1790 if (f->first_write == -1ULL ||
1791 io_u->offset < f->first_write)
1792 f->first_write = io_u->offset;
1793 if (f->last_write == -1ULL ||
1794 ((io_u->offset + bytes) > f->last_write))
1795 f->last_write = io_u->offset + bytes;
1797 if (td->last_write_comp) {
1798 int idx = td->last_write_idx++;
1800 td->last_write_comp[idx] = io_u->offset;
1801 if (td->last_write_idx == td->o.iodepth)
1802 td->last_write_idx = 0;
1806 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1807 td->runstate == TD_VERIFYING))
1808 account_io_completion(td, io_u, icd, ddir, bytes);
1810 icd->bytes_done[ddir] += bytes;
1813 ret = io_u->end_io(td, io_u_ptr);
1815 if (ret && !icd->error)
1818 } else if (io_u->error) {
1819 icd->error = io_u->error;
1820 io_u_log_error(td, io_u);
1823 enum error_type_bit eb = td_error_type(ddir, icd->error);
1825 if (!td_non_fatal_error(td, eb, icd->error))
1829 * If there is a non_fatal error, then add to the error count
1830 * and clear all the errors.
1832 update_error_count(td, icd->error);
1840 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1845 if (!gtod_reduce(td))
1846 fio_gettime(&icd->time, NULL);
1851 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1852 icd->bytes_done[ddir] = 0;
1855 static void ios_completed(struct thread_data *td,
1856 struct io_completion_data *icd)
1861 for (i = 0; i < icd->nr; i++) {
1862 io_u = td->io_ops->event(td, i);
1864 io_completed(td, &io_u, icd);
1872 * Complete a single io_u for the sync engines.
1874 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1876 struct io_completion_data icd;
1879 init_icd(td, &icd, 1);
1880 io_completed(td, &io_u, &icd);
1886 td_verror(td, icd.error, "io_u_sync_complete");
1890 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1891 td->bytes_done[ddir] += icd.bytes_done[ddir];
1897 * Called to complete min_events number of io for the async engines.
1899 int io_u_queued_complete(struct thread_data *td, int min_evts)
1901 struct io_completion_data icd;
1902 struct timespec *tvp = NULL;
1904 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1906 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1910 else if (min_evts > td->cur_depth)
1911 min_evts = td->cur_depth;
1913 /* No worries, td_io_getevents fixes min and max if they are
1914 * set incorrectly */
1915 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1917 td_verror(td, -ret, "td_io_getevents");
1922 init_icd(td, &icd, ret);
1923 ios_completed(td, &icd);
1925 td_verror(td, icd.error, "io_u_queued_complete");
1929 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1930 td->bytes_done[ddir] += icd.bytes_done[ddir];
1936 * Call when io_u is really queued, to update the submission latency.
1938 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1940 if (!td->o.disable_slat) {
1941 unsigned long slat_time;
1943 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1948 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1954 * See if we should reuse the last seed, if dedupe is enabled
1956 static struct frand_state *get_buf_state(struct thread_data *td)
1960 if (!td->o.dedupe_percentage)
1961 return &td->buf_state;
1962 else if (td->o.dedupe_percentage == 100) {
1963 frand_copy(&td->buf_state_prev, &td->buf_state);
1964 return &td->buf_state;
1967 v = rand32_between(&td->dedupe_state, 1, 100);
1969 if (v <= td->o.dedupe_percentage)
1970 return &td->buf_state_prev;
1972 return &td->buf_state;
1975 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1977 if (td->o.dedupe_percentage == 100)
1978 frand_copy(rs, &td->buf_state_prev);
1979 else if (rs == &td->buf_state)
1980 frand_copy(&td->buf_state_prev, rs);
1983 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1984 unsigned int max_bs)
1986 struct thread_options *o = &td->o;
1988 if (o->compress_percentage || o->dedupe_percentage) {
1989 unsigned int perc = td->o.compress_percentage;
1990 struct frand_state *rs;
1991 unsigned int left = max_bs;
1992 unsigned int this_write;
1995 rs = get_buf_state(td);
1997 min_write = min(min_write, left);
2000 this_write = min_not_zero(min_write,
2001 td->o.compress_chunk);
2003 fill_random_buf_percentage(rs, buf, perc,
2004 this_write, this_write,
2006 o->buffer_pattern_bytes);
2008 fill_random_buf(rs, buf, min_write);
2009 this_write = min_write;
2014 save_buf_state(td, rs);
2016 } else if (o->buffer_pattern_bytes)
2017 fill_buffer_pattern(td, buf, max_bs);
2018 else if (o->zero_buffers)
2019 memset(buf, 0, max_bs);
2021 fill_random_buf(get_buf_state(td), buf, max_bs);
2025 * "randomly" fill the buffer contents
2027 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2028 unsigned int min_write, unsigned int max_bs)
2030 io_u->buf_filled_len = 0;
2031 fill_io_buffer(td, io_u->buf, min_write, max_bs);