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 int 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)
93 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
94 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
95 uint64_t frand_max, lastb;
97 lastb = last_block(td, f, ddir);
101 frand_max = rand_max(&td->random_state);
102 r = __rand(&td->random_state);
104 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
106 *b = lastb * (r / ((uint64_t) frand_max + 1.0));
110 assert(fio_file_lfsr(f));
112 if (lfsr_next(&f->lfsr, &off))
119 * if we are not maintaining a random map, we are done.
121 if (!file_randommap(td, f))
125 * calculate map offset and check if it's free
127 if (random_map_free(f, *b))
130 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
131 (unsigned long long) *b);
133 *b = axmap_next_free(f->io_axmap, *b);
134 if (*b == (uint64_t) -1ULL)
140 static int __get_next_rand_offset_zipf(struct thread_data *td,
141 struct fio_file *f, enum fio_ddir ddir,
144 *b = zipf_next(&f->zipf);
148 static int __get_next_rand_offset_pareto(struct thread_data *td,
149 struct fio_file *f, enum fio_ddir ddir,
152 *b = pareto_next(&f->zipf);
156 static int __get_next_rand_offset_gauss(struct thread_data *td,
157 struct fio_file *f, enum fio_ddir ddir,
160 *b = gauss_next(&f->gauss);
165 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
167 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
168 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
170 return r1->off - r2->off;
173 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
174 enum fio_ddir ddir, uint64_t *b)
176 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
177 return __get_next_rand_offset(td, f, ddir, b);
178 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
179 return __get_next_rand_offset_zipf(td, f, ddir, b);
180 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
181 return __get_next_rand_offset_pareto(td, f, ddir, b);
182 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
183 return __get_next_rand_offset_gauss(td, f, ddir, b);
185 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
190 * Sort the reads for a verify phase in batches of verifysort_nr, if
193 static inline int should_sort_io(struct thread_data *td)
195 if (!td->o.verifysort_nr || !td->o.do_verify)
199 if (td->runstate != TD_VERIFYING)
201 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
202 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
208 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
214 if (td->o.perc_rand[ddir] == 100)
217 frand_max = rand_max(&td->seq_rand_state[ddir]);
218 r = __rand(&td->seq_rand_state[ddir]);
219 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
221 return v <= td->o.perc_rand[ddir];
224 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
225 enum fio_ddir ddir, uint64_t *b)
230 if (!should_sort_io(td))
231 return get_off_from_method(td, f, ddir, b);
233 if (!flist_empty(&td->next_rand_list)) {
235 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
242 for (i = 0; i < td->o.verifysort_nr; i++) {
243 r = malloc(sizeof(*r));
245 ret = get_off_from_method(td, f, ddir, &r->off);
251 flist_add(&r->list, &td->next_rand_list);
257 assert(!flist_empty(&td->next_rand_list));
258 flist_sort(NULL, &td->next_rand_list, flist_cmp);
262 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
263 enum fio_ddir ddir, uint64_t *b)
265 if (!get_next_rand_offset(td, f, ddir, b))
268 if (td->o.time_based) {
269 fio_file_reset(td, f);
270 if (!get_next_rand_offset(td, f, ddir, b))
274 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
275 f->file_name, (unsigned long long) f->last_pos[ddir],
276 (unsigned long long) f->real_file_size);
280 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
281 enum fio_ddir ddir, uint64_t *offset)
283 struct thread_options *o = &td->o;
285 assert(ddir_rw(ddir));
287 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
289 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
291 if (f->last_pos[ddir] < f->real_file_size) {
294 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
295 f->last_pos[ddir] = f->real_file_size;
297 pos = f->last_pos[ddir] - f->file_offset;
298 if (pos && o->ddir_seq_add) {
299 pos += o->ddir_seq_add;
302 * If we reach beyond the end of the file
303 * with holed IO, wrap around to the
306 if (pos >= f->real_file_size)
307 pos = f->file_offset;
317 static int get_next_block(struct thread_data *td, struct io_u *io_u,
318 enum fio_ddir ddir, int rw_seq,
319 unsigned int *is_random)
321 struct fio_file *f = io_u->file;
325 assert(ddir_rw(ddir));
331 if (should_do_random(td, ddir)) {
332 ret = get_next_rand_block(td, f, ddir, &b);
336 io_u_set(io_u, IO_U_F_BUSY_OK);
337 ret = get_next_seq_offset(td, f, ddir, &offset);
339 ret = get_next_rand_block(td, f, ddir, &b);
343 ret = get_next_seq_offset(td, f, ddir, &offset);
346 io_u_set(io_u, IO_U_F_BUSY_OK);
349 if (td->o.rw_seq == RW_SEQ_SEQ) {
350 ret = get_next_seq_offset(td, f, ddir, &offset);
352 ret = get_next_rand_block(td, f, ddir, &b);
355 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
356 if (f->last_start[ddir] != -1ULL)
357 offset = f->last_start[ddir] - f->file_offset;
362 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
369 io_u->offset = offset;
371 io_u->offset = b * td->o.ba[ddir];
373 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
382 * For random io, generate a random new block and see if it's used. Repeat
383 * until we find a free one. For sequential io, just return the end of
384 * the last io issued.
386 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
387 unsigned int *is_random)
389 struct fio_file *f = io_u->file;
390 enum fio_ddir ddir = io_u->ddir;
393 assert(ddir_rw(ddir));
395 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
397 td->ddir_seq_nr = td->o.ddir_seq_nr;
400 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
403 if (io_u->offset >= f->io_size) {
404 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
405 (unsigned long long) io_u->offset,
406 (unsigned long long) f->io_size);
410 io_u->offset += f->file_offset;
411 if (io_u->offset >= f->real_file_size) {
412 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
413 (unsigned long long) io_u->offset,
414 (unsigned long long) f->real_file_size);
421 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
422 unsigned int *is_random)
424 if (td->flags & TD_F_PROFILE_OPS) {
425 struct prof_io_ops *ops = &td->prof_io_ops;
427 if (ops->fill_io_u_off)
428 return ops->fill_io_u_off(td, io_u, is_random);
431 return __get_next_offset(td, io_u, is_random);
434 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
437 struct fio_file *f = io_u->file;
439 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
442 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
443 unsigned int is_random)
445 int ddir = io_u->ddir;
446 unsigned int buflen = 0;
447 unsigned int minbs, maxbs;
451 assert(ddir_rw(ddir));
453 if (td->o.bs_is_seq_rand)
454 ddir = is_random ? DDIR_WRITE: DDIR_READ;
456 minbs = td->o.min_bs[ddir];
457 maxbs = td->o.max_bs[ddir];
463 * If we can't satisfy the min block size from here, then fail
465 if (!io_u_fits(td, io_u, minbs))
468 frand_max = rand_max(&td->bsrange_state);
470 r = __rand(&td->bsrange_state);
472 if (!td->o.bssplit_nr[ddir]) {
473 buflen = 1 + (unsigned int) ((double) maxbs *
474 (r / (frand_max + 1.0)));
481 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
482 struct bssplit *bsp = &td->o.bssplit[ddir][i];
486 if ((r <= ((frand_max / 100L) * perc)) &&
487 io_u_fits(td, io_u, buflen))
492 if (td->o.verify != VERIFY_NONE)
493 buflen = (buflen + td->o.verify_interval - 1) &
494 ~(td->o.verify_interval - 1);
496 if (!td->o.bs_unaligned && is_power_of_2(minbs))
497 buflen &= ~(minbs - 1);
499 } while (!io_u_fits(td, io_u, buflen));
504 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
505 unsigned int is_random)
507 if (td->flags & TD_F_PROFILE_OPS) {
508 struct prof_io_ops *ops = &td->prof_io_ops;
510 if (ops->fill_io_u_size)
511 return ops->fill_io_u_size(td, io_u, is_random);
514 return __get_next_buflen(td, io_u, is_random);
517 static void set_rwmix_bytes(struct thread_data *td)
522 * we do time or byte based switch. this is needed because
523 * buffered writes may issue a lot quicker than they complete,
524 * whereas reads do not.
526 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
527 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
530 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
532 uint64_t frand_max = rand_max(&td->rwmix_state);
536 r = __rand(&td->rwmix_state);
537 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
539 if (v <= td->o.rwmix[DDIR_READ])
545 void io_u_quiesce(struct thread_data *td)
548 * We are going to sleep, ensure that we flush anything pending as
549 * not to skew our latency numbers.
551 * Changed to only monitor 'in flight' requests here instead of the
552 * td->cur_depth, b/c td->cur_depth does not accurately represent
553 * io's that have been actually submitted to an async engine,
554 * and cur_depth is meaningless for sync engines.
556 if (td->io_u_queued || td->cur_depth) {
559 ret = td_io_commit(td);
562 while (td->io_u_in_flight) {
565 ret = io_u_queued_complete(td, 1);
569 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
571 enum fio_ddir odir = ddir ^ 1;
574 assert(ddir_rw(ddir));
575 now = utime_since_now(&td->start);
578 * if rate_next_io_time is in the past, need to catch up to rate
580 if (td->rate_next_io_time[ddir] <= now)
584 * We are ahead of rate in this direction. See if we
587 if (td_rw(td) && td->o.rwmix[odir]) {
589 * Other direction is behind rate, switch
591 if (td->rate_next_io_time[odir] <= now)
595 * Both directions are ahead of rate. sleep the min
596 * switch if necissary
598 if (td->rate_next_io_time[ddir] <=
599 td->rate_next_io_time[odir]) {
600 usec = td->rate_next_io_time[ddir] - now;
602 usec = td->rate_next_io_time[odir] - now;
606 usec = td->rate_next_io_time[ddir] - now;
608 if (td->o.io_submit_mode == IO_MODE_INLINE)
611 usec = usec_sleep(td, usec);
617 * Return the data direction for the next io_u. If the job is a
618 * mixed read/write workload, check the rwmix cycle and switch if
621 static enum fio_ddir get_rw_ddir(struct thread_data *td)
626 * see if it's time to fsync
628 if (td->o.fsync_blocks &&
629 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
630 td->io_issues[DDIR_WRITE] && should_fsync(td))
634 * see if it's time to fdatasync
636 if (td->o.fdatasync_blocks &&
637 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
638 td->io_issues[DDIR_WRITE] && should_fsync(td))
639 return DDIR_DATASYNC;
642 * see if it's time to sync_file_range
644 if (td->sync_file_range_nr &&
645 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
646 td->io_issues[DDIR_WRITE] && should_fsync(td))
647 return DDIR_SYNC_FILE_RANGE;
651 * Check if it's time to seed a new data direction.
653 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
655 * Put a top limit on how many bytes we do for
656 * one data direction, to avoid overflowing the
659 ddir = get_rand_ddir(td);
661 if (ddir != td->rwmix_ddir)
664 td->rwmix_ddir = ddir;
666 ddir = td->rwmix_ddir;
667 } else if (td_read(td))
669 else if (td_write(td))
674 td->rwmix_ddir = rate_ddir(td, ddir);
675 return td->rwmix_ddir;
678 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
680 enum fio_ddir ddir = get_rw_ddir(td);
682 if (td_trimwrite(td)) {
683 struct fio_file *f = io_u->file;
684 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
690 io_u->ddir = io_u->acct_ddir = ddir;
692 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
693 td->o.barrier_blocks &&
694 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
695 td->io_issues[DDIR_WRITE])
696 io_u_set(io_u, IO_U_F_BARRIER);
699 void put_file_log(struct thread_data *td, struct fio_file *f)
701 unsigned int ret = put_file(td, f);
704 td_verror(td, ret, "file close");
707 void put_io_u(struct thread_data *td, struct io_u *io_u)
714 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
715 put_file_log(td, io_u->file);
718 io_u_set(io_u, IO_U_F_FREE);
720 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
722 assert(!(td->flags & TD_F_CHILD));
724 io_u_qpush(&td->io_u_freelist, io_u);
726 td_io_u_free_notify(td);
729 void clear_io_u(struct thread_data *td, struct io_u *io_u)
731 io_u_clear(io_u, IO_U_F_FLIGHT);
735 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
737 struct io_u *__io_u = *io_u;
738 enum fio_ddir ddir = acct_ddir(__io_u);
740 dprint(FD_IO, "requeue %p\n", __io_u);
747 io_u_set(__io_u, IO_U_F_FREE);
748 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
749 td->io_issues[ddir]--;
751 io_u_clear(__io_u, IO_U_F_FLIGHT);
752 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
754 assert(!(td->flags & TD_F_CHILD));
757 io_u_rpush(&td->io_u_requeues, __io_u);
759 td_io_u_free_notify(td);
763 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
765 unsigned int is_random;
767 if (td->io_ops->flags & FIO_NOIO)
770 set_rw_ddir(td, io_u);
773 * fsync() or fdatasync() or trim etc, we are done
775 if (!ddir_rw(io_u->ddir))
779 * See if it's time to switch to a new zone
781 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
782 struct fio_file *f = io_u->file;
785 f->file_offset += td->o.zone_range + td->o.zone_skip;
788 * Wrap from the beginning, if we exceed the file size
790 if (f->file_offset >= f->real_file_size)
791 f->file_offset = f->real_file_size - f->file_offset;
792 f->last_pos[io_u->ddir] = f->file_offset;
793 td->io_skip_bytes += td->o.zone_skip;
797 * No log, let the seq/rand engine retrieve the next buflen and
800 if (get_next_offset(td, io_u, &is_random)) {
801 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
805 io_u->buflen = get_next_buflen(td, io_u, is_random);
807 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
811 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
812 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
813 dprint(FD_IO, " off=%llu/%lu > %llu\n",
814 (unsigned long long) io_u->offset, io_u->buflen,
815 (unsigned long long) io_u->file->real_file_size);
820 * mark entry before potentially trimming io_u
822 if (td_random(td) && file_randommap(td, io_u->file))
823 mark_random_map(td, io_u);
826 dprint_io_u(io_u, "fill_io_u");
827 td->zone_bytes += io_u->buflen;
831 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
860 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
862 __io_u_mark_map(td->ts.io_u_submit, nr);
863 td->ts.total_submit++;
866 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
868 __io_u_mark_map(td->ts.io_u_complete, nr);
869 td->ts.total_complete++;
872 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
876 switch (td->cur_depth) {
898 td->ts.io_u_map[idx] += nr;
901 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
938 assert(idx < FIO_IO_U_LAT_U_NR);
939 td->ts.io_u_lat_u[idx]++;
942 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
983 assert(idx < FIO_IO_U_LAT_M_NR);
984 td->ts.io_u_lat_m[idx]++;
987 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
990 io_u_mark_lat_usec(td, usec);
992 io_u_mark_lat_msec(td, usec / 1000);
996 * Get next file to service by choosing one at random
998 static struct fio_file *get_next_file_rand(struct thread_data *td,
999 enum fio_file_flags goodf,
1000 enum fio_file_flags badf)
1002 uint64_t frand_max = rand_max(&td->next_file_state);
1010 r = __rand(&td->next_file_state);
1011 fno = (unsigned int) ((double) td->o.nr_files
1012 * (r / (frand_max + 1.0)));
1015 if (fio_file_done(f))
1018 if (!fio_file_open(f)) {
1021 if (td->nr_open_files >= td->o.open_files)
1022 return ERR_PTR(-EBUSY);
1024 err = td_io_open_file(td, f);
1030 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1031 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1035 td_io_close_file(td, f);
1040 * Get next file to service by doing round robin between all available ones
1042 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1045 unsigned int old_next_file = td->next_file;
1051 f = td->files[td->next_file];
1054 if (td->next_file >= td->o.nr_files)
1057 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1058 if (fio_file_done(f)) {
1063 if (!fio_file_open(f)) {
1066 if (td->nr_open_files >= td->o.open_files)
1067 return ERR_PTR(-EBUSY);
1069 err = td_io_open_file(td, f);
1071 dprint(FD_FILE, "error %d on open of %s\n",
1079 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1081 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1085 td_io_close_file(td, f);
1088 } while (td->next_file != old_next_file);
1090 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1094 static struct fio_file *__get_next_file(struct thread_data *td)
1098 assert(td->o.nr_files <= td->files_index);
1100 if (td->nr_done_files >= td->o.nr_files) {
1101 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1102 " nr_files=%d\n", td->nr_open_files,
1108 f = td->file_service_file;
1109 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1110 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1112 if (td->file_service_left--)
1116 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1117 td->o.file_service_type == FIO_FSERVICE_SEQ)
1118 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1120 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1125 td->file_service_file = f;
1126 td->file_service_left = td->file_service_nr - 1;
1129 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1131 dprint(FD_FILE, "get_next_file: NULL\n");
1135 static struct fio_file *get_next_file(struct thread_data *td)
1137 if (td->flags & TD_F_PROFILE_OPS) {
1138 struct prof_io_ops *ops = &td->prof_io_ops;
1140 if (ops->get_next_file)
1141 return ops->get_next_file(td);
1144 return __get_next_file(td);
1147 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1152 f = get_next_file(td);
1153 if (IS_ERR_OR_NULL(f))
1159 if (!fill_io_u(td, io_u))
1162 put_file_log(td, f);
1163 td_io_close_file(td, f);
1165 fio_file_set_done(f);
1166 td->nr_done_files++;
1167 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1168 td->nr_done_files, td->o.nr_files);
1174 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1175 unsigned long tusec, unsigned long max_usec)
1178 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1179 td_verror(td, ETIMEDOUT, "max latency exceeded");
1180 icd->error = ETIMEDOUT;
1183 static void lat_new_cycle(struct thread_data *td)
1185 fio_gettime(&td->latency_ts, NULL);
1186 td->latency_ios = ddir_rw_sum(td->io_blocks);
1187 td->latency_failed = 0;
1191 * We had an IO outside the latency target. Reduce the queue depth. If we
1192 * are at QD=1, then it's time to give up.
1194 static int __lat_target_failed(struct thread_data *td)
1196 if (td->latency_qd == 1)
1199 td->latency_qd_high = td->latency_qd;
1201 if (td->latency_qd == td->latency_qd_low)
1202 td->latency_qd_low--;
1204 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1206 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1209 * When we ramp QD down, quiesce existing IO to prevent
1210 * a storm of ramp downs due to pending higher depth.
1217 static int lat_target_failed(struct thread_data *td)
1219 if (td->o.latency_percentile.u.f == 100.0)
1220 return __lat_target_failed(td);
1222 td->latency_failed++;
1226 void lat_target_init(struct thread_data *td)
1228 td->latency_end_run = 0;
1230 if (td->o.latency_target) {
1231 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1232 fio_gettime(&td->latency_ts, NULL);
1234 td->latency_qd_high = td->o.iodepth;
1235 td->latency_qd_low = 1;
1236 td->latency_ios = ddir_rw_sum(td->io_blocks);
1238 td->latency_qd = td->o.iodepth;
1241 void lat_target_reset(struct thread_data *td)
1243 if (!td->latency_end_run)
1244 lat_target_init(td);
1247 static void lat_target_success(struct thread_data *td)
1249 const unsigned int qd = td->latency_qd;
1250 struct thread_options *o = &td->o;
1252 td->latency_qd_low = td->latency_qd;
1255 * If we haven't failed yet, we double up to a failing value instead
1256 * of bisecting from highest possible queue depth. If we have set
1257 * a limit other than td->o.iodepth, bisect between that.
1259 if (td->latency_qd_high != o->iodepth)
1260 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1262 td->latency_qd *= 2;
1264 if (td->latency_qd > o->iodepth)
1265 td->latency_qd = o->iodepth;
1267 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1270 * Same as last one, we are done. Let it run a latency cycle, so
1271 * we get only the results from the targeted depth.
1273 if (td->latency_qd == qd) {
1274 if (td->latency_end_run) {
1275 dprint(FD_RATE, "We are done\n");
1278 dprint(FD_RATE, "Quiesce and final run\n");
1280 td->latency_end_run = 1;
1281 reset_all_stats(td);
1290 * Check if we can bump the queue depth
1292 void lat_target_check(struct thread_data *td)
1294 uint64_t usec_window;
1298 usec_window = utime_since_now(&td->latency_ts);
1299 if (usec_window < td->o.latency_window)
1302 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1303 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1304 success_ios *= 100.0;
1306 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1308 if (success_ios >= td->o.latency_percentile.u.f)
1309 lat_target_success(td);
1311 __lat_target_failed(td);
1315 * If latency target is enabled, we might be ramping up or down and not
1316 * using the full queue depth available.
1318 int queue_full(const struct thread_data *td)
1320 const int qempty = io_u_qempty(&td->io_u_freelist);
1324 if (!td->o.latency_target)
1327 return td->cur_depth >= td->latency_qd;
1330 struct io_u *__get_io_u(struct thread_data *td)
1332 struct io_u *io_u = NULL;
1340 if (!io_u_rempty(&td->io_u_requeues))
1341 io_u = io_u_rpop(&td->io_u_requeues);
1342 else if (!queue_full(td)) {
1343 io_u = io_u_qpop(&td->io_u_freelist);
1348 io_u->end_io = NULL;
1352 assert(io_u->flags & IO_U_F_FREE);
1353 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1354 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1358 io_u->acct_ddir = -1;
1360 assert(!(td->flags & TD_F_CHILD));
1361 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1363 } else if (td_async_processing(td)) {
1365 * We ran out, wait for async verify threads to finish and
1368 assert(!(td->flags & TD_F_CHILD));
1369 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1377 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1379 if (!(td->flags & TD_F_TRIM_BACKLOG))
1382 if (td->trim_entries) {
1385 if (td->trim_batch) {
1388 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1389 td->last_ddir != DDIR_READ) {
1390 td->trim_batch = td->o.trim_batch;
1391 if (!td->trim_batch)
1392 td->trim_batch = td->o.trim_backlog;
1396 if (get_trim && !get_next_trim(td, io_u))
1403 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1405 if (!(td->flags & TD_F_VER_BACKLOG))
1408 if (td->io_hist_len) {
1411 if (td->verify_batch)
1413 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1414 td->last_ddir != DDIR_READ) {
1415 td->verify_batch = td->o.verify_batch;
1416 if (!td->verify_batch)
1417 td->verify_batch = td->o.verify_backlog;
1421 if (get_verify && !get_next_verify(td, io_u)) {
1431 * Fill offset and start time into the buffer content, to prevent too
1432 * easy compressible data for simple de-dupe attempts. Do this for every
1433 * 512b block in the range, since that should be the smallest block size
1434 * we can expect from a device.
1436 static void small_content_scramble(struct io_u *io_u)
1438 unsigned int i, nr_blocks = io_u->buflen / 512;
1440 unsigned int offset;
1447 boffset = io_u->offset;
1448 io_u->buf_filled_len = 0;
1450 for (i = 0; i < nr_blocks; i++) {
1452 * Fill the byte offset into a "random" start offset of
1453 * the buffer, given by the product of the usec time
1454 * and the actual offset.
1456 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1457 offset &= ~(sizeof(uint64_t) - 1);
1458 if (offset >= 512 - sizeof(uint64_t))
1459 offset -= sizeof(uint64_t);
1460 memcpy(p + offset, &boffset, sizeof(boffset));
1462 end = p + 512 - sizeof(io_u->start_time);
1463 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1470 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1471 * etc. The returned io_u is fully ready to be prepped and submitted.
1473 struct io_u *get_io_u(struct thread_data *td)
1477 int do_scramble = 0;
1480 io_u = __get_io_u(td);
1482 dprint(FD_IO, "__get_io_u failed\n");
1486 if (check_get_verify(td, io_u))
1488 if (check_get_trim(td, io_u))
1492 * from a requeue, io_u already setup
1498 * If using an iolog, grab next piece if any available.
1500 if (td->flags & TD_F_READ_IOLOG) {
1501 if (read_iolog_get(td, io_u))
1503 } else if (set_io_u_file(td, io_u)) {
1505 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1511 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1515 assert(fio_file_open(f));
1517 if (ddir_rw(io_u->ddir)) {
1518 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1519 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1523 f->last_start[io_u->ddir] = io_u->offset;
1524 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1526 if (io_u->ddir == DDIR_WRITE) {
1527 if (td->flags & TD_F_REFILL_BUFFERS) {
1528 io_u_fill_buffer(td, io_u,
1529 td->o.min_bs[DDIR_WRITE],
1531 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1532 !(td->flags & TD_F_COMPRESS))
1534 if (td->flags & TD_F_VER_NONE) {
1535 populate_verify_io_u(td, io_u);
1538 } else if (io_u->ddir == DDIR_READ) {
1540 * Reset the buf_filled parameters so next time if the
1541 * buffer is used for writes it is refilled.
1543 io_u->buf_filled_len = 0;
1548 * Set io data pointers.
1550 io_u->xfer_buf = io_u->buf;
1551 io_u->xfer_buflen = io_u->buflen;
1555 if (!td_io_prep(td, io_u)) {
1556 if (!td->o.disable_slat)
1557 fio_gettime(&io_u->start_time, NULL);
1559 small_content_scramble(io_u);
1563 dprint(FD_IO, "get_io_u failed\n");
1565 return ERR_PTR(ret);
1568 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1570 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1572 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1575 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1576 io_u->file ? " on file " : "",
1577 io_u->file ? io_u->file->file_name : "",
1578 strerror(io_u->error),
1579 io_ddir_name(io_u->ddir),
1580 io_u->offset, io_u->xfer_buflen);
1582 if (td->io_ops->errdetails) {
1583 char *err = td->io_ops->errdetails(io_u);
1585 log_err("fio: %s\n", err);
1590 td_verror(td, io_u->error, "io_u error");
1593 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1595 __io_u_log_error(td, io_u);
1597 __io_u_log_error(td, io_u);
1600 static inline int gtod_reduce(struct thread_data *td)
1602 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1603 && td->o.disable_bw;
1606 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1607 struct io_completion_data *icd,
1608 const enum fio_ddir idx, unsigned int bytes)
1610 const int no_reduce = !gtod_reduce(td);
1611 unsigned long lusec = 0;
1617 lusec = utime_since(&io_u->issue_time, &icd->time);
1619 if (!td->o.disable_lat) {
1620 unsigned long tusec;
1622 tusec = utime_since(&io_u->start_time, &icd->time);
1623 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1625 if (td->flags & TD_F_PROFILE_OPS) {
1626 struct prof_io_ops *ops = &td->prof_io_ops;
1629 icd->error = ops->io_u_lat(td, tusec);
1632 if (td->o.max_latency && tusec > td->o.max_latency)
1633 lat_fatal(td, icd, tusec, td->o.max_latency);
1634 if (td->o.latency_target && tusec > td->o.latency_target) {
1635 if (lat_target_failed(td))
1636 lat_fatal(td, icd, tusec, td->o.latency_target);
1640 if (!td->o.disable_clat) {
1641 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1642 io_u_mark_latency(td, lusec);
1645 if (!td->o.disable_bw)
1646 add_bw_sample(td, idx, bytes, &icd->time);
1649 add_iops_sample(td, idx, bytes, &icd->time);
1651 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1652 uint32_t *info = io_u_block_info(td, io_u);
1653 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1654 if (io_u->ddir == DDIR_TRIM) {
1655 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1656 BLOCK_INFO_TRIMS(*info) + 1);
1657 } else if (io_u->ddir == DDIR_WRITE) {
1658 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1665 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1666 struct io_completion_data *icd)
1668 struct io_u *io_u = *io_u_ptr;
1669 enum fio_ddir ddir = io_u->ddir;
1670 struct fio_file *f = io_u->file;
1672 dprint_io_u(io_u, "io complete");
1674 assert(io_u->flags & IO_U_F_FLIGHT);
1675 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1678 * Mark IO ok to verify
1682 * Remove errored entry from the verification list
1685 unlog_io_piece(td, io_u);
1687 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1692 if (ddir_sync(ddir)) {
1693 td->last_was_sync = 1;
1695 f->first_write = -1ULL;
1696 f->last_write = -1ULL;
1701 td->last_was_sync = 0;
1702 td->last_ddir = ddir;
1704 if (!io_u->error && ddir_rw(ddir)) {
1705 unsigned int bytes = io_u->buflen - io_u->resid;
1708 td->io_blocks[ddir]++;
1709 td->this_io_blocks[ddir]++;
1710 td->io_bytes[ddir] += bytes;
1712 if (!(io_u->flags & IO_U_F_VER_LIST))
1713 td->this_io_bytes[ddir] += bytes;
1715 if (ddir == DDIR_WRITE) {
1717 if (f->first_write == -1ULL ||
1718 io_u->offset < f->first_write)
1719 f->first_write = io_u->offset;
1720 if (f->last_write == -1ULL ||
1721 ((io_u->offset + bytes) > f->last_write))
1722 f->last_write = io_u->offset + bytes;
1724 if (td->last_write_comp) {
1725 int idx = td->last_write_idx++;
1727 td->last_write_comp[idx] = io_u->offset;
1728 if (td->last_write_idx == td->o.iodepth)
1729 td->last_write_idx = 0;
1733 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1734 td->runstate == TD_VERIFYING))
1735 account_io_completion(td, io_u, icd, ddir, bytes);
1737 icd->bytes_done[ddir] += bytes;
1740 ret = io_u->end_io(td, io_u_ptr);
1742 if (ret && !icd->error)
1745 } else if (io_u->error) {
1746 icd->error = io_u->error;
1747 io_u_log_error(td, io_u);
1750 enum error_type_bit eb = td_error_type(ddir, icd->error);
1752 if (!td_non_fatal_error(td, eb, icd->error))
1756 * If there is a non_fatal error, then add to the error count
1757 * and clear all the errors.
1759 update_error_count(td, icd->error);
1767 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1772 if (!gtod_reduce(td))
1773 fio_gettime(&icd->time, NULL);
1778 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1779 icd->bytes_done[ddir] = 0;
1782 static void ios_completed(struct thread_data *td,
1783 struct io_completion_data *icd)
1788 for (i = 0; i < icd->nr; i++) {
1789 io_u = td->io_ops->event(td, i);
1791 io_completed(td, &io_u, icd);
1799 * Complete a single io_u for the sync engines.
1801 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1803 struct io_completion_data icd;
1806 init_icd(td, &icd, 1);
1807 io_completed(td, &io_u, &icd);
1813 td_verror(td, icd.error, "io_u_sync_complete");
1817 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1818 td->bytes_done[ddir] += icd.bytes_done[ddir];
1824 * Called to complete min_events number of io for the async engines.
1826 int io_u_queued_complete(struct thread_data *td, int min_evts)
1828 struct io_completion_data icd;
1829 struct timespec *tvp = NULL;
1831 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1833 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1837 else if (min_evts > td->cur_depth)
1838 min_evts = td->cur_depth;
1840 /* No worries, td_io_getevents fixes min and max if they are
1841 * set incorrectly */
1842 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1844 td_verror(td, -ret, "td_io_getevents");
1849 init_icd(td, &icd, ret);
1850 ios_completed(td, &icd);
1852 td_verror(td, icd.error, "io_u_queued_complete");
1856 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1857 td->bytes_done[ddir] += icd.bytes_done[ddir];
1863 * Call when io_u is really queued, to update the submission latency.
1865 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1867 if (!td->o.disable_slat) {
1868 unsigned long slat_time;
1870 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1875 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1881 * See if we should reuse the last seed, if dedupe is enabled
1883 static struct frand_state *get_buf_state(struct thread_data *td)
1889 if (!td->o.dedupe_percentage)
1890 return &td->buf_state;
1891 else if (td->o.dedupe_percentage == 100) {
1892 frand_copy(&td->buf_state_prev, &td->buf_state);
1893 return &td->buf_state;
1896 frand_max = rand_max(&td->dedupe_state);
1897 r = __rand(&td->dedupe_state);
1898 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
1900 if (v <= td->o.dedupe_percentage)
1901 return &td->buf_state_prev;
1903 return &td->buf_state;
1906 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1908 if (td->o.dedupe_percentage == 100)
1909 frand_copy(rs, &td->buf_state_prev);
1910 else if (rs == &td->buf_state)
1911 frand_copy(&td->buf_state_prev, rs);
1914 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1915 unsigned int max_bs)
1917 struct thread_options *o = &td->o;
1919 if (o->compress_percentage || o->dedupe_percentage) {
1920 unsigned int perc = td->o.compress_percentage;
1921 struct frand_state *rs;
1922 unsigned int left = max_bs;
1923 unsigned int this_write;
1926 rs = get_buf_state(td);
1928 min_write = min(min_write, left);
1931 this_write = min_not_zero(min_write,
1932 td->o.compress_chunk);
1934 fill_random_buf_percentage(rs, buf, perc,
1935 this_write, this_write,
1937 o->buffer_pattern_bytes);
1939 fill_random_buf(rs, buf, min_write);
1940 this_write = min_write;
1945 save_buf_state(td, rs);
1947 } else if (o->buffer_pattern_bytes)
1948 fill_buffer_pattern(td, buf, max_bs);
1949 else if (o->zero_buffers)
1950 memset(buf, 0, max_bs);
1952 fill_random_buf(get_buf_state(td), buf, max_bs);
1956 * "randomly" fill the buffer contents
1958 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1959 unsigned int min_write, unsigned int max_bs)
1961 io_u->buf_filled_len = 0;
1962 fill_io_buffer(td, io_u->buf, min_write, max_bs);