13 #include "lib/axmap.h"
16 struct io_completion_data {
19 int error; /* output */
20 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
21 struct timeval time; /* output */
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
28 static int random_map_free(struct fio_file *f, const uint64_t block)
30 return !axmap_isset(f->io_axmap, block);
34 * Mark a given offset as used in the map.
36 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned int nr_blocks;
43 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
46 if (!(io_u->flags & IO_U_F_BUSY_OK))
47 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 if ((nr_blocks * min_bs) < io_u->buflen)
50 io_u->buflen = nr_blocks * min_bs;
53 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 assert(ddir_rw(ddir));
62 * Hmm, should we make sure that ->io_size <= ->real_file_size?
64 max_size = f->io_size;
65 if (max_size > f->real_file_size)
66 max_size = f->real_file_size;
69 max_size = td->o.zone_range;
71 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
79 struct flist_head list;
83 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84 enum fio_ddir ddir, uint64_t *b)
88 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
91 lastb = last_block(td, f, ddir);
95 r = __rand(&td->random_state);
97 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
99 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
103 assert(fio_file_lfsr(f));
105 if (lfsr_next(&f->lfsr, &off))
112 * if we are not maintaining a random map, we are done.
114 if (!file_randommap(td, f))
118 * calculate map offset and check if it's free
120 if (random_map_free(f, *b))
123 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
124 (unsigned long long) *b);
126 *b = axmap_next_free(f->io_axmap, *b);
127 if (*b == (uint64_t) -1ULL)
133 static int __get_next_rand_offset_zipf(struct thread_data *td,
134 struct fio_file *f, enum fio_ddir ddir,
137 *b = zipf_next(&f->zipf);
141 static int __get_next_rand_offset_pareto(struct thread_data *td,
142 struct fio_file *f, enum fio_ddir ddir,
145 *b = pareto_next(&f->zipf);
149 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
151 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
152 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
154 return r1->off - r2->off;
157 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
158 enum fio_ddir ddir, uint64_t *b)
160 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
161 return __get_next_rand_offset(td, f, ddir, b);
162 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
163 return __get_next_rand_offset_zipf(td, f, ddir, b);
164 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
165 return __get_next_rand_offset_pareto(td, f, ddir, b);
167 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
172 * Sort the reads for a verify phase in batches of verifysort_nr, if
175 static inline int should_sort_io(struct thread_data *td)
177 if (!td->o.verifysort_nr || !td->o.do_verify)
181 if (td->runstate != TD_VERIFYING)
183 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
189 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
194 if (td->o.perc_rand[ddir] == 100)
197 r = __rand(&td->seq_rand_state[ddir]);
198 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
200 return v <= td->o.perc_rand[ddir];
203 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
204 enum fio_ddir ddir, uint64_t *b)
209 if (!should_sort_io(td))
210 return get_off_from_method(td, f, ddir, b);
212 if (!flist_empty(&td->next_rand_list)) {
214 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
221 for (i = 0; i < td->o.verifysort_nr; i++) {
222 r = malloc(sizeof(*r));
224 ret = get_off_from_method(td, f, ddir, &r->off);
230 flist_add(&r->list, &td->next_rand_list);
236 assert(!flist_empty(&td->next_rand_list));
237 flist_sort(NULL, &td->next_rand_list, flist_cmp);
241 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
242 enum fio_ddir ddir, uint64_t *b)
244 if (!get_next_rand_offset(td, f, ddir, b))
247 if (td->o.time_based) {
248 fio_file_reset(td, f);
249 if (!get_next_rand_offset(td, f, ddir, b))
253 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
254 f->file_name, (unsigned long long) f->last_pos[ddir],
255 (unsigned long long) f->real_file_size);
259 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
260 enum fio_ddir ddir, uint64_t *offset)
262 struct thread_options *o = &td->o;
264 assert(ddir_rw(ddir));
266 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
268 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
270 if (f->last_pos[ddir] < f->real_file_size) {
273 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
274 f->last_pos[ddir] = f->real_file_size;
276 pos = f->last_pos[ddir] - f->file_offset;
277 if (pos && o->ddir_seq_add) {
278 pos += o->ddir_seq_add;
281 * If we reach beyond the end of the file
282 * with holed IO, wrap around to the
285 if (pos >= f->real_file_size)
286 pos = f->file_offset;
296 static int get_next_block(struct thread_data *td, struct io_u *io_u,
297 enum fio_ddir ddir, int rw_seq,
298 unsigned int *is_random)
300 struct fio_file *f = io_u->file;
304 assert(ddir_rw(ddir));
310 if (should_do_random(td, ddir)) {
311 ret = get_next_rand_block(td, f, ddir, &b);
315 io_u->flags |= IO_U_F_BUSY_OK;
316 ret = get_next_seq_offset(td, f, ddir, &offset);
318 ret = get_next_rand_block(td, f, ddir, &b);
322 ret = get_next_seq_offset(td, f, ddir, &offset);
325 io_u->flags |= IO_U_F_BUSY_OK;
328 if (td->o.rw_seq == RW_SEQ_SEQ) {
329 ret = get_next_seq_offset(td, f, ddir, &offset);
331 ret = get_next_rand_block(td, f, ddir, &b);
334 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
335 if (f->last_start[ddir] != -1ULL)
336 offset = f->last_start[ddir] - f->file_offset;
341 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
348 io_u->offset = offset;
350 io_u->offset = b * td->o.ba[ddir];
352 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
361 * For random io, generate a random new block and see if it's used. Repeat
362 * until we find a free one. For sequential io, just return the end of
363 * the last io issued.
365 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
366 unsigned int *is_random)
368 struct fio_file *f = io_u->file;
369 enum fio_ddir ddir = io_u->ddir;
372 assert(ddir_rw(ddir));
374 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
376 td->ddir_seq_nr = td->o.ddir_seq_nr;
379 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
382 if (io_u->offset >= f->io_size) {
383 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
384 (unsigned long long) io_u->offset,
385 (unsigned long long) f->io_size);
389 io_u->offset += f->file_offset;
390 if (io_u->offset >= f->real_file_size) {
391 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
392 (unsigned long long) io_u->offset,
393 (unsigned long long) f->real_file_size);
400 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
401 unsigned int *is_random)
403 if (td->flags & TD_F_PROFILE_OPS) {
404 struct prof_io_ops *ops = &td->prof_io_ops;
406 if (ops->fill_io_u_off)
407 return ops->fill_io_u_off(td, io_u, is_random);
410 return __get_next_offset(td, io_u, is_random);
413 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
416 struct fio_file *f = io_u->file;
418 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
421 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
422 unsigned int is_random)
424 int ddir = io_u->ddir;
425 unsigned int buflen = 0;
426 unsigned int minbs, maxbs;
429 assert(ddir_rw(ddir));
431 if (td->o.bs_is_seq_rand)
432 ddir = is_random ? DDIR_WRITE: DDIR_READ;
434 minbs = td->o.min_bs[ddir];
435 maxbs = td->o.max_bs[ddir];
441 * If we can't satisfy the min block size from here, then fail
443 if (!io_u_fits(td, io_u, minbs))
447 r = __rand(&td->bsrange_state);
449 if (!td->o.bssplit_nr[ddir]) {
450 buflen = 1 + (unsigned int) ((double) maxbs *
451 (r / (FRAND_MAX + 1.0)));
458 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
459 struct bssplit *bsp = &td->o.bssplit[ddir][i];
463 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
464 io_u_fits(td, io_u, buflen))
469 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
470 buflen = (buflen + td->o.verify_interval - 1) &
471 ~(td->o.verify_interval - 1);
473 if (!td->o.bs_unaligned && is_power_of_2(minbs))
474 buflen = (buflen + minbs - 1) & ~(minbs - 1);
476 } while (!io_u_fits(td, io_u, buflen));
481 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
482 unsigned int is_random)
484 if (td->flags & TD_F_PROFILE_OPS) {
485 struct prof_io_ops *ops = &td->prof_io_ops;
487 if (ops->fill_io_u_size)
488 return ops->fill_io_u_size(td, io_u, is_random);
491 return __get_next_buflen(td, io_u, is_random);
494 static void set_rwmix_bytes(struct thread_data *td)
499 * we do time or byte based switch. this is needed because
500 * buffered writes may issue a lot quicker than they complete,
501 * whereas reads do not.
503 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
504 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
507 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
512 r = __rand(&td->rwmix_state);
513 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
515 if (v <= td->o.rwmix[DDIR_READ])
521 void io_u_quiesce(struct thread_data *td)
524 * We are going to sleep, ensure that we flush anything pending as
525 * not to skew our latency numbers.
527 * Changed to only monitor 'in flight' requests here instead of the
528 * td->cur_depth, b/c td->cur_depth does not accurately represent
529 * io's that have been actually submitted to an async engine,
530 * and cur_depth is meaningless for sync engines.
532 while (td->io_u_in_flight) {
535 ret = io_u_queued_complete(td, 1, NULL);
539 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
541 enum fio_ddir odir = ddir ^ 1;
545 assert(ddir_rw(ddir));
547 if (td->rate_pending_usleep[ddir] <= 0)
551 * We have too much pending sleep in this direction. See if we
554 if (td_rw(td) && td->o.rwmix[odir]) {
556 * Other direction does not have too much pending, switch
558 if (td->rate_pending_usleep[odir] < 100000)
562 * Both directions have pending sleep. Sleep the minimum time
563 * and deduct from both.
565 if (td->rate_pending_usleep[ddir] <=
566 td->rate_pending_usleep[odir]) {
567 usec = td->rate_pending_usleep[ddir];
569 usec = td->rate_pending_usleep[odir];
573 usec = td->rate_pending_usleep[ddir];
577 fio_gettime(&t, NULL);
578 usec_sleep(td, usec);
579 usec = utime_since_now(&t);
581 td->rate_pending_usleep[ddir] -= usec;
584 if (td_rw(td) && __should_check_rate(td, odir))
585 td->rate_pending_usleep[odir] -= usec;
587 if (ddir == DDIR_TRIM)
594 * Return the data direction for the next io_u. If the job is a
595 * mixed read/write workload, check the rwmix cycle and switch if
598 static enum fio_ddir get_rw_ddir(struct thread_data *td)
603 * see if it's time to fsync
605 if (td->o.fsync_blocks &&
606 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
607 td->io_issues[DDIR_WRITE] && should_fsync(td))
611 * see if it's time to fdatasync
613 if (td->o.fdatasync_blocks &&
614 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
615 td->io_issues[DDIR_WRITE] && should_fsync(td))
616 return DDIR_DATASYNC;
619 * see if it's time to sync_file_range
621 if (td->sync_file_range_nr &&
622 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
623 td->io_issues[DDIR_WRITE] && should_fsync(td))
624 return DDIR_SYNC_FILE_RANGE;
628 * Check if it's time to seed a new data direction.
630 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
632 * Put a top limit on how many bytes we do for
633 * one data direction, to avoid overflowing the
636 ddir = get_rand_ddir(td);
638 if (ddir != td->rwmix_ddir)
641 td->rwmix_ddir = ddir;
643 ddir = td->rwmix_ddir;
644 } else if (td_read(td))
646 else if (td_write(td))
651 td->rwmix_ddir = rate_ddir(td, ddir);
652 return td->rwmix_ddir;
655 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
657 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
659 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
660 td->o.barrier_blocks &&
661 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
662 td->io_issues[DDIR_WRITE])
663 io_u->flags |= IO_U_F_BARRIER;
666 void put_file_log(struct thread_data *td, struct fio_file *f)
668 unsigned int ret = put_file(td, f);
671 td_verror(td, ret, "file close");
674 void put_io_u(struct thread_data *td, struct io_u *io_u)
678 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
679 put_file_log(td, io_u->file);
682 io_u->flags |= IO_U_F_FREE;
684 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
686 io_u_qpush(&td->io_u_freelist, io_u);
688 td_io_u_free_notify(td);
691 void clear_io_u(struct thread_data *td, struct io_u *io_u)
693 io_u->flags &= ~IO_U_F_FLIGHT;
697 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
699 struct io_u *__io_u = *io_u;
700 enum fio_ddir ddir = acct_ddir(__io_u);
702 dprint(FD_IO, "requeue %p\n", __io_u);
706 __io_u->flags |= IO_U_F_FREE;
707 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
708 td->io_issues[ddir]--;
710 __io_u->flags &= ~IO_U_F_FLIGHT;
711 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
714 io_u_rpush(&td->io_u_requeues, __io_u);
719 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
721 unsigned int is_random;
723 if (td->io_ops->flags & FIO_NOIO)
726 set_rw_ddir(td, io_u);
729 * fsync() or fdatasync() or trim etc, we are done
731 if (!ddir_rw(io_u->ddir))
735 * See if it's time to switch to a new zone
737 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
738 struct fio_file *f = io_u->file;
741 f->file_offset += td->o.zone_range + td->o.zone_skip;
744 * Wrap from the beginning, if we exceed the file size
746 if (f->file_offset >= f->real_file_size)
747 f->file_offset = f->real_file_size - f->file_offset;
748 f->last_pos[io_u->ddir] = f->file_offset;
749 td->io_skip_bytes += td->o.zone_skip;
753 * No log, let the seq/rand engine retrieve the next buflen and
756 if (get_next_offset(td, io_u, &is_random)) {
757 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
761 io_u->buflen = get_next_buflen(td, io_u, is_random);
763 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
767 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
768 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
769 dprint(FD_IO, " off=%llu/%lu > %llu\n",
770 (unsigned long long) io_u->offset, io_u->buflen,
771 (unsigned long long) io_u->file->real_file_size);
776 * mark entry before potentially trimming io_u
778 if (td_random(td) && file_randommap(td, io_u->file))
779 mark_random_map(td, io_u);
782 dprint_io_u(io_u, "fill_io_u");
783 td->zone_bytes += io_u->buflen;
787 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
816 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
818 __io_u_mark_map(td->ts.io_u_submit, nr);
819 td->ts.total_submit++;
822 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
824 __io_u_mark_map(td->ts.io_u_complete, nr);
825 td->ts.total_complete++;
828 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
832 switch (td->cur_depth) {
854 td->ts.io_u_map[idx] += nr;
857 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
894 assert(idx < FIO_IO_U_LAT_U_NR);
895 td->ts.io_u_lat_u[idx]++;
898 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
939 assert(idx < FIO_IO_U_LAT_M_NR);
940 td->ts.io_u_lat_m[idx]++;
943 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
946 io_u_mark_lat_usec(td, usec);
948 io_u_mark_lat_msec(td, usec / 1000);
952 * Get next file to service by choosing one at random
954 static struct fio_file *get_next_file_rand(struct thread_data *td,
955 enum fio_file_flags goodf,
956 enum fio_file_flags badf)
965 r = __rand(&td->next_file_state);
966 fno = (unsigned int) ((double) td->o.nr_files
967 * (r / (FRAND_MAX + 1.0)));
970 if (fio_file_done(f))
973 if (!fio_file_open(f)) {
976 if (td->nr_open_files >= td->o.open_files)
977 return ERR_PTR(-EBUSY);
979 err = td_io_open_file(td, f);
985 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
986 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
990 td_io_close_file(td, f);
995 * Get next file to service by doing round robin between all available ones
997 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1000 unsigned int old_next_file = td->next_file;
1006 f = td->files[td->next_file];
1009 if (td->next_file >= td->o.nr_files)
1012 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1013 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);
1026 dprint(FD_FILE, "error %d on open of %s\n",
1034 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1036 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1040 td_io_close_file(td, f);
1043 } while (td->next_file != old_next_file);
1045 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1049 static struct fio_file *__get_next_file(struct thread_data *td)
1053 assert(td->o.nr_files <= td->files_index);
1055 if (td->nr_done_files >= td->o.nr_files) {
1056 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1057 " nr_files=%d\n", td->nr_open_files,
1063 f = td->file_service_file;
1064 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1065 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1067 if (td->file_service_left--)
1071 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1072 td->o.file_service_type == FIO_FSERVICE_SEQ)
1073 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1075 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1080 td->file_service_file = f;
1081 td->file_service_left = td->file_service_nr - 1;
1084 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1086 dprint(FD_FILE, "get_next_file: NULL\n");
1090 static struct fio_file *get_next_file(struct thread_data *td)
1092 if (td->flags & TD_F_PROFILE_OPS) {
1093 struct prof_io_ops *ops = &td->prof_io_ops;
1095 if (ops->get_next_file)
1096 return ops->get_next_file(td);
1099 return __get_next_file(td);
1102 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1107 f = get_next_file(td);
1108 if (IS_ERR_OR_NULL(f))
1114 if (!fill_io_u(td, io_u))
1117 put_file_log(td, f);
1118 td_io_close_file(td, f);
1120 fio_file_set_done(f);
1121 td->nr_done_files++;
1122 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1123 td->nr_done_files, td->o.nr_files);
1129 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1130 unsigned long tusec, unsigned long max_usec)
1133 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1134 td_verror(td, ETIMEDOUT, "max latency exceeded");
1135 icd->error = ETIMEDOUT;
1138 static void lat_new_cycle(struct thread_data *td)
1140 fio_gettime(&td->latency_ts, NULL);
1141 td->latency_ios = ddir_rw_sum(td->io_blocks);
1142 td->latency_failed = 0;
1146 * We had an IO outside the latency target. Reduce the queue depth. If we
1147 * are at QD=1, then it's time to give up.
1149 static int __lat_target_failed(struct thread_data *td)
1151 if (td->latency_qd == 1)
1154 td->latency_qd_high = td->latency_qd;
1156 if (td->latency_qd == td->latency_qd_low)
1157 td->latency_qd_low--;
1159 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1161 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1164 * When we ramp QD down, quiesce existing IO to prevent
1165 * a storm of ramp downs due to pending higher depth.
1172 static int lat_target_failed(struct thread_data *td)
1174 if (td->o.latency_percentile.u.f == 100.0)
1175 return __lat_target_failed(td);
1177 td->latency_failed++;
1181 void lat_target_init(struct thread_data *td)
1183 td->latency_end_run = 0;
1185 if (td->o.latency_target) {
1186 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1187 fio_gettime(&td->latency_ts, NULL);
1189 td->latency_qd_high = td->o.iodepth;
1190 td->latency_qd_low = 1;
1191 td->latency_ios = ddir_rw_sum(td->io_blocks);
1193 td->latency_qd = td->o.iodepth;
1196 void lat_target_reset(struct thread_data *td)
1198 if (!td->latency_end_run)
1199 lat_target_init(td);
1202 static void lat_target_success(struct thread_data *td)
1204 const unsigned int qd = td->latency_qd;
1205 struct thread_options *o = &td->o;
1207 td->latency_qd_low = td->latency_qd;
1210 * If we haven't failed yet, we double up to a failing value instead
1211 * of bisecting from highest possible queue depth. If we have set
1212 * a limit other than td->o.iodepth, bisect between that.
1214 if (td->latency_qd_high != o->iodepth)
1215 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1217 td->latency_qd *= 2;
1219 if (td->latency_qd > o->iodepth)
1220 td->latency_qd = o->iodepth;
1222 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1225 * Same as last one, we are done. Let it run a latency cycle, so
1226 * we get only the results from the targeted depth.
1228 if (td->latency_qd == qd) {
1229 if (td->latency_end_run) {
1230 dprint(FD_RATE, "We are done\n");
1233 dprint(FD_RATE, "Quiesce and final run\n");
1235 td->latency_end_run = 1;
1236 reset_all_stats(td);
1245 * Check if we can bump the queue depth
1247 void lat_target_check(struct thread_data *td)
1249 uint64_t usec_window;
1253 usec_window = utime_since_now(&td->latency_ts);
1254 if (usec_window < td->o.latency_window)
1257 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1258 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1259 success_ios *= 100.0;
1261 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1263 if (success_ios >= td->o.latency_percentile.u.f)
1264 lat_target_success(td);
1266 __lat_target_failed(td);
1270 * If latency target is enabled, we might be ramping up or down and not
1271 * using the full queue depth available.
1273 int queue_full(const struct thread_data *td)
1275 const int qempty = io_u_qempty(&td->io_u_freelist);
1279 if (!td->o.latency_target)
1282 return td->cur_depth >= td->latency_qd;
1285 struct io_u *__get_io_u(struct thread_data *td)
1287 struct io_u *io_u = NULL;
1295 if (!io_u_rempty(&td->io_u_requeues))
1296 io_u = io_u_rpop(&td->io_u_requeues);
1297 else if (!queue_full(td)) {
1298 io_u = io_u_qpop(&td->io_u_freelist);
1303 io_u->end_io = NULL;
1307 assert(io_u->flags & IO_U_F_FREE);
1308 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1309 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1313 io_u->acct_ddir = -1;
1315 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1317 } else if (td->o.verify_async) {
1319 * We ran out, wait for async verify threads to finish and
1322 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1330 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1332 if (!(td->flags & TD_F_TRIM_BACKLOG))
1335 if (td->trim_entries) {
1338 if (td->trim_batch) {
1341 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1342 td->last_ddir != DDIR_READ) {
1343 td->trim_batch = td->o.trim_batch;
1344 if (!td->trim_batch)
1345 td->trim_batch = td->o.trim_backlog;
1349 if (get_trim && !get_next_trim(td, io_u))
1356 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1358 if (!(td->flags & TD_F_VER_BACKLOG))
1361 if (td->io_hist_len) {
1364 if (td->verify_batch)
1366 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1367 td->last_ddir != DDIR_READ) {
1368 td->verify_batch = td->o.verify_batch;
1369 if (!td->verify_batch)
1370 td->verify_batch = td->o.verify_backlog;
1374 if (get_verify && !get_next_verify(td, io_u)) {
1384 * Fill offset and start time into the buffer content, to prevent too
1385 * easy compressible data for simple de-dupe attempts. Do this for every
1386 * 512b block in the range, since that should be the smallest block size
1387 * we can expect from a device.
1389 static void small_content_scramble(struct io_u *io_u)
1391 unsigned int i, nr_blocks = io_u->buflen / 512;
1393 unsigned int offset;
1400 boffset = io_u->offset;
1401 io_u->buf_filled_len = 0;
1403 for (i = 0; i < nr_blocks; i++) {
1405 * Fill the byte offset into a "random" start offset of
1406 * the buffer, given by the product of the usec time
1407 * and the actual offset.
1409 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1410 offset &= ~(sizeof(uint64_t) - 1);
1411 if (offset >= 512 - sizeof(uint64_t))
1412 offset -= sizeof(uint64_t);
1413 memcpy(p + offset, &boffset, sizeof(boffset));
1415 end = p + 512 - sizeof(io_u->start_time);
1416 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1423 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1424 * etc. The returned io_u is fully ready to be prepped and submitted.
1426 struct io_u *get_io_u(struct thread_data *td)
1430 int do_scramble = 0;
1433 io_u = __get_io_u(td);
1435 dprint(FD_IO, "__get_io_u failed\n");
1439 if (check_get_verify(td, io_u))
1441 if (check_get_trim(td, io_u))
1445 * from a requeue, io_u already setup
1451 * If using an iolog, grab next piece if any available.
1453 if (td->flags & TD_F_READ_IOLOG) {
1454 if (read_iolog_get(td, io_u))
1456 } else if (set_io_u_file(td, io_u)) {
1458 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1464 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1468 assert(fio_file_open(f));
1470 if (ddir_rw(io_u->ddir)) {
1471 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1472 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1476 f->last_start[io_u->ddir] = io_u->offset;
1477 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1479 if (io_u->ddir == DDIR_WRITE) {
1480 if (td->flags & TD_F_REFILL_BUFFERS) {
1481 io_u_fill_buffer(td, io_u,
1482 td->o.min_bs[DDIR_WRITE],
1484 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1485 !(td->flags & TD_F_COMPRESS))
1487 if (td->flags & TD_F_VER_NONE) {
1488 populate_verify_io_u(td, io_u);
1491 } else if (io_u->ddir == DDIR_READ) {
1493 * Reset the buf_filled parameters so next time if the
1494 * buffer is used for writes it is refilled.
1496 io_u->buf_filled_len = 0;
1501 * Set io data pointers.
1503 io_u->xfer_buf = io_u->buf;
1504 io_u->xfer_buflen = io_u->buflen;
1508 if (!td_io_prep(td, io_u)) {
1509 if (!td->o.disable_slat)
1510 fio_gettime(&io_u->start_time, NULL);
1512 small_content_scramble(io_u);
1516 dprint(FD_IO, "get_io_u failed\n");
1518 return ERR_PTR(ret);
1521 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1523 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1525 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1528 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1529 io_u->file ? " on file " : "",
1530 io_u->file ? io_u->file->file_name : "",
1531 strerror(io_u->error),
1532 io_ddir_name(io_u->ddir),
1533 io_u->offset, io_u->xfer_buflen);
1536 td_verror(td, io_u->error, "io_u error");
1539 static inline int gtod_reduce(struct thread_data *td)
1541 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1542 && td->o.disable_bw;
1545 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1546 struct io_completion_data *icd,
1547 const enum fio_ddir idx, unsigned int bytes)
1549 unsigned long lusec = 0;
1551 if (!gtod_reduce(td))
1552 lusec = utime_since(&io_u->issue_time, &icd->time);
1554 if (!td->o.disable_lat) {
1555 unsigned long tusec;
1557 tusec = utime_since(&io_u->start_time, &icd->time);
1558 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1560 if (td->flags & TD_F_PROFILE_OPS) {
1561 struct prof_io_ops *ops = &td->prof_io_ops;
1564 icd->error = ops->io_u_lat(td, tusec);
1567 if (td->o.max_latency && tusec > td->o.max_latency)
1568 lat_fatal(td, icd, tusec, td->o.max_latency);
1569 if (td->o.latency_target && tusec > td->o.latency_target) {
1570 if (lat_target_failed(td))
1571 lat_fatal(td, icd, tusec, td->o.latency_target);
1575 if (!td->o.disable_clat) {
1576 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1577 io_u_mark_latency(td, lusec);
1580 if (!td->o.disable_bw)
1581 add_bw_sample(td, idx, bytes, &icd->time);
1583 if (!gtod_reduce(td))
1584 add_iops_sample(td, idx, bytes, &icd->time);
1587 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1589 uint64_t secs, remainder, bps, bytes;
1591 bytes = td->this_io_bytes[ddir];
1592 bps = td->rate_bps[ddir];
1594 remainder = bytes % bps;
1595 return remainder * 1000000 / bps + secs * 1000000;
1598 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1599 struct io_completion_data *icd)
1601 struct io_u *io_u = *io_u_ptr;
1602 enum fio_ddir ddir = io_u->ddir;
1603 struct fio_file *f = io_u->file;
1605 dprint_io_u(io_u, "io complete");
1608 assert(io_u->flags & IO_U_F_FLIGHT);
1609 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1612 * Mark IO ok to verify
1616 * Remove errored entry from the verification list
1619 unlog_io_piece(td, io_u);
1621 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1628 if (ddir_sync(ddir)) {
1629 td->last_was_sync = 1;
1631 f->first_write = -1ULL;
1632 f->last_write = -1ULL;
1637 td->last_was_sync = 0;
1638 td->last_ddir = ddir;
1640 if (!io_u->error && ddir_rw(ddir)) {
1641 unsigned int bytes = io_u->buflen - io_u->resid;
1642 const enum fio_ddir oddir = ddir ^ 1;
1645 td->io_blocks[ddir]++;
1646 td->this_io_blocks[ddir]++;
1647 td->io_bytes[ddir] += bytes;
1649 if (!(io_u->flags & IO_U_F_VER_LIST))
1650 td->this_io_bytes[ddir] += bytes;
1652 if (ddir == DDIR_WRITE) {
1654 if (f->first_write == -1ULL ||
1655 io_u->offset < f->first_write)
1656 f->first_write = io_u->offset;
1657 if (f->last_write == -1ULL ||
1658 ((io_u->offset + bytes) > f->last_write))
1659 f->last_write = io_u->offset + bytes;
1661 if (td->last_write_comp) {
1662 int idx = td->last_write_idx++;
1664 td->last_write_comp[idx] = io_u->offset;
1665 if (td->last_write_idx == td->o.iodepth)
1666 td->last_write_idx = 0;
1670 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1671 td->runstate == TD_VERIFYING)) {
1672 account_io_completion(td, io_u, icd, ddir, bytes);
1674 if (__should_check_rate(td, ddir)) {
1675 td->rate_pending_usleep[ddir] =
1676 (usec_for_io(td, ddir) -
1677 utime_since_now(&td->start));
1679 if (ddir != DDIR_TRIM &&
1680 __should_check_rate(td, oddir)) {
1681 td->rate_pending_usleep[oddir] =
1682 (usec_for_io(td, oddir) -
1683 utime_since_now(&td->start));
1687 icd->bytes_done[ddir] += bytes;
1690 ret = io_u->end_io(td, io_u_ptr);
1692 if (ret && !icd->error)
1695 } else if (io_u->error) {
1696 icd->error = io_u->error;
1697 io_u_log_error(td, io_u);
1700 enum error_type_bit eb = td_error_type(ddir, icd->error);
1702 if (!td_non_fatal_error(td, eb, icd->error))
1706 * If there is a non_fatal error, then add to the error count
1707 * and clear all the errors.
1709 update_error_count(td, icd->error);
1717 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1722 if (!gtod_reduce(td))
1723 fio_gettime(&icd->time, NULL);
1728 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1729 icd->bytes_done[ddir] = 0;
1732 static void ios_completed(struct thread_data *td,
1733 struct io_completion_data *icd)
1738 for (i = 0; i < icd->nr; i++) {
1739 io_u = td->io_ops->event(td, i);
1741 io_completed(td, &io_u, icd);
1749 * Complete a single io_u for the sync engines.
1751 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1754 struct io_completion_data icd;
1756 init_icd(td, &icd, 1);
1757 io_completed(td, &io_u, &icd);
1763 td_verror(td, icd.error, "io_u_sync_complete");
1770 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1771 bytes[ddir] += icd.bytes_done[ddir];
1778 * Called to complete min_events number of io for the async engines.
1780 int io_u_queued_complete(struct thread_data *td, int min_evts,
1783 struct io_completion_data icd;
1784 struct timespec *tvp = NULL;
1786 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1788 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1792 else if (min_evts > td->cur_depth)
1793 min_evts = td->cur_depth;
1795 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1797 td_verror(td, -ret, "td_io_getevents");
1802 init_icd(td, &icd, ret);
1803 ios_completed(td, &icd);
1805 td_verror(td, icd.error, "io_u_queued_complete");
1812 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1813 bytes[ddir] += icd.bytes_done[ddir];
1820 * Call when io_u is really queued, to update the submission latency.
1822 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1824 if (!td->o.disable_slat) {
1825 unsigned long slat_time;
1827 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1828 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1834 * See if we should reuse the last seed, if dedupe is enabled
1836 static struct frand_state *get_buf_state(struct thread_data *td)
1841 if (!td->o.dedupe_percentage)
1842 return &td->buf_state;
1843 else if (td->o.dedupe_percentage == 100)
1844 return &td->buf_state_prev;
1846 r = __rand(&td->dedupe_state);
1847 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1849 if (v <= td->o.dedupe_percentage)
1850 return &td->buf_state_prev;
1852 return &td->buf_state;
1855 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1857 if (rs == &td->buf_state)
1858 frand_copy(&td->buf_state_prev, rs);
1861 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1862 unsigned int max_bs)
1864 struct thread_options *o = &td->o;
1866 if (o->compress_percentage) {
1867 unsigned int perc = td->o.compress_percentage;
1868 struct frand_state *rs;
1869 unsigned int left = max_bs;
1872 rs = get_buf_state(td);
1874 min_write = min(min_write, left);
1877 unsigned int seg = min_write;
1879 seg = min(min_write, td->o.compress_chunk);
1883 fill_random_buf_percentage(rs, buf, perc, seg,
1884 min_write, o->buffer_pattern,
1885 o->buffer_pattern_bytes);
1887 fill_random_buf(rs, buf, min_write);
1891 save_buf_state(td, rs);
1893 } else if (o->buffer_pattern_bytes)
1894 fill_buffer_pattern(td, buf, max_bs);
1896 memset(buf, 0, max_bs);
1900 * "randomly" fill the buffer contents
1902 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1903 unsigned int min_write, unsigned int max_bs)
1905 io_u->buf_filled_len = 0;
1906 fill_io_buffer(td, io_u->buf, min_write, max_bs);