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 if (td->o.min_bs[ddir] > td->o.ba[ddir])
72 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
74 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
82 struct flist_head list;
86 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir, uint64_t *b)
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
94 lastb = last_block(td, f, ddir);
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
106 assert(fio_file_lfsr(f));
108 if (lfsr_next(&f->lfsr, &off))
115 * if we are not maintaining a random map, we are done.
117 if (!file_randommap(td, f))
121 * calculate map offset and check if it's free
123 if (random_map_free(f, *b))
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
136 static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
140 *b = zipf_next(&f->zipf);
144 static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
148 *b = pareto_next(&f->zipf);
152 static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
156 *b = gauss_next(&f->gauss);
161 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
163 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
164 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
166 return r1->off - r2->off;
169 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
170 enum fio_ddir ddir, uint64_t *b)
172 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
173 return __get_next_rand_offset(td, f, ddir, b);
174 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
175 return __get_next_rand_offset_zipf(td, f, ddir, b);
176 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
177 return __get_next_rand_offset_pareto(td, f, ddir, b);
178 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
179 return __get_next_rand_offset_gauss(td, f, ddir, b);
181 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
186 * Sort the reads for a verify phase in batches of verifysort_nr, if
189 static inline int should_sort_io(struct thread_data *td)
191 if (!td->o.verifysort_nr || !td->o.do_verify)
195 if (td->runstate != TD_VERIFYING)
197 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
203 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
208 if (td->o.perc_rand[ddir] == 100)
211 r = __rand(&td->seq_rand_state[ddir]);
212 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
214 return v <= td->o.perc_rand[ddir];
217 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
218 enum fio_ddir ddir, uint64_t *b)
223 if (!should_sort_io(td))
224 return get_off_from_method(td, f, ddir, b);
226 if (!flist_empty(&td->next_rand_list)) {
228 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
235 for (i = 0; i < td->o.verifysort_nr; i++) {
236 r = malloc(sizeof(*r));
238 ret = get_off_from_method(td, f, ddir, &r->off);
244 flist_add(&r->list, &td->next_rand_list);
250 assert(!flist_empty(&td->next_rand_list));
251 flist_sort(NULL, &td->next_rand_list, flist_cmp);
255 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
256 enum fio_ddir ddir, uint64_t *b)
258 if (!get_next_rand_offset(td, f, ddir, b))
261 if (td->o.time_based) {
262 fio_file_reset(td, f);
263 if (!get_next_rand_offset(td, f, ddir, b))
267 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
268 f->file_name, (unsigned long long) f->last_pos[ddir],
269 (unsigned long long) f->real_file_size);
273 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
274 enum fio_ddir ddir, uint64_t *offset)
276 struct thread_options *o = &td->o;
278 assert(ddir_rw(ddir));
280 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
282 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
284 if (f->last_pos[ddir] < f->real_file_size) {
287 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
288 f->last_pos[ddir] = f->real_file_size;
290 pos = f->last_pos[ddir] - f->file_offset;
291 if (pos && o->ddir_seq_add) {
292 pos += o->ddir_seq_add;
295 * If we reach beyond the end of the file
296 * with holed IO, wrap around to the
299 if (pos >= f->real_file_size)
300 pos = f->file_offset;
310 static int get_next_block(struct thread_data *td, struct io_u *io_u,
311 enum fio_ddir ddir, int rw_seq,
312 unsigned int *is_random)
314 struct fio_file *f = io_u->file;
318 assert(ddir_rw(ddir));
324 if (should_do_random(td, ddir)) {
325 ret = get_next_rand_block(td, f, ddir, &b);
329 io_u->flags |= IO_U_F_BUSY_OK;
330 ret = get_next_seq_offset(td, f, ddir, &offset);
332 ret = get_next_rand_block(td, f, ddir, &b);
336 ret = get_next_seq_offset(td, f, ddir, &offset);
339 io_u->flags |= IO_U_F_BUSY_OK;
342 if (td->o.rw_seq == RW_SEQ_SEQ) {
343 ret = get_next_seq_offset(td, f, ddir, &offset);
345 ret = get_next_rand_block(td, f, ddir, &b);
348 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
349 if (f->last_start[ddir] != -1ULL)
350 offset = f->last_start[ddir] - f->file_offset;
355 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
362 io_u->offset = offset;
364 io_u->offset = b * td->o.ba[ddir];
366 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
375 * For random io, generate a random new block and see if it's used. Repeat
376 * until we find a free one. For sequential io, just return the end of
377 * the last io issued.
379 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
380 unsigned int *is_random)
382 struct fio_file *f = io_u->file;
383 enum fio_ddir ddir = io_u->ddir;
386 assert(ddir_rw(ddir));
388 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
390 td->ddir_seq_nr = td->o.ddir_seq_nr;
393 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
396 if (io_u->offset >= f->io_size) {
397 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
398 (unsigned long long) io_u->offset,
399 (unsigned long long) f->io_size);
403 io_u->offset += f->file_offset;
404 if (io_u->offset >= f->real_file_size) {
405 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
406 (unsigned long long) io_u->offset,
407 (unsigned long long) f->real_file_size);
414 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
415 unsigned int *is_random)
417 if (td->flags & TD_F_PROFILE_OPS) {
418 struct prof_io_ops *ops = &td->prof_io_ops;
420 if (ops->fill_io_u_off)
421 return ops->fill_io_u_off(td, io_u, is_random);
424 return __get_next_offset(td, io_u, is_random);
427 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
430 struct fio_file *f = io_u->file;
432 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
435 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
436 unsigned int is_random)
438 int ddir = io_u->ddir;
439 unsigned int buflen = 0;
440 unsigned int minbs, maxbs;
443 assert(ddir_rw(ddir));
445 if (td->o.bs_is_seq_rand)
446 ddir = is_random ? DDIR_WRITE: DDIR_READ;
448 minbs = td->o.min_bs[ddir];
449 maxbs = td->o.max_bs[ddir];
455 * If we can't satisfy the min block size from here, then fail
457 if (!io_u_fits(td, io_u, minbs))
461 r = __rand(&td->bsrange_state);
463 if (!td->o.bssplit_nr[ddir]) {
464 buflen = 1 + (unsigned int) ((double) maxbs *
465 (r / (FRAND_MAX + 1.0)));
472 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
473 struct bssplit *bsp = &td->o.bssplit[ddir][i];
477 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
478 io_u_fits(td, io_u, buflen))
483 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
484 buflen = (buflen + td->o.verify_interval - 1) &
485 ~(td->o.verify_interval - 1);
487 if (!td->o.bs_unaligned && is_power_of_2(minbs))
488 buflen = (buflen + minbs - 1) & ~(minbs - 1);
490 } while (!io_u_fits(td, io_u, buflen));
495 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
496 unsigned int is_random)
498 if (td->flags & TD_F_PROFILE_OPS) {
499 struct prof_io_ops *ops = &td->prof_io_ops;
501 if (ops->fill_io_u_size)
502 return ops->fill_io_u_size(td, io_u, is_random);
505 return __get_next_buflen(td, io_u, is_random);
508 static void set_rwmix_bytes(struct thread_data *td)
513 * we do time or byte based switch. this is needed because
514 * buffered writes may issue a lot quicker than they complete,
515 * whereas reads do not.
517 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
518 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
521 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
526 r = __rand(&td->rwmix_state);
527 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
529 if (v <= td->o.rwmix[DDIR_READ])
535 void io_u_quiesce(struct thread_data *td)
538 * We are going to sleep, ensure that we flush anything pending as
539 * not to skew our latency numbers.
541 * Changed to only monitor 'in flight' requests here instead of the
542 * td->cur_depth, b/c td->cur_depth does not accurately represent
543 * io's that have been actually submitted to an async engine,
544 * and cur_depth is meaningless for sync engines.
546 if (td->io_u_queued || td->cur_depth) {
549 ret = td_io_commit(td);
552 while (td->io_u_in_flight) {
555 ret = io_u_queued_complete(td, 1, NULL);
559 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
561 enum fio_ddir odir = ddir ^ 1;
564 assert(ddir_rw(ddir));
566 if (td->rate_pending_usleep[ddir] <= 0)
570 * We have too much pending sleep in this direction. See if we
573 if (td_rw(td) && td->o.rwmix[odir]) {
575 * Other direction does not have too much pending, switch
577 if (td->rate_pending_usleep[odir] < 100000)
581 * Both directions have pending sleep. Sleep the minimum time
582 * and deduct from both.
584 if (td->rate_pending_usleep[ddir] <=
585 td->rate_pending_usleep[odir]) {
586 usec = td->rate_pending_usleep[ddir];
588 usec = td->rate_pending_usleep[odir];
592 usec = td->rate_pending_usleep[ddir];
596 usec = usec_sleep(td, usec);
598 td->rate_pending_usleep[ddir] -= usec;
601 if (td_rw(td) && __should_check_rate(td, odir))
602 td->rate_pending_usleep[odir] -= usec;
608 * Return the data direction for the next io_u. If the job is a
609 * mixed read/write workload, check the rwmix cycle and switch if
612 static enum fio_ddir get_rw_ddir(struct thread_data *td)
617 * see if it's time to fsync
619 if (td->o.fsync_blocks &&
620 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
621 td->io_issues[DDIR_WRITE] && should_fsync(td))
625 * see if it's time to fdatasync
627 if (td->o.fdatasync_blocks &&
628 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
629 td->io_issues[DDIR_WRITE] && should_fsync(td))
630 return DDIR_DATASYNC;
633 * see if it's time to sync_file_range
635 if (td->sync_file_range_nr &&
636 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
637 td->io_issues[DDIR_WRITE] && should_fsync(td))
638 return DDIR_SYNC_FILE_RANGE;
642 * Check if it's time to seed a new data direction.
644 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
646 * Put a top limit on how many bytes we do for
647 * one data direction, to avoid overflowing the
650 ddir = get_rand_ddir(td);
652 if (ddir != td->rwmix_ddir)
655 td->rwmix_ddir = ddir;
657 ddir = td->rwmix_ddir;
658 } else if (td_read(td))
660 else if (td_write(td))
665 td->rwmix_ddir = rate_ddir(td, ddir);
666 return td->rwmix_ddir;
669 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
671 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
673 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
674 td->o.barrier_blocks &&
675 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
676 td->io_issues[DDIR_WRITE])
677 io_u->flags |= IO_U_F_BARRIER;
680 void put_file_log(struct thread_data *td, struct fio_file *f)
682 unsigned int ret = put_file(td, f);
685 td_verror(td, ret, "file close");
688 void put_io_u(struct thread_data *td, struct io_u *io_u)
692 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
693 put_file_log(td, io_u->file);
696 io_u->flags |= IO_U_F_FREE;
698 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
700 io_u_qpush(&td->io_u_freelist, io_u);
702 td_io_u_free_notify(td);
705 void clear_io_u(struct thread_data *td, struct io_u *io_u)
707 io_u->flags &= ~IO_U_F_FLIGHT;
711 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
713 struct io_u *__io_u = *io_u;
714 enum fio_ddir ddir = acct_ddir(__io_u);
716 dprint(FD_IO, "requeue %p\n", __io_u);
720 __io_u->flags |= IO_U_F_FREE;
721 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
722 td->io_issues[ddir]--;
724 __io_u->flags &= ~IO_U_F_FLIGHT;
725 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
728 io_u_rpush(&td->io_u_requeues, __io_u);
733 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
735 unsigned int is_random;
737 if (td->io_ops->flags & FIO_NOIO)
740 set_rw_ddir(td, io_u);
743 * fsync() or fdatasync() or trim etc, we are done
745 if (!ddir_rw(io_u->ddir))
749 * See if it's time to switch to a new zone
751 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
752 struct fio_file *f = io_u->file;
755 f->file_offset += td->o.zone_range + td->o.zone_skip;
758 * Wrap from the beginning, if we exceed the file size
760 if (f->file_offset >= f->real_file_size)
761 f->file_offset = f->real_file_size - f->file_offset;
762 f->last_pos[io_u->ddir] = f->file_offset;
763 td->io_skip_bytes += td->o.zone_skip;
767 * No log, let the seq/rand engine retrieve the next buflen and
770 if (get_next_offset(td, io_u, &is_random)) {
771 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
775 io_u->buflen = get_next_buflen(td, io_u, is_random);
777 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
781 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
782 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
783 dprint(FD_IO, " off=%llu/%lu > %llu\n",
784 (unsigned long long) io_u->offset, io_u->buflen,
785 (unsigned long long) io_u->file->real_file_size);
790 * mark entry before potentially trimming io_u
792 if (td_random(td) && file_randommap(td, io_u->file))
793 mark_random_map(td, io_u);
796 dprint_io_u(io_u, "fill_io_u");
797 td->zone_bytes += io_u->buflen;
801 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
830 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
832 __io_u_mark_map(td->ts.io_u_submit, nr);
833 td->ts.total_submit++;
836 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
838 __io_u_mark_map(td->ts.io_u_complete, nr);
839 td->ts.total_complete++;
842 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
846 switch (td->cur_depth) {
868 td->ts.io_u_map[idx] += nr;
871 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
908 assert(idx < FIO_IO_U_LAT_U_NR);
909 td->ts.io_u_lat_u[idx]++;
912 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
953 assert(idx < FIO_IO_U_LAT_M_NR);
954 td->ts.io_u_lat_m[idx]++;
957 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
960 io_u_mark_lat_usec(td, usec);
962 io_u_mark_lat_msec(td, usec / 1000);
966 * Get next file to service by choosing one at random
968 static struct fio_file *get_next_file_rand(struct thread_data *td,
969 enum fio_file_flags goodf,
970 enum fio_file_flags badf)
979 r = __rand(&td->next_file_state);
980 fno = (unsigned int) ((double) td->o.nr_files
981 * (r / (FRAND_MAX + 1.0)));
984 if (fio_file_done(f))
987 if (!fio_file_open(f)) {
990 if (td->nr_open_files >= td->o.open_files)
991 return ERR_PTR(-EBUSY);
993 err = td_io_open_file(td, f);
999 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1000 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1004 td_io_close_file(td, f);
1009 * Get next file to service by doing round robin between all available ones
1011 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1014 unsigned int old_next_file = td->next_file;
1020 f = td->files[td->next_file];
1023 if (td->next_file >= td->o.nr_files)
1026 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1027 if (fio_file_done(f)) {
1032 if (!fio_file_open(f)) {
1035 if (td->nr_open_files >= td->o.open_files)
1036 return ERR_PTR(-EBUSY);
1038 err = td_io_open_file(td, f);
1040 dprint(FD_FILE, "error %d on open of %s\n",
1048 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1050 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1054 td_io_close_file(td, f);
1057 } while (td->next_file != old_next_file);
1059 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1063 static struct fio_file *__get_next_file(struct thread_data *td)
1067 assert(td->o.nr_files <= td->files_index);
1069 if (td->nr_done_files >= td->o.nr_files) {
1070 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1071 " nr_files=%d\n", td->nr_open_files,
1077 f = td->file_service_file;
1078 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1079 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1081 if (td->file_service_left--)
1085 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1086 td->o.file_service_type == FIO_FSERVICE_SEQ)
1087 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1089 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1094 td->file_service_file = f;
1095 td->file_service_left = td->file_service_nr - 1;
1098 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1100 dprint(FD_FILE, "get_next_file: NULL\n");
1104 static struct fio_file *get_next_file(struct thread_data *td)
1106 if (td->flags & TD_F_PROFILE_OPS) {
1107 struct prof_io_ops *ops = &td->prof_io_ops;
1109 if (ops->get_next_file)
1110 return ops->get_next_file(td);
1113 return __get_next_file(td);
1116 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1121 f = get_next_file(td);
1122 if (IS_ERR_OR_NULL(f))
1128 if (!fill_io_u(td, io_u))
1131 put_file_log(td, f);
1132 td_io_close_file(td, f);
1134 fio_file_set_done(f);
1135 td->nr_done_files++;
1136 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1137 td->nr_done_files, td->o.nr_files);
1143 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1144 unsigned long tusec, unsigned long max_usec)
1147 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1148 td_verror(td, ETIMEDOUT, "max latency exceeded");
1149 icd->error = ETIMEDOUT;
1152 static void lat_new_cycle(struct thread_data *td)
1154 fio_gettime(&td->latency_ts, NULL);
1155 td->latency_ios = ddir_rw_sum(td->io_blocks);
1156 td->latency_failed = 0;
1160 * We had an IO outside the latency target. Reduce the queue depth. If we
1161 * are at QD=1, then it's time to give up.
1163 static int __lat_target_failed(struct thread_data *td)
1165 if (td->latency_qd == 1)
1168 td->latency_qd_high = td->latency_qd;
1170 if (td->latency_qd == td->latency_qd_low)
1171 td->latency_qd_low--;
1173 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1175 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1178 * When we ramp QD down, quiesce existing IO to prevent
1179 * a storm of ramp downs due to pending higher depth.
1186 static int lat_target_failed(struct thread_data *td)
1188 if (td->o.latency_percentile.u.f == 100.0)
1189 return __lat_target_failed(td);
1191 td->latency_failed++;
1195 void lat_target_init(struct thread_data *td)
1197 td->latency_end_run = 0;
1199 if (td->o.latency_target) {
1200 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1201 fio_gettime(&td->latency_ts, NULL);
1203 td->latency_qd_high = td->o.iodepth;
1204 td->latency_qd_low = 1;
1205 td->latency_ios = ddir_rw_sum(td->io_blocks);
1207 td->latency_qd = td->o.iodepth;
1210 void lat_target_reset(struct thread_data *td)
1212 if (!td->latency_end_run)
1213 lat_target_init(td);
1216 static void lat_target_success(struct thread_data *td)
1218 const unsigned int qd = td->latency_qd;
1219 struct thread_options *o = &td->o;
1221 td->latency_qd_low = td->latency_qd;
1224 * If we haven't failed yet, we double up to a failing value instead
1225 * of bisecting from highest possible queue depth. If we have set
1226 * a limit other than td->o.iodepth, bisect between that.
1228 if (td->latency_qd_high != o->iodepth)
1229 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1231 td->latency_qd *= 2;
1233 if (td->latency_qd > o->iodepth)
1234 td->latency_qd = o->iodepth;
1236 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1239 * Same as last one, we are done. Let it run a latency cycle, so
1240 * we get only the results from the targeted depth.
1242 if (td->latency_qd == qd) {
1243 if (td->latency_end_run) {
1244 dprint(FD_RATE, "We are done\n");
1247 dprint(FD_RATE, "Quiesce and final run\n");
1249 td->latency_end_run = 1;
1250 reset_all_stats(td);
1259 * Check if we can bump the queue depth
1261 void lat_target_check(struct thread_data *td)
1263 uint64_t usec_window;
1267 usec_window = utime_since_now(&td->latency_ts);
1268 if (usec_window < td->o.latency_window)
1271 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1272 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1273 success_ios *= 100.0;
1275 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1277 if (success_ios >= td->o.latency_percentile.u.f)
1278 lat_target_success(td);
1280 __lat_target_failed(td);
1284 * If latency target is enabled, we might be ramping up or down and not
1285 * using the full queue depth available.
1287 int queue_full(const struct thread_data *td)
1289 const int qempty = io_u_qempty(&td->io_u_freelist);
1293 if (!td->o.latency_target)
1296 return td->cur_depth >= td->latency_qd;
1299 struct io_u *__get_io_u(struct thread_data *td)
1301 struct io_u *io_u = NULL;
1309 if (!io_u_rempty(&td->io_u_requeues))
1310 io_u = io_u_rpop(&td->io_u_requeues);
1311 else if (!queue_full(td)) {
1312 io_u = io_u_qpop(&td->io_u_freelist);
1317 io_u->end_io = NULL;
1321 assert(io_u->flags & IO_U_F_FREE);
1322 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1323 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1327 io_u->acct_ddir = -1;
1329 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1331 } else if (td->o.verify_async) {
1333 * We ran out, wait for async verify threads to finish and
1336 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1344 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1346 if (!(td->flags & TD_F_TRIM_BACKLOG))
1349 if (td->trim_entries) {
1352 if (td->trim_batch) {
1355 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1356 td->last_ddir != DDIR_READ) {
1357 td->trim_batch = td->o.trim_batch;
1358 if (!td->trim_batch)
1359 td->trim_batch = td->o.trim_backlog;
1363 if (get_trim && !get_next_trim(td, io_u))
1370 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1372 if (!(td->flags & TD_F_VER_BACKLOG))
1375 if (td->io_hist_len) {
1378 if (td->verify_batch)
1380 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1381 td->last_ddir != DDIR_READ) {
1382 td->verify_batch = td->o.verify_batch;
1383 if (!td->verify_batch)
1384 td->verify_batch = td->o.verify_backlog;
1388 if (get_verify && !get_next_verify(td, io_u)) {
1398 * Fill offset and start time into the buffer content, to prevent too
1399 * easy compressible data for simple de-dupe attempts. Do this for every
1400 * 512b block in the range, since that should be the smallest block size
1401 * we can expect from a device.
1403 static void small_content_scramble(struct io_u *io_u)
1405 unsigned int i, nr_blocks = io_u->buflen / 512;
1407 unsigned int offset;
1414 boffset = io_u->offset;
1415 io_u->buf_filled_len = 0;
1417 for (i = 0; i < nr_blocks; i++) {
1419 * Fill the byte offset into a "random" start offset of
1420 * the buffer, given by the product of the usec time
1421 * and the actual offset.
1423 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1424 offset &= ~(sizeof(uint64_t) - 1);
1425 if (offset >= 512 - sizeof(uint64_t))
1426 offset -= sizeof(uint64_t);
1427 memcpy(p + offset, &boffset, sizeof(boffset));
1429 end = p + 512 - sizeof(io_u->start_time);
1430 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1437 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1438 * etc. The returned io_u is fully ready to be prepped and submitted.
1440 struct io_u *get_io_u(struct thread_data *td)
1444 int do_scramble = 0;
1447 io_u = __get_io_u(td);
1449 dprint(FD_IO, "__get_io_u failed\n");
1453 if (check_get_verify(td, io_u))
1455 if (check_get_trim(td, io_u))
1459 * from a requeue, io_u already setup
1465 * If using an iolog, grab next piece if any available.
1467 if (td->flags & TD_F_READ_IOLOG) {
1468 if (read_iolog_get(td, io_u))
1470 } else if (set_io_u_file(td, io_u)) {
1472 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1478 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1482 assert(fio_file_open(f));
1484 if (ddir_rw(io_u->ddir)) {
1485 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1486 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1490 f->last_start[io_u->ddir] = io_u->offset;
1491 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1493 if (io_u->ddir == DDIR_WRITE) {
1494 if (td->flags & TD_F_REFILL_BUFFERS) {
1495 io_u_fill_buffer(td, io_u,
1496 td->o.min_bs[DDIR_WRITE],
1498 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1499 !(td->flags & TD_F_COMPRESS))
1501 if (td->flags & TD_F_VER_NONE) {
1502 populate_verify_io_u(td, io_u);
1505 } else if (io_u->ddir == DDIR_READ) {
1507 * Reset the buf_filled parameters so next time if the
1508 * buffer is used for writes it is refilled.
1510 io_u->buf_filled_len = 0;
1515 * Set io data pointers.
1517 io_u->xfer_buf = io_u->buf;
1518 io_u->xfer_buflen = io_u->buflen;
1522 if (!td_io_prep(td, io_u)) {
1523 if (!td->o.disable_slat)
1524 fio_gettime(&io_u->start_time, NULL);
1526 small_content_scramble(io_u);
1530 dprint(FD_IO, "get_io_u failed\n");
1532 return ERR_PTR(ret);
1535 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1537 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1539 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1542 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1543 io_u->file ? " on file " : "",
1544 io_u->file ? io_u->file->file_name : "",
1545 strerror(io_u->error),
1546 io_ddir_name(io_u->ddir),
1547 io_u->offset, io_u->xfer_buflen);
1550 td_verror(td, io_u->error, "io_u error");
1553 static inline int gtod_reduce(struct thread_data *td)
1555 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1556 && td->o.disable_bw;
1559 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1560 struct io_completion_data *icd,
1561 const enum fio_ddir idx, unsigned int bytes)
1563 unsigned long lusec = 0;
1565 if (!gtod_reduce(td))
1566 lusec = utime_since(&io_u->issue_time, &icd->time);
1568 if (!td->o.disable_lat) {
1569 unsigned long tusec;
1571 tusec = utime_since(&io_u->start_time, &icd->time);
1572 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1574 if (td->flags & TD_F_PROFILE_OPS) {
1575 struct prof_io_ops *ops = &td->prof_io_ops;
1578 icd->error = ops->io_u_lat(td, tusec);
1581 if (td->o.max_latency && tusec > td->o.max_latency)
1582 lat_fatal(td, icd, tusec, td->o.max_latency);
1583 if (td->o.latency_target && tusec > td->o.latency_target) {
1584 if (lat_target_failed(td))
1585 lat_fatal(td, icd, tusec, td->o.latency_target);
1589 if (!td->o.disable_clat) {
1590 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1591 io_u_mark_latency(td, lusec);
1594 if (!td->o.disable_bw)
1595 add_bw_sample(td, idx, bytes, &icd->time);
1597 if (!gtod_reduce(td))
1598 add_iops_sample(td, idx, bytes, &icd->time);
1601 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1603 uint64_t secs, remainder, bps, bytes;
1605 bytes = td->this_io_bytes[ddir];
1606 bps = td->rate_bps[ddir];
1608 remainder = bytes % bps;
1609 return remainder * 1000000 / bps + secs * 1000000;
1612 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1613 struct io_completion_data *icd)
1615 struct io_u *io_u = *io_u_ptr;
1616 enum fio_ddir ddir = io_u->ddir;
1617 struct fio_file *f = io_u->file;
1619 dprint_io_u(io_u, "io complete");
1622 assert(io_u->flags & IO_U_F_FLIGHT);
1623 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1626 * Mark IO ok to verify
1630 * Remove errored entry from the verification list
1633 unlog_io_piece(td, io_u);
1635 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1642 if (ddir_sync(ddir)) {
1643 td->last_was_sync = 1;
1645 f->first_write = -1ULL;
1646 f->last_write = -1ULL;
1651 td->last_was_sync = 0;
1652 td->last_ddir = ddir;
1654 if (!io_u->error && ddir_rw(ddir)) {
1655 unsigned int bytes = io_u->buflen - io_u->resid;
1656 const enum fio_ddir oddir = ddir ^ 1;
1659 td->io_blocks[ddir]++;
1660 td->this_io_blocks[ddir]++;
1661 td->io_bytes[ddir] += bytes;
1663 if (!(io_u->flags & IO_U_F_VER_LIST))
1664 td->this_io_bytes[ddir] += bytes;
1666 if (ddir == DDIR_WRITE) {
1668 if (f->first_write == -1ULL ||
1669 io_u->offset < f->first_write)
1670 f->first_write = io_u->offset;
1671 if (f->last_write == -1ULL ||
1672 ((io_u->offset + bytes) > f->last_write))
1673 f->last_write = io_u->offset + bytes;
1675 if (td->last_write_comp) {
1676 int idx = td->last_write_idx++;
1678 td->last_write_comp[idx] = io_u->offset;
1679 if (td->last_write_idx == td->o.iodepth)
1680 td->last_write_idx = 0;
1684 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1685 td->runstate == TD_VERIFYING)) {
1686 account_io_completion(td, io_u, icd, ddir, bytes);
1688 if (__should_check_rate(td, ddir)) {
1689 td->rate_pending_usleep[ddir] =
1690 (usec_for_io(td, ddir) -
1691 utime_since_now(&td->start));
1693 if (ddir != DDIR_TRIM &&
1694 __should_check_rate(td, oddir)) {
1695 td->rate_pending_usleep[oddir] =
1696 (usec_for_io(td, oddir) -
1697 utime_since_now(&td->start));
1701 icd->bytes_done[ddir] += bytes;
1704 ret = io_u->end_io(td, io_u_ptr);
1706 if (ret && !icd->error)
1709 } else if (io_u->error) {
1710 icd->error = io_u->error;
1711 io_u_log_error(td, io_u);
1714 enum error_type_bit eb = td_error_type(ddir, icd->error);
1716 if (!td_non_fatal_error(td, eb, icd->error))
1720 * If there is a non_fatal error, then add to the error count
1721 * and clear all the errors.
1723 update_error_count(td, icd->error);
1731 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1736 if (!gtod_reduce(td))
1737 fio_gettime(&icd->time, NULL);
1742 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1743 icd->bytes_done[ddir] = 0;
1746 static void ios_completed(struct thread_data *td,
1747 struct io_completion_data *icd)
1752 for (i = 0; i < icd->nr; i++) {
1753 io_u = td->io_ops->event(td, i);
1755 io_completed(td, &io_u, icd);
1763 * Complete a single io_u for the sync engines.
1765 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1768 struct io_completion_data icd;
1770 init_icd(td, &icd, 1);
1771 io_completed(td, &io_u, &icd);
1777 td_verror(td, icd.error, "io_u_sync_complete");
1784 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1785 bytes[ddir] += icd.bytes_done[ddir];
1792 * Called to complete min_events number of io for the async engines.
1794 int io_u_queued_complete(struct thread_data *td, int min_evts,
1797 struct io_completion_data icd;
1798 struct timespec *tvp = NULL;
1800 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1802 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1806 else if (min_evts > td->cur_depth)
1807 min_evts = td->cur_depth;
1809 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1811 td_verror(td, -ret, "td_io_getevents");
1816 init_icd(td, &icd, ret);
1817 ios_completed(td, &icd);
1819 td_verror(td, icd.error, "io_u_queued_complete");
1826 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1827 bytes[ddir] += icd.bytes_done[ddir];
1834 * Call when io_u is really queued, to update the submission latency.
1836 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1838 if (!td->o.disable_slat) {
1839 unsigned long slat_time;
1841 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1842 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1848 * See if we should reuse the last seed, if dedupe is enabled
1850 static struct frand_state *get_buf_state(struct thread_data *td)
1855 if (!td->o.dedupe_percentage)
1856 return &td->buf_state;
1857 else if (td->o.dedupe_percentage == 100)
1858 return &td->buf_state_prev;
1860 r = __rand(&td->dedupe_state);
1861 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1863 if (v <= td->o.dedupe_percentage)
1864 return &td->buf_state_prev;
1866 return &td->buf_state;
1869 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1871 if (rs == &td->buf_state)
1872 frand_copy(&td->buf_state_prev, rs);
1875 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1876 unsigned int max_bs)
1878 struct thread_options *o = &td->o;
1880 if (o->compress_percentage || o->dedupe_percentage) {
1881 unsigned int perc = td->o.compress_percentage;
1882 struct frand_state *rs;
1883 unsigned int left = max_bs;
1886 rs = get_buf_state(td);
1888 min_write = min(min_write, left);
1891 unsigned int seg = min_write;
1893 seg = min(min_write, td->o.compress_chunk);
1897 fill_random_buf_percentage(rs, buf, perc, seg,
1898 min_write, o->buffer_pattern,
1899 o->buffer_pattern_bytes);
1901 fill_random_buf(rs, buf, min_write);
1905 save_buf_state(td, rs);
1907 } else if (o->buffer_pattern_bytes)
1908 fill_buffer_pattern(td, buf, max_bs);
1909 else if (o->zero_buffers)
1910 memset(buf, 0, max_bs);
1912 fill_random_buf(get_buf_state(td), buf, max_bs);
1916 * "randomly" fill the buffer contents
1918 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1919 unsigned int min_write, unsigned int max_bs)
1921 io_u->buf_filled_len = 0;
1922 fill_io_buffer(td, io_u->buf, min_write, max_bs);