14 struct io_completion_data {
17 int error; /* output */
18 unsigned long bytes_done[2]; /* output */
19 struct timeval time; /* output */
23 * The ->file_map[] contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 static int random_map_free(struct fio_file *f, const unsigned long long block)
28 unsigned int idx = RAND_MAP_IDX(f, block);
29 unsigned int bit = RAND_MAP_BIT(f, block);
31 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
33 return (f->file_map[idx] & (1UL << bit)) == 0;
37 * Mark a given offset as used in the map.
39 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
41 unsigned int min_bs = td->o.rw_min_bs;
42 struct fio_file *f = io_u->file;
43 unsigned long long block;
44 unsigned int blocks, nr_blocks;
47 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
48 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
50 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
53 unsigned int idx, bit;
54 unsigned long mask, this_blocks;
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
64 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
67 idx = RAND_MAP_IDX(f, block);
68 bit = RAND_MAP_BIT(f, block);
70 fio_assert(td, idx < f->num_maps);
72 this_blocks = nr_blocks;
73 if (this_blocks + bit > BLOCKS_PER_MAP)
74 this_blocks = BLOCKS_PER_MAP - bit;
77 if (this_blocks == BLOCKS_PER_MAP)
80 mask = ((1UL << this_blocks) - 1) << bit;
82 if (!(f->file_map[idx] & mask))
86 } while (this_blocks);
91 f->file_map[idx] |= mask;
92 nr_blocks -= this_blocks;
93 blocks += this_blocks;
97 if ((blocks * min_bs) < io_u->buflen)
98 io_u->buflen = blocks * min_bs;
101 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
104 unsigned long long max_blocks;
105 unsigned long long max_size;
107 assert(ddir_rw(ddir));
110 * Hmm, should we make sure that ->io_size <= ->real_file_size?
112 max_size = f->io_size;
113 if (max_size > f->real_file_size)
114 max_size = f->real_file_size;
116 if (td->o.zone_range)
117 max_size = td->o.zone_range;
119 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
127 * Return the next free block in the map.
129 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
130 enum fio_ddir ddir, unsigned long long *b)
132 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
135 lastb = last_block(td, f, ddir);
139 i = f->last_free_lookup;
140 block = i * BLOCKS_PER_MAP;
141 while (block * min_bs < f->real_file_size &&
142 block * min_bs < f->io_size) {
143 if (f->file_map[i] != -1UL) {
144 block += ffz(f->file_map[i]);
147 f->last_free_lookup = i;
152 block += BLOCKS_PER_MAP;
156 dprint(FD_IO, "failed finding a free block\n");
160 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
161 enum fio_ddir ddir, unsigned long long *b)
163 unsigned long long rmax, r, lastb;
166 lastb = last_block(td, f, ddir);
170 if (f->failed_rands >= 200)
173 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
175 if (td->o.use_os_rand)
176 r = os_random_long(&td->random_state);
178 r = __rand(&td->__random_state);
180 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
182 dprint(FD_RANDOM, "off rand %llu\n", r);
186 * if we are not maintaining a random map, we are done.
188 if (!file_randommap(td, f))
192 * calculate map offset and check if it's free
194 if (random_map_free(f, *b))
197 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
201 if (!f->failed_rands++)
202 f->last_free_lookup = 0;
205 * we get here, if we didn't suceed in looking up a block. generate
206 * a random start offset into the filemap, and find the first free
211 f->last_free_lookup = (f->num_maps - 1) *
212 (r / ((unsigned long long) rmax + 1.0));
213 if (!get_next_free_block(td, f, ddir, b))
216 if (td->o.use_os_rand)
217 r = os_random_long(&td->random_state);
219 r = __rand(&td->__random_state);
223 * that didn't work either, try exhaustive search from the start
225 f->last_free_lookup = 0;
227 if (!get_next_free_block(td, f, ddir, b))
229 f->last_free_lookup = 0;
230 return get_next_free_block(td, f, ddir, b);
237 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, unsigned long long *b)
240 if (!get_next_rand_offset(td, f, ddir, b))
243 if (td->o.time_based) {
245 if (!get_next_rand_offset(td, f, ddir, b))
249 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250 f->file_name, f->last_pos, f->real_file_size);
254 static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
255 enum fio_ddir ddir, unsigned long long *b)
257 assert(ddir_rw(ddir));
259 if (f->last_pos >= f->io_size && td->o.time_based)
260 f->last_pos = f->last_pos - f->io_size;
262 if (f->last_pos < f->real_file_size) {
263 unsigned long long pos;
265 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
266 f->last_pos = f->real_file_size;
268 pos = f->last_pos - f->file_offset;
270 pos += td->o.ddir_seq_add;
272 *b = pos / td->o.min_bs[ddir];
279 static int get_next_block(struct thread_data *td, struct io_u *io_u,
280 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
282 struct fio_file *f = io_u->file;
285 assert(ddir_rw(ddir));
289 ret = get_next_rand_block(td, f, ddir, b);
291 ret = get_next_seq_block(td, f, ddir, b);
293 io_u->flags |= IO_U_F_BUSY_OK;
295 if (td->o.rw_seq == RW_SEQ_SEQ) {
296 ret = get_next_seq_block(td, f, ddir, b);
298 ret = get_next_rand_block(td, f, ddir, b);
299 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
300 if (f->last_start != -1ULL)
301 *b = (f->last_start - f->file_offset)
302 / td->o.min_bs[ddir];
307 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
316 * For random io, generate a random new block and see if it's used. Repeat
317 * until we find a free one. For sequential io, just return the end of
318 * the last io issued.
320 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
322 struct fio_file *f = io_u->file;
323 unsigned long long b;
324 enum fio_ddir ddir = io_u->ddir;
327 assert(ddir_rw(ddir));
329 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
331 td->ddir_seq_nr = td->o.ddir_seq_nr;
334 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
337 io_u->offset = b * td->o.ba[ddir];
338 if (io_u->offset >= f->io_size) {
339 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
340 io_u->offset, f->io_size);
344 io_u->offset += f->file_offset;
345 if (io_u->offset >= f->real_file_size) {
346 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
347 io_u->offset, f->real_file_size);
354 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
356 struct prof_io_ops *ops = &td->prof_io_ops;
358 if (ops->fill_io_u_off)
359 return ops->fill_io_u_off(td, io_u);
361 return __get_next_offset(td, io_u);
364 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
367 struct fio_file *f = io_u->file;
369 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
372 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
374 const int ddir = io_u->ddir;
375 unsigned int uninitialized_var(buflen);
376 unsigned int minbs, maxbs;
377 unsigned long r, rand_max;
379 assert(ddir_rw(ddir));
381 minbs = td->o.min_bs[ddir];
382 maxbs = td->o.max_bs[ddir];
388 * If we can't satisfy the min block size from here, then fail
390 if (!io_u_fits(td, io_u, minbs))
393 if (td->o.use_os_rand)
394 rand_max = OS_RAND_MAX;
396 rand_max = FRAND_MAX;
399 if (td->o.use_os_rand)
400 r = os_random_long(&td->bsrange_state);
402 r = __rand(&td->__bsrange_state);
404 if (!td->o.bssplit_nr[ddir]) {
405 buflen = 1 + (unsigned int) ((double) maxbs *
406 (r / (rand_max + 1.0)));
413 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
414 struct bssplit *bsp = &td->o.bssplit[ddir][i];
418 if ((r <= ((rand_max / 100L) * perc)) &&
419 io_u_fits(td, io_u, buflen))
424 if (!td->o.bs_unaligned && is_power_of_2(minbs))
425 buflen = (buflen + minbs - 1) & ~(minbs - 1);
427 } while (!io_u_fits(td, io_u, buflen));
432 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
434 struct prof_io_ops *ops = &td->prof_io_ops;
436 if (ops->fill_io_u_size)
437 return ops->fill_io_u_size(td, io_u);
439 return __get_next_buflen(td, io_u);
442 static void set_rwmix_bytes(struct thread_data *td)
447 * we do time or byte based switch. this is needed because
448 * buffered writes may issue a lot quicker than they complete,
449 * whereas reads do not.
451 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
452 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
455 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
460 if (td->o.use_os_rand) {
461 r = os_random_long(&td->rwmix_state);
462 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
464 r = __rand(&td->__rwmix_state);
465 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
468 if (v <= td->o.rwmix[DDIR_READ])
474 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
476 enum fio_ddir odir = ddir ^ 1;
480 assert(ddir_rw(ddir));
482 if (td->rate_pending_usleep[ddir] <= 0)
486 * We have too much pending sleep in this direction. See if we
491 * Other direction does not have too much pending, switch
493 if (td->rate_pending_usleep[odir] < 100000)
497 * Both directions have pending sleep. Sleep the minimum time
498 * and deduct from both.
500 if (td->rate_pending_usleep[ddir] <=
501 td->rate_pending_usleep[odir]) {
502 usec = td->rate_pending_usleep[ddir];
504 usec = td->rate_pending_usleep[odir];
508 usec = td->rate_pending_usleep[ddir];
511 * We are going to sleep, ensure that we flush anything pending as
512 * not to skew our latency numbers.
514 * Changed to only monitor 'in flight' requests here instead of the
515 * td->cur_depth, b/c td->cur_depth does not accurately represent
516 * io's that have been actually submitted to an async engine,
517 * and cur_depth is meaningless for sync engines.
519 if (td->io_u_in_flight) {
522 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
525 fio_gettime(&t, NULL);
526 usec_sleep(td, usec);
527 usec = utime_since_now(&t);
529 td->rate_pending_usleep[ddir] -= usec;
532 if (td_rw(td) && __should_check_rate(td, odir))
533 td->rate_pending_usleep[odir] -= usec;
539 * Return the data direction for the next io_u. If the job is a
540 * mixed read/write workload, check the rwmix cycle and switch if
543 static enum fio_ddir get_rw_ddir(struct thread_data *td)
548 * see if it's time to fsync
550 if (td->o.fsync_blocks &&
551 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
552 td->io_issues[DDIR_WRITE] && should_fsync(td))
556 * see if it's time to fdatasync
558 if (td->o.fdatasync_blocks &&
559 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
560 td->io_issues[DDIR_WRITE] && should_fsync(td))
561 return DDIR_DATASYNC;
564 * see if it's time to sync_file_range
566 if (td->sync_file_range_nr &&
567 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
568 td->io_issues[DDIR_WRITE] && should_fsync(td))
569 return DDIR_SYNC_FILE_RANGE;
573 * Check if it's time to seed a new data direction.
575 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
577 * Put a top limit on how many bytes we do for
578 * one data direction, to avoid overflowing the
581 ddir = get_rand_ddir(td);
583 if (ddir != td->rwmix_ddir)
586 td->rwmix_ddir = ddir;
588 ddir = td->rwmix_ddir;
589 } else if (td_read(td))
594 td->rwmix_ddir = rate_ddir(td, ddir);
595 return td->rwmix_ddir;
598 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
600 io_u->ddir = get_rw_ddir(td);
602 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
603 td->o.barrier_blocks &&
604 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
605 td->io_issues[DDIR_WRITE])
606 io_u->flags |= IO_U_F_BARRIER;
609 void put_file_log(struct thread_data *td, struct fio_file *f)
611 int ret = put_file(td, f);
614 td_verror(td, ret, "file close");
617 void put_io_u(struct thread_data *td, struct io_u *io_u)
621 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
622 put_file_log(td, io_u->file);
624 io_u->flags &= ~IO_U_F_FREE_DEF;
625 io_u->flags |= IO_U_F_FREE;
627 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
629 flist_del_init(&io_u->list);
630 flist_add(&io_u->list, &td->io_u_freelist);
632 td_io_u_free_notify(td);
635 void clear_io_u(struct thread_data *td, struct io_u *io_u)
637 io_u->flags &= ~IO_U_F_FLIGHT;
641 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
643 struct io_u *__io_u = *io_u;
645 dprint(FD_IO, "requeue %p\n", __io_u);
649 __io_u->flags |= IO_U_F_FREE;
650 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
651 td->io_issues[__io_u->ddir]--;
653 __io_u->flags &= ~IO_U_F_FLIGHT;
654 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
656 flist_del(&__io_u->list);
657 flist_add_tail(&__io_u->list, &td->io_u_requeues);
662 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
664 if (td->io_ops->flags & FIO_NOIO)
667 set_rw_ddir(td, io_u);
670 * fsync() or fdatasync() or trim etc, we are done
672 if (!ddir_rw(io_u->ddir))
676 * See if it's time to switch to a new zone
678 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
680 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
681 io_u->file->last_pos = io_u->file->file_offset;
682 td->io_skip_bytes += td->o.zone_skip;
686 * No log, let the seq/rand engine retrieve the next buflen and
689 if (get_next_offset(td, io_u)) {
690 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
694 io_u->buflen = get_next_buflen(td, io_u);
696 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
700 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
701 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
702 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
703 io_u->buflen, io_u->file->real_file_size);
708 * mark entry before potentially trimming io_u
710 if (td_random(td) && file_randommap(td, io_u->file))
711 mark_random_map(td, io_u);
714 * If using a write iolog, store this entry.
717 dprint_io_u(io_u, "fill_io_u");
718 td->zone_bytes += io_u->buflen;
723 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
752 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
754 __io_u_mark_map(td->ts.io_u_submit, nr);
755 td->ts.total_submit++;
758 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
760 __io_u_mark_map(td->ts.io_u_complete, nr);
761 td->ts.total_complete++;
764 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
768 switch (td->cur_depth) {
790 td->ts.io_u_map[idx] += nr;
793 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
830 assert(idx < FIO_IO_U_LAT_U_NR);
831 td->ts.io_u_lat_u[idx]++;
834 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
875 assert(idx < FIO_IO_U_LAT_M_NR);
876 td->ts.io_u_lat_m[idx]++;
879 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
882 io_u_mark_lat_usec(td, usec);
884 io_u_mark_lat_msec(td, usec / 1000);
888 * Get next file to service by choosing one at random
890 static struct fio_file *get_next_file_rand(struct thread_data *td,
891 enum fio_file_flags goodf,
892 enum fio_file_flags badf)
901 if (td->o.use_os_rand) {
902 r = os_random_long(&td->next_file_state);
903 fno = (unsigned int) ((double) td->o.nr_files
904 * (r / (OS_RAND_MAX + 1.0)));
906 r = __rand(&td->__next_file_state);
907 fno = (unsigned int) ((double) td->o.nr_files
908 * (r / (FRAND_MAX + 1.0)));
912 if (fio_file_done(f))
915 if (!fio_file_open(f)) {
918 err = td_io_open_file(td, f);
924 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
925 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
929 td_io_close_file(td, f);
934 * Get next file to service by doing round robin between all available ones
936 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
939 unsigned int old_next_file = td->next_file;
945 f = td->files[td->next_file];
948 if (td->next_file >= td->o.nr_files)
951 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
952 if (fio_file_done(f)) {
957 if (!fio_file_open(f)) {
960 err = td_io_open_file(td, f);
962 dprint(FD_FILE, "error %d on open of %s\n",
970 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
972 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
976 td_io_close_file(td, f);
979 } while (td->next_file != old_next_file);
981 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
985 static struct fio_file *__get_next_file(struct thread_data *td)
989 assert(td->o.nr_files <= td->files_index);
991 if (td->nr_done_files >= td->o.nr_files) {
992 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
993 " nr_files=%d\n", td->nr_open_files,
999 f = td->file_service_file;
1000 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1001 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1003 if (td->file_service_left--)
1007 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1008 td->o.file_service_type == FIO_FSERVICE_SEQ)
1009 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1011 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1013 td->file_service_file = f;
1014 td->file_service_left = td->file_service_nr - 1;
1016 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1020 static struct fio_file *get_next_file(struct thread_data *td)
1022 struct prof_io_ops *ops = &td->prof_io_ops;
1024 if (ops->get_next_file)
1025 return ops->get_next_file(td);
1027 return __get_next_file(td);
1030 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1035 f = get_next_file(td);
1042 if (!fill_io_u(td, io_u))
1045 put_file_log(td, f);
1046 td_io_close_file(td, f);
1048 fio_file_set_done(f);
1049 td->nr_done_files++;
1050 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1051 td->nr_done_files, td->o.nr_files);
1058 struct io_u *__get_io_u(struct thread_data *td)
1060 struct io_u *io_u = NULL;
1065 if (!flist_empty(&td->io_u_requeues))
1066 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1067 else if (!queue_full(td)) {
1068 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1073 io_u->end_io = NULL;
1077 assert(io_u->flags & IO_U_F_FREE);
1078 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1079 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1080 io_u->flags &= ~IO_U_F_VER_LIST;
1083 flist_del(&io_u->list);
1084 flist_add(&io_u->list, &td->io_u_busylist);
1086 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1087 } else if (td->o.verify_async) {
1089 * We ran out, wait for async verify threads to finish and
1092 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1100 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1102 if (td->o.trim_backlog && td->trim_entries) {
1105 if (td->trim_batch) {
1108 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1109 td->last_ddir != DDIR_READ) {
1110 td->trim_batch = td->o.trim_batch;
1111 if (!td->trim_batch)
1112 td->trim_batch = td->o.trim_backlog;
1116 if (get_trim && !get_next_trim(td, io_u))
1123 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1125 if (td->o.verify_backlog && td->io_hist_len) {
1128 if (td->verify_batch)
1130 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1131 td->last_ddir != DDIR_READ) {
1132 td->verify_batch = td->o.verify_batch;
1133 if (!td->verify_batch)
1134 td->verify_batch = td->o.verify_backlog;
1138 if (get_verify && !get_next_verify(td, io_u)) {
1148 * Fill offset and start time into the buffer content, to prevent too
1149 * easy compressible data for simple de-dupe attempts. Do this for every
1150 * 512b block in the range, since that should be the smallest block size
1151 * we can expect from a device.
1153 static void small_content_scramble(struct io_u *io_u)
1155 unsigned int i, nr_blocks = io_u->buflen / 512;
1156 unsigned long long boffset;
1157 unsigned int offset;
1164 boffset = io_u->offset;
1165 io_u->buf_filled_len = 0;
1167 for (i = 0; i < nr_blocks; i++) {
1169 * Fill the byte offset into a "random" start offset of
1170 * the buffer, given by the product of the usec time
1171 * and the actual offset.
1173 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1174 offset &= ~(sizeof(unsigned long long) - 1);
1175 if (offset >= 512 - sizeof(unsigned long long))
1176 offset -= sizeof(unsigned long long);
1177 memcpy(p + offset, &boffset, sizeof(boffset));
1179 end = p + 512 - sizeof(io_u->start_time);
1180 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1187 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1188 * etc. The returned io_u is fully ready to be prepped and submitted.
1190 struct io_u *get_io_u(struct thread_data *td)
1194 int do_scramble = 0;
1196 io_u = __get_io_u(td);
1198 dprint(FD_IO, "__get_io_u failed\n");
1202 if (check_get_verify(td, io_u))
1204 if (check_get_trim(td, io_u))
1208 * from a requeue, io_u already setup
1214 * If using an iolog, grab next piece if any available.
1216 if (td->o.read_iolog_file) {
1217 if (read_iolog_get(td, io_u))
1219 } else if (set_io_u_file(td, io_u)) {
1220 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1225 assert(fio_file_open(f));
1227 if (ddir_rw(io_u->ddir)) {
1228 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1229 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1233 f->last_start = io_u->offset;
1234 f->last_pos = io_u->offset + io_u->buflen;
1236 if (io_u->ddir == DDIR_WRITE) {
1237 if (td->o.refill_buffers) {
1238 io_u_fill_buffer(td, io_u,
1239 io_u->xfer_buflen, io_u->xfer_buflen);
1240 } else if (td->o.scramble_buffers)
1242 if (td->o.verify != VERIFY_NONE) {
1243 populate_verify_io_u(td, io_u);
1246 } else if (io_u->ddir == DDIR_READ) {
1248 * Reset the buf_filled parameters so next time if the
1249 * buffer is used for writes it is refilled.
1251 io_u->buf_filled_len = 0;
1256 * Set io data pointers.
1258 io_u->xfer_buf = io_u->buf;
1259 io_u->xfer_buflen = io_u->buflen;
1263 if (!td_io_prep(td, io_u)) {
1264 if (!td->o.disable_slat)
1265 fio_gettime(&io_u->start_time, NULL);
1267 small_content_scramble(io_u);
1271 dprint(FD_IO, "get_io_u failed\n");
1276 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1278 const char *msg[] = { "read", "write", "sync", "datasync",
1279 "sync_file_range", "wait", "trim" };
1283 log_err("fio: io_u error");
1286 log_err(" on file %s", io_u->file->file_name);
1288 log_err(": %s\n", strerror(io_u->error));
1290 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1291 io_u->offset, io_u->xfer_buflen);
1294 td_verror(td, io_u->error, "io_u error");
1297 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1298 struct io_completion_data *icd,
1299 const enum fio_ddir idx, unsigned int bytes)
1301 unsigned long uninitialized_var(lusec);
1303 if (!td->o.disable_clat || !td->o.disable_bw)
1304 lusec = utime_since(&io_u->issue_time, &icd->time);
1306 if (!td->o.disable_lat) {
1307 unsigned long tusec;
1309 tusec = utime_since(&io_u->start_time, &icd->time);
1310 add_lat_sample(td, idx, tusec, bytes);
1313 if (!td->o.disable_clat) {
1314 add_clat_sample(td, idx, lusec, bytes);
1315 io_u_mark_latency(td, lusec);
1318 if (!td->o.disable_bw)
1319 add_bw_sample(td, idx, bytes, &icd->time);
1321 add_iops_sample(td, idx, &icd->time);
1324 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1326 unsigned long long secs, remainder, bps, bytes;
1327 bytes = td->this_io_bytes[ddir];
1328 bps = td->rate_bps[ddir];
1330 remainder = bytes % bps;
1331 return remainder * 1000000 / bps + secs * 1000000;
1334 static void io_completed(struct thread_data *td, struct io_u *io_u,
1335 struct io_completion_data *icd)
1338 * Older gcc's are too dumb to realize that usec is always used
1339 * initialized, silence that warning.
1341 unsigned long uninitialized_var(usec);
1344 dprint_io_u(io_u, "io complete");
1347 assert(io_u->flags & IO_U_F_FLIGHT);
1348 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1351 if (ddir_sync(io_u->ddir)) {
1352 td->last_was_sync = 1;
1355 f->first_write = -1ULL;
1356 f->last_write = -1ULL;
1361 td->last_was_sync = 0;
1362 td->last_ddir = io_u->ddir;
1364 if (!io_u->error && ddir_rw(io_u->ddir)) {
1365 unsigned int bytes = io_u->buflen - io_u->resid;
1366 const enum fio_ddir idx = io_u->ddir;
1367 const enum fio_ddir odx = io_u->ddir ^ 1;
1370 td->io_blocks[idx]++;
1371 td->this_io_blocks[idx]++;
1372 td->io_bytes[idx] += bytes;
1373 td->this_io_bytes[idx] += bytes;
1375 if (idx == DDIR_WRITE) {
1378 if (f->first_write == -1ULL ||
1379 io_u->offset < f->first_write)
1380 f->first_write = io_u->offset;
1381 if (f->last_write == -1ULL ||
1382 ((io_u->offset + bytes) > f->last_write))
1383 f->last_write = io_u->offset + bytes;
1387 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1388 td->runstate == TD_VERIFYING)) {
1389 account_io_completion(td, io_u, icd, idx, bytes);
1391 if (__should_check_rate(td, idx)) {
1392 td->rate_pending_usleep[idx] =
1393 (usec_for_io(td, idx) -
1394 utime_since_now(&td->start));
1396 if (__should_check_rate(td, odx))
1397 td->rate_pending_usleep[odx] =
1398 (usec_for_io(td, odx) -
1399 utime_since_now(&td->start));
1402 if (td_write(td) && idx == DDIR_WRITE &&
1404 td->o.verify != VERIFY_NONE)
1405 log_io_piece(td, io_u);
1407 icd->bytes_done[idx] += bytes;
1410 ret = io_u->end_io(td, io_u);
1411 if (ret && !icd->error)
1414 } else if (io_u->error) {
1415 icd->error = io_u->error;
1416 io_u_log_error(td, io_u);
1418 if (icd->error && td_non_fatal_error(icd->error) &&
1419 (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1421 * If there is a non_fatal error, then add to the error count
1422 * and clear all the errors.
1424 update_error_count(td, icd->error);
1431 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1434 if (!td->o.disable_clat || !td->o.disable_bw)
1435 fio_gettime(&icd->time, NULL);
1440 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1443 static void ios_completed(struct thread_data *td,
1444 struct io_completion_data *icd)
1449 for (i = 0; i < icd->nr; i++) {
1450 io_u = td->io_ops->event(td, i);
1452 io_completed(td, io_u, icd);
1454 if (!(io_u->flags & IO_U_F_FREE_DEF))
1460 * Complete a single io_u for the sync engines.
1462 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1463 unsigned long *bytes)
1465 struct io_completion_data icd;
1467 init_icd(td, &icd, 1);
1468 io_completed(td, io_u, &icd);
1470 if (!(io_u->flags & IO_U_F_FREE_DEF))
1474 td_verror(td, icd.error, "io_u_sync_complete");
1479 bytes[0] += icd.bytes_done[0];
1480 bytes[1] += icd.bytes_done[1];
1487 * Called to complete min_events number of io for the async engines.
1489 int io_u_queued_complete(struct thread_data *td, int min_evts,
1490 unsigned long *bytes)
1492 struct io_completion_data icd;
1493 struct timespec *tvp = NULL;
1495 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1497 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1502 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1504 td_verror(td, -ret, "td_io_getevents");
1509 init_icd(td, &icd, ret);
1510 ios_completed(td, &icd);
1512 td_verror(td, icd.error, "io_u_queued_complete");
1517 bytes[0] += icd.bytes_done[0];
1518 bytes[1] += icd.bytes_done[1];
1525 * Call when io_u is really queued, to update the submission latency.
1527 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1529 if (!td->o.disable_slat) {
1530 unsigned long slat_time;
1532 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1533 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1538 * "randomly" fill the buffer contents
1540 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1541 unsigned int min_write, unsigned int max_bs)
1543 io_u->buf_filled_len = 0;
1545 if (!td->o.zero_buffers) {
1546 unsigned int perc = td->o.compress_percentage;
1549 unsigned int seg = min_write;
1551 seg = min(min_write, td->o.compress_chunk);
1552 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1555 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1557 memset(io_u->buf, 0, max_bs);