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 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
124 * Return the next free block in the map.
126 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
127 enum fio_ddir ddir, unsigned long long *b)
129 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
132 lastb = last_block(td, f, ddir);
136 i = f->last_free_lookup;
137 block = i * BLOCKS_PER_MAP;
138 while (block * min_bs < f->real_file_size &&
139 block * min_bs < f->io_size) {
140 if (f->file_map[i] != -1UL) {
141 block += ffz(f->file_map[i]);
144 f->last_free_lookup = i;
149 block += BLOCKS_PER_MAP;
153 dprint(FD_IO, "failed finding a free block\n");
157 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
158 enum fio_ddir ddir, unsigned long long *b)
160 unsigned long long r, lastb;
163 lastb = last_block(td, f, ddir);
167 if (f->failed_rands >= 200)
171 if (td->o.use_os_rand) {
172 r = os_random_long(&td->random_state);
173 *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
175 r = __rand(&td->__random_state);
176 *b = (lastb - 1) * (r / ((unsigned long long) FRAND_MAX + 1.0));
179 dprint(FD_RANDOM, "off rand %llu\n", r);
183 * if we are not maintaining a random map, we are done.
185 if (!file_randommap(td, f))
189 * calculate map offset and check if it's free
191 if (random_map_free(f, *b))
194 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
198 if (!f->failed_rands++)
199 f->last_free_lookup = 0;
202 * we get here, if we didn't suceed in looking up a block. generate
203 * a random start offset into the filemap, and find the first free
208 f->last_free_lookup = (f->num_maps - 1) *
209 (r / (OS_RAND_MAX + 1.0));
210 if (!get_next_free_block(td, f, ddir, b))
213 if (td->o.use_os_rand)
214 r = os_random_long(&td->random_state);
216 r = __rand(&td->__random_state);
220 * that didn't work either, try exhaustive search from the start
222 f->last_free_lookup = 0;
224 if (!get_next_free_block(td, f, ddir, b))
226 f->last_free_lookup = 0;
227 return get_next_free_block(td, f, ddir, b);
234 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
235 enum fio_ddir ddir, unsigned long long *b)
237 if (get_next_rand_offset(td, f, ddir, b)) {
238 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
239 f->file_name, f->last_pos, f->real_file_size);
246 static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
247 enum fio_ddir ddir, unsigned long long *b)
249 assert(ddir_rw(ddir));
251 if (f->last_pos < f->real_file_size) {
252 unsigned long long pos = f->last_pos - f->file_offset;
255 pos += td->o.ddir_seq_add;
257 *b = pos / td->o.min_bs[ddir];
264 static int get_next_block(struct thread_data *td, struct io_u *io_u,
265 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
267 struct fio_file *f = io_u->file;
270 assert(ddir_rw(ddir));
274 ret = get_next_rand_block(td, f, ddir, b);
276 ret = get_next_seq_block(td, f, ddir, b);
278 io_u->flags |= IO_U_F_BUSY_OK;
280 if (td->o.rw_seq == RW_SEQ_SEQ) {
281 ret = get_next_seq_block(td, f, ddir, b);
283 ret = get_next_rand_block(td, f, ddir, b);
284 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
285 if (f->last_start != -1ULL)
286 *b = (f->last_start - f->file_offset)
287 / td->o.min_bs[ddir];
292 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
301 * For random io, generate a random new block and see if it's used. Repeat
302 * until we find a free one. For sequential io, just return the end of
303 * the last io issued.
305 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
307 struct fio_file *f = io_u->file;
308 unsigned long long b;
309 enum fio_ddir ddir = io_u->ddir;
312 assert(ddir_rw(ddir));
314 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
316 td->ddir_seq_nr = td->o.ddir_seq_nr;
319 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
322 io_u->offset = b * td->o.ba[ddir];
323 if (io_u->offset >= f->io_size) {
324 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
325 io_u->offset, f->io_size);
329 io_u->offset += f->file_offset;
330 if (io_u->offset >= f->real_file_size) {
331 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
332 io_u->offset, f->real_file_size);
339 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
341 struct prof_io_ops *ops = &td->prof_io_ops;
343 if (ops->fill_io_u_off)
344 return ops->fill_io_u_off(td, io_u);
346 return __get_next_offset(td, io_u);
349 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
352 struct fio_file *f = io_u->file;
354 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
357 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
359 const int ddir = io_u->ddir;
360 unsigned int uninitialized_var(buflen);
361 unsigned int minbs, maxbs;
362 unsigned long r, rand_max;
364 assert(ddir_rw(ddir));
366 minbs = td->o.min_bs[ddir];
367 maxbs = td->o.max_bs[ddir];
372 if (td->o.use_os_rand)
373 rand_max = OS_RAND_MAX;
375 rand_max = FRAND_MAX;
378 if (td->o.use_os_rand)
379 r = os_random_long(&td->bsrange_state);
381 r = __rand(&td->__bsrange_state);
383 if (!td->o.bssplit_nr[ddir]) {
384 buflen = 1 + (unsigned int) ((double) maxbs *
385 (r / (rand_max + 1.0)));
392 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
393 struct bssplit *bsp = &td->o.bssplit[ddir][i];
397 if ((r <= ((rand_max / 100L) * perc)) &&
398 io_u_fits(td, io_u, buflen))
403 if (!td->o.bs_unaligned && is_power_of_2(minbs))
404 buflen = (buflen + minbs - 1) & ~(minbs - 1);
406 } while (!io_u_fits(td, io_u, buflen));
411 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
413 struct prof_io_ops *ops = &td->prof_io_ops;
415 if (ops->fill_io_u_size)
416 return ops->fill_io_u_size(td, io_u);
418 return __get_next_buflen(td, io_u);
421 static void set_rwmix_bytes(struct thread_data *td)
426 * we do time or byte based switch. this is needed because
427 * buffered writes may issue a lot quicker than they complete,
428 * whereas reads do not.
430 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
431 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
434 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
439 if (td->o.use_os_rand) {
440 r = os_random_long(&td->rwmix_state);
441 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
443 r = __rand(&td->__rwmix_state);
444 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
447 if (v <= td->o.rwmix[DDIR_READ])
453 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
455 enum fio_ddir odir = ddir ^ 1;
459 assert(ddir_rw(ddir));
461 if (td->rate_pending_usleep[ddir] <= 0)
465 * We have too much pending sleep in this direction. See if we
470 * Other direction does not have too much pending, switch
472 if (td->rate_pending_usleep[odir] < 100000)
476 * Both directions have pending sleep. Sleep the minimum time
477 * and deduct from both.
479 if (td->rate_pending_usleep[ddir] <=
480 td->rate_pending_usleep[odir]) {
481 usec = td->rate_pending_usleep[ddir];
483 usec = td->rate_pending_usleep[odir];
487 usec = td->rate_pending_usleep[ddir];
490 * We are going to sleep, ensure that we flush anything pending as
491 * not to skew our latency numbers
496 ret = io_u_queued_complete(td, td->cur_depth, NULL);
499 fio_gettime(&t, NULL);
500 usec_sleep(td, usec);
501 usec = utime_since_now(&t);
503 td->rate_pending_usleep[ddir] -= usec;
506 if (td_rw(td) && __should_check_rate(td, odir))
507 td->rate_pending_usleep[odir] -= usec;
513 * Return the data direction for the next io_u. If the job is a
514 * mixed read/write workload, check the rwmix cycle and switch if
517 static enum fio_ddir get_rw_ddir(struct thread_data *td)
522 * see if it's time to fsync
524 if (td->o.fsync_blocks &&
525 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
526 td->io_issues[DDIR_WRITE] && should_fsync(td))
530 * see if it's time to fdatasync
532 if (td->o.fdatasync_blocks &&
533 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
534 td->io_issues[DDIR_WRITE] && should_fsync(td))
535 return DDIR_DATASYNC;
538 * see if it's time to sync_file_range
540 if (td->sync_file_range_nr &&
541 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
542 td->io_issues[DDIR_WRITE] && should_fsync(td))
543 return DDIR_SYNC_FILE_RANGE;
547 * Check if it's time to seed a new data direction.
549 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
551 * Put a top limit on how many bytes we do for
552 * one data direction, to avoid overflowing the
555 ddir = get_rand_ddir(td);
557 if (ddir != td->rwmix_ddir)
560 td->rwmix_ddir = ddir;
562 ddir = td->rwmix_ddir;
563 } else if (td_read(td))
568 td->rwmix_ddir = rate_ddir(td, ddir);
569 return td->rwmix_ddir;
572 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
574 io_u->ddir = get_rw_ddir(td);
576 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
577 td->o.barrier_blocks &&
578 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
579 td->io_issues[DDIR_WRITE])
580 io_u->flags |= IO_U_F_BARRIER;
583 void put_file_log(struct thread_data *td, struct fio_file *f)
585 int ret = put_file(td, f);
588 td_verror(td, ret, "file close");
591 void put_io_u(struct thread_data *td, struct io_u *io_u)
595 io_u->flags |= IO_U_F_FREE;
596 io_u->flags &= ~IO_U_F_FREE_DEF;
599 put_file_log(td, io_u->file);
602 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
604 flist_del_init(&io_u->list);
605 flist_add(&io_u->list, &td->io_u_freelist);
607 td_io_u_free_notify(td);
610 void clear_io_u(struct thread_data *td, struct io_u *io_u)
612 io_u->flags &= ~IO_U_F_FLIGHT;
616 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
618 struct io_u *__io_u = *io_u;
620 dprint(FD_IO, "requeue %p\n", __io_u);
624 __io_u->flags |= IO_U_F_FREE;
625 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
626 td->io_issues[__io_u->ddir]--;
628 __io_u->flags &= ~IO_U_F_FLIGHT;
629 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
631 flist_del(&__io_u->list);
632 flist_add_tail(&__io_u->list, &td->io_u_requeues);
637 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
639 if (td->io_ops->flags & FIO_NOIO)
642 set_rw_ddir(td, io_u);
645 * fsync() or fdatasync() or trim etc, we are done
647 if (!ddir_rw(io_u->ddir))
651 * See if it's time to switch to a new zone
653 if (td->zone_bytes >= td->o.zone_size) {
655 io_u->file->last_pos += td->o.zone_skip;
656 td->io_skip_bytes += td->o.zone_skip;
660 * No log, let the seq/rand engine retrieve the next buflen and
663 if (get_next_offset(td, io_u)) {
664 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
668 io_u->buflen = get_next_buflen(td, io_u);
670 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
674 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
675 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
676 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
677 io_u->buflen, io_u->file->real_file_size);
682 * mark entry before potentially trimming io_u
684 if (td_random(td) && file_randommap(td, io_u->file))
685 mark_random_map(td, io_u);
688 * If using a write iolog, store this entry.
691 dprint_io_u(io_u, "fill_io_u");
692 td->zone_bytes += io_u->buflen;
697 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
726 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
728 __io_u_mark_map(td->ts.io_u_submit, nr);
729 td->ts.total_submit++;
732 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
734 __io_u_mark_map(td->ts.io_u_complete, nr);
735 td->ts.total_complete++;
738 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
742 switch (td->cur_depth) {
764 td->ts.io_u_map[idx] += nr;
767 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
804 assert(idx < FIO_IO_U_LAT_U_NR);
805 td->ts.io_u_lat_u[idx]++;
808 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
849 assert(idx < FIO_IO_U_LAT_M_NR);
850 td->ts.io_u_lat_m[idx]++;
853 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
856 io_u_mark_lat_usec(td, usec);
858 io_u_mark_lat_msec(td, usec / 1000);
862 * Get next file to service by choosing one at random
864 static struct fio_file *get_next_file_rand(struct thread_data *td,
865 enum fio_file_flags goodf,
866 enum fio_file_flags badf)
875 if (td->o.use_os_rand) {
876 r = os_random_long(&td->next_file_state);
877 fno = (unsigned int) ((double) td->o.nr_files
878 * (r / (OS_RAND_MAX + 1.0)));
880 r = __rand(&td->__next_file_state);
881 fno = (unsigned int) ((double) td->o.nr_files
882 * (r / (FRAND_MAX + 1.0)));
886 if (fio_file_done(f))
889 if (!fio_file_open(f)) {
892 err = td_io_open_file(td, f);
898 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
899 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
903 td_io_close_file(td, f);
908 * Get next file to service by doing round robin between all available ones
910 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
913 unsigned int old_next_file = td->next_file;
919 f = td->files[td->next_file];
922 if (td->next_file >= td->o.nr_files)
925 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
926 if (fio_file_done(f)) {
931 if (!fio_file_open(f)) {
934 err = td_io_open_file(td, f);
936 dprint(FD_FILE, "error %d on open of %s\n",
944 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
946 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
950 td_io_close_file(td, f);
953 } while (td->next_file != old_next_file);
955 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
959 static struct fio_file *__get_next_file(struct thread_data *td)
963 assert(td->o.nr_files <= td->files_index);
965 if (td->nr_done_files >= td->o.nr_files) {
966 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
967 " nr_files=%d\n", td->nr_open_files,
973 f = td->file_service_file;
974 if (f && fio_file_open(f) && !fio_file_closing(f)) {
975 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
977 if (td->file_service_left--)
981 if (td->o.file_service_type == FIO_FSERVICE_RR ||
982 td->o.file_service_type == FIO_FSERVICE_SEQ)
983 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
985 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
987 td->file_service_file = f;
988 td->file_service_left = td->file_service_nr - 1;
990 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
994 static struct fio_file *get_next_file(struct thread_data *td)
996 struct prof_io_ops *ops = &td->prof_io_ops;
998 if (ops->get_next_file)
999 return ops->get_next_file(td);
1001 return __get_next_file(td);
1004 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1009 f = get_next_file(td);
1016 if (!fill_io_u(td, io_u))
1019 put_file_log(td, f);
1020 td_io_close_file(td, f);
1022 fio_file_set_done(f);
1023 td->nr_done_files++;
1024 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1025 td->nr_done_files, td->o.nr_files);
1032 struct io_u *__get_io_u(struct thread_data *td)
1034 struct io_u *io_u = NULL;
1039 if (!flist_empty(&td->io_u_requeues))
1040 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1041 else if (!queue_full(td)) {
1042 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1047 io_u->end_io = NULL;
1051 assert(io_u->flags & IO_U_F_FREE);
1052 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1053 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1056 flist_del(&io_u->list);
1057 flist_add(&io_u->list, &td->io_u_busylist);
1059 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1060 } else if (td->o.verify_async) {
1062 * We ran out, wait for async verify threads to finish and
1065 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1073 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1075 if (td->o.trim_backlog && td->trim_entries) {
1078 if (td->trim_batch) {
1081 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1082 td->last_ddir != DDIR_READ) {
1083 td->trim_batch = td->o.trim_batch;
1084 if (!td->trim_batch)
1085 td->trim_batch = td->o.trim_backlog;
1089 if (get_trim && !get_next_trim(td, io_u))
1096 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1098 if (td->o.verify_backlog && td->io_hist_len) {
1101 if (td->verify_batch) {
1104 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1105 td->last_ddir != DDIR_READ) {
1106 td->verify_batch = td->o.verify_batch;
1107 if (!td->verify_batch)
1108 td->verify_batch = td->o.verify_backlog;
1112 if (get_verify && !get_next_verify(td, io_u))
1120 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1121 * etc. The returned io_u is fully ready to be prepped and submitted.
1123 struct io_u *get_io_u(struct thread_data *td)
1128 io_u = __get_io_u(td);
1130 dprint(FD_IO, "__get_io_u failed\n");
1134 if (check_get_verify(td, io_u))
1136 if (check_get_trim(td, io_u))
1140 * from a requeue, io_u already setup
1146 * If using an iolog, grab next piece if any available.
1148 if (td->o.read_iolog_file) {
1149 if (read_iolog_get(td, io_u))
1151 } else if (set_io_u_file(td, io_u)) {
1152 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1157 assert(fio_file_open(f));
1159 if (ddir_rw(io_u->ddir)) {
1160 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1161 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1165 f->last_start = io_u->offset;
1166 f->last_pos = io_u->offset + io_u->buflen;
1168 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1169 populate_verify_io_u(td, io_u);
1170 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1171 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1172 else if (io_u->ddir == DDIR_READ) {
1174 * Reset the buf_filled parameters so next time if the
1175 * buffer is used for writes it is refilled.
1177 io_u->buf_filled_len = 0;
1182 * Set io data pointers.
1184 io_u->xfer_buf = io_u->buf;
1185 io_u->xfer_buflen = io_u->buflen;
1189 if (!td_io_prep(td, io_u)) {
1190 if (!td->o.disable_slat)
1191 fio_gettime(&io_u->start_time, NULL);
1195 dprint(FD_IO, "get_io_u failed\n");
1200 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1202 const char *msg[] = { "read", "write", "sync", "datasync",
1203 "sync_file_range", "wait", "trim" };
1207 log_err("fio: io_u error");
1210 log_err(" on file %s", io_u->file->file_name);
1212 log_err(": %s\n", strerror(io_u->error));
1214 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1215 io_u->offset, io_u->xfer_buflen);
1218 td_verror(td, io_u->error, "io_u error");
1221 static void io_completed(struct thread_data *td, struct io_u *io_u,
1222 struct io_completion_data *icd)
1225 * Older gcc's are too dumb to realize that usec is always used
1226 * initialized, silence that warning.
1228 unsigned long uninitialized_var(usec);
1231 dprint_io_u(io_u, "io complete");
1234 assert(io_u->flags & IO_U_F_FLIGHT);
1235 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1238 if (ddir_sync(io_u->ddir)) {
1239 td->last_was_sync = 1;
1242 f->first_write = -1ULL;
1243 f->last_write = -1ULL;
1248 td->last_was_sync = 0;
1249 td->last_ddir = io_u->ddir;
1251 if (!io_u->error && ddir_rw(io_u->ddir)) {
1252 unsigned int bytes = io_u->buflen - io_u->resid;
1253 const enum fio_ddir idx = io_u->ddir;
1254 const enum fio_ddir odx = io_u->ddir ^ 1;
1257 td->io_blocks[idx]++;
1258 td->io_bytes[idx] += bytes;
1259 td->this_io_bytes[idx] += bytes;
1261 if (idx == DDIR_WRITE) {
1264 if (f->first_write == -1ULL ||
1265 io_u->offset < f->first_write)
1266 f->first_write = io_u->offset;
1267 if (f->last_write == -1ULL ||
1268 ((io_u->offset + bytes) > f->last_write))
1269 f->last_write = io_u->offset + bytes;
1273 if (ramp_time_over(td)) {
1274 unsigned long uninitialized_var(lusec);
1276 if (!td->o.disable_clat || !td->o.disable_bw)
1277 lusec = utime_since(&io_u->issue_time,
1279 if (!td->o.disable_lat) {
1280 unsigned long tusec;
1282 tusec = utime_since(&io_u->start_time,
1284 add_lat_sample(td, idx, tusec, bytes);
1286 if (!td->o.disable_clat) {
1287 add_clat_sample(td, idx, lusec, bytes);
1288 io_u_mark_latency(td, lusec);
1290 if (!td->o.disable_bw)
1291 add_bw_sample(td, idx, bytes, &icd->time);
1292 if (__should_check_rate(td, idx)) {
1293 td->rate_pending_usleep[idx] =
1294 ((td->this_io_bytes[idx] *
1295 td->rate_nsec_cycle[idx]) / 1000 -
1296 utime_since_now(&td->start));
1298 if (__should_check_rate(td, idx ^ 1))
1299 td->rate_pending_usleep[odx] =
1300 ((td->this_io_bytes[odx] *
1301 td->rate_nsec_cycle[odx]) / 1000 -
1302 utime_since_now(&td->start));
1305 if (td_write(td) && idx == DDIR_WRITE &&
1307 td->o.verify != VERIFY_NONE)
1308 log_io_piece(td, io_u);
1310 icd->bytes_done[idx] += bytes;
1313 ret = io_u->end_io(td, io_u);
1314 if (ret && !icd->error)
1317 } else if (io_u->error) {
1318 icd->error = io_u->error;
1319 io_u_log_error(td, io_u);
1321 if (td->o.continue_on_error && icd->error &&
1322 td_non_fatal_error(icd->error)) {
1324 * If there is a non_fatal error, then add to the error count
1325 * and clear all the errors.
1327 update_error_count(td, icd->error);
1334 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1337 if (!td->o.disable_clat || !td->o.disable_bw)
1338 fio_gettime(&icd->time, NULL);
1343 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1346 static void ios_completed(struct thread_data *td,
1347 struct io_completion_data *icd)
1352 for (i = 0; i < icd->nr; i++) {
1353 io_u = td->io_ops->event(td, i);
1355 io_completed(td, io_u, icd);
1357 if (!(io_u->flags & IO_U_F_FREE_DEF))
1363 * Complete a single io_u for the sync engines.
1365 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1366 unsigned long *bytes)
1368 struct io_completion_data icd;
1370 init_icd(td, &icd, 1);
1371 io_completed(td, io_u, &icd);
1373 if (!(io_u->flags & IO_U_F_FREE_DEF))
1377 td_verror(td, icd.error, "io_u_sync_complete");
1382 bytes[0] += icd.bytes_done[0];
1383 bytes[1] += icd.bytes_done[1];
1390 * Called to complete min_events number of io for the async engines.
1392 int io_u_queued_complete(struct thread_data *td, int min_evts,
1393 unsigned long *bytes)
1395 struct io_completion_data icd;
1396 struct timespec *tvp = NULL;
1398 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1400 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1405 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1407 td_verror(td, -ret, "td_io_getevents");
1412 init_icd(td, &icd, ret);
1413 ios_completed(td, &icd);
1415 td_verror(td, icd.error, "io_u_queued_complete");
1420 bytes[0] += icd.bytes_done[0];
1421 bytes[1] += icd.bytes_done[1];
1428 * Call when io_u is really queued, to update the submission latency.
1430 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1432 if (!td->o.disable_slat) {
1433 unsigned long slat_time;
1435 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1436 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1441 * "randomly" fill the buffer contents
1443 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1444 unsigned int max_bs)
1446 io_u->buf_filled_len = 0;
1448 if (!td->o.zero_buffers)
1449 fill_random_buf(io_u->buf, max_bs);
1451 memset(io_u->buf, 0, max_bs);