14 struct io_completion_data {
16 int account; /* input */
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 static int random_map_free(struct fio_file *f, const unsigned long long block)
29 unsigned int idx = RAND_MAP_IDX(f, block);
30 unsigned int bit = RAND_MAP_BIT(f, block);
32 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
34 return (f->file_map[idx] & (1UL << bit)) == 0;
38 * Mark a given offset as used in the map.
40 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
42 unsigned int min_bs = td->o.rw_min_bs;
43 struct fio_file *f = io_u->file;
44 unsigned long long block;
45 unsigned int blocks, nr_blocks;
48 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
49 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
51 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
54 unsigned int idx, bit;
55 unsigned long mask, this_blocks;
58 * If we have a mixed random workload, we may
59 * encounter blocks we already did IO to.
65 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
68 idx = RAND_MAP_IDX(f, block);
69 bit = RAND_MAP_BIT(f, block);
71 fio_assert(td, idx < f->num_maps);
73 this_blocks = nr_blocks;
74 if (this_blocks + bit > BLOCKS_PER_MAP)
75 this_blocks = BLOCKS_PER_MAP - bit;
78 if (this_blocks == BLOCKS_PER_MAP)
81 mask = ((1UL << this_blocks) - 1) << bit;
83 if (!(f->file_map[idx] & mask))
87 } while (this_blocks);
92 f->file_map[idx] |= mask;
93 nr_blocks -= this_blocks;
94 blocks += this_blocks;
98 if ((blocks * min_bs) < io_u->buflen)
99 io_u->buflen = blocks * min_bs;
102 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
105 unsigned long long max_blocks;
106 unsigned long long max_size;
108 assert(ddir_rw(ddir));
111 * Hmm, should we make sure that ->io_size <= ->real_file_size?
113 max_size = f->io_size;
114 if (max_size > f->real_file_size)
115 max_size = f->real_file_size;
117 if (td->o.zone_range)
118 max_size = td->o.zone_range;
120 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
128 * Return the next free block in the map.
130 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
131 enum fio_ddir ddir, unsigned long long *b)
133 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
136 lastb = last_block(td, f, ddir);
140 i = f->last_free_lookup;
141 block = i * BLOCKS_PER_MAP;
142 while (block * min_bs < f->real_file_size &&
143 block * min_bs < f->io_size) {
144 if (f->file_map[i] != -1UL) {
145 block += ffz(f->file_map[i]);
148 f->last_free_lookup = i;
153 block += BLOCKS_PER_MAP;
157 dprint(FD_IO, "failed finding a free block\n");
161 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
162 enum fio_ddir ddir, unsigned long long *b)
164 unsigned long long rmax, r, lastb;
167 lastb = last_block(td, f, ddir);
171 if (f->failed_rands >= 200)
174 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
176 if (td->o.use_os_rand)
177 r = os_random_long(&td->random_state);
179 r = __rand(&td->__random_state);
181 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
183 dprint(FD_RANDOM, "off rand %llu\n", r);
187 * if we are not maintaining a random map, we are done.
189 if (!file_randommap(td, f))
193 * calculate map offset and check if it's free
195 if (random_map_free(f, *b))
198 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
202 if (!f->failed_rands++)
203 f->last_free_lookup = 0;
206 * we get here, if we didn't suceed in looking up a block. generate
207 * a random start offset into the filemap, and find the first free
212 f->last_free_lookup = (f->num_maps - 1) *
213 (r / ((unsigned long long) rmax + 1.0));
214 if (!get_next_free_block(td, f, ddir, b))
217 if (td->o.use_os_rand)
218 r = os_random_long(&td->random_state);
220 r = __rand(&td->__random_state);
224 * that didn't work either, try exhaustive search from the start
226 f->last_free_lookup = 0;
228 if (!get_next_free_block(td, f, ddir, b))
230 f->last_free_lookup = 0;
231 return get_next_free_block(td, f, ddir, b);
238 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
239 enum fio_ddir ddir, unsigned long long *b)
241 if (get_next_rand_offset(td, f, ddir, b)) {
242 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
243 f->file_name, f->last_pos, f->real_file_size);
250 static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
251 enum fio_ddir ddir, unsigned long long *b)
253 assert(ddir_rw(ddir));
255 if (f->last_pos < f->real_file_size) {
256 unsigned long long pos;
258 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
259 f->last_pos = f->real_file_size;
261 pos = f->last_pos - f->file_offset;
263 pos += td->o.ddir_seq_add;
265 *b = pos / td->o.min_bs[ddir];
272 static int get_next_block(struct thread_data *td, struct io_u *io_u,
273 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
275 struct fio_file *f = io_u->file;
278 assert(ddir_rw(ddir));
282 ret = get_next_rand_block(td, f, ddir, b);
284 ret = get_next_seq_block(td, f, ddir, b);
286 io_u->flags |= IO_U_F_BUSY_OK;
288 if (td->o.rw_seq == RW_SEQ_SEQ) {
289 ret = get_next_seq_block(td, f, ddir, b);
291 ret = get_next_rand_block(td, f, ddir, b);
292 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
293 if (f->last_start != -1ULL)
294 *b = (f->last_start - f->file_offset)
295 / td->o.min_bs[ddir];
300 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
309 * For random io, generate a random new block and see if it's used. Repeat
310 * until we find a free one. For sequential io, just return the end of
311 * the last io issued.
313 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
315 struct fio_file *f = io_u->file;
316 unsigned long long b;
317 enum fio_ddir ddir = io_u->ddir;
320 assert(ddir_rw(ddir));
322 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
324 td->ddir_seq_nr = td->o.ddir_seq_nr;
327 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
330 io_u->offset = b * td->o.ba[ddir];
331 if (io_u->offset >= f->io_size) {
332 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
333 io_u->offset, f->io_size);
337 io_u->offset += f->file_offset;
338 if (io_u->offset >= f->real_file_size) {
339 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
340 io_u->offset, f->real_file_size);
347 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
349 struct prof_io_ops *ops = &td->prof_io_ops;
351 if (ops->fill_io_u_off)
352 return ops->fill_io_u_off(td, io_u);
354 return __get_next_offset(td, io_u);
357 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
360 struct fio_file *f = io_u->file;
362 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
365 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
367 const int ddir = io_u->ddir;
368 unsigned int uninitialized_var(buflen);
369 unsigned int minbs, maxbs;
370 unsigned long r, rand_max;
372 assert(ddir_rw(ddir));
374 minbs = td->o.min_bs[ddir];
375 maxbs = td->o.max_bs[ddir];
380 if (td->o.use_os_rand)
381 rand_max = OS_RAND_MAX;
383 rand_max = FRAND_MAX;
386 if (td->o.use_os_rand)
387 r = os_random_long(&td->bsrange_state);
389 r = __rand(&td->__bsrange_state);
391 if (!td->o.bssplit_nr[ddir]) {
392 buflen = 1 + (unsigned int) ((double) maxbs *
393 (r / (rand_max + 1.0)));
400 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
401 struct bssplit *bsp = &td->o.bssplit[ddir][i];
405 if ((r <= ((rand_max / 100L) * perc)) &&
406 io_u_fits(td, io_u, buflen))
411 if (!td->o.bs_unaligned && is_power_of_2(minbs))
412 buflen = (buflen + minbs - 1) & ~(minbs - 1);
414 } while (!io_u_fits(td, io_u, buflen));
419 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
421 struct prof_io_ops *ops = &td->prof_io_ops;
423 if (ops->fill_io_u_size)
424 return ops->fill_io_u_size(td, io_u);
426 return __get_next_buflen(td, io_u);
429 static void set_rwmix_bytes(struct thread_data *td)
434 * we do time or byte based switch. this is needed because
435 * buffered writes may issue a lot quicker than they complete,
436 * whereas reads do not.
438 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
439 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
442 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
447 if (td->o.use_os_rand) {
448 r = os_random_long(&td->rwmix_state);
449 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
451 r = __rand(&td->__rwmix_state);
452 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
455 if (v <= td->o.rwmix[DDIR_READ])
461 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
463 enum fio_ddir odir = ddir ^ 1;
467 assert(ddir_rw(ddir));
469 if (td->rate_pending_usleep[ddir] <= 0)
473 * We have too much pending sleep in this direction. See if we
478 * Other direction does not have too much pending, switch
480 if (td->rate_pending_usleep[odir] < 100000)
484 * Both directions have pending sleep. Sleep the minimum time
485 * and deduct from both.
487 if (td->rate_pending_usleep[ddir] <=
488 td->rate_pending_usleep[odir]) {
489 usec = td->rate_pending_usleep[ddir];
491 usec = td->rate_pending_usleep[odir];
495 usec = td->rate_pending_usleep[ddir];
498 * We are going to sleep, ensure that we flush anything pending as
499 * not to skew our latency numbers.
501 * Changed to only monitor 'in flight' requests here instead of the
502 * td->cur_depth, b/c td->cur_depth does not accurately represent
503 * io's that have been actually submitted to an async engine,
504 * and cur_depth is meaningless for sync engines.
506 if (td->io_u_in_flight) {
509 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
512 fio_gettime(&t, NULL);
513 usec_sleep(td, usec);
514 usec = utime_since_now(&t);
516 td->rate_pending_usleep[ddir] -= usec;
519 if (td_rw(td) && __should_check_rate(td, odir))
520 td->rate_pending_usleep[odir] -= usec;
526 * Return the data direction for the next io_u. If the job is a
527 * mixed read/write workload, check the rwmix cycle and switch if
530 static enum fio_ddir get_rw_ddir(struct thread_data *td)
535 * see if it's time to fsync
537 if (td->o.fsync_blocks &&
538 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
539 td->io_issues[DDIR_WRITE] && should_fsync(td))
543 * see if it's time to fdatasync
545 if (td->o.fdatasync_blocks &&
546 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
547 td->io_issues[DDIR_WRITE] && should_fsync(td))
548 return DDIR_DATASYNC;
551 * see if it's time to sync_file_range
553 if (td->sync_file_range_nr &&
554 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
555 td->io_issues[DDIR_WRITE] && should_fsync(td))
556 return DDIR_SYNC_FILE_RANGE;
560 * Check if it's time to seed a new data direction.
562 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
564 * Put a top limit on how many bytes we do for
565 * one data direction, to avoid overflowing the
568 ddir = get_rand_ddir(td);
570 if (ddir != td->rwmix_ddir)
573 td->rwmix_ddir = ddir;
575 ddir = td->rwmix_ddir;
576 } else if (td_read(td))
581 td->rwmix_ddir = rate_ddir(td, ddir);
582 return td->rwmix_ddir;
585 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
587 io_u->ddir = get_rw_ddir(td);
589 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
590 td->o.barrier_blocks &&
591 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
592 td->io_issues[DDIR_WRITE])
593 io_u->flags |= IO_U_F_BARRIER;
596 void put_file_log(struct thread_data *td, struct fio_file *f)
598 int ret = put_file(td, f);
601 td_verror(td, ret, "file close");
604 void put_io_u(struct thread_data *td, struct io_u *io_u)
608 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
609 put_file_log(td, io_u->file);
611 io_u->flags &= ~IO_U_F_FREE_DEF;
612 io_u->flags |= IO_U_F_FREE;
614 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
616 flist_del_init(&io_u->list);
617 flist_add(&io_u->list, &td->io_u_freelist);
619 td_io_u_free_notify(td);
622 void clear_io_u(struct thread_data *td, struct io_u *io_u)
624 io_u->flags &= ~IO_U_F_FLIGHT;
628 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
630 struct io_u *__io_u = *io_u;
632 dprint(FD_IO, "requeue %p\n", __io_u);
636 __io_u->flags |= IO_U_F_FREE;
637 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
638 td->io_issues[__io_u->ddir]--;
640 __io_u->flags &= ~IO_U_F_FLIGHT;
641 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
643 flist_del(&__io_u->list);
644 flist_add_tail(&__io_u->list, &td->io_u_requeues);
649 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
651 if (td->io_ops->flags & FIO_NOIO)
654 set_rw_ddir(td, io_u);
657 * fsync() or fdatasync() or trim etc, we are done
659 if (!ddir_rw(io_u->ddir))
663 * See if it's time to switch to a new zone
665 if (td->zone_bytes >= td->o.zone_size) {
667 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
668 io_u->file->last_pos = io_u->file->file_offset;
669 td->io_skip_bytes += td->o.zone_skip;
673 * No log, let the seq/rand engine retrieve the next buflen and
676 if (get_next_offset(td, io_u)) {
677 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
681 io_u->buflen = get_next_buflen(td, io_u);
683 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
687 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
688 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
689 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
690 io_u->buflen, io_u->file->real_file_size);
695 * mark entry before potentially trimming io_u
697 if (td_random(td) && file_randommap(td, io_u->file))
698 mark_random_map(td, io_u);
701 * If using a write iolog, store this entry.
704 dprint_io_u(io_u, "fill_io_u");
705 td->zone_bytes += io_u->buflen;
710 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
739 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
741 __io_u_mark_map(td->ts.io_u_submit, nr);
742 td->ts.total_submit++;
745 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
747 __io_u_mark_map(td->ts.io_u_complete, nr);
748 td->ts.total_complete++;
751 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
755 switch (td->cur_depth) {
777 td->ts.io_u_map[idx] += nr;
780 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
817 assert(idx < FIO_IO_U_LAT_U_NR);
818 td->ts.io_u_lat_u[idx]++;
821 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
862 assert(idx < FIO_IO_U_LAT_M_NR);
863 td->ts.io_u_lat_m[idx]++;
866 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
869 io_u_mark_lat_usec(td, usec);
871 io_u_mark_lat_msec(td, usec / 1000);
875 * Get next file to service by choosing one at random
877 static struct fio_file *get_next_file_rand(struct thread_data *td,
878 enum fio_file_flags goodf,
879 enum fio_file_flags badf)
888 if (td->o.use_os_rand) {
889 r = os_random_long(&td->next_file_state);
890 fno = (unsigned int) ((double) td->o.nr_files
891 * (r / (OS_RAND_MAX + 1.0)));
893 r = __rand(&td->__next_file_state);
894 fno = (unsigned int) ((double) td->o.nr_files
895 * (r / (FRAND_MAX + 1.0)));
899 if (fio_file_done(f))
902 if (!fio_file_open(f)) {
905 err = td_io_open_file(td, f);
911 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
912 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
916 td_io_close_file(td, f);
921 * Get next file to service by doing round robin between all available ones
923 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
926 unsigned int old_next_file = td->next_file;
932 f = td->files[td->next_file];
935 if (td->next_file >= td->o.nr_files)
938 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
939 if (fio_file_done(f)) {
944 if (!fio_file_open(f)) {
947 err = td_io_open_file(td, f);
949 dprint(FD_FILE, "error %d on open of %s\n",
957 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
959 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
963 td_io_close_file(td, f);
966 } while (td->next_file != old_next_file);
968 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
972 static struct fio_file *__get_next_file(struct thread_data *td)
976 assert(td->o.nr_files <= td->files_index);
978 if (td->nr_done_files >= td->o.nr_files) {
979 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
980 " nr_files=%d\n", td->nr_open_files,
986 f = td->file_service_file;
987 if (f && fio_file_open(f) && !fio_file_closing(f)) {
988 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
990 if (td->file_service_left--)
994 if (td->o.file_service_type == FIO_FSERVICE_RR ||
995 td->o.file_service_type == FIO_FSERVICE_SEQ)
996 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
998 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1000 td->file_service_file = f;
1001 td->file_service_left = td->file_service_nr - 1;
1003 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1007 static struct fio_file *get_next_file(struct thread_data *td)
1009 struct prof_io_ops *ops = &td->prof_io_ops;
1011 if (ops->get_next_file)
1012 return ops->get_next_file(td);
1014 return __get_next_file(td);
1017 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1022 f = get_next_file(td);
1029 if (!fill_io_u(td, io_u))
1032 put_file_log(td, f);
1033 td_io_close_file(td, f);
1035 fio_file_set_done(f);
1036 td->nr_done_files++;
1037 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1038 td->nr_done_files, td->o.nr_files);
1045 struct io_u *__get_io_u(struct thread_data *td)
1047 struct io_u *io_u = NULL;
1052 if (!flist_empty(&td->io_u_requeues))
1053 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1054 else if (!queue_full(td)) {
1055 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1060 io_u->end_io = NULL;
1064 assert(io_u->flags & IO_U_F_FREE);
1065 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1066 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1069 flist_del(&io_u->list);
1070 flist_add(&io_u->list, &td->io_u_busylist);
1072 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1073 } else if (td->o.verify_async) {
1075 * We ran out, wait for async verify threads to finish and
1078 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1086 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1088 if (td->o.trim_backlog && td->trim_entries) {
1091 if (td->trim_batch) {
1094 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1095 td->last_ddir != DDIR_READ) {
1096 td->trim_batch = td->o.trim_batch;
1097 if (!td->trim_batch)
1098 td->trim_batch = td->o.trim_backlog;
1102 if (get_trim && !get_next_trim(td, io_u))
1109 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1111 if (td->o.verify_backlog && td->io_hist_len) {
1114 if (td->verify_batch) {
1117 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1118 td->last_ddir != DDIR_READ) {
1119 td->verify_batch = td->o.verify_batch;
1120 if (!td->verify_batch)
1121 td->verify_batch = td->o.verify_backlog;
1125 if (get_verify && !get_next_verify(td, io_u))
1133 * Fill offset and start time into the buffer content, to prevent too
1134 * easy compressible data for simple de-dupe attempts. Do this for every
1135 * 512b block in the range, since that should be the smallest block size
1136 * we can expect from a device.
1138 static void small_content_scramble(struct io_u *io_u)
1140 unsigned int i, nr_blocks = io_u->buflen / 512;
1141 unsigned long long boffset;
1142 unsigned int offset;
1149 boffset = io_u->offset;
1150 io_u->buf_filled_len = 0;
1152 for (i = 0; i < nr_blocks; i++) {
1154 * Fill the byte offset into a "random" start offset of
1155 * the buffer, given by the product of the usec time
1156 * and the actual offset.
1158 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1159 offset &= ~(sizeof(unsigned long long) - 1);
1160 if (offset >= 512 - sizeof(unsigned long long))
1161 offset -= sizeof(unsigned long long);
1162 memcpy(p + offset, &boffset, sizeof(boffset));
1164 end = p + 512 - sizeof(io_u->start_time);
1165 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1172 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1173 * etc. The returned io_u is fully ready to be prepped and submitted.
1175 struct io_u *get_io_u(struct thread_data *td)
1179 int do_scramble = 0;
1181 io_u = __get_io_u(td);
1183 dprint(FD_IO, "__get_io_u failed\n");
1187 if (check_get_verify(td, io_u))
1189 if (check_get_trim(td, io_u))
1193 * from a requeue, io_u already setup
1199 * If using an iolog, grab next piece if any available.
1201 if (td->o.read_iolog_file) {
1202 if (read_iolog_get(td, io_u))
1204 } else if (set_io_u_file(td, io_u)) {
1205 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1210 assert(fio_file_open(f));
1212 if (ddir_rw(io_u->ddir)) {
1213 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1214 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1218 f->last_start = io_u->offset;
1219 f->last_pos = io_u->offset + io_u->buflen;
1221 if (io_u->ddir == DDIR_WRITE) {
1222 if (td->o.verify != VERIFY_NONE)
1223 populate_verify_io_u(td, io_u);
1224 else if (td->o.refill_buffers)
1225 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1226 else if (td->o.scramble_buffers)
1228 } else if (io_u->ddir == DDIR_READ) {
1230 * Reset the buf_filled parameters so next time if the
1231 * buffer is used for writes it is refilled.
1233 io_u->buf_filled_len = 0;
1238 * Set io data pointers.
1240 io_u->xfer_buf = io_u->buf;
1241 io_u->xfer_buflen = io_u->buflen;
1245 if (!td_io_prep(td, io_u)) {
1246 if (!td->o.disable_slat)
1247 fio_gettime(&io_u->start_time, NULL);
1249 small_content_scramble(io_u);
1253 dprint(FD_IO, "get_io_u failed\n");
1258 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1260 const char *msg[] = { "read", "write", "sync", "datasync",
1261 "sync_file_range", "wait", "trim" };
1265 log_err("fio: io_u error");
1268 log_err(" on file %s", io_u->file->file_name);
1270 log_err(": %s\n", strerror(io_u->error));
1272 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1273 io_u->offset, io_u->xfer_buflen);
1276 td_verror(td, io_u->error, "io_u error");
1279 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1280 struct io_completion_data *icd,
1281 const enum fio_ddir idx, unsigned int bytes)
1283 unsigned long uninitialized_var(lusec);
1288 if (!td->o.disable_clat || !td->o.disable_bw)
1289 lusec = utime_since(&io_u->issue_time, &icd->time);
1291 if (!td->o.disable_lat) {
1292 unsigned long tusec;
1294 tusec = utime_since(&io_u->start_time, &icd->time);
1295 add_lat_sample(td, idx, tusec, bytes);
1298 if (!td->o.disable_clat) {
1299 add_clat_sample(td, idx, lusec, bytes);
1300 io_u_mark_latency(td, lusec);
1303 if (!td->o.disable_bw)
1304 add_bw_sample(td, idx, bytes, &icd->time);
1306 add_iops_sample(td, idx, &icd->time);
1309 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1311 unsigned long long secs, remainder, bps, bytes;
1312 bytes = td->this_io_bytes[ddir];
1313 bps = td->rate_bps[ddir];
1315 remainder = bytes % bps;
1316 return remainder * 1000000 / bps + secs * 1000000;
1319 static void io_completed(struct thread_data *td, struct io_u *io_u,
1320 struct io_completion_data *icd)
1323 * Older gcc's are too dumb to realize that usec is always used
1324 * initialized, silence that warning.
1326 unsigned long uninitialized_var(usec);
1329 dprint_io_u(io_u, "io complete");
1332 assert(io_u->flags & IO_U_F_FLIGHT);
1333 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1336 if (ddir_sync(io_u->ddir)) {
1337 td->last_was_sync = 1;
1340 f->first_write = -1ULL;
1341 f->last_write = -1ULL;
1346 td->last_was_sync = 0;
1347 td->last_ddir = io_u->ddir;
1349 if (!io_u->error && ddir_rw(io_u->ddir)) {
1350 unsigned int bytes = io_u->buflen - io_u->resid;
1351 const enum fio_ddir idx = io_u->ddir;
1352 const enum fio_ddir odx = io_u->ddir ^ 1;
1355 td->io_blocks[idx]++;
1356 td->this_io_blocks[idx]++;
1357 td->io_bytes[idx] += bytes;
1358 td->this_io_bytes[idx] += bytes;
1360 if (idx == DDIR_WRITE) {
1363 if (f->first_write == -1ULL ||
1364 io_u->offset < f->first_write)
1365 f->first_write = io_u->offset;
1366 if (f->last_write == -1ULL ||
1367 ((io_u->offset + bytes) > f->last_write))
1368 f->last_write = io_u->offset + bytes;
1372 if (ramp_time_over(td) && td->runstate == TD_RUNNING) {
1373 account_io_completion(td, io_u, icd, idx, bytes);
1375 if (__should_check_rate(td, idx)) {
1376 td->rate_pending_usleep[idx] =
1377 (usec_for_io(td, idx) -
1378 utime_since_now(&td->start));
1380 if (__should_check_rate(td, odx))
1381 td->rate_pending_usleep[odx] =
1382 (usec_for_io(td, odx) -
1383 utime_since_now(&td->start));
1386 if (td_write(td) && idx == DDIR_WRITE &&
1388 td->o.verify != VERIFY_NONE)
1389 log_io_piece(td, io_u);
1391 icd->bytes_done[idx] += bytes;
1394 ret = io_u->end_io(td, io_u);
1395 if (ret && !icd->error)
1398 } else if (io_u->error) {
1399 icd->error = io_u->error;
1400 io_u_log_error(td, io_u);
1402 if (icd->error && td_non_fatal_error(icd->error) &&
1403 (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1405 * If there is a non_fatal error, then add to the error count
1406 * and clear all the errors.
1408 update_error_count(td, icd->error);
1415 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1418 if (!td->o.disable_clat || !td->o.disable_bw)
1419 fio_gettime(&icd->time, NULL);
1425 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1428 static void ios_completed(struct thread_data *td,
1429 struct io_completion_data *icd)
1434 for (i = 0; i < icd->nr; i++) {
1435 io_u = td->io_ops->event(td, i);
1437 io_completed(td, io_u, icd);
1439 if (!(io_u->flags & IO_U_F_FREE_DEF))
1447 * Complete a single io_u for the sync engines.
1449 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1450 unsigned long *bytes)
1452 struct io_completion_data icd;
1454 init_icd(td, &icd, 1);
1455 io_completed(td, io_u, &icd);
1457 if (!(io_u->flags & IO_U_F_FREE_DEF))
1461 td_verror(td, icd.error, "io_u_sync_complete");
1466 bytes[0] += icd.bytes_done[0];
1467 bytes[1] += icd.bytes_done[1];
1474 * Called to complete min_events number of io for the async engines.
1476 int io_u_queued_complete(struct thread_data *td, int min_evts,
1477 unsigned long *bytes)
1479 struct io_completion_data icd;
1480 struct timespec *tvp = NULL;
1482 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1484 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1489 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1491 td_verror(td, -ret, "td_io_getevents");
1496 init_icd(td, &icd, ret);
1497 ios_completed(td, &icd);
1499 td_verror(td, icd.error, "io_u_queued_complete");
1504 bytes[0] += icd.bytes_done[0];
1505 bytes[1] += icd.bytes_done[1];
1512 * Call when io_u is really queued, to update the submission latency.
1514 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1516 if (!td->o.disable_slat) {
1517 unsigned long slat_time;
1519 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1520 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1525 * "randomly" fill the buffer contents
1527 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1528 unsigned int max_bs)
1530 io_u->buf_filled_len = 0;
1532 if (!td->o.zero_buffers)
1533 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1535 memset(io_u->buf, 0, max_bs);