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 r = os_random_long(&td->random_state);
172 dprint(FD_RANDOM, "off rand %llu\n", r);
173 *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
176 * if we are not maintaining a random map, we are done.
178 if (!file_randommap(td, f))
182 * calculate map offset and check if it's free
184 if (random_map_free(f, *b))
187 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
191 if (!f->failed_rands++)
192 f->last_free_lookup = 0;
195 * we get here, if we didn't suceed in looking up a block. generate
196 * a random start offset into the filemap, and find the first free
201 f->last_free_lookup = (f->num_maps - 1) *
202 (r / (OS_RAND_MAX + 1.0));
203 if (!get_next_free_block(td, f, ddir, b))
206 r = os_random_long(&td->random_state);
210 * that didn't work either, try exhaustive search from the start
212 f->last_free_lookup = 0;
214 if (!get_next_free_block(td, f, ddir, b))
216 f->last_free_lookup = 0;
217 return get_next_free_block(td, f, ddir, b);
224 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
225 enum fio_ddir ddir, unsigned long long *b)
227 if (get_next_rand_offset(td, f, ddir, b)) {
228 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
229 f->file_name, f->last_pos, f->real_file_size);
236 static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
237 enum fio_ddir ddir, unsigned long long *b)
239 assert(ddir_rw(ddir));
241 if (f->last_pos < f->real_file_size) {
242 *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
249 static int get_next_block(struct thread_data *td, struct io_u *io_u,
250 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
252 struct fio_file *f = io_u->file;
255 assert(ddir_rw(ddir));
259 ret = get_next_rand_block(td, f, ddir, b);
261 ret = get_next_seq_block(td, f, ddir, b);
263 io_u->flags |= IO_U_F_BUSY_OK;
265 if (td->o.rw_seq == RW_SEQ_SEQ) {
266 ret = get_next_seq_block(td, f, ddir, b);
268 ret = get_next_rand_block(td, f, ddir, b);
269 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
270 if (f->last_start != -1ULL)
271 *b = (f->last_start - f->file_offset)
272 / td->o.min_bs[ddir];
277 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
286 * For random io, generate a random new block and see if it's used. Repeat
287 * until we find a free one. For sequential io, just return the end of
288 * the last io issued.
290 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
292 struct fio_file *f = io_u->file;
293 unsigned long long b;
294 enum fio_ddir ddir = io_u->ddir;
297 assert(ddir_rw(ddir));
299 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
301 td->ddir_seq_nr = td->o.ddir_seq_nr;
304 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
307 io_u->offset = b * td->o.ba[ddir];
308 if (io_u->offset >= f->io_size) {
309 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
310 io_u->offset, f->io_size);
314 io_u->offset += f->file_offset;
315 if (io_u->offset >= f->real_file_size) {
316 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
317 io_u->offset, f->real_file_size);
324 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
326 struct prof_io_ops *ops = &td->prof_io_ops;
328 if (ops->fill_io_u_off)
329 return ops->fill_io_u_off(td, io_u);
331 return __get_next_offset(td, io_u);
334 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
336 const int ddir = io_u->ddir;
337 unsigned int uninitialized_var(buflen);
338 unsigned int minbs, maxbs;
341 assert(ddir_rw(ddir));
343 minbs = td->o.min_bs[ddir];
344 maxbs = td->o.max_bs[ddir];
349 r = os_random_long(&td->bsrange_state);
350 if (!td->o.bssplit_nr[ddir]) {
351 buflen = 1 + (unsigned int) ((double) maxbs *
352 (r / (OS_RAND_MAX + 1.0)));
359 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
360 struct bssplit *bsp = &td->o.bssplit[ddir][i];
364 if (r <= ((OS_RAND_MAX / 100L) * perc))
368 if (!td->o.bs_unaligned && is_power_of_2(minbs))
369 buflen = (buflen + minbs - 1) & ~(minbs - 1);
372 if (io_u->offset + buflen > io_u->file->real_file_size) {
373 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
381 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
383 struct prof_io_ops *ops = &td->prof_io_ops;
385 if (ops->fill_io_u_size)
386 return ops->fill_io_u_size(td, io_u);
388 return __get_next_buflen(td, io_u);
391 static void set_rwmix_bytes(struct thread_data *td)
396 * we do time or byte based switch. this is needed because
397 * buffered writes may issue a lot quicker than they complete,
398 * whereas reads do not.
400 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
401 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
404 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
409 r = os_random_long(&td->rwmix_state);
410 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
411 if (v <= td->o.rwmix[DDIR_READ])
417 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
419 enum fio_ddir odir = ddir ^ 1;
423 assert(ddir_rw(ddir));
425 if (td->rate_pending_usleep[ddir] <= 0)
429 * We have too much pending sleep in this direction. See if we
434 * Other direction does not have too much pending, switch
436 if (td->rate_pending_usleep[odir] < 100000)
440 * Both directions have pending sleep. Sleep the minimum time
441 * and deduct from both.
443 if (td->rate_pending_usleep[ddir] <=
444 td->rate_pending_usleep[odir]) {
445 usec = td->rate_pending_usleep[ddir];
447 usec = td->rate_pending_usleep[odir];
451 usec = td->rate_pending_usleep[ddir];
453 fio_gettime(&t, NULL);
454 usec_sleep(td, usec);
455 usec = utime_since_now(&t);
457 td->rate_pending_usleep[ddir] -= usec;
460 if (td_rw(td) && __should_check_rate(td, odir))
461 td->rate_pending_usleep[odir] -= usec;
467 * Return the data direction for the next io_u. If the job is a
468 * mixed read/write workload, check the rwmix cycle and switch if
471 static enum fio_ddir get_rw_ddir(struct thread_data *td)
476 * see if it's time to fsync
478 if (td->o.fsync_blocks &&
479 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
480 td->io_issues[DDIR_WRITE] && should_fsync(td))
484 * see if it's time to fdatasync
486 if (td->o.fdatasync_blocks &&
487 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
488 td->io_issues[DDIR_WRITE] && should_fsync(td))
489 return DDIR_DATASYNC;
492 * see if it's time to sync_file_range
494 if (td->sync_file_range_nr &&
495 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
496 td->io_issues[DDIR_WRITE] && should_fsync(td))
497 return DDIR_SYNC_FILE_RANGE;
501 * Check if it's time to seed a new data direction.
503 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
505 * Put a top limit on how many bytes we do for
506 * one data direction, to avoid overflowing the
509 ddir = get_rand_ddir(td);
511 if (ddir != td->rwmix_ddir)
514 td->rwmix_ddir = ddir;
516 ddir = td->rwmix_ddir;
517 } else if (td_read(td))
522 td->rwmix_ddir = rate_ddir(td, ddir);
523 return td->rwmix_ddir;
526 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
528 io_u->ddir = get_rw_ddir(td);
530 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
531 td->o.barrier_blocks &&
532 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
533 td->io_issues[DDIR_WRITE])
534 io_u->flags |= IO_U_F_BARRIER;
537 void put_file_log(struct thread_data *td, struct fio_file *f)
539 int ret = put_file(td, f);
542 td_verror(td, ret, "file close");
545 void put_io_u(struct thread_data *td, struct io_u *io_u)
549 io_u->flags |= IO_U_F_FREE;
550 io_u->flags &= ~IO_U_F_FREE_DEF;
553 put_file_log(td, io_u->file);
556 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
558 flist_del_init(&io_u->list);
559 flist_add(&io_u->list, &td->io_u_freelist);
561 td_io_u_free_notify(td);
564 void clear_io_u(struct thread_data *td, struct io_u *io_u)
566 io_u->flags &= ~IO_U_F_FLIGHT;
570 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
572 struct io_u *__io_u = *io_u;
574 dprint(FD_IO, "requeue %p\n", __io_u);
578 __io_u->flags |= IO_U_F_FREE;
579 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
580 td->io_issues[__io_u->ddir]--;
582 __io_u->flags &= ~IO_U_F_FLIGHT;
583 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
585 flist_del(&__io_u->list);
586 flist_add_tail(&__io_u->list, &td->io_u_requeues);
591 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
593 if (td->io_ops->flags & FIO_NOIO)
596 set_rw_ddir(td, io_u);
599 * fsync() or fdatasync() or trim etc, we are done
601 if (!ddir_rw(io_u->ddir))
605 * See if it's time to switch to a new zone
607 if (td->zone_bytes >= td->o.zone_size) {
609 io_u->file->last_pos += td->o.zone_skip;
610 td->io_skip_bytes += td->o.zone_skip;
614 * No log, let the seq/rand engine retrieve the next buflen and
617 if (get_next_offset(td, io_u)) {
618 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
622 io_u->buflen = get_next_buflen(td, io_u);
624 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
628 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
629 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
630 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
631 io_u->buflen, io_u->file->real_file_size);
636 * mark entry before potentially trimming io_u
638 if (td_random(td) && file_randommap(td, io_u->file))
639 mark_random_map(td, io_u);
642 * If using a write iolog, store this entry.
645 dprint_io_u(io_u, "fill_io_u");
646 td->zone_bytes += io_u->buflen;
651 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
680 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
682 __io_u_mark_map(td->ts.io_u_submit, nr);
683 td->ts.total_submit++;
686 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
688 __io_u_mark_map(td->ts.io_u_complete, nr);
689 td->ts.total_complete++;
692 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
696 switch (td->cur_depth) {
718 td->ts.io_u_map[idx] += nr;
721 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
758 assert(idx < FIO_IO_U_LAT_U_NR);
759 td->ts.io_u_lat_u[idx]++;
762 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
803 assert(idx < FIO_IO_U_LAT_M_NR);
804 td->ts.io_u_lat_m[idx]++;
807 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
810 io_u_mark_lat_usec(td, usec);
812 io_u_mark_lat_msec(td, usec / 1000);
816 * Get next file to service by choosing one at random
818 static struct fio_file *get_next_file_rand(struct thread_data *td,
819 enum fio_file_flags goodf,
820 enum fio_file_flags badf)
826 long r = os_random_long(&td->next_file_state);
829 fno = (unsigned int) ((double) td->o.nr_files
830 * (r / (OS_RAND_MAX + 1.0)));
832 if (fio_file_done(f))
835 if (!fio_file_open(f)) {
838 err = td_io_open_file(td, f);
844 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
845 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
849 td_io_close_file(td, f);
854 * Get next file to service by doing round robin between all available ones
856 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
859 unsigned int old_next_file = td->next_file;
865 f = td->files[td->next_file];
868 if (td->next_file >= td->o.nr_files)
871 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
872 if (fio_file_done(f)) {
877 if (!fio_file_open(f)) {
880 err = td_io_open_file(td, f);
882 dprint(FD_FILE, "error %d on open of %s\n",
890 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
892 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
896 td_io_close_file(td, f);
899 } while (td->next_file != old_next_file);
901 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
905 static struct fio_file *__get_next_file(struct thread_data *td)
909 assert(td->o.nr_files <= td->files_index);
911 if (td->nr_done_files >= td->o.nr_files) {
912 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
913 " nr_files=%d\n", td->nr_open_files,
919 f = td->file_service_file;
920 if (f && fio_file_open(f) && !fio_file_closing(f)) {
921 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
923 if (td->file_service_left--)
927 if (td->o.file_service_type == FIO_FSERVICE_RR ||
928 td->o.file_service_type == FIO_FSERVICE_SEQ)
929 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
931 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
933 td->file_service_file = f;
934 td->file_service_left = td->file_service_nr - 1;
936 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
940 static struct fio_file *get_next_file(struct thread_data *td)
942 struct prof_io_ops *ops = &td->prof_io_ops;
944 if (ops->get_next_file)
945 return ops->get_next_file(td);
947 return __get_next_file(td);
950 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
955 f = get_next_file(td);
962 if (!fill_io_u(td, io_u))
966 td_io_close_file(td, f);
968 fio_file_set_done(f);
970 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
971 td->nr_done_files, td->o.nr_files);
978 struct io_u *__get_io_u(struct thread_data *td)
980 struct io_u *io_u = NULL;
985 if (!flist_empty(&td->io_u_requeues))
986 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
987 else if (!queue_full(td)) {
988 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
997 assert(io_u->flags & IO_U_F_FREE);
998 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
999 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1002 flist_del(&io_u->list);
1003 flist_add(&io_u->list, &td->io_u_busylist);
1005 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1006 } else if (td->o.verify_async) {
1008 * We ran out, wait for async verify threads to finish and
1011 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1019 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1021 if (td->o.trim_backlog && td->trim_entries) {
1024 if (td->trim_batch) {
1027 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1028 td->last_ddir != DDIR_READ) {
1029 td->trim_batch = td->o.trim_batch;
1030 if (!td->trim_batch)
1031 td->trim_batch = td->o.trim_backlog;
1035 if (get_trim && !get_next_trim(td, io_u))
1042 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1044 if (td->o.verify_backlog && td->io_hist_len) {
1047 if (td->verify_batch) {
1050 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1051 td->last_ddir != DDIR_READ) {
1052 td->verify_batch = td->o.verify_batch;
1053 if (!td->verify_batch)
1054 td->verify_batch = td->o.verify_backlog;
1058 if (get_verify && !get_next_verify(td, io_u))
1066 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1067 * etc. The returned io_u is fully ready to be prepped and submitted.
1069 struct io_u *get_io_u(struct thread_data *td)
1074 io_u = __get_io_u(td);
1076 dprint(FD_IO, "__get_io_u failed\n");
1080 if (check_get_verify(td, io_u))
1082 if (check_get_trim(td, io_u))
1086 * from a requeue, io_u already setup
1092 * If using an iolog, grab next piece if any available.
1094 if (td->o.read_iolog_file) {
1095 if (read_iolog_get(td, io_u))
1097 } else if (set_io_u_file(td, io_u)) {
1098 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1103 assert(fio_file_open(f));
1105 if (ddir_rw(io_u->ddir)) {
1106 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1107 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1111 f->last_start = io_u->offset;
1112 f->last_pos = io_u->offset + io_u->buflen;
1114 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1115 populate_verify_io_u(td, io_u);
1116 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1117 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1118 else if (io_u->ddir == DDIR_READ) {
1120 * Reset the buf_filled parameters so next time if the
1121 * buffer is used for writes it is refilled.
1123 io_u->buf_filled_len = 0;
1128 * Set io data pointers.
1130 io_u->xfer_buf = io_u->buf;
1131 io_u->xfer_buflen = io_u->buflen;
1135 if (!td_io_prep(td, io_u)) {
1136 if (!td->o.disable_slat)
1137 fio_gettime(&io_u->start_time, NULL);
1141 dprint(FD_IO, "get_io_u failed\n");
1146 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1148 const char *msg[] = { "read", "write", "sync", "datasync",
1149 "sync_file_range", "wait", "trim" };
1153 log_err("fio: io_u error");
1156 log_err(" on file %s", io_u->file->file_name);
1158 log_err(": %s\n", strerror(io_u->error));
1160 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1161 io_u->offset, io_u->xfer_buflen);
1164 td_verror(td, io_u->error, "io_u error");
1167 static void io_completed(struct thread_data *td, struct io_u *io_u,
1168 struct io_completion_data *icd)
1171 * Older gcc's are too dumb to realize that usec is always used
1172 * initialized, silence that warning.
1174 unsigned long uninitialized_var(usec);
1177 dprint_io_u(io_u, "io complete");
1180 assert(io_u->flags & IO_U_F_FLIGHT);
1181 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1184 if (ddir_sync(io_u->ddir)) {
1185 td->last_was_sync = 1;
1188 f->first_write = -1ULL;
1189 f->last_write = -1ULL;
1194 td->last_was_sync = 0;
1195 td->last_ddir = io_u->ddir;
1197 if (!io_u->error && ddir_rw(io_u->ddir)) {
1198 unsigned int bytes = io_u->buflen - io_u->resid;
1199 const enum fio_ddir idx = io_u->ddir;
1200 const enum fio_ddir odx = io_u->ddir ^ 1;
1203 td->io_blocks[idx]++;
1204 td->io_bytes[idx] += bytes;
1205 td->this_io_bytes[idx] += bytes;
1207 if (idx == DDIR_WRITE) {
1210 if (f->first_write == -1ULL ||
1211 io_u->offset < f->first_write)
1212 f->first_write = io_u->offset;
1213 if (f->last_write == -1ULL ||
1214 ((io_u->offset + bytes) > f->last_write))
1215 f->last_write = io_u->offset + bytes;
1219 if (ramp_time_over(td)) {
1220 unsigned long uninitialized_var(lusec);
1222 if (!td->o.disable_clat || !td->o.disable_bw)
1223 lusec = utime_since(&io_u->issue_time,
1225 if (!td->o.disable_lat) {
1226 unsigned long tusec;
1228 tusec = utime_since(&io_u->start_time,
1230 add_lat_sample(td, idx, tusec, bytes);
1232 if (!td->o.disable_clat) {
1233 add_clat_sample(td, idx, lusec, bytes);
1234 io_u_mark_latency(td, lusec);
1236 if (!td->o.disable_bw)
1237 add_bw_sample(td, idx, bytes, &icd->time);
1238 if (__should_check_rate(td, idx)) {
1239 td->rate_pending_usleep[idx] =
1240 ((td->this_io_bytes[idx] *
1241 td->rate_nsec_cycle[idx]) / 1000 -
1242 utime_since_now(&td->start));
1244 if (__should_check_rate(td, idx ^ 1))
1245 td->rate_pending_usleep[odx] =
1246 ((td->this_io_bytes[odx] *
1247 td->rate_nsec_cycle[odx]) / 1000 -
1248 utime_since_now(&td->start));
1251 if (td_write(td) && idx == DDIR_WRITE &&
1253 td->o.verify != VERIFY_NONE)
1254 log_io_piece(td, io_u);
1256 icd->bytes_done[idx] += bytes;
1259 ret = io_u->end_io(td, io_u);
1260 if (ret && !icd->error)
1263 } else if (io_u->error) {
1264 icd->error = io_u->error;
1265 io_u_log_error(td, io_u);
1267 if (td->o.continue_on_error && icd->error &&
1268 td_non_fatal_error(icd->error)) {
1270 * If there is a non_fatal error, then add to the error count
1271 * and clear all the errors.
1273 update_error_count(td, icd->error);
1280 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1283 if (!td->o.disable_clat || !td->o.disable_bw)
1284 fio_gettime(&icd->time, NULL);
1289 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1292 static void ios_completed(struct thread_data *td,
1293 struct io_completion_data *icd)
1298 for (i = 0; i < icd->nr; i++) {
1299 io_u = td->io_ops->event(td, i);
1301 io_completed(td, io_u, icd);
1303 if (!(io_u->flags & IO_U_F_FREE_DEF))
1309 * Complete a single io_u for the sync engines.
1311 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1312 unsigned long *bytes)
1314 struct io_completion_data icd;
1316 init_icd(td, &icd, 1);
1317 io_completed(td, io_u, &icd);
1319 if (!(io_u->flags & IO_U_F_FREE_DEF))
1323 td_verror(td, icd.error, "io_u_sync_complete");
1328 bytes[0] += icd.bytes_done[0];
1329 bytes[1] += icd.bytes_done[1];
1336 * Called to complete min_events number of io for the async engines.
1338 int io_u_queued_complete(struct thread_data *td, int min_evts,
1339 unsigned long *bytes)
1341 struct io_completion_data icd;
1342 struct timespec *tvp = NULL;
1344 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1346 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1351 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1353 td_verror(td, -ret, "td_io_getevents");
1358 init_icd(td, &icd, ret);
1359 ios_completed(td, &icd);
1361 td_verror(td, icd.error, "io_u_queued_complete");
1366 bytes[0] += icd.bytes_done[0];
1367 bytes[1] += icd.bytes_done[1];
1374 * Call when io_u is really queued, to update the submission latency.
1376 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1378 if (!td->o.disable_slat) {
1379 unsigned long slat_time;
1381 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1382 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1387 * "randomly" fill the buffer contents
1389 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1390 unsigned int max_bs)
1392 io_u->buf_filled_len = 0;
1394 if (!td->o.zero_buffers)
1395 fill_random_buf(io_u->buf, max_bs);
1397 memset(io_u->buf, 0, max_bs);