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] & (1 << 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 this_blocks, mask;
54 unsigned int idx, bit;
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 = ((1U << 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 min_bs = td->o.rw_min_bs, lastb;
132 lastb = last_block(td, f, ddir);
136 i = f->last_free_lookup;
137 *b = (i * BLOCKS_PER_MAP);
138 while ((*b) * min_bs < f->real_file_size &&
139 (*b) * min_bs < f->io_size) {
140 if (f->file_map[i] != (unsigned int) -1) {
141 *b += ffz(f->file_map[i]);
144 f->last_free_lookup = i;
148 *b += BLOCKS_PER_MAP;
152 dprint(FD_IO, "failed finding a free block\n");
156 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
157 enum fio_ddir ddir, unsigned long long *b)
159 unsigned long long r, lastb;
162 lastb = last_block(td, f, ddir);
167 r = os_random_long(&td->random_state);
168 dprint(FD_RANDOM, "off rand %llu\n", r);
169 *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
172 * if we are not maintaining a random map, we are done.
174 if (!file_randommap(td, f))
178 * calculate map offset and check if it's free
180 if (random_map_free(f, *b))
183 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
188 * we get here, if we didn't suceed in looking up a block. generate
189 * a random start offset into the filemap, and find the first free
194 f->last_free_lookup = (f->num_maps - 1) *
195 (r / (OS_RAND_MAX + 1.0));
196 if (!get_next_free_block(td, f, ddir, b))
199 r = os_random_long(&td->random_state);
203 * that didn't work either, try exhaustive search from the start
205 f->last_free_lookup = 0;
206 return get_next_free_block(td, f, ddir, b);
209 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
210 enum fio_ddir ddir, unsigned long long *b)
212 if (get_next_rand_offset(td, f, ddir, b)) {
213 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
214 f->file_name, f->last_pos, f->real_file_size);
221 static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
222 enum fio_ddir ddir, unsigned long long *b)
224 assert(ddir_rw(ddir));
226 if (f->last_pos < f->real_file_size) {
227 *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
234 static int get_next_block(struct thread_data *td, struct io_u *io_u,
235 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
237 struct fio_file *f = io_u->file;
240 assert(ddir_rw(ddir));
244 ret = get_next_rand_block(td, f, ddir, b);
246 ret = get_next_seq_block(td, f, ddir, b);
248 io_u->flags |= IO_U_F_BUSY_OK;
250 if (td->o.rw_seq == RW_SEQ_SEQ) {
251 ret = get_next_seq_block(td, f, ddir, b);
253 ret = get_next_rand_block(td, f, ddir, b);
254 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
255 if (f->last_start != -1ULL)
256 *b = (f->last_start - f->file_offset)
257 / td->o.min_bs[ddir];
262 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
271 * For random io, generate a random new block and see if it's used. Repeat
272 * until we find a free one. For sequential io, just return the end of
273 * the last io issued.
275 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
277 struct fio_file *f = io_u->file;
278 unsigned long long b;
279 enum fio_ddir ddir = io_u->ddir;
282 assert(ddir_rw(ddir));
284 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
286 td->ddir_seq_nr = td->o.ddir_seq_nr;
289 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
292 io_u->offset = b * td->o.ba[ddir];
293 if (io_u->offset >= f->io_size) {
294 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
295 io_u->offset, f->io_size);
299 io_u->offset += f->file_offset;
300 if (io_u->offset >= f->real_file_size) {
301 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
302 io_u->offset, f->real_file_size);
309 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
311 struct prof_io_ops *ops = &td->prof_io_ops;
313 if (ops->fill_io_u_off)
314 return ops->fill_io_u_off(td, io_u);
316 return __get_next_offset(td, io_u);
319 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
321 const int ddir = io_u->ddir;
322 unsigned int uninitialized_var(buflen);
323 unsigned int minbs, maxbs;
326 assert(ddir_rw(ddir));
328 minbs = td->o.min_bs[ddir];
329 maxbs = td->o.max_bs[ddir];
334 r = os_random_long(&td->bsrange_state);
335 if (!td->o.bssplit_nr[ddir]) {
336 buflen = 1 + (unsigned int) ((double) maxbs *
337 (r / (OS_RAND_MAX + 1.0)));
344 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
345 struct bssplit *bsp = &td->o.bssplit[ddir][i];
349 if (r <= ((OS_RAND_MAX / 100L) * perc))
353 if (!td->o.bs_unaligned && is_power_of_2(minbs))
354 buflen = (buflen + minbs - 1) & ~(minbs - 1);
357 if (io_u->offset + buflen > io_u->file->real_file_size) {
358 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
366 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
368 struct prof_io_ops *ops = &td->prof_io_ops;
370 if (ops->fill_io_u_size)
371 return ops->fill_io_u_size(td, io_u);
373 return __get_next_buflen(td, io_u);
376 static void set_rwmix_bytes(struct thread_data *td)
381 * we do time or byte based switch. this is needed because
382 * buffered writes may issue a lot quicker than they complete,
383 * whereas reads do not.
385 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
386 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
389 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
394 r = os_random_long(&td->rwmix_state);
395 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
396 if (v <= td->o.rwmix[DDIR_READ])
402 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
404 enum fio_ddir odir = ddir ^ 1;
408 assert(ddir_rw(ddir));
410 if (td->rate_pending_usleep[ddir] <= 0)
414 * We have too much pending sleep in this direction. See if we
419 * Other direction does not have too much pending, switch
421 if (td->rate_pending_usleep[odir] < 100000)
425 * Both directions have pending sleep. Sleep the minimum time
426 * and deduct from both.
428 if (td->rate_pending_usleep[ddir] <=
429 td->rate_pending_usleep[odir]) {
430 usec = td->rate_pending_usleep[ddir];
432 usec = td->rate_pending_usleep[odir];
436 usec = td->rate_pending_usleep[ddir];
438 fio_gettime(&t, NULL);
439 usec_sleep(td, usec);
440 usec = utime_since_now(&t);
442 td->rate_pending_usleep[ddir] -= usec;
445 if (td_rw(td) && __should_check_rate(td, odir))
446 td->rate_pending_usleep[odir] -= usec;
452 * Return the data direction for the next io_u. If the job is a
453 * mixed read/write workload, check the rwmix cycle and switch if
456 static enum fio_ddir get_rw_ddir(struct thread_data *td)
461 * see if it's time to fsync
463 if (td->o.fsync_blocks &&
464 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
465 td->io_issues[DDIR_WRITE] && should_fsync(td))
469 * see if it's time to fdatasync
471 if (td->o.fdatasync_blocks &&
472 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
473 td->io_issues[DDIR_WRITE] && should_fsync(td))
474 return DDIR_DATASYNC;
477 * see if it's time to sync_file_range
479 if (td->sync_file_range_nr &&
480 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
481 td->io_issues[DDIR_WRITE] && should_fsync(td))
482 return DDIR_SYNC_FILE_RANGE;
486 * Check if it's time to seed a new data direction.
488 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
490 * Put a top limit on how many bytes we do for
491 * one data direction, to avoid overflowing the
494 ddir = get_rand_ddir(td);
496 if (ddir != td->rwmix_ddir)
499 td->rwmix_ddir = ddir;
501 ddir = td->rwmix_ddir;
502 } else if (td_read(td))
507 td->rwmix_ddir = rate_ddir(td, ddir);
508 return td->rwmix_ddir;
511 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
513 io_u->ddir = get_rw_ddir(td);
515 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
516 td->o.barrier_blocks &&
517 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
518 td->io_issues[DDIR_WRITE])
519 io_u->flags |= IO_U_F_BARRIER;
522 void put_file_log(struct thread_data *td, struct fio_file *f)
524 int ret = put_file(td, f);
527 td_verror(td, ret, "file close");
530 void put_io_u(struct thread_data *td, struct io_u *io_u)
534 io_u->flags |= IO_U_F_FREE;
535 io_u->flags &= ~IO_U_F_FREE_DEF;
538 put_file_log(td, io_u->file);
541 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
543 flist_del_init(&io_u->list);
544 flist_add(&io_u->list, &td->io_u_freelist);
546 td_io_u_free_notify(td);
549 void clear_io_u(struct thread_data *td, struct io_u *io_u)
551 io_u->flags &= ~IO_U_F_FLIGHT;
555 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
557 struct io_u *__io_u = *io_u;
559 dprint(FD_IO, "requeue %p\n", __io_u);
563 __io_u->flags |= IO_U_F_FREE;
564 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
565 td->io_issues[__io_u->ddir]--;
567 __io_u->flags &= ~IO_U_F_FLIGHT;
568 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
570 flist_del(&__io_u->list);
571 flist_add_tail(&__io_u->list, &td->io_u_requeues);
576 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
578 if (td->io_ops->flags & FIO_NOIO)
581 set_rw_ddir(td, io_u);
584 * fsync() or fdatasync() or trim etc, we are done
586 if (!ddir_rw(io_u->ddir))
590 * See if it's time to switch to a new zone
592 if (td->zone_bytes >= td->o.zone_size) {
594 io_u->file->last_pos += td->o.zone_skip;
595 td->io_skip_bytes += td->o.zone_skip;
599 * No log, let the seq/rand engine retrieve the next buflen and
602 if (get_next_offset(td, io_u)) {
603 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
607 io_u->buflen = get_next_buflen(td, io_u);
609 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
613 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
614 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
615 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
616 io_u->buflen, io_u->file->real_file_size);
621 * mark entry before potentially trimming io_u
623 if (td_random(td) && file_randommap(td, io_u->file))
624 mark_random_map(td, io_u);
627 * If using a write iolog, store this entry.
630 dprint_io_u(io_u, "fill_io_u");
631 td->zone_bytes += io_u->buflen;
636 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
665 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
667 __io_u_mark_map(td->ts.io_u_submit, nr);
668 td->ts.total_submit++;
671 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
673 __io_u_mark_map(td->ts.io_u_complete, nr);
674 td->ts.total_complete++;
677 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
681 switch (td->cur_depth) {
703 td->ts.io_u_map[index] += nr;
706 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
743 assert(index < FIO_IO_U_LAT_U_NR);
744 td->ts.io_u_lat_u[index]++;
747 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
788 assert(index < FIO_IO_U_LAT_M_NR);
789 td->ts.io_u_lat_m[index]++;
792 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
795 io_u_mark_lat_usec(td, usec);
797 io_u_mark_lat_msec(td, usec / 1000);
801 * Get next file to service by choosing one at random
803 static struct fio_file *get_next_file_rand(struct thread_data *td,
804 enum fio_file_flags goodf,
805 enum fio_file_flags badf)
811 long r = os_random_long(&td->next_file_state);
814 fno = (unsigned int) ((double) td->o.nr_files
815 * (r / (OS_RAND_MAX + 1.0)));
817 if (fio_file_done(f))
820 if (!fio_file_open(f)) {
823 err = td_io_open_file(td, f);
829 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
830 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
834 td_io_close_file(td, f);
839 * Get next file to service by doing round robin between all available ones
841 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
844 unsigned int old_next_file = td->next_file;
850 f = td->files[td->next_file];
853 if (td->next_file >= td->o.nr_files)
856 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
857 if (fio_file_done(f)) {
862 if (!fio_file_open(f)) {
865 err = td_io_open_file(td, f);
867 dprint(FD_FILE, "error %d on open of %s\n",
875 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
877 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
881 td_io_close_file(td, f);
884 } while (td->next_file != old_next_file);
886 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
890 static struct fio_file *__get_next_file(struct thread_data *td)
894 assert(td->o.nr_files <= td->files_index);
896 if (td->nr_done_files >= td->o.nr_files) {
897 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
898 " nr_files=%d\n", td->nr_open_files,
904 f = td->file_service_file;
905 if (f && fio_file_open(f) && !fio_file_closing(f)) {
906 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
908 if (td->file_service_left--)
912 if (td->o.file_service_type == FIO_FSERVICE_RR ||
913 td->o.file_service_type == FIO_FSERVICE_SEQ)
914 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
916 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
918 td->file_service_file = f;
919 td->file_service_left = td->file_service_nr - 1;
921 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
925 static struct fio_file *get_next_file(struct thread_data *td)
927 struct prof_io_ops *ops = &td->prof_io_ops;
929 if (ops->get_next_file)
930 return ops->get_next_file(td);
932 return __get_next_file(td);
935 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
940 f = get_next_file(td);
947 if (!fill_io_u(td, io_u))
951 td_io_close_file(td, f);
953 fio_file_set_done(f);
955 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
956 td->nr_done_files, td->o.nr_files);
963 struct io_u *__get_io_u(struct thread_data *td)
965 struct io_u *io_u = NULL;
970 if (!flist_empty(&td->io_u_requeues))
971 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
972 else if (!queue_full(td)) {
973 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
982 assert(io_u->flags & IO_U_F_FREE);
983 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
984 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
987 flist_del(&io_u->list);
988 flist_add(&io_u->list, &td->io_u_busylist);
990 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
991 } else if (td->o.verify_async) {
993 * We ran out, wait for async verify threads to finish and
996 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1004 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1006 if (td->o.trim_backlog && td->trim_entries) {
1009 if (td->trim_batch) {
1012 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1013 td->last_ddir != DDIR_READ) {
1014 td->trim_batch = td->o.trim_batch;
1015 if (!td->trim_batch)
1016 td->trim_batch = td->o.trim_backlog;
1020 if (get_trim && !get_next_trim(td, io_u))
1027 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1029 if (td->o.verify_backlog && td->io_hist_len) {
1032 if (td->verify_batch) {
1035 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1036 td->last_ddir != DDIR_READ) {
1037 td->verify_batch = td->o.verify_batch;
1038 if (!td->verify_batch)
1039 td->verify_batch = td->o.verify_backlog;
1043 if (get_verify && !get_next_verify(td, io_u))
1051 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1052 * etc. The returned io_u is fully ready to be prepped and submitted.
1054 struct io_u *get_io_u(struct thread_data *td)
1059 io_u = __get_io_u(td);
1061 dprint(FD_IO, "__get_io_u failed\n");
1065 if (check_get_verify(td, io_u))
1067 if (check_get_trim(td, io_u))
1071 * from a requeue, io_u already setup
1077 * If using an iolog, grab next piece if any available.
1079 if (td->o.read_iolog_file) {
1080 if (read_iolog_get(td, io_u))
1082 } else if (set_io_u_file(td, io_u)) {
1083 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1088 assert(fio_file_open(f));
1090 if (ddir_rw(io_u->ddir)) {
1091 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1092 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1096 f->last_start = io_u->offset;
1097 f->last_pos = io_u->offset + io_u->buflen;
1099 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1100 populate_verify_io_u(td, io_u);
1101 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1102 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1103 else if (io_u->ddir == DDIR_READ) {
1105 * Reset the buf_filled parameters so next time if the
1106 * buffer is used for writes it is refilled.
1108 io_u->buf_filled_len = 0;
1113 * Set io data pointers.
1115 io_u->xfer_buf = io_u->buf;
1116 io_u->xfer_buflen = io_u->buflen;
1120 if (!td_io_prep(td, io_u)) {
1121 if (!td->o.disable_slat)
1122 fio_gettime(&io_u->start_time, NULL);
1126 dprint(FD_IO, "get_io_u failed\n");
1131 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1133 const char *msg[] = { "read", "write", "sync", "datasync",
1134 "sync_file_range", "wait", "trim" };
1138 log_err("fio: io_u error");
1141 log_err(" on file %s", io_u->file->file_name);
1143 log_err(": %s\n", strerror(io_u->error));
1145 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1146 io_u->offset, io_u->xfer_buflen);
1149 td_verror(td, io_u->error, "io_u error");
1152 static void io_completed(struct thread_data *td, struct io_u *io_u,
1153 struct io_completion_data *icd)
1156 * Older gcc's are too dumb to realize that usec is always used
1157 * initialized, silence that warning.
1159 unsigned long uninitialized_var(usec);
1162 dprint_io_u(io_u, "io complete");
1165 assert(io_u->flags & IO_U_F_FLIGHT);
1166 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1169 if (ddir_sync(io_u->ddir)) {
1170 td->last_was_sync = 1;
1173 f->first_write = -1ULL;
1174 f->last_write = -1ULL;
1179 td->last_was_sync = 0;
1180 td->last_ddir = io_u->ddir;
1182 if (!io_u->error && ddir_rw(io_u->ddir)) {
1183 unsigned int bytes = io_u->buflen - io_u->resid;
1184 const enum fio_ddir idx = io_u->ddir;
1185 const enum fio_ddir odx = io_u->ddir ^ 1;
1188 td->io_blocks[idx]++;
1189 td->io_bytes[idx] += bytes;
1190 td->this_io_bytes[idx] += bytes;
1192 if (idx == DDIR_WRITE) {
1195 if (f->first_write == -1ULL ||
1196 io_u->offset < f->first_write)
1197 f->first_write = io_u->offset;
1198 if (f->last_write == -1ULL ||
1199 ((io_u->offset + bytes) > f->last_write))
1200 f->last_write = io_u->offset + bytes;
1204 if (ramp_time_over(td)) {
1205 unsigned long uninitialized_var(lusec);
1207 if (!td->o.disable_clat || !td->o.disable_bw)
1208 lusec = utime_since(&io_u->issue_time,
1210 if (!td->o.disable_lat) {
1211 unsigned long tusec;
1213 tusec = utime_since(&io_u->start_time,
1215 add_lat_sample(td, idx, tusec, bytes);
1217 if (!td->o.disable_clat) {
1218 add_clat_sample(td, idx, lusec, bytes);
1219 io_u_mark_latency(td, lusec);
1221 if (!td->o.disable_bw)
1222 add_bw_sample(td, idx, bytes, &icd->time);
1223 if (__should_check_rate(td, idx)) {
1224 td->rate_pending_usleep[idx] =
1225 ((td->this_io_bytes[idx] *
1226 td->rate_nsec_cycle[idx]) / 1000 -
1227 utime_since_now(&td->start));
1229 if (__should_check_rate(td, idx ^ 1))
1230 td->rate_pending_usleep[odx] =
1231 ((td->this_io_bytes[odx] *
1232 td->rate_nsec_cycle[odx]) / 1000 -
1233 utime_since_now(&td->start));
1236 if (td_write(td) && idx == DDIR_WRITE &&
1238 td->o.verify != VERIFY_NONE)
1239 log_io_piece(td, io_u);
1241 icd->bytes_done[idx] += bytes;
1244 ret = io_u->end_io(td, io_u);
1245 if (ret && !icd->error)
1248 } else if (io_u->error) {
1249 icd->error = io_u->error;
1250 io_u_log_error(td, io_u);
1252 if (td->o.continue_on_error && icd->error &&
1253 td_non_fatal_error(icd->error)) {
1255 * If there is a non_fatal error, then add to the error count
1256 * and clear all the errors.
1258 update_error_count(td, icd->error);
1265 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1268 if (!td->o.disable_clat || !td->o.disable_bw)
1269 fio_gettime(&icd->time, NULL);
1274 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1277 static void ios_completed(struct thread_data *td,
1278 struct io_completion_data *icd)
1283 for (i = 0; i < icd->nr; i++) {
1284 io_u = td->io_ops->event(td, i);
1286 io_completed(td, io_u, icd);
1288 if (!(io_u->flags & IO_U_F_FREE_DEF))
1294 * Complete a single io_u for the sync engines.
1296 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1297 unsigned long *bytes)
1299 struct io_completion_data icd;
1301 init_icd(td, &icd, 1);
1302 io_completed(td, io_u, &icd);
1304 if (!(io_u->flags & IO_U_F_FREE_DEF))
1308 td_verror(td, icd.error, "io_u_sync_complete");
1313 bytes[0] += icd.bytes_done[0];
1314 bytes[1] += icd.bytes_done[1];
1321 * Called to complete min_events number of io for the async engines.
1323 int io_u_queued_complete(struct thread_data *td, int min_evts,
1324 unsigned long *bytes)
1326 struct io_completion_data icd;
1327 struct timespec *tvp = NULL;
1329 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1331 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1336 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1338 td_verror(td, -ret, "td_io_getevents");
1343 init_icd(td, &icd, ret);
1344 ios_completed(td, &icd);
1346 td_verror(td, icd.error, "io_u_queued_complete");
1351 bytes[0] += icd.bytes_done[0];
1352 bytes[1] += icd.bytes_done[1];
1359 * Call when io_u is really queued, to update the submission latency.
1361 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1363 if (!td->o.disable_slat) {
1364 unsigned long slat_time;
1366 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1367 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1372 * "randomly" fill the buffer contents
1374 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1375 unsigned int max_bs)
1377 if (!td->o.zero_buffers)
1378 fill_random_buf(io_u->buf, max_bs);
1380 memset(io_u->buf, 0, max_bs);