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;
254 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
255 f->last_pos = f->real_file_size;
257 pos = f->last_pos - f->file_offset;
259 pos += td->o.ddir_seq_add;
261 *b = pos / td->o.min_bs[ddir];
268 static int get_next_block(struct thread_data *td, struct io_u *io_u,
269 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
271 struct fio_file *f = io_u->file;
274 assert(ddir_rw(ddir));
278 ret = get_next_rand_block(td, f, ddir, b);
280 ret = get_next_seq_block(td, f, ddir, b);
282 io_u->flags |= IO_U_F_BUSY_OK;
284 if (td->o.rw_seq == RW_SEQ_SEQ) {
285 ret = get_next_seq_block(td, f, ddir, b);
287 ret = get_next_rand_block(td, f, ddir, b);
288 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
289 if (f->last_start != -1ULL)
290 *b = (f->last_start - f->file_offset)
291 / td->o.min_bs[ddir];
296 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
305 * For random io, generate a random new block and see if it's used. Repeat
306 * until we find a free one. For sequential io, just return the end of
307 * the last io issued.
309 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
311 struct fio_file *f = io_u->file;
312 unsigned long long b;
313 enum fio_ddir ddir = io_u->ddir;
316 assert(ddir_rw(ddir));
318 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
320 td->ddir_seq_nr = td->o.ddir_seq_nr;
323 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
326 io_u->offset = b * td->o.ba[ddir];
327 if (io_u->offset >= f->io_size) {
328 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
329 io_u->offset, f->io_size);
333 io_u->offset += f->file_offset;
334 if (io_u->offset >= f->real_file_size) {
335 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
336 io_u->offset, f->real_file_size);
343 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
345 struct prof_io_ops *ops = &td->prof_io_ops;
347 if (ops->fill_io_u_off)
348 return ops->fill_io_u_off(td, io_u);
350 return __get_next_offset(td, io_u);
353 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
356 struct fio_file *f = io_u->file;
358 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
361 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
363 const int ddir = io_u->ddir;
364 unsigned int uninitialized_var(buflen);
365 unsigned int minbs, maxbs;
366 unsigned long r, rand_max;
368 assert(ddir_rw(ddir));
370 minbs = td->o.min_bs[ddir];
371 maxbs = td->o.max_bs[ddir];
376 if (td->o.use_os_rand)
377 rand_max = OS_RAND_MAX;
379 rand_max = FRAND_MAX;
382 if (td->o.use_os_rand)
383 r = os_random_long(&td->bsrange_state);
385 r = __rand(&td->__bsrange_state);
387 if (!td->o.bssplit_nr[ddir]) {
388 buflen = 1 + (unsigned int) ((double) maxbs *
389 (r / (rand_max + 1.0)));
396 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
397 struct bssplit *bsp = &td->o.bssplit[ddir][i];
401 if ((r <= ((rand_max / 100L) * perc)) &&
402 io_u_fits(td, io_u, buflen))
407 if (!td->o.bs_unaligned && is_power_of_2(minbs))
408 buflen = (buflen + minbs - 1) & ~(minbs - 1);
410 } while (!io_u_fits(td, io_u, buflen));
415 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
417 struct prof_io_ops *ops = &td->prof_io_ops;
419 if (ops->fill_io_u_size)
420 return ops->fill_io_u_size(td, io_u);
422 return __get_next_buflen(td, io_u);
425 static void set_rwmix_bytes(struct thread_data *td)
430 * we do time or byte based switch. this is needed because
431 * buffered writes may issue a lot quicker than they complete,
432 * whereas reads do not.
434 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
435 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
438 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
443 if (td->o.use_os_rand) {
444 r = os_random_long(&td->rwmix_state);
445 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
447 r = __rand(&td->__rwmix_state);
448 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
451 if (v <= td->o.rwmix[DDIR_READ])
457 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
459 enum fio_ddir odir = ddir ^ 1;
463 assert(ddir_rw(ddir));
465 if (td->rate_pending_usleep[ddir] <= 0)
469 * We have too much pending sleep in this direction. See if we
474 * Other direction does not have too much pending, switch
476 if (td->rate_pending_usleep[odir] < 100000)
480 * Both directions have pending sleep. Sleep the minimum time
481 * and deduct from both.
483 if (td->rate_pending_usleep[ddir] <=
484 td->rate_pending_usleep[odir]) {
485 usec = td->rate_pending_usleep[ddir];
487 usec = td->rate_pending_usleep[odir];
491 usec = td->rate_pending_usleep[ddir];
494 * We are going to sleep, ensure that we flush anything pending as
495 * not to skew our latency numbers
500 ret = io_u_queued_complete(td, td->cur_depth, NULL);
503 fio_gettime(&t, NULL);
504 usec_sleep(td, usec);
505 usec = utime_since_now(&t);
507 td->rate_pending_usleep[ddir] -= usec;
510 if (td_rw(td) && __should_check_rate(td, odir))
511 td->rate_pending_usleep[odir] -= usec;
517 * Return the data direction for the next io_u. If the job is a
518 * mixed read/write workload, check the rwmix cycle and switch if
521 static enum fio_ddir get_rw_ddir(struct thread_data *td)
526 * see if it's time to fsync
528 if (td->o.fsync_blocks &&
529 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
530 td->io_issues[DDIR_WRITE] && should_fsync(td))
534 * see if it's time to fdatasync
536 if (td->o.fdatasync_blocks &&
537 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
538 td->io_issues[DDIR_WRITE] && should_fsync(td))
539 return DDIR_DATASYNC;
542 * see if it's time to sync_file_range
544 if (td->sync_file_range_nr &&
545 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
546 td->io_issues[DDIR_WRITE] && should_fsync(td))
547 return DDIR_SYNC_FILE_RANGE;
551 * Check if it's time to seed a new data direction.
553 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
555 * Put a top limit on how many bytes we do for
556 * one data direction, to avoid overflowing the
559 ddir = get_rand_ddir(td);
561 if (ddir != td->rwmix_ddir)
564 td->rwmix_ddir = ddir;
566 ddir = td->rwmix_ddir;
567 } else if (td_read(td))
572 td->rwmix_ddir = rate_ddir(td, ddir);
573 return td->rwmix_ddir;
576 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
578 io_u->ddir = get_rw_ddir(td);
580 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
581 td->o.barrier_blocks &&
582 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
583 td->io_issues[DDIR_WRITE])
584 io_u->flags |= IO_U_F_BARRIER;
587 void put_file_log(struct thread_data *td, struct fio_file *f)
589 int ret = put_file(td, f);
592 td_verror(td, ret, "file close");
595 void put_io_u(struct thread_data *td, struct io_u *io_u)
599 io_u->flags |= IO_U_F_FREE;
600 io_u->flags &= ~IO_U_F_FREE_DEF;
603 put_file_log(td, io_u->file);
606 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
608 flist_del_init(&io_u->list);
609 flist_add(&io_u->list, &td->io_u_freelist);
611 td_io_u_free_notify(td);
614 void clear_io_u(struct thread_data *td, struct io_u *io_u)
616 io_u->flags &= ~IO_U_F_FLIGHT;
620 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
622 struct io_u *__io_u = *io_u;
624 dprint(FD_IO, "requeue %p\n", __io_u);
628 __io_u->flags |= IO_U_F_FREE;
629 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
630 td->io_issues[__io_u->ddir]--;
632 __io_u->flags &= ~IO_U_F_FLIGHT;
633 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
635 flist_del(&__io_u->list);
636 flist_add_tail(&__io_u->list, &td->io_u_requeues);
641 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
643 if (td->io_ops->flags & FIO_NOIO)
646 set_rw_ddir(td, io_u);
649 * fsync() or fdatasync() or trim etc, we are done
651 if (!ddir_rw(io_u->ddir))
655 * See if it's time to switch to a new zone
657 if (td->zone_bytes >= td->o.zone_size) {
659 io_u->file->last_pos += td->o.zone_skip;
660 td->io_skip_bytes += td->o.zone_skip;
664 * No log, let the seq/rand engine retrieve the next buflen and
667 if (get_next_offset(td, io_u)) {
668 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
672 io_u->buflen = get_next_buflen(td, io_u);
674 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
678 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
679 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
680 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
681 io_u->buflen, io_u->file->real_file_size);
686 * mark entry before potentially trimming io_u
688 if (td_random(td) && file_randommap(td, io_u->file))
689 mark_random_map(td, io_u);
692 * If using a write iolog, store this entry.
695 dprint_io_u(io_u, "fill_io_u");
696 td->zone_bytes += io_u->buflen;
701 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
730 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
732 __io_u_mark_map(td->ts.io_u_submit, nr);
733 td->ts.total_submit++;
736 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
738 __io_u_mark_map(td->ts.io_u_complete, nr);
739 td->ts.total_complete++;
742 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
746 switch (td->cur_depth) {
768 td->ts.io_u_map[idx] += nr;
771 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
808 assert(idx < FIO_IO_U_LAT_U_NR);
809 td->ts.io_u_lat_u[idx]++;
812 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
853 assert(idx < FIO_IO_U_LAT_M_NR);
854 td->ts.io_u_lat_m[idx]++;
857 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
860 io_u_mark_lat_usec(td, usec);
862 io_u_mark_lat_msec(td, usec / 1000);
866 * Get next file to service by choosing one at random
868 static struct fio_file *get_next_file_rand(struct thread_data *td,
869 enum fio_file_flags goodf,
870 enum fio_file_flags badf)
879 if (td->o.use_os_rand) {
880 r = os_random_long(&td->next_file_state);
881 fno = (unsigned int) ((double) td->o.nr_files
882 * (r / (OS_RAND_MAX + 1.0)));
884 r = __rand(&td->__next_file_state);
885 fno = (unsigned int) ((double) td->o.nr_files
886 * (r / (FRAND_MAX + 1.0)));
890 if (fio_file_done(f))
893 if (!fio_file_open(f)) {
896 err = td_io_open_file(td, f);
902 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
903 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
907 td_io_close_file(td, f);
912 * Get next file to service by doing round robin between all available ones
914 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
917 unsigned int old_next_file = td->next_file;
923 f = td->files[td->next_file];
926 if (td->next_file >= td->o.nr_files)
929 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
930 if (fio_file_done(f)) {
935 if (!fio_file_open(f)) {
938 err = td_io_open_file(td, f);
940 dprint(FD_FILE, "error %d on open of %s\n",
948 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
950 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
954 td_io_close_file(td, f);
957 } while (td->next_file != old_next_file);
959 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
963 static struct fio_file *__get_next_file(struct thread_data *td)
967 assert(td->o.nr_files <= td->files_index);
969 if (td->nr_done_files >= td->o.nr_files) {
970 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
971 " nr_files=%d\n", td->nr_open_files,
977 f = td->file_service_file;
978 if (f && fio_file_open(f) && !fio_file_closing(f)) {
979 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
981 if (td->file_service_left--)
985 if (td->o.file_service_type == FIO_FSERVICE_RR ||
986 td->o.file_service_type == FIO_FSERVICE_SEQ)
987 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
989 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
991 td->file_service_file = f;
992 td->file_service_left = td->file_service_nr - 1;
994 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
998 static struct fio_file *get_next_file(struct thread_data *td)
1000 struct prof_io_ops *ops = &td->prof_io_ops;
1002 if (ops->get_next_file)
1003 return ops->get_next_file(td);
1005 return __get_next_file(td);
1008 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1013 f = get_next_file(td);
1020 if (!fill_io_u(td, io_u))
1023 put_file_log(td, f);
1024 td_io_close_file(td, f);
1026 fio_file_set_done(f);
1027 td->nr_done_files++;
1028 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1029 td->nr_done_files, td->o.nr_files);
1036 struct io_u *__get_io_u(struct thread_data *td)
1038 struct io_u *io_u = NULL;
1043 if (!flist_empty(&td->io_u_requeues))
1044 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1045 else if (!queue_full(td)) {
1046 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1051 io_u->end_io = NULL;
1055 assert(io_u->flags & IO_U_F_FREE);
1056 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1057 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1060 flist_del(&io_u->list);
1061 flist_add(&io_u->list, &td->io_u_busylist);
1063 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1064 } else if (td->o.verify_async) {
1066 * We ran out, wait for async verify threads to finish and
1069 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1077 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1079 if (td->o.trim_backlog && td->trim_entries) {
1082 if (td->trim_batch) {
1085 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1086 td->last_ddir != DDIR_READ) {
1087 td->trim_batch = td->o.trim_batch;
1088 if (!td->trim_batch)
1089 td->trim_batch = td->o.trim_backlog;
1093 if (get_trim && !get_next_trim(td, io_u))
1100 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1102 if (td->o.verify_backlog && td->io_hist_len) {
1105 if (td->verify_batch) {
1108 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1109 td->last_ddir != DDIR_READ) {
1110 td->verify_batch = td->o.verify_batch;
1111 if (!td->verify_batch)
1112 td->verify_batch = td->o.verify_backlog;
1116 if (get_verify && !get_next_verify(td, io_u))
1124 * Fill offset and start time into the buffer content, to prevent too
1125 * easy compressible data for simple de-dupe attempts. Do this for every
1126 * 512b block in the range, since that should be the smallest block size
1127 * we can expect from a device.
1129 static void small_content_scramble(struct io_u *io_u)
1131 unsigned int i, nr_blocks = io_u->buflen / 512;
1132 unsigned long long boffset;
1133 unsigned int offset;
1140 boffset = io_u->offset;
1142 for (i = 0; i < nr_blocks; i++) {
1144 * Fill the byte offset into a "random" start offset of
1145 * the buffer, given by the product of the usec time
1146 * and the actual offset.
1148 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1149 offset &= ~(sizeof(unsigned long long) - 1);
1150 if (offset >= 512 - sizeof(unsigned long long))
1151 offset -= sizeof(unsigned long long);
1152 memcpy(p + offset, &boffset, sizeof(boffset));
1154 end = p + 512 - sizeof(io_u->start_time);
1155 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1162 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1163 * etc. The returned io_u is fully ready to be prepped and submitted.
1165 struct io_u *get_io_u(struct thread_data *td)
1169 int do_scramble = 0;
1171 io_u = __get_io_u(td);
1173 dprint(FD_IO, "__get_io_u failed\n");
1177 if (check_get_verify(td, io_u))
1179 if (check_get_trim(td, io_u))
1183 * from a requeue, io_u already setup
1189 * If using an iolog, grab next piece if any available.
1191 if (td->o.read_iolog_file) {
1192 if (read_iolog_get(td, io_u))
1194 } else if (set_io_u_file(td, io_u)) {
1195 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1200 assert(fio_file_open(f));
1202 if (ddir_rw(io_u->ddir)) {
1203 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1204 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1208 f->last_start = io_u->offset;
1209 f->last_pos = io_u->offset + io_u->buflen;
1211 if (io_u->ddir == DDIR_WRITE) {
1212 if (td->o.verify != VERIFY_NONE)
1213 populate_verify_io_u(td, io_u);
1214 else if (td->o.refill_buffers)
1215 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1216 else if (td->o.scramble_buffers)
1218 } else if (io_u->ddir == DDIR_READ) {
1220 * Reset the buf_filled parameters so next time if the
1221 * buffer is used for writes it is refilled.
1223 io_u->buf_filled_len = 0;
1228 * Set io data pointers.
1230 io_u->xfer_buf = io_u->buf;
1231 io_u->xfer_buflen = io_u->buflen;
1235 if (!td_io_prep(td, io_u)) {
1236 if (!td->o.disable_slat)
1237 fio_gettime(&io_u->start_time, NULL);
1239 small_content_scramble(io_u);
1243 dprint(FD_IO, "get_io_u failed\n");
1248 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1250 const char *msg[] = { "read", "write", "sync", "datasync",
1251 "sync_file_range", "wait", "trim" };
1255 log_err("fio: io_u error");
1258 log_err(" on file %s", io_u->file->file_name);
1260 log_err(": %s\n", strerror(io_u->error));
1262 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1263 io_u->offset, io_u->xfer_buflen);
1266 td_verror(td, io_u->error, "io_u error");
1269 static void io_completed(struct thread_data *td, struct io_u *io_u,
1270 struct io_completion_data *icd)
1273 * Older gcc's are too dumb to realize that usec is always used
1274 * initialized, silence that warning.
1276 unsigned long uninitialized_var(usec);
1279 dprint_io_u(io_u, "io complete");
1282 assert(io_u->flags & IO_U_F_FLIGHT);
1283 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1286 if (ddir_sync(io_u->ddir)) {
1287 td->last_was_sync = 1;
1290 f->first_write = -1ULL;
1291 f->last_write = -1ULL;
1296 td->last_was_sync = 0;
1297 td->last_ddir = io_u->ddir;
1299 if (!io_u->error && ddir_rw(io_u->ddir)) {
1300 unsigned int bytes = io_u->buflen - io_u->resid;
1301 const enum fio_ddir idx = io_u->ddir;
1302 const enum fio_ddir odx = io_u->ddir ^ 1;
1305 td->io_blocks[idx]++;
1306 td->io_bytes[idx] += bytes;
1307 td->this_io_bytes[idx] += bytes;
1309 if (idx == DDIR_WRITE) {
1312 if (f->first_write == -1ULL ||
1313 io_u->offset < f->first_write)
1314 f->first_write = io_u->offset;
1315 if (f->last_write == -1ULL ||
1316 ((io_u->offset + bytes) > f->last_write))
1317 f->last_write = io_u->offset + bytes;
1321 if (ramp_time_over(td)) {
1322 unsigned long uninitialized_var(lusec);
1324 if (!td->o.disable_clat || !td->o.disable_bw)
1325 lusec = utime_since(&io_u->issue_time,
1327 if (!td->o.disable_lat) {
1328 unsigned long tusec;
1330 tusec = utime_since(&io_u->start_time,
1332 add_lat_sample(td, idx, tusec, bytes);
1334 if (!td->o.disable_clat) {
1335 add_clat_sample(td, idx, lusec, bytes);
1336 io_u_mark_latency(td, lusec);
1338 if (!td->o.disable_bw)
1339 add_bw_sample(td, idx, bytes, &icd->time);
1340 if (__should_check_rate(td, idx)) {
1341 td->rate_pending_usleep[idx] =
1342 ((td->this_io_bytes[idx] *
1343 td->rate_nsec_cycle[idx]) / 1000 -
1344 utime_since_now(&td->start));
1346 if (__should_check_rate(td, idx ^ 1))
1347 td->rate_pending_usleep[odx] =
1348 ((td->this_io_bytes[odx] *
1349 td->rate_nsec_cycle[odx]) / 1000 -
1350 utime_since_now(&td->start));
1353 if (td_write(td) && idx == DDIR_WRITE &&
1355 td->o.verify != VERIFY_NONE)
1356 log_io_piece(td, io_u);
1358 icd->bytes_done[idx] += bytes;
1361 ret = io_u->end_io(td, io_u);
1362 if (ret && !icd->error)
1365 } else if (io_u->error) {
1366 icd->error = io_u->error;
1367 io_u_log_error(td, io_u);
1369 if (td->o.continue_on_error && icd->error &&
1370 td_non_fatal_error(icd->error)) {
1372 * If there is a non_fatal error, then add to the error count
1373 * and clear all the errors.
1375 update_error_count(td, icd->error);
1382 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1385 if (!td->o.disable_clat || !td->o.disable_bw)
1386 fio_gettime(&icd->time, NULL);
1391 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1394 static void ios_completed(struct thread_data *td,
1395 struct io_completion_data *icd)
1400 for (i = 0; i < icd->nr; i++) {
1401 io_u = td->io_ops->event(td, i);
1403 io_completed(td, io_u, icd);
1405 if (!(io_u->flags & IO_U_F_FREE_DEF))
1411 * Complete a single io_u for the sync engines.
1413 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1414 unsigned long *bytes)
1416 struct io_completion_data icd;
1418 init_icd(td, &icd, 1);
1419 io_completed(td, io_u, &icd);
1421 if (!(io_u->flags & IO_U_F_FREE_DEF))
1425 td_verror(td, icd.error, "io_u_sync_complete");
1430 bytes[0] += icd.bytes_done[0];
1431 bytes[1] += icd.bytes_done[1];
1438 * Called to complete min_events number of io for the async engines.
1440 int io_u_queued_complete(struct thread_data *td, int min_evts,
1441 unsigned long *bytes)
1443 struct io_completion_data icd;
1444 struct timespec *tvp = NULL;
1446 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1448 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1453 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1455 td_verror(td, -ret, "td_io_getevents");
1460 init_icd(td, &icd, ret);
1461 ios_completed(td, &icd);
1463 td_verror(td, icd.error, "io_u_queued_complete");
1468 bytes[0] += icd.bytes_done[0];
1469 bytes[1] += icd.bytes_done[1];
1476 * Call when io_u is really queued, to update the submission latency.
1478 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1480 if (!td->o.disable_slat) {
1481 unsigned long slat_time;
1483 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1484 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1489 * "randomly" fill the buffer contents
1491 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1492 unsigned int max_bs)
1494 io_u->buf_filled_len = 0;
1496 if (!td->o.zero_buffers)
1497 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1499 memset(io_u->buf, 0, max_bs);