13 #include "lib/axmap.h"
15 struct io_completion_data {
18 int error; /* output */
19 unsigned long bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timeval time; /* output */
24 * The ->io_axmap 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 uint64_t block)
29 return !axmap_isset(f->io_axmap, block);
33 * Mark a given offset as used in the map.
35 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
37 unsigned int min_bs = td->o.rw_min_bs;
38 struct fio_file *f = io_u->file;
39 unsigned int nr_blocks;
42 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
43 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
45 if (!(io_u->flags & IO_U_F_BUSY_OK))
46 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
48 if ((nr_blocks * min_bs) < io_u->buflen)
49 io_u->buflen = nr_blocks * min_bs;
52 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
58 assert(ddir_rw(ddir));
61 * Hmm, should we make sure that ->io_size <= ->real_file_size?
63 max_size = f->io_size;
64 if (max_size > f->real_file_size)
65 max_size = f->real_file_size;
68 max_size = td->o.zone_range;
70 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
78 struct flist_head list;
82 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
83 enum fio_ddir ddir, uint64_t *b)
87 lastb = last_block(td, f, ddir);
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
94 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
96 if (td->o.use_os_rand) {
98 r = os_random_long(&td->random_state);
101 r = __rand(&td->__random_state);
104 dprint(FD_RANDOM, "off rand %llu\n", r);
106 *b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
110 if (lfsr_next(&f->lfsr, &off, lastb))
117 * if we are not maintaining a random map, we are done.
119 if (!file_randommap(td, f))
123 * calculate map offset and check if it's free
125 if (random_map_free(f, *b))
128 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", *b);
130 *b = axmap_next_free(f->io_axmap, *b);
131 if (*b == (uint64_t) -1ULL)
137 static int __get_next_rand_offset_zipf(struct thread_data *td,
138 struct fio_file *f, enum fio_ddir ddir,
141 *b = zipf_next(&f->zipf);
145 static int __get_next_rand_offset_pareto(struct thread_data *td,
146 struct fio_file *f, enum fio_ddir ddir,
149 *b = pareto_next(&f->zipf);
153 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
155 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
156 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
158 return r1->off - r2->off;
161 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
162 enum fio_ddir ddir, uint64_t *b)
164 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
165 return __get_next_rand_offset(td, f, ddir, b);
166 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
167 return __get_next_rand_offset_zipf(td, f, ddir, b);
168 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
169 return __get_next_rand_offset_pareto(td, f, ddir, b);
171 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
175 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
176 enum fio_ddir ddir, uint64_t *b)
182 * If sort not enabled, or not a pure random read workload without
183 * any stored write metadata, just return a random offset
185 if (!td->o.verifysort_nr || !(ddir == DDIR_READ && td->o.do_verify &&
186 td->o.verify != VERIFY_NONE && td_random(td)))
187 return get_off_from_method(td, f, ddir, b);
189 if (!flist_empty(&td->next_rand_list)) {
192 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
199 for (i = 0; i < td->o.verifysort_nr; i++) {
200 r = malloc(sizeof(*r));
202 ret = get_off_from_method(td, f, ddir, &r->off);
208 flist_add(&r->list, &td->next_rand_list);
214 assert(!flist_empty(&td->next_rand_list));
215 flist_sort(NULL, &td->next_rand_list, flist_cmp);
219 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
220 enum fio_ddir ddir, uint64_t *b)
222 if (!get_next_rand_offset(td, f, ddir, b))
225 if (td->o.time_based) {
227 if (!get_next_rand_offset(td, f, ddir, b))
231 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
232 f->file_name, f->last_pos, f->real_file_size);
236 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
237 enum fio_ddir ddir, uint64_t *offset)
239 assert(ddir_rw(ddir));
241 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
242 f->last_pos = f->last_pos - f->io_size;
244 if (f->last_pos < f->real_file_size) {
247 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
248 f->last_pos = f->real_file_size;
250 pos = f->last_pos - f->file_offset;
252 pos += td->o.ddir_seq_add;
261 static int get_next_block(struct thread_data *td, struct io_u *io_u,
262 enum fio_ddir ddir, int rw_seq)
264 struct fio_file *f = io_u->file;
268 assert(ddir_rw(ddir));
274 ret = get_next_rand_block(td, f, ddir, &b);
276 ret = get_next_seq_offset(td, f, ddir, &offset);
278 io_u->flags |= IO_U_F_BUSY_OK;
280 if (td->o.rw_seq == RW_SEQ_SEQ) {
281 ret = get_next_seq_offset(td, f, ddir, &offset);
283 ret = get_next_rand_block(td, f, ddir, &b);
284 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
285 if (f->last_start != -1ULL)
286 offset = f->last_start - f->file_offset;
291 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
298 io_u->offset = offset;
300 io_u->offset = b * td->o.ba[ddir];
302 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
312 * For random io, generate a random new block and see if it's used. Repeat
313 * until we find a free one. For sequential io, just return the end of
314 * the last io issued.
316 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
318 struct fio_file *f = io_u->file;
319 enum fio_ddir ddir = io_u->ddir;
322 assert(ddir_rw(ddir));
324 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
326 td->ddir_seq_nr = td->o.ddir_seq_nr;
329 if (get_next_block(td, io_u, ddir, rw_seq_hit))
332 if (io_u->offset >= f->io_size) {
333 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
334 io_u->offset, f->io_size);
338 io_u->offset += f->file_offset;
339 if (io_u->offset >= f->real_file_size) {
340 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
341 io_u->offset, f->real_file_size);
348 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
350 if (td->flags & TD_F_PROFILE_OPS) {
351 struct prof_io_ops *ops = &td->prof_io_ops;
353 if (ops->fill_io_u_off)
354 return ops->fill_io_u_off(td, io_u);
357 return __get_next_offset(td, io_u);
360 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
363 struct fio_file *f = io_u->file;
365 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
368 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
370 const int ddir = io_u->ddir;
371 unsigned int buflen = 0;
372 unsigned int minbs, maxbs;
373 unsigned long r, rand_max;
375 assert(ddir_rw(ddir));
377 minbs = td->o.min_bs[ddir];
378 maxbs = td->o.max_bs[ddir];
384 * If we can't satisfy the min block size from here, then fail
386 if (!io_u_fits(td, io_u, minbs))
389 if (td->o.use_os_rand)
390 rand_max = OS_RAND_MAX;
392 rand_max = FRAND_MAX;
395 if (td->o.use_os_rand)
396 r = os_random_long(&td->bsrange_state);
398 r = __rand(&td->__bsrange_state);
400 if (!td->o.bssplit_nr[ddir]) {
401 buflen = 1 + (unsigned int) ((double) maxbs *
402 (r / (rand_max + 1.0)));
409 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
410 struct bssplit *bsp = &td->o.bssplit[ddir][i];
414 if ((r <= ((rand_max / 100L) * perc)) &&
415 io_u_fits(td, io_u, buflen))
420 if (!td->o.bs_unaligned && is_power_of_2(minbs))
421 buflen = (buflen + minbs - 1) & ~(minbs - 1);
423 } while (!io_u_fits(td, io_u, buflen));
428 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
430 if (td->flags & TD_F_PROFILE_OPS) {
431 struct prof_io_ops *ops = &td->prof_io_ops;
433 if (ops->fill_io_u_size)
434 return ops->fill_io_u_size(td, io_u);
437 return __get_next_buflen(td, io_u);
440 static void set_rwmix_bytes(struct thread_data *td)
445 * we do time or byte based switch. this is needed because
446 * buffered writes may issue a lot quicker than they complete,
447 * whereas reads do not.
449 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
450 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
453 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
458 if (td->o.use_os_rand) {
459 r = os_random_long(&td->rwmix_state);
460 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
462 r = __rand(&td->__rwmix_state);
463 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
466 if (v <= td->o.rwmix[DDIR_READ])
472 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
474 enum fio_ddir odir = ddir ^ 1;
478 assert(ddir_rw(ddir));
480 if (td->rate_pending_usleep[ddir] <= 0)
484 * We have too much pending sleep in this direction. See if we
489 * Other direction does not have too much pending, switch
491 if (td->rate_pending_usleep[odir] < 100000)
495 * Both directions have pending sleep. Sleep the minimum time
496 * and deduct from both.
498 if (td->rate_pending_usleep[ddir] <=
499 td->rate_pending_usleep[odir]) {
500 usec = td->rate_pending_usleep[ddir];
502 usec = td->rate_pending_usleep[odir];
506 usec = td->rate_pending_usleep[ddir];
509 * We are going to sleep, ensure that we flush anything pending as
510 * not to skew our latency numbers.
512 * Changed to only monitor 'in flight' requests here instead of the
513 * td->cur_depth, b/c td->cur_depth does not accurately represent
514 * io's that have been actually submitted to an async engine,
515 * and cur_depth is meaningless for sync engines.
517 if (td->io_u_in_flight) {
520 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
523 fio_gettime(&t, NULL);
524 usec_sleep(td, usec);
525 usec = utime_since_now(&t);
527 td->rate_pending_usleep[ddir] -= usec;
530 if (td_rw(td) && __should_check_rate(td, odir))
531 td->rate_pending_usleep[odir] -= usec;
539 * Return the data direction for the next io_u. If the job is a
540 * mixed read/write workload, check the rwmix cycle and switch if
543 static enum fio_ddir get_rw_ddir(struct thread_data *td)
548 * see if it's time to fsync
550 if (td->o.fsync_blocks &&
551 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
552 td->io_issues[DDIR_WRITE] && should_fsync(td))
556 * see if it's time to fdatasync
558 if (td->o.fdatasync_blocks &&
559 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
560 td->io_issues[DDIR_WRITE] && should_fsync(td))
561 return DDIR_DATASYNC;
564 * see if it's time to sync_file_range
566 if (td->sync_file_range_nr &&
567 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
568 td->io_issues[DDIR_WRITE] && should_fsync(td))
569 return DDIR_SYNC_FILE_RANGE;
573 * Check if it's time to seed a new data direction.
575 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
577 * Put a top limit on how many bytes we do for
578 * one data direction, to avoid overflowing the
581 ddir = get_rand_ddir(td);
583 if (ddir != td->rwmix_ddir)
586 td->rwmix_ddir = ddir;
588 ddir = td->rwmix_ddir;
589 } else if (td_read(td))
591 else if (td_write(td))
596 td->rwmix_ddir = rate_ddir(td, ddir);
597 return td->rwmix_ddir;
600 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
602 io_u->ddir = get_rw_ddir(td);
604 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
605 td->o.barrier_blocks &&
606 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
607 td->io_issues[DDIR_WRITE])
608 io_u->flags |= IO_U_F_BARRIER;
611 void put_file_log(struct thread_data *td, struct fio_file *f)
613 int ret = put_file(td, f);
616 td_verror(td, ret, "file close");
619 void put_io_u(struct thread_data *td, struct io_u *io_u)
623 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
624 put_file_log(td, io_u->file);
626 io_u->flags &= ~IO_U_F_FREE_DEF;
627 io_u->flags |= IO_U_F_FREE;
629 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
631 flist_del_init(&io_u->list);
632 flist_add(&io_u->list, &td->io_u_freelist);
634 td_io_u_free_notify(td);
637 void clear_io_u(struct thread_data *td, struct io_u *io_u)
639 io_u->flags &= ~IO_U_F_FLIGHT;
643 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
645 struct io_u *__io_u = *io_u;
647 dprint(FD_IO, "requeue %p\n", __io_u);
651 __io_u->flags |= IO_U_F_FREE;
652 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
653 td->io_issues[__io_u->ddir]--;
655 __io_u->flags &= ~IO_U_F_FLIGHT;
656 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
658 flist_del(&__io_u->list);
659 flist_add_tail(&__io_u->list, &td->io_u_requeues);
664 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
666 if (td->io_ops->flags & FIO_NOIO)
669 set_rw_ddir(td, io_u);
672 * fsync() or fdatasync() or trim etc, we are done
674 if (!ddir_rw(io_u->ddir))
678 * See if it's time to switch to a new zone
680 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
682 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
683 io_u->file->last_pos = io_u->file->file_offset;
684 td->io_skip_bytes += td->o.zone_skip;
688 * No log, let the seq/rand engine retrieve the next buflen and
691 if (get_next_offset(td, io_u)) {
692 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
696 io_u->buflen = get_next_buflen(td, io_u);
698 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
702 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
703 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
704 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
705 io_u->buflen, io_u->file->real_file_size);
710 * mark entry before potentially trimming io_u
712 if (td_random(td) && file_randommap(td, io_u->file))
713 mark_random_map(td, io_u);
716 * If using a write iolog, store this entry.
719 dprint_io_u(io_u, "fill_io_u");
720 td->zone_bytes += io_u->buflen;
725 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
754 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
756 __io_u_mark_map(td->ts.io_u_submit, nr);
757 td->ts.total_submit++;
760 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
762 __io_u_mark_map(td->ts.io_u_complete, nr);
763 td->ts.total_complete++;
766 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
770 switch (td->cur_depth) {
792 td->ts.io_u_map[idx] += nr;
795 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
832 assert(idx < FIO_IO_U_LAT_U_NR);
833 td->ts.io_u_lat_u[idx]++;
836 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
877 assert(idx < FIO_IO_U_LAT_M_NR);
878 td->ts.io_u_lat_m[idx]++;
881 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
884 io_u_mark_lat_usec(td, usec);
886 io_u_mark_lat_msec(td, usec / 1000);
890 * Get next file to service by choosing one at random
892 static struct fio_file *get_next_file_rand(struct thread_data *td,
893 enum fio_file_flags goodf,
894 enum fio_file_flags badf)
903 if (td->o.use_os_rand) {
904 r = os_random_long(&td->next_file_state);
905 fno = (unsigned int) ((double) td->o.nr_files
906 * (r / (OS_RAND_MAX + 1.0)));
908 r = __rand(&td->__next_file_state);
909 fno = (unsigned int) ((double) td->o.nr_files
910 * (r / (FRAND_MAX + 1.0)));
914 if (fio_file_done(f))
917 if (!fio_file_open(f)) {
920 err = td_io_open_file(td, f);
926 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
927 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
931 td_io_close_file(td, f);
936 * Get next file to service by doing round robin between all available ones
938 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
941 unsigned int old_next_file = td->next_file;
947 f = td->files[td->next_file];
950 if (td->next_file >= td->o.nr_files)
953 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
954 if (fio_file_done(f)) {
959 if (!fio_file_open(f)) {
962 err = td_io_open_file(td, f);
964 dprint(FD_FILE, "error %d on open of %s\n",
972 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
974 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
978 td_io_close_file(td, f);
981 } while (td->next_file != old_next_file);
983 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
987 static struct fio_file *__get_next_file(struct thread_data *td)
991 assert(td->o.nr_files <= td->files_index);
993 if (td->nr_done_files >= td->o.nr_files) {
994 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
995 " nr_files=%d\n", td->nr_open_files,
1001 f = td->file_service_file;
1002 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1003 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1005 if (td->file_service_left--)
1009 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1010 td->o.file_service_type == FIO_FSERVICE_SEQ)
1011 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1013 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1015 td->file_service_file = f;
1016 td->file_service_left = td->file_service_nr - 1;
1018 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1022 static struct fio_file *get_next_file(struct thread_data *td)
1024 if (!(td->flags & TD_F_PROFILE_OPS)) {
1025 struct prof_io_ops *ops = &td->prof_io_ops;
1027 if (ops->get_next_file)
1028 return ops->get_next_file(td);
1031 return __get_next_file(td);
1034 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1039 f = get_next_file(td);
1046 if (!fill_io_u(td, io_u))
1049 put_file_log(td, f);
1050 td_io_close_file(td, f);
1052 fio_file_set_done(f);
1053 td->nr_done_files++;
1054 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1055 td->nr_done_files, td->o.nr_files);
1062 struct io_u *__get_io_u(struct thread_data *td)
1064 struct io_u *io_u = NULL;
1069 if (!flist_empty(&td->io_u_requeues))
1070 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1071 else if (!queue_full(td)) {
1072 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1077 io_u->end_io = NULL;
1081 assert(io_u->flags & IO_U_F_FREE);
1082 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1083 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1084 io_u->flags &= ~IO_U_F_VER_LIST;
1087 flist_del(&io_u->list);
1088 flist_add_tail(&io_u->list, &td->io_u_busylist);
1090 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1091 } else if (td->o.verify_async) {
1093 * We ran out, wait for async verify threads to finish and
1096 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1104 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1106 if (!(td->flags & TD_F_TRIM_BACKLOG))
1109 if (td->trim_entries) {
1112 if (td->trim_batch) {
1115 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1116 td->last_ddir != DDIR_READ) {
1117 td->trim_batch = td->o.trim_batch;
1118 if (!td->trim_batch)
1119 td->trim_batch = td->o.trim_backlog;
1123 if (get_trim && !get_next_trim(td, io_u))
1130 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1132 if (!(td->flags & TD_F_VER_BACKLOG))
1135 if (td->io_hist_len) {
1138 if (td->verify_batch)
1140 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1141 td->last_ddir != DDIR_READ) {
1142 td->verify_batch = td->o.verify_batch;
1143 if (!td->verify_batch)
1144 td->verify_batch = td->o.verify_backlog;
1148 if (get_verify && !get_next_verify(td, io_u)) {
1158 * Fill offset and start time into the buffer content, to prevent too
1159 * easy compressible data for simple de-dupe attempts. Do this for every
1160 * 512b block in the range, since that should be the smallest block size
1161 * we can expect from a device.
1163 static void small_content_scramble(struct io_u *io_u)
1165 unsigned int i, nr_blocks = io_u->buflen / 512;
1167 unsigned int offset;
1174 boffset = io_u->offset;
1175 io_u->buf_filled_len = 0;
1177 for (i = 0; i < nr_blocks; i++) {
1179 * Fill the byte offset into a "random" start offset of
1180 * the buffer, given by the product of the usec time
1181 * and the actual offset.
1183 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1184 offset &= ~(sizeof(uint64_t) - 1);
1185 if (offset >= 512 - sizeof(uint64_t))
1186 offset -= sizeof(uint64_t);
1187 memcpy(p + offset, &boffset, sizeof(boffset));
1189 end = p + 512 - sizeof(io_u->start_time);
1190 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1197 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1198 * etc. The returned io_u is fully ready to be prepped and submitted.
1200 struct io_u *get_io_u(struct thread_data *td)
1204 int do_scramble = 0;
1206 io_u = __get_io_u(td);
1208 dprint(FD_IO, "__get_io_u failed\n");
1212 if (check_get_verify(td, io_u))
1214 if (check_get_trim(td, io_u))
1218 * from a requeue, io_u already setup
1224 * If using an iolog, grab next piece if any available.
1226 if (td->flags & TD_F_READ_IOLOG) {
1227 if (read_iolog_get(td, io_u))
1229 } else if (set_io_u_file(td, io_u)) {
1230 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1235 assert(fio_file_open(f));
1237 if (ddir_rw(io_u->ddir)) {
1238 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1239 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1243 f->last_start = io_u->offset;
1244 f->last_pos = io_u->offset + io_u->buflen;
1246 if (io_u->ddir == DDIR_WRITE) {
1247 if (td->flags & TD_F_REFILL_BUFFERS) {
1248 io_u_fill_buffer(td, io_u,
1249 io_u->xfer_buflen, io_u->xfer_buflen);
1250 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1252 if (td->flags & TD_F_VER_NONE) {
1253 populate_verify_io_u(td, io_u);
1256 } else if (io_u->ddir == DDIR_READ) {
1258 * Reset the buf_filled parameters so next time if the
1259 * buffer is used for writes it is refilled.
1261 io_u->buf_filled_len = 0;
1266 * Set io data pointers.
1268 io_u->xfer_buf = io_u->buf;
1269 io_u->xfer_buflen = io_u->buflen;
1273 if (!td_io_prep(td, io_u)) {
1274 if (!td->o.disable_slat)
1275 fio_gettime(&io_u->start_time, NULL);
1277 small_content_scramble(io_u);
1281 dprint(FD_IO, "get_io_u failed\n");
1286 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1288 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1289 const char *msg[] = { "read", "write", "sync", "datasync",
1290 "sync_file_range", "wait", "trim" };
1292 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1295 log_err("fio: io_u error");
1298 log_err(" on file %s", io_u->file->file_name);
1300 log_err(": %s\n", strerror(io_u->error));
1302 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1303 io_u->offset, io_u->xfer_buflen);
1306 td_verror(td, io_u->error, "io_u error");
1309 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1310 struct io_completion_data *icd,
1311 const enum fio_ddir idx, unsigned int bytes)
1313 unsigned long lusec = 0;
1315 if (!td->o.disable_clat || !td->o.disable_bw)
1316 lusec = utime_since(&io_u->issue_time, &icd->time);
1318 if (!td->o.disable_lat) {
1319 unsigned long tusec;
1321 tusec = utime_since(&io_u->start_time, &icd->time);
1322 add_lat_sample(td, idx, tusec, bytes);
1324 if (td->o.max_latency && tusec > td->o.max_latency) {
1326 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1327 td_verror(td, ETIMEDOUT, "max latency exceeded");
1328 icd->error = ETIMEDOUT;
1332 if (!td->o.disable_clat) {
1333 add_clat_sample(td, idx, lusec, bytes);
1334 io_u_mark_latency(td, lusec);
1337 if (!td->o.disable_bw)
1338 add_bw_sample(td, idx, bytes, &icd->time);
1340 add_iops_sample(td, idx, &icd->time);
1343 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1345 uint64_t secs, remainder, bps, bytes;
1347 bytes = td->this_io_bytes[ddir];
1348 bps = td->rate_bps[ddir];
1350 remainder = bytes % bps;
1351 return remainder * 1000000 / bps + secs * 1000000;
1354 static void io_completed(struct thread_data *td, struct io_u *io_u,
1355 struct io_completion_data *icd)
1359 dprint_io_u(io_u, "io complete");
1362 assert(io_u->flags & IO_U_F_FLIGHT);
1363 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1366 if (ddir_sync(io_u->ddir)) {
1367 td->last_was_sync = 1;
1370 f->first_write = -1ULL;
1371 f->last_write = -1ULL;
1376 td->last_was_sync = 0;
1377 td->last_ddir = io_u->ddir;
1379 if (!io_u->error && ddir_rw(io_u->ddir)) {
1380 unsigned int bytes = io_u->buflen - io_u->resid;
1381 const enum fio_ddir idx = io_u->ddir;
1382 const enum fio_ddir odx = io_u->ddir ^ 1;
1385 td->io_blocks[idx]++;
1386 td->this_io_blocks[idx]++;
1387 td->io_bytes[idx] += bytes;
1389 if (!(io_u->flags & IO_U_F_VER_LIST))
1390 td->this_io_bytes[idx] += bytes;
1392 if (idx == DDIR_WRITE) {
1395 if (f->first_write == -1ULL ||
1396 io_u->offset < f->first_write)
1397 f->first_write = io_u->offset;
1398 if (f->last_write == -1ULL ||
1399 ((io_u->offset + bytes) > f->last_write))
1400 f->last_write = io_u->offset + bytes;
1404 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1405 td->runstate == TD_VERIFYING)) {
1406 account_io_completion(td, io_u, icd, idx, bytes);
1408 if (__should_check_rate(td, idx)) {
1409 td->rate_pending_usleep[idx] =
1410 (usec_for_io(td, idx) -
1411 utime_since_now(&td->start));
1413 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1414 td->rate_pending_usleep[odx] =
1415 (usec_for_io(td, odx) -
1416 utime_since_now(&td->start));
1419 if (td_write(td) && idx == DDIR_WRITE &&
1421 td->o.verify != VERIFY_NONE)
1422 log_io_piece(td, io_u);
1424 icd->bytes_done[idx] += bytes;
1427 ret = io_u->end_io(td, io_u);
1428 if (ret && !icd->error)
1431 } else if (io_u->error) {
1432 icd->error = io_u->error;
1433 io_u_log_error(td, io_u);
1436 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1437 if (!td_non_fatal_error(td, eb, icd->error))
1440 * If there is a non_fatal error, then add to the error count
1441 * and clear all the errors.
1443 update_error_count(td, icd->error);
1450 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1454 if (!td->o.disable_clat || !td->o.disable_bw)
1455 fio_gettime(&icd->time, NULL);
1460 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1461 icd->bytes_done[ddir] = 0;
1464 static void ios_completed(struct thread_data *td,
1465 struct io_completion_data *icd)
1470 for (i = 0; i < icd->nr; i++) {
1471 io_u = td->io_ops->event(td, i);
1473 io_completed(td, io_u, icd);
1475 if (!(io_u->flags & IO_U_F_FREE_DEF))
1481 * Complete a single io_u for the sync engines.
1483 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1484 unsigned long *bytes)
1486 struct io_completion_data icd;
1488 init_icd(td, &icd, 1);
1489 io_completed(td, io_u, &icd);
1491 if (!(io_u->flags & IO_U_F_FREE_DEF))
1495 td_verror(td, icd.error, "io_u_sync_complete");
1502 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1503 bytes[ddir] += icd.bytes_done[ddir];
1510 * Called to complete min_events number of io for the async engines.
1512 int io_u_queued_complete(struct thread_data *td, int min_evts,
1513 unsigned long *bytes)
1515 struct io_completion_data icd;
1516 struct timespec *tvp = NULL;
1518 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1520 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1525 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1527 td_verror(td, -ret, "td_io_getevents");
1532 init_icd(td, &icd, ret);
1533 ios_completed(td, &icd);
1535 td_verror(td, icd.error, "io_u_queued_complete");
1542 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1543 bytes[ddir] += icd.bytes_done[ddir];
1550 * Call when io_u is really queued, to update the submission latency.
1552 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1554 if (!td->o.disable_slat) {
1555 unsigned long slat_time;
1557 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1558 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1563 * "randomly" fill the buffer contents
1565 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1566 unsigned int min_write, unsigned int max_bs)
1568 io_u->buf_filled_len = 0;
1570 if (!td->o.zero_buffers) {
1571 unsigned int perc = td->o.compress_percentage;
1574 unsigned int seg = min_write;
1576 seg = min(min_write, td->o.compress_chunk);
1577 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1580 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1582 memset(io_u->buf, 0, max_bs);