13 #include "lib/axmap.h"
15 struct io_completion_data {
18 int error; /* output */
19 uint64_t 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);
176 * Sort the reads for a verify phase in batches of verifysort_nr, if
179 static inline int should_sort_io(struct thread_data *td)
181 if (!td->o.verifysort_nr || !td->o.do_verify)
185 if (td->runstate != TD_VERIFYING)
187 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
193 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
194 enum fio_ddir ddir, uint64_t *b)
199 if (!should_sort_io(td))
200 return get_off_from_method(td, f, ddir, b);
202 if (!flist_empty(&td->next_rand_list)) {
205 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
212 for (i = 0; i < td->o.verifysort_nr; i++) {
213 r = malloc(sizeof(*r));
215 ret = get_off_from_method(td, f, ddir, &r->off);
221 flist_add(&r->list, &td->next_rand_list);
227 assert(!flist_empty(&td->next_rand_list));
228 flist_sort(NULL, &td->next_rand_list, flist_cmp);
232 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
233 enum fio_ddir ddir, uint64_t *b)
235 if (!get_next_rand_offset(td, f, ddir, b))
238 if (td->o.time_based) {
239 fio_file_reset(td, f);
240 if (!get_next_rand_offset(td, f, ddir, b))
244 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
245 f->file_name, f->last_pos, f->real_file_size);
249 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
250 enum fio_ddir ddir, uint64_t *offset)
252 assert(ddir_rw(ddir));
254 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
255 f->last_pos = f->last_pos - f->io_size;
257 if (f->last_pos < f->real_file_size) {
260 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
261 f->last_pos = f->real_file_size;
263 pos = f->last_pos - f->file_offset;
265 pos += td->o.ddir_seq_add;
274 static int get_next_block(struct thread_data *td, struct io_u *io_u,
275 enum fio_ddir ddir, int rw_seq)
277 struct fio_file *f = io_u->file;
281 assert(ddir_rw(ddir));
287 ret = get_next_rand_block(td, f, ddir, &b);
289 ret = get_next_seq_offset(td, f, ddir, &offset);
291 io_u->flags |= IO_U_F_BUSY_OK;
293 if (td->o.rw_seq == RW_SEQ_SEQ) {
294 ret = get_next_seq_offset(td, f, ddir, &offset);
296 ret = get_next_rand_block(td, f, ddir, &b);
297 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
298 if (f->last_start != -1ULL)
299 offset = f->last_start - f->file_offset;
304 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
311 io_u->offset = offset;
313 io_u->offset = b * td->o.ba[ddir];
315 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
325 * For random io, generate a random new block and see if it's used. Repeat
326 * until we find a free one. For sequential io, just return the end of
327 * the last io issued.
329 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
331 struct fio_file *f = io_u->file;
332 enum fio_ddir ddir = io_u->ddir;
335 assert(ddir_rw(ddir));
337 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
339 td->ddir_seq_nr = td->o.ddir_seq_nr;
342 if (get_next_block(td, io_u, ddir, rw_seq_hit))
345 if (io_u->offset >= f->io_size) {
346 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
347 io_u->offset, f->io_size);
351 io_u->offset += f->file_offset;
352 if (io_u->offset >= f->real_file_size) {
353 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
354 io_u->offset, f->real_file_size);
361 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
363 if (td->flags & TD_F_PROFILE_OPS) {
364 struct prof_io_ops *ops = &td->prof_io_ops;
366 if (ops->fill_io_u_off)
367 return ops->fill_io_u_off(td, io_u);
370 return __get_next_offset(td, io_u);
373 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
376 struct fio_file *f = io_u->file;
378 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
381 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
383 const int ddir = io_u->ddir;
384 unsigned int buflen = 0;
385 unsigned int minbs, maxbs;
386 unsigned long r, rand_max;
388 assert(ddir_rw(ddir));
390 minbs = td->o.min_bs[ddir];
391 maxbs = td->o.max_bs[ddir];
397 * If we can't satisfy the min block size from here, then fail
399 if (!io_u_fits(td, io_u, minbs))
402 if (td->o.use_os_rand)
403 rand_max = OS_RAND_MAX;
405 rand_max = FRAND_MAX;
408 if (td->o.use_os_rand)
409 r = os_random_long(&td->bsrange_state);
411 r = __rand(&td->__bsrange_state);
413 if (!td->o.bssplit_nr[ddir]) {
414 buflen = 1 + (unsigned int) ((double) maxbs *
415 (r / (rand_max + 1.0)));
422 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
423 struct bssplit *bsp = &td->o.bssplit[ddir][i];
427 if ((r <= ((rand_max / 100L) * perc)) &&
428 io_u_fits(td, io_u, buflen))
433 if (!td->o.bs_unaligned && is_power_of_2(minbs))
434 buflen = (buflen + minbs - 1) & ~(minbs - 1);
436 } while (!io_u_fits(td, io_u, buflen));
441 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
443 if (td->flags & TD_F_PROFILE_OPS) {
444 struct prof_io_ops *ops = &td->prof_io_ops;
446 if (ops->fill_io_u_size)
447 return ops->fill_io_u_size(td, io_u);
450 return __get_next_buflen(td, io_u);
453 static void set_rwmix_bytes(struct thread_data *td)
458 * we do time or byte based switch. this is needed because
459 * buffered writes may issue a lot quicker than they complete,
460 * whereas reads do not.
462 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
463 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
466 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
471 if (td->o.use_os_rand) {
472 r = os_random_long(&td->rwmix_state);
473 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
475 r = __rand(&td->__rwmix_state);
476 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
479 if (v <= td->o.rwmix[DDIR_READ])
485 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
487 enum fio_ddir odir = ddir ^ 1;
491 assert(ddir_rw(ddir));
493 if (td->rate_pending_usleep[ddir] <= 0)
497 * We have too much pending sleep in this direction. See if we
502 * Other direction does not have too much pending, switch
504 if (td->rate_pending_usleep[odir] < 100000)
508 * Both directions have pending sleep. Sleep the minimum time
509 * and deduct from both.
511 if (td->rate_pending_usleep[ddir] <=
512 td->rate_pending_usleep[odir]) {
513 usec = td->rate_pending_usleep[ddir];
515 usec = td->rate_pending_usleep[odir];
519 usec = td->rate_pending_usleep[ddir];
522 * We are going to sleep, ensure that we flush anything pending as
523 * not to skew our latency numbers.
525 * Changed to only monitor 'in flight' requests here instead of the
526 * td->cur_depth, b/c td->cur_depth does not accurately represent
527 * io's that have been actually submitted to an async engine,
528 * and cur_depth is meaningless for sync engines.
530 if (td->io_u_in_flight) {
533 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
536 fio_gettime(&t, NULL);
537 usec_sleep(td, usec);
538 usec = utime_since_now(&t);
540 td->rate_pending_usleep[ddir] -= usec;
543 if (td_rw(td) && __should_check_rate(td, odir))
544 td->rate_pending_usleep[odir] -= usec;
552 * Return the data direction for the next io_u. If the job is a
553 * mixed read/write workload, check the rwmix cycle and switch if
556 static enum fio_ddir get_rw_ddir(struct thread_data *td)
561 * see if it's time to fsync
563 if (td->o.fsync_blocks &&
564 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
565 td->io_issues[DDIR_WRITE] && should_fsync(td))
569 * see if it's time to fdatasync
571 if (td->o.fdatasync_blocks &&
572 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
573 td->io_issues[DDIR_WRITE] && should_fsync(td))
574 return DDIR_DATASYNC;
577 * see if it's time to sync_file_range
579 if (td->sync_file_range_nr &&
580 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
581 td->io_issues[DDIR_WRITE] && should_fsync(td))
582 return DDIR_SYNC_FILE_RANGE;
586 * Check if it's time to seed a new data direction.
588 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
590 * Put a top limit on how many bytes we do for
591 * one data direction, to avoid overflowing the
594 ddir = get_rand_ddir(td);
596 if (ddir != td->rwmix_ddir)
599 td->rwmix_ddir = ddir;
601 ddir = td->rwmix_ddir;
602 } else if (td_read(td))
604 else if (td_write(td))
609 td->rwmix_ddir = rate_ddir(td, ddir);
610 return td->rwmix_ddir;
613 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
615 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
617 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
618 td->o.barrier_blocks &&
619 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
620 td->io_issues[DDIR_WRITE])
621 io_u->flags |= IO_U_F_BARRIER;
624 void put_file_log(struct thread_data *td, struct fio_file *f)
626 int ret = put_file(td, f);
629 td_verror(td, ret, "file close");
632 void put_io_u(struct thread_data *td, struct io_u *io_u)
636 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
637 put_file_log(td, io_u->file);
639 io_u->flags &= ~IO_U_F_FREE_DEF;
640 io_u->flags |= IO_U_F_FREE;
642 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
644 flist_del_init(&io_u->list);
645 flist_add(&io_u->list, &td->io_u_freelist);
647 td_io_u_free_notify(td);
650 void clear_io_u(struct thread_data *td, struct io_u *io_u)
652 io_u->flags &= ~IO_U_F_FLIGHT;
656 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
658 struct io_u *__io_u = *io_u;
659 enum fio_ddir ddir = acct_ddir(__io_u);
661 dprint(FD_IO, "requeue %p\n", __io_u);
665 __io_u->flags |= IO_U_F_FREE;
666 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
667 td->io_issues[ddir]--;
669 __io_u->flags &= ~IO_U_F_FLIGHT;
670 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
672 flist_del(&__io_u->list);
673 flist_add_tail(&__io_u->list, &td->io_u_requeues);
678 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
680 if (td->io_ops->flags & FIO_NOIO)
683 set_rw_ddir(td, io_u);
686 * fsync() or fdatasync() or trim etc, we are done
688 if (!ddir_rw(io_u->ddir))
692 * See if it's time to switch to a new zone
694 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
696 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
697 io_u->file->last_pos = io_u->file->file_offset;
698 td->io_skip_bytes += td->o.zone_skip;
702 * No log, let the seq/rand engine retrieve the next buflen and
705 if (get_next_offset(td, io_u)) {
706 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
710 io_u->buflen = get_next_buflen(td, io_u);
712 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
716 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
717 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
718 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
719 io_u->buflen, io_u->file->real_file_size);
724 * mark entry before potentially trimming io_u
726 if (td_random(td) && file_randommap(td, io_u->file))
727 mark_random_map(td, io_u);
730 dprint_io_u(io_u, "fill_io_u");
731 td->zone_bytes += io_u->buflen;
735 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
764 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
766 __io_u_mark_map(td->ts.io_u_submit, nr);
767 td->ts.total_submit++;
770 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
772 __io_u_mark_map(td->ts.io_u_complete, nr);
773 td->ts.total_complete++;
776 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
780 switch (td->cur_depth) {
802 td->ts.io_u_map[idx] += nr;
805 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
842 assert(idx < FIO_IO_U_LAT_U_NR);
843 td->ts.io_u_lat_u[idx]++;
846 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
887 assert(idx < FIO_IO_U_LAT_M_NR);
888 td->ts.io_u_lat_m[idx]++;
891 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
894 io_u_mark_lat_usec(td, usec);
896 io_u_mark_lat_msec(td, usec / 1000);
900 * Get next file to service by choosing one at random
902 static struct fio_file *get_next_file_rand(struct thread_data *td,
903 enum fio_file_flags goodf,
904 enum fio_file_flags badf)
913 if (td->o.use_os_rand) {
914 r = os_random_long(&td->next_file_state);
915 fno = (unsigned int) ((double) td->o.nr_files
916 * (r / (OS_RAND_MAX + 1.0)));
918 r = __rand(&td->__next_file_state);
919 fno = (unsigned int) ((double) td->o.nr_files
920 * (r / (FRAND_MAX + 1.0)));
924 if (fio_file_done(f))
927 if (!fio_file_open(f)) {
930 err = td_io_open_file(td, f);
936 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
937 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
941 td_io_close_file(td, f);
946 * Get next file to service by doing round robin between all available ones
948 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
951 unsigned int old_next_file = td->next_file;
957 f = td->files[td->next_file];
960 if (td->next_file >= td->o.nr_files)
963 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
964 if (fio_file_done(f)) {
969 if (!fio_file_open(f)) {
972 err = td_io_open_file(td, f);
974 dprint(FD_FILE, "error %d on open of %s\n",
982 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
984 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
988 td_io_close_file(td, f);
991 } while (td->next_file != old_next_file);
993 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
997 static struct fio_file *__get_next_file(struct thread_data *td)
1001 assert(td->o.nr_files <= td->files_index);
1003 if (td->nr_done_files >= td->o.nr_files) {
1004 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1005 " nr_files=%d\n", td->nr_open_files,
1011 f = td->file_service_file;
1012 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1013 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1015 if (td->file_service_left--)
1019 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1020 td->o.file_service_type == FIO_FSERVICE_SEQ)
1021 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1023 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1025 td->file_service_file = f;
1026 td->file_service_left = td->file_service_nr - 1;
1028 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1032 static struct fio_file *get_next_file(struct thread_data *td)
1034 if (!(td->flags & TD_F_PROFILE_OPS)) {
1035 struct prof_io_ops *ops = &td->prof_io_ops;
1037 if (ops->get_next_file)
1038 return ops->get_next_file(td);
1041 return __get_next_file(td);
1044 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1049 f = get_next_file(td);
1056 if (!fill_io_u(td, io_u))
1059 put_file_log(td, f);
1060 td_io_close_file(td, f);
1062 fio_file_set_done(f);
1063 td->nr_done_files++;
1064 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1065 td->nr_done_files, td->o.nr_files);
1072 struct io_u *__get_io_u(struct thread_data *td)
1074 struct io_u *io_u = NULL;
1079 if (!flist_empty(&td->io_u_requeues))
1080 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1081 else if (!queue_full(td)) {
1082 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1087 io_u->end_io = NULL;
1091 assert(io_u->flags & IO_U_F_FREE);
1092 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1093 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1094 io_u->flags &= ~IO_U_F_VER_LIST;
1097 io_u->acct_ddir = -1;
1098 flist_del(&io_u->list);
1099 flist_add_tail(&io_u->list, &td->io_u_busylist);
1101 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1102 } else if (td->o.verify_async) {
1104 * We ran out, wait for async verify threads to finish and
1107 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1115 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1117 if (!(td->flags & TD_F_TRIM_BACKLOG))
1120 if (td->trim_entries) {
1123 if (td->trim_batch) {
1126 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1127 td->last_ddir != DDIR_READ) {
1128 td->trim_batch = td->o.trim_batch;
1129 if (!td->trim_batch)
1130 td->trim_batch = td->o.trim_backlog;
1134 if (get_trim && !get_next_trim(td, io_u))
1141 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1143 if (!(td->flags & TD_F_VER_BACKLOG))
1146 if (td->io_hist_len) {
1149 if (td->verify_batch)
1151 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1152 td->last_ddir != DDIR_READ) {
1153 td->verify_batch = td->o.verify_batch;
1154 if (!td->verify_batch)
1155 td->verify_batch = td->o.verify_backlog;
1159 if (get_verify && !get_next_verify(td, io_u)) {
1169 * Fill offset and start time into the buffer content, to prevent too
1170 * easy compressible data for simple de-dupe attempts. Do this for every
1171 * 512b block in the range, since that should be the smallest block size
1172 * we can expect from a device.
1174 static void small_content_scramble(struct io_u *io_u)
1176 unsigned int i, nr_blocks = io_u->buflen / 512;
1178 unsigned int offset;
1185 boffset = io_u->offset;
1186 io_u->buf_filled_len = 0;
1188 for (i = 0; i < nr_blocks; i++) {
1190 * Fill the byte offset into a "random" start offset of
1191 * the buffer, given by the product of the usec time
1192 * and the actual offset.
1194 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1195 offset &= ~(sizeof(uint64_t) - 1);
1196 if (offset >= 512 - sizeof(uint64_t))
1197 offset -= sizeof(uint64_t);
1198 memcpy(p + offset, &boffset, sizeof(boffset));
1200 end = p + 512 - sizeof(io_u->start_time);
1201 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1208 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1209 * etc. The returned io_u is fully ready to be prepped and submitted.
1211 struct io_u *get_io_u(struct thread_data *td)
1215 int do_scramble = 0;
1217 io_u = __get_io_u(td);
1219 dprint(FD_IO, "__get_io_u failed\n");
1223 if (check_get_verify(td, io_u))
1225 if (check_get_trim(td, io_u))
1229 * from a requeue, io_u already setup
1235 * If using an iolog, grab next piece if any available.
1237 if (td->flags & TD_F_READ_IOLOG) {
1238 if (read_iolog_get(td, io_u))
1240 } else if (set_io_u_file(td, io_u)) {
1241 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1246 assert(fio_file_open(f));
1248 if (ddir_rw(io_u->ddir)) {
1249 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1250 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1254 f->last_start = io_u->offset;
1255 f->last_pos = io_u->offset + io_u->buflen;
1257 if (io_u->ddir == DDIR_WRITE) {
1258 if (td->flags & TD_F_REFILL_BUFFERS) {
1259 io_u_fill_buffer(td, io_u,
1260 io_u->xfer_buflen, io_u->xfer_buflen);
1261 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1263 if (td->flags & TD_F_VER_NONE) {
1264 populate_verify_io_u(td, io_u);
1267 } else if (io_u->ddir == DDIR_READ) {
1269 * Reset the buf_filled parameters so next time if the
1270 * buffer is used for writes it is refilled.
1272 io_u->buf_filled_len = 0;
1277 * Set io data pointers.
1279 io_u->xfer_buf = io_u->buf;
1280 io_u->xfer_buflen = io_u->buflen;
1284 if (!td_io_prep(td, io_u)) {
1285 if (!td->o.disable_slat)
1286 fio_gettime(&io_u->start_time, NULL);
1288 small_content_scramble(io_u);
1292 dprint(FD_IO, "get_io_u failed\n");
1297 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1299 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1300 const char *msg[] = { "read", "write", "sync", "datasync",
1301 "sync_file_range", "wait", "trim" };
1303 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1306 log_err("fio: io_u error");
1309 log_err(" on file %s", io_u->file->file_name);
1311 log_err(": %s\n", strerror(io_u->error));
1313 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1314 io_u->offset, io_u->xfer_buflen);
1317 td_verror(td, io_u->error, "io_u error");
1320 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1321 struct io_completion_data *icd,
1322 const enum fio_ddir idx, unsigned int bytes)
1324 unsigned long lusec = 0;
1326 if (!td->o.disable_clat || !td->o.disable_bw)
1327 lusec = utime_since(&io_u->issue_time, &icd->time);
1329 if (!td->o.disable_lat) {
1330 unsigned long tusec;
1332 tusec = utime_since(&io_u->start_time, &icd->time);
1333 add_lat_sample(td, idx, tusec, bytes);
1335 if (td->o.max_latency && tusec > td->o.max_latency) {
1337 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1338 td_verror(td, ETIMEDOUT, "max latency exceeded");
1339 icd->error = ETIMEDOUT;
1343 if (!td->o.disable_clat) {
1344 add_clat_sample(td, idx, lusec, bytes);
1345 io_u_mark_latency(td, lusec);
1348 if (!td->o.disable_bw)
1349 add_bw_sample(td, idx, bytes, &icd->time);
1351 add_iops_sample(td, idx, &icd->time);
1354 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1356 uint64_t secs, remainder, bps, bytes;
1358 bytes = td->this_io_bytes[ddir];
1359 bps = td->rate_bps[ddir];
1361 remainder = bytes % bps;
1362 return remainder * 1000000 / bps + secs * 1000000;
1365 static void io_completed(struct thread_data *td, struct io_u *io_u,
1366 struct io_completion_data *icd)
1370 dprint_io_u(io_u, "io complete");
1373 assert(io_u->flags & IO_U_F_FLIGHT);
1374 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1377 if (ddir_sync(io_u->ddir)) {
1378 td->last_was_sync = 1;
1381 f->first_write = -1ULL;
1382 f->last_write = -1ULL;
1387 td->last_was_sync = 0;
1388 td->last_ddir = io_u->ddir;
1390 if (!io_u->error && ddir_rw(io_u->ddir)) {
1391 unsigned int bytes = io_u->buflen - io_u->resid;
1392 const enum fio_ddir idx = io_u->ddir;
1393 const enum fio_ddir odx = io_u->ddir ^ 1;
1396 td->io_blocks[idx]++;
1397 td->this_io_blocks[idx]++;
1398 td->io_bytes[idx] += bytes;
1400 if (!(io_u->flags & IO_U_F_VER_LIST))
1401 td->this_io_bytes[idx] += bytes;
1403 if (idx == DDIR_WRITE) {
1406 if (f->first_write == -1ULL ||
1407 io_u->offset < f->first_write)
1408 f->first_write = io_u->offset;
1409 if (f->last_write == -1ULL ||
1410 ((io_u->offset + bytes) > f->last_write))
1411 f->last_write = io_u->offset + bytes;
1415 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1416 td->runstate == TD_VERIFYING)) {
1417 account_io_completion(td, io_u, icd, idx, bytes);
1419 if (__should_check_rate(td, idx)) {
1420 td->rate_pending_usleep[idx] =
1421 (usec_for_io(td, idx) -
1422 utime_since_now(&td->start));
1424 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1425 td->rate_pending_usleep[odx] =
1426 (usec_for_io(td, odx) -
1427 utime_since_now(&td->start));
1430 if (td_write(td) && idx == DDIR_WRITE &&
1432 td->o.verify != VERIFY_NONE &&
1433 !td->o.experimental_verify)
1434 log_io_piece(td, io_u);
1436 icd->bytes_done[idx] += bytes;
1439 ret = io_u->end_io(td, io_u);
1440 if (ret && !icd->error)
1443 } else if (io_u->error) {
1444 icd->error = io_u->error;
1445 io_u_log_error(td, io_u);
1448 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1449 if (!td_non_fatal_error(td, eb, icd->error))
1452 * If there is a non_fatal error, then add to the error count
1453 * and clear all the errors.
1455 update_error_count(td, icd->error);
1462 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1466 if (!td->o.disable_clat || !td->o.disable_bw)
1467 fio_gettime(&icd->time, NULL);
1472 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1473 icd->bytes_done[ddir] = 0;
1476 static void ios_completed(struct thread_data *td,
1477 struct io_completion_data *icd)
1482 for (i = 0; i < icd->nr; i++) {
1483 io_u = td->io_ops->event(td, i);
1485 io_completed(td, io_u, icd);
1487 if (!(io_u->flags & IO_U_F_FREE_DEF))
1493 * Complete a single io_u for the sync engines.
1495 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1498 struct io_completion_data icd;
1500 init_icd(td, &icd, 1);
1501 io_completed(td, io_u, &icd);
1503 if (!(io_u->flags & IO_U_F_FREE_DEF))
1507 td_verror(td, icd.error, "io_u_sync_complete");
1514 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1515 bytes[ddir] += icd.bytes_done[ddir];
1522 * Called to complete min_events number of io for the async engines.
1524 int io_u_queued_complete(struct thread_data *td, int min_evts,
1527 struct io_completion_data icd;
1528 struct timespec *tvp = NULL;
1530 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1532 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1537 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1539 td_verror(td, -ret, "td_io_getevents");
1544 init_icd(td, &icd, ret);
1545 ios_completed(td, &icd);
1547 td_verror(td, icd.error, "io_u_queued_complete");
1554 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1555 bytes[ddir] += icd.bytes_done[ddir];
1562 * Call when io_u is really queued, to update the submission latency.
1564 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1566 if (!td->o.disable_slat) {
1567 unsigned long slat_time;
1569 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1570 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1575 * "randomly" fill the buffer contents
1577 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1578 unsigned int min_write, unsigned int max_bs)
1580 io_u->buf_filled_len = 0;
1582 if (!td->o.zero_buffers) {
1583 unsigned int perc = td->o.compress_percentage;
1586 unsigned int seg = min_write;
1588 seg = min(min_write, td->o.compress_chunk);
1589 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1592 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1594 memset(io_u->buf, 0, max_bs);