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", (unsigned long long) 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",
129 (unsigned long long) *b);
131 *b = axmap_next_free(f->io_axmap, *b);
132 if (*b == (uint64_t) -1ULL)
138 static int __get_next_rand_offset_zipf(struct thread_data *td,
139 struct fio_file *f, enum fio_ddir ddir,
142 *b = zipf_next(&f->zipf);
146 static int __get_next_rand_offset_pareto(struct thread_data *td,
147 struct fio_file *f, enum fio_ddir ddir,
150 *b = pareto_next(&f->zipf);
154 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
156 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
157 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
159 return r1->off - r2->off;
162 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
163 enum fio_ddir ddir, uint64_t *b)
165 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
166 return __get_next_rand_offset(td, f, ddir, b);
167 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
168 return __get_next_rand_offset_zipf(td, f, ddir, b);
169 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
170 return __get_next_rand_offset_pareto(td, f, ddir, b);
172 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
177 * Sort the reads for a verify phase in batches of verifysort_nr, if
180 static inline int should_sort_io(struct thread_data *td)
182 if (!td->o.verifysort_nr || !td->o.do_verify)
186 if (td->runstate != TD_VERIFYING)
188 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
194 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
195 enum fio_ddir ddir, uint64_t *b)
200 if (!should_sort_io(td))
201 return get_off_from_method(td, f, ddir, b);
203 if (!flist_empty(&td->next_rand_list)) {
206 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
213 for (i = 0; i < td->o.verifysort_nr; i++) {
214 r = malloc(sizeof(*r));
216 ret = get_off_from_method(td, f, ddir, &r->off);
222 flist_add(&r->list, &td->next_rand_list);
228 assert(!flist_empty(&td->next_rand_list));
229 flist_sort(NULL, &td->next_rand_list, flist_cmp);
233 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
234 enum fio_ddir ddir, uint64_t *b)
236 if (!get_next_rand_offset(td, f, ddir, b))
239 if (td->o.time_based) {
240 fio_file_reset(td, f);
241 if (!get_next_rand_offset(td, f, ddir, b))
245 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
246 f->file_name, (unsigned long long) f->last_pos,
247 (unsigned long long) f->real_file_size);
251 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
252 enum fio_ddir ddir, uint64_t *offset)
254 assert(ddir_rw(ddir));
256 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
257 f->last_pos = f->last_pos - f->io_size;
259 if (f->last_pos < f->real_file_size) {
262 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
263 f->last_pos = f->real_file_size;
265 pos = f->last_pos - f->file_offset;
267 pos += td->o.ddir_seq_add;
276 static int get_next_block(struct thread_data *td, struct io_u *io_u,
277 enum fio_ddir ddir, int rw_seq)
279 struct fio_file *f = io_u->file;
283 assert(ddir_rw(ddir));
289 ret = get_next_rand_block(td, f, ddir, &b);
291 ret = get_next_seq_offset(td, f, ddir, &offset);
293 io_u->flags |= IO_U_F_BUSY_OK;
295 if (td->o.rw_seq == RW_SEQ_SEQ) {
296 ret = get_next_seq_offset(td, f, ddir, &offset);
298 ret = get_next_rand_block(td, f, ddir, &b);
299 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
300 if (f->last_start != -1ULL)
301 offset = f->last_start - f->file_offset;
306 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
313 io_u->offset = offset;
315 io_u->offset = b * td->o.ba[ddir];
317 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
326 * For random io, generate a random new block and see if it's used. Repeat
327 * until we find a free one. For sequential io, just return the end of
328 * the last io issued.
330 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
332 struct fio_file *f = io_u->file;
333 enum fio_ddir ddir = io_u->ddir;
336 assert(ddir_rw(ddir));
338 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
340 td->ddir_seq_nr = td->o.ddir_seq_nr;
343 if (get_next_block(td, io_u, ddir, rw_seq_hit))
346 if (io_u->offset >= f->io_size) {
347 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
348 (unsigned long long) io_u->offset,
349 (unsigned long long) f->io_size);
353 io_u->offset += f->file_offset;
354 if (io_u->offset >= f->real_file_size) {
355 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
356 (unsigned long long) io_u->offset,
357 (unsigned long long) f->real_file_size);
364 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
366 if (td->flags & TD_F_PROFILE_OPS) {
367 struct prof_io_ops *ops = &td->prof_io_ops;
369 if (ops->fill_io_u_off)
370 return ops->fill_io_u_off(td, io_u);
373 return __get_next_offset(td, io_u);
376 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
379 struct fio_file *f = io_u->file;
381 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
384 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
386 const int ddir = io_u->ddir;
387 unsigned int buflen = 0;
388 unsigned int minbs, maxbs;
389 unsigned long r, rand_max;
391 assert(ddir_rw(ddir));
393 minbs = td->o.min_bs[ddir];
394 maxbs = td->o.max_bs[ddir];
400 * If we can't satisfy the min block size from here, then fail
402 if (!io_u_fits(td, io_u, minbs))
405 if (td->o.use_os_rand)
406 rand_max = OS_RAND_MAX;
408 rand_max = FRAND_MAX;
411 if (td->o.use_os_rand)
412 r = os_random_long(&td->bsrange_state);
414 r = __rand(&td->__bsrange_state);
416 if (!td->o.bssplit_nr[ddir]) {
417 buflen = 1 + (unsigned int) ((double) maxbs *
418 (r / (rand_max + 1.0)));
425 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
426 struct bssplit *bsp = &td->o.bssplit[ddir][i];
430 if ((r <= ((rand_max / 100L) * perc)) &&
431 io_u_fits(td, io_u, buflen))
436 if (!td->o.bs_unaligned && is_power_of_2(minbs))
437 buflen = (buflen + minbs - 1) & ~(minbs - 1);
439 } while (!io_u_fits(td, io_u, buflen));
444 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
446 if (td->flags & TD_F_PROFILE_OPS) {
447 struct prof_io_ops *ops = &td->prof_io_ops;
449 if (ops->fill_io_u_size)
450 return ops->fill_io_u_size(td, io_u);
453 return __get_next_buflen(td, io_u);
456 static void set_rwmix_bytes(struct thread_data *td)
461 * we do time or byte based switch. this is needed because
462 * buffered writes may issue a lot quicker than they complete,
463 * whereas reads do not.
465 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
466 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
469 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
474 if (td->o.use_os_rand) {
475 r = os_random_long(&td->rwmix_state);
476 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
478 r = __rand(&td->__rwmix_state);
479 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
482 if (v <= td->o.rwmix[DDIR_READ])
488 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
490 enum fio_ddir odir = ddir ^ 1;
494 assert(ddir_rw(ddir));
496 if (td->rate_pending_usleep[ddir] <= 0)
500 * We have too much pending sleep in this direction. See if we
503 if (td_rw(td) && td->o.rwmix[odir]) {
505 * Other direction does not have too much pending, switch
507 if (td->rate_pending_usleep[odir] < 100000)
511 * Both directions have pending sleep. Sleep the minimum time
512 * and deduct from both.
514 if (td->rate_pending_usleep[ddir] <=
515 td->rate_pending_usleep[odir]) {
516 usec = td->rate_pending_usleep[ddir];
518 usec = td->rate_pending_usleep[odir];
522 usec = td->rate_pending_usleep[ddir];
525 * We are going to sleep, ensure that we flush anything pending as
526 * not to skew our latency numbers.
528 * Changed to only monitor 'in flight' requests here instead of the
529 * td->cur_depth, b/c td->cur_depth does not accurately represent
530 * io's that have been actually submitted to an async engine,
531 * and cur_depth is meaningless for sync engines.
533 while (td->io_u_in_flight) {
536 ret = io_u_queued_complete(td, 1, NULL);
539 fio_gettime(&t, NULL);
540 usec_sleep(td, usec);
541 usec = utime_since_now(&t);
543 td->rate_pending_usleep[ddir] -= usec;
546 if (td_rw(td) && __should_check_rate(td, odir))
547 td->rate_pending_usleep[odir] -= usec;
556 * Return the data direction for the next io_u. If the job is a
557 * mixed read/write workload, check the rwmix cycle and switch if
560 static enum fio_ddir get_rw_ddir(struct thread_data *td)
565 * see if it's time to fsync
567 if (td->o.fsync_blocks &&
568 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
569 td->io_issues[DDIR_WRITE] && should_fsync(td))
573 * see if it's time to fdatasync
575 if (td->o.fdatasync_blocks &&
576 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
577 td->io_issues[DDIR_WRITE] && should_fsync(td))
578 return DDIR_DATASYNC;
581 * see if it's time to sync_file_range
583 if (td->sync_file_range_nr &&
584 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
585 td->io_issues[DDIR_WRITE] && should_fsync(td))
586 return DDIR_SYNC_FILE_RANGE;
590 * Check if it's time to seed a new data direction.
592 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
594 * Put a top limit on how many bytes we do for
595 * one data direction, to avoid overflowing the
598 ddir = get_rand_ddir(td);
600 if (ddir != td->rwmix_ddir)
603 td->rwmix_ddir = ddir;
605 ddir = td->rwmix_ddir;
606 } else if (td_read(td))
608 else if (td_write(td))
613 td->rwmix_ddir = rate_ddir(td, ddir);
614 return td->rwmix_ddir;
617 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
619 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
621 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
622 td->o.barrier_blocks &&
623 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
624 td->io_issues[DDIR_WRITE])
625 io_u->flags |= IO_U_F_BARRIER;
628 void put_file_log(struct thread_data *td, struct fio_file *f)
630 int ret = put_file(td, f);
633 td_verror(td, ret, "file close");
636 void put_io_u(struct thread_data *td, struct io_u *io_u)
640 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
641 put_file_log(td, io_u->file);
643 io_u->flags &= ~IO_U_F_FREE_DEF;
644 io_u->flags |= IO_U_F_FREE;
646 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
648 flist_del_init(&io_u->list);
649 flist_add(&io_u->list, &td->io_u_freelist);
651 td_io_u_free_notify(td);
654 void clear_io_u(struct thread_data *td, struct io_u *io_u)
656 io_u->flags &= ~IO_U_F_FLIGHT;
660 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
662 struct io_u *__io_u = *io_u;
663 enum fio_ddir ddir = acct_ddir(__io_u);
665 dprint(FD_IO, "requeue %p\n", __io_u);
669 __io_u->flags |= IO_U_F_FREE;
670 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
671 td->io_issues[ddir]--;
673 __io_u->flags &= ~IO_U_F_FLIGHT;
674 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
676 flist_del(&__io_u->list);
677 flist_add_tail(&__io_u->list, &td->io_u_requeues);
682 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
684 if (td->io_ops->flags & FIO_NOIO)
687 set_rw_ddir(td, io_u);
690 * fsync() or fdatasync() or trim etc, we are done
692 if (!ddir_rw(io_u->ddir))
696 * See if it's time to switch to a new zone
698 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
700 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
701 io_u->file->last_pos = io_u->file->file_offset;
702 td->io_skip_bytes += td->o.zone_skip;
706 * No log, let the seq/rand engine retrieve the next buflen and
709 if (get_next_offset(td, io_u)) {
710 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
714 io_u->buflen = get_next_buflen(td, io_u);
716 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
720 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
721 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
722 dprint(FD_IO, " off=%llu/%lu > %llu\n",
723 (unsigned long long) io_u->offset, io_u->buflen,
724 (unsigned long long) io_u->file->real_file_size);
729 * mark entry before potentially trimming io_u
731 if (td_random(td) && file_randommap(td, io_u->file))
732 mark_random_map(td, io_u);
735 dprint_io_u(io_u, "fill_io_u");
736 td->zone_bytes += io_u->buflen;
740 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
769 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
771 __io_u_mark_map(td->ts.io_u_submit, nr);
772 td->ts.total_submit++;
775 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
777 __io_u_mark_map(td->ts.io_u_complete, nr);
778 td->ts.total_complete++;
781 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
785 switch (td->cur_depth) {
807 td->ts.io_u_map[idx] += nr;
810 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
847 assert(idx < FIO_IO_U_LAT_U_NR);
848 td->ts.io_u_lat_u[idx]++;
851 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
892 assert(idx < FIO_IO_U_LAT_M_NR);
893 td->ts.io_u_lat_m[idx]++;
896 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
899 io_u_mark_lat_usec(td, usec);
901 io_u_mark_lat_msec(td, usec / 1000);
905 * Get next file to service by choosing one at random
907 static struct fio_file *get_next_file_rand(struct thread_data *td,
908 enum fio_file_flags goodf,
909 enum fio_file_flags badf)
918 if (td->o.use_os_rand) {
919 r = os_random_long(&td->next_file_state);
920 fno = (unsigned int) ((double) td->o.nr_files
921 * (r / (OS_RAND_MAX + 1.0)));
923 r = __rand(&td->__next_file_state);
924 fno = (unsigned int) ((double) td->o.nr_files
925 * (r / (FRAND_MAX + 1.0)));
929 if (fio_file_done(f))
932 if (!fio_file_open(f)) {
935 err = td_io_open_file(td, f);
941 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
942 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
946 td_io_close_file(td, f);
951 * Get next file to service by doing round robin between all available ones
953 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
956 unsigned int old_next_file = td->next_file;
962 f = td->files[td->next_file];
965 if (td->next_file >= td->o.nr_files)
968 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
969 if (fio_file_done(f)) {
974 if (!fio_file_open(f)) {
977 err = td_io_open_file(td, f);
979 dprint(FD_FILE, "error %d on open of %s\n",
987 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
989 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
993 td_io_close_file(td, f);
996 } while (td->next_file != old_next_file);
998 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1002 static struct fio_file *__get_next_file(struct thread_data *td)
1006 assert(td->o.nr_files <= td->files_index);
1008 if (td->nr_done_files >= td->o.nr_files) {
1009 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1010 " nr_files=%d\n", td->nr_open_files,
1016 f = td->file_service_file;
1017 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1018 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1020 if (td->file_service_left--)
1024 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1025 td->o.file_service_type == FIO_FSERVICE_SEQ)
1026 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1028 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1030 td->file_service_file = f;
1031 td->file_service_left = td->file_service_nr - 1;
1033 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1037 static struct fio_file *get_next_file(struct thread_data *td)
1039 if (!(td->flags & TD_F_PROFILE_OPS)) {
1040 struct prof_io_ops *ops = &td->prof_io_ops;
1042 if (ops->get_next_file)
1043 return ops->get_next_file(td);
1046 return __get_next_file(td);
1049 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1054 f = get_next_file(td);
1061 if (!fill_io_u(td, io_u))
1064 put_file_log(td, f);
1065 td_io_close_file(td, f);
1067 fio_file_set_done(f);
1068 td->nr_done_files++;
1069 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1070 td->nr_done_files, td->o.nr_files);
1077 struct io_u *__get_io_u(struct thread_data *td)
1079 struct io_u *io_u = NULL;
1084 if (!flist_empty(&td->io_u_requeues))
1085 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1086 else if (!queue_full(td)) {
1087 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1092 io_u->end_io = NULL;
1096 assert(io_u->flags & IO_U_F_FREE);
1097 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1098 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1099 io_u->flags &= ~IO_U_F_VER_LIST;
1102 io_u->acct_ddir = -1;
1103 flist_del(&io_u->list);
1104 flist_add_tail(&io_u->list, &td->io_u_busylist);
1106 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1107 } else if (td->o.verify_async) {
1109 * We ran out, wait for async verify threads to finish and
1112 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1120 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1122 if (!(td->flags & TD_F_TRIM_BACKLOG))
1125 if (td->trim_entries) {
1128 if (td->trim_batch) {
1131 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1132 td->last_ddir != DDIR_READ) {
1133 td->trim_batch = td->o.trim_batch;
1134 if (!td->trim_batch)
1135 td->trim_batch = td->o.trim_backlog;
1139 if (get_trim && !get_next_trim(td, io_u))
1146 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1148 if (!(td->flags & TD_F_VER_BACKLOG))
1151 if (td->io_hist_len) {
1154 if (td->verify_batch)
1156 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1157 td->last_ddir != DDIR_READ) {
1158 td->verify_batch = td->o.verify_batch;
1159 if (!td->verify_batch)
1160 td->verify_batch = td->o.verify_backlog;
1164 if (get_verify && !get_next_verify(td, io_u)) {
1174 * Fill offset and start time into the buffer content, to prevent too
1175 * easy compressible data for simple de-dupe attempts. Do this for every
1176 * 512b block in the range, since that should be the smallest block size
1177 * we can expect from a device.
1179 static void small_content_scramble(struct io_u *io_u)
1181 unsigned int i, nr_blocks = io_u->buflen / 512;
1183 unsigned int offset;
1190 boffset = io_u->offset;
1191 io_u->buf_filled_len = 0;
1193 for (i = 0; i < nr_blocks; i++) {
1195 * Fill the byte offset into a "random" start offset of
1196 * the buffer, given by the product of the usec time
1197 * and the actual offset.
1199 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1200 offset &= ~(sizeof(uint64_t) - 1);
1201 if (offset >= 512 - sizeof(uint64_t))
1202 offset -= sizeof(uint64_t);
1203 memcpy(p + offset, &boffset, sizeof(boffset));
1205 end = p + 512 - sizeof(io_u->start_time);
1206 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1213 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1214 * etc. The returned io_u is fully ready to be prepped and submitted.
1216 struct io_u *get_io_u(struct thread_data *td)
1220 int do_scramble = 0;
1222 io_u = __get_io_u(td);
1224 dprint(FD_IO, "__get_io_u failed\n");
1228 if (check_get_verify(td, io_u))
1230 if (check_get_trim(td, io_u))
1234 * from a requeue, io_u already setup
1240 * If using an iolog, grab next piece if any available.
1242 if (td->flags & TD_F_READ_IOLOG) {
1243 if (read_iolog_get(td, io_u))
1245 } else if (set_io_u_file(td, io_u)) {
1246 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1251 assert(fio_file_open(f));
1253 if (ddir_rw(io_u->ddir)) {
1254 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1255 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1259 f->last_start = io_u->offset;
1260 f->last_pos = io_u->offset + io_u->buflen;
1262 if (io_u->ddir == DDIR_WRITE) {
1263 if (td->flags & TD_F_REFILL_BUFFERS) {
1264 io_u_fill_buffer(td, io_u,
1265 io_u->xfer_buflen, io_u->xfer_buflen);
1266 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1268 if (td->flags & TD_F_VER_NONE) {
1269 populate_verify_io_u(td, io_u);
1272 } else if (io_u->ddir == DDIR_READ) {
1274 * Reset the buf_filled parameters so next time if the
1275 * buffer is used for writes it is refilled.
1277 io_u->buf_filled_len = 0;
1282 * Set io data pointers.
1284 io_u->xfer_buf = io_u->buf;
1285 io_u->xfer_buflen = io_u->buflen;
1289 if (!td_io_prep(td, io_u)) {
1290 if (!td->o.disable_slat)
1291 fio_gettime(&io_u->start_time, NULL);
1293 small_content_scramble(io_u);
1297 dprint(FD_IO, "get_io_u failed\n");
1302 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1304 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1305 const char *msg[] = { "read", "write", "sync", "datasync",
1306 "sync_file_range", "wait", "trim" };
1308 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1311 log_err("fio: io_u error");
1314 log_err(" on file %s", io_u->file->file_name);
1316 log_err(": %s\n", strerror(io_u->error));
1318 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1319 io_u->offset, io_u->xfer_buflen);
1322 td_verror(td, io_u->error, "io_u error");
1325 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1326 struct io_completion_data *icd,
1327 const enum fio_ddir idx, unsigned int bytes)
1329 unsigned long lusec = 0;
1331 if (!td->o.disable_clat || !td->o.disable_bw)
1332 lusec = utime_since(&io_u->issue_time, &icd->time);
1334 if (!td->o.disable_lat) {
1335 unsigned long tusec;
1337 tusec = utime_since(&io_u->start_time, &icd->time);
1338 add_lat_sample(td, idx, tusec, bytes);
1340 if (td->o.max_latency && tusec > td->o.max_latency) {
1342 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1343 td_verror(td, ETIMEDOUT, "max latency exceeded");
1344 icd->error = ETIMEDOUT;
1348 if (!td->o.disable_clat) {
1349 add_clat_sample(td, idx, lusec, bytes);
1350 io_u_mark_latency(td, lusec);
1353 if (!td->o.disable_bw)
1354 add_bw_sample(td, idx, bytes, &icd->time);
1356 add_iops_sample(td, idx, &icd->time);
1359 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1361 uint64_t secs, remainder, bps, bytes;
1363 bytes = td->this_io_bytes[ddir];
1364 bps = td->rate_bps[ddir];
1366 remainder = bytes % bps;
1367 return remainder * 1000000 / bps + secs * 1000000;
1370 static void io_completed(struct thread_data *td, struct io_u *io_u,
1371 struct io_completion_data *icd)
1375 dprint_io_u(io_u, "io complete");
1378 assert(io_u->flags & IO_U_F_FLIGHT);
1379 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1382 if (ddir_sync(io_u->ddir)) {
1383 td->last_was_sync = 1;
1386 f->first_write = -1ULL;
1387 f->last_write = -1ULL;
1392 td->last_was_sync = 0;
1393 td->last_ddir = io_u->ddir;
1395 if (!io_u->error && ddir_rw(io_u->ddir)) {
1396 unsigned int bytes = io_u->buflen - io_u->resid;
1397 const enum fio_ddir idx = io_u->ddir;
1398 const enum fio_ddir odx = io_u->ddir ^ 1;
1401 td->io_blocks[idx]++;
1402 td->this_io_blocks[idx]++;
1403 td->io_bytes[idx] += bytes;
1405 if (!(io_u->flags & IO_U_F_VER_LIST))
1406 td->this_io_bytes[idx] += bytes;
1408 if (idx == DDIR_WRITE) {
1411 if (f->first_write == -1ULL ||
1412 io_u->offset < f->first_write)
1413 f->first_write = io_u->offset;
1414 if (f->last_write == -1ULL ||
1415 ((io_u->offset + bytes) > f->last_write))
1416 f->last_write = io_u->offset + bytes;
1420 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1421 td->runstate == TD_VERIFYING)) {
1422 account_io_completion(td, io_u, icd, idx, bytes);
1424 if (__should_check_rate(td, idx)) {
1425 td->rate_pending_usleep[idx] =
1426 (usec_for_io(td, idx) -
1427 utime_since_now(&td->start));
1429 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1430 td->rate_pending_usleep[odx] =
1431 (usec_for_io(td, odx) -
1432 utime_since_now(&td->start));
1435 if (td_write(td) && idx == DDIR_WRITE &&
1437 td->o.verify != VERIFY_NONE &&
1438 !td->o.experimental_verify)
1439 log_io_piece(td, io_u);
1441 icd->bytes_done[idx] += bytes;
1444 ret = io_u->end_io(td, io_u);
1445 if (ret && !icd->error)
1448 } else if (io_u->error) {
1449 icd->error = io_u->error;
1450 io_u_log_error(td, io_u);
1453 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1454 if (!td_non_fatal_error(td, eb, icd->error))
1457 * If there is a non_fatal error, then add to the error count
1458 * and clear all the errors.
1460 update_error_count(td, icd->error);
1467 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1471 if (!td->o.disable_clat || !td->o.disable_bw)
1472 fio_gettime(&icd->time, NULL);
1477 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1478 icd->bytes_done[ddir] = 0;
1481 static void ios_completed(struct thread_data *td,
1482 struct io_completion_data *icd)
1487 for (i = 0; i < icd->nr; i++) {
1488 io_u = td->io_ops->event(td, i);
1490 io_completed(td, io_u, icd);
1492 if (!(io_u->flags & IO_U_F_FREE_DEF))
1498 * Complete a single io_u for the sync engines.
1500 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1503 struct io_completion_data icd;
1505 init_icd(td, &icd, 1);
1506 io_completed(td, io_u, &icd);
1508 if (!(io_u->flags & IO_U_F_FREE_DEF))
1512 td_verror(td, icd.error, "io_u_sync_complete");
1519 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1520 bytes[ddir] += icd.bytes_done[ddir];
1527 * Called to complete min_events number of io for the async engines.
1529 int io_u_queued_complete(struct thread_data *td, int min_evts,
1532 struct io_completion_data icd;
1533 struct timespec *tvp = NULL;
1535 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1537 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1542 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1544 td_verror(td, -ret, "td_io_getevents");
1549 init_icd(td, &icd, ret);
1550 ios_completed(td, &icd);
1552 td_verror(td, icd.error, "io_u_queued_complete");
1559 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1560 bytes[ddir] += icd.bytes_done[ddir];
1567 * Call when io_u is really queued, to update the submission latency.
1569 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1571 if (!td->o.disable_slat) {
1572 unsigned long slat_time;
1574 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1575 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1580 * "randomly" fill the buffer contents
1582 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1583 unsigned int min_write, unsigned int max_bs)
1585 io_u->buf_filled_len = 0;
1587 if (!td->o.zero_buffers) {
1588 unsigned int perc = td->o.compress_percentage;
1591 unsigned int seg = min_write;
1593 seg = min(min_write, td->o.compress_chunk);
1594 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1597 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1599 memset(io_u->buf, 0, max_bs);