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 should_do_random(struct thread_data *td)
199 if (td->o.perc_rand == 100)
202 if (td->o.use_os_rand) {
203 r = os_random_long(&td->seq_rand_state);
204 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
206 r = __rand(&td->__seq_rand_state);
207 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
210 return v <= td->o.perc_rand;
213 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
214 enum fio_ddir ddir, uint64_t *b)
219 if (!should_sort_io(td))
220 return get_off_from_method(td, f, ddir, b);
222 if (!flist_empty(&td->next_rand_list)) {
225 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
232 for (i = 0; i < td->o.verifysort_nr; i++) {
233 r = malloc(sizeof(*r));
235 ret = get_off_from_method(td, f, ddir, &r->off);
241 flist_add(&r->list, &td->next_rand_list);
247 assert(!flist_empty(&td->next_rand_list));
248 flist_sort(NULL, &td->next_rand_list, flist_cmp);
252 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
253 enum fio_ddir ddir, uint64_t *b)
255 if (!get_next_rand_offset(td, f, ddir, b))
258 if (td->o.time_based) {
259 fio_file_reset(td, f);
260 if (!get_next_rand_offset(td, f, ddir, b))
264 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
265 f->file_name, (unsigned long long) f->last_pos,
266 (unsigned long long) f->real_file_size);
270 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
271 enum fio_ddir ddir, uint64_t *offset)
273 assert(ddir_rw(ddir));
275 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
276 f->last_pos = f->last_pos - f->io_size;
278 if (f->last_pos < f->real_file_size) {
281 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
282 f->last_pos = f->real_file_size;
284 pos = f->last_pos - f->file_offset;
286 pos += td->o.ddir_seq_add;
295 static int get_next_block(struct thread_data *td, struct io_u *io_u,
296 enum fio_ddir ddir, int rw_seq)
298 struct fio_file *f = io_u->file;
302 assert(ddir_rw(ddir));
308 if (should_do_random(td))
309 ret = get_next_rand_block(td, f, ddir, &b);
311 io_u->flags |= IO_U_F_BUSY_OK;
312 ret = get_next_seq_offset(td, f, ddir, &offset);
314 ret = get_next_rand_block(td, f, ddir, &b);
317 ret = get_next_seq_offset(td, f, ddir, &offset);
319 io_u->flags |= IO_U_F_BUSY_OK;
321 if (td->o.rw_seq == RW_SEQ_SEQ) {
322 ret = get_next_seq_offset(td, f, ddir, &offset);
324 ret = get_next_rand_block(td, f, ddir, &b);
325 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
326 if (f->last_start != -1ULL)
327 offset = f->last_start - f->file_offset;
332 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
339 io_u->offset = offset;
341 io_u->offset = b * td->o.ba[ddir];
343 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
352 * For random io, generate a random new block and see if it's used. Repeat
353 * until we find a free one. For sequential io, just return the end of
354 * the last io issued.
356 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
358 struct fio_file *f = io_u->file;
359 enum fio_ddir ddir = io_u->ddir;
362 assert(ddir_rw(ddir));
364 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
366 td->ddir_seq_nr = td->o.ddir_seq_nr;
369 if (get_next_block(td, io_u, ddir, rw_seq_hit))
372 if (io_u->offset >= f->io_size) {
373 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
374 (unsigned long long) io_u->offset,
375 (unsigned long long) f->io_size);
379 io_u->offset += f->file_offset;
380 if (io_u->offset >= f->real_file_size) {
381 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
382 (unsigned long long) io_u->offset,
383 (unsigned long long) f->real_file_size);
390 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
392 if (td->flags & TD_F_PROFILE_OPS) {
393 struct prof_io_ops *ops = &td->prof_io_ops;
395 if (ops->fill_io_u_off)
396 return ops->fill_io_u_off(td, io_u);
399 return __get_next_offset(td, io_u);
402 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
405 struct fio_file *f = io_u->file;
407 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
410 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
412 const int ddir = io_u->ddir;
413 unsigned int buflen = 0;
414 unsigned int minbs, maxbs;
415 unsigned long r, rand_max;
417 assert(ddir_rw(ddir));
419 minbs = td->o.min_bs[ddir];
420 maxbs = td->o.max_bs[ddir];
426 * If we can't satisfy the min block size from here, then fail
428 if (!io_u_fits(td, io_u, minbs))
431 if (td->o.use_os_rand)
432 rand_max = OS_RAND_MAX;
434 rand_max = FRAND_MAX;
437 if (td->o.use_os_rand)
438 r = os_random_long(&td->bsrange_state);
440 r = __rand(&td->__bsrange_state);
442 if (!td->o.bssplit_nr[ddir]) {
443 buflen = 1 + (unsigned int) ((double) maxbs *
444 (r / (rand_max + 1.0)));
451 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
452 struct bssplit *bsp = &td->o.bssplit[ddir][i];
456 if ((r <= ((rand_max / 100L) * perc)) &&
457 io_u_fits(td, io_u, buflen))
462 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
463 buflen = (buflen + td->o.verify_interval - 1) &
464 ~(td->o.verify_interval - 1);
466 if (!td->o.bs_unaligned && is_power_of_2(minbs))
467 buflen = (buflen + minbs - 1) & ~(minbs - 1);
469 } while (!io_u_fits(td, io_u, buflen));
474 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
476 if (td->flags & TD_F_PROFILE_OPS) {
477 struct prof_io_ops *ops = &td->prof_io_ops;
479 if (ops->fill_io_u_size)
480 return ops->fill_io_u_size(td, io_u);
483 return __get_next_buflen(td, io_u);
486 static void set_rwmix_bytes(struct thread_data *td)
491 * we do time or byte based switch. this is needed because
492 * buffered writes may issue a lot quicker than they complete,
493 * whereas reads do not.
495 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
496 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
499 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
504 if (td->o.use_os_rand) {
505 r = os_random_long(&td->rwmix_state);
506 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
508 r = __rand(&td->__rwmix_state);
509 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
512 if (v <= td->o.rwmix[DDIR_READ])
518 void io_u_quiesce(struct thread_data *td)
521 * We are going to sleep, ensure that we flush anything pending as
522 * not to skew our latency numbers.
524 * Changed to only monitor 'in flight' requests here instead of the
525 * td->cur_depth, b/c td->cur_depth does not accurately represent
526 * io's that have been actually submitted to an async engine,
527 * and cur_depth is meaningless for sync engines.
529 while (td->io_u_in_flight) {
532 ret = io_u_queued_complete(td, 1, NULL);
536 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
538 enum fio_ddir odir = ddir ^ 1;
542 assert(ddir_rw(ddir));
544 if (td->rate_pending_usleep[ddir] <= 0)
548 * We have too much pending sleep in this direction. See if we
551 if (td_rw(td) && td->o.rwmix[odir]) {
553 * Other direction does not have too much pending, switch
555 if (td->rate_pending_usleep[odir] < 100000)
559 * Both directions have pending sleep. Sleep the minimum time
560 * and deduct from both.
562 if (td->rate_pending_usleep[ddir] <=
563 td->rate_pending_usleep[odir]) {
564 usec = td->rate_pending_usleep[ddir];
566 usec = td->rate_pending_usleep[odir];
570 usec = td->rate_pending_usleep[ddir];
574 fio_gettime(&t, NULL);
575 usec_sleep(td, usec);
576 usec = utime_since_now(&t);
578 td->rate_pending_usleep[ddir] -= usec;
581 if (td_rw(td) && __should_check_rate(td, odir))
582 td->rate_pending_usleep[odir] -= usec;
591 * Return the data direction for the next io_u. If the job is a
592 * mixed read/write workload, check the rwmix cycle and switch if
595 static enum fio_ddir get_rw_ddir(struct thread_data *td)
600 * see if it's time to fsync
602 if (td->o.fsync_blocks &&
603 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
604 td->io_issues[DDIR_WRITE] && should_fsync(td))
608 * see if it's time to fdatasync
610 if (td->o.fdatasync_blocks &&
611 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
612 td->io_issues[DDIR_WRITE] && should_fsync(td))
613 return DDIR_DATASYNC;
616 * see if it's time to sync_file_range
618 if (td->sync_file_range_nr &&
619 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
620 td->io_issues[DDIR_WRITE] && should_fsync(td))
621 return DDIR_SYNC_FILE_RANGE;
625 * Check if it's time to seed a new data direction.
627 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
629 * Put a top limit on how many bytes we do for
630 * one data direction, to avoid overflowing the
633 ddir = get_rand_ddir(td);
635 if (ddir != td->rwmix_ddir)
638 td->rwmix_ddir = ddir;
640 ddir = td->rwmix_ddir;
641 } else if (td_read(td))
643 else if (td_write(td))
648 td->rwmix_ddir = rate_ddir(td, ddir);
649 return td->rwmix_ddir;
652 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
654 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
656 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
657 td->o.barrier_blocks &&
658 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
659 td->io_issues[DDIR_WRITE])
660 io_u->flags |= IO_U_F_BARRIER;
663 void put_file_log(struct thread_data *td, struct fio_file *f)
665 int ret = put_file(td, f);
668 td_verror(td, ret, "file close");
671 void put_io_u(struct thread_data *td, struct io_u *io_u)
675 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
676 put_file_log(td, io_u->file);
678 io_u->flags &= ~IO_U_F_FREE_DEF;
679 io_u->flags |= IO_U_F_FREE;
681 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
683 io_u_qpush(&td->io_u_freelist, io_u);
685 td_io_u_free_notify(td);
688 void clear_io_u(struct thread_data *td, struct io_u *io_u)
690 io_u->flags &= ~IO_U_F_FLIGHT;
694 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
696 struct io_u *__io_u = *io_u;
697 enum fio_ddir ddir = acct_ddir(__io_u);
699 dprint(FD_IO, "requeue %p\n", __io_u);
703 __io_u->flags |= IO_U_F_FREE;
704 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
705 td->io_issues[ddir]--;
707 __io_u->flags &= ~IO_U_F_FLIGHT;
708 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
711 io_u_rpush(&td->io_u_requeues, __io_u);
716 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
718 if (td->io_ops->flags & FIO_NOIO)
721 set_rw_ddir(td, io_u);
724 * fsync() or fdatasync() or trim etc, we are done
726 if (!ddir_rw(io_u->ddir))
730 * See if it's time to switch to a new zone
732 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
734 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
735 io_u->file->last_pos = io_u->file->file_offset;
736 td->io_skip_bytes += td->o.zone_skip;
740 * No log, let the seq/rand engine retrieve the next buflen and
743 if (get_next_offset(td, io_u)) {
744 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
748 io_u->buflen = get_next_buflen(td, io_u);
750 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
754 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
755 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
756 dprint(FD_IO, " off=%llu/%lu > %llu\n",
757 (unsigned long long) io_u->offset, io_u->buflen,
758 (unsigned long long) io_u->file->real_file_size);
763 * mark entry before potentially trimming io_u
765 if (td_random(td) && file_randommap(td, io_u->file))
766 mark_random_map(td, io_u);
769 dprint_io_u(io_u, "fill_io_u");
770 td->zone_bytes += io_u->buflen;
774 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
803 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
805 __io_u_mark_map(td->ts.io_u_submit, nr);
806 td->ts.total_submit++;
809 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
811 __io_u_mark_map(td->ts.io_u_complete, nr);
812 td->ts.total_complete++;
815 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
819 switch (td->cur_depth) {
841 td->ts.io_u_map[idx] += nr;
844 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
881 assert(idx < FIO_IO_U_LAT_U_NR);
882 td->ts.io_u_lat_u[idx]++;
885 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
926 assert(idx < FIO_IO_U_LAT_M_NR);
927 td->ts.io_u_lat_m[idx]++;
930 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
933 io_u_mark_lat_usec(td, usec);
935 io_u_mark_lat_msec(td, usec / 1000);
939 * Get next file to service by choosing one at random
941 static struct fio_file *get_next_file_rand(struct thread_data *td,
942 enum fio_file_flags goodf,
943 enum fio_file_flags badf)
952 if (td->o.use_os_rand) {
953 r = os_random_long(&td->next_file_state);
954 fno = (unsigned int) ((double) td->o.nr_files
955 * (r / (OS_RAND_MAX + 1.0)));
957 r = __rand(&td->__next_file_state);
958 fno = (unsigned int) ((double) td->o.nr_files
959 * (r / (FRAND_MAX + 1.0)));
963 if (fio_file_done(f))
966 if (!fio_file_open(f)) {
969 err = td_io_open_file(td, f);
975 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
976 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
980 td_io_close_file(td, f);
985 * Get next file to service by doing round robin between all available ones
987 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
990 unsigned int old_next_file = td->next_file;
996 f = td->files[td->next_file];
999 if (td->next_file >= td->o.nr_files)
1002 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1003 if (fio_file_done(f)) {
1008 if (!fio_file_open(f)) {
1011 err = td_io_open_file(td, f);
1013 dprint(FD_FILE, "error %d on open of %s\n",
1021 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1023 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1027 td_io_close_file(td, f);
1030 } while (td->next_file != old_next_file);
1032 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1036 static struct fio_file *__get_next_file(struct thread_data *td)
1040 assert(td->o.nr_files <= td->files_index);
1042 if (td->nr_done_files >= td->o.nr_files) {
1043 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1044 " nr_files=%d\n", td->nr_open_files,
1050 f = td->file_service_file;
1051 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1052 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1054 if (td->file_service_left--)
1058 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1059 td->o.file_service_type == FIO_FSERVICE_SEQ)
1060 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1062 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1064 td->file_service_file = f;
1065 td->file_service_left = td->file_service_nr - 1;
1067 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1071 static struct fio_file *get_next_file(struct thread_data *td)
1073 if (!(td->flags & TD_F_PROFILE_OPS)) {
1074 struct prof_io_ops *ops = &td->prof_io_ops;
1076 if (ops->get_next_file)
1077 return ops->get_next_file(td);
1080 return __get_next_file(td);
1083 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1088 f = get_next_file(td);
1095 if (!fill_io_u(td, io_u))
1098 put_file_log(td, f);
1099 td_io_close_file(td, f);
1101 fio_file_set_done(f);
1102 td->nr_done_files++;
1103 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1104 td->nr_done_files, td->o.nr_files);
1111 struct io_u *__get_io_u(struct thread_data *td)
1118 if (!io_u_rempty(&td->io_u_requeues))
1119 io_u = io_u_rpop(&td->io_u_requeues);
1120 else if (!io_u_qempty(&td->io_u_freelist))
1121 io_u = io_u_qpop(&td->io_u_freelist);
1127 io_u->end_io = NULL;
1131 assert(io_u->flags & IO_U_F_FREE);
1132 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1133 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1134 io_u->flags &= ~IO_U_F_VER_LIST;
1137 io_u->acct_ddir = -1;
1139 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1140 } else if (td->o.verify_async) {
1142 * We ran out, wait for async verify threads to finish and
1145 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1153 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1155 if (!(td->flags & TD_F_TRIM_BACKLOG))
1158 if (td->trim_entries) {
1161 if (td->trim_batch) {
1164 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1165 td->last_ddir != DDIR_READ) {
1166 td->trim_batch = td->o.trim_batch;
1167 if (!td->trim_batch)
1168 td->trim_batch = td->o.trim_backlog;
1172 if (get_trim && !get_next_trim(td, io_u))
1179 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1181 if (!(td->flags & TD_F_VER_BACKLOG))
1184 if (td->io_hist_len) {
1187 if (td->verify_batch)
1189 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1190 td->last_ddir != DDIR_READ) {
1191 td->verify_batch = td->o.verify_batch;
1192 if (!td->verify_batch)
1193 td->verify_batch = td->o.verify_backlog;
1197 if (get_verify && !get_next_verify(td, io_u)) {
1207 * Fill offset and start time into the buffer content, to prevent too
1208 * easy compressible data for simple de-dupe attempts. Do this for every
1209 * 512b block in the range, since that should be the smallest block size
1210 * we can expect from a device.
1212 static void small_content_scramble(struct io_u *io_u)
1214 unsigned int i, nr_blocks = io_u->buflen / 512;
1216 unsigned int offset;
1223 boffset = io_u->offset;
1224 io_u->buf_filled_len = 0;
1226 for (i = 0; i < nr_blocks; i++) {
1228 * Fill the byte offset into a "random" start offset of
1229 * the buffer, given by the product of the usec time
1230 * and the actual offset.
1232 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1233 offset &= ~(sizeof(uint64_t) - 1);
1234 if (offset >= 512 - sizeof(uint64_t))
1235 offset -= sizeof(uint64_t);
1236 memcpy(p + offset, &boffset, sizeof(boffset));
1238 end = p + 512 - sizeof(io_u->start_time);
1239 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1246 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1247 * etc. The returned io_u is fully ready to be prepped and submitted.
1249 struct io_u *get_io_u(struct thread_data *td)
1253 int do_scramble = 0;
1255 io_u = __get_io_u(td);
1257 dprint(FD_IO, "__get_io_u failed\n");
1261 if (check_get_verify(td, io_u))
1263 if (check_get_trim(td, io_u))
1267 * from a requeue, io_u already setup
1273 * If using an iolog, grab next piece if any available.
1275 if (td->flags & TD_F_READ_IOLOG) {
1276 if (read_iolog_get(td, io_u))
1278 } else if (set_io_u_file(td, io_u)) {
1279 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1284 assert(fio_file_open(f));
1286 if (ddir_rw(io_u->ddir)) {
1287 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1288 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1292 f->last_start = io_u->offset;
1293 f->last_pos = io_u->offset + io_u->buflen;
1295 if (io_u->ddir == DDIR_WRITE) {
1296 if (td->flags & TD_F_REFILL_BUFFERS) {
1297 io_u_fill_buffer(td, io_u,
1298 io_u->xfer_buflen, io_u->xfer_buflen);
1299 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1301 if (td->flags & TD_F_VER_NONE) {
1302 populate_verify_io_u(td, io_u);
1305 } else if (io_u->ddir == DDIR_READ) {
1307 * Reset the buf_filled parameters so next time if the
1308 * buffer is used for writes it is refilled.
1310 io_u->buf_filled_len = 0;
1315 * Set io data pointers.
1317 io_u->xfer_buf = io_u->buf;
1318 io_u->xfer_buflen = io_u->buflen;
1322 if (!td_io_prep(td, io_u)) {
1323 if (!td->o.disable_slat)
1324 fio_gettime(&io_u->start_time, NULL);
1326 small_content_scramble(io_u);
1330 dprint(FD_IO, "get_io_u failed\n");
1335 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1337 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1338 const char *msg[] = { "read", "write", "sync", "datasync",
1339 "sync_file_range", "wait", "trim" };
1341 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1344 log_err("fio: io_u error");
1347 log_err(" on file %s", io_u->file->file_name);
1349 log_err(": %s\n", strerror(io_u->error));
1351 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1352 io_u->offset, io_u->xfer_buflen);
1355 td_verror(td, io_u->error, "io_u error");
1358 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1359 struct io_completion_data *icd,
1360 const enum fio_ddir idx, unsigned int bytes)
1362 unsigned long lusec = 0;
1364 if (!td->o.disable_clat || !td->o.disable_bw)
1365 lusec = utime_since(&io_u->issue_time, &icd->time);
1367 if (!td->o.disable_lat) {
1368 unsigned long tusec;
1370 tusec = utime_since(&io_u->start_time, &icd->time);
1371 add_lat_sample(td, idx, tusec, bytes);
1373 if (td->flags & TD_F_PROFILE_OPS) {
1374 struct prof_io_ops *ops = &td->prof_io_ops;
1377 icd->error = ops->io_u_lat(td, tusec);
1380 if (td->o.max_latency && tusec > td->o.max_latency) {
1382 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1383 td_verror(td, ETIMEDOUT, "max latency exceeded");
1384 icd->error = ETIMEDOUT;
1388 if (!td->o.disable_clat) {
1389 add_clat_sample(td, idx, lusec, bytes);
1390 io_u_mark_latency(td, lusec);
1393 if (!td->o.disable_bw)
1394 add_bw_sample(td, idx, bytes, &icd->time);
1396 add_iops_sample(td, idx, &icd->time);
1399 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1401 uint64_t secs, remainder, bps, bytes;
1403 bytes = td->this_io_bytes[ddir];
1404 bps = td->rate_bps[ddir];
1406 remainder = bytes % bps;
1407 return remainder * 1000000 / bps + secs * 1000000;
1410 static void io_completed(struct thread_data *td, struct io_u *io_u,
1411 struct io_completion_data *icd)
1415 dprint_io_u(io_u, "io complete");
1418 assert(io_u->flags & IO_U_F_FLIGHT);
1419 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1422 if (ddir_sync(io_u->ddir)) {
1423 td->last_was_sync = 1;
1426 f->first_write = -1ULL;
1427 f->last_write = -1ULL;
1432 td->last_was_sync = 0;
1433 td->last_ddir = io_u->ddir;
1435 if (!io_u->error && ddir_rw(io_u->ddir)) {
1436 unsigned int bytes = io_u->buflen - io_u->resid;
1437 const enum fio_ddir idx = io_u->ddir;
1438 const enum fio_ddir odx = io_u->ddir ^ 1;
1441 td->io_blocks[idx]++;
1442 td->this_io_blocks[idx]++;
1443 td->io_bytes[idx] += bytes;
1445 if (!(io_u->flags & IO_U_F_VER_LIST))
1446 td->this_io_bytes[idx] += bytes;
1448 if (idx == DDIR_WRITE) {
1451 if (f->first_write == -1ULL ||
1452 io_u->offset < f->first_write)
1453 f->first_write = io_u->offset;
1454 if (f->last_write == -1ULL ||
1455 ((io_u->offset + bytes) > f->last_write))
1456 f->last_write = io_u->offset + bytes;
1460 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1461 td->runstate == TD_VERIFYING)) {
1462 account_io_completion(td, io_u, icd, idx, bytes);
1464 if (__should_check_rate(td, idx)) {
1465 td->rate_pending_usleep[idx] =
1466 (usec_for_io(td, idx) -
1467 utime_since_now(&td->start));
1469 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1470 td->rate_pending_usleep[odx] =
1471 (usec_for_io(td, odx) -
1472 utime_since_now(&td->start));
1475 if (td_write(td) && idx == DDIR_WRITE &&
1477 td->o.verify != VERIFY_NONE &&
1478 !td->o.experimental_verify)
1479 log_io_piece(td, io_u);
1481 icd->bytes_done[idx] += bytes;
1484 ret = io_u->end_io(td, io_u);
1485 if (ret && !icd->error)
1488 } else if (io_u->error) {
1489 icd->error = io_u->error;
1490 io_u_log_error(td, io_u);
1493 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1494 if (!td_non_fatal_error(td, eb, icd->error))
1497 * If there is a non_fatal error, then add to the error count
1498 * and clear all the errors.
1500 update_error_count(td, icd->error);
1507 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1511 if (!td->o.disable_clat || !td->o.disable_bw)
1512 fio_gettime(&icd->time, NULL);
1517 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1518 icd->bytes_done[ddir] = 0;
1521 static void ios_completed(struct thread_data *td,
1522 struct io_completion_data *icd)
1527 for (i = 0; i < icd->nr; i++) {
1528 io_u = td->io_ops->event(td, i);
1530 io_completed(td, io_u, icd);
1532 if (!(io_u->flags & IO_U_F_FREE_DEF))
1538 * Complete a single io_u for the sync engines.
1540 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1543 struct io_completion_data icd;
1545 init_icd(td, &icd, 1);
1546 io_completed(td, io_u, &icd);
1548 if (!(io_u->flags & IO_U_F_FREE_DEF))
1552 td_verror(td, icd.error, "io_u_sync_complete");
1559 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1560 bytes[ddir] += icd.bytes_done[ddir];
1567 * Called to complete min_events number of io for the async engines.
1569 int io_u_queued_complete(struct thread_data *td, int min_evts,
1572 struct io_completion_data icd;
1573 struct timespec *tvp = NULL;
1575 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1577 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1582 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1584 td_verror(td, -ret, "td_io_getevents");
1589 init_icd(td, &icd, ret);
1590 ios_completed(td, &icd);
1592 td_verror(td, icd.error, "io_u_queued_complete");
1599 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1600 bytes[ddir] += icd.bytes_done[ddir];
1607 * Call when io_u is really queued, to update the submission latency.
1609 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1611 if (!td->o.disable_slat) {
1612 unsigned long slat_time;
1614 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1615 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1619 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1620 unsigned int max_bs)
1622 if (!td->o.zero_buffers) {
1623 unsigned int perc = td->o.compress_percentage;
1626 unsigned int seg = min_write;
1628 seg = min(min_write, td->o.compress_chunk);
1632 fill_random_buf_percentage(&td->buf_state, buf,
1635 fill_random_buf(&td->buf_state, buf, max_bs);
1637 memset(buf, 0, max_bs);
1641 * "randomly" fill the buffer contents
1643 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1644 unsigned int min_write, unsigned int max_bs)
1646 io_u->buf_filled_len = 0;
1647 fill_io_buffer(td, io_u->buf, min_write, max_bs);