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 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
188 return get_off_from_method(td, f, ddir, b);
190 if (!flist_empty(&td->next_rand_list)) {
193 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
200 for (i = 0; i < td->o.verifysort_nr; i++) {
201 r = malloc(sizeof(*r));
203 ret = get_off_from_method(td, f, ddir, &r->off);
209 flist_add(&r->list, &td->next_rand_list);
215 assert(!flist_empty(&td->next_rand_list));
216 flist_sort(NULL, &td->next_rand_list, flist_cmp);
220 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
221 enum fio_ddir ddir, uint64_t *b)
223 if (!get_next_rand_offset(td, f, ddir, b))
226 if (td->o.time_based) {
227 fio_file_reset(td, f);
228 if (!get_next_rand_offset(td, f, ddir, b))
232 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
233 f->file_name, f->last_pos, f->real_file_size);
237 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, uint64_t *offset)
240 assert(ddir_rw(ddir));
242 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
243 f->last_pos = f->last_pos - f->io_size;
245 if (f->last_pos < f->real_file_size) {
248 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
249 f->last_pos = f->real_file_size;
251 pos = f->last_pos - f->file_offset;
253 pos += td->o.ddir_seq_add;
262 static int get_next_block(struct thread_data *td, struct io_u *io_u,
263 enum fio_ddir ddir, int rw_seq)
265 struct fio_file *f = io_u->file;
269 assert(ddir_rw(ddir));
275 ret = get_next_rand_block(td, f, ddir, &b);
277 ret = get_next_seq_offset(td, f, ddir, &offset);
279 io_u->flags |= IO_U_F_BUSY_OK;
281 if (td->o.rw_seq == RW_SEQ_SEQ) {
282 ret = get_next_seq_offset(td, f, ddir, &offset);
284 ret = get_next_rand_block(td, f, ddir, &b);
285 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
286 if (f->last_start != -1ULL)
287 offset = f->last_start - f->file_offset;
292 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
299 io_u->offset = offset;
301 io_u->offset = b * td->o.ba[ddir];
303 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
313 * For random io, generate a random new block and see if it's used. Repeat
314 * until we find a free one. For sequential io, just return the end of
315 * the last io issued.
317 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
319 struct fio_file *f = io_u->file;
320 enum fio_ddir ddir = io_u->ddir;
323 assert(ddir_rw(ddir));
325 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
327 td->ddir_seq_nr = td->o.ddir_seq_nr;
330 if (get_next_block(td, io_u, ddir, rw_seq_hit))
333 if (io_u->offset >= f->io_size) {
334 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
335 io_u->offset, f->io_size);
339 io_u->offset += f->file_offset;
340 if (io_u->offset >= f->real_file_size) {
341 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
342 io_u->offset, f->real_file_size);
349 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
351 if (td->flags & TD_F_PROFILE_OPS) {
352 struct prof_io_ops *ops = &td->prof_io_ops;
354 if (ops->fill_io_u_off)
355 return ops->fill_io_u_off(td, io_u);
358 return __get_next_offset(td, io_u);
361 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
364 struct fio_file *f = io_u->file;
366 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
369 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
371 const int ddir = io_u->ddir;
372 unsigned int buflen = 0;
373 unsigned int minbs, maxbs;
374 unsigned long r, rand_max;
376 assert(ddir_rw(ddir));
378 minbs = td->o.min_bs[ddir];
379 maxbs = td->o.max_bs[ddir];
385 * If we can't satisfy the min block size from here, then fail
387 if (!io_u_fits(td, io_u, minbs))
390 if (td->o.use_os_rand)
391 rand_max = OS_RAND_MAX;
393 rand_max = FRAND_MAX;
396 if (td->o.use_os_rand)
397 r = os_random_long(&td->bsrange_state);
399 r = __rand(&td->__bsrange_state);
401 if (!td->o.bssplit_nr[ddir]) {
402 buflen = 1 + (unsigned int) ((double) maxbs *
403 (r / (rand_max + 1.0)));
410 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
411 struct bssplit *bsp = &td->o.bssplit[ddir][i];
415 if ((r <= ((rand_max / 100L) * perc)) &&
416 io_u_fits(td, io_u, buflen))
421 if (!td->o.bs_unaligned && is_power_of_2(minbs))
422 buflen = (buflen + minbs - 1) & ~(minbs - 1);
424 } while (!io_u_fits(td, io_u, buflen));
429 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
431 if (td->flags & TD_F_PROFILE_OPS) {
432 struct prof_io_ops *ops = &td->prof_io_ops;
434 if (ops->fill_io_u_size)
435 return ops->fill_io_u_size(td, io_u);
438 return __get_next_buflen(td, io_u);
441 static void set_rwmix_bytes(struct thread_data *td)
446 * we do time or byte based switch. this is needed because
447 * buffered writes may issue a lot quicker than they complete,
448 * whereas reads do not.
450 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
451 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
454 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
459 if (td->o.use_os_rand) {
460 r = os_random_long(&td->rwmix_state);
461 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
463 r = __rand(&td->__rwmix_state);
464 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
467 if (v <= td->o.rwmix[DDIR_READ])
473 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
475 enum fio_ddir odir = ddir ^ 1;
479 assert(ddir_rw(ddir));
481 if (td->rate_pending_usleep[ddir] <= 0)
485 * We have too much pending sleep in this direction. See if we
490 * Other direction does not have too much pending, switch
492 if (td->rate_pending_usleep[odir] < 100000)
496 * Both directions have pending sleep. Sleep the minimum time
497 * and deduct from both.
499 if (td->rate_pending_usleep[ddir] <=
500 td->rate_pending_usleep[odir]) {
501 usec = td->rate_pending_usleep[ddir];
503 usec = td->rate_pending_usleep[odir];
507 usec = td->rate_pending_usleep[ddir];
510 * We are going to sleep, ensure that we flush anything pending as
511 * not to skew our latency numbers.
513 * Changed to only monitor 'in flight' requests here instead of the
514 * td->cur_depth, b/c td->cur_depth does not accurately represent
515 * io's that have been actually submitted to an async engine,
516 * and cur_depth is meaningless for sync engines.
518 if (td->io_u_in_flight) {
521 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
524 fio_gettime(&t, NULL);
525 usec_sleep(td, usec);
526 usec = utime_since_now(&t);
528 td->rate_pending_usleep[ddir] -= usec;
531 if (td_rw(td) && __should_check_rate(td, odir))
532 td->rate_pending_usleep[odir] -= usec;
540 * Return the data direction for the next io_u. If the job is a
541 * mixed read/write workload, check the rwmix cycle and switch if
544 static enum fio_ddir get_rw_ddir(struct thread_data *td)
549 * If verify phase started, it's always a READ
551 if (td->runstate == TD_VERIFYING)
555 * see if it's time to fsync
557 if (td->o.fsync_blocks &&
558 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
559 td->io_issues[DDIR_WRITE] && should_fsync(td))
563 * see if it's time to fdatasync
565 if (td->o.fdatasync_blocks &&
566 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
567 td->io_issues[DDIR_WRITE] && should_fsync(td))
568 return DDIR_DATASYNC;
571 * see if it's time to sync_file_range
573 if (td->sync_file_range_nr &&
574 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
575 td->io_issues[DDIR_WRITE] && should_fsync(td))
576 return DDIR_SYNC_FILE_RANGE;
580 * Check if it's time to seed a new data direction.
582 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
584 * Put a top limit on how many bytes we do for
585 * one data direction, to avoid overflowing the
588 ddir = get_rand_ddir(td);
590 if (ddir != td->rwmix_ddir)
593 td->rwmix_ddir = ddir;
595 ddir = td->rwmix_ddir;
596 } else if (td_read(td))
598 else if (td_write(td))
603 td->rwmix_ddir = rate_ddir(td, ddir);
604 return td->rwmix_ddir;
607 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
609 io_u->ddir = get_rw_ddir(td);
611 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
612 td->o.barrier_blocks &&
613 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
614 td->io_issues[DDIR_WRITE])
615 io_u->flags |= IO_U_F_BARRIER;
618 void put_file_log(struct thread_data *td, struct fio_file *f)
620 int ret = put_file(td, f);
623 td_verror(td, ret, "file close");
626 void put_io_u(struct thread_data *td, struct io_u *io_u)
630 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
631 put_file_log(td, io_u->file);
633 io_u->flags &= ~IO_U_F_FREE_DEF;
634 io_u->flags |= IO_U_F_FREE;
636 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
638 flist_del_init(&io_u->list);
639 flist_add(&io_u->list, &td->io_u_freelist);
641 td_io_u_free_notify(td);
644 void clear_io_u(struct thread_data *td, struct io_u *io_u)
646 io_u->flags &= ~IO_U_F_FLIGHT;
650 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
652 struct io_u *__io_u = *io_u;
654 dprint(FD_IO, "requeue %p\n", __io_u);
658 __io_u->flags |= IO_U_F_FREE;
659 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
660 td->io_issues[__io_u->ddir]--;
662 __io_u->flags &= ~IO_U_F_FLIGHT;
663 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
665 flist_del(&__io_u->list);
666 flist_add_tail(&__io_u->list, &td->io_u_requeues);
671 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
673 if (td->io_ops->flags & FIO_NOIO)
676 set_rw_ddir(td, io_u);
679 * fsync() or fdatasync() or trim etc, we are done
681 if (!ddir_rw(io_u->ddir))
685 * See if it's time to switch to a new zone
687 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
689 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
690 io_u->file->last_pos = io_u->file->file_offset;
691 td->io_skip_bytes += td->o.zone_skip;
695 * No log, let the seq/rand engine retrieve the next buflen and
698 if (get_next_offset(td, io_u)) {
699 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
703 io_u->buflen = get_next_buflen(td, io_u);
705 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
709 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
710 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
711 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
712 io_u->buflen, io_u->file->real_file_size);
717 * mark entry before potentially trimming io_u
719 if (td_random(td) && file_randommap(td, io_u->file))
720 mark_random_map(td, io_u);
723 * If using a write iolog, store this entry.
726 dprint_io_u(io_u, "fill_io_u");
727 td->zone_bytes += io_u->buflen;
732 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
761 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
763 __io_u_mark_map(td->ts.io_u_submit, nr);
764 td->ts.total_submit++;
767 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
769 __io_u_mark_map(td->ts.io_u_complete, nr);
770 td->ts.total_complete++;
773 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
777 switch (td->cur_depth) {
799 td->ts.io_u_map[idx] += nr;
802 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
839 assert(idx < FIO_IO_U_LAT_U_NR);
840 td->ts.io_u_lat_u[idx]++;
843 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
884 assert(idx < FIO_IO_U_LAT_M_NR);
885 td->ts.io_u_lat_m[idx]++;
888 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
891 io_u_mark_lat_usec(td, usec);
893 io_u_mark_lat_msec(td, usec / 1000);
897 * Get next file to service by choosing one at random
899 static struct fio_file *get_next_file_rand(struct thread_data *td,
900 enum fio_file_flags goodf,
901 enum fio_file_flags badf)
910 if (td->o.use_os_rand) {
911 r = os_random_long(&td->next_file_state);
912 fno = (unsigned int) ((double) td->o.nr_files
913 * (r / (OS_RAND_MAX + 1.0)));
915 r = __rand(&td->__next_file_state);
916 fno = (unsigned int) ((double) td->o.nr_files
917 * (r / (FRAND_MAX + 1.0)));
921 if (fio_file_done(f))
924 if (!fio_file_open(f)) {
927 err = td_io_open_file(td, f);
933 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
934 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
938 td_io_close_file(td, f);
943 * Get next file to service by doing round robin between all available ones
945 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
948 unsigned int old_next_file = td->next_file;
954 f = td->files[td->next_file];
957 if (td->next_file >= td->o.nr_files)
960 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
961 if (fio_file_done(f)) {
966 if (!fio_file_open(f)) {
969 err = td_io_open_file(td, f);
971 dprint(FD_FILE, "error %d on open of %s\n",
979 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
981 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
985 td_io_close_file(td, f);
988 } while (td->next_file != old_next_file);
990 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
994 static struct fio_file *__get_next_file(struct thread_data *td)
998 assert(td->o.nr_files <= td->files_index);
1000 if (td->nr_done_files >= td->o.nr_files) {
1001 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1002 " nr_files=%d\n", td->nr_open_files,
1008 f = td->file_service_file;
1009 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1010 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1012 if (td->file_service_left--)
1016 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1017 td->o.file_service_type == FIO_FSERVICE_SEQ)
1018 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1020 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1022 td->file_service_file = f;
1023 td->file_service_left = td->file_service_nr - 1;
1025 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1029 static struct fio_file *get_next_file(struct thread_data *td)
1031 if (!(td->flags & TD_F_PROFILE_OPS)) {
1032 struct prof_io_ops *ops = &td->prof_io_ops;
1034 if (ops->get_next_file)
1035 return ops->get_next_file(td);
1038 return __get_next_file(td);
1041 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1046 f = get_next_file(td);
1053 if (!fill_io_u(td, io_u))
1056 put_file_log(td, f);
1057 td_io_close_file(td, f);
1059 fio_file_set_done(f);
1060 td->nr_done_files++;
1061 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1062 td->nr_done_files, td->o.nr_files);
1069 struct io_u *__get_io_u(struct thread_data *td)
1071 struct io_u *io_u = NULL;
1076 if (!flist_empty(&td->io_u_requeues))
1077 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1078 else if (!queue_full(td)) {
1079 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1084 io_u->end_io = NULL;
1088 assert(io_u->flags & IO_U_F_FREE);
1089 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1090 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1091 io_u->flags &= ~IO_U_F_VER_LIST;
1094 flist_del(&io_u->list);
1095 flist_add_tail(&io_u->list, &td->io_u_busylist);
1097 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1098 } else if (td->o.verify_async) {
1100 * We ran out, wait for async verify threads to finish and
1103 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1111 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1113 if (!(td->flags & TD_F_TRIM_BACKLOG))
1116 if (td->trim_entries) {
1119 if (td->trim_batch) {
1122 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1123 td->last_ddir != DDIR_READ) {
1124 td->trim_batch = td->o.trim_batch;
1125 if (!td->trim_batch)
1126 td->trim_batch = td->o.trim_backlog;
1130 if (get_trim && !get_next_trim(td, io_u))
1137 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1139 if (!(td->flags & TD_F_VER_BACKLOG))
1142 if (td->io_hist_len) {
1145 if (td->verify_batch)
1147 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1148 td->last_ddir != DDIR_READ) {
1149 td->verify_batch = td->o.verify_batch;
1150 if (!td->verify_batch)
1151 td->verify_batch = td->o.verify_backlog;
1155 if (get_verify && !get_next_verify(td, io_u)) {
1165 * Fill offset and start time into the buffer content, to prevent too
1166 * easy compressible data for simple de-dupe attempts. Do this for every
1167 * 512b block in the range, since that should be the smallest block size
1168 * we can expect from a device.
1170 static void small_content_scramble(struct io_u *io_u)
1172 unsigned int i, nr_blocks = io_u->buflen / 512;
1174 unsigned int offset;
1181 boffset = io_u->offset;
1182 io_u->buf_filled_len = 0;
1184 for (i = 0; i < nr_blocks; i++) {
1186 * Fill the byte offset into a "random" start offset of
1187 * the buffer, given by the product of the usec time
1188 * and the actual offset.
1190 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1191 offset &= ~(sizeof(uint64_t) - 1);
1192 if (offset >= 512 - sizeof(uint64_t))
1193 offset -= sizeof(uint64_t);
1194 memcpy(p + offset, &boffset, sizeof(boffset));
1196 end = p + 512 - sizeof(io_u->start_time);
1197 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1204 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1205 * etc. The returned io_u is fully ready to be prepped and submitted.
1207 struct io_u *get_io_u(struct thread_data *td)
1211 int do_scramble = 0;
1213 io_u = __get_io_u(td);
1215 dprint(FD_IO, "__get_io_u failed\n");
1219 if (check_get_verify(td, io_u))
1221 if (check_get_trim(td, io_u))
1225 * from a requeue, io_u already setup
1231 * If using an iolog, grab next piece if any available.
1233 if (td->flags & TD_F_READ_IOLOG) {
1234 if (read_iolog_get(td, io_u))
1236 } else if (set_io_u_file(td, io_u)) {
1237 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1242 assert(fio_file_open(f));
1244 if (ddir_rw(io_u->ddir)) {
1245 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1246 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1250 f->last_start = io_u->offset;
1251 f->last_pos = io_u->offset + io_u->buflen;
1253 if (io_u->ddir == DDIR_WRITE) {
1254 if (td->flags & TD_F_REFILL_BUFFERS) {
1255 io_u_fill_buffer(td, io_u,
1256 io_u->xfer_buflen, io_u->xfer_buflen);
1257 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1259 if (td->flags & TD_F_VER_NONE) {
1260 populate_verify_io_u(td, io_u);
1263 } else if (io_u->ddir == DDIR_READ) {
1265 * Reset the buf_filled parameters so next time if the
1266 * buffer is used for writes it is refilled.
1268 io_u->buf_filled_len = 0;
1273 * Set io data pointers.
1275 io_u->xfer_buf = io_u->buf;
1276 io_u->xfer_buflen = io_u->buflen;
1280 if (!td_io_prep(td, io_u)) {
1281 if (!td->o.disable_slat)
1282 fio_gettime(&io_u->start_time, NULL);
1284 small_content_scramble(io_u);
1288 dprint(FD_IO, "get_io_u failed\n");
1293 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1295 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1296 const char *msg[] = { "read", "write", "sync", "datasync",
1297 "sync_file_range", "wait", "trim" };
1299 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1302 log_err("fio: io_u error");
1305 log_err(" on file %s", io_u->file->file_name);
1307 log_err(": %s\n", strerror(io_u->error));
1309 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1310 io_u->offset, io_u->xfer_buflen);
1313 td_verror(td, io_u->error, "io_u error");
1316 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1317 struct io_completion_data *icd,
1318 const enum fio_ddir idx, unsigned int bytes)
1320 unsigned long lusec = 0;
1322 if (!td->o.disable_clat || !td->o.disable_bw)
1323 lusec = utime_since(&io_u->issue_time, &icd->time);
1325 if (!td->o.disable_lat) {
1326 unsigned long tusec;
1328 tusec = utime_since(&io_u->start_time, &icd->time);
1329 add_lat_sample(td, idx, tusec, bytes);
1331 if (td->o.max_latency && tusec > td->o.max_latency) {
1333 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1334 td_verror(td, ETIMEDOUT, "max latency exceeded");
1335 icd->error = ETIMEDOUT;
1339 if (!td->o.disable_clat) {
1340 add_clat_sample(td, idx, lusec, bytes);
1341 io_u_mark_latency(td, lusec);
1344 if (!td->o.disable_bw)
1345 add_bw_sample(td, idx, bytes, &icd->time);
1347 add_iops_sample(td, idx, &icd->time);
1350 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1352 uint64_t secs, remainder, bps, bytes;
1354 bytes = td->this_io_bytes[ddir];
1355 bps = td->rate_bps[ddir];
1357 remainder = bytes % bps;
1358 return remainder * 1000000 / bps + secs * 1000000;
1361 static void io_completed(struct thread_data *td, struct io_u *io_u,
1362 struct io_completion_data *icd)
1366 dprint_io_u(io_u, "io complete");
1369 assert(io_u->flags & IO_U_F_FLIGHT);
1370 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1373 if (ddir_sync(io_u->ddir)) {
1374 td->last_was_sync = 1;
1377 f->first_write = -1ULL;
1378 f->last_write = -1ULL;
1383 td->last_was_sync = 0;
1384 td->last_ddir = io_u->ddir;
1386 if (!io_u->error && ddir_rw(io_u->ddir)) {
1387 unsigned int bytes = io_u->buflen - io_u->resid;
1388 const enum fio_ddir idx = io_u->ddir;
1389 const enum fio_ddir odx = io_u->ddir ^ 1;
1392 td->io_blocks[idx]++;
1393 td->this_io_blocks[idx]++;
1394 td->io_bytes[idx] += bytes;
1396 if (!(io_u->flags & IO_U_F_VER_LIST))
1397 td->this_io_bytes[idx] += bytes;
1399 if (idx == DDIR_WRITE) {
1402 if (f->first_write == -1ULL ||
1403 io_u->offset < f->first_write)
1404 f->first_write = io_u->offset;
1405 if (f->last_write == -1ULL ||
1406 ((io_u->offset + bytes) > f->last_write))
1407 f->last_write = io_u->offset + bytes;
1411 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1412 td->runstate == TD_VERIFYING)) {
1413 account_io_completion(td, io_u, icd, idx, bytes);
1415 if (__should_check_rate(td, idx)) {
1416 td->rate_pending_usleep[idx] =
1417 (usec_for_io(td, idx) -
1418 utime_since_now(&td->start));
1420 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1421 td->rate_pending_usleep[odx] =
1422 (usec_for_io(td, odx) -
1423 utime_since_now(&td->start));
1426 if (td_write(td) && idx == DDIR_WRITE &&
1428 td->o.verify != VERIFY_NONE &&
1429 !td->o.experimental_verify)
1430 log_io_piece(td, io_u);
1432 icd->bytes_done[idx] += bytes;
1435 ret = io_u->end_io(td, io_u);
1436 if (ret && !icd->error)
1439 } else if (io_u->error) {
1440 icd->error = io_u->error;
1441 io_u_log_error(td, io_u);
1444 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1445 if (!td_non_fatal_error(td, eb, icd->error))
1448 * If there is a non_fatal error, then add to the error count
1449 * and clear all the errors.
1451 update_error_count(td, icd->error);
1458 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1462 if (!td->o.disable_clat || !td->o.disable_bw)
1463 fio_gettime(&icd->time, NULL);
1468 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1469 icd->bytes_done[ddir] = 0;
1472 static void ios_completed(struct thread_data *td,
1473 struct io_completion_data *icd)
1478 for (i = 0; i < icd->nr; i++) {
1479 io_u = td->io_ops->event(td, i);
1481 io_completed(td, io_u, icd);
1483 if (!(io_u->flags & IO_U_F_FREE_DEF))
1489 * Complete a single io_u for the sync engines.
1491 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1492 unsigned long *bytes)
1494 struct io_completion_data icd;
1496 init_icd(td, &icd, 1);
1497 io_completed(td, io_u, &icd);
1499 if (!(io_u->flags & IO_U_F_FREE_DEF))
1503 td_verror(td, icd.error, "io_u_sync_complete");
1510 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1511 bytes[ddir] += icd.bytes_done[ddir];
1518 * Called to complete min_events number of io for the async engines.
1520 int io_u_queued_complete(struct thread_data *td, int min_evts,
1521 unsigned long *bytes)
1523 struct io_completion_data icd;
1524 struct timespec *tvp = NULL;
1526 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1528 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1533 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1535 td_verror(td, -ret, "td_io_getevents");
1540 init_icd(td, &icd, ret);
1541 ios_completed(td, &icd);
1543 td_verror(td, icd.error, "io_u_queued_complete");
1550 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1551 bytes[ddir] += icd.bytes_done[ddir];
1558 * Call when io_u is really queued, to update the submission latency.
1560 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1562 if (!td->o.disable_slat) {
1563 unsigned long slat_time;
1565 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1566 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1571 * "randomly" fill the buffer contents
1573 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1574 unsigned int min_write, unsigned int max_bs)
1576 io_u->buf_filled_len = 0;
1578 if (!td->o.zero_buffers) {
1579 unsigned int perc = td->o.compress_percentage;
1582 unsigned int seg = min_write;
1584 seg = min(min_write, td->o.compress_chunk);
1585 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1588 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1590 memset(io_u->buf, 0, max_bs);