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.bs_unaligned && is_power_of_2(minbs))
463 buflen = (buflen + minbs - 1) & ~(minbs - 1);
465 } while (!io_u_fits(td, io_u, buflen));
470 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
472 if (td->flags & TD_F_PROFILE_OPS) {
473 struct prof_io_ops *ops = &td->prof_io_ops;
475 if (ops->fill_io_u_size)
476 return ops->fill_io_u_size(td, io_u);
479 return __get_next_buflen(td, io_u);
482 static void set_rwmix_bytes(struct thread_data *td)
487 * we do time or byte based switch. this is needed because
488 * buffered writes may issue a lot quicker than they complete,
489 * whereas reads do not.
491 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
492 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
495 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
500 if (td->o.use_os_rand) {
501 r = os_random_long(&td->rwmix_state);
502 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
504 r = __rand(&td->__rwmix_state);
505 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
508 if (v <= td->o.rwmix[DDIR_READ])
514 void io_u_quiesce(struct thread_data *td)
517 * We are going to sleep, ensure that we flush anything pending as
518 * not to skew our latency numbers.
520 * Changed to only monitor 'in flight' requests here instead of the
521 * td->cur_depth, b/c td->cur_depth does not accurately represent
522 * io's that have been actually submitted to an async engine,
523 * and cur_depth is meaningless for sync engines.
525 while (td->io_u_in_flight) {
528 ret = io_u_queued_complete(td, 1, NULL);
532 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
534 enum fio_ddir odir = ddir ^ 1;
538 assert(ddir_rw(ddir));
540 if (td->rate_pending_usleep[ddir] <= 0)
544 * We have too much pending sleep in this direction. See if we
547 if (td_rw(td) && td->o.rwmix[odir]) {
549 * Other direction does not have too much pending, switch
551 if (td->rate_pending_usleep[odir] < 100000)
555 * Both directions have pending sleep. Sleep the minimum time
556 * and deduct from both.
558 if (td->rate_pending_usleep[ddir] <=
559 td->rate_pending_usleep[odir]) {
560 usec = td->rate_pending_usleep[ddir];
562 usec = td->rate_pending_usleep[odir];
566 usec = td->rate_pending_usleep[ddir];
570 fio_gettime(&t, NULL);
571 usec_sleep(td, usec);
572 usec = utime_since_now(&t);
574 td->rate_pending_usleep[ddir] -= usec;
577 if (td_rw(td) && __should_check_rate(td, odir))
578 td->rate_pending_usleep[odir] -= usec;
587 * Return the data direction for the next io_u. If the job is a
588 * mixed read/write workload, check the rwmix cycle and switch if
591 static enum fio_ddir get_rw_ddir(struct thread_data *td)
596 * see if it's time to fsync
598 if (td->o.fsync_blocks &&
599 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
600 td->io_issues[DDIR_WRITE] && should_fsync(td))
604 * see if it's time to fdatasync
606 if (td->o.fdatasync_blocks &&
607 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
608 td->io_issues[DDIR_WRITE] && should_fsync(td))
609 return DDIR_DATASYNC;
612 * see if it's time to sync_file_range
614 if (td->sync_file_range_nr &&
615 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
616 td->io_issues[DDIR_WRITE] && should_fsync(td))
617 return DDIR_SYNC_FILE_RANGE;
621 * Check if it's time to seed a new data direction.
623 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
625 * Put a top limit on how many bytes we do for
626 * one data direction, to avoid overflowing the
629 ddir = get_rand_ddir(td);
631 if (ddir != td->rwmix_ddir)
634 td->rwmix_ddir = ddir;
636 ddir = td->rwmix_ddir;
637 } else if (td_read(td))
639 else if (td_write(td))
644 td->rwmix_ddir = rate_ddir(td, ddir);
645 return td->rwmix_ddir;
648 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
650 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
652 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
653 td->o.barrier_blocks &&
654 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
655 td->io_issues[DDIR_WRITE])
656 io_u->flags |= IO_U_F_BARRIER;
659 void put_file_log(struct thread_data *td, struct fio_file *f)
661 int ret = put_file(td, f);
664 td_verror(td, ret, "file close");
667 void put_io_u(struct thread_data *td, struct io_u *io_u)
671 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
672 put_file_log(td, io_u->file);
674 io_u->flags &= ~IO_U_F_FREE_DEF;
675 io_u->flags |= IO_U_F_FREE;
677 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
679 flist_del_init(&io_u->list);
680 flist_add(&io_u->list, &td->io_u_freelist);
682 td_io_u_free_notify(td);
685 void clear_io_u(struct thread_data *td, struct io_u *io_u)
687 io_u->flags &= ~IO_U_F_FLIGHT;
691 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
693 struct io_u *__io_u = *io_u;
694 enum fio_ddir ddir = acct_ddir(__io_u);
696 dprint(FD_IO, "requeue %p\n", __io_u);
700 __io_u->flags |= IO_U_F_FREE;
701 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
702 td->io_issues[ddir]--;
704 __io_u->flags &= ~IO_U_F_FLIGHT;
705 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
707 flist_del(&__io_u->list);
708 flist_add_tail(&__io_u->list, &td->io_u_requeues);
713 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
715 if (td->io_ops->flags & FIO_NOIO)
718 set_rw_ddir(td, io_u);
721 * fsync() or fdatasync() or trim etc, we are done
723 if (!ddir_rw(io_u->ddir))
727 * See if it's time to switch to a new zone
729 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
731 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
732 io_u->file->last_pos = io_u->file->file_offset;
733 td->io_skip_bytes += td->o.zone_skip;
737 * No log, let the seq/rand engine retrieve the next buflen and
740 if (get_next_offset(td, io_u)) {
741 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
745 io_u->buflen = get_next_buflen(td, io_u);
747 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
751 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
752 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
753 dprint(FD_IO, " off=%llu/%lu > %llu\n",
754 (unsigned long long) io_u->offset, io_u->buflen,
755 (unsigned long long) io_u->file->real_file_size);
760 * mark entry before potentially trimming io_u
762 if (td_random(td) && file_randommap(td, io_u->file))
763 mark_random_map(td, io_u);
766 dprint_io_u(io_u, "fill_io_u");
767 td->zone_bytes += io_u->buflen;
771 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
800 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
802 __io_u_mark_map(td->ts.io_u_submit, nr);
803 td->ts.total_submit++;
806 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
808 __io_u_mark_map(td->ts.io_u_complete, nr);
809 td->ts.total_complete++;
812 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
816 switch (td->cur_depth) {
838 td->ts.io_u_map[idx] += nr;
841 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
878 assert(idx < FIO_IO_U_LAT_U_NR);
879 td->ts.io_u_lat_u[idx]++;
882 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
923 assert(idx < FIO_IO_U_LAT_M_NR);
924 td->ts.io_u_lat_m[idx]++;
927 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
930 io_u_mark_lat_usec(td, usec);
932 io_u_mark_lat_msec(td, usec / 1000);
936 * Get next file to service by choosing one at random
938 static struct fio_file *get_next_file_rand(struct thread_data *td,
939 enum fio_file_flags goodf,
940 enum fio_file_flags badf)
949 if (td->o.use_os_rand) {
950 r = os_random_long(&td->next_file_state);
951 fno = (unsigned int) ((double) td->o.nr_files
952 * (r / (OS_RAND_MAX + 1.0)));
954 r = __rand(&td->__next_file_state);
955 fno = (unsigned int) ((double) td->o.nr_files
956 * (r / (FRAND_MAX + 1.0)));
960 if (fio_file_done(f))
963 if (!fio_file_open(f)) {
966 err = td_io_open_file(td, f);
972 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
973 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
977 td_io_close_file(td, f);
982 * Get next file to service by doing round robin between all available ones
984 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
987 unsigned int old_next_file = td->next_file;
993 f = td->files[td->next_file];
996 if (td->next_file >= td->o.nr_files)
999 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1000 if (fio_file_done(f)) {
1005 if (!fio_file_open(f)) {
1008 err = td_io_open_file(td, f);
1010 dprint(FD_FILE, "error %d on open of %s\n",
1018 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1020 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1024 td_io_close_file(td, f);
1027 } while (td->next_file != old_next_file);
1029 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1033 static struct fio_file *__get_next_file(struct thread_data *td)
1037 assert(td->o.nr_files <= td->files_index);
1039 if (td->nr_done_files >= td->o.nr_files) {
1040 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1041 " nr_files=%d\n", td->nr_open_files,
1047 f = td->file_service_file;
1048 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1049 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1051 if (td->file_service_left--)
1055 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1056 td->o.file_service_type == FIO_FSERVICE_SEQ)
1057 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1059 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1061 td->file_service_file = f;
1062 td->file_service_left = td->file_service_nr - 1;
1064 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1068 static struct fio_file *get_next_file(struct thread_data *td)
1070 if (!(td->flags & TD_F_PROFILE_OPS)) {
1071 struct prof_io_ops *ops = &td->prof_io_ops;
1073 if (ops->get_next_file)
1074 return ops->get_next_file(td);
1077 return __get_next_file(td);
1080 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1085 f = get_next_file(td);
1092 if (!fill_io_u(td, io_u))
1095 put_file_log(td, f);
1096 td_io_close_file(td, f);
1098 fio_file_set_done(f);
1099 td->nr_done_files++;
1100 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1101 td->nr_done_files, td->o.nr_files);
1108 struct io_u *__get_io_u(struct thread_data *td)
1110 struct io_u *io_u = NULL;
1115 if (!flist_empty(&td->io_u_requeues))
1116 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1117 else if (!queue_full(td)) {
1118 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1123 io_u->end_io = NULL;
1127 assert(io_u->flags & IO_U_F_FREE);
1128 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1129 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1130 io_u->flags &= ~IO_U_F_VER_LIST;
1133 io_u->acct_ddir = -1;
1134 flist_del(&io_u->list);
1135 flist_add_tail(&io_u->list, &td->io_u_busylist);
1137 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1138 } else if (td->o.verify_async) {
1140 * We ran out, wait for async verify threads to finish and
1143 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1151 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1153 if (!(td->flags & TD_F_TRIM_BACKLOG))
1156 if (td->trim_entries) {
1159 if (td->trim_batch) {
1162 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1163 td->last_ddir != DDIR_READ) {
1164 td->trim_batch = td->o.trim_batch;
1165 if (!td->trim_batch)
1166 td->trim_batch = td->o.trim_backlog;
1170 if (get_trim && !get_next_trim(td, io_u))
1177 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1179 if (!(td->flags & TD_F_VER_BACKLOG))
1182 if (td->io_hist_len) {
1185 if (td->verify_batch)
1187 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1188 td->last_ddir != DDIR_READ) {
1189 td->verify_batch = td->o.verify_batch;
1190 if (!td->verify_batch)
1191 td->verify_batch = td->o.verify_backlog;
1195 if (get_verify && !get_next_verify(td, io_u)) {
1205 * Fill offset and start time into the buffer content, to prevent too
1206 * easy compressible data for simple de-dupe attempts. Do this for every
1207 * 512b block in the range, since that should be the smallest block size
1208 * we can expect from a device.
1210 static void small_content_scramble(struct io_u *io_u)
1212 unsigned int i, nr_blocks = io_u->buflen / 512;
1214 unsigned int offset;
1221 boffset = io_u->offset;
1222 io_u->buf_filled_len = 0;
1224 for (i = 0; i < nr_blocks; i++) {
1226 * Fill the byte offset into a "random" start offset of
1227 * the buffer, given by the product of the usec time
1228 * and the actual offset.
1230 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1231 offset &= ~(sizeof(uint64_t) - 1);
1232 if (offset >= 512 - sizeof(uint64_t))
1233 offset -= sizeof(uint64_t);
1234 memcpy(p + offset, &boffset, sizeof(boffset));
1236 end = p + 512 - sizeof(io_u->start_time);
1237 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1244 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1245 * etc. The returned io_u is fully ready to be prepped and submitted.
1247 struct io_u *get_io_u(struct thread_data *td)
1251 int do_scramble = 0;
1253 io_u = __get_io_u(td);
1255 dprint(FD_IO, "__get_io_u failed\n");
1259 if (check_get_verify(td, io_u))
1261 if (check_get_trim(td, io_u))
1265 * from a requeue, io_u already setup
1271 * If using an iolog, grab next piece if any available.
1273 if (td->flags & TD_F_READ_IOLOG) {
1274 if (read_iolog_get(td, io_u))
1276 } else if (set_io_u_file(td, io_u)) {
1277 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1282 assert(fio_file_open(f));
1284 if (ddir_rw(io_u->ddir)) {
1285 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1286 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1290 f->last_start = io_u->offset;
1291 f->last_pos = io_u->offset + io_u->buflen;
1293 if (io_u->ddir == DDIR_WRITE) {
1294 if (td->flags & TD_F_REFILL_BUFFERS) {
1295 io_u_fill_buffer(td, io_u,
1296 io_u->xfer_buflen, io_u->xfer_buflen);
1297 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1299 if (td->flags & TD_F_VER_NONE) {
1300 populate_verify_io_u(td, io_u);
1303 } else if (io_u->ddir == DDIR_READ) {
1305 * Reset the buf_filled parameters so next time if the
1306 * buffer is used for writes it is refilled.
1308 io_u->buf_filled_len = 0;
1313 * Set io data pointers.
1315 io_u->xfer_buf = io_u->buf;
1316 io_u->xfer_buflen = io_u->buflen;
1320 if (!td_io_prep(td, io_u)) {
1321 if (!td->o.disable_slat)
1322 fio_gettime(&io_u->start_time, NULL);
1324 small_content_scramble(io_u);
1328 dprint(FD_IO, "get_io_u failed\n");
1333 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1335 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1336 const char *msg[] = { "read", "write", "sync", "datasync",
1337 "sync_file_range", "wait", "trim" };
1339 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1342 log_err("fio: io_u error");
1345 log_err(" on file %s", io_u->file->file_name);
1347 log_err(": %s\n", strerror(io_u->error));
1349 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1350 io_u->offset, io_u->xfer_buflen);
1353 td_verror(td, io_u->error, "io_u error");
1356 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1357 struct io_completion_data *icd,
1358 const enum fio_ddir idx, unsigned int bytes)
1360 unsigned long lusec = 0;
1362 if (!td->o.disable_clat || !td->o.disable_bw)
1363 lusec = utime_since(&io_u->issue_time, &icd->time);
1365 if (!td->o.disable_lat) {
1366 unsigned long tusec;
1368 tusec = utime_since(&io_u->start_time, &icd->time);
1369 add_lat_sample(td, idx, tusec, bytes);
1371 if (td->flags & TD_F_PROFILE_OPS) {
1372 struct prof_io_ops *ops = &td->prof_io_ops;
1375 icd->error = ops->io_u_lat(td, tusec);
1378 if (td->o.max_latency && tusec > td->o.max_latency) {
1380 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1381 td_verror(td, ETIMEDOUT, "max latency exceeded");
1382 icd->error = ETIMEDOUT;
1386 if (!td->o.disable_clat) {
1387 add_clat_sample(td, idx, lusec, bytes);
1388 io_u_mark_latency(td, lusec);
1391 if (!td->o.disable_bw)
1392 add_bw_sample(td, idx, bytes, &icd->time);
1394 add_iops_sample(td, idx, &icd->time);
1397 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1399 uint64_t secs, remainder, bps, bytes;
1401 bytes = td->this_io_bytes[ddir];
1402 bps = td->rate_bps[ddir];
1404 remainder = bytes % bps;
1405 return remainder * 1000000 / bps + secs * 1000000;
1408 static void io_completed(struct thread_data *td, struct io_u *io_u,
1409 struct io_completion_data *icd)
1413 dprint_io_u(io_u, "io complete");
1416 assert(io_u->flags & IO_U_F_FLIGHT);
1417 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1420 if (ddir_sync(io_u->ddir)) {
1421 td->last_was_sync = 1;
1424 f->first_write = -1ULL;
1425 f->last_write = -1ULL;
1430 td->last_was_sync = 0;
1431 td->last_ddir = io_u->ddir;
1433 if (!io_u->error && ddir_rw(io_u->ddir)) {
1434 unsigned int bytes = io_u->buflen - io_u->resid;
1435 const enum fio_ddir idx = io_u->ddir;
1436 const enum fio_ddir odx = io_u->ddir ^ 1;
1439 td->io_blocks[idx]++;
1440 td->this_io_blocks[idx]++;
1441 td->io_bytes[idx] += bytes;
1443 if (!(io_u->flags & IO_U_F_VER_LIST))
1444 td->this_io_bytes[idx] += bytes;
1446 if (idx == DDIR_WRITE) {
1449 if (f->first_write == -1ULL ||
1450 io_u->offset < f->first_write)
1451 f->first_write = io_u->offset;
1452 if (f->last_write == -1ULL ||
1453 ((io_u->offset + bytes) > f->last_write))
1454 f->last_write = io_u->offset + bytes;
1458 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1459 td->runstate == TD_VERIFYING)) {
1460 account_io_completion(td, io_u, icd, idx, bytes);
1462 if (__should_check_rate(td, idx)) {
1463 td->rate_pending_usleep[idx] =
1464 (usec_for_io(td, idx) -
1465 utime_since_now(&td->start));
1467 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1468 td->rate_pending_usleep[odx] =
1469 (usec_for_io(td, odx) -
1470 utime_since_now(&td->start));
1473 if (td_write(td) && idx == DDIR_WRITE &&
1475 td->o.verify != VERIFY_NONE &&
1476 !td->o.experimental_verify)
1477 log_io_piece(td, io_u);
1479 icd->bytes_done[idx] += bytes;
1482 ret = io_u->end_io(td, io_u);
1483 if (ret && !icd->error)
1486 } else if (io_u->error) {
1487 icd->error = io_u->error;
1488 io_u_log_error(td, io_u);
1491 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1492 if (!td_non_fatal_error(td, eb, icd->error))
1495 * If there is a non_fatal error, then add to the error count
1496 * and clear all the errors.
1498 update_error_count(td, icd->error);
1505 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1509 if (!td->o.disable_clat || !td->o.disable_bw)
1510 fio_gettime(&icd->time, NULL);
1515 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1516 icd->bytes_done[ddir] = 0;
1519 static void ios_completed(struct thread_data *td,
1520 struct io_completion_data *icd)
1525 for (i = 0; i < icd->nr; i++) {
1526 io_u = td->io_ops->event(td, i);
1528 io_completed(td, io_u, icd);
1530 if (!(io_u->flags & IO_U_F_FREE_DEF))
1536 * Complete a single io_u for the sync engines.
1538 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1541 struct io_completion_data icd;
1543 init_icd(td, &icd, 1);
1544 io_completed(td, io_u, &icd);
1546 if (!(io_u->flags & IO_U_F_FREE_DEF))
1550 td_verror(td, icd.error, "io_u_sync_complete");
1557 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1558 bytes[ddir] += icd.bytes_done[ddir];
1565 * Called to complete min_events number of io for the async engines.
1567 int io_u_queued_complete(struct thread_data *td, int min_evts,
1570 struct io_completion_data icd;
1571 struct timespec *tvp = NULL;
1573 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1575 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1580 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1582 td_verror(td, -ret, "td_io_getevents");
1587 init_icd(td, &icd, ret);
1588 ios_completed(td, &icd);
1590 td_verror(td, icd.error, "io_u_queued_complete");
1597 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1598 bytes[ddir] += icd.bytes_done[ddir];
1605 * Call when io_u is really queued, to update the submission latency.
1607 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1609 if (!td->o.disable_slat) {
1610 unsigned long slat_time;
1612 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1613 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1617 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1618 unsigned int max_bs)
1620 if (!td->o.zero_buffers) {
1621 unsigned int perc = td->o.compress_percentage;
1624 unsigned int seg = min_write;
1626 seg = min(min_write, td->o.compress_chunk);
1630 fill_random_buf_percentage(&td->buf_state, buf,
1633 fill_random_buf(&td->buf_state, buf, max_bs);
1635 memset(buf, 0, max_bs);
1639 * "randomly" fill the buffer contents
1641 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1642 unsigned int min_write, unsigned int max_bs)
1644 io_u->buf_filled_len = 0;
1645 fill_io_buffer(td, io_u->buf, min_write, max_bs);