13 #include "lib/axmap.h"
16 struct io_completion_data {
19 int error; /* output */
20 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
21 struct timeval time; /* output */
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
28 static int random_map_free(struct fio_file *f, const uint64_t block)
30 return !axmap_isset(f->io_axmap, block);
34 * Mark a given offset as used in the map.
36 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned int nr_blocks;
43 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
46 if (!(io_u->flags & IO_U_F_BUSY_OK))
47 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 if ((nr_blocks * min_bs) < io_u->buflen)
50 io_u->buflen = nr_blocks * min_bs;
53 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 assert(ddir_rw(ddir));
62 * Hmm, should we make sure that ->io_size <= ->real_file_size?
64 max_size = f->io_size;
65 if (max_size > f->real_file_size)
66 max_size = f->real_file_size;
69 max_size = td->o.zone_range;
71 if (td->o.min_bs[ddir] > td->o.ba[ddir])
72 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
74 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
82 struct flist_head list;
86 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir, uint64_t *b)
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
94 lastb = last_block(td, f, ddir);
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
106 assert(fio_file_lfsr(f));
108 if (lfsr_next(&f->lfsr, &off))
115 * if we are not maintaining a random map, we are done.
117 if (!file_randommap(td, f))
121 * calculate map offset and check if it's free
123 if (random_map_free(f, *b))
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
136 static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
140 *b = zipf_next(&f->zipf);
144 static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
148 *b = pareto_next(&f->zipf);
152 static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
156 *b = gauss_next(&f->gauss);
161 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
163 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
164 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
166 return r1->off - r2->off;
169 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
170 enum fio_ddir ddir, uint64_t *b)
172 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
173 return __get_next_rand_offset(td, f, ddir, b);
174 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
175 return __get_next_rand_offset_zipf(td, f, ddir, b);
176 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
177 return __get_next_rand_offset_pareto(td, f, ddir, b);
178 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
179 return __get_next_rand_offset_gauss(td, f, ddir, b);
181 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
186 * Sort the reads for a verify phase in batches of verifysort_nr, if
189 static inline int should_sort_io(struct thread_data *td)
191 if (!td->o.verifysort_nr || !td->o.do_verify)
195 if (td->runstate != TD_VERIFYING)
197 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
203 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
208 if (td->o.perc_rand[ddir] == 100)
211 r = __rand(&td->seq_rand_state[ddir]);
212 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
214 return v <= td->o.perc_rand[ddir];
217 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
218 enum fio_ddir ddir, uint64_t *b)
223 if (!should_sort_io(td))
224 return get_off_from_method(td, f, ddir, b);
226 if (!flist_empty(&td->next_rand_list)) {
228 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
235 for (i = 0; i < td->o.verifysort_nr; i++) {
236 r = malloc(sizeof(*r));
238 ret = get_off_from_method(td, f, ddir, &r->off);
244 flist_add(&r->list, &td->next_rand_list);
250 assert(!flist_empty(&td->next_rand_list));
251 flist_sort(NULL, &td->next_rand_list, flist_cmp);
255 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
256 enum fio_ddir ddir, uint64_t *b)
258 if (!get_next_rand_offset(td, f, ddir, b))
261 if (td->o.time_based) {
262 fio_file_reset(td, f);
263 if (!get_next_rand_offset(td, f, ddir, b))
267 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
268 f->file_name, (unsigned long long) f->last_pos[ddir],
269 (unsigned long long) f->real_file_size);
273 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
274 enum fio_ddir ddir, uint64_t *offset)
276 struct thread_options *o = &td->o;
278 assert(ddir_rw(ddir));
280 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
282 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
284 if (f->last_pos[ddir] < f->real_file_size) {
287 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
288 f->last_pos[ddir] = f->real_file_size;
290 pos = f->last_pos[ddir] - f->file_offset;
291 if (pos && o->ddir_seq_add) {
292 pos += o->ddir_seq_add;
295 * If we reach beyond the end of the file
296 * with holed IO, wrap around to the
299 if (pos >= f->real_file_size)
300 pos = f->file_offset;
310 static int get_next_block(struct thread_data *td, struct io_u *io_u,
311 enum fio_ddir ddir, int rw_seq,
312 unsigned int *is_random)
314 struct fio_file *f = io_u->file;
318 assert(ddir_rw(ddir));
324 if (should_do_random(td, ddir)) {
325 ret = get_next_rand_block(td, f, ddir, &b);
329 io_u_set(io_u, IO_U_F_BUSY_OK);
330 ret = get_next_seq_offset(td, f, ddir, &offset);
332 ret = get_next_rand_block(td, f, ddir, &b);
336 ret = get_next_seq_offset(td, f, ddir, &offset);
339 io_u_set(io_u, IO_U_F_BUSY_OK);
342 if (td->o.rw_seq == RW_SEQ_SEQ) {
343 ret = get_next_seq_offset(td, f, ddir, &offset);
345 ret = get_next_rand_block(td, f, ddir, &b);
348 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
349 if (f->last_start[ddir] != -1ULL)
350 offset = f->last_start[ddir] - f->file_offset;
355 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
362 io_u->offset = offset;
364 io_u->offset = b * td->o.ba[ddir];
366 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
375 * For random io, generate a random new block and see if it's used. Repeat
376 * until we find a free one. For sequential io, just return the end of
377 * the last io issued.
379 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
380 unsigned int *is_random)
382 struct fio_file *f = io_u->file;
383 enum fio_ddir ddir = io_u->ddir;
386 assert(ddir_rw(ddir));
388 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
390 td->ddir_seq_nr = td->o.ddir_seq_nr;
393 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
396 if (io_u->offset >= f->io_size) {
397 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
398 (unsigned long long) io_u->offset,
399 (unsigned long long) f->io_size);
403 io_u->offset += f->file_offset;
404 if (io_u->offset >= f->real_file_size) {
405 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
406 (unsigned long long) io_u->offset,
407 (unsigned long long) f->real_file_size);
414 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
415 unsigned int *is_random)
417 if (td->flags & TD_F_PROFILE_OPS) {
418 struct prof_io_ops *ops = &td->prof_io_ops;
420 if (ops->fill_io_u_off)
421 return ops->fill_io_u_off(td, io_u, is_random);
424 return __get_next_offset(td, io_u, is_random);
427 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
430 struct fio_file *f = io_u->file;
432 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
435 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
436 unsigned int is_random)
438 int ddir = io_u->ddir;
439 unsigned int buflen = 0;
440 unsigned int minbs, maxbs;
443 assert(ddir_rw(ddir));
445 if (td->o.bs_is_seq_rand)
446 ddir = is_random ? DDIR_WRITE: DDIR_READ;
448 minbs = td->o.min_bs[ddir];
449 maxbs = td->o.max_bs[ddir];
455 * If we can't satisfy the min block size from here, then fail
457 if (!io_u_fits(td, io_u, minbs))
461 r = __rand(&td->bsrange_state);
463 if (!td->o.bssplit_nr[ddir]) {
464 buflen = 1 + (unsigned int) ((double) maxbs *
465 (r / (FRAND_MAX + 1.0)));
472 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
473 struct bssplit *bsp = &td->o.bssplit[ddir][i];
477 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
478 io_u_fits(td, io_u, buflen))
483 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
484 buflen = (buflen + td->o.verify_interval - 1) &
485 ~(td->o.verify_interval - 1);
487 if (!td->o.bs_unaligned && is_power_of_2(minbs))
488 buflen &= ~(minbs - 1);
490 } while (!io_u_fits(td, io_u, buflen));
495 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
496 unsigned int is_random)
498 if (td->flags & TD_F_PROFILE_OPS) {
499 struct prof_io_ops *ops = &td->prof_io_ops;
501 if (ops->fill_io_u_size)
502 return ops->fill_io_u_size(td, io_u, is_random);
505 return __get_next_buflen(td, io_u, is_random);
508 static void set_rwmix_bytes(struct thread_data *td)
513 * we do time or byte based switch. this is needed because
514 * buffered writes may issue a lot quicker than they complete,
515 * whereas reads do not.
517 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
518 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
521 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
526 r = __rand(&td->rwmix_state);
527 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
529 if (v <= td->o.rwmix[DDIR_READ])
535 void io_u_quiesce(struct thread_data *td)
538 * We are going to sleep, ensure that we flush anything pending as
539 * not to skew our latency numbers.
541 * Changed to only monitor 'in flight' requests here instead of the
542 * td->cur_depth, b/c td->cur_depth does not accurately represent
543 * io's that have been actually submitted to an async engine,
544 * and cur_depth is meaningless for sync engines.
546 if (td->io_u_queued || td->cur_depth) {
549 ret = td_io_commit(td);
552 while (td->io_u_in_flight) {
555 ret = io_u_queued_complete(td, 1);
559 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
561 enum fio_ddir odir = ddir ^ 1;
564 assert(ddir_rw(ddir));
566 if (td->rate_pending_usleep[ddir] <= 0)
570 * We have too much pending sleep in this direction. See if we
573 if (td_rw(td) && td->o.rwmix[odir]) {
575 * Other direction does not have too much pending, switch
577 if (td->rate_pending_usleep[odir] < 100000)
581 * Both directions have pending sleep. Sleep the minimum time
582 * and deduct from both.
584 if (td->rate_pending_usleep[ddir] <=
585 td->rate_pending_usleep[odir]) {
586 usec = td->rate_pending_usleep[ddir];
588 usec = td->rate_pending_usleep[odir];
592 usec = td->rate_pending_usleep[ddir];
594 if (td->o.io_submit_mode == IO_MODE_INLINE)
597 usec = usec_sleep(td, usec);
599 td->rate_pending_usleep[ddir] -= usec;
602 if (td_rw(td) && __should_check_rate(td, odir))
603 td->rate_pending_usleep[odir] -= usec;
609 * Return the data direction for the next io_u. If the job is a
610 * mixed read/write workload, check the rwmix cycle and switch if
613 static enum fio_ddir get_rw_ddir(struct thread_data *td)
618 * see if it's time to fsync
620 if (td->o.fsync_blocks &&
621 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
622 td->io_issues[DDIR_WRITE] && should_fsync(td))
626 * see if it's time to fdatasync
628 if (td->o.fdatasync_blocks &&
629 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
630 td->io_issues[DDIR_WRITE] && should_fsync(td))
631 return DDIR_DATASYNC;
634 * see if it's time to sync_file_range
636 if (td->sync_file_range_nr &&
637 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
638 td->io_issues[DDIR_WRITE] && should_fsync(td))
639 return DDIR_SYNC_FILE_RANGE;
643 * Check if it's time to seed a new data direction.
645 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
647 * Put a top limit on how many bytes we do for
648 * one data direction, to avoid overflowing the
651 ddir = get_rand_ddir(td);
653 if (ddir != td->rwmix_ddir)
656 td->rwmix_ddir = ddir;
658 ddir = td->rwmix_ddir;
659 } else if (td_read(td))
661 else if (td_write(td))
666 td->rwmix_ddir = rate_ddir(td, ddir);
667 return td->rwmix_ddir;
670 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
672 enum fio_ddir ddir = get_rw_ddir(td);
674 if (td_trimwrite(td)) {
675 struct fio_file *f = io_u->file;
676 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
682 io_u->ddir = io_u->acct_ddir = ddir;
684 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
685 td->o.barrier_blocks &&
686 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
687 td->io_issues[DDIR_WRITE])
688 io_u_set(io_u, IO_U_F_BARRIER);
691 void put_file_log(struct thread_data *td, struct fio_file *f)
693 unsigned int ret = put_file(td, f);
696 td_verror(td, ret, "file close");
699 void put_io_u(struct thread_data *td, struct io_u *io_u)
706 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
707 put_file_log(td, io_u->file);
710 io_u_set(io_u, IO_U_F_FREE);
712 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
714 assert(!(td->flags & TD_F_CHILD));
716 io_u_qpush(&td->io_u_freelist, io_u);
718 td_io_u_free_notify(td);
721 void clear_io_u(struct thread_data *td, struct io_u *io_u)
723 io_u_clear(io_u, IO_U_F_FLIGHT);
727 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
729 struct io_u *__io_u = *io_u;
730 enum fio_ddir ddir = acct_ddir(__io_u);
732 dprint(FD_IO, "requeue %p\n", __io_u);
739 io_u_set(__io_u, IO_U_F_FREE);
740 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
741 td->io_issues[ddir]--;
743 io_u_clear(__io_u, IO_U_F_FLIGHT);
744 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
746 assert(!(td->flags & TD_F_CHILD));
749 io_u_rpush(&td->io_u_requeues, __io_u);
751 td_io_u_free_notify(td);
755 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
757 unsigned int is_random;
759 if (td->io_ops->flags & FIO_NOIO)
762 set_rw_ddir(td, io_u);
765 * fsync() or fdatasync() or trim etc, we are done
767 if (!ddir_rw(io_u->ddir))
771 * See if it's time to switch to a new zone
773 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
774 struct fio_file *f = io_u->file;
777 f->file_offset += td->o.zone_range + td->o.zone_skip;
780 * Wrap from the beginning, if we exceed the file size
782 if (f->file_offset >= f->real_file_size)
783 f->file_offset = f->real_file_size - f->file_offset;
784 f->last_pos[io_u->ddir] = f->file_offset;
785 td->io_skip_bytes += td->o.zone_skip;
789 * No log, let the seq/rand engine retrieve the next buflen and
792 if (get_next_offset(td, io_u, &is_random)) {
793 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
797 io_u->buflen = get_next_buflen(td, io_u, is_random);
799 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
803 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
804 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
805 dprint(FD_IO, " off=%llu/%lu > %llu\n",
806 (unsigned long long) io_u->offset, io_u->buflen,
807 (unsigned long long) io_u->file->real_file_size);
812 * mark entry before potentially trimming io_u
814 if (td_random(td) && file_randommap(td, io_u->file))
815 mark_random_map(td, io_u);
818 dprint_io_u(io_u, "fill_io_u");
819 td->zone_bytes += io_u->buflen;
823 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
852 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
854 __io_u_mark_map(td->ts.io_u_submit, nr);
855 td->ts.total_submit++;
858 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
860 __io_u_mark_map(td->ts.io_u_complete, nr);
861 td->ts.total_complete++;
864 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
868 switch (td->cur_depth) {
890 td->ts.io_u_map[idx] += nr;
893 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
930 assert(idx < FIO_IO_U_LAT_U_NR);
931 td->ts.io_u_lat_u[idx]++;
934 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
975 assert(idx < FIO_IO_U_LAT_M_NR);
976 td->ts.io_u_lat_m[idx]++;
979 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
982 io_u_mark_lat_usec(td, usec);
984 io_u_mark_lat_msec(td, usec / 1000);
988 * Get next file to service by choosing one at random
990 static struct fio_file *get_next_file_rand(struct thread_data *td,
991 enum fio_file_flags goodf,
992 enum fio_file_flags badf)
1001 r = __rand(&td->next_file_state);
1002 fno = (unsigned int) ((double) td->o.nr_files
1003 * (r / (FRAND_MAX + 1.0)));
1006 if (fio_file_done(f))
1009 if (!fio_file_open(f)) {
1012 if (td->nr_open_files >= td->o.open_files)
1013 return ERR_PTR(-EBUSY);
1015 err = td_io_open_file(td, f);
1021 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1022 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1026 td_io_close_file(td, f);
1031 * Get next file to service by doing round robin between all available ones
1033 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1036 unsigned int old_next_file = td->next_file;
1042 f = td->files[td->next_file];
1045 if (td->next_file >= td->o.nr_files)
1048 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1049 if (fio_file_done(f)) {
1054 if (!fio_file_open(f)) {
1057 if (td->nr_open_files >= td->o.open_files)
1058 return ERR_PTR(-EBUSY);
1060 err = td_io_open_file(td, f);
1062 dprint(FD_FILE, "error %d on open of %s\n",
1070 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1072 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1076 td_io_close_file(td, f);
1079 } while (td->next_file != old_next_file);
1081 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1085 static struct fio_file *__get_next_file(struct thread_data *td)
1089 assert(td->o.nr_files <= td->files_index);
1091 if (td->nr_done_files >= td->o.nr_files) {
1092 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1093 " nr_files=%d\n", td->nr_open_files,
1099 f = td->file_service_file;
1100 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1101 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1103 if (td->file_service_left--)
1107 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1108 td->o.file_service_type == FIO_FSERVICE_SEQ)
1109 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1111 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1116 td->file_service_file = f;
1117 td->file_service_left = td->file_service_nr - 1;
1120 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1122 dprint(FD_FILE, "get_next_file: NULL\n");
1126 static struct fio_file *get_next_file(struct thread_data *td)
1128 if (td->flags & TD_F_PROFILE_OPS) {
1129 struct prof_io_ops *ops = &td->prof_io_ops;
1131 if (ops->get_next_file)
1132 return ops->get_next_file(td);
1135 return __get_next_file(td);
1138 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1143 f = get_next_file(td);
1144 if (IS_ERR_OR_NULL(f))
1150 if (!fill_io_u(td, io_u))
1153 put_file_log(td, f);
1154 td_io_close_file(td, f);
1156 fio_file_set_done(f);
1157 td->nr_done_files++;
1158 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1159 td->nr_done_files, td->o.nr_files);
1165 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1166 unsigned long tusec, unsigned long max_usec)
1169 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1170 td_verror(td, ETIMEDOUT, "max latency exceeded");
1171 icd->error = ETIMEDOUT;
1174 static void lat_new_cycle(struct thread_data *td)
1176 fio_gettime(&td->latency_ts, NULL);
1177 td->latency_ios = ddir_rw_sum(td->io_blocks);
1178 td->latency_failed = 0;
1182 * We had an IO outside the latency target. Reduce the queue depth. If we
1183 * are at QD=1, then it's time to give up.
1185 static int __lat_target_failed(struct thread_data *td)
1187 if (td->latency_qd == 1)
1190 td->latency_qd_high = td->latency_qd;
1192 if (td->latency_qd == td->latency_qd_low)
1193 td->latency_qd_low--;
1195 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1197 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1200 * When we ramp QD down, quiesce existing IO to prevent
1201 * a storm of ramp downs due to pending higher depth.
1208 static int lat_target_failed(struct thread_data *td)
1210 if (td->o.latency_percentile.u.f == 100.0)
1211 return __lat_target_failed(td);
1213 td->latency_failed++;
1217 void lat_target_init(struct thread_data *td)
1219 td->latency_end_run = 0;
1221 if (td->o.latency_target) {
1222 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1223 fio_gettime(&td->latency_ts, NULL);
1225 td->latency_qd_high = td->o.iodepth;
1226 td->latency_qd_low = 1;
1227 td->latency_ios = ddir_rw_sum(td->io_blocks);
1229 td->latency_qd = td->o.iodepth;
1232 void lat_target_reset(struct thread_data *td)
1234 if (!td->latency_end_run)
1235 lat_target_init(td);
1238 static void lat_target_success(struct thread_data *td)
1240 const unsigned int qd = td->latency_qd;
1241 struct thread_options *o = &td->o;
1243 td->latency_qd_low = td->latency_qd;
1246 * If we haven't failed yet, we double up to a failing value instead
1247 * of bisecting from highest possible queue depth. If we have set
1248 * a limit other than td->o.iodepth, bisect between that.
1250 if (td->latency_qd_high != o->iodepth)
1251 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1253 td->latency_qd *= 2;
1255 if (td->latency_qd > o->iodepth)
1256 td->latency_qd = o->iodepth;
1258 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1261 * Same as last one, we are done. Let it run a latency cycle, so
1262 * we get only the results from the targeted depth.
1264 if (td->latency_qd == qd) {
1265 if (td->latency_end_run) {
1266 dprint(FD_RATE, "We are done\n");
1269 dprint(FD_RATE, "Quiesce and final run\n");
1271 td->latency_end_run = 1;
1272 reset_all_stats(td);
1281 * Check if we can bump the queue depth
1283 void lat_target_check(struct thread_data *td)
1285 uint64_t usec_window;
1289 usec_window = utime_since_now(&td->latency_ts);
1290 if (usec_window < td->o.latency_window)
1293 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1294 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1295 success_ios *= 100.0;
1297 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1299 if (success_ios >= td->o.latency_percentile.u.f)
1300 lat_target_success(td);
1302 __lat_target_failed(td);
1306 * If latency target is enabled, we might be ramping up or down and not
1307 * using the full queue depth available.
1309 int queue_full(const struct thread_data *td)
1311 const int qempty = io_u_qempty(&td->io_u_freelist);
1315 if (!td->o.latency_target)
1318 return td->cur_depth >= td->latency_qd;
1321 struct io_u *__get_io_u(struct thread_data *td)
1323 struct io_u *io_u = NULL;
1331 if (!io_u_rempty(&td->io_u_requeues))
1332 io_u = io_u_rpop(&td->io_u_requeues);
1333 else if (!queue_full(td)) {
1334 io_u = io_u_qpop(&td->io_u_freelist);
1339 io_u->end_io = NULL;
1343 assert(io_u->flags & IO_U_F_FREE);
1344 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1345 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1349 io_u->acct_ddir = -1;
1351 assert(!(td->flags & TD_F_CHILD));
1352 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1354 } else if (td_async_processing(td)) {
1356 * We ran out, wait for async verify threads to finish and
1359 assert(!(td->flags & TD_F_CHILD));
1360 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1368 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1370 if (!(td->flags & TD_F_TRIM_BACKLOG))
1373 if (td->trim_entries) {
1376 if (td->trim_batch) {
1379 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1380 td->last_ddir != DDIR_READ) {
1381 td->trim_batch = td->o.trim_batch;
1382 if (!td->trim_batch)
1383 td->trim_batch = td->o.trim_backlog;
1387 if (get_trim && !get_next_trim(td, io_u))
1394 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1396 if (!(td->flags & TD_F_VER_BACKLOG))
1399 if (td->io_hist_len) {
1402 if (td->verify_batch)
1404 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1405 td->last_ddir != DDIR_READ) {
1406 td->verify_batch = td->o.verify_batch;
1407 if (!td->verify_batch)
1408 td->verify_batch = td->o.verify_backlog;
1412 if (get_verify && !get_next_verify(td, io_u)) {
1422 * Fill offset and start time into the buffer content, to prevent too
1423 * easy compressible data for simple de-dupe attempts. Do this for every
1424 * 512b block in the range, since that should be the smallest block size
1425 * we can expect from a device.
1427 static void small_content_scramble(struct io_u *io_u)
1429 unsigned int i, nr_blocks = io_u->buflen / 512;
1431 unsigned int offset;
1438 boffset = io_u->offset;
1439 io_u->buf_filled_len = 0;
1441 for (i = 0; i < nr_blocks; i++) {
1443 * Fill the byte offset into a "random" start offset of
1444 * the buffer, given by the product of the usec time
1445 * and the actual offset.
1447 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1448 offset &= ~(sizeof(uint64_t) - 1);
1449 if (offset >= 512 - sizeof(uint64_t))
1450 offset -= sizeof(uint64_t);
1451 memcpy(p + offset, &boffset, sizeof(boffset));
1453 end = p + 512 - sizeof(io_u->start_time);
1454 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1461 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1462 * etc. The returned io_u is fully ready to be prepped and submitted.
1464 struct io_u *get_io_u(struct thread_data *td)
1468 int do_scramble = 0;
1471 io_u = __get_io_u(td);
1473 dprint(FD_IO, "__get_io_u failed\n");
1477 if (check_get_verify(td, io_u))
1479 if (check_get_trim(td, io_u))
1483 * from a requeue, io_u already setup
1489 * If using an iolog, grab next piece if any available.
1491 if (td->flags & TD_F_READ_IOLOG) {
1492 if (read_iolog_get(td, io_u))
1494 } else if (set_io_u_file(td, io_u)) {
1496 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1502 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1506 assert(fio_file_open(f));
1508 if (ddir_rw(io_u->ddir)) {
1509 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1510 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1514 f->last_start[io_u->ddir] = io_u->offset;
1515 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1517 if (io_u->ddir == DDIR_WRITE) {
1518 if (td->flags & TD_F_REFILL_BUFFERS) {
1519 io_u_fill_buffer(td, io_u,
1520 td->o.min_bs[DDIR_WRITE],
1522 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1523 !(td->flags & TD_F_COMPRESS))
1525 if (td->flags & TD_F_VER_NONE) {
1526 populate_verify_io_u(td, io_u);
1529 } else if (io_u->ddir == DDIR_READ) {
1531 * Reset the buf_filled parameters so next time if the
1532 * buffer is used for writes it is refilled.
1534 io_u->buf_filled_len = 0;
1539 * Set io data pointers.
1541 io_u->xfer_buf = io_u->buf;
1542 io_u->xfer_buflen = io_u->buflen;
1546 if (!td_io_prep(td, io_u)) {
1547 if (!td->o.disable_slat)
1548 fio_gettime(&io_u->start_time, NULL);
1550 small_content_scramble(io_u);
1554 dprint(FD_IO, "get_io_u failed\n");
1556 return ERR_PTR(ret);
1559 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1561 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1563 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1566 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1567 io_u->file ? " on file " : "",
1568 io_u->file ? io_u->file->file_name : "",
1569 strerror(io_u->error),
1570 io_ddir_name(io_u->ddir),
1571 io_u->offset, io_u->xfer_buflen);
1574 td_verror(td, io_u->error, "io_u error");
1577 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1579 __io_u_log_error(td, io_u);
1581 __io_u_log_error(td, io_u);
1584 static inline int gtod_reduce(struct thread_data *td)
1586 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1587 && td->o.disable_bw;
1590 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1591 struct io_completion_data *icd,
1592 const enum fio_ddir idx, unsigned int bytes)
1594 const int no_reduce = !gtod_reduce(td);
1595 unsigned long lusec = 0;
1598 lusec = utime_since(&io_u->issue_time, &icd->time);
1600 if (!td->o.disable_lat) {
1601 unsigned long tusec;
1603 tusec = utime_since(&io_u->start_time, &icd->time);
1604 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1606 if (td->flags & TD_F_PROFILE_OPS) {
1607 struct prof_io_ops *ops = &td->prof_io_ops;
1610 icd->error = ops->io_u_lat(td, tusec);
1613 if (td->o.max_latency && tusec > td->o.max_latency)
1614 lat_fatal(td, icd, tusec, td->o.max_latency);
1615 if (td->o.latency_target && tusec > td->o.latency_target) {
1616 if (lat_target_failed(td))
1617 lat_fatal(td, icd, tusec, td->o.latency_target);
1621 if (!td->o.disable_clat) {
1622 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1623 io_u_mark_latency(td, lusec);
1629 if (!td->o.disable_bw)
1630 add_bw_sample(td, idx, bytes, &icd->time);
1633 add_iops_sample(td, idx, bytes, &icd->time);
1635 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1636 uint32_t *info = io_u_block_info(td, io_u);
1637 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1638 if (io_u->ddir == DDIR_TRIM) {
1639 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1640 BLOCK_INFO_TRIMS(*info) + 1);
1641 } else if (io_u->ddir == DDIR_WRITE) {
1642 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1649 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1651 uint64_t secs, remainder, bps, bytes;
1653 assert(!(td->flags & TD_F_CHILD));
1654 bytes = td->this_io_bytes[ddir];
1655 bps = td->rate_bps[ddir];
1657 remainder = bytes % bps;
1658 return remainder * 1000000 / bps + secs * 1000000;
1661 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1662 struct io_completion_data *icd)
1664 struct io_u *io_u = *io_u_ptr;
1665 enum fio_ddir ddir = io_u->ddir;
1666 struct fio_file *f = io_u->file;
1668 dprint_io_u(io_u, "io complete");
1670 assert(io_u->flags & IO_U_F_FLIGHT);
1671 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1674 * Mark IO ok to verify
1678 * Remove errored entry from the verification list
1681 unlog_io_piece(td, io_u);
1683 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1688 if (ddir_sync(ddir)) {
1689 td->last_was_sync = 1;
1691 f->first_write = -1ULL;
1692 f->last_write = -1ULL;
1697 td->last_was_sync = 0;
1698 td->last_ddir = ddir;
1700 if (!io_u->error && ddir_rw(ddir)) {
1701 unsigned int bytes = io_u->buflen - io_u->resid;
1702 const enum fio_ddir oddir = ddir ^ 1;
1705 td->io_blocks[ddir]++;
1706 td->this_io_blocks[ddir]++;
1707 td->io_bytes[ddir] += bytes;
1709 if (!(io_u->flags & IO_U_F_VER_LIST))
1710 td->this_io_bytes[ddir] += bytes;
1712 if (ddir == DDIR_WRITE) {
1714 if (f->first_write == -1ULL ||
1715 io_u->offset < f->first_write)
1716 f->first_write = io_u->offset;
1717 if (f->last_write == -1ULL ||
1718 ((io_u->offset + bytes) > f->last_write))
1719 f->last_write = io_u->offset + bytes;
1721 if (td->last_write_comp) {
1722 int idx = td->last_write_idx++;
1724 td->last_write_comp[idx] = io_u->offset;
1725 if (td->last_write_idx == td->o.iodepth)
1726 td->last_write_idx = 0;
1730 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1731 td->runstate == TD_VERIFYING)) {
1732 struct thread_data *__td = td;
1734 account_io_completion(td, io_u, icd, ddir, bytes);
1739 if (__should_check_rate(__td, ddir)) {
1740 __td->rate_pending_usleep[ddir] =
1741 (usec_for_io(__td, ddir) -
1742 utime_since_now(&__td->start));
1744 if (ddir != DDIR_TRIM &&
1745 __should_check_rate(__td, oddir)) {
1746 __td->rate_pending_usleep[oddir] =
1747 (usec_for_io(__td, oddir) -
1748 utime_since_now(&__td->start));
1752 icd->bytes_done[ddir] += bytes;
1755 ret = io_u->end_io(td, io_u_ptr);
1757 if (ret && !icd->error)
1760 } else if (io_u->error) {
1761 icd->error = io_u->error;
1762 io_u_log_error(td, io_u);
1765 enum error_type_bit eb = td_error_type(ddir, icd->error);
1767 if (!td_non_fatal_error(td, eb, icd->error))
1771 * If there is a non_fatal error, then add to the error count
1772 * and clear all the errors.
1774 update_error_count(td, icd->error);
1782 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1787 if (!gtod_reduce(td))
1788 fio_gettime(&icd->time, NULL);
1793 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1794 icd->bytes_done[ddir] = 0;
1797 static void ios_completed(struct thread_data *td,
1798 struct io_completion_data *icd)
1803 for (i = 0; i < icd->nr; i++) {
1804 io_u = td->io_ops->event(td, i);
1806 io_completed(td, &io_u, icd);
1814 * Complete a single io_u for the sync engines.
1816 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1818 struct io_completion_data icd;
1821 init_icd(td, &icd, 1);
1822 io_completed(td, &io_u, &icd);
1828 td_verror(td, icd.error, "io_u_sync_complete");
1832 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1833 td->bytes_done[ddir] += icd.bytes_done[ddir];
1839 * Called to complete min_events number of io for the async engines.
1841 int io_u_queued_complete(struct thread_data *td, int min_evts)
1843 struct io_completion_data icd;
1844 struct timespec *tvp = NULL;
1846 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1848 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1852 else if (min_evts > td->cur_depth)
1853 min_evts = td->cur_depth;
1855 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1857 td_verror(td, -ret, "td_io_getevents");
1862 init_icd(td, &icd, ret);
1863 ios_completed(td, &icd);
1865 td_verror(td, icd.error, "io_u_queued_complete");
1869 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1870 td->bytes_done[ddir] += icd.bytes_done[ddir];
1876 * Call when io_u is really queued, to update the submission latency.
1878 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1880 if (!td->o.disable_slat) {
1881 unsigned long slat_time;
1883 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1884 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1890 * See if we should reuse the last seed, if dedupe is enabled
1892 static struct frand_state *get_buf_state(struct thread_data *td)
1897 if (!td->o.dedupe_percentage)
1898 return &td->buf_state;
1899 else if (td->o.dedupe_percentage == 100) {
1900 frand_copy(&td->buf_state_prev, &td->buf_state);
1901 return &td->buf_state;
1904 r = __rand(&td->dedupe_state);
1905 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1907 if (v <= td->o.dedupe_percentage)
1908 return &td->buf_state_prev;
1910 return &td->buf_state;
1913 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1915 if (td->o.dedupe_percentage == 100)
1916 frand_copy(rs, &td->buf_state_prev);
1917 else if (rs == &td->buf_state)
1918 frand_copy(&td->buf_state_prev, rs);
1921 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1922 unsigned int max_bs)
1924 struct thread_options *o = &td->o;
1926 if (o->compress_percentage || o->dedupe_percentage) {
1927 unsigned int perc = td->o.compress_percentage;
1928 struct frand_state *rs;
1929 unsigned int left = max_bs;
1932 rs = get_buf_state(td);
1934 min_write = min(min_write, left);
1937 unsigned int seg = min_write;
1939 seg = min(min_write, td->o.compress_chunk);
1943 fill_random_buf_percentage(rs, buf, perc, seg,
1944 min_write, o->buffer_pattern,
1945 o->buffer_pattern_bytes);
1947 fill_random_buf(rs, buf, min_write);
1951 save_buf_state(td, rs);
1953 } else if (o->buffer_pattern_bytes)
1954 fill_buffer_pattern(td, buf, max_bs);
1955 else if (o->zero_buffers)
1956 memset(buf, 0, max_bs);
1958 fill_random_buf(get_buf_state(td), buf, max_bs);
1962 * "randomly" fill the buffer contents
1964 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1965 unsigned int min_write, unsigned int max_bs)
1967 io_u->buf_filled_len = 0;
1968 fill_io_buffer(td, io_u->buf, min_write, max_bs);