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->flags |= 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->flags |= 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 = (buflen + minbs - 1) & ~(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, NULL);
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];
596 usec = usec_sleep(td, usec);
598 td->rate_pending_usleep[ddir] -= usec;
601 if (td_rw(td) && __should_check_rate(td, odir))
602 td->rate_pending_usleep[odir] -= usec;
608 * Return the data direction for the next io_u. If the job is a
609 * mixed read/write workload, check the rwmix cycle and switch if
612 static enum fio_ddir get_rw_ddir(struct thread_data *td)
617 * see if it's time to fsync
619 if (td->o.fsync_blocks &&
620 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
621 td->io_issues[DDIR_WRITE] && should_fsync(td))
625 * see if it's time to fdatasync
627 if (td->o.fdatasync_blocks &&
628 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
629 td->io_issues[DDIR_WRITE] && should_fsync(td))
630 return DDIR_DATASYNC;
633 * see if it's time to sync_file_range
635 if (td->sync_file_range_nr &&
636 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
637 td->io_issues[DDIR_WRITE] && should_fsync(td))
638 return DDIR_SYNC_FILE_RANGE;
642 * Check if it's time to seed a new data direction.
644 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
646 * Put a top limit on how many bytes we do for
647 * one data direction, to avoid overflowing the
650 ddir = get_rand_ddir(td);
652 if (ddir != td->rwmix_ddir)
655 td->rwmix_ddir = ddir;
657 ddir = td->rwmix_ddir;
658 } else if (td_read(td))
660 else if (td_write(td))
665 td->rwmix_ddir = rate_ddir(td, ddir);
666 return td->rwmix_ddir;
669 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
671 enum fio_ddir ddir = get_rw_ddir(td);
673 if (td_writetrim(td)) {
674 struct fio_file *f = io_u->file;
675 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
681 io_u->ddir = io_u->acct_ddir = ddir;
683 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
684 td->o.barrier_blocks &&
685 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
686 td->io_issues[DDIR_WRITE])
687 io_u->flags |= IO_U_F_BARRIER;
690 void put_file_log(struct thread_data *td, struct fio_file *f)
692 unsigned int ret = put_file(td, f);
695 td_verror(td, ret, "file close");
698 void put_io_u(struct thread_data *td, struct io_u *io_u)
702 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
703 put_file_log(td, io_u->file);
706 io_u->flags |= IO_U_F_FREE;
708 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
710 io_u_qpush(&td->io_u_freelist, io_u);
712 td_io_u_free_notify(td);
715 void clear_io_u(struct thread_data *td, struct io_u *io_u)
717 io_u->flags &= ~IO_U_F_FLIGHT;
721 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
723 struct io_u *__io_u = *io_u;
724 enum fio_ddir ddir = acct_ddir(__io_u);
726 dprint(FD_IO, "requeue %p\n", __io_u);
730 __io_u->flags |= IO_U_F_FREE;
731 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
732 td->io_issues[ddir]--;
734 __io_u->flags &= ~IO_U_F_FLIGHT;
735 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
738 io_u_rpush(&td->io_u_requeues, __io_u);
743 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
745 unsigned int is_random;
747 if (td->io_ops->flags & FIO_NOIO)
750 set_rw_ddir(td, io_u);
753 * fsync() or fdatasync() or trim etc, we are done
755 if (!ddir_rw(io_u->ddir))
759 * See if it's time to switch to a new zone
761 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
762 struct fio_file *f = io_u->file;
765 f->file_offset += td->o.zone_range + td->o.zone_skip;
768 * Wrap from the beginning, if we exceed the file size
770 if (f->file_offset >= f->real_file_size)
771 f->file_offset = f->real_file_size - f->file_offset;
772 f->last_pos[io_u->ddir] = f->file_offset;
773 td->io_skip_bytes += td->o.zone_skip;
777 * No log, let the seq/rand engine retrieve the next buflen and
780 if (get_next_offset(td, io_u, &is_random)) {
781 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
785 io_u->buflen = get_next_buflen(td, io_u, is_random);
787 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
791 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
792 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
793 dprint(FD_IO, " off=%llu/%lu > %llu\n",
794 (unsigned long long) io_u->offset, io_u->buflen,
795 (unsigned long long) io_u->file->real_file_size);
800 * mark entry before potentially trimming io_u
802 if (td_random(td) && file_randommap(td, io_u->file))
803 mark_random_map(td, io_u);
806 dprint_io_u(io_u, "fill_io_u");
807 td->zone_bytes += io_u->buflen;
811 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
840 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
842 __io_u_mark_map(td->ts.io_u_submit, nr);
843 td->ts.total_submit++;
846 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
848 __io_u_mark_map(td->ts.io_u_complete, nr);
849 td->ts.total_complete++;
852 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
856 switch (td->cur_depth) {
878 td->ts.io_u_map[idx] += nr;
881 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
918 assert(idx < FIO_IO_U_LAT_U_NR);
919 td->ts.io_u_lat_u[idx]++;
922 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
963 assert(idx < FIO_IO_U_LAT_M_NR);
964 td->ts.io_u_lat_m[idx]++;
967 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
970 io_u_mark_lat_usec(td, usec);
972 io_u_mark_lat_msec(td, usec / 1000);
976 * Get next file to service by choosing one at random
978 static struct fio_file *get_next_file_rand(struct thread_data *td,
979 enum fio_file_flags goodf,
980 enum fio_file_flags badf)
989 r = __rand(&td->next_file_state);
990 fno = (unsigned int) ((double) td->o.nr_files
991 * (r / (FRAND_MAX + 1.0)));
994 if (fio_file_done(f))
997 if (!fio_file_open(f)) {
1000 if (td->nr_open_files >= td->o.open_files)
1001 return ERR_PTR(-EBUSY);
1003 err = td_io_open_file(td, f);
1009 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1010 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1014 td_io_close_file(td, f);
1019 * Get next file to service by doing round robin between all available ones
1021 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1024 unsigned int old_next_file = td->next_file;
1030 f = td->files[td->next_file];
1033 if (td->next_file >= td->o.nr_files)
1036 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1037 if (fio_file_done(f)) {
1042 if (!fio_file_open(f)) {
1045 if (td->nr_open_files >= td->o.open_files)
1046 return ERR_PTR(-EBUSY);
1048 err = td_io_open_file(td, f);
1050 dprint(FD_FILE, "error %d on open of %s\n",
1058 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1060 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1064 td_io_close_file(td, f);
1067 } while (td->next_file != old_next_file);
1069 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1073 static struct fio_file *__get_next_file(struct thread_data *td)
1077 assert(td->o.nr_files <= td->files_index);
1079 if (td->nr_done_files >= td->o.nr_files) {
1080 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1081 " nr_files=%d\n", td->nr_open_files,
1087 f = td->file_service_file;
1088 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1089 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1091 if (td->file_service_left--)
1095 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1096 td->o.file_service_type == FIO_FSERVICE_SEQ)
1097 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1099 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1104 td->file_service_file = f;
1105 td->file_service_left = td->file_service_nr - 1;
1108 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1110 dprint(FD_FILE, "get_next_file: NULL\n");
1114 static struct fio_file *get_next_file(struct thread_data *td)
1116 if (td->flags & TD_F_PROFILE_OPS) {
1117 struct prof_io_ops *ops = &td->prof_io_ops;
1119 if (ops->get_next_file)
1120 return ops->get_next_file(td);
1123 return __get_next_file(td);
1126 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1131 f = get_next_file(td);
1132 if (IS_ERR_OR_NULL(f))
1138 if (!fill_io_u(td, io_u))
1141 put_file_log(td, f);
1142 td_io_close_file(td, f);
1144 fio_file_set_done(f);
1145 td->nr_done_files++;
1146 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1147 td->nr_done_files, td->o.nr_files);
1153 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1154 unsigned long tusec, unsigned long max_usec)
1157 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1158 td_verror(td, ETIMEDOUT, "max latency exceeded");
1159 icd->error = ETIMEDOUT;
1162 static void lat_new_cycle(struct thread_data *td)
1164 fio_gettime(&td->latency_ts, NULL);
1165 td->latency_ios = ddir_rw_sum(td->io_blocks);
1166 td->latency_failed = 0;
1170 * We had an IO outside the latency target. Reduce the queue depth. If we
1171 * are at QD=1, then it's time to give up.
1173 static int __lat_target_failed(struct thread_data *td)
1175 if (td->latency_qd == 1)
1178 td->latency_qd_high = td->latency_qd;
1180 if (td->latency_qd == td->latency_qd_low)
1181 td->latency_qd_low--;
1183 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1185 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1188 * When we ramp QD down, quiesce existing IO to prevent
1189 * a storm of ramp downs due to pending higher depth.
1196 static int lat_target_failed(struct thread_data *td)
1198 if (td->o.latency_percentile.u.f == 100.0)
1199 return __lat_target_failed(td);
1201 td->latency_failed++;
1205 void lat_target_init(struct thread_data *td)
1207 td->latency_end_run = 0;
1209 if (td->o.latency_target) {
1210 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1211 fio_gettime(&td->latency_ts, NULL);
1213 td->latency_qd_high = td->o.iodepth;
1214 td->latency_qd_low = 1;
1215 td->latency_ios = ddir_rw_sum(td->io_blocks);
1217 td->latency_qd = td->o.iodepth;
1220 void lat_target_reset(struct thread_data *td)
1222 if (!td->latency_end_run)
1223 lat_target_init(td);
1226 static void lat_target_success(struct thread_data *td)
1228 const unsigned int qd = td->latency_qd;
1229 struct thread_options *o = &td->o;
1231 td->latency_qd_low = td->latency_qd;
1234 * If we haven't failed yet, we double up to a failing value instead
1235 * of bisecting from highest possible queue depth. If we have set
1236 * a limit other than td->o.iodepth, bisect between that.
1238 if (td->latency_qd_high != o->iodepth)
1239 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1241 td->latency_qd *= 2;
1243 if (td->latency_qd > o->iodepth)
1244 td->latency_qd = o->iodepth;
1246 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1249 * Same as last one, we are done. Let it run a latency cycle, so
1250 * we get only the results from the targeted depth.
1252 if (td->latency_qd == qd) {
1253 if (td->latency_end_run) {
1254 dprint(FD_RATE, "We are done\n");
1257 dprint(FD_RATE, "Quiesce and final run\n");
1259 td->latency_end_run = 1;
1260 reset_all_stats(td);
1269 * Check if we can bump the queue depth
1271 void lat_target_check(struct thread_data *td)
1273 uint64_t usec_window;
1277 usec_window = utime_since_now(&td->latency_ts);
1278 if (usec_window < td->o.latency_window)
1281 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1282 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1283 success_ios *= 100.0;
1285 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1287 if (success_ios >= td->o.latency_percentile.u.f)
1288 lat_target_success(td);
1290 __lat_target_failed(td);
1294 * If latency target is enabled, we might be ramping up or down and not
1295 * using the full queue depth available.
1297 int queue_full(const struct thread_data *td)
1299 const int qempty = io_u_qempty(&td->io_u_freelist);
1303 if (!td->o.latency_target)
1306 return td->cur_depth >= td->latency_qd;
1309 struct io_u *__get_io_u(struct thread_data *td)
1311 struct io_u *io_u = NULL;
1319 if (!io_u_rempty(&td->io_u_requeues))
1320 io_u = io_u_rpop(&td->io_u_requeues);
1321 else if (!queue_full(td)) {
1322 io_u = io_u_qpop(&td->io_u_freelist);
1327 io_u->end_io = NULL;
1331 assert(io_u->flags & IO_U_F_FREE);
1332 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1333 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1337 io_u->acct_ddir = -1;
1339 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1341 } else if (td->o.verify_async) {
1343 * We ran out, wait for async verify threads to finish and
1346 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1354 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1356 if (!(td->flags & TD_F_TRIM_BACKLOG))
1359 if (td->trim_entries) {
1362 if (td->trim_batch) {
1365 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1366 td->last_ddir != DDIR_READ) {
1367 td->trim_batch = td->o.trim_batch;
1368 if (!td->trim_batch)
1369 td->trim_batch = td->o.trim_backlog;
1373 if (get_trim && !get_next_trim(td, io_u))
1380 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1382 if (!(td->flags & TD_F_VER_BACKLOG))
1385 if (td->io_hist_len) {
1388 if (td->verify_batch)
1390 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1391 td->last_ddir != DDIR_READ) {
1392 td->verify_batch = td->o.verify_batch;
1393 if (!td->verify_batch)
1394 td->verify_batch = td->o.verify_backlog;
1398 if (get_verify && !get_next_verify(td, io_u)) {
1408 * Fill offset and start time into the buffer content, to prevent too
1409 * easy compressible data for simple de-dupe attempts. Do this for every
1410 * 512b block in the range, since that should be the smallest block size
1411 * we can expect from a device.
1413 static void small_content_scramble(struct io_u *io_u)
1415 unsigned int i, nr_blocks = io_u->buflen / 512;
1417 unsigned int offset;
1424 boffset = io_u->offset;
1425 io_u->buf_filled_len = 0;
1427 for (i = 0; i < nr_blocks; i++) {
1429 * Fill the byte offset into a "random" start offset of
1430 * the buffer, given by the product of the usec time
1431 * and the actual offset.
1433 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1434 offset &= ~(sizeof(uint64_t) - 1);
1435 if (offset >= 512 - sizeof(uint64_t))
1436 offset -= sizeof(uint64_t);
1437 memcpy(p + offset, &boffset, sizeof(boffset));
1439 end = p + 512 - sizeof(io_u->start_time);
1440 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1447 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1448 * etc. The returned io_u is fully ready to be prepped and submitted.
1450 struct io_u *get_io_u(struct thread_data *td)
1454 int do_scramble = 0;
1457 io_u = __get_io_u(td);
1459 dprint(FD_IO, "__get_io_u failed\n");
1463 if (check_get_verify(td, io_u))
1465 if (check_get_trim(td, io_u))
1469 * from a requeue, io_u already setup
1475 * If using an iolog, grab next piece if any available.
1477 if (td->flags & TD_F_READ_IOLOG) {
1478 if (read_iolog_get(td, io_u))
1480 } else if (set_io_u_file(td, io_u)) {
1482 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1488 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1492 assert(fio_file_open(f));
1494 if (ddir_rw(io_u->ddir)) {
1495 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1496 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1500 f->last_start[io_u->ddir] = io_u->offset;
1501 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1503 if (io_u->ddir == DDIR_WRITE) {
1504 if (td->flags & TD_F_REFILL_BUFFERS) {
1505 io_u_fill_buffer(td, io_u,
1506 td->o.min_bs[DDIR_WRITE],
1508 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1509 !(td->flags & TD_F_COMPRESS))
1511 if (td->flags & TD_F_VER_NONE) {
1512 populate_verify_io_u(td, io_u);
1515 } else if (io_u->ddir == DDIR_READ) {
1517 * Reset the buf_filled parameters so next time if the
1518 * buffer is used for writes it is refilled.
1520 io_u->buf_filled_len = 0;
1525 * Set io data pointers.
1527 io_u->xfer_buf = io_u->buf;
1528 io_u->xfer_buflen = io_u->buflen;
1532 if (!td_io_prep(td, io_u)) {
1533 if (!td->o.disable_slat)
1534 fio_gettime(&io_u->start_time, NULL);
1536 small_content_scramble(io_u);
1540 dprint(FD_IO, "get_io_u failed\n");
1542 return ERR_PTR(ret);
1545 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1547 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1549 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1552 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1553 io_u->file ? " on file " : "",
1554 io_u->file ? io_u->file->file_name : "",
1555 strerror(io_u->error),
1556 io_ddir_name(io_u->ddir),
1557 io_u->offset, io_u->xfer_buflen);
1560 td_verror(td, io_u->error, "io_u error");
1563 static inline int gtod_reduce(struct thread_data *td)
1565 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1566 && td->o.disable_bw;
1569 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1570 struct io_completion_data *icd,
1571 const enum fio_ddir idx, unsigned int bytes)
1573 unsigned long lusec = 0;
1575 if (!gtod_reduce(td))
1576 lusec = utime_since(&io_u->issue_time, &icd->time);
1578 if (!td->o.disable_lat) {
1579 unsigned long tusec;
1581 tusec = utime_since(&io_u->start_time, &icd->time);
1582 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1584 if (td->flags & TD_F_PROFILE_OPS) {
1585 struct prof_io_ops *ops = &td->prof_io_ops;
1588 icd->error = ops->io_u_lat(td, tusec);
1591 if (td->o.max_latency && tusec > td->o.max_latency)
1592 lat_fatal(td, icd, tusec, td->o.max_latency);
1593 if (td->o.latency_target && tusec > td->o.latency_target) {
1594 if (lat_target_failed(td))
1595 lat_fatal(td, icd, tusec, td->o.latency_target);
1599 if (!td->o.disable_clat) {
1600 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1601 io_u_mark_latency(td, lusec);
1604 if (!td->o.disable_bw)
1605 add_bw_sample(td, idx, bytes, &icd->time);
1607 if (!gtod_reduce(td))
1608 add_iops_sample(td, idx, bytes, &icd->time);
1610 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1611 uint32_t *info = io_u_block_info(td, io_u);
1612 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1613 if (io_u->ddir == DDIR_TRIM) {
1614 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1615 BLOCK_INFO_TRIMS(*info) + 1);
1616 } else if (io_u->ddir == DDIR_WRITE) {
1617 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1624 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1626 uint64_t secs, remainder, bps, bytes;
1628 bytes = td->this_io_bytes[ddir];
1629 bps = td->rate_bps[ddir];
1631 remainder = bytes % bps;
1632 return remainder * 1000000 / bps + secs * 1000000;
1635 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1636 struct io_completion_data *icd)
1638 struct io_u *io_u = *io_u_ptr;
1639 enum fio_ddir ddir = io_u->ddir;
1640 struct fio_file *f = io_u->file;
1642 dprint_io_u(io_u, "io complete");
1645 assert(io_u->flags & IO_U_F_FLIGHT);
1646 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1649 * Mark IO ok to verify
1653 * Remove errored entry from the verification list
1656 unlog_io_piece(td, io_u);
1658 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1665 if (ddir_sync(ddir)) {
1666 td->last_was_sync = 1;
1668 f->first_write = -1ULL;
1669 f->last_write = -1ULL;
1674 td->last_was_sync = 0;
1675 td->last_ddir = ddir;
1677 if (!io_u->error && ddir_rw(ddir)) {
1678 unsigned int bytes = io_u->buflen - io_u->resid;
1679 const enum fio_ddir oddir = ddir ^ 1;
1682 td->io_blocks[ddir]++;
1683 td->this_io_blocks[ddir]++;
1684 td->io_bytes[ddir] += bytes;
1686 if (!(io_u->flags & IO_U_F_VER_LIST))
1687 td->this_io_bytes[ddir] += bytes;
1689 if (ddir == DDIR_WRITE) {
1691 if (f->first_write == -1ULL ||
1692 io_u->offset < f->first_write)
1693 f->first_write = io_u->offset;
1694 if (f->last_write == -1ULL ||
1695 ((io_u->offset + bytes) > f->last_write))
1696 f->last_write = io_u->offset + bytes;
1698 if (td->last_write_comp) {
1699 int idx = td->last_write_idx++;
1701 td->last_write_comp[idx] = io_u->offset;
1702 if (td->last_write_idx == td->o.iodepth)
1703 td->last_write_idx = 0;
1707 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1708 td->runstate == TD_VERIFYING)) {
1709 account_io_completion(td, io_u, icd, ddir, bytes);
1711 if (__should_check_rate(td, ddir)) {
1712 td->rate_pending_usleep[ddir] =
1713 (usec_for_io(td, ddir) -
1714 utime_since_now(&td->start));
1716 if (ddir != DDIR_TRIM &&
1717 __should_check_rate(td, oddir)) {
1718 td->rate_pending_usleep[oddir] =
1719 (usec_for_io(td, oddir) -
1720 utime_since_now(&td->start));
1724 icd->bytes_done[ddir] += bytes;
1727 ret = io_u->end_io(td, io_u_ptr);
1729 if (ret && !icd->error)
1732 } else if (io_u->error) {
1733 icd->error = io_u->error;
1734 io_u_log_error(td, io_u);
1737 enum error_type_bit eb = td_error_type(ddir, icd->error);
1739 if (!td_non_fatal_error(td, eb, icd->error))
1743 * If there is a non_fatal error, then add to the error count
1744 * and clear all the errors.
1746 update_error_count(td, icd->error);
1754 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1759 if (!gtod_reduce(td))
1760 fio_gettime(&icd->time, NULL);
1765 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1766 icd->bytes_done[ddir] = 0;
1769 static void ios_completed(struct thread_data *td,
1770 struct io_completion_data *icd)
1775 for (i = 0; i < icd->nr; i++) {
1776 io_u = td->io_ops->event(td, i);
1778 io_completed(td, &io_u, icd);
1786 * Complete a single io_u for the sync engines.
1788 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1791 struct io_completion_data icd;
1793 init_icd(td, &icd, 1);
1794 io_completed(td, &io_u, &icd);
1800 td_verror(td, icd.error, "io_u_sync_complete");
1807 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1808 bytes[ddir] += icd.bytes_done[ddir];
1815 * Called to complete min_events number of io for the async engines.
1817 int io_u_queued_complete(struct thread_data *td, int min_evts,
1820 struct io_completion_data icd;
1821 struct timespec *tvp = NULL;
1823 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1825 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1829 else if (min_evts > td->cur_depth)
1830 min_evts = td->cur_depth;
1832 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1834 td_verror(td, -ret, "td_io_getevents");
1839 init_icd(td, &icd, ret);
1840 ios_completed(td, &icd);
1842 td_verror(td, icd.error, "io_u_queued_complete");
1849 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1850 bytes[ddir] += icd.bytes_done[ddir];
1857 * Call when io_u is really queued, to update the submission latency.
1859 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1861 if (!td->o.disable_slat) {
1862 unsigned long slat_time;
1864 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1865 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1871 * See if we should reuse the last seed, if dedupe is enabled
1873 static struct frand_state *get_buf_state(struct thread_data *td)
1878 if (!td->o.dedupe_percentage)
1879 return &td->buf_state;
1880 else if (td->o.dedupe_percentage == 100)
1881 return &td->buf_state_prev;
1883 r = __rand(&td->dedupe_state);
1884 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1886 if (v <= td->o.dedupe_percentage)
1887 return &td->buf_state_prev;
1889 return &td->buf_state;
1892 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1894 if (rs == &td->buf_state)
1895 frand_copy(&td->buf_state_prev, rs);
1898 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1899 unsigned int max_bs)
1901 struct thread_options *o = &td->o;
1903 if (o->compress_percentage || o->dedupe_percentage) {
1904 unsigned int perc = td->o.compress_percentage;
1905 struct frand_state *rs;
1906 unsigned int left = max_bs;
1909 rs = get_buf_state(td);
1911 min_write = min(min_write, left);
1914 unsigned int seg = min_write;
1916 seg = min(min_write, td->o.compress_chunk);
1920 fill_random_buf_percentage(rs, buf, perc, seg,
1921 min_write, o->buffer_pattern,
1922 o->buffer_pattern_bytes);
1924 fill_random_buf(rs, buf, min_write);
1928 save_buf_state(td, rs);
1930 } else if (o->buffer_pattern_bytes)
1931 fill_buffer_pattern(td, buf, max_bs);
1932 else if (o->zero_buffers)
1933 memset(buf, 0, max_bs);
1935 fill_random_buf(get_buf_state(td), buf, max_bs);
1939 * "randomly" fill the buffer contents
1941 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1942 unsigned int min_write, unsigned int max_bs)
1944 io_u->buf_filled_len = 0;
1945 fill_io_buffer(td, io_u->buf, min_write, max_bs);