12 struct io_completion_data {
15 int error; /* output */
16 unsigned long bytes_done[2]; /* output */
17 struct timeval time; /* output */
21 * The ->file_map[] contains a map of blocks we have or have not done io
22 * to yet. Used to make sure we cover the entire range in a fair fashion.
24 static int random_map_free(struct fio_file *f, const unsigned long long block)
26 unsigned int idx = RAND_MAP_IDX(f, block);
27 unsigned int bit = RAND_MAP_BIT(f, block);
29 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
31 return (f->file_map[idx] & (1 << bit)) == 0;
35 * Mark a given offset as used in the map.
37 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
39 unsigned int min_bs = td->o.rw_min_bs;
40 struct fio_file *f = io_u->file;
41 unsigned long long block;
42 unsigned int blocks, nr_blocks;
44 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
45 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
49 unsigned int this_blocks, mask;
50 unsigned int idx, bit;
53 * If we have a mixed random workload, we may
54 * encounter blocks we already did IO to.
56 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
62 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
65 fio_assert(td, idx < f->num_maps);
67 this_blocks = nr_blocks;
68 if (this_blocks + bit > BLOCKS_PER_MAP)
69 this_blocks = BLOCKS_PER_MAP - bit;
71 if (this_blocks == BLOCKS_PER_MAP)
74 mask = ((1U << this_blocks) - 1) << bit;
76 f->file_map[idx] |= mask;
77 nr_blocks -= this_blocks;
78 blocks += this_blocks;
82 if ((blocks * min_bs) < io_u->buflen)
83 io_u->buflen = blocks * min_bs;
86 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
89 unsigned long long max_blocks;
90 unsigned long long max_size;
93 * Hmm, should we make sure that ->io_size <= ->real_file_size?
95 max_size = f->io_size;
96 if (max_size > f->real_file_size)
97 max_size = f->real_file_size;
99 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
107 * Return the next free block in the map.
109 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
110 enum fio_ddir ddir, unsigned long long *b)
112 unsigned long long min_bs = td->o.rw_min_bs;
115 i = f->last_free_lookup;
116 *b = (i * BLOCKS_PER_MAP);
117 while ((*b) * min_bs < f->real_file_size &&
118 (*b) * min_bs < f->io_size) {
119 if (f->file_map[i] != (unsigned int) -1) {
120 *b += ffz(f->file_map[i]);
121 if (*b > last_block(td, f, ddir))
123 f->last_free_lookup = i;
127 *b += BLOCKS_PER_MAP;
131 dprint(FD_IO, "failed finding a free block\n");
135 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
136 enum fio_ddir ddir, unsigned long long *b)
138 unsigned long long r;
142 r = os_random_long(&td->random_state);
143 dprint(FD_RANDOM, "off rand %llu\n", r);
144 *b = (last_block(td, f, ddir) - 1)
145 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
148 * if we are not maintaining a random map, we are done.
150 if (!file_randommap(td, f))
154 * calculate map offset and check if it's free
156 if (random_map_free(f, *b))
159 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
164 * we get here, if we didn't suceed in looking up a block. generate
165 * a random start offset into the filemap, and find the first free
170 f->last_free_lookup = (f->num_maps - 1) *
171 (r / (OS_RAND_MAX + 1.0));
172 if (!get_next_free_block(td, f, ddir, b))
175 r = os_random_long(&td->random_state);
179 * that didn't work either, try exhaustive search from the start
181 f->last_free_lookup = 0;
182 return get_next_free_block(td, f, ddir, b);
186 * For random io, generate a random new block and see if it's used. Repeat
187 * until we find a free one. For sequential io, just return the end of
188 * the last io issued.
190 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
192 struct fio_file *f = io_u->file;
193 unsigned long long b;
194 enum fio_ddir ddir = io_u->ddir;
196 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
197 td->ddir_nr = td->o.ddir_nr;
199 if (get_next_rand_offset(td, f, ddir, &b)) {
200 dprint(FD_IO, "%s: getting rand offset failed\n",
205 if (f->last_pos >= f->real_file_size) {
206 if (!td_random(td) ||
207 get_next_rand_offset(td, f, ddir, &b)) {
208 dprint(FD_IO, "%s: pos %llu > size %llu\n",
209 f->file_name, f->last_pos,
214 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
217 io_u->offset = b * td->o.ba[ddir];
218 if (io_u->offset >= f->io_size) {
219 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
220 io_u->offset, f->io_size);
224 io_u->offset += f->file_offset;
225 if (io_u->offset >= f->real_file_size) {
226 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
227 io_u->offset, f->real_file_size);
234 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
236 const int ddir = io_u->ddir;
237 unsigned int uninitialized_var(buflen);
238 unsigned int minbs, maxbs;
241 minbs = td->o.min_bs[ddir];
242 maxbs = td->o.max_bs[ddir];
247 r = os_random_long(&td->bsrange_state);
248 if (!td->o.bssplit_nr[ddir]) {
249 buflen = 1 + (unsigned int) ((double) maxbs *
250 (r / (OS_RAND_MAX + 1.0)));
257 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
258 struct bssplit *bsp = &td->o.bssplit[ddir][i];
262 if (r <= ((OS_RAND_MAX / 100L) * perc))
266 if (!td->o.bs_unaligned && is_power_of_2(minbs))
267 buflen = (buflen + minbs - 1) & ~(minbs - 1);
270 if (io_u->offset + buflen > io_u->file->real_file_size) {
271 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
279 static void set_rwmix_bytes(struct thread_data *td)
284 * we do time or byte based switch. this is needed because
285 * buffered writes may issue a lot quicker than they complete,
286 * whereas reads do not.
288 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
289 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
292 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
297 r = os_random_long(&td->rwmix_state);
298 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
299 if (v <= td->o.rwmix[DDIR_READ])
305 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
307 enum fio_ddir odir = ddir ^ 1;
311 if (td->rate_pending_usleep[ddir] <= 0)
315 * We have too much pending sleep in this direction. See if we
320 * Other direction does not have too much pending, switch
322 if (td->rate_pending_usleep[odir] < 100000)
326 * Both directions have pending sleep. Sleep the minimum time
327 * and deduct from both.
329 if (td->rate_pending_usleep[ddir] <=
330 td->rate_pending_usleep[odir]) {
331 usec = td->rate_pending_usleep[ddir];
333 usec = td->rate_pending_usleep[odir];
337 usec = td->rate_pending_usleep[ddir];
339 fio_gettime(&t, NULL);
340 usec_sleep(td, usec);
341 usec = utime_since_now(&t);
343 td->rate_pending_usleep[ddir] -= usec;
346 if (td_rw(td) && __should_check_rate(td, odir))
347 td->rate_pending_usleep[odir] -= usec;
353 * Return the data direction for the next io_u. If the job is a
354 * mixed read/write workload, check the rwmix cycle and switch if
357 static enum fio_ddir get_rw_ddir(struct thread_data *td)
362 * see if it's time to fsync
364 if (td->o.fsync_blocks &&
365 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
366 td->io_issues[DDIR_WRITE] && should_fsync(td))
370 * see if it's time to fdatasync
372 if (td->o.fdatasync_blocks &&
373 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
374 td->io_issues[DDIR_WRITE] && should_fsync(td))
375 return DDIR_DATASYNC;
379 * Check if it's time to seed a new data direction.
381 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
383 * Put a top limit on how many bytes we do for
384 * one data direction, to avoid overflowing the
387 ddir = get_rand_ddir(td);
389 if (ddir != td->rwmix_ddir)
392 td->rwmix_ddir = ddir;
394 ddir = td->rwmix_ddir;
395 } else if (td_read(td))
400 td->rwmix_ddir = rate_ddir(td, ddir);
401 return td->rwmix_ddir;
404 static void put_file_log(struct thread_data *td, struct fio_file *f)
406 int ret = put_file(td, f);
409 td_verror(td, ret, "file close");
412 void put_io_u(struct thread_data *td, struct io_u *io_u)
414 assert((io_u->flags & IO_U_F_FREE) == 0);
415 io_u->flags |= IO_U_F_FREE;
418 put_file_log(td, io_u->file);
421 flist_del(&io_u->list);
422 flist_add(&io_u->list, &td->io_u_freelist);
426 void clear_io_u(struct thread_data *td, struct io_u *io_u)
428 io_u->flags &= ~IO_U_F_FLIGHT;
432 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
434 struct io_u *__io_u = *io_u;
436 dprint(FD_IO, "requeue %p\n", __io_u);
438 __io_u->flags |= IO_U_F_FREE;
439 if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
440 td->io_issues[__io_u->ddir]--;
442 __io_u->flags &= ~IO_U_F_FLIGHT;
444 flist_del(&__io_u->list);
445 flist_add_tail(&__io_u->list, &td->io_u_requeues);
450 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
452 if (td->io_ops->flags & FIO_NOIO)
455 io_u->ddir = get_rw_ddir(td);
458 * fsync() or fdatasync(), we are done
460 if (ddir_sync(io_u->ddir))
464 * See if it's time to switch to a new zone
466 if (td->zone_bytes >= td->o.zone_size) {
468 io_u->file->last_pos += td->o.zone_skip;
469 td->io_skip_bytes += td->o.zone_skip;
473 * No log, let the seq/rand engine retrieve the next buflen and
476 if (get_next_offset(td, io_u)) {
477 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
481 io_u->buflen = get_next_buflen(td, io_u);
483 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
487 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
488 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
489 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
490 io_u->buflen, io_u->file->real_file_size);
495 * mark entry before potentially trimming io_u
497 if (td_random(td) && file_randommap(td, io_u->file))
498 mark_random_map(td, io_u);
501 * If using a write iolog, store this entry.
504 dprint_io_u(io_u, "fill_io_u");
505 td->zone_bytes += io_u->buflen;
510 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
539 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
541 __io_u_mark_map(td->ts.io_u_submit, nr);
542 td->ts.total_submit++;
545 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
547 __io_u_mark_map(td->ts.io_u_complete, nr);
548 td->ts.total_complete++;
551 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
555 switch (td->cur_depth) {
577 td->ts.io_u_map[index] += nr;
580 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
617 assert(index < FIO_IO_U_LAT_U_NR);
618 td->ts.io_u_lat_u[index]++;
621 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
662 assert(index < FIO_IO_U_LAT_M_NR);
663 td->ts.io_u_lat_m[index]++;
666 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
669 io_u_mark_lat_usec(td, usec);
671 io_u_mark_lat_msec(td, usec / 1000);
675 * Get next file to service by choosing one at random
677 static struct fio_file *get_next_file_rand(struct thread_data *td,
678 enum fio_file_flags goodf,
679 enum fio_file_flags badf)
685 long r = os_random_long(&td->next_file_state);
688 fno = (unsigned int) ((double) td->o.nr_files
689 * (r / (OS_RAND_MAX + 1.0)));
691 if (fio_file_done(f))
694 if (!fio_file_open(f)) {
697 err = td_io_open_file(td, f);
703 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
704 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
708 td_io_close_file(td, f);
713 * Get next file to service by doing round robin between all available ones
715 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
718 unsigned int old_next_file = td->next_file;
724 f = td->files[td->next_file];
727 if (td->next_file >= td->o.nr_files)
730 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
731 if (fio_file_done(f)) {
736 if (!fio_file_open(f)) {
739 err = td_io_open_file(td, f);
741 dprint(FD_FILE, "error %d on open of %s\n",
749 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
750 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
754 td_io_close_file(td, f);
757 } while (td->next_file != old_next_file);
759 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
763 static struct fio_file *get_next_file(struct thread_data *td)
767 assert(td->o.nr_files <= td->files_index);
769 if (td->nr_done_files >= td->o.nr_files) {
770 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
771 " nr_files=%d\n", td->nr_open_files,
777 f = td->file_service_file;
778 if (f && fio_file_open(f) && !fio_file_closing(f)) {
779 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
781 if (td->file_service_left--)
785 if (td->o.file_service_type == FIO_FSERVICE_RR ||
786 td->o.file_service_type == FIO_FSERVICE_SEQ)
787 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
789 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
791 td->file_service_file = f;
792 td->file_service_left = td->file_service_nr - 1;
794 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
798 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
803 f = get_next_file(td);
810 if (!fill_io_u(td, io_u))
814 td_io_close_file(td, f);
816 fio_file_set_done(f);
818 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
825 struct io_u *__get_io_u(struct thread_data *td)
827 struct io_u *io_u = NULL;
829 if (!flist_empty(&td->io_u_requeues))
830 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
831 else if (!queue_full(td)) {
832 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
841 assert(io_u->flags & IO_U_F_FREE);
842 io_u->flags &= ~IO_U_F_FREE;
845 flist_del(&io_u->list);
846 flist_add(&io_u->list, &td->io_u_busylist);
854 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
855 * etc. The returned io_u is fully ready to be prepped and submitted.
857 struct io_u *get_io_u(struct thread_data *td)
862 io_u = __get_io_u(td);
864 dprint(FD_IO, "__get_io_u failed\n");
869 * from a requeue, io_u already setup
875 * If using an iolog, grab next piece if any available.
877 if (td->o.read_iolog_file) {
878 if (read_iolog_get(td, io_u))
880 } else if (set_io_u_file(td, io_u)) {
881 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
886 assert(fio_file_open(f));
888 if (!ddir_sync(io_u->ddir)) {
889 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
890 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
894 f->last_pos = io_u->offset + io_u->buflen;
896 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
897 populate_verify_io_u(td, io_u);
898 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
899 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
903 * Set io data pointers.
905 io_u->xfer_buf = io_u->buf;
906 io_u->xfer_buflen = io_u->buflen;
909 if (!td_io_prep(td, io_u)) {
910 if (!td->o.disable_slat)
911 fio_gettime(&io_u->start_time, NULL);
915 dprint(FD_IO, "get_io_u failed\n");
920 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
922 const char *msg[] = { "read", "write", "sync" };
924 log_err("fio: io_u error");
927 log_err(" on file %s", io_u->file->file_name);
929 log_err(": %s\n", strerror(io_u->error));
931 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
932 io_u->offset, io_u->xfer_buflen);
935 td_verror(td, io_u->error, "io_u error");
938 static void io_completed(struct thread_data *td, struct io_u *io_u,
939 struct io_completion_data *icd)
942 * Older gcc's are too dumb to realize that usec is always used
943 * initialized, silence that warning.
945 unsigned long uninitialized_var(usec);
947 dprint_io_u(io_u, "io complete");
949 assert(io_u->flags & IO_U_F_FLIGHT);
950 io_u->flags &= ~IO_U_F_FLIGHT;
952 if (ddir_sync(io_u->ddir)) {
953 td->last_was_sync = 1;
957 td->last_was_sync = 0;
960 unsigned int bytes = io_u->buflen - io_u->resid;
961 const enum fio_ddir idx = io_u->ddir;
964 td->io_blocks[idx]++;
965 td->io_bytes[idx] += bytes;
966 td->this_io_bytes[idx] += bytes;
968 if (ramp_time_over(td)) {
969 unsigned long uninitialized_var(lusec);
970 unsigned long uninitialized_var(rusec);
972 if (!td->o.disable_clat || !td->o.disable_bw)
973 lusec = utime_since(&io_u->issue_time,
975 if (__should_check_rate(td, idx) ||
976 __should_check_rate(td, idx ^ 1))
977 rusec = utime_since(&io_u->start_time,
980 if (!td->o.disable_clat) {
981 add_clat_sample(td, idx, usec, bytes);
982 io_u_mark_latency(td, lusec);
984 if (!td->o.disable_bw)
985 add_bw_sample(td, idx, bytes, &icd->time);
986 if (__should_check_rate(td, idx)) {
987 td->rate_pending_usleep[idx] +=
988 (long) td->rate_usec_cycle[idx] - rusec;
990 if (__should_check_rate(td, idx ^ 1))
991 td->rate_pending_usleep[idx ^ 1] -= rusec;
994 if (td_write(td) && idx == DDIR_WRITE &&
996 td->o.verify != VERIFY_NONE)
997 log_io_piece(td, io_u);
999 icd->bytes_done[idx] += bytes;
1002 ret = io_u->end_io(td, io_u);
1003 if (ret && !icd->error)
1007 icd->error = io_u->error;
1008 io_u_log_error(td, io_u);
1010 if (td->o.continue_on_error && icd->error &&
1011 td_non_fatal_error(icd->error)) {
1013 * If there is a non_fatal error, then add to the error count
1014 * and clear all the errors.
1016 update_error_count(td, icd->error);
1023 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1026 if (!td->o.disable_clat || !td->o.disable_bw)
1027 fio_gettime(&icd->time, NULL);
1032 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1035 static void ios_completed(struct thread_data *td,
1036 struct io_completion_data *icd)
1041 for (i = 0; i < icd->nr; i++) {
1042 io_u = td->io_ops->event(td, i);
1044 io_completed(td, io_u, icd);
1050 * Complete a single io_u for the sync engines.
1052 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1053 unsigned long *bytes)
1055 struct io_completion_data icd;
1057 init_icd(td, &icd, 1);
1058 io_completed(td, io_u, &icd);
1062 td_verror(td, icd.error, "io_u_sync_complete");
1067 bytes[0] += icd.bytes_done[0];
1068 bytes[1] += icd.bytes_done[1];
1075 * Called to complete min_events number of io for the async engines.
1077 int io_u_queued_complete(struct thread_data *td, int min_evts,
1078 unsigned long *bytes)
1080 struct io_completion_data icd;
1081 struct timespec *tvp = NULL;
1083 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1085 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1090 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1092 td_verror(td, -ret, "td_io_getevents");
1097 init_icd(td, &icd, ret);
1098 ios_completed(td, &icd);
1100 td_verror(td, icd.error, "io_u_queued_complete");
1105 bytes[0] += icd.bytes_done[0];
1106 bytes[1] += icd.bytes_done[1];
1113 * Call when io_u is really queued, to update the submission latency.
1115 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1117 if (!td->o.disable_slat) {
1118 unsigned long slat_time;
1120 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1121 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1126 * "randomly" fill the buffer contents
1128 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1129 unsigned int max_bs)
1131 long *ptr = io_u->buf;
1133 if (!td->o.zero_buffers) {
1134 while ((void *) ptr - io_u->buf < max_bs) {
1135 *ptr = rand() * GOLDEN_RATIO_PRIME;
1139 memset(ptr, 0, max_bs);