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 inline int is_power_of_2(unsigned int val)
236 return (val != 0 && ((val & (val - 1)) == 0));
239 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
241 const int ddir = io_u->ddir;
242 unsigned int uninitialized_var(buflen);
243 unsigned int minbs, maxbs;
246 minbs = td->o.min_bs[ddir];
247 maxbs = td->o.max_bs[ddir];
252 r = os_random_long(&td->bsrange_state);
253 if (!td->o.bssplit_nr[ddir]) {
254 buflen = 1 + (unsigned int) ((double) maxbs *
255 (r / (OS_RAND_MAX + 1.0)));
262 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
263 struct bssplit *bsp = &td->o.bssplit[ddir][i];
267 if (r <= ((OS_RAND_MAX / 100L) * perc))
271 if (!td->o.bs_unaligned && is_power_of_2(minbs))
272 buflen = (buflen + minbs - 1) & ~(minbs - 1);
275 if (io_u->offset + buflen > io_u->file->real_file_size) {
276 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
284 static void set_rwmix_bytes(struct thread_data *td)
289 * we do time or byte based switch. this is needed because
290 * buffered writes may issue a lot quicker than they complete,
291 * whereas reads do not.
293 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
294 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
297 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
302 r = os_random_long(&td->rwmix_state);
303 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
304 if (v <= td->o.rwmix[DDIR_READ])
310 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
312 enum fio_ddir odir = ddir ^ 1;
316 if (td->rate_pending_usleep[ddir] <= 0)
320 * We have too much pending sleep in this direction. See if we
325 * Other direction does not have too much pending, switch
327 if (td->rate_pending_usleep[odir] < 100000)
331 * Both directions have pending sleep. Sleep the minimum time
332 * and deduct from both.
334 if (td->rate_pending_usleep[ddir] <=
335 td->rate_pending_usleep[odir]) {
336 usec = td->rate_pending_usleep[ddir];
338 usec = td->rate_pending_usleep[odir];
342 usec = td->rate_pending_usleep[ddir];
344 fio_gettime(&t, NULL);
345 usec_sleep(td, usec);
346 usec = utime_since_now(&t);
348 td->rate_pending_usleep[ddir] -= usec;
351 if (td_rw(td) && __should_check_rate(td, odir))
352 td->rate_pending_usleep[odir] -= usec;
358 * Return the data direction for the next io_u. If the job is a
359 * mixed read/write workload, check the rwmix cycle and switch if
362 static enum fio_ddir get_rw_ddir(struct thread_data *td)
368 * Check if it's time to seed a new data direction.
370 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
372 * Put a top limit on how many bytes we do for
373 * one data direction, to avoid overflowing the
376 ddir = get_rand_ddir(td);
378 if (ddir != td->rwmix_ddir)
381 td->rwmix_ddir = ddir;
383 ddir = td->rwmix_ddir;
384 } else if (td_read(td))
389 td->rwmix_ddir = rate_ddir(td, ddir);
390 return td->rwmix_ddir;
393 static void put_file_log(struct thread_data *td, struct fio_file *f)
395 int ret = put_file(td, f);
398 td_verror(td, ret, "file close");
401 void put_io_u(struct thread_data *td, struct io_u *io_u)
403 assert((io_u->flags & IO_U_F_FREE) == 0);
404 io_u->flags |= IO_U_F_FREE;
407 put_file_log(td, io_u->file);
410 flist_del(&io_u->list);
411 flist_add(&io_u->list, &td->io_u_freelist);
415 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
417 struct io_u *__io_u = *io_u;
419 dprint(FD_IO, "requeue %p\n", __io_u);
421 __io_u->flags |= IO_U_F_FREE;
422 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
423 td->io_issues[__io_u->ddir]--;
425 __io_u->flags &= ~IO_U_F_FLIGHT;
427 flist_del(&__io_u->list);
428 flist_add_tail(&__io_u->list, &td->io_u_requeues);
433 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
435 if (td->io_ops->flags & FIO_NOIO)
439 * see if it's time to sync
441 if (td->o.fsync_blocks &&
442 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
443 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
444 io_u->ddir = DDIR_SYNC;
448 io_u->ddir = get_rw_ddir(td);
451 * See if it's time to switch to a new zone
453 if (td->zone_bytes >= td->o.zone_size) {
455 io_u->file->last_pos += td->o.zone_skip;
456 td->io_skip_bytes += td->o.zone_skip;
460 * No log, let the seq/rand engine retrieve the next buflen and
463 if (get_next_offset(td, io_u)) {
464 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
468 io_u->buflen = get_next_buflen(td, io_u);
470 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
474 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
475 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
476 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
477 io_u->buflen, io_u->file->real_file_size);
482 * mark entry before potentially trimming io_u
484 if (td_random(td) && file_randommap(td, io_u->file))
485 mark_random_map(td, io_u);
488 * If using a write iolog, store this entry.
491 dprint_io_u(io_u, "fill_io_u");
492 td->zone_bytes += io_u->buflen;
497 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
526 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
528 __io_u_mark_map(td->ts.io_u_submit, nr);
529 td->ts.total_submit++;
532 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
534 __io_u_mark_map(td->ts.io_u_complete, nr);
535 td->ts.total_complete++;
538 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
542 switch (td->cur_depth) {
564 td->ts.io_u_map[index] += nr;
567 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
604 assert(index < FIO_IO_U_LAT_U_NR);
605 td->ts.io_u_lat_u[index]++;
608 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
649 assert(index < FIO_IO_U_LAT_M_NR);
650 td->ts.io_u_lat_m[index]++;
653 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
656 io_u_mark_lat_usec(td, usec);
658 io_u_mark_lat_msec(td, usec / 1000);
662 * Get next file to service by choosing one at random
664 static struct fio_file *get_next_file_rand(struct thread_data *td, enum fio_file_flags goodf,
665 enum fio_file_flags badf)
671 long r = os_random_long(&td->next_file_state);
674 fno = (unsigned int) ((double) td->o.nr_files
675 * (r / (OS_RAND_MAX + 1.0)));
677 if (fio_file_done(f))
680 if (!fio_file_open(f)) {
683 err = td_io_open_file(td, f);
689 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
690 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
694 td_io_close_file(td, f);
699 * Get next file to service by doing round robin between all available ones
701 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
704 unsigned int old_next_file = td->next_file;
710 f = td->files[td->next_file];
713 if (td->next_file >= td->o.nr_files)
716 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
717 if (fio_file_done(f)) {
722 if (!fio_file_open(f)) {
725 err = td_io_open_file(td, f);
727 dprint(FD_FILE, "error %d on open of %s\n",
735 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
736 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
740 td_io_close_file(td, f);
743 } while (td->next_file != old_next_file);
745 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
749 static struct fio_file *get_next_file(struct thread_data *td)
753 assert(td->o.nr_files <= td->files_index);
755 if (td->nr_done_files >= td->o.nr_files) {
756 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
757 " nr_files=%d\n", td->nr_open_files,
763 f = td->file_service_file;
764 if (f && fio_file_open(f) && !fio_file_closing(f)) {
765 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
767 if (td->file_service_left--)
771 if (td->o.file_service_type == FIO_FSERVICE_RR ||
772 td->o.file_service_type == FIO_FSERVICE_SEQ)
773 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
775 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
777 td->file_service_file = f;
778 td->file_service_left = td->file_service_nr - 1;
780 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
784 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
789 f = get_next_file(td);
796 if (!fill_io_u(td, io_u))
800 td_io_close_file(td, f);
802 fio_file_set_done(f);
804 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
811 struct io_u *__get_io_u(struct thread_data *td)
813 struct io_u *io_u = NULL;
815 if (!flist_empty(&td->io_u_requeues))
816 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
817 else if (!queue_full(td)) {
818 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
827 assert(io_u->flags & IO_U_F_FREE);
828 io_u->flags &= ~IO_U_F_FREE;
831 flist_del(&io_u->list);
832 flist_add(&io_u->list, &td->io_u_busylist);
840 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
841 * etc. The returned io_u is fully ready to be prepped and submitted.
843 struct io_u *get_io_u(struct thread_data *td)
848 io_u = __get_io_u(td);
850 dprint(FD_IO, "__get_io_u failed\n");
855 * from a requeue, io_u already setup
861 * If using an iolog, grab next piece if any available.
863 if (td->o.read_iolog_file) {
864 if (read_iolog_get(td, io_u))
866 } else if (set_io_u_file(td, io_u)) {
867 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
872 assert(fio_file_open(f));
874 if (io_u->ddir != DDIR_SYNC) {
875 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
876 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
880 f->last_pos = io_u->offset + io_u->buflen;
882 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
883 populate_verify_io_u(td, io_u);
884 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
885 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
889 * Set io data pointers.
891 io_u->xfer_buf = io_u->buf;
892 io_u->xfer_buflen = io_u->buflen;
895 if (!td_io_prep(td, io_u)) {
896 if (!td->o.disable_slat)
897 fio_gettime(&io_u->start_time, NULL);
901 dprint(FD_IO, "get_io_u failed\n");
906 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
908 const char *msg[] = { "read", "write", "sync" };
910 log_err("fio: io_u error");
913 log_err(" on file %s", io_u->file->file_name);
915 log_err(": %s\n", strerror(io_u->error));
917 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
918 io_u->offset, io_u->xfer_buflen);
921 td_verror(td, io_u->error, "io_u error");
924 static void io_completed(struct thread_data *td, struct io_u *io_u,
925 struct io_completion_data *icd)
928 * Older gcc's are too dumb to realize that usec is always used
929 * initialized, silence that warning.
931 unsigned long uninitialized_var(usec);
933 dprint_io_u(io_u, "io complete");
935 assert(io_u->flags & IO_U_F_FLIGHT);
936 io_u->flags &= ~IO_U_F_FLIGHT;
938 if (io_u->ddir == DDIR_SYNC) {
939 td->last_was_sync = 1;
943 td->last_was_sync = 0;
946 unsigned int bytes = io_u->buflen - io_u->resid;
947 const enum fio_ddir idx = io_u->ddir;
950 td->io_blocks[idx]++;
951 td->io_bytes[idx] += bytes;
952 td->this_io_bytes[idx] += bytes;
954 if (ramp_time_over(td)) {
955 if (!td->o.disable_clat || !td->o.disable_bw ||
956 __should_check_rate(td, idx))
957 usec = utime_since(&io_u->issue_time,
960 if (!td->o.disable_clat) {
961 add_clat_sample(td, idx, usec, bytes);
962 io_u_mark_latency(td, usec);
964 if (!td->o.disable_bw)
965 add_bw_sample(td, idx, bytes, &icd->time);
966 if (__should_check_rate(td, idx))
967 td->rate_pending_usleep[idx] += (long) td->rate_usec_cycle[idx] - usec;
968 if (__should_check_rate(td, idx ^ 1))
969 td->rate_pending_usleep[idx ^ 1] -= usec;
972 if (td_write(td) && idx == DDIR_WRITE &&
974 td->o.verify != VERIFY_NONE)
975 log_io_piece(td, io_u);
977 icd->bytes_done[idx] += bytes;
980 ret = io_u->end_io(td, io_u);
981 if (ret && !icd->error)
985 icd->error = io_u->error;
986 io_u_log_error(td, io_u);
990 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
993 if (!td->o.disable_clat || !td->o.disable_bw)
994 fio_gettime(&icd->time, NULL);
999 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1002 static void ios_completed(struct thread_data *td,
1003 struct io_completion_data *icd)
1008 for (i = 0; i < icd->nr; i++) {
1009 io_u = td->io_ops->event(td, i);
1011 io_completed(td, io_u, icd);
1017 * Complete a single io_u for the sync engines.
1019 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1020 unsigned long *bytes)
1022 struct io_completion_data icd;
1024 init_icd(td, &icd, 1);
1025 io_completed(td, io_u, &icd);
1029 td_verror(td, icd.error, "io_u_sync_complete");
1034 bytes[0] += icd.bytes_done[0];
1035 bytes[1] += icd.bytes_done[1];
1042 * Called to complete min_events number of io for the async engines.
1044 int io_u_queued_complete(struct thread_data *td, int min_evts,
1045 unsigned long *bytes)
1047 struct io_completion_data icd;
1048 struct timespec *tvp = NULL;
1050 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1052 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1057 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1059 td_verror(td, -ret, "td_io_getevents");
1064 init_icd(td, &icd, ret);
1065 ios_completed(td, &icd);
1067 td_verror(td, icd.error, "io_u_queued_complete");
1072 bytes[0] += icd.bytes_done[0];
1073 bytes[1] += icd.bytes_done[1];
1080 * Call when io_u is really queued, to update the submission latency.
1082 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1084 if (!td->o.disable_slat) {
1085 unsigned long slat_time;
1087 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1088 add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
1093 * "randomly" fill the buffer contents
1095 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1096 unsigned int max_bs)
1098 long *ptr = io_u->buf;
1100 if (!td->o.zero_buffers) {
1101 while ((void *) ptr - io_u->buf < max_bs) {
1102 *ptr = rand() * GOLDEN_RATIO_PRIME;
1106 memset(ptr, 0, max_bs);