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])
311 * Return the data direction for the next io_u. If the job is a
312 * mixed read/write workload, check the rwmix cycle and switch if
315 static enum fio_ddir get_rw_ddir(struct thread_data *td)
319 * Check if it's time to seed a new data direction.
321 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
322 unsigned long long max_bytes;
326 * Put a top limit on how many bytes we do for
327 * one data direction, to avoid overflowing the
330 ddir = get_rand_ddir(td);
331 max_bytes = td->this_io_bytes[ddir];
333 (td->o.size * td->o.rwmix[ddir] / 100)) {
334 if (!td->rw_end_set[ddir]) {
335 td->rw_end_set[ddir] = 1;
336 fio_gettime(&td->rw_end[ddir], NULL);
342 if (ddir != td->rwmix_ddir)
345 td->rwmix_ddir = ddir;
347 return td->rwmix_ddir;
348 } else if (td_read(td))
354 static void put_file_log(struct thread_data *td, struct fio_file *f)
356 int ret = put_file(td, f);
359 td_verror(td, ret, "file close");
362 void put_io_u(struct thread_data *td, struct io_u *io_u)
364 assert((io_u->flags & IO_U_F_FREE) == 0);
365 io_u->flags |= IO_U_F_FREE;
368 put_file_log(td, io_u->file);
371 flist_del(&io_u->list);
372 flist_add(&io_u->list, &td->io_u_freelist);
376 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
378 struct io_u *__io_u = *io_u;
380 dprint(FD_IO, "requeue %p\n", __io_u);
382 __io_u->flags |= IO_U_F_FREE;
383 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
384 td->io_issues[__io_u->ddir]--;
386 __io_u->flags &= ~IO_U_F_FLIGHT;
388 flist_del(&__io_u->list);
389 flist_add_tail(&__io_u->list, &td->io_u_requeues);
394 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
396 if (td->io_ops->flags & FIO_NOIO)
400 * see if it's time to sync
402 if (td->o.fsync_blocks &&
403 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
404 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
405 io_u->ddir = DDIR_SYNC;
409 io_u->ddir = get_rw_ddir(td);
412 * See if it's time to switch to a new zone
414 if (td->zone_bytes >= td->o.zone_size) {
416 io_u->file->last_pos += td->o.zone_skip;
417 td->io_skip_bytes += td->o.zone_skip;
421 * No log, let the seq/rand engine retrieve the next buflen and
424 if (get_next_offset(td, io_u)) {
425 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
429 io_u->buflen = get_next_buflen(td, io_u);
431 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
435 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
436 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
437 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
438 io_u->buflen, io_u->file->real_file_size);
443 * mark entry before potentially trimming io_u
445 if (td_random(td) && file_randommap(td, io_u->file))
446 mark_random_map(td, io_u);
449 * If using a write iolog, store this entry.
452 dprint_io_u(io_u, "fill_io_u");
453 td->zone_bytes += io_u->buflen;
458 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
487 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
489 __io_u_mark_map(td->ts.io_u_submit, nr);
490 td->ts.total_submit++;
493 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
495 __io_u_mark_map(td->ts.io_u_complete, nr);
496 td->ts.total_complete++;
499 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
503 switch (td->cur_depth) {
525 td->ts.io_u_map[index] += nr;
528 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
565 assert(index < FIO_IO_U_LAT_U_NR);
566 td->ts.io_u_lat_u[index]++;
569 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
610 assert(index < FIO_IO_U_LAT_M_NR);
611 td->ts.io_u_lat_m[index]++;
614 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
617 io_u_mark_lat_usec(td, usec);
619 io_u_mark_lat_msec(td, usec / 1000);
623 * Get next file to service by choosing one at random
625 static struct fio_file *get_next_file_rand(struct thread_data *td, enum fio_file_flags goodf,
626 enum fio_file_flags badf)
632 long r = os_random_long(&td->next_file_state);
635 fno = (unsigned int) ((double) td->o.nr_files
636 * (r / (OS_RAND_MAX + 1.0)));
638 if (fio_file_done(f))
641 if (!fio_file_open(f)) {
644 err = td_io_open_file(td, f);
650 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
651 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
655 td_io_close_file(td, f);
660 * Get next file to service by doing round robin between all available ones
662 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
665 unsigned int old_next_file = td->next_file;
671 f = td->files[td->next_file];
674 if (td->next_file >= td->o.nr_files)
677 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
678 if (fio_file_done(f)) {
683 if (!fio_file_open(f)) {
686 err = td_io_open_file(td, f);
688 dprint(FD_FILE, "error %d on open of %s\n",
696 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
697 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
701 td_io_close_file(td, f);
704 } while (td->next_file != old_next_file);
706 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
710 static struct fio_file *get_next_file(struct thread_data *td)
714 assert(td->o.nr_files <= td->files_index);
716 if (td->nr_done_files >= td->o.nr_files) {
717 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
718 " nr_files=%d\n", td->nr_open_files,
724 f = td->file_service_file;
725 if (f && fio_file_open(f) && !fio_file_closing(f)) {
726 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
728 if (td->file_service_left--)
732 if (td->o.file_service_type == FIO_FSERVICE_RR ||
733 td->o.file_service_type == FIO_FSERVICE_SEQ)
734 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
736 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
738 td->file_service_file = f;
739 td->file_service_left = td->file_service_nr - 1;
741 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
745 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
750 f = get_next_file(td);
757 if (!fill_io_u(td, io_u))
761 td_io_close_file(td, f);
763 fio_file_set_done(f);
765 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
772 struct io_u *__get_io_u(struct thread_data *td)
774 struct io_u *io_u = NULL;
776 if (!flist_empty(&td->io_u_requeues))
777 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
778 else if (!queue_full(td)) {
779 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
788 assert(io_u->flags & IO_U_F_FREE);
789 io_u->flags &= ~IO_U_F_FREE;
792 flist_del(&io_u->list);
793 flist_add(&io_u->list, &td->io_u_busylist);
801 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
802 * etc. The returned io_u is fully ready to be prepped and submitted.
804 struct io_u *get_io_u(struct thread_data *td)
809 io_u = __get_io_u(td);
811 dprint(FD_IO, "__get_io_u failed\n");
816 * from a requeue, io_u already setup
822 * If using an iolog, grab next piece if any available.
824 if (td->o.read_iolog_file) {
825 if (read_iolog_get(td, io_u))
827 } else if (set_io_u_file(td, io_u)) {
828 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
833 assert(fio_file_open(f));
835 if (io_u->ddir != DDIR_SYNC) {
836 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
837 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
841 f->last_pos = io_u->offset + io_u->buflen;
843 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
844 populate_verify_io_u(td, io_u);
845 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
846 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
850 * Set io data pointers.
852 io_u->xfer_buf = io_u->buf;
853 io_u->xfer_buflen = io_u->buflen;
856 if (!td_io_prep(td, io_u)) {
857 if (!td->o.disable_slat)
858 fio_gettime(&io_u->start_time, NULL);
862 dprint(FD_IO, "get_io_u failed\n");
867 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
869 const char *msg[] = { "read", "write", "sync" };
871 log_err("fio: io_u error");
874 log_err(" on file %s", io_u->file->file_name);
876 log_err(": %s\n", strerror(io_u->error));
878 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
879 io_u->offset, io_u->xfer_buflen);
882 td_verror(td, io_u->error, "io_u error");
885 static void io_completed(struct thread_data *td, struct io_u *io_u,
886 struct io_completion_data *icd)
889 * Older gcc's are too dumb to realize that usec is always used
890 * initialized, silence that warning.
892 unsigned long uninitialized_var(usec);
894 dprint_io_u(io_u, "io complete");
896 assert(io_u->flags & IO_U_F_FLIGHT);
897 io_u->flags &= ~IO_U_F_FLIGHT;
899 if (io_u->ddir == DDIR_SYNC) {
900 td->last_was_sync = 1;
904 td->last_was_sync = 0;
907 unsigned int bytes = io_u->buflen - io_u->resid;
908 const enum fio_ddir idx = io_u->ddir;
911 td->io_blocks[idx]++;
912 td->io_bytes[idx] += bytes;
913 td->this_io_bytes[idx] += bytes;
915 if (ramp_time_over(td)) {
916 if (!td->o.disable_clat || !td->o.disable_bw)
917 usec = utime_since(&io_u->issue_time,
920 if (!td->o.disable_clat) {
921 add_clat_sample(td, idx, usec, bytes);
922 io_u_mark_latency(td, usec);
924 if (!td->o.disable_bw)
925 add_bw_sample(td, idx, bytes, &icd->time);
928 if (td_write(td) && idx == DDIR_WRITE &&
930 td->o.verify != VERIFY_NONE)
931 log_io_piece(td, io_u);
933 icd->bytes_done[idx] += bytes;
936 ret = io_u->end_io(td, io_u);
937 if (ret && !icd->error)
941 icd->error = io_u->error;
942 io_u_log_error(td, io_u);
946 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
949 if (!td->o.disable_clat || !td->o.disable_bw)
950 fio_gettime(&icd->time, NULL);
955 icd->bytes_done[0] = icd->bytes_done[1] = 0;
958 static void ios_completed(struct thread_data *td,
959 struct io_completion_data *icd)
964 for (i = 0; i < icd->nr; i++) {
965 io_u = td->io_ops->event(td, i);
967 io_completed(td, io_u, icd);
973 * Complete a single io_u for the sync engines.
975 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
977 struct io_completion_data icd;
979 init_icd(td, &icd, 1);
980 io_completed(td, io_u, &icd);
984 return icd.bytes_done[0] + icd.bytes_done[1];
986 td_verror(td, icd.error, "io_u_sync_complete");
991 * Called to complete min_events number of io for the async engines.
993 long io_u_queued_complete(struct thread_data *td, int min_evts)
995 struct io_completion_data icd;
996 struct timespec *tvp = NULL;
998 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1000 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1005 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1007 td_verror(td, -ret, "td_io_getevents");
1012 init_icd(td, &icd, ret);
1013 ios_completed(td, &icd);
1015 return icd.bytes_done[0] + icd.bytes_done[1];
1017 td_verror(td, icd.error, "io_u_queued_complete");
1022 * Call when io_u is really queued, to update the submission latency.
1024 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1026 if (!td->o.disable_slat) {
1027 unsigned long slat_time;
1029 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1030 add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
1035 * "randomly" fill the buffer contents
1037 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1038 unsigned int max_bs)
1040 long *ptr = io_u->buf;
1042 if (!td->o.zero_buffers) {
1043 while ((void *) ptr - io_u->buf < max_bs) {
1044 *ptr = rand() * GOLDEN_RATIO_PRIME;
1048 memset(ptr, 0, max_bs);