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) {
325 * Put a top limit on how many bytes we do for
326 * one data direction, to avoid overflowing the
329 ddir = get_rand_ddir(td);
331 if (ddir != td->rwmix_ddir)
334 td->rwmix_ddir = ddir;
336 return td->rwmix_ddir;
337 } else if (td_read(td))
343 static void put_file_log(struct thread_data *td, struct fio_file *f)
345 int ret = put_file(td, f);
348 td_verror(td, ret, "file close");
351 void put_io_u(struct thread_data *td, struct io_u *io_u)
353 assert((io_u->flags & IO_U_F_FREE) == 0);
354 io_u->flags |= IO_U_F_FREE;
357 put_file_log(td, io_u->file);
360 flist_del(&io_u->list);
361 flist_add(&io_u->list, &td->io_u_freelist);
365 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
367 struct io_u *__io_u = *io_u;
369 dprint(FD_IO, "requeue %p\n", __io_u);
371 __io_u->flags |= IO_U_F_FREE;
372 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
373 td->io_issues[__io_u->ddir]--;
375 __io_u->flags &= ~IO_U_F_FLIGHT;
377 flist_del(&__io_u->list);
378 flist_add_tail(&__io_u->list, &td->io_u_requeues);
383 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
385 if (td->io_ops->flags & FIO_NOIO)
389 * see if it's time to sync
391 if (td->o.fsync_blocks &&
392 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
393 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
394 io_u->ddir = DDIR_SYNC;
398 io_u->ddir = get_rw_ddir(td);
401 * See if it's time to switch to a new zone
403 if (td->zone_bytes >= td->o.zone_size) {
405 io_u->file->last_pos += td->o.zone_skip;
406 td->io_skip_bytes += td->o.zone_skip;
410 * No log, let the seq/rand engine retrieve the next buflen and
413 if (get_next_offset(td, io_u)) {
414 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
418 io_u->buflen = get_next_buflen(td, io_u);
420 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
424 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
425 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
426 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
427 io_u->buflen, io_u->file->real_file_size);
432 * mark entry before potentially trimming io_u
434 if (td_random(td) && file_randommap(td, io_u->file))
435 mark_random_map(td, io_u);
438 * If using a write iolog, store this entry.
441 dprint_io_u(io_u, "fill_io_u");
442 td->zone_bytes += io_u->buflen;
447 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
476 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
478 __io_u_mark_map(td->ts.io_u_submit, nr);
479 td->ts.total_submit++;
482 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
484 __io_u_mark_map(td->ts.io_u_complete, nr);
485 td->ts.total_complete++;
488 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
492 switch (td->cur_depth) {
514 td->ts.io_u_map[index] += nr;
517 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
554 assert(index < FIO_IO_U_LAT_U_NR);
555 td->ts.io_u_lat_u[index]++;
558 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
599 assert(index < FIO_IO_U_LAT_M_NR);
600 td->ts.io_u_lat_m[index]++;
603 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
606 io_u_mark_lat_usec(td, usec);
608 io_u_mark_lat_msec(td, usec / 1000);
612 * Get next file to service by choosing one at random
614 static struct fio_file *get_next_file_rand(struct thread_data *td, enum fio_file_flags goodf,
615 enum fio_file_flags badf)
621 long r = os_random_long(&td->next_file_state);
624 fno = (unsigned int) ((double) td->o.nr_files
625 * (r / (OS_RAND_MAX + 1.0)));
627 if (fio_file_done(f))
630 if (!fio_file_open(f)) {
633 err = td_io_open_file(td, f);
639 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
640 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
644 td_io_close_file(td, f);
649 * Get next file to service by doing round robin between all available ones
651 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
654 unsigned int old_next_file = td->next_file;
660 f = td->files[td->next_file];
663 if (td->next_file >= td->o.nr_files)
666 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
667 if (fio_file_done(f)) {
672 if (!fio_file_open(f)) {
675 err = td_io_open_file(td, f);
677 dprint(FD_FILE, "error %d on open of %s\n",
685 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
686 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
690 td_io_close_file(td, f);
693 } while (td->next_file != old_next_file);
695 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
699 static struct fio_file *get_next_file(struct thread_data *td)
703 assert(td->o.nr_files <= td->files_index);
705 if (td->nr_done_files >= td->o.nr_files) {
706 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
707 " nr_files=%d\n", td->nr_open_files,
713 f = td->file_service_file;
714 if (f && fio_file_open(f) && !fio_file_closing(f)) {
715 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
717 if (td->file_service_left--)
721 if (td->o.file_service_type == FIO_FSERVICE_RR ||
722 td->o.file_service_type == FIO_FSERVICE_SEQ)
723 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
725 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
727 td->file_service_file = f;
728 td->file_service_left = td->file_service_nr - 1;
730 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
734 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
739 f = get_next_file(td);
746 if (!fill_io_u(td, io_u))
750 td_io_close_file(td, f);
752 fio_file_set_done(f);
754 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
761 struct io_u *__get_io_u(struct thread_data *td)
763 struct io_u *io_u = NULL;
765 if (!flist_empty(&td->io_u_requeues))
766 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
767 else if (!queue_full(td)) {
768 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
777 assert(io_u->flags & IO_U_F_FREE);
778 io_u->flags &= ~IO_U_F_FREE;
781 flist_del(&io_u->list);
782 flist_add(&io_u->list, &td->io_u_busylist);
790 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
791 * etc. The returned io_u is fully ready to be prepped and submitted.
793 struct io_u *get_io_u(struct thread_data *td)
798 io_u = __get_io_u(td);
800 dprint(FD_IO, "__get_io_u failed\n");
805 * from a requeue, io_u already setup
811 * If using an iolog, grab next piece if any available.
813 if (td->o.read_iolog_file) {
814 if (read_iolog_get(td, io_u))
816 } else if (set_io_u_file(td, io_u)) {
817 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
822 assert(fio_file_open(f));
824 if (io_u->ddir != DDIR_SYNC) {
825 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
826 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
830 f->last_pos = io_u->offset + io_u->buflen;
832 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
833 populate_verify_io_u(td, io_u);
834 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
835 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
839 * Set io data pointers.
841 io_u->xfer_buf = io_u->buf;
842 io_u->xfer_buflen = io_u->buflen;
845 if (!td_io_prep(td, io_u)) {
846 if (!td->o.disable_slat)
847 fio_gettime(&io_u->start_time, NULL);
851 dprint(FD_IO, "get_io_u failed\n");
856 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
858 const char *msg[] = { "read", "write", "sync" };
860 log_err("fio: io_u error");
863 log_err(" on file %s", io_u->file->file_name);
865 log_err(": %s\n", strerror(io_u->error));
867 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
868 io_u->offset, io_u->xfer_buflen);
871 td_verror(td, io_u->error, "io_u error");
874 static void io_completed(struct thread_data *td, struct io_u *io_u,
875 struct io_completion_data *icd)
878 * Older gcc's are too dumb to realize that usec is always used
879 * initialized, silence that warning.
881 unsigned long uninitialized_var(usec);
883 dprint_io_u(io_u, "io complete");
885 assert(io_u->flags & IO_U_F_FLIGHT);
886 io_u->flags &= ~IO_U_F_FLIGHT;
888 if (io_u->ddir == DDIR_SYNC) {
889 td->last_was_sync = 1;
893 td->last_was_sync = 0;
896 unsigned int bytes = io_u->buflen - io_u->resid;
897 const enum fio_ddir idx = io_u->ddir;
900 td->io_blocks[idx]++;
901 td->io_bytes[idx] += bytes;
902 td->this_io_bytes[idx] += bytes;
904 if (ramp_time_over(td)) {
905 if (!td->o.disable_clat || !td->o.disable_bw)
906 usec = utime_since(&io_u->issue_time,
909 if (!td->o.disable_clat) {
910 add_clat_sample(td, idx, usec, bytes);
911 io_u_mark_latency(td, usec);
913 if (!td->o.disable_bw)
914 add_bw_sample(td, idx, bytes, &icd->time);
917 if (td_write(td) && idx == DDIR_WRITE &&
919 td->o.verify != VERIFY_NONE)
920 log_io_piece(td, io_u);
922 icd->bytes_done[idx] += bytes;
925 ret = io_u->end_io(td, io_u);
926 if (ret && !icd->error)
930 icd->error = io_u->error;
931 io_u_log_error(td, io_u);
935 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
938 if (!td->o.disable_clat || !td->o.disable_bw)
939 fio_gettime(&icd->time, NULL);
944 icd->bytes_done[0] = icd->bytes_done[1] = 0;
947 static void ios_completed(struct thread_data *td,
948 struct io_completion_data *icd)
953 for (i = 0; i < icd->nr; i++) {
954 io_u = td->io_ops->event(td, i);
956 io_completed(td, io_u, icd);
962 * Complete a single io_u for the sync engines.
964 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
966 struct io_completion_data icd;
968 init_icd(td, &icd, 1);
969 io_completed(td, io_u, &icd);
973 return icd.bytes_done[0] + icd.bytes_done[1];
975 td_verror(td, icd.error, "io_u_sync_complete");
980 * Called to complete min_events number of io for the async engines.
982 long io_u_queued_complete(struct thread_data *td, int min_evts)
984 struct io_completion_data icd;
985 struct timespec *tvp = NULL;
987 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
989 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
994 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
996 td_verror(td, -ret, "td_io_getevents");
1001 init_icd(td, &icd, ret);
1002 ios_completed(td, &icd);
1004 return icd.bytes_done[0] + icd.bytes_done[1];
1006 td_verror(td, icd.error, "io_u_queued_complete");
1011 * Call when io_u is really queued, to update the submission latency.
1013 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1015 if (!td->o.disable_slat) {
1016 unsigned long slat_time;
1018 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1019 add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
1024 * "randomly" fill the buffer contents
1026 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1027 unsigned int max_bs)
1029 long *ptr = io_u->buf;
1031 if (!td->o.zero_buffers) {
1032 while ((void *) ptr - io_u->buf < max_bs) {
1033 *ptr = rand() * GOLDEN_RATIO_PRIME;
1037 memset(ptr, 0, max_bs);