12 * Change this define to play with the timeout handling
14 #undef FIO_USE_TIMEOUT
16 struct io_completion_data {
19 int error; /* output */
20 unsigned long bytes_done[2]; /* output */
21 struct timeval time; /* output */
25 * The ->file_map[] 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 unsigned long long block)
30 unsigned int idx = RAND_MAP_IDX(f, block);
31 unsigned int bit = RAND_MAP_BIT(f, block);
33 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
35 return (f->file_map[idx] & (1 << bit)) == 0;
39 * Mark a given offset as used in the map.
41 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
43 unsigned int min_bs = td->o.rw_min_bs;
44 struct fio_file *f = io_u->file;
45 unsigned long long block;
46 unsigned int blocks, nr_blocks;
48 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
49 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
53 unsigned int this_blocks, mask;
54 unsigned int idx, bit;
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
60 if ((td->o.ddir_nr == 1) && !random_map_free(f, block))
63 idx = RAND_MAP_IDX(f, block);
64 bit = RAND_MAP_BIT(f, block);
66 fio_assert(td, idx < f->num_maps);
68 this_blocks = nr_blocks;
69 if (this_blocks + bit > BLOCKS_PER_MAP)
70 this_blocks = BLOCKS_PER_MAP - bit;
72 if (this_blocks == BLOCKS_PER_MAP)
75 mask = ((1U << this_blocks) - 1) << bit;
77 fio_assert(td, !(f->file_map[idx] & mask));
78 f->file_map[idx] |= mask;
79 nr_blocks -= this_blocks;
80 blocks += this_blocks;
83 if ((blocks * min_bs) < io_u->buflen)
84 io_u->buflen = blocks * min_bs;
87 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
90 unsigned long long max_blocks;
91 unsigned long long max_size;
94 * Hmm, should we make sure that ->io_size <= ->real_file_size?
96 max_size = f->io_size;
97 if (max_size > f->real_file_size)
98 max_size = f->real_file_size;
100 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
108 * Return the next free block in the map.
110 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
111 enum fio_ddir ddir, unsigned long long *b)
113 unsigned long long min_bs = td->o.rw_min_bs;
116 i = f->last_free_lookup;
117 *b = (i * BLOCKS_PER_MAP);
118 while ((*b) * min_bs < f->real_file_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) 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) * (r / (RAND_MAX+1.0));
171 if (!get_next_free_block(td, f, ddir, b))
174 r = os_random_long(&td->random_state);
178 * that didn't work either, try exhaustive search from the start
180 f->last_free_lookup = 0;
181 return get_next_free_block(td, f, ddir, b);
185 * For random io, generate a random new block and see if it's used. Repeat
186 * until we find a free one. For sequential io, just return the end of
187 * the last io issued.
189 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
191 struct fio_file *f = io_u->file;
192 unsigned long long b;
193 enum fio_ddir ddir = io_u->ddir;
195 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
196 td->ddir_nr = td->o.ddir_nr;
198 if (get_next_rand_offset(td, f, ddir, &b))
201 if (f->last_pos >= f->real_file_size) {
202 if (!td_random(td) ||
203 get_next_rand_offset(td, f, ddir, &b))
206 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
209 io_u->offset = b * td->o.min_bs[ddir];
210 if (io_u->offset >= f->io_size) {
211 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
212 io_u->offset, f->io_size);
216 io_u->offset += f->file_offset;
217 if (io_u->offset >= f->real_file_size) {
218 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
219 io_u->offset, f->real_file_size);
226 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
228 const int ddir = io_u->ddir;
229 unsigned int buflen = buflen; /* silence dumb gcc warning */
232 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
233 buflen = td->o.min_bs[ddir];
235 r = os_random_long(&td->bsrange_state);
236 if (!td->o.bssplit_nr) {
237 buflen = (unsigned int)
238 (1 + (double) (td->o.max_bs[ddir] - 1)
239 * r / (RAND_MAX + 1.0));
244 for (i = 0; i < td->o.bssplit_nr; i++) {
245 struct bssplit *bsp = &td->o.bssplit[i];
249 if (r <= ((LONG_MAX / 100L) * perc))
253 if (!td->o.bs_unaligned) {
254 buflen = (buflen + td->o.min_bs[ddir] - 1)
255 & ~(td->o.min_bs[ddir] - 1);
259 if (io_u->offset + buflen > io_u->file->real_file_size) {
260 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
261 td->o.min_bs[ddir], ddir);
262 buflen = td->o.min_bs[ddir];
268 static void set_rwmix_bytes(struct thread_data *td)
273 * we do time or byte based switch. this is needed because
274 * buffered writes may issue a lot quicker than they complete,
275 * whereas reads do not.
277 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
278 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
281 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
286 r = os_random_long(&td->rwmix_state);
287 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
288 if (v <= td->o.rwmix[DDIR_READ])
295 * Return the data direction for the next io_u. If the job is a
296 * mixed read/write workload, check the rwmix cycle and switch if
299 static enum fio_ddir get_rw_ddir(struct thread_data *td)
303 * Check if it's time to seed a new data direction.
305 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
306 unsigned long long max_bytes;
310 * Put a top limit on how many bytes we do for
311 * one data direction, to avoid overflowing the
314 ddir = get_rand_ddir(td);
315 max_bytes = td->this_io_bytes[ddir];
317 (td->o.size * td->o.rwmix[ddir] / 100)) {
318 if (!td->rw_end_set[ddir])
319 td->rw_end_set[ddir] = 1;
324 if (ddir != td->rwmix_ddir)
327 td->rwmix_ddir = ddir;
329 return td->rwmix_ddir;
330 } else if (td_read(td))
336 static void put_file_log(struct thread_data *td, struct fio_file *f)
338 int ret = put_file(td, f);
341 td_verror(td, ret, "file close");
344 void put_io_u(struct thread_data *td, struct io_u *io_u)
346 assert((io_u->flags & IO_U_F_FREE) == 0);
347 io_u->flags |= IO_U_F_FREE;
350 put_file_log(td, io_u->file);
353 list_del(&io_u->list);
354 list_add(&io_u->list, &td->io_u_freelist);
358 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
360 struct io_u *__io_u = *io_u;
362 dprint(FD_IO, "requeue %p\n", __io_u);
364 __io_u->flags |= IO_U_F_FREE;
365 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
366 td->io_issues[__io_u->ddir]--;
368 __io_u->flags &= ~IO_U_F_FLIGHT;
370 list_del(&__io_u->list);
371 list_add_tail(&__io_u->list, &td->io_u_requeues);
376 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
378 if (td->io_ops->flags & FIO_NOIO)
382 * see if it's time to sync
384 if (td->o.fsync_blocks &&
385 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
386 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
387 io_u->ddir = DDIR_SYNC;
391 io_u->ddir = get_rw_ddir(td);
394 * See if it's time to switch to a new zone
396 if (td->zone_bytes >= td->o.zone_size) {
398 io_u->file->last_pos += td->o.zone_skip;
399 td->io_skip_bytes += td->o.zone_skip;
403 * No log, let the seq/rand engine retrieve the next buflen and
406 if (get_next_offset(td, io_u)) {
407 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
411 io_u->buflen = get_next_buflen(td, io_u);
413 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
417 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
418 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
419 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
420 io_u->buflen, io_u->file->real_file_size);
425 * mark entry before potentially trimming io_u
427 if (td_random(td) && file_randommap(td, io_u->file))
428 mark_random_map(td, io_u);
431 * If using a write iolog, store this entry.
434 dprint_io_u(io_u, "fill_io_u");
435 td->zone_bytes += io_u->buflen;
440 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
469 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
471 __io_u_mark_map(td->ts.io_u_submit, nr);
472 td->ts.total_submit++;
475 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
477 __io_u_mark_map(td->ts.io_u_complete, nr);
478 td->ts.total_complete++;
481 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
485 switch (td->cur_depth) {
507 td->ts.io_u_map[index] += nr;
510 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
547 assert(index < FIO_IO_U_LAT_U_NR);
548 td->ts.io_u_lat_u[index]++;
551 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
592 assert(index < FIO_IO_U_LAT_M_NR);
593 td->ts.io_u_lat_m[index]++;
596 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
599 io_u_mark_lat_usec(td, usec);
601 io_u_mark_lat_msec(td, usec / 1000);
605 * Get next file to service by choosing one at random
607 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
614 long r = os_random_long(&td->next_file_state);
616 fno = (unsigned int) ((double) td->o.nr_files
617 * (r / (RAND_MAX + 1.0)));
619 if (f->flags & FIO_FILE_DONE)
622 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
623 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
630 * Get next file to service by doing round robin between all available ones
632 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
635 unsigned int old_next_file = td->next_file;
639 f = td->files[td->next_file];
642 if (td->next_file >= td->o.nr_files)
645 if (f->flags & FIO_FILE_DONE) {
650 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
654 } while (td->next_file != old_next_file);
656 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
660 static struct fio_file *get_next_file(struct thread_data *td)
664 assert(td->o.nr_files <= td->files_index);
666 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
667 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
668 " nr_files=%d\n", td->nr_open_files,
674 f = td->file_service_file;
675 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
678 if (td->o.file_service_type == FIO_FSERVICE_RR)
679 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
681 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
683 td->file_service_file = f;
684 td->file_service_left = td->file_service_nr - 1;
686 dprint(FD_FILE, "get_next_file: %p\n", f);
690 static struct fio_file *find_next_new_file(struct thread_data *td)
694 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
697 if (td->o.file_service_type == FIO_FSERVICE_RR)
698 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
700 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
705 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
710 f = get_next_file(td);
718 if (!fill_io_u(td, io_u))
722 * optimization to prevent close/open of the same file. This
723 * way we preserve queueing etc.
725 if (td->o.nr_files == 1 && td->o.time_based) {
732 * td_io_close() does a put_file() as well, so no need to
736 td_io_close_file(td, f);
737 f->flags |= FIO_FILE_DONE;
741 * probably not the right place to do this, but see
742 * if we need to open a new file
744 if (td->nr_open_files < td->o.open_files &&
745 td->o.open_files != td->o.nr_files) {
746 f = find_next_new_file(td);
748 if (!f || td_io_open_file(td, f))
759 struct io_u *__get_io_u(struct thread_data *td)
761 struct io_u *io_u = NULL;
763 if (!list_empty(&td->io_u_requeues))
764 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
765 else if (!queue_full(td)) {
766 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
775 assert(io_u->flags & IO_U_F_FREE);
776 io_u->flags &= ~IO_U_F_FREE;
779 list_del(&io_u->list);
780 list_add(&io_u->list, &td->io_u_busylist);
788 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
789 * etc. The returned io_u is fully ready to be prepped and submitted.
791 struct io_u *get_io_u(struct thread_data *td)
796 io_u = __get_io_u(td);
798 dprint(FD_IO, "__get_io_u failed\n");
803 * from a requeue, io_u already setup
809 * If using an iolog, grab next piece if any available.
811 if (td->o.read_iolog_file) {
812 if (read_iolog_get(td, io_u))
814 } else if (set_io_u_file(td, io_u)) {
815 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
820 assert(f->flags & FIO_FILE_OPEN);
822 if (io_u->ddir != DDIR_SYNC) {
823 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
824 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
828 f->last_pos = io_u->offset + io_u->buflen;
830 if (td->o.verify != VERIFY_NONE)
831 populate_verify_io_u(td, io_u);
832 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
833 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
837 * Set io data pointers.
839 io_u->endpos = io_u->offset + io_u->buflen;
840 io_u->xfer_buf = io_u->buf;
841 io_u->xfer_buflen = io_u->buflen;
844 if (!td_io_prep(td, io_u)) {
845 fio_gettime(&io_u->start_time, NULL);
849 dprint(FD_IO, "get_io_u failed\n");
854 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
856 const char *msg[] = { "read", "write", "sync" };
858 log_err("fio: io_u error");
861 log_err(" on file %s", io_u->file->file_name);
863 log_err(": %s\n", strerror(io_u->error));
865 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
866 io_u->offset, io_u->xfer_buflen);
869 td_verror(td, io_u->error, "io_u error");
872 static void io_completed(struct thread_data *td, struct io_u *io_u,
873 struct io_completion_data *icd)
877 dprint_io_u(io_u, "io complete");
879 assert(io_u->flags & IO_U_F_FLIGHT);
880 io_u->flags &= ~IO_U_F_FLIGHT;
882 if (io_u->ddir == DDIR_SYNC) {
883 td->last_was_sync = 1;
887 td->last_was_sync = 0;
890 unsigned int bytes = io_u->buflen - io_u->resid;
891 const enum fio_ddir idx = io_u->ddir;
894 td->io_blocks[idx]++;
895 td->io_bytes[idx] += bytes;
896 td->this_io_bytes[idx] += bytes;
898 usec = utime_since(&io_u->issue_time, &icd->time);
900 add_clat_sample(td, idx, usec);
901 add_bw_sample(td, idx, &icd->time);
902 io_u_mark_latency(td, usec);
904 if (td_write(td) && idx == DDIR_WRITE &&
906 td->o.verify != VERIFY_NONE)
907 log_io_piece(td, io_u);
909 icd->bytes_done[idx] += bytes;
912 ret = io_u->end_io(td, io_u);
913 if (ret && !icd->error)
917 icd->error = io_u->error;
918 io_u_log_error(td, io_u);
922 static void init_icd(struct io_completion_data *icd, int nr)
924 fio_gettime(&icd->time, NULL);
929 icd->bytes_done[0] = icd->bytes_done[1] = 0;
932 static void ios_completed(struct thread_data *td,
933 struct io_completion_data *icd)
938 for (i = 0; i < icd->nr; i++) {
939 io_u = td->io_ops->event(td, i);
941 io_completed(td, io_u, icd);
947 * Complete a single io_u for the sync engines.
949 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
951 struct io_completion_data icd;
954 io_completed(td, io_u, &icd);
958 return icd.bytes_done[0] + icd.bytes_done[1];
960 td_verror(td, icd.error, "io_u_sync_complete");
965 * Called to complete min_events number of io for the async engines.
967 long io_u_queued_complete(struct thread_data *td, int min_events)
969 struct io_completion_data icd;
970 struct timespec *tvp = NULL;
972 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
974 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
979 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
981 td_verror(td, -ret, "td_io_getevents");
987 ios_completed(td, &icd);
989 return icd.bytes_done[0] + icd.bytes_done[1];
991 td_verror(td, icd.error, "io_u_queued_complete");
996 * Call when io_u is really queued, to update the submission latency.
998 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1000 unsigned long slat_time;
1002 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1003 add_slat_sample(td, io_u->ddir, slat_time);
1007 * "randomly" fill the buffer contents
1009 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1010 unsigned int max_bs)
1012 long *ptr = io_u->buf;
1014 if (!td->o.zero_buffers) {
1015 while ((void *) ptr - io_u->buf < max_bs) {
1016 *ptr = rand() * GOLDEN_RATIO_PRIME;
1020 memset(ptr, 0, max_bs);
1023 #ifdef FIO_USE_TIMEOUT
1024 void io_u_set_timeout(struct thread_data *td)
1026 assert(td->cur_depth);
1028 td->timer.it_interval.tv_sec = 0;
1029 td->timer.it_interval.tv_usec = 0;
1030 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1031 td->timer.it_value.tv_usec = 0;
1032 setitimer(ITIMER_REAL, &td->timer, NULL);
1033 fio_gettime(&td->timeout_end, NULL);
1036 static void io_u_dump(struct io_u *io_u)
1038 unsigned long t_start = mtime_since_now(&io_u->start_time);
1039 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1041 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1042 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1043 io_u->xfer_buf, io_u->buflen,
1046 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1049 void io_u_set_timeout(struct thread_data fio_unused *td)
1054 #ifdef FIO_USE_TIMEOUT
1055 static void io_u_timeout_handler(int fio_unused sig)
1057 struct thread_data *td, *__td;
1058 pid_t pid = getpid();
1059 struct list_head *entry;
1063 log_err("fio: io_u timeout\n");
1066 * TLS would be nice...
1069 for_each_td(__td, i) {
1070 if (__td->pid == pid) {
1077 log_err("fio: io_u timeout, can't find job\n");
1081 if (!td->cur_depth) {
1082 log_err("fio: timeout without pending work?\n");
1086 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1088 list_for_each(entry, &td->io_u_busylist) {
1089 io_u = list_entry(entry, struct io_u, list);
1094 td_verror(td, ETIMEDOUT, "io_u timeout");
1099 void io_u_init_timeout(void)
1101 #ifdef FIO_USE_TIMEOUT
1102 signal(SIGALRM, io_u_timeout_handler);