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] & (1UL << 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;
47 unsigned int nr_blocks;
49 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
51 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
53 while (blocks < nr_blocks) {
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 f->file_map[idx] |= (1UL << bit);
73 if ((blocks * min_bs) < io_u->buflen)
74 io_u->buflen = blocks * min_bs;
77 static inline unsigned long long last_block(struct thread_data *td,
81 unsigned long long max_blocks;
82 unsigned long long max_size;
85 * Hmm, should we make sure that ->io_size <= ->real_file_size?
87 max_size = f->io_size;
88 if (max_size > f->real_file_size)
89 max_size = f->real_file_size;
91 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
99 * Return the next free block in the map.
101 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
102 enum fio_ddir ddir, unsigned long long *b)
104 unsigned long long min_bs = td->o.rw_min_bs;
107 i = f->last_free_lookup;
108 *b = (i * BLOCKS_PER_MAP);
109 while ((*b) * min_bs < f->real_file_size) {
110 if (f->file_map[i] != -1UL) {
111 *b += fio_ffz(f->file_map[i]);
112 if (*b > last_block(td, f, ddir))
114 f->last_free_lookup = i;
118 *b += BLOCKS_PER_MAP;
122 dprint(FD_IO, "failed finding a free block\n");
126 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
127 enum fio_ddir ddir, unsigned long long *b)
129 unsigned long long r;
133 r = os_random_long(&td->random_state);
134 dprint(FD_RANDOM, "off rand %llu\n", r);
135 *b = (last_block(td, f, ddir) - 1)
136 * (r / ((unsigned long long) RAND_MAX + 1.0));
139 * if we are not maintaining a random map, we are done.
141 if (!file_randommap(td, f))
145 * calculate map offset and check if it's free
147 if (random_map_free(f, *b))
150 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
155 * we get here, if we didn't suceed in looking up a block. generate
156 * a random start offset into the filemap, and find the first free
161 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
162 if (!get_next_free_block(td, f, ddir, b))
165 r = os_random_long(&td->random_state);
169 * that didn't work either, try exhaustive search from the start
171 f->last_free_lookup = 0;
172 return get_next_free_block(td, f, ddir, b);
176 * For random io, generate a random new block and see if it's used. Repeat
177 * until we find a free one. For sequential io, just return the end of
178 * the last io issued.
180 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
182 struct fio_file *f = io_u->file;
183 unsigned long long b;
184 enum fio_ddir ddir = io_u->ddir;
186 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
187 td->ddir_nr = td->o.ddir_nr;
189 if (get_next_rand_offset(td, f, ddir, &b))
192 if (f->last_pos >= f->real_file_size) {
193 if (!td_random(td) ||
194 get_next_rand_offset(td, f, ddir, &b))
197 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
200 io_u->offset = b * td->o.min_bs[ddir];
201 if (io_u->offset >= f->io_size) {
202 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
203 io_u->offset, f->io_size);
207 io_u->offset += f->file_offset;
208 if (io_u->offset >= f->real_file_size) {
209 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
210 io_u->offset, f->real_file_size);
217 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
219 const int ddir = io_u->ddir;
220 unsigned int buflen = buflen; /* silence dumb gcc warning */
223 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
224 buflen = td->o.min_bs[ddir];
226 r = os_random_long(&td->bsrange_state);
227 if (!td->o.bssplit_nr) {
228 buflen = (unsigned int)
229 (1 + (double) (td->o.max_bs[ddir] - 1)
230 * r / (RAND_MAX + 1.0));
235 for (i = 0; i < td->o.bssplit_nr; i++) {
236 struct bssplit *bsp = &td->o.bssplit[i];
240 if (r <= ((LONG_MAX / 100L) * perc))
244 if (!td->o.bs_unaligned) {
245 buflen = (buflen + td->o.min_bs[ddir] - 1)
246 & ~(td->o.min_bs[ddir] - 1);
250 if (io_u->offset + buflen > io_u->file->real_file_size) {
251 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
252 td->o.min_bs[ddir], ddir);
253 buflen = td->o.min_bs[ddir];
259 static void set_rwmix_bytes(struct thread_data *td)
261 unsigned long issues;
265 * we do time or byte based switch. this is needed because
266 * buffered writes may issue a lot quicker than they complete,
267 * whereas reads do not.
269 issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
270 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
272 td->rwmix_issues = td->io_issues[td->rwmix_ddir]
273 + (issues * ((100 - diff)) / diff);
276 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
281 r = os_random_long(&td->rwmix_state);
282 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
283 if (v < td->o.rwmix[DDIR_READ])
290 * Return the data direction for the next io_u. If the job is a
291 * mixed read/write workload, check the rwmix cycle and switch if
294 static enum fio_ddir get_rw_ddir(struct thread_data *td)
298 * Check if it's time to seed a new data direction.
300 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
301 unsigned long long max_bytes;
305 * Put a top limit on how many bytes we do for
306 * one data direction, to avoid overflowing the
309 ddir = get_rand_ddir(td);
310 max_bytes = td->this_io_bytes[ddir];
312 (td->o.size * td->o.rwmix[ddir] / 100)) {
313 if (!td->rw_end_set[ddir])
314 td->rw_end_set[ddir] = 1;
319 if (ddir != td->rwmix_ddir)
322 td->rwmix_ddir = ddir;
324 return td->rwmix_ddir;
325 } else if (td_read(td))
331 static void put_file_log(struct thread_data *td, struct fio_file *f)
333 int ret = put_file(td, f);
336 td_verror(td, ret, "file close");
339 void put_io_u(struct thread_data *td, struct io_u *io_u)
341 assert((io_u->flags & IO_U_F_FREE) == 0);
342 io_u->flags |= IO_U_F_FREE;
345 put_file_log(td, io_u->file);
348 list_del(&io_u->list);
349 list_add(&io_u->list, &td->io_u_freelist);
353 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
355 struct io_u *__io_u = *io_u;
357 __io_u->flags |= IO_U_F_FREE;
358 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
359 td->io_issues[__io_u->ddir]--;
361 __io_u->flags &= ~IO_U_F_FLIGHT;
363 list_del(&__io_u->list);
364 list_add_tail(&__io_u->list, &td->io_u_requeues);
369 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
371 if (td->io_ops->flags & FIO_NOIO)
375 * see if it's time to sync
377 if (td->o.fsync_blocks &&
378 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
379 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
380 io_u->ddir = DDIR_SYNC;
384 io_u->ddir = get_rw_ddir(td);
387 * See if it's time to switch to a new zone
389 if (td->zone_bytes >= td->o.zone_size) {
391 io_u->file->last_pos += td->o.zone_skip;
392 td->io_skip_bytes += td->o.zone_skip;
396 * No log, let the seq/rand engine retrieve the next buflen and
399 if (get_next_offset(td, io_u)) {
400 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
404 io_u->buflen = get_next_buflen(td, io_u);
406 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
410 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
411 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
412 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
413 io_u->buflen, io_u->file->real_file_size);
418 * mark entry before potentially trimming io_u
420 if (td_random(td) && file_randommap(td, io_u->file))
421 mark_random_map(td, io_u);
424 * If using a write iolog, store this entry.
427 dprint_io_u(io_u, "fill_io_u");
428 td->zone_bytes += io_u->buflen;
433 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
462 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
464 __io_u_mark_map(td->ts.io_u_submit, nr);
465 td->ts.total_submit++;
468 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
470 __io_u_mark_map(td->ts.io_u_complete, nr);
471 td->ts.total_complete++;
474 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
478 switch (td->cur_depth) {
500 td->ts.io_u_map[index] += nr;
503 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
540 assert(index < FIO_IO_U_LAT_U_NR);
541 td->ts.io_u_lat_u[index]++;
544 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
585 assert(index < FIO_IO_U_LAT_M_NR);
586 td->ts.io_u_lat_m[index]++;
589 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
592 io_u_mark_lat_usec(td, usec);
594 io_u_mark_lat_msec(td, usec / 1000);
598 * Get next file to service by choosing one at random
600 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
607 long r = os_random_long(&td->next_file_state);
609 fno = (unsigned int) ((double) td->o.nr_files
610 * (r / (RAND_MAX + 1.0)));
612 if (f->flags & FIO_FILE_DONE)
615 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
616 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
623 * Get next file to service by doing round robin between all available ones
625 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
628 unsigned int old_next_file = td->next_file;
632 f = td->files[td->next_file];
635 if (td->next_file >= td->o.nr_files)
638 if (f->flags & FIO_FILE_DONE) {
643 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
647 } while (td->next_file != old_next_file);
649 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
653 static struct fio_file *get_next_file(struct thread_data *td)
657 assert(td->o.nr_files <= td->files_index);
659 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
660 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
661 " nr_files=%d\n", td->nr_open_files,
667 f = td->file_service_file;
668 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
671 if (td->o.file_service_type == FIO_FSERVICE_RR)
672 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
674 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
676 td->file_service_file = f;
677 td->file_service_left = td->file_service_nr - 1;
679 dprint(FD_FILE, "get_next_file: %p\n", f);
683 static struct fio_file *find_next_new_file(struct thread_data *td)
687 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
690 if (td->o.file_service_type == FIO_FSERVICE_RR)
691 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
693 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
698 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
703 f = get_next_file(td);
711 if (!fill_io_u(td, io_u))
715 * optimization to prevent close/open of the same file. This
716 * way we preserve queueing etc.
718 if (td->o.nr_files == 1 && td->o.time_based) {
725 * td_io_close() does a put_file() as well, so no need to
729 td_io_close_file(td, f);
730 f->flags |= FIO_FILE_DONE;
734 * probably not the right place to do this, but see
735 * if we need to open a new file
737 if (td->nr_open_files < td->o.open_files &&
738 td->o.open_files != td->o.nr_files) {
739 f = find_next_new_file(td);
741 if (!f || td_io_open_file(td, f))
752 struct io_u *__get_io_u(struct thread_data *td)
754 struct io_u *io_u = NULL;
756 if (!list_empty(&td->io_u_requeues))
757 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
758 else if (!queue_full(td)) {
759 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
768 assert(io_u->flags & IO_U_F_FREE);
769 io_u->flags &= ~IO_U_F_FREE;
772 list_del(&io_u->list);
773 list_add(&io_u->list, &td->io_u_busylist);
781 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
782 * etc. The returned io_u is fully ready to be prepped and submitted.
784 struct io_u *get_io_u(struct thread_data *td)
789 io_u = __get_io_u(td);
791 dprint(FD_IO, "__get_io_u failed\n");
796 * from a requeue, io_u already setup
802 * If using an iolog, grab next piece if any available.
804 if (td->o.read_iolog_file) {
805 if (read_iolog_get(td, io_u))
807 } else if (set_io_u_file(td, io_u)) {
808 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
813 assert(f->flags & FIO_FILE_OPEN);
815 if (io_u->ddir != DDIR_SYNC) {
816 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
817 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
821 f->last_pos = io_u->offset + io_u->buflen;
823 if (td->o.verify != VERIFY_NONE)
824 populate_verify_io_u(td, io_u);
828 * Set io data pointers.
830 io_u->endpos = io_u->offset + io_u->buflen;
831 io_u->xfer_buf = io_u->buf;
832 io_u->xfer_buflen = io_u->buflen;
834 if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
835 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
837 if (!td_io_prep(td, io_u)) {
838 fio_gettime(&io_u->start_time, NULL);
842 dprint(FD_IO, "get_io_u failed\n");
847 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
849 const char *msg[] = { "read", "write", "sync" };
851 log_err("fio: io_u error");
854 log_err(" on file %s", io_u->file->file_name);
856 log_err(": %s\n", strerror(io_u->error));
858 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
859 io_u->offset, io_u->xfer_buflen);
862 td_verror(td, io_u->error, "io_u error");
865 static void io_completed(struct thread_data *td, struct io_u *io_u,
866 struct io_completion_data *icd)
870 dprint_io_u(io_u, "io complete");
872 assert(io_u->flags & IO_U_F_FLIGHT);
873 io_u->flags &= ~IO_U_F_FLIGHT;
875 if (io_u->ddir == DDIR_SYNC) {
876 td->last_was_sync = 1;
880 td->last_was_sync = 0;
883 unsigned int bytes = io_u->buflen - io_u->resid;
884 const enum fio_ddir idx = io_u->ddir;
887 td->io_blocks[idx]++;
888 td->io_bytes[idx] += bytes;
889 td->this_io_bytes[idx] += bytes;
891 usec = utime_since(&io_u->issue_time, &icd->time);
893 add_clat_sample(td, idx, usec);
894 add_bw_sample(td, idx, &icd->time);
895 io_u_mark_latency(td, usec);
897 if (td_write(td) && idx == DDIR_WRITE &&
899 td->o.verify != VERIFY_NONE)
900 log_io_piece(td, io_u);
902 icd->bytes_done[idx] += bytes;
905 ret = io_u->end_io(td, io_u);
906 if (ret && !icd->error)
910 icd->error = io_u->error;
911 io_u_log_error(td, io_u);
915 static void init_icd(struct io_completion_data *icd, int nr)
917 fio_gettime(&icd->time, NULL);
922 icd->bytes_done[0] = icd->bytes_done[1] = 0;
925 static void ios_completed(struct thread_data *td,
926 struct io_completion_data *icd)
931 for (i = 0; i < icd->nr; i++) {
932 io_u = td->io_ops->event(td, i);
934 io_completed(td, io_u, icd);
940 * Complete a single io_u for the sync engines.
942 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
944 struct io_completion_data icd;
947 io_completed(td, io_u, &icd);
951 return icd.bytes_done[0] + icd.bytes_done[1];
953 td_verror(td, icd.error, "io_u_sync_complete");
958 * Called to complete min_events number of io for the async engines.
960 long io_u_queued_complete(struct thread_data *td, int min_events)
962 struct io_completion_data icd;
963 struct timespec *tvp = NULL;
965 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
967 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
972 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
974 td_verror(td, -ret, "td_io_getevents");
980 ios_completed(td, &icd);
982 return icd.bytes_done[0] + icd.bytes_done[1];
984 td_verror(td, icd.error, "io_u_queued_complete");
989 * Call when io_u is really queued, to update the submission latency.
991 void io_u_queued(struct thread_data *td, struct io_u *io_u)
993 unsigned long slat_time;
995 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
996 add_slat_sample(td, io_u->ddir, slat_time);
1000 * "randomly" fill the buffer contents
1002 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1003 unsigned int max_bs)
1005 long *ptr = io_u->buf;
1007 if (!td->o.zero_buffers) {
1008 while ((void *) ptr - io_u->buf < max_bs) {
1009 *ptr = rand() * GOLDEN_RATIO_PRIME;
1013 memset(ptr, 0, max_bs);
1016 #ifdef FIO_USE_TIMEOUT
1017 void io_u_set_timeout(struct thread_data *td)
1019 assert(td->cur_depth);
1021 td->timer.it_interval.tv_sec = 0;
1022 td->timer.it_interval.tv_usec = 0;
1023 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1024 td->timer.it_value.tv_usec = 0;
1025 setitimer(ITIMER_REAL, &td->timer, NULL);
1026 fio_gettime(&td->timeout_end, NULL);
1029 static void io_u_dump(struct io_u *io_u)
1031 unsigned long t_start = mtime_since_now(&io_u->start_time);
1032 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1034 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1035 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1036 io_u->xfer_buf, io_u->buflen,
1039 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1042 void io_u_set_timeout(struct thread_data fio_unused *td)
1047 #ifdef FIO_USE_TIMEOUT
1048 static void io_u_timeout_handler(int fio_unused sig)
1050 struct thread_data *td, *__td;
1051 pid_t pid = getpid();
1052 struct list_head *entry;
1056 log_err("fio: io_u timeout\n");
1059 * TLS would be nice...
1062 for_each_td(__td, i) {
1063 if (__td->pid == pid) {
1070 log_err("fio: io_u timeout, can't find job\n");
1074 if (!td->cur_depth) {
1075 log_err("fio: timeout without pending work?\n");
1079 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1081 list_for_each(entry, &td->io_u_busylist) {
1082 io_u = list_entry(entry, struct io_u, list);
1087 td_verror(td, ETIMEDOUT, "io_u timeout");
1092 void io_u_init_timeout(void)
1094 #ifdef FIO_USE_TIMEOUT
1095 signal(SIGALRM, io_u_timeout_handler);