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)) {
66 idx = RAND_MAP_IDX(f, block);
67 bit = RAND_MAP_BIT(f, block);
69 fio_assert(td, idx < f->num_maps);
71 this_blocks = nr_blocks;
72 if (this_blocks + bit > BLOCKS_PER_MAP)
73 this_blocks = BLOCKS_PER_MAP - bit;
75 if (this_blocks == BLOCKS_PER_MAP)
78 mask = ((1U << this_blocks) - 1) << bit;
80 fio_assert(td, !(f->file_map[idx] & mask));
81 f->file_map[idx] |= mask;
82 nr_blocks -= this_blocks;
83 blocks += this_blocks;
87 if ((blocks * min_bs) < io_u->buflen)
88 io_u->buflen = blocks * min_bs;
91 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
94 unsigned long long max_blocks;
95 unsigned long long max_size;
98 * Hmm, should we make sure that ->io_size <= ->real_file_size?
100 max_size = f->io_size;
101 if (max_size > f->real_file_size)
102 max_size = f->real_file_size;
104 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
112 * Return the next free block in the map.
114 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
115 enum fio_ddir ddir, unsigned long long *b)
117 unsigned long long min_bs = td->o.rw_min_bs;
120 i = f->last_free_lookup;
121 *b = (i * BLOCKS_PER_MAP);
122 while ((*b) * min_bs < f->real_file_size) {
123 if (f->file_map[i] != (unsigned int) -1) {
124 *b += ffz(f->file_map[i]);
125 if (*b > last_block(td, f, ddir))
127 f->last_free_lookup = i;
131 *b += BLOCKS_PER_MAP;
135 dprint(FD_IO, "failed finding a free block\n");
139 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
140 enum fio_ddir ddir, unsigned long long *b)
142 unsigned long long r;
146 r = os_random_long(&td->random_state);
147 dprint(FD_RANDOM, "off rand %llu\n", r);
148 *b = (last_block(td, f, ddir) - 1)
149 * (r / ((unsigned long long) RAND_MAX + 1.0));
152 * if we are not maintaining a random map, we are done.
154 if (!file_randommap(td, f))
158 * calculate map offset and check if it's free
160 if (random_map_free(f, *b))
163 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
168 * we get here, if we didn't suceed in looking up a block. generate
169 * a random start offset into the filemap, and find the first free
174 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
175 if (!get_next_free_block(td, f, ddir, b))
178 r = os_random_long(&td->random_state);
182 * that didn't work either, try exhaustive search from the start
184 f->last_free_lookup = 0;
185 return get_next_free_block(td, f, ddir, b);
189 * For random io, generate a random new block and see if it's used. Repeat
190 * until we find a free one. For sequential io, just return the end of
191 * the last io issued.
193 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
195 struct fio_file *f = io_u->file;
196 unsigned long long b;
197 enum fio_ddir ddir = io_u->ddir;
199 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
200 td->ddir_nr = td->o.ddir_nr;
202 if (get_next_rand_offset(td, f, ddir, &b))
205 if (f->last_pos >= f->real_file_size) {
206 if (!td_random(td) ||
207 get_next_rand_offset(td, f, ddir, &b))
210 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
213 io_u->offset = b * td->o.min_bs[ddir];
214 if (io_u->offset >= f->io_size) {
215 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
216 io_u->offset, f->io_size);
220 io_u->offset += f->file_offset;
221 if (io_u->offset >= f->real_file_size) {
222 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
223 io_u->offset, f->real_file_size);
230 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
232 const int ddir = io_u->ddir;
233 unsigned int buflen = buflen; /* silence dumb gcc warning */
236 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
237 buflen = td->o.min_bs[ddir];
239 r = os_random_long(&td->bsrange_state);
240 if (!td->o.bssplit_nr) {
241 buflen = (unsigned int)
242 (1 + (double) (td->o.max_bs[ddir] - 1)
243 * r / (RAND_MAX + 1.0));
248 for (i = 0; i < td->o.bssplit_nr; i++) {
249 struct bssplit *bsp = &td->o.bssplit[i];
253 if (r <= ((LONG_MAX / 100L) * perc))
257 if (!td->o.bs_unaligned) {
258 buflen = (buflen + td->o.min_bs[ddir] - 1)
259 & ~(td->o.min_bs[ddir] - 1);
263 if (io_u->offset + buflen > io_u->file->real_file_size) {
264 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
265 td->o.min_bs[ddir], ddir);
266 buflen = td->o.min_bs[ddir];
272 static void set_rwmix_bytes(struct thread_data *td)
277 * we do time or byte based switch. this is needed because
278 * buffered writes may issue a lot quicker than they complete,
279 * whereas reads do not.
281 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
282 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
285 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
290 r = os_random_long(&td->rwmix_state);
291 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
292 if (v <= td->o.rwmix[DDIR_READ])
299 * Return the data direction for the next io_u. If the job is a
300 * mixed read/write workload, check the rwmix cycle and switch if
303 static enum fio_ddir get_rw_ddir(struct thread_data *td)
307 * Check if it's time to seed a new data direction.
309 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
310 unsigned long long max_bytes;
314 * Put a top limit on how many bytes we do for
315 * one data direction, to avoid overflowing the
318 ddir = get_rand_ddir(td);
319 max_bytes = td->this_io_bytes[ddir];
321 (td->o.size * td->o.rwmix[ddir] / 100)) {
322 if (!td->rw_end_set[ddir])
323 td->rw_end_set[ddir] = 1;
328 if (ddir != td->rwmix_ddir)
331 td->rwmix_ddir = ddir;
333 return td->rwmix_ddir;
334 } else if (td_read(td))
340 static void put_file_log(struct thread_data *td, struct fio_file *f)
342 int ret = put_file(td, f);
345 td_verror(td, ret, "file close");
348 void put_io_u(struct thread_data *td, struct io_u *io_u)
350 assert((io_u->flags & IO_U_F_FREE) == 0);
351 io_u->flags |= IO_U_F_FREE;
354 put_file_log(td, io_u->file);
357 list_del(&io_u->list);
358 list_add(&io_u->list, &td->io_u_freelist);
362 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
364 struct io_u *__io_u = *io_u;
366 dprint(FD_IO, "requeue %p\n", __io_u);
368 __io_u->flags |= IO_U_F_FREE;
369 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
370 td->io_issues[__io_u->ddir]--;
372 __io_u->flags &= ~IO_U_F_FLIGHT;
374 list_del(&__io_u->list);
375 list_add_tail(&__io_u->list, &td->io_u_requeues);
380 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
382 if (td->io_ops->flags & FIO_NOIO)
386 * see if it's time to sync
388 if (td->o.fsync_blocks &&
389 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
390 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
391 io_u->ddir = DDIR_SYNC;
395 io_u->ddir = get_rw_ddir(td);
398 * See if it's time to switch to a new zone
400 if (td->zone_bytes >= td->o.zone_size) {
402 io_u->file->last_pos += td->o.zone_skip;
403 td->io_skip_bytes += td->o.zone_skip;
407 * No log, let the seq/rand engine retrieve the next buflen and
410 if (get_next_offset(td, io_u)) {
411 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
415 io_u->buflen = get_next_buflen(td, io_u);
417 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
421 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
422 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
423 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
424 io_u->buflen, io_u->file->real_file_size);
429 * mark entry before potentially trimming io_u
431 if (td_random(td) && file_randommap(td, io_u->file))
432 mark_random_map(td, io_u);
435 * If using a write iolog, store this entry.
438 dprint_io_u(io_u, "fill_io_u");
439 td->zone_bytes += io_u->buflen;
444 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
473 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
475 __io_u_mark_map(td->ts.io_u_submit, nr);
476 td->ts.total_submit++;
479 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
481 __io_u_mark_map(td->ts.io_u_complete, nr);
482 td->ts.total_complete++;
485 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
489 switch (td->cur_depth) {
511 td->ts.io_u_map[index] += nr;
514 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
551 assert(index < FIO_IO_U_LAT_U_NR);
552 td->ts.io_u_lat_u[index]++;
555 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
596 assert(index < FIO_IO_U_LAT_M_NR);
597 td->ts.io_u_lat_m[index]++;
600 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
603 io_u_mark_lat_usec(td, usec);
605 io_u_mark_lat_msec(td, usec / 1000);
609 * Get next file to service by choosing one at random
611 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
618 long r = os_random_long(&td->next_file_state);
620 fno = (unsigned int) ((double) td->o.nr_files
621 * (r / (RAND_MAX + 1.0)));
623 if (f->flags & FIO_FILE_DONE)
626 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
627 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
634 * Get next file to service by doing round robin between all available ones
636 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
639 unsigned int old_next_file = td->next_file;
643 f = td->files[td->next_file];
646 if (td->next_file >= td->o.nr_files)
649 if (f->flags & FIO_FILE_DONE) {
654 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
658 } while (td->next_file != old_next_file);
660 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
664 static struct fio_file *get_next_file(struct thread_data *td)
668 assert(td->o.nr_files <= td->files_index);
670 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
671 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
672 " nr_files=%d\n", td->nr_open_files,
678 f = td->file_service_file;
679 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
682 if (td->o.file_service_type == FIO_FSERVICE_RR)
683 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
685 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
687 td->file_service_file = f;
688 td->file_service_left = td->file_service_nr - 1;
690 dprint(FD_FILE, "get_next_file: %p\n", f);
694 static struct fio_file *find_next_new_file(struct thread_data *td)
698 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
701 if (td->o.file_service_type == FIO_FSERVICE_RR)
702 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
704 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
709 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
714 f = get_next_file(td);
722 if (!fill_io_u(td, io_u))
726 * optimization to prevent close/open of the same file. This
727 * way we preserve queueing etc.
729 if (td->o.nr_files == 1 && td->o.time_based) {
736 * td_io_close() does a put_file() as well, so no need to
740 td_io_close_file(td, f);
741 f->flags |= FIO_FILE_DONE;
745 * probably not the right place to do this, but see
746 * if we need to open a new file
748 if (td->nr_open_files < td->o.open_files &&
749 td->o.open_files != td->o.nr_files) {
750 f = find_next_new_file(td);
752 if (!f || td_io_open_file(td, f))
763 struct io_u *__get_io_u(struct thread_data *td)
765 struct io_u *io_u = NULL;
767 if (!list_empty(&td->io_u_requeues))
768 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
769 else if (!queue_full(td)) {
770 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
779 assert(io_u->flags & IO_U_F_FREE);
780 io_u->flags &= ~IO_U_F_FREE;
783 list_del(&io_u->list);
784 list_add(&io_u->list, &td->io_u_busylist);
792 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
793 * etc. The returned io_u is fully ready to be prepped and submitted.
795 struct io_u *get_io_u(struct thread_data *td)
800 io_u = __get_io_u(td);
802 dprint(FD_IO, "__get_io_u failed\n");
807 * from a requeue, io_u already setup
813 * If using an iolog, grab next piece if any available.
815 if (td->o.read_iolog_file) {
816 if (read_iolog_get(td, io_u))
818 } else if (set_io_u_file(td, io_u)) {
819 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
824 assert(f->flags & FIO_FILE_OPEN);
826 if (io_u->ddir != DDIR_SYNC) {
827 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
828 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
832 f->last_pos = io_u->offset + io_u->buflen;
834 if (td->o.verify != VERIFY_NONE)
835 populate_verify_io_u(td, io_u);
836 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
837 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
841 * Set io data pointers.
843 io_u->endpos = io_u->offset + io_u->buflen;
844 io_u->xfer_buf = io_u->buf;
845 io_u->xfer_buflen = io_u->buflen;
848 if (!td_io_prep(td, io_u)) {
849 fio_gettime(&io_u->start_time, NULL);
853 dprint(FD_IO, "get_io_u failed\n");
858 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
860 const char *msg[] = { "read", "write", "sync" };
862 log_err("fio: io_u error");
865 log_err(" on file %s", io_u->file->file_name);
867 log_err(": %s\n", strerror(io_u->error));
869 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
870 io_u->offset, io_u->xfer_buflen);
873 td_verror(td, io_u->error, "io_u error");
876 static void io_completed(struct thread_data *td, struct io_u *io_u,
877 struct io_completion_data *icd)
881 dprint_io_u(io_u, "io complete");
883 assert(io_u->flags & IO_U_F_FLIGHT);
884 io_u->flags &= ~IO_U_F_FLIGHT;
886 if (io_u->ddir == DDIR_SYNC) {
887 td->last_was_sync = 1;
891 td->last_was_sync = 0;
894 unsigned int bytes = io_u->buflen - io_u->resid;
895 const enum fio_ddir idx = io_u->ddir;
898 td->io_blocks[idx]++;
899 td->io_bytes[idx] += bytes;
900 td->this_io_bytes[idx] += bytes;
902 usec = utime_since(&io_u->issue_time, &icd->time);
904 add_clat_sample(td, idx, usec);
905 add_bw_sample(td, idx, &icd->time);
906 io_u_mark_latency(td, usec);
908 if (td_write(td) && idx == DDIR_WRITE &&
910 td->o.verify != VERIFY_NONE)
911 log_io_piece(td, io_u);
913 icd->bytes_done[idx] += bytes;
916 ret = io_u->end_io(td, io_u);
917 if (ret && !icd->error)
921 icd->error = io_u->error;
922 io_u_log_error(td, io_u);
926 static void init_icd(struct io_completion_data *icd, int nr)
928 fio_gettime(&icd->time, NULL);
933 icd->bytes_done[0] = icd->bytes_done[1] = 0;
936 static void ios_completed(struct thread_data *td,
937 struct io_completion_data *icd)
942 for (i = 0; i < icd->nr; i++) {
943 io_u = td->io_ops->event(td, i);
945 io_completed(td, io_u, icd);
951 * Complete a single io_u for the sync engines.
953 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
955 struct io_completion_data icd;
958 io_completed(td, io_u, &icd);
962 return icd.bytes_done[0] + icd.bytes_done[1];
964 td_verror(td, icd.error, "io_u_sync_complete");
969 * Called to complete min_events number of io for the async engines.
971 long io_u_queued_complete(struct thread_data *td, int min_events)
973 struct io_completion_data icd;
974 struct timespec *tvp = NULL;
976 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
978 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
983 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
985 td_verror(td, -ret, "td_io_getevents");
991 ios_completed(td, &icd);
993 return icd.bytes_done[0] + icd.bytes_done[1];
995 td_verror(td, icd.error, "io_u_queued_complete");
1000 * Call when io_u is really queued, to update the submission latency.
1002 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1004 unsigned long slat_time;
1006 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1007 add_slat_sample(td, io_u->ddir, slat_time);
1011 * "randomly" fill the buffer contents
1013 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1014 unsigned int max_bs)
1016 long *ptr = io_u->buf;
1018 if (!td->o.zero_buffers) {
1019 while ((void *) ptr - io_u->buf < max_bs) {
1020 *ptr = rand() * GOLDEN_RATIO_PRIME;
1024 memset(ptr, 0, max_bs);
1027 #ifdef FIO_USE_TIMEOUT
1028 void io_u_set_timeout(struct thread_data *td)
1030 assert(td->cur_depth);
1032 td->timer.it_interval.tv_sec = 0;
1033 td->timer.it_interval.tv_usec = 0;
1034 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1035 td->timer.it_value.tv_usec = 0;
1036 setitimer(ITIMER_REAL, &td->timer, NULL);
1037 fio_gettime(&td->timeout_end, NULL);
1040 static void io_u_dump(struct io_u *io_u)
1042 unsigned long t_start = mtime_since_now(&io_u->start_time);
1043 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1045 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1046 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1047 io_u->xfer_buf, io_u->buflen,
1050 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1053 void io_u_set_timeout(struct thread_data fio_unused *td)
1058 #ifdef FIO_USE_TIMEOUT
1059 static void io_u_timeout_handler(int fio_unused sig)
1061 struct thread_data *td, *__td;
1062 pid_t pid = getpid();
1063 struct list_head *entry;
1067 log_err("fio: io_u timeout\n");
1070 * TLS would be nice...
1073 for_each_td(__td, i) {
1074 if (__td->pid == pid) {
1081 log_err("fio: io_u timeout, can't find job\n");
1085 if (!td->cur_depth) {
1086 log_err("fio: timeout without pending work?\n");
1090 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1092 list_for_each(entry, &td->io_u_busylist) {
1093 io_u = list_entry(entry, struct io_u, list);
1098 td_verror(td, ETIMEDOUT, "io_u timeout");
1103 void io_u_init_timeout(void)
1105 #ifdef FIO_USE_TIMEOUT
1106 signal(SIGALRM, io_u_timeout_handler);