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) OS_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) *
175 (r / (OS_RAND_MAX + 1.0));
176 if (!get_next_free_block(td, f, ddir, b))
179 r = os_random_long(&td->random_state);
183 * that didn't work either, try exhaustive search from the start
185 f->last_free_lookup = 0;
186 return get_next_free_block(td, f, ddir, b);
190 * For random io, generate a random new block and see if it's used. Repeat
191 * until we find a free one. For sequential io, just return the end of
192 * the last io issued.
194 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
196 struct fio_file *f = io_u->file;
197 unsigned long long b;
198 enum fio_ddir ddir = io_u->ddir;
200 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
201 td->ddir_nr = td->o.ddir_nr;
203 if (get_next_rand_offset(td, f, ddir, &b))
206 if (f->last_pos >= f->real_file_size) {
207 if (!td_random(td) ||
208 get_next_rand_offset(td, f, ddir, &b))
211 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
214 io_u->offset = b * td->o.min_bs[ddir];
215 if (io_u->offset >= f->io_size) {
216 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
217 io_u->offset, f->io_size);
221 io_u->offset += f->file_offset;
222 if (io_u->offset >= f->real_file_size) {
223 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
224 io_u->offset, f->real_file_size);
231 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
233 const int ddir = io_u->ddir;
234 unsigned int buflen = buflen; /* silence dumb gcc warning */
237 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
238 buflen = td->o.min_bs[ddir];
240 r = os_random_long(&td->bsrange_state);
241 if (!td->o.bssplit_nr) {
242 buflen = (unsigned int)
243 (1 + (double) (td->o.max_bs[ddir] - 1)
244 * r / (OS_RAND_MAX + 1.0));
249 for (i = 0; i < td->o.bssplit_nr; i++) {
250 struct bssplit *bsp = &td->o.bssplit[i];
254 if (r <= ((LONG_MAX / 100L) * perc))
258 if (!td->o.bs_unaligned) {
259 buflen = (buflen + td->o.min_bs[ddir] - 1)
260 & ~(td->o.min_bs[ddir] - 1);
264 if (io_u->offset + buflen > io_u->file->real_file_size) {
265 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
266 td->o.min_bs[ddir], ddir);
267 buflen = td->o.min_bs[ddir];
273 static void set_rwmix_bytes(struct thread_data *td)
278 * we do time or byte based switch. this is needed because
279 * buffered writes may issue a lot quicker than they complete,
280 * whereas reads do not.
282 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
283 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
286 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
291 r = os_random_long(&td->rwmix_state);
292 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
293 if (v <= td->o.rwmix[DDIR_READ])
300 * Return the data direction for the next io_u. If the job is a
301 * mixed read/write workload, check the rwmix cycle and switch if
304 static enum fio_ddir get_rw_ddir(struct thread_data *td)
308 * Check if it's time to seed a new data direction.
310 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
311 unsigned long long max_bytes;
315 * Put a top limit on how many bytes we do for
316 * one data direction, to avoid overflowing the
319 ddir = get_rand_ddir(td);
320 max_bytes = td->this_io_bytes[ddir];
322 (td->o.size * td->o.rwmix[ddir] / 100)) {
323 if (!td->rw_end_set[ddir]) {
324 td->rw_end_set[ddir] = 1;
325 fio_gettime(&td->rw_end[ddir], NULL);
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, int goodf,
621 long r = os_random_long(&td->next_file_state);
623 fno = (unsigned int) ((double) td->o.nr_files
624 * (r / (OS_RAND_MAX + 1.0)));
626 if (f->flags & FIO_FILE_DONE)
629 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
630 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
637 * Get next file to service by doing round robin between all available ones
639 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
642 unsigned int old_next_file = td->next_file;
646 f = td->files[td->next_file];
649 if (td->next_file >= td->o.nr_files)
652 if (f->flags & FIO_FILE_DONE) {
657 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
661 } while (td->next_file != old_next_file);
663 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
667 static struct fio_file *get_next_file(struct thread_data *td)
671 assert(td->o.nr_files <= td->files_index);
673 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
674 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
675 " nr_files=%d\n", td->nr_open_files,
681 f = td->file_service_file;
682 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
685 if (td->o.file_service_type == FIO_FSERVICE_RR)
686 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
688 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
690 td->file_service_file = f;
691 td->file_service_left = td->file_service_nr - 1;
693 dprint(FD_FILE, "get_next_file: %p\n", f);
697 static struct fio_file *find_next_new_file(struct thread_data *td)
701 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
704 if (td->o.file_service_type == FIO_FSERVICE_RR)
705 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
707 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
712 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
717 f = get_next_file(td);
725 if (!fill_io_u(td, io_u))
729 * optimization to prevent close/open of the same file. This
730 * way we preserve queueing etc.
732 if (td->o.nr_files == 1 && td->o.time_based) {
739 * td_io_close() does a put_file() as well, so no need to
743 td_io_close_file(td, f);
744 f->flags |= FIO_FILE_DONE;
748 * probably not the right place to do this, but see
749 * if we need to open a new file
751 if (td->nr_open_files < td->o.open_files &&
752 td->o.open_files != td->o.nr_files) {
753 f = find_next_new_file(td);
755 if (!f || td_io_open_file(td, f))
766 struct io_u *__get_io_u(struct thread_data *td)
768 struct io_u *io_u = NULL;
770 if (!flist_empty(&td->io_u_requeues))
771 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
772 else if (!queue_full(td)) {
773 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
782 assert(io_u->flags & IO_U_F_FREE);
783 io_u->flags &= ~IO_U_F_FREE;
786 flist_del(&io_u->list);
787 flist_add(&io_u->list, &td->io_u_busylist);
795 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
796 * etc. The returned io_u is fully ready to be prepped and submitted.
798 struct io_u *get_io_u(struct thread_data *td)
803 io_u = __get_io_u(td);
805 dprint(FD_IO, "__get_io_u failed\n");
810 * from a requeue, io_u already setup
816 * If using an iolog, grab next piece if any available.
818 if (td->o.read_iolog_file) {
819 if (read_iolog_get(td, io_u))
821 } else if (set_io_u_file(td, io_u)) {
822 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
827 assert(f->flags & FIO_FILE_OPEN);
829 if (io_u->ddir != DDIR_SYNC) {
830 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
831 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
835 f->last_pos = io_u->offset + io_u->buflen;
837 if (td->o.verify != VERIFY_NONE)
838 populate_verify_io_u(td, io_u);
839 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
840 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
844 * Set io data pointers.
846 io_u->endpos = io_u->offset + io_u->buflen;
847 io_u->xfer_buf = io_u->buf;
848 io_u->xfer_buflen = io_u->buflen;
851 if (!td_io_prep(td, io_u)) {
852 fio_gettime(&io_u->start_time, NULL);
856 dprint(FD_IO, "get_io_u failed\n");
861 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
863 const char *msg[] = { "read", "write", "sync" };
865 log_err("fio: io_u error");
868 log_err(" on file %s", io_u->file->file_name);
870 log_err(": %s\n", strerror(io_u->error));
872 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
873 io_u->offset, io_u->xfer_buflen);
876 td_verror(td, io_u->error, "io_u error");
879 static void io_completed(struct thread_data *td, struct io_u *io_u,
880 struct io_completion_data *icd)
884 dprint_io_u(io_u, "io complete");
886 assert(io_u->flags & IO_U_F_FLIGHT);
887 io_u->flags &= ~IO_U_F_FLIGHT;
889 if (io_u->ddir == DDIR_SYNC) {
890 td->last_was_sync = 1;
894 td->last_was_sync = 0;
897 unsigned int bytes = io_u->buflen - io_u->resid;
898 const enum fio_ddir idx = io_u->ddir;
901 td->io_blocks[idx]++;
902 td->io_bytes[idx] += bytes;
903 td->this_io_bytes[idx] += bytes;
905 usec = utime_since(&io_u->issue_time, &icd->time);
907 add_clat_sample(td, idx, usec);
908 add_bw_sample(td, idx, &icd->time);
909 io_u_mark_latency(td, usec);
911 if (td_write(td) && idx == DDIR_WRITE &&
913 td->o.verify != VERIFY_NONE)
914 log_io_piece(td, io_u);
916 icd->bytes_done[idx] += bytes;
919 ret = io_u->end_io(td, io_u);
920 if (ret && !icd->error)
924 icd->error = io_u->error;
925 io_u_log_error(td, io_u);
929 static void init_icd(struct io_completion_data *icd, int nr)
931 fio_gettime(&icd->time, NULL);
936 icd->bytes_done[0] = icd->bytes_done[1] = 0;
939 static void ios_completed(struct thread_data *td,
940 struct io_completion_data *icd)
945 for (i = 0; i < icd->nr; i++) {
946 io_u = td->io_ops->event(td, i);
948 io_completed(td, io_u, icd);
954 * Complete a single io_u for the sync engines.
956 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
958 struct io_completion_data icd;
961 io_completed(td, io_u, &icd);
965 return icd.bytes_done[0] + icd.bytes_done[1];
967 td_verror(td, icd.error, "io_u_sync_complete");
972 * Called to complete min_events number of io for the async engines.
974 long io_u_queued_complete(struct thread_data *td, int min_evts)
976 struct io_completion_data icd;
977 struct timespec *tvp = NULL;
979 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
981 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
986 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
988 td_verror(td, -ret, "td_io_getevents");
994 ios_completed(td, &icd);
996 return icd.bytes_done[0] + icd.bytes_done[1];
998 td_verror(td, icd.error, "io_u_queued_complete");
1003 * Call when io_u is really queued, to update the submission latency.
1005 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1007 unsigned long slat_time;
1009 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1010 add_slat_sample(td, io_u->ddir, slat_time);
1014 * "randomly" fill the buffer contents
1016 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1017 unsigned int max_bs)
1019 long *ptr = io_u->buf;
1021 if (!td->o.zero_buffers) {
1022 while ((void *) ptr - io_u->buf < max_bs) {
1023 *ptr = rand() * GOLDEN_RATIO_PRIME;
1027 memset(ptr, 0, max_bs);
1030 #ifdef FIO_USE_TIMEOUT
1031 void io_u_set_timeout(struct thread_data *td)
1033 assert(td->cur_depth);
1035 td->timer.it_interval.tv_sec = 0;
1036 td->timer.it_interval.tv_usec = 0;
1037 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1038 td->timer.it_value.tv_usec = 0;
1039 setitimer(ITIMER_REAL, &td->timer, NULL);
1040 fio_gettime(&td->timeout_end, NULL);
1043 static void io_u_dump(struct io_u *io_u)
1045 unsigned long t_start = mtime_since_now(&io_u->start_time);
1046 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1048 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1049 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1050 io_u->xfer_buf, io_u->buflen,
1053 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1056 void io_u_set_timeout(struct thread_data fio_unused *td)
1061 #ifdef FIO_USE_TIMEOUT
1062 static void io_u_timeout_handler(int fio_unused sig)
1064 struct thread_data *td, *__td;
1065 pid_t pid = getpid();
1066 struct flist_head *entry;
1070 log_err("fio: io_u timeout\n");
1073 * TLS would be nice...
1076 for_each_td(__td, i) {
1077 if (__td->pid == pid) {
1084 log_err("fio: io_u timeout, can't find job\n");
1088 if (!td->cur_depth) {
1089 log_err("fio: timeout without pending work?\n");
1093 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1095 flist_for_each(entry, &td->io_u_busylist) {
1096 io_u = flist_entry(entry, struct io_u, list);
1101 td_verror(td, ETIMEDOUT, "io_u timeout");
1106 void io_u_init_timeout(void)
1108 #ifdef FIO_USE_TIMEOUT
1109 signal(SIGALRM, io_u_timeout_handler);