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 f->file_map[idx] |= mask;
81 nr_blocks -= this_blocks;
82 blocks += this_blocks;
86 if ((blocks * min_bs) < io_u->buflen)
87 io_u->buflen = blocks * min_bs;
90 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
93 unsigned long long max_blocks;
94 unsigned long long max_size;
97 * Hmm, should we make sure that ->io_size <= ->real_file_size?
99 max_size = f->io_size;
100 if (max_size > f->real_file_size)
101 max_size = f->real_file_size;
103 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
111 * Return the next free block in the map.
113 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
114 enum fio_ddir ddir, unsigned long long *b)
116 unsigned long long min_bs = td->o.rw_min_bs;
119 i = f->last_free_lookup;
120 *b = (i * BLOCKS_PER_MAP);
121 while ((*b) * min_bs < f->real_file_size) {
122 if (f->file_map[i] != (unsigned int) -1) {
123 *b += ffz(f->file_map[i]);
124 if (*b > last_block(td, f, ddir))
126 f->last_free_lookup = i;
130 *b += BLOCKS_PER_MAP;
134 dprint(FD_IO, "failed finding a free block\n");
138 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
139 enum fio_ddir ddir, unsigned long long *b)
141 unsigned long long r;
145 r = os_random_long(&td->random_state);
146 dprint(FD_RANDOM, "off rand %llu\n", r);
147 *b = (last_block(td, f, ddir) - 1)
148 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
151 * if we are not maintaining a random map, we are done.
153 if (!file_randommap(td, f))
157 * calculate map offset and check if it's free
159 if (random_map_free(f, *b))
162 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
167 * we get here, if we didn't suceed in looking up a block. generate
168 * a random start offset into the filemap, and find the first free
173 f->last_free_lookup = (f->num_maps - 1) *
174 (r / (OS_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 inline int is_power_of_2(unsigned int val)
232 return (val != 0 && ((val & (val - 1)) == 0));
235 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
237 const int ddir = io_u->ddir;
238 unsigned int uninitialized_var(buflen);
239 unsigned int minbs, maxbs;
242 minbs = td->o.min_bs[ddir];
243 maxbs = td->o.max_bs[ddir];
248 r = os_random_long(&td->bsrange_state);
249 if (!td->o.bssplit_nr) {
250 buflen = 1 + (unsigned int) ((double) maxbs *
251 (r / (OS_RAND_MAX + 1.0)));
258 for (i = 0; i < td->o.bssplit_nr; i++) {
259 struct bssplit *bsp = &td->o.bssplit[i];
263 if (r <= ((OS_RAND_MAX / 100L) * perc))
267 if (!td->o.bs_unaligned && is_power_of_2(minbs))
268 buflen = (buflen + minbs - 1) & ~(minbs - 1);
271 if (io_u->offset + buflen > io_u->file->real_file_size) {
272 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
280 static void set_rwmix_bytes(struct thread_data *td)
285 * we do time or byte based switch. this is needed because
286 * buffered writes may issue a lot quicker than they complete,
287 * whereas reads do not.
289 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
290 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
293 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
298 r = os_random_long(&td->rwmix_state);
299 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
300 if (v <= td->o.rwmix[DDIR_READ])
307 * Return the data direction for the next io_u. If the job is a
308 * mixed read/write workload, check the rwmix cycle and switch if
311 static enum fio_ddir get_rw_ddir(struct thread_data *td)
315 * Check if it's time to seed a new data direction.
317 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
318 unsigned long long max_bytes;
322 * Put a top limit on how many bytes we do for
323 * one data direction, to avoid overflowing the
326 ddir = get_rand_ddir(td);
327 max_bytes = td->this_io_bytes[ddir];
329 (td->o.size * td->o.rwmix[ddir] / 100)) {
330 if (!td->rw_end_set[ddir]) {
331 td->rw_end_set[ddir] = 1;
332 fio_gettime(&td->rw_end[ddir], NULL);
338 if (ddir != td->rwmix_ddir)
341 td->rwmix_ddir = ddir;
343 return td->rwmix_ddir;
344 } else if (td_read(td))
350 static void put_file_log(struct thread_data *td, struct fio_file *f)
352 int ret = put_file(td, f);
355 td_verror(td, ret, "file close");
358 void put_io_u(struct thread_data *td, struct io_u *io_u)
360 assert((io_u->flags & IO_U_F_FREE) == 0);
361 io_u->flags |= IO_U_F_FREE;
364 put_file_log(td, io_u->file);
367 flist_del(&io_u->list);
368 flist_add(&io_u->list, &td->io_u_freelist);
372 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
374 struct io_u *__io_u = *io_u;
376 dprint(FD_IO, "requeue %p\n", __io_u);
378 __io_u->flags |= IO_U_F_FREE;
379 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
380 td->io_issues[__io_u->ddir]--;
382 __io_u->flags &= ~IO_U_F_FLIGHT;
384 flist_del(&__io_u->list);
385 flist_add_tail(&__io_u->list, &td->io_u_requeues);
390 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
392 if (td->io_ops->flags & FIO_NOIO)
396 * see if it's time to sync
398 if (td->o.fsync_blocks &&
399 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
400 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
401 io_u->ddir = DDIR_SYNC;
405 io_u->ddir = get_rw_ddir(td);
408 * See if it's time to switch to a new zone
410 if (td->zone_bytes >= td->o.zone_size) {
412 io_u->file->last_pos += td->o.zone_skip;
413 td->io_skip_bytes += td->o.zone_skip;
417 * No log, let the seq/rand engine retrieve the next buflen and
420 if (get_next_offset(td, io_u)) {
421 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
425 io_u->buflen = get_next_buflen(td, io_u);
427 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
431 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
432 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
433 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
434 io_u->buflen, io_u->file->real_file_size);
439 * mark entry before potentially trimming io_u
441 if (td_random(td) && file_randommap(td, io_u->file))
442 mark_random_map(td, io_u);
445 * If using a write iolog, store this entry.
448 dprint_io_u(io_u, "fill_io_u");
449 td->zone_bytes += io_u->buflen;
454 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
483 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
485 __io_u_mark_map(td->ts.io_u_submit, nr);
486 td->ts.total_submit++;
489 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
491 __io_u_mark_map(td->ts.io_u_complete, nr);
492 td->ts.total_complete++;
495 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
499 switch (td->cur_depth) {
521 td->ts.io_u_map[index] += nr;
524 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
561 assert(index < FIO_IO_U_LAT_U_NR);
562 td->ts.io_u_lat_u[index]++;
565 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
606 assert(index < FIO_IO_U_LAT_M_NR);
607 td->ts.io_u_lat_m[index]++;
610 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
613 io_u_mark_lat_usec(td, usec);
615 io_u_mark_lat_msec(td, usec / 1000);
619 * Get next file to service by choosing one at random
621 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
628 long r = os_random_long(&td->next_file_state);
630 fno = (unsigned int) ((double) td->o.nr_files
631 * (r / (OS_RAND_MAX + 1.0)));
633 if (f->flags & FIO_FILE_DONE)
636 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
637 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
644 * Get next file to service by doing round robin between all available ones
646 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
649 unsigned int old_next_file = td->next_file;
653 f = td->files[td->next_file];
656 if (td->next_file >= td->o.nr_files)
659 if (f->flags & FIO_FILE_DONE) {
664 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
668 } while (td->next_file != old_next_file);
670 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
674 static struct fio_file *get_next_file(struct thread_data *td)
678 assert(td->o.nr_files <= td->files_index);
680 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
681 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
682 " nr_files=%d\n", td->nr_open_files,
688 f = td->file_service_file;
689 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
692 if (td->o.file_service_type == FIO_FSERVICE_RR)
693 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
695 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
697 td->file_service_file = f;
698 td->file_service_left = td->file_service_nr - 1;
700 dprint(FD_FILE, "get_next_file: %p\n", f);
704 static struct fio_file *find_next_new_file(struct thread_data *td)
708 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
711 if (td->o.file_service_type == FIO_FSERVICE_RR)
712 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
714 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
719 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
724 f = get_next_file(td);
732 if (!fill_io_u(td, io_u))
736 * optimization to prevent close/open of the same file. This
737 * way we preserve queueing etc.
739 if (td->o.nr_files == 1 && td->o.time_based) {
746 * td_io_close() does a put_file() as well, so no need to
750 td_io_close_file(td, f);
751 f->flags |= FIO_FILE_DONE;
755 * probably not the right place to do this, but see
756 * if we need to open a new file
758 if (td->nr_open_files < td->o.open_files &&
759 td->o.open_files != td->o.nr_files) {
760 f = find_next_new_file(td);
762 if (!f || td_io_open_file(td, f))
773 struct io_u *__get_io_u(struct thread_data *td)
775 struct io_u *io_u = NULL;
777 if (!flist_empty(&td->io_u_requeues))
778 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
779 else if (!queue_full(td)) {
780 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
789 assert(io_u->flags & IO_U_F_FREE);
790 io_u->flags &= ~IO_U_F_FREE;
793 flist_del(&io_u->list);
794 flist_add(&io_u->list, &td->io_u_busylist);
802 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
803 * etc. The returned io_u is fully ready to be prepped and submitted.
805 struct io_u *get_io_u(struct thread_data *td)
810 io_u = __get_io_u(td);
812 dprint(FD_IO, "__get_io_u failed\n");
817 * from a requeue, io_u already setup
823 * If using an iolog, grab next piece if any available.
825 if (td->o.read_iolog_file) {
826 if (read_iolog_get(td, io_u))
828 } else if (set_io_u_file(td, io_u)) {
829 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
834 assert(f->flags & FIO_FILE_OPEN);
836 if (io_u->ddir != DDIR_SYNC) {
837 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
838 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
842 f->last_pos = io_u->offset + io_u->buflen;
844 if (td->o.verify != VERIFY_NONE)
845 populate_verify_io_u(td, io_u);
846 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
847 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
851 * Set io data pointers.
853 io_u->endpos = io_u->offset + io_u->buflen;
854 io_u->xfer_buf = io_u->buf;
855 io_u->xfer_buflen = io_u->buflen;
858 if (!td_io_prep(td, io_u)) {
859 fio_gettime(&io_u->start_time, NULL);
863 dprint(FD_IO, "get_io_u failed\n");
868 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
870 const char *msg[] = { "read", "write", "sync" };
872 log_err("fio: io_u error");
875 log_err(" on file %s", io_u->file->file_name);
877 log_err(": %s\n", strerror(io_u->error));
879 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
880 io_u->offset, io_u->xfer_buflen);
883 td_verror(td, io_u->error, "io_u error");
886 static void io_completed(struct thread_data *td, struct io_u *io_u,
887 struct io_completion_data *icd)
891 dprint_io_u(io_u, "io complete");
893 assert(io_u->flags & IO_U_F_FLIGHT);
894 io_u->flags &= ~IO_U_F_FLIGHT;
896 if (io_u->ddir == DDIR_SYNC) {
897 td->last_was_sync = 1;
901 td->last_was_sync = 0;
904 unsigned int bytes = io_u->buflen - io_u->resid;
905 const enum fio_ddir idx = io_u->ddir;
908 ramp_done = ramp_time_over(td);
911 td->io_blocks[idx]++;
912 td->io_bytes[idx] += bytes;
913 td->this_io_bytes[idx] += bytes;
915 usec = utime_since(&io_u->issue_time, &icd->time);
917 add_clat_sample(td, idx, usec);
918 add_bw_sample(td, idx, &icd->time);
919 io_u_mark_latency(td, usec);
922 if (td_write(td) && idx == DDIR_WRITE &&
924 td->o.verify != VERIFY_NONE)
925 log_io_piece(td, io_u);
928 icd->bytes_done[idx] += bytes;
931 ret = io_u->end_io(td, io_u);
932 if (ret && !icd->error)
936 icd->error = io_u->error;
937 io_u_log_error(td, io_u);
941 static void init_icd(struct io_completion_data *icd, int nr)
943 fio_gettime(&icd->time, NULL);
948 icd->bytes_done[0] = icd->bytes_done[1] = 0;
951 static void ios_completed(struct thread_data *td,
952 struct io_completion_data *icd)
957 for (i = 0; i < icd->nr; i++) {
958 io_u = td->io_ops->event(td, i);
960 io_completed(td, io_u, icd);
966 * Complete a single io_u for the sync engines.
968 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
970 struct io_completion_data icd;
973 io_completed(td, io_u, &icd);
977 return icd.bytes_done[0] + icd.bytes_done[1];
979 td_verror(td, icd.error, "io_u_sync_complete");
984 * Called to complete min_events number of io for the async engines.
986 long io_u_queued_complete(struct thread_data *td, int min_evts)
988 struct io_completion_data icd;
989 struct timespec *tvp = NULL;
991 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
993 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
998 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1000 td_verror(td, -ret, "td_io_getevents");
1005 init_icd(&icd, ret);
1006 ios_completed(td, &icd);
1008 return icd.bytes_done[0] + icd.bytes_done[1];
1010 td_verror(td, icd.error, "io_u_queued_complete");
1015 * Call when io_u is really queued, to update the submission latency.
1017 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1019 unsigned long slat_time;
1021 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1022 add_slat_sample(td, io_u->ddir, slat_time);
1026 * "randomly" fill the buffer contents
1028 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1029 unsigned int max_bs)
1031 long *ptr = io_u->buf;
1033 if (!td->o.zero_buffers) {
1034 while ((void *) ptr - io_u->buf < max_bs) {
1035 *ptr = rand() * GOLDEN_RATIO_PRIME;
1039 memset(ptr, 0, max_bs);
1042 #ifdef FIO_USE_TIMEOUT
1043 void io_u_set_timeout(struct thread_data *td)
1045 assert(td->cur_depth);
1047 td->timer.it_interval.tv_sec = 0;
1048 td->timer.it_interval.tv_usec = 0;
1049 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1050 td->timer.it_value.tv_usec = 0;
1051 setitimer(ITIMER_REAL, &td->timer, NULL);
1052 fio_gettime(&td->timeout_end, NULL);
1055 static void io_u_dump(struct io_u *io_u)
1057 unsigned long t_start = mtime_since_now(&io_u->start_time);
1058 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1060 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1061 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1062 io_u->xfer_buf, io_u->buflen,
1065 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1068 void io_u_set_timeout(struct thread_data fio_unused *td)
1073 #ifdef FIO_USE_TIMEOUT
1074 static void io_u_timeout_handler(int fio_unused sig)
1076 struct thread_data *td, *__td;
1077 pid_t pid = getpid();
1078 struct flist_head *entry;
1082 log_err("fio: io_u timeout\n");
1085 * TLS would be nice...
1088 for_each_td(__td, i) {
1089 if (__td->pid == pid) {
1096 log_err("fio: io_u timeout, can't find job\n");
1100 if (!td->cur_depth) {
1101 log_err("fio: timeout without pending work?\n");
1105 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1107 flist_for_each(entry, &td->io_u_busylist) {
1108 io_u = flist_entry(entry, struct io_u, list);
1113 td_verror(td, ETIMEDOUT, "io_u timeout");
1118 void io_u_init_timeout(void)
1120 #ifdef FIO_USE_TIMEOUT
1121 signal(SIGALRM, io_u_timeout_handler);