11 * Change this define to play with the timeout handling
13 #undef FIO_USE_TIMEOUT
15 struct io_completion_data {
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 static int random_map_free(struct thread_data *td, struct fio_file *f,
28 const unsigned long long block)
30 unsigned int idx = RAND_MAP_IDX(td, f, block);
31 unsigned int bit = RAND_MAP_BIT(td, 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(td, f, block))
63 idx = RAND_MAP_IDX(td, f, block);
64 bit = RAND_MAP_BIT(td, 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;
83 max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
87 return max_blocks - 1;
91 * Return the next free block in the map.
93 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
94 enum fio_ddir ddir, unsigned long long *b)
96 unsigned long long min_bs = td->o.rw_min_bs;
99 i = f->last_free_lookup;
100 *b = (i * BLOCKS_PER_MAP);
101 while ((*b) * min_bs < f->real_file_size) {
102 if (f->file_map[i] != -1UL) {
103 *b += fio_ffz(f->file_map[i]);
104 if (*b > last_block(td, f, ddir))
106 f->last_free_lookup = i;
110 *b += BLOCKS_PER_MAP;
114 dprint(FD_IO, "failed finding a free block\n");
118 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
119 enum fio_ddir ddir, unsigned long long *b)
121 unsigned long long r;
125 r = os_random_long(&td->random_state);
126 dprint(FD_RANDOM, "off rand %llu\n", r);
127 *b = (last_block(td, f, ddir) - 1) * (r / ((unsigned long long) RAND_MAX + 1.0));
130 * if we are not maintaining a random map, we are done.
132 if (td->o.norandommap)
136 * calculate map offset and check if it's free
138 if (random_map_free(td, f, *b))
141 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
146 * we get here, if we didn't suceed in looking up a block. generate
147 * a random start offset into the filemap, and find the first free
152 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
153 if (!get_next_free_block(td, f, ddir, b))
156 r = os_random_long(&td->random_state);
160 * that didn't work either, try exhaustive search from the start
162 f->last_free_lookup = 0;
163 return get_next_free_block(td, f, ddir, b);
167 * For random io, generate a random new block and see if it's used. Repeat
168 * until we find a free one. For sequential io, just return the end of
169 * the last io issued.
171 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
173 struct fio_file *f = io_u->file;
174 unsigned long long b;
175 enum fio_ddir ddir = io_u->ddir;
177 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
178 td->ddir_nr = td->o.ddir_nr;
180 if (get_next_rand_offset(td, f, ddir, &b))
183 if (f->last_pos >= f->real_file_size) {
184 if (!td_random(td) ||
185 get_next_rand_offset(td, f, ddir, &b))
188 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
191 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
192 if (io_u->offset >= f->real_file_size) {
193 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
194 io_u->offset, f->real_file_size);
201 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
203 const int ddir = io_u->ddir;
207 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
208 buflen = td->o.min_bs[ddir];
210 r = os_random_long(&td->bsrange_state);
211 if (!td->o.bssplit_nr)
212 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
217 for (i = 0; i < td->o.bssplit_nr; i++) {
218 struct bssplit *bsp = &td->o.bssplit[i];
222 if (r <= ((LONG_MAX / 100L) * perc))
226 if (!td->o.bs_unaligned)
227 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
230 if (io_u->offset + buflen > io_u->file->real_file_size) {
231 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
232 td->o.min_bs[ddir], ddir);
233 buflen = td->o.min_bs[ddir];
239 static void set_rwmix_bytes(struct thread_data *td)
241 unsigned long long rbytes;
245 * we do time or byte based switch. this is needed because
246 * buffered writes may issue a lot quicker than they complete,
247 * whereas reads do not.
249 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
250 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
252 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
255 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
260 r = os_random_long(&td->rwmix_state);
261 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
262 if (v < td->o.rwmix[DDIR_READ])
269 * Return the data direction for the next io_u. If the job is a
270 * mixed read/write workload, check the rwmix cycle and switch if
273 static enum fio_ddir get_rw_ddir(struct thread_data *td)
277 unsigned long elapsed;
280 fio_gettime(&now, NULL);
281 elapsed = mtime_since_now(&td->rwmix_switch);
284 * if this is the first cycle, make it shorter
286 cycle = td->o.rwmixcycle;
287 if (!td->rwmix_bytes)
291 * Check if it's time to seed a new data direction.
293 if (elapsed >= cycle ||
294 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
295 unsigned long long max_bytes;
299 * Put a top limit on how many bytes we do for
300 * one data direction, to avoid overflowing the
303 ddir = get_rand_ddir(td);
304 max_bytes = td->this_io_bytes[ddir];
305 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
306 if (!td->rw_end_set[ddir]) {
307 td->rw_end_set[ddir] = 1;
308 memcpy(&td->rw_end[ddir], &now, sizeof(now));
313 if (ddir != td->rwmix_ddir)
316 td->rwmix_ddir = ddir;
317 memcpy(&td->rwmix_switch, &now, sizeof(now));
319 return td->rwmix_ddir;
320 } else if (td_read(td))
326 void put_io_u(struct thread_data *td, struct io_u *io_u)
328 assert((io_u->flags & IO_U_F_FREE) == 0);
329 io_u->flags |= IO_U_F_FREE;
332 int ret = put_file(td, io_u->file);
335 td_verror(td, ret, "file close");
339 list_del(&io_u->list);
340 list_add(&io_u->list, &td->io_u_freelist);
344 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
346 struct io_u *__io_u = *io_u;
348 __io_u->flags |= IO_U_F_FREE;
349 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
350 td->io_issues[__io_u->ddir]--;
352 __io_u->flags &= ~IO_U_F_FLIGHT;
354 list_del(&__io_u->list);
355 list_add_tail(&__io_u->list, &td->io_u_requeues);
360 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
362 if (td->io_ops->flags & FIO_NOIO)
366 * see if it's time to sync
368 if (td->o.fsync_blocks &&
369 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
370 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
371 io_u->ddir = DDIR_SYNC;
375 io_u->ddir = get_rw_ddir(td);
378 * See if it's time to switch to a new zone
380 if (td->zone_bytes >= td->o.zone_size) {
382 io_u->file->last_pos += td->o.zone_skip;
383 td->io_skip_bytes += td->o.zone_skip;
387 * No log, let the seq/rand engine retrieve the next buflen and
390 if (get_next_offset(td, io_u)) {
391 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
395 io_u->buflen = get_next_buflen(td, io_u);
397 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
401 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
402 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
403 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
404 io_u->buflen, io_u->file->real_file_size);
409 * mark entry before potentially trimming io_u
411 if (td_random(td) && !td->o.norandommap)
412 mark_random_map(td, io_u);
415 * If using a write iolog, store this entry.
418 dprint_io_u(io_u, "fill_io_u");
419 td->zone_bytes += io_u->buflen;
424 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
428 if (io_u->ddir == DDIR_SYNC)
431 switch (td->cur_depth) {
453 td->ts.io_u_map[index]++;
454 td->ts.total_io_u[io_u->ddir]++;
457 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
494 assert(index < FIO_IO_U_LAT_U_NR);
495 td->ts.io_u_lat_u[index]++;
498 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
539 assert(index < FIO_IO_U_LAT_M_NR);
540 td->ts.io_u_lat_m[index]++;
543 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
546 io_u_mark_lat_usec(td, usec);
548 io_u_mark_lat_msec(td, usec / 1000);
552 * Get next file to service by choosing one at random
554 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
561 long r = os_random_long(&td->next_file_state);
563 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
565 if (f->flags & FIO_FILE_DONE)
568 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
569 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
576 * Get next file to service by doing round robin between all available ones
578 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
581 unsigned int old_next_file = td->next_file;
585 f = td->files[td->next_file];
588 if (td->next_file >= td->o.nr_files)
591 if (f->flags & FIO_FILE_DONE) {
596 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
600 } while (td->next_file != old_next_file);
602 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
606 static struct fio_file *get_next_file(struct thread_data *td)
610 assert(td->o.nr_files <= td->files_index);
612 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
613 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", td->nr_open_files, td->nr_done_files, td->o.nr_files);
617 f = td->file_service_file;
618 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
621 if (td->o.file_service_type == FIO_FSERVICE_RR)
622 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
624 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
626 td->file_service_file = f;
627 td->file_service_left = td->file_service_nr - 1;
629 dprint(FD_FILE, "get_next_file: %p\n", f);
633 static struct fio_file *find_next_new_file(struct thread_data *td)
637 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
640 if (td->o.file_service_type == FIO_FSERVICE_RR)
641 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
643 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
648 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
653 f = get_next_file(td);
661 if (!fill_io_u(td, io_u))
665 * td_io_close() does a put_file() as well, so no need to
668 unlock_file(io_u->file);
670 td_io_close_file(td, f);
671 f->flags |= FIO_FILE_DONE;
675 * probably not the right place to do this, but see
676 * if we need to open a new file
678 if (td->nr_open_files < td->o.open_files &&
679 td->o.open_files != td->o.nr_files) {
680 f = find_next_new_file(td);
682 if (!f || td_io_open_file(td, f))
693 struct io_u *__get_io_u(struct thread_data *td)
695 struct io_u *io_u = NULL;
697 if (!list_empty(&td->io_u_requeues))
698 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
699 else if (!queue_full(td)) {
700 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
709 assert(io_u->flags & IO_U_F_FREE);
710 io_u->flags &= ~IO_U_F_FREE;
713 list_del(&io_u->list);
714 list_add(&io_u->list, &td->io_u_busylist);
722 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
723 * etc. The returned io_u is fully ready to be prepped and submitted.
725 struct io_u *get_io_u(struct thread_data *td)
730 io_u = __get_io_u(td);
732 dprint(FD_IO, "__get_io_u failed\n");
737 * from a requeue, io_u already setup
743 * If using an iolog, grab next piece if any available.
745 if (td->o.read_iolog_file) {
746 if (read_iolog_get(td, io_u))
748 } else if (set_io_u_file(td, io_u)) {
749 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
754 assert(f->flags & FIO_FILE_OPEN);
756 if (io_u->ddir != DDIR_SYNC) {
757 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
758 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
762 f->last_pos = io_u->offset + io_u->buflen;
764 if (td->o.verify != VERIFY_NONE)
765 populate_verify_io_u(td, io_u);
769 * Set io data pointers.
771 io_u->endpos = io_u->offset + io_u->buflen;
772 io_u->xfer_buf = io_u->buf;
773 io_u->xfer_buflen = io_u->buflen;
775 if (!td_io_prep(td, io_u)) {
776 fio_gettime(&io_u->start_time, NULL);
780 dprint(FD_IO, "get_io_u failed\n");
785 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
787 const char *msg[] = { "read", "write", "sync" };
789 log_err("fio: io_u error");
792 log_err(" on file %s", io_u->file->file_name);
794 log_err(": %s\n", strerror(io_u->error));
796 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
799 td_verror(td, io_u->error, "io_u error");
802 static void io_completed(struct thread_data *td, struct io_u *io_u,
803 struct io_completion_data *icd)
807 dprint_io_u(io_u, "io complete");
809 assert(io_u->flags & IO_U_F_FLIGHT);
810 io_u->flags &= ~IO_U_F_FLIGHT;
812 if (io_u->ddir == DDIR_SYNC) {
813 td->last_was_sync = 1;
817 td->last_was_sync = 0;
820 unsigned int bytes = io_u->buflen - io_u->resid;
821 const enum fio_ddir idx = io_u->ddir;
824 td->io_blocks[idx]++;
825 td->io_bytes[idx] += bytes;
826 td->this_io_bytes[idx] += bytes;
828 usec = utime_since(&io_u->issue_time, &icd->time);
830 add_clat_sample(td, idx, usec);
831 add_bw_sample(td, idx, &icd->time);
832 io_u_mark_latency(td, usec);
834 if (td_write(td) && idx == DDIR_WRITE &&
836 td->o.verify != VERIFY_NONE)
837 log_io_piece(td, io_u);
839 icd->bytes_done[idx] += bytes;
842 ret = io_u->end_io(td, io_u);
843 if (ret && !icd->error)
847 icd->error = io_u->error;
848 io_u_log_error(td, io_u);
852 static void init_icd(struct io_completion_data *icd, int nr)
854 fio_gettime(&icd->time, NULL);
859 icd->bytes_done[0] = icd->bytes_done[1] = 0;
862 static void ios_completed(struct thread_data *td,
863 struct io_completion_data *icd)
868 for (i = 0; i < icd->nr; i++) {
869 io_u = td->io_ops->event(td, i);
871 io_completed(td, io_u, icd);
877 * Complete a single io_u for the sync engines.
879 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
881 struct io_completion_data icd;
884 io_completed(td, io_u, &icd);
888 return icd.bytes_done[0] + icd.bytes_done[1];
890 td_verror(td, icd.error, "io_u_sync_complete");
895 * Called to complete min_events number of io for the async engines.
897 long io_u_queued_complete(struct thread_data *td, int min_events)
899 struct io_completion_data icd;
900 struct timespec *tvp = NULL;
902 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
904 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
909 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
911 td_verror(td, -ret, "td_io_getevents");
917 ios_completed(td, &icd);
919 return icd.bytes_done[0] + icd.bytes_done[1];
921 td_verror(td, icd.error, "io_u_queued_complete");
926 * Call when io_u is really queued, to update the submission latency.
928 void io_u_queued(struct thread_data *td, struct io_u *io_u)
930 unsigned long slat_time;
932 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
933 add_slat_sample(td, io_u->ddir, slat_time);
936 #ifdef FIO_USE_TIMEOUT
937 void io_u_set_timeout(struct thread_data *td)
939 assert(td->cur_depth);
941 td->timer.it_interval.tv_sec = 0;
942 td->timer.it_interval.tv_usec = 0;
943 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
944 td->timer.it_value.tv_usec = 0;
945 setitimer(ITIMER_REAL, &td->timer, NULL);
946 fio_gettime(&td->timeout_end, NULL);
949 static void io_u_dump(struct io_u *io_u)
951 unsigned long t_start = mtime_since_now(&io_u->start_time);
952 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
954 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
955 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
956 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
959 void io_u_set_timeout(struct thread_data fio_unused *td)
964 #ifdef FIO_USE_TIMEOUT
965 static void io_u_timeout_handler(int fio_unused sig)
967 struct thread_data *td, *__td;
968 pid_t pid = getpid();
969 struct list_head *entry;
973 log_err("fio: io_u timeout\n");
976 * TLS would be nice...
979 for_each_td(__td, i) {
980 if (__td->pid == pid) {
987 log_err("fio: io_u timeout, can't find job\n");
991 if (!td->cur_depth) {
992 log_err("fio: timeout without pending work?\n");
996 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
998 list_for_each(entry, &td->io_u_busylist) {
999 io_u = list_entry(entry, struct io_u, list);
1004 td_verror(td, ETIMEDOUT, "io_u timeout");
1009 void io_u_init_timeout(void)
1011 #ifdef FIO_USE_TIMEOUT
1012 signal(SIGALRM, io_u_timeout_handler);