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))
63 idx = RAND_MAP_IDX(f, block);
64 bit = RAND_MAP_BIT(f, block);
66 fio_assert(td, idx < f->num_maps);
68 this_blocks = nr_blocks;
69 if (this_blocks + bit > BLOCKS_PER_MAP)
70 this_blocks = BLOCKS_PER_MAP - bit;
72 if (this_blocks == BLOCKS_PER_MAP)
75 mask = ((1U << this_blocks) - 1) << bit;
77 fio_assert(td, !(f->file_map[idx] & mask));
78 f->file_map[idx] |= mask;
79 nr_blocks -= this_blocks;
80 blocks += this_blocks;
84 if ((blocks * min_bs) < io_u->buflen)
85 io_u->buflen = blocks * min_bs;
88 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
91 unsigned long long max_blocks;
92 unsigned long long max_size;
95 * Hmm, should we make sure that ->io_size <= ->real_file_size?
97 max_size = f->io_size;
98 if (max_size > f->real_file_size)
99 max_size = f->real_file_size;
101 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
109 * Return the next free block in the map.
111 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
112 enum fio_ddir ddir, unsigned long long *b)
114 unsigned long long min_bs = td->o.rw_min_bs;
117 i = f->last_free_lookup;
118 *b = (i * BLOCKS_PER_MAP);
119 while ((*b) * min_bs < f->real_file_size) {
120 if (f->file_map[i] != (unsigned int) -1) {
121 *b += ffz(f->file_map[i]);
122 if (*b > last_block(td, f, ddir))
124 f->last_free_lookup = i;
128 *b += BLOCKS_PER_MAP;
132 dprint(FD_IO, "failed finding a free block\n");
136 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
137 enum fio_ddir ddir, unsigned long long *b)
139 unsigned long long r;
143 r = os_random_long(&td->random_state);
144 dprint(FD_RANDOM, "off rand %llu\n", r);
145 *b = (last_block(td, f, ddir) - 1)
146 * (r / ((unsigned long long) RAND_MAX + 1.0));
149 * if we are not maintaining a random map, we are done.
151 if (!file_randommap(td, f))
155 * calculate map offset and check if it's free
157 if (random_map_free(f, *b))
160 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
165 * we get here, if we didn't suceed in looking up a block. generate
166 * a random start offset into the filemap, and find the first free
171 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
172 if (!get_next_free_block(td, f, ddir, b))
175 r = os_random_long(&td->random_state);
179 * that didn't work either, try exhaustive search from the start
181 f->last_free_lookup = 0;
182 return get_next_free_block(td, f, ddir, b);
186 * For random io, generate a random new block and see if it's used. Repeat
187 * until we find a free one. For sequential io, just return the end of
188 * the last io issued.
190 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
192 struct fio_file *f = io_u->file;
193 unsigned long long b;
194 enum fio_ddir ddir = io_u->ddir;
196 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
197 td->ddir_nr = td->o.ddir_nr;
199 if (get_next_rand_offset(td, f, ddir, &b))
202 if (f->last_pos >= f->real_file_size) {
203 if (!td_random(td) ||
204 get_next_rand_offset(td, f, ddir, &b))
207 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
210 io_u->offset = b * td->o.min_bs[ddir];
211 if (io_u->offset >= f->io_size) {
212 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
213 io_u->offset, f->io_size);
217 io_u->offset += f->file_offset;
218 if (io_u->offset >= f->real_file_size) {
219 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
220 io_u->offset, f->real_file_size);
227 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
229 const int ddir = io_u->ddir;
230 unsigned int buflen = buflen; /* silence dumb gcc warning */
233 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
234 buflen = td->o.min_bs[ddir];
236 r = os_random_long(&td->bsrange_state);
237 if (!td->o.bssplit_nr) {
238 buflen = (unsigned int)
239 (1 + (double) (td->o.max_bs[ddir] - 1)
240 * r / (RAND_MAX + 1.0));
245 for (i = 0; i < td->o.bssplit_nr; i++) {
246 struct bssplit *bsp = &td->o.bssplit[i];
250 if (r <= ((LONG_MAX / 100L) * perc))
254 if (!td->o.bs_unaligned) {
255 buflen = (buflen + td->o.min_bs[ddir] - 1)
256 & ~(td->o.min_bs[ddir] - 1);
260 if (io_u->offset + buflen > io_u->file->real_file_size) {
261 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
262 td->o.min_bs[ddir], ddir);
263 buflen = td->o.min_bs[ddir];
269 static void set_rwmix_bytes(struct thread_data *td)
274 * we do time or byte based switch. this is needed because
275 * buffered writes may issue a lot quicker than they complete,
276 * whereas reads do not.
278 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
279 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
282 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
287 r = os_random_long(&td->rwmix_state);
288 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
289 if (v <= td->o.rwmix[DDIR_READ])
296 * Return the data direction for the next io_u. If the job is a
297 * mixed read/write workload, check the rwmix cycle and switch if
300 static enum fio_ddir get_rw_ddir(struct thread_data *td)
304 * Check if it's time to seed a new data direction.
306 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
307 unsigned long long max_bytes;
311 * Put a top limit on how many bytes we do for
312 * one data direction, to avoid overflowing the
315 ddir = get_rand_ddir(td);
316 max_bytes = td->this_io_bytes[ddir];
318 (td->o.size * td->o.rwmix[ddir] / 100)) {
319 if (!td->rw_end_set[ddir])
320 td->rw_end_set[ddir] = 1;
325 if (ddir != td->rwmix_ddir)
328 td->rwmix_ddir = ddir;
330 return td->rwmix_ddir;
331 } else if (td_read(td))
337 static void put_file_log(struct thread_data *td, struct fio_file *f)
339 int ret = put_file(td, f);
342 td_verror(td, ret, "file close");
345 void put_io_u(struct thread_data *td, struct io_u *io_u)
347 assert((io_u->flags & IO_U_F_FREE) == 0);
348 io_u->flags |= IO_U_F_FREE;
351 put_file_log(td, io_u->file);
354 list_del(&io_u->list);
355 list_add(&io_u->list, &td->io_u_freelist);
359 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
361 struct io_u *__io_u = *io_u;
363 dprint(FD_IO, "requeue %p\n", __io_u);
365 __io_u->flags |= IO_U_F_FREE;
366 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
367 td->io_issues[__io_u->ddir]--;
369 __io_u->flags &= ~IO_U_F_FLIGHT;
371 list_del(&__io_u->list);
372 list_add_tail(&__io_u->list, &td->io_u_requeues);
377 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
379 if (td->io_ops->flags & FIO_NOIO)
383 * see if it's time to sync
385 if (td->o.fsync_blocks &&
386 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
387 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
388 io_u->ddir = DDIR_SYNC;
392 io_u->ddir = get_rw_ddir(td);
395 * See if it's time to switch to a new zone
397 if (td->zone_bytes >= td->o.zone_size) {
399 io_u->file->last_pos += td->o.zone_skip;
400 td->io_skip_bytes += td->o.zone_skip;
404 * No log, let the seq/rand engine retrieve the next buflen and
407 if (get_next_offset(td, io_u)) {
408 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
412 io_u->buflen = get_next_buflen(td, io_u);
414 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
418 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
419 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
420 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
421 io_u->buflen, io_u->file->real_file_size);
426 * mark entry before potentially trimming io_u
428 if (td_random(td) && file_randommap(td, io_u->file))
429 mark_random_map(td, io_u);
432 * If using a write iolog, store this entry.
435 dprint_io_u(io_u, "fill_io_u");
436 td->zone_bytes += io_u->buflen;
441 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
470 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
472 __io_u_mark_map(td->ts.io_u_submit, nr);
473 td->ts.total_submit++;
476 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
478 __io_u_mark_map(td->ts.io_u_complete, nr);
479 td->ts.total_complete++;
482 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
486 switch (td->cur_depth) {
508 td->ts.io_u_map[index] += nr;
511 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
548 assert(index < FIO_IO_U_LAT_U_NR);
549 td->ts.io_u_lat_u[index]++;
552 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
593 assert(index < FIO_IO_U_LAT_M_NR);
594 td->ts.io_u_lat_m[index]++;
597 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
600 io_u_mark_lat_usec(td, usec);
602 io_u_mark_lat_msec(td, usec / 1000);
606 * Get next file to service by choosing one at random
608 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
615 long r = os_random_long(&td->next_file_state);
617 fno = (unsigned int) ((double) td->o.nr_files
618 * (r / (RAND_MAX + 1.0)));
620 if (f->flags & FIO_FILE_DONE)
623 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
624 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
631 * Get next file to service by doing round robin between all available ones
633 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
636 unsigned int old_next_file = td->next_file;
640 f = td->files[td->next_file];
643 if (td->next_file >= td->o.nr_files)
646 if (f->flags & FIO_FILE_DONE) {
651 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
655 } while (td->next_file != old_next_file);
657 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
661 static struct fio_file *get_next_file(struct thread_data *td)
665 assert(td->o.nr_files <= td->files_index);
667 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
668 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
669 " nr_files=%d\n", td->nr_open_files,
675 f = td->file_service_file;
676 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
679 if (td->o.file_service_type == FIO_FSERVICE_RR)
680 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
682 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
684 td->file_service_file = f;
685 td->file_service_left = td->file_service_nr - 1;
687 dprint(FD_FILE, "get_next_file: %p\n", f);
691 static struct fio_file *find_next_new_file(struct thread_data *td)
695 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
698 if (td->o.file_service_type == FIO_FSERVICE_RR)
699 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
701 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
706 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
711 f = get_next_file(td);
719 if (!fill_io_u(td, io_u))
723 * optimization to prevent close/open of the same file. This
724 * way we preserve queueing etc.
726 if (td->o.nr_files == 1 && td->o.time_based) {
733 * td_io_close() does a put_file() as well, so no need to
737 td_io_close_file(td, f);
738 f->flags |= FIO_FILE_DONE;
742 * probably not the right place to do this, but see
743 * if we need to open a new file
745 if (td->nr_open_files < td->o.open_files &&
746 td->o.open_files != td->o.nr_files) {
747 f = find_next_new_file(td);
749 if (!f || td_io_open_file(td, f))
760 struct io_u *__get_io_u(struct thread_data *td)
762 struct io_u *io_u = NULL;
764 if (!list_empty(&td->io_u_requeues))
765 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
766 else if (!queue_full(td)) {
767 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
776 assert(io_u->flags & IO_U_F_FREE);
777 io_u->flags &= ~IO_U_F_FREE;
780 list_del(&io_u->list);
781 list_add(&io_u->list, &td->io_u_busylist);
789 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
790 * etc. The returned io_u is fully ready to be prepped and submitted.
792 struct io_u *get_io_u(struct thread_data *td)
797 io_u = __get_io_u(td);
799 dprint(FD_IO, "__get_io_u failed\n");
804 * from a requeue, io_u already setup
810 * If using an iolog, grab next piece if any available.
812 if (td->o.read_iolog_file) {
813 if (read_iolog_get(td, io_u))
815 } else if (set_io_u_file(td, io_u)) {
816 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
821 assert(f->flags & FIO_FILE_OPEN);
823 if (io_u->ddir != DDIR_SYNC) {
824 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
825 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
829 f->last_pos = io_u->offset + io_u->buflen;
831 if (td->o.verify != VERIFY_NONE)
832 populate_verify_io_u(td, io_u);
833 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
834 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
838 * Set io data pointers.
840 io_u->endpos = io_u->offset + io_u->buflen;
841 io_u->xfer_buf = io_u->buf;
842 io_u->xfer_buflen = io_u->buflen;
845 if (!td_io_prep(td, io_u)) {
846 fio_gettime(&io_u->start_time, NULL);
850 dprint(FD_IO, "get_io_u failed\n");
855 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
857 const char *msg[] = { "read", "write", "sync" };
859 log_err("fio: io_u error");
862 log_err(" on file %s", io_u->file->file_name);
864 log_err(": %s\n", strerror(io_u->error));
866 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
867 io_u->offset, io_u->xfer_buflen);
870 td_verror(td, io_u->error, "io_u error");
873 static void io_completed(struct thread_data *td, struct io_u *io_u,
874 struct io_completion_data *icd)
878 dprint_io_u(io_u, "io complete");
880 assert(io_u->flags & IO_U_F_FLIGHT);
881 io_u->flags &= ~IO_U_F_FLIGHT;
883 if (io_u->ddir == DDIR_SYNC) {
884 td->last_was_sync = 1;
888 td->last_was_sync = 0;
891 unsigned int bytes = io_u->buflen - io_u->resid;
892 const enum fio_ddir idx = io_u->ddir;
895 td->io_blocks[idx]++;
896 td->io_bytes[idx] += bytes;
897 td->this_io_bytes[idx] += bytes;
899 usec = utime_since(&io_u->issue_time, &icd->time);
901 add_clat_sample(td, idx, usec);
902 add_bw_sample(td, idx, &icd->time);
903 io_u_mark_latency(td, usec);
905 if (td_write(td) && idx == DDIR_WRITE &&
907 td->o.verify != VERIFY_NONE)
908 log_io_piece(td, io_u);
910 icd->bytes_done[idx] += bytes;
913 ret = io_u->end_io(td, io_u);
914 if (ret && !icd->error)
918 icd->error = io_u->error;
919 io_u_log_error(td, io_u);
923 static void init_icd(struct io_completion_data *icd, int nr)
925 fio_gettime(&icd->time, NULL);
930 icd->bytes_done[0] = icd->bytes_done[1] = 0;
933 static void ios_completed(struct thread_data *td,
934 struct io_completion_data *icd)
939 for (i = 0; i < icd->nr; i++) {
940 io_u = td->io_ops->event(td, i);
942 io_completed(td, io_u, icd);
948 * Complete a single io_u for the sync engines.
950 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
952 struct io_completion_data icd;
955 io_completed(td, io_u, &icd);
959 return icd.bytes_done[0] + icd.bytes_done[1];
961 td_verror(td, icd.error, "io_u_sync_complete");
966 * Called to complete min_events number of io for the async engines.
968 long io_u_queued_complete(struct thread_data *td, int min_events)
970 struct io_completion_data icd;
971 struct timespec *tvp = NULL;
973 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
975 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
980 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
982 td_verror(td, -ret, "td_io_getevents");
988 ios_completed(td, &icd);
990 return icd.bytes_done[0] + icd.bytes_done[1];
992 td_verror(td, icd.error, "io_u_queued_complete");
997 * Call when io_u is really queued, to update the submission latency.
999 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1001 unsigned long slat_time;
1003 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1004 add_slat_sample(td, io_u->ddir, slat_time);
1008 * "randomly" fill the buffer contents
1010 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1011 unsigned int max_bs)
1013 long *ptr = io_u->buf;
1015 if (!td->o.zero_buffers) {
1016 while ((void *) ptr - io_u->buf < max_bs) {
1017 *ptr = rand() * GOLDEN_RATIO_PRIME;
1021 memset(ptr, 0, max_bs);
1024 #ifdef FIO_USE_TIMEOUT
1025 void io_u_set_timeout(struct thread_data *td)
1027 assert(td->cur_depth);
1029 td->timer.it_interval.tv_sec = 0;
1030 td->timer.it_interval.tv_usec = 0;
1031 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1032 td->timer.it_value.tv_usec = 0;
1033 setitimer(ITIMER_REAL, &td->timer, NULL);
1034 fio_gettime(&td->timeout_end, NULL);
1037 static void io_u_dump(struct io_u *io_u)
1039 unsigned long t_start = mtime_since_now(&io_u->start_time);
1040 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1042 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1043 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1044 io_u->xfer_buf, io_u->buflen,
1047 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1050 void io_u_set_timeout(struct thread_data fio_unused *td)
1055 #ifdef FIO_USE_TIMEOUT
1056 static void io_u_timeout_handler(int fio_unused sig)
1058 struct thread_data *td, *__td;
1059 pid_t pid = getpid();
1060 struct list_head *entry;
1064 log_err("fio: io_u timeout\n");
1067 * TLS would be nice...
1070 for_each_td(__td, i) {
1071 if (__td->pid == pid) {
1078 log_err("fio: io_u timeout, can't find job\n");
1082 if (!td->cur_depth) {
1083 log_err("fio: timeout without pending work?\n");
1087 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1089 list_for_each(entry, &td->io_u_busylist) {
1090 io_u = list_entry(entry, struct io_u, list);
1095 td_verror(td, ETIMEDOUT, "io_u timeout");
1100 void io_u_init_timeout(void)
1102 #ifdef FIO_USE_TIMEOUT
1103 signal(SIGALRM, io_u_timeout_handler);