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 thread_data *td, struct fio_file *f,
29 unsigned long long block)
31 unsigned int idx = RAND_MAP_IDX(td, f, block);
32 unsigned int bit = RAND_MAP_BIT(td, f, block);
34 return (f->file_map[idx] & (1UL << bit)) == 0;
38 * Mark a given offset as used in the map.
40 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
42 unsigned int min_bs = td->o.rw_min_bs;
43 struct fio_file *f = io_u->file;
44 unsigned long long block;
46 unsigned int nr_blocks;
48 block = io_u->offset / (unsigned long long) min_bs;
50 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
52 while (blocks < nr_blocks) {
53 unsigned int idx, bit;
56 * If we have a mixed random workload, we may
57 * encounter blocks we already did IO to.
59 if (!td->o.ddir_nr && !random_map_free(td, f, block))
62 idx = RAND_MAP_IDX(td, f, block);
63 bit = RAND_MAP_BIT(td, f, block);
65 fio_assert(td, idx < f->num_maps);
67 f->file_map[idx] |= (1UL << bit);
72 if ((blocks * min_bs) < io_u->buflen)
73 io_u->buflen = blocks * min_bs;
77 * Return the next free block in the map.
79 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
80 unsigned long long *b)
84 i = f->last_free_lookup;
85 *b = (i * BLOCKS_PER_MAP);
86 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
87 if (f->file_map[i] != -1UL) {
88 *b += ffz(f->file_map[i]);
89 f->last_free_lookup = i;
100 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
101 int ddir, unsigned long long *b)
103 unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
104 unsigned long long r, rb;
108 r = os_random_long(&td->random_state);
112 *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
113 if (td->o.norandommap)
115 rb = *b + (f->file_offset / td->o.min_bs[ddir]);
117 } while (!random_map_free(td, f, rb) && loops);
120 * if we failed to retrieve a truly random offset within
121 * the loops assigned, see if there are free ones left at all
123 if (!loops && get_next_free_block(td, f, b))
130 * For random io, generate a random new block and see if it's used. Repeat
131 * until we find a free one. For sequential io, just return the end of
132 * the last io issued.
134 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
136 struct fio_file *f = io_u->file;
137 const int ddir = io_u->ddir;
138 unsigned long long b;
140 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
141 td->ddir_nr = td->o.ddir_nr;
143 if (get_next_rand_offset(td, f, ddir, &b))
146 if (f->last_pos >= f->real_file_size)
149 b = f->last_pos / td->o.min_bs[ddir];
152 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
153 if (io_u->offset >= f->real_file_size)
159 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
161 struct fio_file *f = io_u->file;
162 const int ddir = io_u->ddir;
166 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
167 buflen = td->o.min_bs[ddir];
169 r = os_random_long(&td->bsrange_state);
170 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
171 if (!td->o.bs_unaligned)
172 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
175 while (buflen + io_u->offset > f->real_file_size) {
176 if (buflen == td->o.min_bs[ddir]) {
177 if (!td->o.odirect) {
178 assert(io_u->offset <= f->real_file_size);
179 buflen = f->real_file_size - io_u->offset;
185 buflen = td->o.min_bs[ddir];
191 static void set_rwmix_bytes(struct thread_data *td)
193 unsigned long long rbytes;
197 * we do time or byte based switch. this is needed because
198 * buffered writes may issue a lot quicker than they complete,
199 * whereas reads do not.
201 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
202 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
204 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
207 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
212 r = os_random_long(&td->rwmix_state);
213 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
214 if (v < td->o.rwmix[DDIR_READ])
221 * Return the data direction for the next io_u. If the job is a
222 * mixed read/write workload, check the rwmix cycle and switch if
225 static enum fio_ddir get_rw_ddir(struct thread_data *td)
229 unsigned long elapsed;
232 fio_gettime(&now, NULL);
233 elapsed = mtime_since_now(&td->rwmix_switch);
236 * if this is the first cycle, make it shorter
238 cycle = td->o.rwmixcycle;
239 if (!td->rwmix_bytes)
243 * Check if it's time to seed a new data direction.
245 if (elapsed >= cycle ||
246 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
247 unsigned long long max_bytes;
251 * Put a top limit on how many bytes we do for
252 * one data direction, to avoid overflowing the
255 ddir = get_rand_ddir(td);
256 max_bytes = td->this_io_bytes[ddir];
257 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
258 if (!td->rw_end_set[ddir]) {
259 td->rw_end_set[ddir] = 1;
260 memcpy(&td->rw_end[ddir], &now, sizeof(now));
265 if (ddir != td->rwmix_ddir)
268 td->rwmix_ddir = ddir;
269 memcpy(&td->rwmix_switch, &now, sizeof(now));
271 return td->rwmix_ddir;
272 } else if (td_read(td))
278 void put_io_u(struct thread_data *td, struct io_u *io_u)
280 assert((io_u->flags & IO_U_F_FREE) == 0);
281 io_u->flags |= IO_U_F_FREE;
284 list_del(&io_u->list);
285 list_add(&io_u->list, &td->io_u_freelist);
289 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
291 struct io_u *__io_u = *io_u;
293 __io_u->flags |= IO_U_F_FREE;
294 __io_u->flags &= ~IO_U_F_FLIGHT;
296 list_del(&__io_u->list);
297 list_add_tail(&__io_u->list, &td->io_u_requeues);
302 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
305 * If using an iolog, grab next piece if any available.
307 if (td->o.read_iolog)
308 return read_iolog_get(td, io_u);
311 * see if it's time to sync
313 if (td->o.fsync_blocks &&
314 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
315 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
316 io_u->ddir = DDIR_SYNC;
320 io_u->ddir = get_rw_ddir(td);
323 * No log, let the seq/rand engine retrieve the next buflen and
326 if (get_next_offset(td, io_u))
329 io_u->buflen = get_next_buflen(td, io_u);
334 * mark entry before potentially trimming io_u
336 if (td_random(td) && !td->o.norandommap)
337 mark_random_map(td, io_u);
340 * If using a write iolog, store this entry.
343 if (td->o.write_iolog_file)
344 write_iolog_put(td, io_u);
349 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
353 if (io_u->ddir == DDIR_SYNC)
356 switch (td->cur_depth) {
373 td->ts.io_u_map[index]++;
374 td->ts.total_io_u[io_u->ddir]++;
377 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
408 td->ts.io_u_lat[index]++;
412 * Get next file to service by choosing one at random
414 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
421 long r = os_random_long(&td->next_file_state);
423 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
425 if (f->flags & FIO_FILE_DONE)
428 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
434 * Get next file to service by doing round robin between all available ones
436 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
439 unsigned int old_next_file = td->next_file;
443 f = &td->files[td->next_file];
446 if (td->next_file >= td->o.nr_files)
449 if (f->flags & FIO_FILE_DONE)
452 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
456 } while (td->next_file != old_next_file);
461 static struct fio_file *get_next_file(struct thread_data *td)
465 assert(td->o.nr_files <= td->files_index);
467 if (!td->nr_open_files)
470 f = td->file_service_file;
471 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
474 if (td->o.file_service_type == FIO_FSERVICE_RR)
475 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
477 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
479 td->file_service_file = f;
480 td->file_service_left = td->file_service_nr - 1;
484 static struct fio_file *find_next_new_file(struct thread_data *td)
488 if (td->o.file_service_type == FIO_FSERVICE_RR)
489 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
491 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
496 struct io_u *__get_io_u(struct thread_data *td)
498 struct io_u *io_u = NULL;
500 if (!list_empty(&td->io_u_requeues))
501 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
502 else if (!queue_full(td)) {
503 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
512 assert(io_u->flags & IO_U_F_FREE);
513 io_u->flags &= ~IO_U_F_FREE;
516 list_del(&io_u->list);
517 list_add(&io_u->list, &td->io_u_busylist);
525 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
526 * etc. The returned io_u is fully ready to be prepped and submitted.
528 struct io_u *get_io_u(struct thread_data *td)
534 io_u = __get_io_u(td);
539 * from a requeue, io_u already setup
545 f = get_next_file(td);
554 if (!fill_io_u(td, io_u))
558 * No more to do for this file, close it
561 td_io_close_file(td, f);
562 f->flags |= FIO_FILE_DONE;
565 * probably not the right place to do this, but see
566 * if we need to open a new file
568 if (td->nr_open_files < td->o.open_files &&
569 td->o.open_files != td->o.nr_files) {
570 f = find_next_new_file(td);
572 if (!f || (ret = td_io_open_file(td, f))) {
580 if (td->zone_bytes >= td->o.zone_size) {
582 f->last_pos += td->o.zone_skip;
585 if (io_u->buflen + io_u->offset > f->real_file_size) {
586 if (td->io_ops->flags & FIO_RAWIO) {
591 io_u->buflen = f->real_file_size - io_u->offset;
594 if (io_u->ddir != DDIR_SYNC) {
600 f->last_pos = io_u->offset + io_u->buflen;
602 if (td->o.verify != VERIFY_NONE)
603 populate_verify_io_u(td, io_u);
607 * Set io data pointers.
610 io_u->xfer_buf = io_u->buf;
611 io_u->xfer_buflen = io_u->buflen;
613 if (td_io_prep(td, io_u)) {
618 fio_gettime(&io_u->start_time, NULL);
622 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
624 const char *msg[] = { "read", "write", "sync" };
626 log_err("fio: io_u error");
629 log_err(" on file %s", io_u->file->file_name);
631 log_err(": %s\n", strerror(io_u->error));
633 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
636 td_verror(td, io_u->error, "io_u error");
639 static void io_completed(struct thread_data *td, struct io_u *io_u,
640 struct io_completion_data *icd)
644 assert(io_u->flags & IO_U_F_FLIGHT);
645 io_u->flags &= ~IO_U_F_FLIGHT;
647 put_file(td, io_u->file);
649 if (io_u->ddir == DDIR_SYNC) {
650 td->last_was_sync = 1;
654 td->last_was_sync = 0;
657 unsigned int bytes = io_u->buflen - io_u->resid;
658 const enum fio_ddir idx = io_u->ddir;
661 td->io_blocks[idx]++;
662 td->io_bytes[idx] += bytes;
663 td->zone_bytes += bytes;
664 td->this_io_bytes[idx] += bytes;
666 io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
668 msec = mtime_since(&io_u->issue_time, &icd->time);
670 add_clat_sample(td, idx, msec);
671 add_bw_sample(td, idx, &icd->time);
672 io_u_mark_latency(td, msec);
674 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
675 td->o.verify != VERIFY_NONE)
676 log_io_piece(td, io_u);
678 icd->bytes_done[idx] += bytes;
681 ret = io_u->end_io(td, io_u);
682 if (ret && !icd->error)
686 icd->error = io_u->error;
687 io_u_log_error(td, io_u);
691 static void init_icd(struct io_completion_data *icd, int nr)
693 fio_gettime(&icd->time, NULL);
698 icd->bytes_done[0] = icd->bytes_done[1] = 0;
701 static void ios_completed(struct thread_data *td,
702 struct io_completion_data *icd)
707 for (i = 0; i < icd->nr; i++) {
708 io_u = td->io_ops->event(td, i);
710 io_completed(td, io_u, icd);
716 * Complete a single io_u for the sync engines.
718 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
720 struct io_completion_data icd;
723 io_completed(td, io_u, &icd);
727 return icd.bytes_done[0] + icd.bytes_done[1];
729 td_verror(td, icd.error, "io_u_sync_complete");
734 * Called to complete min_events number of io for the async engines.
736 long io_u_queued_complete(struct thread_data *td, int min_events)
738 struct io_completion_data icd;
739 struct timespec *tvp = NULL;
741 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
746 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
748 td_verror(td, -ret, "td_io_getevents");
754 ios_completed(td, &icd);
756 return icd.bytes_done[0] + icd.bytes_done[1];
758 td_verror(td, icd.error, "io_u_queued_complete");
763 * Call when io_u is really queued, to update the submission latency.
765 void io_u_queued(struct thread_data *td, struct io_u *io_u)
767 unsigned long slat_time;
769 slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
770 add_slat_sample(td, io_u->ddir, slat_time);
773 #ifdef FIO_USE_TIMEOUT
774 void io_u_set_timeout(struct thread_data *td)
776 assert(td->cur_depth);
778 td->timer.it_interval.tv_sec = 0;
779 td->timer.it_interval.tv_usec = 0;
780 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
781 td->timer.it_value.tv_usec = 0;
782 setitimer(ITIMER_REAL, &td->timer, NULL);
783 fio_gettime(&td->timeout_end, NULL);
786 static void io_u_dump(struct io_u *io_u)
788 unsigned long t_start = mtime_since_now(&io_u->start_time);
789 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
791 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
792 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);
793 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
796 void io_u_set_timeout(struct thread_data fio_unused *td)
801 #ifdef FIO_USE_TIMEOUT
802 static void io_u_timeout_handler(int fio_unused sig)
804 struct thread_data *td, *__td;
805 pid_t pid = getpid();
806 struct list_head *entry;
810 log_err("fio: io_u timeout\n");
813 * TLS would be nice...
816 for_each_td(__td, i) {
817 if (__td->pid == pid) {
824 log_err("fio: io_u timeout, can't find job\n");
828 if (!td->cur_depth) {
829 log_err("fio: timeout without pending work?\n");
833 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
835 list_for_each(entry, &td->io_u_busylist) {
836 io_u = list_entry(entry, struct io_u, list);
841 td_verror(td, ETIMEDOUT, "io_u timeout");
846 void io_u_init_timeout(void)
848 #ifdef FIO_USE_TIMEOUT
849 signal(SIGALRM, io_u_timeout_handler);