12 * The ->file_map[] contains a map of blocks we have or have not done io
13 * to yet. Used to make sure we cover the entire range in a fair fashion.
15 static int random_map_free(struct thread_data *td, struct fio_file *f,
16 unsigned long long block)
18 unsigned int idx = RAND_MAP_IDX(td, f, block);
19 unsigned int bit = RAND_MAP_BIT(td, f, block);
21 return (f->file_map[idx] & (1UL << bit)) == 0;
25 * Mark a given offset as used in the map.
27 static void mark_random_map(struct thread_data *td, struct fio_file *f,
30 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
31 unsigned int blocks = 0;
33 while (blocks < (io_u->buflen / td->min_bs)) {
34 unsigned int idx, bit;
36 if (!random_map_free(td, f, block))
39 idx = RAND_MAP_IDX(td, f, block);
40 bit = RAND_MAP_BIT(td, f, block);
42 assert(idx < f->num_maps);
44 f->file_map[idx] |= (1UL << bit);
49 if ((blocks * td->min_bs) < io_u->buflen)
50 io_u->buflen = blocks * td->min_bs;
54 * Return the next free block in the map.
56 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
57 unsigned long long *b)
63 while ((*b) * td->min_bs < f->file_size) {
64 if (f->file_map[i] != -1UL) {
65 *b += ffz(f->file_map[i]);
77 * For random io, generate a random new block and see if it's used. Repeat
78 * until we find a free one. For sequential io, just return the end of
81 static int get_next_offset(struct thread_data *td, struct fio_file *f,
82 unsigned long long *offset)
84 unsigned long long b, rb;
87 if (!td->sequential) {
88 unsigned long long max_blocks = td->io_size / td->min_bs;
92 r = os_random_long(&td->random_state);
93 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
96 rb = b + (f->file_offset / td->min_bs);
98 } while (!random_map_free(td, f, rb) && loops);
101 if (get_next_free_block(td, f, &b))
105 b = f->last_pos / td->min_bs;
107 *offset = (b * td->min_bs) + f->file_offset;
108 if (*offset > f->file_size)
114 static unsigned int get_next_buflen(struct thread_data *td)
119 if (td->min_bs == td->max_bs)
122 r = os_random_long(&td->bsrange_state);
123 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
124 if (!td->bs_unaligned)
125 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
128 if (buflen > td->io_size - td->this_io_bytes[td->ddir]) {
130 * if using direct/raw io, we may not be able to
131 * shrink the size. so just fail it.
133 if (td->io_ops->flags & FIO_RAWIO)
136 buflen = td->io_size - td->this_io_bytes[td->ddir];
143 * Return the data direction for the next io_u. If the job is a
144 * mixed read/write workload, check the rwmix cycle and switch if
147 static int get_rw_ddir(struct thread_data *td)
151 unsigned long elapsed;
153 gettimeofday(&now, NULL);
154 elapsed = mtime_since_now(&td->rwmix_switch);
157 * Check if it's time to seed a new data direction.
159 if (elapsed >= td->rwmixcycle) {
163 r = os_random_long(&td->rwmix_state);
164 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
165 if (v < td->rwmixread)
166 td->rwmix_ddir = DDIR_READ;
168 td->rwmix_ddir = DDIR_WRITE;
169 memcpy(&td->rwmix_switch, &now, sizeof(now));
171 return td->rwmix_ddir;
172 } else if (td_read(td))
178 void put_io_u(struct thread_data *td, struct io_u *io_u)
181 list_del(&io_u->list);
182 list_add(&io_u->list, &td->io_u_freelist);
186 static int fill_io_u(struct thread_data *td, struct fio_file *f,
190 * If using an iolog, grab next piece if any available.
193 return read_iolog_get(td, io_u);
196 * see if it's time to sync
198 if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
199 && should_fsync(td)) {
200 io_u->ddir = DDIR_SYNC;
206 * No log, let the seq/rand engine retrieve the next position.
208 if (!get_next_offset(td, f, &io_u->offset)) {
209 io_u->buflen = get_next_buflen(td);
212 io_u->ddir = get_rw_ddir(td);
215 * If using a write iolog, store this entry.
217 if (td->write_iolog_file)
218 write_iolog_put(td, io_u);
228 struct io_u *__get_io_u(struct thread_data *td)
230 struct io_u *io_u = NULL;
232 if (!queue_full(td)) {
233 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
238 list_del(&io_u->list);
239 list_add(&io_u->list, &td->io_u_busylist);
247 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
248 * etc. The returned io_u is fully ready to be prepped and submitted.
250 struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
254 io_u = __get_io_u(td);
258 if (td->zone_bytes >= td->zone_size) {
260 f->last_pos += td->zone_skip;
263 if (fill_io_u(td, f, io_u)) {
268 if (io_u->buflen + io_u->offset > f->file_size) {
269 if (td->io_ops->flags & FIO_RAWIO) {
274 io_u->buflen = f->file_size - io_u->offset;
277 if (io_u->ddir != DDIR_SYNC) {
283 if (!td->read_iolog && !td->sequential && !td->norandommap)
284 mark_random_map(td, f, io_u);
286 f->last_pos += io_u->buflen;
288 if (td->verify != VERIFY_NONE)
289 populate_verify_io_u(td, io_u);
292 if (td_io_prep(td, io_u)) {
297 gettimeofday(&io_u->start_time, NULL);
301 void io_completed(struct thread_data *td, struct io_u *io_u,
302 struct io_completion_data *icd)
307 if (io_u->ddir == DDIR_SYNC) {
308 td->last_was_sync = 1;
312 td->last_was_sync = 0;
314 gettimeofday(&e, NULL);
317 unsigned int bytes = io_u->buflen - io_u->resid;
318 const int idx = io_u->ddir;
320 td->io_blocks[idx]++;
321 td->io_bytes[idx] += bytes;
322 td->zone_bytes += bytes;
323 td->this_io_bytes[idx] += bytes;
325 msec = mtime_since(&io_u->issue_time, &e);
327 add_clat_sample(td, idx, msec);
328 add_bw_sample(td, idx);
330 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
331 log_io_piece(td, io_u);
333 icd->bytes_done[idx] += bytes;
335 icd->error = io_u->error;
338 void ios_completed(struct thread_data *td, struct io_completion_data *icd)
344 icd->bytes_done[0] = icd->bytes_done[1] = 0;
346 for (i = 0; i < icd->nr; i++) {
347 io_u = td->io_ops->event(td, i);
349 io_completed(td, io_u, icd);