Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
6 | #include <assert.h> | |
7 | ||
8 | #include "fio.h" | |
9 | #include "os.h" | |
10 | ||
11 | /* | |
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. | |
14 | */ | |
15 | static int random_map_free(struct thread_data *td, struct fio_file *f, | |
16 | unsigned long long block) | |
17 | { | |
18 | unsigned int idx = RAND_MAP_IDX(td, f, block); | |
19 | unsigned int bit = RAND_MAP_BIT(td, f, block); | |
20 | ||
21 | return (f->file_map[idx] & (1UL << bit)) == 0; | |
22 | } | |
23 | ||
df415585 JA |
24 | /* |
25 | * Mark a given offset as used in the map. | |
26 | */ | |
27 | static void mark_random_map(struct thread_data *td, struct fio_file *f, | |
28 | struct io_u *io_u) | |
29 | { | |
30 | unsigned long long block = io_u->offset / (unsigned long long) td->min_bs; | |
31 | unsigned int blocks = 0; | |
32 | ||
33 | while (blocks < (io_u->buflen / td->min_bs)) { | |
34 | unsigned int idx, bit; | |
35 | ||
36 | if (!random_map_free(td, f, block)) | |
37 | break; | |
38 | ||
39 | idx = RAND_MAP_IDX(td, f, block); | |
40 | bit = RAND_MAP_BIT(td, f, block); | |
41 | ||
42 | assert(idx < f->num_maps); | |
43 | ||
44 | f->file_map[idx] |= (1UL << bit); | |
45 | block++; | |
46 | blocks++; | |
47 | } | |
48 | ||
49 | if ((blocks * td->min_bs) < io_u->buflen) | |
50 | io_u->buflen = blocks * td->min_bs; | |
51 | } | |
52 | ||
10ba535a JA |
53 | /* |
54 | * Return the next free block in the map. | |
55 | */ | |
56 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
57 | unsigned long long *b) | |
58 | { | |
59 | int i; | |
60 | ||
61 | *b = 0; | |
62 | i = 0; | |
63 | while ((*b) * td->min_bs < f->file_size) { | |
64 | if (f->file_map[i] != -1UL) { | |
65 | *b += ffz(f->file_map[i]); | |
66 | return 0; | |
67 | } | |
68 | ||
69 | *b += BLOCKS_PER_MAP; | |
70 | i++; | |
71 | } | |
72 | ||
73 | return 1; | |
74 | } | |
75 | ||
76 | /* | |
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 | |
79 | * the last io issued. | |
80 | */ | |
81 | static int get_next_offset(struct thread_data *td, struct fio_file *f, | |
82 | unsigned long long *offset) | |
83 | { | |
84 | unsigned long long b, rb; | |
85 | long r; | |
86 | ||
87 | if (!td->sequential) { | |
88 | unsigned long long max_blocks = td->io_size / td->min_bs; | |
89 | int loops = 50; | |
90 | ||
91 | do { | |
92 | r = os_random_long(&td->random_state); | |
93 | b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
94 | rb = b + (f->file_offset / td->min_bs); | |
95 | loops--; | |
96 | } while (!random_map_free(td, f, rb) && loops); | |
97 | ||
98 | if (!loops) { | |
99 | if (get_next_free_block(td, f, &b)) | |
100 | return 1; | |
101 | } | |
102 | } else | |
103 | b = f->last_pos / td->min_bs; | |
104 | ||
105 | *offset = (b * td->min_bs) + f->file_offset; | |
106 | if (*offset > f->file_size) | |
107 | return 1; | |
108 | ||
109 | return 0; | |
110 | } | |
111 | ||
112 | static unsigned int get_next_buflen(struct thread_data *td) | |
113 | { | |
114 | unsigned int buflen; | |
115 | long r; | |
116 | ||
117 | if (td->min_bs == td->max_bs) | |
118 | buflen = td->min_bs; | |
119 | else { | |
120 | r = os_random_long(&td->bsrange_state); | |
121 | buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0)); | |
122 | buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1); | |
123 | } | |
124 | ||
125 | if (buflen > td->io_size - td->this_io_bytes[td->ddir]) { | |
126 | /* | |
127 | * if using direct/raw io, we may not be able to | |
128 | * shrink the size. so just fail it. | |
129 | */ | |
130 | if (td->io_ops->flags & FIO_RAWIO) | |
131 | return 0; | |
132 | ||
133 | buflen = td->io_size - td->this_io_bytes[td->ddir]; | |
134 | } | |
135 | ||
136 | return buflen; | |
137 | } | |
138 | ||
139 | /* | |
140 | * Return the data direction for the next io_u. If the job is a | |
141 | * mixed read/write workload, check the rwmix cycle and switch if | |
142 | * necessary. | |
143 | */ | |
144 | static int get_rw_ddir(struct thread_data *td) | |
145 | { | |
146 | if (td_rw(td)) { | |
147 | struct timeval now; | |
148 | unsigned long elapsed; | |
149 | ||
150 | gettimeofday(&now, NULL); | |
151 | elapsed = mtime_since_now(&td->rwmix_switch); | |
152 | ||
153 | /* | |
154 | * Check if it's time to seed a new data direction. | |
155 | */ | |
156 | if (elapsed >= td->rwmixcycle) { | |
157 | unsigned int v; | |
158 | long r; | |
159 | ||
160 | r = os_random_long(&td->rwmix_state); | |
161 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
162 | if (v < td->rwmixread) | |
163 | td->rwmix_ddir = DDIR_READ; | |
164 | else | |
165 | td->rwmix_ddir = DDIR_WRITE; | |
166 | memcpy(&td->rwmix_switch, &now, sizeof(now)); | |
167 | } | |
168 | return td->rwmix_ddir; | |
169 | } else if (td_read(td)) | |
170 | return DDIR_READ; | |
171 | else | |
172 | return DDIR_WRITE; | |
173 | } | |
174 | ||
10ba535a JA |
175 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
176 | { | |
177 | io_u->file = NULL; | |
178 | list_del(&io_u->list); | |
179 | list_add(&io_u->list, &td->io_u_freelist); | |
180 | td->cur_depth--; | |
181 | } | |
182 | ||
183 | static int fill_io_u(struct thread_data *td, struct fio_file *f, | |
184 | struct io_u *io_u) | |
185 | { | |
186 | /* | |
187 | * If using an iolog, grab next piece if any available. | |
188 | */ | |
189 | if (td->read_iolog) | |
190 | return read_iolog_get(td, io_u); | |
191 | ||
87dc1ab1 JA |
192 | /* |
193 | * see if it's time to sync | |
194 | */ | |
195 | if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks) | |
196 | && should_fsync(td)) { | |
197 | io_u->ddir = DDIR_SYNC; | |
198 | io_u->file = f; | |
199 | return 0; | |
200 | } | |
201 | ||
10ba535a JA |
202 | /* |
203 | * No log, let the seq/rand engine retrieve the next position. | |
204 | */ | |
205 | if (!get_next_offset(td, f, &io_u->offset)) { | |
206 | io_u->buflen = get_next_buflen(td); | |
207 | ||
208 | if (io_u->buflen) { | |
209 | io_u->ddir = get_rw_ddir(td); | |
210 | ||
211 | /* | |
212 | * If using a write iolog, store this entry. | |
213 | */ | |
214 | if (td->write_iolog) | |
215 | write_iolog_put(td, io_u); | |
216 | ||
217 | io_u->file = f; | |
218 | return 0; | |
219 | } | |
220 | } | |
221 | ||
222 | return 1; | |
223 | } | |
224 | ||
225 | struct io_u *__get_io_u(struct thread_data *td) | |
226 | { | |
227 | struct io_u *io_u = NULL; | |
228 | ||
229 | if (!queue_full(td)) { | |
230 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); | |
231 | ||
232 | io_u->error = 0; | |
233 | io_u->resid = 0; | |
234 | list_del(&io_u->list); | |
235 | list_add(&io_u->list, &td->io_u_busylist); | |
236 | td->cur_depth++; | |
237 | } | |
238 | ||
239 | return io_u; | |
240 | } | |
241 | ||
242 | /* | |
243 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
244 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
245 | */ | |
246 | struct io_u *get_io_u(struct thread_data *td, struct fio_file *f) | |
247 | { | |
248 | struct io_u *io_u; | |
249 | ||
250 | io_u = __get_io_u(td); | |
251 | if (!io_u) | |
252 | return NULL; | |
253 | ||
254 | if (td->zone_bytes >= td->zone_size) { | |
255 | td->zone_bytes = 0; | |
256 | f->last_pos += td->zone_skip; | |
257 | } | |
258 | ||
259 | if (fill_io_u(td, f, io_u)) { | |
260 | put_io_u(td, io_u); | |
261 | return NULL; | |
262 | } | |
263 | ||
264 | if (io_u->buflen + io_u->offset > f->file_size) { | |
265 | if (td->io_ops->flags & FIO_RAWIO) { | |
266 | put_io_u(td, io_u); | |
267 | return NULL; | |
268 | } | |
269 | ||
270 | io_u->buflen = f->file_size - io_u->offset; | |
271 | } | |
272 | ||
87dc1ab1 JA |
273 | if (io_u->ddir != DDIR_SYNC) { |
274 | if (!io_u->buflen) { | |
275 | put_io_u(td, io_u); | |
276 | return NULL; | |
277 | } | |
10ba535a | 278 | |
87dc1ab1 JA |
279 | if (!td->read_iolog && !td->sequential) |
280 | mark_random_map(td, f, io_u); | |
10ba535a | 281 | |
87dc1ab1 | 282 | f->last_pos += io_u->buflen; |
10ba535a | 283 | |
87dc1ab1 JA |
284 | if (td->verify != VERIFY_NONE) |
285 | populate_verify_io_u(td, io_u); | |
286 | } | |
10ba535a JA |
287 | |
288 | if (td_io_prep(td, io_u)) { | |
289 | put_io_u(td, io_u); | |
290 | return NULL; | |
291 | } | |
292 | ||
293 | gettimeofday(&io_u->start_time, NULL); | |
294 | return io_u; | |
295 | } | |
296 | ||
297 | void io_completed(struct thread_data *td, struct io_u *io_u, | |
298 | struct io_completion_data *icd) | |
299 | { | |
300 | struct timeval e; | |
301 | unsigned long msec; | |
302 | ||
87dc1ab1 JA |
303 | if (io_u->ddir == DDIR_SYNC) { |
304 | td->last_was_sync = 1; | |
305 | return; | |
306 | } | |
307 | ||
308 | td->last_was_sync = 0; | |
309 | ||
10ba535a JA |
310 | gettimeofday(&e, NULL); |
311 | ||
312 | if (!io_u->error) { | |
313 | unsigned int bytes = io_u->buflen - io_u->resid; | |
314 | const int idx = io_u->ddir; | |
315 | ||
316 | td->io_blocks[idx]++; | |
317 | td->io_bytes[idx] += bytes; | |
318 | td->zone_bytes += bytes; | |
319 | td->this_io_bytes[idx] += bytes; | |
320 | ||
321 | msec = mtime_since(&io_u->issue_time, &e); | |
322 | ||
323 | add_clat_sample(td, idx, msec); | |
324 | add_bw_sample(td, idx); | |
325 | ||
326 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) | |
327 | log_io_piece(td, io_u); | |
328 | ||
329 | icd->bytes_done[idx] += bytes; | |
330 | } else | |
331 | icd->error = io_u->error; | |
332 | } | |
333 | ||
334 | void ios_completed(struct thread_data *td, struct io_completion_data *icd) | |
335 | { | |
336 | struct io_u *io_u; | |
337 | int i; | |
338 | ||
339 | icd->error = 0; | |
340 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
341 | ||
342 | for (i = 0; i < icd->nr; i++) { | |
343 | io_u = td->io_ops->event(td, i); | |
344 | ||
345 | io_completed(td, io_u, icd); | |
346 | put_io_u(td, io_u); | |
347 | } | |
348 | } | |
349 | ||
350 |