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 | { | |
afdbe580 | 30 | unsigned int min_bs = td->rw_min_bs; |
a00735e6 JA |
31 | unsigned long long block; |
32 | unsigned int blocks; | |
c685b5b2 | 33 | unsigned int nr_blocks; |
df415585 | 34 | |
a00735e6 JA |
35 | block = io_u->offset / (unsigned long long) min_bs; |
36 | blocks = 0; | |
c685b5b2 JA |
37 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
38 | ||
39 | while (blocks < nr_blocks) { | |
df415585 JA |
40 | unsigned int idx, bit; |
41 | ||
42 | if (!random_map_free(td, f, block)) | |
43 | break; | |
44 | ||
45 | idx = RAND_MAP_IDX(td, f, block); | |
46 | bit = RAND_MAP_BIT(td, f, block); | |
47 | ||
48 | assert(idx < f->num_maps); | |
49 | ||
50 | f->file_map[idx] |= (1UL << bit); | |
51 | block++; | |
52 | blocks++; | |
53 | } | |
54 | ||
a00735e6 JA |
55 | if ((blocks * min_bs) < io_u->buflen) |
56 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
57 | } |
58 | ||
10ba535a JA |
59 | /* |
60 | * Return the next free block in the map. | |
61 | */ | |
62 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
63 | unsigned long long *b) | |
64 | { | |
65 | int i; | |
66 | ||
c685b5b2 JA |
67 | i = f->last_free_lookup; |
68 | *b = (i * BLOCKS_PER_MAP); | |
ee88470c | 69 | while ((*b) * td->rw_min_bs < f->real_file_size) { |
10ba535a JA |
70 | if (f->file_map[i] != -1UL) { |
71 | *b += ffz(f->file_map[i]); | |
c685b5b2 | 72 | f->last_free_lookup = i; |
10ba535a JA |
73 | return 0; |
74 | } | |
75 | ||
76 | *b += BLOCKS_PER_MAP; | |
77 | i++; | |
78 | } | |
79 | ||
80 | return 1; | |
81 | } | |
82 | ||
83 | /* | |
84 | * For random io, generate a random new block and see if it's used. Repeat | |
85 | * until we find a free one. For sequential io, just return the end of | |
86 | * the last io issued. | |
87 | */ | |
88 | static int get_next_offset(struct thread_data *td, struct fio_file *f, | |
c685b5b2 | 89 | struct io_u *io_u) |
10ba535a | 90 | { |
c685b5b2 | 91 | const int ddir = io_u->ddir; |
10ba535a JA |
92 | unsigned long long b, rb; |
93 | long r; | |
94 | ||
95 | if (!td->sequential) { | |
f697125a | 96 | unsigned long long max_blocks = f->file_size / td->min_bs[ddir]; |
c685b5b2 | 97 | int loops = 5; |
10ba535a JA |
98 | |
99 | do { | |
100 | r = os_random_long(&td->random_state); | |
101 | b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
bb8895e0 JA |
102 | if (td->norandommap) |
103 | break; | |
a00735e6 | 104 | rb = b + (f->file_offset / td->min_bs[ddir]); |
10ba535a JA |
105 | loops--; |
106 | } while (!random_map_free(td, f, rb) && loops); | |
107 | ||
108 | if (!loops) { | |
109 | if (get_next_free_block(td, f, &b)) | |
110 | return 1; | |
c685b5b2 JA |
111 | |
112 | b += f->file_offset / td->min_bs[ddir]; | |
10ba535a JA |
113 | } |
114 | } else | |
a00735e6 | 115 | b = f->last_pos / td->min_bs[ddir]; |
10ba535a | 116 | |
c685b5b2 JA |
117 | io_u->offset = (b * td->min_bs[ddir]) + f->file_offset; |
118 | if (io_u->offset > f->real_file_size) | |
10ba535a JA |
119 | return 1; |
120 | ||
121 | return 0; | |
122 | } | |
123 | ||
a00735e6 | 124 | static unsigned int get_next_buflen(struct thread_data *td, int ddir) |
10ba535a JA |
125 | { |
126 | unsigned int buflen; | |
127 | long r; | |
128 | ||
a00735e6 JA |
129 | if (td->min_bs[ddir] == td->max_bs[ddir]) |
130 | buflen = td->min_bs[ddir]; | |
10ba535a JA |
131 | else { |
132 | r = os_random_long(&td->bsrange_state); | |
1e97cce9 | 133 | buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); |
690adba3 | 134 | if (!td->bs_unaligned) |
a00735e6 | 135 | buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1); |
10ba535a JA |
136 | } |
137 | ||
a00735e6 | 138 | if (buflen > td->io_size - td->this_io_bytes[ddir]) { |
10ba535a JA |
139 | /* |
140 | * if using direct/raw io, we may not be able to | |
141 | * shrink the size. so just fail it. | |
142 | */ | |
143 | if (td->io_ops->flags & FIO_RAWIO) | |
144 | return 0; | |
145 | ||
a00735e6 | 146 | buflen = td->io_size - td->this_io_bytes[ddir]; |
10ba535a JA |
147 | } |
148 | ||
149 | return buflen; | |
150 | } | |
151 | ||
152 | /* | |
153 | * Return the data direction for the next io_u. If the job is a | |
154 | * mixed read/write workload, check the rwmix cycle and switch if | |
155 | * necessary. | |
156 | */ | |
1e97cce9 | 157 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
158 | { |
159 | if (td_rw(td)) { | |
160 | struct timeval now; | |
161 | unsigned long elapsed; | |
162 | ||
02bcaa8c | 163 | fio_gettime(&now, NULL); |
10ba535a JA |
164 | elapsed = mtime_since_now(&td->rwmix_switch); |
165 | ||
166 | /* | |
167 | * Check if it's time to seed a new data direction. | |
168 | */ | |
169 | if (elapsed >= td->rwmixcycle) { | |
170 | unsigned int v; | |
171 | long r; | |
172 | ||
173 | r = os_random_long(&td->rwmix_state); | |
174 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
175 | if (v < td->rwmixread) | |
176 | td->rwmix_ddir = DDIR_READ; | |
177 | else | |
178 | td->rwmix_ddir = DDIR_WRITE; | |
179 | memcpy(&td->rwmix_switch, &now, sizeof(now)); | |
180 | } | |
181 | return td->rwmix_ddir; | |
182 | } else if (td_read(td)) | |
183 | return DDIR_READ; | |
184 | else | |
185 | return DDIR_WRITE; | |
186 | } | |
187 | ||
10ba535a JA |
188 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
189 | { | |
190 | io_u->file = NULL; | |
191 | list_del(&io_u->list); | |
192 | list_add(&io_u->list, &td->io_u_freelist); | |
193 | td->cur_depth--; | |
194 | } | |
195 | ||
196 | static int fill_io_u(struct thread_data *td, struct fio_file *f, | |
197 | struct io_u *io_u) | |
198 | { | |
199 | /* | |
200 | * If using an iolog, grab next piece if any available. | |
201 | */ | |
202 | if (td->read_iolog) | |
203 | return read_iolog_get(td, io_u); | |
204 | ||
87dc1ab1 JA |
205 | /* |
206 | * see if it's time to sync | |
207 | */ | |
208 | if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks) | |
209 | && should_fsync(td)) { | |
210 | io_u->ddir = DDIR_SYNC; | |
211 | io_u->file = f; | |
212 | return 0; | |
213 | } | |
214 | ||
a00735e6 JA |
215 | io_u->ddir = get_rw_ddir(td); |
216 | ||
10ba535a | 217 | /* |
c685b5b2 JA |
218 | * No log, let the seq/rand engine retrieve the next buflen and |
219 | * position. | |
10ba535a | 220 | */ |
c685b5b2 JA |
221 | io_u->buflen = get_next_buflen(td, io_u->ddir); |
222 | if (io_u->buflen) { | |
223 | if (!get_next_offset(td, f, io_u)) { | |
224 | /* | |
225 | * mark entry before potentially trimming io_u | |
226 | */ | |
227 | if (!td->read_iolog && !td->sequential && | |
228 | !td->norandommap) | |
229 | mark_random_map(td, f, io_u); | |
230 | ||
10ba535a JA |
231 | /* |
232 | * If using a write iolog, store this entry. | |
233 | */ | |
076efc7c | 234 | if (td->write_iolog_file) |
10ba535a JA |
235 | write_iolog_put(td, io_u); |
236 | ||
237 | io_u->file = f; | |
238 | return 0; | |
239 | } | |
240 | } | |
241 | ||
242 | return 1; | |
243 | } | |
244 | ||
71619dc2 JA |
245 | static void io_u_mark_depth(struct thread_data *td) |
246 | { | |
247 | int index = 0; | |
248 | ||
249 | switch (td->cur_depth) { | |
250 | default: | |
251 | index++; | |
252 | case 32 ... 63: | |
253 | index++; | |
254 | case 16 ... 31: | |
255 | index++; | |
256 | case 8 ... 15: | |
257 | index++; | |
258 | case 4 ... 7: | |
259 | index++; | |
260 | case 2 ... 3: | |
261 | index++; | |
262 | case 1: | |
263 | break; | |
264 | } | |
265 | ||
266 | td->io_u_map[index]++; | |
267 | td->total_io_u++; | |
268 | } | |
269 | ||
10ba535a JA |
270 | struct io_u *__get_io_u(struct thread_data *td) |
271 | { | |
272 | struct io_u *io_u = NULL; | |
273 | ||
274 | if (!queue_full(td)) { | |
275 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); | |
276 | ||
6040dabc | 277 | io_u->buflen = 0; |
10ba535a JA |
278 | io_u->error = 0; |
279 | io_u->resid = 0; | |
280 | list_del(&io_u->list); | |
281 | list_add(&io_u->list, &td->io_u_busylist); | |
282 | td->cur_depth++; | |
71619dc2 | 283 | io_u_mark_depth(td); |
10ba535a JA |
284 | } |
285 | ||
286 | return io_u; | |
287 | } | |
288 | ||
289 | /* | |
290 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
291 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
292 | */ | |
293 | struct io_u *get_io_u(struct thread_data *td, struct fio_file *f) | |
294 | { | |
295 | struct io_u *io_u; | |
296 | ||
297 | io_u = __get_io_u(td); | |
298 | if (!io_u) | |
299 | return NULL; | |
300 | ||
301 | if (td->zone_bytes >= td->zone_size) { | |
302 | td->zone_bytes = 0; | |
303 | f->last_pos += td->zone_skip; | |
304 | } | |
305 | ||
306 | if (fill_io_u(td, f, io_u)) { | |
307 | put_io_u(td, io_u); | |
308 | return NULL; | |
309 | } | |
310 | ||
ee88470c | 311 | if (io_u->buflen + io_u->offset > f->real_file_size) { |
10ba535a JA |
312 | if (td->io_ops->flags & FIO_RAWIO) { |
313 | put_io_u(td, io_u); | |
314 | return NULL; | |
315 | } | |
316 | ||
ee88470c | 317 | io_u->buflen = f->real_file_size - io_u->offset; |
10ba535a JA |
318 | } |
319 | ||
87dc1ab1 JA |
320 | if (io_u->ddir != DDIR_SYNC) { |
321 | if (!io_u->buflen) { | |
322 | put_io_u(td, io_u); | |
323 | return NULL; | |
324 | } | |
10ba535a | 325 | |
87dc1ab1 | 326 | f->last_pos += io_u->buflen; |
10ba535a | 327 | |
87dc1ab1 JA |
328 | if (td->verify != VERIFY_NONE) |
329 | populate_verify_io_u(td, io_u); | |
330 | } | |
10ba535a JA |
331 | |
332 | if (td_io_prep(td, io_u)) { | |
333 | put_io_u(td, io_u); | |
334 | return NULL; | |
335 | } | |
336 | ||
165faf16 JA |
337 | /* |
338 | * Set io data pointers. | |
339 | */ | |
cec6b55d JA |
340 | io_u->xfer_buf = io_u->buf; |
341 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 342 | |
02bcaa8c | 343 | fio_gettime(&io_u->start_time, NULL); |
10ba535a JA |
344 | return io_u; |
345 | } | |
346 | ||
347 | void io_completed(struct thread_data *td, struct io_u *io_u, | |
348 | struct io_completion_data *icd) | |
349 | { | |
10ba535a JA |
350 | unsigned long msec; |
351 | ||
87dc1ab1 JA |
352 | if (io_u->ddir == DDIR_SYNC) { |
353 | td->last_was_sync = 1; | |
354 | return; | |
355 | } | |
356 | ||
357 | td->last_was_sync = 0; | |
358 | ||
10ba535a JA |
359 | if (!io_u->error) { |
360 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 361 | const enum fio_ddir idx = io_u->ddir; |
10ba535a JA |
362 | |
363 | td->io_blocks[idx]++; | |
364 | td->io_bytes[idx] += bytes; | |
365 | td->zone_bytes += bytes; | |
366 | td->this_io_bytes[idx] += bytes; | |
367 | ||
02bcaa8c JA |
368 | io_u->file->last_completed_pos = io_u->offset + io_u->buflen; |
369 | ||
370 | msec = mtime_since(&io_u->issue_time, &icd->time); | |
10ba535a JA |
371 | |
372 | add_clat_sample(td, idx, msec); | |
02bcaa8c | 373 | add_bw_sample(td, idx, &icd->time); |
10ba535a JA |
374 | |
375 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) | |
376 | log_io_piece(td, io_u); | |
377 | ||
378 | icd->bytes_done[idx] += bytes; | |
379 | } else | |
380 | icd->error = io_u->error; | |
381 | } | |
382 | ||
383 | void ios_completed(struct thread_data *td, struct io_completion_data *icd) | |
384 | { | |
385 | struct io_u *io_u; | |
386 | int i; | |
387 | ||
02bcaa8c JA |
388 | fio_gettime(&icd->time, NULL); |
389 | ||
10ba535a JA |
390 | icd->error = 0; |
391 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
392 | ||
393 | for (i = 0; i < icd->nr; i++) { | |
394 | io_u = td->io_ops->event(td, i); | |
395 | ||
396 | io_completed(td, io_u, icd); | |
397 | put_io_u(td, io_u); | |
398 | } | |
399 | } |