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> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
9 | #include "os.h" | |
10 | ||
5945b9b4 JA |
11 | /* |
12 | * Change this define to play with the timeout handling | |
13 | */ | |
14 | #undef FIO_USE_TIMEOUT | |
15 | ||
97601024 JA |
16 | struct io_completion_data { |
17 | int nr; /* input */ | |
97601024 JA |
18 | |
19 | int error; /* output */ | |
20 | unsigned long bytes_done[2]; /* output */ | |
21 | struct timeval time; /* output */ | |
22 | }; | |
23 | ||
10ba535a JA |
24 | /* |
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. | |
27 | */ | |
28 | static int random_map_free(struct thread_data *td, struct fio_file *f, | |
29 | unsigned long long block) | |
30 | { | |
31 | unsigned int idx = RAND_MAP_IDX(td, f, block); | |
32 | unsigned int bit = RAND_MAP_BIT(td, f, block); | |
33 | ||
34 | return (f->file_map[idx] & (1UL << bit)) == 0; | |
35 | } | |
36 | ||
df415585 JA |
37 | /* |
38 | * Mark a given offset as used in the map. | |
39 | */ | |
9bf2061e | 40 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 41 | { |
2dc1bbeb | 42 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 43 | struct fio_file *f = io_u->file; |
a00735e6 JA |
44 | unsigned long long block; |
45 | unsigned int blocks; | |
c685b5b2 | 46 | unsigned int nr_blocks; |
df415585 | 47 | |
a00735e6 JA |
48 | block = io_u->offset / (unsigned long long) min_bs; |
49 | blocks = 0; | |
c685b5b2 JA |
50 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
51 | ||
52 | while (blocks < nr_blocks) { | |
df415585 JA |
53 | unsigned int idx, bit; |
54 | ||
211097b2 | 55 | if (!td->o.ddir_nr && !random_map_free(td, f, block)) |
df415585 JA |
56 | break; |
57 | ||
58 | idx = RAND_MAP_IDX(td, f, block); | |
59 | bit = RAND_MAP_BIT(td, f, block); | |
60 | ||
0032bf9f | 61 | fio_assert(td, idx < f->num_maps); |
df415585 JA |
62 | |
63 | f->file_map[idx] |= (1UL << bit); | |
64 | block++; | |
65 | blocks++; | |
66 | } | |
67 | ||
a00735e6 JA |
68 | if ((blocks * min_bs) < io_u->buflen) |
69 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
70 | } |
71 | ||
10ba535a JA |
72 | /* |
73 | * Return the next free block in the map. | |
74 | */ | |
75 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
76 | unsigned long long *b) | |
77 | { | |
78 | int i; | |
79 | ||
c685b5b2 JA |
80 | i = f->last_free_lookup; |
81 | *b = (i * BLOCKS_PER_MAP); | |
2dc1bbeb | 82 | while ((*b) * td->o.rw_min_bs < f->real_file_size) { |
10ba535a JA |
83 | if (f->file_map[i] != -1UL) { |
84 | *b += ffz(f->file_map[i]); | |
c685b5b2 | 85 | f->last_free_lookup = i; |
10ba535a JA |
86 | return 0; |
87 | } | |
88 | ||
89 | *b += BLOCKS_PER_MAP; | |
90 | i++; | |
91 | } | |
92 | ||
93 | return 1; | |
94 | } | |
95 | ||
96 | /* | |
97 | * For random io, generate a random new block and see if it's used. Repeat | |
98 | * until we find a free one. For sequential io, just return the end of | |
99 | * the last io issued. | |
100 | */ | |
9bf2061e | 101 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 102 | { |
9bf2061e | 103 | struct fio_file *f = io_u->file; |
c685b5b2 | 104 | const int ddir = io_u->ddir; |
10ba535a JA |
105 | unsigned long long b, rb; |
106 | long r; | |
107 | ||
413dd459 | 108 | if (td_random(td)) { |
2dc1bbeb | 109 | unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir]; |
e6433576 | 110 | int loops = 5; |
10ba535a | 111 | |
211097b2 JA |
112 | if (td->o.ddir_nr) { |
113 | if (!td->ddir_nr) | |
114 | td->ddir_nr = td->o.ddir_nr; | |
115 | else if (--td->ddir_nr) { | |
116 | b = f->last_pos / td->o.min_bs[ddir]; | |
117 | goto out; | |
118 | } else | |
119 | td->ddir_nr = td->o.ddir_nr; | |
120 | } | |
121 | ||
10ba535a JA |
122 | do { |
123 | r = os_random_long(&td->random_state); | |
2393acba JA |
124 | if (!max_blocks) |
125 | b = 0; | |
126 | else | |
127 | b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
2dc1bbeb | 128 | if (td->o.norandommap) |
bb8895e0 | 129 | break; |
2dc1bbeb | 130 | rb = b + (f->file_offset / td->o.min_bs[ddir]); |
10ba535a JA |
131 | loops--; |
132 | } while (!random_map_free(td, f, rb) && loops); | |
133 | ||
bca4ed4d JA |
134 | /* |
135 | * if we failed to retrieve a truly random offset within | |
136 | * the loops assigned, see if there are free ones left at all | |
137 | */ | |
138 | if (!loops && get_next_free_block(td, f, &b)) | |
139 | return 1; | |
10ba535a | 140 | } else |
2dc1bbeb | 141 | b = f->last_pos / td->o.min_bs[ddir]; |
10ba535a | 142 | |
211097b2 | 143 | out: |
2dc1bbeb | 144 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
bca4ed4d | 145 | if (io_u->offset >= f->real_file_size) |
10ba535a JA |
146 | return 1; |
147 | ||
148 | return 0; | |
149 | } | |
150 | ||
9bf2061e | 151 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 152 | { |
9bf2061e | 153 | struct fio_file *f = io_u->file; |
bca4ed4d | 154 | const int ddir = io_u->ddir; |
10ba535a JA |
155 | unsigned int buflen; |
156 | long r; | |
157 | ||
2dc1bbeb JA |
158 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
159 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
160 | else { |
161 | r = os_random_long(&td->bsrange_state); | |
2dc1bbeb JA |
162 | buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); |
163 | if (!td->o.bs_unaligned) | |
164 | buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); | |
10ba535a JA |
165 | } |
166 | ||
bca4ed4d | 167 | while (buflen + io_u->offset > f->real_file_size) { |
2dc1bbeb JA |
168 | if (buflen == td->o.min_bs[ddir]) { |
169 | if (!td->o.odirect) { | |
2393acba JA |
170 | assert(io_u->offset <= f->real_file_size); |
171 | buflen = f->real_file_size - io_u->offset; | |
172 | return buflen; | |
173 | } | |
bca4ed4d | 174 | return 0; |
2393acba | 175 | } |
bca4ed4d | 176 | |
2dc1bbeb | 177 | buflen = td->o.min_bs[ddir]; |
10ba535a JA |
178 | } |
179 | ||
180 | return buflen; | |
181 | } | |
182 | ||
afe24a5a JA |
183 | static void set_rwmix_bytes(struct thread_data *td) |
184 | { | |
185 | unsigned long long rbytes; | |
186 | unsigned int diff; | |
187 | ||
188 | /* | |
189 | * we do time or byte based switch. this is needed because | |
190 | * buffered writes may issue a lot quicker than they complete, | |
191 | * whereas reads do not. | |
192 | */ | |
193 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 194 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 195 | |
e47f799f JA |
196 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
197 | } | |
198 | ||
199 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
200 | { | |
201 | unsigned int v; | |
202 | long r; | |
203 | ||
204 | r = os_random_long(&td->rwmix_state); | |
205 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
206 | if (v < td->o.rwmix[DDIR_READ]) | |
207 | return DDIR_READ; | |
208 | ||
209 | return DDIR_WRITE; | |
afe24a5a JA |
210 | } |
211 | ||
10ba535a JA |
212 | /* |
213 | * Return the data direction for the next io_u. If the job is a | |
214 | * mixed read/write workload, check the rwmix cycle and switch if | |
215 | * necessary. | |
216 | */ | |
1e97cce9 | 217 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
218 | { |
219 | if (td_rw(td)) { | |
220 | struct timeval now; | |
221 | unsigned long elapsed; | |
afe24a5a | 222 | unsigned int cycle; |
10ba535a | 223 | |
02bcaa8c | 224 | fio_gettime(&now, NULL); |
10ba535a JA |
225 | elapsed = mtime_since_now(&td->rwmix_switch); |
226 | ||
e47f799f JA |
227 | /* |
228 | * if this is the first cycle, make it shorter | |
229 | */ | |
afe24a5a JA |
230 | cycle = td->o.rwmixcycle; |
231 | if (!td->rwmix_bytes) | |
232 | cycle /= 10; | |
233 | ||
10ba535a JA |
234 | /* |
235 | * Check if it's time to seed a new data direction. | |
236 | */ | |
e47f799f | 237 | if (elapsed >= cycle || |
afe24a5a | 238 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
239 | unsigned long long max_bytes; |
240 | enum fio_ddir ddir; | |
241 | ||
242 | /* | |
243 | * Put a top limit on how many bytes we do for | |
244 | * one data direction, to avoid overflowing the | |
245 | * ranges too much | |
246 | */ | |
247 | ddir = get_rand_ddir(td); | |
248 | max_bytes = td->this_io_bytes[ddir]; | |
38d77cae JA |
249 | if (max_bytes >= (td->io_size * td->o.rwmix[ddir] / 100)) { |
250 | if (!td->rw_end_set[ddir]) { | |
251 | td->rw_end_set[ddir] = 1; | |
252 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
253 | } | |
e47f799f | 254 | ddir ^= 1; |
38d77cae | 255 | } |
e47f799f JA |
256 | |
257 | if (ddir != td->rwmix_ddir) | |
258 | set_rwmix_bytes(td); | |
259 | ||
260 | td->rwmix_ddir = ddir; | |
10ba535a JA |
261 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
262 | } | |
263 | return td->rwmix_ddir; | |
264 | } else if (td_read(td)) | |
265 | return DDIR_READ; | |
266 | else | |
267 | return DDIR_WRITE; | |
268 | } | |
269 | ||
10ba535a JA |
270 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
271 | { | |
0c6e7517 JA |
272 | assert((io_u->flags & IO_U_F_FREE) == 0); |
273 | io_u->flags |= IO_U_F_FREE; | |
274 | ||
10ba535a JA |
275 | io_u->file = NULL; |
276 | list_del(&io_u->list); | |
277 | list_add(&io_u->list, &td->io_u_freelist); | |
278 | td->cur_depth--; | |
279 | } | |
280 | ||
755200a3 JA |
281 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
282 | { | |
283 | struct io_u *__io_u = *io_u; | |
284 | ||
4d2e0f49 JA |
285 | __io_u->flags |= IO_U_F_FREE; |
286 | __io_u->flags &= ~IO_U_F_FLIGHT; | |
287 | ||
755200a3 JA |
288 | list_del(&__io_u->list); |
289 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
290 | td->cur_depth--; | |
291 | *io_u = NULL; | |
292 | } | |
293 | ||
9bf2061e | 294 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a JA |
295 | { |
296 | /* | |
297 | * If using an iolog, grab next piece if any available. | |
298 | */ | |
2dc1bbeb | 299 | if (td->o.read_iolog) |
10ba535a JA |
300 | return read_iolog_get(td, io_u); |
301 | ||
87dc1ab1 JA |
302 | /* |
303 | * see if it's time to sync | |
304 | */ | |
2dc1bbeb JA |
305 | if (td->o.fsync_blocks && |
306 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
307 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 308 | io_u->ddir = DDIR_SYNC; |
87dc1ab1 JA |
309 | return 0; |
310 | } | |
311 | ||
a00735e6 JA |
312 | io_u->ddir = get_rw_ddir(td); |
313 | ||
10ba535a | 314 | /* |
c685b5b2 JA |
315 | * No log, let the seq/rand engine retrieve the next buflen and |
316 | * position. | |
10ba535a | 317 | */ |
9bf2061e | 318 | if (get_next_offset(td, io_u)) |
bca4ed4d | 319 | return 1; |
10ba535a | 320 | |
9bf2061e | 321 | io_u->buflen = get_next_buflen(td, io_u); |
bca4ed4d JA |
322 | if (!io_u->buflen) |
323 | return 1; | |
324 | ||
325 | /* | |
326 | * mark entry before potentially trimming io_u | |
327 | */ | |
2dc1bbeb | 328 | if (!td->o.read_iolog && td_random(td) && !td->o.norandommap) |
9bf2061e | 329 | mark_random_map(td, io_u); |
bca4ed4d JA |
330 | |
331 | /* | |
332 | * If using a write iolog, store this entry. | |
333 | */ | |
2dc1bbeb | 334 | if (td->o.write_iolog_file) |
bca4ed4d JA |
335 | write_iolog_put(td, io_u); |
336 | ||
bca4ed4d | 337 | return 0; |
10ba535a JA |
338 | } |
339 | ||
b3605062 | 340 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
341 | { |
342 | int index = 0; | |
343 | ||
b3605062 JA |
344 | if (io_u->ddir == DDIR_SYNC) |
345 | return; | |
346 | ||
71619dc2 JA |
347 | switch (td->cur_depth) { |
348 | default: | |
349 | index++; | |
350 | case 32 ... 63: | |
351 | index++; | |
352 | case 16 ... 31: | |
353 | index++; | |
354 | case 8 ... 15: | |
355 | index++; | |
356 | case 4 ... 7: | |
357 | index++; | |
358 | case 2 ... 3: | |
359 | index++; | |
360 | case 1: | |
361 | break; | |
362 | } | |
363 | ||
756867bd | 364 | td->ts.io_u_map[index]++; |
b3605062 | 365 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
366 | } |
367 | ||
ec118304 JA |
368 | static void io_u_mark_latency(struct thread_data *td, unsigned long msec) |
369 | { | |
370 | int index = 0; | |
371 | ||
372 | switch (msec) { | |
373 | default: | |
374 | index++; | |
8abdce66 | 375 | case 1000 ... 1999: |
ec118304 | 376 | index++; |
8abdce66 | 377 | case 750 ... 999: |
ec118304 | 378 | index++; |
8abdce66 | 379 | case 500 ... 749: |
ec118304 | 380 | index++; |
8abdce66 | 381 | case 250 ... 499: |
ec118304 | 382 | index++; |
8abdce66 | 383 | case 100 ... 249: |
ec118304 | 384 | index++; |
8abdce66 | 385 | case 50 ... 99: |
ec118304 | 386 | index++; |
8abdce66 | 387 | case 20 ... 49: |
ec118304 | 388 | index++; |
8abdce66 | 389 | case 10 ... 19: |
ec118304 | 390 | index++; |
8abdce66 | 391 | case 4 ... 9: |
ec118304 JA |
392 | index++; |
393 | case 2 ... 3: | |
394 | index++; | |
395 | case 0 ... 1: | |
396 | break; | |
397 | } | |
398 | ||
756867bd | 399 | td->ts.io_u_lat[index]++; |
ec118304 JA |
400 | } |
401 | ||
0aabe160 JA |
402 | /* |
403 | * Get next file to service by choosing one at random | |
404 | */ | |
1c178180 JA |
405 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
406 | int badf) | |
0aabe160 | 407 | { |
0aabe160 | 408 | struct fio_file *f; |
1c178180 | 409 | int fno; |
0aabe160 JA |
410 | |
411 | do { | |
7c83c089 JA |
412 | long r = os_random_long(&td->next_file_state); |
413 | ||
2dc1bbeb | 414 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 JA |
415 | f = &td->files[fno]; |
416 | ||
417 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) | |
0aabe160 JA |
418 | return f; |
419 | } while (1); | |
420 | } | |
421 | ||
422 | /* | |
423 | * Get next file to service by doing round robin between all available ones | |
424 | */ | |
1c178180 JA |
425 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
426 | int badf) | |
3d7c391d JA |
427 | { |
428 | unsigned int old_next_file = td->next_file; | |
429 | struct fio_file *f; | |
430 | ||
431 | do { | |
432 | f = &td->files[td->next_file]; | |
433 | ||
434 | td->next_file++; | |
2dc1bbeb | 435 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
436 | td->next_file = 0; |
437 | ||
1c178180 | 438 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
439 | break; |
440 | ||
441 | f = NULL; | |
442 | } while (td->next_file != old_next_file); | |
443 | ||
444 | return f; | |
445 | } | |
446 | ||
bdb4e2e9 JA |
447 | static struct fio_file *get_next_file(struct thread_data *td) |
448 | { | |
1907dbc6 JA |
449 | struct fio_file *f; |
450 | ||
2dc1bbeb | 451 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 452 | |
bdb4e2e9 JA |
453 | if (!td->nr_open_files) |
454 | return NULL; | |
455 | ||
1907dbc6 | 456 | f = td->file_service_file; |
f11bd94d | 457 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
1907dbc6 JA |
458 | return f; |
459 | ||
2dc1bbeb | 460 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 461 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 462 | else |
1c178180 | 463 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
464 | |
465 | td->file_service_file = f; | |
466 | td->file_service_left = td->file_service_nr - 1; | |
467 | return f; | |
bdb4e2e9 JA |
468 | } |
469 | ||
1c178180 JA |
470 | static struct fio_file *find_next_new_file(struct thread_data *td) |
471 | { | |
472 | struct fio_file *f; | |
473 | ||
2dc1bbeb | 474 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
475 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
476 | else | |
477 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
478 | ||
479 | return f; | |
480 | } | |
481 | ||
10ba535a JA |
482 | struct io_u *__get_io_u(struct thread_data *td) |
483 | { | |
484 | struct io_u *io_u = NULL; | |
485 | ||
755200a3 JA |
486 | if (!list_empty(&td->io_u_requeues)) |
487 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
488 | else if (!queue_full(td)) { | |
10ba535a JA |
489 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
490 | ||
6040dabc | 491 | io_u->buflen = 0; |
10ba535a | 492 | io_u->resid = 0; |
755200a3 | 493 | io_u->file = NULL; |
d7762cf8 | 494 | io_u->end_io = NULL; |
755200a3 JA |
495 | } |
496 | ||
497 | if (io_u) { | |
0c6e7517 JA |
498 | assert(io_u->flags & IO_U_F_FREE); |
499 | io_u->flags &= ~IO_U_F_FREE; | |
500 | ||
755200a3 | 501 | io_u->error = 0; |
10ba535a JA |
502 | list_del(&io_u->list); |
503 | list_add(&io_u->list, &td->io_u_busylist); | |
504 | td->cur_depth++; | |
505 | } | |
506 | ||
507 | return io_u; | |
508 | } | |
509 | ||
510 | /* | |
511 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
512 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
513 | */ | |
3d7c391d | 514 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 515 | { |
3d7c391d | 516 | struct fio_file *f; |
10ba535a | 517 | struct io_u *io_u; |
1c178180 | 518 | int ret; |
10ba535a JA |
519 | |
520 | io_u = __get_io_u(td); | |
521 | if (!io_u) | |
522 | return NULL; | |
523 | ||
755200a3 JA |
524 | /* |
525 | * from a requeue, io_u already setup | |
526 | */ | |
527 | if (io_u->file) | |
77f392bf | 528 | goto out; |
755200a3 | 529 | |
bdb4e2e9 JA |
530 | do { |
531 | f = get_next_file(td); | |
532 | if (!f) { | |
533 | put_io_u(td, io_u); | |
534 | return NULL; | |
535 | } | |
3d7c391d | 536 | |
1c178180 | 537 | set_file: |
bdb4e2e9 | 538 | io_u->file = f; |
3d7c391d | 539 | |
bdb4e2e9 JA |
540 | if (!fill_io_u(td, io_u)) |
541 | break; | |
10ba535a | 542 | |
bdb4e2e9 JA |
543 | /* |
544 | * No more to do for this file, close it | |
545 | */ | |
546 | io_u->file = NULL; | |
b5af8293 JA |
547 | td_io_close_file(td, f); |
548 | ||
549 | /* | |
550 | * probably not the right place to do this, but see | |
551 | * if we need to open a new file | |
552 | */ | |
2dc1bbeb JA |
553 | if (td->nr_open_files < td->o.open_files && |
554 | td->o.open_files != td->o.nr_files) { | |
1c178180 | 555 | f = find_next_new_file(td); |
a978ba68 | 556 | |
1c178180 | 557 | if (!f || (ret = td_io_open_file(td, f))) { |
a978ba68 | 558 | put_io_u(td, io_u); |
e291dc12 | 559 | return NULL; |
a978ba68 | 560 | } |
1c178180 | 561 | goto set_file; |
a978ba68 | 562 | } |
bdb4e2e9 | 563 | } while (1); |
10ba535a | 564 | |
2dc1bbeb | 565 | if (td->zone_bytes >= td->o.zone_size) { |
9bf2061e | 566 | td->zone_bytes = 0; |
2dc1bbeb | 567 | f->last_pos += td->o.zone_skip; |
9bf2061e JA |
568 | } |
569 | ||
ee88470c | 570 | if (io_u->buflen + io_u->offset > f->real_file_size) { |
10ba535a JA |
571 | if (td->io_ops->flags & FIO_RAWIO) { |
572 | put_io_u(td, io_u); | |
573 | return NULL; | |
574 | } | |
575 | ||
ee88470c | 576 | io_u->buflen = f->real_file_size - io_u->offset; |
10ba535a JA |
577 | } |
578 | ||
87dc1ab1 JA |
579 | if (io_u->ddir != DDIR_SYNC) { |
580 | if (!io_u->buflen) { | |
581 | put_io_u(td, io_u); | |
582 | return NULL; | |
583 | } | |
10ba535a | 584 | |
36167d82 | 585 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 586 | |
2dc1bbeb | 587 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
588 | populate_verify_io_u(td, io_u); |
589 | } | |
10ba535a | 590 | |
165faf16 JA |
591 | /* |
592 | * Set io data pointers. | |
593 | */ | |
77f392bf | 594 | out: |
cec6b55d JA |
595 | io_u->xfer_buf = io_u->buf; |
596 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 597 | |
36167d82 JA |
598 | if (td_io_prep(td, io_u)) { |
599 | put_io_u(td, io_u); | |
600 | return NULL; | |
601 | } | |
602 | ||
02bcaa8c | 603 | fio_gettime(&io_u->start_time, NULL); |
10ba535a JA |
604 | return io_u; |
605 | } | |
606 | ||
5451792e JA |
607 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
608 | { | |
609 | const char *msg[] = { "read", "write", "sync" }; | |
610 | ||
611 | log_err("fio: io_u error"); | |
612 | ||
613 | if (io_u->file) | |
614 | log_err(" on file %s", io_u->file->file_name); | |
615 | ||
616 | log_err(": %s\n", strerror(io_u->error)); | |
617 | ||
618 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
619 | ||
620 | if (!td->error) | |
621 | td_verror(td, io_u->error, "io_u error"); | |
622 | } | |
623 | ||
97601024 JA |
624 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
625 | struct io_completion_data *icd) | |
10ba535a | 626 | { |
10ba535a JA |
627 | unsigned long msec; |
628 | ||
0c6e7517 JA |
629 | assert(io_u->flags & IO_U_F_FLIGHT); |
630 | io_u->flags &= ~IO_U_F_FLIGHT; | |
631 | ||
0ad920e7 JA |
632 | put_file(td, io_u->file); |
633 | ||
87dc1ab1 JA |
634 | if (io_u->ddir == DDIR_SYNC) { |
635 | td->last_was_sync = 1; | |
636 | return; | |
637 | } | |
638 | ||
639 | td->last_was_sync = 0; | |
640 | ||
10ba535a JA |
641 | if (!io_u->error) { |
642 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 643 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 644 | int ret; |
10ba535a JA |
645 | |
646 | td->io_blocks[idx]++; | |
647 | td->io_bytes[idx] += bytes; | |
648 | td->zone_bytes += bytes; | |
649 | td->this_io_bytes[idx] += bytes; | |
650 | ||
02bcaa8c JA |
651 | io_u->file->last_completed_pos = io_u->offset + io_u->buflen; |
652 | ||
653 | msec = mtime_since(&io_u->issue_time, &icd->time); | |
10ba535a JA |
654 | |
655 | add_clat_sample(td, idx, msec); | |
02bcaa8c | 656 | add_bw_sample(td, idx, &icd->time); |
ec118304 | 657 | io_u_mark_latency(td, msec); |
10ba535a JA |
658 | |
659 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) | |
660 | log_io_piece(td, io_u); | |
661 | ||
662 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 663 | |
d7762cf8 JA |
664 | if (io_u->end_io) { |
665 | ret = io_u->end_io(io_u); | |
3af6ef39 JA |
666 | if (ret && !icd->error) |
667 | icd->error = ret; | |
668 | } | |
5451792e | 669 | } else { |
10ba535a | 670 | icd->error = io_u->error; |
5451792e JA |
671 | io_u_log_error(td, io_u); |
672 | } | |
10ba535a JA |
673 | } |
674 | ||
d7762cf8 | 675 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 676 | { |
02bcaa8c JA |
677 | fio_gettime(&icd->time, NULL); |
678 | ||
3af6ef39 JA |
679 | icd->nr = nr; |
680 | ||
10ba535a JA |
681 | icd->error = 0; |
682 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
683 | } |
684 | ||
97601024 JA |
685 | static void ios_completed(struct thread_data *td, |
686 | struct io_completion_data *icd) | |
36167d82 JA |
687 | { |
688 | struct io_u *io_u; | |
689 | int i; | |
690 | ||
10ba535a JA |
691 | for (i = 0; i < icd->nr; i++) { |
692 | io_u = td->io_ops->event(td, i); | |
693 | ||
694 | io_completed(td, io_u, icd); | |
695 | put_io_u(td, io_u); | |
696 | } | |
697 | } | |
97601024 | 698 | |
e7e6cfb4 JA |
699 | /* |
700 | * Complete a single io_u for the sync engines. | |
701 | */ | |
d7762cf8 | 702 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
703 | { |
704 | struct io_completion_data icd; | |
705 | ||
d7762cf8 | 706 | init_icd(&icd, 1); |
97601024 JA |
707 | io_completed(td, io_u, &icd); |
708 | put_io_u(td, io_u); | |
709 | ||
710 | if (!icd.error) | |
711 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
712 | ||
37e974a5 | 713 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
714 | return -1; |
715 | } | |
716 | ||
e7e6cfb4 JA |
717 | /* |
718 | * Called to complete min_events number of io for the async engines. | |
719 | */ | |
d7762cf8 | 720 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 721 | { |
97601024 | 722 | struct io_completion_data icd; |
00de55ef | 723 | struct timespec *tvp = NULL; |
97601024 | 724 | int ret; |
4d06a338 | 725 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 726 | |
4d06a338 | 727 | if (!min_events) |
00de55ef | 728 | tvp = &ts; |
97601024 | 729 | |
00de55ef | 730 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 731 | if (ret < 0) { |
e1161c32 | 732 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
733 | return ret; |
734 | } else if (!ret) | |
735 | return ret; | |
736 | ||
d7762cf8 | 737 | init_icd(&icd, ret); |
97601024 JA |
738 | ios_completed(td, &icd); |
739 | if (!icd.error) | |
740 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
741 | ||
37e974a5 | 742 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
743 | return -1; |
744 | } | |
7e77dd02 JA |
745 | |
746 | /* | |
747 | * Call when io_u is really queued, to update the submission latency. | |
748 | */ | |
749 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
750 | { | |
751 | unsigned long slat_time; | |
752 | ||
753 | slat_time = mtime_since(&io_u->start_time, &io_u->issue_time); | |
754 | add_slat_sample(td, io_u->ddir, slat_time); | |
755 | } | |
433afcb4 | 756 | |
55bc9728 | 757 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
758 | void io_u_set_timeout(struct thread_data *td) |
759 | { | |
760 | assert(td->cur_depth); | |
761 | ||
762 | td->timer.it_interval.tv_sec = 0; | |
763 | td->timer.it_interval.tv_usec = 0; | |
764 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
765 | td->timer.it_value.tv_usec = 0; | |
766 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
767 | fio_gettime(&td->timeout_end, NULL); | |
768 | } | |
5945b9b4 JA |
769 | |
770 | static void io_u_dump(struct io_u *io_u) | |
771 | { | |
772 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
773 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
774 | ||
775 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
776 | 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); | |
777 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
778 | } | |
55bc9728 JA |
779 | #else |
780 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
781 | { | |
782 | } | |
783 | #endif | |
433afcb4 | 784 | |
55bc9728 | 785 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
786 | static void io_u_timeout_handler(int fio_unused sig) |
787 | { | |
788 | struct thread_data *td, *__td; | |
789 | pid_t pid = getpid(); | |
5945b9b4 JA |
790 | struct list_head *entry; |
791 | struct io_u *io_u; | |
433afcb4 JA |
792 | int i; |
793 | ||
794 | log_err("fio: io_u timeout\n"); | |
795 | ||
796 | /* | |
797 | * TLS would be nice... | |
798 | */ | |
799 | td = NULL; | |
800 | for_each_td(__td, i) { | |
801 | if (__td->pid == pid) { | |
802 | td = __td; | |
803 | break; | |
804 | } | |
805 | } | |
806 | ||
807 | if (!td) { | |
808 | log_err("fio: io_u timeout, can't find job\n"); | |
809 | exit(1); | |
810 | } | |
811 | ||
812 | if (!td->cur_depth) { | |
813 | log_err("fio: timeout without pending work?\n"); | |
814 | return; | |
815 | } | |
816 | ||
15506d09 | 817 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
818 | |
819 | list_for_each(entry, &td->io_u_busylist) { | |
820 | io_u = list_entry(entry, struct io_u, list); | |
821 | ||
822 | io_u_dump(io_u); | |
823 | } | |
824 | ||
825 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
826 | exit(1); |
827 | } | |
55bc9728 | 828 | #endif |
433afcb4 JA |
829 | |
830 | void io_u_init_timeout(void) | |
831 | { | |
55bc9728 | 832 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 833 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 834 | #endif |
433afcb4 | 835 | } |