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 | ||
1e3d53ac JA |
55 | /* |
56 | * If we have a mixed random workload, we may | |
57 | * encounter blocks we already did IO to. | |
58 | */ | |
211097b2 | 59 | if (!td->o.ddir_nr && !random_map_free(td, f, block)) |
df415585 JA |
60 | break; |
61 | ||
62 | idx = RAND_MAP_IDX(td, f, block); | |
63 | bit = RAND_MAP_BIT(td, f, block); | |
64 | ||
0032bf9f | 65 | fio_assert(td, idx < f->num_maps); |
df415585 JA |
66 | |
67 | f->file_map[idx] |= (1UL << bit); | |
68 | block++; | |
69 | blocks++; | |
70 | } | |
71 | ||
a00735e6 JA |
72 | if ((blocks * min_bs) < io_u->buflen) |
73 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
74 | } |
75 | ||
10ba535a JA |
76 | /* |
77 | * Return the next free block in the map. | |
78 | */ | |
79 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
80 | unsigned long long *b) | |
81 | { | |
82 | int i; | |
83 | ||
c685b5b2 JA |
84 | i = f->last_free_lookup; |
85 | *b = (i * BLOCKS_PER_MAP); | |
2dc1bbeb | 86 | while ((*b) * td->o.rw_min_bs < f->real_file_size) { |
10ba535a JA |
87 | if (f->file_map[i] != -1UL) { |
88 | *b += ffz(f->file_map[i]); | |
c685b5b2 | 89 | f->last_free_lookup = i; |
10ba535a JA |
90 | return 0; |
91 | } | |
92 | ||
93 | *b += BLOCKS_PER_MAP; | |
94 | i++; | |
95 | } | |
96 | ||
97 | return 1; | |
98 | } | |
99 | ||
ec4015da JA |
100 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
101 | int ddir, unsigned long long *b) | |
102 | { | |
7bb48f84 | 103 | unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir]; |
ec4015da JA |
104 | unsigned long long r, rb; |
105 | int loops = 5; | |
106 | ||
107 | do { | |
108 | r = os_random_long(&td->random_state); | |
109 | if (!max_blocks) | |
110 | *b = 0; | |
111 | else | |
112 | *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
113 | if (td->o.norandommap) | |
114 | break; | |
115 | rb = *b + (f->file_offset / td->o.min_bs[ddir]); | |
116 | loops--; | |
117 | } while (!random_map_free(td, f, rb) && loops); | |
118 | ||
119 | /* | |
120 | * if we failed to retrieve a truly random offset within | |
121 | * the loops assigned, see if there are free ones left at all | |
122 | */ | |
123 | if (!loops && get_next_free_block(td, f, b)) | |
124 | return 1; | |
125 | ||
126 | return 0; | |
127 | } | |
128 | ||
10ba535a JA |
129 | /* |
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. | |
133 | */ | |
9bf2061e | 134 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 135 | { |
9bf2061e | 136 | struct fio_file *f = io_u->file; |
c685b5b2 | 137 | const int ddir = io_u->ddir; |
ec4015da | 138 | unsigned long long b; |
10ba535a | 139 | |
ec4015da JA |
140 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
141 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 142 | |
ec4015da | 143 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 144 | return 1; |
43063a1c JA |
145 | } else { |
146 | if (f->last_pos >= f->real_file_size) | |
147 | return 1; | |
148 | ||
2dc1bbeb | 149 | b = f->last_pos / td->o.min_bs[ddir]; |
43063a1c | 150 | } |
10ba535a | 151 | |
2dc1bbeb | 152 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
bca4ed4d | 153 | if (io_u->offset >= f->real_file_size) |
10ba535a JA |
154 | return 1; |
155 | ||
156 | return 0; | |
157 | } | |
158 | ||
9bf2061e | 159 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 160 | { |
9bf2061e | 161 | struct fio_file *f = io_u->file; |
bca4ed4d | 162 | const int ddir = io_u->ddir; |
10ba535a JA |
163 | unsigned int buflen; |
164 | long r; | |
165 | ||
2dc1bbeb JA |
166 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
167 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
168 | else { |
169 | r = os_random_long(&td->bsrange_state); | |
2dc1bbeb JA |
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); | |
10ba535a JA |
173 | } |
174 | ||
bca4ed4d | 175 | while (buflen + io_u->offset > f->real_file_size) { |
2dc1bbeb JA |
176 | if (buflen == td->o.min_bs[ddir]) { |
177 | if (!td->o.odirect) { | |
2393acba JA |
178 | assert(io_u->offset <= f->real_file_size); |
179 | buflen = f->real_file_size - io_u->offset; | |
180 | return buflen; | |
181 | } | |
bca4ed4d | 182 | return 0; |
2393acba | 183 | } |
bca4ed4d | 184 | |
2dc1bbeb | 185 | buflen = td->o.min_bs[ddir]; |
10ba535a JA |
186 | } |
187 | ||
188 | return buflen; | |
189 | } | |
190 | ||
afe24a5a JA |
191 | static void set_rwmix_bytes(struct thread_data *td) |
192 | { | |
193 | unsigned long long rbytes; | |
194 | unsigned int diff; | |
195 | ||
196 | /* | |
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. | |
200 | */ | |
201 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 202 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 203 | |
e47f799f JA |
204 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
205 | } | |
206 | ||
207 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
208 | { | |
209 | unsigned int v; | |
210 | long r; | |
211 | ||
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]) | |
215 | return DDIR_READ; | |
216 | ||
217 | return DDIR_WRITE; | |
afe24a5a JA |
218 | } |
219 | ||
10ba535a JA |
220 | /* |
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 | |
223 | * necessary. | |
224 | */ | |
1e97cce9 | 225 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
226 | { |
227 | if (td_rw(td)) { | |
228 | struct timeval now; | |
229 | unsigned long elapsed; | |
afe24a5a | 230 | unsigned int cycle; |
10ba535a | 231 | |
02bcaa8c | 232 | fio_gettime(&now, NULL); |
10ba535a JA |
233 | elapsed = mtime_since_now(&td->rwmix_switch); |
234 | ||
e47f799f JA |
235 | /* |
236 | * if this is the first cycle, make it shorter | |
237 | */ | |
afe24a5a JA |
238 | cycle = td->o.rwmixcycle; |
239 | if (!td->rwmix_bytes) | |
240 | cycle /= 10; | |
241 | ||
10ba535a JA |
242 | /* |
243 | * Check if it's time to seed a new data direction. | |
244 | */ | |
e47f799f | 245 | if (elapsed >= cycle || |
afe24a5a | 246 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
247 | unsigned long long max_bytes; |
248 | enum fio_ddir ddir; | |
249 | ||
250 | /* | |
251 | * Put a top limit on how many bytes we do for | |
252 | * one data direction, to avoid overflowing the | |
253 | * ranges too much | |
254 | */ | |
255 | ddir = get_rand_ddir(td); | |
256 | max_bytes = td->this_io_bytes[ddir]; | |
7bb48f84 | 257 | if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) { |
38d77cae JA |
258 | if (!td->rw_end_set[ddir]) { |
259 | td->rw_end_set[ddir] = 1; | |
260 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
261 | } | |
e47f799f | 262 | ddir ^= 1; |
38d77cae | 263 | } |
e47f799f JA |
264 | |
265 | if (ddir != td->rwmix_ddir) | |
266 | set_rwmix_bytes(td); | |
267 | ||
268 | td->rwmix_ddir = ddir; | |
10ba535a JA |
269 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
270 | } | |
271 | return td->rwmix_ddir; | |
272 | } else if (td_read(td)) | |
273 | return DDIR_READ; | |
274 | else | |
275 | return DDIR_WRITE; | |
276 | } | |
277 | ||
10ba535a JA |
278 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
279 | { | |
0c6e7517 JA |
280 | assert((io_u->flags & IO_U_F_FREE) == 0); |
281 | io_u->flags |= IO_U_F_FREE; | |
282 | ||
10ba535a JA |
283 | io_u->file = NULL; |
284 | list_del(&io_u->list); | |
285 | list_add(&io_u->list, &td->io_u_freelist); | |
286 | td->cur_depth--; | |
287 | } | |
288 | ||
755200a3 JA |
289 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
290 | { | |
291 | struct io_u *__io_u = *io_u; | |
292 | ||
4d2e0f49 JA |
293 | __io_u->flags |= IO_U_F_FREE; |
294 | __io_u->flags &= ~IO_U_F_FLIGHT; | |
295 | ||
755200a3 JA |
296 | list_del(&__io_u->list); |
297 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
298 | td->cur_depth--; | |
299 | *io_u = NULL; | |
300 | } | |
301 | ||
9bf2061e | 302 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a JA |
303 | { |
304 | /* | |
305 | * If using an iolog, grab next piece if any available. | |
306 | */ | |
2dc1bbeb | 307 | if (td->o.read_iolog) |
10ba535a JA |
308 | return read_iolog_get(td, io_u); |
309 | ||
87dc1ab1 JA |
310 | /* |
311 | * see if it's time to sync | |
312 | */ | |
2dc1bbeb JA |
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)) { | |
87dc1ab1 | 316 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 317 | goto out; |
87dc1ab1 JA |
318 | } |
319 | ||
a00735e6 JA |
320 | io_u->ddir = get_rw_ddir(td); |
321 | ||
10ba535a | 322 | /* |
c685b5b2 JA |
323 | * No log, let the seq/rand engine retrieve the next buflen and |
324 | * position. | |
10ba535a | 325 | */ |
9bf2061e | 326 | if (get_next_offset(td, io_u)) |
bca4ed4d | 327 | return 1; |
10ba535a | 328 | |
9bf2061e | 329 | io_u->buflen = get_next_buflen(td, io_u); |
bca4ed4d JA |
330 | if (!io_u->buflen) |
331 | return 1; | |
332 | ||
333 | /* | |
334 | * mark entry before potentially trimming io_u | |
335 | */ | |
8fd15a9a | 336 | if (td_random(td) && !td->o.norandommap) |
9bf2061e | 337 | mark_random_map(td, io_u); |
bca4ed4d JA |
338 | |
339 | /* | |
340 | * If using a write iolog, store this entry. | |
341 | */ | |
c38e9468 | 342 | out: |
2dc1bbeb | 343 | if (td->o.write_iolog_file) |
bca4ed4d JA |
344 | write_iolog_put(td, io_u); |
345 | ||
bca4ed4d | 346 | return 0; |
10ba535a JA |
347 | } |
348 | ||
b3605062 | 349 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
350 | { |
351 | int index = 0; | |
352 | ||
b3605062 JA |
353 | if (io_u->ddir == DDIR_SYNC) |
354 | return; | |
355 | ||
71619dc2 JA |
356 | switch (td->cur_depth) { |
357 | default: | |
358 | index++; | |
359 | case 32 ... 63: | |
360 | index++; | |
361 | case 16 ... 31: | |
362 | index++; | |
363 | case 8 ... 15: | |
364 | index++; | |
365 | case 4 ... 7: | |
366 | index++; | |
367 | case 2 ... 3: | |
368 | index++; | |
369 | case 1: | |
370 | break; | |
371 | } | |
372 | ||
756867bd | 373 | td->ts.io_u_map[index]++; |
b3605062 | 374 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
375 | } |
376 | ||
ec118304 JA |
377 | static void io_u_mark_latency(struct thread_data *td, unsigned long msec) |
378 | { | |
379 | int index = 0; | |
380 | ||
381 | switch (msec) { | |
382 | default: | |
383 | index++; | |
8abdce66 | 384 | case 1000 ... 1999: |
ec118304 | 385 | index++; |
8abdce66 | 386 | case 750 ... 999: |
ec118304 | 387 | index++; |
8abdce66 | 388 | case 500 ... 749: |
ec118304 | 389 | index++; |
8abdce66 | 390 | case 250 ... 499: |
ec118304 | 391 | index++; |
8abdce66 | 392 | case 100 ... 249: |
ec118304 | 393 | index++; |
8abdce66 | 394 | case 50 ... 99: |
ec118304 | 395 | index++; |
8abdce66 | 396 | case 20 ... 49: |
ec118304 | 397 | index++; |
8abdce66 | 398 | case 10 ... 19: |
ec118304 | 399 | index++; |
8abdce66 | 400 | case 4 ... 9: |
ec118304 JA |
401 | index++; |
402 | case 2 ... 3: | |
403 | index++; | |
404 | case 0 ... 1: | |
405 | break; | |
406 | } | |
407 | ||
756867bd | 408 | td->ts.io_u_lat[index]++; |
ec118304 JA |
409 | } |
410 | ||
0aabe160 JA |
411 | /* |
412 | * Get next file to service by choosing one at random | |
413 | */ | |
1c178180 JA |
414 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
415 | int badf) | |
0aabe160 | 416 | { |
0aabe160 | 417 | struct fio_file *f; |
1c178180 | 418 | int fno; |
0aabe160 JA |
419 | |
420 | do { | |
7c83c089 JA |
421 | long r = os_random_long(&td->next_file_state); |
422 | ||
2dc1bbeb | 423 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 JA |
424 | f = &td->files[fno]; |
425 | ||
426 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) | |
0aabe160 JA |
427 | return f; |
428 | } while (1); | |
429 | } | |
430 | ||
431 | /* | |
432 | * Get next file to service by doing round robin between all available ones | |
433 | */ | |
1c178180 JA |
434 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
435 | int badf) | |
3d7c391d JA |
436 | { |
437 | unsigned int old_next_file = td->next_file; | |
438 | struct fio_file *f; | |
439 | ||
440 | do { | |
441 | f = &td->files[td->next_file]; | |
442 | ||
443 | td->next_file++; | |
2dc1bbeb | 444 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
445 | td->next_file = 0; |
446 | ||
1c178180 | 447 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
448 | break; |
449 | ||
450 | f = NULL; | |
451 | } while (td->next_file != old_next_file); | |
452 | ||
453 | return f; | |
454 | } | |
455 | ||
bdb4e2e9 JA |
456 | static struct fio_file *get_next_file(struct thread_data *td) |
457 | { | |
1907dbc6 JA |
458 | struct fio_file *f; |
459 | ||
2dc1bbeb | 460 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 461 | |
bdb4e2e9 JA |
462 | if (!td->nr_open_files) |
463 | return NULL; | |
464 | ||
1907dbc6 | 465 | f = td->file_service_file; |
f11bd94d | 466 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
1907dbc6 JA |
467 | return f; |
468 | ||
2dc1bbeb | 469 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 470 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 471 | else |
1c178180 | 472 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
473 | |
474 | td->file_service_file = f; | |
475 | td->file_service_left = td->file_service_nr - 1; | |
476 | return f; | |
bdb4e2e9 JA |
477 | } |
478 | ||
1c178180 JA |
479 | static struct fio_file *find_next_new_file(struct thread_data *td) |
480 | { | |
481 | struct fio_file *f; | |
482 | ||
2dc1bbeb | 483 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
484 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
485 | else | |
486 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
487 | ||
488 | return f; | |
489 | } | |
490 | ||
10ba535a JA |
491 | struct io_u *__get_io_u(struct thread_data *td) |
492 | { | |
493 | struct io_u *io_u = NULL; | |
494 | ||
755200a3 JA |
495 | if (!list_empty(&td->io_u_requeues)) |
496 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
497 | else if (!queue_full(td)) { | |
10ba535a JA |
498 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
499 | ||
6040dabc | 500 | io_u->buflen = 0; |
10ba535a | 501 | io_u->resid = 0; |
755200a3 | 502 | io_u->file = NULL; |
d7762cf8 | 503 | io_u->end_io = NULL; |
755200a3 JA |
504 | } |
505 | ||
506 | if (io_u) { | |
0c6e7517 JA |
507 | assert(io_u->flags & IO_U_F_FREE); |
508 | io_u->flags &= ~IO_U_F_FREE; | |
509 | ||
755200a3 | 510 | io_u->error = 0; |
10ba535a JA |
511 | list_del(&io_u->list); |
512 | list_add(&io_u->list, &td->io_u_busylist); | |
513 | td->cur_depth++; | |
514 | } | |
515 | ||
516 | return io_u; | |
517 | } | |
518 | ||
519 | /* | |
520 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
521 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
522 | */ | |
3d7c391d | 523 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 524 | { |
3d7c391d | 525 | struct fio_file *f; |
10ba535a | 526 | struct io_u *io_u; |
1c178180 | 527 | int ret; |
10ba535a JA |
528 | |
529 | io_u = __get_io_u(td); | |
530 | if (!io_u) | |
531 | return NULL; | |
532 | ||
755200a3 JA |
533 | /* |
534 | * from a requeue, io_u already setup | |
535 | */ | |
536 | if (io_u->file) | |
77f392bf | 537 | goto out; |
755200a3 | 538 | |
bdb4e2e9 JA |
539 | do { |
540 | f = get_next_file(td); | |
541 | if (!f) { | |
542 | put_io_u(td, io_u); | |
543 | return NULL; | |
544 | } | |
3d7c391d | 545 | |
1c178180 | 546 | set_file: |
bdb4e2e9 | 547 | io_u->file = f; |
3d7c391d | 548 | |
bdb4e2e9 JA |
549 | if (!fill_io_u(td, io_u)) |
550 | break; | |
10ba535a | 551 | |
bdb4e2e9 JA |
552 | /* |
553 | * No more to do for this file, close it | |
554 | */ | |
555 | io_u->file = NULL; | |
b5af8293 JA |
556 | td_io_close_file(td, f); |
557 | ||
558 | /* | |
559 | * probably not the right place to do this, but see | |
560 | * if we need to open a new file | |
561 | */ | |
2dc1bbeb JA |
562 | if (td->nr_open_files < td->o.open_files && |
563 | td->o.open_files != td->o.nr_files) { | |
1c178180 | 564 | f = find_next_new_file(td); |
a978ba68 | 565 | |
1c178180 | 566 | if (!f || (ret = td_io_open_file(td, f))) { |
a978ba68 | 567 | put_io_u(td, io_u); |
e291dc12 | 568 | return NULL; |
a978ba68 | 569 | } |
1c178180 | 570 | goto set_file; |
a978ba68 | 571 | } |
bdb4e2e9 | 572 | } while (1); |
10ba535a | 573 | |
2dc1bbeb | 574 | if (td->zone_bytes >= td->o.zone_size) { |
9bf2061e | 575 | td->zone_bytes = 0; |
2dc1bbeb | 576 | f->last_pos += td->o.zone_skip; |
9bf2061e JA |
577 | } |
578 | ||
ee88470c | 579 | if (io_u->buflen + io_u->offset > f->real_file_size) { |
10ba535a JA |
580 | if (td->io_ops->flags & FIO_RAWIO) { |
581 | put_io_u(td, io_u); | |
582 | return NULL; | |
583 | } | |
584 | ||
ee88470c | 585 | io_u->buflen = f->real_file_size - io_u->offset; |
10ba535a JA |
586 | } |
587 | ||
87dc1ab1 JA |
588 | if (io_u->ddir != DDIR_SYNC) { |
589 | if (!io_u->buflen) { | |
590 | put_io_u(td, io_u); | |
591 | return NULL; | |
592 | } | |
10ba535a | 593 | |
36167d82 | 594 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 595 | |
2dc1bbeb | 596 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
597 | populate_verify_io_u(td, io_u); |
598 | } | |
10ba535a | 599 | |
165faf16 JA |
600 | /* |
601 | * Set io data pointers. | |
602 | */ | |
77f392bf | 603 | out: |
cec6b55d JA |
604 | io_u->xfer_buf = io_u->buf; |
605 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 606 | |
36167d82 JA |
607 | if (td_io_prep(td, io_u)) { |
608 | put_io_u(td, io_u); | |
609 | return NULL; | |
610 | } | |
611 | ||
02bcaa8c | 612 | fio_gettime(&io_u->start_time, NULL); |
10ba535a JA |
613 | return io_u; |
614 | } | |
615 | ||
5451792e JA |
616 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
617 | { | |
618 | const char *msg[] = { "read", "write", "sync" }; | |
619 | ||
620 | log_err("fio: io_u error"); | |
621 | ||
622 | if (io_u->file) | |
623 | log_err(" on file %s", io_u->file->file_name); | |
624 | ||
625 | log_err(": %s\n", strerror(io_u->error)); | |
626 | ||
627 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
628 | ||
629 | if (!td->error) | |
630 | td_verror(td, io_u->error, "io_u error"); | |
631 | } | |
632 | ||
97601024 JA |
633 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
634 | struct io_completion_data *icd) | |
10ba535a | 635 | { |
10ba535a JA |
636 | unsigned long msec; |
637 | ||
0c6e7517 JA |
638 | assert(io_u->flags & IO_U_F_FLIGHT); |
639 | io_u->flags &= ~IO_U_F_FLIGHT; | |
640 | ||
0ad920e7 JA |
641 | put_file(td, io_u->file); |
642 | ||
87dc1ab1 JA |
643 | if (io_u->ddir == DDIR_SYNC) { |
644 | td->last_was_sync = 1; | |
645 | return; | |
646 | } | |
647 | ||
648 | td->last_was_sync = 0; | |
649 | ||
10ba535a JA |
650 | if (!io_u->error) { |
651 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 652 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 653 | int ret; |
10ba535a JA |
654 | |
655 | td->io_blocks[idx]++; | |
656 | td->io_bytes[idx] += bytes; | |
657 | td->zone_bytes += bytes; | |
658 | td->this_io_bytes[idx] += bytes; | |
659 | ||
02bcaa8c JA |
660 | io_u->file->last_completed_pos = io_u->offset + io_u->buflen; |
661 | ||
662 | msec = mtime_since(&io_u->issue_time, &icd->time); | |
10ba535a JA |
663 | |
664 | add_clat_sample(td, idx, msec); | |
02bcaa8c | 665 | add_bw_sample(td, idx, &icd->time); |
ec118304 | 666 | io_u_mark_latency(td, msec); |
10ba535a | 667 | |
41128405 JA |
668 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE && |
669 | td->o.verify != VERIFY_NONE) | |
10ba535a JA |
670 | log_io_piece(td, io_u); |
671 | ||
672 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 673 | |
d7762cf8 | 674 | if (io_u->end_io) { |
36690c9b | 675 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
676 | if (ret && !icd->error) |
677 | icd->error = ret; | |
678 | } | |
5451792e | 679 | } else { |
10ba535a | 680 | icd->error = io_u->error; |
5451792e JA |
681 | io_u_log_error(td, io_u); |
682 | } | |
10ba535a JA |
683 | } |
684 | ||
d7762cf8 | 685 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 686 | { |
02bcaa8c JA |
687 | fio_gettime(&icd->time, NULL); |
688 | ||
3af6ef39 JA |
689 | icd->nr = nr; |
690 | ||
10ba535a JA |
691 | icd->error = 0; |
692 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
693 | } |
694 | ||
97601024 JA |
695 | static void ios_completed(struct thread_data *td, |
696 | struct io_completion_data *icd) | |
36167d82 JA |
697 | { |
698 | struct io_u *io_u; | |
699 | int i; | |
700 | ||
10ba535a JA |
701 | for (i = 0; i < icd->nr; i++) { |
702 | io_u = td->io_ops->event(td, i); | |
703 | ||
704 | io_completed(td, io_u, icd); | |
705 | put_io_u(td, io_u); | |
706 | } | |
707 | } | |
97601024 | 708 | |
e7e6cfb4 JA |
709 | /* |
710 | * Complete a single io_u for the sync engines. | |
711 | */ | |
d7762cf8 | 712 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
713 | { |
714 | struct io_completion_data icd; | |
715 | ||
d7762cf8 | 716 | init_icd(&icd, 1); |
97601024 JA |
717 | io_completed(td, io_u, &icd); |
718 | put_io_u(td, io_u); | |
719 | ||
720 | if (!icd.error) | |
721 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
722 | ||
37e974a5 | 723 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
724 | return -1; |
725 | } | |
726 | ||
e7e6cfb4 JA |
727 | /* |
728 | * Called to complete min_events number of io for the async engines. | |
729 | */ | |
d7762cf8 | 730 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 731 | { |
97601024 | 732 | struct io_completion_data icd; |
00de55ef | 733 | struct timespec *tvp = NULL; |
97601024 | 734 | int ret; |
4d06a338 | 735 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 736 | |
4d06a338 | 737 | if (!min_events) |
00de55ef | 738 | tvp = &ts; |
97601024 | 739 | |
00de55ef | 740 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 741 | if (ret < 0) { |
e1161c32 | 742 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
743 | return ret; |
744 | } else if (!ret) | |
745 | return ret; | |
746 | ||
d7762cf8 | 747 | init_icd(&icd, ret); |
97601024 JA |
748 | ios_completed(td, &icd); |
749 | if (!icd.error) | |
750 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
751 | ||
37e974a5 | 752 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
753 | return -1; |
754 | } | |
7e77dd02 JA |
755 | |
756 | /* | |
757 | * Call when io_u is really queued, to update the submission latency. | |
758 | */ | |
759 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
760 | { | |
761 | unsigned long slat_time; | |
762 | ||
763 | slat_time = mtime_since(&io_u->start_time, &io_u->issue_time); | |
764 | add_slat_sample(td, io_u->ddir, slat_time); | |
765 | } | |
433afcb4 | 766 | |
55bc9728 | 767 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
768 | void io_u_set_timeout(struct thread_data *td) |
769 | { | |
770 | assert(td->cur_depth); | |
771 | ||
772 | td->timer.it_interval.tv_sec = 0; | |
773 | td->timer.it_interval.tv_usec = 0; | |
774 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
775 | td->timer.it_value.tv_usec = 0; | |
776 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
777 | fio_gettime(&td->timeout_end, NULL); | |
778 | } | |
5945b9b4 JA |
779 | |
780 | static void io_u_dump(struct io_u *io_u) | |
781 | { | |
782 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
783 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
784 | ||
785 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
786 | 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); | |
787 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
788 | } | |
55bc9728 JA |
789 | #else |
790 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
791 | { | |
792 | } | |
793 | #endif | |
433afcb4 | 794 | |
55bc9728 | 795 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
796 | static void io_u_timeout_handler(int fio_unused sig) |
797 | { | |
798 | struct thread_data *td, *__td; | |
799 | pid_t pid = getpid(); | |
5945b9b4 JA |
800 | struct list_head *entry; |
801 | struct io_u *io_u; | |
433afcb4 JA |
802 | int i; |
803 | ||
804 | log_err("fio: io_u timeout\n"); | |
805 | ||
806 | /* | |
807 | * TLS would be nice... | |
808 | */ | |
809 | td = NULL; | |
810 | for_each_td(__td, i) { | |
811 | if (__td->pid == pid) { | |
812 | td = __td; | |
813 | break; | |
814 | } | |
815 | } | |
816 | ||
817 | if (!td) { | |
818 | log_err("fio: io_u timeout, can't find job\n"); | |
819 | exit(1); | |
820 | } | |
821 | ||
822 | if (!td->cur_depth) { | |
823 | log_err("fio: timeout without pending work?\n"); | |
824 | return; | |
825 | } | |
826 | ||
15506d09 | 827 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
828 | |
829 | list_for_each(entry, &td->io_u_busylist) { | |
830 | io_u = list_entry(entry, struct io_u, list); | |
831 | ||
832 | io_u_dump(io_u); | |
833 | } | |
834 | ||
835 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
836 | exit(1); |
837 | } | |
55bc9728 | 838 | #endif |
433afcb4 JA |
839 | |
840 | void io_u_init_timeout(void) | |
841 | { | |
55bc9728 | 842 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 843 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 844 | #endif |
433afcb4 | 845 | } |