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 | ||
100 | /* | |
101 | * For random io, generate a random new block and see if it's used. Repeat | |
102 | * until we find a free one. For sequential io, just return the end of | |
103 | * the last io issued. | |
104 | */ | |
9bf2061e | 105 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 106 | { |
9bf2061e | 107 | struct fio_file *f = io_u->file; |
c685b5b2 | 108 | const int ddir = io_u->ddir; |
10ba535a JA |
109 | unsigned long long b, rb; |
110 | long r; | |
111 | ||
413dd459 | 112 | if (td_random(td)) { |
2dc1bbeb | 113 | unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir]; |
e6433576 | 114 | int loops = 5; |
10ba535a | 115 | |
211097b2 | 116 | if (td->o.ddir_nr) { |
1e3d53ac | 117 | if (!--td->ddir_nr) |
211097b2 | 118 | td->ddir_nr = td->o.ddir_nr; |
1e3d53ac | 119 | else { |
211097b2 JA |
120 | b = f->last_pos / td->o.min_bs[ddir]; |
121 | goto out; | |
1e3d53ac | 122 | } |
211097b2 JA |
123 | } |
124 | ||
10ba535a JA |
125 | do { |
126 | r = os_random_long(&td->random_state); | |
2393acba JA |
127 | if (!max_blocks) |
128 | b = 0; | |
129 | else | |
130 | b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
2dc1bbeb | 131 | if (td->o.norandommap) |
bb8895e0 | 132 | break; |
2dc1bbeb | 133 | rb = b + (f->file_offset / td->o.min_bs[ddir]); |
10ba535a JA |
134 | loops--; |
135 | } while (!random_map_free(td, f, rb) && loops); | |
136 | ||
bca4ed4d JA |
137 | /* |
138 | * if we failed to retrieve a truly random offset within | |
139 | * the loops assigned, see if there are free ones left at all | |
140 | */ | |
141 | if (!loops && get_next_free_block(td, f, &b)) | |
142 | return 1; | |
10ba535a | 143 | } else |
2dc1bbeb | 144 | b = f->last_pos / td->o.min_bs[ddir]; |
10ba535a | 145 | |
211097b2 | 146 | out: |
2dc1bbeb | 147 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
bca4ed4d | 148 | if (io_u->offset >= f->real_file_size) |
10ba535a JA |
149 | return 1; |
150 | ||
151 | return 0; | |
152 | } | |
153 | ||
9bf2061e | 154 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 155 | { |
9bf2061e | 156 | struct fio_file *f = io_u->file; |
bca4ed4d | 157 | const int ddir = io_u->ddir; |
10ba535a JA |
158 | unsigned int buflen; |
159 | long r; | |
160 | ||
2dc1bbeb JA |
161 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
162 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
163 | else { |
164 | r = os_random_long(&td->bsrange_state); | |
2dc1bbeb JA |
165 | buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); |
166 | if (!td->o.bs_unaligned) | |
167 | buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); | |
10ba535a JA |
168 | } |
169 | ||
bca4ed4d | 170 | while (buflen + io_u->offset > f->real_file_size) { |
2dc1bbeb JA |
171 | if (buflen == td->o.min_bs[ddir]) { |
172 | if (!td->o.odirect) { | |
2393acba JA |
173 | assert(io_u->offset <= f->real_file_size); |
174 | buflen = f->real_file_size - io_u->offset; | |
175 | return buflen; | |
176 | } | |
bca4ed4d | 177 | return 0; |
2393acba | 178 | } |
bca4ed4d | 179 | |
2dc1bbeb | 180 | buflen = td->o.min_bs[ddir]; |
10ba535a JA |
181 | } |
182 | ||
183 | return buflen; | |
184 | } | |
185 | ||
afe24a5a JA |
186 | static void set_rwmix_bytes(struct thread_data *td) |
187 | { | |
188 | unsigned long long rbytes; | |
189 | unsigned int diff; | |
190 | ||
191 | /* | |
192 | * we do time or byte based switch. this is needed because | |
193 | * buffered writes may issue a lot quicker than they complete, | |
194 | * whereas reads do not. | |
195 | */ | |
196 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 197 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 198 | |
e47f799f JA |
199 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
200 | } | |
201 | ||
202 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
203 | { | |
204 | unsigned int v; | |
205 | long r; | |
206 | ||
207 | r = os_random_long(&td->rwmix_state); | |
208 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
209 | if (v < td->o.rwmix[DDIR_READ]) | |
210 | return DDIR_READ; | |
211 | ||
212 | return DDIR_WRITE; | |
afe24a5a JA |
213 | } |
214 | ||
10ba535a JA |
215 | /* |
216 | * Return the data direction for the next io_u. If the job is a | |
217 | * mixed read/write workload, check the rwmix cycle and switch if | |
218 | * necessary. | |
219 | */ | |
1e97cce9 | 220 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
221 | { |
222 | if (td_rw(td)) { | |
223 | struct timeval now; | |
224 | unsigned long elapsed; | |
afe24a5a | 225 | unsigned int cycle; |
10ba535a | 226 | |
02bcaa8c | 227 | fio_gettime(&now, NULL); |
10ba535a JA |
228 | elapsed = mtime_since_now(&td->rwmix_switch); |
229 | ||
e47f799f JA |
230 | /* |
231 | * if this is the first cycle, make it shorter | |
232 | */ | |
afe24a5a JA |
233 | cycle = td->o.rwmixcycle; |
234 | if (!td->rwmix_bytes) | |
235 | cycle /= 10; | |
236 | ||
10ba535a JA |
237 | /* |
238 | * Check if it's time to seed a new data direction. | |
239 | */ | |
e47f799f | 240 | if (elapsed >= cycle || |
afe24a5a | 241 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
242 | unsigned long long max_bytes; |
243 | enum fio_ddir ddir; | |
244 | ||
245 | /* | |
246 | * Put a top limit on how many bytes we do for | |
247 | * one data direction, to avoid overflowing the | |
248 | * ranges too much | |
249 | */ | |
250 | ddir = get_rand_ddir(td); | |
251 | max_bytes = td->this_io_bytes[ddir]; | |
38d77cae JA |
252 | if (max_bytes >= (td->io_size * td->o.rwmix[ddir] / 100)) { |
253 | if (!td->rw_end_set[ddir]) { | |
254 | td->rw_end_set[ddir] = 1; | |
255 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
256 | } | |
e47f799f | 257 | ddir ^= 1; |
38d77cae | 258 | } |
e47f799f JA |
259 | |
260 | if (ddir != td->rwmix_ddir) | |
261 | set_rwmix_bytes(td); | |
262 | ||
263 | td->rwmix_ddir = ddir; | |
10ba535a JA |
264 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
265 | } | |
266 | return td->rwmix_ddir; | |
267 | } else if (td_read(td)) | |
268 | return DDIR_READ; | |
269 | else | |
270 | return DDIR_WRITE; | |
271 | } | |
272 | ||
10ba535a JA |
273 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
274 | { | |
0c6e7517 JA |
275 | assert((io_u->flags & IO_U_F_FREE) == 0); |
276 | io_u->flags |= IO_U_F_FREE; | |
277 | ||
10ba535a JA |
278 | io_u->file = NULL; |
279 | list_del(&io_u->list); | |
280 | list_add(&io_u->list, &td->io_u_freelist); | |
281 | td->cur_depth--; | |
282 | } | |
283 | ||
755200a3 JA |
284 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
285 | { | |
286 | struct io_u *__io_u = *io_u; | |
287 | ||
4d2e0f49 JA |
288 | __io_u->flags |= IO_U_F_FREE; |
289 | __io_u->flags &= ~IO_U_F_FLIGHT; | |
290 | ||
755200a3 JA |
291 | list_del(&__io_u->list); |
292 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
293 | td->cur_depth--; | |
294 | *io_u = NULL; | |
295 | } | |
296 | ||
9bf2061e | 297 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a JA |
298 | { |
299 | /* | |
300 | * If using an iolog, grab next piece if any available. | |
301 | */ | |
2dc1bbeb | 302 | if (td->o.read_iolog) |
10ba535a JA |
303 | return read_iolog_get(td, io_u); |
304 | ||
87dc1ab1 JA |
305 | /* |
306 | * see if it's time to sync | |
307 | */ | |
2dc1bbeb JA |
308 | if (td->o.fsync_blocks && |
309 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
310 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 311 | io_u->ddir = DDIR_SYNC; |
87dc1ab1 JA |
312 | return 0; |
313 | } | |
314 | ||
a00735e6 JA |
315 | io_u->ddir = get_rw_ddir(td); |
316 | ||
10ba535a | 317 | /* |
c685b5b2 JA |
318 | * No log, let the seq/rand engine retrieve the next buflen and |
319 | * position. | |
10ba535a | 320 | */ |
9bf2061e | 321 | if (get_next_offset(td, io_u)) |
bca4ed4d | 322 | return 1; |
10ba535a | 323 | |
9bf2061e | 324 | io_u->buflen = get_next_buflen(td, io_u); |
bca4ed4d JA |
325 | if (!io_u->buflen) |
326 | return 1; | |
327 | ||
328 | /* | |
329 | * mark entry before potentially trimming io_u | |
330 | */ | |
2dc1bbeb | 331 | if (!td->o.read_iolog && td_random(td) && !td->o.norandommap) |
9bf2061e | 332 | mark_random_map(td, io_u); |
bca4ed4d JA |
333 | |
334 | /* | |
335 | * If using a write iolog, store this entry. | |
336 | */ | |
2dc1bbeb | 337 | if (td->o.write_iolog_file) |
bca4ed4d JA |
338 | write_iolog_put(td, io_u); |
339 | ||
bca4ed4d | 340 | return 0; |
10ba535a JA |
341 | } |
342 | ||
b3605062 | 343 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
344 | { |
345 | int index = 0; | |
346 | ||
b3605062 JA |
347 | if (io_u->ddir == DDIR_SYNC) |
348 | return; | |
349 | ||
71619dc2 JA |
350 | switch (td->cur_depth) { |
351 | default: | |
352 | index++; | |
353 | case 32 ... 63: | |
354 | index++; | |
355 | case 16 ... 31: | |
356 | index++; | |
357 | case 8 ... 15: | |
358 | index++; | |
359 | case 4 ... 7: | |
360 | index++; | |
361 | case 2 ... 3: | |
362 | index++; | |
363 | case 1: | |
364 | break; | |
365 | } | |
366 | ||
756867bd | 367 | td->ts.io_u_map[index]++; |
b3605062 | 368 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
369 | } |
370 | ||
ec118304 JA |
371 | static void io_u_mark_latency(struct thread_data *td, unsigned long msec) |
372 | { | |
373 | int index = 0; | |
374 | ||
375 | switch (msec) { | |
376 | default: | |
377 | index++; | |
8abdce66 | 378 | case 1000 ... 1999: |
ec118304 | 379 | index++; |
8abdce66 | 380 | case 750 ... 999: |
ec118304 | 381 | index++; |
8abdce66 | 382 | case 500 ... 749: |
ec118304 | 383 | index++; |
8abdce66 | 384 | case 250 ... 499: |
ec118304 | 385 | index++; |
8abdce66 | 386 | case 100 ... 249: |
ec118304 | 387 | index++; |
8abdce66 | 388 | case 50 ... 99: |
ec118304 | 389 | index++; |
8abdce66 | 390 | case 20 ... 49: |
ec118304 | 391 | index++; |
8abdce66 | 392 | case 10 ... 19: |
ec118304 | 393 | index++; |
8abdce66 | 394 | case 4 ... 9: |
ec118304 JA |
395 | index++; |
396 | case 2 ... 3: | |
397 | index++; | |
398 | case 0 ... 1: | |
399 | break; | |
400 | } | |
401 | ||
756867bd | 402 | td->ts.io_u_lat[index]++; |
ec118304 JA |
403 | } |
404 | ||
0aabe160 JA |
405 | /* |
406 | * Get next file to service by choosing one at random | |
407 | */ | |
1c178180 JA |
408 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
409 | int badf) | |
0aabe160 | 410 | { |
0aabe160 | 411 | struct fio_file *f; |
1c178180 | 412 | int fno; |
0aabe160 JA |
413 | |
414 | do { | |
7c83c089 JA |
415 | long r = os_random_long(&td->next_file_state); |
416 | ||
2dc1bbeb | 417 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 JA |
418 | f = &td->files[fno]; |
419 | ||
420 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) | |
0aabe160 JA |
421 | return f; |
422 | } while (1); | |
423 | } | |
424 | ||
425 | /* | |
426 | * Get next file to service by doing round robin between all available ones | |
427 | */ | |
1c178180 JA |
428 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
429 | int badf) | |
3d7c391d JA |
430 | { |
431 | unsigned int old_next_file = td->next_file; | |
432 | struct fio_file *f; | |
433 | ||
434 | do { | |
435 | f = &td->files[td->next_file]; | |
436 | ||
437 | td->next_file++; | |
2dc1bbeb | 438 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
439 | td->next_file = 0; |
440 | ||
1c178180 | 441 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
442 | break; |
443 | ||
444 | f = NULL; | |
445 | } while (td->next_file != old_next_file); | |
446 | ||
447 | return f; | |
448 | } | |
449 | ||
bdb4e2e9 JA |
450 | static struct fio_file *get_next_file(struct thread_data *td) |
451 | { | |
1907dbc6 JA |
452 | struct fio_file *f; |
453 | ||
2dc1bbeb | 454 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 455 | |
bdb4e2e9 JA |
456 | if (!td->nr_open_files) |
457 | return NULL; | |
458 | ||
1907dbc6 | 459 | f = td->file_service_file; |
f11bd94d | 460 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
1907dbc6 JA |
461 | return f; |
462 | ||
2dc1bbeb | 463 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 464 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 465 | else |
1c178180 | 466 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
467 | |
468 | td->file_service_file = f; | |
469 | td->file_service_left = td->file_service_nr - 1; | |
470 | return f; | |
bdb4e2e9 JA |
471 | } |
472 | ||
1c178180 JA |
473 | static struct fio_file *find_next_new_file(struct thread_data *td) |
474 | { | |
475 | struct fio_file *f; | |
476 | ||
2dc1bbeb | 477 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
478 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
479 | else | |
480 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
481 | ||
482 | return f; | |
483 | } | |
484 | ||
10ba535a JA |
485 | struct io_u *__get_io_u(struct thread_data *td) |
486 | { | |
487 | struct io_u *io_u = NULL; | |
488 | ||
755200a3 JA |
489 | if (!list_empty(&td->io_u_requeues)) |
490 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
491 | else if (!queue_full(td)) { | |
10ba535a JA |
492 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
493 | ||
6040dabc | 494 | io_u->buflen = 0; |
10ba535a | 495 | io_u->resid = 0; |
755200a3 | 496 | io_u->file = NULL; |
d7762cf8 | 497 | io_u->end_io = NULL; |
755200a3 JA |
498 | } |
499 | ||
500 | if (io_u) { | |
0c6e7517 JA |
501 | assert(io_u->flags & IO_U_F_FREE); |
502 | io_u->flags &= ~IO_U_F_FREE; | |
503 | ||
755200a3 | 504 | io_u->error = 0; |
10ba535a JA |
505 | list_del(&io_u->list); |
506 | list_add(&io_u->list, &td->io_u_busylist); | |
507 | td->cur_depth++; | |
508 | } | |
509 | ||
510 | return io_u; | |
511 | } | |
512 | ||
513 | /* | |
514 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
515 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
516 | */ | |
3d7c391d | 517 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 518 | { |
3d7c391d | 519 | struct fio_file *f; |
10ba535a | 520 | struct io_u *io_u; |
1c178180 | 521 | int ret; |
10ba535a JA |
522 | |
523 | io_u = __get_io_u(td); | |
524 | if (!io_u) | |
525 | return NULL; | |
526 | ||
755200a3 JA |
527 | /* |
528 | * from a requeue, io_u already setup | |
529 | */ | |
530 | if (io_u->file) | |
77f392bf | 531 | goto out; |
755200a3 | 532 | |
bdb4e2e9 JA |
533 | do { |
534 | f = get_next_file(td); | |
535 | if (!f) { | |
536 | put_io_u(td, io_u); | |
537 | return NULL; | |
538 | } | |
3d7c391d | 539 | |
1c178180 | 540 | set_file: |
bdb4e2e9 | 541 | io_u->file = f; |
3d7c391d | 542 | |
bdb4e2e9 JA |
543 | if (!fill_io_u(td, io_u)) |
544 | break; | |
10ba535a | 545 | |
bdb4e2e9 JA |
546 | /* |
547 | * No more to do for this file, close it | |
548 | */ | |
549 | io_u->file = NULL; | |
b5af8293 JA |
550 | td_io_close_file(td, f); |
551 | ||
552 | /* | |
553 | * probably not the right place to do this, but see | |
554 | * if we need to open a new file | |
555 | */ | |
2dc1bbeb JA |
556 | if (td->nr_open_files < td->o.open_files && |
557 | td->o.open_files != td->o.nr_files) { | |
1c178180 | 558 | f = find_next_new_file(td); |
a978ba68 | 559 | |
1c178180 | 560 | if (!f || (ret = td_io_open_file(td, f))) { |
a978ba68 | 561 | put_io_u(td, io_u); |
e291dc12 | 562 | return NULL; |
a978ba68 | 563 | } |
1c178180 | 564 | goto set_file; |
a978ba68 | 565 | } |
bdb4e2e9 | 566 | } while (1); |
10ba535a | 567 | |
2dc1bbeb | 568 | if (td->zone_bytes >= td->o.zone_size) { |
9bf2061e | 569 | td->zone_bytes = 0; |
2dc1bbeb | 570 | f->last_pos += td->o.zone_skip; |
9bf2061e JA |
571 | } |
572 | ||
ee88470c | 573 | if (io_u->buflen + io_u->offset > f->real_file_size) { |
10ba535a JA |
574 | if (td->io_ops->flags & FIO_RAWIO) { |
575 | put_io_u(td, io_u); | |
576 | return NULL; | |
577 | } | |
578 | ||
ee88470c | 579 | io_u->buflen = f->real_file_size - io_u->offset; |
10ba535a JA |
580 | } |
581 | ||
87dc1ab1 JA |
582 | if (io_u->ddir != DDIR_SYNC) { |
583 | if (!io_u->buflen) { | |
584 | put_io_u(td, io_u); | |
585 | return NULL; | |
586 | } | |
10ba535a | 587 | |
36167d82 | 588 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 589 | |
2dc1bbeb | 590 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
591 | populate_verify_io_u(td, io_u); |
592 | } | |
10ba535a | 593 | |
165faf16 JA |
594 | /* |
595 | * Set io data pointers. | |
596 | */ | |
77f392bf | 597 | out: |
cec6b55d JA |
598 | io_u->xfer_buf = io_u->buf; |
599 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 600 | |
36167d82 JA |
601 | if (td_io_prep(td, io_u)) { |
602 | put_io_u(td, io_u); | |
603 | return NULL; | |
604 | } | |
605 | ||
02bcaa8c | 606 | fio_gettime(&io_u->start_time, NULL); |
10ba535a JA |
607 | return io_u; |
608 | } | |
609 | ||
5451792e JA |
610 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
611 | { | |
612 | const char *msg[] = { "read", "write", "sync" }; | |
613 | ||
614 | log_err("fio: io_u error"); | |
615 | ||
616 | if (io_u->file) | |
617 | log_err(" on file %s", io_u->file->file_name); | |
618 | ||
619 | log_err(": %s\n", strerror(io_u->error)); | |
620 | ||
621 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
622 | ||
623 | if (!td->error) | |
624 | td_verror(td, io_u->error, "io_u error"); | |
625 | } | |
626 | ||
97601024 JA |
627 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
628 | struct io_completion_data *icd) | |
10ba535a | 629 | { |
10ba535a JA |
630 | unsigned long msec; |
631 | ||
0c6e7517 JA |
632 | assert(io_u->flags & IO_U_F_FLIGHT); |
633 | io_u->flags &= ~IO_U_F_FLIGHT; | |
634 | ||
0ad920e7 JA |
635 | put_file(td, io_u->file); |
636 | ||
87dc1ab1 JA |
637 | if (io_u->ddir == DDIR_SYNC) { |
638 | td->last_was_sync = 1; | |
639 | return; | |
640 | } | |
641 | ||
642 | td->last_was_sync = 0; | |
643 | ||
10ba535a JA |
644 | if (!io_u->error) { |
645 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 646 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 647 | int ret; |
10ba535a JA |
648 | |
649 | td->io_blocks[idx]++; | |
650 | td->io_bytes[idx] += bytes; | |
651 | td->zone_bytes += bytes; | |
652 | td->this_io_bytes[idx] += bytes; | |
653 | ||
02bcaa8c JA |
654 | io_u->file->last_completed_pos = io_u->offset + io_u->buflen; |
655 | ||
656 | msec = mtime_since(&io_u->issue_time, &icd->time); | |
10ba535a JA |
657 | |
658 | add_clat_sample(td, idx, msec); | |
02bcaa8c | 659 | add_bw_sample(td, idx, &icd->time); |
ec118304 | 660 | io_u_mark_latency(td, msec); |
10ba535a JA |
661 | |
662 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) | |
663 | log_io_piece(td, io_u); | |
664 | ||
665 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 666 | |
d7762cf8 JA |
667 | if (io_u->end_io) { |
668 | ret = io_u->end_io(io_u); | |
3af6ef39 JA |
669 | if (ret && !icd->error) |
670 | icd->error = ret; | |
671 | } | |
5451792e | 672 | } else { |
10ba535a | 673 | icd->error = io_u->error; |
5451792e JA |
674 | io_u_log_error(td, io_u); |
675 | } | |
10ba535a JA |
676 | } |
677 | ||
d7762cf8 | 678 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 679 | { |
02bcaa8c JA |
680 | fio_gettime(&icd->time, NULL); |
681 | ||
3af6ef39 JA |
682 | icd->nr = nr; |
683 | ||
10ba535a JA |
684 | icd->error = 0; |
685 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
686 | } |
687 | ||
97601024 JA |
688 | static void ios_completed(struct thread_data *td, |
689 | struct io_completion_data *icd) | |
36167d82 JA |
690 | { |
691 | struct io_u *io_u; | |
692 | int i; | |
693 | ||
10ba535a JA |
694 | for (i = 0; i < icd->nr; i++) { |
695 | io_u = td->io_ops->event(td, i); | |
696 | ||
697 | io_completed(td, io_u, icd); | |
698 | put_io_u(td, io_u); | |
699 | } | |
700 | } | |
97601024 | 701 | |
e7e6cfb4 JA |
702 | /* |
703 | * Complete a single io_u for the sync engines. | |
704 | */ | |
d7762cf8 | 705 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
706 | { |
707 | struct io_completion_data icd; | |
708 | ||
d7762cf8 | 709 | init_icd(&icd, 1); |
97601024 JA |
710 | io_completed(td, io_u, &icd); |
711 | put_io_u(td, io_u); | |
712 | ||
713 | if (!icd.error) | |
714 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
715 | ||
37e974a5 | 716 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
717 | return -1; |
718 | } | |
719 | ||
e7e6cfb4 JA |
720 | /* |
721 | * Called to complete min_events number of io for the async engines. | |
722 | */ | |
d7762cf8 | 723 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 724 | { |
97601024 | 725 | struct io_completion_data icd; |
00de55ef | 726 | struct timespec *tvp = NULL; |
97601024 | 727 | int ret; |
4d06a338 | 728 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 729 | |
4d06a338 | 730 | if (!min_events) |
00de55ef | 731 | tvp = &ts; |
97601024 | 732 | |
00de55ef | 733 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 734 | if (ret < 0) { |
e1161c32 | 735 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
736 | return ret; |
737 | } else if (!ret) | |
738 | return ret; | |
739 | ||
d7762cf8 | 740 | init_icd(&icd, ret); |
97601024 JA |
741 | ios_completed(td, &icd); |
742 | if (!icd.error) | |
743 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
744 | ||
37e974a5 | 745 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
746 | return -1; |
747 | } | |
7e77dd02 JA |
748 | |
749 | /* | |
750 | * Call when io_u is really queued, to update the submission latency. | |
751 | */ | |
752 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
753 | { | |
754 | unsigned long slat_time; | |
755 | ||
756 | slat_time = mtime_since(&io_u->start_time, &io_u->issue_time); | |
757 | add_slat_sample(td, io_u->ddir, slat_time); | |
758 | } | |
433afcb4 | 759 | |
55bc9728 | 760 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
761 | void io_u_set_timeout(struct thread_data *td) |
762 | { | |
763 | assert(td->cur_depth); | |
764 | ||
765 | td->timer.it_interval.tv_sec = 0; | |
766 | td->timer.it_interval.tv_usec = 0; | |
767 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
768 | td->timer.it_value.tv_usec = 0; | |
769 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
770 | fio_gettime(&td->timeout_end, NULL); | |
771 | } | |
5945b9b4 JA |
772 | |
773 | static void io_u_dump(struct io_u *io_u) | |
774 | { | |
775 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
776 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
777 | ||
778 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
779 | 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); | |
780 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
781 | } | |
55bc9728 JA |
782 | #else |
783 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
784 | { | |
785 | } | |
786 | #endif | |
433afcb4 | 787 | |
55bc9728 | 788 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
789 | static void io_u_timeout_handler(int fio_unused sig) |
790 | { | |
791 | struct thread_data *td, *__td; | |
792 | pid_t pid = getpid(); | |
5945b9b4 JA |
793 | struct list_head *entry; |
794 | struct io_u *io_u; | |
433afcb4 JA |
795 | int i; |
796 | ||
797 | log_err("fio: io_u timeout\n"); | |
798 | ||
799 | /* | |
800 | * TLS would be nice... | |
801 | */ | |
802 | td = NULL; | |
803 | for_each_td(__td, i) { | |
804 | if (__td->pid == pid) { | |
805 | td = __td; | |
806 | break; | |
807 | } | |
808 | } | |
809 | ||
810 | if (!td) { | |
811 | log_err("fio: io_u timeout, can't find job\n"); | |
812 | exit(1); | |
813 | } | |
814 | ||
815 | if (!td->cur_depth) { | |
816 | log_err("fio: timeout without pending work?\n"); | |
817 | return; | |
818 | } | |
819 | ||
15506d09 | 820 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
821 | |
822 | list_for_each(entry, &td->io_u_busylist) { | |
823 | io_u = list_entry(entry, struct io_u, list); | |
824 | ||
825 | io_u_dump(io_u); | |
826 | } | |
827 | ||
828 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
829 | exit(1); |
830 | } | |
55bc9728 | 831 | #endif |
433afcb4 JA |
832 | |
833 | void io_u_init_timeout(void) | |
834 | { | |
55bc9728 | 835 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 836 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 837 | #endif |
433afcb4 | 838 | } |