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" | |
10ba535a | 9 | |
5945b9b4 JA |
10 | /* |
11 | * Change this define to play with the timeout handling | |
12 | */ | |
13 | #undef FIO_USE_TIMEOUT | |
14 | ||
97601024 JA |
15 | struct io_completion_data { |
16 | int nr; /* input */ | |
97601024 JA |
17 | |
18 | int error; /* output */ | |
19 | unsigned long bytes_done[2]; /* output */ | |
20 | struct timeval time; /* output */ | |
21 | }; | |
22 | ||
10ba535a JA |
23 | /* |
24 | * The ->file_map[] contains a map of blocks we have or have not done io | |
25 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
26 | */ | |
27 | static int random_map_free(struct thread_data *td, struct fio_file *f, | |
28 | unsigned long long block) | |
29 | { | |
30 | unsigned int idx = RAND_MAP_IDX(td, f, block); | |
31 | unsigned int bit = RAND_MAP_BIT(td, f, block); | |
32 | ||
33 | return (f->file_map[idx] & (1UL << bit)) == 0; | |
34 | } | |
35 | ||
df415585 JA |
36 | /* |
37 | * Mark a given offset as used in the map. | |
38 | */ | |
9bf2061e | 39 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 40 | { |
2dc1bbeb | 41 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 42 | struct fio_file *f = io_u->file; |
a00735e6 JA |
43 | unsigned long long block; |
44 | unsigned int blocks; | |
c685b5b2 | 45 | unsigned int nr_blocks; |
df415585 | 46 | |
a00735e6 JA |
47 | block = io_u->offset / (unsigned long long) min_bs; |
48 | blocks = 0; | |
c685b5b2 JA |
49 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
50 | ||
51 | while (blocks < nr_blocks) { | |
df415585 JA |
52 | unsigned int idx, bit; |
53 | ||
1e3d53ac JA |
54 | /* |
55 | * If we have a mixed random workload, we may | |
56 | * encounter blocks we already did IO to. | |
57 | */ | |
211097b2 | 58 | if (!td->o.ddir_nr && !random_map_free(td, f, block)) |
df415585 JA |
59 | break; |
60 | ||
61 | idx = RAND_MAP_IDX(td, f, block); | |
62 | bit = RAND_MAP_BIT(td, f, block); | |
63 | ||
0032bf9f | 64 | fio_assert(td, idx < f->num_maps); |
df415585 JA |
65 | |
66 | f->file_map[idx] |= (1UL << bit); | |
67 | block++; | |
68 | blocks++; | |
69 | } | |
70 | ||
a00735e6 JA |
71 | if ((blocks * min_bs) < io_u->buflen) |
72 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
73 | } |
74 | ||
10ba535a JA |
75 | /* |
76 | * Return the next free block in the map. | |
77 | */ | |
78 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
79 | unsigned long long *b) | |
80 | { | |
81 | int i; | |
82 | ||
c685b5b2 JA |
83 | i = f->last_free_lookup; |
84 | *b = (i * BLOCKS_PER_MAP); | |
2dc1bbeb | 85 | while ((*b) * td->o.rw_min_bs < f->real_file_size) { |
10ba535a | 86 | if (f->file_map[i] != -1UL) { |
b12ebc65 | 87 | *b += fio_ffz(f->file_map[i]); |
c685b5b2 | 88 | f->last_free_lookup = i; |
10ba535a JA |
89 | return 0; |
90 | } | |
91 | ||
92 | *b += BLOCKS_PER_MAP; | |
93 | i++; | |
94 | } | |
95 | ||
96 | return 1; | |
97 | } | |
98 | ||
ec4015da JA |
99 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
100 | int ddir, unsigned long long *b) | |
101 | { | |
7bb48f84 | 102 | unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir]; |
ec4015da JA |
103 | unsigned long long r, rb; |
104 | int loops = 5; | |
105 | ||
106 | do { | |
107 | r = os_random_long(&td->random_state); | |
108 | if (!max_blocks) | |
109 | *b = 0; | |
110 | else | |
111 | *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
43c63a78 JA |
112 | /* |
113 | * if we are not maintaining a random map, we are done. | |
114 | */ | |
ec4015da | 115 | if (td->o.norandommap) |
43c63a78 JA |
116 | return 0; |
117 | ||
118 | /* | |
119 | * calculate map offset and chec if it's free | |
120 | */ | |
ec4015da | 121 | rb = *b + (f->file_offset / td->o.min_bs[ddir]); |
43c63a78 JA |
122 | if (random_map_free(td, f, rb)) |
123 | return 0; | |
124 | ||
125 | } while (--loops); | |
ec4015da JA |
126 | |
127 | /* | |
43c63a78 JA |
128 | * we get here, if we didn't suceed in looking up a block. generate |
129 | * a random start offset into the filemap, and find the first free | |
130 | * block from there. | |
ec4015da | 131 | */ |
43c63a78 JA |
132 | loops = 10; |
133 | do { | |
134 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); | |
135 | if (!get_next_free_block(td, f, b)) | |
136 | return 0; | |
ec4015da | 137 | |
43c63a78 JA |
138 | r = os_random_long(&td->random_state); |
139 | } while (--loops); | |
140 | ||
141 | /* | |
142 | * that didn't work either, try exhaustive search from the start | |
143 | */ | |
144 | f->last_free_lookup = 0; | |
145 | return get_next_free_block(td, f, b); | |
ec4015da JA |
146 | } |
147 | ||
10ba535a JA |
148 | /* |
149 | * For random io, generate a random new block and see if it's used. Repeat | |
150 | * until we find a free one. For sequential io, just return the end of | |
151 | * the last io issued. | |
152 | */ | |
9bf2061e | 153 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 154 | { |
9bf2061e | 155 | struct fio_file *f = io_u->file; |
c685b5b2 | 156 | const int ddir = io_u->ddir; |
ec4015da | 157 | unsigned long long b; |
10ba535a | 158 | |
ec4015da JA |
159 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
160 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 161 | |
ec4015da | 162 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 163 | return 1; |
43063a1c JA |
164 | } else { |
165 | if (f->last_pos >= f->real_file_size) | |
166 | return 1; | |
167 | ||
2dc1bbeb | 168 | b = f->last_pos / td->o.min_bs[ddir]; |
43063a1c | 169 | } |
10ba535a | 170 | |
2dc1bbeb | 171 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
bca4ed4d | 172 | if (io_u->offset >= f->real_file_size) |
10ba535a JA |
173 | return 1; |
174 | ||
175 | return 0; | |
176 | } | |
177 | ||
9bf2061e | 178 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 179 | { |
bca4ed4d | 180 | const int ddir = io_u->ddir; |
10ba535a JA |
181 | unsigned int buflen; |
182 | long r; | |
183 | ||
2dc1bbeb JA |
184 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
185 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
186 | else { |
187 | r = os_random_long(&td->bsrange_state); | |
2dc1bbeb JA |
188 | buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); |
189 | if (!td->o.bs_unaligned) | |
190 | buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); | |
10ba535a JA |
191 | } |
192 | ||
6a5e6884 JA |
193 | if (io_u->offset + buflen > io_u->file->real_file_size) |
194 | buflen = td->o.min_bs[ddir]; | |
195 | ||
10ba535a JA |
196 | return buflen; |
197 | } | |
198 | ||
afe24a5a JA |
199 | static void set_rwmix_bytes(struct thread_data *td) |
200 | { | |
201 | unsigned long long rbytes; | |
202 | unsigned int diff; | |
203 | ||
204 | /* | |
205 | * we do time or byte based switch. this is needed because | |
206 | * buffered writes may issue a lot quicker than they complete, | |
207 | * whereas reads do not. | |
208 | */ | |
209 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 210 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 211 | |
e47f799f JA |
212 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
213 | } | |
214 | ||
215 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
216 | { | |
217 | unsigned int v; | |
218 | long r; | |
219 | ||
220 | r = os_random_long(&td->rwmix_state); | |
221 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
222 | if (v < td->o.rwmix[DDIR_READ]) | |
223 | return DDIR_READ; | |
224 | ||
225 | return DDIR_WRITE; | |
afe24a5a JA |
226 | } |
227 | ||
10ba535a JA |
228 | /* |
229 | * Return the data direction for the next io_u. If the job is a | |
230 | * mixed read/write workload, check the rwmix cycle and switch if | |
231 | * necessary. | |
232 | */ | |
1e97cce9 | 233 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
234 | { |
235 | if (td_rw(td)) { | |
236 | struct timeval now; | |
237 | unsigned long elapsed; | |
afe24a5a | 238 | unsigned int cycle; |
10ba535a | 239 | |
02bcaa8c | 240 | fio_gettime(&now, NULL); |
10ba535a JA |
241 | elapsed = mtime_since_now(&td->rwmix_switch); |
242 | ||
e47f799f JA |
243 | /* |
244 | * if this is the first cycle, make it shorter | |
245 | */ | |
afe24a5a JA |
246 | cycle = td->o.rwmixcycle; |
247 | if (!td->rwmix_bytes) | |
248 | cycle /= 10; | |
249 | ||
10ba535a JA |
250 | /* |
251 | * Check if it's time to seed a new data direction. | |
252 | */ | |
e47f799f | 253 | if (elapsed >= cycle || |
afe24a5a | 254 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
255 | unsigned long long max_bytes; |
256 | enum fio_ddir ddir; | |
257 | ||
258 | /* | |
259 | * Put a top limit on how many bytes we do for | |
260 | * one data direction, to avoid overflowing the | |
261 | * ranges too much | |
262 | */ | |
263 | ddir = get_rand_ddir(td); | |
264 | max_bytes = td->this_io_bytes[ddir]; | |
7bb48f84 | 265 | if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) { |
38d77cae JA |
266 | if (!td->rw_end_set[ddir]) { |
267 | td->rw_end_set[ddir] = 1; | |
268 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
269 | } | |
e47f799f | 270 | ddir ^= 1; |
38d77cae | 271 | } |
e47f799f JA |
272 | |
273 | if (ddir != td->rwmix_ddir) | |
274 | set_rwmix_bytes(td); | |
275 | ||
276 | td->rwmix_ddir = ddir; | |
10ba535a JA |
277 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
278 | } | |
279 | return td->rwmix_ddir; | |
280 | } else if (td_read(td)) | |
281 | return DDIR_READ; | |
282 | else | |
283 | return DDIR_WRITE; | |
284 | } | |
285 | ||
10ba535a JA |
286 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
287 | { | |
0c6e7517 JA |
288 | assert((io_u->flags & IO_U_F_FREE) == 0); |
289 | io_u->flags |= IO_U_F_FREE; | |
290 | ||
2dbdab7e JA |
291 | if (io_u->file) |
292 | put_file(td, io_u->file); | |
293 | ||
10ba535a JA |
294 | io_u->file = NULL; |
295 | list_del(&io_u->list); | |
296 | list_add(&io_u->list, &td->io_u_freelist); | |
297 | td->cur_depth--; | |
298 | } | |
299 | ||
755200a3 JA |
300 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
301 | { | |
302 | struct io_u *__io_u = *io_u; | |
303 | ||
4d2e0f49 JA |
304 | __io_u->flags |= IO_U_F_FREE; |
305 | __io_u->flags &= ~IO_U_F_FLIGHT; | |
306 | ||
755200a3 JA |
307 | list_del(&__io_u->list); |
308 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
309 | td->cur_depth--; | |
310 | *io_u = NULL; | |
311 | } | |
312 | ||
9bf2061e | 313 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 314 | { |
87dc1ab1 JA |
315 | /* |
316 | * see if it's time to sync | |
317 | */ | |
2dc1bbeb JA |
318 | if (td->o.fsync_blocks && |
319 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
320 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 321 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 322 | goto out; |
87dc1ab1 JA |
323 | } |
324 | ||
a00735e6 JA |
325 | io_u->ddir = get_rw_ddir(td); |
326 | ||
48f5abd3 JA |
327 | /* |
328 | * See if it's time to switch to a new zone | |
329 | */ | |
330 | if (td->zone_bytes >= td->o.zone_size) { | |
331 | td->zone_bytes = 0; | |
332 | io_u->file->last_pos += td->o.zone_skip; | |
333 | td->io_skip_bytes += td->o.zone_skip; | |
334 | } | |
335 | ||
10ba535a | 336 | /* |
c685b5b2 JA |
337 | * No log, let the seq/rand engine retrieve the next buflen and |
338 | * position. | |
10ba535a | 339 | */ |
9bf2061e | 340 | if (get_next_offset(td, io_u)) |
bca4ed4d | 341 | return 1; |
10ba535a | 342 | |
9bf2061e | 343 | io_u->buflen = get_next_buflen(td, io_u); |
bca4ed4d JA |
344 | if (!io_u->buflen) |
345 | return 1; | |
346 | ||
6a5e6884 JA |
347 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) |
348 | return 1; | |
349 | ||
bca4ed4d JA |
350 | /* |
351 | * mark entry before potentially trimming io_u | |
352 | */ | |
8fd15a9a | 353 | if (td_random(td) && !td->o.norandommap) |
9bf2061e | 354 | mark_random_map(td, io_u); |
bca4ed4d JA |
355 | |
356 | /* | |
357 | * If using a write iolog, store this entry. | |
358 | */ | |
c38e9468 | 359 | out: |
f29b25a3 | 360 | log_io_u(td, io_u); |
bca4ed4d | 361 | return 0; |
10ba535a JA |
362 | } |
363 | ||
b3605062 | 364 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
365 | { |
366 | int index = 0; | |
367 | ||
b3605062 JA |
368 | if (io_u->ddir == DDIR_SYNC) |
369 | return; | |
370 | ||
71619dc2 JA |
371 | switch (td->cur_depth) { |
372 | default: | |
a783e61a JA |
373 | index = 6; |
374 | break; | |
71619dc2 | 375 | case 32 ... 63: |
a783e61a JA |
376 | index = 5; |
377 | break; | |
71619dc2 | 378 | case 16 ... 31: |
a783e61a JA |
379 | index = 4; |
380 | break; | |
71619dc2 | 381 | case 8 ... 15: |
a783e61a JA |
382 | index = 3; |
383 | break; | |
71619dc2 | 384 | case 4 ... 7: |
a783e61a JA |
385 | index = 2; |
386 | break; | |
71619dc2 | 387 | case 2 ... 3: |
a783e61a | 388 | index = 1; |
71619dc2 JA |
389 | case 1: |
390 | break; | |
391 | } | |
392 | ||
756867bd | 393 | td->ts.io_u_map[index]++; |
b3605062 | 394 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
395 | } |
396 | ||
04a0feae JA |
397 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
398 | { | |
399 | int index = 0; | |
400 | ||
401 | assert(usec < 1000); | |
402 | ||
403 | switch (usec) { | |
404 | case 750 ... 999: | |
405 | index = 9; | |
406 | break; | |
407 | case 500 ... 749: | |
408 | index = 8; | |
409 | break; | |
410 | case 250 ... 499: | |
411 | index = 7; | |
412 | break; | |
413 | case 100 ... 249: | |
414 | index = 6; | |
415 | break; | |
416 | case 50 ... 99: | |
417 | index = 5; | |
418 | break; | |
419 | case 20 ... 49: | |
420 | index = 4; | |
421 | break; | |
422 | case 10 ... 19: | |
423 | index = 3; | |
424 | break; | |
425 | case 4 ... 9: | |
426 | index = 2; | |
427 | break; | |
428 | case 2 ... 3: | |
429 | index = 1; | |
430 | case 0 ... 1: | |
431 | break; | |
432 | } | |
433 | ||
434 | assert(index < FIO_IO_U_LAT_U_NR); | |
435 | td->ts.io_u_lat_u[index]++; | |
436 | } | |
437 | ||
438 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
439 | { |
440 | int index = 0; | |
441 | ||
442 | switch (msec) { | |
443 | default: | |
04a0feae JA |
444 | index = 11; |
445 | break; | |
8abdce66 | 446 | case 1000 ... 1999: |
04a0feae JA |
447 | index = 10; |
448 | break; | |
8abdce66 | 449 | case 750 ... 999: |
04a0feae JA |
450 | index = 9; |
451 | break; | |
8abdce66 | 452 | case 500 ... 749: |
04a0feae JA |
453 | index = 8; |
454 | break; | |
8abdce66 | 455 | case 250 ... 499: |
04a0feae JA |
456 | index = 7; |
457 | break; | |
8abdce66 | 458 | case 100 ... 249: |
04a0feae JA |
459 | index = 6; |
460 | break; | |
8abdce66 | 461 | case 50 ... 99: |
04a0feae JA |
462 | index = 5; |
463 | break; | |
8abdce66 | 464 | case 20 ... 49: |
04a0feae JA |
465 | index = 4; |
466 | break; | |
8abdce66 | 467 | case 10 ... 19: |
04a0feae JA |
468 | index = 3; |
469 | break; | |
8abdce66 | 470 | case 4 ... 9: |
04a0feae JA |
471 | index = 2; |
472 | break; | |
ec118304 | 473 | case 2 ... 3: |
04a0feae | 474 | index = 1; |
ec118304 JA |
475 | case 0 ... 1: |
476 | break; | |
477 | } | |
478 | ||
04a0feae JA |
479 | assert(index < FIO_IO_U_LAT_M_NR); |
480 | td->ts.io_u_lat_m[index]++; | |
481 | } | |
482 | ||
483 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
484 | { | |
485 | if (usec < 1000) | |
486 | io_u_mark_lat_usec(td, usec); | |
487 | else | |
488 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
489 | } |
490 | ||
0aabe160 JA |
491 | /* |
492 | * Get next file to service by choosing one at random | |
493 | */ | |
1c178180 JA |
494 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
495 | int badf) | |
0aabe160 | 496 | { |
0aabe160 | 497 | struct fio_file *f; |
1c178180 | 498 | int fno; |
0aabe160 JA |
499 | |
500 | do { | |
7c83c089 JA |
501 | long r = os_random_long(&td->next_file_state); |
502 | ||
2dc1bbeb | 503 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 | 504 | f = &td->files[fno]; |
059e63c0 JA |
505 | if (f->flags & FIO_FILE_DONE) |
506 | continue; | |
1c178180 JA |
507 | |
508 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) | |
0aabe160 JA |
509 | return f; |
510 | } while (1); | |
511 | } | |
512 | ||
513 | /* | |
514 | * Get next file to service by doing round robin between all available ones | |
515 | */ | |
1c178180 JA |
516 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
517 | int badf) | |
3d7c391d JA |
518 | { |
519 | unsigned int old_next_file = td->next_file; | |
520 | struct fio_file *f; | |
521 | ||
522 | do { | |
523 | f = &td->files[td->next_file]; | |
524 | ||
525 | td->next_file++; | |
2dc1bbeb | 526 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
527 | td->next_file = 0; |
528 | ||
d5ed68ea JA |
529 | if (f->flags & FIO_FILE_DONE) { |
530 | f = NULL; | |
059e63c0 | 531 | continue; |
d5ed68ea | 532 | } |
059e63c0 | 533 | |
1c178180 | 534 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
535 | break; |
536 | ||
537 | f = NULL; | |
538 | } while (td->next_file != old_next_file); | |
539 | ||
540 | return f; | |
541 | } | |
542 | ||
bdb4e2e9 JA |
543 | static struct fio_file *get_next_file(struct thread_data *td) |
544 | { | |
1907dbc6 JA |
545 | struct fio_file *f; |
546 | ||
2dc1bbeb | 547 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 548 | |
1020a139 | 549 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
bdb4e2e9 JA |
550 | return NULL; |
551 | ||
1907dbc6 | 552 | f = td->file_service_file; |
f11bd94d | 553 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
1907dbc6 JA |
554 | return f; |
555 | ||
2dc1bbeb | 556 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 557 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 558 | else |
1c178180 | 559 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
560 | |
561 | td->file_service_file = f; | |
562 | td->file_service_left = td->file_service_nr - 1; | |
563 | return f; | |
bdb4e2e9 JA |
564 | } |
565 | ||
1c178180 JA |
566 | static struct fio_file *find_next_new_file(struct thread_data *td) |
567 | { | |
568 | struct fio_file *f; | |
569 | ||
1020a139 JA |
570 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
571 | return NULL; | |
572 | ||
2dc1bbeb | 573 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
574 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
575 | else | |
576 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
577 | ||
578 | return f; | |
579 | } | |
580 | ||
429f6675 JA |
581 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
582 | { | |
583 | struct fio_file *f; | |
584 | ||
585 | do { | |
586 | f = get_next_file(td); | |
587 | if (!f) | |
588 | return 1; | |
589 | ||
590 | set_file: | |
591 | io_u->file = f; | |
592 | get_file(f); | |
593 | ||
594 | if (!fill_io_u(td, io_u)) | |
595 | break; | |
596 | ||
597 | /* | |
598 | * td_io_close() does a put_file() as well, so no need to | |
599 | * do that here. | |
600 | */ | |
601 | io_u->file = NULL; | |
602 | td_io_close_file(td, f); | |
603 | f->flags |= FIO_FILE_DONE; | |
604 | td->nr_done_files++; | |
605 | ||
606 | /* | |
607 | * probably not the right place to do this, but see | |
608 | * if we need to open a new file | |
609 | */ | |
610 | if (td->nr_open_files < td->o.open_files && | |
611 | td->o.open_files != td->o.nr_files) { | |
612 | f = find_next_new_file(td); | |
613 | ||
614 | if (!f || td_io_open_file(td, f)) | |
615 | return 1; | |
616 | ||
617 | goto set_file; | |
618 | } | |
619 | } while (1); | |
620 | ||
621 | return 0; | |
622 | } | |
623 | ||
624 | ||
10ba535a JA |
625 | struct io_u *__get_io_u(struct thread_data *td) |
626 | { | |
627 | struct io_u *io_u = NULL; | |
628 | ||
755200a3 JA |
629 | if (!list_empty(&td->io_u_requeues)) |
630 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
631 | else if (!queue_full(td)) { | |
10ba535a JA |
632 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
633 | ||
6040dabc | 634 | io_u->buflen = 0; |
10ba535a | 635 | io_u->resid = 0; |
755200a3 | 636 | io_u->file = NULL; |
d7762cf8 | 637 | io_u->end_io = NULL; |
755200a3 JA |
638 | } |
639 | ||
640 | if (io_u) { | |
0c6e7517 JA |
641 | assert(io_u->flags & IO_U_F_FREE); |
642 | io_u->flags &= ~IO_U_F_FREE; | |
643 | ||
755200a3 | 644 | io_u->error = 0; |
10ba535a JA |
645 | list_del(&io_u->list); |
646 | list_add(&io_u->list, &td->io_u_busylist); | |
647 | td->cur_depth++; | |
648 | } | |
649 | ||
650 | return io_u; | |
651 | } | |
652 | ||
653 | /* | |
654 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
655 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
656 | */ | |
3d7c391d | 657 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 658 | { |
3d7c391d | 659 | struct fio_file *f; |
10ba535a JA |
660 | struct io_u *io_u; |
661 | ||
662 | io_u = __get_io_u(td); | |
663 | if (!io_u) | |
664 | return NULL; | |
665 | ||
755200a3 JA |
666 | /* |
667 | * from a requeue, io_u already setup | |
668 | */ | |
669 | if (io_u->file) | |
77f392bf | 670 | goto out; |
755200a3 | 671 | |
429f6675 JA |
672 | /* |
673 | * If using an iolog, grab next piece if any available. | |
674 | */ | |
675 | if (td->o.read_iolog_file) { | |
676 | if (read_iolog_get(td, io_u)) | |
677 | goto err_put; | |
678 | } else if (set_io_u_file(td, io_u)) | |
679 | goto err_put; | |
680 | ||
681 | f = io_u->file; | |
682 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 683 | |
87dc1ab1 | 684 | if (io_u->ddir != DDIR_SYNC) { |
429f6675 JA |
685 | if (!io_u->buflen) |
686 | goto err_put; | |
10ba535a | 687 | |
36167d82 | 688 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 689 | |
2dc1bbeb | 690 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
691 | populate_verify_io_u(td, io_u); |
692 | } | |
10ba535a | 693 | |
165faf16 JA |
694 | /* |
695 | * Set io data pointers. | |
696 | */ | |
d460eb31 | 697 | io_u->endpos = io_u->offset + io_u->buflen; |
77f392bf | 698 | out: |
cec6b55d JA |
699 | io_u->xfer_buf = io_u->buf; |
700 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 701 | |
429f6675 JA |
702 | if (!td_io_prep(td, io_u)) { |
703 | fio_gettime(&io_u->start_time, NULL); | |
704 | return io_u; | |
36167d82 | 705 | } |
429f6675 JA |
706 | err_put: |
707 | put_io_u(td, io_u); | |
708 | return NULL; | |
10ba535a JA |
709 | } |
710 | ||
5451792e JA |
711 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
712 | { | |
713 | const char *msg[] = { "read", "write", "sync" }; | |
714 | ||
715 | log_err("fio: io_u error"); | |
716 | ||
717 | if (io_u->file) | |
718 | log_err(" on file %s", io_u->file->file_name); | |
719 | ||
720 | log_err(": %s\n", strerror(io_u->error)); | |
721 | ||
722 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
723 | ||
724 | if (!td->error) | |
725 | td_verror(td, io_u->error, "io_u error"); | |
726 | } | |
727 | ||
97601024 JA |
728 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
729 | struct io_completion_data *icd) | |
10ba535a | 730 | { |
d85f5118 | 731 | unsigned long usec; |
10ba535a | 732 | |
0c6e7517 JA |
733 | assert(io_u->flags & IO_U_F_FLIGHT); |
734 | io_u->flags &= ~IO_U_F_FLIGHT; | |
735 | ||
87dc1ab1 JA |
736 | if (io_u->ddir == DDIR_SYNC) { |
737 | td->last_was_sync = 1; | |
738 | return; | |
739 | } | |
740 | ||
741 | td->last_was_sync = 0; | |
742 | ||
10ba535a JA |
743 | if (!io_u->error) { |
744 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 745 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 746 | int ret; |
10ba535a JA |
747 | |
748 | td->io_blocks[idx]++; | |
749 | td->io_bytes[idx] += bytes; | |
750 | td->zone_bytes += bytes; | |
751 | td->this_io_bytes[idx] += bytes; | |
752 | ||
d460eb31 | 753 | io_u->file->last_completed_pos = io_u->endpos; |
02bcaa8c | 754 | |
d85f5118 | 755 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 756 | |
d85f5118 | 757 | add_clat_sample(td, idx, usec); |
02bcaa8c | 758 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 759 | io_u_mark_latency(td, usec); |
10ba535a | 760 | |
660a1cb5 | 761 | if (td_write(td) && idx == DDIR_WRITE && |
41128405 | 762 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
763 | log_io_piece(td, io_u); |
764 | ||
765 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 766 | |
d7762cf8 | 767 | if (io_u->end_io) { |
36690c9b | 768 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
769 | if (ret && !icd->error) |
770 | icd->error = ret; | |
771 | } | |
5451792e | 772 | } else { |
10ba535a | 773 | icd->error = io_u->error; |
5451792e JA |
774 | io_u_log_error(td, io_u); |
775 | } | |
10ba535a JA |
776 | } |
777 | ||
d7762cf8 | 778 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 779 | { |
02bcaa8c JA |
780 | fio_gettime(&icd->time, NULL); |
781 | ||
3af6ef39 JA |
782 | icd->nr = nr; |
783 | ||
10ba535a JA |
784 | icd->error = 0; |
785 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
786 | } |
787 | ||
97601024 JA |
788 | static void ios_completed(struct thread_data *td, |
789 | struct io_completion_data *icd) | |
36167d82 JA |
790 | { |
791 | struct io_u *io_u; | |
792 | int i; | |
793 | ||
10ba535a JA |
794 | for (i = 0; i < icd->nr; i++) { |
795 | io_u = td->io_ops->event(td, i); | |
796 | ||
797 | io_completed(td, io_u, icd); | |
798 | put_io_u(td, io_u); | |
799 | } | |
800 | } | |
97601024 | 801 | |
e7e6cfb4 JA |
802 | /* |
803 | * Complete a single io_u for the sync engines. | |
804 | */ | |
d7762cf8 | 805 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
806 | { |
807 | struct io_completion_data icd; | |
808 | ||
d7762cf8 | 809 | init_icd(&icd, 1); |
97601024 JA |
810 | io_completed(td, io_u, &icd); |
811 | put_io_u(td, io_u); | |
812 | ||
813 | if (!icd.error) | |
814 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
815 | ||
37e974a5 | 816 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
817 | return -1; |
818 | } | |
819 | ||
e7e6cfb4 JA |
820 | /* |
821 | * Called to complete min_events number of io for the async engines. | |
822 | */ | |
d7762cf8 | 823 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 824 | { |
97601024 | 825 | struct io_completion_data icd; |
00de55ef | 826 | struct timespec *tvp = NULL; |
97601024 | 827 | int ret; |
4d06a338 | 828 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 829 | |
4d06a338 | 830 | if (!min_events) |
00de55ef | 831 | tvp = &ts; |
97601024 | 832 | |
00de55ef | 833 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 834 | if (ret < 0) { |
e1161c32 | 835 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
836 | return ret; |
837 | } else if (!ret) | |
838 | return ret; | |
839 | ||
d7762cf8 | 840 | init_icd(&icd, ret); |
97601024 JA |
841 | ios_completed(td, &icd); |
842 | if (!icd.error) | |
843 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
844 | ||
37e974a5 | 845 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
846 | return -1; |
847 | } | |
7e77dd02 JA |
848 | |
849 | /* | |
850 | * Call when io_u is really queued, to update the submission latency. | |
851 | */ | |
852 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
853 | { | |
854 | unsigned long slat_time; | |
855 | ||
d85f5118 | 856 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
857 | add_slat_sample(td, io_u->ddir, slat_time); |
858 | } | |
433afcb4 | 859 | |
55bc9728 | 860 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
861 | void io_u_set_timeout(struct thread_data *td) |
862 | { | |
863 | assert(td->cur_depth); | |
864 | ||
865 | td->timer.it_interval.tv_sec = 0; | |
866 | td->timer.it_interval.tv_usec = 0; | |
867 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
868 | td->timer.it_value.tv_usec = 0; | |
869 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
870 | fio_gettime(&td->timeout_end, NULL); | |
871 | } | |
5945b9b4 JA |
872 | |
873 | static void io_u_dump(struct io_u *io_u) | |
874 | { | |
875 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
876 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
877 | ||
878 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
879 | 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); | |
880 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
881 | } | |
55bc9728 JA |
882 | #else |
883 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
884 | { | |
885 | } | |
886 | #endif | |
433afcb4 | 887 | |
55bc9728 | 888 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
889 | static void io_u_timeout_handler(int fio_unused sig) |
890 | { | |
891 | struct thread_data *td, *__td; | |
892 | pid_t pid = getpid(); | |
5945b9b4 JA |
893 | struct list_head *entry; |
894 | struct io_u *io_u; | |
433afcb4 JA |
895 | int i; |
896 | ||
897 | log_err("fio: io_u timeout\n"); | |
898 | ||
899 | /* | |
900 | * TLS would be nice... | |
901 | */ | |
902 | td = NULL; | |
903 | for_each_td(__td, i) { | |
904 | if (__td->pid == pid) { | |
905 | td = __td; | |
906 | break; | |
907 | } | |
908 | } | |
909 | ||
910 | if (!td) { | |
911 | log_err("fio: io_u timeout, can't find job\n"); | |
912 | exit(1); | |
913 | } | |
914 | ||
915 | if (!td->cur_depth) { | |
916 | log_err("fio: timeout without pending work?\n"); | |
917 | return; | |
918 | } | |
919 | ||
15506d09 | 920 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
921 | |
922 | list_for_each(entry, &td->io_u_busylist) { | |
923 | io_u = list_entry(entry, struct io_u, list); | |
924 | ||
925 | io_u_dump(io_u); | |
926 | } | |
927 | ||
928 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
929 | exit(1); |
930 | } | |
55bc9728 | 931 | #endif |
433afcb4 JA |
932 | |
933 | void io_u_init_timeout(void) | |
934 | { | |
55bc9728 | 935 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 936 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 937 | #endif |
433afcb4 | 938 | } |