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 | */ | |
b6a4c7d1 | 58 | if (!td->o.ddir_nr == 1 && !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 | ||
2ba1c290 JA |
75 | static inline unsigned long long last_block(struct thread_data *td, |
76 | struct fio_file *f) | |
77 | { | |
78 | unsigned long long max_blocks; | |
79 | ||
80 | max_blocks = f->io_size / td->o.rw_min_bs; | |
81 | if (!max_blocks) | |
82 | return 0; | |
83 | ||
84 | return max_blocks - 1; | |
85 | } | |
86 | ||
10ba535a JA |
87 | /* |
88 | * Return the next free block in the map. | |
89 | */ | |
90 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
91 | unsigned long long *b) | |
92 | { | |
93 | int i; | |
94 | ||
c685b5b2 JA |
95 | i = f->last_free_lookup; |
96 | *b = (i * BLOCKS_PER_MAP); | |
2dc1bbeb | 97 | while ((*b) * td->o.rw_min_bs < f->real_file_size) { |
10ba535a | 98 | if (f->file_map[i] != -1UL) { |
b12ebc65 | 99 | *b += fio_ffz(f->file_map[i]); |
2ba1c290 JA |
100 | if (*b > last_block(td, f)) |
101 | break; | |
c685b5b2 | 102 | f->last_free_lookup = i; |
10ba535a JA |
103 | return 0; |
104 | } | |
105 | ||
106 | *b += BLOCKS_PER_MAP; | |
107 | i++; | |
108 | } | |
109 | ||
2ba1c290 | 110 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
111 | return 1; |
112 | } | |
113 | ||
ec4015da | 114 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
2ba1c290 | 115 | unsigned long long *b) |
ec4015da | 116 | { |
ec4015da JA |
117 | unsigned long long r, rb; |
118 | int loops = 5; | |
119 | ||
120 | do { | |
121 | r = os_random_long(&td->random_state); | |
2ba1c290 JA |
122 | *b = last_block(td, f); |
123 | ||
43c63a78 JA |
124 | /* |
125 | * if we are not maintaining a random map, we are done. | |
126 | */ | |
ec4015da | 127 | if (td->o.norandommap) |
43c63a78 JA |
128 | return 0; |
129 | ||
130 | /* | |
131 | * calculate map offset and chec if it's free | |
132 | */ | |
273f8c91 | 133 | rb = *b; |
43c63a78 JA |
134 | if (random_map_free(td, f, rb)) |
135 | return 0; | |
136 | ||
137 | } while (--loops); | |
ec4015da JA |
138 | |
139 | /* | |
43c63a78 JA |
140 | * we get here, if we didn't suceed in looking up a block. generate |
141 | * a random start offset into the filemap, and find the first free | |
142 | * block from there. | |
ec4015da | 143 | */ |
43c63a78 JA |
144 | loops = 10; |
145 | do { | |
146 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); | |
147 | if (!get_next_free_block(td, f, b)) | |
148 | return 0; | |
ec4015da | 149 | |
43c63a78 JA |
150 | r = os_random_long(&td->random_state); |
151 | } while (--loops); | |
152 | ||
153 | /* | |
154 | * that didn't work either, try exhaustive search from the start | |
155 | */ | |
156 | f->last_free_lookup = 0; | |
157 | return get_next_free_block(td, f, b); | |
ec4015da JA |
158 | } |
159 | ||
10ba535a JA |
160 | /* |
161 | * For random io, generate a random new block and see if it's used. Repeat | |
162 | * until we find a free one. For sequential io, just return the end of | |
163 | * the last io issued. | |
164 | */ | |
9bf2061e | 165 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 166 | { |
9bf2061e | 167 | struct fio_file *f = io_u->file; |
ec4015da | 168 | unsigned long long b; |
10ba535a | 169 | |
ec4015da JA |
170 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
171 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 172 | |
2ba1c290 | 173 | if (get_next_rand_offset(td, f, &b)) |
bca4ed4d | 174 | return 1; |
43063a1c | 175 | } else { |
0c3d768a | 176 | if (f->last_pos >= f->real_file_size) { |
2ba1c290 | 177 | if (!td_random(td) || get_next_rand_offset(td, f, &b)) |
0c3d768a | 178 | return 1; |
bcdedd0a | 179 | } else |
2ba1c290 | 180 | b = (f->last_pos - f->file_offset) / td->o.rw_min_bs; |
43063a1c | 181 | } |
10ba535a | 182 | |
2ba1c290 JA |
183 | io_u->offset = (b * td->o.rw_min_bs) + f->file_offset; |
184 | if (io_u->offset >= f->real_file_size) { | |
185 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
186 | io_u->offset, f->real_file_size); | |
10ba535a | 187 | return 1; |
2ba1c290 | 188 | } |
10ba535a JA |
189 | |
190 | return 0; | |
191 | } | |
192 | ||
9bf2061e | 193 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 194 | { |
bca4ed4d | 195 | const int ddir = io_u->ddir; |
10ba535a JA |
196 | unsigned int buflen; |
197 | long r; | |
198 | ||
2dc1bbeb JA |
199 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
200 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
201 | else { |
202 | r = os_random_long(&td->bsrange_state); | |
564ca972 JA |
203 | if (!td->o.bssplit_nr) |
204 | buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); | |
205 | else { | |
206 | long perc = 0; | |
207 | unsigned int i; | |
208 | ||
209 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
210 | struct bssplit *bsp = &td->o.bssplit[i]; | |
211 | ||
212 | buflen = bsp->bs; | |
213 | perc += bsp->perc; | |
214 | if (r <= ((LONG_MAX / 100L) * perc)) | |
215 | break; | |
216 | } | |
217 | } | |
2dc1bbeb JA |
218 | if (!td->o.bs_unaligned) |
219 | buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); | |
10ba535a JA |
220 | } |
221 | ||
6a5e6884 JA |
222 | if (io_u->offset + buflen > io_u->file->real_file_size) |
223 | buflen = td->o.min_bs[ddir]; | |
224 | ||
10ba535a JA |
225 | return buflen; |
226 | } | |
227 | ||
afe24a5a JA |
228 | static void set_rwmix_bytes(struct thread_data *td) |
229 | { | |
230 | unsigned long long rbytes; | |
231 | unsigned int diff; | |
232 | ||
233 | /* | |
234 | * we do time or byte based switch. this is needed because | |
235 | * buffered writes may issue a lot quicker than they complete, | |
236 | * whereas reads do not. | |
237 | */ | |
238 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 239 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 240 | |
e47f799f JA |
241 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
242 | } | |
243 | ||
244 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
245 | { | |
246 | unsigned int v; | |
247 | long r; | |
248 | ||
249 | r = os_random_long(&td->rwmix_state); | |
250 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
251 | if (v < td->o.rwmix[DDIR_READ]) | |
252 | return DDIR_READ; | |
253 | ||
254 | return DDIR_WRITE; | |
afe24a5a JA |
255 | } |
256 | ||
10ba535a JA |
257 | /* |
258 | * Return the data direction for the next io_u. If the job is a | |
259 | * mixed read/write workload, check the rwmix cycle and switch if | |
260 | * necessary. | |
261 | */ | |
1e97cce9 | 262 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
263 | { |
264 | if (td_rw(td)) { | |
265 | struct timeval now; | |
266 | unsigned long elapsed; | |
afe24a5a | 267 | unsigned int cycle; |
10ba535a | 268 | |
02bcaa8c | 269 | fio_gettime(&now, NULL); |
10ba535a JA |
270 | elapsed = mtime_since_now(&td->rwmix_switch); |
271 | ||
e47f799f JA |
272 | /* |
273 | * if this is the first cycle, make it shorter | |
274 | */ | |
afe24a5a JA |
275 | cycle = td->o.rwmixcycle; |
276 | if (!td->rwmix_bytes) | |
277 | cycle /= 10; | |
278 | ||
10ba535a JA |
279 | /* |
280 | * Check if it's time to seed a new data direction. | |
281 | */ | |
e47f799f | 282 | if (elapsed >= cycle || |
afe24a5a | 283 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
284 | unsigned long long max_bytes; |
285 | enum fio_ddir ddir; | |
286 | ||
287 | /* | |
288 | * Put a top limit on how many bytes we do for | |
289 | * one data direction, to avoid overflowing the | |
290 | * ranges too much | |
291 | */ | |
292 | ddir = get_rand_ddir(td); | |
293 | max_bytes = td->this_io_bytes[ddir]; | |
7bb48f84 | 294 | if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) { |
38d77cae JA |
295 | if (!td->rw_end_set[ddir]) { |
296 | td->rw_end_set[ddir] = 1; | |
297 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
298 | } | |
e47f799f | 299 | ddir ^= 1; |
38d77cae | 300 | } |
e47f799f JA |
301 | |
302 | if (ddir != td->rwmix_ddir) | |
303 | set_rwmix_bytes(td); | |
304 | ||
305 | td->rwmix_ddir = ddir; | |
10ba535a JA |
306 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
307 | } | |
308 | return td->rwmix_ddir; | |
309 | } else if (td_read(td)) | |
310 | return DDIR_READ; | |
311 | else | |
312 | return DDIR_WRITE; | |
313 | } | |
314 | ||
10ba535a JA |
315 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
316 | { | |
0c6e7517 JA |
317 | assert((io_u->flags & IO_U_F_FREE) == 0); |
318 | io_u->flags |= IO_U_F_FREE; | |
319 | ||
2dbdab7e JA |
320 | if (io_u->file) |
321 | put_file(td, io_u->file); | |
322 | ||
10ba535a JA |
323 | io_u->file = NULL; |
324 | list_del(&io_u->list); | |
325 | list_add(&io_u->list, &td->io_u_freelist); | |
326 | td->cur_depth--; | |
327 | } | |
328 | ||
755200a3 JA |
329 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
330 | { | |
331 | struct io_u *__io_u = *io_u; | |
332 | ||
4d2e0f49 | 333 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
334 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
335 | td->io_issues[__io_u->ddir]--; | |
336 | ||
4d2e0f49 JA |
337 | __io_u->flags &= ~IO_U_F_FLIGHT; |
338 | ||
755200a3 JA |
339 | list_del(&__io_u->list); |
340 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
341 | td->cur_depth--; | |
342 | *io_u = NULL; | |
343 | } | |
344 | ||
9bf2061e | 345 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 346 | { |
b4c5e1ac JA |
347 | if (td->io_ops->flags & FIO_NOIO) |
348 | goto out; | |
349 | ||
87dc1ab1 JA |
350 | /* |
351 | * see if it's time to sync | |
352 | */ | |
2dc1bbeb JA |
353 | if (td->o.fsync_blocks && |
354 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
355 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 356 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 357 | goto out; |
87dc1ab1 JA |
358 | } |
359 | ||
a00735e6 JA |
360 | io_u->ddir = get_rw_ddir(td); |
361 | ||
48f5abd3 JA |
362 | /* |
363 | * See if it's time to switch to a new zone | |
364 | */ | |
365 | if (td->zone_bytes >= td->o.zone_size) { | |
366 | td->zone_bytes = 0; | |
367 | io_u->file->last_pos += td->o.zone_skip; | |
368 | td->io_skip_bytes += td->o.zone_skip; | |
369 | } | |
370 | ||
10ba535a | 371 | /* |
c685b5b2 JA |
372 | * No log, let the seq/rand engine retrieve the next buflen and |
373 | * position. | |
10ba535a | 374 | */ |
2ba1c290 JA |
375 | if (get_next_offset(td, io_u)) { |
376 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 377 | return 1; |
2ba1c290 | 378 | } |
10ba535a | 379 | |
9bf2061e | 380 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
381 | if (!io_u->buflen) { |
382 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 383 | return 1; |
2ba1c290 | 384 | } |
bca4ed4d | 385 | |
2ba1c290 JA |
386 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
387 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
6a5e6884 | 388 | return 1; |
2ba1c290 | 389 | } |
6a5e6884 | 390 | |
bca4ed4d JA |
391 | /* |
392 | * mark entry before potentially trimming io_u | |
393 | */ | |
8fd15a9a | 394 | if (td_random(td) && !td->o.norandommap) |
9bf2061e | 395 | mark_random_map(td, io_u); |
bca4ed4d JA |
396 | |
397 | /* | |
398 | * If using a write iolog, store this entry. | |
399 | */ | |
c38e9468 | 400 | out: |
2ba1c290 | 401 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 402 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 403 | log_io_u(td, io_u); |
bca4ed4d | 404 | return 0; |
10ba535a JA |
405 | } |
406 | ||
b3605062 | 407 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
408 | { |
409 | int index = 0; | |
410 | ||
b3605062 JA |
411 | if (io_u->ddir == DDIR_SYNC) |
412 | return; | |
413 | ||
71619dc2 JA |
414 | switch (td->cur_depth) { |
415 | default: | |
a783e61a JA |
416 | index = 6; |
417 | break; | |
71619dc2 | 418 | case 32 ... 63: |
a783e61a JA |
419 | index = 5; |
420 | break; | |
71619dc2 | 421 | case 16 ... 31: |
a783e61a JA |
422 | index = 4; |
423 | break; | |
71619dc2 | 424 | case 8 ... 15: |
a783e61a JA |
425 | index = 3; |
426 | break; | |
71619dc2 | 427 | case 4 ... 7: |
a783e61a JA |
428 | index = 2; |
429 | break; | |
71619dc2 | 430 | case 2 ... 3: |
a783e61a | 431 | index = 1; |
71619dc2 JA |
432 | case 1: |
433 | break; | |
434 | } | |
435 | ||
756867bd | 436 | td->ts.io_u_map[index]++; |
b3605062 | 437 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
438 | } |
439 | ||
04a0feae JA |
440 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
441 | { | |
442 | int index = 0; | |
443 | ||
444 | assert(usec < 1000); | |
445 | ||
446 | switch (usec) { | |
447 | case 750 ... 999: | |
448 | index = 9; | |
449 | break; | |
450 | case 500 ... 749: | |
451 | index = 8; | |
452 | break; | |
453 | case 250 ... 499: | |
454 | index = 7; | |
455 | break; | |
456 | case 100 ... 249: | |
457 | index = 6; | |
458 | break; | |
459 | case 50 ... 99: | |
460 | index = 5; | |
461 | break; | |
462 | case 20 ... 49: | |
463 | index = 4; | |
464 | break; | |
465 | case 10 ... 19: | |
466 | index = 3; | |
467 | break; | |
468 | case 4 ... 9: | |
469 | index = 2; | |
470 | break; | |
471 | case 2 ... 3: | |
472 | index = 1; | |
473 | case 0 ... 1: | |
474 | break; | |
475 | } | |
476 | ||
477 | assert(index < FIO_IO_U_LAT_U_NR); | |
478 | td->ts.io_u_lat_u[index]++; | |
479 | } | |
480 | ||
481 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
482 | { |
483 | int index = 0; | |
484 | ||
485 | switch (msec) { | |
486 | default: | |
04a0feae JA |
487 | index = 11; |
488 | break; | |
8abdce66 | 489 | case 1000 ... 1999: |
04a0feae JA |
490 | index = 10; |
491 | break; | |
8abdce66 | 492 | case 750 ... 999: |
04a0feae JA |
493 | index = 9; |
494 | break; | |
8abdce66 | 495 | case 500 ... 749: |
04a0feae JA |
496 | index = 8; |
497 | break; | |
8abdce66 | 498 | case 250 ... 499: |
04a0feae JA |
499 | index = 7; |
500 | break; | |
8abdce66 | 501 | case 100 ... 249: |
04a0feae JA |
502 | index = 6; |
503 | break; | |
8abdce66 | 504 | case 50 ... 99: |
04a0feae JA |
505 | index = 5; |
506 | break; | |
8abdce66 | 507 | case 20 ... 49: |
04a0feae JA |
508 | index = 4; |
509 | break; | |
8abdce66 | 510 | case 10 ... 19: |
04a0feae JA |
511 | index = 3; |
512 | break; | |
8abdce66 | 513 | case 4 ... 9: |
04a0feae JA |
514 | index = 2; |
515 | break; | |
ec118304 | 516 | case 2 ... 3: |
04a0feae | 517 | index = 1; |
ec118304 JA |
518 | case 0 ... 1: |
519 | break; | |
520 | } | |
521 | ||
04a0feae JA |
522 | assert(index < FIO_IO_U_LAT_M_NR); |
523 | td->ts.io_u_lat_m[index]++; | |
524 | } | |
525 | ||
526 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
527 | { | |
528 | if (usec < 1000) | |
529 | io_u_mark_lat_usec(td, usec); | |
530 | else | |
531 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
532 | } |
533 | ||
0aabe160 JA |
534 | /* |
535 | * Get next file to service by choosing one at random | |
536 | */ | |
1c178180 JA |
537 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
538 | int badf) | |
0aabe160 | 539 | { |
0aabe160 | 540 | struct fio_file *f; |
1c178180 | 541 | int fno; |
0aabe160 JA |
542 | |
543 | do { | |
7c83c089 JA |
544 | long r = os_random_long(&td->next_file_state); |
545 | ||
2dc1bbeb | 546 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 | 547 | f = &td->files[fno]; |
059e63c0 JA |
548 | if (f->flags & FIO_FILE_DONE) |
549 | continue; | |
1c178180 | 550 | |
2ba1c290 JA |
551 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
552 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 553 | return f; |
2ba1c290 | 554 | } |
0aabe160 JA |
555 | } while (1); |
556 | } | |
557 | ||
558 | /* | |
559 | * Get next file to service by doing round robin between all available ones | |
560 | */ | |
1c178180 JA |
561 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
562 | int badf) | |
3d7c391d JA |
563 | { |
564 | unsigned int old_next_file = td->next_file; | |
565 | struct fio_file *f; | |
566 | ||
567 | do { | |
568 | f = &td->files[td->next_file]; | |
569 | ||
570 | td->next_file++; | |
2dc1bbeb | 571 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
572 | td->next_file = 0; |
573 | ||
d5ed68ea JA |
574 | if (f->flags & FIO_FILE_DONE) { |
575 | f = NULL; | |
059e63c0 | 576 | continue; |
d5ed68ea | 577 | } |
059e63c0 | 578 | |
1c178180 | 579 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
580 | break; |
581 | ||
582 | f = NULL; | |
583 | } while (td->next_file != old_next_file); | |
584 | ||
2ba1c290 | 585 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
586 | return f; |
587 | } | |
588 | ||
bdb4e2e9 JA |
589 | static struct fio_file *get_next_file(struct thread_data *td) |
590 | { | |
1907dbc6 JA |
591 | struct fio_file *f; |
592 | ||
2dc1bbeb | 593 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 594 | |
2ba1c290 JA |
595 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
596 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", td->nr_open_files, td->nr_done_files, td->o.nr_files); | |
bdb4e2e9 | 597 | return NULL; |
2ba1c290 | 598 | } |
bdb4e2e9 | 599 | |
1907dbc6 | 600 | f = td->file_service_file; |
f11bd94d | 601 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 602 | goto out; |
1907dbc6 | 603 | |
2dc1bbeb | 604 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 605 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 606 | else |
1c178180 | 607 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
608 | |
609 | td->file_service_file = f; | |
610 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
611 | out: |
612 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 613 | return f; |
bdb4e2e9 JA |
614 | } |
615 | ||
1c178180 JA |
616 | static struct fio_file *find_next_new_file(struct thread_data *td) |
617 | { | |
618 | struct fio_file *f; | |
619 | ||
1020a139 JA |
620 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
621 | return NULL; | |
622 | ||
2dc1bbeb | 623 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
624 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
625 | else | |
626 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
627 | ||
628 | return f; | |
629 | } | |
630 | ||
429f6675 JA |
631 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
632 | { | |
633 | struct fio_file *f; | |
634 | ||
635 | do { | |
636 | f = get_next_file(td); | |
637 | if (!f) | |
638 | return 1; | |
639 | ||
640 | set_file: | |
641 | io_u->file = f; | |
642 | get_file(f); | |
643 | ||
644 | if (!fill_io_u(td, io_u)) | |
645 | break; | |
646 | ||
647 | /* | |
648 | * td_io_close() does a put_file() as well, so no need to | |
649 | * do that here. | |
650 | */ | |
651 | io_u->file = NULL; | |
652 | td_io_close_file(td, f); | |
653 | f->flags |= FIO_FILE_DONE; | |
654 | td->nr_done_files++; | |
655 | ||
656 | /* | |
657 | * probably not the right place to do this, but see | |
658 | * if we need to open a new file | |
659 | */ | |
660 | if (td->nr_open_files < td->o.open_files && | |
661 | td->o.open_files != td->o.nr_files) { | |
662 | f = find_next_new_file(td); | |
663 | ||
664 | if (!f || td_io_open_file(td, f)) | |
665 | return 1; | |
666 | ||
667 | goto set_file; | |
668 | } | |
669 | } while (1); | |
670 | ||
671 | return 0; | |
672 | } | |
673 | ||
674 | ||
10ba535a JA |
675 | struct io_u *__get_io_u(struct thread_data *td) |
676 | { | |
677 | struct io_u *io_u = NULL; | |
678 | ||
755200a3 JA |
679 | if (!list_empty(&td->io_u_requeues)) |
680 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
681 | else if (!queue_full(td)) { | |
10ba535a JA |
682 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
683 | ||
6040dabc | 684 | io_u->buflen = 0; |
10ba535a | 685 | io_u->resid = 0; |
755200a3 | 686 | io_u->file = NULL; |
d7762cf8 | 687 | io_u->end_io = NULL; |
755200a3 JA |
688 | } |
689 | ||
690 | if (io_u) { | |
0c6e7517 JA |
691 | assert(io_u->flags & IO_U_F_FREE); |
692 | io_u->flags &= ~IO_U_F_FREE; | |
693 | ||
755200a3 | 694 | io_u->error = 0; |
10ba535a JA |
695 | list_del(&io_u->list); |
696 | list_add(&io_u->list, &td->io_u_busylist); | |
697 | td->cur_depth++; | |
698 | } | |
699 | ||
700 | return io_u; | |
701 | } | |
702 | ||
703 | /* | |
704 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
705 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
706 | */ | |
3d7c391d | 707 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 708 | { |
3d7c391d | 709 | struct fio_file *f; |
10ba535a JA |
710 | struct io_u *io_u; |
711 | ||
712 | io_u = __get_io_u(td); | |
2ba1c290 JA |
713 | if (!io_u) { |
714 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 715 | return NULL; |
2ba1c290 | 716 | } |
10ba535a | 717 | |
755200a3 JA |
718 | /* |
719 | * from a requeue, io_u already setup | |
720 | */ | |
721 | if (io_u->file) | |
77f392bf | 722 | goto out; |
755200a3 | 723 | |
429f6675 JA |
724 | /* |
725 | * If using an iolog, grab next piece if any available. | |
726 | */ | |
727 | if (td->o.read_iolog_file) { | |
728 | if (read_iolog_get(td, io_u)) | |
729 | goto err_put; | |
2ba1c290 JA |
730 | } else if (set_io_u_file(td, io_u)) { |
731 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 732 | goto err_put; |
2ba1c290 | 733 | } |
429f6675 JA |
734 | |
735 | f = io_u->file; | |
736 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 737 | |
87dc1ab1 | 738 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 739 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 740 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 741 | goto err_put; |
2ba1c290 | 742 | } |
10ba535a | 743 | |
36167d82 | 744 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 745 | |
2dc1bbeb | 746 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
747 | populate_verify_io_u(td, io_u); |
748 | } | |
10ba535a | 749 | |
165faf16 JA |
750 | /* |
751 | * Set io data pointers. | |
752 | */ | |
d460eb31 | 753 | io_u->endpos = io_u->offset + io_u->buflen; |
77f392bf | 754 | out: |
cec6b55d JA |
755 | io_u->xfer_buf = io_u->buf; |
756 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 757 | |
429f6675 JA |
758 | if (!td_io_prep(td, io_u)) { |
759 | fio_gettime(&io_u->start_time, NULL); | |
760 | return io_u; | |
36167d82 | 761 | } |
429f6675 | 762 | err_put: |
2ba1c290 | 763 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
764 | put_io_u(td, io_u); |
765 | return NULL; | |
10ba535a JA |
766 | } |
767 | ||
5451792e JA |
768 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
769 | { | |
770 | const char *msg[] = { "read", "write", "sync" }; | |
771 | ||
772 | log_err("fio: io_u error"); | |
773 | ||
774 | if (io_u->file) | |
775 | log_err(" on file %s", io_u->file->file_name); | |
776 | ||
777 | log_err(": %s\n", strerror(io_u->error)); | |
778 | ||
779 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
780 | ||
781 | if (!td->error) | |
782 | td_verror(td, io_u->error, "io_u error"); | |
783 | } | |
784 | ||
97601024 JA |
785 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
786 | struct io_completion_data *icd) | |
10ba535a | 787 | { |
d85f5118 | 788 | unsigned long usec; |
10ba535a | 789 | |
2ba1c290 JA |
790 | dprint_io_u(io_u, "io complete"); |
791 | ||
0c6e7517 JA |
792 | assert(io_u->flags & IO_U_F_FLIGHT); |
793 | io_u->flags &= ~IO_U_F_FLIGHT; | |
794 | ||
87dc1ab1 JA |
795 | if (io_u->ddir == DDIR_SYNC) { |
796 | td->last_was_sync = 1; | |
797 | return; | |
798 | } | |
799 | ||
800 | td->last_was_sync = 0; | |
801 | ||
10ba535a JA |
802 | if (!io_u->error) { |
803 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 804 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 805 | int ret; |
10ba535a JA |
806 | |
807 | td->io_blocks[idx]++; | |
808 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
809 | td->this_io_bytes[idx] += bytes; |
810 | ||
d460eb31 | 811 | io_u->file->last_completed_pos = io_u->endpos; |
02bcaa8c | 812 | |
d85f5118 | 813 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 814 | |
d85f5118 | 815 | add_clat_sample(td, idx, usec); |
02bcaa8c | 816 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 817 | io_u_mark_latency(td, usec); |
10ba535a | 818 | |
660a1cb5 | 819 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 820 | td->o.do_verify && |
41128405 | 821 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
822 | log_io_piece(td, io_u); |
823 | ||
824 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 825 | |
d7762cf8 | 826 | if (io_u->end_io) { |
36690c9b | 827 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
828 | if (ret && !icd->error) |
829 | icd->error = ret; | |
830 | } | |
5451792e | 831 | } else { |
10ba535a | 832 | icd->error = io_u->error; |
5451792e JA |
833 | io_u_log_error(td, io_u); |
834 | } | |
10ba535a JA |
835 | } |
836 | ||
d7762cf8 | 837 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 838 | { |
02bcaa8c JA |
839 | fio_gettime(&icd->time, NULL); |
840 | ||
3af6ef39 JA |
841 | icd->nr = nr; |
842 | ||
10ba535a JA |
843 | icd->error = 0; |
844 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
845 | } |
846 | ||
97601024 JA |
847 | static void ios_completed(struct thread_data *td, |
848 | struct io_completion_data *icd) | |
36167d82 JA |
849 | { |
850 | struct io_u *io_u; | |
851 | int i; | |
852 | ||
10ba535a JA |
853 | for (i = 0; i < icd->nr; i++) { |
854 | io_u = td->io_ops->event(td, i); | |
855 | ||
856 | io_completed(td, io_u, icd); | |
857 | put_io_u(td, io_u); | |
858 | } | |
859 | } | |
97601024 | 860 | |
e7e6cfb4 JA |
861 | /* |
862 | * Complete a single io_u for the sync engines. | |
863 | */ | |
d7762cf8 | 864 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
865 | { |
866 | struct io_completion_data icd; | |
867 | ||
d7762cf8 | 868 | init_icd(&icd, 1); |
97601024 JA |
869 | io_completed(td, io_u, &icd); |
870 | put_io_u(td, io_u); | |
871 | ||
872 | if (!icd.error) | |
873 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
874 | ||
37e974a5 | 875 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
876 | return -1; |
877 | } | |
878 | ||
e7e6cfb4 JA |
879 | /* |
880 | * Called to complete min_events number of io for the async engines. | |
881 | */ | |
d7762cf8 | 882 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 883 | { |
97601024 | 884 | struct io_completion_data icd; |
00de55ef | 885 | struct timespec *tvp = NULL; |
97601024 | 886 | int ret; |
4d06a338 | 887 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 888 | |
b271fe62 JA |
889 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
890 | ||
4d06a338 | 891 | if (!min_events) |
00de55ef | 892 | tvp = &ts; |
97601024 | 893 | |
00de55ef | 894 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 895 | if (ret < 0) { |
e1161c32 | 896 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
897 | return ret; |
898 | } else if (!ret) | |
899 | return ret; | |
900 | ||
d7762cf8 | 901 | init_icd(&icd, ret); |
97601024 JA |
902 | ios_completed(td, &icd); |
903 | if (!icd.error) | |
904 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
905 | ||
37e974a5 | 906 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
907 | return -1; |
908 | } | |
7e77dd02 JA |
909 | |
910 | /* | |
911 | * Call when io_u is really queued, to update the submission latency. | |
912 | */ | |
913 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
914 | { | |
915 | unsigned long slat_time; | |
916 | ||
d85f5118 | 917 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
918 | add_slat_sample(td, io_u->ddir, slat_time); |
919 | } | |
433afcb4 | 920 | |
55bc9728 | 921 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
922 | void io_u_set_timeout(struct thread_data *td) |
923 | { | |
924 | assert(td->cur_depth); | |
925 | ||
926 | td->timer.it_interval.tv_sec = 0; | |
927 | td->timer.it_interval.tv_usec = 0; | |
928 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
929 | td->timer.it_value.tv_usec = 0; | |
930 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
931 | fio_gettime(&td->timeout_end, NULL); | |
932 | } | |
5945b9b4 JA |
933 | |
934 | static void io_u_dump(struct io_u *io_u) | |
935 | { | |
936 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
937 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
938 | ||
939 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
940 | 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); | |
941 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
942 | } | |
55bc9728 JA |
943 | #else |
944 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
945 | { | |
946 | } | |
947 | #endif | |
433afcb4 | 948 | |
55bc9728 | 949 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
950 | static void io_u_timeout_handler(int fio_unused sig) |
951 | { | |
952 | struct thread_data *td, *__td; | |
953 | pid_t pid = getpid(); | |
5945b9b4 JA |
954 | struct list_head *entry; |
955 | struct io_u *io_u; | |
433afcb4 JA |
956 | int i; |
957 | ||
958 | log_err("fio: io_u timeout\n"); | |
959 | ||
960 | /* | |
961 | * TLS would be nice... | |
962 | */ | |
963 | td = NULL; | |
964 | for_each_td(__td, i) { | |
965 | if (__td->pid == pid) { | |
966 | td = __td; | |
967 | break; | |
968 | } | |
969 | } | |
970 | ||
971 | if (!td) { | |
972 | log_err("fio: io_u timeout, can't find job\n"); | |
973 | exit(1); | |
974 | } | |
975 | ||
976 | if (!td->cur_depth) { | |
977 | log_err("fio: timeout without pending work?\n"); | |
978 | return; | |
979 | } | |
980 | ||
15506d09 | 981 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
982 | |
983 | list_for_each(entry, &td->io_u_busylist) { | |
984 | io_u = list_entry(entry, struct io_u, list); | |
985 | ||
986 | io_u_dump(io_u); | |
987 | } | |
988 | ||
989 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
990 | exit(1); |
991 | } | |
55bc9728 | 992 | #endif |
433afcb4 JA |
993 | |
994 | void io_u_init_timeout(void) | |
995 | { | |
55bc9728 | 996 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 997 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 998 | #endif |
433afcb4 | 999 | } |