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, | |
84422acd | 28 | const unsigned long long block) |
10ba535a JA |
29 | { |
30 | unsigned int idx = RAND_MAP_IDX(td, f, block); | |
31 | unsigned int bit = RAND_MAP_BIT(td, f, block); | |
32 | ||
84422acd JA |
33 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
34 | ||
10ba535a JA |
35 | return (f->file_map[idx] & (1UL << bit)) == 0; |
36 | } | |
37 | ||
df415585 JA |
38 | /* |
39 | * Mark a given offset as used in the map. | |
40 | */ | |
9bf2061e | 41 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 42 | { |
2dc1bbeb | 43 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 44 | struct fio_file *f = io_u->file; |
a00735e6 JA |
45 | unsigned long long block; |
46 | unsigned int blocks; | |
c685b5b2 | 47 | unsigned int nr_blocks; |
df415585 | 48 | |
b9c5b644 | 49 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
a00735e6 | 50 | blocks = 0; |
c685b5b2 JA |
51 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
52 | ||
53 | while (blocks < nr_blocks) { | |
df415585 JA |
54 | unsigned int idx, bit; |
55 | ||
1e3d53ac JA |
56 | /* |
57 | * If we have a mixed random workload, we may | |
58 | * encounter blocks we already did IO to. | |
59 | */ | |
51b6996e | 60 | if ((td->o.ddir_nr == 1) && !random_map_free(td, f, block)) |
df415585 JA |
61 | break; |
62 | ||
63 | idx = RAND_MAP_IDX(td, f, block); | |
64 | bit = RAND_MAP_BIT(td, f, block); | |
65 | ||
0032bf9f | 66 | fio_assert(td, idx < f->num_maps); |
df415585 JA |
67 | |
68 | f->file_map[idx] |= (1UL << bit); | |
69 | block++; | |
70 | blocks++; | |
71 | } | |
72 | ||
a00735e6 JA |
73 | if ((blocks * min_bs) < io_u->buflen) |
74 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
75 | } |
76 | ||
2ba1c290 | 77 | static inline unsigned long long last_block(struct thread_data *td, |
4ba66134 JA |
78 | struct fio_file *f, |
79 | enum fio_ddir ddir) | |
2ba1c290 JA |
80 | { |
81 | unsigned long long max_blocks; | |
82 | ||
84422acd | 83 | max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir]; |
2ba1c290 JA |
84 | if (!max_blocks) |
85 | return 0; | |
86 | ||
87 | return max_blocks - 1; | |
88 | } | |
89 | ||
10ba535a JA |
90 | /* |
91 | * Return the next free block in the map. | |
92 | */ | |
93 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 94 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 95 | { |
84422acd | 96 | unsigned long long min_bs = td->o.rw_min_bs; |
10ba535a JA |
97 | int i; |
98 | ||
c685b5b2 JA |
99 | i = f->last_free_lookup; |
100 | *b = (i * BLOCKS_PER_MAP); | |
84422acd | 101 | while ((*b) * min_bs < f->real_file_size) { |
10ba535a | 102 | if (f->file_map[i] != -1UL) { |
b12ebc65 | 103 | *b += fio_ffz(f->file_map[i]); |
4ba66134 | 104 | if (*b > last_block(td, f, ddir)) |
2ba1c290 | 105 | break; |
c685b5b2 | 106 | f->last_free_lookup = i; |
10ba535a JA |
107 | return 0; |
108 | } | |
109 | ||
110 | *b += BLOCKS_PER_MAP; | |
111 | i++; | |
112 | } | |
113 | ||
2ba1c290 | 114 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
115 | return 1; |
116 | } | |
117 | ||
ec4015da | 118 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 119 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 120 | { |
84422acd | 121 | unsigned long long r; |
ec4015da JA |
122 | int loops = 5; |
123 | ||
124 | do { | |
125 | r = os_random_long(&td->random_state); | |
84422acd | 126 | dprint(FD_RANDOM, "off rand %llu\n", r); |
5ec10eaa JA |
127 | *b = (last_block(td, f, ddir) - 1) |
128 | * (r / ((unsigned long long) RAND_MAX + 1.0)); | |
2ba1c290 | 129 | |
43c63a78 JA |
130 | /* |
131 | * if we are not maintaining a random map, we are done. | |
132 | */ | |
303032ae | 133 | if (!file_randommap(td, f)) |
43c63a78 JA |
134 | return 0; |
135 | ||
136 | /* | |
84422acd | 137 | * calculate map offset and check if it's free |
43c63a78 | 138 | */ |
84422acd | 139 | if (random_map_free(td, f, *b)) |
43c63a78 JA |
140 | return 0; |
141 | ||
84422acd JA |
142 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
143 | *b); | |
43c63a78 | 144 | } while (--loops); |
ec4015da JA |
145 | |
146 | /* | |
43c63a78 JA |
147 | * we get here, if we didn't suceed in looking up a block. generate |
148 | * a random start offset into the filemap, and find the first free | |
149 | * block from there. | |
ec4015da | 150 | */ |
43c63a78 JA |
151 | loops = 10; |
152 | do { | |
153 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); | |
4ba66134 | 154 | if (!get_next_free_block(td, f, ddir, b)) |
43c63a78 | 155 | return 0; |
ec4015da | 156 | |
43c63a78 JA |
157 | r = os_random_long(&td->random_state); |
158 | } while (--loops); | |
159 | ||
160 | /* | |
161 | * that didn't work either, try exhaustive search from the start | |
162 | */ | |
163 | f->last_free_lookup = 0; | |
4ba66134 | 164 | return get_next_free_block(td, f, ddir, b); |
ec4015da JA |
165 | } |
166 | ||
10ba535a JA |
167 | /* |
168 | * For random io, generate a random new block and see if it's used. Repeat | |
169 | * until we find a free one. For sequential io, just return the end of | |
170 | * the last io issued. | |
171 | */ | |
9bf2061e | 172 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 173 | { |
9bf2061e | 174 | struct fio_file *f = io_u->file; |
ec4015da | 175 | unsigned long long b; |
4ba66134 | 176 | enum fio_ddir ddir = io_u->ddir; |
10ba535a | 177 | |
ec4015da JA |
178 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
179 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 180 | |
4ba66134 | 181 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 182 | return 1; |
43063a1c | 183 | } else { |
0c3d768a | 184 | if (f->last_pos >= f->real_file_size) { |
4ba66134 JA |
185 | if (!td_random(td) || |
186 | get_next_rand_offset(td, f, ddir, &b)) | |
0c3d768a | 187 | return 1; |
bcdedd0a | 188 | } else |
4ba66134 | 189 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 190 | } |
10ba535a | 191 | |
4ba66134 | 192 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
2ba1c290 JA |
193 | if (io_u->offset >= f->real_file_size) { |
194 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
195 | io_u->offset, f->real_file_size); | |
10ba535a | 196 | return 1; |
2ba1c290 | 197 | } |
10ba535a JA |
198 | |
199 | return 0; | |
200 | } | |
201 | ||
9bf2061e | 202 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 203 | { |
bca4ed4d | 204 | const int ddir = io_u->ddir; |
10ba535a JA |
205 | unsigned int buflen; |
206 | long r; | |
207 | ||
2dc1bbeb JA |
208 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
209 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
210 | else { |
211 | r = os_random_long(&td->bsrange_state); | |
5ec10eaa JA |
212 | if (!td->o.bssplit_nr) { |
213 | buflen = (unsigned int) | |
214 | (1 + (double) (td->o.max_bs[ddir] - 1) | |
215 | * r / (RAND_MAX + 1.0)); | |
216 | } else { | |
564ca972 JA |
217 | long perc = 0; |
218 | unsigned int i; | |
219 | ||
220 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
221 | struct bssplit *bsp = &td->o.bssplit[i]; | |
222 | ||
223 | buflen = bsp->bs; | |
224 | perc += bsp->perc; | |
225 | if (r <= ((LONG_MAX / 100L) * perc)) | |
226 | break; | |
227 | } | |
228 | } | |
5ec10eaa JA |
229 | if (!td->o.bs_unaligned) { |
230 | buflen = (buflen + td->o.min_bs[ddir] - 1) | |
231 | & ~(td->o.min_bs[ddir] - 1); | |
232 | } | |
10ba535a JA |
233 | } |
234 | ||
4ba66134 JA |
235 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
236 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
237 | td->o.min_bs[ddir], ddir); | |
6a5e6884 | 238 | buflen = td->o.min_bs[ddir]; |
4ba66134 | 239 | } |
6a5e6884 | 240 | |
10ba535a JA |
241 | return buflen; |
242 | } | |
243 | ||
afe24a5a JA |
244 | static void set_rwmix_bytes(struct thread_data *td) |
245 | { | |
e4928662 | 246 | unsigned long issues; |
afe24a5a JA |
247 | unsigned int diff; |
248 | ||
249 | /* | |
250 | * we do time or byte based switch. this is needed because | |
251 | * buffered writes may issue a lot quicker than they complete, | |
252 | * whereas reads do not. | |
253 | */ | |
e4928662 | 254 | issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues; |
e47f799f | 255 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 256 | |
e4928662 JA |
257 | td->rwmix_issues = td->io_issues[td->rwmix_ddir] |
258 | + (issues * ((100 - diff)) / diff); | |
e47f799f JA |
259 | } |
260 | ||
261 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
262 | { | |
263 | unsigned int v; | |
264 | long r; | |
265 | ||
266 | r = os_random_long(&td->rwmix_state); | |
267 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
268 | if (v < td->o.rwmix[DDIR_READ]) | |
269 | return DDIR_READ; | |
270 | ||
271 | return DDIR_WRITE; | |
afe24a5a JA |
272 | } |
273 | ||
10ba535a JA |
274 | /* |
275 | * Return the data direction for the next io_u. If the job is a | |
276 | * mixed read/write workload, check the rwmix cycle and switch if | |
277 | * necessary. | |
278 | */ | |
1e97cce9 | 279 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
280 | { |
281 | if (td_rw(td)) { | |
10ba535a JA |
282 | /* |
283 | * Check if it's time to seed a new data direction. | |
284 | */ | |
e4928662 | 285 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f | 286 | unsigned long long max_bytes; |
5ec10eaa | 287 | enum fio_ddir ddir; |
e47f799f JA |
288 | |
289 | /* | |
290 | * Put a top limit on how many bytes we do for | |
291 | * one data direction, to avoid overflowing the | |
292 | * ranges too much | |
293 | */ | |
294 | ddir = get_rand_ddir(td); | |
295 | max_bytes = td->this_io_bytes[ddir]; | |
5ec10eaa JA |
296 | if (max_bytes >= |
297 | (td->o.size * td->o.rwmix[ddir] / 100)) { | |
e4928662 | 298 | if (!td->rw_end_set[ddir]) |
38d77cae | 299 | td->rw_end_set[ddir] = 1; |
e4928662 | 300 | |
e47f799f | 301 | ddir ^= 1; |
38d77cae | 302 | } |
e47f799f JA |
303 | |
304 | if (ddir != td->rwmix_ddir) | |
305 | set_rwmix_bytes(td); | |
306 | ||
307 | td->rwmix_ddir = ddir; | |
10ba535a JA |
308 | } |
309 | return td->rwmix_ddir; | |
310 | } else if (td_read(td)) | |
311 | return DDIR_READ; | |
312 | else | |
313 | return DDIR_WRITE; | |
314 | } | |
315 | ||
10ba535a JA |
316 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
317 | { | |
0c6e7517 JA |
318 | assert((io_u->flags & IO_U_F_FREE) == 0); |
319 | io_u->flags |= IO_U_F_FREE; | |
320 | ||
6977bcd0 JA |
321 | if (io_u->file) { |
322 | int ret = put_file(td, io_u->file); | |
323 | ||
324 | if (ret) | |
325 | td_verror(td, ret, "file close"); | |
326 | } | |
2dbdab7e | 327 | |
10ba535a JA |
328 | io_u->file = NULL; |
329 | list_del(&io_u->list); | |
330 | list_add(&io_u->list, &td->io_u_freelist); | |
331 | td->cur_depth--; | |
332 | } | |
333 | ||
755200a3 JA |
334 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
335 | { | |
336 | struct io_u *__io_u = *io_u; | |
337 | ||
4d2e0f49 | 338 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
339 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
340 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 341 | |
4d2e0f49 JA |
342 | __io_u->flags &= ~IO_U_F_FLIGHT; |
343 | ||
755200a3 JA |
344 | list_del(&__io_u->list); |
345 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
346 | td->cur_depth--; | |
347 | *io_u = NULL; | |
348 | } | |
349 | ||
9bf2061e | 350 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 351 | { |
b4c5e1ac JA |
352 | if (td->io_ops->flags & FIO_NOIO) |
353 | goto out; | |
354 | ||
87dc1ab1 JA |
355 | /* |
356 | * see if it's time to sync | |
357 | */ | |
2dc1bbeb JA |
358 | if (td->o.fsync_blocks && |
359 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
360 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 361 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 362 | goto out; |
87dc1ab1 JA |
363 | } |
364 | ||
a00735e6 JA |
365 | io_u->ddir = get_rw_ddir(td); |
366 | ||
48f5abd3 JA |
367 | /* |
368 | * See if it's time to switch to a new zone | |
369 | */ | |
370 | if (td->zone_bytes >= td->o.zone_size) { | |
371 | td->zone_bytes = 0; | |
372 | io_u->file->last_pos += td->o.zone_skip; | |
373 | td->io_skip_bytes += td->o.zone_skip; | |
374 | } | |
375 | ||
10ba535a | 376 | /* |
c685b5b2 JA |
377 | * No log, let the seq/rand engine retrieve the next buflen and |
378 | * position. | |
10ba535a | 379 | */ |
2ba1c290 JA |
380 | if (get_next_offset(td, io_u)) { |
381 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 382 | return 1; |
2ba1c290 | 383 | } |
10ba535a | 384 | |
9bf2061e | 385 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
386 | if (!io_u->buflen) { |
387 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 388 | return 1; |
2ba1c290 | 389 | } |
bca4ed4d | 390 | |
2ba1c290 JA |
391 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
392 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
393 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
394 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 395 | return 1; |
2ba1c290 | 396 | } |
6a5e6884 | 397 | |
bca4ed4d JA |
398 | /* |
399 | * mark entry before potentially trimming io_u | |
400 | */ | |
303032ae | 401 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 402 | mark_random_map(td, io_u); |
bca4ed4d JA |
403 | |
404 | /* | |
405 | * If using a write iolog, store this entry. | |
406 | */ | |
c38e9468 | 407 | out: |
2ba1c290 | 408 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 409 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 410 | log_io_u(td, io_u); |
bca4ed4d | 411 | return 0; |
10ba535a JA |
412 | } |
413 | ||
b3605062 | 414 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
415 | { |
416 | int index = 0; | |
417 | ||
b3605062 JA |
418 | if (io_u->ddir == DDIR_SYNC) |
419 | return; | |
420 | ||
71619dc2 JA |
421 | switch (td->cur_depth) { |
422 | default: | |
a783e61a JA |
423 | index = 6; |
424 | break; | |
71619dc2 | 425 | case 32 ... 63: |
a783e61a JA |
426 | index = 5; |
427 | break; | |
71619dc2 | 428 | case 16 ... 31: |
a783e61a JA |
429 | index = 4; |
430 | break; | |
71619dc2 | 431 | case 8 ... 15: |
a783e61a JA |
432 | index = 3; |
433 | break; | |
71619dc2 | 434 | case 4 ... 7: |
a783e61a JA |
435 | index = 2; |
436 | break; | |
71619dc2 | 437 | case 2 ... 3: |
a783e61a | 438 | index = 1; |
71619dc2 JA |
439 | case 1: |
440 | break; | |
441 | } | |
442 | ||
756867bd | 443 | td->ts.io_u_map[index]++; |
b3605062 | 444 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
445 | } |
446 | ||
04a0feae JA |
447 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
448 | { | |
449 | int index = 0; | |
450 | ||
451 | assert(usec < 1000); | |
452 | ||
453 | switch (usec) { | |
454 | case 750 ... 999: | |
455 | index = 9; | |
456 | break; | |
457 | case 500 ... 749: | |
458 | index = 8; | |
459 | break; | |
460 | case 250 ... 499: | |
461 | index = 7; | |
462 | break; | |
463 | case 100 ... 249: | |
464 | index = 6; | |
465 | break; | |
466 | case 50 ... 99: | |
467 | index = 5; | |
468 | break; | |
469 | case 20 ... 49: | |
470 | index = 4; | |
471 | break; | |
472 | case 10 ... 19: | |
473 | index = 3; | |
474 | break; | |
475 | case 4 ... 9: | |
476 | index = 2; | |
477 | break; | |
478 | case 2 ... 3: | |
479 | index = 1; | |
480 | case 0 ... 1: | |
481 | break; | |
482 | } | |
483 | ||
484 | assert(index < FIO_IO_U_LAT_U_NR); | |
485 | td->ts.io_u_lat_u[index]++; | |
486 | } | |
487 | ||
488 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
489 | { |
490 | int index = 0; | |
491 | ||
492 | switch (msec) { | |
493 | default: | |
04a0feae JA |
494 | index = 11; |
495 | break; | |
8abdce66 | 496 | case 1000 ... 1999: |
04a0feae JA |
497 | index = 10; |
498 | break; | |
8abdce66 | 499 | case 750 ... 999: |
04a0feae JA |
500 | index = 9; |
501 | break; | |
8abdce66 | 502 | case 500 ... 749: |
04a0feae JA |
503 | index = 8; |
504 | break; | |
8abdce66 | 505 | case 250 ... 499: |
04a0feae JA |
506 | index = 7; |
507 | break; | |
8abdce66 | 508 | case 100 ... 249: |
04a0feae JA |
509 | index = 6; |
510 | break; | |
8abdce66 | 511 | case 50 ... 99: |
04a0feae JA |
512 | index = 5; |
513 | break; | |
8abdce66 | 514 | case 20 ... 49: |
04a0feae JA |
515 | index = 4; |
516 | break; | |
8abdce66 | 517 | case 10 ... 19: |
04a0feae JA |
518 | index = 3; |
519 | break; | |
8abdce66 | 520 | case 4 ... 9: |
04a0feae JA |
521 | index = 2; |
522 | break; | |
ec118304 | 523 | case 2 ... 3: |
04a0feae | 524 | index = 1; |
ec118304 JA |
525 | case 0 ... 1: |
526 | break; | |
527 | } | |
528 | ||
04a0feae JA |
529 | assert(index < FIO_IO_U_LAT_M_NR); |
530 | td->ts.io_u_lat_m[index]++; | |
531 | } | |
532 | ||
533 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
534 | { | |
535 | if (usec < 1000) | |
536 | io_u_mark_lat_usec(td, usec); | |
537 | else | |
538 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
539 | } |
540 | ||
0aabe160 JA |
541 | /* |
542 | * Get next file to service by choosing one at random | |
543 | */ | |
1c178180 JA |
544 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
545 | int badf) | |
0aabe160 | 546 | { |
0aabe160 | 547 | struct fio_file *f; |
1c178180 | 548 | int fno; |
0aabe160 JA |
549 | |
550 | do { | |
7c83c089 JA |
551 | long r = os_random_long(&td->next_file_state); |
552 | ||
5ec10eaa JA |
553 | fno = (unsigned int) ((double) td->o.nr_files |
554 | * (r / (RAND_MAX + 1.0))); | |
126d65c6 | 555 | f = td->files[fno]; |
059e63c0 JA |
556 | if (f->flags & FIO_FILE_DONE) |
557 | continue; | |
1c178180 | 558 | |
2ba1c290 JA |
559 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
560 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 561 | return f; |
2ba1c290 | 562 | } |
0aabe160 JA |
563 | } while (1); |
564 | } | |
565 | ||
566 | /* | |
567 | * Get next file to service by doing round robin between all available ones | |
568 | */ | |
1c178180 JA |
569 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
570 | int badf) | |
3d7c391d JA |
571 | { |
572 | unsigned int old_next_file = td->next_file; | |
573 | struct fio_file *f; | |
574 | ||
575 | do { | |
126d65c6 | 576 | f = td->files[td->next_file]; |
3d7c391d JA |
577 | |
578 | td->next_file++; | |
2dc1bbeb | 579 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
580 | td->next_file = 0; |
581 | ||
d5ed68ea JA |
582 | if (f->flags & FIO_FILE_DONE) { |
583 | f = NULL; | |
059e63c0 | 584 | continue; |
d5ed68ea | 585 | } |
059e63c0 | 586 | |
1c178180 | 587 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
588 | break; |
589 | ||
590 | f = NULL; | |
591 | } while (td->next_file != old_next_file); | |
592 | ||
2ba1c290 | 593 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
594 | return f; |
595 | } | |
596 | ||
bdb4e2e9 JA |
597 | static struct fio_file *get_next_file(struct thread_data *td) |
598 | { | |
1907dbc6 JA |
599 | struct fio_file *f; |
600 | ||
2dc1bbeb | 601 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 602 | |
2ba1c290 | 603 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
604 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
605 | " nr_files=%d\n", td->nr_open_files, | |
606 | td->nr_done_files, | |
607 | td->o.nr_files); | |
bdb4e2e9 | 608 | return NULL; |
2ba1c290 | 609 | } |
bdb4e2e9 | 610 | |
1907dbc6 | 611 | f = td->file_service_file; |
f11bd94d | 612 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 613 | goto out; |
1907dbc6 | 614 | |
2dc1bbeb | 615 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 616 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 617 | else |
1c178180 | 618 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
619 | |
620 | td->file_service_file = f; | |
621 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
622 | out: |
623 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 624 | return f; |
bdb4e2e9 JA |
625 | } |
626 | ||
1c178180 JA |
627 | static struct fio_file *find_next_new_file(struct thread_data *td) |
628 | { | |
629 | struct fio_file *f; | |
630 | ||
1020a139 JA |
631 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
632 | return NULL; | |
633 | ||
2dc1bbeb | 634 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
635 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
636 | else | |
637 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
638 | ||
639 | return f; | |
640 | } | |
641 | ||
429f6675 JA |
642 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
643 | { | |
644 | struct fio_file *f; | |
645 | ||
646 | do { | |
647 | f = get_next_file(td); | |
648 | if (!f) | |
649 | return 1; | |
650 | ||
651 | set_file: | |
652 | io_u->file = f; | |
653 | get_file(f); | |
654 | ||
655 | if (!fill_io_u(td, io_u)) | |
656 | break; | |
657 | ||
658 | /* | |
659 | * td_io_close() does a put_file() as well, so no need to | |
660 | * do that here. | |
661 | */ | |
662 | io_u->file = NULL; | |
663 | td_io_close_file(td, f); | |
664 | f->flags |= FIO_FILE_DONE; | |
665 | td->nr_done_files++; | |
666 | ||
667 | /* | |
668 | * probably not the right place to do this, but see | |
669 | * if we need to open a new file | |
670 | */ | |
671 | if (td->nr_open_files < td->o.open_files && | |
672 | td->o.open_files != td->o.nr_files) { | |
673 | f = find_next_new_file(td); | |
674 | ||
675 | if (!f || td_io_open_file(td, f)) | |
676 | return 1; | |
677 | ||
678 | goto set_file; | |
679 | } | |
680 | } while (1); | |
681 | ||
682 | return 0; | |
683 | } | |
684 | ||
685 | ||
10ba535a JA |
686 | struct io_u *__get_io_u(struct thread_data *td) |
687 | { | |
688 | struct io_u *io_u = NULL; | |
689 | ||
755200a3 JA |
690 | if (!list_empty(&td->io_u_requeues)) |
691 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
692 | else if (!queue_full(td)) { | |
10ba535a JA |
693 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
694 | ||
6040dabc | 695 | io_u->buflen = 0; |
10ba535a | 696 | io_u->resid = 0; |
755200a3 | 697 | io_u->file = NULL; |
d7762cf8 | 698 | io_u->end_io = NULL; |
755200a3 JA |
699 | } |
700 | ||
701 | if (io_u) { | |
0c6e7517 JA |
702 | assert(io_u->flags & IO_U_F_FREE); |
703 | io_u->flags &= ~IO_U_F_FREE; | |
704 | ||
755200a3 | 705 | io_u->error = 0; |
10ba535a JA |
706 | list_del(&io_u->list); |
707 | list_add(&io_u->list, &td->io_u_busylist); | |
708 | td->cur_depth++; | |
709 | } | |
710 | ||
711 | return io_u; | |
712 | } | |
713 | ||
714 | /* | |
715 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
716 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
717 | */ | |
3d7c391d | 718 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 719 | { |
3d7c391d | 720 | struct fio_file *f; |
10ba535a JA |
721 | struct io_u *io_u; |
722 | ||
723 | io_u = __get_io_u(td); | |
2ba1c290 JA |
724 | if (!io_u) { |
725 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 726 | return NULL; |
2ba1c290 | 727 | } |
10ba535a | 728 | |
755200a3 JA |
729 | /* |
730 | * from a requeue, io_u already setup | |
731 | */ | |
732 | if (io_u->file) | |
77f392bf | 733 | goto out; |
755200a3 | 734 | |
429f6675 JA |
735 | /* |
736 | * If using an iolog, grab next piece if any available. | |
737 | */ | |
738 | if (td->o.read_iolog_file) { | |
739 | if (read_iolog_get(td, io_u)) | |
740 | goto err_put; | |
2ba1c290 JA |
741 | } else if (set_io_u_file(td, io_u)) { |
742 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 743 | goto err_put; |
2ba1c290 | 744 | } |
5ec10eaa | 745 | |
429f6675 JA |
746 | f = io_u->file; |
747 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 748 | |
87dc1ab1 | 749 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 750 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 751 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 752 | goto err_put; |
2ba1c290 | 753 | } |
10ba535a | 754 | |
36167d82 | 755 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 756 | |
2dc1bbeb | 757 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
758 | populate_verify_io_u(td, io_u); |
759 | } | |
10ba535a | 760 | |
165faf16 JA |
761 | /* |
762 | * Set io data pointers. | |
763 | */ | |
d460eb31 | 764 | io_u->endpos = io_u->offset + io_u->buflen; |
cec6b55d JA |
765 | io_u->xfer_buf = io_u->buf; |
766 | io_u->xfer_buflen = io_u->buflen; | |
6ac7a331 | 767 | out: |
429f6675 JA |
768 | if (!td_io_prep(td, io_u)) { |
769 | fio_gettime(&io_u->start_time, NULL); | |
770 | return io_u; | |
36167d82 | 771 | } |
429f6675 | 772 | err_put: |
2ba1c290 | 773 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
774 | put_io_u(td, io_u); |
775 | return NULL; | |
10ba535a JA |
776 | } |
777 | ||
5451792e JA |
778 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
779 | { | |
780 | const char *msg[] = { "read", "write", "sync" }; | |
781 | ||
782 | log_err("fio: io_u error"); | |
783 | ||
784 | if (io_u->file) | |
785 | log_err(" on file %s", io_u->file->file_name); | |
786 | ||
787 | log_err(": %s\n", strerror(io_u->error)); | |
788 | ||
5ec10eaa JA |
789 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
790 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
791 | |
792 | if (!td->error) | |
793 | td_verror(td, io_u->error, "io_u error"); | |
794 | } | |
795 | ||
97601024 JA |
796 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
797 | struct io_completion_data *icd) | |
10ba535a | 798 | { |
d85f5118 | 799 | unsigned long usec; |
10ba535a | 800 | |
2ba1c290 JA |
801 | dprint_io_u(io_u, "io complete"); |
802 | ||
0c6e7517 JA |
803 | assert(io_u->flags & IO_U_F_FLIGHT); |
804 | io_u->flags &= ~IO_U_F_FLIGHT; | |
805 | ||
87dc1ab1 JA |
806 | if (io_u->ddir == DDIR_SYNC) { |
807 | td->last_was_sync = 1; | |
808 | return; | |
809 | } | |
810 | ||
811 | td->last_was_sync = 0; | |
812 | ||
10ba535a JA |
813 | if (!io_u->error) { |
814 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 815 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 816 | int ret; |
10ba535a JA |
817 | |
818 | td->io_blocks[idx]++; | |
819 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
820 | td->this_io_bytes[idx] += bytes; |
821 | ||
d85f5118 | 822 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 823 | |
d85f5118 | 824 | add_clat_sample(td, idx, usec); |
02bcaa8c | 825 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 826 | io_u_mark_latency(td, usec); |
10ba535a | 827 | |
660a1cb5 | 828 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 829 | td->o.do_verify && |
41128405 | 830 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
831 | log_io_piece(td, io_u); |
832 | ||
833 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 834 | |
d7762cf8 | 835 | if (io_u->end_io) { |
36690c9b | 836 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
837 | if (ret && !icd->error) |
838 | icd->error = ret; | |
839 | } | |
5451792e | 840 | } else { |
10ba535a | 841 | icd->error = io_u->error; |
5451792e JA |
842 | io_u_log_error(td, io_u); |
843 | } | |
10ba535a JA |
844 | } |
845 | ||
d7762cf8 | 846 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 847 | { |
02bcaa8c JA |
848 | fio_gettime(&icd->time, NULL); |
849 | ||
3af6ef39 JA |
850 | icd->nr = nr; |
851 | ||
10ba535a JA |
852 | icd->error = 0; |
853 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
854 | } |
855 | ||
97601024 JA |
856 | static void ios_completed(struct thread_data *td, |
857 | struct io_completion_data *icd) | |
36167d82 JA |
858 | { |
859 | struct io_u *io_u; | |
860 | int i; | |
861 | ||
10ba535a JA |
862 | for (i = 0; i < icd->nr; i++) { |
863 | io_u = td->io_ops->event(td, i); | |
864 | ||
865 | io_completed(td, io_u, icd); | |
866 | put_io_u(td, io_u); | |
867 | } | |
868 | } | |
97601024 | 869 | |
e7e6cfb4 JA |
870 | /* |
871 | * Complete a single io_u for the sync engines. | |
872 | */ | |
d7762cf8 | 873 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
874 | { |
875 | struct io_completion_data icd; | |
876 | ||
d7762cf8 | 877 | init_icd(&icd, 1); |
97601024 JA |
878 | io_completed(td, io_u, &icd); |
879 | put_io_u(td, io_u); | |
880 | ||
881 | if (!icd.error) | |
882 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
883 | ||
37e974a5 | 884 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
885 | return -1; |
886 | } | |
887 | ||
e7e6cfb4 JA |
888 | /* |
889 | * Called to complete min_events number of io for the async engines. | |
890 | */ | |
d7762cf8 | 891 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 892 | { |
97601024 | 893 | struct io_completion_data icd; |
00de55ef | 894 | struct timespec *tvp = NULL; |
97601024 | 895 | int ret; |
4d06a338 | 896 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 897 | |
b271fe62 JA |
898 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
899 | ||
4d06a338 | 900 | if (!min_events) |
00de55ef | 901 | tvp = &ts; |
97601024 | 902 | |
00de55ef | 903 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 904 | if (ret < 0) { |
e1161c32 | 905 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
906 | return ret; |
907 | } else if (!ret) | |
908 | return ret; | |
909 | ||
d7762cf8 | 910 | init_icd(&icd, ret); |
97601024 JA |
911 | ios_completed(td, &icd); |
912 | if (!icd.error) | |
913 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
914 | ||
37e974a5 | 915 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
916 | return -1; |
917 | } | |
7e77dd02 JA |
918 | |
919 | /* | |
920 | * Call when io_u is really queued, to update the submission latency. | |
921 | */ | |
922 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
923 | { | |
924 | unsigned long slat_time; | |
925 | ||
d85f5118 | 926 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
927 | add_slat_sample(td, io_u->ddir, slat_time); |
928 | } | |
433afcb4 | 929 | |
55bc9728 | 930 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
931 | void io_u_set_timeout(struct thread_data *td) |
932 | { | |
933 | assert(td->cur_depth); | |
934 | ||
935 | td->timer.it_interval.tv_sec = 0; | |
936 | td->timer.it_interval.tv_usec = 0; | |
937 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
938 | td->timer.it_value.tv_usec = 0; | |
939 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
940 | fio_gettime(&td->timeout_end, NULL); | |
941 | } | |
5945b9b4 JA |
942 | |
943 | static void io_u_dump(struct io_u *io_u) | |
944 | { | |
945 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
946 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
947 | ||
948 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
5ec10eaa JA |
949 | log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, |
950 | io_u->xfer_buf, io_u->buflen, | |
951 | io_u->xfer_buflen, | |
952 | io_u->offset); | |
5945b9b4 JA |
953 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); |
954 | } | |
55bc9728 JA |
955 | #else |
956 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
957 | { | |
958 | } | |
959 | #endif | |
433afcb4 | 960 | |
55bc9728 | 961 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
962 | static void io_u_timeout_handler(int fio_unused sig) |
963 | { | |
964 | struct thread_data *td, *__td; | |
965 | pid_t pid = getpid(); | |
5945b9b4 JA |
966 | struct list_head *entry; |
967 | struct io_u *io_u; | |
433afcb4 JA |
968 | int i; |
969 | ||
970 | log_err("fio: io_u timeout\n"); | |
971 | ||
972 | /* | |
973 | * TLS would be nice... | |
974 | */ | |
975 | td = NULL; | |
976 | for_each_td(__td, i) { | |
977 | if (__td->pid == pid) { | |
978 | td = __td; | |
979 | break; | |
980 | } | |
981 | } | |
982 | ||
983 | if (!td) { | |
984 | log_err("fio: io_u timeout, can't find job\n"); | |
985 | exit(1); | |
986 | } | |
987 | ||
988 | if (!td->cur_depth) { | |
989 | log_err("fio: timeout without pending work?\n"); | |
990 | return; | |
991 | } | |
992 | ||
15506d09 | 993 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
994 | |
995 | list_for_each(entry, &td->io_u_busylist) { | |
996 | io_u = list_entry(entry, struct io_u, list); | |
997 | ||
998 | io_u_dump(io_u); | |
999 | } | |
1000 | ||
1001 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
1002 | exit(1); |
1003 | } | |
55bc9728 | 1004 | #endif |
433afcb4 JA |
1005 | |
1006 | void io_u_init_timeout(void) | |
1007 | { | |
55bc9728 | 1008 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 1009 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 1010 | #endif |
433afcb4 | 1011 | } |