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 | */ | |
aec2de20 | 27 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 28 | { |
aec2de20 JA |
29 | unsigned int idx = RAND_MAP_IDX(f, block); |
30 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 31 | |
84422acd JA |
32 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
33 | ||
10ba535a JA |
34 | return (f->file_map[idx] & (1UL << bit)) == 0; |
35 | } | |
36 | ||
df415585 JA |
37 | /* |
38 | * Mark a given offset as used in the map. | |
39 | */ | |
9bf2061e | 40 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 41 | { |
2dc1bbeb | 42 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 43 | struct fio_file *f = io_u->file; |
a00735e6 JA |
44 | unsigned long long block; |
45 | unsigned int blocks; | |
c685b5b2 | 46 | unsigned int nr_blocks; |
df415585 | 47 | |
b9c5b644 | 48 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
a00735e6 | 49 | blocks = 0; |
c685b5b2 JA |
50 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
51 | ||
52 | while (blocks < nr_blocks) { | |
df415585 JA |
53 | unsigned int idx, bit; |
54 | ||
1e3d53ac JA |
55 | /* |
56 | * If we have a mixed random workload, we may | |
57 | * encounter blocks we already did IO to. | |
58 | */ | |
aec2de20 | 59 | if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) |
df415585 JA |
60 | break; |
61 | ||
aec2de20 JA |
62 | idx = RAND_MAP_IDX(f, block); |
63 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 64 | |
0032bf9f | 65 | fio_assert(td, idx < f->num_maps); |
df415585 JA |
66 | |
67 | f->file_map[idx] |= (1UL << bit); | |
68 | block++; | |
69 | blocks++; | |
70 | } | |
71 | ||
a00735e6 JA |
72 | if ((blocks * min_bs) < io_u->buflen) |
73 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
74 | } |
75 | ||
2ba1c290 | 76 | static inline unsigned long long last_block(struct thread_data *td, |
4ba66134 JA |
77 | struct fio_file *f, |
78 | enum fio_ddir ddir) | |
2ba1c290 JA |
79 | { |
80 | unsigned long long max_blocks; | |
81 | ||
84422acd | 82 | max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir]; |
2ba1c290 JA |
83 | if (!max_blocks) |
84 | return 0; | |
85 | ||
67778e88 | 86 | return max_blocks; |
2ba1c290 JA |
87 | } |
88 | ||
10ba535a JA |
89 | /* |
90 | * Return the next free block in the map. | |
91 | */ | |
92 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 93 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 94 | { |
84422acd | 95 | unsigned long long min_bs = td->o.rw_min_bs; |
10ba535a JA |
96 | int i; |
97 | ||
c685b5b2 JA |
98 | i = f->last_free_lookup; |
99 | *b = (i * BLOCKS_PER_MAP); | |
84422acd | 100 | while ((*b) * min_bs < f->real_file_size) { |
10ba535a | 101 | if (f->file_map[i] != -1UL) { |
b12ebc65 | 102 | *b += fio_ffz(f->file_map[i]); |
4ba66134 | 103 | if (*b > last_block(td, f, ddir)) |
2ba1c290 | 104 | break; |
c685b5b2 | 105 | f->last_free_lookup = i; |
10ba535a JA |
106 | return 0; |
107 | } | |
108 | ||
109 | *b += BLOCKS_PER_MAP; | |
110 | i++; | |
111 | } | |
112 | ||
2ba1c290 | 113 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
114 | return 1; |
115 | } | |
116 | ||
ec4015da | 117 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 118 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 119 | { |
84422acd | 120 | unsigned long long r; |
ec4015da JA |
121 | int loops = 5; |
122 | ||
123 | do { | |
124 | r = os_random_long(&td->random_state); | |
84422acd | 125 | dprint(FD_RANDOM, "off rand %llu\n", r); |
5ec10eaa JA |
126 | *b = (last_block(td, f, ddir) - 1) |
127 | * (r / ((unsigned long long) RAND_MAX + 1.0)); | |
2ba1c290 | 128 | |
43c63a78 JA |
129 | /* |
130 | * if we are not maintaining a random map, we are done. | |
131 | */ | |
303032ae | 132 | if (!file_randommap(td, f)) |
43c63a78 JA |
133 | return 0; |
134 | ||
135 | /* | |
84422acd | 136 | * calculate map offset and check if it's free |
43c63a78 | 137 | */ |
aec2de20 | 138 | if (random_map_free(f, *b)) |
43c63a78 JA |
139 | return 0; |
140 | ||
84422acd JA |
141 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
142 | *b); | |
43c63a78 | 143 | } while (--loops); |
ec4015da JA |
144 | |
145 | /* | |
43c63a78 JA |
146 | * we get here, if we didn't suceed in looking up a block. generate |
147 | * a random start offset into the filemap, and find the first free | |
148 | * block from there. | |
ec4015da | 149 | */ |
43c63a78 JA |
150 | loops = 10; |
151 | do { | |
152 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); | |
4ba66134 | 153 | if (!get_next_free_block(td, f, ddir, b)) |
43c63a78 | 154 | return 0; |
ec4015da | 155 | |
43c63a78 JA |
156 | r = os_random_long(&td->random_state); |
157 | } while (--loops); | |
158 | ||
159 | /* | |
160 | * that didn't work either, try exhaustive search from the start | |
161 | */ | |
162 | f->last_free_lookup = 0; | |
4ba66134 | 163 | return get_next_free_block(td, f, ddir, b); |
ec4015da JA |
164 | } |
165 | ||
10ba535a JA |
166 | /* |
167 | * For random io, generate a random new block and see if it's used. Repeat | |
168 | * until we find a free one. For sequential io, just return the end of | |
169 | * the last io issued. | |
170 | */ | |
9bf2061e | 171 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 172 | { |
9bf2061e | 173 | struct fio_file *f = io_u->file; |
ec4015da | 174 | unsigned long long b; |
4ba66134 | 175 | enum fio_ddir ddir = io_u->ddir; |
10ba535a | 176 | |
ec4015da JA |
177 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
178 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 179 | |
4ba66134 | 180 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 181 | return 1; |
43063a1c | 182 | } else { |
0c3d768a | 183 | if (f->last_pos >= f->real_file_size) { |
4ba66134 JA |
184 | if (!td_random(td) || |
185 | get_next_rand_offset(td, f, ddir, &b)) | |
0c3d768a | 186 | return 1; |
bcdedd0a | 187 | } else |
4ba66134 | 188 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 189 | } |
10ba535a | 190 | |
009bd847 JA |
191 | io_u->offset = b * td->o.min_bs[ddir]; |
192 | if (io_u->offset >= f->io_size) { | |
193 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
194 | io_u->offset, f->io_size); | |
195 | return 1; | |
196 | } | |
197 | ||
198 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
199 | if (io_u->offset >= f->real_file_size) { |
200 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
201 | io_u->offset, f->real_file_size); | |
10ba535a | 202 | return 1; |
2ba1c290 | 203 | } |
10ba535a JA |
204 | |
205 | return 0; | |
206 | } | |
207 | ||
9bf2061e | 208 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 209 | { |
bca4ed4d | 210 | const int ddir = io_u->ddir; |
10ba535a JA |
211 | unsigned int buflen; |
212 | long r; | |
213 | ||
2dc1bbeb JA |
214 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
215 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
216 | else { |
217 | r = os_random_long(&td->bsrange_state); | |
5ec10eaa JA |
218 | if (!td->o.bssplit_nr) { |
219 | buflen = (unsigned int) | |
220 | (1 + (double) (td->o.max_bs[ddir] - 1) | |
221 | * r / (RAND_MAX + 1.0)); | |
222 | } else { | |
564ca972 JA |
223 | long perc = 0; |
224 | unsigned int i; | |
225 | ||
226 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
227 | struct bssplit *bsp = &td->o.bssplit[i]; | |
228 | ||
229 | buflen = bsp->bs; | |
230 | perc += bsp->perc; | |
231 | if (r <= ((LONG_MAX / 100L) * perc)) | |
232 | break; | |
233 | } | |
234 | } | |
5ec10eaa JA |
235 | if (!td->o.bs_unaligned) { |
236 | buflen = (buflen + td->o.min_bs[ddir] - 1) | |
237 | & ~(td->o.min_bs[ddir] - 1); | |
238 | } | |
10ba535a JA |
239 | } |
240 | ||
4ba66134 JA |
241 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
242 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
243 | td->o.min_bs[ddir], ddir); | |
6a5e6884 | 244 | buflen = td->o.min_bs[ddir]; |
4ba66134 | 245 | } |
6a5e6884 | 246 | |
10ba535a JA |
247 | return buflen; |
248 | } | |
249 | ||
afe24a5a JA |
250 | static void set_rwmix_bytes(struct thread_data *td) |
251 | { | |
e4928662 | 252 | unsigned long issues; |
afe24a5a JA |
253 | unsigned int diff; |
254 | ||
255 | /* | |
256 | * we do time or byte based switch. this is needed because | |
257 | * buffered writes may issue a lot quicker than they complete, | |
258 | * whereas reads do not. | |
259 | */ | |
e4928662 | 260 | issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues; |
e47f799f | 261 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 262 | |
e4928662 JA |
263 | td->rwmix_issues = td->io_issues[td->rwmix_ddir] |
264 | + (issues * ((100 - diff)) / diff); | |
e47f799f JA |
265 | } |
266 | ||
267 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
268 | { | |
269 | unsigned int v; | |
270 | long r; | |
271 | ||
272 | r = os_random_long(&td->rwmix_state); | |
273 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
274 | if (v < td->o.rwmix[DDIR_READ]) | |
275 | return DDIR_READ; | |
276 | ||
277 | return DDIR_WRITE; | |
afe24a5a JA |
278 | } |
279 | ||
10ba535a JA |
280 | /* |
281 | * Return the data direction for the next io_u. If the job is a | |
282 | * mixed read/write workload, check the rwmix cycle and switch if | |
283 | * necessary. | |
284 | */ | |
1e97cce9 | 285 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
286 | { |
287 | if (td_rw(td)) { | |
10ba535a JA |
288 | /* |
289 | * Check if it's time to seed a new data direction. | |
290 | */ | |
e4928662 | 291 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f | 292 | unsigned long long max_bytes; |
5ec10eaa | 293 | enum fio_ddir ddir; |
e47f799f JA |
294 | |
295 | /* | |
296 | * Put a top limit on how many bytes we do for | |
297 | * one data direction, to avoid overflowing the | |
298 | * ranges too much | |
299 | */ | |
300 | ddir = get_rand_ddir(td); | |
301 | max_bytes = td->this_io_bytes[ddir]; | |
5ec10eaa JA |
302 | if (max_bytes >= |
303 | (td->o.size * td->o.rwmix[ddir] / 100)) { | |
e4928662 | 304 | if (!td->rw_end_set[ddir]) |
38d77cae | 305 | td->rw_end_set[ddir] = 1; |
e4928662 | 306 | |
e47f799f | 307 | ddir ^= 1; |
38d77cae | 308 | } |
e47f799f JA |
309 | |
310 | if (ddir != td->rwmix_ddir) | |
311 | set_rwmix_bytes(td); | |
312 | ||
313 | td->rwmix_ddir = ddir; | |
10ba535a JA |
314 | } |
315 | return td->rwmix_ddir; | |
316 | } else if (td_read(td)) | |
317 | return DDIR_READ; | |
318 | else | |
319 | return DDIR_WRITE; | |
320 | } | |
321 | ||
10ba535a JA |
322 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
323 | { | |
0c6e7517 JA |
324 | assert((io_u->flags & IO_U_F_FREE) == 0); |
325 | io_u->flags |= IO_U_F_FREE; | |
326 | ||
6977bcd0 JA |
327 | if (io_u->file) { |
328 | int ret = put_file(td, io_u->file); | |
329 | ||
330 | if (ret) | |
331 | td_verror(td, ret, "file close"); | |
332 | } | |
2dbdab7e | 333 | |
10ba535a JA |
334 | io_u->file = NULL; |
335 | list_del(&io_u->list); | |
336 | list_add(&io_u->list, &td->io_u_freelist); | |
337 | td->cur_depth--; | |
338 | } | |
339 | ||
755200a3 JA |
340 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
341 | { | |
342 | struct io_u *__io_u = *io_u; | |
343 | ||
4d2e0f49 | 344 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
345 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
346 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 347 | |
4d2e0f49 JA |
348 | __io_u->flags &= ~IO_U_F_FLIGHT; |
349 | ||
755200a3 JA |
350 | list_del(&__io_u->list); |
351 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
352 | td->cur_depth--; | |
353 | *io_u = NULL; | |
354 | } | |
355 | ||
9bf2061e | 356 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 357 | { |
b4c5e1ac JA |
358 | if (td->io_ops->flags & FIO_NOIO) |
359 | goto out; | |
360 | ||
87dc1ab1 JA |
361 | /* |
362 | * see if it's time to sync | |
363 | */ | |
2dc1bbeb JA |
364 | if (td->o.fsync_blocks && |
365 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
366 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 367 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 368 | goto out; |
87dc1ab1 JA |
369 | } |
370 | ||
a00735e6 JA |
371 | io_u->ddir = get_rw_ddir(td); |
372 | ||
48f5abd3 JA |
373 | /* |
374 | * See if it's time to switch to a new zone | |
375 | */ | |
376 | if (td->zone_bytes >= td->o.zone_size) { | |
377 | td->zone_bytes = 0; | |
378 | io_u->file->last_pos += td->o.zone_skip; | |
379 | td->io_skip_bytes += td->o.zone_skip; | |
380 | } | |
381 | ||
10ba535a | 382 | /* |
c685b5b2 JA |
383 | * No log, let the seq/rand engine retrieve the next buflen and |
384 | * position. | |
10ba535a | 385 | */ |
2ba1c290 JA |
386 | if (get_next_offset(td, io_u)) { |
387 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 388 | return 1; |
2ba1c290 | 389 | } |
10ba535a | 390 | |
9bf2061e | 391 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
392 | if (!io_u->buflen) { |
393 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 394 | return 1; |
2ba1c290 | 395 | } |
bca4ed4d | 396 | |
2ba1c290 JA |
397 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
398 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
399 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
400 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 401 | return 1; |
2ba1c290 | 402 | } |
6a5e6884 | 403 | |
bca4ed4d JA |
404 | /* |
405 | * mark entry before potentially trimming io_u | |
406 | */ | |
303032ae | 407 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 408 | mark_random_map(td, io_u); |
bca4ed4d JA |
409 | |
410 | /* | |
411 | * If using a write iolog, store this entry. | |
412 | */ | |
c38e9468 | 413 | out: |
2ba1c290 | 414 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 415 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 416 | log_io_u(td, io_u); |
bca4ed4d | 417 | return 0; |
10ba535a JA |
418 | } |
419 | ||
d8005759 | 420 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 JA |
421 | { |
422 | int index = 0; | |
423 | ||
424 | switch (td->cur_depth) { | |
425 | default: | |
a783e61a JA |
426 | index = 6; |
427 | break; | |
71619dc2 | 428 | case 32 ... 63: |
a783e61a JA |
429 | index = 5; |
430 | break; | |
71619dc2 | 431 | case 16 ... 31: |
a783e61a JA |
432 | index = 4; |
433 | break; | |
71619dc2 | 434 | case 8 ... 15: |
a783e61a JA |
435 | index = 3; |
436 | break; | |
71619dc2 | 437 | case 4 ... 7: |
a783e61a JA |
438 | index = 2; |
439 | break; | |
71619dc2 | 440 | case 2 ... 3: |
a783e61a | 441 | index = 1; |
71619dc2 JA |
442 | case 1: |
443 | break; | |
444 | } | |
445 | ||
3bec7ae0 | 446 | td->ts.io_u_map[index] += nr; |
71619dc2 JA |
447 | } |
448 | ||
04a0feae JA |
449 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
450 | { | |
451 | int index = 0; | |
452 | ||
453 | assert(usec < 1000); | |
454 | ||
455 | switch (usec) { | |
456 | case 750 ... 999: | |
457 | index = 9; | |
458 | break; | |
459 | case 500 ... 749: | |
460 | index = 8; | |
461 | break; | |
462 | case 250 ... 499: | |
463 | index = 7; | |
464 | break; | |
465 | case 100 ... 249: | |
466 | index = 6; | |
467 | break; | |
468 | case 50 ... 99: | |
469 | index = 5; | |
470 | break; | |
471 | case 20 ... 49: | |
472 | index = 4; | |
473 | break; | |
474 | case 10 ... 19: | |
475 | index = 3; | |
476 | break; | |
477 | case 4 ... 9: | |
478 | index = 2; | |
479 | break; | |
480 | case 2 ... 3: | |
481 | index = 1; | |
482 | case 0 ... 1: | |
483 | break; | |
484 | } | |
485 | ||
486 | assert(index < FIO_IO_U_LAT_U_NR); | |
487 | td->ts.io_u_lat_u[index]++; | |
488 | } | |
489 | ||
490 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
491 | { |
492 | int index = 0; | |
493 | ||
494 | switch (msec) { | |
495 | default: | |
04a0feae JA |
496 | index = 11; |
497 | break; | |
8abdce66 | 498 | case 1000 ... 1999: |
04a0feae JA |
499 | index = 10; |
500 | break; | |
8abdce66 | 501 | case 750 ... 999: |
04a0feae JA |
502 | index = 9; |
503 | break; | |
8abdce66 | 504 | case 500 ... 749: |
04a0feae JA |
505 | index = 8; |
506 | break; | |
8abdce66 | 507 | case 250 ... 499: |
04a0feae JA |
508 | index = 7; |
509 | break; | |
8abdce66 | 510 | case 100 ... 249: |
04a0feae JA |
511 | index = 6; |
512 | break; | |
8abdce66 | 513 | case 50 ... 99: |
04a0feae JA |
514 | index = 5; |
515 | break; | |
8abdce66 | 516 | case 20 ... 49: |
04a0feae JA |
517 | index = 4; |
518 | break; | |
8abdce66 | 519 | case 10 ... 19: |
04a0feae JA |
520 | index = 3; |
521 | break; | |
8abdce66 | 522 | case 4 ... 9: |
04a0feae JA |
523 | index = 2; |
524 | break; | |
ec118304 | 525 | case 2 ... 3: |
04a0feae | 526 | index = 1; |
ec118304 JA |
527 | case 0 ... 1: |
528 | break; | |
529 | } | |
530 | ||
04a0feae JA |
531 | assert(index < FIO_IO_U_LAT_M_NR); |
532 | td->ts.io_u_lat_m[index]++; | |
533 | } | |
534 | ||
535 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
536 | { | |
537 | if (usec < 1000) | |
538 | io_u_mark_lat_usec(td, usec); | |
539 | else | |
540 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
541 | } |
542 | ||
0aabe160 JA |
543 | /* |
544 | * Get next file to service by choosing one at random | |
545 | */ | |
1c178180 JA |
546 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
547 | int badf) | |
0aabe160 | 548 | { |
0aabe160 | 549 | struct fio_file *f; |
1c178180 | 550 | int fno; |
0aabe160 JA |
551 | |
552 | do { | |
7c83c089 JA |
553 | long r = os_random_long(&td->next_file_state); |
554 | ||
5ec10eaa JA |
555 | fno = (unsigned int) ((double) td->o.nr_files |
556 | * (r / (RAND_MAX + 1.0))); | |
126d65c6 | 557 | f = td->files[fno]; |
059e63c0 JA |
558 | if (f->flags & FIO_FILE_DONE) |
559 | continue; | |
1c178180 | 560 | |
2ba1c290 JA |
561 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
562 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 563 | return f; |
2ba1c290 | 564 | } |
0aabe160 JA |
565 | } while (1); |
566 | } | |
567 | ||
568 | /* | |
569 | * Get next file to service by doing round robin between all available ones | |
570 | */ | |
1c178180 JA |
571 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
572 | int badf) | |
3d7c391d JA |
573 | { |
574 | unsigned int old_next_file = td->next_file; | |
575 | struct fio_file *f; | |
576 | ||
577 | do { | |
126d65c6 | 578 | f = td->files[td->next_file]; |
3d7c391d JA |
579 | |
580 | td->next_file++; | |
2dc1bbeb | 581 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
582 | td->next_file = 0; |
583 | ||
d5ed68ea JA |
584 | if (f->flags & FIO_FILE_DONE) { |
585 | f = NULL; | |
059e63c0 | 586 | continue; |
d5ed68ea | 587 | } |
059e63c0 | 588 | |
1c178180 | 589 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
590 | break; |
591 | ||
592 | f = NULL; | |
593 | } while (td->next_file != old_next_file); | |
594 | ||
2ba1c290 | 595 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
596 | return f; |
597 | } | |
598 | ||
bdb4e2e9 JA |
599 | static struct fio_file *get_next_file(struct thread_data *td) |
600 | { | |
1907dbc6 JA |
601 | struct fio_file *f; |
602 | ||
2dc1bbeb | 603 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 604 | |
2ba1c290 | 605 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
606 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
607 | " nr_files=%d\n", td->nr_open_files, | |
608 | td->nr_done_files, | |
609 | td->o.nr_files); | |
bdb4e2e9 | 610 | return NULL; |
2ba1c290 | 611 | } |
bdb4e2e9 | 612 | |
1907dbc6 | 613 | f = td->file_service_file; |
f11bd94d | 614 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 615 | goto out; |
1907dbc6 | 616 | |
2dc1bbeb | 617 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 618 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 619 | else |
1c178180 | 620 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
621 | |
622 | td->file_service_file = f; | |
623 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
624 | out: |
625 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 626 | return f; |
bdb4e2e9 JA |
627 | } |
628 | ||
1c178180 JA |
629 | static struct fio_file *find_next_new_file(struct thread_data *td) |
630 | { | |
631 | struct fio_file *f; | |
632 | ||
1020a139 JA |
633 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
634 | return NULL; | |
635 | ||
2dc1bbeb | 636 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
637 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
638 | else | |
639 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
640 | ||
641 | return f; | |
642 | } | |
643 | ||
429f6675 JA |
644 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
645 | { | |
646 | struct fio_file *f; | |
647 | ||
648 | do { | |
649 | f = get_next_file(td); | |
650 | if (!f) | |
651 | return 1; | |
652 | ||
653 | set_file: | |
654 | io_u->file = f; | |
655 | get_file(f); | |
656 | ||
657 | if (!fill_io_u(td, io_u)) | |
658 | break; | |
659 | ||
009bd847 JA |
660 | /* |
661 | * optimization to prevent close/open of the same file. This | |
662 | * way we preserve queueing etc. | |
663 | */ | |
664 | if (td->o.nr_files == 1 && td->o.time_based) { | |
665 | put_file(td, f); | |
666 | fio_file_reset(f); | |
667 | goto set_file; | |
668 | } | |
669 | ||
429f6675 JA |
670 | /* |
671 | * td_io_close() does a put_file() as well, so no need to | |
672 | * do that here. | |
673 | */ | |
674 | io_u->file = NULL; | |
675 | td_io_close_file(td, f); | |
676 | f->flags |= FIO_FILE_DONE; | |
677 | td->nr_done_files++; | |
678 | ||
679 | /* | |
680 | * probably not the right place to do this, but see | |
681 | * if we need to open a new file | |
682 | */ | |
683 | if (td->nr_open_files < td->o.open_files && | |
684 | td->o.open_files != td->o.nr_files) { | |
685 | f = find_next_new_file(td); | |
686 | ||
687 | if (!f || td_io_open_file(td, f)) | |
688 | return 1; | |
689 | ||
690 | goto set_file; | |
691 | } | |
692 | } while (1); | |
693 | ||
694 | return 0; | |
695 | } | |
696 | ||
697 | ||
10ba535a JA |
698 | struct io_u *__get_io_u(struct thread_data *td) |
699 | { | |
700 | struct io_u *io_u = NULL; | |
701 | ||
755200a3 JA |
702 | if (!list_empty(&td->io_u_requeues)) |
703 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
704 | else if (!queue_full(td)) { | |
10ba535a JA |
705 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
706 | ||
6040dabc | 707 | io_u->buflen = 0; |
10ba535a | 708 | io_u->resid = 0; |
755200a3 | 709 | io_u->file = NULL; |
d7762cf8 | 710 | io_u->end_io = NULL; |
755200a3 JA |
711 | } |
712 | ||
713 | if (io_u) { | |
0c6e7517 JA |
714 | assert(io_u->flags & IO_U_F_FREE); |
715 | io_u->flags &= ~IO_U_F_FREE; | |
716 | ||
755200a3 | 717 | io_u->error = 0; |
10ba535a JA |
718 | list_del(&io_u->list); |
719 | list_add(&io_u->list, &td->io_u_busylist); | |
720 | td->cur_depth++; | |
721 | } | |
722 | ||
723 | return io_u; | |
724 | } | |
725 | ||
726 | /* | |
727 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
728 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
729 | */ | |
3d7c391d | 730 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 731 | { |
3d7c391d | 732 | struct fio_file *f; |
10ba535a JA |
733 | struct io_u *io_u; |
734 | ||
735 | io_u = __get_io_u(td); | |
2ba1c290 JA |
736 | if (!io_u) { |
737 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 738 | return NULL; |
2ba1c290 | 739 | } |
10ba535a | 740 | |
755200a3 JA |
741 | /* |
742 | * from a requeue, io_u already setup | |
743 | */ | |
744 | if (io_u->file) | |
77f392bf | 745 | goto out; |
755200a3 | 746 | |
429f6675 JA |
747 | /* |
748 | * If using an iolog, grab next piece if any available. | |
749 | */ | |
750 | if (td->o.read_iolog_file) { | |
751 | if (read_iolog_get(td, io_u)) | |
752 | goto err_put; | |
2ba1c290 JA |
753 | } else if (set_io_u_file(td, io_u)) { |
754 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 755 | goto err_put; |
2ba1c290 | 756 | } |
5ec10eaa | 757 | |
429f6675 JA |
758 | f = io_u->file; |
759 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 760 | |
87dc1ab1 | 761 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 762 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 763 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 764 | goto err_put; |
2ba1c290 | 765 | } |
10ba535a | 766 | |
36167d82 | 767 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 768 | |
2dc1bbeb | 769 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
770 | populate_verify_io_u(td, io_u); |
771 | } | |
10ba535a | 772 | |
165faf16 JA |
773 | /* |
774 | * Set io data pointers. | |
775 | */ | |
d460eb31 | 776 | io_u->endpos = io_u->offset + io_u->buflen; |
cec6b55d JA |
777 | io_u->xfer_buf = io_u->buf; |
778 | io_u->xfer_buflen = io_u->buflen; | |
6ac7a331 | 779 | out: |
429f6675 JA |
780 | if (!td_io_prep(td, io_u)) { |
781 | fio_gettime(&io_u->start_time, NULL); | |
782 | return io_u; | |
36167d82 | 783 | } |
429f6675 | 784 | err_put: |
2ba1c290 | 785 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
786 | put_io_u(td, io_u); |
787 | return NULL; | |
10ba535a JA |
788 | } |
789 | ||
5451792e JA |
790 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
791 | { | |
792 | const char *msg[] = { "read", "write", "sync" }; | |
793 | ||
794 | log_err("fio: io_u error"); | |
795 | ||
796 | if (io_u->file) | |
797 | log_err(" on file %s", io_u->file->file_name); | |
798 | ||
799 | log_err(": %s\n", strerror(io_u->error)); | |
800 | ||
5ec10eaa JA |
801 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
802 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
803 | |
804 | if (!td->error) | |
805 | td_verror(td, io_u->error, "io_u error"); | |
806 | } | |
807 | ||
97601024 JA |
808 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
809 | struct io_completion_data *icd) | |
10ba535a | 810 | { |
d85f5118 | 811 | unsigned long usec; |
10ba535a | 812 | |
2ba1c290 JA |
813 | dprint_io_u(io_u, "io complete"); |
814 | ||
0c6e7517 JA |
815 | assert(io_u->flags & IO_U_F_FLIGHT); |
816 | io_u->flags &= ~IO_U_F_FLIGHT; | |
817 | ||
87dc1ab1 JA |
818 | if (io_u->ddir == DDIR_SYNC) { |
819 | td->last_was_sync = 1; | |
820 | return; | |
821 | } | |
822 | ||
823 | td->last_was_sync = 0; | |
824 | ||
10ba535a JA |
825 | if (!io_u->error) { |
826 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 827 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 828 | int ret; |
10ba535a JA |
829 | |
830 | td->io_blocks[idx]++; | |
831 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
832 | td->this_io_bytes[idx] += bytes; |
833 | ||
d85f5118 | 834 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 835 | |
d85f5118 | 836 | add_clat_sample(td, idx, usec); |
02bcaa8c | 837 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 838 | io_u_mark_latency(td, usec); |
10ba535a | 839 | |
660a1cb5 | 840 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 841 | td->o.do_verify && |
41128405 | 842 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
843 | log_io_piece(td, io_u); |
844 | ||
845 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 846 | |
d7762cf8 | 847 | if (io_u->end_io) { |
36690c9b | 848 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
849 | if (ret && !icd->error) |
850 | icd->error = ret; | |
851 | } | |
5451792e | 852 | } else { |
10ba535a | 853 | icd->error = io_u->error; |
5451792e JA |
854 | io_u_log_error(td, io_u); |
855 | } | |
10ba535a JA |
856 | } |
857 | ||
d7762cf8 | 858 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 859 | { |
02bcaa8c JA |
860 | fio_gettime(&icd->time, NULL); |
861 | ||
3af6ef39 JA |
862 | icd->nr = nr; |
863 | ||
10ba535a JA |
864 | icd->error = 0; |
865 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
866 | } |
867 | ||
97601024 JA |
868 | static void ios_completed(struct thread_data *td, |
869 | struct io_completion_data *icd) | |
36167d82 JA |
870 | { |
871 | struct io_u *io_u; | |
872 | int i; | |
873 | ||
10ba535a JA |
874 | for (i = 0; i < icd->nr; i++) { |
875 | io_u = td->io_ops->event(td, i); | |
876 | ||
877 | io_completed(td, io_u, icd); | |
878 | put_io_u(td, io_u); | |
879 | } | |
880 | } | |
97601024 | 881 | |
e7e6cfb4 JA |
882 | /* |
883 | * Complete a single io_u for the sync engines. | |
884 | */ | |
d7762cf8 | 885 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
886 | { |
887 | struct io_completion_data icd; | |
888 | ||
d7762cf8 | 889 | init_icd(&icd, 1); |
97601024 JA |
890 | io_completed(td, io_u, &icd); |
891 | put_io_u(td, io_u); | |
892 | ||
893 | if (!icd.error) | |
894 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
895 | ||
37e974a5 | 896 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
897 | return -1; |
898 | } | |
899 | ||
e7e6cfb4 JA |
900 | /* |
901 | * Called to complete min_events number of io for the async engines. | |
902 | */ | |
d7762cf8 | 903 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 904 | { |
97601024 | 905 | struct io_completion_data icd; |
00de55ef | 906 | struct timespec *tvp = NULL; |
97601024 | 907 | int ret; |
4d06a338 | 908 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 909 | |
b271fe62 JA |
910 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
911 | ||
4d06a338 | 912 | if (!min_events) |
00de55ef | 913 | tvp = &ts; |
97601024 | 914 | |
00de55ef | 915 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 916 | if (ret < 0) { |
e1161c32 | 917 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
918 | return ret; |
919 | } else if (!ret) | |
920 | return ret; | |
921 | ||
d7762cf8 | 922 | init_icd(&icd, ret); |
97601024 JA |
923 | ios_completed(td, &icd); |
924 | if (!icd.error) | |
925 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
926 | ||
37e974a5 | 927 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
928 | return -1; |
929 | } | |
7e77dd02 JA |
930 | |
931 | /* | |
932 | * Call when io_u is really queued, to update the submission latency. | |
933 | */ | |
934 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
935 | { | |
936 | unsigned long slat_time; | |
937 | ||
d85f5118 | 938 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
939 | add_slat_sample(td, io_u->ddir, slat_time); |
940 | } | |
433afcb4 | 941 | |
55bc9728 | 942 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
943 | void io_u_set_timeout(struct thread_data *td) |
944 | { | |
945 | assert(td->cur_depth); | |
946 | ||
947 | td->timer.it_interval.tv_sec = 0; | |
948 | td->timer.it_interval.tv_usec = 0; | |
949 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
950 | td->timer.it_value.tv_usec = 0; | |
951 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
952 | fio_gettime(&td->timeout_end, NULL); | |
953 | } | |
5945b9b4 JA |
954 | |
955 | static void io_u_dump(struct io_u *io_u) | |
956 | { | |
957 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
958 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
959 | ||
960 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
5ec10eaa JA |
961 | log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, |
962 | io_u->xfer_buf, io_u->buflen, | |
963 | io_u->xfer_buflen, | |
964 | io_u->offset); | |
5945b9b4 JA |
965 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); |
966 | } | |
55bc9728 JA |
967 | #else |
968 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
969 | { | |
970 | } | |
971 | #endif | |
433afcb4 | 972 | |
55bc9728 | 973 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
974 | static void io_u_timeout_handler(int fio_unused sig) |
975 | { | |
976 | struct thread_data *td, *__td; | |
977 | pid_t pid = getpid(); | |
5945b9b4 JA |
978 | struct list_head *entry; |
979 | struct io_u *io_u; | |
433afcb4 JA |
980 | int i; |
981 | ||
982 | log_err("fio: io_u timeout\n"); | |
983 | ||
984 | /* | |
985 | * TLS would be nice... | |
986 | */ | |
987 | td = NULL; | |
988 | for_each_td(__td, i) { | |
989 | if (__td->pid == pid) { | |
990 | td = __td; | |
991 | break; | |
992 | } | |
993 | } | |
994 | ||
995 | if (!td) { | |
996 | log_err("fio: io_u timeout, can't find job\n"); | |
997 | exit(1); | |
998 | } | |
999 | ||
1000 | if (!td->cur_depth) { | |
1001 | log_err("fio: timeout without pending work?\n"); | |
1002 | return; | |
1003 | } | |
1004 | ||
15506d09 | 1005 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
1006 | |
1007 | list_for_each(entry, &td->io_u_busylist) { | |
1008 | io_u = list_entry(entry, struct io_u, list); | |
1009 | ||
1010 | io_u_dump(io_u); | |
1011 | } | |
1012 | ||
1013 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
1014 | exit(1); |
1015 | } | |
55bc9728 | 1016 | #endif |
433afcb4 JA |
1017 | |
1018 | void io_u_init_timeout(void) | |
1019 | { | |
55bc9728 | 1020 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 1021 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 1022 | #endif |
433afcb4 | 1023 | } |