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 | 75 | static inline unsigned long long last_block(struct thread_data *td, |
4ba66134 JA |
76 | struct fio_file *f, |
77 | enum fio_ddir ddir) | |
2ba1c290 JA |
78 | { |
79 | unsigned long long max_blocks; | |
80 | ||
4ba66134 | 81 | max_blocks = f->io_size / td->o.min_bs[ddir]; |
2ba1c290 JA |
82 | if (!max_blocks) |
83 | return 0; | |
84 | ||
85 | return max_blocks - 1; | |
86 | } | |
87 | ||
10ba535a JA |
88 | /* |
89 | * Return the next free block in the map. | |
90 | */ | |
91 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 92 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a JA |
93 | { |
94 | int i; | |
95 | ||
c685b5b2 JA |
96 | i = f->last_free_lookup; |
97 | *b = (i * BLOCKS_PER_MAP); | |
2dc1bbeb | 98 | while ((*b) * td->o.rw_min_bs < f->real_file_size) { |
10ba535a | 99 | if (f->file_map[i] != -1UL) { |
b12ebc65 | 100 | *b += fio_ffz(f->file_map[i]); |
4ba66134 | 101 | if (*b > last_block(td, f, ddir)) |
2ba1c290 | 102 | break; |
c685b5b2 | 103 | f->last_free_lookup = i; |
10ba535a JA |
104 | return 0; |
105 | } | |
106 | ||
107 | *b += BLOCKS_PER_MAP; | |
108 | i++; | |
109 | } | |
110 | ||
2ba1c290 | 111 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
112 | return 1; |
113 | } | |
114 | ||
ec4015da | 115 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 116 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 117 | { |
ec4015da JA |
118 | unsigned long long r, rb; |
119 | int loops = 5; | |
120 | ||
121 | do { | |
122 | r = os_random_long(&td->random_state); | |
4ba66134 | 123 | *b = last_block(td, f, ddir); |
2ba1c290 | 124 | |
43c63a78 JA |
125 | /* |
126 | * if we are not maintaining a random map, we are done. | |
127 | */ | |
ec4015da | 128 | if (td->o.norandommap) |
43c63a78 JA |
129 | return 0; |
130 | ||
131 | /* | |
132 | * calculate map offset and chec if it's free | |
133 | */ | |
273f8c91 | 134 | rb = *b; |
43c63a78 JA |
135 | if (random_map_free(td, f, rb)) |
136 | return 0; | |
137 | ||
138 | } while (--loops); | |
ec4015da JA |
139 | |
140 | /* | |
43c63a78 JA |
141 | * we get here, if we didn't suceed in looking up a block. generate |
142 | * a random start offset into the filemap, and find the first free | |
143 | * block from there. | |
ec4015da | 144 | */ |
43c63a78 JA |
145 | loops = 10; |
146 | do { | |
147 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); | |
4ba66134 | 148 | if (!get_next_free_block(td, f, ddir, b)) |
43c63a78 | 149 | return 0; |
ec4015da | 150 | |
43c63a78 JA |
151 | r = os_random_long(&td->random_state); |
152 | } while (--loops); | |
153 | ||
154 | /* | |
155 | * that didn't work either, try exhaustive search from the start | |
156 | */ | |
157 | f->last_free_lookup = 0; | |
4ba66134 | 158 | return get_next_free_block(td, f, ddir, b); |
ec4015da JA |
159 | } |
160 | ||
10ba535a JA |
161 | /* |
162 | * For random io, generate a random new block and see if it's used. Repeat | |
163 | * until we find a free one. For sequential io, just return the end of | |
164 | * the last io issued. | |
165 | */ | |
9bf2061e | 166 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 167 | { |
9bf2061e | 168 | struct fio_file *f = io_u->file; |
ec4015da | 169 | unsigned long long b; |
4ba66134 | 170 | enum fio_ddir ddir = io_u->ddir; |
10ba535a | 171 | |
ec4015da JA |
172 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
173 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 174 | |
4ba66134 | 175 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 176 | return 1; |
43063a1c | 177 | } else { |
0c3d768a | 178 | if (f->last_pos >= f->real_file_size) { |
4ba66134 JA |
179 | if (!td_random(td) || |
180 | get_next_rand_offset(td, f, ddir, &b)) | |
0c3d768a | 181 | return 1; |
bcdedd0a | 182 | } else |
4ba66134 | 183 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 184 | } |
10ba535a | 185 | |
4ba66134 | 186 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
2ba1c290 JA |
187 | if (io_u->offset >= f->real_file_size) { |
188 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
189 | io_u->offset, f->real_file_size); | |
10ba535a | 190 | return 1; |
2ba1c290 | 191 | } |
10ba535a JA |
192 | |
193 | return 0; | |
194 | } | |
195 | ||
9bf2061e | 196 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 197 | { |
bca4ed4d | 198 | const int ddir = io_u->ddir; |
10ba535a JA |
199 | unsigned int buflen; |
200 | long r; | |
201 | ||
2dc1bbeb JA |
202 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
203 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
204 | else { |
205 | r = os_random_long(&td->bsrange_state); | |
564ca972 JA |
206 | if (!td->o.bssplit_nr) |
207 | buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); | |
208 | else { | |
209 | long perc = 0; | |
210 | unsigned int i; | |
211 | ||
212 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
213 | struct bssplit *bsp = &td->o.bssplit[i]; | |
214 | ||
215 | buflen = bsp->bs; | |
216 | perc += bsp->perc; | |
217 | if (r <= ((LONG_MAX / 100L) * perc)) | |
218 | break; | |
219 | } | |
220 | } | |
2dc1bbeb JA |
221 | if (!td->o.bs_unaligned) |
222 | buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); | |
10ba535a JA |
223 | } |
224 | ||
4ba66134 JA |
225 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
226 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
227 | td->o.min_bs[ddir], ddir); | |
6a5e6884 | 228 | buflen = td->o.min_bs[ddir]; |
4ba66134 | 229 | } |
6a5e6884 | 230 | |
10ba535a JA |
231 | return buflen; |
232 | } | |
233 | ||
afe24a5a JA |
234 | static void set_rwmix_bytes(struct thread_data *td) |
235 | { | |
236 | unsigned long long rbytes; | |
237 | unsigned int diff; | |
238 | ||
239 | /* | |
240 | * we do time or byte based switch. this is needed because | |
241 | * buffered writes may issue a lot quicker than they complete, | |
242 | * whereas reads do not. | |
243 | */ | |
244 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 245 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 246 | |
e47f799f JA |
247 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
248 | } | |
249 | ||
250 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
251 | { | |
252 | unsigned int v; | |
253 | long r; | |
254 | ||
255 | r = os_random_long(&td->rwmix_state); | |
256 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
257 | if (v < td->o.rwmix[DDIR_READ]) | |
258 | return DDIR_READ; | |
259 | ||
260 | return DDIR_WRITE; | |
afe24a5a JA |
261 | } |
262 | ||
10ba535a JA |
263 | /* |
264 | * Return the data direction for the next io_u. If the job is a | |
265 | * mixed read/write workload, check the rwmix cycle and switch if | |
266 | * necessary. | |
267 | */ | |
1e97cce9 | 268 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
269 | { |
270 | if (td_rw(td)) { | |
271 | struct timeval now; | |
272 | unsigned long elapsed; | |
afe24a5a | 273 | unsigned int cycle; |
10ba535a | 274 | |
02bcaa8c | 275 | fio_gettime(&now, NULL); |
10ba535a JA |
276 | elapsed = mtime_since_now(&td->rwmix_switch); |
277 | ||
e47f799f JA |
278 | /* |
279 | * if this is the first cycle, make it shorter | |
280 | */ | |
afe24a5a JA |
281 | cycle = td->o.rwmixcycle; |
282 | if (!td->rwmix_bytes) | |
283 | cycle /= 10; | |
284 | ||
10ba535a JA |
285 | /* |
286 | * Check if it's time to seed a new data direction. | |
287 | */ | |
e47f799f | 288 | if (elapsed >= cycle || |
afe24a5a | 289 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
290 | unsigned long long max_bytes; |
291 | enum fio_ddir ddir; | |
292 | ||
293 | /* | |
294 | * Put a top limit on how many bytes we do for | |
295 | * one data direction, to avoid overflowing the | |
296 | * ranges too much | |
297 | */ | |
298 | ddir = get_rand_ddir(td); | |
299 | max_bytes = td->this_io_bytes[ddir]; | |
7bb48f84 | 300 | if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) { |
38d77cae JA |
301 | if (!td->rw_end_set[ddir]) { |
302 | td->rw_end_set[ddir] = 1; | |
303 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
304 | } | |
e47f799f | 305 | ddir ^= 1; |
38d77cae | 306 | } |
e47f799f JA |
307 | |
308 | if (ddir != td->rwmix_ddir) | |
309 | set_rwmix_bytes(td); | |
310 | ||
311 | td->rwmix_ddir = ddir; | |
10ba535a JA |
312 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
313 | } | |
314 | return td->rwmix_ddir; | |
315 | } else if (td_read(td)) | |
316 | return DDIR_READ; | |
317 | else | |
318 | return DDIR_WRITE; | |
319 | } | |
320 | ||
10ba535a JA |
321 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
322 | { | |
0c6e7517 JA |
323 | assert((io_u->flags & IO_U_F_FREE) == 0); |
324 | io_u->flags |= IO_U_F_FREE; | |
325 | ||
2dbdab7e JA |
326 | if (io_u->file) |
327 | put_file(td, io_u->file); | |
328 | ||
10ba535a JA |
329 | io_u->file = NULL; |
330 | list_del(&io_u->list); | |
331 | list_add(&io_u->list, &td->io_u_freelist); | |
332 | td->cur_depth--; | |
333 | } | |
334 | ||
755200a3 JA |
335 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
336 | { | |
337 | struct io_u *__io_u = *io_u; | |
338 | ||
4d2e0f49 | 339 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
340 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
341 | td->io_issues[__io_u->ddir]--; | |
342 | ||
4d2e0f49 JA |
343 | __io_u->flags &= ~IO_U_F_FLIGHT; |
344 | ||
755200a3 JA |
345 | list_del(&__io_u->list); |
346 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
347 | td->cur_depth--; | |
348 | *io_u = NULL; | |
349 | } | |
350 | ||
9bf2061e | 351 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 352 | { |
b4c5e1ac JA |
353 | if (td->io_ops->flags & FIO_NOIO) |
354 | goto out; | |
355 | ||
87dc1ab1 JA |
356 | /* |
357 | * see if it's time to sync | |
358 | */ | |
2dc1bbeb JA |
359 | if (td->o.fsync_blocks && |
360 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
361 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 362 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 363 | goto out; |
87dc1ab1 JA |
364 | } |
365 | ||
a00735e6 JA |
366 | io_u->ddir = get_rw_ddir(td); |
367 | ||
48f5abd3 JA |
368 | /* |
369 | * See if it's time to switch to a new zone | |
370 | */ | |
371 | if (td->zone_bytes >= td->o.zone_size) { | |
372 | td->zone_bytes = 0; | |
373 | io_u->file->last_pos += td->o.zone_skip; | |
374 | td->io_skip_bytes += td->o.zone_skip; | |
375 | } | |
376 | ||
10ba535a | 377 | /* |
c685b5b2 JA |
378 | * No log, let the seq/rand engine retrieve the next buflen and |
379 | * position. | |
10ba535a | 380 | */ |
2ba1c290 JA |
381 | if (get_next_offset(td, io_u)) { |
382 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 383 | return 1; |
2ba1c290 | 384 | } |
10ba535a | 385 | |
9bf2061e | 386 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
387 | if (!io_u->buflen) { |
388 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 389 | return 1; |
2ba1c290 | 390 | } |
bca4ed4d | 391 | |
2ba1c290 JA |
392 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
393 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
394 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
395 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 396 | return 1; |
2ba1c290 | 397 | } |
6a5e6884 | 398 | |
bca4ed4d JA |
399 | /* |
400 | * mark entry before potentially trimming io_u | |
401 | */ | |
8fd15a9a | 402 | if (td_random(td) && !td->o.norandommap) |
9bf2061e | 403 | mark_random_map(td, io_u); |
bca4ed4d JA |
404 | |
405 | /* | |
406 | * If using a write iolog, store this entry. | |
407 | */ | |
c38e9468 | 408 | out: |
2ba1c290 | 409 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 410 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 411 | log_io_u(td, io_u); |
bca4ed4d | 412 | return 0; |
10ba535a JA |
413 | } |
414 | ||
b3605062 | 415 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
416 | { |
417 | int index = 0; | |
418 | ||
b3605062 JA |
419 | if (io_u->ddir == DDIR_SYNC) |
420 | return; | |
421 | ||
71619dc2 JA |
422 | switch (td->cur_depth) { |
423 | default: | |
a783e61a JA |
424 | index = 6; |
425 | break; | |
71619dc2 | 426 | case 32 ... 63: |
a783e61a JA |
427 | index = 5; |
428 | break; | |
71619dc2 | 429 | case 16 ... 31: |
a783e61a JA |
430 | index = 4; |
431 | break; | |
71619dc2 | 432 | case 8 ... 15: |
a783e61a JA |
433 | index = 3; |
434 | break; | |
71619dc2 | 435 | case 4 ... 7: |
a783e61a JA |
436 | index = 2; |
437 | break; | |
71619dc2 | 438 | case 2 ... 3: |
a783e61a | 439 | index = 1; |
71619dc2 JA |
440 | case 1: |
441 | break; | |
442 | } | |
443 | ||
756867bd | 444 | td->ts.io_u_map[index]++; |
b3605062 | 445 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
446 | } |
447 | ||
04a0feae JA |
448 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
449 | { | |
450 | int index = 0; | |
451 | ||
452 | assert(usec < 1000); | |
453 | ||
454 | switch (usec) { | |
455 | case 750 ... 999: | |
456 | index = 9; | |
457 | break; | |
458 | case 500 ... 749: | |
459 | index = 8; | |
460 | break; | |
461 | case 250 ... 499: | |
462 | index = 7; | |
463 | break; | |
464 | case 100 ... 249: | |
465 | index = 6; | |
466 | break; | |
467 | case 50 ... 99: | |
468 | index = 5; | |
469 | break; | |
470 | case 20 ... 49: | |
471 | index = 4; | |
472 | break; | |
473 | case 10 ... 19: | |
474 | index = 3; | |
475 | break; | |
476 | case 4 ... 9: | |
477 | index = 2; | |
478 | break; | |
479 | case 2 ... 3: | |
480 | index = 1; | |
481 | case 0 ... 1: | |
482 | break; | |
483 | } | |
484 | ||
485 | assert(index < FIO_IO_U_LAT_U_NR); | |
486 | td->ts.io_u_lat_u[index]++; | |
487 | } | |
488 | ||
489 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
490 | { |
491 | int index = 0; | |
492 | ||
493 | switch (msec) { | |
494 | default: | |
04a0feae JA |
495 | index = 11; |
496 | break; | |
8abdce66 | 497 | case 1000 ... 1999: |
04a0feae JA |
498 | index = 10; |
499 | break; | |
8abdce66 | 500 | case 750 ... 999: |
04a0feae JA |
501 | index = 9; |
502 | break; | |
8abdce66 | 503 | case 500 ... 749: |
04a0feae JA |
504 | index = 8; |
505 | break; | |
8abdce66 | 506 | case 250 ... 499: |
04a0feae JA |
507 | index = 7; |
508 | break; | |
8abdce66 | 509 | case 100 ... 249: |
04a0feae JA |
510 | index = 6; |
511 | break; | |
8abdce66 | 512 | case 50 ... 99: |
04a0feae JA |
513 | index = 5; |
514 | break; | |
8abdce66 | 515 | case 20 ... 49: |
04a0feae JA |
516 | index = 4; |
517 | break; | |
8abdce66 | 518 | case 10 ... 19: |
04a0feae JA |
519 | index = 3; |
520 | break; | |
8abdce66 | 521 | case 4 ... 9: |
04a0feae JA |
522 | index = 2; |
523 | break; | |
ec118304 | 524 | case 2 ... 3: |
04a0feae | 525 | index = 1; |
ec118304 JA |
526 | case 0 ... 1: |
527 | break; | |
528 | } | |
529 | ||
04a0feae JA |
530 | assert(index < FIO_IO_U_LAT_M_NR); |
531 | td->ts.io_u_lat_m[index]++; | |
532 | } | |
533 | ||
534 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
535 | { | |
536 | if (usec < 1000) | |
537 | io_u_mark_lat_usec(td, usec); | |
538 | else | |
539 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
540 | } |
541 | ||
0aabe160 JA |
542 | /* |
543 | * Get next file to service by choosing one at random | |
544 | */ | |
1c178180 JA |
545 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
546 | int badf) | |
0aabe160 | 547 | { |
0aabe160 | 548 | struct fio_file *f; |
1c178180 | 549 | int fno; |
0aabe160 JA |
550 | |
551 | do { | |
7c83c089 JA |
552 | long r = os_random_long(&td->next_file_state); |
553 | ||
2dc1bbeb | 554 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 | 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 { | |
576 | f = &td->files[td->next_file]; | |
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 JA |
603 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
604 | 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 | 605 | return NULL; |
2ba1c290 | 606 | } |
bdb4e2e9 | 607 | |
1907dbc6 | 608 | f = td->file_service_file; |
f11bd94d | 609 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 610 | goto out; |
1907dbc6 | 611 | |
2dc1bbeb | 612 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 613 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 614 | else |
1c178180 | 615 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
616 | |
617 | td->file_service_file = f; | |
618 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
619 | out: |
620 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 621 | return f; |
bdb4e2e9 JA |
622 | } |
623 | ||
1c178180 JA |
624 | static struct fio_file *find_next_new_file(struct thread_data *td) |
625 | { | |
626 | struct fio_file *f; | |
627 | ||
1020a139 JA |
628 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
629 | return NULL; | |
630 | ||
2dc1bbeb | 631 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
632 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
633 | else | |
634 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
635 | ||
636 | return f; | |
637 | } | |
638 | ||
429f6675 JA |
639 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
640 | { | |
641 | struct fio_file *f; | |
642 | ||
643 | do { | |
644 | f = get_next_file(td); | |
645 | if (!f) | |
646 | return 1; | |
647 | ||
648 | set_file: | |
649 | io_u->file = f; | |
650 | get_file(f); | |
651 | ||
652 | if (!fill_io_u(td, io_u)) | |
653 | break; | |
654 | ||
655 | /* | |
656 | * td_io_close() does a put_file() as well, so no need to | |
657 | * do that here. | |
658 | */ | |
659 | io_u->file = NULL; | |
660 | td_io_close_file(td, f); | |
661 | f->flags |= FIO_FILE_DONE; | |
662 | td->nr_done_files++; | |
663 | ||
664 | /* | |
665 | * probably not the right place to do this, but see | |
666 | * if we need to open a new file | |
667 | */ | |
668 | if (td->nr_open_files < td->o.open_files && | |
669 | td->o.open_files != td->o.nr_files) { | |
670 | f = find_next_new_file(td); | |
671 | ||
672 | if (!f || td_io_open_file(td, f)) | |
673 | return 1; | |
674 | ||
675 | goto set_file; | |
676 | } | |
677 | } while (1); | |
678 | ||
679 | return 0; | |
680 | } | |
681 | ||
682 | ||
10ba535a JA |
683 | struct io_u *__get_io_u(struct thread_data *td) |
684 | { | |
685 | struct io_u *io_u = NULL; | |
686 | ||
755200a3 JA |
687 | if (!list_empty(&td->io_u_requeues)) |
688 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
689 | else if (!queue_full(td)) { | |
10ba535a JA |
690 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
691 | ||
6040dabc | 692 | io_u->buflen = 0; |
10ba535a | 693 | io_u->resid = 0; |
755200a3 | 694 | io_u->file = NULL; |
d7762cf8 | 695 | io_u->end_io = NULL; |
755200a3 JA |
696 | } |
697 | ||
698 | if (io_u) { | |
0c6e7517 JA |
699 | assert(io_u->flags & IO_U_F_FREE); |
700 | io_u->flags &= ~IO_U_F_FREE; | |
701 | ||
755200a3 | 702 | io_u->error = 0; |
10ba535a JA |
703 | list_del(&io_u->list); |
704 | list_add(&io_u->list, &td->io_u_busylist); | |
705 | td->cur_depth++; | |
706 | } | |
707 | ||
708 | return io_u; | |
709 | } | |
710 | ||
711 | /* | |
712 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
713 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
714 | */ | |
3d7c391d | 715 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 716 | { |
3d7c391d | 717 | struct fio_file *f; |
10ba535a JA |
718 | struct io_u *io_u; |
719 | ||
720 | io_u = __get_io_u(td); | |
2ba1c290 JA |
721 | if (!io_u) { |
722 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 723 | return NULL; |
2ba1c290 | 724 | } |
10ba535a | 725 | |
755200a3 JA |
726 | /* |
727 | * from a requeue, io_u already setup | |
728 | */ | |
729 | if (io_u->file) | |
77f392bf | 730 | goto out; |
755200a3 | 731 | |
429f6675 JA |
732 | /* |
733 | * If using an iolog, grab next piece if any available. | |
734 | */ | |
735 | if (td->o.read_iolog_file) { | |
736 | if (read_iolog_get(td, io_u)) | |
737 | goto err_put; | |
2ba1c290 JA |
738 | } else if (set_io_u_file(td, io_u)) { |
739 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 740 | goto err_put; |
2ba1c290 | 741 | } |
429f6675 JA |
742 | |
743 | f = io_u->file; | |
744 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 745 | |
87dc1ab1 | 746 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 747 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 748 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 749 | goto err_put; |
2ba1c290 | 750 | } |
10ba535a | 751 | |
36167d82 | 752 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 753 | |
2dc1bbeb | 754 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
755 | populate_verify_io_u(td, io_u); |
756 | } | |
10ba535a | 757 | |
165faf16 JA |
758 | /* |
759 | * Set io data pointers. | |
760 | */ | |
d460eb31 | 761 | io_u->endpos = io_u->offset + io_u->buflen; |
77f392bf | 762 | out: |
cec6b55d JA |
763 | io_u->xfer_buf = io_u->buf; |
764 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 765 | |
429f6675 JA |
766 | if (!td_io_prep(td, io_u)) { |
767 | fio_gettime(&io_u->start_time, NULL); | |
768 | return io_u; | |
36167d82 | 769 | } |
429f6675 | 770 | err_put: |
2ba1c290 | 771 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
772 | put_io_u(td, io_u); |
773 | return NULL; | |
10ba535a JA |
774 | } |
775 | ||
5451792e JA |
776 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
777 | { | |
778 | const char *msg[] = { "read", "write", "sync" }; | |
779 | ||
780 | log_err("fio: io_u error"); | |
781 | ||
782 | if (io_u->file) | |
783 | log_err(" on file %s", io_u->file->file_name); | |
784 | ||
785 | log_err(": %s\n", strerror(io_u->error)); | |
786 | ||
787 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
788 | ||
789 | if (!td->error) | |
790 | td_verror(td, io_u->error, "io_u error"); | |
791 | } | |
792 | ||
97601024 JA |
793 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
794 | struct io_completion_data *icd) | |
10ba535a | 795 | { |
d85f5118 | 796 | unsigned long usec; |
10ba535a | 797 | |
2ba1c290 JA |
798 | dprint_io_u(io_u, "io complete"); |
799 | ||
0c6e7517 JA |
800 | assert(io_u->flags & IO_U_F_FLIGHT); |
801 | io_u->flags &= ~IO_U_F_FLIGHT; | |
802 | ||
87dc1ab1 JA |
803 | if (io_u->ddir == DDIR_SYNC) { |
804 | td->last_was_sync = 1; | |
805 | return; | |
806 | } | |
807 | ||
808 | td->last_was_sync = 0; | |
809 | ||
10ba535a JA |
810 | if (!io_u->error) { |
811 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 812 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 813 | int ret; |
10ba535a JA |
814 | |
815 | td->io_blocks[idx]++; | |
816 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
817 | td->this_io_bytes[idx] += bytes; |
818 | ||
d460eb31 | 819 | io_u->file->last_completed_pos = io_u->endpos; |
02bcaa8c | 820 | |
d85f5118 | 821 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 822 | |
d85f5118 | 823 | add_clat_sample(td, idx, usec); |
02bcaa8c | 824 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 825 | io_u_mark_latency(td, usec); |
10ba535a | 826 | |
660a1cb5 | 827 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 828 | td->o.do_verify && |
41128405 | 829 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
830 | log_io_piece(td, io_u); |
831 | ||
832 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 833 | |
d7762cf8 | 834 | if (io_u->end_io) { |
36690c9b | 835 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
836 | if (ret && !icd->error) |
837 | icd->error = ret; | |
838 | } | |
5451792e | 839 | } else { |
10ba535a | 840 | icd->error = io_u->error; |
5451792e JA |
841 | io_u_log_error(td, io_u); |
842 | } | |
10ba535a JA |
843 | } |
844 | ||
d7762cf8 | 845 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 846 | { |
02bcaa8c JA |
847 | fio_gettime(&icd->time, NULL); |
848 | ||
3af6ef39 JA |
849 | icd->nr = nr; |
850 | ||
10ba535a JA |
851 | icd->error = 0; |
852 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
853 | } |
854 | ||
97601024 JA |
855 | static void ios_completed(struct thread_data *td, |
856 | struct io_completion_data *icd) | |
36167d82 JA |
857 | { |
858 | struct io_u *io_u; | |
859 | int i; | |
860 | ||
10ba535a JA |
861 | for (i = 0; i < icd->nr; i++) { |
862 | io_u = td->io_ops->event(td, i); | |
863 | ||
864 | io_completed(td, io_u, icd); | |
865 | put_io_u(td, io_u); | |
866 | } | |
867 | } | |
97601024 | 868 | |
e7e6cfb4 JA |
869 | /* |
870 | * Complete a single io_u for the sync engines. | |
871 | */ | |
d7762cf8 | 872 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
873 | { |
874 | struct io_completion_data icd; | |
875 | ||
d7762cf8 | 876 | init_icd(&icd, 1); |
97601024 JA |
877 | io_completed(td, io_u, &icd); |
878 | put_io_u(td, io_u); | |
879 | ||
880 | if (!icd.error) | |
881 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
882 | ||
37e974a5 | 883 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
884 | return -1; |
885 | } | |
886 | ||
e7e6cfb4 JA |
887 | /* |
888 | * Called to complete min_events number of io for the async engines. | |
889 | */ | |
d7762cf8 | 890 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 891 | { |
97601024 | 892 | struct io_completion_data icd; |
00de55ef | 893 | struct timespec *tvp = NULL; |
97601024 | 894 | int ret; |
4d06a338 | 895 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 896 | |
b271fe62 JA |
897 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
898 | ||
4d06a338 | 899 | if (!min_events) |
00de55ef | 900 | tvp = &ts; |
97601024 | 901 | |
00de55ef | 902 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 903 | if (ret < 0) { |
e1161c32 | 904 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
905 | return ret; |
906 | } else if (!ret) | |
907 | return ret; | |
908 | ||
d7762cf8 | 909 | init_icd(&icd, ret); |
97601024 JA |
910 | ios_completed(td, &icd); |
911 | if (!icd.error) | |
912 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
913 | ||
37e974a5 | 914 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
915 | return -1; |
916 | } | |
7e77dd02 JA |
917 | |
918 | /* | |
919 | * Call when io_u is really queued, to update the submission latency. | |
920 | */ | |
921 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
922 | { | |
923 | unsigned long slat_time; | |
924 | ||
d85f5118 | 925 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
926 | add_slat_sample(td, io_u->ddir, slat_time); |
927 | } | |
433afcb4 | 928 | |
55bc9728 | 929 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
930 | void io_u_set_timeout(struct thread_data *td) |
931 | { | |
932 | assert(td->cur_depth); | |
933 | ||
934 | td->timer.it_interval.tv_sec = 0; | |
935 | td->timer.it_interval.tv_usec = 0; | |
936 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
937 | td->timer.it_value.tv_usec = 0; | |
938 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
939 | fio_gettime(&td->timeout_end, NULL); | |
940 | } | |
5945b9b4 JA |
941 | |
942 | static void io_u_dump(struct io_u *io_u) | |
943 | { | |
944 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
945 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
946 | ||
947 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
948 | 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); | |
949 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
950 | } | |
55bc9728 JA |
951 | #else |
952 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
953 | { | |
954 | } | |
955 | #endif | |
433afcb4 | 956 | |
55bc9728 | 957 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
958 | static void io_u_timeout_handler(int fio_unused sig) |
959 | { | |
960 | struct thread_data *td, *__td; | |
961 | pid_t pid = getpid(); | |
5945b9b4 JA |
962 | struct list_head *entry; |
963 | struct io_u *io_u; | |
433afcb4 JA |
964 | int i; |
965 | ||
966 | log_err("fio: io_u timeout\n"); | |
967 | ||
968 | /* | |
969 | * TLS would be nice... | |
970 | */ | |
971 | td = NULL; | |
972 | for_each_td(__td, i) { | |
973 | if (__td->pid == pid) { | |
974 | td = __td; | |
975 | break; | |
976 | } | |
977 | } | |
978 | ||
979 | if (!td) { | |
980 | log_err("fio: io_u timeout, can't find job\n"); | |
981 | exit(1); | |
982 | } | |
983 | ||
984 | if (!td->cur_depth) { | |
985 | log_err("fio: timeout without pending work?\n"); | |
986 | return; | |
987 | } | |
988 | ||
15506d09 | 989 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
990 | |
991 | list_for_each(entry, &td->io_u_busylist) { | |
992 | io_u = list_entry(entry, struct io_u, list); | |
993 | ||
994 | io_u_dump(io_u); | |
995 | } | |
996 | ||
997 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
998 | exit(1); |
999 | } | |
55bc9728 | 1000 | #endif |
433afcb4 JA |
1001 | |
1002 | void io_u_init_timeout(void) | |
1003 | { | |
55bc9728 | 1004 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 1005 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 1006 | #endif |
433afcb4 | 1007 | } |