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 JA |
126 | dprint(FD_RANDOM, "off rand %llu\n", r); |
127 | *b = (last_block(td, f, ddir) - 1) * (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 | */ | |
ec4015da | 132 | if (td->o.norandommap) |
43c63a78 JA |
133 | return 0; |
134 | ||
135 | /* | |
84422acd | 136 | * calculate map offset and check if it's free |
43c63a78 | 137 | */ |
84422acd | 138 | if (random_map_free(td, 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 | |
4ba66134 | 191 | io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; |
2ba1c290 JA |
192 | if (io_u->offset >= f->real_file_size) { |
193 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
194 | io_u->offset, f->real_file_size); | |
10ba535a | 195 | return 1; |
2ba1c290 | 196 | } |
10ba535a JA |
197 | |
198 | return 0; | |
199 | } | |
200 | ||
9bf2061e | 201 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 202 | { |
bca4ed4d | 203 | const int ddir = io_u->ddir; |
10ba535a JA |
204 | unsigned int buflen; |
205 | long r; | |
206 | ||
2dc1bbeb JA |
207 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
208 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
209 | else { |
210 | r = os_random_long(&td->bsrange_state); | |
564ca972 JA |
211 | if (!td->o.bssplit_nr) |
212 | buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); | |
213 | else { | |
214 | long perc = 0; | |
215 | unsigned int i; | |
216 | ||
217 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
218 | struct bssplit *bsp = &td->o.bssplit[i]; | |
219 | ||
220 | buflen = bsp->bs; | |
221 | perc += bsp->perc; | |
222 | if (r <= ((LONG_MAX / 100L) * perc)) | |
223 | break; | |
224 | } | |
225 | } | |
2dc1bbeb JA |
226 | if (!td->o.bs_unaligned) |
227 | buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); | |
10ba535a JA |
228 | } |
229 | ||
4ba66134 JA |
230 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
231 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
232 | td->o.min_bs[ddir], ddir); | |
6a5e6884 | 233 | buflen = td->o.min_bs[ddir]; |
4ba66134 | 234 | } |
6a5e6884 | 235 | |
10ba535a JA |
236 | return buflen; |
237 | } | |
238 | ||
afe24a5a JA |
239 | static void set_rwmix_bytes(struct thread_data *td) |
240 | { | |
241 | unsigned long long rbytes; | |
242 | unsigned int diff; | |
243 | ||
244 | /* | |
245 | * we do time or byte based switch. this is needed because | |
246 | * buffered writes may issue a lot quicker than they complete, | |
247 | * whereas reads do not. | |
248 | */ | |
249 | rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; | |
e47f799f | 250 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 251 | |
e47f799f JA |
252 | td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); |
253 | } | |
254 | ||
255 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
256 | { | |
257 | unsigned int v; | |
258 | long r; | |
259 | ||
260 | r = os_random_long(&td->rwmix_state); | |
261 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
262 | if (v < td->o.rwmix[DDIR_READ]) | |
263 | return DDIR_READ; | |
264 | ||
265 | return DDIR_WRITE; | |
afe24a5a JA |
266 | } |
267 | ||
10ba535a JA |
268 | /* |
269 | * Return the data direction for the next io_u. If the job is a | |
270 | * mixed read/write workload, check the rwmix cycle and switch if | |
271 | * necessary. | |
272 | */ | |
1e97cce9 | 273 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
274 | { |
275 | if (td_rw(td)) { | |
276 | struct timeval now; | |
277 | unsigned long elapsed; | |
afe24a5a | 278 | unsigned int cycle; |
10ba535a | 279 | |
02bcaa8c | 280 | fio_gettime(&now, NULL); |
10ba535a JA |
281 | elapsed = mtime_since_now(&td->rwmix_switch); |
282 | ||
e47f799f JA |
283 | /* |
284 | * if this is the first cycle, make it shorter | |
285 | */ | |
afe24a5a JA |
286 | cycle = td->o.rwmixcycle; |
287 | if (!td->rwmix_bytes) | |
288 | cycle /= 10; | |
289 | ||
10ba535a JA |
290 | /* |
291 | * Check if it's time to seed a new data direction. | |
292 | */ | |
e47f799f | 293 | if (elapsed >= cycle || |
afe24a5a | 294 | td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { |
e47f799f JA |
295 | unsigned long long max_bytes; |
296 | enum fio_ddir ddir; | |
297 | ||
298 | /* | |
299 | * Put a top limit on how many bytes we do for | |
300 | * one data direction, to avoid overflowing the | |
301 | * ranges too much | |
302 | */ | |
303 | ddir = get_rand_ddir(td); | |
304 | max_bytes = td->this_io_bytes[ddir]; | |
7bb48f84 | 305 | if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) { |
38d77cae JA |
306 | if (!td->rw_end_set[ddir]) { |
307 | td->rw_end_set[ddir] = 1; | |
308 | memcpy(&td->rw_end[ddir], &now, sizeof(now)); | |
309 | } | |
e47f799f | 310 | ddir ^= 1; |
38d77cae | 311 | } |
e47f799f JA |
312 | |
313 | if (ddir != td->rwmix_ddir) | |
314 | set_rwmix_bytes(td); | |
315 | ||
316 | td->rwmix_ddir = ddir; | |
10ba535a JA |
317 | memcpy(&td->rwmix_switch, &now, sizeof(now)); |
318 | } | |
319 | return td->rwmix_ddir; | |
320 | } else if (td_read(td)) | |
321 | return DDIR_READ; | |
322 | else | |
323 | return DDIR_WRITE; | |
324 | } | |
325 | ||
10ba535a JA |
326 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
327 | { | |
0c6e7517 JA |
328 | assert((io_u->flags & IO_U_F_FREE) == 0); |
329 | io_u->flags |= IO_U_F_FREE; | |
330 | ||
2dbdab7e JA |
331 | if (io_u->file) |
332 | put_file(td, io_u->file); | |
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]--; | |
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 | */ | |
8fd15a9a | 407 | if (td_random(td) && !td->o.norandommap) |
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 | ||
b3605062 | 420 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
421 | { |
422 | int index = 0; | |
423 | ||
b3605062 JA |
424 | if (io_u->ddir == DDIR_SYNC) |
425 | return; | |
426 | ||
71619dc2 JA |
427 | switch (td->cur_depth) { |
428 | default: | |
a783e61a JA |
429 | index = 6; |
430 | break; | |
71619dc2 | 431 | case 32 ... 63: |
a783e61a JA |
432 | index = 5; |
433 | break; | |
71619dc2 | 434 | case 16 ... 31: |
a783e61a JA |
435 | index = 4; |
436 | break; | |
71619dc2 | 437 | case 8 ... 15: |
a783e61a JA |
438 | index = 3; |
439 | break; | |
71619dc2 | 440 | case 4 ... 7: |
a783e61a JA |
441 | index = 2; |
442 | break; | |
71619dc2 | 443 | case 2 ... 3: |
a783e61a | 444 | index = 1; |
71619dc2 JA |
445 | case 1: |
446 | break; | |
447 | } | |
448 | ||
756867bd | 449 | td->ts.io_u_map[index]++; |
b3605062 | 450 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
451 | } |
452 | ||
04a0feae JA |
453 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
454 | { | |
455 | int index = 0; | |
456 | ||
457 | assert(usec < 1000); | |
458 | ||
459 | switch (usec) { | |
460 | case 750 ... 999: | |
461 | index = 9; | |
462 | break; | |
463 | case 500 ... 749: | |
464 | index = 8; | |
465 | break; | |
466 | case 250 ... 499: | |
467 | index = 7; | |
468 | break; | |
469 | case 100 ... 249: | |
470 | index = 6; | |
471 | break; | |
472 | case 50 ... 99: | |
473 | index = 5; | |
474 | break; | |
475 | case 20 ... 49: | |
476 | index = 4; | |
477 | break; | |
478 | case 10 ... 19: | |
479 | index = 3; | |
480 | break; | |
481 | case 4 ... 9: | |
482 | index = 2; | |
483 | break; | |
484 | case 2 ... 3: | |
485 | index = 1; | |
486 | case 0 ... 1: | |
487 | break; | |
488 | } | |
489 | ||
490 | assert(index < FIO_IO_U_LAT_U_NR); | |
491 | td->ts.io_u_lat_u[index]++; | |
492 | } | |
493 | ||
494 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
495 | { |
496 | int index = 0; | |
497 | ||
498 | switch (msec) { | |
499 | default: | |
04a0feae JA |
500 | index = 11; |
501 | break; | |
8abdce66 | 502 | case 1000 ... 1999: |
04a0feae JA |
503 | index = 10; |
504 | break; | |
8abdce66 | 505 | case 750 ... 999: |
04a0feae JA |
506 | index = 9; |
507 | break; | |
8abdce66 | 508 | case 500 ... 749: |
04a0feae JA |
509 | index = 8; |
510 | break; | |
8abdce66 | 511 | case 250 ... 499: |
04a0feae JA |
512 | index = 7; |
513 | break; | |
8abdce66 | 514 | case 100 ... 249: |
04a0feae JA |
515 | index = 6; |
516 | break; | |
8abdce66 | 517 | case 50 ... 99: |
04a0feae JA |
518 | index = 5; |
519 | break; | |
8abdce66 | 520 | case 20 ... 49: |
04a0feae JA |
521 | index = 4; |
522 | break; | |
8abdce66 | 523 | case 10 ... 19: |
04a0feae JA |
524 | index = 3; |
525 | break; | |
8abdce66 | 526 | case 4 ... 9: |
04a0feae JA |
527 | index = 2; |
528 | break; | |
ec118304 | 529 | case 2 ... 3: |
04a0feae | 530 | index = 1; |
ec118304 JA |
531 | case 0 ... 1: |
532 | break; | |
533 | } | |
534 | ||
04a0feae JA |
535 | assert(index < FIO_IO_U_LAT_M_NR); |
536 | td->ts.io_u_lat_m[index]++; | |
537 | } | |
538 | ||
539 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
540 | { | |
541 | if (usec < 1000) | |
542 | io_u_mark_lat_usec(td, usec); | |
543 | else | |
544 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
545 | } |
546 | ||
0aabe160 JA |
547 | /* |
548 | * Get next file to service by choosing one at random | |
549 | */ | |
1c178180 JA |
550 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
551 | int badf) | |
0aabe160 | 552 | { |
0aabe160 | 553 | struct fio_file *f; |
1c178180 | 554 | int fno; |
0aabe160 JA |
555 | |
556 | do { | |
7c83c089 JA |
557 | long r = os_random_long(&td->next_file_state); |
558 | ||
2dc1bbeb | 559 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
1c178180 | 560 | f = &td->files[fno]; |
059e63c0 JA |
561 | if (f->flags & FIO_FILE_DONE) |
562 | continue; | |
1c178180 | 563 | |
2ba1c290 JA |
564 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
565 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 566 | return f; |
2ba1c290 | 567 | } |
0aabe160 JA |
568 | } while (1); |
569 | } | |
570 | ||
571 | /* | |
572 | * Get next file to service by doing round robin between all available ones | |
573 | */ | |
1c178180 JA |
574 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
575 | int badf) | |
3d7c391d JA |
576 | { |
577 | unsigned int old_next_file = td->next_file; | |
578 | struct fio_file *f; | |
579 | ||
580 | do { | |
581 | f = &td->files[td->next_file]; | |
582 | ||
583 | td->next_file++; | |
2dc1bbeb | 584 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
585 | td->next_file = 0; |
586 | ||
d5ed68ea JA |
587 | if (f->flags & FIO_FILE_DONE) { |
588 | f = NULL; | |
059e63c0 | 589 | continue; |
d5ed68ea | 590 | } |
059e63c0 | 591 | |
1c178180 | 592 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
593 | break; |
594 | ||
595 | f = NULL; | |
596 | } while (td->next_file != old_next_file); | |
597 | ||
2ba1c290 | 598 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
599 | return f; |
600 | } | |
601 | ||
bdb4e2e9 JA |
602 | static struct fio_file *get_next_file(struct thread_data *td) |
603 | { | |
1907dbc6 JA |
604 | struct fio_file *f; |
605 | ||
2dc1bbeb | 606 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 607 | |
2ba1c290 JA |
608 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
609 | 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 | 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 | ||
660 | /* | |
661 | * td_io_close() does a put_file() as well, so no need to | |
662 | * do that here. | |
663 | */ | |
664 | io_u->file = NULL; | |
665 | td_io_close_file(td, f); | |
666 | f->flags |= FIO_FILE_DONE; | |
667 | td->nr_done_files++; | |
668 | ||
669 | /* | |
670 | * probably not the right place to do this, but see | |
671 | * if we need to open a new file | |
672 | */ | |
673 | if (td->nr_open_files < td->o.open_files && | |
674 | td->o.open_files != td->o.nr_files) { | |
675 | f = find_next_new_file(td); | |
676 | ||
677 | if (!f || td_io_open_file(td, f)) | |
678 | return 1; | |
679 | ||
680 | goto set_file; | |
681 | } | |
682 | } while (1); | |
683 | ||
684 | return 0; | |
685 | } | |
686 | ||
687 | ||
10ba535a JA |
688 | struct io_u *__get_io_u(struct thread_data *td) |
689 | { | |
690 | struct io_u *io_u = NULL; | |
691 | ||
755200a3 JA |
692 | if (!list_empty(&td->io_u_requeues)) |
693 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
694 | else if (!queue_full(td)) { | |
10ba535a JA |
695 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
696 | ||
6040dabc | 697 | io_u->buflen = 0; |
10ba535a | 698 | io_u->resid = 0; |
755200a3 | 699 | io_u->file = NULL; |
d7762cf8 | 700 | io_u->end_io = NULL; |
755200a3 JA |
701 | } |
702 | ||
703 | if (io_u) { | |
0c6e7517 JA |
704 | assert(io_u->flags & IO_U_F_FREE); |
705 | io_u->flags &= ~IO_U_F_FREE; | |
706 | ||
755200a3 | 707 | io_u->error = 0; |
10ba535a JA |
708 | list_del(&io_u->list); |
709 | list_add(&io_u->list, &td->io_u_busylist); | |
710 | td->cur_depth++; | |
711 | } | |
712 | ||
713 | return io_u; | |
714 | } | |
715 | ||
716 | /* | |
717 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
718 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
719 | */ | |
3d7c391d | 720 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 721 | { |
3d7c391d | 722 | struct fio_file *f; |
10ba535a JA |
723 | struct io_u *io_u; |
724 | ||
725 | io_u = __get_io_u(td); | |
2ba1c290 JA |
726 | if (!io_u) { |
727 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 728 | return NULL; |
2ba1c290 | 729 | } |
10ba535a | 730 | |
755200a3 JA |
731 | /* |
732 | * from a requeue, io_u already setup | |
733 | */ | |
734 | if (io_u->file) | |
77f392bf | 735 | goto out; |
755200a3 | 736 | |
429f6675 JA |
737 | /* |
738 | * If using an iolog, grab next piece if any available. | |
739 | */ | |
740 | if (td->o.read_iolog_file) { | |
741 | if (read_iolog_get(td, io_u)) | |
742 | goto err_put; | |
2ba1c290 JA |
743 | } else if (set_io_u_file(td, io_u)) { |
744 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 745 | goto err_put; |
2ba1c290 | 746 | } |
429f6675 JA |
747 | |
748 | f = io_u->file; | |
749 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 750 | |
87dc1ab1 | 751 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 752 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 753 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 754 | goto err_put; |
2ba1c290 | 755 | } |
10ba535a | 756 | |
36167d82 | 757 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 758 | |
2dc1bbeb | 759 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
760 | populate_verify_io_u(td, io_u); |
761 | } | |
10ba535a | 762 | |
165faf16 JA |
763 | /* |
764 | * Set io data pointers. | |
765 | */ | |
d460eb31 | 766 | io_u->endpos = io_u->offset + io_u->buflen; |
77f392bf | 767 | out: |
cec6b55d JA |
768 | io_u->xfer_buf = io_u->buf; |
769 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 770 | |
429f6675 JA |
771 | if (!td_io_prep(td, io_u)) { |
772 | fio_gettime(&io_u->start_time, NULL); | |
773 | return io_u; | |
36167d82 | 774 | } |
429f6675 | 775 | err_put: |
2ba1c290 | 776 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
777 | put_io_u(td, io_u); |
778 | return NULL; | |
10ba535a JA |
779 | } |
780 | ||
5451792e JA |
781 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
782 | { | |
783 | const char *msg[] = { "read", "write", "sync" }; | |
784 | ||
785 | log_err("fio: io_u error"); | |
786 | ||
787 | if (io_u->file) | |
788 | log_err(" on file %s", io_u->file->file_name); | |
789 | ||
790 | log_err(": %s\n", strerror(io_u->error)); | |
791 | ||
792 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
793 | ||
794 | if (!td->error) | |
795 | td_verror(td, io_u->error, "io_u error"); | |
796 | } | |
797 | ||
97601024 JA |
798 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
799 | struct io_completion_data *icd) | |
10ba535a | 800 | { |
d85f5118 | 801 | unsigned long usec; |
10ba535a | 802 | |
2ba1c290 JA |
803 | dprint_io_u(io_u, "io complete"); |
804 | ||
0c6e7517 JA |
805 | assert(io_u->flags & IO_U_F_FLIGHT); |
806 | io_u->flags &= ~IO_U_F_FLIGHT; | |
807 | ||
87dc1ab1 JA |
808 | if (io_u->ddir == DDIR_SYNC) { |
809 | td->last_was_sync = 1; | |
810 | return; | |
811 | } | |
812 | ||
813 | td->last_was_sync = 0; | |
814 | ||
10ba535a JA |
815 | if (!io_u->error) { |
816 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 817 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 818 | int ret; |
10ba535a JA |
819 | |
820 | td->io_blocks[idx]++; | |
821 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
822 | td->this_io_bytes[idx] += bytes; |
823 | ||
d460eb31 | 824 | io_u->file->last_completed_pos = io_u->endpos; |
02bcaa8c | 825 | |
d85f5118 | 826 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 827 | |
d85f5118 | 828 | add_clat_sample(td, idx, usec); |
02bcaa8c | 829 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 830 | io_u_mark_latency(td, usec); |
10ba535a | 831 | |
660a1cb5 | 832 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 833 | td->o.do_verify && |
41128405 | 834 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
835 | log_io_piece(td, io_u); |
836 | ||
837 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 838 | |
d7762cf8 | 839 | if (io_u->end_io) { |
36690c9b | 840 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
841 | if (ret && !icd->error) |
842 | icd->error = ret; | |
843 | } | |
5451792e | 844 | } else { |
10ba535a | 845 | icd->error = io_u->error; |
5451792e JA |
846 | io_u_log_error(td, io_u); |
847 | } | |
10ba535a JA |
848 | } |
849 | ||
d7762cf8 | 850 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 851 | { |
02bcaa8c JA |
852 | fio_gettime(&icd->time, NULL); |
853 | ||
3af6ef39 JA |
854 | icd->nr = nr; |
855 | ||
10ba535a JA |
856 | icd->error = 0; |
857 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
858 | } |
859 | ||
97601024 JA |
860 | static void ios_completed(struct thread_data *td, |
861 | struct io_completion_data *icd) | |
36167d82 JA |
862 | { |
863 | struct io_u *io_u; | |
864 | int i; | |
865 | ||
10ba535a JA |
866 | for (i = 0; i < icd->nr; i++) { |
867 | io_u = td->io_ops->event(td, i); | |
868 | ||
869 | io_completed(td, io_u, icd); | |
870 | put_io_u(td, io_u); | |
871 | } | |
872 | } | |
97601024 | 873 | |
e7e6cfb4 JA |
874 | /* |
875 | * Complete a single io_u for the sync engines. | |
876 | */ | |
d7762cf8 | 877 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
878 | { |
879 | struct io_completion_data icd; | |
880 | ||
d7762cf8 | 881 | init_icd(&icd, 1); |
97601024 JA |
882 | io_completed(td, io_u, &icd); |
883 | put_io_u(td, io_u); | |
884 | ||
885 | if (!icd.error) | |
886 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
887 | ||
37e974a5 | 888 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
889 | return -1; |
890 | } | |
891 | ||
e7e6cfb4 JA |
892 | /* |
893 | * Called to complete min_events number of io for the async engines. | |
894 | */ | |
d7762cf8 | 895 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 896 | { |
97601024 | 897 | struct io_completion_data icd; |
00de55ef | 898 | struct timespec *tvp = NULL; |
97601024 | 899 | int ret; |
4d06a338 | 900 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 901 | |
b271fe62 JA |
902 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
903 | ||
4d06a338 | 904 | if (!min_events) |
00de55ef | 905 | tvp = &ts; |
97601024 | 906 | |
00de55ef | 907 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 908 | if (ret < 0) { |
e1161c32 | 909 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
910 | return ret; |
911 | } else if (!ret) | |
912 | return ret; | |
913 | ||
d7762cf8 | 914 | init_icd(&icd, ret); |
97601024 JA |
915 | ios_completed(td, &icd); |
916 | if (!icd.error) | |
917 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
918 | ||
37e974a5 | 919 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
920 | return -1; |
921 | } | |
7e77dd02 JA |
922 | |
923 | /* | |
924 | * Call when io_u is really queued, to update the submission latency. | |
925 | */ | |
926 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
927 | { | |
928 | unsigned long slat_time; | |
929 | ||
d85f5118 | 930 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
931 | add_slat_sample(td, io_u->ddir, slat_time); |
932 | } | |
433afcb4 | 933 | |
55bc9728 | 934 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
935 | void io_u_set_timeout(struct thread_data *td) |
936 | { | |
937 | assert(td->cur_depth); | |
938 | ||
939 | td->timer.it_interval.tv_sec = 0; | |
940 | td->timer.it_interval.tv_usec = 0; | |
941 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
942 | td->timer.it_value.tv_usec = 0; | |
943 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
944 | fio_gettime(&td->timeout_end, NULL); | |
945 | } | |
5945b9b4 JA |
946 | |
947 | static void io_u_dump(struct io_u *io_u) | |
948 | { | |
949 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
950 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
951 | ||
952 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
953 | 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); | |
954 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
955 | } | |
55bc9728 JA |
956 | #else |
957 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
958 | { | |
959 | } | |
960 | #endif | |
433afcb4 | 961 | |
55bc9728 | 962 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
963 | static void io_u_timeout_handler(int fio_unused sig) |
964 | { | |
965 | struct thread_data *td, *__td; | |
966 | pid_t pid = getpid(); | |
5945b9b4 JA |
967 | struct list_head *entry; |
968 | struct io_u *io_u; | |
433afcb4 JA |
969 | int i; |
970 | ||
971 | log_err("fio: io_u timeout\n"); | |
972 | ||
973 | /* | |
974 | * TLS would be nice... | |
975 | */ | |
976 | td = NULL; | |
977 | for_each_td(__td, i) { | |
978 | if (__td->pid == pid) { | |
979 | td = __td; | |
980 | break; | |
981 | } | |
982 | } | |
983 | ||
984 | if (!td) { | |
985 | log_err("fio: io_u timeout, can't find job\n"); | |
986 | exit(1); | |
987 | } | |
988 | ||
989 | if (!td->cur_depth) { | |
990 | log_err("fio: timeout without pending work?\n"); | |
991 | return; | |
992 | } | |
993 | ||
15506d09 | 994 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
995 | |
996 | list_for_each(entry, &td->io_u_busylist) { | |
997 | io_u = list_entry(entry, struct io_u, list); | |
998 | ||
999 | io_u_dump(io_u); | |
1000 | } | |
1001 | ||
1002 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
1003 | exit(1); |
1004 | } | |
55bc9728 | 1005 | #endif |
433afcb4 JA |
1006 | |
1007 | void io_u_init_timeout(void) | |
1008 | { | |
55bc9728 | 1009 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 1010 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 1011 | #endif |
433afcb4 | 1012 | } |