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 | ||
6977bcd0 JA |
331 | if (io_u->file) { |
332 | int ret = put_file(td, io_u->file); | |
333 | ||
334 | if (ret) | |
335 | td_verror(td, ret, "file close"); | |
336 | } | |
2dbdab7e | 337 | |
10ba535a JA |
338 | io_u->file = NULL; |
339 | list_del(&io_u->list); | |
340 | list_add(&io_u->list, &td->io_u_freelist); | |
341 | td->cur_depth--; | |
342 | } | |
343 | ||
755200a3 JA |
344 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
345 | { | |
346 | struct io_u *__io_u = *io_u; | |
347 | ||
4d2e0f49 | 348 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
349 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
350 | td->io_issues[__io_u->ddir]--; | |
351 | ||
4d2e0f49 JA |
352 | __io_u->flags &= ~IO_U_F_FLIGHT; |
353 | ||
755200a3 JA |
354 | list_del(&__io_u->list); |
355 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
356 | td->cur_depth--; | |
357 | *io_u = NULL; | |
358 | } | |
359 | ||
9bf2061e | 360 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 361 | { |
b4c5e1ac JA |
362 | if (td->io_ops->flags & FIO_NOIO) |
363 | goto out; | |
364 | ||
87dc1ab1 JA |
365 | /* |
366 | * see if it's time to sync | |
367 | */ | |
2dc1bbeb JA |
368 | if (td->o.fsync_blocks && |
369 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
370 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 371 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 372 | goto out; |
87dc1ab1 JA |
373 | } |
374 | ||
a00735e6 JA |
375 | io_u->ddir = get_rw_ddir(td); |
376 | ||
48f5abd3 JA |
377 | /* |
378 | * See if it's time to switch to a new zone | |
379 | */ | |
380 | if (td->zone_bytes >= td->o.zone_size) { | |
381 | td->zone_bytes = 0; | |
382 | io_u->file->last_pos += td->o.zone_skip; | |
383 | td->io_skip_bytes += td->o.zone_skip; | |
384 | } | |
385 | ||
10ba535a | 386 | /* |
c685b5b2 JA |
387 | * No log, let the seq/rand engine retrieve the next buflen and |
388 | * position. | |
10ba535a | 389 | */ |
2ba1c290 JA |
390 | if (get_next_offset(td, io_u)) { |
391 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 392 | return 1; |
2ba1c290 | 393 | } |
10ba535a | 394 | |
9bf2061e | 395 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
396 | if (!io_u->buflen) { |
397 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 398 | return 1; |
2ba1c290 | 399 | } |
bca4ed4d | 400 | |
2ba1c290 JA |
401 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
402 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
403 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
404 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 405 | return 1; |
2ba1c290 | 406 | } |
6a5e6884 | 407 | |
bca4ed4d JA |
408 | /* |
409 | * mark entry before potentially trimming io_u | |
410 | */ | |
8fd15a9a | 411 | if (td_random(td) && !td->o.norandommap) |
9bf2061e | 412 | mark_random_map(td, io_u); |
bca4ed4d JA |
413 | |
414 | /* | |
415 | * If using a write iolog, store this entry. | |
416 | */ | |
c38e9468 | 417 | out: |
2ba1c290 | 418 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 419 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 420 | log_io_u(td, io_u); |
bca4ed4d | 421 | return 0; |
10ba535a JA |
422 | } |
423 | ||
b3605062 | 424 | void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) |
71619dc2 JA |
425 | { |
426 | int index = 0; | |
427 | ||
b3605062 JA |
428 | if (io_u->ddir == DDIR_SYNC) |
429 | return; | |
430 | ||
71619dc2 JA |
431 | switch (td->cur_depth) { |
432 | default: | |
a783e61a JA |
433 | index = 6; |
434 | break; | |
71619dc2 | 435 | case 32 ... 63: |
a783e61a JA |
436 | index = 5; |
437 | break; | |
71619dc2 | 438 | case 16 ... 31: |
a783e61a JA |
439 | index = 4; |
440 | break; | |
71619dc2 | 441 | case 8 ... 15: |
a783e61a JA |
442 | index = 3; |
443 | break; | |
71619dc2 | 444 | case 4 ... 7: |
a783e61a JA |
445 | index = 2; |
446 | break; | |
71619dc2 | 447 | case 2 ... 3: |
a783e61a | 448 | index = 1; |
71619dc2 JA |
449 | case 1: |
450 | break; | |
451 | } | |
452 | ||
756867bd | 453 | td->ts.io_u_map[index]++; |
b3605062 | 454 | td->ts.total_io_u[io_u->ddir]++; |
71619dc2 JA |
455 | } |
456 | ||
04a0feae JA |
457 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
458 | { | |
459 | int index = 0; | |
460 | ||
461 | assert(usec < 1000); | |
462 | ||
463 | switch (usec) { | |
464 | case 750 ... 999: | |
465 | index = 9; | |
466 | break; | |
467 | case 500 ... 749: | |
468 | index = 8; | |
469 | break; | |
470 | case 250 ... 499: | |
471 | index = 7; | |
472 | break; | |
473 | case 100 ... 249: | |
474 | index = 6; | |
475 | break; | |
476 | case 50 ... 99: | |
477 | index = 5; | |
478 | break; | |
479 | case 20 ... 49: | |
480 | index = 4; | |
481 | break; | |
482 | case 10 ... 19: | |
483 | index = 3; | |
484 | break; | |
485 | case 4 ... 9: | |
486 | index = 2; | |
487 | break; | |
488 | case 2 ... 3: | |
489 | index = 1; | |
490 | case 0 ... 1: | |
491 | break; | |
492 | } | |
493 | ||
494 | assert(index < FIO_IO_U_LAT_U_NR); | |
495 | td->ts.io_u_lat_u[index]++; | |
496 | } | |
497 | ||
498 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
499 | { |
500 | int index = 0; | |
501 | ||
502 | switch (msec) { | |
503 | default: | |
04a0feae JA |
504 | index = 11; |
505 | break; | |
8abdce66 | 506 | case 1000 ... 1999: |
04a0feae JA |
507 | index = 10; |
508 | break; | |
8abdce66 | 509 | case 750 ... 999: |
04a0feae JA |
510 | index = 9; |
511 | break; | |
8abdce66 | 512 | case 500 ... 749: |
04a0feae JA |
513 | index = 8; |
514 | break; | |
8abdce66 | 515 | case 250 ... 499: |
04a0feae JA |
516 | index = 7; |
517 | break; | |
8abdce66 | 518 | case 100 ... 249: |
04a0feae JA |
519 | index = 6; |
520 | break; | |
8abdce66 | 521 | case 50 ... 99: |
04a0feae JA |
522 | index = 5; |
523 | break; | |
8abdce66 | 524 | case 20 ... 49: |
04a0feae JA |
525 | index = 4; |
526 | break; | |
8abdce66 | 527 | case 10 ... 19: |
04a0feae JA |
528 | index = 3; |
529 | break; | |
8abdce66 | 530 | case 4 ... 9: |
04a0feae JA |
531 | index = 2; |
532 | break; | |
ec118304 | 533 | case 2 ... 3: |
04a0feae | 534 | index = 1; |
ec118304 JA |
535 | case 0 ... 1: |
536 | break; | |
537 | } | |
538 | ||
04a0feae JA |
539 | assert(index < FIO_IO_U_LAT_M_NR); |
540 | td->ts.io_u_lat_m[index]++; | |
541 | } | |
542 | ||
543 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
544 | { | |
545 | if (usec < 1000) | |
546 | io_u_mark_lat_usec(td, usec); | |
547 | else | |
548 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
549 | } |
550 | ||
0aabe160 JA |
551 | /* |
552 | * Get next file to service by choosing one at random | |
553 | */ | |
1c178180 JA |
554 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
555 | int badf) | |
0aabe160 | 556 | { |
0aabe160 | 557 | struct fio_file *f; |
1c178180 | 558 | int fno; |
0aabe160 JA |
559 | |
560 | do { | |
7c83c089 JA |
561 | long r = os_random_long(&td->next_file_state); |
562 | ||
2dc1bbeb | 563 | fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0))); |
126d65c6 | 564 | f = td->files[fno]; |
059e63c0 JA |
565 | if (f->flags & FIO_FILE_DONE) |
566 | continue; | |
1c178180 | 567 | |
2ba1c290 JA |
568 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
569 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 570 | return f; |
2ba1c290 | 571 | } |
0aabe160 JA |
572 | } while (1); |
573 | } | |
574 | ||
575 | /* | |
576 | * Get next file to service by doing round robin between all available ones | |
577 | */ | |
1c178180 JA |
578 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
579 | int badf) | |
3d7c391d JA |
580 | { |
581 | unsigned int old_next_file = td->next_file; | |
582 | struct fio_file *f; | |
583 | ||
584 | do { | |
126d65c6 | 585 | f = td->files[td->next_file]; |
3d7c391d JA |
586 | |
587 | td->next_file++; | |
2dc1bbeb | 588 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
589 | td->next_file = 0; |
590 | ||
d5ed68ea JA |
591 | if (f->flags & FIO_FILE_DONE) { |
592 | f = NULL; | |
059e63c0 | 593 | continue; |
d5ed68ea | 594 | } |
059e63c0 | 595 | |
1c178180 | 596 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
597 | break; |
598 | ||
599 | f = NULL; | |
600 | } while (td->next_file != old_next_file); | |
601 | ||
2ba1c290 | 602 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
603 | return f; |
604 | } | |
605 | ||
bdb4e2e9 JA |
606 | static struct fio_file *get_next_file(struct thread_data *td) |
607 | { | |
1907dbc6 JA |
608 | struct fio_file *f; |
609 | ||
2dc1bbeb | 610 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 611 | |
2ba1c290 JA |
612 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
613 | 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 | 614 | return NULL; |
2ba1c290 | 615 | } |
bdb4e2e9 | 616 | |
1907dbc6 | 617 | f = td->file_service_file; |
f11bd94d | 618 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 619 | goto out; |
1907dbc6 | 620 | |
2dc1bbeb | 621 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 622 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 623 | else |
1c178180 | 624 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
625 | |
626 | td->file_service_file = f; | |
627 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
628 | out: |
629 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 630 | return f; |
bdb4e2e9 JA |
631 | } |
632 | ||
1c178180 JA |
633 | static struct fio_file *find_next_new_file(struct thread_data *td) |
634 | { | |
635 | struct fio_file *f; | |
636 | ||
1020a139 JA |
637 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
638 | return NULL; | |
639 | ||
2dc1bbeb | 640 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
641 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
642 | else | |
643 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
644 | ||
645 | return f; | |
646 | } | |
647 | ||
429f6675 JA |
648 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
649 | { | |
650 | struct fio_file *f; | |
651 | ||
652 | do { | |
653 | f = get_next_file(td); | |
654 | if (!f) | |
655 | return 1; | |
656 | ||
657 | set_file: | |
658 | io_u->file = f; | |
659 | get_file(f); | |
660 | ||
661 | if (!fill_io_u(td, io_u)) | |
662 | break; | |
663 | ||
664 | /* | |
665 | * td_io_close() does a put_file() as well, so no need to | |
666 | * do that here. | |
667 | */ | |
668 | io_u->file = NULL; | |
669 | td_io_close_file(td, f); | |
670 | f->flags |= FIO_FILE_DONE; | |
671 | td->nr_done_files++; | |
672 | ||
673 | /* | |
674 | * probably not the right place to do this, but see | |
675 | * if we need to open a new file | |
676 | */ | |
677 | if (td->nr_open_files < td->o.open_files && | |
678 | td->o.open_files != td->o.nr_files) { | |
679 | f = find_next_new_file(td); | |
680 | ||
681 | if (!f || td_io_open_file(td, f)) | |
682 | return 1; | |
683 | ||
684 | goto set_file; | |
685 | } | |
686 | } while (1); | |
687 | ||
688 | return 0; | |
689 | } | |
690 | ||
691 | ||
10ba535a JA |
692 | struct io_u *__get_io_u(struct thread_data *td) |
693 | { | |
694 | struct io_u *io_u = NULL; | |
695 | ||
755200a3 JA |
696 | if (!list_empty(&td->io_u_requeues)) |
697 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
698 | else if (!queue_full(td)) { | |
10ba535a JA |
699 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
700 | ||
6040dabc | 701 | io_u->buflen = 0; |
10ba535a | 702 | io_u->resid = 0; |
755200a3 | 703 | io_u->file = NULL; |
d7762cf8 | 704 | io_u->end_io = NULL; |
755200a3 JA |
705 | } |
706 | ||
707 | if (io_u) { | |
0c6e7517 JA |
708 | assert(io_u->flags & IO_U_F_FREE); |
709 | io_u->flags &= ~IO_U_F_FREE; | |
710 | ||
755200a3 | 711 | io_u->error = 0; |
10ba535a JA |
712 | list_del(&io_u->list); |
713 | list_add(&io_u->list, &td->io_u_busylist); | |
714 | td->cur_depth++; | |
715 | } | |
716 | ||
717 | return io_u; | |
718 | } | |
719 | ||
720 | /* | |
721 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
722 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
723 | */ | |
3d7c391d | 724 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 725 | { |
3d7c391d | 726 | struct fio_file *f; |
10ba535a JA |
727 | struct io_u *io_u; |
728 | ||
729 | io_u = __get_io_u(td); | |
2ba1c290 JA |
730 | if (!io_u) { |
731 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 732 | return NULL; |
2ba1c290 | 733 | } |
10ba535a | 734 | |
755200a3 JA |
735 | /* |
736 | * from a requeue, io_u already setup | |
737 | */ | |
738 | if (io_u->file) | |
77f392bf | 739 | goto out; |
755200a3 | 740 | |
429f6675 JA |
741 | /* |
742 | * If using an iolog, grab next piece if any available. | |
743 | */ | |
744 | if (td->o.read_iolog_file) { | |
745 | if (read_iolog_get(td, io_u)) | |
746 | goto err_put; | |
2ba1c290 JA |
747 | } else if (set_io_u_file(td, io_u)) { |
748 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 749 | goto err_put; |
2ba1c290 | 750 | } |
429f6675 JA |
751 | |
752 | f = io_u->file; | |
753 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 754 | |
87dc1ab1 | 755 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 756 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 757 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 758 | goto err_put; |
2ba1c290 | 759 | } |
10ba535a | 760 | |
36167d82 | 761 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 762 | |
2dc1bbeb | 763 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
764 | populate_verify_io_u(td, io_u); |
765 | } | |
10ba535a | 766 | |
165faf16 JA |
767 | /* |
768 | * Set io data pointers. | |
769 | */ | |
d460eb31 | 770 | io_u->endpos = io_u->offset + io_u->buflen; |
cec6b55d JA |
771 | io_u->xfer_buf = io_u->buf; |
772 | io_u->xfer_buflen = io_u->buflen; | |
6ac7a331 | 773 | out: |
429f6675 JA |
774 | if (!td_io_prep(td, io_u)) { |
775 | fio_gettime(&io_u->start_time, NULL); | |
776 | return io_u; | |
36167d82 | 777 | } |
429f6675 | 778 | err_put: |
2ba1c290 | 779 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
780 | put_io_u(td, io_u); |
781 | return NULL; | |
10ba535a JA |
782 | } |
783 | ||
5451792e JA |
784 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
785 | { | |
786 | const char *msg[] = { "read", "write", "sync" }; | |
787 | ||
788 | log_err("fio: io_u error"); | |
789 | ||
790 | if (io_u->file) | |
791 | log_err(" on file %s", io_u->file->file_name); | |
792 | ||
793 | log_err(": %s\n", strerror(io_u->error)); | |
794 | ||
795 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); | |
796 | ||
797 | if (!td->error) | |
798 | td_verror(td, io_u->error, "io_u error"); | |
799 | } | |
800 | ||
97601024 JA |
801 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
802 | struct io_completion_data *icd) | |
10ba535a | 803 | { |
d85f5118 | 804 | unsigned long usec; |
10ba535a | 805 | |
2ba1c290 JA |
806 | dprint_io_u(io_u, "io complete"); |
807 | ||
0c6e7517 JA |
808 | assert(io_u->flags & IO_U_F_FLIGHT); |
809 | io_u->flags &= ~IO_U_F_FLIGHT; | |
810 | ||
87dc1ab1 JA |
811 | if (io_u->ddir == DDIR_SYNC) { |
812 | td->last_was_sync = 1; | |
813 | return; | |
814 | } | |
815 | ||
816 | td->last_was_sync = 0; | |
817 | ||
10ba535a JA |
818 | if (!io_u->error) { |
819 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 820 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 821 | int ret; |
10ba535a JA |
822 | |
823 | td->io_blocks[idx]++; | |
824 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
825 | td->this_io_bytes[idx] += bytes; |
826 | ||
d85f5118 | 827 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 828 | |
d85f5118 | 829 | add_clat_sample(td, idx, usec); |
02bcaa8c | 830 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 831 | io_u_mark_latency(td, usec); |
10ba535a | 832 | |
660a1cb5 | 833 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 834 | td->o.do_verify && |
41128405 | 835 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
836 | log_io_piece(td, io_u); |
837 | ||
838 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 839 | |
d7762cf8 | 840 | if (io_u->end_io) { |
36690c9b | 841 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
842 | if (ret && !icd->error) |
843 | icd->error = ret; | |
844 | } | |
5451792e | 845 | } else { |
10ba535a | 846 | icd->error = io_u->error; |
5451792e JA |
847 | io_u_log_error(td, io_u); |
848 | } | |
10ba535a JA |
849 | } |
850 | ||
d7762cf8 | 851 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 852 | { |
02bcaa8c JA |
853 | fio_gettime(&icd->time, NULL); |
854 | ||
3af6ef39 JA |
855 | icd->nr = nr; |
856 | ||
10ba535a JA |
857 | icd->error = 0; |
858 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
859 | } |
860 | ||
97601024 JA |
861 | static void ios_completed(struct thread_data *td, |
862 | struct io_completion_data *icd) | |
36167d82 JA |
863 | { |
864 | struct io_u *io_u; | |
865 | int i; | |
866 | ||
10ba535a JA |
867 | for (i = 0; i < icd->nr; i++) { |
868 | io_u = td->io_ops->event(td, i); | |
869 | ||
870 | io_completed(td, io_u, icd); | |
871 | put_io_u(td, io_u); | |
872 | } | |
873 | } | |
97601024 | 874 | |
e7e6cfb4 JA |
875 | /* |
876 | * Complete a single io_u for the sync engines. | |
877 | */ | |
d7762cf8 | 878 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
879 | { |
880 | struct io_completion_data icd; | |
881 | ||
d7762cf8 | 882 | init_icd(&icd, 1); |
97601024 JA |
883 | io_completed(td, io_u, &icd); |
884 | put_io_u(td, io_u); | |
885 | ||
886 | if (!icd.error) | |
887 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
888 | ||
37e974a5 | 889 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
890 | return -1; |
891 | } | |
892 | ||
e7e6cfb4 JA |
893 | /* |
894 | * Called to complete min_events number of io for the async engines. | |
895 | */ | |
d7762cf8 | 896 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 897 | { |
97601024 | 898 | struct io_completion_data icd; |
00de55ef | 899 | struct timespec *tvp = NULL; |
97601024 | 900 | int ret; |
4d06a338 | 901 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 902 | |
b271fe62 JA |
903 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
904 | ||
4d06a338 | 905 | if (!min_events) |
00de55ef | 906 | tvp = &ts; |
97601024 | 907 | |
00de55ef | 908 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 909 | if (ret < 0) { |
e1161c32 | 910 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
911 | return ret; |
912 | } else if (!ret) | |
913 | return ret; | |
914 | ||
d7762cf8 | 915 | init_icd(&icd, ret); |
97601024 JA |
916 | ios_completed(td, &icd); |
917 | if (!icd.error) | |
918 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
919 | ||
37e974a5 | 920 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
921 | return -1; |
922 | } | |
7e77dd02 JA |
923 | |
924 | /* | |
925 | * Call when io_u is really queued, to update the submission latency. | |
926 | */ | |
927 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
928 | { | |
929 | unsigned long slat_time; | |
930 | ||
d85f5118 | 931 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
932 | add_slat_sample(td, io_u->ddir, slat_time); |
933 | } | |
433afcb4 | 934 | |
55bc9728 | 935 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
936 | void io_u_set_timeout(struct thread_data *td) |
937 | { | |
938 | assert(td->cur_depth); | |
939 | ||
940 | td->timer.it_interval.tv_sec = 0; | |
941 | td->timer.it_interval.tv_usec = 0; | |
942 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
943 | td->timer.it_value.tv_usec = 0; | |
944 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
945 | fio_gettime(&td->timeout_end, NULL); | |
946 | } | |
5945b9b4 JA |
947 | |
948 | static void io_u_dump(struct io_u *io_u) | |
949 | { | |
950 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
951 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
952 | ||
953 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
954 | 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); | |
955 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); | |
956 | } | |
55bc9728 JA |
957 | #else |
958 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
959 | { | |
960 | } | |
961 | #endif | |
433afcb4 | 962 | |
55bc9728 | 963 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
964 | static void io_u_timeout_handler(int fio_unused sig) |
965 | { | |
966 | struct thread_data *td, *__td; | |
967 | pid_t pid = getpid(); | |
5945b9b4 JA |
968 | struct list_head *entry; |
969 | struct io_u *io_u; | |
433afcb4 JA |
970 | int i; |
971 | ||
972 | log_err("fio: io_u timeout\n"); | |
973 | ||
974 | /* | |
975 | * TLS would be nice... | |
976 | */ | |
977 | td = NULL; | |
978 | for_each_td(__td, i) { | |
979 | if (__td->pid == pid) { | |
980 | td = __td; | |
981 | break; | |
982 | } | |
983 | } | |
984 | ||
985 | if (!td) { | |
986 | log_err("fio: io_u timeout, can't find job\n"); | |
987 | exit(1); | |
988 | } | |
989 | ||
990 | if (!td->cur_depth) { | |
991 | log_err("fio: timeout without pending work?\n"); | |
992 | return; | |
993 | } | |
994 | ||
15506d09 | 995 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
996 | |
997 | list_for_each(entry, &td->io_u_busylist) { | |
998 | io_u = list_entry(entry, struct io_u, list); | |
999 | ||
1000 | io_u_dump(io_u); | |
1001 | } | |
1002 | ||
1003 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
1004 | exit(1); |
1005 | } | |
55bc9728 | 1006 | #endif |
433afcb4 JA |
1007 | |
1008 | void io_u_init_timeout(void) | |
1009 | { | |
55bc9728 | 1010 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 1011 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 1012 | #endif |
433afcb4 | 1013 | } |