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" | |
5973cafb | 9 | #include "hash.h" |
10ba535a | 10 | |
5945b9b4 JA |
11 | /* |
12 | * Change this define to play with the timeout handling | |
13 | */ | |
14 | #undef FIO_USE_TIMEOUT | |
15 | ||
97601024 JA |
16 | struct io_completion_data { |
17 | int nr; /* input */ | |
97601024 JA |
18 | |
19 | int error; /* output */ | |
20 | unsigned long bytes_done[2]; /* output */ | |
21 | struct timeval time; /* output */ | |
22 | }; | |
23 | ||
10ba535a JA |
24 | /* |
25 | * The ->file_map[] contains a map of blocks we have or have not done io | |
26 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
27 | */ | |
aec2de20 | 28 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 29 | { |
aec2de20 JA |
30 | unsigned int idx = RAND_MAP_IDX(f, block); |
31 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 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 | */ | |
aec2de20 | 60 | if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) |
df415585 JA |
61 | break; |
62 | ||
aec2de20 JA |
63 | idx = RAND_MAP_IDX(f, block); |
64 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 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 | ||
67778e88 | 87 | return max_blocks; |
2ba1c290 JA |
88 | } |
89 | ||
10ba535a JA |
90 | /* |
91 | * Return the next free block in the map. | |
92 | */ | |
93 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 94 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 95 | { |
84422acd | 96 | unsigned long long min_bs = td->o.rw_min_bs; |
10ba535a JA |
97 | int i; |
98 | ||
c685b5b2 JA |
99 | i = f->last_free_lookup; |
100 | *b = (i * BLOCKS_PER_MAP); | |
84422acd | 101 | while ((*b) * min_bs < f->real_file_size) { |
10ba535a | 102 | if (f->file_map[i] != -1UL) { |
b12ebc65 | 103 | *b += fio_ffz(f->file_map[i]); |
4ba66134 | 104 | if (*b > last_block(td, f, ddir)) |
2ba1c290 | 105 | break; |
c685b5b2 | 106 | f->last_free_lookup = i; |
10ba535a JA |
107 | return 0; |
108 | } | |
109 | ||
110 | *b += BLOCKS_PER_MAP; | |
111 | i++; | |
112 | } | |
113 | ||
2ba1c290 | 114 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
115 | return 1; |
116 | } | |
117 | ||
ec4015da | 118 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 119 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 120 | { |
84422acd | 121 | unsigned long long r; |
ec4015da JA |
122 | int loops = 5; |
123 | ||
124 | do { | |
125 | r = os_random_long(&td->random_state); | |
84422acd | 126 | dprint(FD_RANDOM, "off rand %llu\n", r); |
5ec10eaa JA |
127 | *b = (last_block(td, f, ddir) - 1) |
128 | * (r / ((unsigned long long) RAND_MAX + 1.0)); | |
2ba1c290 | 129 | |
43c63a78 JA |
130 | /* |
131 | * if we are not maintaining a random map, we are done. | |
132 | */ | |
303032ae | 133 | if (!file_randommap(td, f)) |
43c63a78 JA |
134 | return 0; |
135 | ||
136 | /* | |
84422acd | 137 | * calculate map offset and check if it's free |
43c63a78 | 138 | */ |
aec2de20 | 139 | if (random_map_free(f, *b)) |
43c63a78 JA |
140 | return 0; |
141 | ||
84422acd JA |
142 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
143 | *b); | |
43c63a78 | 144 | } while (--loops); |
ec4015da JA |
145 | |
146 | /* | |
43c63a78 JA |
147 | * we get here, if we didn't suceed in looking up a block. generate |
148 | * a random start offset into the filemap, and find the first free | |
149 | * block from there. | |
ec4015da | 150 | */ |
43c63a78 JA |
151 | loops = 10; |
152 | do { | |
153 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); | |
4ba66134 | 154 | if (!get_next_free_block(td, f, ddir, b)) |
43c63a78 | 155 | return 0; |
ec4015da | 156 | |
43c63a78 JA |
157 | r = os_random_long(&td->random_state); |
158 | } while (--loops); | |
159 | ||
160 | /* | |
161 | * that didn't work either, try exhaustive search from the start | |
162 | */ | |
163 | f->last_free_lookup = 0; | |
4ba66134 | 164 | return get_next_free_block(td, f, ddir, b); |
ec4015da JA |
165 | } |
166 | ||
10ba535a JA |
167 | /* |
168 | * For random io, generate a random new block and see if it's used. Repeat | |
169 | * until we find a free one. For sequential io, just return the end of | |
170 | * the last io issued. | |
171 | */ | |
9bf2061e | 172 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 173 | { |
9bf2061e | 174 | struct fio_file *f = io_u->file; |
ec4015da | 175 | unsigned long long b; |
4ba66134 | 176 | enum fio_ddir ddir = io_u->ddir; |
10ba535a | 177 | |
ec4015da JA |
178 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
179 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 180 | |
4ba66134 | 181 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 182 | return 1; |
43063a1c | 183 | } else { |
0c3d768a | 184 | if (f->last_pos >= f->real_file_size) { |
4ba66134 JA |
185 | if (!td_random(td) || |
186 | get_next_rand_offset(td, f, ddir, &b)) | |
0c3d768a | 187 | return 1; |
bcdedd0a | 188 | } else |
4ba66134 | 189 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 190 | } |
10ba535a | 191 | |
009bd847 JA |
192 | io_u->offset = b * td->o.min_bs[ddir]; |
193 | if (io_u->offset >= f->io_size) { | |
194 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
195 | io_u->offset, f->io_size); | |
196 | return 1; | |
197 | } | |
198 | ||
199 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
200 | if (io_u->offset >= f->real_file_size) { |
201 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
202 | io_u->offset, f->real_file_size); | |
10ba535a | 203 | return 1; |
2ba1c290 | 204 | } |
10ba535a JA |
205 | |
206 | return 0; | |
207 | } | |
208 | ||
9bf2061e | 209 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 210 | { |
bca4ed4d | 211 | const int ddir = io_u->ddir; |
10ba535a JA |
212 | unsigned int buflen; |
213 | long r; | |
214 | ||
2dc1bbeb JA |
215 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
216 | buflen = td->o.min_bs[ddir]; | |
10ba535a JA |
217 | else { |
218 | r = os_random_long(&td->bsrange_state); | |
5ec10eaa JA |
219 | if (!td->o.bssplit_nr) { |
220 | buflen = (unsigned int) | |
221 | (1 + (double) (td->o.max_bs[ddir] - 1) | |
222 | * r / (RAND_MAX + 1.0)); | |
223 | } else { | |
564ca972 JA |
224 | long perc = 0; |
225 | unsigned int i; | |
226 | ||
227 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
228 | struct bssplit *bsp = &td->o.bssplit[i]; | |
229 | ||
230 | buflen = bsp->bs; | |
231 | perc += bsp->perc; | |
232 | if (r <= ((LONG_MAX / 100L) * perc)) | |
233 | break; | |
234 | } | |
235 | } | |
5ec10eaa JA |
236 | if (!td->o.bs_unaligned) { |
237 | buflen = (buflen + td->o.min_bs[ddir] - 1) | |
238 | & ~(td->o.min_bs[ddir] - 1); | |
239 | } | |
10ba535a JA |
240 | } |
241 | ||
4ba66134 JA |
242 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
243 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
244 | td->o.min_bs[ddir], ddir); | |
6a5e6884 | 245 | buflen = td->o.min_bs[ddir]; |
4ba66134 | 246 | } |
6a5e6884 | 247 | |
10ba535a JA |
248 | return buflen; |
249 | } | |
250 | ||
afe24a5a JA |
251 | static void set_rwmix_bytes(struct thread_data *td) |
252 | { | |
e4928662 | 253 | unsigned long issues; |
afe24a5a JA |
254 | unsigned int diff; |
255 | ||
256 | /* | |
257 | * we do time or byte based switch. this is needed because | |
258 | * buffered writes may issue a lot quicker than they complete, | |
259 | * whereas reads do not. | |
260 | */ | |
e4928662 | 261 | issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues; |
e47f799f | 262 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
afe24a5a | 263 | |
e4928662 JA |
264 | td->rwmix_issues = td->io_issues[td->rwmix_ddir] |
265 | + (issues * ((100 - diff)) / diff); | |
e47f799f JA |
266 | } |
267 | ||
268 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
269 | { | |
270 | unsigned int v; | |
271 | long r; | |
272 | ||
273 | r = os_random_long(&td->rwmix_state); | |
274 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
275 | if (v < td->o.rwmix[DDIR_READ]) | |
276 | return DDIR_READ; | |
277 | ||
278 | return DDIR_WRITE; | |
afe24a5a JA |
279 | } |
280 | ||
10ba535a JA |
281 | /* |
282 | * Return the data direction for the next io_u. If the job is a | |
283 | * mixed read/write workload, check the rwmix cycle and switch if | |
284 | * necessary. | |
285 | */ | |
1e97cce9 | 286 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
287 | { |
288 | if (td_rw(td)) { | |
10ba535a JA |
289 | /* |
290 | * Check if it's time to seed a new data direction. | |
291 | */ | |
e4928662 | 292 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f | 293 | unsigned long long max_bytes; |
5ec10eaa | 294 | enum fio_ddir ddir; |
e47f799f JA |
295 | |
296 | /* | |
297 | * Put a top limit on how many bytes we do for | |
298 | * one data direction, to avoid overflowing the | |
299 | * ranges too much | |
300 | */ | |
301 | ddir = get_rand_ddir(td); | |
302 | max_bytes = td->this_io_bytes[ddir]; | |
5ec10eaa JA |
303 | if (max_bytes >= |
304 | (td->o.size * td->o.rwmix[ddir] / 100)) { | |
e4928662 | 305 | if (!td->rw_end_set[ddir]) |
38d77cae | 306 | td->rw_end_set[ddir] = 1; |
e4928662 | 307 | |
e47f799f | 308 | ddir ^= 1; |
38d77cae | 309 | } |
e47f799f JA |
310 | |
311 | if (ddir != td->rwmix_ddir) | |
312 | set_rwmix_bytes(td); | |
313 | ||
314 | td->rwmix_ddir = ddir; | |
10ba535a JA |
315 | } |
316 | return td->rwmix_ddir; | |
317 | } else if (td_read(td)) | |
318 | return DDIR_READ; | |
319 | else | |
320 | return DDIR_WRITE; | |
321 | } | |
322 | ||
60f2c658 JA |
323 | static void put_file_log(struct thread_data *td, struct fio_file *f) |
324 | { | |
325 | int ret = put_file(td, f); | |
326 | ||
327 | if (ret) | |
328 | td_verror(td, ret, "file close"); | |
329 | } | |
330 | ||
10ba535a JA |
331 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
332 | { | |
0c6e7517 JA |
333 | assert((io_u->flags & IO_U_F_FREE) == 0); |
334 | io_u->flags |= IO_U_F_FREE; | |
335 | ||
60f2c658 JA |
336 | if (io_u->file) |
337 | put_file_log(td, io_u->file); | |
2dbdab7e | 338 | |
10ba535a JA |
339 | io_u->file = NULL; |
340 | list_del(&io_u->list); | |
341 | list_add(&io_u->list, &td->io_u_freelist); | |
342 | td->cur_depth--; | |
343 | } | |
344 | ||
755200a3 JA |
345 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
346 | { | |
347 | struct io_u *__io_u = *io_u; | |
348 | ||
4d2e0f49 | 349 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
350 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
351 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 352 | |
4d2e0f49 JA |
353 | __io_u->flags &= ~IO_U_F_FLIGHT; |
354 | ||
755200a3 JA |
355 | list_del(&__io_u->list); |
356 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
357 | td->cur_depth--; | |
358 | *io_u = NULL; | |
359 | } | |
360 | ||
9bf2061e | 361 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 362 | { |
b4c5e1ac JA |
363 | if (td->io_ops->flags & FIO_NOIO) |
364 | goto out; | |
365 | ||
87dc1ab1 JA |
366 | /* |
367 | * see if it's time to sync | |
368 | */ | |
2dc1bbeb JA |
369 | if (td->o.fsync_blocks && |
370 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
371 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 372 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 373 | goto out; |
87dc1ab1 JA |
374 | } |
375 | ||
a00735e6 JA |
376 | io_u->ddir = get_rw_ddir(td); |
377 | ||
48f5abd3 JA |
378 | /* |
379 | * See if it's time to switch to a new zone | |
380 | */ | |
381 | if (td->zone_bytes >= td->o.zone_size) { | |
382 | td->zone_bytes = 0; | |
383 | io_u->file->last_pos += td->o.zone_skip; | |
384 | td->io_skip_bytes += td->o.zone_skip; | |
385 | } | |
386 | ||
10ba535a | 387 | /* |
c685b5b2 JA |
388 | * No log, let the seq/rand engine retrieve the next buflen and |
389 | * position. | |
10ba535a | 390 | */ |
2ba1c290 JA |
391 | if (get_next_offset(td, io_u)) { |
392 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 393 | return 1; |
2ba1c290 | 394 | } |
10ba535a | 395 | |
9bf2061e | 396 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
397 | if (!io_u->buflen) { |
398 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 399 | return 1; |
2ba1c290 | 400 | } |
bca4ed4d | 401 | |
2ba1c290 JA |
402 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
403 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
404 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
405 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 406 | return 1; |
2ba1c290 | 407 | } |
6a5e6884 | 408 | |
bca4ed4d JA |
409 | /* |
410 | * mark entry before potentially trimming io_u | |
411 | */ | |
303032ae | 412 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 413 | mark_random_map(td, io_u); |
bca4ed4d JA |
414 | |
415 | /* | |
416 | * If using a write iolog, store this entry. | |
417 | */ | |
c38e9468 | 418 | out: |
2ba1c290 | 419 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 420 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 421 | log_io_u(td, io_u); |
bca4ed4d | 422 | return 0; |
10ba535a JA |
423 | } |
424 | ||
d8005759 | 425 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 JA |
426 | { |
427 | int index = 0; | |
428 | ||
429 | switch (td->cur_depth) { | |
430 | default: | |
a783e61a JA |
431 | index = 6; |
432 | break; | |
71619dc2 | 433 | case 32 ... 63: |
a783e61a JA |
434 | index = 5; |
435 | break; | |
71619dc2 | 436 | case 16 ... 31: |
a783e61a JA |
437 | index = 4; |
438 | break; | |
71619dc2 | 439 | case 8 ... 15: |
a783e61a JA |
440 | index = 3; |
441 | break; | |
71619dc2 | 442 | case 4 ... 7: |
a783e61a JA |
443 | index = 2; |
444 | break; | |
71619dc2 | 445 | case 2 ... 3: |
a783e61a | 446 | index = 1; |
71619dc2 JA |
447 | case 1: |
448 | break; | |
449 | } | |
450 | ||
3bec7ae0 | 451 | td->ts.io_u_map[index] += nr; |
71619dc2 JA |
452 | } |
453 | ||
04a0feae JA |
454 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
455 | { | |
456 | int index = 0; | |
457 | ||
458 | assert(usec < 1000); | |
459 | ||
460 | switch (usec) { | |
461 | case 750 ... 999: | |
462 | index = 9; | |
463 | break; | |
464 | case 500 ... 749: | |
465 | index = 8; | |
466 | break; | |
467 | case 250 ... 499: | |
468 | index = 7; | |
469 | break; | |
470 | case 100 ... 249: | |
471 | index = 6; | |
472 | break; | |
473 | case 50 ... 99: | |
474 | index = 5; | |
475 | break; | |
476 | case 20 ... 49: | |
477 | index = 4; | |
478 | break; | |
479 | case 10 ... 19: | |
480 | index = 3; | |
481 | break; | |
482 | case 4 ... 9: | |
483 | index = 2; | |
484 | break; | |
485 | case 2 ... 3: | |
486 | index = 1; | |
487 | case 0 ... 1: | |
488 | break; | |
489 | } | |
490 | ||
491 | assert(index < FIO_IO_U_LAT_U_NR); | |
492 | td->ts.io_u_lat_u[index]++; | |
493 | } | |
494 | ||
495 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
496 | { |
497 | int index = 0; | |
498 | ||
499 | switch (msec) { | |
500 | default: | |
04a0feae JA |
501 | index = 11; |
502 | break; | |
8abdce66 | 503 | case 1000 ... 1999: |
04a0feae JA |
504 | index = 10; |
505 | break; | |
8abdce66 | 506 | case 750 ... 999: |
04a0feae JA |
507 | index = 9; |
508 | break; | |
8abdce66 | 509 | case 500 ... 749: |
04a0feae JA |
510 | index = 8; |
511 | break; | |
8abdce66 | 512 | case 250 ... 499: |
04a0feae JA |
513 | index = 7; |
514 | break; | |
8abdce66 | 515 | case 100 ... 249: |
04a0feae JA |
516 | index = 6; |
517 | break; | |
8abdce66 | 518 | case 50 ... 99: |
04a0feae JA |
519 | index = 5; |
520 | break; | |
8abdce66 | 521 | case 20 ... 49: |
04a0feae JA |
522 | index = 4; |
523 | break; | |
8abdce66 | 524 | case 10 ... 19: |
04a0feae JA |
525 | index = 3; |
526 | break; | |
8abdce66 | 527 | case 4 ... 9: |
04a0feae JA |
528 | index = 2; |
529 | break; | |
ec118304 | 530 | case 2 ... 3: |
04a0feae | 531 | index = 1; |
ec118304 JA |
532 | case 0 ... 1: |
533 | break; | |
534 | } | |
535 | ||
04a0feae JA |
536 | assert(index < FIO_IO_U_LAT_M_NR); |
537 | td->ts.io_u_lat_m[index]++; | |
538 | } | |
539 | ||
540 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
541 | { | |
542 | if (usec < 1000) | |
543 | io_u_mark_lat_usec(td, usec); | |
544 | else | |
545 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
546 | } |
547 | ||
0aabe160 JA |
548 | /* |
549 | * Get next file to service by choosing one at random | |
550 | */ | |
1c178180 JA |
551 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
552 | int badf) | |
0aabe160 | 553 | { |
0aabe160 | 554 | struct fio_file *f; |
1c178180 | 555 | int fno; |
0aabe160 JA |
556 | |
557 | do { | |
7c83c089 JA |
558 | long r = os_random_long(&td->next_file_state); |
559 | ||
5ec10eaa JA |
560 | fno = (unsigned int) ((double) td->o.nr_files |
561 | * (r / (RAND_MAX + 1.0))); | |
126d65c6 | 562 | f = td->files[fno]; |
059e63c0 JA |
563 | if (f->flags & FIO_FILE_DONE) |
564 | continue; | |
1c178180 | 565 | |
2ba1c290 JA |
566 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
567 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 568 | return f; |
2ba1c290 | 569 | } |
0aabe160 JA |
570 | } while (1); |
571 | } | |
572 | ||
573 | /* | |
574 | * Get next file to service by doing round robin between all available ones | |
575 | */ | |
1c178180 JA |
576 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
577 | int badf) | |
3d7c391d JA |
578 | { |
579 | unsigned int old_next_file = td->next_file; | |
580 | struct fio_file *f; | |
581 | ||
582 | do { | |
126d65c6 | 583 | f = td->files[td->next_file]; |
3d7c391d JA |
584 | |
585 | td->next_file++; | |
2dc1bbeb | 586 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
587 | td->next_file = 0; |
588 | ||
d5ed68ea JA |
589 | if (f->flags & FIO_FILE_DONE) { |
590 | f = NULL; | |
059e63c0 | 591 | continue; |
d5ed68ea | 592 | } |
059e63c0 | 593 | |
1c178180 | 594 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
595 | break; |
596 | ||
597 | f = NULL; | |
598 | } while (td->next_file != old_next_file); | |
599 | ||
2ba1c290 | 600 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
601 | return f; |
602 | } | |
603 | ||
bdb4e2e9 JA |
604 | static struct fio_file *get_next_file(struct thread_data *td) |
605 | { | |
1907dbc6 JA |
606 | struct fio_file *f; |
607 | ||
2dc1bbeb | 608 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 609 | |
2ba1c290 | 610 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
611 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
612 | " nr_files=%d\n", td->nr_open_files, | |
613 | td->nr_done_files, | |
614 | td->o.nr_files); | |
bdb4e2e9 | 615 | return NULL; |
2ba1c290 | 616 | } |
bdb4e2e9 | 617 | |
1907dbc6 | 618 | f = td->file_service_file; |
f11bd94d | 619 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 620 | goto out; |
1907dbc6 | 621 | |
2dc1bbeb | 622 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 623 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 624 | else |
1c178180 | 625 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
626 | |
627 | td->file_service_file = f; | |
628 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
629 | out: |
630 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 631 | return f; |
bdb4e2e9 JA |
632 | } |
633 | ||
1c178180 JA |
634 | static struct fio_file *find_next_new_file(struct thread_data *td) |
635 | { | |
636 | struct fio_file *f; | |
637 | ||
1020a139 JA |
638 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
639 | return NULL; | |
640 | ||
2dc1bbeb | 641 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
642 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
643 | else | |
644 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
645 | ||
646 | return f; | |
647 | } | |
648 | ||
429f6675 JA |
649 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
650 | { | |
651 | struct fio_file *f; | |
652 | ||
653 | do { | |
654 | f = get_next_file(td); | |
655 | if (!f) | |
656 | return 1; | |
657 | ||
658 | set_file: | |
659 | io_u->file = f; | |
660 | get_file(f); | |
661 | ||
662 | if (!fill_io_u(td, io_u)) | |
663 | break; | |
664 | ||
009bd847 JA |
665 | /* |
666 | * optimization to prevent close/open of the same file. This | |
667 | * way we preserve queueing etc. | |
668 | */ | |
669 | if (td->o.nr_files == 1 && td->o.time_based) { | |
60f2c658 | 670 | put_file_log(td, f); |
009bd847 JA |
671 | fio_file_reset(f); |
672 | goto set_file; | |
673 | } | |
674 | ||
429f6675 JA |
675 | /* |
676 | * td_io_close() does a put_file() as well, so no need to | |
677 | * do that here. | |
678 | */ | |
679 | io_u->file = NULL; | |
680 | td_io_close_file(td, f); | |
681 | f->flags |= FIO_FILE_DONE; | |
682 | td->nr_done_files++; | |
683 | ||
684 | /* | |
685 | * probably not the right place to do this, but see | |
686 | * if we need to open a new file | |
687 | */ | |
688 | if (td->nr_open_files < td->o.open_files && | |
689 | td->o.open_files != td->o.nr_files) { | |
690 | f = find_next_new_file(td); | |
691 | ||
692 | if (!f || td_io_open_file(td, f)) | |
693 | return 1; | |
694 | ||
695 | goto set_file; | |
696 | } | |
697 | } while (1); | |
698 | ||
699 | return 0; | |
700 | } | |
701 | ||
702 | ||
10ba535a JA |
703 | struct io_u *__get_io_u(struct thread_data *td) |
704 | { | |
705 | struct io_u *io_u = NULL; | |
706 | ||
755200a3 JA |
707 | if (!list_empty(&td->io_u_requeues)) |
708 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
709 | else if (!queue_full(td)) { | |
10ba535a JA |
710 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
711 | ||
6040dabc | 712 | io_u->buflen = 0; |
10ba535a | 713 | io_u->resid = 0; |
755200a3 | 714 | io_u->file = NULL; |
d7762cf8 | 715 | io_u->end_io = NULL; |
755200a3 JA |
716 | } |
717 | ||
718 | if (io_u) { | |
0c6e7517 JA |
719 | assert(io_u->flags & IO_U_F_FREE); |
720 | io_u->flags &= ~IO_U_F_FREE; | |
721 | ||
755200a3 | 722 | io_u->error = 0; |
10ba535a JA |
723 | list_del(&io_u->list); |
724 | list_add(&io_u->list, &td->io_u_busylist); | |
725 | td->cur_depth++; | |
726 | } | |
727 | ||
728 | return io_u; | |
729 | } | |
730 | ||
731 | /* | |
732 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
733 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
734 | */ | |
3d7c391d | 735 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 736 | { |
3d7c391d | 737 | struct fio_file *f; |
10ba535a JA |
738 | struct io_u *io_u; |
739 | ||
740 | io_u = __get_io_u(td); | |
2ba1c290 JA |
741 | if (!io_u) { |
742 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 743 | return NULL; |
2ba1c290 | 744 | } |
10ba535a | 745 | |
755200a3 JA |
746 | /* |
747 | * from a requeue, io_u already setup | |
748 | */ | |
749 | if (io_u->file) | |
77f392bf | 750 | goto out; |
755200a3 | 751 | |
429f6675 JA |
752 | /* |
753 | * If using an iolog, grab next piece if any available. | |
754 | */ | |
755 | if (td->o.read_iolog_file) { | |
756 | if (read_iolog_get(td, io_u)) | |
757 | goto err_put; | |
2ba1c290 JA |
758 | } else if (set_io_u_file(td, io_u)) { |
759 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 760 | goto err_put; |
2ba1c290 | 761 | } |
5ec10eaa | 762 | |
429f6675 JA |
763 | f = io_u->file; |
764 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 765 | |
87dc1ab1 | 766 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 767 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 768 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 769 | goto err_put; |
2ba1c290 | 770 | } |
10ba535a | 771 | |
36167d82 | 772 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 773 | |
2dc1bbeb | 774 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 JA |
775 | populate_verify_io_u(td, io_u); |
776 | } | |
10ba535a | 777 | |
165faf16 JA |
778 | /* |
779 | * Set io data pointers. | |
780 | */ | |
d460eb31 | 781 | io_u->endpos = io_u->offset + io_u->buflen; |
cec6b55d JA |
782 | io_u->xfer_buf = io_u->buf; |
783 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb JA |
784 | |
785 | if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) | |
786 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
6ac7a331 | 787 | out: |
429f6675 JA |
788 | if (!td_io_prep(td, io_u)) { |
789 | fio_gettime(&io_u->start_time, NULL); | |
790 | return io_u; | |
36167d82 | 791 | } |
429f6675 | 792 | err_put: |
2ba1c290 | 793 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
794 | put_io_u(td, io_u); |
795 | return NULL; | |
10ba535a JA |
796 | } |
797 | ||
5451792e JA |
798 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
799 | { | |
800 | const char *msg[] = { "read", "write", "sync" }; | |
801 | ||
802 | log_err("fio: io_u error"); | |
803 | ||
804 | if (io_u->file) | |
805 | log_err(" on file %s", io_u->file->file_name); | |
806 | ||
807 | log_err(": %s\n", strerror(io_u->error)); | |
808 | ||
5ec10eaa JA |
809 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
810 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
811 | |
812 | if (!td->error) | |
813 | td_verror(td, io_u->error, "io_u error"); | |
814 | } | |
815 | ||
97601024 JA |
816 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
817 | struct io_completion_data *icd) | |
10ba535a | 818 | { |
d85f5118 | 819 | unsigned long usec; |
10ba535a | 820 | |
2ba1c290 JA |
821 | dprint_io_u(io_u, "io complete"); |
822 | ||
0c6e7517 JA |
823 | assert(io_u->flags & IO_U_F_FLIGHT); |
824 | io_u->flags &= ~IO_U_F_FLIGHT; | |
825 | ||
87dc1ab1 JA |
826 | if (io_u->ddir == DDIR_SYNC) { |
827 | td->last_was_sync = 1; | |
828 | return; | |
829 | } | |
830 | ||
831 | td->last_was_sync = 0; | |
832 | ||
10ba535a JA |
833 | if (!io_u->error) { |
834 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 835 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 836 | int ret; |
10ba535a JA |
837 | |
838 | td->io_blocks[idx]++; | |
839 | td->io_bytes[idx] += bytes; | |
10ba535a JA |
840 | td->this_io_bytes[idx] += bytes; |
841 | ||
d85f5118 | 842 | usec = utime_since(&io_u->issue_time, &icd->time); |
10ba535a | 843 | |
d85f5118 | 844 | add_clat_sample(td, idx, usec); |
02bcaa8c | 845 | add_bw_sample(td, idx, &icd->time); |
04a0feae | 846 | io_u_mark_latency(td, usec); |
10ba535a | 847 | |
660a1cb5 | 848 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 849 | td->o.do_verify && |
41128405 | 850 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
851 | log_io_piece(td, io_u); |
852 | ||
853 | icd->bytes_done[idx] += bytes; | |
3af6ef39 | 854 | |
d7762cf8 | 855 | if (io_u->end_io) { |
36690c9b | 856 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
857 | if (ret && !icd->error) |
858 | icd->error = ret; | |
859 | } | |
5451792e | 860 | } else { |
10ba535a | 861 | icd->error = io_u->error; |
5451792e JA |
862 | io_u_log_error(td, io_u); |
863 | } | |
10ba535a JA |
864 | } |
865 | ||
d7762cf8 | 866 | static void init_icd(struct io_completion_data *icd, int nr) |
10ba535a | 867 | { |
02bcaa8c JA |
868 | fio_gettime(&icd->time, NULL); |
869 | ||
3af6ef39 JA |
870 | icd->nr = nr; |
871 | ||
10ba535a JA |
872 | icd->error = 0; |
873 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
874 | } |
875 | ||
97601024 JA |
876 | static void ios_completed(struct thread_data *td, |
877 | struct io_completion_data *icd) | |
36167d82 JA |
878 | { |
879 | struct io_u *io_u; | |
880 | int i; | |
881 | ||
10ba535a JA |
882 | for (i = 0; i < icd->nr; i++) { |
883 | io_u = td->io_ops->event(td, i); | |
884 | ||
885 | io_completed(td, io_u, icd); | |
886 | put_io_u(td, io_u); | |
887 | } | |
888 | } | |
97601024 | 889 | |
e7e6cfb4 JA |
890 | /* |
891 | * Complete a single io_u for the sync engines. | |
892 | */ | |
d7762cf8 | 893 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
894 | { |
895 | struct io_completion_data icd; | |
896 | ||
d7762cf8 | 897 | init_icd(&icd, 1); |
97601024 JA |
898 | io_completed(td, io_u, &icd); |
899 | put_io_u(td, io_u); | |
900 | ||
901 | if (!icd.error) | |
902 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
903 | ||
37e974a5 | 904 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
905 | return -1; |
906 | } | |
907 | ||
e7e6cfb4 JA |
908 | /* |
909 | * Called to complete min_events number of io for the async engines. | |
910 | */ | |
d7762cf8 | 911 | long io_u_queued_complete(struct thread_data *td, int min_events) |
97601024 | 912 | { |
97601024 | 913 | struct io_completion_data icd; |
00de55ef | 914 | struct timespec *tvp = NULL; |
97601024 | 915 | int ret; |
4d06a338 | 916 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 917 | |
b271fe62 JA |
918 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
919 | ||
4d06a338 | 920 | if (!min_events) |
00de55ef | 921 | tvp = &ts; |
97601024 | 922 | |
00de55ef | 923 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 | 924 | if (ret < 0) { |
e1161c32 | 925 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
926 | return ret; |
927 | } else if (!ret) | |
928 | return ret; | |
929 | ||
d7762cf8 | 930 | init_icd(&icd, ret); |
97601024 JA |
931 | ios_completed(td, &icd); |
932 | if (!icd.error) | |
933 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
934 | ||
37e974a5 | 935 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
936 | return -1; |
937 | } | |
7e77dd02 JA |
938 | |
939 | /* | |
940 | * Call when io_u is really queued, to update the submission latency. | |
941 | */ | |
942 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
943 | { | |
944 | unsigned long slat_time; | |
945 | ||
d85f5118 | 946 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
7e77dd02 JA |
947 | add_slat_sample(td, io_u->ddir, slat_time); |
948 | } | |
433afcb4 | 949 | |
5973cafb JA |
950 | /* |
951 | * "randomly" fill the buffer contents | |
952 | */ | |
953 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
954 | unsigned int max_bs) | |
955 | { | |
956 | long *ptr = io_u->buf; | |
957 | ||
958 | if (!td->o.zero_buffers) { | |
959 | while ((void *) ptr - io_u->buf < max_bs) { | |
960 | *ptr = rand() * GOLDEN_RATIO_PRIME; | |
961 | ptr++; | |
962 | } | |
963 | } else | |
964 | memset(ptr, 0, max_bs); | |
965 | } | |
966 | ||
55bc9728 | 967 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
968 | void io_u_set_timeout(struct thread_data *td) |
969 | { | |
970 | assert(td->cur_depth); | |
971 | ||
972 | td->timer.it_interval.tv_sec = 0; | |
973 | td->timer.it_interval.tv_usec = 0; | |
974 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; | |
975 | td->timer.it_value.tv_usec = 0; | |
976 | setitimer(ITIMER_REAL, &td->timer, NULL); | |
977 | fio_gettime(&td->timeout_end, NULL); | |
978 | } | |
5945b9b4 JA |
979 | |
980 | static void io_u_dump(struct io_u *io_u) | |
981 | { | |
982 | unsigned long t_start = mtime_since_now(&io_u->start_time); | |
983 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); | |
984 | ||
985 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); | |
5ec10eaa JA |
986 | log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, |
987 | io_u->xfer_buf, io_u->buflen, | |
988 | io_u->xfer_buflen, | |
989 | io_u->offset); | |
5945b9b4 JA |
990 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); |
991 | } | |
55bc9728 JA |
992 | #else |
993 | void io_u_set_timeout(struct thread_data fio_unused *td) | |
994 | { | |
995 | } | |
996 | #endif | |
433afcb4 | 997 | |
55bc9728 | 998 | #ifdef FIO_USE_TIMEOUT |
433afcb4 JA |
999 | static void io_u_timeout_handler(int fio_unused sig) |
1000 | { | |
1001 | struct thread_data *td, *__td; | |
1002 | pid_t pid = getpid(); | |
5945b9b4 JA |
1003 | struct list_head *entry; |
1004 | struct io_u *io_u; | |
433afcb4 JA |
1005 | int i; |
1006 | ||
1007 | log_err("fio: io_u timeout\n"); | |
1008 | ||
1009 | /* | |
1010 | * TLS would be nice... | |
1011 | */ | |
1012 | td = NULL; | |
1013 | for_each_td(__td, i) { | |
1014 | if (__td->pid == pid) { | |
1015 | td = __td; | |
1016 | break; | |
1017 | } | |
1018 | } | |
1019 | ||
1020 | if (!td) { | |
1021 | log_err("fio: io_u timeout, can't find job\n"); | |
1022 | exit(1); | |
1023 | } | |
1024 | ||
1025 | if (!td->cur_depth) { | |
1026 | log_err("fio: timeout without pending work?\n"); | |
1027 | return; | |
1028 | } | |
1029 | ||
15506d09 | 1030 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
5945b9b4 JA |
1031 | |
1032 | list_for_each(entry, &td->io_u_busylist) { | |
1033 | io_u = list_entry(entry, struct io_u, list); | |
1034 | ||
1035 | io_u_dump(io_u); | |
1036 | } | |
1037 | ||
1038 | td_verror(td, ETIMEDOUT, "io_u timeout"); | |
433afcb4 JA |
1039 | exit(1); |
1040 | } | |
55bc9728 | 1041 | #endif |
433afcb4 JA |
1042 | |
1043 | void io_u_init_timeout(void) | |
1044 | { | |
55bc9728 | 1045 | #ifdef FIO_USE_TIMEOUT |
433afcb4 | 1046 | signal(SIGALRM, io_u_timeout_handler); |
55bc9728 | 1047 | #endif |
433afcb4 | 1048 | } |