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 | |
97601024 JA |
11 | struct io_completion_data { |
12 | int nr; /* input */ | |
97601024 JA |
13 | |
14 | int error; /* output */ | |
15 | unsigned long bytes_done[2]; /* output */ | |
16 | struct timeval time; /* output */ | |
17 | }; | |
18 | ||
10ba535a JA |
19 | /* |
20 | * The ->file_map[] contains a map of blocks we have or have not done io | |
21 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
22 | */ | |
aec2de20 | 23 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 24 | { |
aec2de20 JA |
25 | unsigned int idx = RAND_MAP_IDX(f, block); |
26 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 27 | |
84422acd JA |
28 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
29 | ||
de605666 | 30 | return (f->file_map[idx] & (1 << bit)) == 0; |
10ba535a JA |
31 | } |
32 | ||
df415585 JA |
33 | /* |
34 | * Mark a given offset as used in the map. | |
35 | */ | |
9bf2061e | 36 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 37 | { |
2dc1bbeb | 38 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 39 | struct fio_file *f = io_u->file; |
a00735e6 | 40 | unsigned long long block; |
3e3357b1 | 41 | unsigned int blocks, nr_blocks; |
df415585 | 42 | |
b9c5b644 | 43 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
c685b5b2 | 44 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
3e3357b1 | 45 | blocks = 0; |
c685b5b2 | 46 | |
3e3357b1 JA |
47 | while (nr_blocks) { |
48 | unsigned int this_blocks, mask; | |
df415585 JA |
49 | unsigned int idx, bit; |
50 | ||
1e3d53ac JA |
51 | /* |
52 | * If we have a mixed random workload, we may | |
53 | * encounter blocks we already did IO to. | |
54 | */ | |
8b113b4f JA |
55 | if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) { |
56 | if (!blocks) | |
57 | blocks = 1; | |
df415585 | 58 | break; |
8b113b4f | 59 | } |
df415585 | 60 | |
aec2de20 JA |
61 | idx = RAND_MAP_IDX(f, block); |
62 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 63 | |
0032bf9f | 64 | fio_assert(td, idx < f->num_maps); |
df415585 | 65 | |
3e3357b1 JA |
66 | this_blocks = nr_blocks; |
67 | if (this_blocks + bit > BLOCKS_PER_MAP) | |
68 | this_blocks = BLOCKS_PER_MAP - bit; | |
69 | ||
70 | if (this_blocks == BLOCKS_PER_MAP) | |
71 | mask = -1U; | |
72 | else | |
73 | mask = ((1U << this_blocks) - 1) << bit; | |
74 | ||
3e3357b1 JA |
75 | f->file_map[idx] |= mask; |
76 | nr_blocks -= this_blocks; | |
77 | blocks += this_blocks; | |
c9dd34b2 | 78 | block += this_blocks; |
df415585 JA |
79 | } |
80 | ||
a00735e6 JA |
81 | if ((blocks * min_bs) < io_u->buflen) |
82 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
83 | } |
84 | ||
3e3357b1 JA |
85 | static unsigned long long last_block(struct thread_data *td, struct fio_file *f, |
86 | enum fio_ddir ddir) | |
2ba1c290 JA |
87 | { |
88 | unsigned long long max_blocks; | |
d9dd70f7 | 89 | unsigned long long max_size; |
2ba1c290 | 90 | |
d9dd70f7 JA |
91 | /* |
92 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
93 | */ | |
94 | max_size = f->io_size; | |
95 | if (max_size > f->real_file_size) | |
96 | max_size = f->real_file_size; | |
97 | ||
98 | max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir]; | |
2ba1c290 JA |
99 | if (!max_blocks) |
100 | return 0; | |
101 | ||
67778e88 | 102 | return max_blocks; |
2ba1c290 JA |
103 | } |
104 | ||
10ba535a JA |
105 | /* |
106 | * Return the next free block in the map. | |
107 | */ | |
108 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 109 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 110 | { |
84422acd | 111 | unsigned long long min_bs = td->o.rw_min_bs; |
10ba535a JA |
112 | int i; |
113 | ||
c685b5b2 JA |
114 | i = f->last_free_lookup; |
115 | *b = (i * BLOCKS_PER_MAP); | |
84422acd | 116 | while ((*b) * min_bs < f->real_file_size) { |
de605666 | 117 | if (f->file_map[i] != (unsigned int) -1) { |
697a606c | 118 | *b += ffz(f->file_map[i]); |
4ba66134 | 119 | if (*b > last_block(td, f, ddir)) |
2ba1c290 | 120 | break; |
c685b5b2 | 121 | f->last_free_lookup = i; |
10ba535a JA |
122 | return 0; |
123 | } | |
124 | ||
125 | *b += BLOCKS_PER_MAP; | |
126 | i++; | |
127 | } | |
128 | ||
2ba1c290 | 129 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
130 | return 1; |
131 | } | |
132 | ||
ec4015da | 133 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 134 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 135 | { |
84422acd | 136 | unsigned long long r; |
ec4015da JA |
137 | int loops = 5; |
138 | ||
139 | do { | |
140 | r = os_random_long(&td->random_state); | |
84422acd | 141 | dprint(FD_RANDOM, "off rand %llu\n", r); |
5ec10eaa | 142 | *b = (last_block(td, f, ddir) - 1) |
dc873b6f | 143 | * (r / ((unsigned long long) OS_RAND_MAX + 1.0)); |
2ba1c290 | 144 | |
43c63a78 JA |
145 | /* |
146 | * if we are not maintaining a random map, we are done. | |
147 | */ | |
303032ae | 148 | if (!file_randommap(td, f)) |
43c63a78 JA |
149 | return 0; |
150 | ||
151 | /* | |
84422acd | 152 | * calculate map offset and check if it's free |
43c63a78 | 153 | */ |
aec2de20 | 154 | if (random_map_free(f, *b)) |
43c63a78 JA |
155 | return 0; |
156 | ||
84422acd JA |
157 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
158 | *b); | |
43c63a78 | 159 | } while (--loops); |
ec4015da JA |
160 | |
161 | /* | |
43c63a78 JA |
162 | * we get here, if we didn't suceed in looking up a block. generate |
163 | * a random start offset into the filemap, and find the first free | |
164 | * block from there. | |
ec4015da | 165 | */ |
43c63a78 JA |
166 | loops = 10; |
167 | do { | |
dc873b6f JA |
168 | f->last_free_lookup = (f->num_maps - 1) * |
169 | (r / (OS_RAND_MAX + 1.0)); | |
4ba66134 | 170 | if (!get_next_free_block(td, f, ddir, b)) |
43c63a78 | 171 | return 0; |
ec4015da | 172 | |
43c63a78 JA |
173 | r = os_random_long(&td->random_state); |
174 | } while (--loops); | |
175 | ||
176 | /* | |
177 | * that didn't work either, try exhaustive search from the start | |
178 | */ | |
179 | f->last_free_lookup = 0; | |
4ba66134 | 180 | return get_next_free_block(td, f, ddir, b); |
ec4015da JA |
181 | } |
182 | ||
10ba535a JA |
183 | /* |
184 | * For random io, generate a random new block and see if it's used. Repeat | |
185 | * until we find a free one. For sequential io, just return the end of | |
186 | * the last io issued. | |
187 | */ | |
9bf2061e | 188 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 189 | { |
9bf2061e | 190 | struct fio_file *f = io_u->file; |
ec4015da | 191 | unsigned long long b; |
4ba66134 | 192 | enum fio_ddir ddir = io_u->ddir; |
10ba535a | 193 | |
ec4015da JA |
194 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
195 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 196 | |
4ba66134 | 197 | if (get_next_rand_offset(td, f, ddir, &b)) |
bca4ed4d | 198 | return 1; |
43063a1c | 199 | } else { |
0c3d768a | 200 | if (f->last_pos >= f->real_file_size) { |
4ba66134 JA |
201 | if (!td_random(td) || |
202 | get_next_rand_offset(td, f, ddir, &b)) | |
0c3d768a | 203 | return 1; |
bcdedd0a | 204 | } else |
4ba66134 | 205 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 206 | } |
10ba535a | 207 | |
009bd847 JA |
208 | io_u->offset = b * td->o.min_bs[ddir]; |
209 | if (io_u->offset >= f->io_size) { | |
210 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
211 | io_u->offset, f->io_size); | |
212 | return 1; | |
213 | } | |
214 | ||
215 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
216 | if (io_u->offset >= f->real_file_size) { |
217 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
218 | io_u->offset, f->real_file_size); | |
10ba535a | 219 | return 1; |
2ba1c290 | 220 | } |
10ba535a JA |
221 | |
222 | return 0; | |
223 | } | |
224 | ||
f3059de1 JA |
225 | static inline int is_power_of_2(unsigned int val) |
226 | { | |
227 | return (val != 0 && ((val & (val - 1)) == 0)); | |
228 | } | |
229 | ||
9bf2061e | 230 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 231 | { |
bca4ed4d | 232 | const int ddir = io_u->ddir; |
f3f552b9 | 233 | unsigned int uninitialized_var(buflen); |
f3059de1 | 234 | unsigned int minbs, maxbs; |
10ba535a JA |
235 | long r; |
236 | ||
f3059de1 JA |
237 | minbs = td->o.min_bs[ddir]; |
238 | maxbs = td->o.max_bs[ddir]; | |
239 | ||
240 | if (minbs == maxbs) | |
241 | buflen = minbs; | |
10ba535a JA |
242 | else { |
243 | r = os_random_long(&td->bsrange_state); | |
5ec10eaa | 244 | if (!td->o.bssplit_nr) { |
f3059de1 JA |
245 | buflen = 1 + (unsigned int) ((double) maxbs * |
246 | (r / (OS_RAND_MAX + 1.0))); | |
247 | if (buflen < minbs) | |
248 | buflen = minbs; | |
5ec10eaa | 249 | } else { |
564ca972 JA |
250 | long perc = 0; |
251 | unsigned int i; | |
252 | ||
253 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
254 | struct bssplit *bsp = &td->o.bssplit[i]; | |
255 | ||
256 | buflen = bsp->bs; | |
257 | perc += bsp->perc; | |
f3059de1 | 258 | if (r <= ((OS_RAND_MAX / 100L) * perc)) |
564ca972 JA |
259 | break; |
260 | } | |
261 | } | |
f3059de1 JA |
262 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
263 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a JA |
264 | } |
265 | ||
4ba66134 JA |
266 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
267 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
f3059de1 JA |
268 | minbs, ddir); |
269 | buflen = minbs; | |
4ba66134 | 270 | } |
6a5e6884 | 271 | |
10ba535a JA |
272 | return buflen; |
273 | } | |
274 | ||
afe24a5a JA |
275 | static void set_rwmix_bytes(struct thread_data *td) |
276 | { | |
afe24a5a JA |
277 | unsigned int diff; |
278 | ||
279 | /* | |
280 | * we do time or byte based switch. this is needed because | |
281 | * buffered writes may issue a lot quicker than they complete, | |
282 | * whereas reads do not. | |
283 | */ | |
e47f799f | 284 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 285 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
286 | } |
287 | ||
288 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
289 | { | |
290 | unsigned int v; | |
291 | long r; | |
292 | ||
293 | r = os_random_long(&td->rwmix_state); | |
dc873b6f | 294 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); |
04c540d9 | 295 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
296 | return DDIR_READ; |
297 | ||
298 | return DDIR_WRITE; | |
afe24a5a JA |
299 | } |
300 | ||
10ba535a JA |
301 | /* |
302 | * Return the data direction for the next io_u. If the job is a | |
303 | * mixed read/write workload, check the rwmix cycle and switch if | |
304 | * necessary. | |
305 | */ | |
1e97cce9 | 306 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
307 | { |
308 | if (td_rw(td)) { | |
10ba535a JA |
309 | /* |
310 | * Check if it's time to seed a new data direction. | |
311 | */ | |
e4928662 | 312 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f | 313 | unsigned long long max_bytes; |
5ec10eaa | 314 | enum fio_ddir ddir; |
e47f799f JA |
315 | |
316 | /* | |
317 | * Put a top limit on how many bytes we do for | |
318 | * one data direction, to avoid overflowing the | |
319 | * ranges too much | |
320 | */ | |
321 | ddir = get_rand_ddir(td); | |
322 | max_bytes = td->this_io_bytes[ddir]; | |
5ec10eaa JA |
323 | if (max_bytes >= |
324 | (td->o.size * td->o.rwmix[ddir] / 100)) { | |
6162f524 | 325 | if (!td->rw_end_set[ddir]) { |
38d77cae | 326 | td->rw_end_set[ddir] = 1; |
6162f524 JA |
327 | fio_gettime(&td->rw_end[ddir], NULL); |
328 | } | |
e4928662 | 329 | |
e47f799f | 330 | ddir ^= 1; |
38d77cae | 331 | } |
e47f799f JA |
332 | |
333 | if (ddir != td->rwmix_ddir) | |
334 | set_rwmix_bytes(td); | |
335 | ||
336 | td->rwmix_ddir = ddir; | |
10ba535a JA |
337 | } |
338 | return td->rwmix_ddir; | |
339 | } else if (td_read(td)) | |
340 | return DDIR_READ; | |
341 | else | |
342 | return DDIR_WRITE; | |
343 | } | |
344 | ||
60f2c658 JA |
345 | static void put_file_log(struct thread_data *td, struct fio_file *f) |
346 | { | |
347 | int ret = put_file(td, f); | |
348 | ||
349 | if (ret) | |
350 | td_verror(td, ret, "file close"); | |
351 | } | |
352 | ||
10ba535a JA |
353 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
354 | { | |
0c6e7517 JA |
355 | assert((io_u->flags & IO_U_F_FREE) == 0); |
356 | io_u->flags |= IO_U_F_FREE; | |
357 | ||
60f2c658 JA |
358 | if (io_u->file) |
359 | put_file_log(td, io_u->file); | |
2dbdab7e | 360 | |
10ba535a | 361 | io_u->file = NULL; |
01743ee1 JA |
362 | flist_del(&io_u->list); |
363 | flist_add(&io_u->list, &td->io_u_freelist); | |
10ba535a JA |
364 | td->cur_depth--; |
365 | } | |
366 | ||
755200a3 JA |
367 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
368 | { | |
369 | struct io_u *__io_u = *io_u; | |
370 | ||
465221b0 JA |
371 | dprint(FD_IO, "requeue %p\n", __io_u); |
372 | ||
4d2e0f49 | 373 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
374 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
375 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 376 | |
4d2e0f49 JA |
377 | __io_u->flags &= ~IO_U_F_FLIGHT; |
378 | ||
01743ee1 JA |
379 | flist_del(&__io_u->list); |
380 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
755200a3 JA |
381 | td->cur_depth--; |
382 | *io_u = NULL; | |
383 | } | |
384 | ||
9bf2061e | 385 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 386 | { |
b4c5e1ac JA |
387 | if (td->io_ops->flags & FIO_NOIO) |
388 | goto out; | |
389 | ||
87dc1ab1 JA |
390 | /* |
391 | * see if it's time to sync | |
392 | */ | |
2dc1bbeb JA |
393 | if (td->o.fsync_blocks && |
394 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
395 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 396 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 397 | goto out; |
87dc1ab1 JA |
398 | } |
399 | ||
a00735e6 JA |
400 | io_u->ddir = get_rw_ddir(td); |
401 | ||
48f5abd3 JA |
402 | /* |
403 | * See if it's time to switch to a new zone | |
404 | */ | |
405 | if (td->zone_bytes >= td->o.zone_size) { | |
406 | td->zone_bytes = 0; | |
407 | io_u->file->last_pos += td->o.zone_skip; | |
408 | td->io_skip_bytes += td->o.zone_skip; | |
409 | } | |
410 | ||
10ba535a | 411 | /* |
c685b5b2 JA |
412 | * No log, let the seq/rand engine retrieve the next buflen and |
413 | * position. | |
10ba535a | 414 | */ |
2ba1c290 JA |
415 | if (get_next_offset(td, io_u)) { |
416 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 417 | return 1; |
2ba1c290 | 418 | } |
10ba535a | 419 | |
9bf2061e | 420 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
421 | if (!io_u->buflen) { |
422 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 423 | return 1; |
2ba1c290 | 424 | } |
bca4ed4d | 425 | |
2ba1c290 JA |
426 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
427 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
428 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
429 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 430 | return 1; |
2ba1c290 | 431 | } |
6a5e6884 | 432 | |
bca4ed4d JA |
433 | /* |
434 | * mark entry before potentially trimming io_u | |
435 | */ | |
303032ae | 436 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 437 | mark_random_map(td, io_u); |
bca4ed4d JA |
438 | |
439 | /* | |
440 | * If using a write iolog, store this entry. | |
441 | */ | |
c38e9468 | 442 | out: |
2ba1c290 | 443 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 444 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 445 | log_io_u(td, io_u); |
bca4ed4d | 446 | return 0; |
10ba535a JA |
447 | } |
448 | ||
838bc709 JA |
449 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
450 | { | |
451 | int index = 0; | |
452 | ||
453 | switch (nr) { | |
454 | default: | |
455 | index = 6; | |
456 | break; | |
457 | case 33 ... 64: | |
458 | index = 5; | |
459 | break; | |
460 | case 17 ... 32: | |
461 | index = 4; | |
462 | break; | |
463 | case 9 ... 16: | |
464 | index = 3; | |
465 | break; | |
466 | case 5 ... 8: | |
467 | index = 2; | |
468 | break; | |
469 | case 1 ... 4: | |
470 | index = 1; | |
471 | case 0: | |
472 | break; | |
473 | } | |
474 | ||
475 | map[index]++; | |
476 | } | |
477 | ||
478 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
479 | { | |
480 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
481 | td->ts.total_submit++; | |
482 | } | |
483 | ||
484 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
485 | { | |
486 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
487 | td->ts.total_complete++; | |
488 | } | |
489 | ||
d8005759 | 490 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 JA |
491 | { |
492 | int index = 0; | |
493 | ||
494 | switch (td->cur_depth) { | |
495 | default: | |
a783e61a JA |
496 | index = 6; |
497 | break; | |
71619dc2 | 498 | case 32 ... 63: |
a783e61a JA |
499 | index = 5; |
500 | break; | |
71619dc2 | 501 | case 16 ... 31: |
a783e61a JA |
502 | index = 4; |
503 | break; | |
71619dc2 | 504 | case 8 ... 15: |
a783e61a JA |
505 | index = 3; |
506 | break; | |
71619dc2 | 507 | case 4 ... 7: |
a783e61a JA |
508 | index = 2; |
509 | break; | |
71619dc2 | 510 | case 2 ... 3: |
a783e61a | 511 | index = 1; |
71619dc2 JA |
512 | case 1: |
513 | break; | |
514 | } | |
515 | ||
3bec7ae0 | 516 | td->ts.io_u_map[index] += nr; |
71619dc2 JA |
517 | } |
518 | ||
04a0feae JA |
519 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
520 | { | |
521 | int index = 0; | |
522 | ||
523 | assert(usec < 1000); | |
524 | ||
525 | switch (usec) { | |
526 | case 750 ... 999: | |
527 | index = 9; | |
528 | break; | |
529 | case 500 ... 749: | |
530 | index = 8; | |
531 | break; | |
532 | case 250 ... 499: | |
533 | index = 7; | |
534 | break; | |
535 | case 100 ... 249: | |
536 | index = 6; | |
537 | break; | |
538 | case 50 ... 99: | |
539 | index = 5; | |
540 | break; | |
541 | case 20 ... 49: | |
542 | index = 4; | |
543 | break; | |
544 | case 10 ... 19: | |
545 | index = 3; | |
546 | break; | |
547 | case 4 ... 9: | |
548 | index = 2; | |
549 | break; | |
550 | case 2 ... 3: | |
551 | index = 1; | |
552 | case 0 ... 1: | |
553 | break; | |
554 | } | |
555 | ||
556 | assert(index < FIO_IO_U_LAT_U_NR); | |
557 | td->ts.io_u_lat_u[index]++; | |
558 | } | |
559 | ||
560 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
561 | { |
562 | int index = 0; | |
563 | ||
564 | switch (msec) { | |
565 | default: | |
04a0feae JA |
566 | index = 11; |
567 | break; | |
8abdce66 | 568 | case 1000 ... 1999: |
04a0feae JA |
569 | index = 10; |
570 | break; | |
8abdce66 | 571 | case 750 ... 999: |
04a0feae JA |
572 | index = 9; |
573 | break; | |
8abdce66 | 574 | case 500 ... 749: |
04a0feae JA |
575 | index = 8; |
576 | break; | |
8abdce66 | 577 | case 250 ... 499: |
04a0feae JA |
578 | index = 7; |
579 | break; | |
8abdce66 | 580 | case 100 ... 249: |
04a0feae JA |
581 | index = 6; |
582 | break; | |
8abdce66 | 583 | case 50 ... 99: |
04a0feae JA |
584 | index = 5; |
585 | break; | |
8abdce66 | 586 | case 20 ... 49: |
04a0feae JA |
587 | index = 4; |
588 | break; | |
8abdce66 | 589 | case 10 ... 19: |
04a0feae JA |
590 | index = 3; |
591 | break; | |
8abdce66 | 592 | case 4 ... 9: |
04a0feae JA |
593 | index = 2; |
594 | break; | |
ec118304 | 595 | case 2 ... 3: |
04a0feae | 596 | index = 1; |
ec118304 JA |
597 | case 0 ... 1: |
598 | break; | |
599 | } | |
600 | ||
04a0feae JA |
601 | assert(index < FIO_IO_U_LAT_M_NR); |
602 | td->ts.io_u_lat_m[index]++; | |
603 | } | |
604 | ||
605 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
606 | { | |
607 | if (usec < 1000) | |
608 | io_u_mark_lat_usec(td, usec); | |
609 | else | |
610 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
611 | } |
612 | ||
0aabe160 JA |
613 | /* |
614 | * Get next file to service by choosing one at random | |
615 | */ | |
1c178180 JA |
616 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
617 | int badf) | |
0aabe160 | 618 | { |
0aabe160 | 619 | struct fio_file *f; |
1c178180 | 620 | int fno; |
0aabe160 JA |
621 | |
622 | do { | |
7c83c089 JA |
623 | long r = os_random_long(&td->next_file_state); |
624 | ||
5ec10eaa | 625 | fno = (unsigned int) ((double) td->o.nr_files |
dc873b6f | 626 | * (r / (OS_RAND_MAX + 1.0))); |
126d65c6 | 627 | f = td->files[fno]; |
059e63c0 JA |
628 | if (f->flags & FIO_FILE_DONE) |
629 | continue; | |
1c178180 | 630 | |
2ba1c290 JA |
631 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
632 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 633 | return f; |
2ba1c290 | 634 | } |
0aabe160 JA |
635 | } while (1); |
636 | } | |
637 | ||
638 | /* | |
639 | * Get next file to service by doing round robin between all available ones | |
640 | */ | |
1c178180 JA |
641 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
642 | int badf) | |
3d7c391d JA |
643 | { |
644 | unsigned int old_next_file = td->next_file; | |
645 | struct fio_file *f; | |
646 | ||
647 | do { | |
126d65c6 | 648 | f = td->files[td->next_file]; |
3d7c391d JA |
649 | |
650 | td->next_file++; | |
2dc1bbeb | 651 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
652 | td->next_file = 0; |
653 | ||
d5ed68ea JA |
654 | if (f->flags & FIO_FILE_DONE) { |
655 | f = NULL; | |
059e63c0 | 656 | continue; |
d5ed68ea | 657 | } |
059e63c0 | 658 | |
1c178180 | 659 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
660 | break; |
661 | ||
662 | f = NULL; | |
663 | } while (td->next_file != old_next_file); | |
664 | ||
2ba1c290 | 665 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
666 | return f; |
667 | } | |
668 | ||
bdb4e2e9 JA |
669 | static struct fio_file *get_next_file(struct thread_data *td) |
670 | { | |
1907dbc6 JA |
671 | struct fio_file *f; |
672 | ||
2dc1bbeb | 673 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 674 | |
2ba1c290 | 675 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
676 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
677 | " nr_files=%d\n", td->nr_open_files, | |
678 | td->nr_done_files, | |
679 | td->o.nr_files); | |
bdb4e2e9 | 680 | return NULL; |
2ba1c290 | 681 | } |
bdb4e2e9 | 682 | |
1907dbc6 | 683 | f = td->file_service_file; |
f11bd94d | 684 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
2ba1c290 | 685 | goto out; |
1907dbc6 | 686 | |
2dc1bbeb | 687 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 | 688 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 689 | else |
1c178180 | 690 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
691 | |
692 | td->file_service_file = f; | |
693 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 JA |
694 | out: |
695 | dprint(FD_FILE, "get_next_file: %p\n", f); | |
1907dbc6 | 696 | return f; |
bdb4e2e9 JA |
697 | } |
698 | ||
1c178180 JA |
699 | static struct fio_file *find_next_new_file(struct thread_data *td) |
700 | { | |
701 | struct fio_file *f; | |
702 | ||
1020a139 JA |
703 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
704 | return NULL; | |
705 | ||
2dc1bbeb | 706 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
1c178180 JA |
707 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
708 | else | |
709 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); | |
710 | ||
711 | return f; | |
712 | } | |
713 | ||
429f6675 JA |
714 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
715 | { | |
716 | struct fio_file *f; | |
717 | ||
718 | do { | |
719 | f = get_next_file(td); | |
720 | if (!f) | |
721 | return 1; | |
722 | ||
723 | set_file: | |
724 | io_u->file = f; | |
725 | get_file(f); | |
726 | ||
727 | if (!fill_io_u(td, io_u)) | |
728 | break; | |
729 | ||
009bd847 JA |
730 | /* |
731 | * optimization to prevent close/open of the same file. This | |
732 | * way we preserve queueing etc. | |
733 | */ | |
734 | if (td->o.nr_files == 1 && td->o.time_based) { | |
60f2c658 | 735 | put_file_log(td, f); |
009bd847 JA |
736 | fio_file_reset(f); |
737 | goto set_file; | |
738 | } | |
739 | ||
429f6675 JA |
740 | /* |
741 | * td_io_close() does a put_file() as well, so no need to | |
742 | * do that here. | |
743 | */ | |
744 | io_u->file = NULL; | |
745 | td_io_close_file(td, f); | |
746 | f->flags |= FIO_FILE_DONE; | |
747 | td->nr_done_files++; | |
748 | ||
749 | /* | |
750 | * probably not the right place to do this, but see | |
751 | * if we need to open a new file | |
752 | */ | |
753 | if (td->nr_open_files < td->o.open_files && | |
754 | td->o.open_files != td->o.nr_files) { | |
755 | f = find_next_new_file(td); | |
756 | ||
757 | if (!f || td_io_open_file(td, f)) | |
758 | return 1; | |
759 | ||
760 | goto set_file; | |
761 | } | |
762 | } while (1); | |
763 | ||
764 | return 0; | |
765 | } | |
766 | ||
767 | ||
10ba535a JA |
768 | struct io_u *__get_io_u(struct thread_data *td) |
769 | { | |
770 | struct io_u *io_u = NULL; | |
771 | ||
01743ee1 JA |
772 | if (!flist_empty(&td->io_u_requeues)) |
773 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 774 | else if (!queue_full(td)) { |
01743ee1 | 775 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 776 | |
6040dabc | 777 | io_u->buflen = 0; |
10ba535a | 778 | io_u->resid = 0; |
755200a3 | 779 | io_u->file = NULL; |
d7762cf8 | 780 | io_u->end_io = NULL; |
755200a3 JA |
781 | } |
782 | ||
783 | if (io_u) { | |
0c6e7517 JA |
784 | assert(io_u->flags & IO_U_F_FREE); |
785 | io_u->flags &= ~IO_U_F_FREE; | |
786 | ||
755200a3 | 787 | io_u->error = 0; |
01743ee1 JA |
788 | flist_del(&io_u->list); |
789 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a JA |
790 | td->cur_depth++; |
791 | } | |
792 | ||
793 | return io_u; | |
794 | } | |
795 | ||
796 | /* | |
797 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
798 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
799 | */ | |
3d7c391d | 800 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 801 | { |
3d7c391d | 802 | struct fio_file *f; |
10ba535a JA |
803 | struct io_u *io_u; |
804 | ||
805 | io_u = __get_io_u(td); | |
2ba1c290 JA |
806 | if (!io_u) { |
807 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 808 | return NULL; |
2ba1c290 | 809 | } |
10ba535a | 810 | |
755200a3 JA |
811 | /* |
812 | * from a requeue, io_u already setup | |
813 | */ | |
814 | if (io_u->file) | |
77f392bf | 815 | goto out; |
755200a3 | 816 | |
429f6675 JA |
817 | /* |
818 | * If using an iolog, grab next piece if any available. | |
819 | */ | |
820 | if (td->o.read_iolog_file) { | |
821 | if (read_iolog_get(td, io_u)) | |
822 | goto err_put; | |
2ba1c290 JA |
823 | } else if (set_io_u_file(td, io_u)) { |
824 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 825 | goto err_put; |
2ba1c290 | 826 | } |
5ec10eaa | 827 | |
429f6675 JA |
828 | f = io_u->file; |
829 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 830 | |
87dc1ab1 | 831 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 832 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 833 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 834 | goto err_put; |
2ba1c290 | 835 | } |
10ba535a | 836 | |
36167d82 | 837 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 838 | |
2dc1bbeb | 839 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 | 840 | populate_verify_io_u(td, io_u); |
e4dad9c6 JA |
841 | else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) |
842 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
87dc1ab1 | 843 | } |
10ba535a | 844 | |
165faf16 JA |
845 | /* |
846 | * Set io data pointers. | |
847 | */ | |
d460eb31 | 848 | io_u->endpos = io_u->offset + io_u->buflen; |
cec6b55d JA |
849 | io_u->xfer_buf = io_u->buf; |
850 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 851 | |
6ac7a331 | 852 | out: |
429f6675 | 853 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
854 | if (!td->o.disable_slat) |
855 | fio_gettime(&io_u->start_time, NULL); | |
429f6675 | 856 | return io_u; |
36167d82 | 857 | } |
429f6675 | 858 | err_put: |
2ba1c290 | 859 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
860 | put_io_u(td, io_u); |
861 | return NULL; | |
10ba535a JA |
862 | } |
863 | ||
5451792e JA |
864 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
865 | { | |
866 | const char *msg[] = { "read", "write", "sync" }; | |
867 | ||
868 | log_err("fio: io_u error"); | |
869 | ||
870 | if (io_u->file) | |
871 | log_err(" on file %s", io_u->file->file_name); | |
872 | ||
873 | log_err(": %s\n", strerror(io_u->error)); | |
874 | ||
5ec10eaa JA |
875 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
876 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
877 | |
878 | if (!td->error) | |
879 | td_verror(td, io_u->error, "io_u error"); | |
880 | } | |
881 | ||
97601024 JA |
882 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
883 | struct io_completion_data *icd) | |
10ba535a | 884 | { |
dbad30b0 JA |
885 | /* |
886 | * Older gcc's are too dumb to realize that usec is always used | |
887 | * initialized, silence that warning. | |
888 | */ | |
889 | unsigned long uninitialized_var(usec); | |
10ba535a | 890 | |
2ba1c290 JA |
891 | dprint_io_u(io_u, "io complete"); |
892 | ||
0c6e7517 JA |
893 | assert(io_u->flags & IO_U_F_FLIGHT); |
894 | io_u->flags &= ~IO_U_F_FLIGHT; | |
895 | ||
87dc1ab1 JA |
896 | if (io_u->ddir == DDIR_SYNC) { |
897 | td->last_was_sync = 1; | |
898 | return; | |
899 | } | |
900 | ||
901 | td->last_was_sync = 0; | |
902 | ||
10ba535a JA |
903 | if (!io_u->error) { |
904 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 905 | const enum fio_ddir idx = io_u->ddir; |
b29ee5b3 | 906 | int ret; |
10ba535a | 907 | |
b29ee5b3 JA |
908 | td->io_blocks[idx]++; |
909 | td->io_bytes[idx] += bytes; | |
910 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 911 | |
b29ee5b3 | 912 | if (ramp_time_over(td)) { |
9520ebb9 JA |
913 | if (!td->o.disable_clat || !td->o.disable_bw) |
914 | usec = utime_since(&io_u->issue_time, | |
915 | &icd->time); | |
721938ae | 916 | |
9520ebb9 JA |
917 | if (!td->o.disable_clat) { |
918 | add_clat_sample(td, idx, usec); | |
919 | io_u_mark_latency(td, usec); | |
920 | } | |
921 | if (!td->o.disable_bw) | |
922 | add_bw_sample(td, idx, &icd->time); | |
721938ae | 923 | } |
10ba535a | 924 | |
660a1cb5 | 925 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 926 | td->o.do_verify && |
41128405 | 927 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
928 | log_io_piece(td, io_u); |
929 | ||
b29ee5b3 | 930 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 931 | |
d7762cf8 | 932 | if (io_u->end_io) { |
36690c9b | 933 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
934 | if (ret && !icd->error) |
935 | icd->error = ret; | |
936 | } | |
5451792e | 937 | } else { |
10ba535a | 938 | icd->error = io_u->error; |
5451792e JA |
939 | io_u_log_error(td, io_u); |
940 | } | |
10ba535a JA |
941 | } |
942 | ||
9520ebb9 JA |
943 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
944 | int nr) | |
10ba535a | 945 | { |
9520ebb9 JA |
946 | if (!td->o.disable_clat || !td->o.disable_bw) |
947 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 948 | |
3af6ef39 JA |
949 | icd->nr = nr; |
950 | ||
10ba535a JA |
951 | icd->error = 0; |
952 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
953 | } |
954 | ||
97601024 JA |
955 | static void ios_completed(struct thread_data *td, |
956 | struct io_completion_data *icd) | |
36167d82 JA |
957 | { |
958 | struct io_u *io_u; | |
959 | int i; | |
960 | ||
10ba535a JA |
961 | for (i = 0; i < icd->nr; i++) { |
962 | io_u = td->io_ops->event(td, i); | |
963 | ||
964 | io_completed(td, io_u, icd); | |
965 | put_io_u(td, io_u); | |
966 | } | |
967 | } | |
97601024 | 968 | |
e7e6cfb4 JA |
969 | /* |
970 | * Complete a single io_u for the sync engines. | |
971 | */ | |
d7762cf8 | 972 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
973 | { |
974 | struct io_completion_data icd; | |
975 | ||
9520ebb9 | 976 | init_icd(td, &icd, 1); |
97601024 JA |
977 | io_completed(td, io_u, &icd); |
978 | put_io_u(td, io_u); | |
979 | ||
980 | if (!icd.error) | |
981 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
982 | ||
37e974a5 | 983 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
984 | return -1; |
985 | } | |
986 | ||
e7e6cfb4 JA |
987 | /* |
988 | * Called to complete min_events number of io for the async engines. | |
989 | */ | |
4950421a | 990 | long io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 991 | { |
97601024 | 992 | struct io_completion_data icd; |
00de55ef | 993 | struct timespec *tvp = NULL; |
97601024 | 994 | int ret; |
4d06a338 | 995 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 996 | |
4950421a | 997 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 998 | |
4950421a | 999 | if (!min_evts) |
00de55ef | 1000 | tvp = &ts; |
97601024 | 1001 | |
4950421a | 1002 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1003 | if (ret < 0) { |
e1161c32 | 1004 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1005 | return ret; |
1006 | } else if (!ret) | |
1007 | return ret; | |
1008 | ||
9520ebb9 | 1009 | init_icd(td, &icd, ret); |
97601024 JA |
1010 | ios_completed(td, &icd); |
1011 | if (!icd.error) | |
1012 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
1013 | ||
37e974a5 | 1014 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
1015 | return -1; |
1016 | } | |
7e77dd02 JA |
1017 | |
1018 | /* | |
1019 | * Call when io_u is really queued, to update the submission latency. | |
1020 | */ | |
1021 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1022 | { | |
9520ebb9 JA |
1023 | if (!td->o.disable_slat) { |
1024 | unsigned long slat_time; | |
7e77dd02 | 1025 | |
9520ebb9 JA |
1026 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
1027 | add_slat_sample(td, io_u->ddir, slat_time); | |
1028 | } | |
7e77dd02 | 1029 | } |
433afcb4 | 1030 | |
5973cafb JA |
1031 | /* |
1032 | * "randomly" fill the buffer contents | |
1033 | */ | |
1034 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1035 | unsigned int max_bs) | |
1036 | { | |
1037 | long *ptr = io_u->buf; | |
1038 | ||
1039 | if (!td->o.zero_buffers) { | |
1040 | while ((void *) ptr - io_u->buf < max_bs) { | |
1041 | *ptr = rand() * GOLDEN_RATIO_PRIME; | |
1042 | ptr++; | |
1043 | } | |
1044 | } else | |
1045 | memset(ptr, 0, max_bs); | |
1046 | } |