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" |
4f5af7b2 | 10 | #include "verify.h" |
10ba535a | 11 | |
97601024 JA |
12 | struct io_completion_data { |
13 | int nr; /* input */ | |
97601024 JA |
14 | |
15 | int error; /* output */ | |
16 | unsigned long bytes_done[2]; /* output */ | |
17 | struct timeval time; /* output */ | |
18 | }; | |
19 | ||
10ba535a JA |
20 | /* |
21 | * The ->file_map[] contains a map of blocks we have or have not done io | |
22 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
23 | */ | |
aec2de20 | 24 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 25 | { |
aec2de20 JA |
26 | unsigned int idx = RAND_MAP_IDX(f, block); |
27 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 28 | |
84422acd JA |
29 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
30 | ||
de605666 | 31 | return (f->file_map[idx] & (1 << bit)) == 0; |
10ba535a JA |
32 | } |
33 | ||
df415585 JA |
34 | /* |
35 | * Mark a given offset as used in the map. | |
36 | */ | |
9bf2061e | 37 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 38 | { |
2dc1bbeb | 39 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 40 | struct fio_file *f = io_u->file; |
a00735e6 | 41 | unsigned long long block; |
3e3357b1 | 42 | unsigned int blocks, nr_blocks; |
df415585 | 43 | |
b9c5b644 | 44 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
c685b5b2 | 45 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
3e3357b1 | 46 | blocks = 0; |
c685b5b2 | 47 | |
3e3357b1 JA |
48 | while (nr_blocks) { |
49 | unsigned int this_blocks, mask; | |
df415585 JA |
50 | unsigned int idx, bit; |
51 | ||
1e3d53ac JA |
52 | /* |
53 | * If we have a mixed random workload, we may | |
54 | * encounter blocks we already did IO to. | |
55 | */ | |
8b113b4f JA |
56 | if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) { |
57 | if (!blocks) | |
58 | blocks = 1; | |
df415585 | 59 | break; |
8b113b4f | 60 | } |
df415585 | 61 | |
aec2de20 JA |
62 | idx = RAND_MAP_IDX(f, block); |
63 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 64 | |
0032bf9f | 65 | fio_assert(td, idx < f->num_maps); |
df415585 | 66 | |
3e3357b1 JA |
67 | this_blocks = nr_blocks; |
68 | if (this_blocks + bit > BLOCKS_PER_MAP) | |
69 | this_blocks = BLOCKS_PER_MAP - bit; | |
70 | ||
71 | if (this_blocks == BLOCKS_PER_MAP) | |
72 | mask = -1U; | |
73 | else | |
74 | mask = ((1U << this_blocks) - 1) << bit; | |
75 | ||
3e3357b1 JA |
76 | f->file_map[idx] |= mask; |
77 | nr_blocks -= this_blocks; | |
78 | blocks += this_blocks; | |
c9dd34b2 | 79 | block += this_blocks; |
df415585 JA |
80 | } |
81 | ||
a00735e6 JA |
82 | if ((blocks * min_bs) < io_u->buflen) |
83 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
84 | } |
85 | ||
3e3357b1 JA |
86 | static unsigned long long last_block(struct thread_data *td, struct fio_file *f, |
87 | enum fio_ddir ddir) | |
2ba1c290 JA |
88 | { |
89 | unsigned long long max_blocks; | |
d9dd70f7 | 90 | unsigned long long max_size; |
2ba1c290 | 91 | |
d9dd70f7 JA |
92 | /* |
93 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
94 | */ | |
95 | max_size = f->io_size; | |
96 | if (max_size > f->real_file_size) | |
97 | max_size = f->real_file_size; | |
98 | ||
2b7a01d0 | 99 | max_blocks = max_size / (unsigned long long) td->o.ba[ddir]; |
2ba1c290 JA |
100 | if (!max_blocks) |
101 | return 0; | |
102 | ||
67778e88 | 103 | return max_blocks; |
2ba1c290 JA |
104 | } |
105 | ||
10ba535a JA |
106 | /* |
107 | * Return the next free block in the map. | |
108 | */ | |
109 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 110 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 111 | { |
84422acd | 112 | unsigned long long min_bs = td->o.rw_min_bs; |
10ba535a JA |
113 | int i; |
114 | ||
c685b5b2 JA |
115 | i = f->last_free_lookup; |
116 | *b = (i * BLOCKS_PER_MAP); | |
899c29f2 RR |
117 | while ((*b) * min_bs < f->real_file_size && |
118 | (*b) * min_bs < f->io_size) { | |
de605666 | 119 | if (f->file_map[i] != (unsigned int) -1) { |
697a606c | 120 | *b += ffz(f->file_map[i]); |
4ba66134 | 121 | if (*b > last_block(td, f, ddir)) |
2ba1c290 | 122 | break; |
c685b5b2 | 123 | f->last_free_lookup = i; |
10ba535a JA |
124 | return 0; |
125 | } | |
126 | ||
127 | *b += BLOCKS_PER_MAP; | |
128 | i++; | |
129 | } | |
130 | ||
2ba1c290 | 131 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
132 | return 1; |
133 | } | |
134 | ||
ec4015da | 135 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 136 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 137 | { |
84422acd | 138 | unsigned long long r; |
ec4015da JA |
139 | int loops = 5; |
140 | ||
141 | do { | |
142 | r = os_random_long(&td->random_state); | |
84422acd | 143 | dprint(FD_RANDOM, "off rand %llu\n", r); |
5ec10eaa | 144 | *b = (last_block(td, f, ddir) - 1) |
dc873b6f | 145 | * (r / ((unsigned long long) OS_RAND_MAX + 1.0)); |
2ba1c290 | 146 | |
43c63a78 JA |
147 | /* |
148 | * if we are not maintaining a random map, we are done. | |
149 | */ | |
303032ae | 150 | if (!file_randommap(td, f)) |
43c63a78 JA |
151 | return 0; |
152 | ||
153 | /* | |
84422acd | 154 | * calculate map offset and check if it's free |
43c63a78 | 155 | */ |
aec2de20 | 156 | if (random_map_free(f, *b)) |
43c63a78 JA |
157 | return 0; |
158 | ||
84422acd JA |
159 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
160 | *b); | |
43c63a78 | 161 | } while (--loops); |
ec4015da JA |
162 | |
163 | /* | |
43c63a78 JA |
164 | * we get here, if we didn't suceed in looking up a block. generate |
165 | * a random start offset into the filemap, and find the first free | |
166 | * block from there. | |
ec4015da | 167 | */ |
43c63a78 JA |
168 | loops = 10; |
169 | do { | |
dc873b6f JA |
170 | f->last_free_lookup = (f->num_maps - 1) * |
171 | (r / (OS_RAND_MAX + 1.0)); | |
4ba66134 | 172 | if (!get_next_free_block(td, f, ddir, b)) |
43c63a78 | 173 | return 0; |
ec4015da | 174 | |
43c63a78 JA |
175 | r = os_random_long(&td->random_state); |
176 | } while (--loops); | |
177 | ||
178 | /* | |
179 | * that didn't work either, try exhaustive search from the start | |
180 | */ | |
181 | f->last_free_lookup = 0; | |
4ba66134 | 182 | return get_next_free_block(td, f, ddir, b); |
ec4015da JA |
183 | } |
184 | ||
10ba535a JA |
185 | /* |
186 | * For random io, generate a random new block and see if it's used. Repeat | |
187 | * until we find a free one. For sequential io, just return the end of | |
188 | * the last io issued. | |
189 | */ | |
9bf2061e | 190 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 191 | { |
9bf2061e | 192 | struct fio_file *f = io_u->file; |
ec4015da | 193 | unsigned long long b; |
4ba66134 | 194 | enum fio_ddir ddir = io_u->ddir; |
10ba535a | 195 | |
ec4015da JA |
196 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
197 | td->ddir_nr = td->o.ddir_nr; | |
211097b2 | 198 | |
683023e8 JA |
199 | if (get_next_rand_offset(td, f, ddir, &b)) { |
200 | dprint(FD_IO, "%s: getting rand offset failed\n", | |
201 | f->file_name); | |
bca4ed4d | 202 | return 1; |
683023e8 | 203 | } |
43063a1c | 204 | } else { |
0c3d768a | 205 | if (f->last_pos >= f->real_file_size) { |
4ba66134 | 206 | if (!td_random(td) || |
683023e8 JA |
207 | get_next_rand_offset(td, f, ddir, &b)) { |
208 | dprint(FD_IO, "%s: pos %llu > size %llu\n", | |
209 | f->file_name, f->last_pos, | |
210 | f->real_file_size); | |
0c3d768a | 211 | return 1; |
683023e8 | 212 | } |
bcdedd0a | 213 | } else |
4ba66134 | 214 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 215 | } |
10ba535a | 216 | |
2b7a01d0 | 217 | io_u->offset = b * td->o.ba[ddir]; |
009bd847 JA |
218 | if (io_u->offset >= f->io_size) { |
219 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
220 | io_u->offset, f->io_size); | |
221 | return 1; | |
222 | } | |
223 | ||
224 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
225 | if (io_u->offset >= f->real_file_size) { |
226 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
227 | io_u->offset, f->real_file_size); | |
10ba535a | 228 | return 1; |
2ba1c290 | 229 | } |
10ba535a JA |
230 | |
231 | return 0; | |
232 | } | |
233 | ||
f3059de1 JA |
234 | static inline int is_power_of_2(unsigned int val) |
235 | { | |
236 | return (val != 0 && ((val & (val - 1)) == 0)); | |
237 | } | |
238 | ||
9bf2061e | 239 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 240 | { |
bca4ed4d | 241 | const int ddir = io_u->ddir; |
f3f552b9 | 242 | unsigned int uninitialized_var(buflen); |
f3059de1 | 243 | unsigned int minbs, maxbs; |
10ba535a JA |
244 | long r; |
245 | ||
f3059de1 JA |
246 | minbs = td->o.min_bs[ddir]; |
247 | maxbs = td->o.max_bs[ddir]; | |
248 | ||
249 | if (minbs == maxbs) | |
250 | buflen = minbs; | |
10ba535a JA |
251 | else { |
252 | r = os_random_long(&td->bsrange_state); | |
720e84ad | 253 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 JA |
254 | buflen = 1 + (unsigned int) ((double) maxbs * |
255 | (r / (OS_RAND_MAX + 1.0))); | |
256 | if (buflen < minbs) | |
257 | buflen = minbs; | |
5ec10eaa | 258 | } else { |
564ca972 JA |
259 | long perc = 0; |
260 | unsigned int i; | |
261 | ||
720e84ad JA |
262 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
263 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
264 | |
265 | buflen = bsp->bs; | |
266 | perc += bsp->perc; | |
f3059de1 | 267 | if (r <= ((OS_RAND_MAX / 100L) * perc)) |
564ca972 JA |
268 | break; |
269 | } | |
270 | } | |
f3059de1 JA |
271 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
272 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a JA |
273 | } |
274 | ||
4ba66134 JA |
275 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
276 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
f3059de1 JA |
277 | minbs, ddir); |
278 | buflen = minbs; | |
4ba66134 | 279 | } |
6a5e6884 | 280 | |
10ba535a JA |
281 | return buflen; |
282 | } | |
283 | ||
afe24a5a JA |
284 | static void set_rwmix_bytes(struct thread_data *td) |
285 | { | |
afe24a5a JA |
286 | unsigned int diff; |
287 | ||
288 | /* | |
289 | * we do time or byte based switch. this is needed because | |
290 | * buffered writes may issue a lot quicker than they complete, | |
291 | * whereas reads do not. | |
292 | */ | |
e47f799f | 293 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 294 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
295 | } |
296 | ||
297 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
298 | { | |
299 | unsigned int v; | |
300 | long r; | |
301 | ||
302 | r = os_random_long(&td->rwmix_state); | |
dc873b6f | 303 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); |
04c540d9 | 304 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
305 | return DDIR_READ; |
306 | ||
307 | return DDIR_WRITE; | |
afe24a5a JA |
308 | } |
309 | ||
10ba535a JA |
310 | /* |
311 | * Return the data direction for the next io_u. If the job is a | |
312 | * mixed read/write workload, check the rwmix cycle and switch if | |
313 | * necessary. | |
314 | */ | |
1e97cce9 | 315 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
316 | { |
317 | if (td_rw(td)) { | |
10ba535a JA |
318 | /* |
319 | * Check if it's time to seed a new data direction. | |
320 | */ | |
e4928662 | 321 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f | 322 | unsigned long long max_bytes; |
5ec10eaa | 323 | enum fio_ddir ddir; |
e47f799f JA |
324 | |
325 | /* | |
326 | * Put a top limit on how many bytes we do for | |
327 | * one data direction, to avoid overflowing the | |
328 | * ranges too much | |
329 | */ | |
330 | ddir = get_rand_ddir(td); | |
331 | max_bytes = td->this_io_bytes[ddir]; | |
5ec10eaa JA |
332 | if (max_bytes >= |
333 | (td->o.size * td->o.rwmix[ddir] / 100)) { | |
6162f524 | 334 | if (!td->rw_end_set[ddir]) { |
38d77cae | 335 | td->rw_end_set[ddir] = 1; |
6162f524 JA |
336 | fio_gettime(&td->rw_end[ddir], NULL); |
337 | } | |
e4928662 | 338 | |
e47f799f | 339 | ddir ^= 1; |
38d77cae | 340 | } |
e47f799f JA |
341 | |
342 | if (ddir != td->rwmix_ddir) | |
343 | set_rwmix_bytes(td); | |
344 | ||
345 | td->rwmix_ddir = ddir; | |
10ba535a JA |
346 | } |
347 | return td->rwmix_ddir; | |
348 | } else if (td_read(td)) | |
349 | return DDIR_READ; | |
350 | else | |
351 | return DDIR_WRITE; | |
352 | } | |
353 | ||
60f2c658 JA |
354 | static void put_file_log(struct thread_data *td, struct fio_file *f) |
355 | { | |
356 | int ret = put_file(td, f); | |
357 | ||
358 | if (ret) | |
359 | td_verror(td, ret, "file close"); | |
360 | } | |
361 | ||
10ba535a JA |
362 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
363 | { | |
0c6e7517 JA |
364 | assert((io_u->flags & IO_U_F_FREE) == 0); |
365 | io_u->flags |= IO_U_F_FREE; | |
366 | ||
60f2c658 JA |
367 | if (io_u->file) |
368 | put_file_log(td, io_u->file); | |
2dbdab7e | 369 | |
10ba535a | 370 | io_u->file = NULL; |
01743ee1 JA |
371 | flist_del(&io_u->list); |
372 | flist_add(&io_u->list, &td->io_u_freelist); | |
10ba535a JA |
373 | td->cur_depth--; |
374 | } | |
375 | ||
755200a3 JA |
376 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
377 | { | |
378 | struct io_u *__io_u = *io_u; | |
379 | ||
465221b0 JA |
380 | dprint(FD_IO, "requeue %p\n", __io_u); |
381 | ||
4d2e0f49 | 382 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
383 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
384 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 385 | |
4d2e0f49 JA |
386 | __io_u->flags &= ~IO_U_F_FLIGHT; |
387 | ||
01743ee1 JA |
388 | flist_del(&__io_u->list); |
389 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
755200a3 JA |
390 | td->cur_depth--; |
391 | *io_u = NULL; | |
392 | } | |
393 | ||
9bf2061e | 394 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 395 | { |
b4c5e1ac JA |
396 | if (td->io_ops->flags & FIO_NOIO) |
397 | goto out; | |
398 | ||
87dc1ab1 JA |
399 | /* |
400 | * see if it's time to sync | |
401 | */ | |
2dc1bbeb JA |
402 | if (td->o.fsync_blocks && |
403 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
404 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 405 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 406 | goto out; |
87dc1ab1 JA |
407 | } |
408 | ||
a00735e6 JA |
409 | io_u->ddir = get_rw_ddir(td); |
410 | ||
48f5abd3 JA |
411 | /* |
412 | * See if it's time to switch to a new zone | |
413 | */ | |
414 | if (td->zone_bytes >= td->o.zone_size) { | |
415 | td->zone_bytes = 0; | |
416 | io_u->file->last_pos += td->o.zone_skip; | |
417 | td->io_skip_bytes += td->o.zone_skip; | |
418 | } | |
419 | ||
10ba535a | 420 | /* |
c685b5b2 JA |
421 | * No log, let the seq/rand engine retrieve the next buflen and |
422 | * position. | |
10ba535a | 423 | */ |
2ba1c290 JA |
424 | if (get_next_offset(td, io_u)) { |
425 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 426 | return 1; |
2ba1c290 | 427 | } |
10ba535a | 428 | |
9bf2061e | 429 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
430 | if (!io_u->buflen) { |
431 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 432 | return 1; |
2ba1c290 | 433 | } |
bca4ed4d | 434 | |
2ba1c290 JA |
435 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
436 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
437 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
438 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 439 | return 1; |
2ba1c290 | 440 | } |
6a5e6884 | 441 | |
bca4ed4d JA |
442 | /* |
443 | * mark entry before potentially trimming io_u | |
444 | */ | |
303032ae | 445 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 446 | mark_random_map(td, io_u); |
bca4ed4d JA |
447 | |
448 | /* | |
449 | * If using a write iolog, store this entry. | |
450 | */ | |
c38e9468 | 451 | out: |
2ba1c290 | 452 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 453 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 454 | log_io_u(td, io_u); |
bca4ed4d | 455 | return 0; |
10ba535a JA |
456 | } |
457 | ||
838bc709 JA |
458 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
459 | { | |
460 | int index = 0; | |
461 | ||
462 | switch (nr) { | |
463 | default: | |
464 | index = 6; | |
465 | break; | |
466 | case 33 ... 64: | |
467 | index = 5; | |
468 | break; | |
469 | case 17 ... 32: | |
470 | index = 4; | |
471 | break; | |
472 | case 9 ... 16: | |
473 | index = 3; | |
474 | break; | |
475 | case 5 ... 8: | |
476 | index = 2; | |
477 | break; | |
478 | case 1 ... 4: | |
479 | index = 1; | |
480 | case 0: | |
481 | break; | |
482 | } | |
483 | ||
484 | map[index]++; | |
485 | } | |
486 | ||
487 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
488 | { | |
489 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
490 | td->ts.total_submit++; | |
491 | } | |
492 | ||
493 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
494 | { | |
495 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
496 | td->ts.total_complete++; | |
497 | } | |
498 | ||
d8005759 | 499 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 JA |
500 | { |
501 | int index = 0; | |
502 | ||
503 | switch (td->cur_depth) { | |
504 | default: | |
a783e61a JA |
505 | index = 6; |
506 | break; | |
71619dc2 | 507 | case 32 ... 63: |
a783e61a JA |
508 | index = 5; |
509 | break; | |
71619dc2 | 510 | case 16 ... 31: |
a783e61a JA |
511 | index = 4; |
512 | break; | |
71619dc2 | 513 | case 8 ... 15: |
a783e61a JA |
514 | index = 3; |
515 | break; | |
71619dc2 | 516 | case 4 ... 7: |
a783e61a JA |
517 | index = 2; |
518 | break; | |
71619dc2 | 519 | case 2 ... 3: |
a783e61a | 520 | index = 1; |
71619dc2 JA |
521 | case 1: |
522 | break; | |
523 | } | |
524 | ||
3bec7ae0 | 525 | td->ts.io_u_map[index] += nr; |
71619dc2 JA |
526 | } |
527 | ||
04a0feae JA |
528 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
529 | { | |
530 | int index = 0; | |
531 | ||
532 | assert(usec < 1000); | |
533 | ||
534 | switch (usec) { | |
535 | case 750 ... 999: | |
536 | index = 9; | |
537 | break; | |
538 | case 500 ... 749: | |
539 | index = 8; | |
540 | break; | |
541 | case 250 ... 499: | |
542 | index = 7; | |
543 | break; | |
544 | case 100 ... 249: | |
545 | index = 6; | |
546 | break; | |
547 | case 50 ... 99: | |
548 | index = 5; | |
549 | break; | |
550 | case 20 ... 49: | |
551 | index = 4; | |
552 | break; | |
553 | case 10 ... 19: | |
554 | index = 3; | |
555 | break; | |
556 | case 4 ... 9: | |
557 | index = 2; | |
558 | break; | |
559 | case 2 ... 3: | |
560 | index = 1; | |
561 | case 0 ... 1: | |
562 | break; | |
563 | } | |
564 | ||
565 | assert(index < FIO_IO_U_LAT_U_NR); | |
566 | td->ts.io_u_lat_u[index]++; | |
567 | } | |
568 | ||
569 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
570 | { |
571 | int index = 0; | |
572 | ||
573 | switch (msec) { | |
574 | default: | |
04a0feae JA |
575 | index = 11; |
576 | break; | |
8abdce66 | 577 | case 1000 ... 1999: |
04a0feae JA |
578 | index = 10; |
579 | break; | |
8abdce66 | 580 | case 750 ... 999: |
04a0feae JA |
581 | index = 9; |
582 | break; | |
8abdce66 | 583 | case 500 ... 749: |
04a0feae JA |
584 | index = 8; |
585 | break; | |
8abdce66 | 586 | case 250 ... 499: |
04a0feae JA |
587 | index = 7; |
588 | break; | |
8abdce66 | 589 | case 100 ... 249: |
04a0feae JA |
590 | index = 6; |
591 | break; | |
8abdce66 | 592 | case 50 ... 99: |
04a0feae JA |
593 | index = 5; |
594 | break; | |
8abdce66 | 595 | case 20 ... 49: |
04a0feae JA |
596 | index = 4; |
597 | break; | |
8abdce66 | 598 | case 10 ... 19: |
04a0feae JA |
599 | index = 3; |
600 | break; | |
8abdce66 | 601 | case 4 ... 9: |
04a0feae JA |
602 | index = 2; |
603 | break; | |
ec118304 | 604 | case 2 ... 3: |
04a0feae | 605 | index = 1; |
ec118304 JA |
606 | case 0 ... 1: |
607 | break; | |
608 | } | |
609 | ||
04a0feae JA |
610 | assert(index < FIO_IO_U_LAT_M_NR); |
611 | td->ts.io_u_lat_m[index]++; | |
612 | } | |
613 | ||
614 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
615 | { | |
616 | if (usec < 1000) | |
617 | io_u_mark_lat_usec(td, usec); | |
618 | else | |
619 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
620 | } |
621 | ||
0aabe160 JA |
622 | /* |
623 | * Get next file to service by choosing one at random | |
624 | */ | |
d6aed795 JA |
625 | static struct fio_file *get_next_file_rand(struct thread_data *td, enum fio_file_flags goodf, |
626 | enum fio_file_flags badf) | |
0aabe160 | 627 | { |
0aabe160 | 628 | struct fio_file *f; |
1c178180 | 629 | int fno; |
0aabe160 JA |
630 | |
631 | do { | |
7c83c089 | 632 | long r = os_random_long(&td->next_file_state); |
87b10676 | 633 | int opened = 0; |
7c83c089 | 634 | |
5ec10eaa | 635 | fno = (unsigned int) ((double) td->o.nr_files |
dc873b6f | 636 | * (r / (OS_RAND_MAX + 1.0))); |
126d65c6 | 637 | f = td->files[fno]; |
d6aed795 | 638 | if (fio_file_done(f)) |
059e63c0 | 639 | continue; |
1c178180 | 640 | |
d6aed795 | 641 | if (!fio_file_open(f)) { |
87b10676 JA |
642 | int err; |
643 | ||
644 | err = td_io_open_file(td, f); | |
645 | if (err) | |
646 | continue; | |
647 | opened = 1; | |
648 | } | |
649 | ||
2ba1c290 JA |
650 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
651 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 652 | return f; |
2ba1c290 | 653 | } |
87b10676 JA |
654 | if (opened) |
655 | td_io_close_file(td, f); | |
0aabe160 JA |
656 | } while (1); |
657 | } | |
658 | ||
659 | /* | |
660 | * Get next file to service by doing round robin between all available ones | |
661 | */ | |
1c178180 JA |
662 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
663 | int badf) | |
3d7c391d JA |
664 | { |
665 | unsigned int old_next_file = td->next_file; | |
666 | struct fio_file *f; | |
667 | ||
668 | do { | |
87b10676 JA |
669 | int opened = 0; |
670 | ||
126d65c6 | 671 | f = td->files[td->next_file]; |
3d7c391d JA |
672 | |
673 | td->next_file++; | |
2dc1bbeb | 674 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
675 | td->next_file = 0; |
676 | ||
87b10676 | 677 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 678 | if (fio_file_done(f)) { |
d5ed68ea | 679 | f = NULL; |
059e63c0 | 680 | continue; |
d5ed68ea | 681 | } |
059e63c0 | 682 | |
d6aed795 | 683 | if (!fio_file_open(f)) { |
87b10676 JA |
684 | int err; |
685 | ||
686 | err = td_io_open_file(td, f); | |
b5696bfc JA |
687 | if (err) { |
688 | dprint(FD_FILE, "error %d on open of %s\n", | |
689 | err, f->file_name); | |
87c27b45 | 690 | f = NULL; |
87b10676 | 691 | continue; |
b5696bfc | 692 | } |
87b10676 JA |
693 | opened = 1; |
694 | } | |
695 | ||
b5696bfc | 696 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags); |
1c178180 | 697 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
698 | break; |
699 | ||
87b10676 JA |
700 | if (opened) |
701 | td_io_close_file(td, f); | |
702 | ||
3d7c391d JA |
703 | f = NULL; |
704 | } while (td->next_file != old_next_file); | |
705 | ||
2ba1c290 | 706 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
707 | return f; |
708 | } | |
709 | ||
bdb4e2e9 JA |
710 | static struct fio_file *get_next_file(struct thread_data *td) |
711 | { | |
1907dbc6 JA |
712 | struct fio_file *f; |
713 | ||
2dc1bbeb | 714 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 715 | |
b5696bfc | 716 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
717 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
718 | " nr_files=%d\n", td->nr_open_files, | |
719 | td->nr_done_files, | |
720 | td->o.nr_files); | |
bdb4e2e9 | 721 | return NULL; |
2ba1c290 | 722 | } |
bdb4e2e9 | 723 | |
1907dbc6 | 724 | f = td->file_service_file; |
d6aed795 | 725 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
726 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
727 | goto out; | |
728 | if (td->file_service_left--) | |
729 | goto out; | |
730 | } | |
1907dbc6 | 731 | |
a086c257 JA |
732 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
733 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 734 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 735 | else |
d6aed795 | 736 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 JA |
737 | |
738 | td->file_service_file = f; | |
739 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 740 | out: |
683023e8 | 741 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 742 | return f; |
bdb4e2e9 JA |
743 | } |
744 | ||
429f6675 JA |
745 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
746 | { | |
747 | struct fio_file *f; | |
748 | ||
749 | do { | |
750 | f = get_next_file(td); | |
751 | if (!f) | |
752 | return 1; | |
753 | ||
429f6675 JA |
754 | io_u->file = f; |
755 | get_file(f); | |
756 | ||
757 | if (!fill_io_u(td, io_u)) | |
758 | break; | |
759 | ||
b5696bfc | 760 | put_file_log(td, f); |
429f6675 | 761 | td_io_close_file(td, f); |
b5696bfc | 762 | io_u->file = NULL; |
d6aed795 | 763 | fio_file_set_done(f); |
429f6675 | 764 | td->nr_done_files++; |
87b10676 | 765 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files); |
429f6675 JA |
766 | } while (1); |
767 | ||
768 | return 0; | |
769 | } | |
770 | ||
771 | ||
10ba535a JA |
772 | struct io_u *__get_io_u(struct thread_data *td) |
773 | { | |
774 | struct io_u *io_u = NULL; | |
775 | ||
01743ee1 JA |
776 | if (!flist_empty(&td->io_u_requeues)) |
777 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 778 | else if (!queue_full(td)) { |
01743ee1 | 779 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 780 | |
6040dabc | 781 | io_u->buflen = 0; |
10ba535a | 782 | io_u->resid = 0; |
755200a3 | 783 | io_u->file = NULL; |
d7762cf8 | 784 | io_u->end_io = NULL; |
755200a3 JA |
785 | } |
786 | ||
787 | if (io_u) { | |
0c6e7517 JA |
788 | assert(io_u->flags & IO_U_F_FREE); |
789 | io_u->flags &= ~IO_U_F_FREE; | |
790 | ||
755200a3 | 791 | io_u->error = 0; |
01743ee1 JA |
792 | flist_del(&io_u->list); |
793 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a JA |
794 | td->cur_depth++; |
795 | } | |
796 | ||
797 | return io_u; | |
798 | } | |
799 | ||
800 | /* | |
801 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
802 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
803 | */ | |
3d7c391d | 804 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 805 | { |
3d7c391d | 806 | struct fio_file *f; |
10ba535a JA |
807 | struct io_u *io_u; |
808 | ||
809 | io_u = __get_io_u(td); | |
2ba1c290 JA |
810 | if (!io_u) { |
811 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 812 | return NULL; |
2ba1c290 | 813 | } |
10ba535a | 814 | |
755200a3 JA |
815 | /* |
816 | * from a requeue, io_u already setup | |
817 | */ | |
818 | if (io_u->file) | |
77f392bf | 819 | goto out; |
755200a3 | 820 | |
429f6675 JA |
821 | /* |
822 | * If using an iolog, grab next piece if any available. | |
823 | */ | |
824 | if (td->o.read_iolog_file) { | |
825 | if (read_iolog_get(td, io_u)) | |
826 | goto err_put; | |
2ba1c290 JA |
827 | } else if (set_io_u_file(td, io_u)) { |
828 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 829 | goto err_put; |
2ba1c290 | 830 | } |
5ec10eaa | 831 | |
429f6675 | 832 | f = io_u->file; |
d6aed795 | 833 | assert(fio_file_open(f)); |
97af62ce | 834 | |
87dc1ab1 | 835 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 836 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 837 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 838 | goto err_put; |
2ba1c290 | 839 | } |
10ba535a | 840 | |
36167d82 | 841 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 842 | |
c311cd2a | 843 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE) |
87dc1ab1 | 844 | populate_verify_io_u(td, io_u); |
e4dad9c6 JA |
845 | else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) |
846 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
87dc1ab1 | 847 | } |
10ba535a | 848 | |
165faf16 JA |
849 | /* |
850 | * Set io data pointers. | |
851 | */ | |
cec6b55d JA |
852 | io_u->xfer_buf = io_u->buf; |
853 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 854 | |
6ac7a331 | 855 | out: |
429f6675 | 856 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
857 | if (!td->o.disable_slat) |
858 | fio_gettime(&io_u->start_time, NULL); | |
429f6675 | 859 | return io_u; |
36167d82 | 860 | } |
429f6675 | 861 | err_put: |
2ba1c290 | 862 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
863 | put_io_u(td, io_u); |
864 | return NULL; | |
10ba535a JA |
865 | } |
866 | ||
5451792e JA |
867 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
868 | { | |
869 | const char *msg[] = { "read", "write", "sync" }; | |
870 | ||
871 | log_err("fio: io_u error"); | |
872 | ||
873 | if (io_u->file) | |
874 | log_err(" on file %s", io_u->file->file_name); | |
875 | ||
876 | log_err(": %s\n", strerror(io_u->error)); | |
877 | ||
5ec10eaa JA |
878 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
879 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
880 | |
881 | if (!td->error) | |
882 | td_verror(td, io_u->error, "io_u error"); | |
883 | } | |
884 | ||
97601024 JA |
885 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
886 | struct io_completion_data *icd) | |
10ba535a | 887 | { |
dbad30b0 JA |
888 | /* |
889 | * Older gcc's are too dumb to realize that usec is always used | |
890 | * initialized, silence that warning. | |
891 | */ | |
892 | unsigned long uninitialized_var(usec); | |
10ba535a | 893 | |
2ba1c290 JA |
894 | dprint_io_u(io_u, "io complete"); |
895 | ||
0c6e7517 JA |
896 | assert(io_u->flags & IO_U_F_FLIGHT); |
897 | io_u->flags &= ~IO_U_F_FLIGHT; | |
898 | ||
87dc1ab1 JA |
899 | if (io_u->ddir == DDIR_SYNC) { |
900 | td->last_was_sync = 1; | |
901 | return; | |
902 | } | |
903 | ||
904 | td->last_was_sync = 0; | |
905 | ||
10ba535a JA |
906 | if (!io_u->error) { |
907 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 908 | const enum fio_ddir idx = io_u->ddir; |
b29ee5b3 | 909 | int ret; |
10ba535a | 910 | |
b29ee5b3 JA |
911 | td->io_blocks[idx]++; |
912 | td->io_bytes[idx] += bytes; | |
913 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 914 | |
b29ee5b3 | 915 | if (ramp_time_over(td)) { |
9520ebb9 JA |
916 | if (!td->o.disable_clat || !td->o.disable_bw) |
917 | usec = utime_since(&io_u->issue_time, | |
918 | &icd->time); | |
721938ae | 919 | |
9520ebb9 | 920 | if (!td->o.disable_clat) { |
306ddc97 | 921 | add_clat_sample(td, idx, usec, bytes); |
9520ebb9 JA |
922 | io_u_mark_latency(td, usec); |
923 | } | |
924 | if (!td->o.disable_bw) | |
306ddc97 | 925 | add_bw_sample(td, idx, bytes, &icd->time); |
721938ae | 926 | } |
10ba535a | 927 | |
660a1cb5 | 928 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 929 | td->o.do_verify && |
41128405 | 930 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
931 | log_io_piece(td, io_u); |
932 | ||
b29ee5b3 | 933 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 934 | |
d7762cf8 | 935 | if (io_u->end_io) { |
36690c9b | 936 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
937 | if (ret && !icd->error) |
938 | icd->error = ret; | |
939 | } | |
5451792e | 940 | } else { |
10ba535a | 941 | icd->error = io_u->error; |
5451792e JA |
942 | io_u_log_error(td, io_u); |
943 | } | |
10ba535a JA |
944 | } |
945 | ||
9520ebb9 JA |
946 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
947 | int nr) | |
10ba535a | 948 | { |
9520ebb9 JA |
949 | if (!td->o.disable_clat || !td->o.disable_bw) |
950 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 951 | |
3af6ef39 JA |
952 | icd->nr = nr; |
953 | ||
10ba535a JA |
954 | icd->error = 0; |
955 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
956 | } |
957 | ||
97601024 JA |
958 | static void ios_completed(struct thread_data *td, |
959 | struct io_completion_data *icd) | |
36167d82 JA |
960 | { |
961 | struct io_u *io_u; | |
962 | int i; | |
963 | ||
10ba535a JA |
964 | for (i = 0; i < icd->nr; i++) { |
965 | io_u = td->io_ops->event(td, i); | |
966 | ||
967 | io_completed(td, io_u, icd); | |
968 | put_io_u(td, io_u); | |
969 | } | |
970 | } | |
97601024 | 971 | |
e7e6cfb4 JA |
972 | /* |
973 | * Complete a single io_u for the sync engines. | |
974 | */ | |
d7762cf8 | 975 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
976 | { |
977 | struct io_completion_data icd; | |
978 | ||
9520ebb9 | 979 | init_icd(td, &icd, 1); |
97601024 JA |
980 | io_completed(td, io_u, &icd); |
981 | put_io_u(td, io_u); | |
982 | ||
983 | if (!icd.error) | |
984 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
985 | ||
37e974a5 | 986 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
987 | return -1; |
988 | } | |
989 | ||
e7e6cfb4 JA |
990 | /* |
991 | * Called to complete min_events number of io for the async engines. | |
992 | */ | |
4950421a | 993 | long io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 994 | { |
97601024 | 995 | struct io_completion_data icd; |
00de55ef | 996 | struct timespec *tvp = NULL; |
97601024 | 997 | int ret; |
4d06a338 | 998 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 999 | |
4950421a | 1000 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1001 | |
4950421a | 1002 | if (!min_evts) |
00de55ef | 1003 | tvp = &ts; |
97601024 | 1004 | |
4950421a | 1005 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1006 | if (ret < 0) { |
e1161c32 | 1007 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1008 | return ret; |
1009 | } else if (!ret) | |
1010 | return ret; | |
1011 | ||
9520ebb9 | 1012 | init_icd(td, &icd, ret); |
97601024 JA |
1013 | ios_completed(td, &icd); |
1014 | if (!icd.error) | |
1015 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
1016 | ||
37e974a5 | 1017 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
1018 | return -1; |
1019 | } | |
7e77dd02 JA |
1020 | |
1021 | /* | |
1022 | * Call when io_u is really queued, to update the submission latency. | |
1023 | */ | |
1024 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1025 | { | |
9520ebb9 JA |
1026 | if (!td->o.disable_slat) { |
1027 | unsigned long slat_time; | |
7e77dd02 | 1028 | |
9520ebb9 | 1029 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
306ddc97 | 1030 | add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time); |
9520ebb9 | 1031 | } |
7e77dd02 | 1032 | } |
433afcb4 | 1033 | |
5973cafb JA |
1034 | /* |
1035 | * "randomly" fill the buffer contents | |
1036 | */ | |
1037 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1038 | unsigned int max_bs) | |
1039 | { | |
1040 | long *ptr = io_u->buf; | |
1041 | ||
1042 | if (!td->o.zero_buffers) { | |
1043 | while ((void *) ptr - io_u->buf < max_bs) { | |
1044 | *ptr = rand() * GOLDEN_RATIO_PRIME; | |
1045 | ptr++; | |
1046 | } | |
1047 | } else | |
1048 | memset(ptr, 0, max_bs); | |
1049 | } |