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 | |
683023e8 JA |
197 | if (get_next_rand_offset(td, f, ddir, &b)) { |
198 | dprint(FD_IO, "%s: getting rand offset failed\n", | |
199 | f->file_name); | |
bca4ed4d | 200 | return 1; |
683023e8 | 201 | } |
43063a1c | 202 | } else { |
0c3d768a | 203 | if (f->last_pos >= f->real_file_size) { |
4ba66134 | 204 | if (!td_random(td) || |
683023e8 JA |
205 | get_next_rand_offset(td, f, ddir, &b)) { |
206 | dprint(FD_IO, "%s: pos %llu > size %llu\n", | |
207 | f->file_name, f->last_pos, | |
208 | f->real_file_size); | |
0c3d768a | 209 | return 1; |
683023e8 | 210 | } |
bcdedd0a | 211 | } else |
4ba66134 | 212 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
43063a1c | 213 | } |
10ba535a | 214 | |
009bd847 JA |
215 | io_u->offset = b * td->o.min_bs[ddir]; |
216 | if (io_u->offset >= f->io_size) { | |
217 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
218 | io_u->offset, f->io_size); | |
219 | return 1; | |
220 | } | |
221 | ||
222 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
223 | if (io_u->offset >= f->real_file_size) { |
224 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
225 | io_u->offset, f->real_file_size); | |
10ba535a | 226 | return 1; |
2ba1c290 | 227 | } |
10ba535a JA |
228 | |
229 | return 0; | |
230 | } | |
231 | ||
f3059de1 JA |
232 | static inline int is_power_of_2(unsigned int val) |
233 | { | |
234 | return (val != 0 && ((val & (val - 1)) == 0)); | |
235 | } | |
236 | ||
9bf2061e | 237 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 238 | { |
bca4ed4d | 239 | const int ddir = io_u->ddir; |
f3f552b9 | 240 | unsigned int uninitialized_var(buflen); |
f3059de1 | 241 | unsigned int minbs, maxbs; |
10ba535a JA |
242 | long r; |
243 | ||
f3059de1 JA |
244 | minbs = td->o.min_bs[ddir]; |
245 | maxbs = td->o.max_bs[ddir]; | |
246 | ||
247 | if (minbs == maxbs) | |
248 | buflen = minbs; | |
10ba535a JA |
249 | else { |
250 | r = os_random_long(&td->bsrange_state); | |
5ec10eaa | 251 | if (!td->o.bssplit_nr) { |
f3059de1 JA |
252 | buflen = 1 + (unsigned int) ((double) maxbs * |
253 | (r / (OS_RAND_MAX + 1.0))); | |
254 | if (buflen < minbs) | |
255 | buflen = minbs; | |
5ec10eaa | 256 | } else { |
564ca972 JA |
257 | long perc = 0; |
258 | unsigned int i; | |
259 | ||
260 | for (i = 0; i < td->o.bssplit_nr; i++) { | |
261 | struct bssplit *bsp = &td->o.bssplit[i]; | |
262 | ||
263 | buflen = bsp->bs; | |
264 | perc += bsp->perc; | |
f3059de1 | 265 | if (r <= ((OS_RAND_MAX / 100L) * perc)) |
564ca972 JA |
266 | break; |
267 | } | |
268 | } | |
f3059de1 JA |
269 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
270 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a JA |
271 | } |
272 | ||
4ba66134 JA |
273 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
274 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, | |
f3059de1 JA |
275 | minbs, ddir); |
276 | buflen = minbs; | |
4ba66134 | 277 | } |
6a5e6884 | 278 | |
10ba535a JA |
279 | return buflen; |
280 | } | |
281 | ||
afe24a5a JA |
282 | static void set_rwmix_bytes(struct thread_data *td) |
283 | { | |
afe24a5a JA |
284 | unsigned int diff; |
285 | ||
286 | /* | |
287 | * we do time or byte based switch. this is needed because | |
288 | * buffered writes may issue a lot quicker than they complete, | |
289 | * whereas reads do not. | |
290 | */ | |
e47f799f | 291 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 292 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
293 | } |
294 | ||
295 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
296 | { | |
297 | unsigned int v; | |
298 | long r; | |
299 | ||
300 | r = os_random_long(&td->rwmix_state); | |
dc873b6f | 301 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); |
04c540d9 | 302 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
303 | return DDIR_READ; |
304 | ||
305 | return DDIR_WRITE; | |
afe24a5a JA |
306 | } |
307 | ||
10ba535a JA |
308 | /* |
309 | * Return the data direction for the next io_u. If the job is a | |
310 | * mixed read/write workload, check the rwmix cycle and switch if | |
311 | * necessary. | |
312 | */ | |
1e97cce9 | 313 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
314 | { |
315 | if (td_rw(td)) { | |
10ba535a JA |
316 | /* |
317 | * Check if it's time to seed a new data direction. | |
318 | */ | |
e4928662 | 319 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f | 320 | unsigned long long max_bytes; |
5ec10eaa | 321 | enum fio_ddir ddir; |
e47f799f JA |
322 | |
323 | /* | |
324 | * Put a top limit on how many bytes we do for | |
325 | * one data direction, to avoid overflowing the | |
326 | * ranges too much | |
327 | */ | |
328 | ddir = get_rand_ddir(td); | |
329 | max_bytes = td->this_io_bytes[ddir]; | |
5ec10eaa JA |
330 | if (max_bytes >= |
331 | (td->o.size * td->o.rwmix[ddir] / 100)) { | |
6162f524 | 332 | if (!td->rw_end_set[ddir]) { |
38d77cae | 333 | td->rw_end_set[ddir] = 1; |
6162f524 JA |
334 | fio_gettime(&td->rw_end[ddir], NULL); |
335 | } | |
e4928662 | 336 | |
e47f799f | 337 | ddir ^= 1; |
38d77cae | 338 | } |
e47f799f JA |
339 | |
340 | if (ddir != td->rwmix_ddir) | |
341 | set_rwmix_bytes(td); | |
342 | ||
343 | td->rwmix_ddir = ddir; | |
10ba535a JA |
344 | } |
345 | return td->rwmix_ddir; | |
346 | } else if (td_read(td)) | |
347 | return DDIR_READ; | |
348 | else | |
349 | return DDIR_WRITE; | |
350 | } | |
351 | ||
60f2c658 JA |
352 | static void put_file_log(struct thread_data *td, struct fio_file *f) |
353 | { | |
354 | int ret = put_file(td, f); | |
355 | ||
356 | if (ret) | |
357 | td_verror(td, ret, "file close"); | |
358 | } | |
359 | ||
10ba535a JA |
360 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
361 | { | |
0c6e7517 JA |
362 | assert((io_u->flags & IO_U_F_FREE) == 0); |
363 | io_u->flags |= IO_U_F_FREE; | |
364 | ||
60f2c658 JA |
365 | if (io_u->file) |
366 | put_file_log(td, io_u->file); | |
2dbdab7e | 367 | |
10ba535a | 368 | io_u->file = NULL; |
01743ee1 JA |
369 | flist_del(&io_u->list); |
370 | flist_add(&io_u->list, &td->io_u_freelist); | |
10ba535a JA |
371 | td->cur_depth--; |
372 | } | |
373 | ||
755200a3 JA |
374 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
375 | { | |
376 | struct io_u *__io_u = *io_u; | |
377 | ||
465221b0 JA |
378 | dprint(FD_IO, "requeue %p\n", __io_u); |
379 | ||
4d2e0f49 | 380 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
381 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
382 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 383 | |
4d2e0f49 JA |
384 | __io_u->flags &= ~IO_U_F_FLIGHT; |
385 | ||
01743ee1 JA |
386 | flist_del(&__io_u->list); |
387 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
755200a3 JA |
388 | td->cur_depth--; |
389 | *io_u = NULL; | |
390 | } | |
391 | ||
9bf2061e | 392 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 393 | { |
b4c5e1ac JA |
394 | if (td->io_ops->flags & FIO_NOIO) |
395 | goto out; | |
396 | ||
87dc1ab1 JA |
397 | /* |
398 | * see if it's time to sync | |
399 | */ | |
2dc1bbeb JA |
400 | if (td->o.fsync_blocks && |
401 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
402 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 403 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 404 | goto out; |
87dc1ab1 JA |
405 | } |
406 | ||
a00735e6 JA |
407 | io_u->ddir = get_rw_ddir(td); |
408 | ||
48f5abd3 JA |
409 | /* |
410 | * See if it's time to switch to a new zone | |
411 | */ | |
412 | if (td->zone_bytes >= td->o.zone_size) { | |
413 | td->zone_bytes = 0; | |
414 | io_u->file->last_pos += td->o.zone_skip; | |
415 | td->io_skip_bytes += td->o.zone_skip; | |
416 | } | |
417 | ||
10ba535a | 418 | /* |
c685b5b2 JA |
419 | * No log, let the seq/rand engine retrieve the next buflen and |
420 | * position. | |
10ba535a | 421 | */ |
2ba1c290 JA |
422 | if (get_next_offset(td, io_u)) { |
423 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 424 | return 1; |
2ba1c290 | 425 | } |
10ba535a | 426 | |
9bf2061e | 427 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
428 | if (!io_u->buflen) { |
429 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 430 | return 1; |
2ba1c290 | 431 | } |
bca4ed4d | 432 | |
2ba1c290 JA |
433 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
434 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
435 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
436 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 437 | return 1; |
2ba1c290 | 438 | } |
6a5e6884 | 439 | |
bca4ed4d JA |
440 | /* |
441 | * mark entry before potentially trimming io_u | |
442 | */ | |
303032ae | 443 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 444 | mark_random_map(td, io_u); |
bca4ed4d JA |
445 | |
446 | /* | |
447 | * If using a write iolog, store this entry. | |
448 | */ | |
c38e9468 | 449 | out: |
2ba1c290 | 450 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 451 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 452 | log_io_u(td, io_u); |
bca4ed4d | 453 | return 0; |
10ba535a JA |
454 | } |
455 | ||
838bc709 JA |
456 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
457 | { | |
458 | int index = 0; | |
459 | ||
460 | switch (nr) { | |
461 | default: | |
462 | index = 6; | |
463 | break; | |
464 | case 33 ... 64: | |
465 | index = 5; | |
466 | break; | |
467 | case 17 ... 32: | |
468 | index = 4; | |
469 | break; | |
470 | case 9 ... 16: | |
471 | index = 3; | |
472 | break; | |
473 | case 5 ... 8: | |
474 | index = 2; | |
475 | break; | |
476 | case 1 ... 4: | |
477 | index = 1; | |
478 | case 0: | |
479 | break; | |
480 | } | |
481 | ||
482 | map[index]++; | |
483 | } | |
484 | ||
485 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
486 | { | |
487 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
488 | td->ts.total_submit++; | |
489 | } | |
490 | ||
491 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
492 | { | |
493 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
494 | td->ts.total_complete++; | |
495 | } | |
496 | ||
d8005759 | 497 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 JA |
498 | { |
499 | int index = 0; | |
500 | ||
501 | switch (td->cur_depth) { | |
502 | default: | |
a783e61a JA |
503 | index = 6; |
504 | break; | |
71619dc2 | 505 | case 32 ... 63: |
a783e61a JA |
506 | index = 5; |
507 | break; | |
71619dc2 | 508 | case 16 ... 31: |
a783e61a JA |
509 | index = 4; |
510 | break; | |
71619dc2 | 511 | case 8 ... 15: |
a783e61a JA |
512 | index = 3; |
513 | break; | |
71619dc2 | 514 | case 4 ... 7: |
a783e61a JA |
515 | index = 2; |
516 | break; | |
71619dc2 | 517 | case 2 ... 3: |
a783e61a | 518 | index = 1; |
71619dc2 JA |
519 | case 1: |
520 | break; | |
521 | } | |
522 | ||
3bec7ae0 | 523 | td->ts.io_u_map[index] += nr; |
71619dc2 JA |
524 | } |
525 | ||
04a0feae JA |
526 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
527 | { | |
528 | int index = 0; | |
529 | ||
530 | assert(usec < 1000); | |
531 | ||
532 | switch (usec) { | |
533 | case 750 ... 999: | |
534 | index = 9; | |
535 | break; | |
536 | case 500 ... 749: | |
537 | index = 8; | |
538 | break; | |
539 | case 250 ... 499: | |
540 | index = 7; | |
541 | break; | |
542 | case 100 ... 249: | |
543 | index = 6; | |
544 | break; | |
545 | case 50 ... 99: | |
546 | index = 5; | |
547 | break; | |
548 | case 20 ... 49: | |
549 | index = 4; | |
550 | break; | |
551 | case 10 ... 19: | |
552 | index = 3; | |
553 | break; | |
554 | case 4 ... 9: | |
555 | index = 2; | |
556 | break; | |
557 | case 2 ... 3: | |
558 | index = 1; | |
559 | case 0 ... 1: | |
560 | break; | |
561 | } | |
562 | ||
563 | assert(index < FIO_IO_U_LAT_U_NR); | |
564 | td->ts.io_u_lat_u[index]++; | |
565 | } | |
566 | ||
567 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
568 | { |
569 | int index = 0; | |
570 | ||
571 | switch (msec) { | |
572 | default: | |
04a0feae JA |
573 | index = 11; |
574 | break; | |
8abdce66 | 575 | case 1000 ... 1999: |
04a0feae JA |
576 | index = 10; |
577 | break; | |
8abdce66 | 578 | case 750 ... 999: |
04a0feae JA |
579 | index = 9; |
580 | break; | |
8abdce66 | 581 | case 500 ... 749: |
04a0feae JA |
582 | index = 8; |
583 | break; | |
8abdce66 | 584 | case 250 ... 499: |
04a0feae JA |
585 | index = 7; |
586 | break; | |
8abdce66 | 587 | case 100 ... 249: |
04a0feae JA |
588 | index = 6; |
589 | break; | |
8abdce66 | 590 | case 50 ... 99: |
04a0feae JA |
591 | index = 5; |
592 | break; | |
8abdce66 | 593 | case 20 ... 49: |
04a0feae JA |
594 | index = 4; |
595 | break; | |
8abdce66 | 596 | case 10 ... 19: |
04a0feae JA |
597 | index = 3; |
598 | break; | |
8abdce66 | 599 | case 4 ... 9: |
04a0feae JA |
600 | index = 2; |
601 | break; | |
ec118304 | 602 | case 2 ... 3: |
04a0feae | 603 | index = 1; |
ec118304 JA |
604 | case 0 ... 1: |
605 | break; | |
606 | } | |
607 | ||
04a0feae JA |
608 | assert(index < FIO_IO_U_LAT_M_NR); |
609 | td->ts.io_u_lat_m[index]++; | |
610 | } | |
611 | ||
612 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
613 | { | |
614 | if (usec < 1000) | |
615 | io_u_mark_lat_usec(td, usec); | |
616 | else | |
617 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
618 | } |
619 | ||
0aabe160 JA |
620 | /* |
621 | * Get next file to service by choosing one at random | |
622 | */ | |
1c178180 JA |
623 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
624 | int badf) | |
0aabe160 | 625 | { |
0aabe160 | 626 | struct fio_file *f; |
1c178180 | 627 | int fno; |
0aabe160 JA |
628 | |
629 | do { | |
7c83c089 | 630 | long r = os_random_long(&td->next_file_state); |
87b10676 | 631 | int opened = 0; |
7c83c089 | 632 | |
5ec10eaa | 633 | fno = (unsigned int) ((double) td->o.nr_files |
dc873b6f | 634 | * (r / (OS_RAND_MAX + 1.0))); |
126d65c6 | 635 | f = td->files[fno]; |
059e63c0 JA |
636 | if (f->flags & FIO_FILE_DONE) |
637 | continue; | |
1c178180 | 638 | |
87b10676 JA |
639 | if (!(f->flags & FIO_FILE_OPEN)) { |
640 | int err; | |
641 | ||
642 | err = td_io_open_file(td, f); | |
643 | if (err) | |
644 | continue; | |
645 | opened = 1; | |
646 | } | |
647 | ||
2ba1c290 JA |
648 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
649 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 650 | return f; |
2ba1c290 | 651 | } |
87b10676 JA |
652 | if (opened) |
653 | td_io_close_file(td, f); | |
0aabe160 JA |
654 | } while (1); |
655 | } | |
656 | ||
657 | /* | |
658 | * Get next file to service by doing round robin between all available ones | |
659 | */ | |
1c178180 JA |
660 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
661 | int badf) | |
3d7c391d JA |
662 | { |
663 | unsigned int old_next_file = td->next_file; | |
664 | struct fio_file *f; | |
665 | ||
666 | do { | |
87b10676 JA |
667 | int opened = 0; |
668 | ||
126d65c6 | 669 | f = td->files[td->next_file]; |
3d7c391d JA |
670 | |
671 | td->next_file++; | |
2dc1bbeb | 672 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
673 | td->next_file = 0; |
674 | ||
87b10676 | 675 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d5ed68ea JA |
676 | if (f->flags & FIO_FILE_DONE) { |
677 | f = NULL; | |
059e63c0 | 678 | continue; |
d5ed68ea | 679 | } |
059e63c0 | 680 | |
87b10676 JA |
681 | if (!(f->flags & FIO_FILE_OPEN)) { |
682 | int err; | |
683 | ||
684 | err = td_io_open_file(td, f); | |
b5696bfc JA |
685 | if (err) { |
686 | dprint(FD_FILE, "error %d on open of %s\n", | |
687 | err, f->file_name); | |
87b10676 | 688 | continue; |
b5696bfc | 689 | } |
87b10676 JA |
690 | opened = 1; |
691 | } | |
692 | ||
b5696bfc | 693 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags); |
1c178180 | 694 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
695 | break; |
696 | ||
87b10676 JA |
697 | if (opened) |
698 | td_io_close_file(td, f); | |
699 | ||
3d7c391d JA |
700 | f = NULL; |
701 | } while (td->next_file != old_next_file); | |
702 | ||
2ba1c290 | 703 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
704 | return f; |
705 | } | |
706 | ||
bdb4e2e9 JA |
707 | static struct fio_file *get_next_file(struct thread_data *td) |
708 | { | |
1907dbc6 JA |
709 | struct fio_file *f; |
710 | ||
2dc1bbeb | 711 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 712 | |
b5696bfc | 713 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
714 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
715 | " nr_files=%d\n", td->nr_open_files, | |
716 | td->nr_done_files, | |
717 | td->o.nr_files); | |
bdb4e2e9 | 718 | return NULL; |
2ba1c290 | 719 | } |
bdb4e2e9 | 720 | |
1907dbc6 | 721 | f = td->file_service_file; |
683023e8 | 722 | if (f && (f->flags & FIO_FILE_OPEN) && !(f->flags & FIO_FILE_CLOSING)) { |
a086c257 JA |
723 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
724 | goto out; | |
725 | if (td->file_service_left--) | |
726 | goto out; | |
727 | } | |
1907dbc6 | 728 | |
a086c257 JA |
729 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
730 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
1c178180 | 731 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
bdb4e2e9 | 732 | else |
1c178180 | 733 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
1907dbc6 JA |
734 | |
735 | td->file_service_file = f; | |
736 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 737 | out: |
683023e8 | 738 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 739 | return f; |
bdb4e2e9 JA |
740 | } |
741 | ||
429f6675 JA |
742 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
743 | { | |
744 | struct fio_file *f; | |
745 | ||
746 | do { | |
747 | f = get_next_file(td); | |
748 | if (!f) | |
749 | return 1; | |
750 | ||
751 | set_file: | |
752 | io_u->file = f; | |
753 | get_file(f); | |
754 | ||
755 | if (!fill_io_u(td, io_u)) | |
756 | break; | |
757 | ||
009bd847 JA |
758 | /* |
759 | * optimization to prevent close/open of the same file. This | |
760 | * way we preserve queueing etc. | |
761 | */ | |
762 | if (td->o.nr_files == 1 && td->o.time_based) { | |
60f2c658 | 763 | put_file_log(td, f); |
009bd847 JA |
764 | fio_file_reset(f); |
765 | goto set_file; | |
766 | } | |
767 | ||
b5696bfc | 768 | put_file_log(td, f); |
429f6675 | 769 | td_io_close_file(td, f); |
b5696bfc | 770 | io_u->file = NULL; |
429f6675 JA |
771 | f->flags |= FIO_FILE_DONE; |
772 | td->nr_done_files++; | |
87b10676 | 773 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files); |
429f6675 JA |
774 | } while (1); |
775 | ||
776 | return 0; | |
777 | } | |
778 | ||
779 | ||
10ba535a JA |
780 | struct io_u *__get_io_u(struct thread_data *td) |
781 | { | |
782 | struct io_u *io_u = NULL; | |
783 | ||
01743ee1 JA |
784 | if (!flist_empty(&td->io_u_requeues)) |
785 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 786 | else if (!queue_full(td)) { |
01743ee1 | 787 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 788 | |
6040dabc | 789 | io_u->buflen = 0; |
10ba535a | 790 | io_u->resid = 0; |
755200a3 | 791 | io_u->file = NULL; |
d7762cf8 | 792 | io_u->end_io = NULL; |
755200a3 JA |
793 | } |
794 | ||
795 | if (io_u) { | |
0c6e7517 JA |
796 | assert(io_u->flags & IO_U_F_FREE); |
797 | io_u->flags &= ~IO_U_F_FREE; | |
798 | ||
755200a3 | 799 | io_u->error = 0; |
01743ee1 JA |
800 | flist_del(&io_u->list); |
801 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a JA |
802 | td->cur_depth++; |
803 | } | |
804 | ||
805 | return io_u; | |
806 | } | |
807 | ||
808 | /* | |
809 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
810 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
811 | */ | |
3d7c391d | 812 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 813 | { |
3d7c391d | 814 | struct fio_file *f; |
10ba535a JA |
815 | struct io_u *io_u; |
816 | ||
817 | io_u = __get_io_u(td); | |
2ba1c290 JA |
818 | if (!io_u) { |
819 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 820 | return NULL; |
2ba1c290 | 821 | } |
10ba535a | 822 | |
755200a3 JA |
823 | /* |
824 | * from a requeue, io_u already setup | |
825 | */ | |
826 | if (io_u->file) | |
77f392bf | 827 | goto out; |
755200a3 | 828 | |
429f6675 JA |
829 | /* |
830 | * If using an iolog, grab next piece if any available. | |
831 | */ | |
832 | if (td->o.read_iolog_file) { | |
833 | if (read_iolog_get(td, io_u)) | |
834 | goto err_put; | |
2ba1c290 JA |
835 | } else if (set_io_u_file(td, io_u)) { |
836 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 837 | goto err_put; |
2ba1c290 | 838 | } |
5ec10eaa | 839 | |
429f6675 JA |
840 | f = io_u->file; |
841 | assert(f->flags & FIO_FILE_OPEN); | |
97af62ce | 842 | |
87dc1ab1 | 843 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 844 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 845 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 846 | goto err_put; |
2ba1c290 | 847 | } |
10ba535a | 848 | |
36167d82 | 849 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 850 | |
2dc1bbeb | 851 | if (td->o.verify != VERIFY_NONE) |
87dc1ab1 | 852 | populate_verify_io_u(td, io_u); |
e4dad9c6 JA |
853 | else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) |
854 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
87dc1ab1 | 855 | } |
10ba535a | 856 | |
165faf16 JA |
857 | /* |
858 | * Set io data pointers. | |
859 | */ | |
d460eb31 | 860 | io_u->endpos = io_u->offset + io_u->buflen; |
cec6b55d JA |
861 | io_u->xfer_buf = io_u->buf; |
862 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 863 | |
6ac7a331 | 864 | out: |
429f6675 | 865 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
866 | if (!td->o.disable_slat) |
867 | fio_gettime(&io_u->start_time, NULL); | |
429f6675 | 868 | return io_u; |
36167d82 | 869 | } |
429f6675 | 870 | err_put: |
2ba1c290 | 871 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
872 | put_io_u(td, io_u); |
873 | return NULL; | |
10ba535a JA |
874 | } |
875 | ||
5451792e JA |
876 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
877 | { | |
878 | const char *msg[] = { "read", "write", "sync" }; | |
879 | ||
880 | log_err("fio: io_u error"); | |
881 | ||
882 | if (io_u->file) | |
883 | log_err(" on file %s", io_u->file->file_name); | |
884 | ||
885 | log_err(": %s\n", strerror(io_u->error)); | |
886 | ||
5ec10eaa JA |
887 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
888 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
889 | |
890 | if (!td->error) | |
891 | td_verror(td, io_u->error, "io_u error"); | |
892 | } | |
893 | ||
97601024 JA |
894 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
895 | struct io_completion_data *icd) | |
10ba535a | 896 | { |
dbad30b0 JA |
897 | /* |
898 | * Older gcc's are too dumb to realize that usec is always used | |
899 | * initialized, silence that warning. | |
900 | */ | |
901 | unsigned long uninitialized_var(usec); | |
10ba535a | 902 | |
2ba1c290 JA |
903 | dprint_io_u(io_u, "io complete"); |
904 | ||
0c6e7517 JA |
905 | assert(io_u->flags & IO_U_F_FLIGHT); |
906 | io_u->flags &= ~IO_U_F_FLIGHT; | |
907 | ||
87dc1ab1 JA |
908 | if (io_u->ddir == DDIR_SYNC) { |
909 | td->last_was_sync = 1; | |
910 | return; | |
911 | } | |
912 | ||
913 | td->last_was_sync = 0; | |
914 | ||
10ba535a JA |
915 | if (!io_u->error) { |
916 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 917 | const enum fio_ddir idx = io_u->ddir; |
b29ee5b3 | 918 | int ret; |
10ba535a | 919 | |
b29ee5b3 JA |
920 | td->io_blocks[idx]++; |
921 | td->io_bytes[idx] += bytes; | |
922 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 923 | |
b29ee5b3 | 924 | if (ramp_time_over(td)) { |
9520ebb9 JA |
925 | if (!td->o.disable_clat || !td->o.disable_bw) |
926 | usec = utime_since(&io_u->issue_time, | |
927 | &icd->time); | |
721938ae | 928 | |
9520ebb9 JA |
929 | if (!td->o.disable_clat) { |
930 | add_clat_sample(td, idx, usec); | |
931 | io_u_mark_latency(td, usec); | |
932 | } | |
933 | if (!td->o.disable_bw) | |
934 | add_bw_sample(td, idx, &icd->time); | |
721938ae | 935 | } |
10ba535a | 936 | |
660a1cb5 | 937 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 938 | td->o.do_verify && |
41128405 | 939 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
940 | log_io_piece(td, io_u); |
941 | ||
b29ee5b3 | 942 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 943 | |
d7762cf8 | 944 | if (io_u->end_io) { |
36690c9b | 945 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
946 | if (ret && !icd->error) |
947 | icd->error = ret; | |
948 | } | |
5451792e | 949 | } else { |
10ba535a | 950 | icd->error = io_u->error; |
5451792e JA |
951 | io_u_log_error(td, io_u); |
952 | } | |
10ba535a JA |
953 | } |
954 | ||
9520ebb9 JA |
955 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
956 | int nr) | |
10ba535a | 957 | { |
9520ebb9 JA |
958 | if (!td->o.disable_clat || !td->o.disable_bw) |
959 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 960 | |
3af6ef39 JA |
961 | icd->nr = nr; |
962 | ||
10ba535a JA |
963 | icd->error = 0; |
964 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
965 | } |
966 | ||
97601024 JA |
967 | static void ios_completed(struct thread_data *td, |
968 | struct io_completion_data *icd) | |
36167d82 JA |
969 | { |
970 | struct io_u *io_u; | |
971 | int i; | |
972 | ||
10ba535a JA |
973 | for (i = 0; i < icd->nr; i++) { |
974 | io_u = td->io_ops->event(td, i); | |
975 | ||
976 | io_completed(td, io_u, icd); | |
977 | put_io_u(td, io_u); | |
978 | } | |
979 | } | |
97601024 | 980 | |
e7e6cfb4 JA |
981 | /* |
982 | * Complete a single io_u for the sync engines. | |
983 | */ | |
d7762cf8 | 984 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
985 | { |
986 | struct io_completion_data icd; | |
987 | ||
9520ebb9 | 988 | init_icd(td, &icd, 1); |
97601024 JA |
989 | io_completed(td, io_u, &icd); |
990 | put_io_u(td, io_u); | |
991 | ||
992 | if (!icd.error) | |
993 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
994 | ||
37e974a5 | 995 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
996 | return -1; |
997 | } | |
998 | ||
e7e6cfb4 JA |
999 | /* |
1000 | * Called to complete min_events number of io for the async engines. | |
1001 | */ | |
4950421a | 1002 | long io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 1003 | { |
97601024 | 1004 | struct io_completion_data icd; |
00de55ef | 1005 | struct timespec *tvp = NULL; |
97601024 | 1006 | int ret; |
4d06a338 | 1007 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1008 | |
4950421a | 1009 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1010 | |
4950421a | 1011 | if (!min_evts) |
00de55ef | 1012 | tvp = &ts; |
97601024 | 1013 | |
4950421a | 1014 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1015 | if (ret < 0) { |
e1161c32 | 1016 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1017 | return ret; |
1018 | } else if (!ret) | |
1019 | return ret; | |
1020 | ||
9520ebb9 | 1021 | init_icd(td, &icd, ret); |
97601024 JA |
1022 | ios_completed(td, &icd); |
1023 | if (!icd.error) | |
1024 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
1025 | ||
37e974a5 | 1026 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
1027 | return -1; |
1028 | } | |
7e77dd02 JA |
1029 | |
1030 | /* | |
1031 | * Call when io_u is really queued, to update the submission latency. | |
1032 | */ | |
1033 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1034 | { | |
9520ebb9 JA |
1035 | if (!td->o.disable_slat) { |
1036 | unsigned long slat_time; | |
7e77dd02 | 1037 | |
9520ebb9 JA |
1038 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
1039 | add_slat_sample(td, io_u->ddir, slat_time); | |
1040 | } | |
7e77dd02 | 1041 | } |
433afcb4 | 1042 | |
5973cafb JA |
1043 | /* |
1044 | * "randomly" fill the buffer contents | |
1045 | */ | |
1046 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1047 | unsigned int max_bs) | |
1048 | { | |
1049 | long *ptr = io_u->buf; | |
1050 | ||
1051 | if (!td->o.zero_buffers) { | |
1052 | while ((void *) ptr - io_u->buf < max_bs) { | |
1053 | *ptr = rand() * GOLDEN_RATIO_PRIME; | |
1054 | ptr++; | |
1055 | } | |
1056 | } else | |
1057 | memset(ptr, 0, max_bs); | |
1058 | } |