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) { |
5ec10eaa | 322 | enum fio_ddir ddir; |
e47f799f JA |
323 | |
324 | /* | |
325 | * Put a top limit on how many bytes we do for | |
326 | * one data direction, to avoid overflowing the | |
327 | * ranges too much | |
328 | */ | |
329 | ddir = get_rand_ddir(td); | |
e47f799f JA |
330 | |
331 | if (ddir != td->rwmix_ddir) | |
332 | set_rwmix_bytes(td); | |
333 | ||
334 | td->rwmix_ddir = ddir; | |
10ba535a JA |
335 | } |
336 | return td->rwmix_ddir; | |
337 | } else if (td_read(td)) | |
338 | return DDIR_READ; | |
339 | else | |
340 | return DDIR_WRITE; | |
341 | } | |
342 | ||
60f2c658 JA |
343 | static void put_file_log(struct thread_data *td, struct fio_file *f) |
344 | { | |
345 | int ret = put_file(td, f); | |
346 | ||
347 | if (ret) | |
348 | td_verror(td, ret, "file close"); | |
349 | } | |
350 | ||
10ba535a JA |
351 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
352 | { | |
0c6e7517 JA |
353 | assert((io_u->flags & IO_U_F_FREE) == 0); |
354 | io_u->flags |= IO_U_F_FREE; | |
355 | ||
60f2c658 JA |
356 | if (io_u->file) |
357 | put_file_log(td, io_u->file); | |
2dbdab7e | 358 | |
10ba535a | 359 | io_u->file = NULL; |
01743ee1 JA |
360 | flist_del(&io_u->list); |
361 | flist_add(&io_u->list, &td->io_u_freelist); | |
10ba535a JA |
362 | td->cur_depth--; |
363 | } | |
364 | ||
755200a3 JA |
365 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
366 | { | |
367 | struct io_u *__io_u = *io_u; | |
368 | ||
465221b0 JA |
369 | dprint(FD_IO, "requeue %p\n", __io_u); |
370 | ||
4d2e0f49 | 371 | __io_u->flags |= IO_U_F_FREE; |
e4f54adb JA |
372 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
373 | td->io_issues[__io_u->ddir]--; | |
5ec10eaa | 374 | |
4d2e0f49 JA |
375 | __io_u->flags &= ~IO_U_F_FLIGHT; |
376 | ||
01743ee1 JA |
377 | flist_del(&__io_u->list); |
378 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
755200a3 JA |
379 | td->cur_depth--; |
380 | *io_u = NULL; | |
381 | } | |
382 | ||
9bf2061e | 383 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 384 | { |
b4c5e1ac JA |
385 | if (td->io_ops->flags & FIO_NOIO) |
386 | goto out; | |
387 | ||
87dc1ab1 JA |
388 | /* |
389 | * see if it's time to sync | |
390 | */ | |
2dc1bbeb JA |
391 | if (td->o.fsync_blocks && |
392 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
393 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 | 394 | io_u->ddir = DDIR_SYNC; |
c38e9468 | 395 | goto out; |
87dc1ab1 JA |
396 | } |
397 | ||
a00735e6 JA |
398 | io_u->ddir = get_rw_ddir(td); |
399 | ||
48f5abd3 JA |
400 | /* |
401 | * See if it's time to switch to a new zone | |
402 | */ | |
403 | if (td->zone_bytes >= td->o.zone_size) { | |
404 | td->zone_bytes = 0; | |
405 | io_u->file->last_pos += td->o.zone_skip; | |
406 | td->io_skip_bytes += td->o.zone_skip; | |
407 | } | |
408 | ||
10ba535a | 409 | /* |
c685b5b2 JA |
410 | * No log, let the seq/rand engine retrieve the next buflen and |
411 | * position. | |
10ba535a | 412 | */ |
2ba1c290 JA |
413 | if (get_next_offset(td, io_u)) { |
414 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 415 | return 1; |
2ba1c290 | 416 | } |
10ba535a | 417 | |
9bf2061e | 418 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
419 | if (!io_u->buflen) { |
420 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 421 | return 1; |
2ba1c290 | 422 | } |
bca4ed4d | 423 | |
2ba1c290 JA |
424 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
425 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
426 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
427 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 428 | return 1; |
2ba1c290 | 429 | } |
6a5e6884 | 430 | |
bca4ed4d JA |
431 | /* |
432 | * mark entry before potentially trimming io_u | |
433 | */ | |
303032ae | 434 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 435 | mark_random_map(td, io_u); |
bca4ed4d JA |
436 | |
437 | /* | |
438 | * If using a write iolog, store this entry. | |
439 | */ | |
c38e9468 | 440 | out: |
2ba1c290 | 441 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 442 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 443 | log_io_u(td, io_u); |
bca4ed4d | 444 | return 0; |
10ba535a JA |
445 | } |
446 | ||
838bc709 JA |
447 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
448 | { | |
449 | int index = 0; | |
450 | ||
451 | switch (nr) { | |
452 | default: | |
453 | index = 6; | |
454 | break; | |
455 | case 33 ... 64: | |
456 | index = 5; | |
457 | break; | |
458 | case 17 ... 32: | |
459 | index = 4; | |
460 | break; | |
461 | case 9 ... 16: | |
462 | index = 3; | |
463 | break; | |
464 | case 5 ... 8: | |
465 | index = 2; | |
466 | break; | |
467 | case 1 ... 4: | |
468 | index = 1; | |
469 | case 0: | |
470 | break; | |
471 | } | |
472 | ||
473 | map[index]++; | |
474 | } | |
475 | ||
476 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
477 | { | |
478 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
479 | td->ts.total_submit++; | |
480 | } | |
481 | ||
482 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
483 | { | |
484 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
485 | td->ts.total_complete++; | |
486 | } | |
487 | ||
d8005759 | 488 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 JA |
489 | { |
490 | int index = 0; | |
491 | ||
492 | switch (td->cur_depth) { | |
493 | default: | |
a783e61a JA |
494 | index = 6; |
495 | break; | |
71619dc2 | 496 | case 32 ... 63: |
a783e61a JA |
497 | index = 5; |
498 | break; | |
71619dc2 | 499 | case 16 ... 31: |
a783e61a JA |
500 | index = 4; |
501 | break; | |
71619dc2 | 502 | case 8 ... 15: |
a783e61a JA |
503 | index = 3; |
504 | break; | |
71619dc2 | 505 | case 4 ... 7: |
a783e61a JA |
506 | index = 2; |
507 | break; | |
71619dc2 | 508 | case 2 ... 3: |
a783e61a | 509 | index = 1; |
71619dc2 JA |
510 | case 1: |
511 | break; | |
512 | } | |
513 | ||
3bec7ae0 | 514 | td->ts.io_u_map[index] += nr; |
71619dc2 JA |
515 | } |
516 | ||
04a0feae JA |
517 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
518 | { | |
519 | int index = 0; | |
520 | ||
521 | assert(usec < 1000); | |
522 | ||
523 | switch (usec) { | |
524 | case 750 ... 999: | |
525 | index = 9; | |
526 | break; | |
527 | case 500 ... 749: | |
528 | index = 8; | |
529 | break; | |
530 | case 250 ... 499: | |
531 | index = 7; | |
532 | break; | |
533 | case 100 ... 249: | |
534 | index = 6; | |
535 | break; | |
536 | case 50 ... 99: | |
537 | index = 5; | |
538 | break; | |
539 | case 20 ... 49: | |
540 | index = 4; | |
541 | break; | |
542 | case 10 ... 19: | |
543 | index = 3; | |
544 | break; | |
545 | case 4 ... 9: | |
546 | index = 2; | |
547 | break; | |
548 | case 2 ... 3: | |
549 | index = 1; | |
550 | case 0 ... 1: | |
551 | break; | |
552 | } | |
553 | ||
554 | assert(index < FIO_IO_U_LAT_U_NR); | |
555 | td->ts.io_u_lat_u[index]++; | |
556 | } | |
557 | ||
558 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 JA |
559 | { |
560 | int index = 0; | |
561 | ||
562 | switch (msec) { | |
563 | default: | |
04a0feae JA |
564 | index = 11; |
565 | break; | |
8abdce66 | 566 | case 1000 ... 1999: |
04a0feae JA |
567 | index = 10; |
568 | break; | |
8abdce66 | 569 | case 750 ... 999: |
04a0feae JA |
570 | index = 9; |
571 | break; | |
8abdce66 | 572 | case 500 ... 749: |
04a0feae JA |
573 | index = 8; |
574 | break; | |
8abdce66 | 575 | case 250 ... 499: |
04a0feae JA |
576 | index = 7; |
577 | break; | |
8abdce66 | 578 | case 100 ... 249: |
04a0feae JA |
579 | index = 6; |
580 | break; | |
8abdce66 | 581 | case 50 ... 99: |
04a0feae JA |
582 | index = 5; |
583 | break; | |
8abdce66 | 584 | case 20 ... 49: |
04a0feae JA |
585 | index = 4; |
586 | break; | |
8abdce66 | 587 | case 10 ... 19: |
04a0feae JA |
588 | index = 3; |
589 | break; | |
8abdce66 | 590 | case 4 ... 9: |
04a0feae JA |
591 | index = 2; |
592 | break; | |
ec118304 | 593 | case 2 ... 3: |
04a0feae | 594 | index = 1; |
ec118304 JA |
595 | case 0 ... 1: |
596 | break; | |
597 | } | |
598 | ||
04a0feae JA |
599 | assert(index < FIO_IO_U_LAT_M_NR); |
600 | td->ts.io_u_lat_m[index]++; | |
601 | } | |
602 | ||
603 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
604 | { | |
605 | if (usec < 1000) | |
606 | io_u_mark_lat_usec(td, usec); | |
607 | else | |
608 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
609 | } |
610 | ||
0aabe160 JA |
611 | /* |
612 | * Get next file to service by choosing one at random | |
613 | */ | |
d6aed795 JA |
614 | static struct fio_file *get_next_file_rand(struct thread_data *td, enum fio_file_flags goodf, |
615 | enum fio_file_flags badf) | |
0aabe160 | 616 | { |
0aabe160 | 617 | struct fio_file *f; |
1c178180 | 618 | int fno; |
0aabe160 JA |
619 | |
620 | do { | |
7c83c089 | 621 | long r = os_random_long(&td->next_file_state); |
87b10676 | 622 | int opened = 0; |
7c83c089 | 623 | |
5ec10eaa | 624 | fno = (unsigned int) ((double) td->o.nr_files |
dc873b6f | 625 | * (r / (OS_RAND_MAX + 1.0))); |
126d65c6 | 626 | f = td->files[fno]; |
d6aed795 | 627 | if (fio_file_done(f)) |
059e63c0 | 628 | continue; |
1c178180 | 629 | |
d6aed795 | 630 | if (!fio_file_open(f)) { |
87b10676 JA |
631 | int err; |
632 | ||
633 | err = td_io_open_file(td, f); | |
634 | if (err) | |
635 | continue; | |
636 | opened = 1; | |
637 | } | |
638 | ||
2ba1c290 JA |
639 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
640 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 641 | return f; |
2ba1c290 | 642 | } |
87b10676 JA |
643 | if (opened) |
644 | td_io_close_file(td, f); | |
0aabe160 JA |
645 | } while (1); |
646 | } | |
647 | ||
648 | /* | |
649 | * Get next file to service by doing round robin between all available ones | |
650 | */ | |
1c178180 JA |
651 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
652 | int badf) | |
3d7c391d JA |
653 | { |
654 | unsigned int old_next_file = td->next_file; | |
655 | struct fio_file *f; | |
656 | ||
657 | do { | |
87b10676 JA |
658 | int opened = 0; |
659 | ||
126d65c6 | 660 | f = td->files[td->next_file]; |
3d7c391d JA |
661 | |
662 | td->next_file++; | |
2dc1bbeb | 663 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
664 | td->next_file = 0; |
665 | ||
87b10676 | 666 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 667 | if (fio_file_done(f)) { |
d5ed68ea | 668 | f = NULL; |
059e63c0 | 669 | continue; |
d5ed68ea | 670 | } |
059e63c0 | 671 | |
d6aed795 | 672 | if (!fio_file_open(f)) { |
87b10676 JA |
673 | int err; |
674 | ||
675 | err = td_io_open_file(td, f); | |
b5696bfc JA |
676 | if (err) { |
677 | dprint(FD_FILE, "error %d on open of %s\n", | |
678 | err, f->file_name); | |
87c27b45 | 679 | f = NULL; |
87b10676 | 680 | continue; |
b5696bfc | 681 | } |
87b10676 JA |
682 | opened = 1; |
683 | } | |
684 | ||
b5696bfc | 685 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags); |
1c178180 | 686 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
687 | break; |
688 | ||
87b10676 JA |
689 | if (opened) |
690 | td_io_close_file(td, f); | |
691 | ||
3d7c391d JA |
692 | f = NULL; |
693 | } while (td->next_file != old_next_file); | |
694 | ||
2ba1c290 | 695 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
696 | return f; |
697 | } | |
698 | ||
bdb4e2e9 JA |
699 | static struct fio_file *get_next_file(struct thread_data *td) |
700 | { | |
1907dbc6 JA |
701 | struct fio_file *f; |
702 | ||
2dc1bbeb | 703 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 704 | |
b5696bfc | 705 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
706 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
707 | " nr_files=%d\n", td->nr_open_files, | |
708 | td->nr_done_files, | |
709 | td->o.nr_files); | |
bdb4e2e9 | 710 | return NULL; |
2ba1c290 | 711 | } |
bdb4e2e9 | 712 | |
1907dbc6 | 713 | f = td->file_service_file; |
d6aed795 | 714 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
715 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
716 | goto out; | |
717 | if (td->file_service_left--) | |
718 | goto out; | |
719 | } | |
1907dbc6 | 720 | |
a086c257 JA |
721 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
722 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 723 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 724 | else |
d6aed795 | 725 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 JA |
726 | |
727 | td->file_service_file = f; | |
728 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 729 | out: |
683023e8 | 730 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 731 | return f; |
bdb4e2e9 JA |
732 | } |
733 | ||
429f6675 JA |
734 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
735 | { | |
736 | struct fio_file *f; | |
737 | ||
738 | do { | |
739 | f = get_next_file(td); | |
740 | if (!f) | |
741 | return 1; | |
742 | ||
429f6675 JA |
743 | io_u->file = f; |
744 | get_file(f); | |
745 | ||
746 | if (!fill_io_u(td, io_u)) | |
747 | break; | |
748 | ||
b5696bfc | 749 | put_file_log(td, f); |
429f6675 | 750 | td_io_close_file(td, f); |
b5696bfc | 751 | io_u->file = NULL; |
d6aed795 | 752 | fio_file_set_done(f); |
429f6675 | 753 | td->nr_done_files++; |
87b10676 | 754 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files); |
429f6675 JA |
755 | } while (1); |
756 | ||
757 | return 0; | |
758 | } | |
759 | ||
760 | ||
10ba535a JA |
761 | struct io_u *__get_io_u(struct thread_data *td) |
762 | { | |
763 | struct io_u *io_u = NULL; | |
764 | ||
01743ee1 JA |
765 | if (!flist_empty(&td->io_u_requeues)) |
766 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 767 | else if (!queue_full(td)) { |
01743ee1 | 768 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 769 | |
6040dabc | 770 | io_u->buflen = 0; |
10ba535a | 771 | io_u->resid = 0; |
755200a3 | 772 | io_u->file = NULL; |
d7762cf8 | 773 | io_u->end_io = NULL; |
755200a3 JA |
774 | } |
775 | ||
776 | if (io_u) { | |
0c6e7517 JA |
777 | assert(io_u->flags & IO_U_F_FREE); |
778 | io_u->flags &= ~IO_U_F_FREE; | |
779 | ||
755200a3 | 780 | io_u->error = 0; |
01743ee1 JA |
781 | flist_del(&io_u->list); |
782 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a JA |
783 | td->cur_depth++; |
784 | } | |
785 | ||
786 | return io_u; | |
787 | } | |
788 | ||
789 | /* | |
790 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
791 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
792 | */ | |
3d7c391d | 793 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 794 | { |
3d7c391d | 795 | struct fio_file *f; |
10ba535a JA |
796 | struct io_u *io_u; |
797 | ||
798 | io_u = __get_io_u(td); | |
2ba1c290 JA |
799 | if (!io_u) { |
800 | dprint(FD_IO, "__get_io_u failed\n"); | |
10ba535a | 801 | return NULL; |
2ba1c290 | 802 | } |
10ba535a | 803 | |
755200a3 JA |
804 | /* |
805 | * from a requeue, io_u already setup | |
806 | */ | |
807 | if (io_u->file) | |
77f392bf | 808 | goto out; |
755200a3 | 809 | |
429f6675 JA |
810 | /* |
811 | * If using an iolog, grab next piece if any available. | |
812 | */ | |
813 | if (td->o.read_iolog_file) { | |
814 | if (read_iolog_get(td, io_u)) | |
815 | goto err_put; | |
2ba1c290 JA |
816 | } else if (set_io_u_file(td, io_u)) { |
817 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 818 | goto err_put; |
2ba1c290 | 819 | } |
5ec10eaa | 820 | |
429f6675 | 821 | f = io_u->file; |
d6aed795 | 822 | assert(fio_file_open(f)); |
97af62ce | 823 | |
87dc1ab1 | 824 | if (io_u->ddir != DDIR_SYNC) { |
d0656a93 | 825 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 826 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 827 | goto err_put; |
2ba1c290 | 828 | } |
10ba535a | 829 | |
36167d82 | 830 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 831 | |
c311cd2a | 832 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE) |
87dc1ab1 | 833 | populate_verify_io_u(td, io_u); |
e4dad9c6 JA |
834 | else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) |
835 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
87dc1ab1 | 836 | } |
10ba535a | 837 | |
165faf16 JA |
838 | /* |
839 | * Set io data pointers. | |
840 | */ | |
cec6b55d JA |
841 | io_u->xfer_buf = io_u->buf; |
842 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 843 | |
6ac7a331 | 844 | out: |
429f6675 | 845 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
846 | if (!td->o.disable_slat) |
847 | fio_gettime(&io_u->start_time, NULL); | |
429f6675 | 848 | return io_u; |
36167d82 | 849 | } |
429f6675 | 850 | err_put: |
2ba1c290 | 851 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
852 | put_io_u(td, io_u); |
853 | return NULL; | |
10ba535a JA |
854 | } |
855 | ||
5451792e JA |
856 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
857 | { | |
858 | const char *msg[] = { "read", "write", "sync" }; | |
859 | ||
860 | log_err("fio: io_u error"); | |
861 | ||
862 | if (io_u->file) | |
863 | log_err(" on file %s", io_u->file->file_name); | |
864 | ||
865 | log_err(": %s\n", strerror(io_u->error)); | |
866 | ||
5ec10eaa JA |
867 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
868 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
869 | |
870 | if (!td->error) | |
871 | td_verror(td, io_u->error, "io_u error"); | |
872 | } | |
873 | ||
97601024 JA |
874 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
875 | struct io_completion_data *icd) | |
10ba535a | 876 | { |
dbad30b0 JA |
877 | /* |
878 | * Older gcc's are too dumb to realize that usec is always used | |
879 | * initialized, silence that warning. | |
880 | */ | |
881 | unsigned long uninitialized_var(usec); | |
10ba535a | 882 | |
2ba1c290 JA |
883 | dprint_io_u(io_u, "io complete"); |
884 | ||
0c6e7517 JA |
885 | assert(io_u->flags & IO_U_F_FLIGHT); |
886 | io_u->flags &= ~IO_U_F_FLIGHT; | |
887 | ||
87dc1ab1 JA |
888 | if (io_u->ddir == DDIR_SYNC) { |
889 | td->last_was_sync = 1; | |
890 | return; | |
891 | } | |
892 | ||
893 | td->last_was_sync = 0; | |
894 | ||
10ba535a JA |
895 | if (!io_u->error) { |
896 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 897 | const enum fio_ddir idx = io_u->ddir; |
b29ee5b3 | 898 | int ret; |
10ba535a | 899 | |
b29ee5b3 JA |
900 | td->io_blocks[idx]++; |
901 | td->io_bytes[idx] += bytes; | |
902 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 903 | |
b29ee5b3 | 904 | if (ramp_time_over(td)) { |
9520ebb9 JA |
905 | if (!td->o.disable_clat || !td->o.disable_bw) |
906 | usec = utime_since(&io_u->issue_time, | |
907 | &icd->time); | |
721938ae | 908 | |
9520ebb9 | 909 | if (!td->o.disable_clat) { |
306ddc97 | 910 | add_clat_sample(td, idx, usec, bytes); |
9520ebb9 JA |
911 | io_u_mark_latency(td, usec); |
912 | } | |
913 | if (!td->o.disable_bw) | |
306ddc97 | 914 | add_bw_sample(td, idx, bytes, &icd->time); |
721938ae | 915 | } |
10ba535a | 916 | |
660a1cb5 | 917 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 918 | td->o.do_verify && |
41128405 | 919 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
920 | log_io_piece(td, io_u); |
921 | ||
b29ee5b3 | 922 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 923 | |
d7762cf8 | 924 | if (io_u->end_io) { |
36690c9b | 925 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
926 | if (ret && !icd->error) |
927 | icd->error = ret; | |
928 | } | |
5451792e | 929 | } else { |
10ba535a | 930 | icd->error = io_u->error; |
5451792e JA |
931 | io_u_log_error(td, io_u); |
932 | } | |
10ba535a JA |
933 | } |
934 | ||
9520ebb9 JA |
935 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
936 | int nr) | |
10ba535a | 937 | { |
9520ebb9 JA |
938 | if (!td->o.disable_clat || !td->o.disable_bw) |
939 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 940 | |
3af6ef39 JA |
941 | icd->nr = nr; |
942 | ||
10ba535a JA |
943 | icd->error = 0; |
944 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
945 | } |
946 | ||
97601024 JA |
947 | static void ios_completed(struct thread_data *td, |
948 | struct io_completion_data *icd) | |
36167d82 JA |
949 | { |
950 | struct io_u *io_u; | |
951 | int i; | |
952 | ||
10ba535a JA |
953 | for (i = 0; i < icd->nr; i++) { |
954 | io_u = td->io_ops->event(td, i); | |
955 | ||
956 | io_completed(td, io_u, icd); | |
957 | put_io_u(td, io_u); | |
958 | } | |
959 | } | |
97601024 | 960 | |
e7e6cfb4 JA |
961 | /* |
962 | * Complete a single io_u for the sync engines. | |
963 | */ | |
d7762cf8 | 964 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
965 | { |
966 | struct io_completion_data icd; | |
967 | ||
9520ebb9 | 968 | init_icd(td, &icd, 1); |
97601024 JA |
969 | io_completed(td, io_u, &icd); |
970 | put_io_u(td, io_u); | |
971 | ||
972 | if (!icd.error) | |
973 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
974 | ||
37e974a5 | 975 | td_verror(td, icd.error, "io_u_sync_complete"); |
97601024 JA |
976 | return -1; |
977 | } | |
978 | ||
e7e6cfb4 JA |
979 | /* |
980 | * Called to complete min_events number of io for the async engines. | |
981 | */ | |
4950421a | 982 | long io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 983 | { |
97601024 | 984 | struct io_completion_data icd; |
00de55ef | 985 | struct timespec *tvp = NULL; |
97601024 | 986 | int ret; |
4d06a338 | 987 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 988 | |
4950421a | 989 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 990 | |
4950421a | 991 | if (!min_evts) |
00de55ef | 992 | tvp = &ts; |
97601024 | 993 | |
4950421a | 994 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 995 | if (ret < 0) { |
e1161c32 | 996 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
997 | return ret; |
998 | } else if (!ret) | |
999 | return ret; | |
1000 | ||
9520ebb9 | 1001 | init_icd(td, &icd, ret); |
97601024 JA |
1002 | ios_completed(td, &icd); |
1003 | if (!icd.error) | |
1004 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
1005 | ||
37e974a5 | 1006 | td_verror(td, icd.error, "io_u_queued_complete"); |
97601024 JA |
1007 | return -1; |
1008 | } | |
7e77dd02 JA |
1009 | |
1010 | /* | |
1011 | * Call when io_u is really queued, to update the submission latency. | |
1012 | */ | |
1013 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1014 | { | |
9520ebb9 JA |
1015 | if (!td->o.disable_slat) { |
1016 | unsigned long slat_time; | |
7e77dd02 | 1017 | |
9520ebb9 | 1018 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
306ddc97 | 1019 | add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time); |
9520ebb9 | 1020 | } |
7e77dd02 | 1021 | } |
433afcb4 | 1022 | |
5973cafb JA |
1023 | /* |
1024 | * "randomly" fill the buffer contents | |
1025 | */ | |
1026 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1027 | unsigned int max_bs) | |
1028 | { | |
1029 | long *ptr = io_u->buf; | |
1030 | ||
1031 | if (!td->o.zero_buffers) { | |
1032 | while ((void *) ptr - io_u->buf < max_bs) { | |
1033 | *ptr = rand() * GOLDEN_RATIO_PRIME; | |
1034 | ptr++; | |
1035 | } | |
1036 | } else | |
1037 | memset(ptr, 0, max_bs); | |
1038 | } |