Commit | Line | Data |
---|---|---|
10ba535a | 1 | #include <unistd.h> |
10ba535a | 2 | #include <string.h> |
0c6e7517 | 3 | #include <assert.h> |
10ba535a JA |
4 | |
5 | #include "fio.h" | |
4f5af7b2 | 6 | #include "verify.h" |
0d29de83 | 7 | #include "trim.h" |
1fbbf72e | 8 | #include "lib/rand.h" |
7ebd796f | 9 | #include "lib/axmap.h" |
002fe734 | 10 | #include "err.h" |
0f38bbef | 11 | #include "lib/pow2.h" |
4b157ac6 | 12 | #include "minmax.h" |
10ba535a | 13 | |
97601024 JA |
14 | struct io_completion_data { |
15 | int nr; /* input */ | |
97601024 JA |
16 | |
17 | int error; /* output */ | |
100f49f1 | 18 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
8b6a404c | 19 | struct timespec time; /* output */ |
97601024 JA |
20 | }; |
21 | ||
10ba535a | 22 | /* |
7ebd796f | 23 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
24 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
25 | */ | |
e39c0676 | 26 | static bool random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 27 | { |
7ebd796f | 28 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
29 | } |
30 | ||
df415585 JA |
31 | /* |
32 | * Mark a given offset as used in the map. | |
33 | */ | |
6cc1a3d1 BVA |
34 | static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u, |
35 | uint64_t offset, uint64_t buflen) | |
df415585 | 36 | { |
5fff9543 | 37 | unsigned long long min_bs = td->o.min_bs[io_u->ddir]; |
9bf2061e | 38 | struct fio_file *f = io_u->file; |
5fff9543 | 39 | unsigned long long nr_blocks; |
1ae83d45 | 40 | uint64_t block; |
df415585 | 41 | |
6cc1a3d1 BVA |
42 | block = (offset - f->file_offset) / (uint64_t) min_bs; |
43 | nr_blocks = (buflen + min_bs - 1) / min_bs; | |
bd6b959a | 44 | assert(nr_blocks > 0); |
df415585 | 45 | |
bd6b959a | 46 | if (!(io_u->flags & IO_U_F_BUSY_OK)) { |
7ebd796f | 47 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
bd6b959a BVA |
48 | assert(nr_blocks > 0); |
49 | } | |
df415585 | 50 | |
6cc1a3d1 BVA |
51 | if ((nr_blocks * min_bs) < buflen) |
52 | buflen = nr_blocks * min_bs; | |
53 | ||
54 | return buflen; | |
df415585 JA |
55 | } |
56 | ||
74776733 JA |
57 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
58 | enum fio_ddir ddir) | |
2ba1c290 | 59 | { |
74776733 JA |
60 | uint64_t max_blocks; |
61 | uint64_t max_size; | |
2ba1c290 | 62 | |
ff58fced JA |
63 | assert(ddir_rw(ddir)); |
64 | ||
d9dd70f7 JA |
65 | /* |
66 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
79591fa9 | 67 | * -> not for now since there is code assuming it could go either. |
d9dd70f7 JA |
68 | */ |
69 | max_size = f->io_size; | |
70 | if (max_size > f->real_file_size) | |
71 | max_size = f->real_file_size; | |
72 | ||
7b865a2f | 73 | if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range) |
ed335855 SN |
74 | max_size = td->o.zone_range; |
75 | ||
0412d12e JE |
76 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
77 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
78 | ||
1ae83d45 | 79 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
80 | if (!max_blocks) |
81 | return 0; | |
82 | ||
67778e88 | 83 | return max_blocks; |
2ba1c290 JA |
84 | } |
85 | ||
e25839d4 | 86 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
e0a04ac1 JA |
87 | enum fio_ddir ddir, uint64_t *b, |
88 | uint64_t lastb) | |
ec4015da | 89 | { |
6f49f8bc | 90 | uint64_t r; |
5e0baa7f | 91 | |
c3546b53 JA |
92 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
93 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { | |
6f49f8bc | 94 | |
d6b72507 | 95 | r = __rand(&td->random_state); |
8055e41d | 96 | |
4b91ee8f | 97 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 98 | |
e0a04ac1 | 99 | *b = lastb * (r / (rand_max(&td->random_state) + 1.0)); |
51ede0b1 | 100 | } else { |
8055e41d | 101 | uint64_t off = 0; |
43c63a78 | 102 | |
967d1b63 JA |
103 | assert(fio_file_lfsr(f)); |
104 | ||
6f49f8bc | 105 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 106 | return 1; |
ec4015da | 107 | |
8055e41d JA |
108 | *b = off; |
109 | } | |
0ce8b119 | 110 | |
ec4015da | 111 | /* |
51ede0b1 | 112 | * if we are not maintaining a random map, we are done. |
ec4015da | 113 | */ |
51ede0b1 JA |
114 | if (!file_randommap(td, f)) |
115 | goto ret; | |
43c63a78 JA |
116 | |
117 | /* | |
51ede0b1 | 118 | * calculate map offset and check if it's free |
43c63a78 | 119 | */ |
51ede0b1 JA |
120 | if (random_map_free(f, *b)) |
121 | goto ret; | |
122 | ||
4b91ee8f JA |
123 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
124 | (unsigned long long) *b); | |
51ede0b1 | 125 | |
7ebd796f | 126 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
127 | if (*b == (uint64_t) -1ULL) |
128 | return 1; | |
0ce8b119 JA |
129 | ret: |
130 | return 0; | |
ec4015da JA |
131 | } |
132 | ||
925fee33 JA |
133 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
134 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 135 | uint64_t *b) |
e25839d4 | 136 | { |
9c6f6316 | 137 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
138 | return 0; |
139 | } | |
140 | ||
925fee33 JA |
141 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
142 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 143 | uint64_t *b) |
925fee33 | 144 | { |
9c6f6316 | 145 | *b = pareto_next(&f->zipf); |
925fee33 JA |
146 | return 0; |
147 | } | |
148 | ||
56d9fa4b JA |
149 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
150 | struct fio_file *f, enum fio_ddir ddir, | |
151 | uint64_t *b) | |
152 | { | |
153 | *b = gauss_next(&f->gauss); | |
154 | return 0; | |
155 | } | |
156 | ||
59466396 JA |
157 | static int __get_next_rand_offset_zoned_abs(struct thread_data *td, |
158 | struct fio_file *f, | |
159 | enum fio_ddir ddir, uint64_t *b) | |
160 | { | |
161 | struct zone_split_index *zsi; | |
e3c8c108 | 162 | uint64_t lastb, send, stotal; |
59466396 JA |
163 | unsigned int v; |
164 | ||
165 | lastb = last_block(td, f, ddir); | |
166 | if (!lastb) | |
167 | return 1; | |
168 | ||
169 | if (!td->o.zone_split_nr[ddir]) { | |
170 | bail: | |
171 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
172 | } | |
173 | ||
174 | /* | |
175 | * Generate a value, v, between 1 and 100, both inclusive | |
176 | */ | |
1bd5d213 | 177 | v = rand_between(&td->zone_state, 1, 100); |
59466396 | 178 | |
e3c8c108 JA |
179 | /* |
180 | * Find our generated table. 'send' is the end block of this zone, | |
181 | * 'stotal' is our start offset. | |
182 | */ | |
59466396 JA |
183 | zsi = &td->zone_state_index[ddir][v - 1]; |
184 | stotal = zsi->size_prev / td->o.ba[ddir]; | |
185 | send = zsi->size / td->o.ba[ddir]; | |
186 | ||
187 | /* | |
188 | * Should never happen | |
189 | */ | |
190 | if (send == -1U) { | |
264e3d30 | 191 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
59466396 | 192 | log_err("fio: bug in zoned generation\n"); |
59466396 JA |
193 | goto bail; |
194 | } else if (send > lastb) { | |
195 | /* | |
196 | * This happens if the user specifies ranges that exceed | |
197 | * the file/device size. We can't handle that gracefully, | |
198 | * so error and exit. | |
199 | */ | |
200 | log_err("fio: zoned_abs sizes exceed file size\n"); | |
201 | return 1; | |
202 | } | |
203 | ||
204 | /* | |
e3c8c108 | 205 | * Generate index from 0..send-stotal |
59466396 | 206 | */ |
e3c8c108 | 207 | if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1) |
59466396 JA |
208 | return 1; |
209 | ||
e3c8c108 | 210 | *b += stotal; |
59466396 JA |
211 | return 0; |
212 | } | |
213 | ||
e0a04ac1 JA |
214 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
215 | struct fio_file *f, enum fio_ddir ddir, | |
216 | uint64_t *b) | |
217 | { | |
218 | unsigned int v, send, stotal; | |
219 | uint64_t offset, lastb; | |
e0a04ac1 JA |
220 | struct zone_split_index *zsi; |
221 | ||
222 | lastb = last_block(td, f, ddir); | |
223 | if (!lastb) | |
224 | return 1; | |
225 | ||
226 | if (!td->o.zone_split_nr[ddir]) { | |
227 | bail: | |
228 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
229 | } | |
230 | ||
231 | /* | |
232 | * Generate a value, v, between 1 and 100, both inclusive | |
233 | */ | |
1bd5d213 | 234 | v = rand_between(&td->zone_state, 1, 100); |
e0a04ac1 JA |
235 | |
236 | zsi = &td->zone_state_index[ddir][v - 1]; | |
237 | stotal = zsi->size_perc_prev; | |
238 | send = zsi->size_perc; | |
239 | ||
240 | /* | |
241 | * Should never happen | |
242 | */ | |
243 | if (send == -1U) { | |
264e3d30 | 244 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
e0a04ac1 | 245 | log_err("fio: bug in zoned generation\n"); |
e0a04ac1 JA |
246 | goto bail; |
247 | } | |
248 | ||
249 | /* | |
250 | * 'send' is some percentage below or equal to 100 that | |
251 | * marks the end of the current IO range. 'stotal' marks | |
252 | * the start, in percent. | |
253 | */ | |
254 | if (stotal) | |
255 | offset = stotal * lastb / 100ULL; | |
256 | else | |
257 | offset = 0; | |
258 | ||
259 | lastb = lastb * (send - stotal) / 100ULL; | |
260 | ||
261 | /* | |
262 | * Generate index from 0..send-of-lastb | |
263 | */ | |
264 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
265 | return 1; | |
266 | ||
267 | /* | |
268 | * Add our start offset, if any | |
269 | */ | |
270 | if (offset) | |
271 | *b += offset; | |
272 | ||
273 | return 0; | |
274 | } | |
56d9fa4b | 275 | |
f31feaa2 JA |
276 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
277 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 | 278 | { |
e0a04ac1 JA |
279 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
280 | uint64_t lastb; | |
281 | ||
282 | lastb = last_block(td, f, ddir); | |
283 | if (!lastb) | |
284 | return 1; | |
285 | ||
286 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
287 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
e25839d4 | 288 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
925fee33 JA |
289 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
290 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
291 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
292 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e0a04ac1 JA |
293 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
294 | return __get_next_rand_offset_zoned(td, f, ddir, b); | |
59466396 JA |
295 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS) |
296 | return __get_next_rand_offset_zoned_abs(td, f, ddir, b); | |
e25839d4 JA |
297 | |
298 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
299 | return 1; | |
300 | } | |
301 | ||
e39c0676 | 302 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
303 | { |
304 | unsigned int v; | |
211c9b89 | 305 | |
d9472271 | 306 | if (td->o.perc_rand[ddir] == 100) |
e39c0676 | 307 | return true; |
211c9b89 | 308 | |
1bd5d213 | 309 | v = rand_between(&td->seq_rand_state[ddir], 1, 100); |
211c9b89 | 310 | |
d9472271 | 311 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
312 | } |
313 | ||
0bcf41cd JA |
314 | static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f) |
315 | { | |
316 | struct thread_options *o = &td->o; | |
317 | ||
318 | if (o->invalidate_cache && !o->odirect) { | |
319 | int fio_unused ret; | |
320 | ||
321 | ret = file_invalidate_cache(td, f); | |
322 | } | |
323 | } | |
324 | ||
38dad62d | 325 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 326 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 327 | { |
c04e4661 DE |
328 | if (!get_next_rand_offset(td, f, ddir, b)) |
329 | return 0; | |
330 | ||
8c07860d JA |
331 | if (td->o.time_based || |
332 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { | |
33c48814 | 333 | fio_file_reset(td, f); |
80f02150 | 334 | loop_cache_invalidate(td, f); |
c04e4661 DE |
335 | if (!get_next_rand_offset(td, f, ddir, b)) |
336 | return 0; | |
38dad62d JA |
337 | } |
338 | ||
c04e4661 | 339 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 340 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 341 | (unsigned long long) f->real_file_size); |
c04e4661 | 342 | return 1; |
38dad62d JA |
343 | } |
344 | ||
37cf9e3c | 345 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 346 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 347 | { |
ac002339 JA |
348 | struct thread_options *o = &td->o; |
349 | ||
ff58fced JA |
350 | assert(ddir_rw(ddir)); |
351 | ||
17373ce2 JA |
352 | /* |
353 | * If we reach the end for a time based run, reset us back to 0 | |
354 | * and invalidate the cache, if we need to. | |
355 | */ | |
f1dfb668 | 356 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
19ddc35b | 357 | o->time_based) { |
c89daa4a | 358 | f->last_pos[ddir] = f->file_offset; |
0bcf41cd | 359 | loop_cache_invalidate(td, f); |
19ddc35b | 360 | } |
c04e4661 | 361 | |
f1dfb668 | 362 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 363 | uint64_t pos; |
059b0802 | 364 | |
69b98f11 JA |
365 | /* |
366 | * Only rewind if we already hit the end | |
367 | */ | |
368 | if (f->last_pos[ddir] == f->file_offset && | |
369 | f->file_offset && o->ddir_seq_add < 0) { | |
c22825bb JA |
370 | if (f->real_file_size > f->io_size) |
371 | f->last_pos[ddir] = f->io_size; | |
372 | else | |
373 | f->last_pos[ddir] = f->real_file_size; | |
374 | } | |
a66da7a2 | 375 | |
f1dfb668 | 376 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
377 | if (pos && o->ddir_seq_add) { |
378 | pos += o->ddir_seq_add; | |
379 | ||
380 | /* | |
381 | * If we reach beyond the end of the file | |
382 | * with holed IO, wrap around to the | |
b0a84f48 JA |
383 | * beginning again. If we're doing backwards IO, |
384 | * wrap to the end. | |
ac002339 | 385 | */ |
b0a84f48 JA |
386 | if (pos >= f->real_file_size) { |
387 | if (o->ddir_seq_add > 0) | |
388 | pos = f->file_offset; | |
c22825bb JA |
389 | else { |
390 | if (f->real_file_size > f->io_size) | |
391 | pos = f->io_size; | |
392 | else | |
393 | pos = f->real_file_size; | |
394 | ||
395 | pos += o->ddir_seq_add; | |
396 | } | |
b0a84f48 | 397 | } |
ac002339 | 398 | } |
059b0802 | 399 | |
37cf9e3c | 400 | *offset = pos; |
38dad62d JA |
401 | return 0; |
402 | } | |
403 | ||
404 | return 1; | |
405 | } | |
406 | ||
407 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d | 408 | enum fio_ddir ddir, int rw_seq, |
ec370c22 | 409 | bool *is_random) |
38dad62d JA |
410 | { |
411 | struct fio_file *f = io_u->file; | |
1ae83d45 | 412 | uint64_t b, offset; |
38dad62d JA |
413 | int ret; |
414 | ||
ff58fced JA |
415 | assert(ddir_rw(ddir)); |
416 | ||
37cf9e3c JA |
417 | b = offset = -1ULL; |
418 | ||
38dad62d | 419 | if (rw_seq) { |
211c9b89 | 420 | if (td_random(td)) { |
6aca9b3d | 421 | if (should_do_random(td, ddir)) { |
211c9b89 | 422 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 423 | *is_random = true; |
6aca9b3d | 424 | } else { |
ec370c22 | 425 | *is_random = false; |
1651e431 | 426 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
427 | ret = get_next_seq_offset(td, f, ddir, &offset); |
428 | if (ret) | |
429 | ret = get_next_rand_block(td, f, ddir, &b); | |
430 | } | |
6aca9b3d | 431 | } else { |
ec370c22 | 432 | *is_random = false; |
37cf9e3c | 433 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 434 | } |
38dad62d | 435 | } else { |
1651e431 | 436 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
ec370c22 | 437 | *is_random = false; |
38dad62d JA |
438 | |
439 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 440 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 441 | if (ret) { |
37cf9e3c | 442 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 443 | *is_random = false; |
6aca9b3d | 444 | } |
38dad62d | 445 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
446 | if (f->last_start[ddir] != -1ULL) |
447 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 448 | else |
37cf9e3c | 449 | offset = 0; |
38dad62d JA |
450 | ret = 0; |
451 | } else { | |
452 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
453 | ret = 1; | |
454 | } | |
455 | } | |
6d68b997 | 456 | |
37cf9e3c JA |
457 | if (!ret) { |
458 | if (offset != -1ULL) | |
459 | io_u->offset = offset; | |
460 | else if (b != -1ULL) | |
461 | io_u->offset = b * td->o.ba[ddir]; | |
462 | else { | |
4e0a8fa2 | 463 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
464 | ret = 1; |
465 | } | |
466 | } | |
467 | ||
38dad62d JA |
468 | return ret; |
469 | } | |
470 | ||
10ba535a JA |
471 | /* |
472 | * For random io, generate a random new block and see if it's used. Repeat | |
473 | * until we find a free one. For sequential io, just return the end of | |
474 | * the last io issued. | |
475 | */ | |
e8fb335e | 476 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 477 | bool *is_random) |
10ba535a | 478 | { |
9bf2061e | 479 | struct fio_file *f = io_u->file; |
4ba66134 | 480 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 481 | int rw_seq_hit = 0; |
10ba535a | 482 | |
ff58fced JA |
483 | assert(ddir_rw(ddir)); |
484 | ||
38dad62d JA |
485 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
486 | rw_seq_hit = 1; | |
5736c10d | 487 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 488 | } |
211097b2 | 489 | |
6aca9b3d | 490 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 491 | return 1; |
10ba535a | 492 | |
009bd847 JA |
493 | if (io_u->offset >= f->io_size) { |
494 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
495 | (unsigned long long) io_u->offset, |
496 | (unsigned long long) f->io_size); | |
009bd847 JA |
497 | return 1; |
498 | } | |
499 | ||
500 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
501 | if (io_u->offset >= f->real_file_size) { |
502 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
503 | (unsigned long long) io_u->offset, |
504 | (unsigned long long) f->real_file_size); | |
10ba535a | 505 | return 1; |
2ba1c290 | 506 | } |
10ba535a JA |
507 | |
508 | return 0; | |
509 | } | |
510 | ||
e39c0676 | 511 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
5fff9543 | 512 | unsigned long long buflen) |
79944128 JA |
513 | { |
514 | struct fio_file *f = io_u->file; | |
515 | ||
bedc9dc2 | 516 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
517 | } |
518 | ||
5fff9543 | 519 | static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 520 | bool is_random) |
10ba535a | 521 | { |
6aca9b3d | 522 | int ddir = io_u->ddir; |
5fff9543 JF |
523 | unsigned long long buflen = 0; |
524 | unsigned long long minbs, maxbs; | |
3dd29f7c | 525 | uint64_t frand_max, r; |
7c961359 | 526 | bool power_2; |
10ba535a | 527 | |
9ee1c647 | 528 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
529 | |
530 | if (td->o.bs_is_seq_rand) | |
ec370c22 | 531 | ddir = is_random ? DDIR_WRITE : DDIR_READ; |
ff58fced | 532 | |
f3059de1 JA |
533 | minbs = td->o.min_bs[ddir]; |
534 | maxbs = td->o.max_bs[ddir]; | |
535 | ||
79944128 JA |
536 | if (minbs == maxbs) |
537 | return minbs; | |
538 | ||
52c58027 JA |
539 | /* |
540 | * If we can't satisfy the min block size from here, then fail | |
541 | */ | |
542 | if (!io_u_fits(td, io_u, minbs)) | |
543 | return 0; | |
544 | ||
2f282cec | 545 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 546 | do { |
2f282cec | 547 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 548 | |
720e84ad | 549 | if (!td->o.bssplit_nr[ddir]) { |
5fff9543 | 550 | buflen = minbs + (unsigned long long) ((double) maxbs * |
c3546b53 | 551 | (r / (frand_max + 1.0))); |
5ec10eaa | 552 | } else { |
3dd29f7c | 553 | long long perc = 0; |
564ca972 JA |
554 | unsigned int i; |
555 | ||
720e84ad JA |
556 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
557 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
558 | |
559 | buflen = bsp->bs; | |
560 | perc += bsp->perc; | |
3dd29f7c JA |
561 | if (!perc) |
562 | break; | |
563 | if ((r / perc <= frand_max / 100ULL) && | |
79944128 | 564 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
565 | break; |
566 | } | |
567 | } | |
79944128 | 568 | |
17a6b702 PL |
569 | power_2 = is_power_of_2(minbs); |
570 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 571 | buflen &= ~(minbs - 1); |
17a6b702 PL |
572 | else if (!td->o.bs_unaligned && !power_2) |
573 | buflen -= buflen % minbs; | |
79944128 | 574 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 575 | |
10ba535a JA |
576 | return buflen; |
577 | } | |
578 | ||
afe24a5a JA |
579 | static void set_rwmix_bytes(struct thread_data *td) |
580 | { | |
afe24a5a JA |
581 | unsigned int diff; |
582 | ||
583 | /* | |
584 | * we do time or byte based switch. this is needed because | |
585 | * buffered writes may issue a lot quicker than they complete, | |
586 | * whereas reads do not. | |
587 | */ | |
e47f799f | 588 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 589 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
590 | } |
591 | ||
592 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
593 | { | |
594 | unsigned int v; | |
e47f799f | 595 | |
1bd5d213 | 596 | v = rand_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 597 | |
04c540d9 | 598 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
599 | return DDIR_READ; |
600 | ||
601 | return DDIR_WRITE; | |
afe24a5a JA |
602 | } |
603 | ||
0d593542 | 604 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 605 | { |
0d593542 JA |
606 | int completed = 0; |
607 | ||
002e7183 JA |
608 | /* |
609 | * We are going to sleep, ensure that we flush anything pending as | |
610 | * not to skew our latency numbers. | |
611 | * | |
612 | * Changed to only monitor 'in flight' requests here instead of the | |
613 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
614 | * io's that have been actually submitted to an async engine, | |
615 | * and cur_depth is meaningless for sync engines. | |
616 | */ | |
a80cb54b BVA |
617 | if (td->io_u_queued || td->cur_depth) |
618 | td_io_commit(td); | |
9cc80b6d | 619 | |
002e7183 | 620 | while (td->io_u_in_flight) { |
89fd0573 | 621 | int ret; |
002e7183 | 622 | |
55312f9f | 623 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
624 | if (ret > 0) |
625 | completed += ret; | |
002e7183 | 626 | } |
0d593542 | 627 | |
6be06c46 JA |
628 | if (td->flags & TD_F_REGROW_LOGS) |
629 | regrow_logs(td); | |
630 | ||
0d593542 | 631 | return completed; |
002e7183 JA |
632 | } |
633 | ||
581e7141 JA |
634 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
635 | { | |
636 | enum fio_ddir odir = ddir ^ 1; | |
b5407f8b | 637 | uint64_t usec; |
90eff1c9 | 638 | uint64_t now; |
581e7141 | 639 | |
ff58fced | 640 | assert(ddir_rw(ddir)); |
50a8ce86 | 641 | now = utime_since_now(&td->start); |
ff58fced | 642 | |
50a8ce86 D |
643 | /* |
644 | * if rate_next_io_time is in the past, need to catch up to rate | |
645 | */ | |
646 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
647 | return ddir; |
648 | ||
649 | /* | |
50a8ce86 | 650 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
651 | * should switch. |
652 | */ | |
315fcfec | 653 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 654 | /* |
50a8ce86 | 655 | * Other direction is behind rate, switch |
581e7141 | 656 | */ |
50a8ce86 | 657 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
658 | return odir; |
659 | ||
660 | /* | |
395feabb JA |
661 | * Both directions are ahead of rate. sleep the min, |
662 | * switch if necessary | |
581e7141 | 663 | */ |
50a8ce86 | 664 | if (td->rate_next_io_time[ddir] <= |
395feabb | 665 | td->rate_next_io_time[odir]) { |
50a8ce86 | 666 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 667 | } else { |
50a8ce86 | 668 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
669 | ddir = odir; |
670 | } | |
671 | } else | |
50a8ce86 | 672 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 673 | |
a9da8ab2 JA |
674 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
675 | io_u_quiesce(td); | |
78c1eda5 | 676 | |
1a9bf814 | 677 | usec_sleep(td, usec); |
581e7141 JA |
678 | return ddir; |
679 | } | |
680 | ||
10ba535a JA |
681 | /* |
682 | * Return the data direction for the next io_u. If the job is a | |
683 | * mixed read/write workload, check the rwmix cycle and switch if | |
684 | * necessary. | |
685 | */ | |
1e97cce9 | 686 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 687 | { |
581e7141 JA |
688 | enum fio_ddir ddir; |
689 | ||
5f9099ea | 690 | /* |
f6860972 TK |
691 | * See if it's time to fsync/fdatasync/sync_file_range first, |
692 | * and if not then move on to check regular I/Os. | |
5f9099ea | 693 | */ |
f6860972 TK |
694 | if (should_fsync(td)) { |
695 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
696 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
697 | return DDIR_SYNC; | |
698 | ||
699 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
700 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
701 | return DDIR_DATASYNC; | |
702 | ||
703 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
704 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
705 | return DDIR_SYNC_FILE_RANGE; | |
706 | } | |
44f29692 | 707 | |
10ba535a | 708 | if (td_rw(td)) { |
10ba535a JA |
709 | /* |
710 | * Check if it's time to seed a new data direction. | |
711 | */ | |
e4928662 | 712 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
713 | /* |
714 | * Put a top limit on how many bytes we do for | |
715 | * one data direction, to avoid overflowing the | |
716 | * ranges too much | |
717 | */ | |
718 | ddir = get_rand_ddir(td); | |
e47f799f JA |
719 | |
720 | if (ddir != td->rwmix_ddir) | |
721 | set_rwmix_bytes(td); | |
722 | ||
723 | td->rwmix_ddir = ddir; | |
10ba535a | 724 | } |
581e7141 | 725 | ddir = td->rwmix_ddir; |
10ba535a | 726 | } else if (td_read(td)) |
581e7141 | 727 | ddir = DDIR_READ; |
6eaf09d6 | 728 | else if (td_write(td)) |
581e7141 | 729 | ddir = DDIR_WRITE; |
5c8e84ca | 730 | else if (td_trim(td)) |
6eaf09d6 | 731 | ddir = DDIR_TRIM; |
5c8e84ca TK |
732 | else |
733 | ddir = DDIR_INVAL; | |
581e7141 JA |
734 | |
735 | td->rwmix_ddir = rate_ddir(td, ddir); | |
736 | return td->rwmix_ddir; | |
10ba535a JA |
737 | } |
738 | ||
1ef2b6be JA |
739 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
740 | { | |
0e4dd95c DE |
741 | enum fio_ddir ddir = get_rw_ddir(td); |
742 | ||
82a90686 | 743 | if (td_trimwrite(td)) { |
0e4dd95c DE |
744 | struct fio_file *f = io_u->file; |
745 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
746 | ddir = DDIR_TRIM; | |
747 | else | |
748 | ddir = DDIR_WRITE; | |
749 | } | |
750 | ||
751 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 752 | |
9b87f09b | 753 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
754 | td->o.barrier_blocks && |
755 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
756 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 757 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
758 | } |
759 | ||
e8462bd8 | 760 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 761 | { |
71b84caa | 762 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
763 | |
764 | if (ret) | |
765 | td_verror(td, ret, "file close"); | |
766 | } | |
767 | ||
10ba535a JA |
768 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
769 | { | |
99952ca7 BVA |
770 | if (io_u->post_submit) { |
771 | io_u->post_submit(io_u, io_u->error == 0); | |
772 | io_u->post_submit = NULL; | |
773 | } | |
774 | ||
a9da8ab2 JA |
775 | if (td->parent) |
776 | td = td->parent; | |
777 | ||
e8462bd8 JA |
778 | td_io_u_lock(td); |
779 | ||
f8b0bd10 | 780 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 781 | put_file_log(td, io_u->file); |
f8b0bd10 | 782 | |
10ba535a | 783 | io_u->file = NULL; |
1651e431 | 784 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 785 | |
a9da8ab2 | 786 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 787 | td->cur_depth--; |
a9da8ab2 JA |
788 | assert(!(td->flags & TD_F_CHILD)); |
789 | } | |
2ae0b204 | 790 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 | 791 | td_io_u_free_notify(td); |
564de8d1 | 792 | td_io_u_unlock(td); |
10ba535a JA |
793 | } |
794 | ||
f2bba182 RR |
795 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
796 | { | |
1651e431 | 797 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
798 | put_io_u(td, io_u); |
799 | } | |
800 | ||
755200a3 JA |
801 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
802 | { | |
803 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 804 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 805 | |
465221b0 JA |
806 | dprint(FD_IO, "requeue %p\n", __io_u); |
807 | ||
a9da8ab2 JA |
808 | if (td->parent) |
809 | td = td->parent; | |
810 | ||
e8462bd8 JA |
811 | td_io_u_lock(td); |
812 | ||
1651e431 | 813 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
814 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
815 | td->io_issues[ddir]--; | |
5ec10eaa | 816 | |
1651e431 | 817 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 818 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 819 | td->cur_depth--; |
a9da8ab2 JA |
820 | assert(!(td->flags & TD_F_CHILD)); |
821 | } | |
2ae0b204 JA |
822 | |
823 | io_u_rpush(&td->io_u_requeues, __io_u); | |
a9da8ab2 | 824 | td_io_u_free_notify(td); |
564de8d1 | 825 | td_io_u_unlock(td); |
755200a3 JA |
826 | *io_u = NULL; |
827 | } | |
828 | ||
7b865a2f | 829 | static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u) |
224b3093 | 830 | { |
831 | struct fio_file *f = io_u->file; | |
832 | ||
7b865a2f BVA |
833 | assert(td->o.zone_mode == ZONE_MODE_STRIDED); |
834 | assert(td->o.zone_size); | |
835 | assert(td->o.zone_range); | |
836 | ||
224b3093 | 837 | /* |
838 | * See if it's time to switch to a new zone | |
839 | */ | |
840 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { | |
841 | td->zone_bytes = 0; | |
842 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
843 | ||
844 | /* | |
845 | * Wrap from the beginning, if we exceed the file size | |
846 | */ | |
847 | if (f->file_offset >= f->real_file_size) | |
04bc85a1 JA |
848 | f->file_offset = get_start_offset(td, f); |
849 | ||
224b3093 | 850 | f->last_pos[io_u->ddir] = f->file_offset; |
851 | td->io_skip_bytes += td->o.zone_skip; | |
852 | } | |
853 | ||
854 | /* | |
04bc85a1 JA |
855 | * If zone_size > zone_range, then maintain the same zone until |
856 | * zone_bytes >= zone_size. | |
857 | */ | |
224b3093 | 858 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
859 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
860 | f->file_offset, f->last_pos[io_u->ddir]); | |
861 | f->last_pos[io_u->ddir] = f->file_offset; | |
862 | } | |
863 | ||
864 | /* | |
865 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
866 | * map needs to be reset as it's done with zone_range everytime. | |
867 | */ | |
04bc85a1 | 868 | if ((td->zone_bytes % td->o.zone_range) == 0) |
224b3093 | 869 | fio_file_reset(td, f); |
224b3093 | 870 | } |
871 | ||
9bf2061e | 872 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 873 | { |
ec370c22 | 874 | bool is_random; |
6aca9b3d | 875 | |
9b87f09b | 876 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
877 | goto out; |
878 | ||
1ef2b6be | 879 | set_rw_ddir(td, io_u); |
5f9099ea | 880 | |
87dc1ab1 | 881 | /* |
ff58fced | 882 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 883 | */ |
ff58fced | 884 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 885 | goto out; |
a00735e6 | 886 | |
7b865a2f BVA |
887 | if (td->o.zone_mode == ZONE_MODE_STRIDED) |
888 | setup_strided_zone_mode(td, io_u); | |
48f5abd3 | 889 | |
10ba535a | 890 | /* |
c685b5b2 JA |
891 | * No log, let the seq/rand engine retrieve the next buflen and |
892 | * position. | |
10ba535a | 893 | */ |
6aca9b3d | 894 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 895 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 896 | return 1; |
2ba1c290 | 897 | } |
10ba535a | 898 | |
6aca9b3d | 899 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
900 | if (!io_u->buflen) { |
901 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 902 | return 1; |
2ba1c290 | 903 | } |
bca4ed4d | 904 | |
2ba1c290 | 905 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
5fff9543 | 906 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n", |
e5f9a813 | 907 | io_u, |
4b91ee8f JA |
908 | (unsigned long long) io_u->offset, io_u->buflen, |
909 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 910 | return 1; |
2ba1c290 | 911 | } |
6a5e6884 | 912 | |
bca4ed4d JA |
913 | /* |
914 | * mark entry before potentially trimming io_u | |
915 | */ | |
303032ae | 916 | if (td_random(td) && file_randommap(td, io_u->file)) |
6cc1a3d1 BVA |
917 | io_u->buflen = mark_random_map(td, io_u, io_u->offset, |
918 | io_u->buflen); | |
bca4ed4d | 919 | |
c38e9468 | 920 | out: |
e5f9a813 | 921 | dprint_io_u(io_u, "fill"); |
d9d91e39 | 922 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 923 | return 0; |
10ba535a JA |
924 | } |
925 | ||
6cc0e5aa | 926 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
838bc709 | 927 | { |
2b13e716 | 928 | int idx = 0; |
838bc709 JA |
929 | |
930 | switch (nr) { | |
931 | default: | |
2b13e716 | 932 | idx = 6; |
838bc709 JA |
933 | break; |
934 | case 33 ... 64: | |
2b13e716 | 935 | idx = 5; |
838bc709 JA |
936 | break; |
937 | case 17 ... 32: | |
2b13e716 | 938 | idx = 4; |
838bc709 JA |
939 | break; |
940 | case 9 ... 16: | |
2b13e716 | 941 | idx = 3; |
838bc709 JA |
942 | break; |
943 | case 5 ... 8: | |
2b13e716 | 944 | idx = 2; |
838bc709 JA |
945 | break; |
946 | case 1 ... 4: | |
2b13e716 | 947 | idx = 1; |
838bc709 JA |
948 | case 0: |
949 | break; | |
950 | } | |
951 | ||
2b13e716 | 952 | map[idx]++; |
838bc709 JA |
953 | } |
954 | ||
955 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
956 | { | |
957 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
958 | td->ts.total_submit++; | |
959 | } | |
960 | ||
961 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
962 | { | |
963 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
964 | td->ts.total_complete++; | |
965 | } | |
966 | ||
d8005759 | 967 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 968 | { |
2b13e716 | 969 | int idx = 0; |
71619dc2 JA |
970 | |
971 | switch (td->cur_depth) { | |
972 | default: | |
2b13e716 | 973 | idx = 6; |
a783e61a | 974 | break; |
71619dc2 | 975 | case 32 ... 63: |
2b13e716 | 976 | idx = 5; |
a783e61a | 977 | break; |
71619dc2 | 978 | case 16 ... 31: |
2b13e716 | 979 | idx = 4; |
a783e61a | 980 | break; |
71619dc2 | 981 | case 8 ... 15: |
2b13e716 | 982 | idx = 3; |
a783e61a | 983 | break; |
71619dc2 | 984 | case 4 ... 7: |
2b13e716 | 985 | idx = 2; |
a783e61a | 986 | break; |
71619dc2 | 987 | case 2 ... 3: |
2b13e716 | 988 | idx = 1; |
71619dc2 JA |
989 | case 1: |
990 | break; | |
991 | } | |
992 | ||
2b13e716 | 993 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
994 | } |
995 | ||
d6bb626e | 996 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 997 | { |
2b13e716 | 998 | int idx = 0; |
04a0feae | 999 | |
d6bb626e VF |
1000 | assert(nsec < 1000); |
1001 | ||
1002 | switch (nsec) { | |
1003 | case 750 ... 999: | |
1004 | idx = 9; | |
1005 | break; | |
1006 | case 500 ... 749: | |
1007 | idx = 8; | |
1008 | break; | |
1009 | case 250 ... 499: | |
1010 | idx = 7; | |
1011 | break; | |
1012 | case 100 ... 249: | |
1013 | idx = 6; | |
1014 | break; | |
1015 | case 50 ... 99: | |
1016 | idx = 5; | |
1017 | break; | |
1018 | case 20 ... 49: | |
1019 | idx = 4; | |
1020 | break; | |
1021 | case 10 ... 19: | |
1022 | idx = 3; | |
1023 | break; | |
1024 | case 4 ... 9: | |
1025 | idx = 2; | |
1026 | break; | |
1027 | case 2 ... 3: | |
1028 | idx = 1; | |
1029 | case 0 ... 1: | |
1030 | break; | |
1031 | } | |
1032 | ||
1033 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1034 | td->ts.io_u_lat_n[idx]++; | |
1035 | } | |
1036 | ||
1037 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1038 | { | |
1039 | int idx = 0; | |
1040 | ||
1041 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1042 | |
1043 | switch (usec) { | |
1044 | case 750 ... 999: | |
2b13e716 | 1045 | idx = 9; |
04a0feae JA |
1046 | break; |
1047 | case 500 ... 749: | |
2b13e716 | 1048 | idx = 8; |
04a0feae JA |
1049 | break; |
1050 | case 250 ... 499: | |
2b13e716 | 1051 | idx = 7; |
04a0feae JA |
1052 | break; |
1053 | case 100 ... 249: | |
2b13e716 | 1054 | idx = 6; |
04a0feae JA |
1055 | break; |
1056 | case 50 ... 99: | |
2b13e716 | 1057 | idx = 5; |
04a0feae JA |
1058 | break; |
1059 | case 20 ... 49: | |
2b13e716 | 1060 | idx = 4; |
04a0feae JA |
1061 | break; |
1062 | case 10 ... 19: | |
2b13e716 | 1063 | idx = 3; |
04a0feae JA |
1064 | break; |
1065 | case 4 ... 9: | |
2b13e716 | 1066 | idx = 2; |
04a0feae JA |
1067 | break; |
1068 | case 2 ... 3: | |
2b13e716 | 1069 | idx = 1; |
04a0feae JA |
1070 | case 0 ... 1: |
1071 | break; | |
1072 | } | |
1073 | ||
2b13e716 JA |
1074 | assert(idx < FIO_IO_U_LAT_U_NR); |
1075 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1076 | } |
1077 | ||
d6bb626e | 1078 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1079 | { |
2b13e716 | 1080 | int idx = 0; |
ec118304 | 1081 | |
d6bb626e VF |
1082 | assert(msec >= 1); |
1083 | ||
ec118304 JA |
1084 | switch (msec) { |
1085 | default: | |
2b13e716 | 1086 | idx = 11; |
04a0feae | 1087 | break; |
8abdce66 | 1088 | case 1000 ... 1999: |
2b13e716 | 1089 | idx = 10; |
04a0feae | 1090 | break; |
8abdce66 | 1091 | case 750 ... 999: |
2b13e716 | 1092 | idx = 9; |
04a0feae | 1093 | break; |
8abdce66 | 1094 | case 500 ... 749: |
2b13e716 | 1095 | idx = 8; |
04a0feae | 1096 | break; |
8abdce66 | 1097 | case 250 ... 499: |
2b13e716 | 1098 | idx = 7; |
04a0feae | 1099 | break; |
8abdce66 | 1100 | case 100 ... 249: |
2b13e716 | 1101 | idx = 6; |
04a0feae | 1102 | break; |
8abdce66 | 1103 | case 50 ... 99: |
2b13e716 | 1104 | idx = 5; |
04a0feae | 1105 | break; |
8abdce66 | 1106 | case 20 ... 49: |
2b13e716 | 1107 | idx = 4; |
04a0feae | 1108 | break; |
8abdce66 | 1109 | case 10 ... 19: |
2b13e716 | 1110 | idx = 3; |
04a0feae | 1111 | break; |
8abdce66 | 1112 | case 4 ... 9: |
2b13e716 | 1113 | idx = 2; |
04a0feae | 1114 | break; |
ec118304 | 1115 | case 2 ... 3: |
2b13e716 | 1116 | idx = 1; |
ec118304 JA |
1117 | case 0 ... 1: |
1118 | break; | |
1119 | } | |
1120 | ||
2b13e716 JA |
1121 | assert(idx < FIO_IO_U_LAT_M_NR); |
1122 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1123 | } |
1124 | ||
d6bb626e | 1125 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1126 | { |
d6bb626e VF |
1127 | if (nsec < 1000) |
1128 | io_u_mark_lat_nsec(td, nsec); | |
1129 | else if (nsec < 1000000) | |
1130 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1131 | else |
d6bb626e | 1132 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1133 | } |
1134 | ||
8c07860d JA |
1135 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1136 | { | |
1137 | unsigned long fileno; | |
1138 | ||
1139 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1140 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1141 | unsigned long r; | |
1142 | ||
1143 | r = __rand(&td->next_file_state); | |
1144 | return (unsigned int) ((double) td->o.nr_files | |
1145 | * (r / (frand_max + 1.0))); | |
1146 | } | |
1147 | ||
1148 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1149 | fileno = zipf_next(&td->next_file_zipf); | |
1150 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1151 | fileno = pareto_next(&td->next_file_zipf); | |
1152 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1153 | fileno = gauss_next(&td->next_file_gauss); | |
1154 | else { | |
1155 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1156 | assert(0); | |
1157 | return 0; | |
1158 | } | |
1159 | ||
1160 | return fileno >> FIO_FSERVICE_SHIFT; | |
1161 | } | |
1162 | ||
0aabe160 JA |
1163 | /* |
1164 | * Get next file to service by choosing one at random | |
1165 | */ | |
2cc52930 JA |
1166 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1167 | enum fio_file_flags goodf, | |
d6aed795 | 1168 | enum fio_file_flags badf) |
0aabe160 | 1169 | { |
0aabe160 | 1170 | struct fio_file *f; |
1c178180 | 1171 | int fno; |
0aabe160 JA |
1172 | |
1173 | do { | |
87b10676 | 1174 | int opened = 0; |
4c07ad86 | 1175 | |
8c07860d | 1176 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1177 | |
126d65c6 | 1178 | f = td->files[fno]; |
d6aed795 | 1179 | if (fio_file_done(f)) |
059e63c0 | 1180 | continue; |
1c178180 | 1181 | |
d6aed795 | 1182 | if (!fio_file_open(f)) { |
87b10676 JA |
1183 | int err; |
1184 | ||
002fe734 JA |
1185 | if (td->nr_open_files >= td->o.open_files) |
1186 | return ERR_PTR(-EBUSY); | |
1187 | ||
87b10676 JA |
1188 | err = td_io_open_file(td, f); |
1189 | if (err) | |
1190 | continue; | |
1191 | opened = 1; | |
1192 | } | |
1193 | ||
2ba1c290 JA |
1194 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1195 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1196 | return f; |
2ba1c290 | 1197 | } |
87b10676 JA |
1198 | if (opened) |
1199 | td_io_close_file(td, f); | |
0aabe160 JA |
1200 | } while (1); |
1201 | } | |
1202 | ||
1203 | /* | |
1204 | * Get next file to service by doing round robin between all available ones | |
1205 | */ | |
1c178180 JA |
1206 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1207 | int badf) | |
3d7c391d JA |
1208 | { |
1209 | unsigned int old_next_file = td->next_file; | |
1210 | struct fio_file *f; | |
1211 | ||
1212 | do { | |
87b10676 JA |
1213 | int opened = 0; |
1214 | ||
126d65c6 | 1215 | f = td->files[td->next_file]; |
3d7c391d JA |
1216 | |
1217 | td->next_file++; | |
2dc1bbeb | 1218 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1219 | td->next_file = 0; |
1220 | ||
87b10676 | 1221 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1222 | if (fio_file_done(f)) { |
d5ed68ea | 1223 | f = NULL; |
059e63c0 | 1224 | continue; |
d5ed68ea | 1225 | } |
059e63c0 | 1226 | |
d6aed795 | 1227 | if (!fio_file_open(f)) { |
87b10676 JA |
1228 | int err; |
1229 | ||
002fe734 JA |
1230 | if (td->nr_open_files >= td->o.open_files) |
1231 | return ERR_PTR(-EBUSY); | |
1232 | ||
87b10676 | 1233 | err = td_io_open_file(td, f); |
b5696bfc JA |
1234 | if (err) { |
1235 | dprint(FD_FILE, "error %d on open of %s\n", | |
1236 | err, f->file_name); | |
87c27b45 | 1237 | f = NULL; |
87b10676 | 1238 | continue; |
b5696bfc | 1239 | } |
87b10676 JA |
1240 | opened = 1; |
1241 | } | |
1242 | ||
0b9d69ec JA |
1243 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1244 | f->flags); | |
1c178180 | 1245 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1246 | break; |
1247 | ||
87b10676 JA |
1248 | if (opened) |
1249 | td_io_close_file(td, f); | |
1250 | ||
3d7c391d JA |
1251 | f = NULL; |
1252 | } while (td->next_file != old_next_file); | |
1253 | ||
2ba1c290 | 1254 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1255 | return f; |
1256 | } | |
1257 | ||
7eb36574 | 1258 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1259 | { |
1907dbc6 JA |
1260 | struct fio_file *f; |
1261 | ||
2dc1bbeb | 1262 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1263 | |
b5696bfc | 1264 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1265 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1266 | " nr_files=%d\n", td->nr_open_files, | |
1267 | td->nr_done_files, | |
1268 | td->o.nr_files); | |
bdb4e2e9 | 1269 | return NULL; |
2ba1c290 | 1270 | } |
bdb4e2e9 | 1271 | |
1907dbc6 | 1272 | f = td->file_service_file; |
d6aed795 | 1273 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1274 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1275 | goto out; | |
1276 | if (td->file_service_left--) | |
1277 | goto out; | |
1278 | } | |
1907dbc6 | 1279 | |
a086c257 JA |
1280 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1281 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1282 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1283 | else |
d6aed795 | 1284 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1285 | |
002fe734 JA |
1286 | if (IS_ERR(f)) |
1287 | return f; | |
1288 | ||
1907dbc6 JA |
1289 | td->file_service_file = f; |
1290 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1291 | out: |
0dac421f JA |
1292 | if (f) |
1293 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1294 | else | |
1295 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1296 | return f; |
bdb4e2e9 JA |
1297 | } |
1298 | ||
7eb36574 JA |
1299 | static struct fio_file *get_next_file(struct thread_data *td) |
1300 | { | |
7eb36574 JA |
1301 | return __get_next_file(td); |
1302 | } | |
1303 | ||
002fe734 | 1304 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1305 | { |
1306 | struct fio_file *f; | |
1307 | ||
1308 | do { | |
1309 | f = get_next_file(td); | |
002fe734 JA |
1310 | if (IS_ERR_OR_NULL(f)) |
1311 | return PTR_ERR(f); | |
429f6675 | 1312 | |
429f6675 JA |
1313 | io_u->file = f; |
1314 | get_file(f); | |
1315 | ||
1316 | if (!fill_io_u(td, io_u)) | |
1317 | break; | |
1318 | ||
99952ca7 BVA |
1319 | if (io_u->post_submit) { |
1320 | io_u->post_submit(io_u, false); | |
1321 | io_u->post_submit = NULL; | |
1322 | } | |
1323 | ||
b5696bfc | 1324 | put_file_log(td, f); |
429f6675 | 1325 | td_io_close_file(td, f); |
b5696bfc | 1326 | io_u->file = NULL; |
8c07860d JA |
1327 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1328 | fio_file_reset(td, f); | |
1329 | else { | |
1330 | fio_file_set_done(f); | |
1331 | td->nr_done_files++; | |
1332 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1333 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1334 | } |
429f6675 JA |
1335 | } while (1); |
1336 | ||
1337 | return 0; | |
1338 | } | |
1339 | ||
3e260a46 | 1340 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
c3a32714 | 1341 | unsigned long long tnsec, unsigned long long max_nsec) |
3e260a46 JA |
1342 | { |
1343 | if (!td->error) | |
c3a32714 | 1344 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
3e260a46 JA |
1345 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
1346 | icd->error = ETIMEDOUT; | |
1347 | } | |
1348 | ||
1349 | static void lat_new_cycle(struct thread_data *td) | |
1350 | { | |
1351 | fio_gettime(&td->latency_ts, NULL); | |
1352 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1353 | td->latency_failed = 0; | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1358 | * are at QD=1, then it's time to give up. | |
1359 | */ | |
e39c0676 | 1360 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1361 | { |
1362 | if (td->latency_qd == 1) | |
e39c0676 | 1363 | return true; |
3e260a46 JA |
1364 | |
1365 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1366 | |
1367 | if (td->latency_qd == td->latency_qd_low) | |
1368 | td->latency_qd_low--; | |
1369 | ||
3e260a46 JA |
1370 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1371 | ||
1372 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1373 | ||
1374 | /* | |
1375 | * When we ramp QD down, quiesce existing IO to prevent | |
1376 | * a storm of ramp downs due to pending higher depth. | |
1377 | */ | |
1378 | io_u_quiesce(td); | |
1379 | lat_new_cycle(td); | |
e39c0676 | 1380 | return false; |
3e260a46 JA |
1381 | } |
1382 | ||
e39c0676 | 1383 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1384 | { |
1385 | if (td->o.latency_percentile.u.f == 100.0) | |
1386 | return __lat_target_failed(td); | |
1387 | ||
1388 | td->latency_failed++; | |
e39c0676 | 1389 | return false; |
3e260a46 JA |
1390 | } |
1391 | ||
1392 | void lat_target_init(struct thread_data *td) | |
1393 | { | |
6bb58215 JA |
1394 | td->latency_end_run = 0; |
1395 | ||
3e260a46 JA |
1396 | if (td->o.latency_target) { |
1397 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1398 | fio_gettime(&td->latency_ts, NULL); | |
1399 | td->latency_qd = 1; | |
1400 | td->latency_qd_high = td->o.iodepth; | |
1401 | td->latency_qd_low = 1; | |
1402 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1403 | } else | |
1404 | td->latency_qd = td->o.iodepth; | |
1405 | } | |
1406 | ||
6bb58215 JA |
1407 | void lat_target_reset(struct thread_data *td) |
1408 | { | |
1409 | if (!td->latency_end_run) | |
1410 | lat_target_init(td); | |
1411 | } | |
1412 | ||
3e260a46 JA |
1413 | static void lat_target_success(struct thread_data *td) |
1414 | { | |
1415 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1416 | struct thread_options *o = &td->o; |
3e260a46 JA |
1417 | |
1418 | td->latency_qd_low = td->latency_qd; | |
1419 | ||
1420 | /* | |
1421 | * If we haven't failed yet, we double up to a failing value instead | |
1422 | * of bisecting from highest possible queue depth. If we have set | |
1423 | * a limit other than td->o.iodepth, bisect between that. | |
1424 | */ | |
6bb58215 | 1425 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1426 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1427 | else | |
1428 | td->latency_qd *= 2; | |
1429 | ||
6bb58215 JA |
1430 | if (td->latency_qd > o->iodepth) |
1431 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1432 | |
1433 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1434 | |
3e260a46 | 1435 | /* |
6bb58215 JA |
1436 | * Same as last one, we are done. Let it run a latency cycle, so |
1437 | * we get only the results from the targeted depth. | |
3e260a46 | 1438 | */ |
6bb58215 JA |
1439 | if (td->latency_qd == qd) { |
1440 | if (td->latency_end_run) { | |
1441 | dprint(FD_RATE, "We are done\n"); | |
1442 | td->done = 1; | |
1443 | } else { | |
1444 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1445 | io_u_quiesce(td); | |
1446 | td->latency_end_run = 1; | |
1447 | reset_all_stats(td); | |
1448 | reset_io_stats(td); | |
1449 | } | |
1450 | } | |
3e260a46 JA |
1451 | |
1452 | lat_new_cycle(td); | |
1453 | } | |
1454 | ||
1455 | /* | |
1456 | * Check if we can bump the queue depth | |
1457 | */ | |
1458 | void lat_target_check(struct thread_data *td) | |
1459 | { | |
1460 | uint64_t usec_window; | |
1461 | uint64_t ios; | |
1462 | double success_ios; | |
1463 | ||
1464 | usec_window = utime_since_now(&td->latency_ts); | |
1465 | if (usec_window < td->o.latency_window) | |
1466 | return; | |
1467 | ||
1468 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1469 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1470 | success_ios *= 100.0; | |
1471 | ||
1472 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1473 | ||
1474 | if (success_ios >= td->o.latency_percentile.u.f) | |
1475 | lat_target_success(td); | |
1476 | else | |
1477 | __lat_target_failed(td); | |
1478 | } | |
1479 | ||
1480 | /* | |
1481 | * If latency target is enabled, we might be ramping up or down and not | |
1482 | * using the full queue depth available. | |
1483 | */ | |
e39c0676 | 1484 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1485 | { |
1486 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1487 | ||
1488 | if (qempty) | |
e39c0676 | 1489 | return true; |
3e260a46 | 1490 | if (!td->o.latency_target) |
e39c0676 | 1491 | return false; |
3e260a46 JA |
1492 | |
1493 | return td->cur_depth >= td->latency_qd; | |
1494 | } | |
429f6675 | 1495 | |
10ba535a JA |
1496 | struct io_u *__get_io_u(struct thread_data *td) |
1497 | { | |
0cae66f6 | 1498 | struct io_u *io_u = NULL; |
93b45bb2 | 1499 | int ret; |
10ba535a | 1500 | |
ca09be4b JA |
1501 | if (td->stop_io) |
1502 | return NULL; | |
1503 | ||
e8462bd8 JA |
1504 | td_io_u_lock(td); |
1505 | ||
1506 | again: | |
2ae0b204 JA |
1507 | if (!io_u_rempty(&td->io_u_requeues)) |
1508 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1509 | else if (!queue_full(td)) { |
2ae0b204 | 1510 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1511 | |
225ba9e3 | 1512 | io_u->file = NULL; |
6040dabc | 1513 | io_u->buflen = 0; |
10ba535a | 1514 | io_u->resid = 0; |
d7762cf8 | 1515 | io_u->end_io = NULL; |
755200a3 JA |
1516 | } |
1517 | ||
1518 | if (io_u) { | |
0c6e7517 | 1519 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1520 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1521 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1522 | IO_U_F_VER_LIST); | |
0c6e7517 | 1523 | |
755200a3 | 1524 | io_u->error = 0; |
bcd5abfa | 1525 | io_u->acct_ddir = -1; |
10ba535a | 1526 | td->cur_depth++; |
a9da8ab2 | 1527 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1528 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1529 | io_u->ipo = NULL; |
a9da8ab2 | 1530 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1531 | /* |
1532 | * We ran out, wait for async verify threads to finish and | |
1533 | * return one | |
1534 | */ | |
a9da8ab2 | 1535 | assert(!(td->flags & TD_F_CHILD)); |
93b45bb2 BVA |
1536 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
1537 | assert(ret == 0); | |
1dec3e07 | 1538 | goto again; |
10ba535a JA |
1539 | } |
1540 | ||
e8462bd8 | 1541 | td_io_u_unlock(td); |
10ba535a JA |
1542 | return io_u; |
1543 | } | |
1544 | ||
e39c0676 | 1545 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1546 | { |
d72be545 | 1547 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1548 | return false; |
c9a73054 JA |
1549 | if (!td->trim_entries) |
1550 | return false; | |
d72be545 | 1551 | |
c9a73054 JA |
1552 | if (td->trim_batch) { |
1553 | td->trim_batch--; | |
1554 | if (get_next_trim(td, io_u)) | |
1555 | return true; | |
1556 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1557 | td->last_ddir != DDIR_READ) { | |
1558 | td->trim_batch = td->o.trim_batch; | |
1559 | if (!td->trim_batch) | |
1560 | td->trim_batch = td->o.trim_backlog; | |
1561 | if (get_next_trim(td, io_u)) | |
e39c0676 | 1562 | return true; |
2ba1c290 | 1563 | } |
10ba535a | 1564 | |
e39c0676 | 1565 | return false; |
0d29de83 JA |
1566 | } |
1567 | ||
e39c0676 | 1568 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1569 | { |
d72be545 | 1570 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1571 | return false; |
d72be545 JA |
1572 | |
1573 | if (td->io_hist_len) { | |
9e144189 JA |
1574 | int get_verify = 0; |
1575 | ||
d1ece0c7 | 1576 | if (td->verify_batch) |
9e144189 | 1577 | get_verify = 1; |
d1ece0c7 | 1578 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1579 | td->last_ddir != DDIR_READ) { |
1580 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1581 | if (!td->verify_batch) |
1582 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1583 | get_verify = 1; |
1584 | } | |
1585 | ||
d1ece0c7 JA |
1586 | if (get_verify && !get_next_verify(td, io_u)) { |
1587 | td->verify_batch--; | |
e39c0676 | 1588 | return true; |
d1ece0c7 | 1589 | } |
9e144189 JA |
1590 | } |
1591 | ||
e39c0676 | 1592 | return false; |
0d29de83 JA |
1593 | } |
1594 | ||
de789769 JA |
1595 | /* |
1596 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1597 | * easy compressible data for simple de-dupe attempts. Do this for every |
1598 | * 512b block in the range, since that should be the smallest block size | |
1599 | * we can expect from a device. | |
de789769 JA |
1600 | */ |
1601 | static void small_content_scramble(struct io_u *io_u) | |
1602 | { | |
5fff9543 | 1603 | unsigned long long i, nr_blocks = io_u->buflen >> 9; |
23f394d5 | 1604 | unsigned int offset; |
319b073f JA |
1605 | uint64_t boffset, *iptr; |
1606 | char *p; | |
de789769 | 1607 | |
23f394d5 JA |
1608 | if (!nr_blocks) |
1609 | return; | |
1610 | ||
1611 | p = io_u->xfer_buf; | |
fba76ee8 | 1612 | boffset = io_u->offset; |
319b073f JA |
1613 | |
1614 | if (io_u->buf_filled_len) | |
1615 | io_u->buf_filled_len = 0; | |
1616 | ||
1617 | /* | |
1618 | * Generate random index between 0..7. We do chunks of 512b, if | |
1619 | * we assume a cacheline is 64 bytes, then we have 8 of those. | |
1620 | * Scramble content within the blocks in the same cacheline to | |
1621 | * speed things up. | |
1622 | */ | |
1623 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; | |
fad82f76 | 1624 | |
23f394d5 JA |
1625 | for (i = 0; i < nr_blocks; i++) { |
1626 | /* | |
319b073f JA |
1627 | * Fill offset into start of cacheline, time into end |
1628 | * of cacheline | |
23f394d5 | 1629 | */ |
319b073f JA |
1630 | iptr = (void *) p + (offset << 6); |
1631 | *iptr = boffset; | |
1632 | ||
1633 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); | |
1634 | iptr[0] = io_u->start_time.tv_sec; | |
1635 | iptr[1] = io_u->start_time.tv_nsec; | |
23f394d5 | 1636 | |
23f394d5 | 1637 | p += 512; |
fad82f76 | 1638 | boffset += 512; |
23f394d5 | 1639 | } |
de789769 JA |
1640 | } |
1641 | ||
0d29de83 JA |
1642 | /* |
1643 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
5c5c33c1 | 1644 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
0d29de83 JA |
1645 | */ |
1646 | struct io_u *get_io_u(struct thread_data *td) | |
1647 | { | |
1648 | struct fio_file *f; | |
1649 | struct io_u *io_u; | |
de789769 | 1650 | int do_scramble = 0; |
002fe734 | 1651 | long ret = 0; |
0d29de83 JA |
1652 | |
1653 | io_u = __get_io_u(td); | |
1654 | if (!io_u) { | |
1655 | dprint(FD_IO, "__get_io_u failed\n"); | |
1656 | return NULL; | |
1657 | } | |
1658 | ||
1659 | if (check_get_verify(td, io_u)) | |
1660 | goto out; | |
1661 | if (check_get_trim(td, io_u)) | |
1662 | goto out; | |
1663 | ||
755200a3 JA |
1664 | /* |
1665 | * from a requeue, io_u already setup | |
1666 | */ | |
1667 | if (io_u->file) | |
77f392bf | 1668 | goto out; |
755200a3 | 1669 | |
429f6675 JA |
1670 | /* |
1671 | * If using an iolog, grab next piece if any available. | |
1672 | */ | |
d72be545 | 1673 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1674 | if (read_iolog_get(td, io_u)) |
1675 | goto err_put; | |
2ba1c290 | 1676 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1677 | ret = -EBUSY; |
2ba1c290 | 1678 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1679 | goto err_put; |
2ba1c290 | 1680 | } |
5ec10eaa | 1681 | |
429f6675 | 1682 | f = io_u->file; |
002fe734 JA |
1683 | if (!f) { |
1684 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1685 | goto err_put; | |
1686 | } | |
1687 | ||
d6aed795 | 1688 | assert(fio_file_open(f)); |
97af62ce | 1689 | |
ff58fced | 1690 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1691 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1692 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1693 | goto err_put; |
2ba1c290 | 1694 | } |
10ba535a | 1695 | |
f1dfb668 JA |
1696 | f->last_start[io_u->ddir] = io_u->offset; |
1697 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1698 | |
fd68418e | 1699 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1700 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1701 | io_u_fill_buffer(td, io_u, |
1066358a | 1702 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1703 | io_u->buflen); |
ff441ae8 | 1704 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
5c5c33c1 BVA |
1705 | !(td->flags & TD_F_COMPRESS) && |
1706 | !(td->flags & TD_F_DO_VERIFY)) | |
fd68418e JA |
1707 | do_scramble = 1; |
1708 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1709 | /* |
1710 | * Reset the buf_filled parameters so next time if the | |
1711 | * buffer is used for writes it is refilled. | |
1712 | */ | |
cbe8d756 RR |
1713 | io_u->buf_filled_len = 0; |
1714 | } | |
87dc1ab1 | 1715 | } |
10ba535a | 1716 | |
165faf16 JA |
1717 | /* |
1718 | * Set io data pointers. | |
1719 | */ | |
cec6b55d JA |
1720 | io_u->xfer_buf = io_u->buf; |
1721 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1722 | |
6ac7a331 | 1723 | out: |
0d29de83 | 1724 | assert(io_u->file); |
429f6675 | 1725 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1726 | if (!td->o.disable_lat) |
993bf48b | 1727 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1728 | |
de789769 JA |
1729 | if (do_scramble) |
1730 | small_content_scramble(io_u); | |
03553853 | 1731 | |
429f6675 | 1732 | return io_u; |
36167d82 | 1733 | } |
429f6675 | 1734 | err_put: |
2ba1c290 | 1735 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1736 | put_io_u(td, io_u); |
002fe734 | 1737 | return ERR_PTR(ret); |
10ba535a JA |
1738 | } |
1739 | ||
a9da8ab2 | 1740 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1741 | { |
8b28bd41 | 1742 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1743 | |
8b28bd41 DM |
1744 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1745 | return; | |
5451792e | 1746 | |
5fff9543 | 1747 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n", |
709c8313 RE |
1748 | io_u->file ? " on file " : "", |
1749 | io_u->file ? io_u->file->file_name : "", | |
1750 | strerror(io_u->error), | |
1751 | io_ddir_name(io_u->ddir), | |
1752 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1753 | |
5ad7be56 KD |
1754 | if (td->io_ops->errdetails) { |
1755 | char *err = td->io_ops->errdetails(io_u); | |
1756 | ||
1757 | log_err("fio: %s\n", err); | |
1758 | free(err); | |
1759 | } | |
1760 | ||
5451792e JA |
1761 | if (!td->error) |
1762 | td_verror(td, io_u->error, "io_u error"); | |
1763 | } | |
1764 | ||
a9da8ab2 JA |
1765 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1766 | { | |
1767 | __io_u_log_error(td, io_u); | |
1768 | if (td->parent) | |
094e66cb | 1769 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1770 | } |
1771 | ||
e39c0676 | 1772 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1773 | { |
3ecc8c67 JA |
1774 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1775 | || td->o.gtod_reduce; | |
aba6c951 JA |
1776 | } |
1777 | ||
c8eeb9df JA |
1778 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1779 | struct io_completion_data *icd, | |
1780 | const enum fio_ddir idx, unsigned int bytes) | |
1781 | { | |
a9da8ab2 | 1782 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1783 | unsigned long long llnsec = 0; |
c8eeb9df | 1784 | |
75dc383e JA |
1785 | if (td->parent) |
1786 | td = td->parent; | |
1787 | ||
132b1ee4 | 1788 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1789 | return; |
1790 | ||
a9da8ab2 | 1791 | if (no_reduce) |
d6bb626e | 1792 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1793 | |
1794 | if (!td->o.disable_lat) { | |
c3a32714 | 1795 | unsigned long long tnsec; |
c8eeb9df | 1796 | |
d6bb626e VF |
1797 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
1798 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); | |
15501535 | 1799 | |
d4afedfd JA |
1800 | if (td->flags & TD_F_PROFILE_OPS) { |
1801 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1802 | ||
1803 | if (ops->io_u_lat) | |
c3a32714 | 1804 | icd->error = ops->io_u_lat(td, tnsec); |
d4afedfd JA |
1805 | } |
1806 | ||
c3a32714 JA |
1807 | if (td->o.max_latency && tnsec > td->o.max_latency) |
1808 | lat_fatal(td, icd, tnsec, td->o.max_latency); | |
1809 | if (td->o.latency_target && tnsec > td->o.latency_target) { | |
3e260a46 | 1810 | if (lat_target_failed(td)) |
c3a32714 | 1811 | lat_fatal(td, icd, tnsec, td->o.latency_target); |
15501535 | 1812 | } |
c8eeb9df JA |
1813 | } |
1814 | ||
a47591e4 JA |
1815 | if (ddir_rw(idx)) { |
1816 | if (!td->o.disable_clat) { | |
d6bb626e VF |
1817 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
1818 | io_u_mark_latency(td, llnsec); | |
a47591e4 | 1819 | } |
c8eeb9df | 1820 | |
a47591e4 | 1821 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1822 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1823 | |
a47591e4 JA |
1824 | if (no_reduce && per_unit_log(td->iops_log)) |
1825 | add_iops_sample(td, io_u, bytes); | |
b2b3eefe JA |
1826 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
1827 | add_sync_clat_sample(&td->ts, llnsec); | |
66347cfa DE |
1828 | |
1829 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1830 | uint32_t *info = io_u_block_info(td, io_u); | |
1831 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1832 | if (io_u->ddir == DDIR_TRIM) { | |
1833 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1834 | BLOCK_INFO_TRIMS(*info) + 1); | |
1835 | } else if (io_u->ddir == DDIR_WRITE) { | |
1836 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1837 | *info); | |
1838 | } | |
1839 | } | |
1840 | } | |
c8eeb9df JA |
1841 | } |
1842 | ||
94a6e1bb JA |
1843 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1844 | uint64_t offset, unsigned int bytes) | |
1845 | { | |
1846 | int idx; | |
1847 | ||
639ad1ea JA |
1848 | if (!f) |
1849 | return; | |
1850 | ||
94a6e1bb JA |
1851 | if (f->first_write == -1ULL || offset < f->first_write) |
1852 | f->first_write = offset; | |
1853 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1854 | f->last_write = offset + bytes; | |
1855 | ||
1856 | if (!f->last_write_comp) | |
1857 | return; | |
1858 | ||
1859 | idx = f->last_write_idx++; | |
1860 | f->last_write_comp[idx] = offset; | |
1861 | if (f->last_write_idx == td->o.iodepth) | |
1862 | f->last_write_idx = 0; | |
1863 | } | |
1864 | ||
b2b3eefe JA |
1865 | static bool should_account(struct thread_data *td) |
1866 | { | |
1867 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || | |
1868 | td->runstate == TD_VERIFYING); | |
1869 | } | |
1870 | ||
f8b0bd10 | 1871 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1872 | struct io_completion_data *icd) |
10ba535a | 1873 | { |
f8b0bd10 JA |
1874 | struct io_u *io_u = *io_u_ptr; |
1875 | enum fio_ddir ddir = io_u->ddir; | |
1876 | struct fio_file *f = io_u->file; | |
10ba535a | 1877 | |
e5f9a813 | 1878 | dprint_io_u(io_u, "complete"); |
2ba1c290 | 1879 | |
0c6e7517 | 1880 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1881 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1882 | |
1883 | /* | |
1884 | * Mark IO ok to verify | |
1885 | */ | |
1886 | if (io_u->ipo) { | |
890b6656 JA |
1887 | /* |
1888 | * Remove errored entry from the verification list | |
1889 | */ | |
1890 | if (io_u->error) | |
1891 | unlog_io_piece(td, io_u); | |
1892 | else { | |
1893 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1894 | write_barrier(); | |
1895 | } | |
f9401285 JA |
1896 | } |
1897 | ||
f8b0bd10 | 1898 | if (ddir_sync(ddir)) { |
2ea93f98 | 1899 | td->last_was_sync = true; |
44f29692 JA |
1900 | if (f) { |
1901 | f->first_write = -1ULL; | |
1902 | f->last_write = -1ULL; | |
1903 | } | |
b2b3eefe JA |
1904 | if (should_account(td)) |
1905 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); | |
87dc1ab1 JA |
1906 | return; |
1907 | } | |
1908 | ||
2ea93f98 | 1909 | td->last_was_sync = false; |
f8b0bd10 | 1910 | td->last_ddir = ddir; |
87dc1ab1 | 1911 | |
f8b0bd10 | 1912 | if (!io_u->error && ddir_rw(ddir)) { |
5fff9543 | 1913 | unsigned long long bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1914 | int ret; |
10ba535a | 1915 | |
f8b0bd10 | 1916 | td->io_blocks[ddir]++; |
f8b0bd10 | 1917 | td->io_bytes[ddir] += bytes; |
ae2fafc8 | 1918 | |
e1c325d2 JA |
1919 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
1920 | td->this_io_blocks[ddir]++; | |
f8b0bd10 | 1921 | td->this_io_bytes[ddir] += bytes; |
e1c325d2 | 1922 | } |
f8b0bd10 | 1923 | |
639ad1ea | 1924 | if (ddir == DDIR_WRITE) |
94a6e1bb | 1925 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 1926 | |
b2b3eefe | 1927 | if (should_account(td)) |
f8b0bd10 | 1928 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1929 | |
f8b0bd10 | 1930 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1931 | |
d7762cf8 | 1932 | if (io_u->end_io) { |
f8b0bd10 JA |
1933 | ret = io_u->end_io(td, io_u_ptr); |
1934 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1935 | if (ret && !icd->error) |
1936 | icd->error = ret; | |
1937 | } | |
ff58fced | 1938 | } else if (io_u->error) { |
10ba535a | 1939 | icd->error = io_u->error; |
5451792e JA |
1940 | io_u_log_error(td, io_u); |
1941 | } | |
8b28bd41 | 1942 | if (icd->error) { |
f8b0bd10 JA |
1943 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1944 | ||
8b28bd41 DM |
1945 | if (!td_non_fatal_error(td, eb, icd->error)) |
1946 | return; | |
f8b0bd10 | 1947 | |
f2bba182 RR |
1948 | /* |
1949 | * If there is a non_fatal error, then add to the error count | |
1950 | * and clear all the errors. | |
1951 | */ | |
1952 | update_error_count(td, icd->error); | |
1953 | td_clear_error(td); | |
1954 | icd->error = 0; | |
f8b0bd10 JA |
1955 | if (io_u) |
1956 | io_u->error = 0; | |
f2bba182 | 1957 | } |
10ba535a JA |
1958 | } |
1959 | ||
9520ebb9 JA |
1960 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1961 | int nr) | |
10ba535a | 1962 | { |
6eaf09d6 | 1963 | int ddir; |
aba6c951 JA |
1964 | |
1965 | if (!gtod_reduce(td)) | |
9520ebb9 | 1966 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1967 | |
3af6ef39 JA |
1968 | icd->nr = nr; |
1969 | ||
10ba535a | 1970 | icd->error = 0; |
c1f50f76 | 1971 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 1972 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
1973 | } |
1974 | ||
97601024 JA |
1975 | static void ios_completed(struct thread_data *td, |
1976 | struct io_completion_data *icd) | |
36167d82 JA |
1977 | { |
1978 | struct io_u *io_u; | |
1979 | int i; | |
1980 | ||
10ba535a JA |
1981 | for (i = 0; i < icd->nr; i++) { |
1982 | io_u = td->io_ops->event(td, i); | |
1983 | ||
f8b0bd10 | 1984 | io_completed(td, &io_u, icd); |
e8462bd8 | 1985 | |
f8b0bd10 | 1986 | if (io_u) |
e8462bd8 | 1987 | put_io_u(td, io_u); |
10ba535a JA |
1988 | } |
1989 | } | |
97601024 | 1990 | |
e7e6cfb4 JA |
1991 | /* |
1992 | * Complete a single io_u for the sync engines. | |
1993 | */ | |
55312f9f | 1994 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1995 | { |
1996 | struct io_completion_data icd; | |
55312f9f | 1997 | int ddir; |
97601024 | 1998 | |
9520ebb9 | 1999 | init_icd(td, &icd, 1); |
f8b0bd10 | 2000 | io_completed(td, &io_u, &icd); |
e8462bd8 | 2001 | |
f8b0bd10 | 2002 | if (io_u) |
e8462bd8 | 2003 | put_io_u(td, io_u); |
97601024 | 2004 | |
581e7141 JA |
2005 | if (icd.error) { |
2006 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2007 | return -1; | |
2008 | } | |
97601024 | 2009 | |
c1f50f76 | 2010 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2011 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
2012 | |
2013 | return 0; | |
97601024 JA |
2014 | } |
2015 | ||
e7e6cfb4 JA |
2016 | /* |
2017 | * Called to complete min_events number of io for the async engines. | |
2018 | */ | |
55312f9f | 2019 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2020 | { |
97601024 | 2021 | struct io_completion_data icd; |
00de55ef | 2022 | struct timespec *tvp = NULL; |
55312f9f | 2023 | int ret, ddir; |
4d06a338 | 2024 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2025 | |
12bb8569 | 2026 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2027 | |
4950421a | 2028 | if (!min_evts) |
00de55ef | 2029 | tvp = &ts; |
5fb4b366 RE |
2030 | else if (min_evts > td->cur_depth) |
2031 | min_evts = td->cur_depth; | |
97601024 | 2032 | |
82407585 RP |
2033 | /* No worries, td_io_getevents fixes min and max if they are |
2034 | * set incorrectly */ | |
2035 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2036 | if (ret < 0) { |
e1161c32 | 2037 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2038 | return ret; |
2039 | } else if (!ret) | |
2040 | return ret; | |
2041 | ||
9520ebb9 | 2042 | init_icd(td, &icd, ret); |
97601024 | 2043 | ios_completed(td, &icd); |
581e7141 JA |
2044 | if (icd.error) { |
2045 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2046 | return -1; | |
2047 | } | |
97601024 | 2048 | |
c1f50f76 | 2049 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2050 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2051 | |
0d593542 | 2052 | return ret; |
97601024 | 2053 | } |
7e77dd02 JA |
2054 | |
2055 | /* | |
2056 | * Call when io_u is really queued, to update the submission latency. | |
2057 | */ | |
2058 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2059 | { | |
8243be59 | 2060 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2061 | unsigned long slat_time; |
7e77dd02 | 2062 | |
d6bb626e | 2063 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2064 | |
2065 | if (td->parent) | |
2066 | td = td->parent; | |
2067 | ||
ae588852 JA |
2068 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
2069 | io_u->offset); | |
9520ebb9 | 2070 | } |
7e77dd02 | 2071 | } |
433afcb4 | 2072 | |
5c94b008 JA |
2073 | /* |
2074 | * See if we should reuse the last seed, if dedupe is enabled | |
2075 | */ | |
9451b93e | 2076 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2077 | { |
2078 | unsigned int v; | |
5c94b008 JA |
2079 | |
2080 | if (!td->o.dedupe_percentage) | |
2081 | return &td->buf_state; | |
732eedd0 | 2082 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2083 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2084 | return &td->buf_state; | |
732eedd0 | 2085 | } |
5c94b008 | 2086 | |
1bd5d213 | 2087 | v = rand_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2088 | |
2089 | if (v <= td->o.dedupe_percentage) | |
2090 | return &td->buf_state_prev; | |
2091 | ||
2092 | return &td->buf_state; | |
2093 | } | |
2094 | ||
9451b93e | 2095 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2096 | { |
9451b93e JA |
2097 | if (td->o.dedupe_percentage == 100) |
2098 | frand_copy(rs, &td->buf_state_prev); | |
2099 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2100 | frand_copy(&td->buf_state_prev, rs); |
2101 | } | |
2102 | ||
5fff9543 JF |
2103 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write, |
2104 | unsigned long long max_bs) | |
5973cafb | 2105 | { |
d1af2894 JA |
2106 | struct thread_options *o = &td->o; |
2107 | ||
15600335 JA |
2108 | if (o->mem_type == MEM_CUDA_MALLOC) |
2109 | return; | |
03553853 | 2110 | |
4eff3e57 | 2111 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2112 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2113 | struct frand_state *rs; |
5fff9543 JF |
2114 | unsigned long long left = max_bs; |
2115 | unsigned long long this_write; | |
5c94b008 | 2116 | |
1066358a | 2117 | do { |
9451b93e | 2118 | rs = get_buf_state(td); |
9c42684e | 2119 | |
1066358a | 2120 | min_write = min(min_write, left); |
f97a43a1 | 2121 | |
1066358a | 2122 | if (perc) { |
4b157ac6 | 2123 | this_write = min_not_zero(min_write, |
5fff9543 | 2124 | (unsigned long long) td->o.compress_chunk); |
1e7f82e2 JA |
2125 | |
2126 | fill_random_buf_percentage(rs, buf, perc, | |
2127 | this_write, this_write, | |
2128 | o->buffer_pattern, | |
2129 | o->buffer_pattern_bytes); | |
2130 | } else { | |
1066358a | 2131 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2132 | this_write = min_write; |
2133 | } | |
1066358a | 2134 | |
1e7f82e2 JA |
2135 | buf += this_write; |
2136 | left -= this_write; | |
9451b93e | 2137 | save_buf_state(td, rs); |
1066358a | 2138 | } while (left); |
d1af2894 JA |
2139 | } else if (o->buffer_pattern_bytes) |
2140 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2141 | else if (o->zero_buffers) |
cc86c395 | 2142 | memset(buf, 0, max_bs); |
999d245e | 2143 | else |
9451b93e | 2144 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2145 | } |
2146 | ||
2147 | /* | |
2148 | * "randomly" fill the buffer contents | |
2149 | */ | |
2150 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
5fff9543 | 2151 | unsigned long long min_write, unsigned long long max_bs) |
cc86c395 JA |
2152 | { |
2153 | io_u->buf_filled_len = 0; | |
2154 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2155 | } |
e2c75fc4 TK |
2156 | |
2157 | static int do_sync_file_range(const struct thread_data *td, | |
2158 | struct fio_file *f) | |
2159 | { | |
2160 | off64_t offset, nbytes; | |
2161 | ||
2162 | offset = f->first_write; | |
2163 | nbytes = f->last_write - f->first_write; | |
2164 | ||
2165 | if (!nbytes) | |
2166 | return 0; | |
2167 | ||
2168 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2169 | } | |
2170 | ||
2171 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2172 | { | |
2173 | int ret; | |
2174 | ||
2175 | if (io_u->ddir == DDIR_SYNC) { | |
2176 | ret = fsync(io_u->file->fd); | |
2177 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2178 | #ifdef CONFIG_FDATASYNC | |
2179 | ret = fdatasync(io_u->file->fd); | |
2180 | #else | |
2181 | ret = io_u->xfer_buflen; | |
2182 | io_u->error = EINVAL; | |
2183 | #endif | |
2184 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2185 | ret = do_sync_file_range(td, io_u->file); | |
2186 | else { | |
2187 | ret = io_u->xfer_buflen; | |
2188 | io_u->error = EINVAL; | |
2189 | } | |
2190 | ||
2191 | if (ret < 0) | |
2192 | io_u->error = errno; | |
2193 | ||
2194 | return ret; | |
2195 | } | |
2196 | ||
2197 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2198 | { | |
2199 | #ifndef FIO_HAVE_TRIM | |
2200 | io_u->error = EINVAL; | |
2201 | return 0; | |
2202 | #else | |
2203 | struct fio_file *f = io_u->file; | |
2204 | int ret; | |
2205 | ||
496b1f9e | 2206 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2207 | if (!ret) |
2208 | return io_u->xfer_buflen; | |
2209 | ||
2210 | io_u->error = ret; | |
2211 | return 0; | |
2212 | #endif | |
2213 | } |