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; | |
df415585 | 44 | |
2ab9e98b | 45 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 46 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 47 | |
6cc1a3d1 BVA |
48 | if ((nr_blocks * min_bs) < buflen) |
49 | buflen = nr_blocks * min_bs; | |
50 | ||
51 | return buflen; | |
df415585 JA |
52 | } |
53 | ||
74776733 JA |
54 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
55 | enum fio_ddir ddir) | |
2ba1c290 | 56 | { |
74776733 JA |
57 | uint64_t max_blocks; |
58 | uint64_t max_size; | |
2ba1c290 | 59 | |
ff58fced JA |
60 | assert(ddir_rw(ddir)); |
61 | ||
d9dd70f7 JA |
62 | /* |
63 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
79591fa9 | 64 | * -> not for now since there is code assuming it could go either. |
d9dd70f7 JA |
65 | */ |
66 | max_size = f->io_size; | |
67 | if (max_size > f->real_file_size) | |
68 | max_size = f->real_file_size; | |
69 | ||
7b865a2f | 70 | if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range) |
ed335855 SN |
71 | max_size = td->o.zone_range; |
72 | ||
0412d12e JE |
73 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
74 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
75 | ||
1ae83d45 | 76 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
77 | if (!max_blocks) |
78 | return 0; | |
79 | ||
67778e88 | 80 | return max_blocks; |
2ba1c290 JA |
81 | } |
82 | ||
e25839d4 | 83 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
e0a04ac1 JA |
84 | enum fio_ddir ddir, uint64_t *b, |
85 | uint64_t lastb) | |
ec4015da | 86 | { |
6f49f8bc | 87 | uint64_t r; |
5e0baa7f | 88 | |
c3546b53 JA |
89 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
90 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { | |
6f49f8bc | 91 | |
d6b72507 | 92 | r = __rand(&td->random_state); |
8055e41d | 93 | |
4b91ee8f | 94 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 95 | |
e0a04ac1 | 96 | *b = lastb * (r / (rand_max(&td->random_state) + 1.0)); |
51ede0b1 | 97 | } else { |
8055e41d | 98 | uint64_t off = 0; |
43c63a78 | 99 | |
967d1b63 JA |
100 | assert(fio_file_lfsr(f)); |
101 | ||
6f49f8bc | 102 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 103 | return 1; |
ec4015da | 104 | |
8055e41d JA |
105 | *b = off; |
106 | } | |
0ce8b119 | 107 | |
ec4015da | 108 | /* |
51ede0b1 | 109 | * if we are not maintaining a random map, we are done. |
ec4015da | 110 | */ |
51ede0b1 JA |
111 | if (!file_randommap(td, f)) |
112 | goto ret; | |
43c63a78 JA |
113 | |
114 | /* | |
51ede0b1 | 115 | * calculate map offset and check if it's free |
43c63a78 | 116 | */ |
51ede0b1 JA |
117 | if (random_map_free(f, *b)) |
118 | goto ret; | |
119 | ||
4b91ee8f JA |
120 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
121 | (unsigned long long) *b); | |
51ede0b1 | 122 | |
7ebd796f | 123 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
124 | if (*b == (uint64_t) -1ULL) |
125 | return 1; | |
0ce8b119 JA |
126 | ret: |
127 | return 0; | |
ec4015da JA |
128 | } |
129 | ||
925fee33 JA |
130 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
131 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 132 | uint64_t *b) |
e25839d4 | 133 | { |
9c6f6316 | 134 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
135 | return 0; |
136 | } | |
137 | ||
925fee33 JA |
138 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
139 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 140 | uint64_t *b) |
925fee33 | 141 | { |
9c6f6316 | 142 | *b = pareto_next(&f->zipf); |
925fee33 JA |
143 | return 0; |
144 | } | |
145 | ||
56d9fa4b JA |
146 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
147 | struct fio_file *f, enum fio_ddir ddir, | |
148 | uint64_t *b) | |
149 | { | |
150 | *b = gauss_next(&f->gauss); | |
151 | return 0; | |
152 | } | |
153 | ||
59466396 JA |
154 | static int __get_next_rand_offset_zoned_abs(struct thread_data *td, |
155 | struct fio_file *f, | |
156 | enum fio_ddir ddir, uint64_t *b) | |
157 | { | |
158 | struct zone_split_index *zsi; | |
e3c8c108 | 159 | uint64_t lastb, send, stotal; |
59466396 JA |
160 | unsigned int v; |
161 | ||
162 | lastb = last_block(td, f, ddir); | |
163 | if (!lastb) | |
164 | return 1; | |
165 | ||
166 | if (!td->o.zone_split_nr[ddir]) { | |
167 | bail: | |
168 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
169 | } | |
170 | ||
171 | /* | |
172 | * Generate a value, v, between 1 and 100, both inclusive | |
173 | */ | |
1bd5d213 | 174 | v = rand_between(&td->zone_state, 1, 100); |
59466396 | 175 | |
e3c8c108 JA |
176 | /* |
177 | * Find our generated table. 'send' is the end block of this zone, | |
178 | * 'stotal' is our start offset. | |
179 | */ | |
59466396 JA |
180 | zsi = &td->zone_state_index[ddir][v - 1]; |
181 | stotal = zsi->size_prev / td->o.ba[ddir]; | |
182 | send = zsi->size / td->o.ba[ddir]; | |
183 | ||
184 | /* | |
185 | * Should never happen | |
186 | */ | |
187 | if (send == -1U) { | |
264e3d30 | 188 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
59466396 | 189 | log_err("fio: bug in zoned generation\n"); |
59466396 JA |
190 | goto bail; |
191 | } else if (send > lastb) { | |
192 | /* | |
193 | * This happens if the user specifies ranges that exceed | |
194 | * the file/device size. We can't handle that gracefully, | |
195 | * so error and exit. | |
196 | */ | |
197 | log_err("fio: zoned_abs sizes exceed file size\n"); | |
198 | return 1; | |
199 | } | |
200 | ||
201 | /* | |
e3c8c108 | 202 | * Generate index from 0..send-stotal |
59466396 | 203 | */ |
e3c8c108 | 204 | if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1) |
59466396 JA |
205 | return 1; |
206 | ||
e3c8c108 | 207 | *b += stotal; |
59466396 JA |
208 | return 0; |
209 | } | |
210 | ||
e0a04ac1 JA |
211 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
212 | struct fio_file *f, enum fio_ddir ddir, | |
213 | uint64_t *b) | |
214 | { | |
215 | unsigned int v, send, stotal; | |
216 | uint64_t offset, lastb; | |
e0a04ac1 JA |
217 | struct zone_split_index *zsi; |
218 | ||
219 | lastb = last_block(td, f, ddir); | |
220 | if (!lastb) | |
221 | return 1; | |
222 | ||
223 | if (!td->o.zone_split_nr[ddir]) { | |
224 | bail: | |
225 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
226 | } | |
227 | ||
228 | /* | |
229 | * Generate a value, v, between 1 and 100, both inclusive | |
230 | */ | |
1bd5d213 | 231 | v = rand_between(&td->zone_state, 1, 100); |
e0a04ac1 JA |
232 | |
233 | zsi = &td->zone_state_index[ddir][v - 1]; | |
234 | stotal = zsi->size_perc_prev; | |
235 | send = zsi->size_perc; | |
236 | ||
237 | /* | |
238 | * Should never happen | |
239 | */ | |
240 | if (send == -1U) { | |
264e3d30 | 241 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
e0a04ac1 | 242 | log_err("fio: bug in zoned generation\n"); |
e0a04ac1 JA |
243 | goto bail; |
244 | } | |
245 | ||
246 | /* | |
247 | * 'send' is some percentage below or equal to 100 that | |
248 | * marks the end of the current IO range. 'stotal' marks | |
249 | * the start, in percent. | |
250 | */ | |
251 | if (stotal) | |
252 | offset = stotal * lastb / 100ULL; | |
253 | else | |
254 | offset = 0; | |
255 | ||
256 | lastb = lastb * (send - stotal) / 100ULL; | |
257 | ||
258 | /* | |
259 | * Generate index from 0..send-of-lastb | |
260 | */ | |
261 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
262 | return 1; | |
263 | ||
264 | /* | |
265 | * Add our start offset, if any | |
266 | */ | |
267 | if (offset) | |
268 | *b += offset; | |
269 | ||
270 | return 0; | |
271 | } | |
56d9fa4b | 272 | |
f31feaa2 JA |
273 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
274 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 | 275 | { |
e0a04ac1 JA |
276 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
277 | uint64_t lastb; | |
278 | ||
279 | lastb = last_block(td, f, ddir); | |
280 | if (!lastb) | |
281 | return 1; | |
282 | ||
283 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
284 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
e25839d4 | 285 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
925fee33 JA |
286 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
287 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
288 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
289 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e0a04ac1 JA |
290 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
291 | return __get_next_rand_offset_zoned(td, f, ddir, b); | |
59466396 JA |
292 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS) |
293 | return __get_next_rand_offset_zoned_abs(td, f, ddir, b); | |
e25839d4 JA |
294 | |
295 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
296 | return 1; | |
297 | } | |
298 | ||
e39c0676 | 299 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
300 | { |
301 | unsigned int v; | |
211c9b89 | 302 | |
d9472271 | 303 | if (td->o.perc_rand[ddir] == 100) |
e39c0676 | 304 | return true; |
211c9b89 | 305 | |
1bd5d213 | 306 | v = rand_between(&td->seq_rand_state[ddir], 1, 100); |
211c9b89 | 307 | |
d9472271 | 308 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
309 | } |
310 | ||
0bcf41cd JA |
311 | static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f) |
312 | { | |
313 | struct thread_options *o = &td->o; | |
314 | ||
315 | if (o->invalidate_cache && !o->odirect) { | |
316 | int fio_unused ret; | |
317 | ||
318 | ret = file_invalidate_cache(td, f); | |
319 | } | |
320 | } | |
321 | ||
38dad62d | 322 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 323 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 324 | { |
c04e4661 DE |
325 | if (!get_next_rand_offset(td, f, ddir, b)) |
326 | return 0; | |
327 | ||
8c07860d JA |
328 | if (td->o.time_based || |
329 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { | |
33c48814 | 330 | fio_file_reset(td, f); |
80f02150 | 331 | loop_cache_invalidate(td, f); |
c04e4661 DE |
332 | if (!get_next_rand_offset(td, f, ddir, b)) |
333 | return 0; | |
38dad62d JA |
334 | } |
335 | ||
c04e4661 | 336 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 337 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 338 | (unsigned long long) f->real_file_size); |
c04e4661 | 339 | return 1; |
38dad62d JA |
340 | } |
341 | ||
37cf9e3c | 342 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 343 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 344 | { |
ac002339 JA |
345 | struct thread_options *o = &td->o; |
346 | ||
ff58fced JA |
347 | assert(ddir_rw(ddir)); |
348 | ||
17373ce2 JA |
349 | /* |
350 | * If we reach the end for a time based run, reset us back to 0 | |
351 | * and invalidate the cache, if we need to. | |
352 | */ | |
f1dfb668 | 353 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
19ddc35b | 354 | o->time_based) { |
c89daa4a | 355 | f->last_pos[ddir] = f->file_offset; |
0bcf41cd | 356 | loop_cache_invalidate(td, f); |
19ddc35b | 357 | } |
c04e4661 | 358 | |
f1dfb668 | 359 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 360 | uint64_t pos; |
059b0802 | 361 | |
69b98f11 JA |
362 | /* |
363 | * Only rewind if we already hit the end | |
364 | */ | |
365 | if (f->last_pos[ddir] == f->file_offset && | |
366 | f->file_offset && o->ddir_seq_add < 0) { | |
c22825bb JA |
367 | if (f->real_file_size > f->io_size) |
368 | f->last_pos[ddir] = f->io_size; | |
369 | else | |
370 | f->last_pos[ddir] = f->real_file_size; | |
371 | } | |
a66da7a2 | 372 | |
f1dfb668 | 373 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
374 | if (pos && o->ddir_seq_add) { |
375 | pos += o->ddir_seq_add; | |
376 | ||
377 | /* | |
378 | * If we reach beyond the end of the file | |
379 | * with holed IO, wrap around to the | |
b0a84f48 JA |
380 | * beginning again. If we're doing backwards IO, |
381 | * wrap to the end. | |
ac002339 | 382 | */ |
b0a84f48 JA |
383 | if (pos >= f->real_file_size) { |
384 | if (o->ddir_seq_add > 0) | |
385 | pos = f->file_offset; | |
c22825bb JA |
386 | else { |
387 | if (f->real_file_size > f->io_size) | |
388 | pos = f->io_size; | |
389 | else | |
390 | pos = f->real_file_size; | |
391 | ||
392 | pos += o->ddir_seq_add; | |
393 | } | |
b0a84f48 | 394 | } |
ac002339 | 395 | } |
059b0802 | 396 | |
37cf9e3c | 397 | *offset = pos; |
38dad62d JA |
398 | return 0; |
399 | } | |
400 | ||
401 | return 1; | |
402 | } | |
403 | ||
404 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d | 405 | enum fio_ddir ddir, int rw_seq, |
ec370c22 | 406 | bool *is_random) |
38dad62d JA |
407 | { |
408 | struct fio_file *f = io_u->file; | |
1ae83d45 | 409 | uint64_t b, offset; |
38dad62d JA |
410 | int ret; |
411 | ||
ff58fced JA |
412 | assert(ddir_rw(ddir)); |
413 | ||
37cf9e3c JA |
414 | b = offset = -1ULL; |
415 | ||
38dad62d | 416 | if (rw_seq) { |
211c9b89 | 417 | if (td_random(td)) { |
6aca9b3d | 418 | if (should_do_random(td, ddir)) { |
211c9b89 | 419 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 420 | *is_random = true; |
6aca9b3d | 421 | } else { |
ec370c22 | 422 | *is_random = false; |
1651e431 | 423 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
424 | ret = get_next_seq_offset(td, f, ddir, &offset); |
425 | if (ret) | |
426 | ret = get_next_rand_block(td, f, ddir, &b); | |
427 | } | |
6aca9b3d | 428 | } else { |
ec370c22 | 429 | *is_random = false; |
37cf9e3c | 430 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 431 | } |
38dad62d | 432 | } else { |
1651e431 | 433 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
ec370c22 | 434 | *is_random = false; |
38dad62d JA |
435 | |
436 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 437 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 438 | if (ret) { |
37cf9e3c | 439 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 440 | *is_random = false; |
6aca9b3d | 441 | } |
38dad62d | 442 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
443 | if (f->last_start[ddir] != -1ULL) |
444 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 445 | else |
37cf9e3c | 446 | offset = 0; |
38dad62d JA |
447 | ret = 0; |
448 | } else { | |
449 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
450 | ret = 1; | |
451 | } | |
452 | } | |
6d68b997 | 453 | |
37cf9e3c JA |
454 | if (!ret) { |
455 | if (offset != -1ULL) | |
456 | io_u->offset = offset; | |
457 | else if (b != -1ULL) | |
458 | io_u->offset = b * td->o.ba[ddir]; | |
459 | else { | |
4e0a8fa2 | 460 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
461 | ret = 1; |
462 | } | |
463 | } | |
464 | ||
38dad62d JA |
465 | return ret; |
466 | } | |
467 | ||
10ba535a JA |
468 | /* |
469 | * For random io, generate a random new block and see if it's used. Repeat | |
470 | * until we find a free one. For sequential io, just return the end of | |
471 | * the last io issued. | |
472 | */ | |
e8fb335e | 473 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 474 | bool *is_random) |
10ba535a | 475 | { |
9bf2061e | 476 | struct fio_file *f = io_u->file; |
4ba66134 | 477 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 478 | int rw_seq_hit = 0; |
10ba535a | 479 | |
ff58fced JA |
480 | assert(ddir_rw(ddir)); |
481 | ||
38dad62d JA |
482 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
483 | rw_seq_hit = 1; | |
5736c10d | 484 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 485 | } |
211097b2 | 486 | |
6aca9b3d | 487 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 488 | return 1; |
10ba535a | 489 | |
009bd847 JA |
490 | if (io_u->offset >= f->io_size) { |
491 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
492 | (unsigned long long) io_u->offset, |
493 | (unsigned long long) f->io_size); | |
009bd847 JA |
494 | return 1; |
495 | } | |
496 | ||
497 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
498 | if (io_u->offset >= f->real_file_size) { |
499 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
500 | (unsigned long long) io_u->offset, |
501 | (unsigned long long) f->real_file_size); | |
10ba535a | 502 | return 1; |
2ba1c290 | 503 | } |
10ba535a JA |
504 | |
505 | return 0; | |
506 | } | |
507 | ||
e39c0676 | 508 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
5fff9543 | 509 | unsigned long long buflen) |
79944128 JA |
510 | { |
511 | struct fio_file *f = io_u->file; | |
512 | ||
bedc9dc2 | 513 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
514 | } |
515 | ||
5fff9543 | 516 | static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 517 | bool is_random) |
10ba535a | 518 | { |
6aca9b3d | 519 | int ddir = io_u->ddir; |
5fff9543 JF |
520 | unsigned long long buflen = 0; |
521 | unsigned long long minbs, maxbs; | |
3dd29f7c | 522 | uint64_t frand_max, r; |
7c961359 | 523 | bool power_2; |
10ba535a | 524 | |
9ee1c647 | 525 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
526 | |
527 | if (td->o.bs_is_seq_rand) | |
ec370c22 | 528 | ddir = is_random ? DDIR_WRITE : DDIR_READ; |
ff58fced | 529 | |
f3059de1 JA |
530 | minbs = td->o.min_bs[ddir]; |
531 | maxbs = td->o.max_bs[ddir]; | |
532 | ||
79944128 JA |
533 | if (minbs == maxbs) |
534 | return minbs; | |
535 | ||
52c58027 JA |
536 | /* |
537 | * If we can't satisfy the min block size from here, then fail | |
538 | */ | |
539 | if (!io_u_fits(td, io_u, minbs)) | |
540 | return 0; | |
541 | ||
2f282cec | 542 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 543 | do { |
2f282cec | 544 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 545 | |
720e84ad | 546 | if (!td->o.bssplit_nr[ddir]) { |
5fff9543 | 547 | buflen = minbs + (unsigned long long) ((double) maxbs * |
c3546b53 | 548 | (r / (frand_max + 1.0))); |
5ec10eaa | 549 | } else { |
3dd29f7c | 550 | long long perc = 0; |
564ca972 JA |
551 | unsigned int i; |
552 | ||
720e84ad JA |
553 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
554 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
555 | |
556 | buflen = bsp->bs; | |
557 | perc += bsp->perc; | |
3dd29f7c JA |
558 | if (!perc) |
559 | break; | |
560 | if ((r / perc <= frand_max / 100ULL) && | |
79944128 | 561 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
562 | break; |
563 | } | |
564 | } | |
79944128 | 565 | |
17a6b702 PL |
566 | power_2 = is_power_of_2(minbs); |
567 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 568 | buflen &= ~(minbs - 1); |
17a6b702 PL |
569 | else if (!td->o.bs_unaligned && !power_2) |
570 | buflen -= buflen % minbs; | |
79944128 | 571 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 572 | |
10ba535a JA |
573 | return buflen; |
574 | } | |
575 | ||
afe24a5a JA |
576 | static void set_rwmix_bytes(struct thread_data *td) |
577 | { | |
afe24a5a JA |
578 | unsigned int diff; |
579 | ||
580 | /* | |
581 | * we do time or byte based switch. this is needed because | |
582 | * buffered writes may issue a lot quicker than they complete, | |
583 | * whereas reads do not. | |
584 | */ | |
e47f799f | 585 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 586 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
587 | } |
588 | ||
589 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
590 | { | |
591 | unsigned int v; | |
e47f799f | 592 | |
1bd5d213 | 593 | v = rand_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 594 | |
04c540d9 | 595 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
596 | return DDIR_READ; |
597 | ||
598 | return DDIR_WRITE; | |
afe24a5a JA |
599 | } |
600 | ||
0d593542 | 601 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 602 | { |
0d593542 JA |
603 | int completed = 0; |
604 | ||
002e7183 JA |
605 | /* |
606 | * We are going to sleep, ensure that we flush anything pending as | |
607 | * not to skew our latency numbers. | |
608 | * | |
609 | * Changed to only monitor 'in flight' requests here instead of the | |
610 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
611 | * io's that have been actually submitted to an async engine, | |
612 | * and cur_depth is meaningless for sync engines. | |
613 | */ | |
a80cb54b BVA |
614 | if (td->io_u_queued || td->cur_depth) |
615 | td_io_commit(td); | |
9cc80b6d | 616 | |
002e7183 | 617 | while (td->io_u_in_flight) { |
89fd0573 | 618 | int ret; |
002e7183 | 619 | |
55312f9f | 620 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
621 | if (ret > 0) |
622 | completed += ret; | |
002e7183 | 623 | } |
0d593542 | 624 | |
6be06c46 JA |
625 | if (td->flags & TD_F_REGROW_LOGS) |
626 | regrow_logs(td); | |
627 | ||
0d593542 | 628 | return completed; |
002e7183 JA |
629 | } |
630 | ||
581e7141 JA |
631 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
632 | { | |
633 | enum fio_ddir odir = ddir ^ 1; | |
b5407f8b | 634 | uint64_t usec; |
90eff1c9 | 635 | uint64_t now; |
581e7141 | 636 | |
ff58fced | 637 | assert(ddir_rw(ddir)); |
50a8ce86 | 638 | now = utime_since_now(&td->start); |
ff58fced | 639 | |
50a8ce86 D |
640 | /* |
641 | * if rate_next_io_time is in the past, need to catch up to rate | |
642 | */ | |
643 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
644 | return ddir; |
645 | ||
646 | /* | |
50a8ce86 | 647 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
648 | * should switch. |
649 | */ | |
315fcfec | 650 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 651 | /* |
50a8ce86 | 652 | * Other direction is behind rate, switch |
581e7141 | 653 | */ |
50a8ce86 | 654 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
655 | return odir; |
656 | ||
657 | /* | |
395feabb JA |
658 | * Both directions are ahead of rate. sleep the min, |
659 | * switch if necessary | |
581e7141 | 660 | */ |
50a8ce86 | 661 | if (td->rate_next_io_time[ddir] <= |
395feabb | 662 | td->rate_next_io_time[odir]) { |
50a8ce86 | 663 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 664 | } else { |
50a8ce86 | 665 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
666 | ddir = odir; |
667 | } | |
668 | } else | |
50a8ce86 | 669 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 670 | |
a9da8ab2 JA |
671 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
672 | io_u_quiesce(td); | |
78c1eda5 | 673 | |
1a9bf814 | 674 | usec_sleep(td, usec); |
581e7141 JA |
675 | return ddir; |
676 | } | |
677 | ||
10ba535a JA |
678 | /* |
679 | * Return the data direction for the next io_u. If the job is a | |
680 | * mixed read/write workload, check the rwmix cycle and switch if | |
681 | * necessary. | |
682 | */ | |
1e97cce9 | 683 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 684 | { |
581e7141 JA |
685 | enum fio_ddir ddir; |
686 | ||
5f9099ea | 687 | /* |
f6860972 TK |
688 | * See if it's time to fsync/fdatasync/sync_file_range first, |
689 | * and if not then move on to check regular I/Os. | |
5f9099ea | 690 | */ |
f6860972 TK |
691 | if (should_fsync(td)) { |
692 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
693 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
694 | return DDIR_SYNC; | |
695 | ||
696 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
697 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
698 | return DDIR_DATASYNC; | |
699 | ||
700 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
701 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
702 | return DDIR_SYNC_FILE_RANGE; | |
703 | } | |
44f29692 | 704 | |
10ba535a | 705 | if (td_rw(td)) { |
10ba535a JA |
706 | /* |
707 | * Check if it's time to seed a new data direction. | |
708 | */ | |
e4928662 | 709 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
710 | /* |
711 | * Put a top limit on how many bytes we do for | |
712 | * one data direction, to avoid overflowing the | |
713 | * ranges too much | |
714 | */ | |
715 | ddir = get_rand_ddir(td); | |
e47f799f JA |
716 | |
717 | if (ddir != td->rwmix_ddir) | |
718 | set_rwmix_bytes(td); | |
719 | ||
720 | td->rwmix_ddir = ddir; | |
10ba535a | 721 | } |
581e7141 | 722 | ddir = td->rwmix_ddir; |
10ba535a | 723 | } else if (td_read(td)) |
581e7141 | 724 | ddir = DDIR_READ; |
6eaf09d6 | 725 | else if (td_write(td)) |
581e7141 | 726 | ddir = DDIR_WRITE; |
5c8e84ca | 727 | else if (td_trim(td)) |
6eaf09d6 | 728 | ddir = DDIR_TRIM; |
5c8e84ca TK |
729 | else |
730 | ddir = DDIR_INVAL; | |
581e7141 JA |
731 | |
732 | td->rwmix_ddir = rate_ddir(td, ddir); | |
733 | return td->rwmix_ddir; | |
10ba535a JA |
734 | } |
735 | ||
1ef2b6be JA |
736 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
737 | { | |
0e4dd95c DE |
738 | enum fio_ddir ddir = get_rw_ddir(td); |
739 | ||
82a90686 | 740 | if (td_trimwrite(td)) { |
0e4dd95c DE |
741 | struct fio_file *f = io_u->file; |
742 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
743 | ddir = DDIR_TRIM; | |
744 | else | |
745 | ddir = DDIR_WRITE; | |
746 | } | |
747 | ||
748 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 749 | |
9b87f09b | 750 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
751 | td->o.barrier_blocks && |
752 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
753 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 754 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
755 | } |
756 | ||
e8462bd8 | 757 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 758 | { |
71b84caa | 759 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
760 | |
761 | if (ret) | |
762 | td_verror(td, ret, "file close"); | |
763 | } | |
764 | ||
10ba535a JA |
765 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
766 | { | |
99952ca7 BVA |
767 | if (io_u->post_submit) { |
768 | io_u->post_submit(io_u, io_u->error == 0); | |
769 | io_u->post_submit = NULL; | |
770 | } | |
771 | ||
a9da8ab2 JA |
772 | if (td->parent) |
773 | td = td->parent; | |
774 | ||
e8462bd8 JA |
775 | td_io_u_lock(td); |
776 | ||
f8b0bd10 | 777 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 778 | put_file_log(td, io_u->file); |
f8b0bd10 | 779 | |
10ba535a | 780 | io_u->file = NULL; |
1651e431 | 781 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 782 | |
a9da8ab2 | 783 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 784 | td->cur_depth--; |
a9da8ab2 JA |
785 | assert(!(td->flags & TD_F_CHILD)); |
786 | } | |
2ae0b204 | 787 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 | 788 | td_io_u_free_notify(td); |
564de8d1 | 789 | td_io_u_unlock(td); |
10ba535a JA |
790 | } |
791 | ||
f2bba182 RR |
792 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
793 | { | |
1651e431 | 794 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
795 | put_io_u(td, io_u); |
796 | } | |
797 | ||
755200a3 JA |
798 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
799 | { | |
800 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 801 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 802 | |
465221b0 JA |
803 | dprint(FD_IO, "requeue %p\n", __io_u); |
804 | ||
a9da8ab2 JA |
805 | if (td->parent) |
806 | td = td->parent; | |
807 | ||
e8462bd8 JA |
808 | td_io_u_lock(td); |
809 | ||
1651e431 | 810 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
811 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
812 | td->io_issues[ddir]--; | |
5ec10eaa | 813 | |
1651e431 | 814 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 815 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 816 | td->cur_depth--; |
a9da8ab2 JA |
817 | assert(!(td->flags & TD_F_CHILD)); |
818 | } | |
2ae0b204 JA |
819 | |
820 | io_u_rpush(&td->io_u_requeues, __io_u); | |
a9da8ab2 | 821 | td_io_u_free_notify(td); |
564de8d1 | 822 | td_io_u_unlock(td); |
755200a3 JA |
823 | *io_u = NULL; |
824 | } | |
825 | ||
7b865a2f | 826 | static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u) |
224b3093 | 827 | { |
828 | struct fio_file *f = io_u->file; | |
829 | ||
7b865a2f BVA |
830 | assert(td->o.zone_mode == ZONE_MODE_STRIDED); |
831 | assert(td->o.zone_size); | |
832 | assert(td->o.zone_range); | |
833 | ||
224b3093 | 834 | /* |
835 | * See if it's time to switch to a new zone | |
836 | */ | |
837 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { | |
838 | td->zone_bytes = 0; | |
839 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
840 | ||
841 | /* | |
842 | * Wrap from the beginning, if we exceed the file size | |
843 | */ | |
844 | if (f->file_offset >= f->real_file_size) | |
04bc85a1 JA |
845 | f->file_offset = get_start_offset(td, f); |
846 | ||
224b3093 | 847 | f->last_pos[io_u->ddir] = f->file_offset; |
848 | td->io_skip_bytes += td->o.zone_skip; | |
849 | } | |
850 | ||
851 | /* | |
04bc85a1 JA |
852 | * If zone_size > zone_range, then maintain the same zone until |
853 | * zone_bytes >= zone_size. | |
854 | */ | |
224b3093 | 855 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
856 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
857 | f->file_offset, f->last_pos[io_u->ddir]); | |
858 | f->last_pos[io_u->ddir] = f->file_offset; | |
859 | } | |
860 | ||
861 | /* | |
862 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
863 | * map needs to be reset as it's done with zone_range everytime. | |
864 | */ | |
04bc85a1 | 865 | if ((td->zone_bytes % td->o.zone_range) == 0) |
224b3093 | 866 | fio_file_reset(td, f); |
224b3093 | 867 | } |
868 | ||
9bf2061e | 869 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 870 | { |
ec370c22 | 871 | bool is_random; |
6aca9b3d | 872 | |
9b87f09b | 873 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
874 | goto out; |
875 | ||
1ef2b6be | 876 | set_rw_ddir(td, io_u); |
5f9099ea | 877 | |
87dc1ab1 | 878 | /* |
ff58fced | 879 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 880 | */ |
ff58fced | 881 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 882 | goto out; |
a00735e6 | 883 | |
7b865a2f BVA |
884 | if (td->o.zone_mode == ZONE_MODE_STRIDED) |
885 | setup_strided_zone_mode(td, io_u); | |
48f5abd3 | 886 | |
10ba535a | 887 | /* |
c685b5b2 JA |
888 | * No log, let the seq/rand engine retrieve the next buflen and |
889 | * position. | |
10ba535a | 890 | */ |
6aca9b3d | 891 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 892 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 893 | return 1; |
2ba1c290 | 894 | } |
10ba535a | 895 | |
6aca9b3d | 896 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
897 | if (!io_u->buflen) { |
898 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 899 | return 1; |
2ba1c290 | 900 | } |
bca4ed4d | 901 | |
2ba1c290 | 902 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
5fff9543 | 903 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n", |
e5f9a813 | 904 | io_u, |
4b91ee8f JA |
905 | (unsigned long long) io_u->offset, io_u->buflen, |
906 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 907 | return 1; |
2ba1c290 | 908 | } |
6a5e6884 | 909 | |
bca4ed4d JA |
910 | /* |
911 | * mark entry before potentially trimming io_u | |
912 | */ | |
303032ae | 913 | if (td_random(td) && file_randommap(td, io_u->file)) |
6cc1a3d1 BVA |
914 | io_u->buflen = mark_random_map(td, io_u, io_u->offset, |
915 | io_u->buflen); | |
bca4ed4d | 916 | |
c38e9468 | 917 | out: |
e5f9a813 | 918 | dprint_io_u(io_u, "fill"); |
d9d91e39 | 919 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 920 | return 0; |
10ba535a JA |
921 | } |
922 | ||
6cc0e5aa | 923 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
838bc709 | 924 | { |
2b13e716 | 925 | int idx = 0; |
838bc709 JA |
926 | |
927 | switch (nr) { | |
928 | default: | |
2b13e716 | 929 | idx = 6; |
838bc709 JA |
930 | break; |
931 | case 33 ... 64: | |
2b13e716 | 932 | idx = 5; |
838bc709 JA |
933 | break; |
934 | case 17 ... 32: | |
2b13e716 | 935 | idx = 4; |
838bc709 JA |
936 | break; |
937 | case 9 ... 16: | |
2b13e716 | 938 | idx = 3; |
838bc709 JA |
939 | break; |
940 | case 5 ... 8: | |
2b13e716 | 941 | idx = 2; |
838bc709 JA |
942 | break; |
943 | case 1 ... 4: | |
2b13e716 | 944 | idx = 1; |
838bc709 JA |
945 | case 0: |
946 | break; | |
947 | } | |
948 | ||
2b13e716 | 949 | map[idx]++; |
838bc709 JA |
950 | } |
951 | ||
952 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
953 | { | |
954 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
955 | td->ts.total_submit++; | |
956 | } | |
957 | ||
958 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
959 | { | |
960 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
961 | td->ts.total_complete++; | |
962 | } | |
963 | ||
d8005759 | 964 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 965 | { |
2b13e716 | 966 | int idx = 0; |
71619dc2 JA |
967 | |
968 | switch (td->cur_depth) { | |
969 | default: | |
2b13e716 | 970 | idx = 6; |
a783e61a | 971 | break; |
71619dc2 | 972 | case 32 ... 63: |
2b13e716 | 973 | idx = 5; |
a783e61a | 974 | break; |
71619dc2 | 975 | case 16 ... 31: |
2b13e716 | 976 | idx = 4; |
a783e61a | 977 | break; |
71619dc2 | 978 | case 8 ... 15: |
2b13e716 | 979 | idx = 3; |
a783e61a | 980 | break; |
71619dc2 | 981 | case 4 ... 7: |
2b13e716 | 982 | idx = 2; |
a783e61a | 983 | break; |
71619dc2 | 984 | case 2 ... 3: |
2b13e716 | 985 | idx = 1; |
71619dc2 JA |
986 | case 1: |
987 | break; | |
988 | } | |
989 | ||
2b13e716 | 990 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
991 | } |
992 | ||
d6bb626e | 993 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 994 | { |
2b13e716 | 995 | int idx = 0; |
04a0feae | 996 | |
d6bb626e VF |
997 | assert(nsec < 1000); |
998 | ||
999 | switch (nsec) { | |
1000 | case 750 ... 999: | |
1001 | idx = 9; | |
1002 | break; | |
1003 | case 500 ... 749: | |
1004 | idx = 8; | |
1005 | break; | |
1006 | case 250 ... 499: | |
1007 | idx = 7; | |
1008 | break; | |
1009 | case 100 ... 249: | |
1010 | idx = 6; | |
1011 | break; | |
1012 | case 50 ... 99: | |
1013 | idx = 5; | |
1014 | break; | |
1015 | case 20 ... 49: | |
1016 | idx = 4; | |
1017 | break; | |
1018 | case 10 ... 19: | |
1019 | idx = 3; | |
1020 | break; | |
1021 | case 4 ... 9: | |
1022 | idx = 2; | |
1023 | break; | |
1024 | case 2 ... 3: | |
1025 | idx = 1; | |
1026 | case 0 ... 1: | |
1027 | break; | |
1028 | } | |
1029 | ||
1030 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1031 | td->ts.io_u_lat_n[idx]++; | |
1032 | } | |
1033 | ||
1034 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1035 | { | |
1036 | int idx = 0; | |
1037 | ||
1038 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1039 | |
1040 | switch (usec) { | |
1041 | case 750 ... 999: | |
2b13e716 | 1042 | idx = 9; |
04a0feae JA |
1043 | break; |
1044 | case 500 ... 749: | |
2b13e716 | 1045 | idx = 8; |
04a0feae JA |
1046 | break; |
1047 | case 250 ... 499: | |
2b13e716 | 1048 | idx = 7; |
04a0feae JA |
1049 | break; |
1050 | case 100 ... 249: | |
2b13e716 | 1051 | idx = 6; |
04a0feae JA |
1052 | break; |
1053 | case 50 ... 99: | |
2b13e716 | 1054 | idx = 5; |
04a0feae JA |
1055 | break; |
1056 | case 20 ... 49: | |
2b13e716 | 1057 | idx = 4; |
04a0feae JA |
1058 | break; |
1059 | case 10 ... 19: | |
2b13e716 | 1060 | idx = 3; |
04a0feae JA |
1061 | break; |
1062 | case 4 ... 9: | |
2b13e716 | 1063 | idx = 2; |
04a0feae JA |
1064 | break; |
1065 | case 2 ... 3: | |
2b13e716 | 1066 | idx = 1; |
04a0feae JA |
1067 | case 0 ... 1: |
1068 | break; | |
1069 | } | |
1070 | ||
2b13e716 JA |
1071 | assert(idx < FIO_IO_U_LAT_U_NR); |
1072 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1073 | } |
1074 | ||
d6bb626e | 1075 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1076 | { |
2b13e716 | 1077 | int idx = 0; |
ec118304 | 1078 | |
d6bb626e VF |
1079 | assert(msec >= 1); |
1080 | ||
ec118304 JA |
1081 | switch (msec) { |
1082 | default: | |
2b13e716 | 1083 | idx = 11; |
04a0feae | 1084 | break; |
8abdce66 | 1085 | case 1000 ... 1999: |
2b13e716 | 1086 | idx = 10; |
04a0feae | 1087 | break; |
8abdce66 | 1088 | case 750 ... 999: |
2b13e716 | 1089 | idx = 9; |
04a0feae | 1090 | break; |
8abdce66 | 1091 | case 500 ... 749: |
2b13e716 | 1092 | idx = 8; |
04a0feae | 1093 | break; |
8abdce66 | 1094 | case 250 ... 499: |
2b13e716 | 1095 | idx = 7; |
04a0feae | 1096 | break; |
8abdce66 | 1097 | case 100 ... 249: |
2b13e716 | 1098 | idx = 6; |
04a0feae | 1099 | break; |
8abdce66 | 1100 | case 50 ... 99: |
2b13e716 | 1101 | idx = 5; |
04a0feae | 1102 | break; |
8abdce66 | 1103 | case 20 ... 49: |
2b13e716 | 1104 | idx = 4; |
04a0feae | 1105 | break; |
8abdce66 | 1106 | case 10 ... 19: |
2b13e716 | 1107 | idx = 3; |
04a0feae | 1108 | break; |
8abdce66 | 1109 | case 4 ... 9: |
2b13e716 | 1110 | idx = 2; |
04a0feae | 1111 | break; |
ec118304 | 1112 | case 2 ... 3: |
2b13e716 | 1113 | idx = 1; |
ec118304 JA |
1114 | case 0 ... 1: |
1115 | break; | |
1116 | } | |
1117 | ||
2b13e716 JA |
1118 | assert(idx < FIO_IO_U_LAT_M_NR); |
1119 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1120 | } |
1121 | ||
d6bb626e | 1122 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1123 | { |
d6bb626e VF |
1124 | if (nsec < 1000) |
1125 | io_u_mark_lat_nsec(td, nsec); | |
1126 | else if (nsec < 1000000) | |
1127 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1128 | else |
d6bb626e | 1129 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1130 | } |
1131 | ||
8c07860d JA |
1132 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1133 | { | |
1134 | unsigned long fileno; | |
1135 | ||
1136 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1137 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1138 | unsigned long r; | |
1139 | ||
1140 | r = __rand(&td->next_file_state); | |
1141 | return (unsigned int) ((double) td->o.nr_files | |
1142 | * (r / (frand_max + 1.0))); | |
1143 | } | |
1144 | ||
1145 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1146 | fileno = zipf_next(&td->next_file_zipf); | |
1147 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1148 | fileno = pareto_next(&td->next_file_zipf); | |
1149 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1150 | fileno = gauss_next(&td->next_file_gauss); | |
1151 | else { | |
1152 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1153 | assert(0); | |
1154 | return 0; | |
1155 | } | |
1156 | ||
1157 | return fileno >> FIO_FSERVICE_SHIFT; | |
1158 | } | |
1159 | ||
0aabe160 JA |
1160 | /* |
1161 | * Get next file to service by choosing one at random | |
1162 | */ | |
2cc52930 JA |
1163 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1164 | enum fio_file_flags goodf, | |
d6aed795 | 1165 | enum fio_file_flags badf) |
0aabe160 | 1166 | { |
0aabe160 | 1167 | struct fio_file *f; |
1c178180 | 1168 | int fno; |
0aabe160 JA |
1169 | |
1170 | do { | |
87b10676 | 1171 | int opened = 0; |
4c07ad86 | 1172 | |
8c07860d | 1173 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1174 | |
126d65c6 | 1175 | f = td->files[fno]; |
d6aed795 | 1176 | if (fio_file_done(f)) |
059e63c0 | 1177 | continue; |
1c178180 | 1178 | |
d6aed795 | 1179 | if (!fio_file_open(f)) { |
87b10676 JA |
1180 | int err; |
1181 | ||
002fe734 JA |
1182 | if (td->nr_open_files >= td->o.open_files) |
1183 | return ERR_PTR(-EBUSY); | |
1184 | ||
87b10676 JA |
1185 | err = td_io_open_file(td, f); |
1186 | if (err) | |
1187 | continue; | |
1188 | opened = 1; | |
1189 | } | |
1190 | ||
2ba1c290 JA |
1191 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1192 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1193 | return f; |
2ba1c290 | 1194 | } |
87b10676 JA |
1195 | if (opened) |
1196 | td_io_close_file(td, f); | |
0aabe160 JA |
1197 | } while (1); |
1198 | } | |
1199 | ||
1200 | /* | |
1201 | * Get next file to service by doing round robin between all available ones | |
1202 | */ | |
1c178180 JA |
1203 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1204 | int badf) | |
3d7c391d JA |
1205 | { |
1206 | unsigned int old_next_file = td->next_file; | |
1207 | struct fio_file *f; | |
1208 | ||
1209 | do { | |
87b10676 JA |
1210 | int opened = 0; |
1211 | ||
126d65c6 | 1212 | f = td->files[td->next_file]; |
3d7c391d JA |
1213 | |
1214 | td->next_file++; | |
2dc1bbeb | 1215 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1216 | td->next_file = 0; |
1217 | ||
87b10676 | 1218 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1219 | if (fio_file_done(f)) { |
d5ed68ea | 1220 | f = NULL; |
059e63c0 | 1221 | continue; |
d5ed68ea | 1222 | } |
059e63c0 | 1223 | |
d6aed795 | 1224 | if (!fio_file_open(f)) { |
87b10676 JA |
1225 | int err; |
1226 | ||
002fe734 JA |
1227 | if (td->nr_open_files >= td->o.open_files) |
1228 | return ERR_PTR(-EBUSY); | |
1229 | ||
87b10676 | 1230 | err = td_io_open_file(td, f); |
b5696bfc JA |
1231 | if (err) { |
1232 | dprint(FD_FILE, "error %d on open of %s\n", | |
1233 | err, f->file_name); | |
87c27b45 | 1234 | f = NULL; |
87b10676 | 1235 | continue; |
b5696bfc | 1236 | } |
87b10676 JA |
1237 | opened = 1; |
1238 | } | |
1239 | ||
0b9d69ec JA |
1240 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1241 | f->flags); | |
1c178180 | 1242 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1243 | break; |
1244 | ||
87b10676 JA |
1245 | if (opened) |
1246 | td_io_close_file(td, f); | |
1247 | ||
3d7c391d JA |
1248 | f = NULL; |
1249 | } while (td->next_file != old_next_file); | |
1250 | ||
2ba1c290 | 1251 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1252 | return f; |
1253 | } | |
1254 | ||
7eb36574 | 1255 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1256 | { |
1907dbc6 JA |
1257 | struct fio_file *f; |
1258 | ||
2dc1bbeb | 1259 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1260 | |
b5696bfc | 1261 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1262 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1263 | " nr_files=%d\n", td->nr_open_files, | |
1264 | td->nr_done_files, | |
1265 | td->o.nr_files); | |
bdb4e2e9 | 1266 | return NULL; |
2ba1c290 | 1267 | } |
bdb4e2e9 | 1268 | |
1907dbc6 | 1269 | f = td->file_service_file; |
d6aed795 | 1270 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1271 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1272 | goto out; | |
1273 | if (td->file_service_left--) | |
1274 | goto out; | |
1275 | } | |
1907dbc6 | 1276 | |
a086c257 JA |
1277 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1278 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1279 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1280 | else |
d6aed795 | 1281 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1282 | |
002fe734 JA |
1283 | if (IS_ERR(f)) |
1284 | return f; | |
1285 | ||
1907dbc6 JA |
1286 | td->file_service_file = f; |
1287 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1288 | out: |
0dac421f JA |
1289 | if (f) |
1290 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1291 | else | |
1292 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1293 | return f; |
bdb4e2e9 JA |
1294 | } |
1295 | ||
7eb36574 JA |
1296 | static struct fio_file *get_next_file(struct thread_data *td) |
1297 | { | |
7eb36574 JA |
1298 | return __get_next_file(td); |
1299 | } | |
1300 | ||
002fe734 | 1301 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1302 | { |
1303 | struct fio_file *f; | |
1304 | ||
1305 | do { | |
1306 | f = get_next_file(td); | |
002fe734 JA |
1307 | if (IS_ERR_OR_NULL(f)) |
1308 | return PTR_ERR(f); | |
429f6675 | 1309 | |
429f6675 JA |
1310 | io_u->file = f; |
1311 | get_file(f); | |
1312 | ||
1313 | if (!fill_io_u(td, io_u)) | |
1314 | break; | |
1315 | ||
99952ca7 BVA |
1316 | if (io_u->post_submit) { |
1317 | io_u->post_submit(io_u, false); | |
1318 | io_u->post_submit = NULL; | |
1319 | } | |
1320 | ||
b5696bfc | 1321 | put_file_log(td, f); |
429f6675 | 1322 | td_io_close_file(td, f); |
b5696bfc | 1323 | io_u->file = NULL; |
8c07860d JA |
1324 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1325 | fio_file_reset(td, f); | |
1326 | else { | |
1327 | fio_file_set_done(f); | |
1328 | td->nr_done_files++; | |
1329 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1330 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1331 | } |
429f6675 JA |
1332 | } while (1); |
1333 | ||
1334 | return 0; | |
1335 | } | |
1336 | ||
3e260a46 | 1337 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
c3a32714 | 1338 | unsigned long long tnsec, unsigned long long max_nsec) |
3e260a46 JA |
1339 | { |
1340 | if (!td->error) | |
c3a32714 | 1341 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
3e260a46 JA |
1342 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
1343 | icd->error = ETIMEDOUT; | |
1344 | } | |
1345 | ||
1346 | static void lat_new_cycle(struct thread_data *td) | |
1347 | { | |
1348 | fio_gettime(&td->latency_ts, NULL); | |
1349 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1350 | td->latency_failed = 0; | |
1351 | } | |
1352 | ||
1353 | /* | |
1354 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1355 | * are at QD=1, then it's time to give up. | |
1356 | */ | |
e39c0676 | 1357 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1358 | { |
1359 | if (td->latency_qd == 1) | |
e39c0676 | 1360 | return true; |
3e260a46 JA |
1361 | |
1362 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1363 | |
1364 | if (td->latency_qd == td->latency_qd_low) | |
1365 | td->latency_qd_low--; | |
1366 | ||
3e260a46 JA |
1367 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1368 | ||
1369 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1370 | ||
1371 | /* | |
1372 | * When we ramp QD down, quiesce existing IO to prevent | |
1373 | * a storm of ramp downs due to pending higher depth. | |
1374 | */ | |
1375 | io_u_quiesce(td); | |
1376 | lat_new_cycle(td); | |
e39c0676 | 1377 | return false; |
3e260a46 JA |
1378 | } |
1379 | ||
e39c0676 | 1380 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1381 | { |
1382 | if (td->o.latency_percentile.u.f == 100.0) | |
1383 | return __lat_target_failed(td); | |
1384 | ||
1385 | td->latency_failed++; | |
e39c0676 | 1386 | return false; |
3e260a46 JA |
1387 | } |
1388 | ||
1389 | void lat_target_init(struct thread_data *td) | |
1390 | { | |
6bb58215 JA |
1391 | td->latency_end_run = 0; |
1392 | ||
3e260a46 JA |
1393 | if (td->o.latency_target) { |
1394 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1395 | fio_gettime(&td->latency_ts, NULL); | |
1396 | td->latency_qd = 1; | |
1397 | td->latency_qd_high = td->o.iodepth; | |
1398 | td->latency_qd_low = 1; | |
1399 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1400 | } else | |
1401 | td->latency_qd = td->o.iodepth; | |
1402 | } | |
1403 | ||
6bb58215 JA |
1404 | void lat_target_reset(struct thread_data *td) |
1405 | { | |
1406 | if (!td->latency_end_run) | |
1407 | lat_target_init(td); | |
1408 | } | |
1409 | ||
3e260a46 JA |
1410 | static void lat_target_success(struct thread_data *td) |
1411 | { | |
1412 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1413 | struct thread_options *o = &td->o; |
3e260a46 JA |
1414 | |
1415 | td->latency_qd_low = td->latency_qd; | |
1416 | ||
1417 | /* | |
1418 | * If we haven't failed yet, we double up to a failing value instead | |
1419 | * of bisecting from highest possible queue depth. If we have set | |
1420 | * a limit other than td->o.iodepth, bisect between that. | |
1421 | */ | |
6bb58215 | 1422 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1423 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1424 | else | |
1425 | td->latency_qd *= 2; | |
1426 | ||
6bb58215 JA |
1427 | if (td->latency_qd > o->iodepth) |
1428 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1429 | |
1430 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1431 | |
3e260a46 | 1432 | /* |
6bb58215 JA |
1433 | * Same as last one, we are done. Let it run a latency cycle, so |
1434 | * we get only the results from the targeted depth. | |
3e260a46 | 1435 | */ |
6bb58215 JA |
1436 | if (td->latency_qd == qd) { |
1437 | if (td->latency_end_run) { | |
1438 | dprint(FD_RATE, "We are done\n"); | |
1439 | td->done = 1; | |
1440 | } else { | |
1441 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1442 | io_u_quiesce(td); | |
1443 | td->latency_end_run = 1; | |
1444 | reset_all_stats(td); | |
1445 | reset_io_stats(td); | |
1446 | } | |
1447 | } | |
3e260a46 JA |
1448 | |
1449 | lat_new_cycle(td); | |
1450 | } | |
1451 | ||
1452 | /* | |
1453 | * Check if we can bump the queue depth | |
1454 | */ | |
1455 | void lat_target_check(struct thread_data *td) | |
1456 | { | |
1457 | uint64_t usec_window; | |
1458 | uint64_t ios; | |
1459 | double success_ios; | |
1460 | ||
1461 | usec_window = utime_since_now(&td->latency_ts); | |
1462 | if (usec_window < td->o.latency_window) | |
1463 | return; | |
1464 | ||
1465 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1466 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1467 | success_ios *= 100.0; | |
1468 | ||
1469 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1470 | ||
1471 | if (success_ios >= td->o.latency_percentile.u.f) | |
1472 | lat_target_success(td); | |
1473 | else | |
1474 | __lat_target_failed(td); | |
1475 | } | |
1476 | ||
1477 | /* | |
1478 | * If latency target is enabled, we might be ramping up or down and not | |
1479 | * using the full queue depth available. | |
1480 | */ | |
e39c0676 | 1481 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1482 | { |
1483 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1484 | ||
1485 | if (qempty) | |
e39c0676 | 1486 | return true; |
3e260a46 | 1487 | if (!td->o.latency_target) |
e39c0676 | 1488 | return false; |
3e260a46 JA |
1489 | |
1490 | return td->cur_depth >= td->latency_qd; | |
1491 | } | |
429f6675 | 1492 | |
10ba535a JA |
1493 | struct io_u *__get_io_u(struct thread_data *td) |
1494 | { | |
0cae66f6 | 1495 | struct io_u *io_u = NULL; |
93b45bb2 | 1496 | int ret; |
10ba535a | 1497 | |
ca09be4b JA |
1498 | if (td->stop_io) |
1499 | return NULL; | |
1500 | ||
e8462bd8 JA |
1501 | td_io_u_lock(td); |
1502 | ||
1503 | again: | |
2ae0b204 JA |
1504 | if (!io_u_rempty(&td->io_u_requeues)) |
1505 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1506 | else if (!queue_full(td)) { |
2ae0b204 | 1507 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1508 | |
225ba9e3 | 1509 | io_u->file = NULL; |
6040dabc | 1510 | io_u->buflen = 0; |
10ba535a | 1511 | io_u->resid = 0; |
d7762cf8 | 1512 | io_u->end_io = NULL; |
755200a3 JA |
1513 | } |
1514 | ||
1515 | if (io_u) { | |
0c6e7517 | 1516 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1517 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1518 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1519 | IO_U_F_VER_LIST); | |
0c6e7517 | 1520 | |
755200a3 | 1521 | io_u->error = 0; |
bcd5abfa | 1522 | io_u->acct_ddir = -1; |
10ba535a | 1523 | td->cur_depth++; |
a9da8ab2 | 1524 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1525 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1526 | io_u->ipo = NULL; |
a9da8ab2 | 1527 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1528 | /* |
1529 | * We ran out, wait for async verify threads to finish and | |
1530 | * return one | |
1531 | */ | |
a9da8ab2 | 1532 | assert(!(td->flags & TD_F_CHILD)); |
93b45bb2 BVA |
1533 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
1534 | assert(ret == 0); | |
1dec3e07 | 1535 | goto again; |
10ba535a JA |
1536 | } |
1537 | ||
e8462bd8 | 1538 | td_io_u_unlock(td); |
10ba535a JA |
1539 | return io_u; |
1540 | } | |
1541 | ||
e39c0676 | 1542 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1543 | { |
d72be545 | 1544 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1545 | return false; |
c9a73054 JA |
1546 | if (!td->trim_entries) |
1547 | return false; | |
d72be545 | 1548 | |
c9a73054 JA |
1549 | if (td->trim_batch) { |
1550 | td->trim_batch--; | |
1551 | if (get_next_trim(td, io_u)) | |
1552 | return true; | |
1553 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1554 | td->last_ddir != DDIR_READ) { | |
1555 | td->trim_batch = td->o.trim_batch; | |
1556 | if (!td->trim_batch) | |
1557 | td->trim_batch = td->o.trim_backlog; | |
1558 | if (get_next_trim(td, io_u)) | |
e39c0676 | 1559 | return true; |
2ba1c290 | 1560 | } |
10ba535a | 1561 | |
e39c0676 | 1562 | return false; |
0d29de83 JA |
1563 | } |
1564 | ||
e39c0676 | 1565 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1566 | { |
d72be545 | 1567 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1568 | return false; |
d72be545 JA |
1569 | |
1570 | if (td->io_hist_len) { | |
9e144189 JA |
1571 | int get_verify = 0; |
1572 | ||
d1ece0c7 | 1573 | if (td->verify_batch) |
9e144189 | 1574 | get_verify = 1; |
d1ece0c7 | 1575 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1576 | td->last_ddir != DDIR_READ) { |
1577 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1578 | if (!td->verify_batch) |
1579 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1580 | get_verify = 1; |
1581 | } | |
1582 | ||
d1ece0c7 JA |
1583 | if (get_verify && !get_next_verify(td, io_u)) { |
1584 | td->verify_batch--; | |
e39c0676 | 1585 | return true; |
d1ece0c7 | 1586 | } |
9e144189 JA |
1587 | } |
1588 | ||
e39c0676 | 1589 | return false; |
0d29de83 JA |
1590 | } |
1591 | ||
de789769 JA |
1592 | /* |
1593 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1594 | * easy compressible data for simple de-dupe attempts. Do this for every |
1595 | * 512b block in the range, since that should be the smallest block size | |
1596 | * we can expect from a device. | |
de789769 JA |
1597 | */ |
1598 | static void small_content_scramble(struct io_u *io_u) | |
1599 | { | |
5fff9543 | 1600 | unsigned long long i, nr_blocks = io_u->buflen >> 9; |
23f394d5 | 1601 | unsigned int offset; |
319b073f JA |
1602 | uint64_t boffset, *iptr; |
1603 | char *p; | |
de789769 | 1604 | |
23f394d5 JA |
1605 | if (!nr_blocks) |
1606 | return; | |
1607 | ||
1608 | p = io_u->xfer_buf; | |
fba76ee8 | 1609 | boffset = io_u->offset; |
319b073f JA |
1610 | |
1611 | if (io_u->buf_filled_len) | |
1612 | io_u->buf_filled_len = 0; | |
1613 | ||
1614 | /* | |
1615 | * Generate random index between 0..7. We do chunks of 512b, if | |
1616 | * we assume a cacheline is 64 bytes, then we have 8 of those. | |
1617 | * Scramble content within the blocks in the same cacheline to | |
1618 | * speed things up. | |
1619 | */ | |
1620 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; | |
fad82f76 | 1621 | |
23f394d5 JA |
1622 | for (i = 0; i < nr_blocks; i++) { |
1623 | /* | |
319b073f JA |
1624 | * Fill offset into start of cacheline, time into end |
1625 | * of cacheline | |
23f394d5 | 1626 | */ |
319b073f JA |
1627 | iptr = (void *) p + (offset << 6); |
1628 | *iptr = boffset; | |
1629 | ||
1630 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); | |
1631 | iptr[0] = io_u->start_time.tv_sec; | |
1632 | iptr[1] = io_u->start_time.tv_nsec; | |
23f394d5 | 1633 | |
23f394d5 | 1634 | p += 512; |
fad82f76 | 1635 | boffset += 512; |
23f394d5 | 1636 | } |
de789769 JA |
1637 | } |
1638 | ||
0d29de83 JA |
1639 | /* |
1640 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
5c5c33c1 | 1641 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
0d29de83 JA |
1642 | */ |
1643 | struct io_u *get_io_u(struct thread_data *td) | |
1644 | { | |
1645 | struct fio_file *f; | |
1646 | struct io_u *io_u; | |
de789769 | 1647 | int do_scramble = 0; |
002fe734 | 1648 | long ret = 0; |
0d29de83 JA |
1649 | |
1650 | io_u = __get_io_u(td); | |
1651 | if (!io_u) { | |
1652 | dprint(FD_IO, "__get_io_u failed\n"); | |
1653 | return NULL; | |
1654 | } | |
1655 | ||
1656 | if (check_get_verify(td, io_u)) | |
1657 | goto out; | |
1658 | if (check_get_trim(td, io_u)) | |
1659 | goto out; | |
1660 | ||
755200a3 JA |
1661 | /* |
1662 | * from a requeue, io_u already setup | |
1663 | */ | |
1664 | if (io_u->file) | |
77f392bf | 1665 | goto out; |
755200a3 | 1666 | |
429f6675 JA |
1667 | /* |
1668 | * If using an iolog, grab next piece if any available. | |
1669 | */ | |
d72be545 | 1670 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1671 | if (read_iolog_get(td, io_u)) |
1672 | goto err_put; | |
2ba1c290 | 1673 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1674 | ret = -EBUSY; |
2ba1c290 | 1675 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1676 | goto err_put; |
2ba1c290 | 1677 | } |
5ec10eaa | 1678 | |
429f6675 | 1679 | f = io_u->file; |
002fe734 JA |
1680 | if (!f) { |
1681 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1682 | goto err_put; | |
1683 | } | |
1684 | ||
d6aed795 | 1685 | assert(fio_file_open(f)); |
97af62ce | 1686 | |
ff58fced | 1687 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1688 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1689 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1690 | goto err_put; |
2ba1c290 | 1691 | } |
10ba535a | 1692 | |
f1dfb668 JA |
1693 | f->last_start[io_u->ddir] = io_u->offset; |
1694 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1695 | |
fd68418e | 1696 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1697 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1698 | io_u_fill_buffer(td, io_u, |
1066358a | 1699 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1700 | io_u->buflen); |
ff441ae8 | 1701 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
5c5c33c1 BVA |
1702 | !(td->flags & TD_F_COMPRESS) && |
1703 | !(td->flags & TD_F_DO_VERIFY)) | |
fd68418e JA |
1704 | do_scramble = 1; |
1705 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1706 | /* |
1707 | * Reset the buf_filled parameters so next time if the | |
1708 | * buffer is used for writes it is refilled. | |
1709 | */ | |
cbe8d756 RR |
1710 | io_u->buf_filled_len = 0; |
1711 | } | |
87dc1ab1 | 1712 | } |
10ba535a | 1713 | |
165faf16 JA |
1714 | /* |
1715 | * Set io data pointers. | |
1716 | */ | |
cec6b55d JA |
1717 | io_u->xfer_buf = io_u->buf; |
1718 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1719 | |
6ac7a331 | 1720 | out: |
0d29de83 | 1721 | assert(io_u->file); |
429f6675 | 1722 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1723 | if (!td->o.disable_lat) |
993bf48b | 1724 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1725 | |
de789769 JA |
1726 | if (do_scramble) |
1727 | small_content_scramble(io_u); | |
03553853 | 1728 | |
429f6675 | 1729 | return io_u; |
36167d82 | 1730 | } |
429f6675 | 1731 | err_put: |
2ba1c290 | 1732 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1733 | put_io_u(td, io_u); |
002fe734 | 1734 | return ERR_PTR(ret); |
10ba535a JA |
1735 | } |
1736 | ||
a9da8ab2 | 1737 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1738 | { |
8b28bd41 | 1739 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1740 | |
8b28bd41 DM |
1741 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1742 | return; | |
5451792e | 1743 | |
5fff9543 | 1744 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n", |
709c8313 RE |
1745 | io_u->file ? " on file " : "", |
1746 | io_u->file ? io_u->file->file_name : "", | |
1747 | strerror(io_u->error), | |
1748 | io_ddir_name(io_u->ddir), | |
1749 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1750 | |
5ad7be56 KD |
1751 | if (td->io_ops->errdetails) { |
1752 | char *err = td->io_ops->errdetails(io_u); | |
1753 | ||
1754 | log_err("fio: %s\n", err); | |
1755 | free(err); | |
1756 | } | |
1757 | ||
5451792e JA |
1758 | if (!td->error) |
1759 | td_verror(td, io_u->error, "io_u error"); | |
1760 | } | |
1761 | ||
a9da8ab2 JA |
1762 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1763 | { | |
1764 | __io_u_log_error(td, io_u); | |
1765 | if (td->parent) | |
094e66cb | 1766 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1767 | } |
1768 | ||
e39c0676 | 1769 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1770 | { |
3ecc8c67 JA |
1771 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1772 | || td->o.gtod_reduce; | |
aba6c951 JA |
1773 | } |
1774 | ||
c8eeb9df JA |
1775 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1776 | struct io_completion_data *icd, | |
1777 | const enum fio_ddir idx, unsigned int bytes) | |
1778 | { | |
a9da8ab2 | 1779 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1780 | unsigned long long llnsec = 0; |
c8eeb9df | 1781 | |
75dc383e JA |
1782 | if (td->parent) |
1783 | td = td->parent; | |
1784 | ||
132b1ee4 | 1785 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1786 | return; |
1787 | ||
a9da8ab2 | 1788 | if (no_reduce) |
d6bb626e | 1789 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1790 | |
1791 | if (!td->o.disable_lat) { | |
c3a32714 | 1792 | unsigned long long tnsec; |
c8eeb9df | 1793 | |
d6bb626e VF |
1794 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
1795 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); | |
15501535 | 1796 | |
d4afedfd JA |
1797 | if (td->flags & TD_F_PROFILE_OPS) { |
1798 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1799 | ||
1800 | if (ops->io_u_lat) | |
c3a32714 | 1801 | icd->error = ops->io_u_lat(td, tnsec); |
d4afedfd JA |
1802 | } |
1803 | ||
c3a32714 JA |
1804 | if (td->o.max_latency && tnsec > td->o.max_latency) |
1805 | lat_fatal(td, icd, tnsec, td->o.max_latency); | |
1806 | if (td->o.latency_target && tnsec > td->o.latency_target) { | |
3e260a46 | 1807 | if (lat_target_failed(td)) |
c3a32714 | 1808 | lat_fatal(td, icd, tnsec, td->o.latency_target); |
15501535 | 1809 | } |
c8eeb9df JA |
1810 | } |
1811 | ||
a47591e4 JA |
1812 | if (ddir_rw(idx)) { |
1813 | if (!td->o.disable_clat) { | |
d6bb626e VF |
1814 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
1815 | io_u_mark_latency(td, llnsec); | |
a47591e4 | 1816 | } |
c8eeb9df | 1817 | |
a47591e4 | 1818 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1819 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1820 | |
a47591e4 JA |
1821 | if (no_reduce && per_unit_log(td->iops_log)) |
1822 | add_iops_sample(td, io_u, bytes); | |
b2b3eefe JA |
1823 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
1824 | add_sync_clat_sample(&td->ts, llnsec); | |
66347cfa DE |
1825 | |
1826 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1827 | uint32_t *info = io_u_block_info(td, io_u); | |
1828 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1829 | if (io_u->ddir == DDIR_TRIM) { | |
1830 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1831 | BLOCK_INFO_TRIMS(*info) + 1); | |
1832 | } else if (io_u->ddir == DDIR_WRITE) { | |
1833 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1834 | *info); | |
1835 | } | |
1836 | } | |
1837 | } | |
c8eeb9df JA |
1838 | } |
1839 | ||
94a6e1bb JA |
1840 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1841 | uint64_t offset, unsigned int bytes) | |
1842 | { | |
1843 | int idx; | |
1844 | ||
639ad1ea JA |
1845 | if (!f) |
1846 | return; | |
1847 | ||
94a6e1bb JA |
1848 | if (f->first_write == -1ULL || offset < f->first_write) |
1849 | f->first_write = offset; | |
1850 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1851 | f->last_write = offset + bytes; | |
1852 | ||
1853 | if (!f->last_write_comp) | |
1854 | return; | |
1855 | ||
1856 | idx = f->last_write_idx++; | |
1857 | f->last_write_comp[idx] = offset; | |
1858 | if (f->last_write_idx == td->o.iodepth) | |
1859 | f->last_write_idx = 0; | |
1860 | } | |
1861 | ||
b2b3eefe JA |
1862 | static bool should_account(struct thread_data *td) |
1863 | { | |
1864 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || | |
1865 | td->runstate == TD_VERIFYING); | |
1866 | } | |
1867 | ||
f8b0bd10 | 1868 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1869 | struct io_completion_data *icd) |
10ba535a | 1870 | { |
f8b0bd10 JA |
1871 | struct io_u *io_u = *io_u_ptr; |
1872 | enum fio_ddir ddir = io_u->ddir; | |
1873 | struct fio_file *f = io_u->file; | |
10ba535a | 1874 | |
e5f9a813 | 1875 | dprint_io_u(io_u, "complete"); |
2ba1c290 | 1876 | |
0c6e7517 | 1877 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1878 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1879 | |
1880 | /* | |
1881 | * Mark IO ok to verify | |
1882 | */ | |
1883 | if (io_u->ipo) { | |
890b6656 JA |
1884 | /* |
1885 | * Remove errored entry from the verification list | |
1886 | */ | |
1887 | if (io_u->error) | |
1888 | unlog_io_piece(td, io_u); | |
1889 | else { | |
1890 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1891 | write_barrier(); | |
1892 | } | |
f9401285 JA |
1893 | } |
1894 | ||
f8b0bd10 | 1895 | if (ddir_sync(ddir)) { |
2ea93f98 | 1896 | td->last_was_sync = true; |
44f29692 JA |
1897 | if (f) { |
1898 | f->first_write = -1ULL; | |
1899 | f->last_write = -1ULL; | |
1900 | } | |
b2b3eefe JA |
1901 | if (should_account(td)) |
1902 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); | |
87dc1ab1 JA |
1903 | return; |
1904 | } | |
1905 | ||
2ea93f98 | 1906 | td->last_was_sync = false; |
f8b0bd10 | 1907 | td->last_ddir = ddir; |
87dc1ab1 | 1908 | |
f8b0bd10 | 1909 | if (!io_u->error && ddir_rw(ddir)) { |
5fff9543 | 1910 | unsigned long long bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1911 | int ret; |
10ba535a | 1912 | |
f8b0bd10 | 1913 | td->io_blocks[ddir]++; |
f8b0bd10 | 1914 | td->io_bytes[ddir] += bytes; |
ae2fafc8 | 1915 | |
e1c325d2 JA |
1916 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
1917 | td->this_io_blocks[ddir]++; | |
f8b0bd10 | 1918 | td->this_io_bytes[ddir] += bytes; |
e1c325d2 | 1919 | } |
f8b0bd10 | 1920 | |
639ad1ea | 1921 | if (ddir == DDIR_WRITE) |
94a6e1bb | 1922 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 1923 | |
b2b3eefe | 1924 | if (should_account(td)) |
f8b0bd10 | 1925 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1926 | |
f8b0bd10 | 1927 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1928 | |
d7762cf8 | 1929 | if (io_u->end_io) { |
f8b0bd10 JA |
1930 | ret = io_u->end_io(td, io_u_ptr); |
1931 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1932 | if (ret && !icd->error) |
1933 | icd->error = ret; | |
1934 | } | |
ff58fced | 1935 | } else if (io_u->error) { |
10ba535a | 1936 | icd->error = io_u->error; |
5451792e JA |
1937 | io_u_log_error(td, io_u); |
1938 | } | |
8b28bd41 | 1939 | if (icd->error) { |
f8b0bd10 JA |
1940 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1941 | ||
8b28bd41 DM |
1942 | if (!td_non_fatal_error(td, eb, icd->error)) |
1943 | return; | |
f8b0bd10 | 1944 | |
f2bba182 RR |
1945 | /* |
1946 | * If there is a non_fatal error, then add to the error count | |
1947 | * and clear all the errors. | |
1948 | */ | |
1949 | update_error_count(td, icd->error); | |
1950 | td_clear_error(td); | |
1951 | icd->error = 0; | |
f8b0bd10 JA |
1952 | if (io_u) |
1953 | io_u->error = 0; | |
f2bba182 | 1954 | } |
10ba535a JA |
1955 | } |
1956 | ||
9520ebb9 JA |
1957 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1958 | int nr) | |
10ba535a | 1959 | { |
6eaf09d6 | 1960 | int ddir; |
aba6c951 JA |
1961 | |
1962 | if (!gtod_reduce(td)) | |
9520ebb9 | 1963 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1964 | |
3af6ef39 JA |
1965 | icd->nr = nr; |
1966 | ||
10ba535a | 1967 | icd->error = 0; |
c1f50f76 | 1968 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 1969 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
1970 | } |
1971 | ||
97601024 JA |
1972 | static void ios_completed(struct thread_data *td, |
1973 | struct io_completion_data *icd) | |
36167d82 JA |
1974 | { |
1975 | struct io_u *io_u; | |
1976 | int i; | |
1977 | ||
10ba535a JA |
1978 | for (i = 0; i < icd->nr; i++) { |
1979 | io_u = td->io_ops->event(td, i); | |
1980 | ||
f8b0bd10 | 1981 | io_completed(td, &io_u, icd); |
e8462bd8 | 1982 | |
f8b0bd10 | 1983 | if (io_u) |
e8462bd8 | 1984 | put_io_u(td, io_u); |
10ba535a JA |
1985 | } |
1986 | } | |
97601024 | 1987 | |
e7e6cfb4 JA |
1988 | /* |
1989 | * Complete a single io_u for the sync engines. | |
1990 | */ | |
55312f9f | 1991 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1992 | { |
1993 | struct io_completion_data icd; | |
55312f9f | 1994 | int ddir; |
97601024 | 1995 | |
9520ebb9 | 1996 | init_icd(td, &icd, 1); |
f8b0bd10 | 1997 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1998 | |
f8b0bd10 | 1999 | if (io_u) |
e8462bd8 | 2000 | put_io_u(td, io_u); |
97601024 | 2001 | |
581e7141 JA |
2002 | if (icd.error) { |
2003 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2004 | return -1; | |
2005 | } | |
97601024 | 2006 | |
c1f50f76 | 2007 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2008 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
2009 | |
2010 | return 0; | |
97601024 JA |
2011 | } |
2012 | ||
e7e6cfb4 JA |
2013 | /* |
2014 | * Called to complete min_events number of io for the async engines. | |
2015 | */ | |
55312f9f | 2016 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2017 | { |
97601024 | 2018 | struct io_completion_data icd; |
00de55ef | 2019 | struct timespec *tvp = NULL; |
55312f9f | 2020 | int ret, ddir; |
4d06a338 | 2021 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2022 | |
12bb8569 | 2023 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2024 | |
4950421a | 2025 | if (!min_evts) |
00de55ef | 2026 | tvp = &ts; |
5fb4b366 RE |
2027 | else if (min_evts > td->cur_depth) |
2028 | min_evts = td->cur_depth; | |
97601024 | 2029 | |
82407585 RP |
2030 | /* No worries, td_io_getevents fixes min and max if they are |
2031 | * set incorrectly */ | |
2032 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2033 | if (ret < 0) { |
e1161c32 | 2034 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2035 | return ret; |
2036 | } else if (!ret) | |
2037 | return ret; | |
2038 | ||
9520ebb9 | 2039 | init_icd(td, &icd, ret); |
97601024 | 2040 | ios_completed(td, &icd); |
581e7141 JA |
2041 | if (icd.error) { |
2042 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2043 | return -1; | |
2044 | } | |
97601024 | 2045 | |
c1f50f76 | 2046 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2047 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2048 | |
0d593542 | 2049 | return ret; |
97601024 | 2050 | } |
7e77dd02 JA |
2051 | |
2052 | /* | |
2053 | * Call when io_u is really queued, to update the submission latency. | |
2054 | */ | |
2055 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2056 | { | |
8243be59 | 2057 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2058 | unsigned long slat_time; |
7e77dd02 | 2059 | |
d6bb626e | 2060 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2061 | |
2062 | if (td->parent) | |
2063 | td = td->parent; | |
2064 | ||
ae588852 JA |
2065 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
2066 | io_u->offset); | |
9520ebb9 | 2067 | } |
7e77dd02 | 2068 | } |
433afcb4 | 2069 | |
5c94b008 JA |
2070 | /* |
2071 | * See if we should reuse the last seed, if dedupe is enabled | |
2072 | */ | |
9451b93e | 2073 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2074 | { |
2075 | unsigned int v; | |
5c94b008 JA |
2076 | |
2077 | if (!td->o.dedupe_percentage) | |
2078 | return &td->buf_state; | |
732eedd0 | 2079 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2080 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2081 | return &td->buf_state; | |
732eedd0 | 2082 | } |
5c94b008 | 2083 | |
1bd5d213 | 2084 | v = rand_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2085 | |
2086 | if (v <= td->o.dedupe_percentage) | |
2087 | return &td->buf_state_prev; | |
2088 | ||
2089 | return &td->buf_state; | |
2090 | } | |
2091 | ||
9451b93e | 2092 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2093 | { |
9451b93e JA |
2094 | if (td->o.dedupe_percentage == 100) |
2095 | frand_copy(rs, &td->buf_state_prev); | |
2096 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2097 | frand_copy(&td->buf_state_prev, rs); |
2098 | } | |
2099 | ||
5fff9543 JF |
2100 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write, |
2101 | unsigned long long max_bs) | |
5973cafb | 2102 | { |
d1af2894 JA |
2103 | struct thread_options *o = &td->o; |
2104 | ||
15600335 JA |
2105 | if (o->mem_type == MEM_CUDA_MALLOC) |
2106 | return; | |
03553853 | 2107 | |
4eff3e57 | 2108 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2109 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2110 | struct frand_state *rs; |
5fff9543 JF |
2111 | unsigned long long left = max_bs; |
2112 | unsigned long long this_write; | |
5c94b008 | 2113 | |
1066358a | 2114 | do { |
9451b93e | 2115 | rs = get_buf_state(td); |
9c42684e | 2116 | |
1066358a | 2117 | min_write = min(min_write, left); |
f97a43a1 | 2118 | |
1066358a | 2119 | if (perc) { |
4b157ac6 | 2120 | this_write = min_not_zero(min_write, |
5fff9543 | 2121 | (unsigned long long) td->o.compress_chunk); |
1e7f82e2 JA |
2122 | |
2123 | fill_random_buf_percentage(rs, buf, perc, | |
2124 | this_write, this_write, | |
2125 | o->buffer_pattern, | |
2126 | o->buffer_pattern_bytes); | |
2127 | } else { | |
1066358a | 2128 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2129 | this_write = min_write; |
2130 | } | |
1066358a | 2131 | |
1e7f82e2 JA |
2132 | buf += this_write; |
2133 | left -= this_write; | |
9451b93e | 2134 | save_buf_state(td, rs); |
1066358a | 2135 | } while (left); |
d1af2894 JA |
2136 | } else if (o->buffer_pattern_bytes) |
2137 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2138 | else if (o->zero_buffers) |
cc86c395 | 2139 | memset(buf, 0, max_bs); |
999d245e | 2140 | else |
9451b93e | 2141 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2142 | } |
2143 | ||
2144 | /* | |
2145 | * "randomly" fill the buffer contents | |
2146 | */ | |
2147 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
5fff9543 | 2148 | unsigned long long min_write, unsigned long long max_bs) |
cc86c395 JA |
2149 | { |
2150 | io_u->buf_filled_len = 0; | |
2151 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2152 | } |
e2c75fc4 TK |
2153 | |
2154 | static int do_sync_file_range(const struct thread_data *td, | |
2155 | struct fio_file *f) | |
2156 | { | |
2157 | off64_t offset, nbytes; | |
2158 | ||
2159 | offset = f->first_write; | |
2160 | nbytes = f->last_write - f->first_write; | |
2161 | ||
2162 | if (!nbytes) | |
2163 | return 0; | |
2164 | ||
2165 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2166 | } | |
2167 | ||
2168 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2169 | { | |
2170 | int ret; | |
2171 | ||
2172 | if (io_u->ddir == DDIR_SYNC) { | |
2173 | ret = fsync(io_u->file->fd); | |
2174 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2175 | #ifdef CONFIG_FDATASYNC | |
2176 | ret = fdatasync(io_u->file->fd); | |
2177 | #else | |
2178 | ret = io_u->xfer_buflen; | |
2179 | io_u->error = EINVAL; | |
2180 | #endif | |
2181 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2182 | ret = do_sync_file_range(td, io_u->file); | |
2183 | else { | |
2184 | ret = io_u->xfer_buflen; | |
2185 | io_u->error = EINVAL; | |
2186 | } | |
2187 | ||
2188 | if (ret < 0) | |
2189 | io_u->error = errno; | |
2190 | ||
2191 | return ret; | |
2192 | } | |
2193 | ||
2194 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2195 | { | |
2196 | #ifndef FIO_HAVE_TRIM | |
2197 | io_u->error = EINVAL; | |
2198 | return 0; | |
2199 | #else | |
2200 | struct fio_file *f = io_u->file; | |
2201 | int ret; | |
2202 | ||
496b1f9e | 2203 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2204 | if (!ret) |
2205 | return io_u->xfer_buflen; | |
2206 | ||
2207 | io_u->error = ret; | |
2208 | return 0; | |
2209 | #endif | |
2210 | } |