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 | */ | |
9bf2061e | 34 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 35 | { |
5fff9543 | 36 | unsigned long long min_bs = td->o.min_bs[io_u->ddir]; |
9bf2061e | 37 | struct fio_file *f = io_u->file; |
5fff9543 | 38 | unsigned long long nr_blocks; |
1ae83d45 | 39 | uint64_t block; |
df415585 | 40 | |
1ae83d45 | 41 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 42 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 43 | |
2ab9e98b | 44 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 45 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 46 | |
51ede0b1 JA |
47 | if ((nr_blocks * min_bs) < io_u->buflen) |
48 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
49 | } |
50 | ||
74776733 JA |
51 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
52 | enum fio_ddir ddir) | |
2ba1c290 | 53 | { |
74776733 JA |
54 | uint64_t max_blocks; |
55 | uint64_t max_size; | |
2ba1c290 | 56 | |
ff58fced JA |
57 | assert(ddir_rw(ddir)); |
58 | ||
d9dd70f7 JA |
59 | /* |
60 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
79591fa9 | 61 | * -> not for now since there is code assuming it could go either. |
d9dd70f7 JA |
62 | */ |
63 | max_size = f->io_size; | |
64 | if (max_size > f->real_file_size) | |
65 | max_size = f->real_file_size; | |
66 | ||
7b865a2f | 67 | if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range) |
ed335855 SN |
68 | max_size = td->o.zone_range; |
69 | ||
0412d12e JE |
70 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
71 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
72 | ||
1ae83d45 | 73 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
74 | if (!max_blocks) |
75 | return 0; | |
76 | ||
67778e88 | 77 | return max_blocks; |
2ba1c290 JA |
78 | } |
79 | ||
e25839d4 | 80 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
e0a04ac1 JA |
81 | enum fio_ddir ddir, uint64_t *b, |
82 | uint64_t lastb) | |
ec4015da | 83 | { |
6f49f8bc | 84 | uint64_t r; |
5e0baa7f | 85 | |
c3546b53 JA |
86 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
87 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { | |
6f49f8bc | 88 | |
d6b72507 | 89 | r = __rand(&td->random_state); |
8055e41d | 90 | |
4b91ee8f | 91 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 92 | |
e0a04ac1 | 93 | *b = lastb * (r / (rand_max(&td->random_state) + 1.0)); |
51ede0b1 | 94 | } else { |
8055e41d | 95 | uint64_t off = 0; |
43c63a78 | 96 | |
967d1b63 JA |
97 | assert(fio_file_lfsr(f)); |
98 | ||
6f49f8bc | 99 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 100 | return 1; |
ec4015da | 101 | |
8055e41d JA |
102 | *b = off; |
103 | } | |
0ce8b119 | 104 | |
ec4015da | 105 | /* |
51ede0b1 | 106 | * if we are not maintaining a random map, we are done. |
ec4015da | 107 | */ |
51ede0b1 JA |
108 | if (!file_randommap(td, f)) |
109 | goto ret; | |
43c63a78 JA |
110 | |
111 | /* | |
51ede0b1 | 112 | * calculate map offset and check if it's free |
43c63a78 | 113 | */ |
51ede0b1 JA |
114 | if (random_map_free(f, *b)) |
115 | goto ret; | |
116 | ||
4b91ee8f JA |
117 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
118 | (unsigned long long) *b); | |
51ede0b1 | 119 | |
7ebd796f | 120 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
121 | if (*b == (uint64_t) -1ULL) |
122 | return 1; | |
0ce8b119 JA |
123 | ret: |
124 | return 0; | |
ec4015da JA |
125 | } |
126 | ||
925fee33 JA |
127 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
128 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 129 | uint64_t *b) |
e25839d4 | 130 | { |
9c6f6316 | 131 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
132 | return 0; |
133 | } | |
134 | ||
925fee33 JA |
135 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
136 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 137 | uint64_t *b) |
925fee33 | 138 | { |
9c6f6316 | 139 | *b = pareto_next(&f->zipf); |
925fee33 JA |
140 | return 0; |
141 | } | |
142 | ||
56d9fa4b JA |
143 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
144 | struct fio_file *f, enum fio_ddir ddir, | |
145 | uint64_t *b) | |
146 | { | |
147 | *b = gauss_next(&f->gauss); | |
148 | return 0; | |
149 | } | |
150 | ||
59466396 JA |
151 | static int __get_next_rand_offset_zoned_abs(struct thread_data *td, |
152 | struct fio_file *f, | |
153 | enum fio_ddir ddir, uint64_t *b) | |
154 | { | |
155 | struct zone_split_index *zsi; | |
e3c8c108 | 156 | uint64_t lastb, send, stotal; |
59466396 JA |
157 | unsigned int v; |
158 | ||
159 | lastb = last_block(td, f, ddir); | |
160 | if (!lastb) | |
161 | return 1; | |
162 | ||
163 | if (!td->o.zone_split_nr[ddir]) { | |
164 | bail: | |
165 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
166 | } | |
167 | ||
168 | /* | |
169 | * Generate a value, v, between 1 and 100, both inclusive | |
170 | */ | |
1bd5d213 | 171 | v = rand_between(&td->zone_state, 1, 100); |
59466396 | 172 | |
e3c8c108 JA |
173 | /* |
174 | * Find our generated table. 'send' is the end block of this zone, | |
175 | * 'stotal' is our start offset. | |
176 | */ | |
59466396 JA |
177 | zsi = &td->zone_state_index[ddir][v - 1]; |
178 | stotal = zsi->size_prev / td->o.ba[ddir]; | |
179 | send = zsi->size / td->o.ba[ddir]; | |
180 | ||
181 | /* | |
182 | * Should never happen | |
183 | */ | |
184 | if (send == -1U) { | |
264e3d30 | 185 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
59466396 | 186 | log_err("fio: bug in zoned generation\n"); |
59466396 JA |
187 | goto bail; |
188 | } else if (send > lastb) { | |
189 | /* | |
190 | * This happens if the user specifies ranges that exceed | |
191 | * the file/device size. We can't handle that gracefully, | |
192 | * so error and exit. | |
193 | */ | |
194 | log_err("fio: zoned_abs sizes exceed file size\n"); | |
195 | return 1; | |
196 | } | |
197 | ||
198 | /* | |
e3c8c108 | 199 | * Generate index from 0..send-stotal |
59466396 | 200 | */ |
e3c8c108 | 201 | if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1) |
59466396 JA |
202 | return 1; |
203 | ||
e3c8c108 | 204 | *b += stotal; |
59466396 JA |
205 | return 0; |
206 | } | |
207 | ||
e0a04ac1 JA |
208 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
209 | struct fio_file *f, enum fio_ddir ddir, | |
210 | uint64_t *b) | |
211 | { | |
212 | unsigned int v, send, stotal; | |
213 | uint64_t offset, lastb; | |
e0a04ac1 JA |
214 | struct zone_split_index *zsi; |
215 | ||
216 | lastb = last_block(td, f, ddir); | |
217 | if (!lastb) | |
218 | return 1; | |
219 | ||
220 | if (!td->o.zone_split_nr[ddir]) { | |
221 | bail: | |
222 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
223 | } | |
224 | ||
225 | /* | |
226 | * Generate a value, v, between 1 and 100, both inclusive | |
227 | */ | |
1bd5d213 | 228 | v = rand_between(&td->zone_state, 1, 100); |
e0a04ac1 JA |
229 | |
230 | zsi = &td->zone_state_index[ddir][v - 1]; | |
231 | stotal = zsi->size_perc_prev; | |
232 | send = zsi->size_perc; | |
233 | ||
234 | /* | |
235 | * Should never happen | |
236 | */ | |
237 | if (send == -1U) { | |
264e3d30 | 238 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
e0a04ac1 | 239 | log_err("fio: bug in zoned generation\n"); |
e0a04ac1 JA |
240 | goto bail; |
241 | } | |
242 | ||
243 | /* | |
244 | * 'send' is some percentage below or equal to 100 that | |
245 | * marks the end of the current IO range. 'stotal' marks | |
246 | * the start, in percent. | |
247 | */ | |
248 | if (stotal) | |
249 | offset = stotal * lastb / 100ULL; | |
250 | else | |
251 | offset = 0; | |
252 | ||
253 | lastb = lastb * (send - stotal) / 100ULL; | |
254 | ||
255 | /* | |
256 | * Generate index from 0..send-of-lastb | |
257 | */ | |
258 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
259 | return 1; | |
260 | ||
261 | /* | |
262 | * Add our start offset, if any | |
263 | */ | |
264 | if (offset) | |
265 | *b += offset; | |
266 | ||
267 | return 0; | |
268 | } | |
56d9fa4b | 269 | |
f31feaa2 JA |
270 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
271 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 | 272 | { |
e0a04ac1 JA |
273 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
274 | uint64_t lastb; | |
275 | ||
276 | lastb = last_block(td, f, ddir); | |
277 | if (!lastb) | |
278 | return 1; | |
279 | ||
280 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
281 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
e25839d4 | 282 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
925fee33 JA |
283 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
284 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
285 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
286 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e0a04ac1 JA |
287 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
288 | return __get_next_rand_offset_zoned(td, f, ddir, b); | |
59466396 JA |
289 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS) |
290 | return __get_next_rand_offset_zoned_abs(td, f, ddir, b); | |
e25839d4 JA |
291 | |
292 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
293 | return 1; | |
294 | } | |
295 | ||
e39c0676 | 296 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
297 | { |
298 | unsigned int v; | |
211c9b89 | 299 | |
d9472271 | 300 | if (td->o.perc_rand[ddir] == 100) |
e39c0676 | 301 | return true; |
211c9b89 | 302 | |
1bd5d213 | 303 | v = rand_between(&td->seq_rand_state[ddir], 1, 100); |
211c9b89 | 304 | |
d9472271 | 305 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
306 | } |
307 | ||
0bcf41cd JA |
308 | static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f) |
309 | { | |
310 | struct thread_options *o = &td->o; | |
311 | ||
312 | if (o->invalidate_cache && !o->odirect) { | |
313 | int fio_unused ret; | |
314 | ||
315 | ret = file_invalidate_cache(td, f); | |
316 | } | |
317 | } | |
318 | ||
38dad62d | 319 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 320 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 321 | { |
c04e4661 DE |
322 | if (!get_next_rand_offset(td, f, ddir, b)) |
323 | return 0; | |
324 | ||
8c07860d JA |
325 | if (td->o.time_based || |
326 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { | |
33c48814 | 327 | fio_file_reset(td, f); |
80f02150 | 328 | loop_cache_invalidate(td, f); |
c04e4661 DE |
329 | if (!get_next_rand_offset(td, f, ddir, b)) |
330 | return 0; | |
38dad62d JA |
331 | } |
332 | ||
c04e4661 | 333 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 334 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 335 | (unsigned long long) f->real_file_size); |
c04e4661 | 336 | return 1; |
38dad62d JA |
337 | } |
338 | ||
37cf9e3c | 339 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 340 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 341 | { |
ac002339 JA |
342 | struct thread_options *o = &td->o; |
343 | ||
ff58fced JA |
344 | assert(ddir_rw(ddir)); |
345 | ||
17373ce2 JA |
346 | /* |
347 | * If we reach the end for a time based run, reset us back to 0 | |
348 | * and invalidate the cache, if we need to. | |
349 | */ | |
f1dfb668 | 350 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
19ddc35b | 351 | o->time_based) { |
c89daa4a | 352 | f->last_pos[ddir] = f->file_offset; |
0bcf41cd | 353 | loop_cache_invalidate(td, f); |
19ddc35b | 354 | } |
c04e4661 | 355 | |
f1dfb668 | 356 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 357 | uint64_t pos; |
059b0802 | 358 | |
69b98f11 JA |
359 | /* |
360 | * Only rewind if we already hit the end | |
361 | */ | |
362 | if (f->last_pos[ddir] == f->file_offset && | |
363 | f->file_offset && o->ddir_seq_add < 0) { | |
c22825bb JA |
364 | if (f->real_file_size > f->io_size) |
365 | f->last_pos[ddir] = f->io_size; | |
366 | else | |
367 | f->last_pos[ddir] = f->real_file_size; | |
368 | } | |
a66da7a2 | 369 | |
f1dfb668 | 370 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
371 | if (pos && o->ddir_seq_add) { |
372 | pos += o->ddir_seq_add; | |
373 | ||
374 | /* | |
375 | * If we reach beyond the end of the file | |
376 | * with holed IO, wrap around to the | |
b0a84f48 JA |
377 | * beginning again. If we're doing backwards IO, |
378 | * wrap to the end. | |
ac002339 | 379 | */ |
b0a84f48 JA |
380 | if (pos >= f->real_file_size) { |
381 | if (o->ddir_seq_add > 0) | |
382 | pos = f->file_offset; | |
c22825bb JA |
383 | else { |
384 | if (f->real_file_size > f->io_size) | |
385 | pos = f->io_size; | |
386 | else | |
387 | pos = f->real_file_size; | |
388 | ||
389 | pos += o->ddir_seq_add; | |
390 | } | |
b0a84f48 | 391 | } |
ac002339 | 392 | } |
059b0802 | 393 | |
37cf9e3c | 394 | *offset = pos; |
38dad62d JA |
395 | return 0; |
396 | } | |
397 | ||
398 | return 1; | |
399 | } | |
400 | ||
401 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d | 402 | enum fio_ddir ddir, int rw_seq, |
ec370c22 | 403 | bool *is_random) |
38dad62d JA |
404 | { |
405 | struct fio_file *f = io_u->file; | |
1ae83d45 | 406 | uint64_t b, offset; |
38dad62d JA |
407 | int ret; |
408 | ||
ff58fced JA |
409 | assert(ddir_rw(ddir)); |
410 | ||
37cf9e3c JA |
411 | b = offset = -1ULL; |
412 | ||
38dad62d | 413 | if (rw_seq) { |
211c9b89 | 414 | if (td_random(td)) { |
6aca9b3d | 415 | if (should_do_random(td, ddir)) { |
211c9b89 | 416 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 417 | *is_random = true; |
6aca9b3d | 418 | } else { |
ec370c22 | 419 | *is_random = false; |
1651e431 | 420 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
421 | ret = get_next_seq_offset(td, f, ddir, &offset); |
422 | if (ret) | |
423 | ret = get_next_rand_block(td, f, ddir, &b); | |
424 | } | |
6aca9b3d | 425 | } else { |
ec370c22 | 426 | *is_random = false; |
37cf9e3c | 427 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 428 | } |
38dad62d | 429 | } else { |
1651e431 | 430 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
ec370c22 | 431 | *is_random = false; |
38dad62d JA |
432 | |
433 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 434 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 435 | if (ret) { |
37cf9e3c | 436 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 437 | *is_random = false; |
6aca9b3d | 438 | } |
38dad62d | 439 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
440 | if (f->last_start[ddir] != -1ULL) |
441 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 442 | else |
37cf9e3c | 443 | offset = 0; |
38dad62d JA |
444 | ret = 0; |
445 | } else { | |
446 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
447 | ret = 1; | |
448 | } | |
449 | } | |
6d68b997 | 450 | |
37cf9e3c JA |
451 | if (!ret) { |
452 | if (offset != -1ULL) | |
453 | io_u->offset = offset; | |
454 | else if (b != -1ULL) | |
455 | io_u->offset = b * td->o.ba[ddir]; | |
456 | else { | |
4e0a8fa2 | 457 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
458 | ret = 1; |
459 | } | |
460 | } | |
461 | ||
38dad62d JA |
462 | return ret; |
463 | } | |
464 | ||
10ba535a JA |
465 | /* |
466 | * For random io, generate a random new block and see if it's used. Repeat | |
467 | * until we find a free one. For sequential io, just return the end of | |
468 | * the last io issued. | |
469 | */ | |
e8fb335e | 470 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 471 | bool *is_random) |
10ba535a | 472 | { |
9bf2061e | 473 | struct fio_file *f = io_u->file; |
4ba66134 | 474 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 475 | int rw_seq_hit = 0; |
10ba535a | 476 | |
ff58fced JA |
477 | assert(ddir_rw(ddir)); |
478 | ||
38dad62d JA |
479 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
480 | rw_seq_hit = 1; | |
5736c10d | 481 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 482 | } |
211097b2 | 483 | |
6aca9b3d | 484 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 485 | return 1; |
10ba535a | 486 | |
009bd847 JA |
487 | if (io_u->offset >= f->io_size) { |
488 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
489 | (unsigned long long) io_u->offset, |
490 | (unsigned long long) f->io_size); | |
009bd847 JA |
491 | return 1; |
492 | } | |
493 | ||
494 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
495 | if (io_u->offset >= f->real_file_size) { |
496 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
497 | (unsigned long long) io_u->offset, |
498 | (unsigned long long) f->real_file_size); | |
10ba535a | 499 | return 1; |
2ba1c290 | 500 | } |
10ba535a JA |
501 | |
502 | return 0; | |
503 | } | |
504 | ||
e39c0676 | 505 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
5fff9543 | 506 | unsigned long long buflen) |
79944128 JA |
507 | { |
508 | struct fio_file *f = io_u->file; | |
509 | ||
bedc9dc2 | 510 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
511 | } |
512 | ||
5fff9543 | 513 | static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 514 | bool is_random) |
10ba535a | 515 | { |
6aca9b3d | 516 | int ddir = io_u->ddir; |
5fff9543 JF |
517 | unsigned long long buflen = 0; |
518 | unsigned long long minbs, maxbs; | |
3dd29f7c | 519 | uint64_t frand_max, r; |
7c961359 | 520 | bool power_2; |
10ba535a | 521 | |
9ee1c647 | 522 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
523 | |
524 | if (td->o.bs_is_seq_rand) | |
ec370c22 | 525 | ddir = is_random ? DDIR_WRITE : DDIR_READ; |
ff58fced | 526 | |
f3059de1 JA |
527 | minbs = td->o.min_bs[ddir]; |
528 | maxbs = td->o.max_bs[ddir]; | |
529 | ||
79944128 JA |
530 | if (minbs == maxbs) |
531 | return minbs; | |
532 | ||
52c58027 JA |
533 | /* |
534 | * If we can't satisfy the min block size from here, then fail | |
535 | */ | |
536 | if (!io_u_fits(td, io_u, minbs)) | |
537 | return 0; | |
538 | ||
2f282cec | 539 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 540 | do { |
2f282cec | 541 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 542 | |
720e84ad | 543 | if (!td->o.bssplit_nr[ddir]) { |
5fff9543 | 544 | buflen = minbs + (unsigned long long) ((double) maxbs * |
c3546b53 | 545 | (r / (frand_max + 1.0))); |
5ec10eaa | 546 | } else { |
3dd29f7c | 547 | long long perc = 0; |
564ca972 JA |
548 | unsigned int i; |
549 | ||
720e84ad JA |
550 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
551 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
552 | |
553 | buflen = bsp->bs; | |
554 | perc += bsp->perc; | |
3dd29f7c JA |
555 | if (!perc) |
556 | break; | |
557 | if ((r / perc <= frand_max / 100ULL) && | |
79944128 | 558 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
559 | break; |
560 | } | |
561 | } | |
79944128 | 562 | |
17a6b702 PL |
563 | power_2 = is_power_of_2(minbs); |
564 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 565 | buflen &= ~(minbs - 1); |
17a6b702 PL |
566 | else if (!td->o.bs_unaligned && !power_2) |
567 | buflen -= buflen % minbs; | |
79944128 | 568 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 569 | |
10ba535a JA |
570 | return buflen; |
571 | } | |
572 | ||
afe24a5a JA |
573 | static void set_rwmix_bytes(struct thread_data *td) |
574 | { | |
afe24a5a JA |
575 | unsigned int diff; |
576 | ||
577 | /* | |
578 | * we do time or byte based switch. this is needed because | |
579 | * buffered writes may issue a lot quicker than they complete, | |
580 | * whereas reads do not. | |
581 | */ | |
e47f799f | 582 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 583 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
584 | } |
585 | ||
586 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
587 | { | |
588 | unsigned int v; | |
e47f799f | 589 | |
1bd5d213 | 590 | v = rand_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 591 | |
04c540d9 | 592 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
593 | return DDIR_READ; |
594 | ||
595 | return DDIR_WRITE; | |
afe24a5a JA |
596 | } |
597 | ||
0d593542 | 598 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 599 | { |
0d593542 JA |
600 | int completed = 0; |
601 | ||
002e7183 JA |
602 | /* |
603 | * We are going to sleep, ensure that we flush anything pending as | |
604 | * not to skew our latency numbers. | |
605 | * | |
606 | * Changed to only monitor 'in flight' requests here instead of the | |
607 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
608 | * io's that have been actually submitted to an async engine, | |
609 | * and cur_depth is meaningless for sync engines. | |
610 | */ | |
a80cb54b BVA |
611 | if (td->io_u_queued || td->cur_depth) |
612 | td_io_commit(td); | |
9cc80b6d | 613 | |
002e7183 | 614 | while (td->io_u_in_flight) { |
89fd0573 | 615 | int ret; |
002e7183 | 616 | |
55312f9f | 617 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
618 | if (ret > 0) |
619 | completed += ret; | |
002e7183 | 620 | } |
0d593542 | 621 | |
6be06c46 JA |
622 | if (td->flags & TD_F_REGROW_LOGS) |
623 | regrow_logs(td); | |
624 | ||
0d593542 | 625 | return completed; |
002e7183 JA |
626 | } |
627 | ||
581e7141 JA |
628 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
629 | { | |
630 | enum fio_ddir odir = ddir ^ 1; | |
b5407f8b | 631 | uint64_t usec; |
90eff1c9 | 632 | uint64_t now; |
581e7141 | 633 | |
ff58fced | 634 | assert(ddir_rw(ddir)); |
50a8ce86 | 635 | now = utime_since_now(&td->start); |
ff58fced | 636 | |
50a8ce86 D |
637 | /* |
638 | * if rate_next_io_time is in the past, need to catch up to rate | |
639 | */ | |
640 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
641 | return ddir; |
642 | ||
643 | /* | |
50a8ce86 | 644 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
645 | * should switch. |
646 | */ | |
315fcfec | 647 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 648 | /* |
50a8ce86 | 649 | * Other direction is behind rate, switch |
581e7141 | 650 | */ |
50a8ce86 | 651 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
652 | return odir; |
653 | ||
654 | /* | |
395feabb JA |
655 | * Both directions are ahead of rate. sleep the min, |
656 | * switch if necessary | |
581e7141 | 657 | */ |
50a8ce86 | 658 | if (td->rate_next_io_time[ddir] <= |
395feabb | 659 | td->rate_next_io_time[odir]) { |
50a8ce86 | 660 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 661 | } else { |
50a8ce86 | 662 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
663 | ddir = odir; |
664 | } | |
665 | } else | |
50a8ce86 | 666 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 667 | |
a9da8ab2 JA |
668 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
669 | io_u_quiesce(td); | |
78c1eda5 | 670 | |
1a9bf814 | 671 | usec_sleep(td, usec); |
581e7141 JA |
672 | return ddir; |
673 | } | |
674 | ||
10ba535a JA |
675 | /* |
676 | * Return the data direction for the next io_u. If the job is a | |
677 | * mixed read/write workload, check the rwmix cycle and switch if | |
678 | * necessary. | |
679 | */ | |
1e97cce9 | 680 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 681 | { |
581e7141 JA |
682 | enum fio_ddir ddir; |
683 | ||
5f9099ea | 684 | /* |
f6860972 TK |
685 | * See if it's time to fsync/fdatasync/sync_file_range first, |
686 | * and if not then move on to check regular I/Os. | |
5f9099ea | 687 | */ |
f6860972 TK |
688 | if (should_fsync(td)) { |
689 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
690 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
691 | return DDIR_SYNC; | |
692 | ||
693 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
694 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
695 | return DDIR_DATASYNC; | |
696 | ||
697 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
698 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
699 | return DDIR_SYNC_FILE_RANGE; | |
700 | } | |
44f29692 | 701 | |
10ba535a | 702 | if (td_rw(td)) { |
10ba535a JA |
703 | /* |
704 | * Check if it's time to seed a new data direction. | |
705 | */ | |
e4928662 | 706 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
707 | /* |
708 | * Put a top limit on how many bytes we do for | |
709 | * one data direction, to avoid overflowing the | |
710 | * ranges too much | |
711 | */ | |
712 | ddir = get_rand_ddir(td); | |
e47f799f JA |
713 | |
714 | if (ddir != td->rwmix_ddir) | |
715 | set_rwmix_bytes(td); | |
716 | ||
717 | td->rwmix_ddir = ddir; | |
10ba535a | 718 | } |
581e7141 | 719 | ddir = td->rwmix_ddir; |
10ba535a | 720 | } else if (td_read(td)) |
581e7141 | 721 | ddir = DDIR_READ; |
6eaf09d6 | 722 | else if (td_write(td)) |
581e7141 | 723 | ddir = DDIR_WRITE; |
5c8e84ca | 724 | else if (td_trim(td)) |
6eaf09d6 | 725 | ddir = DDIR_TRIM; |
5c8e84ca TK |
726 | else |
727 | ddir = DDIR_INVAL; | |
581e7141 JA |
728 | |
729 | td->rwmix_ddir = rate_ddir(td, ddir); | |
730 | return td->rwmix_ddir; | |
10ba535a JA |
731 | } |
732 | ||
1ef2b6be JA |
733 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
734 | { | |
0e4dd95c DE |
735 | enum fio_ddir ddir = get_rw_ddir(td); |
736 | ||
82a90686 | 737 | if (td_trimwrite(td)) { |
0e4dd95c DE |
738 | struct fio_file *f = io_u->file; |
739 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
740 | ddir = DDIR_TRIM; | |
741 | else | |
742 | ddir = DDIR_WRITE; | |
743 | } | |
744 | ||
745 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 746 | |
9b87f09b | 747 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
748 | td->o.barrier_blocks && |
749 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
750 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 751 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
752 | } |
753 | ||
e8462bd8 | 754 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 755 | { |
71b84caa | 756 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
757 | |
758 | if (ret) | |
759 | td_verror(td, ret, "file close"); | |
760 | } | |
761 | ||
10ba535a JA |
762 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
763 | { | |
99952ca7 BVA |
764 | if (io_u->post_submit) { |
765 | io_u->post_submit(io_u, io_u->error == 0); | |
766 | io_u->post_submit = NULL; | |
767 | } | |
768 | ||
a9da8ab2 JA |
769 | if (td->parent) |
770 | td = td->parent; | |
771 | ||
e8462bd8 JA |
772 | td_io_u_lock(td); |
773 | ||
f8b0bd10 | 774 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 775 | put_file_log(td, io_u->file); |
f8b0bd10 | 776 | |
10ba535a | 777 | io_u->file = NULL; |
1651e431 | 778 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 779 | |
a9da8ab2 | 780 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 781 | td->cur_depth--; |
a9da8ab2 JA |
782 | assert(!(td->flags & TD_F_CHILD)); |
783 | } | |
2ae0b204 | 784 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 | 785 | td_io_u_free_notify(td); |
564de8d1 | 786 | td_io_u_unlock(td); |
10ba535a JA |
787 | } |
788 | ||
f2bba182 RR |
789 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
790 | { | |
1651e431 | 791 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
792 | put_io_u(td, io_u); |
793 | } | |
794 | ||
755200a3 JA |
795 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
796 | { | |
797 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 798 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 799 | |
465221b0 JA |
800 | dprint(FD_IO, "requeue %p\n", __io_u); |
801 | ||
a9da8ab2 JA |
802 | if (td->parent) |
803 | td = td->parent; | |
804 | ||
e8462bd8 JA |
805 | td_io_u_lock(td); |
806 | ||
1651e431 | 807 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
808 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
809 | td->io_issues[ddir]--; | |
5ec10eaa | 810 | |
1651e431 | 811 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 812 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 813 | td->cur_depth--; |
a9da8ab2 JA |
814 | assert(!(td->flags & TD_F_CHILD)); |
815 | } | |
2ae0b204 JA |
816 | |
817 | io_u_rpush(&td->io_u_requeues, __io_u); | |
a9da8ab2 | 818 | td_io_u_free_notify(td); |
564de8d1 | 819 | td_io_u_unlock(td); |
755200a3 JA |
820 | *io_u = NULL; |
821 | } | |
822 | ||
7b865a2f | 823 | static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u) |
224b3093 | 824 | { |
825 | struct fio_file *f = io_u->file; | |
826 | ||
7b865a2f BVA |
827 | assert(td->o.zone_mode == ZONE_MODE_STRIDED); |
828 | assert(td->o.zone_size); | |
829 | assert(td->o.zone_range); | |
830 | ||
224b3093 | 831 | /* |
832 | * See if it's time to switch to a new zone | |
833 | */ | |
834 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { | |
835 | td->zone_bytes = 0; | |
836 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
837 | ||
838 | /* | |
839 | * Wrap from the beginning, if we exceed the file size | |
840 | */ | |
841 | if (f->file_offset >= f->real_file_size) | |
04bc85a1 JA |
842 | f->file_offset = get_start_offset(td, f); |
843 | ||
224b3093 | 844 | f->last_pos[io_u->ddir] = f->file_offset; |
845 | td->io_skip_bytes += td->o.zone_skip; | |
846 | } | |
847 | ||
848 | /* | |
04bc85a1 JA |
849 | * If zone_size > zone_range, then maintain the same zone until |
850 | * zone_bytes >= zone_size. | |
851 | */ | |
224b3093 | 852 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
853 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
854 | f->file_offset, f->last_pos[io_u->ddir]); | |
855 | f->last_pos[io_u->ddir] = f->file_offset; | |
856 | } | |
857 | ||
858 | /* | |
859 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
860 | * map needs to be reset as it's done with zone_range everytime. | |
861 | */ | |
04bc85a1 | 862 | if ((td->zone_bytes % td->o.zone_range) == 0) |
224b3093 | 863 | fio_file_reset(td, f); |
224b3093 | 864 | } |
865 | ||
9bf2061e | 866 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 867 | { |
ec370c22 | 868 | bool is_random; |
6aca9b3d | 869 | |
9b87f09b | 870 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
871 | goto out; |
872 | ||
1ef2b6be | 873 | set_rw_ddir(td, io_u); |
5f9099ea | 874 | |
87dc1ab1 | 875 | /* |
ff58fced | 876 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 877 | */ |
ff58fced | 878 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 879 | goto out; |
a00735e6 | 880 | |
7b865a2f BVA |
881 | if (td->o.zone_mode == ZONE_MODE_STRIDED) |
882 | setup_strided_zone_mode(td, io_u); | |
48f5abd3 | 883 | |
10ba535a | 884 | /* |
c685b5b2 JA |
885 | * No log, let the seq/rand engine retrieve the next buflen and |
886 | * position. | |
10ba535a | 887 | */ |
6aca9b3d | 888 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 889 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 890 | return 1; |
2ba1c290 | 891 | } |
10ba535a | 892 | |
6aca9b3d | 893 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
894 | if (!io_u->buflen) { |
895 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 896 | return 1; |
2ba1c290 | 897 | } |
bca4ed4d | 898 | |
2ba1c290 | 899 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
5fff9543 | 900 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n", |
e5f9a813 | 901 | io_u, |
4b91ee8f JA |
902 | (unsigned long long) io_u->offset, io_u->buflen, |
903 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 904 | return 1; |
2ba1c290 | 905 | } |
6a5e6884 | 906 | |
bca4ed4d JA |
907 | /* |
908 | * mark entry before potentially trimming io_u | |
909 | */ | |
303032ae | 910 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 911 | mark_random_map(td, io_u); |
bca4ed4d | 912 | |
c38e9468 | 913 | out: |
e5f9a813 | 914 | dprint_io_u(io_u, "fill"); |
d9d91e39 | 915 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 916 | return 0; |
10ba535a JA |
917 | } |
918 | ||
6cc0e5aa | 919 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
838bc709 | 920 | { |
2b13e716 | 921 | int idx = 0; |
838bc709 JA |
922 | |
923 | switch (nr) { | |
924 | default: | |
2b13e716 | 925 | idx = 6; |
838bc709 JA |
926 | break; |
927 | case 33 ... 64: | |
2b13e716 | 928 | idx = 5; |
838bc709 JA |
929 | break; |
930 | case 17 ... 32: | |
2b13e716 | 931 | idx = 4; |
838bc709 JA |
932 | break; |
933 | case 9 ... 16: | |
2b13e716 | 934 | idx = 3; |
838bc709 JA |
935 | break; |
936 | case 5 ... 8: | |
2b13e716 | 937 | idx = 2; |
838bc709 JA |
938 | break; |
939 | case 1 ... 4: | |
2b13e716 | 940 | idx = 1; |
838bc709 JA |
941 | case 0: |
942 | break; | |
943 | } | |
944 | ||
2b13e716 | 945 | map[idx]++; |
838bc709 JA |
946 | } |
947 | ||
948 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
949 | { | |
950 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
951 | td->ts.total_submit++; | |
952 | } | |
953 | ||
954 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
955 | { | |
956 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
957 | td->ts.total_complete++; | |
958 | } | |
959 | ||
d8005759 | 960 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 961 | { |
2b13e716 | 962 | int idx = 0; |
71619dc2 JA |
963 | |
964 | switch (td->cur_depth) { | |
965 | default: | |
2b13e716 | 966 | idx = 6; |
a783e61a | 967 | break; |
71619dc2 | 968 | case 32 ... 63: |
2b13e716 | 969 | idx = 5; |
a783e61a | 970 | break; |
71619dc2 | 971 | case 16 ... 31: |
2b13e716 | 972 | idx = 4; |
a783e61a | 973 | break; |
71619dc2 | 974 | case 8 ... 15: |
2b13e716 | 975 | idx = 3; |
a783e61a | 976 | break; |
71619dc2 | 977 | case 4 ... 7: |
2b13e716 | 978 | idx = 2; |
a783e61a | 979 | break; |
71619dc2 | 980 | case 2 ... 3: |
2b13e716 | 981 | idx = 1; |
71619dc2 JA |
982 | case 1: |
983 | break; | |
984 | } | |
985 | ||
2b13e716 | 986 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
987 | } |
988 | ||
d6bb626e | 989 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 990 | { |
2b13e716 | 991 | int idx = 0; |
04a0feae | 992 | |
d6bb626e VF |
993 | assert(nsec < 1000); |
994 | ||
995 | switch (nsec) { | |
996 | case 750 ... 999: | |
997 | idx = 9; | |
998 | break; | |
999 | case 500 ... 749: | |
1000 | idx = 8; | |
1001 | break; | |
1002 | case 250 ... 499: | |
1003 | idx = 7; | |
1004 | break; | |
1005 | case 100 ... 249: | |
1006 | idx = 6; | |
1007 | break; | |
1008 | case 50 ... 99: | |
1009 | idx = 5; | |
1010 | break; | |
1011 | case 20 ... 49: | |
1012 | idx = 4; | |
1013 | break; | |
1014 | case 10 ... 19: | |
1015 | idx = 3; | |
1016 | break; | |
1017 | case 4 ... 9: | |
1018 | idx = 2; | |
1019 | break; | |
1020 | case 2 ... 3: | |
1021 | idx = 1; | |
1022 | case 0 ... 1: | |
1023 | break; | |
1024 | } | |
1025 | ||
1026 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1027 | td->ts.io_u_lat_n[idx]++; | |
1028 | } | |
1029 | ||
1030 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1031 | { | |
1032 | int idx = 0; | |
1033 | ||
1034 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1035 | |
1036 | switch (usec) { | |
1037 | case 750 ... 999: | |
2b13e716 | 1038 | idx = 9; |
04a0feae JA |
1039 | break; |
1040 | case 500 ... 749: | |
2b13e716 | 1041 | idx = 8; |
04a0feae JA |
1042 | break; |
1043 | case 250 ... 499: | |
2b13e716 | 1044 | idx = 7; |
04a0feae JA |
1045 | break; |
1046 | case 100 ... 249: | |
2b13e716 | 1047 | idx = 6; |
04a0feae JA |
1048 | break; |
1049 | case 50 ... 99: | |
2b13e716 | 1050 | idx = 5; |
04a0feae JA |
1051 | break; |
1052 | case 20 ... 49: | |
2b13e716 | 1053 | idx = 4; |
04a0feae JA |
1054 | break; |
1055 | case 10 ... 19: | |
2b13e716 | 1056 | idx = 3; |
04a0feae JA |
1057 | break; |
1058 | case 4 ... 9: | |
2b13e716 | 1059 | idx = 2; |
04a0feae JA |
1060 | break; |
1061 | case 2 ... 3: | |
2b13e716 | 1062 | idx = 1; |
04a0feae JA |
1063 | case 0 ... 1: |
1064 | break; | |
1065 | } | |
1066 | ||
2b13e716 JA |
1067 | assert(idx < FIO_IO_U_LAT_U_NR); |
1068 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1069 | } |
1070 | ||
d6bb626e | 1071 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1072 | { |
2b13e716 | 1073 | int idx = 0; |
ec118304 | 1074 | |
d6bb626e VF |
1075 | assert(msec >= 1); |
1076 | ||
ec118304 JA |
1077 | switch (msec) { |
1078 | default: | |
2b13e716 | 1079 | idx = 11; |
04a0feae | 1080 | break; |
8abdce66 | 1081 | case 1000 ... 1999: |
2b13e716 | 1082 | idx = 10; |
04a0feae | 1083 | break; |
8abdce66 | 1084 | case 750 ... 999: |
2b13e716 | 1085 | idx = 9; |
04a0feae | 1086 | break; |
8abdce66 | 1087 | case 500 ... 749: |
2b13e716 | 1088 | idx = 8; |
04a0feae | 1089 | break; |
8abdce66 | 1090 | case 250 ... 499: |
2b13e716 | 1091 | idx = 7; |
04a0feae | 1092 | break; |
8abdce66 | 1093 | case 100 ... 249: |
2b13e716 | 1094 | idx = 6; |
04a0feae | 1095 | break; |
8abdce66 | 1096 | case 50 ... 99: |
2b13e716 | 1097 | idx = 5; |
04a0feae | 1098 | break; |
8abdce66 | 1099 | case 20 ... 49: |
2b13e716 | 1100 | idx = 4; |
04a0feae | 1101 | break; |
8abdce66 | 1102 | case 10 ... 19: |
2b13e716 | 1103 | idx = 3; |
04a0feae | 1104 | break; |
8abdce66 | 1105 | case 4 ... 9: |
2b13e716 | 1106 | idx = 2; |
04a0feae | 1107 | break; |
ec118304 | 1108 | case 2 ... 3: |
2b13e716 | 1109 | idx = 1; |
ec118304 JA |
1110 | case 0 ... 1: |
1111 | break; | |
1112 | } | |
1113 | ||
2b13e716 JA |
1114 | assert(idx < FIO_IO_U_LAT_M_NR); |
1115 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1116 | } |
1117 | ||
d6bb626e | 1118 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1119 | { |
d6bb626e VF |
1120 | if (nsec < 1000) |
1121 | io_u_mark_lat_nsec(td, nsec); | |
1122 | else if (nsec < 1000000) | |
1123 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1124 | else |
d6bb626e | 1125 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1126 | } |
1127 | ||
8c07860d JA |
1128 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1129 | { | |
1130 | unsigned long fileno; | |
1131 | ||
1132 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1133 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1134 | unsigned long r; | |
1135 | ||
1136 | r = __rand(&td->next_file_state); | |
1137 | return (unsigned int) ((double) td->o.nr_files | |
1138 | * (r / (frand_max + 1.0))); | |
1139 | } | |
1140 | ||
1141 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1142 | fileno = zipf_next(&td->next_file_zipf); | |
1143 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1144 | fileno = pareto_next(&td->next_file_zipf); | |
1145 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1146 | fileno = gauss_next(&td->next_file_gauss); | |
1147 | else { | |
1148 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1149 | assert(0); | |
1150 | return 0; | |
1151 | } | |
1152 | ||
1153 | return fileno >> FIO_FSERVICE_SHIFT; | |
1154 | } | |
1155 | ||
0aabe160 JA |
1156 | /* |
1157 | * Get next file to service by choosing one at random | |
1158 | */ | |
2cc52930 JA |
1159 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1160 | enum fio_file_flags goodf, | |
d6aed795 | 1161 | enum fio_file_flags badf) |
0aabe160 | 1162 | { |
0aabe160 | 1163 | struct fio_file *f; |
1c178180 | 1164 | int fno; |
0aabe160 JA |
1165 | |
1166 | do { | |
87b10676 | 1167 | int opened = 0; |
4c07ad86 | 1168 | |
8c07860d | 1169 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1170 | |
126d65c6 | 1171 | f = td->files[fno]; |
d6aed795 | 1172 | if (fio_file_done(f)) |
059e63c0 | 1173 | continue; |
1c178180 | 1174 | |
d6aed795 | 1175 | if (!fio_file_open(f)) { |
87b10676 JA |
1176 | int err; |
1177 | ||
002fe734 JA |
1178 | if (td->nr_open_files >= td->o.open_files) |
1179 | return ERR_PTR(-EBUSY); | |
1180 | ||
87b10676 JA |
1181 | err = td_io_open_file(td, f); |
1182 | if (err) | |
1183 | continue; | |
1184 | opened = 1; | |
1185 | } | |
1186 | ||
2ba1c290 JA |
1187 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1188 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1189 | return f; |
2ba1c290 | 1190 | } |
87b10676 JA |
1191 | if (opened) |
1192 | td_io_close_file(td, f); | |
0aabe160 JA |
1193 | } while (1); |
1194 | } | |
1195 | ||
1196 | /* | |
1197 | * Get next file to service by doing round robin between all available ones | |
1198 | */ | |
1c178180 JA |
1199 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1200 | int badf) | |
3d7c391d JA |
1201 | { |
1202 | unsigned int old_next_file = td->next_file; | |
1203 | struct fio_file *f; | |
1204 | ||
1205 | do { | |
87b10676 JA |
1206 | int opened = 0; |
1207 | ||
126d65c6 | 1208 | f = td->files[td->next_file]; |
3d7c391d JA |
1209 | |
1210 | td->next_file++; | |
2dc1bbeb | 1211 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1212 | td->next_file = 0; |
1213 | ||
87b10676 | 1214 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1215 | if (fio_file_done(f)) { |
d5ed68ea | 1216 | f = NULL; |
059e63c0 | 1217 | continue; |
d5ed68ea | 1218 | } |
059e63c0 | 1219 | |
d6aed795 | 1220 | if (!fio_file_open(f)) { |
87b10676 JA |
1221 | int err; |
1222 | ||
002fe734 JA |
1223 | if (td->nr_open_files >= td->o.open_files) |
1224 | return ERR_PTR(-EBUSY); | |
1225 | ||
87b10676 | 1226 | err = td_io_open_file(td, f); |
b5696bfc JA |
1227 | if (err) { |
1228 | dprint(FD_FILE, "error %d on open of %s\n", | |
1229 | err, f->file_name); | |
87c27b45 | 1230 | f = NULL; |
87b10676 | 1231 | continue; |
b5696bfc | 1232 | } |
87b10676 JA |
1233 | opened = 1; |
1234 | } | |
1235 | ||
0b9d69ec JA |
1236 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1237 | f->flags); | |
1c178180 | 1238 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1239 | break; |
1240 | ||
87b10676 JA |
1241 | if (opened) |
1242 | td_io_close_file(td, f); | |
1243 | ||
3d7c391d JA |
1244 | f = NULL; |
1245 | } while (td->next_file != old_next_file); | |
1246 | ||
2ba1c290 | 1247 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1248 | return f; |
1249 | } | |
1250 | ||
7eb36574 | 1251 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1252 | { |
1907dbc6 JA |
1253 | struct fio_file *f; |
1254 | ||
2dc1bbeb | 1255 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1256 | |
b5696bfc | 1257 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1258 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1259 | " nr_files=%d\n", td->nr_open_files, | |
1260 | td->nr_done_files, | |
1261 | td->o.nr_files); | |
bdb4e2e9 | 1262 | return NULL; |
2ba1c290 | 1263 | } |
bdb4e2e9 | 1264 | |
1907dbc6 | 1265 | f = td->file_service_file; |
d6aed795 | 1266 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1267 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1268 | goto out; | |
1269 | if (td->file_service_left--) | |
1270 | goto out; | |
1271 | } | |
1907dbc6 | 1272 | |
a086c257 JA |
1273 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1274 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1275 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1276 | else |
d6aed795 | 1277 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1278 | |
002fe734 JA |
1279 | if (IS_ERR(f)) |
1280 | return f; | |
1281 | ||
1907dbc6 JA |
1282 | td->file_service_file = f; |
1283 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1284 | out: |
0dac421f JA |
1285 | if (f) |
1286 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1287 | else | |
1288 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1289 | return f; |
bdb4e2e9 JA |
1290 | } |
1291 | ||
7eb36574 JA |
1292 | static struct fio_file *get_next_file(struct thread_data *td) |
1293 | { | |
7eb36574 JA |
1294 | return __get_next_file(td); |
1295 | } | |
1296 | ||
002fe734 | 1297 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1298 | { |
1299 | struct fio_file *f; | |
1300 | ||
1301 | do { | |
1302 | f = get_next_file(td); | |
002fe734 JA |
1303 | if (IS_ERR_OR_NULL(f)) |
1304 | return PTR_ERR(f); | |
429f6675 | 1305 | |
429f6675 JA |
1306 | io_u->file = f; |
1307 | get_file(f); | |
1308 | ||
1309 | if (!fill_io_u(td, io_u)) | |
1310 | break; | |
1311 | ||
99952ca7 BVA |
1312 | if (io_u->post_submit) { |
1313 | io_u->post_submit(io_u, false); | |
1314 | io_u->post_submit = NULL; | |
1315 | } | |
1316 | ||
b5696bfc | 1317 | put_file_log(td, f); |
429f6675 | 1318 | td_io_close_file(td, f); |
b5696bfc | 1319 | io_u->file = NULL; |
8c07860d JA |
1320 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1321 | fio_file_reset(td, f); | |
1322 | else { | |
1323 | fio_file_set_done(f); | |
1324 | td->nr_done_files++; | |
1325 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1326 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1327 | } |
429f6675 JA |
1328 | } while (1); |
1329 | ||
1330 | return 0; | |
1331 | } | |
1332 | ||
3e260a46 | 1333 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
c3a32714 | 1334 | unsigned long long tnsec, unsigned long long max_nsec) |
3e260a46 JA |
1335 | { |
1336 | if (!td->error) | |
c3a32714 | 1337 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
3e260a46 JA |
1338 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
1339 | icd->error = ETIMEDOUT; | |
1340 | } | |
1341 | ||
1342 | static void lat_new_cycle(struct thread_data *td) | |
1343 | { | |
1344 | fio_gettime(&td->latency_ts, NULL); | |
1345 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1346 | td->latency_failed = 0; | |
1347 | } | |
1348 | ||
1349 | /* | |
1350 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1351 | * are at QD=1, then it's time to give up. | |
1352 | */ | |
e39c0676 | 1353 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1354 | { |
1355 | if (td->latency_qd == 1) | |
e39c0676 | 1356 | return true; |
3e260a46 JA |
1357 | |
1358 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1359 | |
1360 | if (td->latency_qd == td->latency_qd_low) | |
1361 | td->latency_qd_low--; | |
1362 | ||
3e260a46 JA |
1363 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1364 | ||
1365 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1366 | ||
1367 | /* | |
1368 | * When we ramp QD down, quiesce existing IO to prevent | |
1369 | * a storm of ramp downs due to pending higher depth. | |
1370 | */ | |
1371 | io_u_quiesce(td); | |
1372 | lat_new_cycle(td); | |
e39c0676 | 1373 | return false; |
3e260a46 JA |
1374 | } |
1375 | ||
e39c0676 | 1376 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1377 | { |
1378 | if (td->o.latency_percentile.u.f == 100.0) | |
1379 | return __lat_target_failed(td); | |
1380 | ||
1381 | td->latency_failed++; | |
e39c0676 | 1382 | return false; |
3e260a46 JA |
1383 | } |
1384 | ||
1385 | void lat_target_init(struct thread_data *td) | |
1386 | { | |
6bb58215 JA |
1387 | td->latency_end_run = 0; |
1388 | ||
3e260a46 JA |
1389 | if (td->o.latency_target) { |
1390 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1391 | fio_gettime(&td->latency_ts, NULL); | |
1392 | td->latency_qd = 1; | |
1393 | td->latency_qd_high = td->o.iodepth; | |
1394 | td->latency_qd_low = 1; | |
1395 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1396 | } else | |
1397 | td->latency_qd = td->o.iodepth; | |
1398 | } | |
1399 | ||
6bb58215 JA |
1400 | void lat_target_reset(struct thread_data *td) |
1401 | { | |
1402 | if (!td->latency_end_run) | |
1403 | lat_target_init(td); | |
1404 | } | |
1405 | ||
3e260a46 JA |
1406 | static void lat_target_success(struct thread_data *td) |
1407 | { | |
1408 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1409 | struct thread_options *o = &td->o; |
3e260a46 JA |
1410 | |
1411 | td->latency_qd_low = td->latency_qd; | |
1412 | ||
1413 | /* | |
1414 | * If we haven't failed yet, we double up to a failing value instead | |
1415 | * of bisecting from highest possible queue depth. If we have set | |
1416 | * a limit other than td->o.iodepth, bisect between that. | |
1417 | */ | |
6bb58215 | 1418 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1419 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1420 | else | |
1421 | td->latency_qd *= 2; | |
1422 | ||
6bb58215 JA |
1423 | if (td->latency_qd > o->iodepth) |
1424 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1425 | |
1426 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1427 | |
3e260a46 | 1428 | /* |
6bb58215 JA |
1429 | * Same as last one, we are done. Let it run a latency cycle, so |
1430 | * we get only the results from the targeted depth. | |
3e260a46 | 1431 | */ |
6bb58215 JA |
1432 | if (td->latency_qd == qd) { |
1433 | if (td->latency_end_run) { | |
1434 | dprint(FD_RATE, "We are done\n"); | |
1435 | td->done = 1; | |
1436 | } else { | |
1437 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1438 | io_u_quiesce(td); | |
1439 | td->latency_end_run = 1; | |
1440 | reset_all_stats(td); | |
1441 | reset_io_stats(td); | |
1442 | } | |
1443 | } | |
3e260a46 JA |
1444 | |
1445 | lat_new_cycle(td); | |
1446 | } | |
1447 | ||
1448 | /* | |
1449 | * Check if we can bump the queue depth | |
1450 | */ | |
1451 | void lat_target_check(struct thread_data *td) | |
1452 | { | |
1453 | uint64_t usec_window; | |
1454 | uint64_t ios; | |
1455 | double success_ios; | |
1456 | ||
1457 | usec_window = utime_since_now(&td->latency_ts); | |
1458 | if (usec_window < td->o.latency_window) | |
1459 | return; | |
1460 | ||
1461 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1462 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1463 | success_ios *= 100.0; | |
1464 | ||
1465 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1466 | ||
1467 | if (success_ios >= td->o.latency_percentile.u.f) | |
1468 | lat_target_success(td); | |
1469 | else | |
1470 | __lat_target_failed(td); | |
1471 | } | |
1472 | ||
1473 | /* | |
1474 | * If latency target is enabled, we might be ramping up or down and not | |
1475 | * using the full queue depth available. | |
1476 | */ | |
e39c0676 | 1477 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1478 | { |
1479 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1480 | ||
1481 | if (qempty) | |
e39c0676 | 1482 | return true; |
3e260a46 | 1483 | if (!td->o.latency_target) |
e39c0676 | 1484 | return false; |
3e260a46 JA |
1485 | |
1486 | return td->cur_depth >= td->latency_qd; | |
1487 | } | |
429f6675 | 1488 | |
10ba535a JA |
1489 | struct io_u *__get_io_u(struct thread_data *td) |
1490 | { | |
0cae66f6 | 1491 | struct io_u *io_u = NULL; |
93b45bb2 | 1492 | int ret; |
10ba535a | 1493 | |
ca09be4b JA |
1494 | if (td->stop_io) |
1495 | return NULL; | |
1496 | ||
e8462bd8 JA |
1497 | td_io_u_lock(td); |
1498 | ||
1499 | again: | |
2ae0b204 JA |
1500 | if (!io_u_rempty(&td->io_u_requeues)) |
1501 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1502 | else if (!queue_full(td)) { |
2ae0b204 | 1503 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1504 | |
225ba9e3 | 1505 | io_u->file = NULL; |
6040dabc | 1506 | io_u->buflen = 0; |
10ba535a | 1507 | io_u->resid = 0; |
d7762cf8 | 1508 | io_u->end_io = NULL; |
755200a3 JA |
1509 | } |
1510 | ||
1511 | if (io_u) { | |
0c6e7517 | 1512 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1513 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1514 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1515 | IO_U_F_VER_LIST); | |
0c6e7517 | 1516 | |
755200a3 | 1517 | io_u->error = 0; |
bcd5abfa | 1518 | io_u->acct_ddir = -1; |
10ba535a | 1519 | td->cur_depth++; |
a9da8ab2 | 1520 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1521 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1522 | io_u->ipo = NULL; |
a9da8ab2 | 1523 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1524 | /* |
1525 | * We ran out, wait for async verify threads to finish and | |
1526 | * return one | |
1527 | */ | |
a9da8ab2 | 1528 | assert(!(td->flags & TD_F_CHILD)); |
93b45bb2 BVA |
1529 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
1530 | assert(ret == 0); | |
1dec3e07 | 1531 | goto again; |
10ba535a JA |
1532 | } |
1533 | ||
e8462bd8 | 1534 | td_io_u_unlock(td); |
10ba535a JA |
1535 | return io_u; |
1536 | } | |
1537 | ||
e39c0676 | 1538 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1539 | { |
d72be545 | 1540 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1541 | return false; |
c9a73054 JA |
1542 | if (!td->trim_entries) |
1543 | return false; | |
d72be545 | 1544 | |
c9a73054 JA |
1545 | if (td->trim_batch) { |
1546 | td->trim_batch--; | |
1547 | if (get_next_trim(td, io_u)) | |
1548 | return true; | |
1549 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1550 | td->last_ddir != DDIR_READ) { | |
1551 | td->trim_batch = td->o.trim_batch; | |
1552 | if (!td->trim_batch) | |
1553 | td->trim_batch = td->o.trim_backlog; | |
1554 | if (get_next_trim(td, io_u)) | |
e39c0676 | 1555 | return true; |
2ba1c290 | 1556 | } |
10ba535a | 1557 | |
e39c0676 | 1558 | return false; |
0d29de83 JA |
1559 | } |
1560 | ||
e39c0676 | 1561 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1562 | { |
d72be545 | 1563 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1564 | return false; |
d72be545 JA |
1565 | |
1566 | if (td->io_hist_len) { | |
9e144189 JA |
1567 | int get_verify = 0; |
1568 | ||
d1ece0c7 | 1569 | if (td->verify_batch) |
9e144189 | 1570 | get_verify = 1; |
d1ece0c7 | 1571 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1572 | td->last_ddir != DDIR_READ) { |
1573 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1574 | if (!td->verify_batch) |
1575 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1576 | get_verify = 1; |
1577 | } | |
1578 | ||
d1ece0c7 JA |
1579 | if (get_verify && !get_next_verify(td, io_u)) { |
1580 | td->verify_batch--; | |
e39c0676 | 1581 | return true; |
d1ece0c7 | 1582 | } |
9e144189 JA |
1583 | } |
1584 | ||
e39c0676 | 1585 | return false; |
0d29de83 JA |
1586 | } |
1587 | ||
de789769 JA |
1588 | /* |
1589 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1590 | * easy compressible data for simple de-dupe attempts. Do this for every |
1591 | * 512b block in the range, since that should be the smallest block size | |
1592 | * we can expect from a device. | |
de789769 JA |
1593 | */ |
1594 | static void small_content_scramble(struct io_u *io_u) | |
1595 | { | |
5fff9543 | 1596 | unsigned long long i, nr_blocks = io_u->buflen >> 9; |
23f394d5 | 1597 | unsigned int offset; |
319b073f JA |
1598 | uint64_t boffset, *iptr; |
1599 | char *p; | |
de789769 | 1600 | |
23f394d5 JA |
1601 | if (!nr_blocks) |
1602 | return; | |
1603 | ||
1604 | p = io_u->xfer_buf; | |
fba76ee8 | 1605 | boffset = io_u->offset; |
319b073f JA |
1606 | |
1607 | if (io_u->buf_filled_len) | |
1608 | io_u->buf_filled_len = 0; | |
1609 | ||
1610 | /* | |
1611 | * Generate random index between 0..7. We do chunks of 512b, if | |
1612 | * we assume a cacheline is 64 bytes, then we have 8 of those. | |
1613 | * Scramble content within the blocks in the same cacheline to | |
1614 | * speed things up. | |
1615 | */ | |
1616 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; | |
fad82f76 | 1617 | |
23f394d5 JA |
1618 | for (i = 0; i < nr_blocks; i++) { |
1619 | /* | |
319b073f JA |
1620 | * Fill offset into start of cacheline, time into end |
1621 | * of cacheline | |
23f394d5 | 1622 | */ |
319b073f JA |
1623 | iptr = (void *) p + (offset << 6); |
1624 | *iptr = boffset; | |
1625 | ||
1626 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); | |
1627 | iptr[0] = io_u->start_time.tv_sec; | |
1628 | iptr[1] = io_u->start_time.tv_nsec; | |
23f394d5 | 1629 | |
23f394d5 | 1630 | p += 512; |
fad82f76 | 1631 | boffset += 512; |
23f394d5 | 1632 | } |
de789769 JA |
1633 | } |
1634 | ||
0d29de83 JA |
1635 | /* |
1636 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
5c5c33c1 | 1637 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
0d29de83 JA |
1638 | */ |
1639 | struct io_u *get_io_u(struct thread_data *td) | |
1640 | { | |
1641 | struct fio_file *f; | |
1642 | struct io_u *io_u; | |
de789769 | 1643 | int do_scramble = 0; |
002fe734 | 1644 | long ret = 0; |
0d29de83 JA |
1645 | |
1646 | io_u = __get_io_u(td); | |
1647 | if (!io_u) { | |
1648 | dprint(FD_IO, "__get_io_u failed\n"); | |
1649 | return NULL; | |
1650 | } | |
1651 | ||
1652 | if (check_get_verify(td, io_u)) | |
1653 | goto out; | |
1654 | if (check_get_trim(td, io_u)) | |
1655 | goto out; | |
1656 | ||
755200a3 JA |
1657 | /* |
1658 | * from a requeue, io_u already setup | |
1659 | */ | |
1660 | if (io_u->file) | |
77f392bf | 1661 | goto out; |
755200a3 | 1662 | |
429f6675 JA |
1663 | /* |
1664 | * If using an iolog, grab next piece if any available. | |
1665 | */ | |
d72be545 | 1666 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1667 | if (read_iolog_get(td, io_u)) |
1668 | goto err_put; | |
2ba1c290 | 1669 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1670 | ret = -EBUSY; |
2ba1c290 | 1671 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1672 | goto err_put; |
2ba1c290 | 1673 | } |
5ec10eaa | 1674 | |
429f6675 | 1675 | f = io_u->file; |
002fe734 JA |
1676 | if (!f) { |
1677 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1678 | goto err_put; | |
1679 | } | |
1680 | ||
d6aed795 | 1681 | assert(fio_file_open(f)); |
97af62ce | 1682 | |
ff58fced | 1683 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1684 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1685 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1686 | goto err_put; |
2ba1c290 | 1687 | } |
10ba535a | 1688 | |
f1dfb668 JA |
1689 | f->last_start[io_u->ddir] = io_u->offset; |
1690 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1691 | |
fd68418e | 1692 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1693 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1694 | io_u_fill_buffer(td, io_u, |
1066358a | 1695 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1696 | io_u->buflen); |
ff441ae8 | 1697 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
5c5c33c1 BVA |
1698 | !(td->flags & TD_F_COMPRESS) && |
1699 | !(td->flags & TD_F_DO_VERIFY)) | |
fd68418e JA |
1700 | do_scramble = 1; |
1701 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1702 | /* |
1703 | * Reset the buf_filled parameters so next time if the | |
1704 | * buffer is used for writes it is refilled. | |
1705 | */ | |
cbe8d756 RR |
1706 | io_u->buf_filled_len = 0; |
1707 | } | |
87dc1ab1 | 1708 | } |
10ba535a | 1709 | |
165faf16 JA |
1710 | /* |
1711 | * Set io data pointers. | |
1712 | */ | |
cec6b55d JA |
1713 | io_u->xfer_buf = io_u->buf; |
1714 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1715 | |
6ac7a331 | 1716 | out: |
0d29de83 | 1717 | assert(io_u->file); |
429f6675 | 1718 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1719 | if (!td->o.disable_lat) |
993bf48b | 1720 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1721 | |
de789769 JA |
1722 | if (do_scramble) |
1723 | small_content_scramble(io_u); | |
03553853 | 1724 | |
429f6675 | 1725 | return io_u; |
36167d82 | 1726 | } |
429f6675 | 1727 | err_put: |
2ba1c290 | 1728 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1729 | put_io_u(td, io_u); |
002fe734 | 1730 | return ERR_PTR(ret); |
10ba535a JA |
1731 | } |
1732 | ||
a9da8ab2 | 1733 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1734 | { |
8b28bd41 | 1735 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1736 | |
8b28bd41 DM |
1737 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1738 | return; | |
5451792e | 1739 | |
5fff9543 | 1740 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n", |
709c8313 RE |
1741 | io_u->file ? " on file " : "", |
1742 | io_u->file ? io_u->file->file_name : "", | |
1743 | strerror(io_u->error), | |
1744 | io_ddir_name(io_u->ddir), | |
1745 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1746 | |
5ad7be56 KD |
1747 | if (td->io_ops->errdetails) { |
1748 | char *err = td->io_ops->errdetails(io_u); | |
1749 | ||
1750 | log_err("fio: %s\n", err); | |
1751 | free(err); | |
1752 | } | |
1753 | ||
5451792e JA |
1754 | if (!td->error) |
1755 | td_verror(td, io_u->error, "io_u error"); | |
1756 | } | |
1757 | ||
a9da8ab2 JA |
1758 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1759 | { | |
1760 | __io_u_log_error(td, io_u); | |
1761 | if (td->parent) | |
094e66cb | 1762 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1763 | } |
1764 | ||
e39c0676 | 1765 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1766 | { |
3ecc8c67 JA |
1767 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1768 | || td->o.gtod_reduce; | |
aba6c951 JA |
1769 | } |
1770 | ||
c8eeb9df JA |
1771 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1772 | struct io_completion_data *icd, | |
1773 | const enum fio_ddir idx, unsigned int bytes) | |
1774 | { | |
a9da8ab2 | 1775 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1776 | unsigned long long llnsec = 0; |
c8eeb9df | 1777 | |
75dc383e JA |
1778 | if (td->parent) |
1779 | td = td->parent; | |
1780 | ||
132b1ee4 | 1781 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1782 | return; |
1783 | ||
a9da8ab2 | 1784 | if (no_reduce) |
d6bb626e | 1785 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1786 | |
1787 | if (!td->o.disable_lat) { | |
c3a32714 | 1788 | unsigned long long tnsec; |
c8eeb9df | 1789 | |
d6bb626e VF |
1790 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
1791 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); | |
15501535 | 1792 | |
d4afedfd JA |
1793 | if (td->flags & TD_F_PROFILE_OPS) { |
1794 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1795 | ||
1796 | if (ops->io_u_lat) | |
c3a32714 | 1797 | icd->error = ops->io_u_lat(td, tnsec); |
d4afedfd JA |
1798 | } |
1799 | ||
c3a32714 JA |
1800 | if (td->o.max_latency && tnsec > td->o.max_latency) |
1801 | lat_fatal(td, icd, tnsec, td->o.max_latency); | |
1802 | if (td->o.latency_target && tnsec > td->o.latency_target) { | |
3e260a46 | 1803 | if (lat_target_failed(td)) |
c3a32714 | 1804 | lat_fatal(td, icd, tnsec, td->o.latency_target); |
15501535 | 1805 | } |
c8eeb9df JA |
1806 | } |
1807 | ||
a47591e4 JA |
1808 | if (ddir_rw(idx)) { |
1809 | if (!td->o.disable_clat) { | |
d6bb626e VF |
1810 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
1811 | io_u_mark_latency(td, llnsec); | |
a47591e4 | 1812 | } |
c8eeb9df | 1813 | |
a47591e4 | 1814 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1815 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1816 | |
a47591e4 JA |
1817 | if (no_reduce && per_unit_log(td->iops_log)) |
1818 | add_iops_sample(td, io_u, bytes); | |
b2b3eefe JA |
1819 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
1820 | add_sync_clat_sample(&td->ts, llnsec); | |
66347cfa DE |
1821 | |
1822 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1823 | uint32_t *info = io_u_block_info(td, io_u); | |
1824 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1825 | if (io_u->ddir == DDIR_TRIM) { | |
1826 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1827 | BLOCK_INFO_TRIMS(*info) + 1); | |
1828 | } else if (io_u->ddir == DDIR_WRITE) { | |
1829 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1830 | *info); | |
1831 | } | |
1832 | } | |
1833 | } | |
c8eeb9df JA |
1834 | } |
1835 | ||
94a6e1bb JA |
1836 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1837 | uint64_t offset, unsigned int bytes) | |
1838 | { | |
1839 | int idx; | |
1840 | ||
639ad1ea JA |
1841 | if (!f) |
1842 | return; | |
1843 | ||
94a6e1bb JA |
1844 | if (f->first_write == -1ULL || offset < f->first_write) |
1845 | f->first_write = offset; | |
1846 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1847 | f->last_write = offset + bytes; | |
1848 | ||
1849 | if (!f->last_write_comp) | |
1850 | return; | |
1851 | ||
1852 | idx = f->last_write_idx++; | |
1853 | f->last_write_comp[idx] = offset; | |
1854 | if (f->last_write_idx == td->o.iodepth) | |
1855 | f->last_write_idx = 0; | |
1856 | } | |
1857 | ||
b2b3eefe JA |
1858 | static bool should_account(struct thread_data *td) |
1859 | { | |
1860 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || | |
1861 | td->runstate == TD_VERIFYING); | |
1862 | } | |
1863 | ||
f8b0bd10 | 1864 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1865 | struct io_completion_data *icd) |
10ba535a | 1866 | { |
f8b0bd10 JA |
1867 | struct io_u *io_u = *io_u_ptr; |
1868 | enum fio_ddir ddir = io_u->ddir; | |
1869 | struct fio_file *f = io_u->file; | |
10ba535a | 1870 | |
e5f9a813 | 1871 | dprint_io_u(io_u, "complete"); |
2ba1c290 | 1872 | |
0c6e7517 | 1873 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1874 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1875 | |
1876 | /* | |
1877 | * Mark IO ok to verify | |
1878 | */ | |
1879 | if (io_u->ipo) { | |
890b6656 JA |
1880 | /* |
1881 | * Remove errored entry from the verification list | |
1882 | */ | |
1883 | if (io_u->error) | |
1884 | unlog_io_piece(td, io_u); | |
1885 | else { | |
1886 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1887 | write_barrier(); | |
1888 | } | |
f9401285 JA |
1889 | } |
1890 | ||
f8b0bd10 | 1891 | if (ddir_sync(ddir)) { |
2ea93f98 | 1892 | td->last_was_sync = true; |
44f29692 JA |
1893 | if (f) { |
1894 | f->first_write = -1ULL; | |
1895 | f->last_write = -1ULL; | |
1896 | } | |
b2b3eefe JA |
1897 | if (should_account(td)) |
1898 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); | |
87dc1ab1 JA |
1899 | return; |
1900 | } | |
1901 | ||
2ea93f98 | 1902 | td->last_was_sync = false; |
f8b0bd10 | 1903 | td->last_ddir = ddir; |
87dc1ab1 | 1904 | |
f8b0bd10 | 1905 | if (!io_u->error && ddir_rw(ddir)) { |
5fff9543 | 1906 | unsigned long long bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1907 | int ret; |
10ba535a | 1908 | |
f8b0bd10 | 1909 | td->io_blocks[ddir]++; |
f8b0bd10 | 1910 | td->io_bytes[ddir] += bytes; |
ae2fafc8 | 1911 | |
e1c325d2 JA |
1912 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
1913 | td->this_io_blocks[ddir]++; | |
f8b0bd10 | 1914 | td->this_io_bytes[ddir] += bytes; |
e1c325d2 | 1915 | } |
f8b0bd10 | 1916 | |
639ad1ea | 1917 | if (ddir == DDIR_WRITE) |
94a6e1bb | 1918 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 1919 | |
b2b3eefe | 1920 | if (should_account(td)) |
f8b0bd10 | 1921 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1922 | |
f8b0bd10 | 1923 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1924 | |
d7762cf8 | 1925 | if (io_u->end_io) { |
f8b0bd10 JA |
1926 | ret = io_u->end_io(td, io_u_ptr); |
1927 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1928 | if (ret && !icd->error) |
1929 | icd->error = ret; | |
1930 | } | |
ff58fced | 1931 | } else if (io_u->error) { |
10ba535a | 1932 | icd->error = io_u->error; |
5451792e JA |
1933 | io_u_log_error(td, io_u); |
1934 | } | |
8b28bd41 | 1935 | if (icd->error) { |
f8b0bd10 JA |
1936 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1937 | ||
8b28bd41 DM |
1938 | if (!td_non_fatal_error(td, eb, icd->error)) |
1939 | return; | |
f8b0bd10 | 1940 | |
f2bba182 RR |
1941 | /* |
1942 | * If there is a non_fatal error, then add to the error count | |
1943 | * and clear all the errors. | |
1944 | */ | |
1945 | update_error_count(td, icd->error); | |
1946 | td_clear_error(td); | |
1947 | icd->error = 0; | |
f8b0bd10 JA |
1948 | if (io_u) |
1949 | io_u->error = 0; | |
f2bba182 | 1950 | } |
10ba535a JA |
1951 | } |
1952 | ||
9520ebb9 JA |
1953 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1954 | int nr) | |
10ba535a | 1955 | { |
6eaf09d6 | 1956 | int ddir; |
aba6c951 JA |
1957 | |
1958 | if (!gtod_reduce(td)) | |
9520ebb9 | 1959 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1960 | |
3af6ef39 JA |
1961 | icd->nr = nr; |
1962 | ||
10ba535a | 1963 | icd->error = 0; |
c1f50f76 | 1964 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 1965 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
1966 | } |
1967 | ||
97601024 JA |
1968 | static void ios_completed(struct thread_data *td, |
1969 | struct io_completion_data *icd) | |
36167d82 JA |
1970 | { |
1971 | struct io_u *io_u; | |
1972 | int i; | |
1973 | ||
10ba535a JA |
1974 | for (i = 0; i < icd->nr; i++) { |
1975 | io_u = td->io_ops->event(td, i); | |
1976 | ||
f8b0bd10 | 1977 | io_completed(td, &io_u, icd); |
e8462bd8 | 1978 | |
f8b0bd10 | 1979 | if (io_u) |
e8462bd8 | 1980 | put_io_u(td, io_u); |
10ba535a JA |
1981 | } |
1982 | } | |
97601024 | 1983 | |
e7e6cfb4 JA |
1984 | /* |
1985 | * Complete a single io_u for the sync engines. | |
1986 | */ | |
55312f9f | 1987 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1988 | { |
1989 | struct io_completion_data icd; | |
55312f9f | 1990 | int ddir; |
97601024 | 1991 | |
9520ebb9 | 1992 | init_icd(td, &icd, 1); |
f8b0bd10 | 1993 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1994 | |
f8b0bd10 | 1995 | if (io_u) |
e8462bd8 | 1996 | put_io_u(td, io_u); |
97601024 | 1997 | |
581e7141 JA |
1998 | if (icd.error) { |
1999 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2000 | return -1; | |
2001 | } | |
97601024 | 2002 | |
c1f50f76 | 2003 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2004 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
2005 | |
2006 | return 0; | |
97601024 JA |
2007 | } |
2008 | ||
e7e6cfb4 JA |
2009 | /* |
2010 | * Called to complete min_events number of io for the async engines. | |
2011 | */ | |
55312f9f | 2012 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2013 | { |
97601024 | 2014 | struct io_completion_data icd; |
00de55ef | 2015 | struct timespec *tvp = NULL; |
55312f9f | 2016 | int ret, ddir; |
4d06a338 | 2017 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2018 | |
12bb8569 | 2019 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2020 | |
4950421a | 2021 | if (!min_evts) |
00de55ef | 2022 | tvp = &ts; |
5fb4b366 RE |
2023 | else if (min_evts > td->cur_depth) |
2024 | min_evts = td->cur_depth; | |
97601024 | 2025 | |
82407585 RP |
2026 | /* No worries, td_io_getevents fixes min and max if they are |
2027 | * set incorrectly */ | |
2028 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2029 | if (ret < 0) { |
e1161c32 | 2030 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2031 | return ret; |
2032 | } else if (!ret) | |
2033 | return ret; | |
2034 | ||
9520ebb9 | 2035 | init_icd(td, &icd, ret); |
97601024 | 2036 | ios_completed(td, &icd); |
581e7141 JA |
2037 | if (icd.error) { |
2038 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2039 | return -1; | |
2040 | } | |
97601024 | 2041 | |
c1f50f76 | 2042 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2043 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2044 | |
0d593542 | 2045 | return ret; |
97601024 | 2046 | } |
7e77dd02 JA |
2047 | |
2048 | /* | |
2049 | * Call when io_u is really queued, to update the submission latency. | |
2050 | */ | |
2051 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2052 | { | |
8243be59 | 2053 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2054 | unsigned long slat_time; |
7e77dd02 | 2055 | |
d6bb626e | 2056 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2057 | |
2058 | if (td->parent) | |
2059 | td = td->parent; | |
2060 | ||
ae588852 JA |
2061 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
2062 | io_u->offset); | |
9520ebb9 | 2063 | } |
7e77dd02 | 2064 | } |
433afcb4 | 2065 | |
5c94b008 JA |
2066 | /* |
2067 | * See if we should reuse the last seed, if dedupe is enabled | |
2068 | */ | |
9451b93e | 2069 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2070 | { |
2071 | unsigned int v; | |
5c94b008 JA |
2072 | |
2073 | if (!td->o.dedupe_percentage) | |
2074 | return &td->buf_state; | |
732eedd0 | 2075 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2076 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2077 | return &td->buf_state; | |
732eedd0 | 2078 | } |
5c94b008 | 2079 | |
1bd5d213 | 2080 | v = rand_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2081 | |
2082 | if (v <= td->o.dedupe_percentage) | |
2083 | return &td->buf_state_prev; | |
2084 | ||
2085 | return &td->buf_state; | |
2086 | } | |
2087 | ||
9451b93e | 2088 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2089 | { |
9451b93e JA |
2090 | if (td->o.dedupe_percentage == 100) |
2091 | frand_copy(rs, &td->buf_state_prev); | |
2092 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2093 | frand_copy(&td->buf_state_prev, rs); |
2094 | } | |
2095 | ||
5fff9543 JF |
2096 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write, |
2097 | unsigned long long max_bs) | |
5973cafb | 2098 | { |
d1af2894 JA |
2099 | struct thread_options *o = &td->o; |
2100 | ||
15600335 JA |
2101 | if (o->mem_type == MEM_CUDA_MALLOC) |
2102 | return; | |
03553853 | 2103 | |
4eff3e57 | 2104 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2105 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2106 | struct frand_state *rs; |
5fff9543 JF |
2107 | unsigned long long left = max_bs; |
2108 | unsigned long long this_write; | |
5c94b008 | 2109 | |
1066358a | 2110 | do { |
9451b93e | 2111 | rs = get_buf_state(td); |
9c42684e | 2112 | |
1066358a | 2113 | min_write = min(min_write, left); |
f97a43a1 | 2114 | |
1066358a | 2115 | if (perc) { |
4b157ac6 | 2116 | this_write = min_not_zero(min_write, |
5fff9543 | 2117 | (unsigned long long) td->o.compress_chunk); |
1e7f82e2 JA |
2118 | |
2119 | fill_random_buf_percentage(rs, buf, perc, | |
2120 | this_write, this_write, | |
2121 | o->buffer_pattern, | |
2122 | o->buffer_pattern_bytes); | |
2123 | } else { | |
1066358a | 2124 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2125 | this_write = min_write; |
2126 | } | |
1066358a | 2127 | |
1e7f82e2 JA |
2128 | buf += this_write; |
2129 | left -= this_write; | |
9451b93e | 2130 | save_buf_state(td, rs); |
1066358a | 2131 | } while (left); |
d1af2894 JA |
2132 | } else if (o->buffer_pattern_bytes) |
2133 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2134 | else if (o->zero_buffers) |
cc86c395 | 2135 | memset(buf, 0, max_bs); |
999d245e | 2136 | else |
9451b93e | 2137 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2138 | } |
2139 | ||
2140 | /* | |
2141 | * "randomly" fill the buffer contents | |
2142 | */ | |
2143 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
5fff9543 | 2144 | unsigned long long min_write, unsigned long long max_bs) |
cc86c395 JA |
2145 | { |
2146 | io_u->buf_filled_len = 0; | |
2147 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2148 | } |
e2c75fc4 TK |
2149 | |
2150 | static int do_sync_file_range(const struct thread_data *td, | |
2151 | struct fio_file *f) | |
2152 | { | |
2153 | off64_t offset, nbytes; | |
2154 | ||
2155 | offset = f->first_write; | |
2156 | nbytes = f->last_write - f->first_write; | |
2157 | ||
2158 | if (!nbytes) | |
2159 | return 0; | |
2160 | ||
2161 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2162 | } | |
2163 | ||
2164 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2165 | { | |
2166 | int ret; | |
2167 | ||
2168 | if (io_u->ddir == DDIR_SYNC) { | |
2169 | ret = fsync(io_u->file->fd); | |
2170 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2171 | #ifdef CONFIG_FDATASYNC | |
2172 | ret = fdatasync(io_u->file->fd); | |
2173 | #else | |
2174 | ret = io_u->xfer_buflen; | |
2175 | io_u->error = EINVAL; | |
2176 | #endif | |
2177 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2178 | ret = do_sync_file_range(td, io_u->file); | |
2179 | else { | |
2180 | ret = io_u->xfer_buflen; | |
2181 | io_u->error = EINVAL; | |
2182 | } | |
2183 | ||
2184 | if (ret < 0) | |
2185 | io_u->error = errno; | |
2186 | ||
2187 | return ret; | |
2188 | } | |
2189 | ||
2190 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2191 | { | |
2192 | #ifndef FIO_HAVE_TRIM | |
2193 | io_u->error = EINVAL; | |
2194 | return 0; | |
2195 | #else | |
2196 | struct fio_file *f = io_u->file; | |
2197 | int ret; | |
2198 | ||
496b1f9e | 2199 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2200 | if (!ret) |
2201 | return io_u->xfer_buflen; | |
2202 | ||
2203 | io_u->error = ret; | |
2204 | return 0; | |
2205 | #endif | |
2206 | } |