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 | { | |
a9da8ab2 JA |
764 | if (td->parent) |
765 | td = td->parent; | |
766 | ||
e8462bd8 JA |
767 | td_io_u_lock(td); |
768 | ||
f8b0bd10 | 769 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 770 | put_file_log(td, io_u->file); |
f8b0bd10 | 771 | |
10ba535a | 772 | io_u->file = NULL; |
1651e431 | 773 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 774 | |
a9da8ab2 | 775 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 776 | td->cur_depth--; |
a9da8ab2 JA |
777 | assert(!(td->flags & TD_F_CHILD)); |
778 | } | |
2ae0b204 | 779 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 | 780 | td_io_u_free_notify(td); |
564de8d1 | 781 | td_io_u_unlock(td); |
10ba535a JA |
782 | } |
783 | ||
f2bba182 RR |
784 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
785 | { | |
1651e431 | 786 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
787 | put_io_u(td, io_u); |
788 | } | |
789 | ||
755200a3 JA |
790 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
791 | { | |
792 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 793 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 794 | |
465221b0 JA |
795 | dprint(FD_IO, "requeue %p\n", __io_u); |
796 | ||
a9da8ab2 JA |
797 | if (td->parent) |
798 | td = td->parent; | |
799 | ||
e8462bd8 JA |
800 | td_io_u_lock(td); |
801 | ||
1651e431 | 802 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
803 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
804 | td->io_issues[ddir]--; | |
5ec10eaa | 805 | |
1651e431 | 806 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 807 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 808 | td->cur_depth--; |
a9da8ab2 JA |
809 | assert(!(td->flags & TD_F_CHILD)); |
810 | } | |
2ae0b204 JA |
811 | |
812 | io_u_rpush(&td->io_u_requeues, __io_u); | |
a9da8ab2 | 813 | td_io_u_free_notify(td); |
564de8d1 | 814 | td_io_u_unlock(td); |
755200a3 JA |
815 | *io_u = NULL; |
816 | } | |
817 | ||
7b865a2f | 818 | static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u) |
224b3093 | 819 | { |
820 | struct fio_file *f = io_u->file; | |
821 | ||
7b865a2f BVA |
822 | assert(td->o.zone_mode == ZONE_MODE_STRIDED); |
823 | assert(td->o.zone_size); | |
824 | assert(td->o.zone_range); | |
825 | ||
224b3093 | 826 | /* |
827 | * See if it's time to switch to a new zone | |
828 | */ | |
829 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { | |
830 | td->zone_bytes = 0; | |
831 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
832 | ||
833 | /* | |
834 | * Wrap from the beginning, if we exceed the file size | |
835 | */ | |
836 | if (f->file_offset >= f->real_file_size) | |
04bc85a1 JA |
837 | f->file_offset = get_start_offset(td, f); |
838 | ||
224b3093 | 839 | f->last_pos[io_u->ddir] = f->file_offset; |
840 | td->io_skip_bytes += td->o.zone_skip; | |
841 | } | |
842 | ||
843 | /* | |
04bc85a1 JA |
844 | * If zone_size > zone_range, then maintain the same zone until |
845 | * zone_bytes >= zone_size. | |
846 | */ | |
224b3093 | 847 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
848 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
849 | f->file_offset, f->last_pos[io_u->ddir]); | |
850 | f->last_pos[io_u->ddir] = f->file_offset; | |
851 | } | |
852 | ||
853 | /* | |
854 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
855 | * map needs to be reset as it's done with zone_range everytime. | |
856 | */ | |
04bc85a1 | 857 | if ((td->zone_bytes % td->o.zone_range) == 0) |
224b3093 | 858 | fio_file_reset(td, f); |
224b3093 | 859 | } |
860 | ||
9bf2061e | 861 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 862 | { |
ec370c22 | 863 | bool is_random; |
6aca9b3d | 864 | |
9b87f09b | 865 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
866 | goto out; |
867 | ||
1ef2b6be | 868 | set_rw_ddir(td, io_u); |
5f9099ea | 869 | |
87dc1ab1 | 870 | /* |
ff58fced | 871 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 872 | */ |
ff58fced | 873 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 874 | goto out; |
a00735e6 | 875 | |
7b865a2f BVA |
876 | if (td->o.zone_mode == ZONE_MODE_STRIDED) |
877 | setup_strided_zone_mode(td, io_u); | |
48f5abd3 | 878 | |
10ba535a | 879 | /* |
c685b5b2 JA |
880 | * No log, let the seq/rand engine retrieve the next buflen and |
881 | * position. | |
10ba535a | 882 | */ |
6aca9b3d | 883 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 884 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 885 | return 1; |
2ba1c290 | 886 | } |
10ba535a | 887 | |
6aca9b3d | 888 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
889 | if (!io_u->buflen) { |
890 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 891 | return 1; |
2ba1c290 | 892 | } |
bca4ed4d | 893 | |
2ba1c290 | 894 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
5fff9543 | 895 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n", |
e5f9a813 | 896 | io_u, |
4b91ee8f JA |
897 | (unsigned long long) io_u->offset, io_u->buflen, |
898 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 899 | return 1; |
2ba1c290 | 900 | } |
6a5e6884 | 901 | |
bca4ed4d JA |
902 | /* |
903 | * mark entry before potentially trimming io_u | |
904 | */ | |
303032ae | 905 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 906 | mark_random_map(td, io_u); |
bca4ed4d | 907 | |
c38e9468 | 908 | out: |
e5f9a813 | 909 | dprint_io_u(io_u, "fill"); |
d9d91e39 | 910 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 911 | return 0; |
10ba535a JA |
912 | } |
913 | ||
6cc0e5aa | 914 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
838bc709 | 915 | { |
2b13e716 | 916 | int idx = 0; |
838bc709 JA |
917 | |
918 | switch (nr) { | |
919 | default: | |
2b13e716 | 920 | idx = 6; |
838bc709 JA |
921 | break; |
922 | case 33 ... 64: | |
2b13e716 | 923 | idx = 5; |
838bc709 JA |
924 | break; |
925 | case 17 ... 32: | |
2b13e716 | 926 | idx = 4; |
838bc709 JA |
927 | break; |
928 | case 9 ... 16: | |
2b13e716 | 929 | idx = 3; |
838bc709 JA |
930 | break; |
931 | case 5 ... 8: | |
2b13e716 | 932 | idx = 2; |
838bc709 JA |
933 | break; |
934 | case 1 ... 4: | |
2b13e716 | 935 | idx = 1; |
838bc709 JA |
936 | case 0: |
937 | break; | |
938 | } | |
939 | ||
2b13e716 | 940 | map[idx]++; |
838bc709 JA |
941 | } |
942 | ||
943 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
944 | { | |
945 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
946 | td->ts.total_submit++; | |
947 | } | |
948 | ||
949 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
950 | { | |
951 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
952 | td->ts.total_complete++; | |
953 | } | |
954 | ||
d8005759 | 955 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 956 | { |
2b13e716 | 957 | int idx = 0; |
71619dc2 JA |
958 | |
959 | switch (td->cur_depth) { | |
960 | default: | |
2b13e716 | 961 | idx = 6; |
a783e61a | 962 | break; |
71619dc2 | 963 | case 32 ... 63: |
2b13e716 | 964 | idx = 5; |
a783e61a | 965 | break; |
71619dc2 | 966 | case 16 ... 31: |
2b13e716 | 967 | idx = 4; |
a783e61a | 968 | break; |
71619dc2 | 969 | case 8 ... 15: |
2b13e716 | 970 | idx = 3; |
a783e61a | 971 | break; |
71619dc2 | 972 | case 4 ... 7: |
2b13e716 | 973 | idx = 2; |
a783e61a | 974 | break; |
71619dc2 | 975 | case 2 ... 3: |
2b13e716 | 976 | idx = 1; |
71619dc2 JA |
977 | case 1: |
978 | break; | |
979 | } | |
980 | ||
2b13e716 | 981 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
982 | } |
983 | ||
d6bb626e | 984 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 985 | { |
2b13e716 | 986 | int idx = 0; |
04a0feae | 987 | |
d6bb626e VF |
988 | assert(nsec < 1000); |
989 | ||
990 | switch (nsec) { | |
991 | case 750 ... 999: | |
992 | idx = 9; | |
993 | break; | |
994 | case 500 ... 749: | |
995 | idx = 8; | |
996 | break; | |
997 | case 250 ... 499: | |
998 | idx = 7; | |
999 | break; | |
1000 | case 100 ... 249: | |
1001 | idx = 6; | |
1002 | break; | |
1003 | case 50 ... 99: | |
1004 | idx = 5; | |
1005 | break; | |
1006 | case 20 ... 49: | |
1007 | idx = 4; | |
1008 | break; | |
1009 | case 10 ... 19: | |
1010 | idx = 3; | |
1011 | break; | |
1012 | case 4 ... 9: | |
1013 | idx = 2; | |
1014 | break; | |
1015 | case 2 ... 3: | |
1016 | idx = 1; | |
1017 | case 0 ... 1: | |
1018 | break; | |
1019 | } | |
1020 | ||
1021 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1022 | td->ts.io_u_lat_n[idx]++; | |
1023 | } | |
1024 | ||
1025 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1026 | { | |
1027 | int idx = 0; | |
1028 | ||
1029 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1030 | |
1031 | switch (usec) { | |
1032 | case 750 ... 999: | |
2b13e716 | 1033 | idx = 9; |
04a0feae JA |
1034 | break; |
1035 | case 500 ... 749: | |
2b13e716 | 1036 | idx = 8; |
04a0feae JA |
1037 | break; |
1038 | case 250 ... 499: | |
2b13e716 | 1039 | idx = 7; |
04a0feae JA |
1040 | break; |
1041 | case 100 ... 249: | |
2b13e716 | 1042 | idx = 6; |
04a0feae JA |
1043 | break; |
1044 | case 50 ... 99: | |
2b13e716 | 1045 | idx = 5; |
04a0feae JA |
1046 | break; |
1047 | case 20 ... 49: | |
2b13e716 | 1048 | idx = 4; |
04a0feae JA |
1049 | break; |
1050 | case 10 ... 19: | |
2b13e716 | 1051 | idx = 3; |
04a0feae JA |
1052 | break; |
1053 | case 4 ... 9: | |
2b13e716 | 1054 | idx = 2; |
04a0feae JA |
1055 | break; |
1056 | case 2 ... 3: | |
2b13e716 | 1057 | idx = 1; |
04a0feae JA |
1058 | case 0 ... 1: |
1059 | break; | |
1060 | } | |
1061 | ||
2b13e716 JA |
1062 | assert(idx < FIO_IO_U_LAT_U_NR); |
1063 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1064 | } |
1065 | ||
d6bb626e | 1066 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1067 | { |
2b13e716 | 1068 | int idx = 0; |
ec118304 | 1069 | |
d6bb626e VF |
1070 | assert(msec >= 1); |
1071 | ||
ec118304 JA |
1072 | switch (msec) { |
1073 | default: | |
2b13e716 | 1074 | idx = 11; |
04a0feae | 1075 | break; |
8abdce66 | 1076 | case 1000 ... 1999: |
2b13e716 | 1077 | idx = 10; |
04a0feae | 1078 | break; |
8abdce66 | 1079 | case 750 ... 999: |
2b13e716 | 1080 | idx = 9; |
04a0feae | 1081 | break; |
8abdce66 | 1082 | case 500 ... 749: |
2b13e716 | 1083 | idx = 8; |
04a0feae | 1084 | break; |
8abdce66 | 1085 | case 250 ... 499: |
2b13e716 | 1086 | idx = 7; |
04a0feae | 1087 | break; |
8abdce66 | 1088 | case 100 ... 249: |
2b13e716 | 1089 | idx = 6; |
04a0feae | 1090 | break; |
8abdce66 | 1091 | case 50 ... 99: |
2b13e716 | 1092 | idx = 5; |
04a0feae | 1093 | break; |
8abdce66 | 1094 | case 20 ... 49: |
2b13e716 | 1095 | idx = 4; |
04a0feae | 1096 | break; |
8abdce66 | 1097 | case 10 ... 19: |
2b13e716 | 1098 | idx = 3; |
04a0feae | 1099 | break; |
8abdce66 | 1100 | case 4 ... 9: |
2b13e716 | 1101 | idx = 2; |
04a0feae | 1102 | break; |
ec118304 | 1103 | case 2 ... 3: |
2b13e716 | 1104 | idx = 1; |
ec118304 JA |
1105 | case 0 ... 1: |
1106 | break; | |
1107 | } | |
1108 | ||
2b13e716 JA |
1109 | assert(idx < FIO_IO_U_LAT_M_NR); |
1110 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1111 | } |
1112 | ||
d6bb626e | 1113 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1114 | { |
d6bb626e VF |
1115 | if (nsec < 1000) |
1116 | io_u_mark_lat_nsec(td, nsec); | |
1117 | else if (nsec < 1000000) | |
1118 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1119 | else |
d6bb626e | 1120 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1121 | } |
1122 | ||
8c07860d JA |
1123 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1124 | { | |
1125 | unsigned long fileno; | |
1126 | ||
1127 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1128 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1129 | unsigned long r; | |
1130 | ||
1131 | r = __rand(&td->next_file_state); | |
1132 | return (unsigned int) ((double) td->o.nr_files | |
1133 | * (r / (frand_max + 1.0))); | |
1134 | } | |
1135 | ||
1136 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1137 | fileno = zipf_next(&td->next_file_zipf); | |
1138 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1139 | fileno = pareto_next(&td->next_file_zipf); | |
1140 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1141 | fileno = gauss_next(&td->next_file_gauss); | |
1142 | else { | |
1143 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1144 | assert(0); | |
1145 | return 0; | |
1146 | } | |
1147 | ||
1148 | return fileno >> FIO_FSERVICE_SHIFT; | |
1149 | } | |
1150 | ||
0aabe160 JA |
1151 | /* |
1152 | * Get next file to service by choosing one at random | |
1153 | */ | |
2cc52930 JA |
1154 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1155 | enum fio_file_flags goodf, | |
d6aed795 | 1156 | enum fio_file_flags badf) |
0aabe160 | 1157 | { |
0aabe160 | 1158 | struct fio_file *f; |
1c178180 | 1159 | int fno; |
0aabe160 JA |
1160 | |
1161 | do { | |
87b10676 | 1162 | int opened = 0; |
4c07ad86 | 1163 | |
8c07860d | 1164 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1165 | |
126d65c6 | 1166 | f = td->files[fno]; |
d6aed795 | 1167 | if (fio_file_done(f)) |
059e63c0 | 1168 | continue; |
1c178180 | 1169 | |
d6aed795 | 1170 | if (!fio_file_open(f)) { |
87b10676 JA |
1171 | int err; |
1172 | ||
002fe734 JA |
1173 | if (td->nr_open_files >= td->o.open_files) |
1174 | return ERR_PTR(-EBUSY); | |
1175 | ||
87b10676 JA |
1176 | err = td_io_open_file(td, f); |
1177 | if (err) | |
1178 | continue; | |
1179 | opened = 1; | |
1180 | } | |
1181 | ||
2ba1c290 JA |
1182 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1183 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1184 | return f; |
2ba1c290 | 1185 | } |
87b10676 JA |
1186 | if (opened) |
1187 | td_io_close_file(td, f); | |
0aabe160 JA |
1188 | } while (1); |
1189 | } | |
1190 | ||
1191 | /* | |
1192 | * Get next file to service by doing round robin between all available ones | |
1193 | */ | |
1c178180 JA |
1194 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1195 | int badf) | |
3d7c391d JA |
1196 | { |
1197 | unsigned int old_next_file = td->next_file; | |
1198 | struct fio_file *f; | |
1199 | ||
1200 | do { | |
87b10676 JA |
1201 | int opened = 0; |
1202 | ||
126d65c6 | 1203 | f = td->files[td->next_file]; |
3d7c391d JA |
1204 | |
1205 | td->next_file++; | |
2dc1bbeb | 1206 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1207 | td->next_file = 0; |
1208 | ||
87b10676 | 1209 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1210 | if (fio_file_done(f)) { |
d5ed68ea | 1211 | f = NULL; |
059e63c0 | 1212 | continue; |
d5ed68ea | 1213 | } |
059e63c0 | 1214 | |
d6aed795 | 1215 | if (!fio_file_open(f)) { |
87b10676 JA |
1216 | int err; |
1217 | ||
002fe734 JA |
1218 | if (td->nr_open_files >= td->o.open_files) |
1219 | return ERR_PTR(-EBUSY); | |
1220 | ||
87b10676 | 1221 | err = td_io_open_file(td, f); |
b5696bfc JA |
1222 | if (err) { |
1223 | dprint(FD_FILE, "error %d on open of %s\n", | |
1224 | err, f->file_name); | |
87c27b45 | 1225 | f = NULL; |
87b10676 | 1226 | continue; |
b5696bfc | 1227 | } |
87b10676 JA |
1228 | opened = 1; |
1229 | } | |
1230 | ||
0b9d69ec JA |
1231 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1232 | f->flags); | |
1c178180 | 1233 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1234 | break; |
1235 | ||
87b10676 JA |
1236 | if (opened) |
1237 | td_io_close_file(td, f); | |
1238 | ||
3d7c391d JA |
1239 | f = NULL; |
1240 | } while (td->next_file != old_next_file); | |
1241 | ||
2ba1c290 | 1242 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1243 | return f; |
1244 | } | |
1245 | ||
7eb36574 | 1246 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1247 | { |
1907dbc6 JA |
1248 | struct fio_file *f; |
1249 | ||
2dc1bbeb | 1250 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1251 | |
b5696bfc | 1252 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1253 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1254 | " nr_files=%d\n", td->nr_open_files, | |
1255 | td->nr_done_files, | |
1256 | td->o.nr_files); | |
bdb4e2e9 | 1257 | return NULL; |
2ba1c290 | 1258 | } |
bdb4e2e9 | 1259 | |
1907dbc6 | 1260 | f = td->file_service_file; |
d6aed795 | 1261 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1262 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1263 | goto out; | |
1264 | if (td->file_service_left--) | |
1265 | goto out; | |
1266 | } | |
1907dbc6 | 1267 | |
a086c257 JA |
1268 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1269 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1270 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1271 | else |
d6aed795 | 1272 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1273 | |
002fe734 JA |
1274 | if (IS_ERR(f)) |
1275 | return f; | |
1276 | ||
1907dbc6 JA |
1277 | td->file_service_file = f; |
1278 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1279 | out: |
0dac421f JA |
1280 | if (f) |
1281 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1282 | else | |
1283 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1284 | return f; |
bdb4e2e9 JA |
1285 | } |
1286 | ||
7eb36574 JA |
1287 | static struct fio_file *get_next_file(struct thread_data *td) |
1288 | { | |
7eb36574 JA |
1289 | return __get_next_file(td); |
1290 | } | |
1291 | ||
002fe734 | 1292 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1293 | { |
1294 | struct fio_file *f; | |
1295 | ||
1296 | do { | |
1297 | f = get_next_file(td); | |
002fe734 JA |
1298 | if (IS_ERR_OR_NULL(f)) |
1299 | return PTR_ERR(f); | |
429f6675 | 1300 | |
429f6675 JA |
1301 | io_u->file = f; |
1302 | get_file(f); | |
1303 | ||
1304 | if (!fill_io_u(td, io_u)) | |
1305 | break; | |
1306 | ||
b5696bfc | 1307 | put_file_log(td, f); |
429f6675 | 1308 | td_io_close_file(td, f); |
b5696bfc | 1309 | io_u->file = NULL; |
8c07860d JA |
1310 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1311 | fio_file_reset(td, f); | |
1312 | else { | |
1313 | fio_file_set_done(f); | |
1314 | td->nr_done_files++; | |
1315 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1316 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1317 | } |
429f6675 JA |
1318 | } while (1); |
1319 | ||
1320 | return 0; | |
1321 | } | |
1322 | ||
3e260a46 | 1323 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
c3a32714 | 1324 | unsigned long long tnsec, unsigned long long max_nsec) |
3e260a46 JA |
1325 | { |
1326 | if (!td->error) | |
c3a32714 | 1327 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
3e260a46 JA |
1328 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
1329 | icd->error = ETIMEDOUT; | |
1330 | } | |
1331 | ||
1332 | static void lat_new_cycle(struct thread_data *td) | |
1333 | { | |
1334 | fio_gettime(&td->latency_ts, NULL); | |
1335 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1336 | td->latency_failed = 0; | |
1337 | } | |
1338 | ||
1339 | /* | |
1340 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1341 | * are at QD=1, then it's time to give up. | |
1342 | */ | |
e39c0676 | 1343 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1344 | { |
1345 | if (td->latency_qd == 1) | |
e39c0676 | 1346 | return true; |
3e260a46 JA |
1347 | |
1348 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1349 | |
1350 | if (td->latency_qd == td->latency_qd_low) | |
1351 | td->latency_qd_low--; | |
1352 | ||
3e260a46 JA |
1353 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1354 | ||
1355 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1356 | ||
1357 | /* | |
1358 | * When we ramp QD down, quiesce existing IO to prevent | |
1359 | * a storm of ramp downs due to pending higher depth. | |
1360 | */ | |
1361 | io_u_quiesce(td); | |
1362 | lat_new_cycle(td); | |
e39c0676 | 1363 | return false; |
3e260a46 JA |
1364 | } |
1365 | ||
e39c0676 | 1366 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1367 | { |
1368 | if (td->o.latency_percentile.u.f == 100.0) | |
1369 | return __lat_target_failed(td); | |
1370 | ||
1371 | td->latency_failed++; | |
e39c0676 | 1372 | return false; |
3e260a46 JA |
1373 | } |
1374 | ||
1375 | void lat_target_init(struct thread_data *td) | |
1376 | { | |
6bb58215 JA |
1377 | td->latency_end_run = 0; |
1378 | ||
3e260a46 JA |
1379 | if (td->o.latency_target) { |
1380 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1381 | fio_gettime(&td->latency_ts, NULL); | |
1382 | td->latency_qd = 1; | |
1383 | td->latency_qd_high = td->o.iodepth; | |
1384 | td->latency_qd_low = 1; | |
1385 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1386 | } else | |
1387 | td->latency_qd = td->o.iodepth; | |
1388 | } | |
1389 | ||
6bb58215 JA |
1390 | void lat_target_reset(struct thread_data *td) |
1391 | { | |
1392 | if (!td->latency_end_run) | |
1393 | lat_target_init(td); | |
1394 | } | |
1395 | ||
3e260a46 JA |
1396 | static void lat_target_success(struct thread_data *td) |
1397 | { | |
1398 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1399 | struct thread_options *o = &td->o; |
3e260a46 JA |
1400 | |
1401 | td->latency_qd_low = td->latency_qd; | |
1402 | ||
1403 | /* | |
1404 | * If we haven't failed yet, we double up to a failing value instead | |
1405 | * of bisecting from highest possible queue depth. If we have set | |
1406 | * a limit other than td->o.iodepth, bisect between that. | |
1407 | */ | |
6bb58215 | 1408 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1409 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1410 | else | |
1411 | td->latency_qd *= 2; | |
1412 | ||
6bb58215 JA |
1413 | if (td->latency_qd > o->iodepth) |
1414 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1415 | |
1416 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1417 | |
3e260a46 | 1418 | /* |
6bb58215 JA |
1419 | * Same as last one, we are done. Let it run a latency cycle, so |
1420 | * we get only the results from the targeted depth. | |
3e260a46 | 1421 | */ |
6bb58215 JA |
1422 | if (td->latency_qd == qd) { |
1423 | if (td->latency_end_run) { | |
1424 | dprint(FD_RATE, "We are done\n"); | |
1425 | td->done = 1; | |
1426 | } else { | |
1427 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1428 | io_u_quiesce(td); | |
1429 | td->latency_end_run = 1; | |
1430 | reset_all_stats(td); | |
1431 | reset_io_stats(td); | |
1432 | } | |
1433 | } | |
3e260a46 JA |
1434 | |
1435 | lat_new_cycle(td); | |
1436 | } | |
1437 | ||
1438 | /* | |
1439 | * Check if we can bump the queue depth | |
1440 | */ | |
1441 | void lat_target_check(struct thread_data *td) | |
1442 | { | |
1443 | uint64_t usec_window; | |
1444 | uint64_t ios; | |
1445 | double success_ios; | |
1446 | ||
1447 | usec_window = utime_since_now(&td->latency_ts); | |
1448 | if (usec_window < td->o.latency_window) | |
1449 | return; | |
1450 | ||
1451 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1452 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1453 | success_ios *= 100.0; | |
1454 | ||
1455 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1456 | ||
1457 | if (success_ios >= td->o.latency_percentile.u.f) | |
1458 | lat_target_success(td); | |
1459 | else | |
1460 | __lat_target_failed(td); | |
1461 | } | |
1462 | ||
1463 | /* | |
1464 | * If latency target is enabled, we might be ramping up or down and not | |
1465 | * using the full queue depth available. | |
1466 | */ | |
e39c0676 | 1467 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1468 | { |
1469 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1470 | ||
1471 | if (qempty) | |
e39c0676 | 1472 | return true; |
3e260a46 | 1473 | if (!td->o.latency_target) |
e39c0676 | 1474 | return false; |
3e260a46 JA |
1475 | |
1476 | return td->cur_depth >= td->latency_qd; | |
1477 | } | |
429f6675 | 1478 | |
10ba535a JA |
1479 | struct io_u *__get_io_u(struct thread_data *td) |
1480 | { | |
0cae66f6 | 1481 | struct io_u *io_u = NULL; |
93b45bb2 | 1482 | int ret; |
10ba535a | 1483 | |
ca09be4b JA |
1484 | if (td->stop_io) |
1485 | return NULL; | |
1486 | ||
e8462bd8 JA |
1487 | td_io_u_lock(td); |
1488 | ||
1489 | again: | |
2ae0b204 JA |
1490 | if (!io_u_rempty(&td->io_u_requeues)) |
1491 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1492 | else if (!queue_full(td)) { |
2ae0b204 | 1493 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1494 | |
225ba9e3 | 1495 | io_u->file = NULL; |
6040dabc | 1496 | io_u->buflen = 0; |
10ba535a | 1497 | io_u->resid = 0; |
d7762cf8 | 1498 | io_u->end_io = NULL; |
755200a3 JA |
1499 | } |
1500 | ||
1501 | if (io_u) { | |
0c6e7517 | 1502 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1503 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1504 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1505 | IO_U_F_VER_LIST); | |
0c6e7517 | 1506 | |
755200a3 | 1507 | io_u->error = 0; |
bcd5abfa | 1508 | io_u->acct_ddir = -1; |
10ba535a | 1509 | td->cur_depth++; |
a9da8ab2 | 1510 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1511 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1512 | io_u->ipo = NULL; |
a9da8ab2 | 1513 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1514 | /* |
1515 | * We ran out, wait for async verify threads to finish and | |
1516 | * return one | |
1517 | */ | |
a9da8ab2 | 1518 | assert(!(td->flags & TD_F_CHILD)); |
93b45bb2 BVA |
1519 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
1520 | assert(ret == 0); | |
1dec3e07 | 1521 | goto again; |
10ba535a JA |
1522 | } |
1523 | ||
e8462bd8 | 1524 | td_io_u_unlock(td); |
10ba535a JA |
1525 | return io_u; |
1526 | } | |
1527 | ||
e39c0676 | 1528 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1529 | { |
d72be545 | 1530 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1531 | return false; |
c9a73054 JA |
1532 | if (!td->trim_entries) |
1533 | return false; | |
d72be545 | 1534 | |
c9a73054 JA |
1535 | if (td->trim_batch) { |
1536 | td->trim_batch--; | |
1537 | if (get_next_trim(td, io_u)) | |
1538 | return true; | |
1539 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1540 | td->last_ddir != DDIR_READ) { | |
1541 | td->trim_batch = td->o.trim_batch; | |
1542 | if (!td->trim_batch) | |
1543 | td->trim_batch = td->o.trim_backlog; | |
1544 | if (get_next_trim(td, io_u)) | |
e39c0676 | 1545 | return true; |
2ba1c290 | 1546 | } |
10ba535a | 1547 | |
e39c0676 | 1548 | return false; |
0d29de83 JA |
1549 | } |
1550 | ||
e39c0676 | 1551 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1552 | { |
d72be545 | 1553 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1554 | return false; |
d72be545 JA |
1555 | |
1556 | if (td->io_hist_len) { | |
9e144189 JA |
1557 | int get_verify = 0; |
1558 | ||
d1ece0c7 | 1559 | if (td->verify_batch) |
9e144189 | 1560 | get_verify = 1; |
d1ece0c7 | 1561 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1562 | td->last_ddir != DDIR_READ) { |
1563 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1564 | if (!td->verify_batch) |
1565 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1566 | get_verify = 1; |
1567 | } | |
1568 | ||
d1ece0c7 JA |
1569 | if (get_verify && !get_next_verify(td, io_u)) { |
1570 | td->verify_batch--; | |
e39c0676 | 1571 | return true; |
d1ece0c7 | 1572 | } |
9e144189 JA |
1573 | } |
1574 | ||
e39c0676 | 1575 | return false; |
0d29de83 JA |
1576 | } |
1577 | ||
de789769 JA |
1578 | /* |
1579 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1580 | * easy compressible data for simple de-dupe attempts. Do this for every |
1581 | * 512b block in the range, since that should be the smallest block size | |
1582 | * we can expect from a device. | |
de789769 JA |
1583 | */ |
1584 | static void small_content_scramble(struct io_u *io_u) | |
1585 | { | |
5fff9543 | 1586 | unsigned long long i, nr_blocks = io_u->buflen >> 9; |
23f394d5 | 1587 | unsigned int offset; |
319b073f JA |
1588 | uint64_t boffset, *iptr; |
1589 | char *p; | |
de789769 | 1590 | |
23f394d5 JA |
1591 | if (!nr_blocks) |
1592 | return; | |
1593 | ||
1594 | p = io_u->xfer_buf; | |
fba76ee8 | 1595 | boffset = io_u->offset; |
319b073f JA |
1596 | |
1597 | if (io_u->buf_filled_len) | |
1598 | io_u->buf_filled_len = 0; | |
1599 | ||
1600 | /* | |
1601 | * Generate random index between 0..7. We do chunks of 512b, if | |
1602 | * we assume a cacheline is 64 bytes, then we have 8 of those. | |
1603 | * Scramble content within the blocks in the same cacheline to | |
1604 | * speed things up. | |
1605 | */ | |
1606 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; | |
fad82f76 | 1607 | |
23f394d5 JA |
1608 | for (i = 0; i < nr_blocks; i++) { |
1609 | /* | |
319b073f JA |
1610 | * Fill offset into start of cacheline, time into end |
1611 | * of cacheline | |
23f394d5 | 1612 | */ |
319b073f JA |
1613 | iptr = (void *) p + (offset << 6); |
1614 | *iptr = boffset; | |
1615 | ||
1616 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); | |
1617 | iptr[0] = io_u->start_time.tv_sec; | |
1618 | iptr[1] = io_u->start_time.tv_nsec; | |
23f394d5 | 1619 | |
23f394d5 | 1620 | p += 512; |
fad82f76 | 1621 | boffset += 512; |
23f394d5 | 1622 | } |
de789769 JA |
1623 | } |
1624 | ||
0d29de83 JA |
1625 | /* |
1626 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
5c5c33c1 | 1627 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
0d29de83 JA |
1628 | */ |
1629 | struct io_u *get_io_u(struct thread_data *td) | |
1630 | { | |
1631 | struct fio_file *f; | |
1632 | struct io_u *io_u; | |
de789769 | 1633 | int do_scramble = 0; |
002fe734 | 1634 | long ret = 0; |
0d29de83 JA |
1635 | |
1636 | io_u = __get_io_u(td); | |
1637 | if (!io_u) { | |
1638 | dprint(FD_IO, "__get_io_u failed\n"); | |
1639 | return NULL; | |
1640 | } | |
1641 | ||
1642 | if (check_get_verify(td, io_u)) | |
1643 | goto out; | |
1644 | if (check_get_trim(td, io_u)) | |
1645 | goto out; | |
1646 | ||
755200a3 JA |
1647 | /* |
1648 | * from a requeue, io_u already setup | |
1649 | */ | |
1650 | if (io_u->file) | |
77f392bf | 1651 | goto out; |
755200a3 | 1652 | |
429f6675 JA |
1653 | /* |
1654 | * If using an iolog, grab next piece if any available. | |
1655 | */ | |
d72be545 | 1656 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1657 | if (read_iolog_get(td, io_u)) |
1658 | goto err_put; | |
2ba1c290 | 1659 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1660 | ret = -EBUSY; |
2ba1c290 | 1661 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1662 | goto err_put; |
2ba1c290 | 1663 | } |
5ec10eaa | 1664 | |
429f6675 | 1665 | f = io_u->file; |
002fe734 JA |
1666 | if (!f) { |
1667 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1668 | goto err_put; | |
1669 | } | |
1670 | ||
d6aed795 | 1671 | assert(fio_file_open(f)); |
97af62ce | 1672 | |
ff58fced | 1673 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1674 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1675 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1676 | goto err_put; |
2ba1c290 | 1677 | } |
10ba535a | 1678 | |
f1dfb668 JA |
1679 | f->last_start[io_u->ddir] = io_u->offset; |
1680 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1681 | |
fd68418e | 1682 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1683 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1684 | io_u_fill_buffer(td, io_u, |
1066358a | 1685 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1686 | io_u->buflen); |
ff441ae8 | 1687 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
5c5c33c1 BVA |
1688 | !(td->flags & TD_F_COMPRESS) && |
1689 | !(td->flags & TD_F_DO_VERIFY)) | |
fd68418e JA |
1690 | do_scramble = 1; |
1691 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1692 | /* |
1693 | * Reset the buf_filled parameters so next time if the | |
1694 | * buffer is used for writes it is refilled. | |
1695 | */ | |
cbe8d756 RR |
1696 | io_u->buf_filled_len = 0; |
1697 | } | |
87dc1ab1 | 1698 | } |
10ba535a | 1699 | |
165faf16 JA |
1700 | /* |
1701 | * Set io data pointers. | |
1702 | */ | |
cec6b55d JA |
1703 | io_u->xfer_buf = io_u->buf; |
1704 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1705 | |
6ac7a331 | 1706 | out: |
0d29de83 | 1707 | assert(io_u->file); |
429f6675 | 1708 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1709 | if (!td->o.disable_lat) |
993bf48b | 1710 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1711 | |
de789769 JA |
1712 | if (do_scramble) |
1713 | small_content_scramble(io_u); | |
03553853 | 1714 | |
429f6675 | 1715 | return io_u; |
36167d82 | 1716 | } |
429f6675 | 1717 | err_put: |
2ba1c290 | 1718 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1719 | put_io_u(td, io_u); |
002fe734 | 1720 | return ERR_PTR(ret); |
10ba535a JA |
1721 | } |
1722 | ||
a9da8ab2 | 1723 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1724 | { |
8b28bd41 | 1725 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1726 | |
8b28bd41 DM |
1727 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1728 | return; | |
5451792e | 1729 | |
5fff9543 | 1730 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n", |
709c8313 RE |
1731 | io_u->file ? " on file " : "", |
1732 | io_u->file ? io_u->file->file_name : "", | |
1733 | strerror(io_u->error), | |
1734 | io_ddir_name(io_u->ddir), | |
1735 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1736 | |
5ad7be56 KD |
1737 | if (td->io_ops->errdetails) { |
1738 | char *err = td->io_ops->errdetails(io_u); | |
1739 | ||
1740 | log_err("fio: %s\n", err); | |
1741 | free(err); | |
1742 | } | |
1743 | ||
5451792e JA |
1744 | if (!td->error) |
1745 | td_verror(td, io_u->error, "io_u error"); | |
1746 | } | |
1747 | ||
a9da8ab2 JA |
1748 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1749 | { | |
1750 | __io_u_log_error(td, io_u); | |
1751 | if (td->parent) | |
094e66cb | 1752 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1753 | } |
1754 | ||
e39c0676 | 1755 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1756 | { |
3ecc8c67 JA |
1757 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1758 | || td->o.gtod_reduce; | |
aba6c951 JA |
1759 | } |
1760 | ||
c8eeb9df JA |
1761 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1762 | struct io_completion_data *icd, | |
1763 | const enum fio_ddir idx, unsigned int bytes) | |
1764 | { | |
a9da8ab2 | 1765 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1766 | unsigned long long llnsec = 0; |
c8eeb9df | 1767 | |
75dc383e JA |
1768 | if (td->parent) |
1769 | td = td->parent; | |
1770 | ||
132b1ee4 | 1771 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1772 | return; |
1773 | ||
a9da8ab2 | 1774 | if (no_reduce) |
d6bb626e | 1775 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1776 | |
1777 | if (!td->o.disable_lat) { | |
c3a32714 | 1778 | unsigned long long tnsec; |
c8eeb9df | 1779 | |
d6bb626e VF |
1780 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
1781 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); | |
15501535 | 1782 | |
d4afedfd JA |
1783 | if (td->flags & TD_F_PROFILE_OPS) { |
1784 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1785 | ||
1786 | if (ops->io_u_lat) | |
c3a32714 | 1787 | icd->error = ops->io_u_lat(td, tnsec); |
d4afedfd JA |
1788 | } |
1789 | ||
c3a32714 JA |
1790 | if (td->o.max_latency && tnsec > td->o.max_latency) |
1791 | lat_fatal(td, icd, tnsec, td->o.max_latency); | |
1792 | if (td->o.latency_target && tnsec > td->o.latency_target) { | |
3e260a46 | 1793 | if (lat_target_failed(td)) |
c3a32714 | 1794 | lat_fatal(td, icd, tnsec, td->o.latency_target); |
15501535 | 1795 | } |
c8eeb9df JA |
1796 | } |
1797 | ||
a47591e4 JA |
1798 | if (ddir_rw(idx)) { |
1799 | if (!td->o.disable_clat) { | |
d6bb626e VF |
1800 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
1801 | io_u_mark_latency(td, llnsec); | |
a47591e4 | 1802 | } |
c8eeb9df | 1803 | |
a47591e4 | 1804 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1805 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1806 | |
a47591e4 JA |
1807 | if (no_reduce && per_unit_log(td->iops_log)) |
1808 | add_iops_sample(td, io_u, bytes); | |
b2b3eefe JA |
1809 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
1810 | add_sync_clat_sample(&td->ts, llnsec); | |
66347cfa DE |
1811 | |
1812 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1813 | uint32_t *info = io_u_block_info(td, io_u); | |
1814 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1815 | if (io_u->ddir == DDIR_TRIM) { | |
1816 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1817 | BLOCK_INFO_TRIMS(*info) + 1); | |
1818 | } else if (io_u->ddir == DDIR_WRITE) { | |
1819 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1820 | *info); | |
1821 | } | |
1822 | } | |
1823 | } | |
c8eeb9df JA |
1824 | } |
1825 | ||
94a6e1bb JA |
1826 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1827 | uint64_t offset, unsigned int bytes) | |
1828 | { | |
1829 | int idx; | |
1830 | ||
639ad1ea JA |
1831 | if (!f) |
1832 | return; | |
1833 | ||
94a6e1bb JA |
1834 | if (f->first_write == -1ULL || offset < f->first_write) |
1835 | f->first_write = offset; | |
1836 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1837 | f->last_write = offset + bytes; | |
1838 | ||
1839 | if (!f->last_write_comp) | |
1840 | return; | |
1841 | ||
1842 | idx = f->last_write_idx++; | |
1843 | f->last_write_comp[idx] = offset; | |
1844 | if (f->last_write_idx == td->o.iodepth) | |
1845 | f->last_write_idx = 0; | |
1846 | } | |
1847 | ||
b2b3eefe JA |
1848 | static bool should_account(struct thread_data *td) |
1849 | { | |
1850 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || | |
1851 | td->runstate == TD_VERIFYING); | |
1852 | } | |
1853 | ||
f8b0bd10 | 1854 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1855 | struct io_completion_data *icd) |
10ba535a | 1856 | { |
f8b0bd10 JA |
1857 | struct io_u *io_u = *io_u_ptr; |
1858 | enum fio_ddir ddir = io_u->ddir; | |
1859 | struct fio_file *f = io_u->file; | |
10ba535a | 1860 | |
e5f9a813 | 1861 | dprint_io_u(io_u, "complete"); |
2ba1c290 | 1862 | |
0c6e7517 | 1863 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1864 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1865 | |
1866 | /* | |
1867 | * Mark IO ok to verify | |
1868 | */ | |
1869 | if (io_u->ipo) { | |
890b6656 JA |
1870 | /* |
1871 | * Remove errored entry from the verification list | |
1872 | */ | |
1873 | if (io_u->error) | |
1874 | unlog_io_piece(td, io_u); | |
1875 | else { | |
1876 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1877 | write_barrier(); | |
1878 | } | |
f9401285 JA |
1879 | } |
1880 | ||
f8b0bd10 | 1881 | if (ddir_sync(ddir)) { |
2ea93f98 | 1882 | td->last_was_sync = true; |
44f29692 JA |
1883 | if (f) { |
1884 | f->first_write = -1ULL; | |
1885 | f->last_write = -1ULL; | |
1886 | } | |
b2b3eefe JA |
1887 | if (should_account(td)) |
1888 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); | |
87dc1ab1 JA |
1889 | return; |
1890 | } | |
1891 | ||
2ea93f98 | 1892 | td->last_was_sync = false; |
f8b0bd10 | 1893 | td->last_ddir = ddir; |
87dc1ab1 | 1894 | |
f8b0bd10 | 1895 | if (!io_u->error && ddir_rw(ddir)) { |
5fff9543 | 1896 | unsigned long long bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1897 | int ret; |
10ba535a | 1898 | |
f8b0bd10 | 1899 | td->io_blocks[ddir]++; |
f8b0bd10 | 1900 | td->io_bytes[ddir] += bytes; |
ae2fafc8 | 1901 | |
e1c325d2 JA |
1902 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
1903 | td->this_io_blocks[ddir]++; | |
f8b0bd10 | 1904 | td->this_io_bytes[ddir] += bytes; |
e1c325d2 | 1905 | } |
f8b0bd10 | 1906 | |
639ad1ea | 1907 | if (ddir == DDIR_WRITE) |
94a6e1bb | 1908 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 1909 | |
b2b3eefe | 1910 | if (should_account(td)) |
f8b0bd10 | 1911 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1912 | |
f8b0bd10 | 1913 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1914 | |
d7762cf8 | 1915 | if (io_u->end_io) { |
f8b0bd10 JA |
1916 | ret = io_u->end_io(td, io_u_ptr); |
1917 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1918 | if (ret && !icd->error) |
1919 | icd->error = ret; | |
1920 | } | |
ff58fced | 1921 | } else if (io_u->error) { |
10ba535a | 1922 | icd->error = io_u->error; |
5451792e JA |
1923 | io_u_log_error(td, io_u); |
1924 | } | |
8b28bd41 | 1925 | if (icd->error) { |
f8b0bd10 JA |
1926 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1927 | ||
8b28bd41 DM |
1928 | if (!td_non_fatal_error(td, eb, icd->error)) |
1929 | return; | |
f8b0bd10 | 1930 | |
f2bba182 RR |
1931 | /* |
1932 | * If there is a non_fatal error, then add to the error count | |
1933 | * and clear all the errors. | |
1934 | */ | |
1935 | update_error_count(td, icd->error); | |
1936 | td_clear_error(td); | |
1937 | icd->error = 0; | |
f8b0bd10 JA |
1938 | if (io_u) |
1939 | io_u->error = 0; | |
f2bba182 | 1940 | } |
10ba535a JA |
1941 | } |
1942 | ||
9520ebb9 JA |
1943 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1944 | int nr) | |
10ba535a | 1945 | { |
6eaf09d6 | 1946 | int ddir; |
aba6c951 JA |
1947 | |
1948 | if (!gtod_reduce(td)) | |
9520ebb9 | 1949 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1950 | |
3af6ef39 JA |
1951 | icd->nr = nr; |
1952 | ||
10ba535a | 1953 | icd->error = 0; |
c1f50f76 | 1954 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 1955 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
1956 | } |
1957 | ||
97601024 JA |
1958 | static void ios_completed(struct thread_data *td, |
1959 | struct io_completion_data *icd) | |
36167d82 JA |
1960 | { |
1961 | struct io_u *io_u; | |
1962 | int i; | |
1963 | ||
10ba535a JA |
1964 | for (i = 0; i < icd->nr; i++) { |
1965 | io_u = td->io_ops->event(td, i); | |
1966 | ||
f8b0bd10 | 1967 | io_completed(td, &io_u, icd); |
e8462bd8 | 1968 | |
f8b0bd10 | 1969 | if (io_u) |
e8462bd8 | 1970 | put_io_u(td, io_u); |
10ba535a JA |
1971 | } |
1972 | } | |
97601024 | 1973 | |
e7e6cfb4 JA |
1974 | /* |
1975 | * Complete a single io_u for the sync engines. | |
1976 | */ | |
55312f9f | 1977 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1978 | { |
1979 | struct io_completion_data icd; | |
55312f9f | 1980 | int ddir; |
97601024 | 1981 | |
9520ebb9 | 1982 | init_icd(td, &icd, 1); |
f8b0bd10 | 1983 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1984 | |
f8b0bd10 | 1985 | if (io_u) |
e8462bd8 | 1986 | put_io_u(td, io_u); |
97601024 | 1987 | |
581e7141 JA |
1988 | if (icd.error) { |
1989 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1990 | return -1; | |
1991 | } | |
97601024 | 1992 | |
c1f50f76 | 1993 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 1994 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
1995 | |
1996 | return 0; | |
97601024 JA |
1997 | } |
1998 | ||
e7e6cfb4 JA |
1999 | /* |
2000 | * Called to complete min_events number of io for the async engines. | |
2001 | */ | |
55312f9f | 2002 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2003 | { |
97601024 | 2004 | struct io_completion_data icd; |
00de55ef | 2005 | struct timespec *tvp = NULL; |
55312f9f | 2006 | int ret, ddir; |
4d06a338 | 2007 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2008 | |
12bb8569 | 2009 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2010 | |
4950421a | 2011 | if (!min_evts) |
00de55ef | 2012 | tvp = &ts; |
5fb4b366 RE |
2013 | else if (min_evts > td->cur_depth) |
2014 | min_evts = td->cur_depth; | |
97601024 | 2015 | |
82407585 RP |
2016 | /* No worries, td_io_getevents fixes min and max if they are |
2017 | * set incorrectly */ | |
2018 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2019 | if (ret < 0) { |
e1161c32 | 2020 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2021 | return ret; |
2022 | } else if (!ret) | |
2023 | return ret; | |
2024 | ||
9520ebb9 | 2025 | init_icd(td, &icd, ret); |
97601024 | 2026 | ios_completed(td, &icd); |
581e7141 JA |
2027 | if (icd.error) { |
2028 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2029 | return -1; | |
2030 | } | |
97601024 | 2031 | |
c1f50f76 | 2032 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2033 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2034 | |
0d593542 | 2035 | return ret; |
97601024 | 2036 | } |
7e77dd02 JA |
2037 | |
2038 | /* | |
2039 | * Call when io_u is really queued, to update the submission latency. | |
2040 | */ | |
2041 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2042 | { | |
8243be59 | 2043 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2044 | unsigned long slat_time; |
7e77dd02 | 2045 | |
d6bb626e | 2046 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2047 | |
2048 | if (td->parent) | |
2049 | td = td->parent; | |
2050 | ||
ae588852 JA |
2051 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
2052 | io_u->offset); | |
9520ebb9 | 2053 | } |
7e77dd02 | 2054 | } |
433afcb4 | 2055 | |
5c94b008 JA |
2056 | /* |
2057 | * See if we should reuse the last seed, if dedupe is enabled | |
2058 | */ | |
9451b93e | 2059 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2060 | { |
2061 | unsigned int v; | |
5c94b008 JA |
2062 | |
2063 | if (!td->o.dedupe_percentage) | |
2064 | return &td->buf_state; | |
732eedd0 | 2065 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2066 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2067 | return &td->buf_state; | |
732eedd0 | 2068 | } |
5c94b008 | 2069 | |
1bd5d213 | 2070 | v = rand_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2071 | |
2072 | if (v <= td->o.dedupe_percentage) | |
2073 | return &td->buf_state_prev; | |
2074 | ||
2075 | return &td->buf_state; | |
2076 | } | |
2077 | ||
9451b93e | 2078 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2079 | { |
9451b93e JA |
2080 | if (td->o.dedupe_percentage == 100) |
2081 | frand_copy(rs, &td->buf_state_prev); | |
2082 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2083 | frand_copy(&td->buf_state_prev, rs); |
2084 | } | |
2085 | ||
5fff9543 JF |
2086 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write, |
2087 | unsigned long long max_bs) | |
5973cafb | 2088 | { |
d1af2894 JA |
2089 | struct thread_options *o = &td->o; |
2090 | ||
15600335 JA |
2091 | if (o->mem_type == MEM_CUDA_MALLOC) |
2092 | return; | |
03553853 | 2093 | |
4eff3e57 | 2094 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2095 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2096 | struct frand_state *rs; |
5fff9543 JF |
2097 | unsigned long long left = max_bs; |
2098 | unsigned long long this_write; | |
5c94b008 | 2099 | |
1066358a | 2100 | do { |
9451b93e | 2101 | rs = get_buf_state(td); |
9c42684e | 2102 | |
1066358a | 2103 | min_write = min(min_write, left); |
f97a43a1 | 2104 | |
1066358a | 2105 | if (perc) { |
4b157ac6 | 2106 | this_write = min_not_zero(min_write, |
5fff9543 | 2107 | (unsigned long long) td->o.compress_chunk); |
1e7f82e2 JA |
2108 | |
2109 | fill_random_buf_percentage(rs, buf, perc, | |
2110 | this_write, this_write, | |
2111 | o->buffer_pattern, | |
2112 | o->buffer_pattern_bytes); | |
2113 | } else { | |
1066358a | 2114 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2115 | this_write = min_write; |
2116 | } | |
1066358a | 2117 | |
1e7f82e2 JA |
2118 | buf += this_write; |
2119 | left -= this_write; | |
9451b93e | 2120 | save_buf_state(td, rs); |
1066358a | 2121 | } while (left); |
d1af2894 JA |
2122 | } else if (o->buffer_pattern_bytes) |
2123 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2124 | else if (o->zero_buffers) |
cc86c395 | 2125 | memset(buf, 0, max_bs); |
999d245e | 2126 | else |
9451b93e | 2127 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2128 | } |
2129 | ||
2130 | /* | |
2131 | * "randomly" fill the buffer contents | |
2132 | */ | |
2133 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
5fff9543 | 2134 | unsigned long long min_write, unsigned long long max_bs) |
cc86c395 JA |
2135 | { |
2136 | io_u->buf_filled_len = 0; | |
2137 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2138 | } |
e2c75fc4 TK |
2139 | |
2140 | static int do_sync_file_range(const struct thread_data *td, | |
2141 | struct fio_file *f) | |
2142 | { | |
2143 | off64_t offset, nbytes; | |
2144 | ||
2145 | offset = f->first_write; | |
2146 | nbytes = f->last_write - f->first_write; | |
2147 | ||
2148 | if (!nbytes) | |
2149 | return 0; | |
2150 | ||
2151 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2152 | } | |
2153 | ||
2154 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2155 | { | |
2156 | int ret; | |
2157 | ||
2158 | if (io_u->ddir == DDIR_SYNC) { | |
2159 | ret = fsync(io_u->file->fd); | |
2160 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2161 | #ifdef CONFIG_FDATASYNC | |
2162 | ret = fdatasync(io_u->file->fd); | |
2163 | #else | |
2164 | ret = io_u->xfer_buflen; | |
2165 | io_u->error = EINVAL; | |
2166 | #endif | |
2167 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2168 | ret = do_sync_file_range(td, io_u->file); | |
2169 | else { | |
2170 | ret = io_u->xfer_buflen; | |
2171 | io_u->error = EINVAL; | |
2172 | } | |
2173 | ||
2174 | if (ret < 0) | |
2175 | io_u->error = errno; | |
2176 | ||
2177 | return ret; | |
2178 | } | |
2179 | ||
2180 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2181 | { | |
2182 | #ifndef FIO_HAVE_TRIM | |
2183 | io_u->error = EINVAL; | |
2184 | return 0; | |
2185 | #else | |
2186 | struct fio_file *f = io_u->file; | |
2187 | int ret; | |
2188 | ||
496b1f9e | 2189 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2190 | if (!ret) |
2191 | return io_u->xfer_buflen; | |
2192 | ||
2193 | io_u->error = ret; | |
2194 | return 0; | |
2195 | #endif | |
2196 | } |