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" |
bfbdd35b | 13 | #include "zbd.h" |
10ba535a | 14 | |
97601024 JA |
15 | struct io_completion_data { |
16 | int nr; /* input */ | |
97601024 JA |
17 | |
18 | int error; /* output */ | |
100f49f1 | 19 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
8b6a404c | 20 | struct timespec time; /* output */ |
97601024 JA |
21 | }; |
22 | ||
10ba535a | 23 | /* |
7ebd796f | 24 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
25 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
26 | */ | |
e39c0676 | 27 | static bool random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 28 | { |
7ebd796f | 29 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
30 | } |
31 | ||
df415585 JA |
32 | /* |
33 | * Mark a given offset as used in the map. | |
34 | */ | |
6cc1a3d1 BVA |
35 | static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u, |
36 | uint64_t offset, uint64_t buflen) | |
df415585 | 37 | { |
5fff9543 | 38 | unsigned long long min_bs = td->o.min_bs[io_u->ddir]; |
9bf2061e | 39 | struct fio_file *f = io_u->file; |
5fff9543 | 40 | unsigned long long nr_blocks; |
1ae83d45 | 41 | uint64_t block; |
df415585 | 42 | |
6cc1a3d1 BVA |
43 | block = (offset - f->file_offset) / (uint64_t) min_bs; |
44 | nr_blocks = (buflen + min_bs - 1) / min_bs; | |
bd6b959a | 45 | assert(nr_blocks > 0); |
df415585 | 46 | |
bd6b959a | 47 | if (!(io_u->flags & IO_U_F_BUSY_OK)) { |
7ebd796f | 48 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
bd6b959a BVA |
49 | assert(nr_blocks > 0); |
50 | } | |
df415585 | 51 | |
6cc1a3d1 BVA |
52 | if ((nr_blocks * min_bs) < buflen) |
53 | buflen = nr_blocks * min_bs; | |
54 | ||
55 | return buflen; | |
df415585 JA |
56 | } |
57 | ||
74776733 JA |
58 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
59 | enum fio_ddir ddir) | |
2ba1c290 | 60 | { |
74776733 JA |
61 | uint64_t max_blocks; |
62 | uint64_t max_size; | |
2ba1c290 | 63 | |
ff58fced JA |
64 | assert(ddir_rw(ddir)); |
65 | ||
d9dd70f7 JA |
66 | /* |
67 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
79591fa9 | 68 | * -> not for now since there is code assuming it could go either. |
d9dd70f7 JA |
69 | */ |
70 | max_size = f->io_size; | |
71 | if (max_size > f->real_file_size) | |
72 | max_size = f->real_file_size; | |
73 | ||
7b865a2f | 74 | if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range) |
ed335855 SN |
75 | max_size = td->o.zone_range; |
76 | ||
0412d12e JE |
77 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
78 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
79 | ||
1ae83d45 | 80 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
81 | if (!max_blocks) |
82 | return 0; | |
83 | ||
67778e88 | 84 | return max_blocks; |
2ba1c290 JA |
85 | } |
86 | ||
e25839d4 | 87 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
e0a04ac1 JA |
88 | enum fio_ddir ddir, uint64_t *b, |
89 | uint64_t lastb) | |
ec4015da | 90 | { |
6f49f8bc | 91 | uint64_t r; |
5e0baa7f | 92 | |
c3546b53 JA |
93 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
94 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { | |
6f49f8bc | 95 | |
d6b72507 | 96 | r = __rand(&td->random_state); |
8055e41d | 97 | |
4b91ee8f | 98 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 99 | |
e0a04ac1 | 100 | *b = lastb * (r / (rand_max(&td->random_state) + 1.0)); |
51ede0b1 | 101 | } else { |
8055e41d | 102 | uint64_t off = 0; |
43c63a78 | 103 | |
967d1b63 JA |
104 | assert(fio_file_lfsr(f)); |
105 | ||
6f49f8bc | 106 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 107 | return 1; |
ec4015da | 108 | |
8055e41d JA |
109 | *b = off; |
110 | } | |
0ce8b119 | 111 | |
ec4015da | 112 | /* |
51ede0b1 | 113 | * if we are not maintaining a random map, we are done. |
ec4015da | 114 | */ |
51ede0b1 JA |
115 | if (!file_randommap(td, f)) |
116 | goto ret; | |
43c63a78 JA |
117 | |
118 | /* | |
51ede0b1 | 119 | * calculate map offset and check if it's free |
43c63a78 | 120 | */ |
51ede0b1 JA |
121 | if (random_map_free(f, *b)) |
122 | goto ret; | |
123 | ||
4b91ee8f JA |
124 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
125 | (unsigned long long) *b); | |
51ede0b1 | 126 | |
7ebd796f | 127 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
128 | if (*b == (uint64_t) -1ULL) |
129 | return 1; | |
0ce8b119 JA |
130 | ret: |
131 | return 0; | |
ec4015da JA |
132 | } |
133 | ||
925fee33 JA |
134 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
135 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 136 | uint64_t *b) |
e25839d4 | 137 | { |
9c6f6316 | 138 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
139 | return 0; |
140 | } | |
141 | ||
925fee33 JA |
142 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
143 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 144 | uint64_t *b) |
925fee33 | 145 | { |
9c6f6316 | 146 | *b = pareto_next(&f->zipf); |
925fee33 JA |
147 | return 0; |
148 | } | |
149 | ||
56d9fa4b JA |
150 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
151 | struct fio_file *f, enum fio_ddir ddir, | |
152 | uint64_t *b) | |
153 | { | |
154 | *b = gauss_next(&f->gauss); | |
155 | return 0; | |
156 | } | |
157 | ||
59466396 JA |
158 | static int __get_next_rand_offset_zoned_abs(struct thread_data *td, |
159 | struct fio_file *f, | |
160 | enum fio_ddir ddir, uint64_t *b) | |
161 | { | |
162 | struct zone_split_index *zsi; | |
e3c8c108 | 163 | uint64_t lastb, send, stotal; |
59466396 JA |
164 | unsigned int v; |
165 | ||
166 | lastb = last_block(td, f, ddir); | |
167 | if (!lastb) | |
168 | return 1; | |
169 | ||
170 | if (!td->o.zone_split_nr[ddir]) { | |
171 | bail: | |
172 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
173 | } | |
174 | ||
175 | /* | |
176 | * Generate a value, v, between 1 and 100, both inclusive | |
177 | */ | |
1bd5d213 | 178 | v = rand_between(&td->zone_state, 1, 100); |
59466396 | 179 | |
e3c8c108 JA |
180 | /* |
181 | * Find our generated table. 'send' is the end block of this zone, | |
182 | * 'stotal' is our start offset. | |
183 | */ | |
59466396 JA |
184 | zsi = &td->zone_state_index[ddir][v - 1]; |
185 | stotal = zsi->size_prev / td->o.ba[ddir]; | |
186 | send = zsi->size / td->o.ba[ddir]; | |
187 | ||
188 | /* | |
189 | * Should never happen | |
190 | */ | |
191 | if (send == -1U) { | |
264e3d30 | 192 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
59466396 | 193 | log_err("fio: bug in zoned generation\n"); |
59466396 JA |
194 | goto bail; |
195 | } else if (send > lastb) { | |
196 | /* | |
197 | * This happens if the user specifies ranges that exceed | |
198 | * the file/device size. We can't handle that gracefully, | |
199 | * so error and exit. | |
200 | */ | |
201 | log_err("fio: zoned_abs sizes exceed file size\n"); | |
202 | return 1; | |
203 | } | |
204 | ||
205 | /* | |
e3c8c108 | 206 | * Generate index from 0..send-stotal |
59466396 | 207 | */ |
e3c8c108 | 208 | if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1) |
59466396 JA |
209 | return 1; |
210 | ||
e3c8c108 | 211 | *b += stotal; |
59466396 JA |
212 | return 0; |
213 | } | |
214 | ||
e0a04ac1 JA |
215 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
216 | struct fio_file *f, enum fio_ddir ddir, | |
217 | uint64_t *b) | |
218 | { | |
219 | unsigned int v, send, stotal; | |
220 | uint64_t offset, lastb; | |
e0a04ac1 JA |
221 | struct zone_split_index *zsi; |
222 | ||
223 | lastb = last_block(td, f, ddir); | |
224 | if (!lastb) | |
225 | return 1; | |
226 | ||
227 | if (!td->o.zone_split_nr[ddir]) { | |
228 | bail: | |
229 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
230 | } | |
231 | ||
232 | /* | |
233 | * Generate a value, v, between 1 and 100, both inclusive | |
234 | */ | |
1bd5d213 | 235 | v = rand_between(&td->zone_state, 1, 100); |
e0a04ac1 JA |
236 | |
237 | zsi = &td->zone_state_index[ddir][v - 1]; | |
238 | stotal = zsi->size_perc_prev; | |
239 | send = zsi->size_perc; | |
240 | ||
241 | /* | |
242 | * Should never happen | |
243 | */ | |
244 | if (send == -1U) { | |
264e3d30 | 245 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
e0a04ac1 | 246 | log_err("fio: bug in zoned generation\n"); |
e0a04ac1 JA |
247 | goto bail; |
248 | } | |
249 | ||
250 | /* | |
251 | * 'send' is some percentage below or equal to 100 that | |
252 | * marks the end of the current IO range. 'stotal' marks | |
253 | * the start, in percent. | |
254 | */ | |
255 | if (stotal) | |
256 | offset = stotal * lastb / 100ULL; | |
257 | else | |
258 | offset = 0; | |
259 | ||
260 | lastb = lastb * (send - stotal) / 100ULL; | |
261 | ||
262 | /* | |
263 | * Generate index from 0..send-of-lastb | |
264 | */ | |
265 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
266 | return 1; | |
267 | ||
268 | /* | |
269 | * Add our start offset, if any | |
270 | */ | |
271 | if (offset) | |
272 | *b += offset; | |
273 | ||
274 | return 0; | |
275 | } | |
56d9fa4b | 276 | |
f31feaa2 JA |
277 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
278 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 | 279 | { |
e0a04ac1 JA |
280 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
281 | uint64_t lastb; | |
282 | ||
283 | lastb = last_block(td, f, ddir); | |
284 | if (!lastb) | |
285 | return 1; | |
286 | ||
287 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
288 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
e25839d4 | 289 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
925fee33 JA |
290 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
291 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
292 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
293 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e0a04ac1 JA |
294 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
295 | return __get_next_rand_offset_zoned(td, f, ddir, b); | |
59466396 JA |
296 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS) |
297 | return __get_next_rand_offset_zoned_abs(td, f, ddir, b); | |
e25839d4 JA |
298 | |
299 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
300 | return 1; | |
301 | } | |
302 | ||
e39c0676 | 303 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
304 | { |
305 | unsigned int v; | |
211c9b89 | 306 | |
d9472271 | 307 | if (td->o.perc_rand[ddir] == 100) |
e39c0676 | 308 | return true; |
211c9b89 | 309 | |
1bd5d213 | 310 | v = rand_between(&td->seq_rand_state[ddir], 1, 100); |
211c9b89 | 311 | |
d9472271 | 312 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
313 | } |
314 | ||
0bcf41cd JA |
315 | static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f) |
316 | { | |
317 | struct thread_options *o = &td->o; | |
318 | ||
319 | if (o->invalidate_cache && !o->odirect) { | |
320 | int fio_unused ret; | |
321 | ||
322 | ret = file_invalidate_cache(td, f); | |
323 | } | |
324 | } | |
325 | ||
38dad62d | 326 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 327 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 328 | { |
c04e4661 DE |
329 | if (!get_next_rand_offset(td, f, ddir, b)) |
330 | return 0; | |
331 | ||
8c07860d JA |
332 | if (td->o.time_based || |
333 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { | |
33c48814 | 334 | fio_file_reset(td, f); |
80f02150 | 335 | loop_cache_invalidate(td, f); |
c04e4661 DE |
336 | if (!get_next_rand_offset(td, f, ddir, b)) |
337 | return 0; | |
38dad62d JA |
338 | } |
339 | ||
c04e4661 | 340 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 341 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 342 | (unsigned long long) f->real_file_size); |
c04e4661 | 343 | return 1; |
38dad62d JA |
344 | } |
345 | ||
37cf9e3c | 346 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 347 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 348 | { |
ac002339 JA |
349 | struct thread_options *o = &td->o; |
350 | ||
ff58fced JA |
351 | assert(ddir_rw(ddir)); |
352 | ||
17373ce2 JA |
353 | /* |
354 | * If we reach the end for a time based run, reset us back to 0 | |
355 | * and invalidate the cache, if we need to. | |
356 | */ | |
f1dfb668 | 357 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
19ddc35b | 358 | o->time_based) { |
c89daa4a | 359 | f->last_pos[ddir] = f->file_offset; |
0bcf41cd | 360 | loop_cache_invalidate(td, f); |
19ddc35b | 361 | } |
c04e4661 | 362 | |
f1dfb668 | 363 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 364 | uint64_t pos; |
059b0802 | 365 | |
69b98f11 JA |
366 | /* |
367 | * Only rewind if we already hit the end | |
368 | */ | |
369 | if (f->last_pos[ddir] == f->file_offset && | |
370 | f->file_offset && o->ddir_seq_add < 0) { | |
c22825bb JA |
371 | if (f->real_file_size > f->io_size) |
372 | f->last_pos[ddir] = f->io_size; | |
373 | else | |
374 | f->last_pos[ddir] = f->real_file_size; | |
375 | } | |
a66da7a2 | 376 | |
f1dfb668 | 377 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
378 | if (pos && o->ddir_seq_add) { |
379 | pos += o->ddir_seq_add; | |
380 | ||
381 | /* | |
382 | * If we reach beyond the end of the file | |
383 | * with holed IO, wrap around to the | |
b0a84f48 JA |
384 | * beginning again. If we're doing backwards IO, |
385 | * wrap to the end. | |
ac002339 | 386 | */ |
b0a84f48 JA |
387 | if (pos >= f->real_file_size) { |
388 | if (o->ddir_seq_add > 0) | |
389 | pos = f->file_offset; | |
c22825bb JA |
390 | else { |
391 | if (f->real_file_size > f->io_size) | |
392 | pos = f->io_size; | |
393 | else | |
394 | pos = f->real_file_size; | |
395 | ||
396 | pos += o->ddir_seq_add; | |
397 | } | |
b0a84f48 | 398 | } |
ac002339 | 399 | } |
059b0802 | 400 | |
37cf9e3c | 401 | *offset = pos; |
38dad62d JA |
402 | return 0; |
403 | } | |
404 | ||
405 | return 1; | |
406 | } | |
407 | ||
408 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d | 409 | enum fio_ddir ddir, int rw_seq, |
ec370c22 | 410 | bool *is_random) |
38dad62d JA |
411 | { |
412 | struct fio_file *f = io_u->file; | |
1ae83d45 | 413 | uint64_t b, offset; |
38dad62d JA |
414 | int ret; |
415 | ||
ff58fced JA |
416 | assert(ddir_rw(ddir)); |
417 | ||
37cf9e3c JA |
418 | b = offset = -1ULL; |
419 | ||
38dad62d | 420 | if (rw_seq) { |
211c9b89 | 421 | if (td_random(td)) { |
6aca9b3d | 422 | if (should_do_random(td, ddir)) { |
211c9b89 | 423 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 424 | *is_random = true; |
6aca9b3d | 425 | } else { |
ec370c22 | 426 | *is_random = false; |
1651e431 | 427 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
428 | ret = get_next_seq_offset(td, f, ddir, &offset); |
429 | if (ret) | |
430 | ret = get_next_rand_block(td, f, ddir, &b); | |
431 | } | |
6aca9b3d | 432 | } else { |
ec370c22 | 433 | *is_random = false; |
37cf9e3c | 434 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 435 | } |
38dad62d | 436 | } else { |
1651e431 | 437 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
ec370c22 | 438 | *is_random = false; |
38dad62d JA |
439 | |
440 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 441 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 442 | if (ret) { |
37cf9e3c | 443 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 444 | *is_random = false; |
6aca9b3d | 445 | } |
38dad62d | 446 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
447 | if (f->last_start[ddir] != -1ULL) |
448 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 449 | else |
37cf9e3c | 450 | offset = 0; |
38dad62d JA |
451 | ret = 0; |
452 | } else { | |
453 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
454 | ret = 1; | |
455 | } | |
456 | } | |
6d68b997 | 457 | |
37cf9e3c JA |
458 | if (!ret) { |
459 | if (offset != -1ULL) | |
460 | io_u->offset = offset; | |
461 | else if (b != -1ULL) | |
462 | io_u->offset = b * td->o.ba[ddir]; | |
463 | else { | |
4e0a8fa2 | 464 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
465 | ret = 1; |
466 | } | |
4fff54cc | 467 | io_u->verify_offset = io_u->offset; |
37cf9e3c JA |
468 | } |
469 | ||
38dad62d JA |
470 | return ret; |
471 | } | |
472 | ||
10ba535a JA |
473 | /* |
474 | * For random io, generate a random new block and see if it's used. Repeat | |
475 | * until we find a free one. For sequential io, just return the end of | |
476 | * the last io issued. | |
477 | */ | |
e8fb335e | 478 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 479 | bool *is_random) |
10ba535a | 480 | { |
9bf2061e | 481 | struct fio_file *f = io_u->file; |
4ba66134 | 482 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 483 | int rw_seq_hit = 0; |
10ba535a | 484 | |
ff58fced JA |
485 | assert(ddir_rw(ddir)); |
486 | ||
38dad62d JA |
487 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
488 | rw_seq_hit = 1; | |
5736c10d | 489 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 490 | } |
211097b2 | 491 | |
6aca9b3d | 492 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 493 | return 1; |
10ba535a | 494 | |
009bd847 JA |
495 | if (io_u->offset >= f->io_size) { |
496 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
497 | (unsigned long long) io_u->offset, |
498 | (unsigned long long) f->io_size); | |
009bd847 JA |
499 | return 1; |
500 | } | |
501 | ||
502 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
503 | if (io_u->offset >= f->real_file_size) { |
504 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
505 | (unsigned long long) io_u->offset, |
506 | (unsigned long long) f->real_file_size); | |
10ba535a | 507 | return 1; |
2ba1c290 | 508 | } |
10ba535a JA |
509 | |
510 | return 0; | |
511 | } | |
512 | ||
e39c0676 | 513 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
5fff9543 | 514 | unsigned long long buflen) |
79944128 JA |
515 | { |
516 | struct fio_file *f = io_u->file; | |
517 | ||
bedc9dc2 | 518 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
519 | } |
520 | ||
5fff9543 | 521 | static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 522 | bool is_random) |
10ba535a | 523 | { |
6aca9b3d | 524 | int ddir = io_u->ddir; |
5fff9543 JF |
525 | unsigned long long buflen = 0; |
526 | unsigned long long minbs, maxbs; | |
3dd29f7c | 527 | uint64_t frand_max, r; |
7c961359 | 528 | bool power_2; |
10ba535a | 529 | |
9ee1c647 | 530 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
531 | |
532 | if (td->o.bs_is_seq_rand) | |
ec370c22 | 533 | ddir = is_random ? DDIR_WRITE : DDIR_READ; |
ff58fced | 534 | |
f3059de1 JA |
535 | minbs = td->o.min_bs[ddir]; |
536 | maxbs = td->o.max_bs[ddir]; | |
537 | ||
79944128 JA |
538 | if (minbs == maxbs) |
539 | return minbs; | |
540 | ||
52c58027 JA |
541 | /* |
542 | * If we can't satisfy the min block size from here, then fail | |
543 | */ | |
544 | if (!io_u_fits(td, io_u, minbs)) | |
545 | return 0; | |
546 | ||
2f282cec | 547 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 548 | do { |
2f282cec | 549 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 550 | |
720e84ad | 551 | if (!td->o.bssplit_nr[ddir]) { |
5fff9543 | 552 | buflen = minbs + (unsigned long long) ((double) maxbs * |
c3546b53 | 553 | (r / (frand_max + 1.0))); |
5ec10eaa | 554 | } else { |
3dd29f7c | 555 | long long perc = 0; |
564ca972 JA |
556 | unsigned int i; |
557 | ||
720e84ad JA |
558 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
559 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 | 560 | |
cc5e4cb0 DLM |
561 | if (!bsp->perc) |
562 | continue; | |
564ca972 JA |
563 | buflen = bsp->bs; |
564 | perc += bsp->perc; | |
3dd29f7c | 565 | if ((r / perc <= frand_max / 100ULL) && |
79944128 | 566 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
567 | break; |
568 | } | |
569 | } | |
79944128 | 570 | |
17a6b702 PL |
571 | power_2 = is_power_of_2(minbs); |
572 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 573 | buflen &= ~(minbs - 1); |
8135fe4d JA |
574 | else if (!td->o.bs_unaligned && !power_2) |
575 | buflen -= buflen % minbs; | |
576 | if (buflen > maxbs) | |
577 | buflen = maxbs; | |
79944128 | 578 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 579 | |
10ba535a JA |
580 | return buflen; |
581 | } | |
582 | ||
afe24a5a JA |
583 | static void set_rwmix_bytes(struct thread_data *td) |
584 | { | |
afe24a5a JA |
585 | unsigned int diff; |
586 | ||
587 | /* | |
588 | * we do time or byte based switch. this is needed because | |
589 | * buffered writes may issue a lot quicker than they complete, | |
590 | * whereas reads do not. | |
591 | */ | |
e47f799f | 592 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 593 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
594 | } |
595 | ||
596 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
597 | { | |
598 | unsigned int v; | |
e47f799f | 599 | |
1bd5d213 | 600 | v = rand_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 601 | |
04c540d9 | 602 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
603 | return DDIR_READ; |
604 | ||
605 | return DDIR_WRITE; | |
afe24a5a JA |
606 | } |
607 | ||
0d593542 | 608 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 609 | { |
2dcfdeb9 | 610 | int ret = 0, completed = 0, err = 0; |
0d593542 | 611 | |
002e7183 JA |
612 | /* |
613 | * We are going to sleep, ensure that we flush anything pending as | |
614 | * not to skew our latency numbers. | |
615 | * | |
616 | * Changed to only monitor 'in flight' requests here instead of the | |
617 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
618 | * io's that have been actually submitted to an async engine, | |
619 | * and cur_depth is meaningless for sync engines. | |
620 | */ | |
a80cb54b BVA |
621 | if (td->io_u_queued || td->cur_depth) |
622 | td_io_commit(td); | |
9cc80b6d | 623 | |
002e7183 | 624 | while (td->io_u_in_flight) { |
55312f9f | 625 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
626 | if (ret > 0) |
627 | completed += ret; | |
d28174f0 | 628 | else if (ret < 0) |
2dcfdeb9 | 629 | err = ret; |
002e7183 | 630 | } |
0d593542 | 631 | |
6be06c46 JA |
632 | if (td->flags & TD_F_REGROW_LOGS) |
633 | regrow_logs(td); | |
634 | ||
d28174f0 JA |
635 | if (completed) |
636 | return completed; | |
637 | ||
2dcfdeb9 | 638 | return err; |
002e7183 JA |
639 | } |
640 | ||
581e7141 JA |
641 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
642 | { | |
643 | enum fio_ddir odir = ddir ^ 1; | |
b5407f8b | 644 | uint64_t usec; |
90eff1c9 | 645 | uint64_t now; |
581e7141 | 646 | |
ff58fced | 647 | assert(ddir_rw(ddir)); |
970d9000 | 648 | now = utime_since_now(&td->epoch); |
ff58fced | 649 | |
50a8ce86 D |
650 | /* |
651 | * if rate_next_io_time is in the past, need to catch up to rate | |
652 | */ | |
653 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
654 | return ddir; |
655 | ||
656 | /* | |
50a8ce86 | 657 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
658 | * should switch. |
659 | */ | |
315fcfec | 660 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 661 | /* |
50a8ce86 | 662 | * Other direction is behind rate, switch |
581e7141 | 663 | */ |
50a8ce86 | 664 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
665 | return odir; |
666 | ||
667 | /* | |
395feabb JA |
668 | * Both directions are ahead of rate. sleep the min, |
669 | * switch if necessary | |
581e7141 | 670 | */ |
50a8ce86 | 671 | if (td->rate_next_io_time[ddir] <= |
395feabb | 672 | td->rate_next_io_time[odir]) { |
50a8ce86 | 673 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 674 | } else { |
50a8ce86 | 675 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
676 | ddir = odir; |
677 | } | |
678 | } else | |
50a8ce86 | 679 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 680 | |
a9da8ab2 JA |
681 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
682 | io_u_quiesce(td); | |
78c1eda5 | 683 | |
df06a036 | 684 | if (td->o.timeout && ((usec + now) > td->o.timeout)) { |
685 | /* | |
686 | * check if the usec is capable of taking negative values | |
687 | */ | |
688 | if (now > td->o.timeout) { | |
689 | ddir = DDIR_INVAL; | |
690 | return ddir; | |
691 | } | |
692 | usec = td->o.timeout - now; | |
693 | } | |
1a9bf814 | 694 | usec_sleep(td, usec); |
df06a036 | 695 | |
696 | now = utime_since_now(&td->epoch); | |
697 | if ((td->o.timeout && (now > td->o.timeout)) || td->terminate) | |
698 | ddir = DDIR_INVAL; | |
699 | ||
581e7141 JA |
700 | return ddir; |
701 | } | |
702 | ||
10ba535a JA |
703 | /* |
704 | * Return the data direction for the next io_u. If the job is a | |
705 | * mixed read/write workload, check the rwmix cycle and switch if | |
706 | * necessary. | |
707 | */ | |
1e97cce9 | 708 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 709 | { |
581e7141 JA |
710 | enum fio_ddir ddir; |
711 | ||
5f9099ea | 712 | /* |
f6860972 TK |
713 | * See if it's time to fsync/fdatasync/sync_file_range first, |
714 | * and if not then move on to check regular I/Os. | |
5f9099ea | 715 | */ |
f6860972 TK |
716 | if (should_fsync(td)) { |
717 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
718 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
719 | return DDIR_SYNC; | |
720 | ||
721 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
722 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
723 | return DDIR_DATASYNC; | |
724 | ||
725 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
726 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
727 | return DDIR_SYNC_FILE_RANGE; | |
728 | } | |
44f29692 | 729 | |
10ba535a | 730 | if (td_rw(td)) { |
10ba535a JA |
731 | /* |
732 | * Check if it's time to seed a new data direction. | |
733 | */ | |
e4928662 | 734 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
735 | /* |
736 | * Put a top limit on how many bytes we do for | |
737 | * one data direction, to avoid overflowing the | |
738 | * ranges too much | |
739 | */ | |
740 | ddir = get_rand_ddir(td); | |
e47f799f JA |
741 | |
742 | if (ddir != td->rwmix_ddir) | |
743 | set_rwmix_bytes(td); | |
744 | ||
745 | td->rwmix_ddir = ddir; | |
10ba535a | 746 | } |
581e7141 | 747 | ddir = td->rwmix_ddir; |
10ba535a | 748 | } else if (td_read(td)) |
581e7141 | 749 | ddir = DDIR_READ; |
6eaf09d6 | 750 | else if (td_write(td)) |
581e7141 | 751 | ddir = DDIR_WRITE; |
5c8e84ca | 752 | else if (td_trim(td)) |
6eaf09d6 | 753 | ddir = DDIR_TRIM; |
5c8e84ca TK |
754 | else |
755 | ddir = DDIR_INVAL; | |
581e7141 JA |
756 | |
757 | td->rwmix_ddir = rate_ddir(td, ddir); | |
758 | return td->rwmix_ddir; | |
10ba535a JA |
759 | } |
760 | ||
1ef2b6be JA |
761 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
762 | { | |
0e4dd95c DE |
763 | enum fio_ddir ddir = get_rw_ddir(td); |
764 | ||
731461cc DLM |
765 | if (td->o.zone_mode == ZONE_MODE_ZBD) |
766 | ddir = zbd_adjust_ddir(td, io_u, ddir); | |
c65057f9 | 767 | |
82a90686 | 768 | if (td_trimwrite(td)) { |
0e4dd95c DE |
769 | struct fio_file *f = io_u->file; |
770 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
771 | ddir = DDIR_TRIM; | |
772 | else | |
773 | ddir = DDIR_WRITE; | |
774 | } | |
775 | ||
776 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 777 | |
9b87f09b | 778 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
779 | td->o.barrier_blocks && |
780 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
781 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 782 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
783 | } |
784 | ||
e8462bd8 | 785 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 786 | { |
71b84caa | 787 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
788 | |
789 | if (ret) | |
790 | td_verror(td, ret, "file close"); | |
791 | } | |
792 | ||
10ba535a JA |
793 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
794 | { | |
26b3a188 JA |
795 | const bool needs_lock = td_async_processing(td); |
796 | ||
d9ed3e63 | 797 | zbd_put_io_u(io_u); |
99952ca7 | 798 | |
a9da8ab2 JA |
799 | if (td->parent) |
800 | td = td->parent; | |
801 | ||
26b3a188 JA |
802 | if (needs_lock) |
803 | __td_io_u_lock(td); | |
e8462bd8 | 804 | |
f8b0bd10 | 805 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 806 | put_file_log(td, io_u->file); |
f8b0bd10 | 807 | |
10ba535a | 808 | io_u->file = NULL; |
1651e431 | 809 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 810 | |
a9da8ab2 | 811 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 812 | td->cur_depth--; |
a9da8ab2 JA |
813 | assert(!(td->flags & TD_F_CHILD)); |
814 | } | |
2ae0b204 | 815 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 | 816 | td_io_u_free_notify(td); |
26b3a188 JA |
817 | |
818 | if (needs_lock) | |
819 | __td_io_u_unlock(td); | |
10ba535a JA |
820 | } |
821 | ||
f2bba182 RR |
822 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
823 | { | |
1651e431 | 824 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
825 | put_io_u(td, io_u); |
826 | } | |
827 | ||
755200a3 JA |
828 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
829 | { | |
26b3a188 | 830 | const bool needs_lock = td_async_processing(td); |
755200a3 | 831 | struct io_u *__io_u = *io_u; |
bcd5abfa | 832 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 833 | |
465221b0 JA |
834 | dprint(FD_IO, "requeue %p\n", __io_u); |
835 | ||
a9da8ab2 JA |
836 | if (td->parent) |
837 | td = td->parent; | |
838 | ||
26b3a188 JA |
839 | if (needs_lock) |
840 | __td_io_u_lock(td); | |
e8462bd8 | 841 | |
1651e431 | 842 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
843 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
844 | td->io_issues[ddir]--; | |
5ec10eaa | 845 | |
1651e431 | 846 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 847 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 848 | td->cur_depth--; |
a9da8ab2 JA |
849 | assert(!(td->flags & TD_F_CHILD)); |
850 | } | |
2ae0b204 JA |
851 | |
852 | io_u_rpush(&td->io_u_requeues, __io_u); | |
a9da8ab2 | 853 | td_io_u_free_notify(td); |
26b3a188 JA |
854 | |
855 | if (needs_lock) | |
856 | __td_io_u_unlock(td); | |
857 | ||
755200a3 JA |
858 | *io_u = NULL; |
859 | } | |
860 | ||
7b865a2f | 861 | static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u) |
224b3093 | 862 | { |
863 | struct fio_file *f = io_u->file; | |
864 | ||
7b865a2f BVA |
865 | assert(td->o.zone_mode == ZONE_MODE_STRIDED); |
866 | assert(td->o.zone_size); | |
867 | assert(td->o.zone_range); | |
868 | ||
224b3093 | 869 | /* |
870 | * See if it's time to switch to a new zone | |
871 | */ | |
fa9fd914 | 872 | if (td->zone_bytes >= td->o.zone_size) { |
224b3093 | 873 | td->zone_bytes = 0; |
874 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
875 | ||
876 | /* | |
877 | * Wrap from the beginning, if we exceed the file size | |
878 | */ | |
879 | if (f->file_offset >= f->real_file_size) | |
04bc85a1 JA |
880 | f->file_offset = get_start_offset(td, f); |
881 | ||
224b3093 | 882 | f->last_pos[io_u->ddir] = f->file_offset; |
883 | td->io_skip_bytes += td->o.zone_skip; | |
884 | } | |
885 | ||
886 | /* | |
04bc85a1 JA |
887 | * If zone_size > zone_range, then maintain the same zone until |
888 | * zone_bytes >= zone_size. | |
889 | */ | |
224b3093 | 890 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
891 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
892 | f->file_offset, f->last_pos[io_u->ddir]); | |
893 | f->last_pos[io_u->ddir] = f->file_offset; | |
894 | } | |
895 | ||
896 | /* | |
897 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
898 | * map needs to be reset as it's done with zone_range everytime. | |
899 | */ | |
04bc85a1 | 900 | if ((td->zone_bytes % td->o.zone_range) == 0) |
224b3093 | 901 | fio_file_reset(td, f); |
224b3093 | 902 | } |
903 | ||
9bf2061e | 904 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 905 | { |
ec370c22 | 906 | bool is_random; |
bfbdd35b BVA |
907 | uint64_t offset; |
908 | enum io_u_action ret; | |
6aca9b3d | 909 | |
9b87f09b | 910 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
911 | goto out; |
912 | ||
1ef2b6be | 913 | set_rw_ddir(td, io_u); |
5f9099ea | 914 | |
df06a036 | 915 | if (io_u->ddir == DDIR_INVAL) { |
916 | dprint(FD_IO, "invalid direction received ddir = %d", io_u->ddir); | |
917 | return 1; | |
918 | } | |
87dc1ab1 | 919 | /* |
ff58fced | 920 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 921 | */ |
ff58fced | 922 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 923 | goto out; |
a00735e6 | 924 | |
7b865a2f BVA |
925 | if (td->o.zone_mode == ZONE_MODE_STRIDED) |
926 | setup_strided_zone_mode(td, io_u); | |
4d37720a DLM |
927 | else if (td->o.zone_mode == ZONE_MODE_ZBD) |
928 | setup_zbd_zone_mode(td, io_u); | |
48f5abd3 | 929 | |
10ba535a | 930 | /* |
c685b5b2 JA |
931 | * No log, let the seq/rand engine retrieve the next buflen and |
932 | * position. | |
10ba535a | 933 | */ |
6aca9b3d | 934 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 935 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 936 | return 1; |
2ba1c290 | 937 | } |
10ba535a | 938 | |
6aca9b3d | 939 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
940 | if (!io_u->buflen) { |
941 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 942 | return 1; |
2ba1c290 | 943 | } |
bca4ed4d | 944 | |
bfbdd35b BVA |
945 | offset = io_u->offset; |
946 | if (td->o.zone_mode == ZONE_MODE_ZBD) { | |
947 | ret = zbd_adjust_block(td, io_u); | |
948 | if (ret == io_u_eof) | |
949 | return 1; | |
950 | } | |
951 | ||
2ba1c290 | 952 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
5fff9543 | 953 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n", |
e5f9a813 | 954 | io_u, |
4b91ee8f JA |
955 | (unsigned long long) io_u->offset, io_u->buflen, |
956 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 957 | return 1; |
2ba1c290 | 958 | } |
6a5e6884 | 959 | |
bca4ed4d JA |
960 | /* |
961 | * mark entry before potentially trimming io_u | |
962 | */ | |
303032ae | 963 | if (td_random(td) && file_randommap(td, io_u->file)) |
bfbdd35b | 964 | io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen); |
bca4ed4d | 965 | |
c38e9468 | 966 | out: |
e5f9a813 | 967 | dprint_io_u(io_u, "fill"); |
d9d91e39 | 968 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 969 | return 0; |
10ba535a JA |
970 | } |
971 | ||
6cc0e5aa | 972 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
838bc709 | 973 | { |
2b13e716 | 974 | int idx = 0; |
838bc709 JA |
975 | |
976 | switch (nr) { | |
977 | default: | |
2b13e716 | 978 | idx = 6; |
838bc709 JA |
979 | break; |
980 | case 33 ... 64: | |
2b13e716 | 981 | idx = 5; |
838bc709 JA |
982 | break; |
983 | case 17 ... 32: | |
2b13e716 | 984 | idx = 4; |
838bc709 JA |
985 | break; |
986 | case 9 ... 16: | |
2b13e716 | 987 | idx = 3; |
838bc709 JA |
988 | break; |
989 | case 5 ... 8: | |
2b13e716 | 990 | idx = 2; |
838bc709 JA |
991 | break; |
992 | case 1 ... 4: | |
2b13e716 | 993 | idx = 1; |
838bc709 JA |
994 | case 0: |
995 | break; | |
996 | } | |
997 | ||
2b13e716 | 998 | map[idx]++; |
838bc709 JA |
999 | } |
1000 | ||
1001 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
1002 | { | |
1003 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
1004 | td->ts.total_submit++; | |
1005 | } | |
1006 | ||
1007 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
1008 | { | |
1009 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
1010 | td->ts.total_complete++; | |
1011 | } | |
1012 | ||
d8005759 | 1013 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 1014 | { |
2b13e716 | 1015 | int idx = 0; |
71619dc2 JA |
1016 | |
1017 | switch (td->cur_depth) { | |
1018 | default: | |
2b13e716 | 1019 | idx = 6; |
a783e61a | 1020 | break; |
71619dc2 | 1021 | case 32 ... 63: |
2b13e716 | 1022 | idx = 5; |
a783e61a | 1023 | break; |
71619dc2 | 1024 | case 16 ... 31: |
2b13e716 | 1025 | idx = 4; |
a783e61a | 1026 | break; |
71619dc2 | 1027 | case 8 ... 15: |
2b13e716 | 1028 | idx = 3; |
a783e61a | 1029 | break; |
71619dc2 | 1030 | case 4 ... 7: |
2b13e716 | 1031 | idx = 2; |
a783e61a | 1032 | break; |
71619dc2 | 1033 | case 2 ... 3: |
2b13e716 | 1034 | idx = 1; |
71619dc2 JA |
1035 | case 1: |
1036 | break; | |
1037 | } | |
1038 | ||
2b13e716 | 1039 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
1040 | } |
1041 | ||
d6bb626e | 1042 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1043 | { |
2b13e716 | 1044 | int idx = 0; |
04a0feae | 1045 | |
d6bb626e VF |
1046 | assert(nsec < 1000); |
1047 | ||
1048 | switch (nsec) { | |
1049 | case 750 ... 999: | |
1050 | idx = 9; | |
1051 | break; | |
1052 | case 500 ... 749: | |
1053 | idx = 8; | |
1054 | break; | |
1055 | case 250 ... 499: | |
1056 | idx = 7; | |
1057 | break; | |
1058 | case 100 ... 249: | |
1059 | idx = 6; | |
1060 | break; | |
1061 | case 50 ... 99: | |
1062 | idx = 5; | |
1063 | break; | |
1064 | case 20 ... 49: | |
1065 | idx = 4; | |
1066 | break; | |
1067 | case 10 ... 19: | |
1068 | idx = 3; | |
1069 | break; | |
1070 | case 4 ... 9: | |
1071 | idx = 2; | |
1072 | break; | |
1073 | case 2 ... 3: | |
1074 | idx = 1; | |
1075 | case 0 ... 1: | |
1076 | break; | |
1077 | } | |
1078 | ||
1079 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1080 | td->ts.io_u_lat_n[idx]++; | |
1081 | } | |
1082 | ||
1083 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1084 | { | |
1085 | int idx = 0; | |
1086 | ||
1087 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1088 | |
1089 | switch (usec) { | |
1090 | case 750 ... 999: | |
2b13e716 | 1091 | idx = 9; |
04a0feae JA |
1092 | break; |
1093 | case 500 ... 749: | |
2b13e716 | 1094 | idx = 8; |
04a0feae JA |
1095 | break; |
1096 | case 250 ... 499: | |
2b13e716 | 1097 | idx = 7; |
04a0feae JA |
1098 | break; |
1099 | case 100 ... 249: | |
2b13e716 | 1100 | idx = 6; |
04a0feae JA |
1101 | break; |
1102 | case 50 ... 99: | |
2b13e716 | 1103 | idx = 5; |
04a0feae JA |
1104 | break; |
1105 | case 20 ... 49: | |
2b13e716 | 1106 | idx = 4; |
04a0feae JA |
1107 | break; |
1108 | case 10 ... 19: | |
2b13e716 | 1109 | idx = 3; |
04a0feae JA |
1110 | break; |
1111 | case 4 ... 9: | |
2b13e716 | 1112 | idx = 2; |
04a0feae JA |
1113 | break; |
1114 | case 2 ... 3: | |
2b13e716 | 1115 | idx = 1; |
04a0feae JA |
1116 | case 0 ... 1: |
1117 | break; | |
1118 | } | |
1119 | ||
2b13e716 JA |
1120 | assert(idx < FIO_IO_U_LAT_U_NR); |
1121 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1122 | } |
1123 | ||
d6bb626e | 1124 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1125 | { |
2b13e716 | 1126 | int idx = 0; |
ec118304 | 1127 | |
d6bb626e VF |
1128 | assert(msec >= 1); |
1129 | ||
ec118304 JA |
1130 | switch (msec) { |
1131 | default: | |
2b13e716 | 1132 | idx = 11; |
04a0feae | 1133 | break; |
8abdce66 | 1134 | case 1000 ... 1999: |
2b13e716 | 1135 | idx = 10; |
04a0feae | 1136 | break; |
8abdce66 | 1137 | case 750 ... 999: |
2b13e716 | 1138 | idx = 9; |
04a0feae | 1139 | break; |
8abdce66 | 1140 | case 500 ... 749: |
2b13e716 | 1141 | idx = 8; |
04a0feae | 1142 | break; |
8abdce66 | 1143 | case 250 ... 499: |
2b13e716 | 1144 | idx = 7; |
04a0feae | 1145 | break; |
8abdce66 | 1146 | case 100 ... 249: |
2b13e716 | 1147 | idx = 6; |
04a0feae | 1148 | break; |
8abdce66 | 1149 | case 50 ... 99: |
2b13e716 | 1150 | idx = 5; |
04a0feae | 1151 | break; |
8abdce66 | 1152 | case 20 ... 49: |
2b13e716 | 1153 | idx = 4; |
04a0feae | 1154 | break; |
8abdce66 | 1155 | case 10 ... 19: |
2b13e716 | 1156 | idx = 3; |
04a0feae | 1157 | break; |
8abdce66 | 1158 | case 4 ... 9: |
2b13e716 | 1159 | idx = 2; |
04a0feae | 1160 | break; |
ec118304 | 1161 | case 2 ... 3: |
2b13e716 | 1162 | idx = 1; |
ec118304 JA |
1163 | case 0 ... 1: |
1164 | break; | |
1165 | } | |
1166 | ||
2b13e716 JA |
1167 | assert(idx < FIO_IO_U_LAT_M_NR); |
1168 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1169 | } |
1170 | ||
d6bb626e | 1171 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1172 | { |
d6bb626e VF |
1173 | if (nsec < 1000) |
1174 | io_u_mark_lat_nsec(td, nsec); | |
1175 | else if (nsec < 1000000) | |
1176 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1177 | else |
d6bb626e | 1178 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1179 | } |
1180 | ||
8c07860d JA |
1181 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1182 | { | |
1183 | unsigned long fileno; | |
1184 | ||
1185 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1186 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1187 | unsigned long r; | |
1188 | ||
1189 | r = __rand(&td->next_file_state); | |
1190 | return (unsigned int) ((double) td->o.nr_files | |
1191 | * (r / (frand_max + 1.0))); | |
1192 | } | |
1193 | ||
1194 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1195 | fileno = zipf_next(&td->next_file_zipf); | |
1196 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1197 | fileno = pareto_next(&td->next_file_zipf); | |
1198 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1199 | fileno = gauss_next(&td->next_file_gauss); | |
1200 | else { | |
1201 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1202 | assert(0); | |
1203 | return 0; | |
1204 | } | |
1205 | ||
1206 | return fileno >> FIO_FSERVICE_SHIFT; | |
1207 | } | |
1208 | ||
0aabe160 JA |
1209 | /* |
1210 | * Get next file to service by choosing one at random | |
1211 | */ | |
2cc52930 JA |
1212 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1213 | enum fio_file_flags goodf, | |
d6aed795 | 1214 | enum fio_file_flags badf) |
0aabe160 | 1215 | { |
0aabe160 | 1216 | struct fio_file *f; |
1c178180 | 1217 | int fno; |
0aabe160 JA |
1218 | |
1219 | do { | |
87b10676 | 1220 | int opened = 0; |
4c07ad86 | 1221 | |
8c07860d | 1222 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1223 | |
126d65c6 | 1224 | f = td->files[fno]; |
d6aed795 | 1225 | if (fio_file_done(f)) |
059e63c0 | 1226 | continue; |
1c178180 | 1227 | |
d6aed795 | 1228 | if (!fio_file_open(f)) { |
87b10676 JA |
1229 | int err; |
1230 | ||
002fe734 JA |
1231 | if (td->nr_open_files >= td->o.open_files) |
1232 | return ERR_PTR(-EBUSY); | |
1233 | ||
87b10676 JA |
1234 | err = td_io_open_file(td, f); |
1235 | if (err) | |
1236 | continue; | |
1237 | opened = 1; | |
1238 | } | |
1239 | ||
2ba1c290 JA |
1240 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1241 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1242 | return f; |
2ba1c290 | 1243 | } |
87b10676 JA |
1244 | if (opened) |
1245 | td_io_close_file(td, f); | |
0aabe160 JA |
1246 | } while (1); |
1247 | } | |
1248 | ||
1249 | /* | |
1250 | * Get next file to service by doing round robin between all available ones | |
1251 | */ | |
1c178180 JA |
1252 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1253 | int badf) | |
3d7c391d JA |
1254 | { |
1255 | unsigned int old_next_file = td->next_file; | |
1256 | struct fio_file *f; | |
1257 | ||
1258 | do { | |
87b10676 JA |
1259 | int opened = 0; |
1260 | ||
126d65c6 | 1261 | f = td->files[td->next_file]; |
3d7c391d JA |
1262 | |
1263 | td->next_file++; | |
2dc1bbeb | 1264 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1265 | td->next_file = 0; |
1266 | ||
87b10676 | 1267 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1268 | if (fio_file_done(f)) { |
d5ed68ea | 1269 | f = NULL; |
059e63c0 | 1270 | continue; |
d5ed68ea | 1271 | } |
059e63c0 | 1272 | |
d6aed795 | 1273 | if (!fio_file_open(f)) { |
87b10676 JA |
1274 | int err; |
1275 | ||
002fe734 JA |
1276 | if (td->nr_open_files >= td->o.open_files) |
1277 | return ERR_PTR(-EBUSY); | |
1278 | ||
87b10676 | 1279 | err = td_io_open_file(td, f); |
b5696bfc JA |
1280 | if (err) { |
1281 | dprint(FD_FILE, "error %d on open of %s\n", | |
1282 | err, f->file_name); | |
87c27b45 | 1283 | f = NULL; |
87b10676 | 1284 | continue; |
b5696bfc | 1285 | } |
87b10676 JA |
1286 | opened = 1; |
1287 | } | |
1288 | ||
0b9d69ec JA |
1289 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1290 | f->flags); | |
1c178180 | 1291 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1292 | break; |
1293 | ||
87b10676 JA |
1294 | if (opened) |
1295 | td_io_close_file(td, f); | |
1296 | ||
3d7c391d JA |
1297 | f = NULL; |
1298 | } while (td->next_file != old_next_file); | |
1299 | ||
2ba1c290 | 1300 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1301 | return f; |
1302 | } | |
1303 | ||
7eb36574 | 1304 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1305 | { |
1907dbc6 JA |
1306 | struct fio_file *f; |
1307 | ||
2dc1bbeb | 1308 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1309 | |
b5696bfc | 1310 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1311 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1312 | " nr_files=%d\n", td->nr_open_files, | |
1313 | td->nr_done_files, | |
1314 | td->o.nr_files); | |
bdb4e2e9 | 1315 | return NULL; |
2ba1c290 | 1316 | } |
bdb4e2e9 | 1317 | |
1907dbc6 | 1318 | f = td->file_service_file; |
d6aed795 | 1319 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1320 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1321 | goto out; | |
1322 | if (td->file_service_left--) | |
1323 | goto out; | |
1324 | } | |
1907dbc6 | 1325 | |
a086c257 JA |
1326 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1327 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1328 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1329 | else |
d6aed795 | 1330 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1331 | |
002fe734 JA |
1332 | if (IS_ERR(f)) |
1333 | return f; | |
1334 | ||
1907dbc6 JA |
1335 | td->file_service_file = f; |
1336 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1337 | out: |
0dac421f JA |
1338 | if (f) |
1339 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1340 | else | |
1341 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1342 | return f; |
bdb4e2e9 JA |
1343 | } |
1344 | ||
7eb36574 JA |
1345 | static struct fio_file *get_next_file(struct thread_data *td) |
1346 | { | |
7eb36574 JA |
1347 | return __get_next_file(td); |
1348 | } | |
1349 | ||
002fe734 | 1350 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1351 | { |
1352 | struct fio_file *f; | |
1353 | ||
1354 | do { | |
1355 | f = get_next_file(td); | |
002fe734 JA |
1356 | if (IS_ERR_OR_NULL(f)) |
1357 | return PTR_ERR(f); | |
429f6675 | 1358 | |
429f6675 JA |
1359 | io_u->file = f; |
1360 | get_file(f); | |
1361 | ||
1362 | if (!fill_io_u(td, io_u)) | |
1363 | break; | |
1364 | ||
d9ed3e63 | 1365 | zbd_put_io_u(io_u); |
99952ca7 | 1366 | |
b5696bfc | 1367 | put_file_log(td, f); |
429f6675 | 1368 | td_io_close_file(td, f); |
b5696bfc | 1369 | io_u->file = NULL; |
8c07860d JA |
1370 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1371 | fio_file_reset(td, f); | |
1372 | else { | |
1373 | fio_file_set_done(f); | |
1374 | td->nr_done_files++; | |
1375 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1376 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1377 | } |
429f6675 JA |
1378 | } while (1); |
1379 | ||
1380 | return 0; | |
1381 | } | |
1382 | ||
3e260a46 | 1383 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
c3a32714 | 1384 | unsigned long long tnsec, unsigned long long max_nsec) |
3e260a46 JA |
1385 | { |
1386 | if (!td->error) | |
c3a32714 | 1387 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
3e260a46 JA |
1388 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
1389 | icd->error = ETIMEDOUT; | |
1390 | } | |
1391 | ||
1392 | static void lat_new_cycle(struct thread_data *td) | |
1393 | { | |
1394 | fio_gettime(&td->latency_ts, NULL); | |
1395 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1396 | td->latency_failed = 0; | |
1397 | } | |
1398 | ||
1399 | /* | |
1400 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1401 | * are at QD=1, then it's time to give up. | |
1402 | */ | |
e39c0676 | 1403 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1404 | { |
1405 | if (td->latency_qd == 1) | |
e39c0676 | 1406 | return true; |
3e260a46 JA |
1407 | |
1408 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1409 | |
1410 | if (td->latency_qd == td->latency_qd_low) | |
1411 | td->latency_qd_low--; | |
1412 | ||
3e260a46 | 1413 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
e1bcd541 | 1414 | td->latency_stable_count = 0; |
3e260a46 JA |
1415 | |
1416 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1417 | ||
1418 | /* | |
1419 | * When we ramp QD down, quiesce existing IO to prevent | |
1420 | * a storm of ramp downs due to pending higher depth. | |
1421 | */ | |
1422 | io_u_quiesce(td); | |
1423 | lat_new_cycle(td); | |
e39c0676 | 1424 | return false; |
3e260a46 JA |
1425 | } |
1426 | ||
e39c0676 | 1427 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1428 | { |
1429 | if (td->o.latency_percentile.u.f == 100.0) | |
1430 | return __lat_target_failed(td); | |
1431 | ||
1432 | td->latency_failed++; | |
e39c0676 | 1433 | return false; |
3e260a46 JA |
1434 | } |
1435 | ||
1436 | void lat_target_init(struct thread_data *td) | |
1437 | { | |
6bb58215 JA |
1438 | td->latency_end_run = 0; |
1439 | ||
3e260a46 JA |
1440 | if (td->o.latency_target) { |
1441 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1442 | fio_gettime(&td->latency_ts, NULL); | |
1443 | td->latency_qd = 1; | |
1444 | td->latency_qd_high = td->o.iodepth; | |
1445 | td->latency_qd_low = 1; | |
1446 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1447 | } else | |
1448 | td->latency_qd = td->o.iodepth; | |
1449 | } | |
1450 | ||
6bb58215 JA |
1451 | void lat_target_reset(struct thread_data *td) |
1452 | { | |
1453 | if (!td->latency_end_run) | |
1454 | lat_target_init(td); | |
1455 | } | |
1456 | ||
3e260a46 JA |
1457 | static void lat_target_success(struct thread_data *td) |
1458 | { | |
1459 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1460 | struct thread_options *o = &td->o; |
3e260a46 JA |
1461 | |
1462 | td->latency_qd_low = td->latency_qd; | |
1463 | ||
e1bcd541 SL |
1464 | if (td->latency_qd + 1 == td->latency_qd_high) { |
1465 | /* | |
1466 | * latency_qd will not incease on lat_target_success(), so | |
1467 | * called stable. If we stick with this queue depth, the | |
1468 | * final latency is likely lower than latency_target. Fix | |
1469 | * this by increasing latency_qd_high slowly. Use a naive | |
1470 | * heuristic here. If we get lat_target_success() 3 times | |
1471 | * in a row, increase latency_qd_high by 1. | |
1472 | */ | |
1473 | if (++td->latency_stable_count >= 3) { | |
1474 | td->latency_qd_high++; | |
1475 | td->latency_stable_count = 0; | |
1476 | } | |
1477 | } | |
1478 | ||
3e260a46 JA |
1479 | /* |
1480 | * If we haven't failed yet, we double up to a failing value instead | |
1481 | * of bisecting from highest possible queue depth. If we have set | |
1482 | * a limit other than td->o.iodepth, bisect between that. | |
1483 | */ | |
6bb58215 | 1484 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1485 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1486 | else | |
1487 | td->latency_qd *= 2; | |
1488 | ||
6bb58215 JA |
1489 | if (td->latency_qd > o->iodepth) |
1490 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1491 | |
1492 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1493 | |
3e260a46 | 1494 | /* |
6bb58215 JA |
1495 | * Same as last one, we are done. Let it run a latency cycle, so |
1496 | * we get only the results from the targeted depth. | |
3e260a46 | 1497 | */ |
e1bcd541 | 1498 | if (!o->latency_run && td->latency_qd == qd) { |
6bb58215 JA |
1499 | if (td->latency_end_run) { |
1500 | dprint(FD_RATE, "We are done\n"); | |
1501 | td->done = 1; | |
1502 | } else { | |
1503 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1504 | io_u_quiesce(td); | |
1505 | td->latency_end_run = 1; | |
1506 | reset_all_stats(td); | |
1507 | reset_io_stats(td); | |
1508 | } | |
1509 | } | |
3e260a46 JA |
1510 | |
1511 | lat_new_cycle(td); | |
1512 | } | |
1513 | ||
1514 | /* | |
1515 | * Check if we can bump the queue depth | |
1516 | */ | |
1517 | void lat_target_check(struct thread_data *td) | |
1518 | { | |
1519 | uint64_t usec_window; | |
1520 | uint64_t ios; | |
1521 | double success_ios; | |
1522 | ||
1523 | usec_window = utime_since_now(&td->latency_ts); | |
1524 | if (usec_window < td->o.latency_window) | |
1525 | return; | |
1526 | ||
1527 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1528 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1529 | success_ios *= 100.0; | |
1530 | ||
1531 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1532 | ||
1533 | if (success_ios >= td->o.latency_percentile.u.f) | |
1534 | lat_target_success(td); | |
1535 | else | |
1536 | __lat_target_failed(td); | |
1537 | } | |
1538 | ||
1539 | /* | |
1540 | * If latency target is enabled, we might be ramping up or down and not | |
1541 | * using the full queue depth available. | |
1542 | */ | |
e39c0676 | 1543 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1544 | { |
1545 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1546 | ||
1547 | if (qempty) | |
e39c0676 | 1548 | return true; |
3e260a46 | 1549 | if (!td->o.latency_target) |
e39c0676 | 1550 | return false; |
3e260a46 JA |
1551 | |
1552 | return td->cur_depth >= td->latency_qd; | |
1553 | } | |
429f6675 | 1554 | |
10ba535a JA |
1555 | struct io_u *__get_io_u(struct thread_data *td) |
1556 | { | |
26b3a188 | 1557 | const bool needs_lock = td_async_processing(td); |
0cae66f6 | 1558 | struct io_u *io_u = NULL; |
93b45bb2 | 1559 | int ret; |
10ba535a | 1560 | |
ca09be4b JA |
1561 | if (td->stop_io) |
1562 | return NULL; | |
1563 | ||
26b3a188 JA |
1564 | if (needs_lock) |
1565 | __td_io_u_lock(td); | |
e8462bd8 JA |
1566 | |
1567 | again: | |
bba6b14f | 1568 | if (!io_u_rempty(&td->io_u_requeues)) { |
2ae0b204 | 1569 | io_u = io_u_rpop(&td->io_u_requeues); |
bba6b14f JA |
1570 | io_u->resid = 0; |
1571 | } else if (!queue_full(td)) { | |
2ae0b204 | 1572 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1573 | |
225ba9e3 | 1574 | io_u->file = NULL; |
6040dabc | 1575 | io_u->buflen = 0; |
10ba535a | 1576 | io_u->resid = 0; |
d7762cf8 | 1577 | io_u->end_io = NULL; |
755200a3 JA |
1578 | } |
1579 | ||
1580 | if (io_u) { | |
0c6e7517 | 1581 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1582 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 | 1583 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
b2a432bf | 1584 | IO_U_F_VER_LIST | IO_U_F_PRIORITY); |
0c6e7517 | 1585 | |
755200a3 | 1586 | io_u->error = 0; |
bcd5abfa | 1587 | io_u->acct_ddir = -1; |
10ba535a | 1588 | td->cur_depth++; |
a9da8ab2 | 1589 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1590 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1591 | io_u->ipo = NULL; |
a9da8ab2 | 1592 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1593 | /* |
1594 | * We ran out, wait for async verify threads to finish and | |
1595 | * return one | |
1596 | */ | |
a9da8ab2 | 1597 | assert(!(td->flags & TD_F_CHILD)); |
93b45bb2 BVA |
1598 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
1599 | assert(ret == 0); | |
ee5e08ad JA |
1600 | if (!td->error) |
1601 | goto again; | |
10ba535a JA |
1602 | } |
1603 | ||
26b3a188 JA |
1604 | if (needs_lock) |
1605 | __td_io_u_unlock(td); | |
1606 | ||
10ba535a JA |
1607 | return io_u; |
1608 | } | |
1609 | ||
e39c0676 | 1610 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1611 | { |
d72be545 | 1612 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1613 | return false; |
c9a73054 JA |
1614 | if (!td->trim_entries) |
1615 | return false; | |
d72be545 | 1616 | |
c9a73054 JA |
1617 | if (td->trim_batch) { |
1618 | td->trim_batch--; | |
1619 | if (get_next_trim(td, io_u)) | |
1620 | return true; | |
1621 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1622 | td->last_ddir != DDIR_READ) { | |
1623 | td->trim_batch = td->o.trim_batch; | |
1624 | if (!td->trim_batch) | |
1625 | td->trim_batch = td->o.trim_backlog; | |
1626 | if (get_next_trim(td, io_u)) | |
e39c0676 | 1627 | return true; |
2ba1c290 | 1628 | } |
10ba535a | 1629 | |
e39c0676 | 1630 | return false; |
0d29de83 JA |
1631 | } |
1632 | ||
e39c0676 | 1633 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1634 | { |
d72be545 | 1635 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1636 | return false; |
d72be545 JA |
1637 | |
1638 | if (td->io_hist_len) { | |
9e144189 JA |
1639 | int get_verify = 0; |
1640 | ||
d1ece0c7 | 1641 | if (td->verify_batch) |
9e144189 | 1642 | get_verify = 1; |
d1ece0c7 | 1643 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1644 | td->last_ddir != DDIR_READ) { |
1645 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1646 | if (!td->verify_batch) |
1647 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1648 | get_verify = 1; |
1649 | } | |
1650 | ||
d1ece0c7 JA |
1651 | if (get_verify && !get_next_verify(td, io_u)) { |
1652 | td->verify_batch--; | |
e39c0676 | 1653 | return true; |
d1ece0c7 | 1654 | } |
9e144189 JA |
1655 | } |
1656 | ||
e39c0676 | 1657 | return false; |
0d29de83 JA |
1658 | } |
1659 | ||
de789769 JA |
1660 | /* |
1661 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1662 | * easy compressible data for simple de-dupe attempts. Do this for every |
1663 | * 512b block in the range, since that should be the smallest block size | |
1664 | * we can expect from a device. | |
de789769 JA |
1665 | */ |
1666 | static void small_content_scramble(struct io_u *io_u) | |
1667 | { | |
5fff9543 | 1668 | unsigned long long i, nr_blocks = io_u->buflen >> 9; |
23f394d5 | 1669 | unsigned int offset; |
319b073f JA |
1670 | uint64_t boffset, *iptr; |
1671 | char *p; | |
de789769 | 1672 | |
23f394d5 JA |
1673 | if (!nr_blocks) |
1674 | return; | |
1675 | ||
1676 | p = io_u->xfer_buf; | |
fba76ee8 | 1677 | boffset = io_u->offset; |
319b073f JA |
1678 | |
1679 | if (io_u->buf_filled_len) | |
1680 | io_u->buf_filled_len = 0; | |
1681 | ||
1682 | /* | |
1683 | * Generate random index between 0..7. We do chunks of 512b, if | |
1684 | * we assume a cacheline is 64 bytes, then we have 8 of those. | |
1685 | * Scramble content within the blocks in the same cacheline to | |
1686 | * speed things up. | |
1687 | */ | |
1688 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; | |
fad82f76 | 1689 | |
23f394d5 JA |
1690 | for (i = 0; i < nr_blocks; i++) { |
1691 | /* | |
319b073f JA |
1692 | * Fill offset into start of cacheline, time into end |
1693 | * of cacheline | |
23f394d5 | 1694 | */ |
319b073f JA |
1695 | iptr = (void *) p + (offset << 6); |
1696 | *iptr = boffset; | |
1697 | ||
1698 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); | |
1699 | iptr[0] = io_u->start_time.tv_sec; | |
1700 | iptr[1] = io_u->start_time.tv_nsec; | |
23f394d5 | 1701 | |
23f394d5 | 1702 | p += 512; |
fad82f76 | 1703 | boffset += 512; |
23f394d5 | 1704 | } |
de789769 JA |
1705 | } |
1706 | ||
0d29de83 JA |
1707 | /* |
1708 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
5c5c33c1 | 1709 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
0d29de83 JA |
1710 | */ |
1711 | struct io_u *get_io_u(struct thread_data *td) | |
1712 | { | |
1713 | struct fio_file *f; | |
1714 | struct io_u *io_u; | |
de789769 | 1715 | int do_scramble = 0; |
002fe734 | 1716 | long ret = 0; |
0d29de83 JA |
1717 | |
1718 | io_u = __get_io_u(td); | |
1719 | if (!io_u) { | |
1720 | dprint(FD_IO, "__get_io_u failed\n"); | |
1721 | return NULL; | |
1722 | } | |
1723 | ||
1724 | if (check_get_verify(td, io_u)) | |
1725 | goto out; | |
1726 | if (check_get_trim(td, io_u)) | |
1727 | goto out; | |
1728 | ||
755200a3 JA |
1729 | /* |
1730 | * from a requeue, io_u already setup | |
1731 | */ | |
1732 | if (io_u->file) | |
77f392bf | 1733 | goto out; |
755200a3 | 1734 | |
429f6675 JA |
1735 | /* |
1736 | * If using an iolog, grab next piece if any available. | |
1737 | */ | |
d72be545 | 1738 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1739 | if (read_iolog_get(td, io_u)) |
1740 | goto err_put; | |
2ba1c290 | 1741 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1742 | ret = -EBUSY; |
2ba1c290 | 1743 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1744 | goto err_put; |
2ba1c290 | 1745 | } |
5ec10eaa | 1746 | |
429f6675 | 1747 | f = io_u->file; |
002fe734 JA |
1748 | if (!f) { |
1749 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1750 | goto err_put; | |
1751 | } | |
1752 | ||
d6aed795 | 1753 | assert(fio_file_open(f)); |
97af62ce | 1754 | |
ff58fced | 1755 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1756 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1757 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1758 | goto err_put; |
2ba1c290 | 1759 | } |
10ba535a | 1760 | |
f1dfb668 JA |
1761 | f->last_start[io_u->ddir] = io_u->offset; |
1762 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1763 | |
fd68418e | 1764 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1765 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1766 | io_u_fill_buffer(td, io_u, |
1066358a | 1767 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1768 | io_u->buflen); |
ff441ae8 | 1769 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
5c5c33c1 BVA |
1770 | !(td->flags & TD_F_COMPRESS) && |
1771 | !(td->flags & TD_F_DO_VERIFY)) | |
fd68418e JA |
1772 | do_scramble = 1; |
1773 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1774 | /* |
1775 | * Reset the buf_filled parameters so next time if the | |
1776 | * buffer is used for writes it is refilled. | |
1777 | */ | |
cbe8d756 RR |
1778 | io_u->buf_filled_len = 0; |
1779 | } | |
87dc1ab1 | 1780 | } |
10ba535a | 1781 | |
165faf16 JA |
1782 | /* |
1783 | * Set io data pointers. | |
1784 | */ | |
cec6b55d JA |
1785 | io_u->xfer_buf = io_u->buf; |
1786 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1787 | |
6ac7a331 | 1788 | out: |
0d29de83 | 1789 | assert(io_u->file); |
429f6675 | 1790 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1791 | if (!td->o.disable_lat) |
993bf48b | 1792 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1793 | |
de789769 JA |
1794 | if (do_scramble) |
1795 | small_content_scramble(io_u); | |
03553853 | 1796 | |
429f6675 | 1797 | return io_u; |
36167d82 | 1798 | } |
429f6675 | 1799 | err_put: |
2ba1c290 | 1800 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1801 | put_io_u(td, io_u); |
002fe734 | 1802 | return ERR_PTR(ret); |
10ba535a JA |
1803 | } |
1804 | ||
a9da8ab2 | 1805 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1806 | { |
8b28bd41 | 1807 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1808 | |
8b28bd41 DM |
1809 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1810 | return; | |
5451792e | 1811 | |
5fff9543 | 1812 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n", |
709c8313 RE |
1813 | io_u->file ? " on file " : "", |
1814 | io_u->file ? io_u->file->file_name : "", | |
1815 | strerror(io_u->error), | |
1816 | io_ddir_name(io_u->ddir), | |
1817 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1818 | |
5ad7be56 KD |
1819 | if (td->io_ops->errdetails) { |
1820 | char *err = td->io_ops->errdetails(io_u); | |
1821 | ||
1822 | log_err("fio: %s\n", err); | |
1823 | free(err); | |
1824 | } | |
1825 | ||
5451792e JA |
1826 | if (!td->error) |
1827 | td_verror(td, io_u->error, "io_u error"); | |
1828 | } | |
1829 | ||
a9da8ab2 JA |
1830 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1831 | { | |
1832 | __io_u_log_error(td, io_u); | |
1833 | if (td->parent) | |
094e66cb | 1834 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1835 | } |
1836 | ||
e39c0676 | 1837 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1838 | { |
3ecc8c67 JA |
1839 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1840 | || td->o.gtod_reduce; | |
aba6c951 JA |
1841 | } |
1842 | ||
d7e92306 JA |
1843 | static void trim_block_info(struct thread_data *td, struct io_u *io_u) |
1844 | { | |
1845 | uint32_t *info = io_u_block_info(td, io_u); | |
1846 | ||
1847 | if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE) | |
1848 | return; | |
1849 | ||
1850 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1); | |
1851 | } | |
1852 | ||
c8eeb9df JA |
1853 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1854 | struct io_completion_data *icd, | |
1855 | const enum fio_ddir idx, unsigned int bytes) | |
1856 | { | |
a9da8ab2 | 1857 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1858 | unsigned long long llnsec = 0; |
c8eeb9df | 1859 | |
75dc383e JA |
1860 | if (td->parent) |
1861 | td = td->parent; | |
1862 | ||
132b1ee4 | 1863 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1864 | return; |
1865 | ||
a9da8ab2 | 1866 | if (no_reduce) |
d6bb626e | 1867 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1868 | |
1869 | if (!td->o.disable_lat) { | |
c3a32714 | 1870 | unsigned long long tnsec; |
c8eeb9df | 1871 | |
d6bb626e | 1872 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
b2a432bf | 1873 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset, io_u_is_prio(io_u)); |
15501535 | 1874 | |
d4afedfd JA |
1875 | if (td->flags & TD_F_PROFILE_OPS) { |
1876 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1877 | ||
1878 | if (ops->io_u_lat) | |
c3a32714 | 1879 | icd->error = ops->io_u_lat(td, tnsec); |
d4afedfd JA |
1880 | } |
1881 | ||
c3a32714 JA |
1882 | if (td->o.max_latency && tnsec > td->o.max_latency) |
1883 | lat_fatal(td, icd, tnsec, td->o.max_latency); | |
1884 | if (td->o.latency_target && tnsec > td->o.latency_target) { | |
3e260a46 | 1885 | if (lat_target_failed(td)) |
c3a32714 | 1886 | lat_fatal(td, icd, tnsec, td->o.latency_target); |
15501535 | 1887 | } |
c8eeb9df JA |
1888 | } |
1889 | ||
a47591e4 JA |
1890 | if (ddir_rw(idx)) { |
1891 | if (!td->o.disable_clat) { | |
b2a432bf | 1892 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset, io_u_is_prio(io_u)); |
d6bb626e | 1893 | io_u_mark_latency(td, llnsec); |
a47591e4 | 1894 | } |
c8eeb9df | 1895 | |
a47591e4 | 1896 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1897 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1898 | |
a47591e4 JA |
1899 | if (no_reduce && per_unit_log(td->iops_log)) |
1900 | add_iops_sample(td, io_u, bytes); | |
b2b3eefe JA |
1901 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
1902 | add_sync_clat_sample(&td->ts, llnsec); | |
66347cfa | 1903 | |
d7e92306 JA |
1904 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) |
1905 | trim_block_info(td, io_u); | |
c8eeb9df JA |
1906 | } |
1907 | ||
94a6e1bb JA |
1908 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1909 | uint64_t offset, unsigned int bytes) | |
1910 | { | |
1911 | int idx; | |
1912 | ||
639ad1ea JA |
1913 | if (!f) |
1914 | return; | |
1915 | ||
94a6e1bb JA |
1916 | if (f->first_write == -1ULL || offset < f->first_write) |
1917 | f->first_write = offset; | |
1918 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1919 | f->last_write = offset + bytes; | |
1920 | ||
1921 | if (!f->last_write_comp) | |
1922 | return; | |
1923 | ||
1924 | idx = f->last_write_idx++; | |
1925 | f->last_write_comp[idx] = offset; | |
1926 | if (f->last_write_idx == td->o.iodepth) | |
1927 | f->last_write_idx = 0; | |
1928 | } | |
1929 | ||
b2b3eefe JA |
1930 | static bool should_account(struct thread_data *td) |
1931 | { | |
1932 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || | |
1933 | td->runstate == TD_VERIFYING); | |
1934 | } | |
1935 | ||
f8b0bd10 | 1936 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1937 | struct io_completion_data *icd) |
10ba535a | 1938 | { |
f8b0bd10 JA |
1939 | struct io_u *io_u = *io_u_ptr; |
1940 | enum fio_ddir ddir = io_u->ddir; | |
1941 | struct fio_file *f = io_u->file; | |
10ba535a | 1942 | |
e5f9a813 | 1943 | dprint_io_u(io_u, "complete"); |
2ba1c290 | 1944 | |
0c6e7517 | 1945 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1946 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1947 | |
1948 | /* | |
1949 | * Mark IO ok to verify | |
1950 | */ | |
1951 | if (io_u->ipo) { | |
890b6656 JA |
1952 | /* |
1953 | * Remove errored entry from the verification list | |
1954 | */ | |
1955 | if (io_u->error) | |
1956 | unlog_io_piece(td, io_u); | |
1957 | else { | |
0337173e BVA |
1958 | atomic_store_release(&io_u->ipo->flags, |
1959 | io_u->ipo->flags & ~IP_F_IN_FLIGHT); | |
890b6656 | 1960 | } |
f9401285 JA |
1961 | } |
1962 | ||
f8b0bd10 | 1963 | if (ddir_sync(ddir)) { |
2ea93f98 | 1964 | td->last_was_sync = true; |
44f29692 JA |
1965 | if (f) { |
1966 | f->first_write = -1ULL; | |
1967 | f->last_write = -1ULL; | |
1968 | } | |
b2b3eefe JA |
1969 | if (should_account(td)) |
1970 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); | |
87dc1ab1 JA |
1971 | return; |
1972 | } | |
1973 | ||
2ea93f98 | 1974 | td->last_was_sync = false; |
f8b0bd10 | 1975 | td->last_ddir = ddir; |
87dc1ab1 | 1976 | |
f8b0bd10 | 1977 | if (!io_u->error && ddir_rw(ddir)) { |
5fff9543 | 1978 | unsigned long long bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1979 | int ret; |
10ba535a | 1980 | |
7d33649c JA |
1981 | /* |
1982 | * Make sure we notice short IO from here, and requeue them | |
1983 | * appropriately! | |
1984 | */ | |
1985 | if (io_u->resid) { | |
1986 | io_u->xfer_buflen = io_u->resid; | |
1987 | io_u->xfer_buf += bytes; | |
1988 | io_u->offset += bytes; | |
1989 | td->ts.short_io_u[io_u->ddir]++; | |
1990 | if (io_u->offset < io_u->file->real_file_size) { | |
1991 | requeue_io_u(td, io_u_ptr); | |
1992 | return; | |
1993 | } | |
1994 | } | |
1995 | ||
f8b0bd10 | 1996 | td->io_blocks[ddir]++; |
f8b0bd10 | 1997 | td->io_bytes[ddir] += bytes; |
ae2fafc8 | 1998 | |
e1c325d2 JA |
1999 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
2000 | td->this_io_blocks[ddir]++; | |
f8b0bd10 | 2001 | td->this_io_bytes[ddir] += bytes; |
e1c325d2 | 2002 | } |
f8b0bd10 | 2003 | |
639ad1ea | 2004 | if (ddir == DDIR_WRITE) |
94a6e1bb | 2005 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 2006 | |
b2b3eefe | 2007 | if (should_account(td)) |
f8b0bd10 | 2008 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 2009 | |
f8b0bd10 | 2010 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 2011 | |
d7762cf8 | 2012 | if (io_u->end_io) { |
f8b0bd10 JA |
2013 | ret = io_u->end_io(td, io_u_ptr); |
2014 | io_u = *io_u_ptr; | |
3af6ef39 JA |
2015 | if (ret && !icd->error) |
2016 | icd->error = ret; | |
2017 | } | |
ff58fced | 2018 | } else if (io_u->error) { |
10ba535a | 2019 | icd->error = io_u->error; |
5451792e JA |
2020 | io_u_log_error(td, io_u); |
2021 | } | |
8b28bd41 | 2022 | if (icd->error) { |
f8b0bd10 JA |
2023 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
2024 | ||
8b28bd41 DM |
2025 | if (!td_non_fatal_error(td, eb, icd->error)) |
2026 | return; | |
f8b0bd10 | 2027 | |
f2bba182 RR |
2028 | /* |
2029 | * If there is a non_fatal error, then add to the error count | |
2030 | * and clear all the errors. | |
2031 | */ | |
2032 | update_error_count(td, icd->error); | |
2033 | td_clear_error(td); | |
2034 | icd->error = 0; | |
f8b0bd10 JA |
2035 | if (io_u) |
2036 | io_u->error = 0; | |
f2bba182 | 2037 | } |
10ba535a JA |
2038 | } |
2039 | ||
9520ebb9 JA |
2040 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
2041 | int nr) | |
10ba535a | 2042 | { |
6eaf09d6 | 2043 | int ddir; |
aba6c951 JA |
2044 | |
2045 | if (!gtod_reduce(td)) | |
9520ebb9 | 2046 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 2047 | |
3af6ef39 JA |
2048 | icd->nr = nr; |
2049 | ||
10ba535a | 2050 | icd->error = 0; |
c1f50f76 | 2051 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 2052 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
2053 | } |
2054 | ||
97601024 JA |
2055 | static void ios_completed(struct thread_data *td, |
2056 | struct io_completion_data *icd) | |
36167d82 JA |
2057 | { |
2058 | struct io_u *io_u; | |
2059 | int i; | |
2060 | ||
10ba535a JA |
2061 | for (i = 0; i < icd->nr; i++) { |
2062 | io_u = td->io_ops->event(td, i); | |
2063 | ||
f8b0bd10 | 2064 | io_completed(td, &io_u, icd); |
e8462bd8 | 2065 | |
f8b0bd10 | 2066 | if (io_u) |
e8462bd8 | 2067 | put_io_u(td, io_u); |
10ba535a JA |
2068 | } |
2069 | } | |
97601024 | 2070 | |
e7e6cfb4 JA |
2071 | /* |
2072 | * Complete a single io_u for the sync engines. | |
2073 | */ | |
55312f9f | 2074 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
2075 | { |
2076 | struct io_completion_data icd; | |
55312f9f | 2077 | int ddir; |
97601024 | 2078 | |
9520ebb9 | 2079 | init_icd(td, &icd, 1); |
f8b0bd10 | 2080 | io_completed(td, &io_u, &icd); |
e8462bd8 | 2081 | |
f8b0bd10 | 2082 | if (io_u) |
e8462bd8 | 2083 | put_io_u(td, io_u); |
97601024 | 2084 | |
581e7141 JA |
2085 | if (icd.error) { |
2086 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2087 | return -1; | |
2088 | } | |
97601024 | 2089 | |
c1f50f76 | 2090 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2091 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
2092 | |
2093 | return 0; | |
97601024 JA |
2094 | } |
2095 | ||
e7e6cfb4 JA |
2096 | /* |
2097 | * Called to complete min_events number of io for the async engines. | |
2098 | */ | |
55312f9f | 2099 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2100 | { |
97601024 | 2101 | struct io_completion_data icd; |
00de55ef | 2102 | struct timespec *tvp = NULL; |
55312f9f | 2103 | int ret, ddir; |
4d06a338 | 2104 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2105 | |
12bb8569 | 2106 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2107 | |
4950421a | 2108 | if (!min_evts) |
00de55ef | 2109 | tvp = &ts; |
5fb4b366 RE |
2110 | else if (min_evts > td->cur_depth) |
2111 | min_evts = td->cur_depth; | |
97601024 | 2112 | |
82407585 RP |
2113 | /* No worries, td_io_getevents fixes min and max if they are |
2114 | * set incorrectly */ | |
2115 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2116 | if (ret < 0) { |
e1161c32 | 2117 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2118 | return ret; |
2119 | } else if (!ret) | |
2120 | return ret; | |
2121 | ||
9520ebb9 | 2122 | init_icd(td, &icd, ret); |
97601024 | 2123 | ios_completed(td, &icd); |
581e7141 JA |
2124 | if (icd.error) { |
2125 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2126 | return -1; | |
2127 | } | |
97601024 | 2128 | |
c1f50f76 | 2129 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2130 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2131 | |
0d593542 | 2132 | return ret; |
97601024 | 2133 | } |
7e77dd02 JA |
2134 | |
2135 | /* | |
2136 | * Call when io_u is really queued, to update the submission latency. | |
2137 | */ | |
2138 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2139 | { | |
8243be59 | 2140 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2141 | unsigned long slat_time; |
7e77dd02 | 2142 | |
d6bb626e | 2143 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2144 | |
2145 | if (td->parent) | |
2146 | td = td->parent; | |
2147 | ||
ae588852 | 2148 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
b2a432bf | 2149 | io_u->offset, io_u_is_prio(io_u)); |
9520ebb9 | 2150 | } |
7e77dd02 | 2151 | } |
433afcb4 | 2152 | |
5c94b008 JA |
2153 | /* |
2154 | * See if we should reuse the last seed, if dedupe is enabled | |
2155 | */ | |
9451b93e | 2156 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2157 | { |
2158 | unsigned int v; | |
5c94b008 JA |
2159 | |
2160 | if (!td->o.dedupe_percentage) | |
2161 | return &td->buf_state; | |
732eedd0 | 2162 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2163 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2164 | return &td->buf_state; | |
732eedd0 | 2165 | } |
5c94b008 | 2166 | |
1bd5d213 | 2167 | v = rand_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2168 | |
2169 | if (v <= td->o.dedupe_percentage) | |
2170 | return &td->buf_state_prev; | |
2171 | ||
2172 | return &td->buf_state; | |
2173 | } | |
2174 | ||
9451b93e | 2175 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2176 | { |
9451b93e JA |
2177 | if (td->o.dedupe_percentage == 100) |
2178 | frand_copy(rs, &td->buf_state_prev); | |
2179 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2180 | frand_copy(&td->buf_state_prev, rs); |
2181 | } | |
2182 | ||
5fff9543 JF |
2183 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write, |
2184 | unsigned long long max_bs) | |
5973cafb | 2185 | { |
d1af2894 JA |
2186 | struct thread_options *o = &td->o; |
2187 | ||
15600335 JA |
2188 | if (o->mem_type == MEM_CUDA_MALLOC) |
2189 | return; | |
03553853 | 2190 | |
4eff3e57 | 2191 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2192 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2193 | struct frand_state *rs; |
5fff9543 JF |
2194 | unsigned long long left = max_bs; |
2195 | unsigned long long this_write; | |
5c94b008 | 2196 | |
1066358a | 2197 | do { |
9451b93e | 2198 | rs = get_buf_state(td); |
9c42684e | 2199 | |
1066358a | 2200 | min_write = min(min_write, left); |
f97a43a1 | 2201 | |
1066358a | 2202 | if (perc) { |
4b157ac6 | 2203 | this_write = min_not_zero(min_write, |
5fff9543 | 2204 | (unsigned long long) td->o.compress_chunk); |
1e7f82e2 JA |
2205 | |
2206 | fill_random_buf_percentage(rs, buf, perc, | |
2207 | this_write, this_write, | |
2208 | o->buffer_pattern, | |
2209 | o->buffer_pattern_bytes); | |
2210 | } else { | |
1066358a | 2211 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2212 | this_write = min_write; |
2213 | } | |
1066358a | 2214 | |
1e7f82e2 JA |
2215 | buf += this_write; |
2216 | left -= this_write; | |
9451b93e | 2217 | save_buf_state(td, rs); |
1066358a | 2218 | } while (left); |
d1af2894 JA |
2219 | } else if (o->buffer_pattern_bytes) |
2220 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2221 | else if (o->zero_buffers) |
cc86c395 | 2222 | memset(buf, 0, max_bs); |
999d245e | 2223 | else |
9451b93e | 2224 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2225 | } |
2226 | ||
2227 | /* | |
2228 | * "randomly" fill the buffer contents | |
2229 | */ | |
2230 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
5fff9543 | 2231 | unsigned long long min_write, unsigned long long max_bs) |
cc86c395 JA |
2232 | { |
2233 | io_u->buf_filled_len = 0; | |
2234 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2235 | } |
e2c75fc4 TK |
2236 | |
2237 | static int do_sync_file_range(const struct thread_data *td, | |
2238 | struct fio_file *f) | |
2239 | { | |
7e2a317e | 2240 | uint64_t offset, nbytes; |
e2c75fc4 TK |
2241 | |
2242 | offset = f->first_write; | |
2243 | nbytes = f->last_write - f->first_write; | |
2244 | ||
2245 | if (!nbytes) | |
2246 | return 0; | |
2247 | ||
2248 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2249 | } | |
2250 | ||
2251 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2252 | { | |
2253 | int ret; | |
2254 | ||
2255 | if (io_u->ddir == DDIR_SYNC) { | |
2256 | ret = fsync(io_u->file->fd); | |
2257 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2258 | #ifdef CONFIG_FDATASYNC | |
2259 | ret = fdatasync(io_u->file->fd); | |
2260 | #else | |
2261 | ret = io_u->xfer_buflen; | |
2262 | io_u->error = EINVAL; | |
2263 | #endif | |
2264 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2265 | ret = do_sync_file_range(td, io_u->file); | |
2266 | else { | |
2267 | ret = io_u->xfer_buflen; | |
2268 | io_u->error = EINVAL; | |
2269 | } | |
2270 | ||
2271 | if (ret < 0) | |
2272 | io_u->error = errno; | |
2273 | ||
2274 | return ret; | |
2275 | } | |
2276 | ||
2277 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2278 | { | |
2279 | #ifndef FIO_HAVE_TRIM | |
2280 | io_u->error = EINVAL; | |
2281 | return 0; | |
2282 | #else | |
2283 | struct fio_file *f = io_u->file; | |
2284 | int ret; | |
2285 | ||
496b1f9e | 2286 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2287 | if (!ret) |
2288 | return io_u->xfer_buflen; | |
2289 | ||
2290 | io_u->error = ret; | |
2291 | return 0; | |
2292 | #endif | |
2293 | } |