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) && |
46ec8270 | 358 | o->time_based && o->nr_files == 1) { |
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 | ||
f9f1e137 VF |
420 | if (td_randtrimwrite(td) && ddir == DDIR_WRITE) { |
421 | /* don't mark randommap for these writes */ | |
422 | io_u_set(td, io_u, IO_U_F_BUSY_OK); | |
423 | offset = f->last_start[DDIR_TRIM]; | |
424 | *is_random = true; | |
425 | ret = 0; | |
426 | } else if (rw_seq) { | |
211c9b89 | 427 | if (td_random(td)) { |
6aca9b3d | 428 | if (should_do_random(td, ddir)) { |
211c9b89 | 429 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 430 | *is_random = true; |
6aca9b3d | 431 | } else { |
ec370c22 | 432 | *is_random = false; |
1651e431 | 433 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
434 | ret = get_next_seq_offset(td, f, ddir, &offset); |
435 | if (ret) | |
436 | ret = get_next_rand_block(td, f, ddir, &b); | |
437 | } | |
6aca9b3d | 438 | } else { |
ec370c22 | 439 | *is_random = false; |
37cf9e3c | 440 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 441 | } |
38dad62d | 442 | } else { |
1651e431 | 443 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
ec370c22 | 444 | *is_random = false; |
38dad62d JA |
445 | |
446 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 447 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 448 | if (ret) { |
37cf9e3c | 449 | ret = get_next_rand_block(td, f, ddir, &b); |
ec370c22 | 450 | *is_random = false; |
6aca9b3d | 451 | } |
38dad62d | 452 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
453 | if (f->last_start[ddir] != -1ULL) |
454 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 455 | else |
37cf9e3c | 456 | offset = 0; |
38dad62d JA |
457 | ret = 0; |
458 | } else { | |
459 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
460 | ret = 1; | |
461 | } | |
462 | } | |
6d68b997 | 463 | |
37cf9e3c JA |
464 | if (!ret) { |
465 | if (offset != -1ULL) | |
466 | io_u->offset = offset; | |
467 | else if (b != -1ULL) | |
468 | io_u->offset = b * td->o.ba[ddir]; | |
469 | else { | |
4e0a8fa2 | 470 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
471 | ret = 1; |
472 | } | |
4fff54cc | 473 | io_u->verify_offset = io_u->offset; |
37cf9e3c JA |
474 | } |
475 | ||
38dad62d JA |
476 | return ret; |
477 | } | |
478 | ||
10ba535a JA |
479 | /* |
480 | * For random io, generate a random new block and see if it's used. Repeat | |
481 | * until we find a free one. For sequential io, just return the end of | |
482 | * the last io issued. | |
483 | */ | |
e8fb335e | 484 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 485 | bool *is_random) |
10ba535a | 486 | { |
9bf2061e | 487 | struct fio_file *f = io_u->file; |
4ba66134 | 488 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 489 | int rw_seq_hit = 0; |
10ba535a | 490 | |
ff58fced JA |
491 | assert(ddir_rw(ddir)); |
492 | ||
38dad62d JA |
493 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
494 | rw_seq_hit = 1; | |
5736c10d | 495 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 496 | } |
211097b2 | 497 | |
6aca9b3d | 498 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 499 | return 1; |
10ba535a | 500 | |
009bd847 JA |
501 | if (io_u->offset >= f->io_size) { |
502 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
503 | (unsigned long long) io_u->offset, |
504 | (unsigned long long) f->io_size); | |
009bd847 JA |
505 | return 1; |
506 | } | |
507 | ||
508 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
509 | if (io_u->offset >= f->real_file_size) { |
510 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
511 | (unsigned long long) io_u->offset, |
512 | (unsigned long long) f->real_file_size); | |
10ba535a | 513 | return 1; |
2ba1c290 | 514 | } |
10ba535a | 515 | |
5e0863da VF |
516 | /* |
517 | * For randtrimwrite, we decide whether to issue a trim or a write | |
518 | * based on whether the offsets for the most recent trim and write | |
519 | * operations match. If they don't match that means we just issued a | |
520 | * new trim and the next operation should be a write. If they *do* | |
521 | * match that means we just completed a trim+write pair and the next | |
522 | * command should be a trim. | |
523 | * | |
524 | * This works fine for sequential workloads but for random workloads | |
525 | * it's possible to complete a trim+write pair and then have the next | |
526 | * randomly generated offset match the previous offset. If that happens | |
527 | * we need to alter the offset for the last write operation in order | |
528 | * to ensure that we issue a write operation the next time through. | |
529 | */ | |
530 | if (td_randtrimwrite(td) && ddir == DDIR_TRIM && | |
531 | f->last_start[DDIR_TRIM] == io_u->offset) | |
793b8686 | 532 | f->last_start[DDIR_WRITE]--; |
5e0863da | 533 | |
9bf46532 | 534 | io_u->verify_offset = io_u->offset; |
10ba535a JA |
535 | return 0; |
536 | } | |
537 | ||
e39c0676 | 538 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
5fff9543 | 539 | unsigned long long buflen) |
79944128 JA |
540 | { |
541 | struct fio_file *f = io_u->file; | |
542 | ||
bedc9dc2 | 543 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
544 | } |
545 | ||
5fff9543 | 546 | static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u, |
ec370c22 | 547 | bool is_random) |
10ba535a | 548 | { |
6aca9b3d | 549 | int ddir = io_u->ddir; |
5fff9543 JF |
550 | unsigned long long buflen = 0; |
551 | unsigned long long minbs, maxbs; | |
3dd29f7c | 552 | uint64_t frand_max, r; |
7c961359 | 553 | bool power_2; |
10ba535a | 554 | |
9ee1c647 | 555 | assert(ddir_rw(ddir)); |
6aca9b3d | 556 | |
f9f1e137 VF |
557 | if (td_randtrimwrite(td) && ddir == DDIR_WRITE) { |
558 | struct fio_file *f = io_u->file; | |
559 | ||
560 | return f->last_pos[DDIR_TRIM] - f->last_start[DDIR_TRIM]; | |
561 | } | |
562 | ||
6aca9b3d | 563 | if (td->o.bs_is_seq_rand) |
ec370c22 | 564 | ddir = is_random ? DDIR_WRITE : DDIR_READ; |
ff58fced | 565 | |
f3059de1 JA |
566 | minbs = td->o.min_bs[ddir]; |
567 | maxbs = td->o.max_bs[ddir]; | |
568 | ||
79944128 JA |
569 | if (minbs == maxbs) |
570 | return minbs; | |
571 | ||
52c58027 JA |
572 | /* |
573 | * If we can't satisfy the min block size from here, then fail | |
574 | */ | |
575 | if (!io_u_fits(td, io_u, minbs)) | |
576 | return 0; | |
577 | ||
2f282cec | 578 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 579 | do { |
2f282cec | 580 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 581 | |
720e84ad | 582 | if (!td->o.bssplit_nr[ddir]) { |
5fff9543 | 583 | buflen = minbs + (unsigned long long) ((double) maxbs * |
c3546b53 | 584 | (r / (frand_max + 1.0))); |
5ec10eaa | 585 | } else { |
3dd29f7c | 586 | long long perc = 0; |
564ca972 JA |
587 | unsigned int i; |
588 | ||
720e84ad JA |
589 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
590 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 | 591 | |
cc5e4cb0 DLM |
592 | if (!bsp->perc) |
593 | continue; | |
564ca972 JA |
594 | buflen = bsp->bs; |
595 | perc += bsp->perc; | |
3dd29f7c | 596 | if ((r / perc <= frand_max / 100ULL) && |
79944128 | 597 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
598 | break; |
599 | } | |
600 | } | |
79944128 | 601 | |
17a6b702 PL |
602 | power_2 = is_power_of_2(minbs); |
603 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 604 | buflen &= ~(minbs - 1); |
8135fe4d JA |
605 | else if (!td->o.bs_unaligned && !power_2) |
606 | buflen -= buflen % minbs; | |
607 | if (buflen > maxbs) | |
608 | buflen = maxbs; | |
79944128 | 609 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 610 | |
10ba535a JA |
611 | return buflen; |
612 | } | |
613 | ||
afe24a5a JA |
614 | static void set_rwmix_bytes(struct thread_data *td) |
615 | { | |
afe24a5a JA |
616 | unsigned int diff; |
617 | ||
618 | /* | |
619 | * we do time or byte based switch. this is needed because | |
620 | * buffered writes may issue a lot quicker than they complete, | |
621 | * whereas reads do not. | |
622 | */ | |
e47f799f | 623 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 624 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
625 | } |
626 | ||
627 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
628 | { | |
629 | unsigned int v; | |
e47f799f | 630 | |
1bd5d213 | 631 | v = rand_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 632 | |
04c540d9 | 633 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
634 | return DDIR_READ; |
635 | ||
636 | return DDIR_WRITE; | |
afe24a5a JA |
637 | } |
638 | ||
0d593542 | 639 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 640 | { |
2dcfdeb9 | 641 | int ret = 0, completed = 0, err = 0; |
0d593542 | 642 | |
002e7183 JA |
643 | /* |
644 | * We are going to sleep, ensure that we flush anything pending as | |
645 | * not to skew our latency numbers. | |
646 | * | |
647 | * Changed to only monitor 'in flight' requests here instead of the | |
648 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
649 | * io's that have been actually submitted to an async engine, | |
650 | * and cur_depth is meaningless for sync engines. | |
651 | */ | |
a80cb54b BVA |
652 | if (td->io_u_queued || td->cur_depth) |
653 | td_io_commit(td); | |
9cc80b6d | 654 | |
002e7183 | 655 | while (td->io_u_in_flight) { |
55312f9f | 656 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
657 | if (ret > 0) |
658 | completed += ret; | |
d28174f0 | 659 | else if (ret < 0) |
2dcfdeb9 | 660 | err = ret; |
002e7183 | 661 | } |
0d593542 | 662 | |
6be06c46 JA |
663 | if (td->flags & TD_F_REGROW_LOGS) |
664 | regrow_logs(td); | |
665 | ||
d28174f0 JA |
666 | if (completed) |
667 | return completed; | |
668 | ||
2dcfdeb9 | 669 | return err; |
002e7183 JA |
670 | } |
671 | ||
581e7141 JA |
672 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
673 | { | |
674 | enum fio_ddir odir = ddir ^ 1; | |
b5407f8b | 675 | uint64_t usec; |
90eff1c9 | 676 | uint64_t now; |
581e7141 | 677 | |
ff58fced | 678 | assert(ddir_rw(ddir)); |
970d9000 | 679 | now = utime_since_now(&td->epoch); |
ff58fced | 680 | |
50a8ce86 D |
681 | /* |
682 | * if rate_next_io_time is in the past, need to catch up to rate | |
683 | */ | |
684 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
685 | return ddir; |
686 | ||
687 | /* | |
50a8ce86 | 688 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
689 | * should switch. |
690 | */ | |
315fcfec | 691 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 692 | /* |
50a8ce86 | 693 | * Other direction is behind rate, switch |
581e7141 | 694 | */ |
50a8ce86 | 695 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
696 | return odir; |
697 | ||
698 | /* | |
395feabb JA |
699 | * Both directions are ahead of rate. sleep the min, |
700 | * switch if necessary | |
581e7141 | 701 | */ |
50a8ce86 | 702 | if (td->rate_next_io_time[ddir] <= |
395feabb | 703 | td->rate_next_io_time[odir]) { |
50a8ce86 | 704 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 705 | } else { |
50a8ce86 | 706 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
707 | ddir = odir; |
708 | } | |
709 | } else | |
50a8ce86 | 710 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 711 | |
a9da8ab2 JA |
712 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
713 | io_u_quiesce(td); | |
78c1eda5 | 714 | |
df06a036 | 715 | if (td->o.timeout && ((usec + now) > td->o.timeout)) { |
716 | /* | |
717 | * check if the usec is capable of taking negative values | |
718 | */ | |
719 | if (now > td->o.timeout) { | |
720 | ddir = DDIR_INVAL; | |
721 | return ddir; | |
722 | } | |
723 | usec = td->o.timeout - now; | |
724 | } | |
1a9bf814 | 725 | usec_sleep(td, usec); |
df06a036 | 726 | |
727 | now = utime_since_now(&td->epoch); | |
728 | if ((td->o.timeout && (now > td->o.timeout)) || td->terminate) | |
729 | ddir = DDIR_INVAL; | |
730 | ||
581e7141 JA |
731 | return ddir; |
732 | } | |
733 | ||
10ba535a JA |
734 | /* |
735 | * Return the data direction for the next io_u. If the job is a | |
736 | * mixed read/write workload, check the rwmix cycle and switch if | |
737 | * necessary. | |
738 | */ | |
1e97cce9 | 739 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 740 | { |
581e7141 JA |
741 | enum fio_ddir ddir; |
742 | ||
5f9099ea | 743 | /* |
f6860972 TK |
744 | * See if it's time to fsync/fdatasync/sync_file_range first, |
745 | * and if not then move on to check regular I/Os. | |
5f9099ea | 746 | */ |
f6860972 TK |
747 | if (should_fsync(td)) { |
748 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
749 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
750 | return DDIR_SYNC; | |
751 | ||
752 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
753 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
754 | return DDIR_DATASYNC; | |
755 | ||
756 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
757 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
758 | return DDIR_SYNC_FILE_RANGE; | |
759 | } | |
44f29692 | 760 | |
10ba535a | 761 | if (td_rw(td)) { |
10ba535a JA |
762 | /* |
763 | * Check if it's time to seed a new data direction. | |
764 | */ | |
e4928662 | 765 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
766 | /* |
767 | * Put a top limit on how many bytes we do for | |
768 | * one data direction, to avoid overflowing the | |
769 | * ranges too much | |
770 | */ | |
771 | ddir = get_rand_ddir(td); | |
e47f799f JA |
772 | |
773 | if (ddir != td->rwmix_ddir) | |
774 | set_rwmix_bytes(td); | |
775 | ||
776 | td->rwmix_ddir = ddir; | |
10ba535a | 777 | } |
581e7141 | 778 | ddir = td->rwmix_ddir; |
10ba535a | 779 | } else if (td_read(td)) |
581e7141 | 780 | ddir = DDIR_READ; |
6eaf09d6 | 781 | else if (td_write(td)) |
581e7141 | 782 | ddir = DDIR_WRITE; |
5c8e84ca | 783 | else if (td_trim(td)) |
6eaf09d6 | 784 | ddir = DDIR_TRIM; |
5c8e84ca TK |
785 | else |
786 | ddir = DDIR_INVAL; | |
581e7141 | 787 | |
2f5102a5 H |
788 | if (!should_check_rate(td)) { |
789 | /* | |
790 | * avoid time-consuming call to utime_since_now() if rate checking | |
791 | * isn't being used. this imrpoves IOPs 50%. See: | |
792 | * https://github.com/axboe/fio/issues/1501#issuecomment-1418327049 | |
793 | */ | |
794 | td->rwmix_ddir = ddir; | |
795 | } else | |
796 | td->rwmix_ddir = rate_ddir(td, ddir); | |
581e7141 | 797 | return td->rwmix_ddir; |
10ba535a JA |
798 | } |
799 | ||
1ef2b6be JA |
800 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
801 | { | |
0e4dd95c DE |
802 | enum fio_ddir ddir = get_rw_ddir(td); |
803 | ||
731461cc DLM |
804 | if (td->o.zone_mode == ZONE_MODE_ZBD) |
805 | ddir = zbd_adjust_ddir(td, io_u, ddir); | |
c65057f9 | 806 | |
82a90686 | 807 | if (td_trimwrite(td)) { |
0e4dd95c | 808 | struct fio_file *f = io_u->file; |
793b8686 | 809 | if (f->last_start[DDIR_WRITE] == f->last_start[DDIR_TRIM]) |
0e4dd95c DE |
810 | ddir = DDIR_TRIM; |
811 | else | |
812 | ddir = DDIR_WRITE; | |
813 | } | |
814 | ||
815 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 816 | |
9b87f09b | 817 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
818 | td->o.barrier_blocks && |
819 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
820 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 821 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
822 | } |
823 | ||
e8462bd8 | 824 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 825 | { |
71b84caa | 826 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
827 | |
828 | if (ret) | |
829 | td_verror(td, ret, "file close"); | |
830 | } | |
831 | ||
10ba535a JA |
832 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
833 | { | |
26b3a188 JA |
834 | const bool needs_lock = td_async_processing(td); |
835 | ||
b2da58c4 | 836 | zbd_put_io_u(td, io_u); |
99952ca7 | 837 | |
a9da8ab2 JA |
838 | if (td->parent) |
839 | td = td->parent; | |
840 | ||
26b3a188 JA |
841 | if (needs_lock) |
842 | __td_io_u_lock(td); | |
e8462bd8 | 843 | |
f8b0bd10 | 844 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 845 | put_file_log(td, io_u->file); |
f8b0bd10 | 846 | |
10ba535a | 847 | io_u->file = NULL; |
1651e431 | 848 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 849 | |
a9da8ab2 | 850 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 851 | td->cur_depth--; |
a9da8ab2 JA |
852 | assert(!(td->flags & TD_F_CHILD)); |
853 | } | |
2ae0b204 | 854 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 | 855 | td_io_u_free_notify(td); |
26b3a188 JA |
856 | |
857 | if (needs_lock) | |
858 | __td_io_u_unlock(td); | |
10ba535a JA |
859 | } |
860 | ||
f2bba182 RR |
861 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
862 | { | |
1651e431 | 863 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
864 | put_io_u(td, io_u); |
865 | } | |
866 | ||
755200a3 JA |
867 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
868 | { | |
26b3a188 | 869 | const bool needs_lock = td_async_processing(td); |
755200a3 | 870 | struct io_u *__io_u = *io_u; |
bcd5abfa | 871 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 872 | |
465221b0 JA |
873 | dprint(FD_IO, "requeue %p\n", __io_u); |
874 | ||
a9da8ab2 JA |
875 | if (td->parent) |
876 | td = td->parent; | |
877 | ||
26b3a188 JA |
878 | if (needs_lock) |
879 | __td_io_u_lock(td); | |
e8462bd8 | 880 | |
1651e431 | 881 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
882 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
883 | td->io_issues[ddir]--; | |
5ec10eaa | 884 | |
1651e431 | 885 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 886 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 887 | td->cur_depth--; |
a9da8ab2 JA |
888 | assert(!(td->flags & TD_F_CHILD)); |
889 | } | |
2ae0b204 JA |
890 | |
891 | io_u_rpush(&td->io_u_requeues, __io_u); | |
a9da8ab2 | 892 | td_io_u_free_notify(td); |
26b3a188 JA |
893 | |
894 | if (needs_lock) | |
895 | __td_io_u_unlock(td); | |
896 | ||
755200a3 JA |
897 | *io_u = NULL; |
898 | } | |
899 | ||
7b865a2f | 900 | static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u) |
224b3093 | 901 | { |
902 | struct fio_file *f = io_u->file; | |
903 | ||
7b865a2f BVA |
904 | assert(td->o.zone_mode == ZONE_MODE_STRIDED); |
905 | assert(td->o.zone_size); | |
906 | assert(td->o.zone_range); | |
907 | ||
224b3093 | 908 | /* |
909 | * See if it's time to switch to a new zone | |
910 | */ | |
fa9fd914 | 911 | if (td->zone_bytes >= td->o.zone_size) { |
224b3093 | 912 | td->zone_bytes = 0; |
913 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
914 | ||
915 | /* | |
916 | * Wrap from the beginning, if we exceed the file size | |
917 | */ | |
918 | if (f->file_offset >= f->real_file_size) | |
04bc85a1 JA |
919 | f->file_offset = get_start_offset(td, f); |
920 | ||
224b3093 | 921 | f->last_pos[io_u->ddir] = f->file_offset; |
922 | td->io_skip_bytes += td->o.zone_skip; | |
923 | } | |
924 | ||
925 | /* | |
04bc85a1 JA |
926 | * If zone_size > zone_range, then maintain the same zone until |
927 | * zone_bytes >= zone_size. | |
928 | */ | |
224b3093 | 929 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
930 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
931 | f->file_offset, f->last_pos[io_u->ddir]); | |
932 | f->last_pos[io_u->ddir] = f->file_offset; | |
933 | } | |
934 | ||
935 | /* | |
936 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
937 | * map needs to be reset as it's done with zone_range everytime. | |
938 | */ | |
04bc85a1 | 939 | if ((td->zone_bytes % td->o.zone_range) == 0) |
224b3093 | 940 | fio_file_reset(td, f); |
224b3093 | 941 | } |
942 | ||
9bf2061e | 943 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 944 | { |
ec370c22 | 945 | bool is_random; |
bfbdd35b BVA |
946 | uint64_t offset; |
947 | enum io_u_action ret; | |
6aca9b3d | 948 | |
9b87f09b | 949 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
950 | goto out; |
951 | ||
1ef2b6be | 952 | set_rw_ddir(td, io_u); |
5f9099ea | 953 | |
df06a036 | 954 | if (io_u->ddir == DDIR_INVAL) { |
955 | dprint(FD_IO, "invalid direction received ddir = %d", io_u->ddir); | |
956 | return 1; | |
957 | } | |
87dc1ab1 | 958 | /* |
ff58fced | 959 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 960 | */ |
ff58fced | 961 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 962 | goto out; |
a00735e6 | 963 | |
7b865a2f BVA |
964 | if (td->o.zone_mode == ZONE_MODE_STRIDED) |
965 | setup_strided_zone_mode(td, io_u); | |
4d37720a DLM |
966 | else if (td->o.zone_mode == ZONE_MODE_ZBD) |
967 | setup_zbd_zone_mode(td, io_u); | |
48f5abd3 | 968 | |
10ba535a | 969 | /* |
c685b5b2 JA |
970 | * No log, let the seq/rand engine retrieve the next buflen and |
971 | * position. | |
10ba535a | 972 | */ |
6aca9b3d | 973 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 974 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 975 | return 1; |
2ba1c290 | 976 | } |
10ba535a | 977 | |
6aca9b3d | 978 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
979 | if (!io_u->buflen) { |
980 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 981 | return 1; |
2ba1c290 | 982 | } |
bca4ed4d | 983 | |
bfbdd35b BVA |
984 | offset = io_u->offset; |
985 | if (td->o.zone_mode == ZONE_MODE_ZBD) { | |
986 | ret = zbd_adjust_block(td, io_u); | |
629eda50 BVA |
987 | if (ret == io_u_eof) { |
988 | dprint(FD_IO, "zbd_adjust_block() returned io_u_eof\n"); | |
bfbdd35b | 989 | return 1; |
629eda50 | 990 | } |
bfbdd35b BVA |
991 | } |
992 | ||
a7e8aae0 KB |
993 | if (td->o.fdp) |
994 | fdp_fill_dspec_data(td, io_u); | |
995 | ||
2ba1c290 | 996 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
5fff9543 | 997 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n", |
e5f9a813 | 998 | io_u, |
4b91ee8f JA |
999 | (unsigned long long) io_u->offset, io_u->buflen, |
1000 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 1001 | return 1; |
2ba1c290 | 1002 | } |
6a5e6884 | 1003 | |
bca4ed4d JA |
1004 | /* |
1005 | * mark entry before potentially trimming io_u | |
1006 | */ | |
303032ae | 1007 | if (td_random(td) && file_randommap(td, io_u->file)) |
bfbdd35b | 1008 | io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen); |
bca4ed4d | 1009 | |
c38e9468 | 1010 | out: |
e5f9a813 | 1011 | dprint_io_u(io_u, "fill"); |
cb7d7abb | 1012 | io_u->verify_offset = io_u->offset; |
d9d91e39 | 1013 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 1014 | return 0; |
10ba535a JA |
1015 | } |
1016 | ||
6cc0e5aa | 1017 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
838bc709 | 1018 | { |
2b13e716 | 1019 | int idx = 0; |
838bc709 JA |
1020 | |
1021 | switch (nr) { | |
1022 | default: | |
2b13e716 | 1023 | idx = 6; |
838bc709 JA |
1024 | break; |
1025 | case 33 ... 64: | |
2b13e716 | 1026 | idx = 5; |
838bc709 JA |
1027 | break; |
1028 | case 17 ... 32: | |
2b13e716 | 1029 | idx = 4; |
838bc709 JA |
1030 | break; |
1031 | case 9 ... 16: | |
2b13e716 | 1032 | idx = 3; |
838bc709 JA |
1033 | break; |
1034 | case 5 ... 8: | |
2b13e716 | 1035 | idx = 2; |
838bc709 JA |
1036 | break; |
1037 | case 1 ... 4: | |
2b13e716 | 1038 | idx = 1; |
87933e32 | 1039 | fio_fallthrough; |
838bc709 JA |
1040 | case 0: |
1041 | break; | |
1042 | } | |
1043 | ||
2b13e716 | 1044 | map[idx]++; |
838bc709 JA |
1045 | } |
1046 | ||
1047 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
1048 | { | |
1049 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
1050 | td->ts.total_submit++; | |
1051 | } | |
1052 | ||
1053 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
1054 | { | |
1055 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
1056 | td->ts.total_complete++; | |
1057 | } | |
1058 | ||
d8005759 | 1059 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 1060 | { |
2b13e716 | 1061 | int idx = 0; |
71619dc2 JA |
1062 | |
1063 | switch (td->cur_depth) { | |
1064 | default: | |
2b13e716 | 1065 | idx = 6; |
a783e61a | 1066 | break; |
71619dc2 | 1067 | case 32 ... 63: |
2b13e716 | 1068 | idx = 5; |
a783e61a | 1069 | break; |
71619dc2 | 1070 | case 16 ... 31: |
2b13e716 | 1071 | idx = 4; |
a783e61a | 1072 | break; |
71619dc2 | 1073 | case 8 ... 15: |
2b13e716 | 1074 | idx = 3; |
a783e61a | 1075 | break; |
71619dc2 | 1076 | case 4 ... 7: |
2b13e716 | 1077 | idx = 2; |
a783e61a | 1078 | break; |
71619dc2 | 1079 | case 2 ... 3: |
2b13e716 | 1080 | idx = 1; |
87933e32 | 1081 | fio_fallthrough; |
71619dc2 JA |
1082 | case 1: |
1083 | break; | |
1084 | } | |
1085 | ||
2b13e716 | 1086 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
1087 | } |
1088 | ||
d6bb626e | 1089 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1090 | { |
2b13e716 | 1091 | int idx = 0; |
04a0feae | 1092 | |
d6bb626e VF |
1093 | assert(nsec < 1000); |
1094 | ||
1095 | switch (nsec) { | |
1096 | case 750 ... 999: | |
1097 | idx = 9; | |
1098 | break; | |
1099 | case 500 ... 749: | |
1100 | idx = 8; | |
1101 | break; | |
1102 | case 250 ... 499: | |
1103 | idx = 7; | |
1104 | break; | |
1105 | case 100 ... 249: | |
1106 | idx = 6; | |
1107 | break; | |
1108 | case 50 ... 99: | |
1109 | idx = 5; | |
1110 | break; | |
1111 | case 20 ... 49: | |
1112 | idx = 4; | |
1113 | break; | |
1114 | case 10 ... 19: | |
1115 | idx = 3; | |
1116 | break; | |
1117 | case 4 ... 9: | |
1118 | idx = 2; | |
1119 | break; | |
1120 | case 2 ... 3: | |
1121 | idx = 1; | |
87933e32 | 1122 | fio_fallthrough; |
d6bb626e VF |
1123 | case 0 ... 1: |
1124 | break; | |
1125 | } | |
1126 | ||
1127 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1128 | td->ts.io_u_lat_n[idx]++; | |
1129 | } | |
1130 | ||
1131 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1132 | { | |
1133 | int idx = 0; | |
1134 | ||
1135 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1136 | |
1137 | switch (usec) { | |
1138 | case 750 ... 999: | |
2b13e716 | 1139 | idx = 9; |
04a0feae JA |
1140 | break; |
1141 | case 500 ... 749: | |
2b13e716 | 1142 | idx = 8; |
04a0feae JA |
1143 | break; |
1144 | case 250 ... 499: | |
2b13e716 | 1145 | idx = 7; |
04a0feae JA |
1146 | break; |
1147 | case 100 ... 249: | |
2b13e716 | 1148 | idx = 6; |
04a0feae JA |
1149 | break; |
1150 | case 50 ... 99: | |
2b13e716 | 1151 | idx = 5; |
04a0feae JA |
1152 | break; |
1153 | case 20 ... 49: | |
2b13e716 | 1154 | idx = 4; |
04a0feae JA |
1155 | break; |
1156 | case 10 ... 19: | |
2b13e716 | 1157 | idx = 3; |
04a0feae JA |
1158 | break; |
1159 | case 4 ... 9: | |
2b13e716 | 1160 | idx = 2; |
04a0feae JA |
1161 | break; |
1162 | case 2 ... 3: | |
2b13e716 | 1163 | idx = 1; |
87933e32 | 1164 | fio_fallthrough; |
04a0feae JA |
1165 | case 0 ... 1: |
1166 | break; | |
1167 | } | |
1168 | ||
2b13e716 JA |
1169 | assert(idx < FIO_IO_U_LAT_U_NR); |
1170 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1171 | } |
1172 | ||
d6bb626e | 1173 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1174 | { |
2b13e716 | 1175 | int idx = 0; |
ec118304 | 1176 | |
d6bb626e VF |
1177 | assert(msec >= 1); |
1178 | ||
ec118304 JA |
1179 | switch (msec) { |
1180 | default: | |
2b13e716 | 1181 | idx = 11; |
04a0feae | 1182 | break; |
8abdce66 | 1183 | case 1000 ... 1999: |
2b13e716 | 1184 | idx = 10; |
04a0feae | 1185 | break; |
8abdce66 | 1186 | case 750 ... 999: |
2b13e716 | 1187 | idx = 9; |
04a0feae | 1188 | break; |
8abdce66 | 1189 | case 500 ... 749: |
2b13e716 | 1190 | idx = 8; |
04a0feae | 1191 | break; |
8abdce66 | 1192 | case 250 ... 499: |
2b13e716 | 1193 | idx = 7; |
04a0feae | 1194 | break; |
8abdce66 | 1195 | case 100 ... 249: |
2b13e716 | 1196 | idx = 6; |
04a0feae | 1197 | break; |
8abdce66 | 1198 | case 50 ... 99: |
2b13e716 | 1199 | idx = 5; |
04a0feae | 1200 | break; |
8abdce66 | 1201 | case 20 ... 49: |
2b13e716 | 1202 | idx = 4; |
04a0feae | 1203 | break; |
8abdce66 | 1204 | case 10 ... 19: |
2b13e716 | 1205 | idx = 3; |
04a0feae | 1206 | break; |
8abdce66 | 1207 | case 4 ... 9: |
2b13e716 | 1208 | idx = 2; |
04a0feae | 1209 | break; |
ec118304 | 1210 | case 2 ... 3: |
2b13e716 | 1211 | idx = 1; |
87933e32 | 1212 | fio_fallthrough; |
ec118304 JA |
1213 | case 0 ... 1: |
1214 | break; | |
1215 | } | |
1216 | ||
2b13e716 JA |
1217 | assert(idx < FIO_IO_U_LAT_M_NR); |
1218 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1219 | } |
1220 | ||
d6bb626e | 1221 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1222 | { |
d6bb626e VF |
1223 | if (nsec < 1000) |
1224 | io_u_mark_lat_nsec(td, nsec); | |
1225 | else if (nsec < 1000000) | |
1226 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1227 | else |
d6bb626e | 1228 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1229 | } |
1230 | ||
8c07860d JA |
1231 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1232 | { | |
1233 | unsigned long fileno; | |
1234 | ||
1235 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1236 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1237 | unsigned long r; | |
1238 | ||
1239 | r = __rand(&td->next_file_state); | |
1240 | return (unsigned int) ((double) td->o.nr_files | |
1241 | * (r / (frand_max + 1.0))); | |
1242 | } | |
1243 | ||
1244 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1245 | fileno = zipf_next(&td->next_file_zipf); | |
1246 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1247 | fileno = pareto_next(&td->next_file_zipf); | |
1248 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1249 | fileno = gauss_next(&td->next_file_gauss); | |
1250 | else { | |
1251 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1252 | assert(0); | |
1253 | return 0; | |
1254 | } | |
1255 | ||
1256 | return fileno >> FIO_FSERVICE_SHIFT; | |
1257 | } | |
1258 | ||
0aabe160 JA |
1259 | /* |
1260 | * Get next file to service by choosing one at random | |
1261 | */ | |
2cc52930 JA |
1262 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1263 | enum fio_file_flags goodf, | |
d6aed795 | 1264 | enum fio_file_flags badf) |
0aabe160 | 1265 | { |
0aabe160 | 1266 | struct fio_file *f; |
1c178180 | 1267 | int fno; |
0aabe160 JA |
1268 | |
1269 | do { | |
87b10676 | 1270 | int opened = 0; |
4c07ad86 | 1271 | |
8c07860d | 1272 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1273 | |
126d65c6 | 1274 | f = td->files[fno]; |
d6aed795 | 1275 | if (fio_file_done(f)) |
059e63c0 | 1276 | continue; |
1c178180 | 1277 | |
d6aed795 | 1278 | if (!fio_file_open(f)) { |
87b10676 JA |
1279 | int err; |
1280 | ||
002fe734 JA |
1281 | if (td->nr_open_files >= td->o.open_files) |
1282 | return ERR_PTR(-EBUSY); | |
1283 | ||
87b10676 JA |
1284 | err = td_io_open_file(td, f); |
1285 | if (err) | |
1286 | continue; | |
1287 | opened = 1; | |
1288 | } | |
1289 | ||
2ba1c290 JA |
1290 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1291 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1292 | return f; |
2ba1c290 | 1293 | } |
87b10676 JA |
1294 | if (opened) |
1295 | td_io_close_file(td, f); | |
0aabe160 JA |
1296 | } while (1); |
1297 | } | |
1298 | ||
1299 | /* | |
1300 | * Get next file to service by doing round robin between all available ones | |
1301 | */ | |
1c178180 JA |
1302 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1303 | int badf) | |
3d7c391d JA |
1304 | { |
1305 | unsigned int old_next_file = td->next_file; | |
1306 | struct fio_file *f; | |
1307 | ||
1308 | do { | |
87b10676 JA |
1309 | int opened = 0; |
1310 | ||
126d65c6 | 1311 | f = td->files[td->next_file]; |
3d7c391d JA |
1312 | |
1313 | td->next_file++; | |
2dc1bbeb | 1314 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1315 | td->next_file = 0; |
1316 | ||
87b10676 | 1317 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1318 | if (fio_file_done(f)) { |
d5ed68ea | 1319 | f = NULL; |
059e63c0 | 1320 | continue; |
d5ed68ea | 1321 | } |
059e63c0 | 1322 | |
d6aed795 | 1323 | if (!fio_file_open(f)) { |
87b10676 JA |
1324 | int err; |
1325 | ||
002fe734 JA |
1326 | if (td->nr_open_files >= td->o.open_files) |
1327 | return ERR_PTR(-EBUSY); | |
1328 | ||
87b10676 | 1329 | err = td_io_open_file(td, f); |
b5696bfc JA |
1330 | if (err) { |
1331 | dprint(FD_FILE, "error %d on open of %s\n", | |
1332 | err, f->file_name); | |
87c27b45 | 1333 | f = NULL; |
87b10676 | 1334 | continue; |
b5696bfc | 1335 | } |
87b10676 JA |
1336 | opened = 1; |
1337 | } | |
1338 | ||
0b9d69ec JA |
1339 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1340 | f->flags); | |
1c178180 | 1341 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1342 | break; |
1343 | ||
87b10676 JA |
1344 | if (opened) |
1345 | td_io_close_file(td, f); | |
1346 | ||
3d7c391d JA |
1347 | f = NULL; |
1348 | } while (td->next_file != old_next_file); | |
1349 | ||
2ba1c290 | 1350 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1351 | return f; |
1352 | } | |
1353 | ||
7eb36574 | 1354 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1355 | { |
1907dbc6 JA |
1356 | struct fio_file *f; |
1357 | ||
2dc1bbeb | 1358 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1359 | |
b5696bfc | 1360 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1361 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1362 | " nr_files=%d\n", td->nr_open_files, | |
1363 | td->nr_done_files, | |
1364 | td->o.nr_files); | |
bdb4e2e9 | 1365 | return NULL; |
2ba1c290 | 1366 | } |
bdb4e2e9 | 1367 | |
1907dbc6 | 1368 | f = td->file_service_file; |
d6aed795 | 1369 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1370 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1371 | goto out; | |
4ef1562a | 1372 | if (td->file_service_left) { |
2d8c2709 JA |
1373 | td->file_service_left--; |
1374 | goto out; | |
4ef1562a | 1375 | } |
a086c257 | 1376 | } |
1907dbc6 | 1377 | |
a086c257 JA |
1378 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1379 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1380 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1381 | else |
d6aed795 | 1382 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1383 | |
002fe734 JA |
1384 | if (IS_ERR(f)) |
1385 | return f; | |
1386 | ||
1907dbc6 JA |
1387 | td->file_service_file = f; |
1388 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1389 | out: |
0dac421f JA |
1390 | if (f) |
1391 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1392 | else | |
1393 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1394 | return f; |
bdb4e2e9 JA |
1395 | } |
1396 | ||
7eb36574 JA |
1397 | static struct fio_file *get_next_file(struct thread_data *td) |
1398 | { | |
7eb36574 JA |
1399 | return __get_next_file(td); |
1400 | } | |
1401 | ||
002fe734 | 1402 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1403 | { |
1404 | struct fio_file *f; | |
1405 | ||
1406 | do { | |
1407 | f = get_next_file(td); | |
002fe734 JA |
1408 | if (IS_ERR_OR_NULL(f)) |
1409 | return PTR_ERR(f); | |
429f6675 | 1410 | |
429f6675 JA |
1411 | io_u->file = f; |
1412 | get_file(f); | |
1413 | ||
1414 | if (!fill_io_u(td, io_u)) | |
1415 | break; | |
1416 | ||
b2da58c4 | 1417 | zbd_put_io_u(td, io_u); |
99952ca7 | 1418 | |
b5696bfc | 1419 | put_file_log(td, f); |
429f6675 | 1420 | td_io_close_file(td, f); |
b5696bfc | 1421 | io_u->file = NULL; |
8c07860d JA |
1422 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1423 | fio_file_reset(td, f); | |
1424 | else { | |
1425 | fio_file_set_done(f); | |
1426 | td->nr_done_files++; | |
1427 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1428 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1429 | } |
429f6675 JA |
1430 | } while (1); |
1431 | ||
1432 | return 0; | |
1433 | } | |
1434 | ||
f7cf63bf | 1435 | static void lat_fatal(struct thread_data *td, struct io_u *io_u, struct io_completion_data *icd, |
c3a32714 | 1436 | unsigned long long tnsec, unsigned long long max_nsec) |
3e260a46 | 1437 | { |
f7cf63bf VR |
1438 | if (!td->error) { |
1439 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec): %s %s %llu %llu\n", | |
1440 | tnsec, max_nsec, | |
1441 | io_u->file->file_name, | |
1442 | io_ddir_name(io_u->ddir), | |
1443 | io_u->offset, io_u->buflen); | |
1444 | } | |
3e260a46 JA |
1445 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
1446 | icd->error = ETIMEDOUT; | |
1447 | } | |
1448 | ||
1449 | static void lat_new_cycle(struct thread_data *td) | |
1450 | { | |
1451 | fio_gettime(&td->latency_ts, NULL); | |
1452 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1453 | td->latency_failed = 0; | |
1454 | } | |
1455 | ||
1456 | /* | |
1457 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1458 | * are at QD=1, then it's time to give up. | |
1459 | */ | |
e39c0676 | 1460 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1461 | { |
1462 | if (td->latency_qd == 1) | |
e39c0676 | 1463 | return true; |
3e260a46 JA |
1464 | |
1465 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1466 | |
1467 | if (td->latency_qd == td->latency_qd_low) | |
1468 | td->latency_qd_low--; | |
1469 | ||
3e260a46 | 1470 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
e1bcd541 | 1471 | td->latency_stable_count = 0; |
3e260a46 JA |
1472 | |
1473 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1474 | ||
1475 | /* | |
1476 | * When we ramp QD down, quiesce existing IO to prevent | |
1477 | * a storm of ramp downs due to pending higher depth. | |
1478 | */ | |
1479 | io_u_quiesce(td); | |
1480 | lat_new_cycle(td); | |
e39c0676 | 1481 | return false; |
3e260a46 JA |
1482 | } |
1483 | ||
e39c0676 | 1484 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1485 | { |
1486 | if (td->o.latency_percentile.u.f == 100.0) | |
1487 | return __lat_target_failed(td); | |
1488 | ||
1489 | td->latency_failed++; | |
e39c0676 | 1490 | return false; |
3e260a46 JA |
1491 | } |
1492 | ||
1493 | void lat_target_init(struct thread_data *td) | |
1494 | { | |
6bb58215 JA |
1495 | td->latency_end_run = 0; |
1496 | ||
3e260a46 JA |
1497 | if (td->o.latency_target) { |
1498 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1499 | fio_gettime(&td->latency_ts, NULL); | |
1500 | td->latency_qd = 1; | |
1501 | td->latency_qd_high = td->o.iodepth; | |
1502 | td->latency_qd_low = 1; | |
1503 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1504 | } else | |
1505 | td->latency_qd = td->o.iodepth; | |
1506 | } | |
1507 | ||
6bb58215 JA |
1508 | void lat_target_reset(struct thread_data *td) |
1509 | { | |
1510 | if (!td->latency_end_run) | |
1511 | lat_target_init(td); | |
1512 | } | |
1513 | ||
3e260a46 JA |
1514 | static void lat_target_success(struct thread_data *td) |
1515 | { | |
1516 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1517 | struct thread_options *o = &td->o; |
3e260a46 JA |
1518 | |
1519 | td->latency_qd_low = td->latency_qd; | |
1520 | ||
e1bcd541 SL |
1521 | if (td->latency_qd + 1 == td->latency_qd_high) { |
1522 | /* | |
1523 | * latency_qd will not incease on lat_target_success(), so | |
1524 | * called stable. If we stick with this queue depth, the | |
1525 | * final latency is likely lower than latency_target. Fix | |
1526 | * this by increasing latency_qd_high slowly. Use a naive | |
1527 | * heuristic here. If we get lat_target_success() 3 times | |
1528 | * in a row, increase latency_qd_high by 1. | |
1529 | */ | |
1530 | if (++td->latency_stable_count >= 3) { | |
1531 | td->latency_qd_high++; | |
1532 | td->latency_stable_count = 0; | |
1533 | } | |
1534 | } | |
1535 | ||
3e260a46 JA |
1536 | /* |
1537 | * If we haven't failed yet, we double up to a failing value instead | |
1538 | * of bisecting from highest possible queue depth. If we have set | |
1539 | * a limit other than td->o.iodepth, bisect between that. | |
1540 | */ | |
6bb58215 | 1541 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1542 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1543 | else | |
1544 | td->latency_qd *= 2; | |
1545 | ||
6bb58215 JA |
1546 | if (td->latency_qd > o->iodepth) |
1547 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1548 | |
1549 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1550 | |
3e260a46 | 1551 | /* |
6bb58215 JA |
1552 | * Same as last one, we are done. Let it run a latency cycle, so |
1553 | * we get only the results from the targeted depth. | |
3e260a46 | 1554 | */ |
e1bcd541 | 1555 | if (!o->latency_run && td->latency_qd == qd) { |
6bb58215 JA |
1556 | if (td->latency_end_run) { |
1557 | dprint(FD_RATE, "We are done\n"); | |
1558 | td->done = 1; | |
1559 | } else { | |
1560 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1561 | io_u_quiesce(td); | |
1562 | td->latency_end_run = 1; | |
1563 | reset_all_stats(td); | |
1564 | reset_io_stats(td); | |
1565 | } | |
1566 | } | |
3e260a46 JA |
1567 | |
1568 | lat_new_cycle(td); | |
1569 | } | |
1570 | ||
1571 | /* | |
1572 | * Check if we can bump the queue depth | |
1573 | */ | |
1574 | void lat_target_check(struct thread_data *td) | |
1575 | { | |
1576 | uint64_t usec_window; | |
1577 | uint64_t ios; | |
1578 | double success_ios; | |
1579 | ||
1580 | usec_window = utime_since_now(&td->latency_ts); | |
1581 | if (usec_window < td->o.latency_window) | |
1582 | return; | |
1583 | ||
1584 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1585 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1586 | success_ios *= 100.0; | |
1587 | ||
1588 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1589 | ||
1590 | if (success_ios >= td->o.latency_percentile.u.f) | |
1591 | lat_target_success(td); | |
1592 | else | |
1593 | __lat_target_failed(td); | |
1594 | } | |
1595 | ||
1596 | /* | |
1597 | * If latency target is enabled, we might be ramping up or down and not | |
1598 | * using the full queue depth available. | |
1599 | */ | |
e39c0676 | 1600 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1601 | { |
1602 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1603 | ||
1604 | if (qempty) | |
e39c0676 | 1605 | return true; |
3e260a46 | 1606 | if (!td->o.latency_target) |
e39c0676 | 1607 | return false; |
3e260a46 JA |
1608 | |
1609 | return td->cur_depth >= td->latency_qd; | |
1610 | } | |
429f6675 | 1611 | |
10ba535a JA |
1612 | struct io_u *__get_io_u(struct thread_data *td) |
1613 | { | |
26b3a188 | 1614 | const bool needs_lock = td_async_processing(td); |
0cae66f6 | 1615 | struct io_u *io_u = NULL; |
10ba535a | 1616 | |
ca09be4b JA |
1617 | if (td->stop_io) |
1618 | return NULL; | |
1619 | ||
26b3a188 JA |
1620 | if (needs_lock) |
1621 | __td_io_u_lock(td); | |
e8462bd8 JA |
1622 | |
1623 | again: | |
bba6b14f | 1624 | if (!io_u_rempty(&td->io_u_requeues)) { |
2ae0b204 | 1625 | io_u = io_u_rpop(&td->io_u_requeues); |
bba6b14f JA |
1626 | io_u->resid = 0; |
1627 | } else if (!queue_full(td)) { | |
2ae0b204 | 1628 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1629 | |
225ba9e3 | 1630 | io_u->file = NULL; |
6040dabc | 1631 | io_u->buflen = 0; |
10ba535a | 1632 | io_u->resid = 0; |
d7762cf8 | 1633 | io_u->end_io = NULL; |
755200a3 JA |
1634 | } |
1635 | ||
1636 | if (io_u) { | |
0c6e7517 | 1637 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1638 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 | 1639 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
692dec0c | 1640 | IO_U_F_VER_LIST); |
0c6e7517 | 1641 | |
755200a3 | 1642 | io_u->error = 0; |
bcd5abfa | 1643 | io_u->acct_ddir = -1; |
10ba535a | 1644 | td->cur_depth++; |
a9da8ab2 | 1645 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1646 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1647 | io_u->ipo = NULL; |
a9da8ab2 | 1648 | } else if (td_async_processing(td)) { |
83276370 | 1649 | int ret; |
1dec3e07 JA |
1650 | /* |
1651 | * We ran out, wait for async verify threads to finish and | |
1652 | * return one | |
1653 | */ | |
a9da8ab2 | 1654 | assert(!(td->flags & TD_F_CHILD)); |
93b45bb2 | 1655 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
83276370 DP |
1656 | if (fio_unlikely(ret != 0)) { |
1657 | td->error = errno; | |
1658 | } else if (!td->error) | |
ee5e08ad | 1659 | goto again; |
10ba535a JA |
1660 | } |
1661 | ||
26b3a188 JA |
1662 | if (needs_lock) |
1663 | __td_io_u_unlock(td); | |
1664 | ||
10ba535a JA |
1665 | return io_u; |
1666 | } | |
1667 | ||
e39c0676 | 1668 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1669 | { |
d72be545 | 1670 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1671 | return false; |
c9a73054 JA |
1672 | if (!td->trim_entries) |
1673 | return false; | |
d72be545 | 1674 | |
c9a73054 JA |
1675 | if (td->trim_batch) { |
1676 | td->trim_batch--; | |
1677 | if (get_next_trim(td, io_u)) | |
1678 | return true; | |
1679 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1680 | td->last_ddir != DDIR_READ) { | |
1681 | td->trim_batch = td->o.trim_batch; | |
1682 | if (!td->trim_batch) | |
1683 | td->trim_batch = td->o.trim_backlog; | |
1684 | if (get_next_trim(td, io_u)) | |
e39c0676 | 1685 | return true; |
2ba1c290 | 1686 | } |
10ba535a | 1687 | |
e39c0676 | 1688 | return false; |
0d29de83 JA |
1689 | } |
1690 | ||
e39c0676 | 1691 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1692 | { |
d72be545 | 1693 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1694 | return false; |
d72be545 JA |
1695 | |
1696 | if (td->io_hist_len) { | |
9e144189 JA |
1697 | int get_verify = 0; |
1698 | ||
d1ece0c7 | 1699 | if (td->verify_batch) |
9e144189 | 1700 | get_verify = 1; |
d1ece0c7 | 1701 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1702 | td->last_ddir != DDIR_READ) { |
1703 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1704 | if (!td->verify_batch) |
1705 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1706 | get_verify = 1; |
1707 | } | |
1708 | ||
d1ece0c7 JA |
1709 | if (get_verify && !get_next_verify(td, io_u)) { |
1710 | td->verify_batch--; | |
e39c0676 | 1711 | return true; |
d1ece0c7 | 1712 | } |
9e144189 JA |
1713 | } |
1714 | ||
e39c0676 | 1715 | return false; |
0d29de83 JA |
1716 | } |
1717 | ||
de789769 JA |
1718 | /* |
1719 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1720 | * easy compressible data for simple de-dupe attempts. Do this for every |
1721 | * 512b block in the range, since that should be the smallest block size | |
1722 | * we can expect from a device. | |
de789769 JA |
1723 | */ |
1724 | static void small_content_scramble(struct io_u *io_u) | |
1725 | { | |
5fff9543 | 1726 | unsigned long long i, nr_blocks = io_u->buflen >> 9; |
23f394d5 | 1727 | unsigned int offset; |
319b073f JA |
1728 | uint64_t boffset, *iptr; |
1729 | char *p; | |
de789769 | 1730 | |
23f394d5 JA |
1731 | if (!nr_blocks) |
1732 | return; | |
1733 | ||
1734 | p = io_u->xfer_buf; | |
fba76ee8 | 1735 | boffset = io_u->offset; |
319b073f JA |
1736 | |
1737 | if (io_u->buf_filled_len) | |
1738 | io_u->buf_filled_len = 0; | |
1739 | ||
1740 | /* | |
1741 | * Generate random index between 0..7. We do chunks of 512b, if | |
1742 | * we assume a cacheline is 64 bytes, then we have 8 of those. | |
1743 | * Scramble content within the blocks in the same cacheline to | |
1744 | * speed things up. | |
1745 | */ | |
1746 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; | |
fad82f76 | 1747 | |
23f394d5 JA |
1748 | for (i = 0; i < nr_blocks; i++) { |
1749 | /* | |
319b073f JA |
1750 | * Fill offset into start of cacheline, time into end |
1751 | * of cacheline | |
23f394d5 | 1752 | */ |
319b073f JA |
1753 | iptr = (void *) p + (offset << 6); |
1754 | *iptr = boffset; | |
1755 | ||
1756 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); | |
1757 | iptr[0] = io_u->start_time.tv_sec; | |
1758 | iptr[1] = io_u->start_time.tv_nsec; | |
23f394d5 | 1759 | |
23f394d5 | 1760 | p += 512; |
fad82f76 | 1761 | boffset += 512; |
23f394d5 | 1762 | } |
de789769 JA |
1763 | } |
1764 | ||
0d29de83 JA |
1765 | /* |
1766 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
5c5c33c1 | 1767 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
0d29de83 JA |
1768 | */ |
1769 | struct io_u *get_io_u(struct thread_data *td) | |
1770 | { | |
1771 | struct fio_file *f; | |
1772 | struct io_u *io_u; | |
de789769 | 1773 | int do_scramble = 0; |
002fe734 | 1774 | long ret = 0; |
0d29de83 JA |
1775 | |
1776 | io_u = __get_io_u(td); | |
1777 | if (!io_u) { | |
1778 | dprint(FD_IO, "__get_io_u failed\n"); | |
1779 | return NULL; | |
1780 | } | |
1781 | ||
1782 | if (check_get_verify(td, io_u)) | |
1783 | goto out; | |
1784 | if (check_get_trim(td, io_u)) | |
1785 | goto out; | |
1786 | ||
755200a3 JA |
1787 | /* |
1788 | * from a requeue, io_u already setup | |
1789 | */ | |
1790 | if (io_u->file) | |
77f392bf | 1791 | goto out; |
755200a3 | 1792 | |
429f6675 JA |
1793 | /* |
1794 | * If using an iolog, grab next piece if any available. | |
1795 | */ | |
d72be545 | 1796 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1797 | if (read_iolog_get(td, io_u)) |
1798 | goto err_put; | |
2ba1c290 | 1799 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1800 | ret = -EBUSY; |
2ba1c290 | 1801 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1802 | goto err_put; |
2ba1c290 | 1803 | } |
5ec10eaa | 1804 | |
429f6675 | 1805 | f = io_u->file; |
002fe734 JA |
1806 | if (!f) { |
1807 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1808 | goto err_put; | |
1809 | } | |
1810 | ||
d6aed795 | 1811 | assert(fio_file_open(f)); |
97af62ce | 1812 | |
ff58fced | 1813 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1814 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1815 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1816 | goto err_put; |
2ba1c290 | 1817 | } |
10ba535a | 1818 | |
f1dfb668 JA |
1819 | f->last_start[io_u->ddir] = io_u->offset; |
1820 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1821 | |
fd68418e | 1822 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1823 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1824 | io_u_fill_buffer(td, io_u, |
1066358a | 1825 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1826 | io_u->buflen); |
ff441ae8 | 1827 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
5c5c33c1 BVA |
1828 | !(td->flags & TD_F_COMPRESS) && |
1829 | !(td->flags & TD_F_DO_VERIFY)) | |
fd68418e JA |
1830 | do_scramble = 1; |
1831 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1832 | /* |
1833 | * Reset the buf_filled parameters so next time if the | |
1834 | * buffer is used for writes it is refilled. | |
1835 | */ | |
cbe8d756 RR |
1836 | io_u->buf_filled_len = 0; |
1837 | } | |
87dc1ab1 | 1838 | } |
10ba535a | 1839 | |
165faf16 JA |
1840 | /* |
1841 | * Set io data pointers. | |
1842 | */ | |
cec6b55d JA |
1843 | io_u->xfer_buf = io_u->buf; |
1844 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1845 | |
03ec570f DLM |
1846 | /* |
1847 | * Remember the issuing context priority. The IO engine may change this. | |
1848 | */ | |
1849 | io_u->ioprio = td->ioprio; | |
f0547200 | 1850 | io_u->clat_prio_index = 0; |
6ac7a331 | 1851 | out: |
0d29de83 | 1852 | assert(io_u->file); |
429f6675 | 1853 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1854 | if (!td->o.disable_lat) |
993bf48b | 1855 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1856 | |
de789769 JA |
1857 | if (do_scramble) |
1858 | small_content_scramble(io_u); | |
03553853 | 1859 | |
429f6675 | 1860 | return io_u; |
36167d82 | 1861 | } |
429f6675 | 1862 | err_put: |
2ba1c290 | 1863 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1864 | put_io_u(td, io_u); |
002fe734 | 1865 | return ERR_PTR(ret); |
10ba535a JA |
1866 | } |
1867 | ||
a9da8ab2 | 1868 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1869 | { |
8b28bd41 | 1870 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1871 | |
8b28bd41 DM |
1872 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1873 | return; | |
5451792e | 1874 | |
5fff9543 | 1875 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n", |
709c8313 RE |
1876 | io_u->file ? " on file " : "", |
1877 | io_u->file ? io_u->file->file_name : "", | |
1878 | strerror(io_u->error), | |
1879 | io_ddir_name(io_u->ddir), | |
1880 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1881 | |
5ad7be56 KD |
1882 | if (td->io_ops->errdetails) { |
1883 | char *err = td->io_ops->errdetails(io_u); | |
1884 | ||
1885 | log_err("fio: %s\n", err); | |
1886 | free(err); | |
1887 | } | |
1888 | ||
5451792e JA |
1889 | if (!td->error) |
1890 | td_verror(td, io_u->error, "io_u error"); | |
1891 | } | |
1892 | ||
a9da8ab2 JA |
1893 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1894 | { | |
1895 | __io_u_log_error(td, io_u); | |
1896 | if (td->parent) | |
094e66cb | 1897 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1898 | } |
1899 | ||
e39c0676 | 1900 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1901 | { |
3ecc8c67 JA |
1902 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1903 | || td->o.gtod_reduce; | |
aba6c951 JA |
1904 | } |
1905 | ||
d7e92306 JA |
1906 | static void trim_block_info(struct thread_data *td, struct io_u *io_u) |
1907 | { | |
1908 | uint32_t *info = io_u_block_info(td, io_u); | |
1909 | ||
1910 | if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE) | |
1911 | return; | |
1912 | ||
1913 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1); | |
1914 | } | |
1915 | ||
c8eeb9df JA |
1916 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1917 | struct io_completion_data *icd, | |
1918 | const enum fio_ddir idx, unsigned int bytes) | |
1919 | { | |
a9da8ab2 | 1920 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1921 | unsigned long long llnsec = 0; |
c8eeb9df | 1922 | |
75dc383e JA |
1923 | if (td->parent) |
1924 | td = td->parent; | |
1925 | ||
132b1ee4 | 1926 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1927 | return; |
1928 | ||
a9da8ab2 | 1929 | if (no_reduce) |
d6bb626e | 1930 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1931 | |
1932 | if (!td->o.disable_lat) { | |
c3a32714 | 1933 | unsigned long long tnsec; |
c8eeb9df | 1934 | |
d6bb626e | 1935 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
03ec570f | 1936 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset, |
692dec0c | 1937 | io_u->ioprio, io_u->clat_prio_index); |
15501535 | 1938 | |
d4afedfd JA |
1939 | if (td->flags & TD_F_PROFILE_OPS) { |
1940 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1941 | ||
1942 | if (ops->io_u_lat) | |
c3a32714 | 1943 | icd->error = ops->io_u_lat(td, tnsec); |
d4afedfd JA |
1944 | } |
1945 | ||
f7cf63bf VR |
1946 | if (ddir_rw(idx)) { |
1947 | if (td->o.max_latency[idx] && tnsec > td->o.max_latency[idx]) | |
1948 | lat_fatal(td, io_u, icd, tnsec, td->o.max_latency[idx]); | |
1949 | if (td->o.latency_target && tnsec > td->o.latency_target) { | |
1950 | if (lat_target_failed(td)) | |
1951 | lat_fatal(td, io_u, icd, tnsec, td->o.latency_target); | |
1952 | } | |
15501535 | 1953 | } |
c8eeb9df JA |
1954 | } |
1955 | ||
a47591e4 JA |
1956 | if (ddir_rw(idx)) { |
1957 | if (!td->o.disable_clat) { | |
03ec570f | 1958 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset, |
692dec0c | 1959 | io_u->ioprio, io_u->clat_prio_index); |
d6bb626e | 1960 | io_u_mark_latency(td, llnsec); |
a47591e4 | 1961 | } |
c8eeb9df | 1962 | |
a47591e4 | 1963 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1964 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1965 | |
a47591e4 JA |
1966 | if (no_reduce && per_unit_log(td->iops_log)) |
1967 | add_iops_sample(td, io_u, bytes); | |
b2b3eefe JA |
1968 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
1969 | add_sync_clat_sample(&td->ts, llnsec); | |
66347cfa | 1970 | |
d7e92306 JA |
1971 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) |
1972 | trim_block_info(td, io_u); | |
c8eeb9df JA |
1973 | } |
1974 | ||
94a6e1bb JA |
1975 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1976 | uint64_t offset, unsigned int bytes) | |
1977 | { | |
1978 | int idx; | |
1979 | ||
639ad1ea JA |
1980 | if (!f) |
1981 | return; | |
1982 | ||
94a6e1bb JA |
1983 | if (f->first_write == -1ULL || offset < f->first_write) |
1984 | f->first_write = offset; | |
1985 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1986 | f->last_write = offset + bytes; | |
1987 | ||
1988 | if (!f->last_write_comp) | |
1989 | return; | |
1990 | ||
1991 | idx = f->last_write_idx++; | |
1992 | f->last_write_comp[idx] = offset; | |
1993 | if (f->last_write_idx == td->o.iodepth) | |
1994 | f->last_write_idx = 0; | |
1995 | } | |
1996 | ||
b2b3eefe JA |
1997 | static bool should_account(struct thread_data *td) |
1998 | { | |
1999 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || | |
2000 | td->runstate == TD_VERIFYING); | |
2001 | } | |
2002 | ||
f8b0bd10 | 2003 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 2004 | struct io_completion_data *icd) |
10ba535a | 2005 | { |
f8b0bd10 JA |
2006 | struct io_u *io_u = *io_u_ptr; |
2007 | enum fio_ddir ddir = io_u->ddir; | |
2008 | struct fio_file *f = io_u->file; | |
10ba535a | 2009 | |
e5f9a813 | 2010 | dprint_io_u(io_u, "complete"); |
2ba1c290 | 2011 | |
0c6e7517 | 2012 | assert(io_u->flags & IO_U_F_FLIGHT); |
ad21b61d | 2013 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK | IO_U_F_PATTERN_DONE); |
f9401285 JA |
2014 | |
2015 | /* | |
2016 | * Mark IO ok to verify | |
2017 | */ | |
2018 | if (io_u->ipo) { | |
890b6656 JA |
2019 | /* |
2020 | * Remove errored entry from the verification list | |
2021 | */ | |
2022 | if (io_u->error) | |
2023 | unlog_io_piece(td, io_u); | |
2024 | else { | |
0337173e BVA |
2025 | atomic_store_release(&io_u->ipo->flags, |
2026 | io_u->ipo->flags & ~IP_F_IN_FLIGHT); | |
890b6656 | 2027 | } |
f9401285 JA |
2028 | } |
2029 | ||
f8b0bd10 | 2030 | if (ddir_sync(ddir)) { |
870ea002 JA |
2031 | if (io_u->error) |
2032 | goto error; | |
2ea93f98 | 2033 | td->last_was_sync = true; |
44f29692 JA |
2034 | if (f) { |
2035 | f->first_write = -1ULL; | |
2036 | f->last_write = -1ULL; | |
2037 | } | |
b2b3eefe JA |
2038 | if (should_account(td)) |
2039 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); | |
87dc1ab1 JA |
2040 | return; |
2041 | } | |
2042 | ||
2ea93f98 | 2043 | td->last_was_sync = false; |
f8b0bd10 | 2044 | td->last_ddir = ddir; |
87dc1ab1 | 2045 | |
f8b0bd10 | 2046 | if (!io_u->error && ddir_rw(ddir)) { |
84abeef7 | 2047 | unsigned long long bytes = io_u->xfer_buflen - io_u->resid; |
b29ee5b3 | 2048 | int ret; |
10ba535a | 2049 | |
7d33649c JA |
2050 | /* |
2051 | * Make sure we notice short IO from here, and requeue them | |
2052 | * appropriately! | |
2053 | */ | |
c4cb947e | 2054 | if (bytes && io_u->resid) { |
7d33649c JA |
2055 | io_u->xfer_buflen = io_u->resid; |
2056 | io_u->xfer_buf += bytes; | |
2057 | io_u->offset += bytes; | |
2058 | td->ts.short_io_u[io_u->ddir]++; | |
2059 | if (io_u->offset < io_u->file->real_file_size) { | |
2060 | requeue_io_u(td, io_u_ptr); | |
2061 | return; | |
2062 | } | |
2063 | } | |
2064 | ||
f8b0bd10 | 2065 | td->io_blocks[ddir]++; |
f8b0bd10 | 2066 | td->io_bytes[ddir] += bytes; |
ae2fafc8 | 2067 | |
e1c325d2 JA |
2068 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
2069 | td->this_io_blocks[ddir]++; | |
f8b0bd10 | 2070 | td->this_io_bytes[ddir] += bytes; |
e1c325d2 | 2071 | } |
f8b0bd10 | 2072 | |
639ad1ea | 2073 | if (ddir == DDIR_WRITE) |
94a6e1bb | 2074 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 2075 | |
b2b3eefe | 2076 | if (should_account(td)) |
f8b0bd10 | 2077 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 2078 | |
f8b0bd10 | 2079 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 2080 | |
d7762cf8 | 2081 | if (io_u->end_io) { |
f8b0bd10 JA |
2082 | ret = io_u->end_io(td, io_u_ptr); |
2083 | io_u = *io_u_ptr; | |
3af6ef39 JA |
2084 | if (ret && !icd->error) |
2085 | icd->error = ret; | |
2086 | } | |
ff58fced | 2087 | } else if (io_u->error) { |
870ea002 | 2088 | error: |
10ba535a | 2089 | icd->error = io_u->error; |
5451792e JA |
2090 | io_u_log_error(td, io_u); |
2091 | } | |
8b28bd41 | 2092 | if (icd->error) { |
f8b0bd10 JA |
2093 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
2094 | ||
8b28bd41 DM |
2095 | if (!td_non_fatal_error(td, eb, icd->error)) |
2096 | return; | |
f8b0bd10 | 2097 | |
f2bba182 RR |
2098 | /* |
2099 | * If there is a non_fatal error, then add to the error count | |
2100 | * and clear all the errors. | |
2101 | */ | |
2102 | update_error_count(td, icd->error); | |
2103 | td_clear_error(td); | |
2104 | icd->error = 0; | |
f8b0bd10 JA |
2105 | if (io_u) |
2106 | io_u->error = 0; | |
f2bba182 | 2107 | } |
10ba535a JA |
2108 | } |
2109 | ||
9520ebb9 JA |
2110 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
2111 | int nr) | |
10ba535a | 2112 | { |
6eaf09d6 | 2113 | int ddir; |
aba6c951 JA |
2114 | |
2115 | if (!gtod_reduce(td)) | |
9520ebb9 | 2116 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 2117 | |
3af6ef39 JA |
2118 | icd->nr = nr; |
2119 | ||
10ba535a | 2120 | icd->error = 0; |
c1f50f76 | 2121 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 2122 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
2123 | } |
2124 | ||
97601024 JA |
2125 | static void ios_completed(struct thread_data *td, |
2126 | struct io_completion_data *icd) | |
36167d82 JA |
2127 | { |
2128 | struct io_u *io_u; | |
2129 | int i; | |
2130 | ||
10ba535a JA |
2131 | for (i = 0; i < icd->nr; i++) { |
2132 | io_u = td->io_ops->event(td, i); | |
2133 | ||
f8b0bd10 | 2134 | io_completed(td, &io_u, icd); |
e8462bd8 | 2135 | |
f8b0bd10 | 2136 | if (io_u) |
e8462bd8 | 2137 | put_io_u(td, io_u); |
10ba535a JA |
2138 | } |
2139 | } | |
97601024 | 2140 | |
191d6634 SK |
2141 | static void io_u_update_bytes_done(struct thread_data *td, |
2142 | struct io_completion_data *icd) | |
2143 | { | |
2144 | int ddir; | |
2145 | ||
2146 | if (td->runstate == TD_VERIFYING) { | |
2147 | td->bytes_verified += icd->bytes_done[DDIR_READ]; | |
2148 | return; | |
2149 | } | |
2150 | ||
2151 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) | |
2152 | td->bytes_done[ddir] += icd->bytes_done[ddir]; | |
2153 | } | |
2154 | ||
e7e6cfb4 JA |
2155 | /* |
2156 | * Complete a single io_u for the sync engines. | |
2157 | */ | |
55312f9f | 2158 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
2159 | { |
2160 | struct io_completion_data icd; | |
2161 | ||
9520ebb9 | 2162 | init_icd(td, &icd, 1); |
f8b0bd10 | 2163 | io_completed(td, &io_u, &icd); |
e8462bd8 | 2164 | |
f8b0bd10 | 2165 | if (io_u) |
e8462bd8 | 2166 | put_io_u(td, io_u); |
97601024 | 2167 | |
581e7141 JA |
2168 | if (icd.error) { |
2169 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2170 | return -1; | |
2171 | } | |
97601024 | 2172 | |
191d6634 | 2173 | io_u_update_bytes_done(td, &icd); |
581e7141 JA |
2174 | |
2175 | return 0; | |
97601024 JA |
2176 | } |
2177 | ||
e7e6cfb4 JA |
2178 | /* |
2179 | * Called to complete min_events number of io for the async engines. | |
2180 | */ | |
55312f9f | 2181 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2182 | { |
97601024 | 2183 | struct io_completion_data icd; |
00de55ef | 2184 | struct timespec *tvp = NULL; |
191d6634 | 2185 | int ret; |
4d06a338 | 2186 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2187 | |
12bb8569 | 2188 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2189 | |
4950421a | 2190 | if (!min_evts) |
00de55ef | 2191 | tvp = &ts; |
5fb4b366 RE |
2192 | else if (min_evts > td->cur_depth) |
2193 | min_evts = td->cur_depth; | |
97601024 | 2194 | |
82407585 RP |
2195 | /* No worries, td_io_getevents fixes min and max if they are |
2196 | * set incorrectly */ | |
2197 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2198 | if (ret < 0) { |
e1161c32 | 2199 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2200 | return ret; |
2201 | } else if (!ret) | |
2202 | return ret; | |
2203 | ||
9520ebb9 | 2204 | init_icd(td, &icd, ret); |
97601024 | 2205 | ios_completed(td, &icd); |
581e7141 JA |
2206 | if (icd.error) { |
2207 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2208 | return -1; | |
2209 | } | |
97601024 | 2210 | |
191d6634 | 2211 | io_u_update_bytes_done(td, &icd); |
581e7141 | 2212 | |
0d593542 | 2213 | return ret; |
97601024 | 2214 | } |
7e77dd02 JA |
2215 | |
2216 | /* | |
2217 | * Call when io_u is really queued, to update the submission latency. | |
2218 | */ | |
2219 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2220 | { | |
8243be59 | 2221 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2222 | unsigned long slat_time; |
7e77dd02 | 2223 | |
d6bb626e | 2224 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2225 | |
2226 | if (td->parent) | |
2227 | td = td->parent; | |
2228 | ||
ae588852 | 2229 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
03ec570f | 2230 | io_u->offset, io_u->ioprio); |
9520ebb9 | 2231 | } |
7e77dd02 | 2232 | } |
433afcb4 | 2233 | |
5c94b008 JA |
2234 | /* |
2235 | * See if we should reuse the last seed, if dedupe is enabled | |
2236 | */ | |
9451b93e | 2237 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2238 | { |
2239 | unsigned int v; | |
0d71aa98 | 2240 | unsigned long long i; |
5c94b008 JA |
2241 | |
2242 | if (!td->o.dedupe_percentage) | |
2243 | return &td->buf_state; | |
732eedd0 | 2244 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2245 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2246 | return &td->buf_state; | |
732eedd0 | 2247 | } |
5c94b008 | 2248 | |
1bd5d213 | 2249 | v = rand_between(&td->dedupe_state, 1, 100); |
5c94b008 | 2250 | |
0d71aa98 BD |
2251 | if (v <= td->o.dedupe_percentage) |
2252 | switch (td->o.dedupe_mode) { | |
2253 | case DEDUPE_MODE_REPEAT: | |
2254 | /* | |
2255 | * The caller advances the returned frand_state. | |
2256 | * A copy of prev should be returned instead since | |
2257 | * a subsequent intention to generate a deduped buffer | |
2258 | * might result in generating a unique one | |
2259 | */ | |
2260 | frand_copy(&td->buf_state_ret, &td->buf_state_prev); | |
2261 | return &td->buf_state_ret; | |
2262 | case DEDUPE_MODE_WORKING_SET: | |
2263 | i = rand_between(&td->dedupe_working_set_index_state, 0, td->num_unique_pages - 1); | |
2264 | frand_copy(&td->buf_state_ret, &td->dedupe_working_set_states[i]); | |
2265 | return &td->buf_state_ret; | |
2266 | default: | |
2267 | log_err("unexpected dedupe mode %u\n", td->o.dedupe_mode); | |
2268 | assert(0); | |
2269 | } | |
5c94b008 JA |
2270 | |
2271 | return &td->buf_state; | |
2272 | } | |
2273 | ||
9451b93e | 2274 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2275 | { |
9451b93e JA |
2276 | if (td->o.dedupe_percentage == 100) |
2277 | frand_copy(rs, &td->buf_state_prev); | |
2278 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2279 | frand_copy(&td->buf_state_prev, rs); |
2280 | } | |
2281 | ||
5fff9543 JF |
2282 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write, |
2283 | unsigned long long max_bs) | |
5973cafb | 2284 | { |
d1af2894 JA |
2285 | struct thread_options *o = &td->o; |
2286 | ||
15600335 JA |
2287 | if (o->mem_type == MEM_CUDA_MALLOC) |
2288 | return; | |
03553853 | 2289 | |
4eff3e57 | 2290 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2291 | unsigned int perc = td->o.compress_percentage; |
eb57e710 | 2292 | struct frand_state *rs = NULL; |
5fff9543 JF |
2293 | unsigned long long left = max_bs; |
2294 | unsigned long long this_write; | |
5c94b008 | 2295 | |
1066358a | 2296 | do { |
eb57e710 BD |
2297 | /* |
2298 | * Buffers are either entirely dedupe-able or not. | |
2299 | * If we choose to dedup, the buffer should undergo | |
2300 | * the same manipulation as the original write. Which | |
2301 | * means we should retrack the steps we took for compression | |
2302 | * as well. | |
2303 | */ | |
2304 | if (!rs) | |
2305 | rs = get_buf_state(td); | |
9c42684e | 2306 | |
1066358a | 2307 | min_write = min(min_write, left); |
f97a43a1 | 2308 | |
eb57e710 BD |
2309 | this_write = min_not_zero(min_write, |
2310 | (unsigned long long) td->o.compress_chunk); | |
1e7f82e2 | 2311 | |
eb57e710 BD |
2312 | fill_random_buf_percentage(rs, buf, perc, |
2313 | this_write, this_write, | |
2314 | o->buffer_pattern, | |
2315 | o->buffer_pattern_bytes); | |
1066358a | 2316 | |
1e7f82e2 JA |
2317 | buf += this_write; |
2318 | left -= this_write; | |
9451b93e | 2319 | save_buf_state(td, rs); |
1066358a | 2320 | } while (left); |
d1af2894 JA |
2321 | } else if (o->buffer_pattern_bytes) |
2322 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2323 | else if (o->zero_buffers) |
cc86c395 | 2324 | memset(buf, 0, max_bs); |
999d245e | 2325 | else |
9451b93e | 2326 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2327 | } |
2328 | ||
2329 | /* | |
2330 | * "randomly" fill the buffer contents | |
2331 | */ | |
2332 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
5fff9543 | 2333 | unsigned long long min_write, unsigned long long max_bs) |
cc86c395 JA |
2334 | { |
2335 | io_u->buf_filled_len = 0; | |
2336 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2337 | } |
e2c75fc4 TK |
2338 | |
2339 | static int do_sync_file_range(const struct thread_data *td, | |
2340 | struct fio_file *f) | |
2341 | { | |
7e2a317e | 2342 | uint64_t offset, nbytes; |
e2c75fc4 TK |
2343 | |
2344 | offset = f->first_write; | |
2345 | nbytes = f->last_write - f->first_write; | |
2346 | ||
2347 | if (!nbytes) | |
2348 | return 0; | |
2349 | ||
2350 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2351 | } | |
2352 | ||
2353 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2354 | { | |
2355 | int ret; | |
2356 | ||
2357 | if (io_u->ddir == DDIR_SYNC) { | |
a04e0665 | 2358 | #ifdef CONFIG_FCNTL_SYNC |
c99c81ad | 2359 | ret = fcntl(io_u->file->fd, F_FULLFSYNC); |
a04e0665 | 2360 | #else |
e2c75fc4 | 2361 | ret = fsync(io_u->file->fd); |
a04e0665 | 2362 | #endif |
e2c75fc4 TK |
2363 | } else if (io_u->ddir == DDIR_DATASYNC) { |
2364 | #ifdef CONFIG_FDATASYNC | |
2365 | ret = fdatasync(io_u->file->fd); | |
2366 | #else | |
2367 | ret = io_u->xfer_buflen; | |
2368 | io_u->error = EINVAL; | |
2369 | #endif | |
2370 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2371 | ret = do_sync_file_range(td, io_u->file); | |
2372 | else { | |
2373 | ret = io_u->xfer_buflen; | |
2374 | io_u->error = EINVAL; | |
2375 | } | |
2376 | ||
2377 | if (ret < 0) | |
2378 | io_u->error = errno; | |
2379 | ||
2380 | return ret; | |
2381 | } | |
2382 | ||
67282020 | 2383 | int do_io_u_trim(struct thread_data *td, struct io_u *io_u) |
e2c75fc4 TK |
2384 | { |
2385 | #ifndef FIO_HAVE_TRIM | |
2386 | io_u->error = EINVAL; | |
2387 | return 0; | |
2388 | #else | |
2389 | struct fio_file *f = io_u->file; | |
2390 | int ret; | |
2391 | ||
e3be810b SK |
2392 | if (td->o.zone_mode == ZONE_MODE_ZBD) { |
2393 | ret = zbd_do_io_u_trim(td, io_u); | |
2394 | if (ret == io_u_completed) | |
2395 | return io_u->xfer_buflen; | |
2396 | if (ret) | |
2397 | goto err; | |
2398 | } | |
2399 | ||
496b1f9e | 2400 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2401 | if (!ret) |
2402 | return io_u->xfer_buflen; | |
2403 | ||
e3be810b | 2404 | err: |
e2c75fc4 TK |
2405 | io_u->error = ret; |
2406 | return 0; | |
2407 | #endif | |
2408 | } |