Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
5973cafb | 9 | #include "hash.h" |
4f5af7b2 | 10 | #include "verify.h" |
0d29de83 | 11 | #include "trim.h" |
1fbbf72e | 12 | #include "lib/rand.h" |
7ebd796f | 13 | #include "lib/axmap.h" |
002fe734 | 14 | #include "err.h" |
10ba535a | 15 | |
97601024 JA |
16 | struct io_completion_data { |
17 | int nr; /* input */ | |
97601024 JA |
18 | |
19 | int error; /* output */ | |
100f49f1 | 20 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
97601024 JA |
21 | struct timeval time; /* output */ |
22 | }; | |
23 | ||
10ba535a | 24 | /* |
7ebd796f | 25 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
26 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
27 | */ | |
1ae83d45 | 28 | static int random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 29 | { |
7ebd796f | 30 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
31 | } |
32 | ||
df415585 JA |
33 | /* |
34 | * Mark a given offset as used in the map. | |
35 | */ | |
9bf2061e | 36 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 37 | { |
2dc1bbeb | 38 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 39 | struct fio_file *f = io_u->file; |
51ede0b1 | 40 | unsigned int nr_blocks; |
1ae83d45 | 41 | uint64_t block; |
df415585 | 42 | |
1ae83d45 | 43 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 44 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 45 | |
2ab9e98b | 46 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 47 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 48 | |
51ede0b1 JA |
49 | if ((nr_blocks * min_bs) < io_u->buflen) |
50 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
51 | } |
52 | ||
74776733 JA |
53 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
54 | enum fio_ddir ddir) | |
2ba1c290 | 55 | { |
74776733 JA |
56 | uint64_t max_blocks; |
57 | uint64_t max_size; | |
2ba1c290 | 58 | |
ff58fced JA |
59 | assert(ddir_rw(ddir)); |
60 | ||
d9dd70f7 JA |
61 | /* |
62 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
63 | */ | |
64 | max_size = f->io_size; | |
65 | if (max_size > f->real_file_size) | |
66 | max_size = f->real_file_size; | |
67 | ||
ed335855 SN |
68 | if (td->o.zone_range) |
69 | max_size = td->o.zone_range; | |
70 | ||
1ae83d45 | 71 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
72 | if (!max_blocks) |
73 | return 0; | |
74 | ||
67778e88 | 75 | return max_blocks; |
2ba1c290 JA |
76 | } |
77 | ||
1ae83d45 JA |
78 | struct rand_off { |
79 | struct flist_head list; | |
80 | uint64_t off; | |
81 | }; | |
82 | ||
e25839d4 | 83 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 84 | enum fio_ddir ddir, uint64_t *b) |
ec4015da | 85 | { |
6f49f8bc | 86 | uint64_t r; |
5e0baa7f | 87 | |
74776733 | 88 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) { |
6f49f8bc JA |
89 | uint64_t lastb; |
90 | ||
91 | lastb = last_block(td, f, ddir); | |
92 | if (!lastb) | |
93 | return 1; | |
94 | ||
d6b72507 | 95 | r = __rand(&td->random_state); |
8055e41d | 96 | |
4b91ee8f | 97 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 98 | |
54a21917 | 99 | *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0)); |
51ede0b1 | 100 | } else { |
8055e41d | 101 | uint64_t off = 0; |
43c63a78 | 102 | |
6f49f8bc | 103 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 104 | return 1; |
ec4015da | 105 | |
8055e41d JA |
106 | *b = off; |
107 | } | |
0ce8b119 | 108 | |
ec4015da | 109 | /* |
51ede0b1 | 110 | * if we are not maintaining a random map, we are done. |
ec4015da | 111 | */ |
51ede0b1 JA |
112 | if (!file_randommap(td, f)) |
113 | goto ret; | |
43c63a78 JA |
114 | |
115 | /* | |
51ede0b1 | 116 | * calculate map offset and check if it's free |
43c63a78 | 117 | */ |
51ede0b1 JA |
118 | if (random_map_free(f, *b)) |
119 | goto ret; | |
120 | ||
4b91ee8f JA |
121 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
122 | (unsigned long long) *b); | |
51ede0b1 | 123 | |
7ebd796f | 124 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
125 | if (*b == (uint64_t) -1ULL) |
126 | return 1; | |
0ce8b119 JA |
127 | ret: |
128 | return 0; | |
ec4015da JA |
129 | } |
130 | ||
925fee33 JA |
131 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
132 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 133 | uint64_t *b) |
e25839d4 | 134 | { |
9c6f6316 | 135 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
136 | return 0; |
137 | } | |
138 | ||
925fee33 JA |
139 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
140 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 141 | uint64_t *b) |
925fee33 | 142 | { |
9c6f6316 | 143 | *b = pareto_next(&f->zipf); |
925fee33 JA |
144 | return 0; |
145 | } | |
146 | ||
1ae83d45 JA |
147 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
148 | { | |
149 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
150 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
151 | ||
152 | return r1->off - r2->off; | |
153 | } | |
154 | ||
155 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
156 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
157 | { |
158 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
159 | return __get_next_rand_offset(td, f, ddir, b); | |
160 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
161 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
162 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
163 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
e25839d4 JA |
164 | |
165 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
166 | return 1; | |
167 | } | |
168 | ||
bcd5abfa JA |
169 | /* |
170 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
171 | * specified. | |
172 | */ | |
173 | static inline int should_sort_io(struct thread_data *td) | |
174 | { | |
175 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
176 | return 0; | |
177 | if (!td_random(td)) | |
178 | return 0; | |
179 | if (td->runstate != TD_VERIFYING) | |
180 | return 0; | |
181 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
182 | return 0; | |
183 | ||
184 | return 1; | |
185 | } | |
186 | ||
d9472271 | 187 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
188 | { |
189 | unsigned int v; | |
190 | unsigned long r; | |
191 | ||
d9472271 | 192 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
193 | return 1; |
194 | ||
d6b72507 | 195 | r = __rand(&td->seq_rand_state[ddir]); |
54a21917 | 196 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
211c9b89 | 197 | |
d9472271 | 198 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
199 | } |
200 | ||
1ae83d45 JA |
201 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
202 | enum fio_ddir ddir, uint64_t *b) | |
203 | { | |
204 | struct rand_off *r; | |
205 | int i, ret = 1; | |
206 | ||
bcd5abfa | 207 | if (!should_sort_io(td)) |
1ae83d45 JA |
208 | return get_off_from_method(td, f, ddir, b); |
209 | ||
210 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 211 | fetch: |
9342d5f8 | 212 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
213 | flist_del(&r->list); |
214 | *b = r->off; | |
215 | free(r); | |
216 | return 0; | |
217 | } | |
218 | ||
219 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
220 | r = malloc(sizeof(*r)); | |
221 | ||
222 | ret = get_off_from_method(td, f, ddir, &r->off); | |
223 | if (ret) { | |
224 | free(r); | |
225 | break; | |
226 | } | |
227 | ||
228 | flist_add(&r->list, &td->next_rand_list); | |
229 | } | |
230 | ||
231 | if (ret && !i) | |
232 | return ret; | |
233 | ||
234 | assert(!flist_empty(&td->next_rand_list)); | |
235 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
236 | goto fetch; | |
237 | } | |
238 | ||
38dad62d | 239 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 240 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 241 | { |
c04e4661 DE |
242 | if (!get_next_rand_offset(td, f, ddir, b)) |
243 | return 0; | |
244 | ||
245 | if (td->o.time_based) { | |
33c48814 | 246 | fio_file_reset(td, f); |
c04e4661 DE |
247 | if (!get_next_rand_offset(td, f, ddir, b)) |
248 | return 0; | |
38dad62d JA |
249 | } |
250 | ||
c04e4661 | 251 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
4b91ee8f JA |
252 | f->file_name, (unsigned long long) f->last_pos, |
253 | (unsigned long long) f->real_file_size); | |
c04e4661 | 254 | return 1; |
38dad62d JA |
255 | } |
256 | ||
37cf9e3c | 257 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 258 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 259 | { |
ac002339 JA |
260 | struct thread_options *o = &td->o; |
261 | ||
ff58fced JA |
262 | assert(ddir_rw(ddir)); |
263 | ||
ac002339 JA |
264 | if (f->last_pos >= f->io_size + get_start_offset(td, f) && |
265 | o->time_based) | |
c04e4661 DE |
266 | f->last_pos = f->last_pos - f->io_size; |
267 | ||
38dad62d | 268 | if (f->last_pos < f->real_file_size) { |
1ae83d45 | 269 | uint64_t pos; |
059b0802 | 270 | |
ac002339 | 271 | if (f->last_pos == f->file_offset && o->ddir_seq_add < 0) |
a66da7a2 JA |
272 | f->last_pos = f->real_file_size; |
273 | ||
274 | pos = f->last_pos - f->file_offset; | |
ac002339 JA |
275 | if (pos && o->ddir_seq_add) { |
276 | pos += o->ddir_seq_add; | |
277 | ||
278 | /* | |
279 | * If we reach beyond the end of the file | |
280 | * with holed IO, wrap around to the | |
281 | * beginning again. | |
282 | */ | |
283 | if (pos >= f->real_file_size) | |
284 | pos = f->file_offset; | |
285 | } | |
059b0802 | 286 | |
37cf9e3c | 287 | *offset = pos; |
38dad62d JA |
288 | return 0; |
289 | } | |
290 | ||
291 | return 1; | |
292 | } | |
293 | ||
294 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
295 | enum fio_ddir ddir, int rw_seq, |
296 | unsigned int *is_random) | |
38dad62d JA |
297 | { |
298 | struct fio_file *f = io_u->file; | |
1ae83d45 | 299 | uint64_t b, offset; |
38dad62d JA |
300 | int ret; |
301 | ||
ff58fced JA |
302 | assert(ddir_rw(ddir)); |
303 | ||
37cf9e3c JA |
304 | b = offset = -1ULL; |
305 | ||
38dad62d | 306 | if (rw_seq) { |
211c9b89 | 307 | if (td_random(td)) { |
6aca9b3d | 308 | if (should_do_random(td, ddir)) { |
211c9b89 | 309 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
310 | *is_random = 1; |
311 | } else { | |
312 | *is_random = 0; | |
211c9b89 JA |
313 | io_u->flags |= IO_U_F_BUSY_OK; |
314 | ret = get_next_seq_offset(td, f, ddir, &offset); | |
315 | if (ret) | |
316 | ret = get_next_rand_block(td, f, ddir, &b); | |
317 | } | |
6aca9b3d JA |
318 | } else { |
319 | *is_random = 0; | |
37cf9e3c | 320 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 321 | } |
38dad62d JA |
322 | } else { |
323 | io_u->flags |= IO_U_F_BUSY_OK; | |
6aca9b3d | 324 | *is_random = 0; |
38dad62d JA |
325 | |
326 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 327 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 328 | if (ret) { |
37cf9e3c | 329 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
330 | *is_random = 0; |
331 | } | |
38dad62d JA |
332 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
333 | if (f->last_start != -1ULL) | |
37cf9e3c | 334 | offset = f->last_start - f->file_offset; |
38dad62d | 335 | else |
37cf9e3c | 336 | offset = 0; |
38dad62d JA |
337 | ret = 0; |
338 | } else { | |
339 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
340 | ret = 1; | |
341 | } | |
342 | } | |
6d68b997 | 343 | |
37cf9e3c JA |
344 | if (!ret) { |
345 | if (offset != -1ULL) | |
346 | io_u->offset = offset; | |
347 | else if (b != -1ULL) | |
348 | io_u->offset = b * td->o.ba[ddir]; | |
349 | else { | |
4e0a8fa2 | 350 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
351 | ret = 1; |
352 | } | |
353 | } | |
354 | ||
38dad62d JA |
355 | return ret; |
356 | } | |
357 | ||
10ba535a JA |
358 | /* |
359 | * For random io, generate a random new block and see if it's used. Repeat | |
360 | * until we find a free one. For sequential io, just return the end of | |
361 | * the last io issued. | |
362 | */ | |
6aca9b3d JA |
363 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
364 | unsigned int *is_random) | |
10ba535a | 365 | { |
9bf2061e | 366 | struct fio_file *f = io_u->file; |
4ba66134 | 367 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 368 | int rw_seq_hit = 0; |
10ba535a | 369 | |
ff58fced JA |
370 | assert(ddir_rw(ddir)); |
371 | ||
38dad62d JA |
372 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
373 | rw_seq_hit = 1; | |
5736c10d | 374 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 375 | } |
211097b2 | 376 | |
6aca9b3d | 377 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 378 | return 1; |
10ba535a | 379 | |
009bd847 JA |
380 | if (io_u->offset >= f->io_size) { |
381 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
382 | (unsigned long long) io_u->offset, |
383 | (unsigned long long) f->io_size); | |
009bd847 JA |
384 | return 1; |
385 | } | |
386 | ||
387 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
388 | if (io_u->offset >= f->real_file_size) { |
389 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
390 | (unsigned long long) io_u->offset, |
391 | (unsigned long long) f->real_file_size); | |
10ba535a | 392 | return 1; |
2ba1c290 | 393 | } |
10ba535a JA |
394 | |
395 | return 0; | |
396 | } | |
397 | ||
6aca9b3d JA |
398 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
399 | unsigned int *is_random) | |
15dc1934 | 400 | { |
d72be545 JA |
401 | if (td->flags & TD_F_PROFILE_OPS) { |
402 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 403 | |
d72be545 | 404 | if (ops->fill_io_u_off) |
6aca9b3d | 405 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 406 | } |
15dc1934 | 407 | |
6aca9b3d | 408 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
409 | } |
410 | ||
79944128 JA |
411 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
412 | unsigned int buflen) | |
413 | { | |
414 | struct fio_file *f = io_u->file; | |
415 | ||
bedc9dc2 | 416 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
417 | } |
418 | ||
6aca9b3d JA |
419 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
420 | unsigned int is_random) | |
10ba535a | 421 | { |
6aca9b3d | 422 | int ddir = io_u->ddir; |
24d23ca7 | 423 | unsigned int buflen = 0; |
f3059de1 | 424 | unsigned int minbs, maxbs; |
54a21917 | 425 | unsigned long r; |
10ba535a | 426 | |
9ee1c647 | 427 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
428 | |
429 | if (td->o.bs_is_seq_rand) | |
430 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 431 | |
f3059de1 JA |
432 | minbs = td->o.min_bs[ddir]; |
433 | maxbs = td->o.max_bs[ddir]; | |
434 | ||
79944128 JA |
435 | if (minbs == maxbs) |
436 | return minbs; | |
437 | ||
52c58027 JA |
438 | /* |
439 | * If we can't satisfy the min block size from here, then fail | |
440 | */ | |
441 | if (!io_u_fits(td, io_u, minbs)) | |
442 | return 0; | |
443 | ||
79944128 | 444 | do { |
d6b72507 | 445 | r = __rand(&td->bsrange_state); |
4c07ad86 | 446 | |
720e84ad | 447 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 448 | buflen = 1 + (unsigned int) ((double) maxbs * |
54a21917 | 449 | (r / (FRAND_MAX + 1.0))); |
f3059de1 JA |
450 | if (buflen < minbs) |
451 | buflen = minbs; | |
5ec10eaa | 452 | } else { |
564ca972 JA |
453 | long perc = 0; |
454 | unsigned int i; | |
455 | ||
720e84ad JA |
456 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
457 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
458 | |
459 | buflen = bsp->bs; | |
460 | perc += bsp->perc; | |
54a21917 | 461 | if ((r <= ((FRAND_MAX / 100L) * perc)) && |
79944128 | 462 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
463 | break; |
464 | } | |
465 | } | |
79944128 | 466 | |
a9f70b1f JB |
467 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
468 | buflen = (buflen + td->o.verify_interval - 1) & | |
469 | ~(td->o.verify_interval - 1); | |
470 | ||
f3059de1 JA |
471 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
472 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 473 | |
79944128 | 474 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 475 | |
10ba535a JA |
476 | return buflen; |
477 | } | |
478 | ||
6aca9b3d JA |
479 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
480 | unsigned int is_random) | |
15dc1934 | 481 | { |
d72be545 JA |
482 | if (td->flags & TD_F_PROFILE_OPS) { |
483 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 484 | |
d72be545 | 485 | if (ops->fill_io_u_size) |
6aca9b3d | 486 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 487 | } |
15dc1934 | 488 | |
6aca9b3d | 489 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
490 | } |
491 | ||
afe24a5a JA |
492 | static void set_rwmix_bytes(struct thread_data *td) |
493 | { | |
afe24a5a JA |
494 | unsigned int diff; |
495 | ||
496 | /* | |
497 | * we do time or byte based switch. this is needed because | |
498 | * buffered writes may issue a lot quicker than they complete, | |
499 | * whereas reads do not. | |
500 | */ | |
e47f799f | 501 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 502 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
503 | } |
504 | ||
505 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
506 | { | |
507 | unsigned int v; | |
1294c3ec | 508 | unsigned long r; |
e47f799f | 509 | |
d6b72507 | 510 | r = __rand(&td->rwmix_state); |
54a21917 | 511 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
4c07ad86 | 512 | |
04c540d9 | 513 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
514 | return DDIR_READ; |
515 | ||
516 | return DDIR_WRITE; | |
afe24a5a JA |
517 | } |
518 | ||
002e7183 JA |
519 | void io_u_quiesce(struct thread_data *td) |
520 | { | |
521 | /* | |
522 | * We are going to sleep, ensure that we flush anything pending as | |
523 | * not to skew our latency numbers. | |
524 | * | |
525 | * Changed to only monitor 'in flight' requests here instead of the | |
526 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
527 | * io's that have been actually submitted to an async engine, | |
528 | * and cur_depth is meaningless for sync engines. | |
529 | */ | |
530 | while (td->io_u_in_flight) { | |
531 | int fio_unused ret; | |
532 | ||
533 | ret = io_u_queued_complete(td, 1, NULL); | |
534 | } | |
535 | } | |
536 | ||
581e7141 JA |
537 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
538 | { | |
539 | enum fio_ddir odir = ddir ^ 1; | |
540 | struct timeval t; | |
541 | long usec; | |
542 | ||
ff58fced JA |
543 | assert(ddir_rw(ddir)); |
544 | ||
315fcfec | 545 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
546 | return ddir; |
547 | ||
548 | /* | |
549 | * We have too much pending sleep in this direction. See if we | |
550 | * should switch. | |
551 | */ | |
315fcfec | 552 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
553 | /* |
554 | * Other direction does not have too much pending, switch | |
555 | */ | |
556 | if (td->rate_pending_usleep[odir] < 100000) | |
557 | return odir; | |
558 | ||
559 | /* | |
560 | * Both directions have pending sleep. Sleep the minimum time | |
561 | * and deduct from both. | |
562 | */ | |
563 | if (td->rate_pending_usleep[ddir] <= | |
564 | td->rate_pending_usleep[odir]) { | |
565 | usec = td->rate_pending_usleep[ddir]; | |
566 | } else { | |
567 | usec = td->rate_pending_usleep[odir]; | |
568 | ddir = odir; | |
569 | } | |
570 | } else | |
571 | usec = td->rate_pending_usleep[ddir]; | |
572 | ||
002e7183 | 573 | io_u_quiesce(td); |
78c1eda5 | 574 | |
581e7141 JA |
575 | fio_gettime(&t, NULL); |
576 | usec_sleep(td, usec); | |
577 | usec = utime_since_now(&t); | |
578 | ||
579 | td->rate_pending_usleep[ddir] -= usec; | |
580 | ||
581 | odir = ddir ^ 1; | |
582 | if (td_rw(td) && __should_check_rate(td, odir)) | |
583 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 584 | |
77b451cc JA |
585 | if (ddir == DDIR_TRIM) |
586 | return DDIR_TRIM; | |
e0224c6b | 587 | |
581e7141 JA |
588 | return ddir; |
589 | } | |
590 | ||
10ba535a JA |
591 | /* |
592 | * Return the data direction for the next io_u. If the job is a | |
593 | * mixed read/write workload, check the rwmix cycle and switch if | |
594 | * necessary. | |
595 | */ | |
1e97cce9 | 596 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 597 | { |
581e7141 JA |
598 | enum fio_ddir ddir; |
599 | ||
5f9099ea JA |
600 | /* |
601 | * see if it's time to fsync | |
602 | */ | |
603 | if (td->o.fsync_blocks && | |
604 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
605 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
606 | return DDIR_SYNC; | |
607 | ||
608 | /* | |
609 | * see if it's time to fdatasync | |
610 | */ | |
611 | if (td->o.fdatasync_blocks && | |
612 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
613 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
614 | return DDIR_DATASYNC; | |
615 | ||
44f29692 JA |
616 | /* |
617 | * see if it's time to sync_file_range | |
618 | */ | |
619 | if (td->sync_file_range_nr && | |
620 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
621 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
622 | return DDIR_SYNC_FILE_RANGE; | |
623 | ||
10ba535a | 624 | if (td_rw(td)) { |
10ba535a JA |
625 | /* |
626 | * Check if it's time to seed a new data direction. | |
627 | */ | |
e4928662 | 628 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
629 | /* |
630 | * Put a top limit on how many bytes we do for | |
631 | * one data direction, to avoid overflowing the | |
632 | * ranges too much | |
633 | */ | |
634 | ddir = get_rand_ddir(td); | |
e47f799f JA |
635 | |
636 | if (ddir != td->rwmix_ddir) | |
637 | set_rwmix_bytes(td); | |
638 | ||
639 | td->rwmix_ddir = ddir; | |
10ba535a | 640 | } |
581e7141 | 641 | ddir = td->rwmix_ddir; |
10ba535a | 642 | } else if (td_read(td)) |
581e7141 | 643 | ddir = DDIR_READ; |
6eaf09d6 | 644 | else if (td_write(td)) |
581e7141 | 645 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
646 | else |
647 | ddir = DDIR_TRIM; | |
581e7141 JA |
648 | |
649 | td->rwmix_ddir = rate_ddir(td, ddir); | |
650 | return td->rwmix_ddir; | |
10ba535a JA |
651 | } |
652 | ||
1ef2b6be JA |
653 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
654 | { | |
bcd5abfa | 655 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
656 | |
657 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
658 | td->o.barrier_blocks && | |
659 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
660 | td->io_issues[DDIR_WRITE]) | |
661 | io_u->flags |= IO_U_F_BARRIER; | |
662 | } | |
663 | ||
e8462bd8 | 664 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 665 | { |
71b84caa | 666 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
667 | |
668 | if (ret) | |
669 | td_verror(td, ret, "file close"); | |
670 | } | |
671 | ||
10ba535a JA |
672 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
673 | { | |
e8462bd8 JA |
674 | td_io_u_lock(td); |
675 | ||
f8b0bd10 | 676 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 677 | put_file_log(td, io_u->file); |
f8b0bd10 | 678 | |
10ba535a | 679 | io_u->file = NULL; |
d7ee2a7d SL |
680 | io_u->flags |= IO_U_F_FREE; |
681 | ||
0c41214f RR |
682 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
683 | td->cur_depth--; | |
2ae0b204 | 684 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
685 | td_io_u_unlock(td); |
686 | td_io_u_free_notify(td); | |
10ba535a JA |
687 | } |
688 | ||
f2bba182 RR |
689 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
690 | { | |
691 | io_u->flags &= ~IO_U_F_FLIGHT; | |
692 | put_io_u(td, io_u); | |
693 | } | |
694 | ||
755200a3 JA |
695 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
696 | { | |
697 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 698 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 699 | |
465221b0 JA |
700 | dprint(FD_IO, "requeue %p\n", __io_u); |
701 | ||
e8462bd8 JA |
702 | td_io_u_lock(td); |
703 | ||
4d2e0f49 | 704 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
705 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
706 | td->io_issues[ddir]--; | |
5ec10eaa | 707 | |
4d2e0f49 | 708 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
709 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
710 | td->cur_depth--; | |
2ae0b204 JA |
711 | |
712 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 713 | td_io_u_unlock(td); |
755200a3 JA |
714 | *io_u = NULL; |
715 | } | |
716 | ||
9bf2061e | 717 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 718 | { |
6aca9b3d JA |
719 | unsigned int is_random; |
720 | ||
b4c5e1ac JA |
721 | if (td->io_ops->flags & FIO_NOIO) |
722 | goto out; | |
723 | ||
1ef2b6be | 724 | set_rw_ddir(td, io_u); |
5f9099ea | 725 | |
87dc1ab1 | 726 | /* |
ff58fced | 727 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 728 | */ |
ff58fced | 729 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 730 | goto out; |
a00735e6 | 731 | |
48f5abd3 JA |
732 | /* |
733 | * See if it's time to switch to a new zone | |
734 | */ | |
13af05ae | 735 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
736 | struct fio_file *f = io_u->file; |
737 | ||
48f5abd3 | 738 | td->zone_bytes = 0; |
4c8be5b1 JA |
739 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
740 | ||
741 | /* | |
742 | * Wrap from the beginning, if we exceed the file size | |
743 | */ | |
744 | if (f->file_offset >= f->real_file_size) | |
745 | f->file_offset = f->real_file_size - f->file_offset; | |
746 | f->last_pos = f->file_offset; | |
48f5abd3 JA |
747 | td->io_skip_bytes += td->o.zone_skip; |
748 | } | |
749 | ||
10ba535a | 750 | /* |
c685b5b2 JA |
751 | * No log, let the seq/rand engine retrieve the next buflen and |
752 | * position. | |
10ba535a | 753 | */ |
6aca9b3d | 754 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 755 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 756 | return 1; |
2ba1c290 | 757 | } |
10ba535a | 758 | |
6aca9b3d | 759 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
760 | if (!io_u->buflen) { |
761 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 762 | return 1; |
2ba1c290 | 763 | } |
bca4ed4d | 764 | |
2ba1c290 JA |
765 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
766 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
767 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
768 | (unsigned long long) io_u->offset, io_u->buflen, | |
769 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 770 | return 1; |
2ba1c290 | 771 | } |
6a5e6884 | 772 | |
bca4ed4d JA |
773 | /* |
774 | * mark entry before potentially trimming io_u | |
775 | */ | |
303032ae | 776 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 777 | mark_random_map(td, io_u); |
bca4ed4d | 778 | |
c38e9468 | 779 | out: |
2ba1c290 | 780 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 781 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 782 | return 0; |
10ba535a JA |
783 | } |
784 | ||
838bc709 JA |
785 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
786 | { | |
2b13e716 | 787 | int idx = 0; |
838bc709 JA |
788 | |
789 | switch (nr) { | |
790 | default: | |
2b13e716 | 791 | idx = 6; |
838bc709 JA |
792 | break; |
793 | case 33 ... 64: | |
2b13e716 | 794 | idx = 5; |
838bc709 JA |
795 | break; |
796 | case 17 ... 32: | |
2b13e716 | 797 | idx = 4; |
838bc709 JA |
798 | break; |
799 | case 9 ... 16: | |
2b13e716 | 800 | idx = 3; |
838bc709 JA |
801 | break; |
802 | case 5 ... 8: | |
2b13e716 | 803 | idx = 2; |
838bc709 JA |
804 | break; |
805 | case 1 ... 4: | |
2b13e716 | 806 | idx = 1; |
838bc709 JA |
807 | case 0: |
808 | break; | |
809 | } | |
810 | ||
2b13e716 | 811 | map[idx]++; |
838bc709 JA |
812 | } |
813 | ||
814 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
815 | { | |
816 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
817 | td->ts.total_submit++; | |
818 | } | |
819 | ||
820 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
821 | { | |
822 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
823 | td->ts.total_complete++; | |
824 | } | |
825 | ||
d8005759 | 826 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 827 | { |
2b13e716 | 828 | int idx = 0; |
71619dc2 JA |
829 | |
830 | switch (td->cur_depth) { | |
831 | default: | |
2b13e716 | 832 | idx = 6; |
a783e61a | 833 | break; |
71619dc2 | 834 | case 32 ... 63: |
2b13e716 | 835 | idx = 5; |
a783e61a | 836 | break; |
71619dc2 | 837 | case 16 ... 31: |
2b13e716 | 838 | idx = 4; |
a783e61a | 839 | break; |
71619dc2 | 840 | case 8 ... 15: |
2b13e716 | 841 | idx = 3; |
a783e61a | 842 | break; |
71619dc2 | 843 | case 4 ... 7: |
2b13e716 | 844 | idx = 2; |
a783e61a | 845 | break; |
71619dc2 | 846 | case 2 ... 3: |
2b13e716 | 847 | idx = 1; |
71619dc2 JA |
848 | case 1: |
849 | break; | |
850 | } | |
851 | ||
2b13e716 | 852 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
853 | } |
854 | ||
04a0feae JA |
855 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
856 | { | |
2b13e716 | 857 | int idx = 0; |
04a0feae JA |
858 | |
859 | assert(usec < 1000); | |
860 | ||
861 | switch (usec) { | |
862 | case 750 ... 999: | |
2b13e716 | 863 | idx = 9; |
04a0feae JA |
864 | break; |
865 | case 500 ... 749: | |
2b13e716 | 866 | idx = 8; |
04a0feae JA |
867 | break; |
868 | case 250 ... 499: | |
2b13e716 | 869 | idx = 7; |
04a0feae JA |
870 | break; |
871 | case 100 ... 249: | |
2b13e716 | 872 | idx = 6; |
04a0feae JA |
873 | break; |
874 | case 50 ... 99: | |
2b13e716 | 875 | idx = 5; |
04a0feae JA |
876 | break; |
877 | case 20 ... 49: | |
2b13e716 | 878 | idx = 4; |
04a0feae JA |
879 | break; |
880 | case 10 ... 19: | |
2b13e716 | 881 | idx = 3; |
04a0feae JA |
882 | break; |
883 | case 4 ... 9: | |
2b13e716 | 884 | idx = 2; |
04a0feae JA |
885 | break; |
886 | case 2 ... 3: | |
2b13e716 | 887 | idx = 1; |
04a0feae JA |
888 | case 0 ... 1: |
889 | break; | |
890 | } | |
891 | ||
2b13e716 JA |
892 | assert(idx < FIO_IO_U_LAT_U_NR); |
893 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
894 | } |
895 | ||
896 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 897 | { |
2b13e716 | 898 | int idx = 0; |
ec118304 JA |
899 | |
900 | switch (msec) { | |
901 | default: | |
2b13e716 | 902 | idx = 11; |
04a0feae | 903 | break; |
8abdce66 | 904 | case 1000 ... 1999: |
2b13e716 | 905 | idx = 10; |
04a0feae | 906 | break; |
8abdce66 | 907 | case 750 ... 999: |
2b13e716 | 908 | idx = 9; |
04a0feae | 909 | break; |
8abdce66 | 910 | case 500 ... 749: |
2b13e716 | 911 | idx = 8; |
04a0feae | 912 | break; |
8abdce66 | 913 | case 250 ... 499: |
2b13e716 | 914 | idx = 7; |
04a0feae | 915 | break; |
8abdce66 | 916 | case 100 ... 249: |
2b13e716 | 917 | idx = 6; |
04a0feae | 918 | break; |
8abdce66 | 919 | case 50 ... 99: |
2b13e716 | 920 | idx = 5; |
04a0feae | 921 | break; |
8abdce66 | 922 | case 20 ... 49: |
2b13e716 | 923 | idx = 4; |
04a0feae | 924 | break; |
8abdce66 | 925 | case 10 ... 19: |
2b13e716 | 926 | idx = 3; |
04a0feae | 927 | break; |
8abdce66 | 928 | case 4 ... 9: |
2b13e716 | 929 | idx = 2; |
04a0feae | 930 | break; |
ec118304 | 931 | case 2 ... 3: |
2b13e716 | 932 | idx = 1; |
ec118304 JA |
933 | case 0 ... 1: |
934 | break; | |
935 | } | |
936 | ||
2b13e716 JA |
937 | assert(idx < FIO_IO_U_LAT_M_NR); |
938 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
939 | } |
940 | ||
941 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
942 | { | |
943 | if (usec < 1000) | |
944 | io_u_mark_lat_usec(td, usec); | |
945 | else | |
946 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
947 | } |
948 | ||
0aabe160 JA |
949 | /* |
950 | * Get next file to service by choosing one at random | |
951 | */ | |
2cc52930 JA |
952 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
953 | enum fio_file_flags goodf, | |
d6aed795 | 954 | enum fio_file_flags badf) |
0aabe160 | 955 | { |
0aabe160 | 956 | struct fio_file *f; |
1c178180 | 957 | int fno; |
0aabe160 JA |
958 | |
959 | do { | |
87b10676 | 960 | int opened = 0; |
1294c3ec | 961 | unsigned long r; |
4c07ad86 | 962 | |
d6b72507 | 963 | r = __rand(&td->next_file_state); |
54a21917 | 964 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 965 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 966 | |
126d65c6 | 967 | f = td->files[fno]; |
d6aed795 | 968 | if (fio_file_done(f)) |
059e63c0 | 969 | continue; |
1c178180 | 970 | |
d6aed795 | 971 | if (!fio_file_open(f)) { |
87b10676 JA |
972 | int err; |
973 | ||
002fe734 JA |
974 | if (td->nr_open_files >= td->o.open_files) |
975 | return ERR_PTR(-EBUSY); | |
976 | ||
87b10676 JA |
977 | err = td_io_open_file(td, f); |
978 | if (err) | |
979 | continue; | |
980 | opened = 1; | |
981 | } | |
982 | ||
2ba1c290 JA |
983 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
984 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 985 | return f; |
2ba1c290 | 986 | } |
87b10676 JA |
987 | if (opened) |
988 | td_io_close_file(td, f); | |
0aabe160 JA |
989 | } while (1); |
990 | } | |
991 | ||
992 | /* | |
993 | * Get next file to service by doing round robin between all available ones | |
994 | */ | |
1c178180 JA |
995 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
996 | int badf) | |
3d7c391d JA |
997 | { |
998 | unsigned int old_next_file = td->next_file; | |
999 | struct fio_file *f; | |
1000 | ||
1001 | do { | |
87b10676 JA |
1002 | int opened = 0; |
1003 | ||
126d65c6 | 1004 | f = td->files[td->next_file]; |
3d7c391d JA |
1005 | |
1006 | td->next_file++; | |
2dc1bbeb | 1007 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1008 | td->next_file = 0; |
1009 | ||
87b10676 | 1010 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1011 | if (fio_file_done(f)) { |
d5ed68ea | 1012 | f = NULL; |
059e63c0 | 1013 | continue; |
d5ed68ea | 1014 | } |
059e63c0 | 1015 | |
d6aed795 | 1016 | if (!fio_file_open(f)) { |
87b10676 JA |
1017 | int err; |
1018 | ||
002fe734 JA |
1019 | if (td->nr_open_files >= td->o.open_files) |
1020 | return ERR_PTR(-EBUSY); | |
1021 | ||
87b10676 | 1022 | err = td_io_open_file(td, f); |
b5696bfc JA |
1023 | if (err) { |
1024 | dprint(FD_FILE, "error %d on open of %s\n", | |
1025 | err, f->file_name); | |
87c27b45 | 1026 | f = NULL; |
87b10676 | 1027 | continue; |
b5696bfc | 1028 | } |
87b10676 JA |
1029 | opened = 1; |
1030 | } | |
1031 | ||
0b9d69ec JA |
1032 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1033 | f->flags); | |
1c178180 | 1034 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1035 | break; |
1036 | ||
87b10676 JA |
1037 | if (opened) |
1038 | td_io_close_file(td, f); | |
1039 | ||
3d7c391d JA |
1040 | f = NULL; |
1041 | } while (td->next_file != old_next_file); | |
1042 | ||
2ba1c290 | 1043 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1044 | return f; |
1045 | } | |
1046 | ||
7eb36574 | 1047 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1048 | { |
1907dbc6 JA |
1049 | struct fio_file *f; |
1050 | ||
2dc1bbeb | 1051 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1052 | |
b5696bfc | 1053 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1054 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1055 | " nr_files=%d\n", td->nr_open_files, | |
1056 | td->nr_done_files, | |
1057 | td->o.nr_files); | |
bdb4e2e9 | 1058 | return NULL; |
2ba1c290 | 1059 | } |
bdb4e2e9 | 1060 | |
1907dbc6 | 1061 | f = td->file_service_file; |
d6aed795 | 1062 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1063 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1064 | goto out; | |
1065 | if (td->file_service_left--) | |
1066 | goto out; | |
1067 | } | |
1907dbc6 | 1068 | |
a086c257 JA |
1069 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1070 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1071 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1072 | else |
d6aed795 | 1073 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1074 | |
002fe734 JA |
1075 | if (IS_ERR(f)) |
1076 | return f; | |
1077 | ||
1907dbc6 JA |
1078 | td->file_service_file = f; |
1079 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1080 | out: |
0dac421f JA |
1081 | if (f) |
1082 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1083 | else | |
1084 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1085 | return f; |
bdb4e2e9 JA |
1086 | } |
1087 | ||
7eb36574 JA |
1088 | static struct fio_file *get_next_file(struct thread_data *td) |
1089 | { | |
372d8962 | 1090 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1091 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1092 | |
d72be545 JA |
1093 | if (ops->get_next_file) |
1094 | return ops->get_next_file(td); | |
1095 | } | |
7eb36574 JA |
1096 | |
1097 | return __get_next_file(td); | |
1098 | } | |
1099 | ||
002fe734 | 1100 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1101 | { |
1102 | struct fio_file *f; | |
1103 | ||
1104 | do { | |
1105 | f = get_next_file(td); | |
002fe734 JA |
1106 | if (IS_ERR_OR_NULL(f)) |
1107 | return PTR_ERR(f); | |
429f6675 | 1108 | |
429f6675 JA |
1109 | io_u->file = f; |
1110 | get_file(f); | |
1111 | ||
1112 | if (!fill_io_u(td, io_u)) | |
1113 | break; | |
1114 | ||
b5696bfc | 1115 | put_file_log(td, f); |
429f6675 | 1116 | td_io_close_file(td, f); |
b5696bfc | 1117 | io_u->file = NULL; |
d6aed795 | 1118 | fio_file_set_done(f); |
429f6675 | 1119 | td->nr_done_files++; |
0b9d69ec JA |
1120 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1121 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1122 | } while (1); |
1123 | ||
1124 | return 0; | |
1125 | } | |
1126 | ||
3e260a46 JA |
1127 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1128 | unsigned long tusec, unsigned long max_usec) | |
1129 | { | |
1130 | if (!td->error) | |
1131 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1132 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1133 | icd->error = ETIMEDOUT; | |
1134 | } | |
1135 | ||
1136 | static void lat_new_cycle(struct thread_data *td) | |
1137 | { | |
1138 | fio_gettime(&td->latency_ts, NULL); | |
1139 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1140 | td->latency_failed = 0; | |
1141 | } | |
1142 | ||
1143 | /* | |
1144 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1145 | * are at QD=1, then it's time to give up. | |
1146 | */ | |
1147 | static int __lat_target_failed(struct thread_data *td) | |
1148 | { | |
1149 | if (td->latency_qd == 1) | |
1150 | return 1; | |
1151 | ||
1152 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1153 | |
1154 | if (td->latency_qd == td->latency_qd_low) | |
1155 | td->latency_qd_low--; | |
1156 | ||
3e260a46 JA |
1157 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1158 | ||
1159 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1160 | ||
1161 | /* | |
1162 | * When we ramp QD down, quiesce existing IO to prevent | |
1163 | * a storm of ramp downs due to pending higher depth. | |
1164 | */ | |
1165 | io_u_quiesce(td); | |
1166 | lat_new_cycle(td); | |
1167 | return 0; | |
1168 | } | |
1169 | ||
1170 | static int lat_target_failed(struct thread_data *td) | |
1171 | { | |
1172 | if (td->o.latency_percentile.u.f == 100.0) | |
1173 | return __lat_target_failed(td); | |
1174 | ||
1175 | td->latency_failed++; | |
1176 | return 0; | |
1177 | } | |
1178 | ||
1179 | void lat_target_init(struct thread_data *td) | |
1180 | { | |
6bb58215 JA |
1181 | td->latency_end_run = 0; |
1182 | ||
3e260a46 JA |
1183 | if (td->o.latency_target) { |
1184 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1185 | fio_gettime(&td->latency_ts, NULL); | |
1186 | td->latency_qd = 1; | |
1187 | td->latency_qd_high = td->o.iodepth; | |
1188 | td->latency_qd_low = 1; | |
1189 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1190 | } else | |
1191 | td->latency_qd = td->o.iodepth; | |
1192 | } | |
1193 | ||
6bb58215 JA |
1194 | void lat_target_reset(struct thread_data *td) |
1195 | { | |
1196 | if (!td->latency_end_run) | |
1197 | lat_target_init(td); | |
1198 | } | |
1199 | ||
3e260a46 JA |
1200 | static void lat_target_success(struct thread_data *td) |
1201 | { | |
1202 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1203 | struct thread_options *o = &td->o; |
3e260a46 JA |
1204 | |
1205 | td->latency_qd_low = td->latency_qd; | |
1206 | ||
1207 | /* | |
1208 | * If we haven't failed yet, we double up to a failing value instead | |
1209 | * of bisecting from highest possible queue depth. If we have set | |
1210 | * a limit other than td->o.iodepth, bisect between that. | |
1211 | */ | |
6bb58215 | 1212 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1213 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1214 | else | |
1215 | td->latency_qd *= 2; | |
1216 | ||
6bb58215 JA |
1217 | if (td->latency_qd > o->iodepth) |
1218 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1219 | |
1220 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1221 | |
3e260a46 | 1222 | /* |
6bb58215 JA |
1223 | * Same as last one, we are done. Let it run a latency cycle, so |
1224 | * we get only the results from the targeted depth. | |
3e260a46 | 1225 | */ |
6bb58215 JA |
1226 | if (td->latency_qd == qd) { |
1227 | if (td->latency_end_run) { | |
1228 | dprint(FD_RATE, "We are done\n"); | |
1229 | td->done = 1; | |
1230 | } else { | |
1231 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1232 | io_u_quiesce(td); | |
1233 | td->latency_end_run = 1; | |
1234 | reset_all_stats(td); | |
1235 | reset_io_stats(td); | |
1236 | } | |
1237 | } | |
3e260a46 JA |
1238 | |
1239 | lat_new_cycle(td); | |
1240 | } | |
1241 | ||
1242 | /* | |
1243 | * Check if we can bump the queue depth | |
1244 | */ | |
1245 | void lat_target_check(struct thread_data *td) | |
1246 | { | |
1247 | uint64_t usec_window; | |
1248 | uint64_t ios; | |
1249 | double success_ios; | |
1250 | ||
1251 | usec_window = utime_since_now(&td->latency_ts); | |
1252 | if (usec_window < td->o.latency_window) | |
1253 | return; | |
1254 | ||
1255 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1256 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1257 | success_ios *= 100.0; | |
1258 | ||
1259 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1260 | ||
1261 | if (success_ios >= td->o.latency_percentile.u.f) | |
1262 | lat_target_success(td); | |
1263 | else | |
1264 | __lat_target_failed(td); | |
1265 | } | |
1266 | ||
1267 | /* | |
1268 | * If latency target is enabled, we might be ramping up or down and not | |
1269 | * using the full queue depth available. | |
1270 | */ | |
effd6ff0 | 1271 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1272 | { |
1273 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1274 | ||
1275 | if (qempty) | |
1276 | return 1; | |
1277 | if (!td->o.latency_target) | |
1278 | return 0; | |
1279 | ||
1280 | return td->cur_depth >= td->latency_qd; | |
1281 | } | |
429f6675 | 1282 | |
10ba535a JA |
1283 | struct io_u *__get_io_u(struct thread_data *td) |
1284 | { | |
0cae66f6 | 1285 | struct io_u *io_u = NULL; |
10ba535a | 1286 | |
ca09be4b JA |
1287 | if (td->stop_io) |
1288 | return NULL; | |
1289 | ||
e8462bd8 JA |
1290 | td_io_u_lock(td); |
1291 | ||
1292 | again: | |
2ae0b204 JA |
1293 | if (!io_u_rempty(&td->io_u_requeues)) |
1294 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1295 | else if (!queue_full(td)) { |
2ae0b204 | 1296 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1297 | |
225ba9e3 | 1298 | io_u->file = NULL; |
6040dabc | 1299 | io_u->buflen = 0; |
10ba535a | 1300 | io_u->resid = 0; |
d7762cf8 | 1301 | io_u->end_io = NULL; |
755200a3 JA |
1302 | } |
1303 | ||
1304 | if (io_u) { | |
0c6e7517 | 1305 | assert(io_u->flags & IO_U_F_FREE); |
f8b0bd10 JA |
1306 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
1307 | IO_U_F_TRIMMED | IO_U_F_BARRIER | | |
1308 | IO_U_F_VER_LIST); | |
0c6e7517 | 1309 | |
755200a3 | 1310 | io_u->error = 0; |
bcd5abfa | 1311 | io_u->acct_ddir = -1; |
10ba535a | 1312 | td->cur_depth++; |
0c41214f | 1313 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1314 | io_u->ipo = NULL; |
1dec3e07 JA |
1315 | } else if (td->o.verify_async) { |
1316 | /* | |
1317 | * We ran out, wait for async verify threads to finish and | |
1318 | * return one | |
1319 | */ | |
1320 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1321 | goto again; | |
10ba535a JA |
1322 | } |
1323 | ||
e8462bd8 | 1324 | td_io_u_unlock(td); |
10ba535a JA |
1325 | return io_u; |
1326 | } | |
1327 | ||
0d29de83 | 1328 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1329 | { |
d72be545 JA |
1330 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1331 | return 0; | |
1332 | ||
1333 | if (td->trim_entries) { | |
0d29de83 | 1334 | int get_trim = 0; |
10ba535a | 1335 | |
0d29de83 JA |
1336 | if (td->trim_batch) { |
1337 | td->trim_batch--; | |
1338 | get_trim = 1; | |
1339 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1340 | td->last_ddir != DDIR_READ) { | |
1341 | td->trim_batch = td->o.trim_batch; | |
1342 | if (!td->trim_batch) | |
1343 | td->trim_batch = td->o.trim_backlog; | |
1344 | get_trim = 1; | |
1345 | } | |
1346 | ||
1347 | if (get_trim && !get_next_trim(td, io_u)) | |
1348 | return 1; | |
2ba1c290 | 1349 | } |
10ba535a | 1350 | |
0d29de83 JA |
1351 | return 0; |
1352 | } | |
1353 | ||
1354 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1355 | { | |
d72be545 JA |
1356 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1357 | return 0; | |
1358 | ||
1359 | if (td->io_hist_len) { | |
9e144189 JA |
1360 | int get_verify = 0; |
1361 | ||
d1ece0c7 | 1362 | if (td->verify_batch) |
9e144189 | 1363 | get_verify = 1; |
d1ece0c7 | 1364 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1365 | td->last_ddir != DDIR_READ) { |
1366 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1367 | if (!td->verify_batch) |
1368 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1369 | get_verify = 1; |
1370 | } | |
1371 | ||
d1ece0c7 JA |
1372 | if (get_verify && !get_next_verify(td, io_u)) { |
1373 | td->verify_batch--; | |
0d29de83 | 1374 | return 1; |
d1ece0c7 | 1375 | } |
9e144189 JA |
1376 | } |
1377 | ||
0d29de83 JA |
1378 | return 0; |
1379 | } | |
1380 | ||
de789769 JA |
1381 | /* |
1382 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1383 | * easy compressible data for simple de-dupe attempts. Do this for every |
1384 | * 512b block in the range, since that should be the smallest block size | |
1385 | * we can expect from a device. | |
de789769 JA |
1386 | */ |
1387 | static void small_content_scramble(struct io_u *io_u) | |
1388 | { | |
23f394d5 | 1389 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1390 | uint64_t boffset; |
23f394d5 JA |
1391 | unsigned int offset; |
1392 | void *p, *end; | |
de789769 | 1393 | |
23f394d5 JA |
1394 | if (!nr_blocks) |
1395 | return; | |
1396 | ||
1397 | p = io_u->xfer_buf; | |
fba76ee8 | 1398 | boffset = io_u->offset; |
81f0366c | 1399 | io_u->buf_filled_len = 0; |
fad82f76 | 1400 | |
23f394d5 JA |
1401 | for (i = 0; i < nr_blocks; i++) { |
1402 | /* | |
1403 | * Fill the byte offset into a "random" start offset of | |
1404 | * the buffer, given by the product of the usec time | |
1405 | * and the actual offset. | |
1406 | */ | |
fad82f76 | 1407 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1408 | offset &= ~(sizeof(uint64_t) - 1); |
1409 | if (offset >= 512 - sizeof(uint64_t)) | |
1410 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1411 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1412 | |
1413 | end = p + 512 - sizeof(io_u->start_time); | |
1414 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1415 | p += 512; | |
fad82f76 | 1416 | boffset += 512; |
23f394d5 | 1417 | } |
de789769 JA |
1418 | } |
1419 | ||
0d29de83 JA |
1420 | /* |
1421 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1422 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1423 | */ | |
1424 | struct io_u *get_io_u(struct thread_data *td) | |
1425 | { | |
1426 | struct fio_file *f; | |
1427 | struct io_u *io_u; | |
de789769 | 1428 | int do_scramble = 0; |
002fe734 | 1429 | long ret = 0; |
0d29de83 JA |
1430 | |
1431 | io_u = __get_io_u(td); | |
1432 | if (!io_u) { | |
1433 | dprint(FD_IO, "__get_io_u failed\n"); | |
1434 | return NULL; | |
1435 | } | |
1436 | ||
1437 | if (check_get_verify(td, io_u)) | |
1438 | goto out; | |
1439 | if (check_get_trim(td, io_u)) | |
1440 | goto out; | |
1441 | ||
755200a3 JA |
1442 | /* |
1443 | * from a requeue, io_u already setup | |
1444 | */ | |
1445 | if (io_u->file) | |
77f392bf | 1446 | goto out; |
755200a3 | 1447 | |
429f6675 JA |
1448 | /* |
1449 | * If using an iolog, grab next piece if any available. | |
1450 | */ | |
d72be545 | 1451 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1452 | if (read_iolog_get(td, io_u)) |
1453 | goto err_put; | |
2ba1c290 | 1454 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1455 | ret = -EBUSY; |
2ba1c290 | 1456 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1457 | goto err_put; |
2ba1c290 | 1458 | } |
5ec10eaa | 1459 | |
429f6675 | 1460 | f = io_u->file; |
002fe734 JA |
1461 | if (!f) { |
1462 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1463 | goto err_put; | |
1464 | } | |
1465 | ||
d6aed795 | 1466 | assert(fio_file_open(f)); |
97af62ce | 1467 | |
ff58fced | 1468 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1469 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1470 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1471 | goto err_put; |
2ba1c290 | 1472 | } |
10ba535a | 1473 | |
38dad62d | 1474 | f->last_start = io_u->offset; |
36167d82 | 1475 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1476 | |
fd68418e | 1477 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1478 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1479 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1480 | td->o.min_bs[DDIR_WRITE], |
1481 | io_u->xfer_buflen); | |
bedc9dc2 JA |
1482 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
1483 | !(td->flags & TD_F_COMPRESS)) | |
fd68418e | 1484 | do_scramble = 1; |
d72be545 | 1485 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1486 | populate_verify_io_u(td, io_u); |
1487 | do_scramble = 0; | |
1488 | } | |
fd68418e | 1489 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1490 | /* |
1491 | * Reset the buf_filled parameters so next time if the | |
1492 | * buffer is used for writes it is refilled. | |
1493 | */ | |
cbe8d756 RR |
1494 | io_u->buf_filled_len = 0; |
1495 | } | |
87dc1ab1 | 1496 | } |
10ba535a | 1497 | |
165faf16 JA |
1498 | /* |
1499 | * Set io data pointers. | |
1500 | */ | |
cec6b55d JA |
1501 | io_u->xfer_buf = io_u->buf; |
1502 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1503 | |
6ac7a331 | 1504 | out: |
0d29de83 | 1505 | assert(io_u->file); |
429f6675 | 1506 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1507 | if (!td->o.disable_slat) |
1508 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1509 | if (do_scramble) |
1510 | small_content_scramble(io_u); | |
429f6675 | 1511 | return io_u; |
36167d82 | 1512 | } |
429f6675 | 1513 | err_put: |
2ba1c290 | 1514 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1515 | put_io_u(td, io_u); |
002fe734 | 1516 | return ERR_PTR(ret); |
10ba535a JA |
1517 | } |
1518 | ||
5451792e JA |
1519 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1520 | { | |
8b28bd41 | 1521 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1522 | |
8b28bd41 DM |
1523 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1524 | return; | |
5451792e | 1525 | |
709c8313 RE |
1526 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1527 | io_u->file ? " on file " : "", | |
1528 | io_u->file ? io_u->file->file_name : "", | |
1529 | strerror(io_u->error), | |
1530 | io_ddir_name(io_u->ddir), | |
1531 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1532 | |
1533 | if (!td->error) | |
1534 | td_verror(td, io_u->error, "io_u error"); | |
1535 | } | |
1536 | ||
aba6c951 JA |
1537 | static inline int gtod_reduce(struct thread_data *td) |
1538 | { | |
729fe3af | 1539 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1540 | && td->o.disable_bw; |
aba6c951 JA |
1541 | } |
1542 | ||
c8eeb9df JA |
1543 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1544 | struct io_completion_data *icd, | |
1545 | const enum fio_ddir idx, unsigned int bytes) | |
1546 | { | |
24d23ca7 | 1547 | unsigned long lusec = 0; |
c8eeb9df | 1548 | |
aba6c951 | 1549 | if (!gtod_reduce(td)) |
c8eeb9df JA |
1550 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1551 | ||
1552 | if (!td->o.disable_lat) { | |
1553 | unsigned long tusec; | |
1554 | ||
1555 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1556 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1557 | |
d4afedfd JA |
1558 | if (td->flags & TD_F_PROFILE_OPS) { |
1559 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1560 | ||
1561 | if (ops->io_u_lat) | |
1562 | icd->error = ops->io_u_lat(td, tusec); | |
1563 | } | |
1564 | ||
3e260a46 JA |
1565 | if (td->o.max_latency && tusec > td->o.max_latency) |
1566 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1567 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1568 | if (lat_target_failed(td)) | |
1569 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1570 | } |
c8eeb9df JA |
1571 | } |
1572 | ||
1573 | if (!td->o.disable_clat) { | |
ae588852 | 1574 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1575 | io_u_mark_latency(td, lusec); |
1576 | } | |
1577 | ||
1578 | if (!td->o.disable_bw) | |
1579 | add_bw_sample(td, idx, bytes, &icd->time); | |
1580 | ||
aba6c951 JA |
1581 | if (!gtod_reduce(td)) |
1582 | add_iops_sample(td, idx, bytes, &icd->time); | |
c8eeb9df JA |
1583 | } |
1584 | ||
1b8dbf25 SL |
1585 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1586 | { | |
1ae83d45 JA |
1587 | uint64_t secs, remainder, bps, bytes; |
1588 | ||
1b8dbf25 SL |
1589 | bytes = td->this_io_bytes[ddir]; |
1590 | bps = td->rate_bps[ddir]; | |
1591 | secs = bytes / bps; | |
1592 | remainder = bytes % bps; | |
1593 | return remainder * 1000000 / bps + secs * 1000000; | |
1594 | } | |
1595 | ||
f8b0bd10 | 1596 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1597 | struct io_completion_data *icd) |
10ba535a | 1598 | { |
f8b0bd10 JA |
1599 | struct io_u *io_u = *io_u_ptr; |
1600 | enum fio_ddir ddir = io_u->ddir; | |
1601 | struct fio_file *f = io_u->file; | |
10ba535a | 1602 | |
2ba1c290 JA |
1603 | dprint_io_u(io_u, "io complete"); |
1604 | ||
2ecc1b57 | 1605 | td_io_u_lock(td); |
0c6e7517 | 1606 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1607 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1608 | |
1609 | /* | |
1610 | * Mark IO ok to verify | |
1611 | */ | |
1612 | if (io_u->ipo) { | |
890b6656 JA |
1613 | /* |
1614 | * Remove errored entry from the verification list | |
1615 | */ | |
1616 | if (io_u->error) | |
1617 | unlog_io_piece(td, io_u); | |
1618 | else { | |
1619 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1620 | write_barrier(); | |
1621 | } | |
f9401285 JA |
1622 | } |
1623 | ||
2ecc1b57 | 1624 | td_io_u_unlock(td); |
0c6e7517 | 1625 | |
f8b0bd10 | 1626 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1627 | td->last_was_sync = 1; |
44f29692 JA |
1628 | if (f) { |
1629 | f->first_write = -1ULL; | |
1630 | f->last_write = -1ULL; | |
1631 | } | |
87dc1ab1 JA |
1632 | return; |
1633 | } | |
1634 | ||
1635 | td->last_was_sync = 0; | |
f8b0bd10 | 1636 | td->last_ddir = ddir; |
87dc1ab1 | 1637 | |
f8b0bd10 | 1638 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1639 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1640 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1641 | int ret; |
10ba535a | 1642 | |
f8b0bd10 JA |
1643 | td->io_blocks[ddir]++; |
1644 | td->this_io_blocks[ddir]++; | |
1645 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1646 | |
1647 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1648 | td->this_io_bytes[ddir] += bytes; |
1649 | ||
ca09be4b JA |
1650 | if (ddir == DDIR_WRITE) { |
1651 | if (f) { | |
1652 | if (f->first_write == -1ULL || | |
1653 | io_u->offset < f->first_write) | |
1654 | f->first_write = io_u->offset; | |
1655 | if (f->last_write == -1ULL || | |
1656 | ((io_u->offset + bytes) > f->last_write)) | |
1657 | f->last_write = io_u->offset + bytes; | |
1658 | } | |
1659 | if (td->last_write_comp) { | |
1660 | int idx = td->last_write_idx++; | |
1661 | ||
1662 | td->last_write_comp[idx] = io_u->offset; | |
1663 | if (td->last_write_idx == td->o.iodepth) | |
1664 | td->last_write_idx = 0; | |
1665 | } | |
44f29692 JA |
1666 | } |
1667 | ||
6b1190fd SL |
1668 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1669 | td->runstate == TD_VERIFYING)) { | |
f8b0bd10 | 1670 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1671 | |
f8b0bd10 JA |
1672 | if (__should_check_rate(td, ddir)) { |
1673 | td->rate_pending_usleep[ddir] = | |
1674 | (usec_for_io(td, ddir) - | |
ba3e4e0c | 1675 | utime_since_now(&td->start)); |
b23b6a2f | 1676 | } |
f8b0bd10 JA |
1677 | if (ddir != DDIR_TRIM && |
1678 | __should_check_rate(td, oddir)) { | |
1679 | td->rate_pending_usleep[oddir] = | |
1680 | (usec_for_io(td, oddir) - | |
ba3e4e0c | 1681 | utime_since_now(&td->start)); |
f8b0bd10 | 1682 | } |
721938ae | 1683 | } |
10ba535a | 1684 | |
f8b0bd10 | 1685 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1686 | |
d7762cf8 | 1687 | if (io_u->end_io) { |
f8b0bd10 JA |
1688 | ret = io_u->end_io(td, io_u_ptr); |
1689 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1690 | if (ret && !icd->error) |
1691 | icd->error = ret; | |
1692 | } | |
ff58fced | 1693 | } else if (io_u->error) { |
10ba535a | 1694 | icd->error = io_u->error; |
5451792e JA |
1695 | io_u_log_error(td, io_u); |
1696 | } | |
8b28bd41 | 1697 | if (icd->error) { |
f8b0bd10 JA |
1698 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1699 | ||
8b28bd41 DM |
1700 | if (!td_non_fatal_error(td, eb, icd->error)) |
1701 | return; | |
f8b0bd10 | 1702 | |
f2bba182 RR |
1703 | /* |
1704 | * If there is a non_fatal error, then add to the error count | |
1705 | * and clear all the errors. | |
1706 | */ | |
1707 | update_error_count(td, icd->error); | |
1708 | td_clear_error(td); | |
1709 | icd->error = 0; | |
f8b0bd10 JA |
1710 | if (io_u) |
1711 | io_u->error = 0; | |
f2bba182 | 1712 | } |
10ba535a JA |
1713 | } |
1714 | ||
9520ebb9 JA |
1715 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1716 | int nr) | |
10ba535a | 1717 | { |
6eaf09d6 | 1718 | int ddir; |
aba6c951 JA |
1719 | |
1720 | if (!gtod_reduce(td)) | |
9520ebb9 | 1721 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1722 | |
3af6ef39 JA |
1723 | icd->nr = nr; |
1724 | ||
10ba535a | 1725 | icd->error = 0; |
6eaf09d6 SL |
1726 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1727 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1728 | } |
1729 | ||
97601024 JA |
1730 | static void ios_completed(struct thread_data *td, |
1731 | struct io_completion_data *icd) | |
36167d82 JA |
1732 | { |
1733 | struct io_u *io_u; | |
1734 | int i; | |
1735 | ||
10ba535a JA |
1736 | for (i = 0; i < icd->nr; i++) { |
1737 | io_u = td->io_ops->event(td, i); | |
1738 | ||
f8b0bd10 | 1739 | io_completed(td, &io_u, icd); |
e8462bd8 | 1740 | |
f8b0bd10 | 1741 | if (io_u) |
e8462bd8 | 1742 | put_io_u(td, io_u); |
10ba535a JA |
1743 | } |
1744 | } | |
97601024 | 1745 | |
e7e6cfb4 JA |
1746 | /* |
1747 | * Complete a single io_u for the sync engines. | |
1748 | */ | |
581e7141 | 1749 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1750 | uint64_t *bytes) |
97601024 JA |
1751 | { |
1752 | struct io_completion_data icd; | |
1753 | ||
9520ebb9 | 1754 | init_icd(td, &icd, 1); |
f8b0bd10 | 1755 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1756 | |
f8b0bd10 | 1757 | if (io_u) |
e8462bd8 | 1758 | put_io_u(td, io_u); |
97601024 | 1759 | |
581e7141 JA |
1760 | if (icd.error) { |
1761 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1762 | return -1; | |
1763 | } | |
97601024 | 1764 | |
581e7141 | 1765 | if (bytes) { |
6eaf09d6 SL |
1766 | int ddir; |
1767 | ||
1768 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1769 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1770 | } |
1771 | ||
1772 | return 0; | |
97601024 JA |
1773 | } |
1774 | ||
e7e6cfb4 JA |
1775 | /* |
1776 | * Called to complete min_events number of io for the async engines. | |
1777 | */ | |
581e7141 | 1778 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1779 | uint64_t *bytes) |
97601024 | 1780 | { |
97601024 | 1781 | struct io_completion_data icd; |
00de55ef | 1782 | struct timespec *tvp = NULL; |
97601024 | 1783 | int ret; |
4d06a338 | 1784 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1785 | |
4950421a | 1786 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1787 | |
4950421a | 1788 | if (!min_evts) |
00de55ef | 1789 | tvp = &ts; |
5fb4b366 RE |
1790 | else if (min_evts > td->cur_depth) |
1791 | min_evts = td->cur_depth; | |
97601024 | 1792 | |
4950421a | 1793 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1794 | if (ret < 0) { |
e1161c32 | 1795 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1796 | return ret; |
1797 | } else if (!ret) | |
1798 | return ret; | |
1799 | ||
9520ebb9 | 1800 | init_icd(td, &icd, ret); |
97601024 | 1801 | ios_completed(td, &icd); |
581e7141 JA |
1802 | if (icd.error) { |
1803 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1804 | return -1; | |
1805 | } | |
97601024 | 1806 | |
581e7141 | 1807 | if (bytes) { |
6eaf09d6 SL |
1808 | int ddir; |
1809 | ||
1810 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1811 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1812 | } |
1813 | ||
1814 | return 0; | |
97601024 | 1815 | } |
7e77dd02 JA |
1816 | |
1817 | /* | |
1818 | * Call when io_u is really queued, to update the submission latency. | |
1819 | */ | |
1820 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1821 | { | |
9520ebb9 JA |
1822 | if (!td->o.disable_slat) { |
1823 | unsigned long slat_time; | |
7e77dd02 | 1824 | |
9520ebb9 | 1825 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1826 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1827 | io_u->offset); | |
9520ebb9 | 1828 | } |
7e77dd02 | 1829 | } |
433afcb4 | 1830 | |
5c94b008 JA |
1831 | /* |
1832 | * See if we should reuse the last seed, if dedupe is enabled | |
1833 | */ | |
1834 | static struct frand_state *get_buf_state(struct thread_data *td) | |
1835 | { | |
1836 | unsigned int v; | |
1837 | unsigned long r; | |
1838 | ||
1839 | if (!td->o.dedupe_percentage) | |
1840 | return &td->buf_state; | |
c0b69b92 JA |
1841 | else if (td->o.dedupe_percentage == 100) |
1842 | return &td->buf_state_prev; | |
5c94b008 JA |
1843 | |
1844 | r = __rand(&td->dedupe_state); | |
1845 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1846 | ||
1847 | if (v <= td->o.dedupe_percentage) | |
1848 | return &td->buf_state_prev; | |
1849 | ||
1850 | return &td->buf_state; | |
1851 | } | |
1852 | ||
1853 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) | |
1854 | { | |
1855 | if (rs == &td->buf_state) | |
1856 | frand_copy(&td->buf_state_prev, rs); | |
1857 | } | |
1858 | ||
cc86c395 JA |
1859 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1860 | unsigned int max_bs) | |
5973cafb | 1861 | { |
d1af2894 JA |
1862 | struct thread_options *o = &td->o; |
1863 | ||
1864 | if (o->compress_percentage) { | |
9c42684e | 1865 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1866 | struct frand_state *rs; |
1066358a | 1867 | unsigned int left = max_bs; |
5c94b008 | 1868 | |
1066358a JA |
1869 | do { |
1870 | rs = get_buf_state(td); | |
9c42684e | 1871 | |
1066358a | 1872 | min_write = min(min_write, left); |
f97a43a1 | 1873 | |
1066358a JA |
1874 | if (perc) { |
1875 | unsigned int seg = min_write; | |
cc86c395 | 1876 | |
1066358a JA |
1877 | seg = min(min_write, td->o.compress_chunk); |
1878 | if (!seg) | |
1879 | seg = min_write; | |
1880 | ||
1881 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1882 | min_write, o->buffer_pattern, |
1883 | o->buffer_pattern_bytes); | |
1066358a JA |
1884 | } else |
1885 | fill_random_buf(rs, buf, min_write); | |
1886 | ||
1887 | buf += min_write; | |
1888 | left -= min_write; | |
5c94b008 | 1889 | save_buf_state(td, rs); |
1066358a | 1890 | } while (left); |
d1af2894 JA |
1891 | } else if (o->buffer_pattern_bytes) |
1892 | fill_buffer_pattern(td, buf, max_bs); | |
1893 | else | |
cc86c395 JA |
1894 | memset(buf, 0, max_bs); |
1895 | } | |
1896 | ||
1897 | /* | |
1898 | * "randomly" fill the buffer contents | |
1899 | */ | |
1900 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1901 | unsigned int min_write, unsigned int max_bs) | |
1902 | { | |
1903 | io_u->buf_filled_len = 0; | |
1904 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1905 | } |