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 | |
967d1b63 JA |
103 | assert(fio_file_lfsr(f)); |
104 | ||
6f49f8bc | 105 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 106 | return 1; |
ec4015da | 107 | |
8055e41d JA |
108 | *b = off; |
109 | } | |
0ce8b119 | 110 | |
ec4015da | 111 | /* |
51ede0b1 | 112 | * if we are not maintaining a random map, we are done. |
ec4015da | 113 | */ |
51ede0b1 JA |
114 | if (!file_randommap(td, f)) |
115 | goto ret; | |
43c63a78 JA |
116 | |
117 | /* | |
51ede0b1 | 118 | * calculate map offset and check if it's free |
43c63a78 | 119 | */ |
51ede0b1 JA |
120 | if (random_map_free(f, *b)) |
121 | goto ret; | |
122 | ||
4b91ee8f JA |
123 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
124 | (unsigned long long) *b); | |
51ede0b1 | 125 | |
7ebd796f | 126 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
127 | if (*b == (uint64_t) -1ULL) |
128 | return 1; | |
0ce8b119 JA |
129 | ret: |
130 | return 0; | |
ec4015da JA |
131 | } |
132 | ||
925fee33 JA |
133 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
134 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 135 | uint64_t *b) |
e25839d4 | 136 | { |
9c6f6316 | 137 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
138 | return 0; |
139 | } | |
140 | ||
925fee33 JA |
141 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
142 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 143 | uint64_t *b) |
925fee33 | 144 | { |
9c6f6316 | 145 | *b = pareto_next(&f->zipf); |
925fee33 JA |
146 | return 0; |
147 | } | |
148 | ||
1ae83d45 JA |
149 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
150 | { | |
151 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
152 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
153 | ||
154 | return r1->off - r2->off; | |
155 | } | |
156 | ||
157 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
158 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
159 | { |
160 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
161 | return __get_next_rand_offset(td, f, ddir, b); | |
162 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
163 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
164 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
165 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
e25839d4 JA |
166 | |
167 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
168 | return 1; | |
169 | } | |
170 | ||
bcd5abfa JA |
171 | /* |
172 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
173 | * specified. | |
174 | */ | |
175 | static inline int should_sort_io(struct thread_data *td) | |
176 | { | |
177 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
178 | return 0; | |
179 | if (!td_random(td)) | |
180 | return 0; | |
181 | if (td->runstate != TD_VERIFYING) | |
182 | return 0; | |
183 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
184 | return 0; | |
185 | ||
186 | return 1; | |
187 | } | |
188 | ||
d9472271 | 189 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
190 | { |
191 | unsigned int v; | |
192 | unsigned long r; | |
193 | ||
d9472271 | 194 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
195 | return 1; |
196 | ||
d6b72507 | 197 | r = __rand(&td->seq_rand_state[ddir]); |
54a21917 | 198 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
211c9b89 | 199 | |
d9472271 | 200 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
201 | } |
202 | ||
1ae83d45 JA |
203 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
204 | enum fio_ddir ddir, uint64_t *b) | |
205 | { | |
206 | struct rand_off *r; | |
207 | int i, ret = 1; | |
208 | ||
bcd5abfa | 209 | if (!should_sort_io(td)) |
1ae83d45 JA |
210 | return get_off_from_method(td, f, ddir, b); |
211 | ||
212 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 213 | fetch: |
9342d5f8 | 214 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
215 | flist_del(&r->list); |
216 | *b = r->off; | |
217 | free(r); | |
218 | return 0; | |
219 | } | |
220 | ||
221 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
222 | r = malloc(sizeof(*r)); | |
223 | ||
224 | ret = get_off_from_method(td, f, ddir, &r->off); | |
225 | if (ret) { | |
226 | free(r); | |
227 | break; | |
228 | } | |
229 | ||
230 | flist_add(&r->list, &td->next_rand_list); | |
231 | } | |
232 | ||
233 | if (ret && !i) | |
234 | return ret; | |
235 | ||
236 | assert(!flist_empty(&td->next_rand_list)); | |
237 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
238 | goto fetch; | |
239 | } | |
240 | ||
38dad62d | 241 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 242 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 243 | { |
c04e4661 DE |
244 | if (!get_next_rand_offset(td, f, ddir, b)) |
245 | return 0; | |
246 | ||
247 | if (td->o.time_based) { | |
33c48814 | 248 | fio_file_reset(td, f); |
c04e4661 DE |
249 | if (!get_next_rand_offset(td, f, ddir, b)) |
250 | return 0; | |
38dad62d JA |
251 | } |
252 | ||
c04e4661 | 253 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 254 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 255 | (unsigned long long) f->real_file_size); |
c04e4661 | 256 | return 1; |
38dad62d JA |
257 | } |
258 | ||
37cf9e3c | 259 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 260 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 261 | { |
ac002339 JA |
262 | struct thread_options *o = &td->o; |
263 | ||
ff58fced JA |
264 | assert(ddir_rw(ddir)); |
265 | ||
f1dfb668 | 266 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
ac002339 | 267 | o->time_based) |
f1dfb668 | 268 | f->last_pos[ddir] = f->last_pos[ddir] - f->io_size; |
c04e4661 | 269 | |
f1dfb668 | 270 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 271 | uint64_t pos; |
059b0802 | 272 | |
f1dfb668 JA |
273 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) |
274 | f->last_pos[ddir] = f->real_file_size; | |
a66da7a2 | 275 | |
f1dfb668 | 276 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
277 | if (pos && o->ddir_seq_add) { |
278 | pos += o->ddir_seq_add; | |
279 | ||
280 | /* | |
281 | * If we reach beyond the end of the file | |
282 | * with holed IO, wrap around to the | |
283 | * beginning again. | |
284 | */ | |
285 | if (pos >= f->real_file_size) | |
286 | pos = f->file_offset; | |
287 | } | |
059b0802 | 288 | |
37cf9e3c | 289 | *offset = pos; |
38dad62d JA |
290 | return 0; |
291 | } | |
292 | ||
293 | return 1; | |
294 | } | |
295 | ||
296 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
297 | enum fio_ddir ddir, int rw_seq, |
298 | unsigned int *is_random) | |
38dad62d JA |
299 | { |
300 | struct fio_file *f = io_u->file; | |
1ae83d45 | 301 | uint64_t b, offset; |
38dad62d JA |
302 | int ret; |
303 | ||
ff58fced JA |
304 | assert(ddir_rw(ddir)); |
305 | ||
37cf9e3c JA |
306 | b = offset = -1ULL; |
307 | ||
38dad62d | 308 | if (rw_seq) { |
211c9b89 | 309 | if (td_random(td)) { |
6aca9b3d | 310 | if (should_do_random(td, ddir)) { |
211c9b89 | 311 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
312 | *is_random = 1; |
313 | } else { | |
314 | *is_random = 0; | |
211c9b89 JA |
315 | io_u->flags |= IO_U_F_BUSY_OK; |
316 | ret = get_next_seq_offset(td, f, ddir, &offset); | |
317 | if (ret) | |
318 | ret = get_next_rand_block(td, f, ddir, &b); | |
319 | } | |
6aca9b3d JA |
320 | } else { |
321 | *is_random = 0; | |
37cf9e3c | 322 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 323 | } |
38dad62d JA |
324 | } else { |
325 | io_u->flags |= IO_U_F_BUSY_OK; | |
6aca9b3d | 326 | *is_random = 0; |
38dad62d JA |
327 | |
328 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 329 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 330 | if (ret) { |
37cf9e3c | 331 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
332 | *is_random = 0; |
333 | } | |
38dad62d | 334 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
335 | if (f->last_start[ddir] != -1ULL) |
336 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 337 | else |
37cf9e3c | 338 | offset = 0; |
38dad62d JA |
339 | ret = 0; |
340 | } else { | |
341 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
342 | ret = 1; | |
343 | } | |
344 | } | |
6d68b997 | 345 | |
37cf9e3c JA |
346 | if (!ret) { |
347 | if (offset != -1ULL) | |
348 | io_u->offset = offset; | |
349 | else if (b != -1ULL) | |
350 | io_u->offset = b * td->o.ba[ddir]; | |
351 | else { | |
4e0a8fa2 | 352 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
353 | ret = 1; |
354 | } | |
355 | } | |
356 | ||
38dad62d JA |
357 | return ret; |
358 | } | |
359 | ||
10ba535a JA |
360 | /* |
361 | * For random io, generate a random new block and see if it's used. Repeat | |
362 | * until we find a free one. For sequential io, just return the end of | |
363 | * the last io issued. | |
364 | */ | |
6aca9b3d JA |
365 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
366 | unsigned int *is_random) | |
10ba535a | 367 | { |
9bf2061e | 368 | struct fio_file *f = io_u->file; |
4ba66134 | 369 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 370 | int rw_seq_hit = 0; |
10ba535a | 371 | |
ff58fced JA |
372 | assert(ddir_rw(ddir)); |
373 | ||
38dad62d JA |
374 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
375 | rw_seq_hit = 1; | |
5736c10d | 376 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 377 | } |
211097b2 | 378 | |
6aca9b3d | 379 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 380 | return 1; |
10ba535a | 381 | |
009bd847 JA |
382 | if (io_u->offset >= f->io_size) { |
383 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
384 | (unsigned long long) io_u->offset, |
385 | (unsigned long long) f->io_size); | |
009bd847 JA |
386 | return 1; |
387 | } | |
388 | ||
389 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
390 | if (io_u->offset >= f->real_file_size) { |
391 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
392 | (unsigned long long) io_u->offset, |
393 | (unsigned long long) f->real_file_size); | |
10ba535a | 394 | return 1; |
2ba1c290 | 395 | } |
10ba535a JA |
396 | |
397 | return 0; | |
398 | } | |
399 | ||
6aca9b3d JA |
400 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
401 | unsigned int *is_random) | |
15dc1934 | 402 | { |
d72be545 JA |
403 | if (td->flags & TD_F_PROFILE_OPS) { |
404 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 405 | |
d72be545 | 406 | if (ops->fill_io_u_off) |
6aca9b3d | 407 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 408 | } |
15dc1934 | 409 | |
6aca9b3d | 410 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
411 | } |
412 | ||
79944128 JA |
413 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
414 | unsigned int buflen) | |
415 | { | |
416 | struct fio_file *f = io_u->file; | |
417 | ||
bedc9dc2 | 418 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
419 | } |
420 | ||
6aca9b3d JA |
421 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
422 | unsigned int is_random) | |
10ba535a | 423 | { |
6aca9b3d | 424 | int ddir = io_u->ddir; |
24d23ca7 | 425 | unsigned int buflen = 0; |
f3059de1 | 426 | unsigned int minbs, maxbs; |
54a21917 | 427 | unsigned long r; |
10ba535a | 428 | |
9ee1c647 | 429 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
430 | |
431 | if (td->o.bs_is_seq_rand) | |
432 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 433 | |
f3059de1 JA |
434 | minbs = td->o.min_bs[ddir]; |
435 | maxbs = td->o.max_bs[ddir]; | |
436 | ||
79944128 JA |
437 | if (minbs == maxbs) |
438 | return minbs; | |
439 | ||
52c58027 JA |
440 | /* |
441 | * If we can't satisfy the min block size from here, then fail | |
442 | */ | |
443 | if (!io_u_fits(td, io_u, minbs)) | |
444 | return 0; | |
445 | ||
79944128 | 446 | do { |
d6b72507 | 447 | r = __rand(&td->bsrange_state); |
4c07ad86 | 448 | |
720e84ad | 449 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 450 | buflen = 1 + (unsigned int) ((double) maxbs * |
54a21917 | 451 | (r / (FRAND_MAX + 1.0))); |
f3059de1 JA |
452 | if (buflen < minbs) |
453 | buflen = minbs; | |
5ec10eaa | 454 | } else { |
564ca972 JA |
455 | long perc = 0; |
456 | unsigned int i; | |
457 | ||
720e84ad JA |
458 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
459 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
460 | |
461 | buflen = bsp->bs; | |
462 | perc += bsp->perc; | |
54a21917 | 463 | if ((r <= ((FRAND_MAX / 100L) * perc)) && |
79944128 | 464 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
465 | break; |
466 | } | |
467 | } | |
79944128 | 468 | |
a9f70b1f JB |
469 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
470 | buflen = (buflen + td->o.verify_interval - 1) & | |
471 | ~(td->o.verify_interval - 1); | |
472 | ||
f3059de1 JA |
473 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
474 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 475 | |
79944128 | 476 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 477 | |
10ba535a JA |
478 | return buflen; |
479 | } | |
480 | ||
6aca9b3d JA |
481 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
482 | unsigned int is_random) | |
15dc1934 | 483 | { |
d72be545 JA |
484 | if (td->flags & TD_F_PROFILE_OPS) { |
485 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 486 | |
d72be545 | 487 | if (ops->fill_io_u_size) |
6aca9b3d | 488 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 489 | } |
15dc1934 | 490 | |
6aca9b3d | 491 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
492 | } |
493 | ||
afe24a5a JA |
494 | static void set_rwmix_bytes(struct thread_data *td) |
495 | { | |
afe24a5a JA |
496 | unsigned int diff; |
497 | ||
498 | /* | |
499 | * we do time or byte based switch. this is needed because | |
500 | * buffered writes may issue a lot quicker than they complete, | |
501 | * whereas reads do not. | |
502 | */ | |
e47f799f | 503 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 504 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
505 | } |
506 | ||
507 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
508 | { | |
509 | unsigned int v; | |
1294c3ec | 510 | unsigned long r; |
e47f799f | 511 | |
d6b72507 | 512 | r = __rand(&td->rwmix_state); |
54a21917 | 513 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
4c07ad86 | 514 | |
04c540d9 | 515 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
516 | return DDIR_READ; |
517 | ||
518 | return DDIR_WRITE; | |
afe24a5a JA |
519 | } |
520 | ||
002e7183 JA |
521 | void io_u_quiesce(struct thread_data *td) |
522 | { | |
523 | /* | |
524 | * We are going to sleep, ensure that we flush anything pending as | |
525 | * not to skew our latency numbers. | |
526 | * | |
527 | * Changed to only monitor 'in flight' requests here instead of the | |
528 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
529 | * io's that have been actually submitted to an async engine, | |
530 | * and cur_depth is meaningless for sync engines. | |
531 | */ | |
9cc80b6d JA |
532 | if (td->io_u_queued || td->cur_depth) { |
533 | int fio_unused ret; | |
534 | ||
535 | ret = td_io_commit(td); | |
536 | } | |
537 | ||
002e7183 JA |
538 | while (td->io_u_in_flight) { |
539 | int fio_unused ret; | |
540 | ||
541 | ret = io_u_queued_complete(td, 1, NULL); | |
542 | } | |
543 | } | |
544 | ||
581e7141 JA |
545 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
546 | { | |
547 | enum fio_ddir odir = ddir ^ 1; | |
581e7141 JA |
548 | long usec; |
549 | ||
ff58fced JA |
550 | assert(ddir_rw(ddir)); |
551 | ||
315fcfec | 552 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
553 | return ddir; |
554 | ||
555 | /* | |
556 | * We have too much pending sleep in this direction. See if we | |
557 | * should switch. | |
558 | */ | |
315fcfec | 559 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
560 | /* |
561 | * Other direction does not have too much pending, switch | |
562 | */ | |
563 | if (td->rate_pending_usleep[odir] < 100000) | |
564 | return odir; | |
565 | ||
566 | /* | |
567 | * Both directions have pending sleep. Sleep the minimum time | |
568 | * and deduct from both. | |
569 | */ | |
570 | if (td->rate_pending_usleep[ddir] <= | |
571 | td->rate_pending_usleep[odir]) { | |
572 | usec = td->rate_pending_usleep[ddir]; | |
573 | } else { | |
574 | usec = td->rate_pending_usleep[odir]; | |
575 | ddir = odir; | |
576 | } | |
577 | } else | |
578 | usec = td->rate_pending_usleep[ddir]; | |
579 | ||
002e7183 | 580 | io_u_quiesce(td); |
78c1eda5 | 581 | |
9cc80b6d | 582 | usec = usec_sleep(td, usec); |
581e7141 JA |
583 | |
584 | td->rate_pending_usleep[ddir] -= usec; | |
585 | ||
586 | odir = ddir ^ 1; | |
587 | if (td_rw(td) && __should_check_rate(td, odir)) | |
588 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 589 | |
77b451cc JA |
590 | if (ddir == DDIR_TRIM) |
591 | return DDIR_TRIM; | |
e0224c6b | 592 | |
581e7141 JA |
593 | return ddir; |
594 | } | |
595 | ||
10ba535a JA |
596 | /* |
597 | * Return the data direction for the next io_u. If the job is a | |
598 | * mixed read/write workload, check the rwmix cycle and switch if | |
599 | * necessary. | |
600 | */ | |
1e97cce9 | 601 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 602 | { |
581e7141 JA |
603 | enum fio_ddir ddir; |
604 | ||
5f9099ea JA |
605 | /* |
606 | * see if it's time to fsync | |
607 | */ | |
608 | if (td->o.fsync_blocks && | |
609 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
610 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
611 | return DDIR_SYNC; | |
612 | ||
613 | /* | |
614 | * see if it's time to fdatasync | |
615 | */ | |
616 | if (td->o.fdatasync_blocks && | |
617 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
618 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
619 | return DDIR_DATASYNC; | |
620 | ||
44f29692 JA |
621 | /* |
622 | * see if it's time to sync_file_range | |
623 | */ | |
624 | if (td->sync_file_range_nr && | |
625 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
626 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
627 | return DDIR_SYNC_FILE_RANGE; | |
628 | ||
10ba535a | 629 | if (td_rw(td)) { |
10ba535a JA |
630 | /* |
631 | * Check if it's time to seed a new data direction. | |
632 | */ | |
e4928662 | 633 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
634 | /* |
635 | * Put a top limit on how many bytes we do for | |
636 | * one data direction, to avoid overflowing the | |
637 | * ranges too much | |
638 | */ | |
639 | ddir = get_rand_ddir(td); | |
e47f799f JA |
640 | |
641 | if (ddir != td->rwmix_ddir) | |
642 | set_rwmix_bytes(td); | |
643 | ||
644 | td->rwmix_ddir = ddir; | |
10ba535a | 645 | } |
581e7141 | 646 | ddir = td->rwmix_ddir; |
10ba535a | 647 | } else if (td_read(td)) |
581e7141 | 648 | ddir = DDIR_READ; |
6eaf09d6 | 649 | else if (td_write(td)) |
581e7141 | 650 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
651 | else |
652 | ddir = DDIR_TRIM; | |
581e7141 JA |
653 | |
654 | td->rwmix_ddir = rate_ddir(td, ddir); | |
655 | return td->rwmix_ddir; | |
10ba535a JA |
656 | } |
657 | ||
1ef2b6be JA |
658 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
659 | { | |
bcd5abfa | 660 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
661 | |
662 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
663 | td->o.barrier_blocks && | |
664 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
665 | td->io_issues[DDIR_WRITE]) | |
666 | io_u->flags |= IO_U_F_BARRIER; | |
667 | } | |
668 | ||
e8462bd8 | 669 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 670 | { |
71b84caa | 671 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
672 | |
673 | if (ret) | |
674 | td_verror(td, ret, "file close"); | |
675 | } | |
676 | ||
10ba535a JA |
677 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
678 | { | |
e8462bd8 JA |
679 | td_io_u_lock(td); |
680 | ||
f8b0bd10 | 681 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 682 | put_file_log(td, io_u->file); |
f8b0bd10 | 683 | |
10ba535a | 684 | io_u->file = NULL; |
d7ee2a7d SL |
685 | io_u->flags |= IO_U_F_FREE; |
686 | ||
0c41214f RR |
687 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
688 | td->cur_depth--; | |
2ae0b204 | 689 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
690 | td_io_u_unlock(td); |
691 | td_io_u_free_notify(td); | |
10ba535a JA |
692 | } |
693 | ||
f2bba182 RR |
694 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
695 | { | |
696 | io_u->flags &= ~IO_U_F_FLIGHT; | |
697 | put_io_u(td, io_u); | |
698 | } | |
699 | ||
755200a3 JA |
700 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
701 | { | |
702 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 703 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 704 | |
465221b0 JA |
705 | dprint(FD_IO, "requeue %p\n", __io_u); |
706 | ||
e8462bd8 JA |
707 | td_io_u_lock(td); |
708 | ||
4d2e0f49 | 709 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
710 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
711 | td->io_issues[ddir]--; | |
5ec10eaa | 712 | |
4d2e0f49 | 713 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
714 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
715 | td->cur_depth--; | |
2ae0b204 JA |
716 | |
717 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 718 | td_io_u_unlock(td); |
755200a3 JA |
719 | *io_u = NULL; |
720 | } | |
721 | ||
9bf2061e | 722 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 723 | { |
6aca9b3d JA |
724 | unsigned int is_random; |
725 | ||
b4c5e1ac JA |
726 | if (td->io_ops->flags & FIO_NOIO) |
727 | goto out; | |
728 | ||
1ef2b6be | 729 | set_rw_ddir(td, io_u); |
5f9099ea | 730 | |
87dc1ab1 | 731 | /* |
ff58fced | 732 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 733 | */ |
ff58fced | 734 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 735 | goto out; |
a00735e6 | 736 | |
48f5abd3 JA |
737 | /* |
738 | * See if it's time to switch to a new zone | |
739 | */ | |
13af05ae | 740 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
741 | struct fio_file *f = io_u->file; |
742 | ||
48f5abd3 | 743 | td->zone_bytes = 0; |
4c8be5b1 JA |
744 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
745 | ||
746 | /* | |
747 | * Wrap from the beginning, if we exceed the file size | |
748 | */ | |
749 | if (f->file_offset >= f->real_file_size) | |
750 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 751 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
752 | td->io_skip_bytes += td->o.zone_skip; |
753 | } | |
754 | ||
10ba535a | 755 | /* |
c685b5b2 JA |
756 | * No log, let the seq/rand engine retrieve the next buflen and |
757 | * position. | |
10ba535a | 758 | */ |
6aca9b3d | 759 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 760 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 761 | return 1; |
2ba1c290 | 762 | } |
10ba535a | 763 | |
6aca9b3d | 764 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
765 | if (!io_u->buflen) { |
766 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 767 | return 1; |
2ba1c290 | 768 | } |
bca4ed4d | 769 | |
2ba1c290 JA |
770 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
771 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
772 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
773 | (unsigned long long) io_u->offset, io_u->buflen, | |
774 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 775 | return 1; |
2ba1c290 | 776 | } |
6a5e6884 | 777 | |
bca4ed4d JA |
778 | /* |
779 | * mark entry before potentially trimming io_u | |
780 | */ | |
303032ae | 781 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 782 | mark_random_map(td, io_u); |
bca4ed4d | 783 | |
c38e9468 | 784 | out: |
2ba1c290 | 785 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 786 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 787 | return 0; |
10ba535a JA |
788 | } |
789 | ||
838bc709 JA |
790 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
791 | { | |
2b13e716 | 792 | int idx = 0; |
838bc709 JA |
793 | |
794 | switch (nr) { | |
795 | default: | |
2b13e716 | 796 | idx = 6; |
838bc709 JA |
797 | break; |
798 | case 33 ... 64: | |
2b13e716 | 799 | idx = 5; |
838bc709 JA |
800 | break; |
801 | case 17 ... 32: | |
2b13e716 | 802 | idx = 4; |
838bc709 JA |
803 | break; |
804 | case 9 ... 16: | |
2b13e716 | 805 | idx = 3; |
838bc709 JA |
806 | break; |
807 | case 5 ... 8: | |
2b13e716 | 808 | idx = 2; |
838bc709 JA |
809 | break; |
810 | case 1 ... 4: | |
2b13e716 | 811 | idx = 1; |
838bc709 JA |
812 | case 0: |
813 | break; | |
814 | } | |
815 | ||
2b13e716 | 816 | map[idx]++; |
838bc709 JA |
817 | } |
818 | ||
819 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
820 | { | |
821 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
822 | td->ts.total_submit++; | |
823 | } | |
824 | ||
825 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
826 | { | |
827 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
828 | td->ts.total_complete++; | |
829 | } | |
830 | ||
d8005759 | 831 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 832 | { |
2b13e716 | 833 | int idx = 0; |
71619dc2 JA |
834 | |
835 | switch (td->cur_depth) { | |
836 | default: | |
2b13e716 | 837 | idx = 6; |
a783e61a | 838 | break; |
71619dc2 | 839 | case 32 ... 63: |
2b13e716 | 840 | idx = 5; |
a783e61a | 841 | break; |
71619dc2 | 842 | case 16 ... 31: |
2b13e716 | 843 | idx = 4; |
a783e61a | 844 | break; |
71619dc2 | 845 | case 8 ... 15: |
2b13e716 | 846 | idx = 3; |
a783e61a | 847 | break; |
71619dc2 | 848 | case 4 ... 7: |
2b13e716 | 849 | idx = 2; |
a783e61a | 850 | break; |
71619dc2 | 851 | case 2 ... 3: |
2b13e716 | 852 | idx = 1; |
71619dc2 JA |
853 | case 1: |
854 | break; | |
855 | } | |
856 | ||
2b13e716 | 857 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
858 | } |
859 | ||
04a0feae JA |
860 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
861 | { | |
2b13e716 | 862 | int idx = 0; |
04a0feae JA |
863 | |
864 | assert(usec < 1000); | |
865 | ||
866 | switch (usec) { | |
867 | case 750 ... 999: | |
2b13e716 | 868 | idx = 9; |
04a0feae JA |
869 | break; |
870 | case 500 ... 749: | |
2b13e716 | 871 | idx = 8; |
04a0feae JA |
872 | break; |
873 | case 250 ... 499: | |
2b13e716 | 874 | idx = 7; |
04a0feae JA |
875 | break; |
876 | case 100 ... 249: | |
2b13e716 | 877 | idx = 6; |
04a0feae JA |
878 | break; |
879 | case 50 ... 99: | |
2b13e716 | 880 | idx = 5; |
04a0feae JA |
881 | break; |
882 | case 20 ... 49: | |
2b13e716 | 883 | idx = 4; |
04a0feae JA |
884 | break; |
885 | case 10 ... 19: | |
2b13e716 | 886 | idx = 3; |
04a0feae JA |
887 | break; |
888 | case 4 ... 9: | |
2b13e716 | 889 | idx = 2; |
04a0feae JA |
890 | break; |
891 | case 2 ... 3: | |
2b13e716 | 892 | idx = 1; |
04a0feae JA |
893 | case 0 ... 1: |
894 | break; | |
895 | } | |
896 | ||
2b13e716 JA |
897 | assert(idx < FIO_IO_U_LAT_U_NR); |
898 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
899 | } |
900 | ||
901 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 902 | { |
2b13e716 | 903 | int idx = 0; |
ec118304 JA |
904 | |
905 | switch (msec) { | |
906 | default: | |
2b13e716 | 907 | idx = 11; |
04a0feae | 908 | break; |
8abdce66 | 909 | case 1000 ... 1999: |
2b13e716 | 910 | idx = 10; |
04a0feae | 911 | break; |
8abdce66 | 912 | case 750 ... 999: |
2b13e716 | 913 | idx = 9; |
04a0feae | 914 | break; |
8abdce66 | 915 | case 500 ... 749: |
2b13e716 | 916 | idx = 8; |
04a0feae | 917 | break; |
8abdce66 | 918 | case 250 ... 499: |
2b13e716 | 919 | idx = 7; |
04a0feae | 920 | break; |
8abdce66 | 921 | case 100 ... 249: |
2b13e716 | 922 | idx = 6; |
04a0feae | 923 | break; |
8abdce66 | 924 | case 50 ... 99: |
2b13e716 | 925 | idx = 5; |
04a0feae | 926 | break; |
8abdce66 | 927 | case 20 ... 49: |
2b13e716 | 928 | idx = 4; |
04a0feae | 929 | break; |
8abdce66 | 930 | case 10 ... 19: |
2b13e716 | 931 | idx = 3; |
04a0feae | 932 | break; |
8abdce66 | 933 | case 4 ... 9: |
2b13e716 | 934 | idx = 2; |
04a0feae | 935 | break; |
ec118304 | 936 | case 2 ... 3: |
2b13e716 | 937 | idx = 1; |
ec118304 JA |
938 | case 0 ... 1: |
939 | break; | |
940 | } | |
941 | ||
2b13e716 JA |
942 | assert(idx < FIO_IO_U_LAT_M_NR); |
943 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
944 | } |
945 | ||
946 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
947 | { | |
948 | if (usec < 1000) | |
949 | io_u_mark_lat_usec(td, usec); | |
950 | else | |
951 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
952 | } |
953 | ||
0aabe160 JA |
954 | /* |
955 | * Get next file to service by choosing one at random | |
956 | */ | |
2cc52930 JA |
957 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
958 | enum fio_file_flags goodf, | |
d6aed795 | 959 | enum fio_file_flags badf) |
0aabe160 | 960 | { |
0aabe160 | 961 | struct fio_file *f; |
1c178180 | 962 | int fno; |
0aabe160 JA |
963 | |
964 | do { | |
87b10676 | 965 | int opened = 0; |
1294c3ec | 966 | unsigned long r; |
4c07ad86 | 967 | |
d6b72507 | 968 | r = __rand(&td->next_file_state); |
54a21917 | 969 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 970 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 971 | |
126d65c6 | 972 | f = td->files[fno]; |
d6aed795 | 973 | if (fio_file_done(f)) |
059e63c0 | 974 | continue; |
1c178180 | 975 | |
d6aed795 | 976 | if (!fio_file_open(f)) { |
87b10676 JA |
977 | int err; |
978 | ||
002fe734 JA |
979 | if (td->nr_open_files >= td->o.open_files) |
980 | return ERR_PTR(-EBUSY); | |
981 | ||
87b10676 JA |
982 | err = td_io_open_file(td, f); |
983 | if (err) | |
984 | continue; | |
985 | opened = 1; | |
986 | } | |
987 | ||
2ba1c290 JA |
988 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
989 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 990 | return f; |
2ba1c290 | 991 | } |
87b10676 JA |
992 | if (opened) |
993 | td_io_close_file(td, f); | |
0aabe160 JA |
994 | } while (1); |
995 | } | |
996 | ||
997 | /* | |
998 | * Get next file to service by doing round robin between all available ones | |
999 | */ | |
1c178180 JA |
1000 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1001 | int badf) | |
3d7c391d JA |
1002 | { |
1003 | unsigned int old_next_file = td->next_file; | |
1004 | struct fio_file *f; | |
1005 | ||
1006 | do { | |
87b10676 JA |
1007 | int opened = 0; |
1008 | ||
126d65c6 | 1009 | f = td->files[td->next_file]; |
3d7c391d JA |
1010 | |
1011 | td->next_file++; | |
2dc1bbeb | 1012 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1013 | td->next_file = 0; |
1014 | ||
87b10676 | 1015 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1016 | if (fio_file_done(f)) { |
d5ed68ea | 1017 | f = NULL; |
059e63c0 | 1018 | continue; |
d5ed68ea | 1019 | } |
059e63c0 | 1020 | |
d6aed795 | 1021 | if (!fio_file_open(f)) { |
87b10676 JA |
1022 | int err; |
1023 | ||
002fe734 JA |
1024 | if (td->nr_open_files >= td->o.open_files) |
1025 | return ERR_PTR(-EBUSY); | |
1026 | ||
87b10676 | 1027 | err = td_io_open_file(td, f); |
b5696bfc JA |
1028 | if (err) { |
1029 | dprint(FD_FILE, "error %d on open of %s\n", | |
1030 | err, f->file_name); | |
87c27b45 | 1031 | f = NULL; |
87b10676 | 1032 | continue; |
b5696bfc | 1033 | } |
87b10676 JA |
1034 | opened = 1; |
1035 | } | |
1036 | ||
0b9d69ec JA |
1037 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1038 | f->flags); | |
1c178180 | 1039 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1040 | break; |
1041 | ||
87b10676 JA |
1042 | if (opened) |
1043 | td_io_close_file(td, f); | |
1044 | ||
3d7c391d JA |
1045 | f = NULL; |
1046 | } while (td->next_file != old_next_file); | |
1047 | ||
2ba1c290 | 1048 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1049 | return f; |
1050 | } | |
1051 | ||
7eb36574 | 1052 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1053 | { |
1907dbc6 JA |
1054 | struct fio_file *f; |
1055 | ||
2dc1bbeb | 1056 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1057 | |
b5696bfc | 1058 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1059 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1060 | " nr_files=%d\n", td->nr_open_files, | |
1061 | td->nr_done_files, | |
1062 | td->o.nr_files); | |
bdb4e2e9 | 1063 | return NULL; |
2ba1c290 | 1064 | } |
bdb4e2e9 | 1065 | |
1907dbc6 | 1066 | f = td->file_service_file; |
d6aed795 | 1067 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1068 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1069 | goto out; | |
1070 | if (td->file_service_left--) | |
1071 | goto out; | |
1072 | } | |
1907dbc6 | 1073 | |
a086c257 JA |
1074 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1075 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1076 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1077 | else |
d6aed795 | 1078 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1079 | |
002fe734 JA |
1080 | if (IS_ERR(f)) |
1081 | return f; | |
1082 | ||
1907dbc6 JA |
1083 | td->file_service_file = f; |
1084 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1085 | out: |
0dac421f JA |
1086 | if (f) |
1087 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1088 | else | |
1089 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1090 | return f; |
bdb4e2e9 JA |
1091 | } |
1092 | ||
7eb36574 JA |
1093 | static struct fio_file *get_next_file(struct thread_data *td) |
1094 | { | |
372d8962 | 1095 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1096 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1097 | |
d72be545 JA |
1098 | if (ops->get_next_file) |
1099 | return ops->get_next_file(td); | |
1100 | } | |
7eb36574 JA |
1101 | |
1102 | return __get_next_file(td); | |
1103 | } | |
1104 | ||
002fe734 | 1105 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1106 | { |
1107 | struct fio_file *f; | |
1108 | ||
1109 | do { | |
1110 | f = get_next_file(td); | |
002fe734 JA |
1111 | if (IS_ERR_OR_NULL(f)) |
1112 | return PTR_ERR(f); | |
429f6675 | 1113 | |
429f6675 JA |
1114 | io_u->file = f; |
1115 | get_file(f); | |
1116 | ||
1117 | if (!fill_io_u(td, io_u)) | |
1118 | break; | |
1119 | ||
b5696bfc | 1120 | put_file_log(td, f); |
429f6675 | 1121 | td_io_close_file(td, f); |
b5696bfc | 1122 | io_u->file = NULL; |
d6aed795 | 1123 | fio_file_set_done(f); |
429f6675 | 1124 | td->nr_done_files++; |
0b9d69ec JA |
1125 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1126 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1127 | } while (1); |
1128 | ||
1129 | return 0; | |
1130 | } | |
1131 | ||
3e260a46 JA |
1132 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1133 | unsigned long tusec, unsigned long max_usec) | |
1134 | { | |
1135 | if (!td->error) | |
1136 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1137 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1138 | icd->error = ETIMEDOUT; | |
1139 | } | |
1140 | ||
1141 | static void lat_new_cycle(struct thread_data *td) | |
1142 | { | |
1143 | fio_gettime(&td->latency_ts, NULL); | |
1144 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1145 | td->latency_failed = 0; | |
1146 | } | |
1147 | ||
1148 | /* | |
1149 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1150 | * are at QD=1, then it's time to give up. | |
1151 | */ | |
1152 | static int __lat_target_failed(struct thread_data *td) | |
1153 | { | |
1154 | if (td->latency_qd == 1) | |
1155 | return 1; | |
1156 | ||
1157 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1158 | |
1159 | if (td->latency_qd == td->latency_qd_low) | |
1160 | td->latency_qd_low--; | |
1161 | ||
3e260a46 JA |
1162 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1163 | ||
1164 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1165 | ||
1166 | /* | |
1167 | * When we ramp QD down, quiesce existing IO to prevent | |
1168 | * a storm of ramp downs due to pending higher depth. | |
1169 | */ | |
1170 | io_u_quiesce(td); | |
1171 | lat_new_cycle(td); | |
1172 | return 0; | |
1173 | } | |
1174 | ||
1175 | static int lat_target_failed(struct thread_data *td) | |
1176 | { | |
1177 | if (td->o.latency_percentile.u.f == 100.0) | |
1178 | return __lat_target_failed(td); | |
1179 | ||
1180 | td->latency_failed++; | |
1181 | return 0; | |
1182 | } | |
1183 | ||
1184 | void lat_target_init(struct thread_data *td) | |
1185 | { | |
6bb58215 JA |
1186 | td->latency_end_run = 0; |
1187 | ||
3e260a46 JA |
1188 | if (td->o.latency_target) { |
1189 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1190 | fio_gettime(&td->latency_ts, NULL); | |
1191 | td->latency_qd = 1; | |
1192 | td->latency_qd_high = td->o.iodepth; | |
1193 | td->latency_qd_low = 1; | |
1194 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1195 | } else | |
1196 | td->latency_qd = td->o.iodepth; | |
1197 | } | |
1198 | ||
6bb58215 JA |
1199 | void lat_target_reset(struct thread_data *td) |
1200 | { | |
1201 | if (!td->latency_end_run) | |
1202 | lat_target_init(td); | |
1203 | } | |
1204 | ||
3e260a46 JA |
1205 | static void lat_target_success(struct thread_data *td) |
1206 | { | |
1207 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1208 | struct thread_options *o = &td->o; |
3e260a46 JA |
1209 | |
1210 | td->latency_qd_low = td->latency_qd; | |
1211 | ||
1212 | /* | |
1213 | * If we haven't failed yet, we double up to a failing value instead | |
1214 | * of bisecting from highest possible queue depth. If we have set | |
1215 | * a limit other than td->o.iodepth, bisect between that. | |
1216 | */ | |
6bb58215 | 1217 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1218 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1219 | else | |
1220 | td->latency_qd *= 2; | |
1221 | ||
6bb58215 JA |
1222 | if (td->latency_qd > o->iodepth) |
1223 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1224 | |
1225 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1226 | |
3e260a46 | 1227 | /* |
6bb58215 JA |
1228 | * Same as last one, we are done. Let it run a latency cycle, so |
1229 | * we get only the results from the targeted depth. | |
3e260a46 | 1230 | */ |
6bb58215 JA |
1231 | if (td->latency_qd == qd) { |
1232 | if (td->latency_end_run) { | |
1233 | dprint(FD_RATE, "We are done\n"); | |
1234 | td->done = 1; | |
1235 | } else { | |
1236 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1237 | io_u_quiesce(td); | |
1238 | td->latency_end_run = 1; | |
1239 | reset_all_stats(td); | |
1240 | reset_io_stats(td); | |
1241 | } | |
1242 | } | |
3e260a46 JA |
1243 | |
1244 | lat_new_cycle(td); | |
1245 | } | |
1246 | ||
1247 | /* | |
1248 | * Check if we can bump the queue depth | |
1249 | */ | |
1250 | void lat_target_check(struct thread_data *td) | |
1251 | { | |
1252 | uint64_t usec_window; | |
1253 | uint64_t ios; | |
1254 | double success_ios; | |
1255 | ||
1256 | usec_window = utime_since_now(&td->latency_ts); | |
1257 | if (usec_window < td->o.latency_window) | |
1258 | return; | |
1259 | ||
1260 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1261 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1262 | success_ios *= 100.0; | |
1263 | ||
1264 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1265 | ||
1266 | if (success_ios >= td->o.latency_percentile.u.f) | |
1267 | lat_target_success(td); | |
1268 | else | |
1269 | __lat_target_failed(td); | |
1270 | } | |
1271 | ||
1272 | /* | |
1273 | * If latency target is enabled, we might be ramping up or down and not | |
1274 | * using the full queue depth available. | |
1275 | */ | |
effd6ff0 | 1276 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1277 | { |
1278 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1279 | ||
1280 | if (qempty) | |
1281 | return 1; | |
1282 | if (!td->o.latency_target) | |
1283 | return 0; | |
1284 | ||
1285 | return td->cur_depth >= td->latency_qd; | |
1286 | } | |
429f6675 | 1287 | |
10ba535a JA |
1288 | struct io_u *__get_io_u(struct thread_data *td) |
1289 | { | |
0cae66f6 | 1290 | struct io_u *io_u = NULL; |
10ba535a | 1291 | |
ca09be4b JA |
1292 | if (td->stop_io) |
1293 | return NULL; | |
1294 | ||
e8462bd8 JA |
1295 | td_io_u_lock(td); |
1296 | ||
1297 | again: | |
2ae0b204 JA |
1298 | if (!io_u_rempty(&td->io_u_requeues)) |
1299 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1300 | else if (!queue_full(td)) { |
2ae0b204 | 1301 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1302 | |
225ba9e3 | 1303 | io_u->file = NULL; |
6040dabc | 1304 | io_u->buflen = 0; |
10ba535a | 1305 | io_u->resid = 0; |
d7762cf8 | 1306 | io_u->end_io = NULL; |
755200a3 JA |
1307 | } |
1308 | ||
1309 | if (io_u) { | |
0c6e7517 | 1310 | assert(io_u->flags & IO_U_F_FREE); |
f8b0bd10 JA |
1311 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
1312 | IO_U_F_TRIMMED | IO_U_F_BARRIER | | |
1313 | IO_U_F_VER_LIST); | |
0c6e7517 | 1314 | |
755200a3 | 1315 | io_u->error = 0; |
bcd5abfa | 1316 | io_u->acct_ddir = -1; |
10ba535a | 1317 | td->cur_depth++; |
0c41214f | 1318 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1319 | io_u->ipo = NULL; |
1dec3e07 JA |
1320 | } else if (td->o.verify_async) { |
1321 | /* | |
1322 | * We ran out, wait for async verify threads to finish and | |
1323 | * return one | |
1324 | */ | |
1325 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1326 | goto again; | |
10ba535a JA |
1327 | } |
1328 | ||
e8462bd8 | 1329 | td_io_u_unlock(td); |
10ba535a JA |
1330 | return io_u; |
1331 | } | |
1332 | ||
0d29de83 | 1333 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1334 | { |
d72be545 JA |
1335 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1336 | return 0; | |
1337 | ||
1338 | if (td->trim_entries) { | |
0d29de83 | 1339 | int get_trim = 0; |
10ba535a | 1340 | |
0d29de83 JA |
1341 | if (td->trim_batch) { |
1342 | td->trim_batch--; | |
1343 | get_trim = 1; | |
1344 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1345 | td->last_ddir != DDIR_READ) { | |
1346 | td->trim_batch = td->o.trim_batch; | |
1347 | if (!td->trim_batch) | |
1348 | td->trim_batch = td->o.trim_backlog; | |
1349 | get_trim = 1; | |
1350 | } | |
1351 | ||
1352 | if (get_trim && !get_next_trim(td, io_u)) | |
1353 | return 1; | |
2ba1c290 | 1354 | } |
10ba535a | 1355 | |
0d29de83 JA |
1356 | return 0; |
1357 | } | |
1358 | ||
1359 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1360 | { | |
d72be545 JA |
1361 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1362 | return 0; | |
1363 | ||
1364 | if (td->io_hist_len) { | |
9e144189 JA |
1365 | int get_verify = 0; |
1366 | ||
d1ece0c7 | 1367 | if (td->verify_batch) |
9e144189 | 1368 | get_verify = 1; |
d1ece0c7 | 1369 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1370 | td->last_ddir != DDIR_READ) { |
1371 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1372 | if (!td->verify_batch) |
1373 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1374 | get_verify = 1; |
1375 | } | |
1376 | ||
d1ece0c7 JA |
1377 | if (get_verify && !get_next_verify(td, io_u)) { |
1378 | td->verify_batch--; | |
0d29de83 | 1379 | return 1; |
d1ece0c7 | 1380 | } |
9e144189 JA |
1381 | } |
1382 | ||
0d29de83 JA |
1383 | return 0; |
1384 | } | |
1385 | ||
de789769 JA |
1386 | /* |
1387 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1388 | * easy compressible data for simple de-dupe attempts. Do this for every |
1389 | * 512b block in the range, since that should be the smallest block size | |
1390 | * we can expect from a device. | |
de789769 JA |
1391 | */ |
1392 | static void small_content_scramble(struct io_u *io_u) | |
1393 | { | |
23f394d5 | 1394 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1395 | uint64_t boffset; |
23f394d5 JA |
1396 | unsigned int offset; |
1397 | void *p, *end; | |
de789769 | 1398 | |
23f394d5 JA |
1399 | if (!nr_blocks) |
1400 | return; | |
1401 | ||
1402 | p = io_u->xfer_buf; | |
fba76ee8 | 1403 | boffset = io_u->offset; |
81f0366c | 1404 | io_u->buf_filled_len = 0; |
fad82f76 | 1405 | |
23f394d5 JA |
1406 | for (i = 0; i < nr_blocks; i++) { |
1407 | /* | |
1408 | * Fill the byte offset into a "random" start offset of | |
1409 | * the buffer, given by the product of the usec time | |
1410 | * and the actual offset. | |
1411 | */ | |
fad82f76 | 1412 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1413 | offset &= ~(sizeof(uint64_t) - 1); |
1414 | if (offset >= 512 - sizeof(uint64_t)) | |
1415 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1416 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1417 | |
1418 | end = p + 512 - sizeof(io_u->start_time); | |
1419 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1420 | p += 512; | |
fad82f76 | 1421 | boffset += 512; |
23f394d5 | 1422 | } |
de789769 JA |
1423 | } |
1424 | ||
0d29de83 JA |
1425 | /* |
1426 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1427 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1428 | */ | |
1429 | struct io_u *get_io_u(struct thread_data *td) | |
1430 | { | |
1431 | struct fio_file *f; | |
1432 | struct io_u *io_u; | |
de789769 | 1433 | int do_scramble = 0; |
002fe734 | 1434 | long ret = 0; |
0d29de83 JA |
1435 | |
1436 | io_u = __get_io_u(td); | |
1437 | if (!io_u) { | |
1438 | dprint(FD_IO, "__get_io_u failed\n"); | |
1439 | return NULL; | |
1440 | } | |
1441 | ||
1442 | if (check_get_verify(td, io_u)) | |
1443 | goto out; | |
1444 | if (check_get_trim(td, io_u)) | |
1445 | goto out; | |
1446 | ||
755200a3 JA |
1447 | /* |
1448 | * from a requeue, io_u already setup | |
1449 | */ | |
1450 | if (io_u->file) | |
77f392bf | 1451 | goto out; |
755200a3 | 1452 | |
429f6675 JA |
1453 | /* |
1454 | * If using an iolog, grab next piece if any available. | |
1455 | */ | |
d72be545 | 1456 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1457 | if (read_iolog_get(td, io_u)) |
1458 | goto err_put; | |
2ba1c290 | 1459 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1460 | ret = -EBUSY; |
2ba1c290 | 1461 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1462 | goto err_put; |
2ba1c290 | 1463 | } |
5ec10eaa | 1464 | |
429f6675 | 1465 | f = io_u->file; |
002fe734 JA |
1466 | if (!f) { |
1467 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1468 | goto err_put; | |
1469 | } | |
1470 | ||
d6aed795 | 1471 | assert(fio_file_open(f)); |
97af62ce | 1472 | |
ff58fced | 1473 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1474 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1475 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1476 | goto err_put; |
2ba1c290 | 1477 | } |
10ba535a | 1478 | |
f1dfb668 JA |
1479 | f->last_start[io_u->ddir] = io_u->offset; |
1480 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1481 | |
fd68418e | 1482 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1483 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1484 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1485 | td->o.min_bs[DDIR_WRITE], |
1486 | io_u->xfer_buflen); | |
bedc9dc2 JA |
1487 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
1488 | !(td->flags & TD_F_COMPRESS)) | |
fd68418e | 1489 | do_scramble = 1; |
d72be545 | 1490 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1491 | populate_verify_io_u(td, io_u); |
1492 | do_scramble = 0; | |
1493 | } | |
fd68418e | 1494 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1495 | /* |
1496 | * Reset the buf_filled parameters so next time if the | |
1497 | * buffer is used for writes it is refilled. | |
1498 | */ | |
cbe8d756 RR |
1499 | io_u->buf_filled_len = 0; |
1500 | } | |
87dc1ab1 | 1501 | } |
10ba535a | 1502 | |
165faf16 JA |
1503 | /* |
1504 | * Set io data pointers. | |
1505 | */ | |
cec6b55d JA |
1506 | io_u->xfer_buf = io_u->buf; |
1507 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1508 | |
6ac7a331 | 1509 | out: |
0d29de83 | 1510 | assert(io_u->file); |
429f6675 | 1511 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1512 | if (!td->o.disable_slat) |
1513 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1514 | if (do_scramble) |
1515 | small_content_scramble(io_u); | |
429f6675 | 1516 | return io_u; |
36167d82 | 1517 | } |
429f6675 | 1518 | err_put: |
2ba1c290 | 1519 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1520 | put_io_u(td, io_u); |
002fe734 | 1521 | return ERR_PTR(ret); |
10ba535a JA |
1522 | } |
1523 | ||
5451792e JA |
1524 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1525 | { | |
8b28bd41 | 1526 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1527 | |
8b28bd41 DM |
1528 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1529 | return; | |
5451792e | 1530 | |
709c8313 RE |
1531 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1532 | io_u->file ? " on file " : "", | |
1533 | io_u->file ? io_u->file->file_name : "", | |
1534 | strerror(io_u->error), | |
1535 | io_ddir_name(io_u->ddir), | |
1536 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1537 | |
1538 | if (!td->error) | |
1539 | td_verror(td, io_u->error, "io_u error"); | |
1540 | } | |
1541 | ||
aba6c951 JA |
1542 | static inline int gtod_reduce(struct thread_data *td) |
1543 | { | |
729fe3af | 1544 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1545 | && td->o.disable_bw; |
aba6c951 JA |
1546 | } |
1547 | ||
c8eeb9df JA |
1548 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1549 | struct io_completion_data *icd, | |
1550 | const enum fio_ddir idx, unsigned int bytes) | |
1551 | { | |
24d23ca7 | 1552 | unsigned long lusec = 0; |
c8eeb9df | 1553 | |
aba6c951 | 1554 | if (!gtod_reduce(td)) |
c8eeb9df JA |
1555 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1556 | ||
1557 | if (!td->o.disable_lat) { | |
1558 | unsigned long tusec; | |
1559 | ||
1560 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1561 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1562 | |
d4afedfd JA |
1563 | if (td->flags & TD_F_PROFILE_OPS) { |
1564 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1565 | ||
1566 | if (ops->io_u_lat) | |
1567 | icd->error = ops->io_u_lat(td, tusec); | |
1568 | } | |
1569 | ||
3e260a46 JA |
1570 | if (td->o.max_latency && tusec > td->o.max_latency) |
1571 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1572 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1573 | if (lat_target_failed(td)) | |
1574 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1575 | } |
c8eeb9df JA |
1576 | } |
1577 | ||
1578 | if (!td->o.disable_clat) { | |
ae588852 | 1579 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1580 | io_u_mark_latency(td, lusec); |
1581 | } | |
1582 | ||
1583 | if (!td->o.disable_bw) | |
1584 | add_bw_sample(td, idx, bytes, &icd->time); | |
1585 | ||
aba6c951 JA |
1586 | if (!gtod_reduce(td)) |
1587 | add_iops_sample(td, idx, bytes, &icd->time); | |
c8eeb9df JA |
1588 | } |
1589 | ||
1b8dbf25 SL |
1590 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1591 | { | |
1ae83d45 JA |
1592 | uint64_t secs, remainder, bps, bytes; |
1593 | ||
1b8dbf25 SL |
1594 | bytes = td->this_io_bytes[ddir]; |
1595 | bps = td->rate_bps[ddir]; | |
1596 | secs = bytes / bps; | |
1597 | remainder = bytes % bps; | |
1598 | return remainder * 1000000 / bps + secs * 1000000; | |
1599 | } | |
1600 | ||
f8b0bd10 | 1601 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1602 | struct io_completion_data *icd) |
10ba535a | 1603 | { |
f8b0bd10 JA |
1604 | struct io_u *io_u = *io_u_ptr; |
1605 | enum fio_ddir ddir = io_u->ddir; | |
1606 | struct fio_file *f = io_u->file; | |
10ba535a | 1607 | |
2ba1c290 JA |
1608 | dprint_io_u(io_u, "io complete"); |
1609 | ||
2ecc1b57 | 1610 | td_io_u_lock(td); |
0c6e7517 | 1611 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1612 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1613 | |
1614 | /* | |
1615 | * Mark IO ok to verify | |
1616 | */ | |
1617 | if (io_u->ipo) { | |
890b6656 JA |
1618 | /* |
1619 | * Remove errored entry from the verification list | |
1620 | */ | |
1621 | if (io_u->error) | |
1622 | unlog_io_piece(td, io_u); | |
1623 | else { | |
1624 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1625 | write_barrier(); | |
1626 | } | |
f9401285 JA |
1627 | } |
1628 | ||
2ecc1b57 | 1629 | td_io_u_unlock(td); |
0c6e7517 | 1630 | |
f8b0bd10 | 1631 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1632 | td->last_was_sync = 1; |
44f29692 JA |
1633 | if (f) { |
1634 | f->first_write = -1ULL; | |
1635 | f->last_write = -1ULL; | |
1636 | } | |
87dc1ab1 JA |
1637 | return; |
1638 | } | |
1639 | ||
1640 | td->last_was_sync = 0; | |
f8b0bd10 | 1641 | td->last_ddir = ddir; |
87dc1ab1 | 1642 | |
f8b0bd10 | 1643 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1644 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1645 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1646 | int ret; |
10ba535a | 1647 | |
f8b0bd10 JA |
1648 | td->io_blocks[ddir]++; |
1649 | td->this_io_blocks[ddir]++; | |
1650 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1651 | |
1652 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1653 | td->this_io_bytes[ddir] += bytes; |
1654 | ||
ca09be4b JA |
1655 | if (ddir == DDIR_WRITE) { |
1656 | if (f) { | |
1657 | if (f->first_write == -1ULL || | |
1658 | io_u->offset < f->first_write) | |
1659 | f->first_write = io_u->offset; | |
1660 | if (f->last_write == -1ULL || | |
1661 | ((io_u->offset + bytes) > f->last_write)) | |
1662 | f->last_write = io_u->offset + bytes; | |
1663 | } | |
1664 | if (td->last_write_comp) { | |
1665 | int idx = td->last_write_idx++; | |
1666 | ||
1667 | td->last_write_comp[idx] = io_u->offset; | |
1668 | if (td->last_write_idx == td->o.iodepth) | |
1669 | td->last_write_idx = 0; | |
1670 | } | |
44f29692 JA |
1671 | } |
1672 | ||
6b1190fd SL |
1673 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1674 | td->runstate == TD_VERIFYING)) { | |
f8b0bd10 | 1675 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1676 | |
f8b0bd10 JA |
1677 | if (__should_check_rate(td, ddir)) { |
1678 | td->rate_pending_usleep[ddir] = | |
1679 | (usec_for_io(td, ddir) - | |
ba3e4e0c | 1680 | utime_since_now(&td->start)); |
b23b6a2f | 1681 | } |
f8b0bd10 JA |
1682 | if (ddir != DDIR_TRIM && |
1683 | __should_check_rate(td, oddir)) { | |
1684 | td->rate_pending_usleep[oddir] = | |
1685 | (usec_for_io(td, oddir) - | |
ba3e4e0c | 1686 | utime_since_now(&td->start)); |
f8b0bd10 | 1687 | } |
721938ae | 1688 | } |
10ba535a | 1689 | |
f8b0bd10 | 1690 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1691 | |
d7762cf8 | 1692 | if (io_u->end_io) { |
f8b0bd10 JA |
1693 | ret = io_u->end_io(td, io_u_ptr); |
1694 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1695 | if (ret && !icd->error) |
1696 | icd->error = ret; | |
1697 | } | |
ff58fced | 1698 | } else if (io_u->error) { |
10ba535a | 1699 | icd->error = io_u->error; |
5451792e JA |
1700 | io_u_log_error(td, io_u); |
1701 | } | |
8b28bd41 | 1702 | if (icd->error) { |
f8b0bd10 JA |
1703 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1704 | ||
8b28bd41 DM |
1705 | if (!td_non_fatal_error(td, eb, icd->error)) |
1706 | return; | |
f8b0bd10 | 1707 | |
f2bba182 RR |
1708 | /* |
1709 | * If there is a non_fatal error, then add to the error count | |
1710 | * and clear all the errors. | |
1711 | */ | |
1712 | update_error_count(td, icd->error); | |
1713 | td_clear_error(td); | |
1714 | icd->error = 0; | |
f8b0bd10 JA |
1715 | if (io_u) |
1716 | io_u->error = 0; | |
f2bba182 | 1717 | } |
10ba535a JA |
1718 | } |
1719 | ||
9520ebb9 JA |
1720 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1721 | int nr) | |
10ba535a | 1722 | { |
6eaf09d6 | 1723 | int ddir; |
aba6c951 JA |
1724 | |
1725 | if (!gtod_reduce(td)) | |
9520ebb9 | 1726 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1727 | |
3af6ef39 JA |
1728 | icd->nr = nr; |
1729 | ||
10ba535a | 1730 | icd->error = 0; |
6eaf09d6 SL |
1731 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1732 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1733 | } |
1734 | ||
97601024 JA |
1735 | static void ios_completed(struct thread_data *td, |
1736 | struct io_completion_data *icd) | |
36167d82 JA |
1737 | { |
1738 | struct io_u *io_u; | |
1739 | int i; | |
1740 | ||
10ba535a JA |
1741 | for (i = 0; i < icd->nr; i++) { |
1742 | io_u = td->io_ops->event(td, i); | |
1743 | ||
f8b0bd10 | 1744 | io_completed(td, &io_u, icd); |
e8462bd8 | 1745 | |
f8b0bd10 | 1746 | if (io_u) |
e8462bd8 | 1747 | put_io_u(td, io_u); |
10ba535a JA |
1748 | } |
1749 | } | |
97601024 | 1750 | |
e7e6cfb4 JA |
1751 | /* |
1752 | * Complete a single io_u for the sync engines. | |
1753 | */ | |
581e7141 | 1754 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1755 | uint64_t *bytes) |
97601024 JA |
1756 | { |
1757 | struct io_completion_data icd; | |
1758 | ||
9520ebb9 | 1759 | init_icd(td, &icd, 1); |
f8b0bd10 | 1760 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1761 | |
f8b0bd10 | 1762 | if (io_u) |
e8462bd8 | 1763 | put_io_u(td, io_u); |
97601024 | 1764 | |
581e7141 JA |
1765 | if (icd.error) { |
1766 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1767 | return -1; | |
1768 | } | |
97601024 | 1769 | |
581e7141 | 1770 | if (bytes) { |
6eaf09d6 SL |
1771 | int ddir; |
1772 | ||
1773 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1774 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1775 | } |
1776 | ||
1777 | return 0; | |
97601024 JA |
1778 | } |
1779 | ||
e7e6cfb4 JA |
1780 | /* |
1781 | * Called to complete min_events number of io for the async engines. | |
1782 | */ | |
581e7141 | 1783 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1784 | uint64_t *bytes) |
97601024 | 1785 | { |
97601024 | 1786 | struct io_completion_data icd; |
00de55ef | 1787 | struct timespec *tvp = NULL; |
97601024 | 1788 | int ret; |
4d06a338 | 1789 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1790 | |
4950421a | 1791 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1792 | |
4950421a | 1793 | if (!min_evts) |
00de55ef | 1794 | tvp = &ts; |
5fb4b366 RE |
1795 | else if (min_evts > td->cur_depth) |
1796 | min_evts = td->cur_depth; | |
97601024 | 1797 | |
4950421a | 1798 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1799 | if (ret < 0) { |
e1161c32 | 1800 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1801 | return ret; |
1802 | } else if (!ret) | |
1803 | return ret; | |
1804 | ||
9520ebb9 | 1805 | init_icd(td, &icd, ret); |
97601024 | 1806 | ios_completed(td, &icd); |
581e7141 JA |
1807 | if (icd.error) { |
1808 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1809 | return -1; | |
1810 | } | |
97601024 | 1811 | |
581e7141 | 1812 | if (bytes) { |
6eaf09d6 SL |
1813 | int ddir; |
1814 | ||
1815 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1816 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1817 | } |
1818 | ||
1819 | return 0; | |
97601024 | 1820 | } |
7e77dd02 JA |
1821 | |
1822 | /* | |
1823 | * Call when io_u is really queued, to update the submission latency. | |
1824 | */ | |
1825 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1826 | { | |
9520ebb9 JA |
1827 | if (!td->o.disable_slat) { |
1828 | unsigned long slat_time; | |
7e77dd02 | 1829 | |
9520ebb9 | 1830 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1831 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1832 | io_u->offset); | |
9520ebb9 | 1833 | } |
7e77dd02 | 1834 | } |
433afcb4 | 1835 | |
5c94b008 JA |
1836 | /* |
1837 | * See if we should reuse the last seed, if dedupe is enabled | |
1838 | */ | |
1839 | static struct frand_state *get_buf_state(struct thread_data *td) | |
1840 | { | |
1841 | unsigned int v; | |
1842 | unsigned long r; | |
1843 | ||
1844 | if (!td->o.dedupe_percentage) | |
1845 | return &td->buf_state; | |
c0b69b92 JA |
1846 | else if (td->o.dedupe_percentage == 100) |
1847 | return &td->buf_state_prev; | |
5c94b008 JA |
1848 | |
1849 | r = __rand(&td->dedupe_state); | |
1850 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1851 | ||
1852 | if (v <= td->o.dedupe_percentage) | |
1853 | return &td->buf_state_prev; | |
1854 | ||
1855 | return &td->buf_state; | |
1856 | } | |
1857 | ||
1858 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) | |
1859 | { | |
1860 | if (rs == &td->buf_state) | |
1861 | frand_copy(&td->buf_state_prev, rs); | |
1862 | } | |
1863 | ||
cc86c395 JA |
1864 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1865 | unsigned int max_bs) | |
5973cafb | 1866 | { |
d1af2894 JA |
1867 | struct thread_options *o = &td->o; |
1868 | ||
1869 | if (o->compress_percentage) { | |
9c42684e | 1870 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1871 | struct frand_state *rs; |
1066358a | 1872 | unsigned int left = max_bs; |
5c94b008 | 1873 | |
1066358a JA |
1874 | do { |
1875 | rs = get_buf_state(td); | |
9c42684e | 1876 | |
1066358a | 1877 | min_write = min(min_write, left); |
f97a43a1 | 1878 | |
1066358a JA |
1879 | if (perc) { |
1880 | unsigned int seg = min_write; | |
cc86c395 | 1881 | |
1066358a JA |
1882 | seg = min(min_write, td->o.compress_chunk); |
1883 | if (!seg) | |
1884 | seg = min_write; | |
1885 | ||
1886 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1887 | min_write, o->buffer_pattern, |
1888 | o->buffer_pattern_bytes); | |
1066358a JA |
1889 | } else |
1890 | fill_random_buf(rs, buf, min_write); | |
1891 | ||
1892 | buf += min_write; | |
1893 | left -= min_write; | |
5c94b008 | 1894 | save_buf_state(td, rs); |
1066358a | 1895 | } while (left); |
d1af2894 JA |
1896 | } else if (o->buffer_pattern_bytes) |
1897 | fill_buffer_pattern(td, buf, max_bs); | |
1898 | else | |
cc86c395 JA |
1899 | memset(buf, 0, max_bs); |
1900 | } | |
1901 | ||
1902 | /* | |
1903 | * "randomly" fill the buffer contents | |
1904 | */ | |
1905 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1906 | unsigned int min_write, unsigned int max_bs) | |
1907 | { | |
1908 | io_u->buf_filled_len = 0; | |
1909 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1910 | } |