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 | { |
1ae83d45 | 86 | uint64_t r, lastb; |
ec4015da | 87 | |
74776733 JA |
88 | lastb = last_block(td, f, ddir); |
89 | if (!lastb) | |
90 | return 1; | |
5e0baa7f | 91 | |
74776733 | 92 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) { |
d6b72507 | 93 | r = __rand(&td->random_state); |
8055e41d | 94 | |
4b91ee8f | 95 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 96 | |
54a21917 | 97 | *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0)); |
51ede0b1 | 98 | } else { |
8055e41d | 99 | uint64_t off = 0; |
43c63a78 | 100 | |
74776733 | 101 | if (lfsr_next(&f->lfsr, &off, lastb)) |
8055e41d | 102 | return 1; |
ec4015da | 103 | |
8055e41d JA |
104 | *b = off; |
105 | } | |
0ce8b119 | 106 | |
ec4015da | 107 | /* |
51ede0b1 | 108 | * if we are not maintaining a random map, we are done. |
ec4015da | 109 | */ |
51ede0b1 JA |
110 | if (!file_randommap(td, f)) |
111 | goto ret; | |
43c63a78 JA |
112 | |
113 | /* | |
51ede0b1 | 114 | * calculate map offset and check if it's free |
43c63a78 | 115 | */ |
51ede0b1 JA |
116 | if (random_map_free(f, *b)) |
117 | goto ret; | |
118 | ||
4b91ee8f JA |
119 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
120 | (unsigned long long) *b); | |
51ede0b1 | 121 | |
7ebd796f | 122 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
123 | if (*b == (uint64_t) -1ULL) |
124 | return 1; | |
0ce8b119 JA |
125 | ret: |
126 | return 0; | |
ec4015da JA |
127 | } |
128 | ||
925fee33 JA |
129 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
130 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 131 | uint64_t *b) |
e25839d4 | 132 | { |
9c6f6316 | 133 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
134 | return 0; |
135 | } | |
136 | ||
925fee33 JA |
137 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
138 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 139 | uint64_t *b) |
925fee33 | 140 | { |
9c6f6316 | 141 | *b = pareto_next(&f->zipf); |
925fee33 JA |
142 | return 0; |
143 | } | |
144 | ||
1ae83d45 JA |
145 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
146 | { | |
147 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
148 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
149 | ||
150 | return r1->off - r2->off; | |
151 | } | |
152 | ||
153 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
154 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
155 | { |
156 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
157 | return __get_next_rand_offset(td, f, ddir, b); | |
158 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
159 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
160 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
161 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
e25839d4 JA |
162 | |
163 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
164 | return 1; | |
165 | } | |
166 | ||
bcd5abfa JA |
167 | /* |
168 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
169 | * specified. | |
170 | */ | |
171 | static inline int should_sort_io(struct thread_data *td) | |
172 | { | |
173 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
174 | return 0; | |
175 | if (!td_random(td)) | |
176 | return 0; | |
177 | if (td->runstate != TD_VERIFYING) | |
178 | return 0; | |
179 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
180 | return 0; | |
181 | ||
182 | return 1; | |
183 | } | |
184 | ||
d9472271 | 185 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
186 | { |
187 | unsigned int v; | |
188 | unsigned long r; | |
189 | ||
d9472271 | 190 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
191 | return 1; |
192 | ||
d6b72507 | 193 | r = __rand(&td->seq_rand_state[ddir]); |
54a21917 | 194 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
211c9b89 | 195 | |
d9472271 | 196 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
197 | } |
198 | ||
1ae83d45 JA |
199 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
200 | enum fio_ddir ddir, uint64_t *b) | |
201 | { | |
202 | struct rand_off *r; | |
203 | int i, ret = 1; | |
204 | ||
bcd5abfa | 205 | if (!should_sort_io(td)) |
1ae83d45 JA |
206 | return get_off_from_method(td, f, ddir, b); |
207 | ||
208 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 209 | fetch: |
9342d5f8 | 210 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
211 | flist_del(&r->list); |
212 | *b = r->off; | |
213 | free(r); | |
214 | return 0; | |
215 | } | |
216 | ||
217 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
218 | r = malloc(sizeof(*r)); | |
219 | ||
220 | ret = get_off_from_method(td, f, ddir, &r->off); | |
221 | if (ret) { | |
222 | free(r); | |
223 | break; | |
224 | } | |
225 | ||
226 | flist_add(&r->list, &td->next_rand_list); | |
227 | } | |
228 | ||
229 | if (ret && !i) | |
230 | return ret; | |
231 | ||
232 | assert(!flist_empty(&td->next_rand_list)); | |
233 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
234 | goto fetch; | |
235 | } | |
236 | ||
38dad62d | 237 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 238 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 239 | { |
c04e4661 DE |
240 | if (!get_next_rand_offset(td, f, ddir, b)) |
241 | return 0; | |
242 | ||
243 | if (td->o.time_based) { | |
33c48814 | 244 | fio_file_reset(td, f); |
c04e4661 DE |
245 | if (!get_next_rand_offset(td, f, ddir, b)) |
246 | return 0; | |
38dad62d JA |
247 | } |
248 | ||
c04e4661 | 249 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
4b91ee8f JA |
250 | f->file_name, (unsigned long long) f->last_pos, |
251 | (unsigned long long) f->real_file_size); | |
c04e4661 | 252 | return 1; |
38dad62d JA |
253 | } |
254 | ||
37cf9e3c | 255 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 256 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 257 | { |
ac002339 JA |
258 | struct thread_options *o = &td->o; |
259 | ||
ff58fced JA |
260 | assert(ddir_rw(ddir)); |
261 | ||
ac002339 JA |
262 | if (f->last_pos >= f->io_size + get_start_offset(td, f) && |
263 | o->time_based) | |
c04e4661 DE |
264 | f->last_pos = f->last_pos - f->io_size; |
265 | ||
38dad62d | 266 | if (f->last_pos < f->real_file_size) { |
1ae83d45 | 267 | uint64_t pos; |
059b0802 | 268 | |
ac002339 | 269 | if (f->last_pos == f->file_offset && o->ddir_seq_add < 0) |
a66da7a2 JA |
270 | f->last_pos = f->real_file_size; |
271 | ||
272 | pos = f->last_pos - f->file_offset; | |
ac002339 JA |
273 | if (pos && o->ddir_seq_add) { |
274 | pos += o->ddir_seq_add; | |
275 | ||
276 | /* | |
277 | * If we reach beyond the end of the file | |
278 | * with holed IO, wrap around to the | |
279 | * beginning again. | |
280 | */ | |
281 | if (pos >= f->real_file_size) | |
282 | pos = f->file_offset; | |
283 | } | |
059b0802 | 284 | |
37cf9e3c | 285 | *offset = pos; |
38dad62d JA |
286 | return 0; |
287 | } | |
288 | ||
289 | return 1; | |
290 | } | |
291 | ||
292 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
293 | enum fio_ddir ddir, int rw_seq, |
294 | unsigned int *is_random) | |
38dad62d JA |
295 | { |
296 | struct fio_file *f = io_u->file; | |
1ae83d45 | 297 | uint64_t b, offset; |
38dad62d JA |
298 | int ret; |
299 | ||
ff58fced JA |
300 | assert(ddir_rw(ddir)); |
301 | ||
37cf9e3c JA |
302 | b = offset = -1ULL; |
303 | ||
38dad62d | 304 | if (rw_seq) { |
211c9b89 | 305 | if (td_random(td)) { |
6aca9b3d | 306 | if (should_do_random(td, ddir)) { |
211c9b89 | 307 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
308 | *is_random = 1; |
309 | } else { | |
310 | *is_random = 0; | |
211c9b89 JA |
311 | io_u->flags |= IO_U_F_BUSY_OK; |
312 | ret = get_next_seq_offset(td, f, ddir, &offset); | |
313 | if (ret) | |
314 | ret = get_next_rand_block(td, f, ddir, &b); | |
315 | } | |
6aca9b3d JA |
316 | } else { |
317 | *is_random = 0; | |
37cf9e3c | 318 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 319 | } |
38dad62d JA |
320 | } else { |
321 | io_u->flags |= IO_U_F_BUSY_OK; | |
6aca9b3d | 322 | *is_random = 0; |
38dad62d JA |
323 | |
324 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 325 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 326 | if (ret) { |
37cf9e3c | 327 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
328 | *is_random = 0; |
329 | } | |
38dad62d JA |
330 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
331 | if (f->last_start != -1ULL) | |
37cf9e3c | 332 | offset = f->last_start - f->file_offset; |
38dad62d | 333 | else |
37cf9e3c | 334 | offset = 0; |
38dad62d JA |
335 | ret = 0; |
336 | } else { | |
337 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
338 | ret = 1; | |
339 | } | |
340 | } | |
6d68b997 | 341 | |
37cf9e3c JA |
342 | if (!ret) { |
343 | if (offset != -1ULL) | |
344 | io_u->offset = offset; | |
345 | else if (b != -1ULL) | |
346 | io_u->offset = b * td->o.ba[ddir]; | |
347 | else { | |
4e0a8fa2 | 348 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
349 | ret = 1; |
350 | } | |
351 | } | |
352 | ||
38dad62d JA |
353 | return ret; |
354 | } | |
355 | ||
10ba535a JA |
356 | /* |
357 | * For random io, generate a random new block and see if it's used. Repeat | |
358 | * until we find a free one. For sequential io, just return the end of | |
359 | * the last io issued. | |
360 | */ | |
6aca9b3d JA |
361 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
362 | unsigned int *is_random) | |
10ba535a | 363 | { |
9bf2061e | 364 | struct fio_file *f = io_u->file; |
4ba66134 | 365 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 366 | int rw_seq_hit = 0; |
10ba535a | 367 | |
ff58fced JA |
368 | assert(ddir_rw(ddir)); |
369 | ||
38dad62d JA |
370 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
371 | rw_seq_hit = 1; | |
5736c10d | 372 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 373 | } |
211097b2 | 374 | |
6aca9b3d | 375 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 376 | return 1; |
10ba535a | 377 | |
009bd847 JA |
378 | if (io_u->offset >= f->io_size) { |
379 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
380 | (unsigned long long) io_u->offset, |
381 | (unsigned long long) f->io_size); | |
009bd847 JA |
382 | return 1; |
383 | } | |
384 | ||
385 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
386 | if (io_u->offset >= f->real_file_size) { |
387 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
388 | (unsigned long long) io_u->offset, |
389 | (unsigned long long) f->real_file_size); | |
10ba535a | 390 | return 1; |
2ba1c290 | 391 | } |
10ba535a JA |
392 | |
393 | return 0; | |
394 | } | |
395 | ||
6aca9b3d JA |
396 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
397 | unsigned int *is_random) | |
15dc1934 | 398 | { |
d72be545 JA |
399 | if (td->flags & TD_F_PROFILE_OPS) { |
400 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 401 | |
d72be545 | 402 | if (ops->fill_io_u_off) |
6aca9b3d | 403 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 404 | } |
15dc1934 | 405 | |
6aca9b3d | 406 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
407 | } |
408 | ||
79944128 JA |
409 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
410 | unsigned int buflen) | |
411 | { | |
412 | struct fio_file *f = io_u->file; | |
413 | ||
bedc9dc2 | 414 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
415 | } |
416 | ||
6aca9b3d JA |
417 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
418 | unsigned int is_random) | |
10ba535a | 419 | { |
6aca9b3d | 420 | int ddir = io_u->ddir; |
24d23ca7 | 421 | unsigned int buflen = 0; |
f3059de1 | 422 | unsigned int minbs, maxbs; |
54a21917 | 423 | unsigned long r; |
10ba535a | 424 | |
9ee1c647 | 425 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
426 | |
427 | if (td->o.bs_is_seq_rand) | |
428 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 429 | |
f3059de1 JA |
430 | minbs = td->o.min_bs[ddir]; |
431 | maxbs = td->o.max_bs[ddir]; | |
432 | ||
79944128 JA |
433 | if (minbs == maxbs) |
434 | return minbs; | |
435 | ||
52c58027 JA |
436 | /* |
437 | * If we can't satisfy the min block size from here, then fail | |
438 | */ | |
439 | if (!io_u_fits(td, io_u, minbs)) | |
440 | return 0; | |
441 | ||
79944128 | 442 | do { |
d6b72507 | 443 | r = __rand(&td->bsrange_state); |
4c07ad86 | 444 | |
720e84ad | 445 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 446 | buflen = 1 + (unsigned int) ((double) maxbs * |
54a21917 | 447 | (r / (FRAND_MAX + 1.0))); |
f3059de1 JA |
448 | if (buflen < minbs) |
449 | buflen = minbs; | |
5ec10eaa | 450 | } else { |
564ca972 JA |
451 | long perc = 0; |
452 | unsigned int i; | |
453 | ||
720e84ad JA |
454 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
455 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
456 | |
457 | buflen = bsp->bs; | |
458 | perc += bsp->perc; | |
54a21917 | 459 | if ((r <= ((FRAND_MAX / 100L) * perc)) && |
79944128 | 460 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
461 | break; |
462 | } | |
463 | } | |
79944128 | 464 | |
a9f70b1f JB |
465 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
466 | buflen = (buflen + td->o.verify_interval - 1) & | |
467 | ~(td->o.verify_interval - 1); | |
468 | ||
f3059de1 JA |
469 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
470 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 471 | |
79944128 | 472 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 473 | |
10ba535a JA |
474 | return buflen; |
475 | } | |
476 | ||
6aca9b3d JA |
477 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
478 | unsigned int is_random) | |
15dc1934 | 479 | { |
d72be545 JA |
480 | if (td->flags & TD_F_PROFILE_OPS) { |
481 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 482 | |
d72be545 | 483 | if (ops->fill_io_u_size) |
6aca9b3d | 484 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 485 | } |
15dc1934 | 486 | |
6aca9b3d | 487 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
488 | } |
489 | ||
afe24a5a JA |
490 | static void set_rwmix_bytes(struct thread_data *td) |
491 | { | |
afe24a5a JA |
492 | unsigned int diff; |
493 | ||
494 | /* | |
495 | * we do time or byte based switch. this is needed because | |
496 | * buffered writes may issue a lot quicker than they complete, | |
497 | * whereas reads do not. | |
498 | */ | |
e47f799f | 499 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 500 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
501 | } |
502 | ||
503 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
504 | { | |
505 | unsigned int v; | |
1294c3ec | 506 | unsigned long r; |
e47f799f | 507 | |
d6b72507 | 508 | r = __rand(&td->rwmix_state); |
54a21917 | 509 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
4c07ad86 | 510 | |
04c540d9 | 511 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
512 | return DDIR_READ; |
513 | ||
514 | return DDIR_WRITE; | |
afe24a5a JA |
515 | } |
516 | ||
002e7183 JA |
517 | void io_u_quiesce(struct thread_data *td) |
518 | { | |
519 | /* | |
520 | * We are going to sleep, ensure that we flush anything pending as | |
521 | * not to skew our latency numbers. | |
522 | * | |
523 | * Changed to only monitor 'in flight' requests here instead of the | |
524 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
525 | * io's that have been actually submitted to an async engine, | |
526 | * and cur_depth is meaningless for sync engines. | |
527 | */ | |
528 | while (td->io_u_in_flight) { | |
529 | int fio_unused ret; | |
530 | ||
531 | ret = io_u_queued_complete(td, 1, NULL); | |
532 | } | |
533 | } | |
534 | ||
581e7141 JA |
535 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
536 | { | |
537 | enum fio_ddir odir = ddir ^ 1; | |
538 | struct timeval t; | |
539 | long usec; | |
540 | ||
ff58fced JA |
541 | assert(ddir_rw(ddir)); |
542 | ||
315fcfec | 543 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
544 | return ddir; |
545 | ||
546 | /* | |
547 | * We have too much pending sleep in this direction. See if we | |
548 | * should switch. | |
549 | */ | |
315fcfec | 550 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
551 | /* |
552 | * Other direction does not have too much pending, switch | |
553 | */ | |
554 | if (td->rate_pending_usleep[odir] < 100000) | |
555 | return odir; | |
556 | ||
557 | /* | |
558 | * Both directions have pending sleep. Sleep the minimum time | |
559 | * and deduct from both. | |
560 | */ | |
561 | if (td->rate_pending_usleep[ddir] <= | |
562 | td->rate_pending_usleep[odir]) { | |
563 | usec = td->rate_pending_usleep[ddir]; | |
564 | } else { | |
565 | usec = td->rate_pending_usleep[odir]; | |
566 | ddir = odir; | |
567 | } | |
568 | } else | |
569 | usec = td->rate_pending_usleep[ddir]; | |
570 | ||
002e7183 | 571 | io_u_quiesce(td); |
78c1eda5 | 572 | |
581e7141 JA |
573 | fio_gettime(&t, NULL); |
574 | usec_sleep(td, usec); | |
575 | usec = utime_since_now(&t); | |
576 | ||
577 | td->rate_pending_usleep[ddir] -= usec; | |
578 | ||
579 | odir = ddir ^ 1; | |
580 | if (td_rw(td) && __should_check_rate(td, odir)) | |
581 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 582 | |
6eaf09d6 SL |
583 | if (ddir_trim(ddir)) |
584 | return ddir; | |
e0224c6b | 585 | |
581e7141 JA |
586 | return ddir; |
587 | } | |
588 | ||
10ba535a JA |
589 | /* |
590 | * Return the data direction for the next io_u. If the job is a | |
591 | * mixed read/write workload, check the rwmix cycle and switch if | |
592 | * necessary. | |
593 | */ | |
1e97cce9 | 594 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 595 | { |
581e7141 JA |
596 | enum fio_ddir ddir; |
597 | ||
5f9099ea JA |
598 | /* |
599 | * see if it's time to fsync | |
600 | */ | |
601 | if (td->o.fsync_blocks && | |
602 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
603 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
604 | return DDIR_SYNC; | |
605 | ||
606 | /* | |
607 | * see if it's time to fdatasync | |
608 | */ | |
609 | if (td->o.fdatasync_blocks && | |
610 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
611 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
612 | return DDIR_DATASYNC; | |
613 | ||
44f29692 JA |
614 | /* |
615 | * see if it's time to sync_file_range | |
616 | */ | |
617 | if (td->sync_file_range_nr && | |
618 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
619 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
620 | return DDIR_SYNC_FILE_RANGE; | |
621 | ||
10ba535a | 622 | if (td_rw(td)) { |
10ba535a JA |
623 | /* |
624 | * Check if it's time to seed a new data direction. | |
625 | */ | |
e4928662 | 626 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
627 | /* |
628 | * Put a top limit on how many bytes we do for | |
629 | * one data direction, to avoid overflowing the | |
630 | * ranges too much | |
631 | */ | |
632 | ddir = get_rand_ddir(td); | |
e47f799f JA |
633 | |
634 | if (ddir != td->rwmix_ddir) | |
635 | set_rwmix_bytes(td); | |
636 | ||
637 | td->rwmix_ddir = ddir; | |
10ba535a | 638 | } |
581e7141 | 639 | ddir = td->rwmix_ddir; |
10ba535a | 640 | } else if (td_read(td)) |
581e7141 | 641 | ddir = DDIR_READ; |
6eaf09d6 | 642 | else if (td_write(td)) |
581e7141 | 643 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
644 | else |
645 | ddir = DDIR_TRIM; | |
581e7141 JA |
646 | |
647 | td->rwmix_ddir = rate_ddir(td, ddir); | |
648 | return td->rwmix_ddir; | |
10ba535a JA |
649 | } |
650 | ||
1ef2b6be JA |
651 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
652 | { | |
bcd5abfa | 653 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
654 | |
655 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
656 | td->o.barrier_blocks && | |
657 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
658 | td->io_issues[DDIR_WRITE]) | |
659 | io_u->flags |= IO_U_F_BARRIER; | |
660 | } | |
661 | ||
e8462bd8 | 662 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 663 | { |
71b84caa | 664 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
665 | |
666 | if (ret) | |
667 | td_verror(td, ret, "file close"); | |
668 | } | |
669 | ||
10ba535a JA |
670 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
671 | { | |
e8462bd8 JA |
672 | td_io_u_lock(td); |
673 | ||
f8b0bd10 | 674 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 675 | put_file_log(td, io_u->file); |
f8b0bd10 | 676 | |
10ba535a | 677 | io_u->file = NULL; |
d7ee2a7d SL |
678 | io_u->flags |= IO_U_F_FREE; |
679 | ||
0c41214f RR |
680 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
681 | td->cur_depth--; | |
2ae0b204 | 682 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
683 | td_io_u_unlock(td); |
684 | td_io_u_free_notify(td); | |
10ba535a JA |
685 | } |
686 | ||
f2bba182 RR |
687 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
688 | { | |
689 | io_u->flags &= ~IO_U_F_FLIGHT; | |
690 | put_io_u(td, io_u); | |
691 | } | |
692 | ||
755200a3 JA |
693 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
694 | { | |
695 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 696 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 697 | |
465221b0 JA |
698 | dprint(FD_IO, "requeue %p\n", __io_u); |
699 | ||
e8462bd8 JA |
700 | td_io_u_lock(td); |
701 | ||
4d2e0f49 | 702 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
703 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
704 | td->io_issues[ddir]--; | |
5ec10eaa | 705 | |
4d2e0f49 | 706 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
707 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
708 | td->cur_depth--; | |
2ae0b204 JA |
709 | |
710 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 711 | td_io_u_unlock(td); |
755200a3 JA |
712 | *io_u = NULL; |
713 | } | |
714 | ||
9bf2061e | 715 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 716 | { |
6aca9b3d JA |
717 | unsigned int is_random; |
718 | ||
b4c5e1ac JA |
719 | if (td->io_ops->flags & FIO_NOIO) |
720 | goto out; | |
721 | ||
1ef2b6be | 722 | set_rw_ddir(td, io_u); |
5f9099ea | 723 | |
87dc1ab1 | 724 | /* |
ff58fced | 725 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 726 | */ |
ff58fced | 727 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 728 | goto out; |
a00735e6 | 729 | |
48f5abd3 JA |
730 | /* |
731 | * See if it's time to switch to a new zone | |
732 | */ | |
13af05ae | 733 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
734 | struct fio_file *f = io_u->file; |
735 | ||
48f5abd3 | 736 | td->zone_bytes = 0; |
4c8be5b1 JA |
737 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
738 | ||
739 | /* | |
740 | * Wrap from the beginning, if we exceed the file size | |
741 | */ | |
742 | if (f->file_offset >= f->real_file_size) | |
743 | f->file_offset = f->real_file_size - f->file_offset; | |
744 | f->last_pos = f->file_offset; | |
48f5abd3 JA |
745 | td->io_skip_bytes += td->o.zone_skip; |
746 | } | |
747 | ||
10ba535a | 748 | /* |
c685b5b2 JA |
749 | * No log, let the seq/rand engine retrieve the next buflen and |
750 | * position. | |
10ba535a | 751 | */ |
6aca9b3d | 752 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 753 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 754 | return 1; |
2ba1c290 | 755 | } |
10ba535a | 756 | |
6aca9b3d | 757 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
758 | if (!io_u->buflen) { |
759 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 760 | return 1; |
2ba1c290 | 761 | } |
bca4ed4d | 762 | |
2ba1c290 JA |
763 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
764 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
765 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
766 | (unsigned long long) io_u->offset, io_u->buflen, | |
767 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 768 | return 1; |
2ba1c290 | 769 | } |
6a5e6884 | 770 | |
bca4ed4d JA |
771 | /* |
772 | * mark entry before potentially trimming io_u | |
773 | */ | |
303032ae | 774 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 775 | mark_random_map(td, io_u); |
bca4ed4d | 776 | |
c38e9468 | 777 | out: |
2ba1c290 | 778 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 779 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 780 | return 0; |
10ba535a JA |
781 | } |
782 | ||
838bc709 JA |
783 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
784 | { | |
2b13e716 | 785 | int idx = 0; |
838bc709 JA |
786 | |
787 | switch (nr) { | |
788 | default: | |
2b13e716 | 789 | idx = 6; |
838bc709 JA |
790 | break; |
791 | case 33 ... 64: | |
2b13e716 | 792 | idx = 5; |
838bc709 JA |
793 | break; |
794 | case 17 ... 32: | |
2b13e716 | 795 | idx = 4; |
838bc709 JA |
796 | break; |
797 | case 9 ... 16: | |
2b13e716 | 798 | idx = 3; |
838bc709 JA |
799 | break; |
800 | case 5 ... 8: | |
2b13e716 | 801 | idx = 2; |
838bc709 JA |
802 | break; |
803 | case 1 ... 4: | |
2b13e716 | 804 | idx = 1; |
838bc709 JA |
805 | case 0: |
806 | break; | |
807 | } | |
808 | ||
2b13e716 | 809 | map[idx]++; |
838bc709 JA |
810 | } |
811 | ||
812 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
813 | { | |
814 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
815 | td->ts.total_submit++; | |
816 | } | |
817 | ||
818 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
819 | { | |
820 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
821 | td->ts.total_complete++; | |
822 | } | |
823 | ||
d8005759 | 824 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 825 | { |
2b13e716 | 826 | int idx = 0; |
71619dc2 JA |
827 | |
828 | switch (td->cur_depth) { | |
829 | default: | |
2b13e716 | 830 | idx = 6; |
a783e61a | 831 | break; |
71619dc2 | 832 | case 32 ... 63: |
2b13e716 | 833 | idx = 5; |
a783e61a | 834 | break; |
71619dc2 | 835 | case 16 ... 31: |
2b13e716 | 836 | idx = 4; |
a783e61a | 837 | break; |
71619dc2 | 838 | case 8 ... 15: |
2b13e716 | 839 | idx = 3; |
a783e61a | 840 | break; |
71619dc2 | 841 | case 4 ... 7: |
2b13e716 | 842 | idx = 2; |
a783e61a | 843 | break; |
71619dc2 | 844 | case 2 ... 3: |
2b13e716 | 845 | idx = 1; |
71619dc2 JA |
846 | case 1: |
847 | break; | |
848 | } | |
849 | ||
2b13e716 | 850 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
851 | } |
852 | ||
04a0feae JA |
853 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
854 | { | |
2b13e716 | 855 | int idx = 0; |
04a0feae JA |
856 | |
857 | assert(usec < 1000); | |
858 | ||
859 | switch (usec) { | |
860 | case 750 ... 999: | |
2b13e716 | 861 | idx = 9; |
04a0feae JA |
862 | break; |
863 | case 500 ... 749: | |
2b13e716 | 864 | idx = 8; |
04a0feae JA |
865 | break; |
866 | case 250 ... 499: | |
2b13e716 | 867 | idx = 7; |
04a0feae JA |
868 | break; |
869 | case 100 ... 249: | |
2b13e716 | 870 | idx = 6; |
04a0feae JA |
871 | break; |
872 | case 50 ... 99: | |
2b13e716 | 873 | idx = 5; |
04a0feae JA |
874 | break; |
875 | case 20 ... 49: | |
2b13e716 | 876 | idx = 4; |
04a0feae JA |
877 | break; |
878 | case 10 ... 19: | |
2b13e716 | 879 | idx = 3; |
04a0feae JA |
880 | break; |
881 | case 4 ... 9: | |
2b13e716 | 882 | idx = 2; |
04a0feae JA |
883 | break; |
884 | case 2 ... 3: | |
2b13e716 | 885 | idx = 1; |
04a0feae JA |
886 | case 0 ... 1: |
887 | break; | |
888 | } | |
889 | ||
2b13e716 JA |
890 | assert(idx < FIO_IO_U_LAT_U_NR); |
891 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
892 | } |
893 | ||
894 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 895 | { |
2b13e716 | 896 | int idx = 0; |
ec118304 JA |
897 | |
898 | switch (msec) { | |
899 | default: | |
2b13e716 | 900 | idx = 11; |
04a0feae | 901 | break; |
8abdce66 | 902 | case 1000 ... 1999: |
2b13e716 | 903 | idx = 10; |
04a0feae | 904 | break; |
8abdce66 | 905 | case 750 ... 999: |
2b13e716 | 906 | idx = 9; |
04a0feae | 907 | break; |
8abdce66 | 908 | case 500 ... 749: |
2b13e716 | 909 | idx = 8; |
04a0feae | 910 | break; |
8abdce66 | 911 | case 250 ... 499: |
2b13e716 | 912 | idx = 7; |
04a0feae | 913 | break; |
8abdce66 | 914 | case 100 ... 249: |
2b13e716 | 915 | idx = 6; |
04a0feae | 916 | break; |
8abdce66 | 917 | case 50 ... 99: |
2b13e716 | 918 | idx = 5; |
04a0feae | 919 | break; |
8abdce66 | 920 | case 20 ... 49: |
2b13e716 | 921 | idx = 4; |
04a0feae | 922 | break; |
8abdce66 | 923 | case 10 ... 19: |
2b13e716 | 924 | idx = 3; |
04a0feae | 925 | break; |
8abdce66 | 926 | case 4 ... 9: |
2b13e716 | 927 | idx = 2; |
04a0feae | 928 | break; |
ec118304 | 929 | case 2 ... 3: |
2b13e716 | 930 | idx = 1; |
ec118304 JA |
931 | case 0 ... 1: |
932 | break; | |
933 | } | |
934 | ||
2b13e716 JA |
935 | assert(idx < FIO_IO_U_LAT_M_NR); |
936 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
937 | } |
938 | ||
939 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
940 | { | |
941 | if (usec < 1000) | |
942 | io_u_mark_lat_usec(td, usec); | |
943 | else | |
944 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
945 | } |
946 | ||
0aabe160 JA |
947 | /* |
948 | * Get next file to service by choosing one at random | |
949 | */ | |
2cc52930 JA |
950 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
951 | enum fio_file_flags goodf, | |
d6aed795 | 952 | enum fio_file_flags badf) |
0aabe160 | 953 | { |
0aabe160 | 954 | struct fio_file *f; |
1c178180 | 955 | int fno; |
0aabe160 JA |
956 | |
957 | do { | |
87b10676 | 958 | int opened = 0; |
1294c3ec | 959 | unsigned long r; |
4c07ad86 | 960 | |
d6b72507 | 961 | r = __rand(&td->next_file_state); |
54a21917 | 962 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 963 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 964 | |
126d65c6 | 965 | f = td->files[fno]; |
d6aed795 | 966 | if (fio_file_done(f)) |
059e63c0 | 967 | continue; |
1c178180 | 968 | |
d6aed795 | 969 | if (!fio_file_open(f)) { |
87b10676 JA |
970 | int err; |
971 | ||
002fe734 JA |
972 | if (td->nr_open_files >= td->o.open_files) |
973 | return ERR_PTR(-EBUSY); | |
974 | ||
87b10676 JA |
975 | err = td_io_open_file(td, f); |
976 | if (err) | |
977 | continue; | |
978 | opened = 1; | |
979 | } | |
980 | ||
2ba1c290 JA |
981 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
982 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 983 | return f; |
2ba1c290 | 984 | } |
87b10676 JA |
985 | if (opened) |
986 | td_io_close_file(td, f); | |
0aabe160 JA |
987 | } while (1); |
988 | } | |
989 | ||
990 | /* | |
991 | * Get next file to service by doing round robin between all available ones | |
992 | */ | |
1c178180 JA |
993 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
994 | int badf) | |
3d7c391d JA |
995 | { |
996 | unsigned int old_next_file = td->next_file; | |
997 | struct fio_file *f; | |
998 | ||
999 | do { | |
87b10676 JA |
1000 | int opened = 0; |
1001 | ||
126d65c6 | 1002 | f = td->files[td->next_file]; |
3d7c391d JA |
1003 | |
1004 | td->next_file++; | |
2dc1bbeb | 1005 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1006 | td->next_file = 0; |
1007 | ||
87b10676 | 1008 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1009 | if (fio_file_done(f)) { |
d5ed68ea | 1010 | f = NULL; |
059e63c0 | 1011 | continue; |
d5ed68ea | 1012 | } |
059e63c0 | 1013 | |
d6aed795 | 1014 | if (!fio_file_open(f)) { |
87b10676 JA |
1015 | int err; |
1016 | ||
002fe734 JA |
1017 | if (td->nr_open_files >= td->o.open_files) |
1018 | return ERR_PTR(-EBUSY); | |
1019 | ||
87b10676 | 1020 | err = td_io_open_file(td, f); |
b5696bfc JA |
1021 | if (err) { |
1022 | dprint(FD_FILE, "error %d on open of %s\n", | |
1023 | err, f->file_name); | |
87c27b45 | 1024 | f = NULL; |
87b10676 | 1025 | continue; |
b5696bfc | 1026 | } |
87b10676 JA |
1027 | opened = 1; |
1028 | } | |
1029 | ||
0b9d69ec JA |
1030 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1031 | f->flags); | |
1c178180 | 1032 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1033 | break; |
1034 | ||
87b10676 JA |
1035 | if (opened) |
1036 | td_io_close_file(td, f); | |
1037 | ||
3d7c391d JA |
1038 | f = NULL; |
1039 | } while (td->next_file != old_next_file); | |
1040 | ||
2ba1c290 | 1041 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1042 | return f; |
1043 | } | |
1044 | ||
7eb36574 | 1045 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1046 | { |
1907dbc6 JA |
1047 | struct fio_file *f; |
1048 | ||
2dc1bbeb | 1049 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1050 | |
b5696bfc | 1051 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1052 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1053 | " nr_files=%d\n", td->nr_open_files, | |
1054 | td->nr_done_files, | |
1055 | td->o.nr_files); | |
bdb4e2e9 | 1056 | return NULL; |
2ba1c290 | 1057 | } |
bdb4e2e9 | 1058 | |
1907dbc6 | 1059 | f = td->file_service_file; |
d6aed795 | 1060 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1061 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1062 | goto out; | |
1063 | if (td->file_service_left--) | |
1064 | goto out; | |
1065 | } | |
1907dbc6 | 1066 | |
a086c257 JA |
1067 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1068 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1069 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1070 | else |
d6aed795 | 1071 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1072 | |
002fe734 JA |
1073 | if (IS_ERR(f)) |
1074 | return f; | |
1075 | ||
1907dbc6 JA |
1076 | td->file_service_file = f; |
1077 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1078 | out: |
0dac421f JA |
1079 | if (f) |
1080 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1081 | else | |
1082 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1083 | return f; |
bdb4e2e9 JA |
1084 | } |
1085 | ||
7eb36574 JA |
1086 | static struct fio_file *get_next_file(struct thread_data *td) |
1087 | { | |
372d8962 | 1088 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1089 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1090 | |
d72be545 JA |
1091 | if (ops->get_next_file) |
1092 | return ops->get_next_file(td); | |
1093 | } | |
7eb36574 JA |
1094 | |
1095 | return __get_next_file(td); | |
1096 | } | |
1097 | ||
002fe734 | 1098 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1099 | { |
1100 | struct fio_file *f; | |
1101 | ||
1102 | do { | |
1103 | f = get_next_file(td); | |
002fe734 JA |
1104 | if (IS_ERR_OR_NULL(f)) |
1105 | return PTR_ERR(f); | |
429f6675 | 1106 | |
429f6675 JA |
1107 | io_u->file = f; |
1108 | get_file(f); | |
1109 | ||
1110 | if (!fill_io_u(td, io_u)) | |
1111 | break; | |
1112 | ||
b5696bfc | 1113 | put_file_log(td, f); |
429f6675 | 1114 | td_io_close_file(td, f); |
b5696bfc | 1115 | io_u->file = NULL; |
d6aed795 | 1116 | fio_file_set_done(f); |
429f6675 | 1117 | td->nr_done_files++; |
0b9d69ec JA |
1118 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1119 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1120 | } while (1); |
1121 | ||
1122 | return 0; | |
1123 | } | |
1124 | ||
3e260a46 JA |
1125 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1126 | unsigned long tusec, unsigned long max_usec) | |
1127 | { | |
1128 | if (!td->error) | |
1129 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1130 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1131 | icd->error = ETIMEDOUT; | |
1132 | } | |
1133 | ||
1134 | static void lat_new_cycle(struct thread_data *td) | |
1135 | { | |
1136 | fio_gettime(&td->latency_ts, NULL); | |
1137 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1138 | td->latency_failed = 0; | |
1139 | } | |
1140 | ||
1141 | /* | |
1142 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1143 | * are at QD=1, then it's time to give up. | |
1144 | */ | |
1145 | static int __lat_target_failed(struct thread_data *td) | |
1146 | { | |
1147 | if (td->latency_qd == 1) | |
1148 | return 1; | |
1149 | ||
1150 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1151 | |
1152 | if (td->latency_qd == td->latency_qd_low) | |
1153 | td->latency_qd_low--; | |
1154 | ||
3e260a46 JA |
1155 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1156 | ||
1157 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1158 | ||
1159 | /* | |
1160 | * When we ramp QD down, quiesce existing IO to prevent | |
1161 | * a storm of ramp downs due to pending higher depth. | |
1162 | */ | |
1163 | io_u_quiesce(td); | |
1164 | lat_new_cycle(td); | |
1165 | return 0; | |
1166 | } | |
1167 | ||
1168 | static int lat_target_failed(struct thread_data *td) | |
1169 | { | |
1170 | if (td->o.latency_percentile.u.f == 100.0) | |
1171 | return __lat_target_failed(td); | |
1172 | ||
1173 | td->latency_failed++; | |
1174 | return 0; | |
1175 | } | |
1176 | ||
1177 | void lat_target_init(struct thread_data *td) | |
1178 | { | |
6bb58215 JA |
1179 | td->latency_end_run = 0; |
1180 | ||
3e260a46 JA |
1181 | if (td->o.latency_target) { |
1182 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1183 | fio_gettime(&td->latency_ts, NULL); | |
1184 | td->latency_qd = 1; | |
1185 | td->latency_qd_high = td->o.iodepth; | |
1186 | td->latency_qd_low = 1; | |
1187 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1188 | } else | |
1189 | td->latency_qd = td->o.iodepth; | |
1190 | } | |
1191 | ||
6bb58215 JA |
1192 | void lat_target_reset(struct thread_data *td) |
1193 | { | |
1194 | if (!td->latency_end_run) | |
1195 | lat_target_init(td); | |
1196 | } | |
1197 | ||
3e260a46 JA |
1198 | static void lat_target_success(struct thread_data *td) |
1199 | { | |
1200 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1201 | struct thread_options *o = &td->o; |
3e260a46 JA |
1202 | |
1203 | td->latency_qd_low = td->latency_qd; | |
1204 | ||
1205 | /* | |
1206 | * If we haven't failed yet, we double up to a failing value instead | |
1207 | * of bisecting from highest possible queue depth. If we have set | |
1208 | * a limit other than td->o.iodepth, bisect between that. | |
1209 | */ | |
6bb58215 | 1210 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1211 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1212 | else | |
1213 | td->latency_qd *= 2; | |
1214 | ||
6bb58215 JA |
1215 | if (td->latency_qd > o->iodepth) |
1216 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1217 | |
1218 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1219 | |
3e260a46 | 1220 | /* |
6bb58215 JA |
1221 | * Same as last one, we are done. Let it run a latency cycle, so |
1222 | * we get only the results from the targeted depth. | |
3e260a46 | 1223 | */ |
6bb58215 JA |
1224 | if (td->latency_qd == qd) { |
1225 | if (td->latency_end_run) { | |
1226 | dprint(FD_RATE, "We are done\n"); | |
1227 | td->done = 1; | |
1228 | } else { | |
1229 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1230 | io_u_quiesce(td); | |
1231 | td->latency_end_run = 1; | |
1232 | reset_all_stats(td); | |
1233 | reset_io_stats(td); | |
1234 | } | |
1235 | } | |
3e260a46 JA |
1236 | |
1237 | lat_new_cycle(td); | |
1238 | } | |
1239 | ||
1240 | /* | |
1241 | * Check if we can bump the queue depth | |
1242 | */ | |
1243 | void lat_target_check(struct thread_data *td) | |
1244 | { | |
1245 | uint64_t usec_window; | |
1246 | uint64_t ios; | |
1247 | double success_ios; | |
1248 | ||
1249 | usec_window = utime_since_now(&td->latency_ts); | |
1250 | if (usec_window < td->o.latency_window) | |
1251 | return; | |
1252 | ||
1253 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1254 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1255 | success_ios *= 100.0; | |
1256 | ||
1257 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1258 | ||
1259 | if (success_ios >= td->o.latency_percentile.u.f) | |
1260 | lat_target_success(td); | |
1261 | else | |
1262 | __lat_target_failed(td); | |
1263 | } | |
1264 | ||
1265 | /* | |
1266 | * If latency target is enabled, we might be ramping up or down and not | |
1267 | * using the full queue depth available. | |
1268 | */ | |
effd6ff0 | 1269 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1270 | { |
1271 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1272 | ||
1273 | if (qempty) | |
1274 | return 1; | |
1275 | if (!td->o.latency_target) | |
1276 | return 0; | |
1277 | ||
1278 | return td->cur_depth >= td->latency_qd; | |
1279 | } | |
429f6675 | 1280 | |
10ba535a JA |
1281 | struct io_u *__get_io_u(struct thread_data *td) |
1282 | { | |
0cae66f6 | 1283 | struct io_u *io_u = NULL; |
10ba535a | 1284 | |
e8462bd8 JA |
1285 | td_io_u_lock(td); |
1286 | ||
1287 | again: | |
2ae0b204 JA |
1288 | if (!io_u_rempty(&td->io_u_requeues)) |
1289 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1290 | else if (!queue_full(td)) { |
2ae0b204 | 1291 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1292 | |
225ba9e3 | 1293 | io_u->file = NULL; |
6040dabc | 1294 | io_u->buflen = 0; |
10ba535a | 1295 | io_u->resid = 0; |
d7762cf8 | 1296 | io_u->end_io = NULL; |
755200a3 JA |
1297 | } |
1298 | ||
1299 | if (io_u) { | |
0c6e7517 | 1300 | assert(io_u->flags & IO_U_F_FREE); |
f8b0bd10 JA |
1301 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
1302 | IO_U_F_TRIMMED | IO_U_F_BARRIER | | |
1303 | IO_U_F_VER_LIST); | |
0c6e7517 | 1304 | |
755200a3 | 1305 | io_u->error = 0; |
bcd5abfa | 1306 | io_u->acct_ddir = -1; |
10ba535a | 1307 | td->cur_depth++; |
0c41214f | 1308 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1309 | io_u->ipo = NULL; |
1dec3e07 JA |
1310 | } else if (td->o.verify_async) { |
1311 | /* | |
1312 | * We ran out, wait for async verify threads to finish and | |
1313 | * return one | |
1314 | */ | |
1315 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1316 | goto again; | |
10ba535a JA |
1317 | } |
1318 | ||
e8462bd8 | 1319 | td_io_u_unlock(td); |
10ba535a JA |
1320 | return io_u; |
1321 | } | |
1322 | ||
0d29de83 | 1323 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1324 | { |
d72be545 JA |
1325 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1326 | return 0; | |
1327 | ||
1328 | if (td->trim_entries) { | |
0d29de83 | 1329 | int get_trim = 0; |
10ba535a | 1330 | |
0d29de83 JA |
1331 | if (td->trim_batch) { |
1332 | td->trim_batch--; | |
1333 | get_trim = 1; | |
1334 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1335 | td->last_ddir != DDIR_READ) { | |
1336 | td->trim_batch = td->o.trim_batch; | |
1337 | if (!td->trim_batch) | |
1338 | td->trim_batch = td->o.trim_backlog; | |
1339 | get_trim = 1; | |
1340 | } | |
1341 | ||
1342 | if (get_trim && !get_next_trim(td, io_u)) | |
1343 | return 1; | |
2ba1c290 | 1344 | } |
10ba535a | 1345 | |
0d29de83 JA |
1346 | return 0; |
1347 | } | |
1348 | ||
1349 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1350 | { | |
d72be545 JA |
1351 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1352 | return 0; | |
1353 | ||
1354 | if (td->io_hist_len) { | |
9e144189 JA |
1355 | int get_verify = 0; |
1356 | ||
d1ece0c7 | 1357 | if (td->verify_batch) |
9e144189 | 1358 | get_verify = 1; |
d1ece0c7 | 1359 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1360 | td->last_ddir != DDIR_READ) { |
1361 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1362 | if (!td->verify_batch) |
1363 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1364 | get_verify = 1; |
1365 | } | |
1366 | ||
d1ece0c7 JA |
1367 | if (get_verify && !get_next_verify(td, io_u)) { |
1368 | td->verify_batch--; | |
0d29de83 | 1369 | return 1; |
d1ece0c7 | 1370 | } |
9e144189 JA |
1371 | } |
1372 | ||
0d29de83 JA |
1373 | return 0; |
1374 | } | |
1375 | ||
de789769 JA |
1376 | /* |
1377 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1378 | * easy compressible data for simple de-dupe attempts. Do this for every |
1379 | * 512b block in the range, since that should be the smallest block size | |
1380 | * we can expect from a device. | |
de789769 JA |
1381 | */ |
1382 | static void small_content_scramble(struct io_u *io_u) | |
1383 | { | |
23f394d5 | 1384 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1385 | uint64_t boffset; |
23f394d5 JA |
1386 | unsigned int offset; |
1387 | void *p, *end; | |
de789769 | 1388 | |
23f394d5 JA |
1389 | if (!nr_blocks) |
1390 | return; | |
1391 | ||
1392 | p = io_u->xfer_buf; | |
fba76ee8 | 1393 | boffset = io_u->offset; |
81f0366c | 1394 | io_u->buf_filled_len = 0; |
fad82f76 | 1395 | |
23f394d5 JA |
1396 | for (i = 0; i < nr_blocks; i++) { |
1397 | /* | |
1398 | * Fill the byte offset into a "random" start offset of | |
1399 | * the buffer, given by the product of the usec time | |
1400 | * and the actual offset. | |
1401 | */ | |
fad82f76 | 1402 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1403 | offset &= ~(sizeof(uint64_t) - 1); |
1404 | if (offset >= 512 - sizeof(uint64_t)) | |
1405 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1406 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1407 | |
1408 | end = p + 512 - sizeof(io_u->start_time); | |
1409 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1410 | p += 512; | |
fad82f76 | 1411 | boffset += 512; |
23f394d5 | 1412 | } |
de789769 JA |
1413 | } |
1414 | ||
0d29de83 JA |
1415 | /* |
1416 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1417 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1418 | */ | |
1419 | struct io_u *get_io_u(struct thread_data *td) | |
1420 | { | |
1421 | struct fio_file *f; | |
1422 | struct io_u *io_u; | |
de789769 | 1423 | int do_scramble = 0; |
002fe734 | 1424 | long ret = 0; |
0d29de83 JA |
1425 | |
1426 | io_u = __get_io_u(td); | |
1427 | if (!io_u) { | |
1428 | dprint(FD_IO, "__get_io_u failed\n"); | |
1429 | return NULL; | |
1430 | } | |
1431 | ||
1432 | if (check_get_verify(td, io_u)) | |
1433 | goto out; | |
1434 | if (check_get_trim(td, io_u)) | |
1435 | goto out; | |
1436 | ||
755200a3 JA |
1437 | /* |
1438 | * from a requeue, io_u already setup | |
1439 | */ | |
1440 | if (io_u->file) | |
77f392bf | 1441 | goto out; |
755200a3 | 1442 | |
429f6675 JA |
1443 | /* |
1444 | * If using an iolog, grab next piece if any available. | |
1445 | */ | |
d72be545 | 1446 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1447 | if (read_iolog_get(td, io_u)) |
1448 | goto err_put; | |
2ba1c290 | 1449 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1450 | ret = -EBUSY; |
2ba1c290 | 1451 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1452 | goto err_put; |
2ba1c290 | 1453 | } |
5ec10eaa | 1454 | |
429f6675 | 1455 | f = io_u->file; |
002fe734 JA |
1456 | if (!f) { |
1457 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1458 | goto err_put; | |
1459 | } | |
1460 | ||
d6aed795 | 1461 | assert(fio_file_open(f)); |
97af62ce | 1462 | |
ff58fced | 1463 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1464 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1465 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1466 | goto err_put; |
2ba1c290 | 1467 | } |
10ba535a | 1468 | |
38dad62d | 1469 | f->last_start = io_u->offset; |
36167d82 | 1470 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1471 | |
fd68418e | 1472 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1473 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1474 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1475 | td->o.min_bs[DDIR_WRITE], |
1476 | io_u->xfer_buflen); | |
bedc9dc2 JA |
1477 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
1478 | !(td->flags & TD_F_COMPRESS)) | |
fd68418e | 1479 | do_scramble = 1; |
d72be545 | 1480 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1481 | populate_verify_io_u(td, io_u); |
1482 | do_scramble = 0; | |
1483 | } | |
fd68418e | 1484 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1485 | /* |
1486 | * Reset the buf_filled parameters so next time if the | |
1487 | * buffer is used for writes it is refilled. | |
1488 | */ | |
cbe8d756 RR |
1489 | io_u->buf_filled_len = 0; |
1490 | } | |
87dc1ab1 | 1491 | } |
10ba535a | 1492 | |
165faf16 JA |
1493 | /* |
1494 | * Set io data pointers. | |
1495 | */ | |
cec6b55d JA |
1496 | io_u->xfer_buf = io_u->buf; |
1497 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1498 | |
6ac7a331 | 1499 | out: |
0d29de83 | 1500 | assert(io_u->file); |
429f6675 | 1501 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1502 | if (!td->o.disable_slat) |
1503 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1504 | if (do_scramble) |
1505 | small_content_scramble(io_u); | |
429f6675 | 1506 | return io_u; |
36167d82 | 1507 | } |
429f6675 | 1508 | err_put: |
2ba1c290 | 1509 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1510 | put_io_u(td, io_u); |
002fe734 | 1511 | return ERR_PTR(ret); |
10ba535a JA |
1512 | } |
1513 | ||
5451792e JA |
1514 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1515 | { | |
8b28bd41 | 1516 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1517 | |
8b28bd41 DM |
1518 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1519 | return; | |
5451792e | 1520 | |
709c8313 RE |
1521 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1522 | io_u->file ? " on file " : "", | |
1523 | io_u->file ? io_u->file->file_name : "", | |
1524 | strerror(io_u->error), | |
1525 | io_ddir_name(io_u->ddir), | |
1526 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1527 | |
1528 | if (!td->error) | |
1529 | td_verror(td, io_u->error, "io_u error"); | |
1530 | } | |
1531 | ||
aba6c951 JA |
1532 | static inline int gtod_reduce(struct thread_data *td) |
1533 | { | |
729fe3af | 1534 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1535 | && td->o.disable_bw; |
aba6c951 JA |
1536 | } |
1537 | ||
c8eeb9df JA |
1538 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1539 | struct io_completion_data *icd, | |
1540 | const enum fio_ddir idx, unsigned int bytes) | |
1541 | { | |
24d23ca7 | 1542 | unsigned long lusec = 0; |
c8eeb9df | 1543 | |
aba6c951 | 1544 | if (!gtod_reduce(td)) |
c8eeb9df JA |
1545 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1546 | ||
1547 | if (!td->o.disable_lat) { | |
1548 | unsigned long tusec; | |
1549 | ||
1550 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1551 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1552 | |
d4afedfd JA |
1553 | if (td->flags & TD_F_PROFILE_OPS) { |
1554 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1555 | ||
1556 | if (ops->io_u_lat) | |
1557 | icd->error = ops->io_u_lat(td, tusec); | |
1558 | } | |
1559 | ||
3e260a46 JA |
1560 | if (td->o.max_latency && tusec > td->o.max_latency) |
1561 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1562 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1563 | if (lat_target_failed(td)) | |
1564 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1565 | } |
c8eeb9df JA |
1566 | } |
1567 | ||
1568 | if (!td->o.disable_clat) { | |
ae588852 | 1569 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1570 | io_u_mark_latency(td, lusec); |
1571 | } | |
1572 | ||
1573 | if (!td->o.disable_bw) | |
1574 | add_bw_sample(td, idx, bytes, &icd->time); | |
1575 | ||
aba6c951 JA |
1576 | if (!gtod_reduce(td)) |
1577 | add_iops_sample(td, idx, bytes, &icd->time); | |
c8eeb9df JA |
1578 | } |
1579 | ||
1b8dbf25 SL |
1580 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1581 | { | |
1ae83d45 JA |
1582 | uint64_t secs, remainder, bps, bytes; |
1583 | ||
1b8dbf25 SL |
1584 | bytes = td->this_io_bytes[ddir]; |
1585 | bps = td->rate_bps[ddir]; | |
1586 | secs = bytes / bps; | |
1587 | remainder = bytes % bps; | |
1588 | return remainder * 1000000 / bps + secs * 1000000; | |
1589 | } | |
1590 | ||
f8b0bd10 | 1591 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1592 | struct io_completion_data *icd) |
10ba535a | 1593 | { |
f8b0bd10 JA |
1594 | struct io_u *io_u = *io_u_ptr; |
1595 | enum fio_ddir ddir = io_u->ddir; | |
1596 | struct fio_file *f = io_u->file; | |
10ba535a | 1597 | |
2ba1c290 JA |
1598 | dprint_io_u(io_u, "io complete"); |
1599 | ||
2ecc1b57 | 1600 | td_io_u_lock(td); |
0c6e7517 | 1601 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1602 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1603 | |
1604 | /* | |
1605 | * Mark IO ok to verify | |
1606 | */ | |
1607 | if (io_u->ipo) { | |
890b6656 JA |
1608 | /* |
1609 | * Remove errored entry from the verification list | |
1610 | */ | |
1611 | if (io_u->error) | |
1612 | unlog_io_piece(td, io_u); | |
1613 | else { | |
1614 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1615 | write_barrier(); | |
1616 | } | |
f9401285 JA |
1617 | } |
1618 | ||
2ecc1b57 | 1619 | td_io_u_unlock(td); |
0c6e7517 | 1620 | |
f8b0bd10 | 1621 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1622 | td->last_was_sync = 1; |
44f29692 JA |
1623 | if (f) { |
1624 | f->first_write = -1ULL; | |
1625 | f->last_write = -1ULL; | |
1626 | } | |
87dc1ab1 JA |
1627 | return; |
1628 | } | |
1629 | ||
1630 | td->last_was_sync = 0; | |
f8b0bd10 | 1631 | td->last_ddir = ddir; |
87dc1ab1 | 1632 | |
f8b0bd10 | 1633 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1634 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1635 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1636 | int ret; |
10ba535a | 1637 | |
f8b0bd10 JA |
1638 | td->io_blocks[ddir]++; |
1639 | td->this_io_blocks[ddir]++; | |
1640 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1641 | |
1642 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1643 | td->this_io_bytes[ddir] += bytes; |
1644 | ||
1645 | if (ddir == DDIR_WRITE && f) { | |
1646 | if (f->first_write == -1ULL || | |
1647 | io_u->offset < f->first_write) | |
1648 | f->first_write = io_u->offset; | |
1649 | if (f->last_write == -1ULL || | |
1650 | ((io_u->offset + bytes) > f->last_write)) | |
1651 | f->last_write = io_u->offset + bytes; | |
44f29692 JA |
1652 | } |
1653 | ||
6b1190fd SL |
1654 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1655 | td->runstate == TD_VERIFYING)) { | |
f8b0bd10 | 1656 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1657 | |
f8b0bd10 JA |
1658 | if (__should_check_rate(td, ddir)) { |
1659 | td->rate_pending_usleep[ddir] = | |
1660 | (usec_for_io(td, ddir) - | |
ba3e4e0c | 1661 | utime_since_now(&td->start)); |
b23b6a2f | 1662 | } |
f8b0bd10 JA |
1663 | if (ddir != DDIR_TRIM && |
1664 | __should_check_rate(td, oddir)) { | |
1665 | td->rate_pending_usleep[oddir] = | |
1666 | (usec_for_io(td, oddir) - | |
ba3e4e0c | 1667 | utime_since_now(&td->start)); |
f8b0bd10 | 1668 | } |
721938ae | 1669 | } |
10ba535a | 1670 | |
f8b0bd10 | 1671 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1672 | |
d7762cf8 | 1673 | if (io_u->end_io) { |
f8b0bd10 JA |
1674 | ret = io_u->end_io(td, io_u_ptr); |
1675 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1676 | if (ret && !icd->error) |
1677 | icd->error = ret; | |
1678 | } | |
ff58fced | 1679 | } else if (io_u->error) { |
10ba535a | 1680 | icd->error = io_u->error; |
5451792e JA |
1681 | io_u_log_error(td, io_u); |
1682 | } | |
8b28bd41 | 1683 | if (icd->error) { |
f8b0bd10 JA |
1684 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1685 | ||
8b28bd41 DM |
1686 | if (!td_non_fatal_error(td, eb, icd->error)) |
1687 | return; | |
f8b0bd10 | 1688 | |
f2bba182 RR |
1689 | /* |
1690 | * If there is a non_fatal error, then add to the error count | |
1691 | * and clear all the errors. | |
1692 | */ | |
1693 | update_error_count(td, icd->error); | |
1694 | td_clear_error(td); | |
1695 | icd->error = 0; | |
f8b0bd10 JA |
1696 | if (io_u) |
1697 | io_u->error = 0; | |
f2bba182 | 1698 | } |
10ba535a JA |
1699 | } |
1700 | ||
9520ebb9 JA |
1701 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1702 | int nr) | |
10ba535a | 1703 | { |
6eaf09d6 | 1704 | int ddir; |
aba6c951 JA |
1705 | |
1706 | if (!gtod_reduce(td)) | |
9520ebb9 | 1707 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1708 | |
3af6ef39 JA |
1709 | icd->nr = nr; |
1710 | ||
10ba535a | 1711 | icd->error = 0; |
6eaf09d6 SL |
1712 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1713 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1714 | } |
1715 | ||
97601024 JA |
1716 | static void ios_completed(struct thread_data *td, |
1717 | struct io_completion_data *icd) | |
36167d82 JA |
1718 | { |
1719 | struct io_u *io_u; | |
1720 | int i; | |
1721 | ||
10ba535a JA |
1722 | for (i = 0; i < icd->nr; i++) { |
1723 | io_u = td->io_ops->event(td, i); | |
1724 | ||
f8b0bd10 | 1725 | io_completed(td, &io_u, icd); |
e8462bd8 | 1726 | |
f8b0bd10 | 1727 | if (io_u) |
e8462bd8 | 1728 | put_io_u(td, io_u); |
10ba535a JA |
1729 | } |
1730 | } | |
97601024 | 1731 | |
e7e6cfb4 JA |
1732 | /* |
1733 | * Complete a single io_u for the sync engines. | |
1734 | */ | |
581e7141 | 1735 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1736 | uint64_t *bytes) |
97601024 JA |
1737 | { |
1738 | struct io_completion_data icd; | |
1739 | ||
9520ebb9 | 1740 | init_icd(td, &icd, 1); |
f8b0bd10 | 1741 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1742 | |
f8b0bd10 | 1743 | if (io_u) |
e8462bd8 | 1744 | put_io_u(td, io_u); |
97601024 | 1745 | |
581e7141 JA |
1746 | if (icd.error) { |
1747 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1748 | return -1; | |
1749 | } | |
97601024 | 1750 | |
581e7141 | 1751 | if (bytes) { |
6eaf09d6 SL |
1752 | int ddir; |
1753 | ||
1754 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1755 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1756 | } |
1757 | ||
1758 | return 0; | |
97601024 JA |
1759 | } |
1760 | ||
e7e6cfb4 JA |
1761 | /* |
1762 | * Called to complete min_events number of io for the async engines. | |
1763 | */ | |
581e7141 | 1764 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1765 | uint64_t *bytes) |
97601024 | 1766 | { |
97601024 | 1767 | struct io_completion_data icd; |
00de55ef | 1768 | struct timespec *tvp = NULL; |
97601024 | 1769 | int ret; |
4d06a338 | 1770 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1771 | |
4950421a | 1772 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1773 | |
4950421a | 1774 | if (!min_evts) |
00de55ef | 1775 | tvp = &ts; |
5fb4b366 RE |
1776 | else if (min_evts > td->cur_depth) |
1777 | min_evts = td->cur_depth; | |
97601024 | 1778 | |
4950421a | 1779 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1780 | if (ret < 0) { |
e1161c32 | 1781 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1782 | return ret; |
1783 | } else if (!ret) | |
1784 | return ret; | |
1785 | ||
9520ebb9 | 1786 | init_icd(td, &icd, ret); |
97601024 | 1787 | ios_completed(td, &icd); |
581e7141 JA |
1788 | if (icd.error) { |
1789 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1790 | return -1; | |
1791 | } | |
97601024 | 1792 | |
581e7141 | 1793 | if (bytes) { |
6eaf09d6 SL |
1794 | int ddir; |
1795 | ||
1796 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1797 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1798 | } |
1799 | ||
1800 | return 0; | |
97601024 | 1801 | } |
7e77dd02 JA |
1802 | |
1803 | /* | |
1804 | * Call when io_u is really queued, to update the submission latency. | |
1805 | */ | |
1806 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1807 | { | |
9520ebb9 JA |
1808 | if (!td->o.disable_slat) { |
1809 | unsigned long slat_time; | |
7e77dd02 | 1810 | |
9520ebb9 | 1811 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1812 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1813 | io_u->offset); | |
9520ebb9 | 1814 | } |
7e77dd02 | 1815 | } |
433afcb4 | 1816 | |
5c94b008 JA |
1817 | /* |
1818 | * See if we should reuse the last seed, if dedupe is enabled | |
1819 | */ | |
1820 | static struct frand_state *get_buf_state(struct thread_data *td) | |
1821 | { | |
1822 | unsigned int v; | |
1823 | unsigned long r; | |
1824 | ||
1825 | if (!td->o.dedupe_percentage) | |
1826 | return &td->buf_state; | |
c0b69b92 JA |
1827 | else if (td->o.dedupe_percentage == 100) |
1828 | return &td->buf_state_prev; | |
5c94b008 JA |
1829 | |
1830 | r = __rand(&td->dedupe_state); | |
1831 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1832 | ||
1833 | if (v <= td->o.dedupe_percentage) | |
1834 | return &td->buf_state_prev; | |
1835 | ||
1836 | return &td->buf_state; | |
1837 | } | |
1838 | ||
1839 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) | |
1840 | { | |
1841 | if (rs == &td->buf_state) | |
1842 | frand_copy(&td->buf_state_prev, rs); | |
1843 | } | |
1844 | ||
cc86c395 JA |
1845 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1846 | unsigned int max_bs) | |
5973cafb | 1847 | { |
ce35b1ec JA |
1848 | if (td->o.buffer_pattern_bytes) |
1849 | fill_buffer_pattern(td, buf, max_bs); | |
1850 | else if (!td->o.zero_buffers) { | |
9c42684e | 1851 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1852 | struct frand_state *rs; |
1066358a | 1853 | unsigned int left = max_bs; |
5c94b008 | 1854 | |
1066358a JA |
1855 | do { |
1856 | rs = get_buf_state(td); | |
9c42684e | 1857 | |
1066358a | 1858 | min_write = min(min_write, left); |
f97a43a1 | 1859 | |
1066358a JA |
1860 | if (perc) { |
1861 | unsigned int seg = min_write; | |
cc86c395 | 1862 | |
1066358a JA |
1863 | seg = min(min_write, td->o.compress_chunk); |
1864 | if (!seg) | |
1865 | seg = min_write; | |
1866 | ||
1867 | fill_random_buf_percentage(rs, buf, perc, seg, | |
1868 | min_write); | |
1869 | } else | |
1870 | fill_random_buf(rs, buf, min_write); | |
1871 | ||
1872 | buf += min_write; | |
1873 | left -= min_write; | |
5c94b008 | 1874 | save_buf_state(td, rs); |
1066358a | 1875 | } while (left); |
9c42684e | 1876 | } else |
cc86c395 JA |
1877 | memset(buf, 0, max_bs); |
1878 | } | |
1879 | ||
1880 | /* | |
1881 | * "randomly" fill the buffer contents | |
1882 | */ | |
1883 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1884 | unsigned int min_write, unsigned int max_bs) | |
1885 | { | |
1886 | io_u->buf_filled_len = 0; | |
1887 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1888 | } |