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 | ||
0412d12e JE |
71 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
72 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
73 | ||
1ae83d45 | 74 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
75 | if (!max_blocks) |
76 | return 0; | |
77 | ||
67778e88 | 78 | return max_blocks; |
2ba1c290 JA |
79 | } |
80 | ||
1ae83d45 JA |
81 | struct rand_off { |
82 | struct flist_head list; | |
83 | uint64_t off; | |
84 | }; | |
85 | ||
e25839d4 | 86 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 87 | enum fio_ddir ddir, uint64_t *b) |
ec4015da | 88 | { |
6f49f8bc | 89 | uint64_t r; |
5e0baa7f | 90 | |
74776733 | 91 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) { |
6f49f8bc JA |
92 | uint64_t lastb; |
93 | ||
94 | lastb = last_block(td, f, ddir); | |
95 | if (!lastb) | |
96 | return 1; | |
97 | ||
d6b72507 | 98 | r = __rand(&td->random_state); |
8055e41d | 99 | |
4b91ee8f | 100 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 101 | |
54a21917 | 102 | *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0)); |
51ede0b1 | 103 | } else { |
8055e41d | 104 | uint64_t off = 0; |
43c63a78 | 105 | |
967d1b63 JA |
106 | assert(fio_file_lfsr(f)); |
107 | ||
6f49f8bc | 108 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 109 | return 1; |
ec4015da | 110 | |
8055e41d JA |
111 | *b = off; |
112 | } | |
0ce8b119 | 113 | |
ec4015da | 114 | /* |
51ede0b1 | 115 | * if we are not maintaining a random map, we are done. |
ec4015da | 116 | */ |
51ede0b1 JA |
117 | if (!file_randommap(td, f)) |
118 | goto ret; | |
43c63a78 JA |
119 | |
120 | /* | |
51ede0b1 | 121 | * calculate map offset and check if it's free |
43c63a78 | 122 | */ |
51ede0b1 JA |
123 | if (random_map_free(f, *b)) |
124 | goto ret; | |
125 | ||
4b91ee8f JA |
126 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
127 | (unsigned long long) *b); | |
51ede0b1 | 128 | |
7ebd796f | 129 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
130 | if (*b == (uint64_t) -1ULL) |
131 | return 1; | |
0ce8b119 JA |
132 | ret: |
133 | return 0; | |
ec4015da JA |
134 | } |
135 | ||
925fee33 JA |
136 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
137 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 138 | uint64_t *b) |
e25839d4 | 139 | { |
9c6f6316 | 140 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
141 | return 0; |
142 | } | |
143 | ||
925fee33 JA |
144 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
145 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 146 | uint64_t *b) |
925fee33 | 147 | { |
9c6f6316 | 148 | *b = pareto_next(&f->zipf); |
925fee33 JA |
149 | return 0; |
150 | } | |
151 | ||
56d9fa4b JA |
152 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
153 | struct fio_file *f, enum fio_ddir ddir, | |
154 | uint64_t *b) | |
155 | { | |
156 | *b = gauss_next(&f->gauss); | |
157 | return 0; | |
158 | } | |
159 | ||
160 | ||
1ae83d45 JA |
161 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
162 | { | |
163 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
164 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
165 | ||
166 | return r1->off - r2->off; | |
167 | } | |
168 | ||
169 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
170 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
171 | { |
172 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
173 | return __get_next_rand_offset(td, f, ddir, b); | |
174 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
175 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
176 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
177 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
178 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
179 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e25839d4 JA |
180 | |
181 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
182 | return 1; | |
183 | } | |
184 | ||
bcd5abfa JA |
185 | /* |
186 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
187 | * specified. | |
188 | */ | |
189 | static inline int should_sort_io(struct thread_data *td) | |
190 | { | |
191 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
192 | return 0; | |
193 | if (!td_random(td)) | |
194 | return 0; | |
195 | if (td->runstate != TD_VERIFYING) | |
196 | return 0; | |
197 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
198 | return 0; | |
199 | ||
200 | return 1; | |
201 | } | |
202 | ||
d9472271 | 203 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
204 | { |
205 | unsigned int v; | |
206 | unsigned long r; | |
207 | ||
d9472271 | 208 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
209 | return 1; |
210 | ||
d6b72507 | 211 | r = __rand(&td->seq_rand_state[ddir]); |
54a21917 | 212 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
211c9b89 | 213 | |
d9472271 | 214 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
215 | } |
216 | ||
1ae83d45 JA |
217 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
218 | enum fio_ddir ddir, uint64_t *b) | |
219 | { | |
220 | struct rand_off *r; | |
221 | int i, ret = 1; | |
222 | ||
bcd5abfa | 223 | if (!should_sort_io(td)) |
1ae83d45 JA |
224 | return get_off_from_method(td, f, ddir, b); |
225 | ||
226 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 227 | fetch: |
9342d5f8 | 228 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
229 | flist_del(&r->list); |
230 | *b = r->off; | |
231 | free(r); | |
232 | return 0; | |
233 | } | |
234 | ||
235 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
236 | r = malloc(sizeof(*r)); | |
237 | ||
238 | ret = get_off_from_method(td, f, ddir, &r->off); | |
239 | if (ret) { | |
240 | free(r); | |
241 | break; | |
242 | } | |
243 | ||
244 | flist_add(&r->list, &td->next_rand_list); | |
245 | } | |
246 | ||
247 | if (ret && !i) | |
248 | return ret; | |
249 | ||
250 | assert(!flist_empty(&td->next_rand_list)); | |
251 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
252 | goto fetch; | |
253 | } | |
254 | ||
38dad62d | 255 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 256 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 257 | { |
c04e4661 DE |
258 | if (!get_next_rand_offset(td, f, ddir, b)) |
259 | return 0; | |
260 | ||
261 | if (td->o.time_based) { | |
33c48814 | 262 | fio_file_reset(td, f); |
c04e4661 DE |
263 | if (!get_next_rand_offset(td, f, ddir, b)) |
264 | return 0; | |
38dad62d JA |
265 | } |
266 | ||
c04e4661 | 267 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 268 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 269 | (unsigned long long) f->real_file_size); |
c04e4661 | 270 | return 1; |
38dad62d JA |
271 | } |
272 | ||
37cf9e3c | 273 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 274 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 275 | { |
ac002339 JA |
276 | struct thread_options *o = &td->o; |
277 | ||
ff58fced JA |
278 | assert(ddir_rw(ddir)); |
279 | ||
f1dfb668 | 280 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
ac002339 | 281 | o->time_based) |
f1dfb668 | 282 | f->last_pos[ddir] = f->last_pos[ddir] - f->io_size; |
c04e4661 | 283 | |
f1dfb668 | 284 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 285 | uint64_t pos; |
059b0802 | 286 | |
f1dfb668 JA |
287 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) |
288 | f->last_pos[ddir] = f->real_file_size; | |
a66da7a2 | 289 | |
f1dfb668 | 290 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
291 | if (pos && o->ddir_seq_add) { |
292 | pos += o->ddir_seq_add; | |
293 | ||
294 | /* | |
295 | * If we reach beyond the end of the file | |
296 | * with holed IO, wrap around to the | |
297 | * beginning again. | |
298 | */ | |
299 | if (pos >= f->real_file_size) | |
300 | pos = f->file_offset; | |
301 | } | |
059b0802 | 302 | |
37cf9e3c | 303 | *offset = pos; |
38dad62d JA |
304 | return 0; |
305 | } | |
306 | ||
307 | return 1; | |
308 | } | |
309 | ||
310 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
311 | enum fio_ddir ddir, int rw_seq, |
312 | unsigned int *is_random) | |
38dad62d JA |
313 | { |
314 | struct fio_file *f = io_u->file; | |
1ae83d45 | 315 | uint64_t b, offset; |
38dad62d JA |
316 | int ret; |
317 | ||
ff58fced JA |
318 | assert(ddir_rw(ddir)); |
319 | ||
37cf9e3c JA |
320 | b = offset = -1ULL; |
321 | ||
38dad62d | 322 | if (rw_seq) { |
211c9b89 | 323 | if (td_random(td)) { |
6aca9b3d | 324 | if (should_do_random(td, ddir)) { |
211c9b89 | 325 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
326 | *is_random = 1; |
327 | } else { | |
328 | *is_random = 0; | |
a9da8ab2 | 329 | io_u_set(io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
330 | ret = get_next_seq_offset(td, f, ddir, &offset); |
331 | if (ret) | |
332 | ret = get_next_rand_block(td, f, ddir, &b); | |
333 | } | |
6aca9b3d JA |
334 | } else { |
335 | *is_random = 0; | |
37cf9e3c | 336 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 337 | } |
38dad62d | 338 | } else { |
a9da8ab2 | 339 | io_u_set(io_u, IO_U_F_BUSY_OK); |
6aca9b3d | 340 | *is_random = 0; |
38dad62d JA |
341 | |
342 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 343 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 344 | if (ret) { |
37cf9e3c | 345 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
346 | *is_random = 0; |
347 | } | |
38dad62d | 348 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
349 | if (f->last_start[ddir] != -1ULL) |
350 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 351 | else |
37cf9e3c | 352 | offset = 0; |
38dad62d JA |
353 | ret = 0; |
354 | } else { | |
355 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
356 | ret = 1; | |
357 | } | |
358 | } | |
6d68b997 | 359 | |
37cf9e3c JA |
360 | if (!ret) { |
361 | if (offset != -1ULL) | |
362 | io_u->offset = offset; | |
363 | else if (b != -1ULL) | |
364 | io_u->offset = b * td->o.ba[ddir]; | |
365 | else { | |
4e0a8fa2 | 366 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
367 | ret = 1; |
368 | } | |
369 | } | |
370 | ||
38dad62d JA |
371 | return ret; |
372 | } | |
373 | ||
10ba535a JA |
374 | /* |
375 | * For random io, generate a random new block and see if it's used. Repeat | |
376 | * until we find a free one. For sequential io, just return the end of | |
377 | * the last io issued. | |
378 | */ | |
6aca9b3d JA |
379 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
380 | unsigned int *is_random) | |
10ba535a | 381 | { |
9bf2061e | 382 | struct fio_file *f = io_u->file; |
4ba66134 | 383 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 384 | int rw_seq_hit = 0; |
10ba535a | 385 | |
ff58fced JA |
386 | assert(ddir_rw(ddir)); |
387 | ||
38dad62d JA |
388 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
389 | rw_seq_hit = 1; | |
5736c10d | 390 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 391 | } |
211097b2 | 392 | |
6aca9b3d | 393 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 394 | return 1; |
10ba535a | 395 | |
009bd847 JA |
396 | if (io_u->offset >= f->io_size) { |
397 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
398 | (unsigned long long) io_u->offset, |
399 | (unsigned long long) f->io_size); | |
009bd847 JA |
400 | return 1; |
401 | } | |
402 | ||
403 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
404 | if (io_u->offset >= f->real_file_size) { |
405 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
406 | (unsigned long long) io_u->offset, |
407 | (unsigned long long) f->real_file_size); | |
10ba535a | 408 | return 1; |
2ba1c290 | 409 | } |
10ba535a JA |
410 | |
411 | return 0; | |
412 | } | |
413 | ||
6aca9b3d JA |
414 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
415 | unsigned int *is_random) | |
15dc1934 | 416 | { |
d72be545 JA |
417 | if (td->flags & TD_F_PROFILE_OPS) { |
418 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 419 | |
d72be545 | 420 | if (ops->fill_io_u_off) |
6aca9b3d | 421 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 422 | } |
15dc1934 | 423 | |
6aca9b3d | 424 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
425 | } |
426 | ||
79944128 JA |
427 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
428 | unsigned int buflen) | |
429 | { | |
430 | struct fio_file *f = io_u->file; | |
431 | ||
bedc9dc2 | 432 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
433 | } |
434 | ||
6aca9b3d JA |
435 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
436 | unsigned int is_random) | |
10ba535a | 437 | { |
6aca9b3d | 438 | int ddir = io_u->ddir; |
24d23ca7 | 439 | unsigned int buflen = 0; |
f3059de1 | 440 | unsigned int minbs, maxbs; |
54a21917 | 441 | unsigned long r; |
10ba535a | 442 | |
9ee1c647 | 443 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
444 | |
445 | if (td->o.bs_is_seq_rand) | |
446 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 447 | |
f3059de1 JA |
448 | minbs = td->o.min_bs[ddir]; |
449 | maxbs = td->o.max_bs[ddir]; | |
450 | ||
79944128 JA |
451 | if (minbs == maxbs) |
452 | return minbs; | |
453 | ||
52c58027 JA |
454 | /* |
455 | * If we can't satisfy the min block size from here, then fail | |
456 | */ | |
457 | if (!io_u_fits(td, io_u, minbs)) | |
458 | return 0; | |
459 | ||
79944128 | 460 | do { |
d6b72507 | 461 | r = __rand(&td->bsrange_state); |
4c07ad86 | 462 | |
720e84ad | 463 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 464 | buflen = 1 + (unsigned int) ((double) maxbs * |
54a21917 | 465 | (r / (FRAND_MAX + 1.0))); |
f3059de1 JA |
466 | if (buflen < minbs) |
467 | buflen = minbs; | |
5ec10eaa | 468 | } else { |
564ca972 JA |
469 | long perc = 0; |
470 | unsigned int i; | |
471 | ||
720e84ad JA |
472 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
473 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
474 | |
475 | buflen = bsp->bs; | |
476 | perc += bsp->perc; | |
54a21917 | 477 | if ((r <= ((FRAND_MAX / 100L) * perc)) && |
79944128 | 478 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
479 | break; |
480 | } | |
481 | } | |
79944128 | 482 | |
a9f70b1f JB |
483 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
484 | buflen = (buflen + td->o.verify_interval - 1) & | |
485 | ~(td->o.verify_interval - 1); | |
486 | ||
f3059de1 | 487 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
7c306bb1 | 488 | buflen &= ~(minbs - 1); |
10ba535a | 489 | |
79944128 | 490 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 491 | |
10ba535a JA |
492 | return buflen; |
493 | } | |
494 | ||
6aca9b3d JA |
495 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
496 | unsigned int is_random) | |
15dc1934 | 497 | { |
d72be545 JA |
498 | if (td->flags & TD_F_PROFILE_OPS) { |
499 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 500 | |
d72be545 | 501 | if (ops->fill_io_u_size) |
6aca9b3d | 502 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 503 | } |
15dc1934 | 504 | |
6aca9b3d | 505 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
506 | } |
507 | ||
afe24a5a JA |
508 | static void set_rwmix_bytes(struct thread_data *td) |
509 | { | |
afe24a5a JA |
510 | unsigned int diff; |
511 | ||
512 | /* | |
513 | * we do time or byte based switch. this is needed because | |
514 | * buffered writes may issue a lot quicker than they complete, | |
515 | * whereas reads do not. | |
516 | */ | |
e47f799f | 517 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 518 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
519 | } |
520 | ||
521 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
522 | { | |
523 | unsigned int v; | |
1294c3ec | 524 | unsigned long r; |
e47f799f | 525 | |
d6b72507 | 526 | r = __rand(&td->rwmix_state); |
54a21917 | 527 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
4c07ad86 | 528 | |
04c540d9 | 529 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
530 | return DDIR_READ; |
531 | ||
532 | return DDIR_WRITE; | |
afe24a5a JA |
533 | } |
534 | ||
002e7183 JA |
535 | void io_u_quiesce(struct thread_data *td) |
536 | { | |
537 | /* | |
538 | * We are going to sleep, ensure that we flush anything pending as | |
539 | * not to skew our latency numbers. | |
540 | * | |
541 | * Changed to only monitor 'in flight' requests here instead of the | |
542 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
543 | * io's that have been actually submitted to an async engine, | |
544 | * and cur_depth is meaningless for sync engines. | |
545 | */ | |
9cc80b6d JA |
546 | if (td->io_u_queued || td->cur_depth) { |
547 | int fio_unused ret; | |
548 | ||
549 | ret = td_io_commit(td); | |
550 | } | |
551 | ||
002e7183 JA |
552 | while (td->io_u_in_flight) { |
553 | int fio_unused ret; | |
554 | ||
55312f9f | 555 | ret = io_u_queued_complete(td, 1); |
002e7183 JA |
556 | } |
557 | } | |
558 | ||
581e7141 JA |
559 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
560 | { | |
561 | enum fio_ddir odir = ddir ^ 1; | |
581e7141 JA |
562 | long usec; |
563 | ||
ff58fced JA |
564 | assert(ddir_rw(ddir)); |
565 | ||
315fcfec | 566 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
567 | return ddir; |
568 | ||
569 | /* | |
570 | * We have too much pending sleep in this direction. See if we | |
571 | * should switch. | |
572 | */ | |
315fcfec | 573 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
574 | /* |
575 | * Other direction does not have too much pending, switch | |
576 | */ | |
577 | if (td->rate_pending_usleep[odir] < 100000) | |
578 | return odir; | |
579 | ||
580 | /* | |
581 | * Both directions have pending sleep. Sleep the minimum time | |
582 | * and deduct from both. | |
583 | */ | |
584 | if (td->rate_pending_usleep[ddir] <= | |
585 | td->rate_pending_usleep[odir]) { | |
586 | usec = td->rate_pending_usleep[ddir]; | |
587 | } else { | |
588 | usec = td->rate_pending_usleep[odir]; | |
589 | ddir = odir; | |
590 | } | |
591 | } else | |
592 | usec = td->rate_pending_usleep[ddir]; | |
593 | ||
a9da8ab2 JA |
594 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
595 | io_u_quiesce(td); | |
78c1eda5 | 596 | |
9cc80b6d | 597 | usec = usec_sleep(td, usec); |
581e7141 JA |
598 | |
599 | td->rate_pending_usleep[ddir] -= usec; | |
600 | ||
601 | odir = ddir ^ 1; | |
602 | if (td_rw(td) && __should_check_rate(td, odir)) | |
603 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 604 | |
581e7141 JA |
605 | return ddir; |
606 | } | |
607 | ||
10ba535a JA |
608 | /* |
609 | * Return the data direction for the next io_u. If the job is a | |
610 | * mixed read/write workload, check the rwmix cycle and switch if | |
611 | * necessary. | |
612 | */ | |
1e97cce9 | 613 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 614 | { |
581e7141 JA |
615 | enum fio_ddir ddir; |
616 | ||
5f9099ea JA |
617 | /* |
618 | * see if it's time to fsync | |
619 | */ | |
620 | if (td->o.fsync_blocks && | |
621 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
622 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
623 | return DDIR_SYNC; | |
624 | ||
625 | /* | |
626 | * see if it's time to fdatasync | |
627 | */ | |
628 | if (td->o.fdatasync_blocks && | |
629 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
630 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
631 | return DDIR_DATASYNC; | |
632 | ||
44f29692 JA |
633 | /* |
634 | * see if it's time to sync_file_range | |
635 | */ | |
636 | if (td->sync_file_range_nr && | |
637 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
638 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
639 | return DDIR_SYNC_FILE_RANGE; | |
640 | ||
10ba535a | 641 | if (td_rw(td)) { |
10ba535a JA |
642 | /* |
643 | * Check if it's time to seed a new data direction. | |
644 | */ | |
e4928662 | 645 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
646 | /* |
647 | * Put a top limit on how many bytes we do for | |
648 | * one data direction, to avoid overflowing the | |
649 | * ranges too much | |
650 | */ | |
651 | ddir = get_rand_ddir(td); | |
e47f799f JA |
652 | |
653 | if (ddir != td->rwmix_ddir) | |
654 | set_rwmix_bytes(td); | |
655 | ||
656 | td->rwmix_ddir = ddir; | |
10ba535a | 657 | } |
581e7141 | 658 | ddir = td->rwmix_ddir; |
10ba535a | 659 | } else if (td_read(td)) |
581e7141 | 660 | ddir = DDIR_READ; |
6eaf09d6 | 661 | else if (td_write(td)) |
581e7141 | 662 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
663 | else |
664 | ddir = DDIR_TRIM; | |
581e7141 JA |
665 | |
666 | td->rwmix_ddir = rate_ddir(td, ddir); | |
667 | return td->rwmix_ddir; | |
10ba535a JA |
668 | } |
669 | ||
1ef2b6be JA |
670 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
671 | { | |
0e4dd95c DE |
672 | enum fio_ddir ddir = get_rw_ddir(td); |
673 | ||
82a90686 | 674 | if (td_trimwrite(td)) { |
0e4dd95c DE |
675 | struct fio_file *f = io_u->file; |
676 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
677 | ddir = DDIR_TRIM; | |
678 | else | |
679 | ddir = DDIR_WRITE; | |
680 | } | |
681 | ||
682 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be JA |
683 | |
684 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
685 | td->o.barrier_blocks && | |
686 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
687 | td->io_issues[DDIR_WRITE]) | |
a9da8ab2 | 688 | io_u_set(io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
689 | } |
690 | ||
e8462bd8 | 691 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 692 | { |
71b84caa | 693 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
694 | |
695 | if (ret) | |
696 | td_verror(td, ret, "file close"); | |
697 | } | |
698 | ||
10ba535a JA |
699 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
700 | { | |
a9da8ab2 JA |
701 | if (td->parent) |
702 | td = td->parent; | |
703 | ||
e8462bd8 JA |
704 | td_io_u_lock(td); |
705 | ||
f8b0bd10 | 706 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 707 | put_file_log(td, io_u->file); |
f8b0bd10 | 708 | |
10ba535a | 709 | io_u->file = NULL; |
a9da8ab2 | 710 | io_u_set(io_u, IO_U_F_FREE); |
d7ee2a7d | 711 | |
a9da8ab2 | 712 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 713 | td->cur_depth--; |
a9da8ab2 JA |
714 | assert(!(td->flags & TD_F_CHILD)); |
715 | } | |
2ae0b204 | 716 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
717 | td_io_u_unlock(td); |
718 | td_io_u_free_notify(td); | |
10ba535a JA |
719 | } |
720 | ||
f2bba182 RR |
721 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
722 | { | |
a9da8ab2 | 723 | io_u_clear(io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
724 | put_io_u(td, io_u); |
725 | } | |
726 | ||
755200a3 JA |
727 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
728 | { | |
729 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 730 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 731 | |
465221b0 JA |
732 | dprint(FD_IO, "requeue %p\n", __io_u); |
733 | ||
a9da8ab2 JA |
734 | if (td->parent) |
735 | td = td->parent; | |
736 | ||
e8462bd8 JA |
737 | td_io_u_lock(td); |
738 | ||
a9da8ab2 | 739 | io_u_set(__io_u, IO_U_F_FREE); |
bcd5abfa JA |
740 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
741 | td->io_issues[ddir]--; | |
5ec10eaa | 742 | |
a9da8ab2 JA |
743 | io_u_clear(__io_u, IO_U_F_FLIGHT); |
744 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { | |
0c41214f | 745 | td->cur_depth--; |
a9da8ab2 JA |
746 | assert(!(td->flags & TD_F_CHILD)); |
747 | } | |
2ae0b204 JA |
748 | |
749 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 750 | td_io_u_unlock(td); |
a9da8ab2 | 751 | td_io_u_free_notify(td); |
755200a3 JA |
752 | *io_u = NULL; |
753 | } | |
754 | ||
9bf2061e | 755 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 756 | { |
6aca9b3d JA |
757 | unsigned int is_random; |
758 | ||
b4c5e1ac JA |
759 | if (td->io_ops->flags & FIO_NOIO) |
760 | goto out; | |
761 | ||
1ef2b6be | 762 | set_rw_ddir(td, io_u); |
5f9099ea | 763 | |
87dc1ab1 | 764 | /* |
ff58fced | 765 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 766 | */ |
ff58fced | 767 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 768 | goto out; |
a00735e6 | 769 | |
48f5abd3 JA |
770 | /* |
771 | * See if it's time to switch to a new zone | |
772 | */ | |
13af05ae | 773 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
774 | struct fio_file *f = io_u->file; |
775 | ||
48f5abd3 | 776 | td->zone_bytes = 0; |
4c8be5b1 JA |
777 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
778 | ||
779 | /* | |
780 | * Wrap from the beginning, if we exceed the file size | |
781 | */ | |
782 | if (f->file_offset >= f->real_file_size) | |
783 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 784 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
785 | td->io_skip_bytes += td->o.zone_skip; |
786 | } | |
787 | ||
10ba535a | 788 | /* |
c685b5b2 JA |
789 | * No log, let the seq/rand engine retrieve the next buflen and |
790 | * position. | |
10ba535a | 791 | */ |
6aca9b3d | 792 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 793 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 794 | return 1; |
2ba1c290 | 795 | } |
10ba535a | 796 | |
6aca9b3d | 797 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
798 | if (!io_u->buflen) { |
799 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 800 | return 1; |
2ba1c290 | 801 | } |
bca4ed4d | 802 | |
2ba1c290 JA |
803 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
804 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
805 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
806 | (unsigned long long) io_u->offset, io_u->buflen, | |
807 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 808 | return 1; |
2ba1c290 | 809 | } |
6a5e6884 | 810 | |
bca4ed4d JA |
811 | /* |
812 | * mark entry before potentially trimming io_u | |
813 | */ | |
303032ae | 814 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 815 | mark_random_map(td, io_u); |
bca4ed4d | 816 | |
c38e9468 | 817 | out: |
2ba1c290 | 818 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 819 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 820 | return 0; |
10ba535a JA |
821 | } |
822 | ||
838bc709 JA |
823 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
824 | { | |
2b13e716 | 825 | int idx = 0; |
838bc709 JA |
826 | |
827 | switch (nr) { | |
828 | default: | |
2b13e716 | 829 | idx = 6; |
838bc709 JA |
830 | break; |
831 | case 33 ... 64: | |
2b13e716 | 832 | idx = 5; |
838bc709 JA |
833 | break; |
834 | case 17 ... 32: | |
2b13e716 | 835 | idx = 4; |
838bc709 JA |
836 | break; |
837 | case 9 ... 16: | |
2b13e716 | 838 | idx = 3; |
838bc709 JA |
839 | break; |
840 | case 5 ... 8: | |
2b13e716 | 841 | idx = 2; |
838bc709 JA |
842 | break; |
843 | case 1 ... 4: | |
2b13e716 | 844 | idx = 1; |
838bc709 JA |
845 | case 0: |
846 | break; | |
847 | } | |
848 | ||
2b13e716 | 849 | map[idx]++; |
838bc709 JA |
850 | } |
851 | ||
852 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
853 | { | |
854 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
855 | td->ts.total_submit++; | |
856 | } | |
857 | ||
858 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
859 | { | |
860 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
861 | td->ts.total_complete++; | |
862 | } | |
863 | ||
d8005759 | 864 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 865 | { |
2b13e716 | 866 | int idx = 0; |
71619dc2 JA |
867 | |
868 | switch (td->cur_depth) { | |
869 | default: | |
2b13e716 | 870 | idx = 6; |
a783e61a | 871 | break; |
71619dc2 | 872 | case 32 ... 63: |
2b13e716 | 873 | idx = 5; |
a783e61a | 874 | break; |
71619dc2 | 875 | case 16 ... 31: |
2b13e716 | 876 | idx = 4; |
a783e61a | 877 | break; |
71619dc2 | 878 | case 8 ... 15: |
2b13e716 | 879 | idx = 3; |
a783e61a | 880 | break; |
71619dc2 | 881 | case 4 ... 7: |
2b13e716 | 882 | idx = 2; |
a783e61a | 883 | break; |
71619dc2 | 884 | case 2 ... 3: |
2b13e716 | 885 | idx = 1; |
71619dc2 JA |
886 | case 1: |
887 | break; | |
888 | } | |
889 | ||
2b13e716 | 890 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
891 | } |
892 | ||
04a0feae JA |
893 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
894 | { | |
2b13e716 | 895 | int idx = 0; |
04a0feae JA |
896 | |
897 | assert(usec < 1000); | |
898 | ||
899 | switch (usec) { | |
900 | case 750 ... 999: | |
2b13e716 | 901 | idx = 9; |
04a0feae JA |
902 | break; |
903 | case 500 ... 749: | |
2b13e716 | 904 | idx = 8; |
04a0feae JA |
905 | break; |
906 | case 250 ... 499: | |
2b13e716 | 907 | idx = 7; |
04a0feae JA |
908 | break; |
909 | case 100 ... 249: | |
2b13e716 | 910 | idx = 6; |
04a0feae JA |
911 | break; |
912 | case 50 ... 99: | |
2b13e716 | 913 | idx = 5; |
04a0feae JA |
914 | break; |
915 | case 20 ... 49: | |
2b13e716 | 916 | idx = 4; |
04a0feae JA |
917 | break; |
918 | case 10 ... 19: | |
2b13e716 | 919 | idx = 3; |
04a0feae JA |
920 | break; |
921 | case 4 ... 9: | |
2b13e716 | 922 | idx = 2; |
04a0feae JA |
923 | break; |
924 | case 2 ... 3: | |
2b13e716 | 925 | idx = 1; |
04a0feae JA |
926 | case 0 ... 1: |
927 | break; | |
928 | } | |
929 | ||
2b13e716 JA |
930 | assert(idx < FIO_IO_U_LAT_U_NR); |
931 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
932 | } |
933 | ||
934 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 935 | { |
2b13e716 | 936 | int idx = 0; |
ec118304 JA |
937 | |
938 | switch (msec) { | |
939 | default: | |
2b13e716 | 940 | idx = 11; |
04a0feae | 941 | break; |
8abdce66 | 942 | case 1000 ... 1999: |
2b13e716 | 943 | idx = 10; |
04a0feae | 944 | break; |
8abdce66 | 945 | case 750 ... 999: |
2b13e716 | 946 | idx = 9; |
04a0feae | 947 | break; |
8abdce66 | 948 | case 500 ... 749: |
2b13e716 | 949 | idx = 8; |
04a0feae | 950 | break; |
8abdce66 | 951 | case 250 ... 499: |
2b13e716 | 952 | idx = 7; |
04a0feae | 953 | break; |
8abdce66 | 954 | case 100 ... 249: |
2b13e716 | 955 | idx = 6; |
04a0feae | 956 | break; |
8abdce66 | 957 | case 50 ... 99: |
2b13e716 | 958 | idx = 5; |
04a0feae | 959 | break; |
8abdce66 | 960 | case 20 ... 49: |
2b13e716 | 961 | idx = 4; |
04a0feae | 962 | break; |
8abdce66 | 963 | case 10 ... 19: |
2b13e716 | 964 | idx = 3; |
04a0feae | 965 | break; |
8abdce66 | 966 | case 4 ... 9: |
2b13e716 | 967 | idx = 2; |
04a0feae | 968 | break; |
ec118304 | 969 | case 2 ... 3: |
2b13e716 | 970 | idx = 1; |
ec118304 JA |
971 | case 0 ... 1: |
972 | break; | |
973 | } | |
974 | ||
2b13e716 JA |
975 | assert(idx < FIO_IO_U_LAT_M_NR); |
976 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
977 | } |
978 | ||
979 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
980 | { | |
981 | if (usec < 1000) | |
982 | io_u_mark_lat_usec(td, usec); | |
983 | else | |
984 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
985 | } |
986 | ||
0aabe160 JA |
987 | /* |
988 | * Get next file to service by choosing one at random | |
989 | */ | |
2cc52930 JA |
990 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
991 | enum fio_file_flags goodf, | |
d6aed795 | 992 | enum fio_file_flags badf) |
0aabe160 | 993 | { |
0aabe160 | 994 | struct fio_file *f; |
1c178180 | 995 | int fno; |
0aabe160 JA |
996 | |
997 | do { | |
87b10676 | 998 | int opened = 0; |
1294c3ec | 999 | unsigned long r; |
4c07ad86 | 1000 | |
d6b72507 | 1001 | r = __rand(&td->next_file_state); |
54a21917 | 1002 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 1003 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 1004 | |
126d65c6 | 1005 | f = td->files[fno]; |
d6aed795 | 1006 | if (fio_file_done(f)) |
059e63c0 | 1007 | continue; |
1c178180 | 1008 | |
d6aed795 | 1009 | if (!fio_file_open(f)) { |
87b10676 JA |
1010 | int err; |
1011 | ||
002fe734 JA |
1012 | if (td->nr_open_files >= td->o.open_files) |
1013 | return ERR_PTR(-EBUSY); | |
1014 | ||
87b10676 JA |
1015 | err = td_io_open_file(td, f); |
1016 | if (err) | |
1017 | continue; | |
1018 | opened = 1; | |
1019 | } | |
1020 | ||
2ba1c290 JA |
1021 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1022 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1023 | return f; |
2ba1c290 | 1024 | } |
87b10676 JA |
1025 | if (opened) |
1026 | td_io_close_file(td, f); | |
0aabe160 JA |
1027 | } while (1); |
1028 | } | |
1029 | ||
1030 | /* | |
1031 | * Get next file to service by doing round robin between all available ones | |
1032 | */ | |
1c178180 JA |
1033 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1034 | int badf) | |
3d7c391d JA |
1035 | { |
1036 | unsigned int old_next_file = td->next_file; | |
1037 | struct fio_file *f; | |
1038 | ||
1039 | do { | |
87b10676 JA |
1040 | int opened = 0; |
1041 | ||
126d65c6 | 1042 | f = td->files[td->next_file]; |
3d7c391d JA |
1043 | |
1044 | td->next_file++; | |
2dc1bbeb | 1045 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1046 | td->next_file = 0; |
1047 | ||
87b10676 | 1048 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1049 | if (fio_file_done(f)) { |
d5ed68ea | 1050 | f = NULL; |
059e63c0 | 1051 | continue; |
d5ed68ea | 1052 | } |
059e63c0 | 1053 | |
d6aed795 | 1054 | if (!fio_file_open(f)) { |
87b10676 JA |
1055 | int err; |
1056 | ||
002fe734 JA |
1057 | if (td->nr_open_files >= td->o.open_files) |
1058 | return ERR_PTR(-EBUSY); | |
1059 | ||
87b10676 | 1060 | err = td_io_open_file(td, f); |
b5696bfc JA |
1061 | if (err) { |
1062 | dprint(FD_FILE, "error %d on open of %s\n", | |
1063 | err, f->file_name); | |
87c27b45 | 1064 | f = NULL; |
87b10676 | 1065 | continue; |
b5696bfc | 1066 | } |
87b10676 JA |
1067 | opened = 1; |
1068 | } | |
1069 | ||
0b9d69ec JA |
1070 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1071 | f->flags); | |
1c178180 | 1072 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1073 | break; |
1074 | ||
87b10676 JA |
1075 | if (opened) |
1076 | td_io_close_file(td, f); | |
1077 | ||
3d7c391d JA |
1078 | f = NULL; |
1079 | } while (td->next_file != old_next_file); | |
1080 | ||
2ba1c290 | 1081 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1082 | return f; |
1083 | } | |
1084 | ||
7eb36574 | 1085 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1086 | { |
1907dbc6 JA |
1087 | struct fio_file *f; |
1088 | ||
2dc1bbeb | 1089 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1090 | |
b5696bfc | 1091 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1092 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1093 | " nr_files=%d\n", td->nr_open_files, | |
1094 | td->nr_done_files, | |
1095 | td->o.nr_files); | |
bdb4e2e9 | 1096 | return NULL; |
2ba1c290 | 1097 | } |
bdb4e2e9 | 1098 | |
1907dbc6 | 1099 | f = td->file_service_file; |
d6aed795 | 1100 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1101 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1102 | goto out; | |
1103 | if (td->file_service_left--) | |
1104 | goto out; | |
1105 | } | |
1907dbc6 | 1106 | |
a086c257 JA |
1107 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1108 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1109 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1110 | else |
d6aed795 | 1111 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1112 | |
002fe734 JA |
1113 | if (IS_ERR(f)) |
1114 | return f; | |
1115 | ||
1907dbc6 JA |
1116 | td->file_service_file = f; |
1117 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1118 | out: |
0dac421f JA |
1119 | if (f) |
1120 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1121 | else | |
1122 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1123 | return f; |
bdb4e2e9 JA |
1124 | } |
1125 | ||
7eb36574 JA |
1126 | static struct fio_file *get_next_file(struct thread_data *td) |
1127 | { | |
372d8962 | 1128 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1129 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1130 | |
d72be545 JA |
1131 | if (ops->get_next_file) |
1132 | return ops->get_next_file(td); | |
1133 | } | |
7eb36574 JA |
1134 | |
1135 | return __get_next_file(td); | |
1136 | } | |
1137 | ||
002fe734 | 1138 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1139 | { |
1140 | struct fio_file *f; | |
1141 | ||
1142 | do { | |
1143 | f = get_next_file(td); | |
002fe734 JA |
1144 | if (IS_ERR_OR_NULL(f)) |
1145 | return PTR_ERR(f); | |
429f6675 | 1146 | |
429f6675 JA |
1147 | io_u->file = f; |
1148 | get_file(f); | |
1149 | ||
1150 | if (!fill_io_u(td, io_u)) | |
1151 | break; | |
1152 | ||
b5696bfc | 1153 | put_file_log(td, f); |
429f6675 | 1154 | td_io_close_file(td, f); |
b5696bfc | 1155 | io_u->file = NULL; |
d6aed795 | 1156 | fio_file_set_done(f); |
429f6675 | 1157 | td->nr_done_files++; |
0b9d69ec JA |
1158 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1159 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1160 | } while (1); |
1161 | ||
1162 | return 0; | |
1163 | } | |
1164 | ||
3e260a46 JA |
1165 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1166 | unsigned long tusec, unsigned long max_usec) | |
1167 | { | |
1168 | if (!td->error) | |
1169 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1170 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1171 | icd->error = ETIMEDOUT; | |
1172 | } | |
1173 | ||
1174 | static void lat_new_cycle(struct thread_data *td) | |
1175 | { | |
1176 | fio_gettime(&td->latency_ts, NULL); | |
1177 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1178 | td->latency_failed = 0; | |
1179 | } | |
1180 | ||
1181 | /* | |
1182 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1183 | * are at QD=1, then it's time to give up. | |
1184 | */ | |
1185 | static int __lat_target_failed(struct thread_data *td) | |
1186 | { | |
1187 | if (td->latency_qd == 1) | |
1188 | return 1; | |
1189 | ||
1190 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1191 | |
1192 | if (td->latency_qd == td->latency_qd_low) | |
1193 | td->latency_qd_low--; | |
1194 | ||
3e260a46 JA |
1195 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1196 | ||
1197 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1198 | ||
1199 | /* | |
1200 | * When we ramp QD down, quiesce existing IO to prevent | |
1201 | * a storm of ramp downs due to pending higher depth. | |
1202 | */ | |
1203 | io_u_quiesce(td); | |
1204 | lat_new_cycle(td); | |
1205 | return 0; | |
1206 | } | |
1207 | ||
1208 | static int lat_target_failed(struct thread_data *td) | |
1209 | { | |
1210 | if (td->o.latency_percentile.u.f == 100.0) | |
1211 | return __lat_target_failed(td); | |
1212 | ||
1213 | td->latency_failed++; | |
1214 | return 0; | |
1215 | } | |
1216 | ||
1217 | void lat_target_init(struct thread_data *td) | |
1218 | { | |
6bb58215 JA |
1219 | td->latency_end_run = 0; |
1220 | ||
3e260a46 JA |
1221 | if (td->o.latency_target) { |
1222 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1223 | fio_gettime(&td->latency_ts, NULL); | |
1224 | td->latency_qd = 1; | |
1225 | td->latency_qd_high = td->o.iodepth; | |
1226 | td->latency_qd_low = 1; | |
1227 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1228 | } else | |
1229 | td->latency_qd = td->o.iodepth; | |
1230 | } | |
1231 | ||
6bb58215 JA |
1232 | void lat_target_reset(struct thread_data *td) |
1233 | { | |
1234 | if (!td->latency_end_run) | |
1235 | lat_target_init(td); | |
1236 | } | |
1237 | ||
3e260a46 JA |
1238 | static void lat_target_success(struct thread_data *td) |
1239 | { | |
1240 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1241 | struct thread_options *o = &td->o; |
3e260a46 JA |
1242 | |
1243 | td->latency_qd_low = td->latency_qd; | |
1244 | ||
1245 | /* | |
1246 | * If we haven't failed yet, we double up to a failing value instead | |
1247 | * of bisecting from highest possible queue depth. If we have set | |
1248 | * a limit other than td->o.iodepth, bisect between that. | |
1249 | */ | |
6bb58215 | 1250 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1251 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1252 | else | |
1253 | td->latency_qd *= 2; | |
1254 | ||
6bb58215 JA |
1255 | if (td->latency_qd > o->iodepth) |
1256 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1257 | |
1258 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1259 | |
3e260a46 | 1260 | /* |
6bb58215 JA |
1261 | * Same as last one, we are done. Let it run a latency cycle, so |
1262 | * we get only the results from the targeted depth. | |
3e260a46 | 1263 | */ |
6bb58215 JA |
1264 | if (td->latency_qd == qd) { |
1265 | if (td->latency_end_run) { | |
1266 | dprint(FD_RATE, "We are done\n"); | |
1267 | td->done = 1; | |
1268 | } else { | |
1269 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1270 | io_u_quiesce(td); | |
1271 | td->latency_end_run = 1; | |
1272 | reset_all_stats(td); | |
1273 | reset_io_stats(td); | |
1274 | } | |
1275 | } | |
3e260a46 JA |
1276 | |
1277 | lat_new_cycle(td); | |
1278 | } | |
1279 | ||
1280 | /* | |
1281 | * Check if we can bump the queue depth | |
1282 | */ | |
1283 | void lat_target_check(struct thread_data *td) | |
1284 | { | |
1285 | uint64_t usec_window; | |
1286 | uint64_t ios; | |
1287 | double success_ios; | |
1288 | ||
1289 | usec_window = utime_since_now(&td->latency_ts); | |
1290 | if (usec_window < td->o.latency_window) | |
1291 | return; | |
1292 | ||
1293 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1294 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1295 | success_ios *= 100.0; | |
1296 | ||
1297 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1298 | ||
1299 | if (success_ios >= td->o.latency_percentile.u.f) | |
1300 | lat_target_success(td); | |
1301 | else | |
1302 | __lat_target_failed(td); | |
1303 | } | |
1304 | ||
1305 | /* | |
1306 | * If latency target is enabled, we might be ramping up or down and not | |
1307 | * using the full queue depth available. | |
1308 | */ | |
effd6ff0 | 1309 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1310 | { |
1311 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1312 | ||
1313 | if (qempty) | |
1314 | return 1; | |
1315 | if (!td->o.latency_target) | |
1316 | return 0; | |
1317 | ||
1318 | return td->cur_depth >= td->latency_qd; | |
1319 | } | |
429f6675 | 1320 | |
10ba535a JA |
1321 | struct io_u *__get_io_u(struct thread_data *td) |
1322 | { | |
0cae66f6 | 1323 | struct io_u *io_u = NULL; |
10ba535a | 1324 | |
ca09be4b JA |
1325 | if (td->stop_io) |
1326 | return NULL; | |
1327 | ||
e8462bd8 JA |
1328 | td_io_u_lock(td); |
1329 | ||
1330 | again: | |
2ae0b204 JA |
1331 | if (!io_u_rempty(&td->io_u_requeues)) |
1332 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1333 | else if (!queue_full(td)) { |
2ae0b204 | 1334 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1335 | |
225ba9e3 | 1336 | io_u->file = NULL; |
6040dabc | 1337 | io_u->buflen = 0; |
10ba535a | 1338 | io_u->resid = 0; |
d7762cf8 | 1339 | io_u->end_io = NULL; |
755200a3 JA |
1340 | } |
1341 | ||
1342 | if (io_u) { | |
0c6e7517 | 1343 | assert(io_u->flags & IO_U_F_FREE); |
a9da8ab2 | 1344 | io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1345 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1346 | IO_U_F_VER_LIST); | |
0c6e7517 | 1347 | |
755200a3 | 1348 | io_u->error = 0; |
bcd5abfa | 1349 | io_u->acct_ddir = -1; |
10ba535a | 1350 | td->cur_depth++; |
a9da8ab2 JA |
1351 | assert(!(td->flags & TD_F_CHILD)); |
1352 | io_u_set(io_u, IO_U_F_IN_CUR_DEPTH); | |
f9401285 | 1353 | io_u->ipo = NULL; |
a9da8ab2 | 1354 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1355 | /* |
1356 | * We ran out, wait for async verify threads to finish and | |
1357 | * return one | |
1358 | */ | |
a9da8ab2 JA |
1359 | assert(!(td->flags & TD_F_CHILD)); |
1360 | assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock)); | |
1dec3e07 | 1361 | goto again; |
10ba535a JA |
1362 | } |
1363 | ||
e8462bd8 | 1364 | td_io_u_unlock(td); |
10ba535a JA |
1365 | return io_u; |
1366 | } | |
1367 | ||
0d29de83 | 1368 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1369 | { |
d72be545 JA |
1370 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1371 | return 0; | |
1372 | ||
1373 | if (td->trim_entries) { | |
0d29de83 | 1374 | int get_trim = 0; |
10ba535a | 1375 | |
0d29de83 JA |
1376 | if (td->trim_batch) { |
1377 | td->trim_batch--; | |
1378 | get_trim = 1; | |
1379 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1380 | td->last_ddir != DDIR_READ) { | |
1381 | td->trim_batch = td->o.trim_batch; | |
1382 | if (!td->trim_batch) | |
1383 | td->trim_batch = td->o.trim_backlog; | |
1384 | get_trim = 1; | |
1385 | } | |
1386 | ||
1387 | if (get_trim && !get_next_trim(td, io_u)) | |
1388 | return 1; | |
2ba1c290 | 1389 | } |
10ba535a | 1390 | |
0d29de83 JA |
1391 | return 0; |
1392 | } | |
1393 | ||
1394 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1395 | { | |
d72be545 JA |
1396 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1397 | return 0; | |
1398 | ||
1399 | if (td->io_hist_len) { | |
9e144189 JA |
1400 | int get_verify = 0; |
1401 | ||
d1ece0c7 | 1402 | if (td->verify_batch) |
9e144189 | 1403 | get_verify = 1; |
d1ece0c7 | 1404 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1405 | td->last_ddir != DDIR_READ) { |
1406 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1407 | if (!td->verify_batch) |
1408 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1409 | get_verify = 1; |
1410 | } | |
1411 | ||
d1ece0c7 JA |
1412 | if (get_verify && !get_next_verify(td, io_u)) { |
1413 | td->verify_batch--; | |
0d29de83 | 1414 | return 1; |
d1ece0c7 | 1415 | } |
9e144189 JA |
1416 | } |
1417 | ||
0d29de83 JA |
1418 | return 0; |
1419 | } | |
1420 | ||
de789769 JA |
1421 | /* |
1422 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1423 | * easy compressible data for simple de-dupe attempts. Do this for every |
1424 | * 512b block in the range, since that should be the smallest block size | |
1425 | * we can expect from a device. | |
de789769 JA |
1426 | */ |
1427 | static void small_content_scramble(struct io_u *io_u) | |
1428 | { | |
23f394d5 | 1429 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1430 | uint64_t boffset; |
23f394d5 JA |
1431 | unsigned int offset; |
1432 | void *p, *end; | |
de789769 | 1433 | |
23f394d5 JA |
1434 | if (!nr_blocks) |
1435 | return; | |
1436 | ||
1437 | p = io_u->xfer_buf; | |
fba76ee8 | 1438 | boffset = io_u->offset; |
81f0366c | 1439 | io_u->buf_filled_len = 0; |
fad82f76 | 1440 | |
23f394d5 JA |
1441 | for (i = 0; i < nr_blocks; i++) { |
1442 | /* | |
1443 | * Fill the byte offset into a "random" start offset of | |
1444 | * the buffer, given by the product of the usec time | |
1445 | * and the actual offset. | |
1446 | */ | |
fad82f76 | 1447 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1448 | offset &= ~(sizeof(uint64_t) - 1); |
1449 | if (offset >= 512 - sizeof(uint64_t)) | |
1450 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1451 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1452 | |
1453 | end = p + 512 - sizeof(io_u->start_time); | |
1454 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1455 | p += 512; | |
fad82f76 | 1456 | boffset += 512; |
23f394d5 | 1457 | } |
de789769 JA |
1458 | } |
1459 | ||
0d29de83 JA |
1460 | /* |
1461 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1462 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1463 | */ | |
1464 | struct io_u *get_io_u(struct thread_data *td) | |
1465 | { | |
1466 | struct fio_file *f; | |
1467 | struct io_u *io_u; | |
de789769 | 1468 | int do_scramble = 0; |
002fe734 | 1469 | long ret = 0; |
0d29de83 JA |
1470 | |
1471 | io_u = __get_io_u(td); | |
1472 | if (!io_u) { | |
1473 | dprint(FD_IO, "__get_io_u failed\n"); | |
1474 | return NULL; | |
1475 | } | |
1476 | ||
1477 | if (check_get_verify(td, io_u)) | |
1478 | goto out; | |
1479 | if (check_get_trim(td, io_u)) | |
1480 | goto out; | |
1481 | ||
755200a3 JA |
1482 | /* |
1483 | * from a requeue, io_u already setup | |
1484 | */ | |
1485 | if (io_u->file) | |
77f392bf | 1486 | goto out; |
755200a3 | 1487 | |
429f6675 JA |
1488 | /* |
1489 | * If using an iolog, grab next piece if any available. | |
1490 | */ | |
d72be545 | 1491 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1492 | if (read_iolog_get(td, io_u)) |
1493 | goto err_put; | |
2ba1c290 | 1494 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1495 | ret = -EBUSY; |
2ba1c290 | 1496 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1497 | goto err_put; |
2ba1c290 | 1498 | } |
5ec10eaa | 1499 | |
429f6675 | 1500 | f = io_u->file; |
002fe734 JA |
1501 | if (!f) { |
1502 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1503 | goto err_put; | |
1504 | } | |
1505 | ||
d6aed795 | 1506 | assert(fio_file_open(f)); |
97af62ce | 1507 | |
ff58fced | 1508 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1509 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1510 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1511 | goto err_put; |
2ba1c290 | 1512 | } |
10ba535a | 1513 | |
f1dfb668 JA |
1514 | f->last_start[io_u->ddir] = io_u->offset; |
1515 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1516 | |
fd68418e | 1517 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1518 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1519 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1520 | td->o.min_bs[DDIR_WRITE], |
1521 | io_u->xfer_buflen); | |
ff441ae8 | 1522 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1523 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1524 | do_scramble = 1; |
d72be545 | 1525 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1526 | populate_verify_io_u(td, io_u); |
1527 | do_scramble = 0; | |
1528 | } | |
fd68418e | 1529 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1530 | /* |
1531 | * Reset the buf_filled parameters so next time if the | |
1532 | * buffer is used for writes it is refilled. | |
1533 | */ | |
cbe8d756 RR |
1534 | io_u->buf_filled_len = 0; |
1535 | } | |
87dc1ab1 | 1536 | } |
10ba535a | 1537 | |
165faf16 JA |
1538 | /* |
1539 | * Set io data pointers. | |
1540 | */ | |
cec6b55d JA |
1541 | io_u->xfer_buf = io_u->buf; |
1542 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1543 | |
6ac7a331 | 1544 | out: |
0d29de83 | 1545 | assert(io_u->file); |
429f6675 | 1546 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1547 | if (!td->o.disable_slat) |
1548 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1549 | if (do_scramble) |
1550 | small_content_scramble(io_u); | |
429f6675 | 1551 | return io_u; |
36167d82 | 1552 | } |
429f6675 | 1553 | err_put: |
2ba1c290 | 1554 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1555 | put_io_u(td, io_u); |
002fe734 | 1556 | return ERR_PTR(ret); |
10ba535a JA |
1557 | } |
1558 | ||
a9da8ab2 | 1559 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1560 | { |
8b28bd41 | 1561 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1562 | |
8b28bd41 DM |
1563 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1564 | return; | |
5451792e | 1565 | |
709c8313 RE |
1566 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1567 | io_u->file ? " on file " : "", | |
1568 | io_u->file ? io_u->file->file_name : "", | |
1569 | strerror(io_u->error), | |
1570 | io_ddir_name(io_u->ddir), | |
1571 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1572 | |
1573 | if (!td->error) | |
1574 | td_verror(td, io_u->error, "io_u error"); | |
1575 | } | |
1576 | ||
a9da8ab2 JA |
1577 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1578 | { | |
1579 | __io_u_log_error(td, io_u); | |
1580 | if (td->parent) | |
1581 | __io_u_log_error(td, io_u); | |
1582 | } | |
1583 | ||
aba6c951 JA |
1584 | static inline int gtod_reduce(struct thread_data *td) |
1585 | { | |
729fe3af | 1586 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1587 | && td->o.disable_bw; |
aba6c951 JA |
1588 | } |
1589 | ||
c8eeb9df JA |
1590 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1591 | struct io_completion_data *icd, | |
1592 | const enum fio_ddir idx, unsigned int bytes) | |
1593 | { | |
a9da8ab2 | 1594 | const int no_reduce = !gtod_reduce(td); |
24d23ca7 | 1595 | unsigned long lusec = 0; |
c8eeb9df | 1596 | |
a9da8ab2 | 1597 | if (no_reduce) |
c8eeb9df JA |
1598 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1599 | ||
1600 | if (!td->o.disable_lat) { | |
1601 | unsigned long tusec; | |
1602 | ||
1603 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1604 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1605 | |
d4afedfd JA |
1606 | if (td->flags & TD_F_PROFILE_OPS) { |
1607 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1608 | ||
1609 | if (ops->io_u_lat) | |
1610 | icd->error = ops->io_u_lat(td, tusec); | |
1611 | } | |
1612 | ||
3e260a46 JA |
1613 | if (td->o.max_latency && tusec > td->o.max_latency) |
1614 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1615 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1616 | if (lat_target_failed(td)) | |
1617 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1618 | } |
c8eeb9df JA |
1619 | } |
1620 | ||
1621 | if (!td->o.disable_clat) { | |
ae588852 | 1622 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1623 | io_u_mark_latency(td, lusec); |
1624 | } | |
1625 | ||
a9da8ab2 JA |
1626 | if (td->parent) |
1627 | td = td->parent; | |
1628 | ||
c8eeb9df JA |
1629 | if (!td->o.disable_bw) |
1630 | add_bw_sample(td, idx, bytes, &icd->time); | |
1631 | ||
a9da8ab2 | 1632 | if (no_reduce) |
aba6c951 | 1633 | add_iops_sample(td, idx, bytes, &icd->time); |
66347cfa DE |
1634 | |
1635 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1636 | uint32_t *info = io_u_block_info(td, io_u); | |
1637 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1638 | if (io_u->ddir == DDIR_TRIM) { | |
1639 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1640 | BLOCK_INFO_TRIMS(*info) + 1); | |
1641 | } else if (io_u->ddir == DDIR_WRITE) { | |
1642 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1643 | *info); | |
1644 | } | |
1645 | } | |
1646 | } | |
c8eeb9df JA |
1647 | } |
1648 | ||
1b8dbf25 SL |
1649 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1650 | { | |
1ae83d45 JA |
1651 | uint64_t secs, remainder, bps, bytes; |
1652 | ||
a9da8ab2 | 1653 | assert(!(td->flags & TD_F_CHILD)); |
1b8dbf25 SL |
1654 | bytes = td->this_io_bytes[ddir]; |
1655 | bps = td->rate_bps[ddir]; | |
1656 | secs = bytes / bps; | |
1657 | remainder = bytes % bps; | |
1658 | return remainder * 1000000 / bps + secs * 1000000; | |
1659 | } | |
1660 | ||
f8b0bd10 | 1661 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1662 | struct io_completion_data *icd) |
10ba535a | 1663 | { |
f8b0bd10 JA |
1664 | struct io_u *io_u = *io_u_ptr; |
1665 | enum fio_ddir ddir = io_u->ddir; | |
1666 | struct fio_file *f = io_u->file; | |
10ba535a | 1667 | |
2ba1c290 JA |
1668 | dprint_io_u(io_u, "io complete"); |
1669 | ||
0c6e7517 | 1670 | assert(io_u->flags & IO_U_F_FLIGHT); |
a9da8ab2 | 1671 | io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1672 | |
1673 | /* | |
1674 | * Mark IO ok to verify | |
1675 | */ | |
1676 | if (io_u->ipo) { | |
890b6656 JA |
1677 | /* |
1678 | * Remove errored entry from the verification list | |
1679 | */ | |
1680 | if (io_u->error) | |
1681 | unlog_io_piece(td, io_u); | |
1682 | else { | |
1683 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1684 | write_barrier(); | |
1685 | } | |
f9401285 JA |
1686 | } |
1687 | ||
f8b0bd10 | 1688 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1689 | td->last_was_sync = 1; |
44f29692 JA |
1690 | if (f) { |
1691 | f->first_write = -1ULL; | |
1692 | f->last_write = -1ULL; | |
1693 | } | |
87dc1ab1 JA |
1694 | return; |
1695 | } | |
1696 | ||
1697 | td->last_was_sync = 0; | |
f8b0bd10 | 1698 | td->last_ddir = ddir; |
87dc1ab1 | 1699 | |
f8b0bd10 | 1700 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1701 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1702 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1703 | int ret; |
10ba535a | 1704 | |
f8b0bd10 JA |
1705 | td->io_blocks[ddir]++; |
1706 | td->this_io_blocks[ddir]++; | |
1707 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1708 | |
1709 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1710 | td->this_io_bytes[ddir] += bytes; |
1711 | ||
ca09be4b JA |
1712 | if (ddir == DDIR_WRITE) { |
1713 | if (f) { | |
1714 | if (f->first_write == -1ULL || | |
1715 | io_u->offset < f->first_write) | |
1716 | f->first_write = io_u->offset; | |
1717 | if (f->last_write == -1ULL || | |
1718 | ((io_u->offset + bytes) > f->last_write)) | |
1719 | f->last_write = io_u->offset + bytes; | |
1720 | } | |
1721 | if (td->last_write_comp) { | |
1722 | int idx = td->last_write_idx++; | |
1723 | ||
1724 | td->last_write_comp[idx] = io_u->offset; | |
1725 | if (td->last_write_idx == td->o.iodepth) | |
1726 | td->last_write_idx = 0; | |
1727 | } | |
44f29692 JA |
1728 | } |
1729 | ||
6b1190fd SL |
1730 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1731 | td->runstate == TD_VERIFYING)) { | |
a9da8ab2 JA |
1732 | struct thread_data *__td = td; |
1733 | ||
f8b0bd10 | 1734 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1735 | |
a9da8ab2 JA |
1736 | if (td->parent) |
1737 | __td = td->parent; | |
1738 | ||
1739 | if (__should_check_rate(__td, ddir)) { | |
1740 | __td->rate_pending_usleep[ddir] = | |
1741 | (usec_for_io(__td, ddir) - | |
1742 | utime_since_now(&__td->start)); | |
b23b6a2f | 1743 | } |
f8b0bd10 | 1744 | if (ddir != DDIR_TRIM && |
a9da8ab2 JA |
1745 | __should_check_rate(__td, oddir)) { |
1746 | __td->rate_pending_usleep[oddir] = | |
1747 | (usec_for_io(__td, oddir) - | |
1748 | utime_since_now(&__td->start)); | |
f8b0bd10 | 1749 | } |
721938ae | 1750 | } |
10ba535a | 1751 | |
f8b0bd10 | 1752 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1753 | |
d7762cf8 | 1754 | if (io_u->end_io) { |
f8b0bd10 JA |
1755 | ret = io_u->end_io(td, io_u_ptr); |
1756 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1757 | if (ret && !icd->error) |
1758 | icd->error = ret; | |
1759 | } | |
ff58fced | 1760 | } else if (io_u->error) { |
10ba535a | 1761 | icd->error = io_u->error; |
5451792e JA |
1762 | io_u_log_error(td, io_u); |
1763 | } | |
8b28bd41 | 1764 | if (icd->error) { |
f8b0bd10 JA |
1765 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1766 | ||
8b28bd41 DM |
1767 | if (!td_non_fatal_error(td, eb, icd->error)) |
1768 | return; | |
f8b0bd10 | 1769 | |
f2bba182 RR |
1770 | /* |
1771 | * If there is a non_fatal error, then add to the error count | |
1772 | * and clear all the errors. | |
1773 | */ | |
1774 | update_error_count(td, icd->error); | |
1775 | td_clear_error(td); | |
1776 | icd->error = 0; | |
f8b0bd10 JA |
1777 | if (io_u) |
1778 | io_u->error = 0; | |
f2bba182 | 1779 | } |
10ba535a JA |
1780 | } |
1781 | ||
9520ebb9 JA |
1782 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1783 | int nr) | |
10ba535a | 1784 | { |
6eaf09d6 | 1785 | int ddir; |
aba6c951 JA |
1786 | |
1787 | if (!gtod_reduce(td)) | |
9520ebb9 | 1788 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1789 | |
3af6ef39 JA |
1790 | icd->nr = nr; |
1791 | ||
10ba535a | 1792 | icd->error = 0; |
6eaf09d6 SL |
1793 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1794 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1795 | } |
1796 | ||
97601024 JA |
1797 | static void ios_completed(struct thread_data *td, |
1798 | struct io_completion_data *icd) | |
36167d82 JA |
1799 | { |
1800 | struct io_u *io_u; | |
1801 | int i; | |
1802 | ||
10ba535a JA |
1803 | for (i = 0; i < icd->nr; i++) { |
1804 | io_u = td->io_ops->event(td, i); | |
1805 | ||
f8b0bd10 | 1806 | io_completed(td, &io_u, icd); |
e8462bd8 | 1807 | |
f8b0bd10 | 1808 | if (io_u) |
e8462bd8 | 1809 | put_io_u(td, io_u); |
10ba535a JA |
1810 | } |
1811 | } | |
97601024 | 1812 | |
e7e6cfb4 JA |
1813 | /* |
1814 | * Complete a single io_u for the sync engines. | |
1815 | */ | |
55312f9f | 1816 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1817 | { |
1818 | struct io_completion_data icd; | |
55312f9f | 1819 | int ddir; |
97601024 | 1820 | |
9520ebb9 | 1821 | init_icd(td, &icd, 1); |
f8b0bd10 | 1822 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1823 | |
f8b0bd10 | 1824 | if (io_u) |
e8462bd8 | 1825 | put_io_u(td, io_u); |
97601024 | 1826 | |
581e7141 JA |
1827 | if (icd.error) { |
1828 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1829 | return -1; | |
1830 | } | |
97601024 | 1831 | |
55312f9f JA |
1832 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1833 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1834 | |
1835 | return 0; | |
97601024 JA |
1836 | } |
1837 | ||
e7e6cfb4 JA |
1838 | /* |
1839 | * Called to complete min_events number of io for the async engines. | |
1840 | */ | |
55312f9f | 1841 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 1842 | { |
97601024 | 1843 | struct io_completion_data icd; |
00de55ef | 1844 | struct timespec *tvp = NULL; |
55312f9f | 1845 | int ret, ddir; |
4d06a338 | 1846 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1847 | |
4950421a | 1848 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1849 | |
4950421a | 1850 | if (!min_evts) |
00de55ef | 1851 | tvp = &ts; |
5fb4b366 RE |
1852 | else if (min_evts > td->cur_depth) |
1853 | min_evts = td->cur_depth; | |
97601024 | 1854 | |
4950421a | 1855 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1856 | if (ret < 0) { |
e1161c32 | 1857 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1858 | return ret; |
1859 | } else if (!ret) | |
1860 | return ret; | |
1861 | ||
9520ebb9 | 1862 | init_icd(td, &icd, ret); |
97601024 | 1863 | ios_completed(td, &icd); |
581e7141 JA |
1864 | if (icd.error) { |
1865 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1866 | return -1; | |
1867 | } | |
97601024 | 1868 | |
55312f9f JA |
1869 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1870 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1871 | |
1872 | return 0; | |
97601024 | 1873 | } |
7e77dd02 JA |
1874 | |
1875 | /* | |
1876 | * Call when io_u is really queued, to update the submission latency. | |
1877 | */ | |
1878 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1879 | { | |
9520ebb9 JA |
1880 | if (!td->o.disable_slat) { |
1881 | unsigned long slat_time; | |
7e77dd02 | 1882 | |
9520ebb9 | 1883 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1884 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1885 | io_u->offset); | |
9520ebb9 | 1886 | } |
7e77dd02 | 1887 | } |
433afcb4 | 1888 | |
5c94b008 JA |
1889 | /* |
1890 | * See if we should reuse the last seed, if dedupe is enabled | |
1891 | */ | |
9451b93e | 1892 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
1893 | { |
1894 | unsigned int v; | |
1895 | unsigned long r; | |
1896 | ||
1897 | if (!td->o.dedupe_percentage) | |
1898 | return &td->buf_state; | |
732eedd0 | 1899 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
1900 | frand_copy(&td->buf_state_prev, &td->buf_state); |
1901 | return &td->buf_state; | |
732eedd0 | 1902 | } |
5c94b008 JA |
1903 | |
1904 | r = __rand(&td->dedupe_state); | |
1905 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1906 | ||
1907 | if (v <= td->o.dedupe_percentage) | |
1908 | return &td->buf_state_prev; | |
1909 | ||
1910 | return &td->buf_state; | |
1911 | } | |
1912 | ||
9451b93e | 1913 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 1914 | { |
9451b93e JA |
1915 | if (td->o.dedupe_percentage == 100) |
1916 | frand_copy(rs, &td->buf_state_prev); | |
1917 | else if (rs == &td->buf_state) | |
5c94b008 JA |
1918 | frand_copy(&td->buf_state_prev, rs); |
1919 | } | |
1920 | ||
cc86c395 JA |
1921 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1922 | unsigned int max_bs) | |
5973cafb | 1923 | { |
d1af2894 JA |
1924 | struct thread_options *o = &td->o; |
1925 | ||
4eff3e57 | 1926 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 1927 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1928 | struct frand_state *rs; |
1066358a | 1929 | unsigned int left = max_bs; |
5c94b008 | 1930 | |
1066358a | 1931 | do { |
9451b93e | 1932 | rs = get_buf_state(td); |
9c42684e | 1933 | |
1066358a | 1934 | min_write = min(min_write, left); |
f97a43a1 | 1935 | |
1066358a JA |
1936 | if (perc) { |
1937 | unsigned int seg = min_write; | |
cc86c395 | 1938 | |
1066358a JA |
1939 | seg = min(min_write, td->o.compress_chunk); |
1940 | if (!seg) | |
1941 | seg = min_write; | |
1942 | ||
1943 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1944 | min_write, o->buffer_pattern, |
1945 | o->buffer_pattern_bytes); | |
1066358a JA |
1946 | } else |
1947 | fill_random_buf(rs, buf, min_write); | |
1948 | ||
1949 | buf += min_write; | |
1950 | left -= min_write; | |
9451b93e | 1951 | save_buf_state(td, rs); |
1066358a | 1952 | } while (left); |
d1af2894 JA |
1953 | } else if (o->buffer_pattern_bytes) |
1954 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 1955 | else if (o->zero_buffers) |
cc86c395 | 1956 | memset(buf, 0, max_bs); |
999d245e | 1957 | else |
9451b93e | 1958 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
1959 | } |
1960 | ||
1961 | /* | |
1962 | * "randomly" fill the buffer contents | |
1963 | */ | |
1964 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1965 | unsigned int min_write, unsigned int max_bs) | |
1966 | { | |
1967 | io_u->buf_filled_len = 0; | |
1968 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1969 | } |