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" |
0f38bbef | 15 | #include "lib/pow2.h" |
10ba535a | 16 | |
97601024 JA |
17 | struct io_completion_data { |
18 | int nr; /* input */ | |
97601024 JA |
19 | |
20 | int error; /* output */ | |
100f49f1 | 21 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
97601024 JA |
22 | struct timeval time; /* output */ |
23 | }; | |
24 | ||
10ba535a | 25 | /* |
7ebd796f | 26 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
27 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
28 | */ | |
1ae83d45 | 29 | static int random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 30 | { |
7ebd796f | 31 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
32 | } |
33 | ||
df415585 JA |
34 | /* |
35 | * Mark a given offset as used in the map. | |
36 | */ | |
9bf2061e | 37 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 38 | { |
2dc1bbeb | 39 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 40 | struct fio_file *f = io_u->file; |
51ede0b1 | 41 | unsigned int nr_blocks; |
1ae83d45 | 42 | uint64_t block; |
df415585 | 43 | |
1ae83d45 | 44 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 45 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 46 | |
2ab9e98b | 47 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 48 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 49 | |
51ede0b1 JA |
50 | if ((nr_blocks * min_bs) < io_u->buflen) |
51 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
52 | } |
53 | ||
74776733 JA |
54 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
55 | enum fio_ddir ddir) | |
2ba1c290 | 56 | { |
74776733 JA |
57 | uint64_t max_blocks; |
58 | uint64_t max_size; | |
2ba1c290 | 59 | |
ff58fced JA |
60 | assert(ddir_rw(ddir)); |
61 | ||
d9dd70f7 JA |
62 | /* |
63 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
64 | */ | |
65 | max_size = f->io_size; | |
66 | if (max_size > f->real_file_size) | |
67 | max_size = f->real_file_size; | |
68 | ||
ed335855 SN |
69 | if (td->o.zone_range) |
70 | max_size = td->o.zone_range; | |
71 | ||
0412d12e JE |
72 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
73 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
74 | ||
1ae83d45 | 75 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
76 | if (!max_blocks) |
77 | return 0; | |
78 | ||
67778e88 | 79 | return max_blocks; |
2ba1c290 JA |
80 | } |
81 | ||
1ae83d45 JA |
82 | struct rand_off { |
83 | struct flist_head list; | |
84 | uint64_t off; | |
85 | }; | |
86 | ||
e25839d4 | 87 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 88 | enum fio_ddir ddir, uint64_t *b) |
ec4015da | 89 | { |
6f49f8bc | 90 | uint64_t r; |
5e0baa7f | 91 | |
74776733 | 92 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) { |
6f49f8bc JA |
93 | uint64_t lastb; |
94 | ||
95 | lastb = last_block(td, f, ddir); | |
96 | if (!lastb) | |
97 | return 1; | |
98 | ||
d6b72507 | 99 | r = __rand(&td->random_state); |
8055e41d | 100 | |
4b91ee8f | 101 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 102 | |
54a21917 | 103 | *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0)); |
51ede0b1 | 104 | } else { |
8055e41d | 105 | uint64_t off = 0; |
43c63a78 | 106 | |
967d1b63 JA |
107 | assert(fio_file_lfsr(f)); |
108 | ||
6f49f8bc | 109 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 110 | return 1; |
ec4015da | 111 | |
8055e41d JA |
112 | *b = off; |
113 | } | |
0ce8b119 | 114 | |
ec4015da | 115 | /* |
51ede0b1 | 116 | * if we are not maintaining a random map, we are done. |
ec4015da | 117 | */ |
51ede0b1 JA |
118 | if (!file_randommap(td, f)) |
119 | goto ret; | |
43c63a78 JA |
120 | |
121 | /* | |
51ede0b1 | 122 | * calculate map offset and check if it's free |
43c63a78 | 123 | */ |
51ede0b1 JA |
124 | if (random_map_free(f, *b)) |
125 | goto ret; | |
126 | ||
4b91ee8f JA |
127 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
128 | (unsigned long long) *b); | |
51ede0b1 | 129 | |
7ebd796f | 130 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
131 | if (*b == (uint64_t) -1ULL) |
132 | return 1; | |
0ce8b119 JA |
133 | ret: |
134 | return 0; | |
ec4015da JA |
135 | } |
136 | ||
925fee33 JA |
137 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
138 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 139 | uint64_t *b) |
e25839d4 | 140 | { |
9c6f6316 | 141 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
142 | return 0; |
143 | } | |
144 | ||
925fee33 JA |
145 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
146 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 147 | uint64_t *b) |
925fee33 | 148 | { |
9c6f6316 | 149 | *b = pareto_next(&f->zipf); |
925fee33 JA |
150 | return 0; |
151 | } | |
152 | ||
56d9fa4b JA |
153 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
154 | struct fio_file *f, enum fio_ddir ddir, | |
155 | uint64_t *b) | |
156 | { | |
157 | *b = gauss_next(&f->gauss); | |
158 | return 0; | |
159 | } | |
160 | ||
161 | ||
1ae83d45 JA |
162 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
163 | { | |
164 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
165 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
166 | ||
167 | return r1->off - r2->off; | |
168 | } | |
169 | ||
170 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
171 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
172 | { |
173 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
174 | return __get_next_rand_offset(td, f, ddir, b); | |
175 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
176 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
177 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
178 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
179 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
180 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e25839d4 JA |
181 | |
182 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
183 | return 1; | |
184 | } | |
185 | ||
bcd5abfa JA |
186 | /* |
187 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
188 | * specified. | |
189 | */ | |
190 | static inline int should_sort_io(struct thread_data *td) | |
191 | { | |
192 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
193 | return 0; | |
194 | if (!td_random(td)) | |
195 | return 0; | |
196 | if (td->runstate != TD_VERIFYING) | |
197 | return 0; | |
198 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
199 | return 0; | |
200 | ||
201 | return 1; | |
202 | } | |
203 | ||
d9472271 | 204 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
205 | { |
206 | unsigned int v; | |
207 | unsigned long r; | |
208 | ||
d9472271 | 209 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
210 | return 1; |
211 | ||
d6b72507 | 212 | r = __rand(&td->seq_rand_state[ddir]); |
54a21917 | 213 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
211c9b89 | 214 | |
d9472271 | 215 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
216 | } |
217 | ||
1ae83d45 JA |
218 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
219 | enum fio_ddir ddir, uint64_t *b) | |
220 | { | |
221 | struct rand_off *r; | |
222 | int i, ret = 1; | |
223 | ||
bcd5abfa | 224 | if (!should_sort_io(td)) |
1ae83d45 JA |
225 | return get_off_from_method(td, f, ddir, b); |
226 | ||
227 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 228 | fetch: |
9342d5f8 | 229 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
230 | flist_del(&r->list); |
231 | *b = r->off; | |
232 | free(r); | |
233 | return 0; | |
234 | } | |
235 | ||
236 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
237 | r = malloc(sizeof(*r)); | |
238 | ||
239 | ret = get_off_from_method(td, f, ddir, &r->off); | |
240 | if (ret) { | |
241 | free(r); | |
242 | break; | |
243 | } | |
244 | ||
245 | flist_add(&r->list, &td->next_rand_list); | |
246 | } | |
247 | ||
248 | if (ret && !i) | |
249 | return ret; | |
250 | ||
251 | assert(!flist_empty(&td->next_rand_list)); | |
252 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
253 | goto fetch; | |
254 | } | |
255 | ||
38dad62d | 256 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 257 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 258 | { |
c04e4661 DE |
259 | if (!get_next_rand_offset(td, f, ddir, b)) |
260 | return 0; | |
261 | ||
262 | if (td->o.time_based) { | |
33c48814 | 263 | fio_file_reset(td, f); |
c04e4661 DE |
264 | if (!get_next_rand_offset(td, f, ddir, b)) |
265 | return 0; | |
38dad62d JA |
266 | } |
267 | ||
c04e4661 | 268 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 269 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 270 | (unsigned long long) f->real_file_size); |
c04e4661 | 271 | return 1; |
38dad62d JA |
272 | } |
273 | ||
37cf9e3c | 274 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 275 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 276 | { |
ac002339 JA |
277 | struct thread_options *o = &td->o; |
278 | ||
ff58fced JA |
279 | assert(ddir_rw(ddir)); |
280 | ||
f1dfb668 | 281 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
ac002339 | 282 | o->time_based) |
f1dfb668 | 283 | f->last_pos[ddir] = f->last_pos[ddir] - f->io_size; |
c04e4661 | 284 | |
f1dfb668 | 285 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 286 | uint64_t pos; |
059b0802 | 287 | |
f1dfb668 JA |
288 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) |
289 | f->last_pos[ddir] = f->real_file_size; | |
a66da7a2 | 290 | |
f1dfb668 | 291 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
292 | if (pos && o->ddir_seq_add) { |
293 | pos += o->ddir_seq_add; | |
294 | ||
295 | /* | |
296 | * If we reach beyond the end of the file | |
297 | * with holed IO, wrap around to the | |
298 | * beginning again. | |
299 | */ | |
300 | if (pos >= f->real_file_size) | |
301 | pos = f->file_offset; | |
302 | } | |
059b0802 | 303 | |
37cf9e3c | 304 | *offset = pos; |
38dad62d JA |
305 | return 0; |
306 | } | |
307 | ||
308 | return 1; | |
309 | } | |
310 | ||
311 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
312 | enum fio_ddir ddir, int rw_seq, |
313 | unsigned int *is_random) | |
38dad62d JA |
314 | { |
315 | struct fio_file *f = io_u->file; | |
1ae83d45 | 316 | uint64_t b, offset; |
38dad62d JA |
317 | int ret; |
318 | ||
ff58fced JA |
319 | assert(ddir_rw(ddir)); |
320 | ||
37cf9e3c JA |
321 | b = offset = -1ULL; |
322 | ||
38dad62d | 323 | if (rw_seq) { |
211c9b89 | 324 | if (td_random(td)) { |
6aca9b3d | 325 | if (should_do_random(td, ddir)) { |
211c9b89 | 326 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
327 | *is_random = 1; |
328 | } else { | |
329 | *is_random = 0; | |
a9da8ab2 | 330 | io_u_set(io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
331 | ret = get_next_seq_offset(td, f, ddir, &offset); |
332 | if (ret) | |
333 | ret = get_next_rand_block(td, f, ddir, &b); | |
334 | } | |
6aca9b3d JA |
335 | } else { |
336 | *is_random = 0; | |
37cf9e3c | 337 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 338 | } |
38dad62d | 339 | } else { |
a9da8ab2 | 340 | io_u_set(io_u, IO_U_F_BUSY_OK); |
6aca9b3d | 341 | *is_random = 0; |
38dad62d JA |
342 | |
343 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 344 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 345 | if (ret) { |
37cf9e3c | 346 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
347 | *is_random = 0; |
348 | } | |
38dad62d | 349 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
350 | if (f->last_start[ddir] != -1ULL) |
351 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 352 | else |
37cf9e3c | 353 | offset = 0; |
38dad62d JA |
354 | ret = 0; |
355 | } else { | |
356 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
357 | ret = 1; | |
358 | } | |
359 | } | |
6d68b997 | 360 | |
37cf9e3c JA |
361 | if (!ret) { |
362 | if (offset != -1ULL) | |
363 | io_u->offset = offset; | |
364 | else if (b != -1ULL) | |
365 | io_u->offset = b * td->o.ba[ddir]; | |
366 | else { | |
4e0a8fa2 | 367 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
368 | ret = 1; |
369 | } | |
370 | } | |
371 | ||
38dad62d JA |
372 | return ret; |
373 | } | |
374 | ||
10ba535a JA |
375 | /* |
376 | * For random io, generate a random new block and see if it's used. Repeat | |
377 | * until we find a free one. For sequential io, just return the end of | |
378 | * the last io issued. | |
379 | */ | |
6aca9b3d JA |
380 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
381 | unsigned int *is_random) | |
10ba535a | 382 | { |
9bf2061e | 383 | struct fio_file *f = io_u->file; |
4ba66134 | 384 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 385 | int rw_seq_hit = 0; |
10ba535a | 386 | |
ff58fced JA |
387 | assert(ddir_rw(ddir)); |
388 | ||
38dad62d JA |
389 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
390 | rw_seq_hit = 1; | |
5736c10d | 391 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 392 | } |
211097b2 | 393 | |
6aca9b3d | 394 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 395 | return 1; |
10ba535a | 396 | |
009bd847 JA |
397 | if (io_u->offset >= f->io_size) { |
398 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
399 | (unsigned long long) io_u->offset, |
400 | (unsigned long long) f->io_size); | |
009bd847 JA |
401 | return 1; |
402 | } | |
403 | ||
404 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
405 | if (io_u->offset >= f->real_file_size) { |
406 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
407 | (unsigned long long) io_u->offset, |
408 | (unsigned long long) f->real_file_size); | |
10ba535a | 409 | return 1; |
2ba1c290 | 410 | } |
10ba535a JA |
411 | |
412 | return 0; | |
413 | } | |
414 | ||
6aca9b3d JA |
415 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
416 | unsigned int *is_random) | |
15dc1934 | 417 | { |
d72be545 JA |
418 | if (td->flags & TD_F_PROFILE_OPS) { |
419 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 420 | |
d72be545 | 421 | if (ops->fill_io_u_off) |
6aca9b3d | 422 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 423 | } |
15dc1934 | 424 | |
6aca9b3d | 425 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
426 | } |
427 | ||
79944128 JA |
428 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
429 | unsigned int buflen) | |
430 | { | |
431 | struct fio_file *f = io_u->file; | |
432 | ||
bedc9dc2 | 433 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
434 | } |
435 | ||
6aca9b3d JA |
436 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
437 | unsigned int is_random) | |
10ba535a | 438 | { |
6aca9b3d | 439 | int ddir = io_u->ddir; |
24d23ca7 | 440 | unsigned int buflen = 0; |
f3059de1 | 441 | unsigned int minbs, maxbs; |
54a21917 | 442 | unsigned long r; |
10ba535a | 443 | |
9ee1c647 | 444 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
445 | |
446 | if (td->o.bs_is_seq_rand) | |
447 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 448 | |
f3059de1 JA |
449 | minbs = td->o.min_bs[ddir]; |
450 | maxbs = td->o.max_bs[ddir]; | |
451 | ||
79944128 JA |
452 | if (minbs == maxbs) |
453 | return minbs; | |
454 | ||
52c58027 JA |
455 | /* |
456 | * If we can't satisfy the min block size from here, then fail | |
457 | */ | |
458 | if (!io_u_fits(td, io_u, minbs)) | |
459 | return 0; | |
460 | ||
79944128 | 461 | do { |
d6b72507 | 462 | r = __rand(&td->bsrange_state); |
4c07ad86 | 463 | |
720e84ad | 464 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 465 | buflen = 1 + (unsigned int) ((double) maxbs * |
54a21917 | 466 | (r / (FRAND_MAX + 1.0))); |
f3059de1 JA |
467 | if (buflen < minbs) |
468 | buflen = minbs; | |
5ec10eaa | 469 | } else { |
564ca972 JA |
470 | long perc = 0; |
471 | unsigned int i; | |
472 | ||
720e84ad JA |
473 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
474 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
475 | |
476 | buflen = bsp->bs; | |
477 | perc += bsp->perc; | |
54a21917 | 478 | if ((r <= ((FRAND_MAX / 100L) * perc)) && |
79944128 | 479 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
480 | break; |
481 | } | |
482 | } | |
79944128 | 483 | |
a9f70b1f JB |
484 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
485 | buflen = (buflen + td->o.verify_interval - 1) & | |
486 | ~(td->o.verify_interval - 1); | |
487 | ||
f3059de1 | 488 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
7c306bb1 | 489 | buflen &= ~(minbs - 1); |
10ba535a | 490 | |
79944128 | 491 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 492 | |
10ba535a JA |
493 | return buflen; |
494 | } | |
495 | ||
6aca9b3d JA |
496 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
497 | unsigned int is_random) | |
15dc1934 | 498 | { |
d72be545 JA |
499 | if (td->flags & TD_F_PROFILE_OPS) { |
500 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 501 | |
d72be545 | 502 | if (ops->fill_io_u_size) |
6aca9b3d | 503 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 504 | } |
15dc1934 | 505 | |
6aca9b3d | 506 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
507 | } |
508 | ||
afe24a5a JA |
509 | static void set_rwmix_bytes(struct thread_data *td) |
510 | { | |
afe24a5a JA |
511 | unsigned int diff; |
512 | ||
513 | /* | |
514 | * we do time or byte based switch. this is needed because | |
515 | * buffered writes may issue a lot quicker than they complete, | |
516 | * whereas reads do not. | |
517 | */ | |
e47f799f | 518 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 519 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
520 | } |
521 | ||
522 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
523 | { | |
524 | unsigned int v; | |
1294c3ec | 525 | unsigned long r; |
e47f799f | 526 | |
d6b72507 | 527 | r = __rand(&td->rwmix_state); |
54a21917 | 528 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
4c07ad86 | 529 | |
04c540d9 | 530 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
531 | return DDIR_READ; |
532 | ||
533 | return DDIR_WRITE; | |
afe24a5a JA |
534 | } |
535 | ||
002e7183 JA |
536 | void io_u_quiesce(struct thread_data *td) |
537 | { | |
538 | /* | |
539 | * We are going to sleep, ensure that we flush anything pending as | |
540 | * not to skew our latency numbers. | |
541 | * | |
542 | * Changed to only monitor 'in flight' requests here instead of the | |
543 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
544 | * io's that have been actually submitted to an async engine, | |
545 | * and cur_depth is meaningless for sync engines. | |
546 | */ | |
9cc80b6d JA |
547 | if (td->io_u_queued || td->cur_depth) { |
548 | int fio_unused ret; | |
549 | ||
550 | ret = td_io_commit(td); | |
551 | } | |
552 | ||
002e7183 JA |
553 | while (td->io_u_in_flight) { |
554 | int fio_unused ret; | |
555 | ||
55312f9f | 556 | ret = io_u_queued_complete(td, 1); |
002e7183 JA |
557 | } |
558 | } | |
559 | ||
581e7141 JA |
560 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
561 | { | |
562 | enum fio_ddir odir = ddir ^ 1; | |
581e7141 JA |
563 | long usec; |
564 | ||
ff58fced JA |
565 | assert(ddir_rw(ddir)); |
566 | ||
315fcfec | 567 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
568 | return ddir; |
569 | ||
570 | /* | |
571 | * We have too much pending sleep in this direction. See if we | |
572 | * should switch. | |
573 | */ | |
315fcfec | 574 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
575 | /* |
576 | * Other direction does not have too much pending, switch | |
577 | */ | |
578 | if (td->rate_pending_usleep[odir] < 100000) | |
579 | return odir; | |
580 | ||
581 | /* | |
582 | * Both directions have pending sleep. Sleep the minimum time | |
583 | * and deduct from both. | |
584 | */ | |
585 | if (td->rate_pending_usleep[ddir] <= | |
586 | td->rate_pending_usleep[odir]) { | |
587 | usec = td->rate_pending_usleep[ddir]; | |
588 | } else { | |
589 | usec = td->rate_pending_usleep[odir]; | |
590 | ddir = odir; | |
591 | } | |
592 | } else | |
593 | usec = td->rate_pending_usleep[ddir]; | |
594 | ||
a9da8ab2 JA |
595 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
596 | io_u_quiesce(td); | |
78c1eda5 | 597 | |
9cc80b6d | 598 | usec = usec_sleep(td, usec); |
581e7141 JA |
599 | |
600 | td->rate_pending_usleep[ddir] -= usec; | |
601 | ||
602 | odir = ddir ^ 1; | |
603 | if (td_rw(td) && __should_check_rate(td, odir)) | |
604 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 605 | |
581e7141 JA |
606 | return ddir; |
607 | } | |
608 | ||
10ba535a JA |
609 | /* |
610 | * Return the data direction for the next io_u. If the job is a | |
611 | * mixed read/write workload, check the rwmix cycle and switch if | |
612 | * necessary. | |
613 | */ | |
1e97cce9 | 614 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 615 | { |
581e7141 JA |
616 | enum fio_ddir ddir; |
617 | ||
5f9099ea JA |
618 | /* |
619 | * see if it's time to fsync | |
620 | */ | |
621 | if (td->o.fsync_blocks && | |
622 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
623 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
624 | return DDIR_SYNC; | |
625 | ||
626 | /* | |
627 | * see if it's time to fdatasync | |
628 | */ | |
629 | if (td->o.fdatasync_blocks && | |
630 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
631 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
632 | return DDIR_DATASYNC; | |
633 | ||
44f29692 JA |
634 | /* |
635 | * see if it's time to sync_file_range | |
636 | */ | |
637 | if (td->sync_file_range_nr && | |
638 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
639 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
640 | return DDIR_SYNC_FILE_RANGE; | |
641 | ||
10ba535a | 642 | if (td_rw(td)) { |
10ba535a JA |
643 | /* |
644 | * Check if it's time to seed a new data direction. | |
645 | */ | |
e4928662 | 646 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
647 | /* |
648 | * Put a top limit on how many bytes we do for | |
649 | * one data direction, to avoid overflowing the | |
650 | * ranges too much | |
651 | */ | |
652 | ddir = get_rand_ddir(td); | |
e47f799f JA |
653 | |
654 | if (ddir != td->rwmix_ddir) | |
655 | set_rwmix_bytes(td); | |
656 | ||
657 | td->rwmix_ddir = ddir; | |
10ba535a | 658 | } |
581e7141 | 659 | ddir = td->rwmix_ddir; |
10ba535a | 660 | } else if (td_read(td)) |
581e7141 | 661 | ddir = DDIR_READ; |
6eaf09d6 | 662 | else if (td_write(td)) |
581e7141 | 663 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
664 | else |
665 | ddir = DDIR_TRIM; | |
581e7141 JA |
666 | |
667 | td->rwmix_ddir = rate_ddir(td, ddir); | |
668 | return td->rwmix_ddir; | |
10ba535a JA |
669 | } |
670 | ||
1ef2b6be JA |
671 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
672 | { | |
0e4dd95c DE |
673 | enum fio_ddir ddir = get_rw_ddir(td); |
674 | ||
82a90686 | 675 | if (td_trimwrite(td)) { |
0e4dd95c DE |
676 | struct fio_file *f = io_u->file; |
677 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
678 | ddir = DDIR_TRIM; | |
679 | else | |
680 | ddir = DDIR_WRITE; | |
681 | } | |
682 | ||
683 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be JA |
684 | |
685 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
686 | td->o.barrier_blocks && | |
687 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
688 | td->io_issues[DDIR_WRITE]) | |
a9da8ab2 | 689 | io_u_set(io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
690 | } |
691 | ||
e8462bd8 | 692 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 693 | { |
71b84caa | 694 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
695 | |
696 | if (ret) | |
697 | td_verror(td, ret, "file close"); | |
698 | } | |
699 | ||
10ba535a JA |
700 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
701 | { | |
a9da8ab2 JA |
702 | if (td->parent) |
703 | td = td->parent; | |
704 | ||
e8462bd8 JA |
705 | td_io_u_lock(td); |
706 | ||
f8b0bd10 | 707 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 708 | put_file_log(td, io_u->file); |
f8b0bd10 | 709 | |
10ba535a | 710 | io_u->file = NULL; |
a9da8ab2 | 711 | io_u_set(io_u, IO_U_F_FREE); |
d7ee2a7d | 712 | |
a9da8ab2 | 713 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 714 | td->cur_depth--; |
a9da8ab2 JA |
715 | assert(!(td->flags & TD_F_CHILD)); |
716 | } | |
2ae0b204 | 717 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
718 | td_io_u_unlock(td); |
719 | td_io_u_free_notify(td); | |
10ba535a JA |
720 | } |
721 | ||
f2bba182 RR |
722 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
723 | { | |
a9da8ab2 | 724 | io_u_clear(io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
725 | put_io_u(td, io_u); |
726 | } | |
727 | ||
755200a3 JA |
728 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
729 | { | |
730 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 731 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 732 | |
465221b0 JA |
733 | dprint(FD_IO, "requeue %p\n", __io_u); |
734 | ||
a9da8ab2 JA |
735 | if (td->parent) |
736 | td = td->parent; | |
737 | ||
e8462bd8 JA |
738 | td_io_u_lock(td); |
739 | ||
a9da8ab2 | 740 | io_u_set(__io_u, IO_U_F_FREE); |
bcd5abfa JA |
741 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
742 | td->io_issues[ddir]--; | |
5ec10eaa | 743 | |
a9da8ab2 JA |
744 | io_u_clear(__io_u, IO_U_F_FLIGHT); |
745 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { | |
0c41214f | 746 | td->cur_depth--; |
a9da8ab2 JA |
747 | assert(!(td->flags & TD_F_CHILD)); |
748 | } | |
2ae0b204 JA |
749 | |
750 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 751 | td_io_u_unlock(td); |
a9da8ab2 | 752 | td_io_u_free_notify(td); |
755200a3 JA |
753 | *io_u = NULL; |
754 | } | |
755 | ||
9bf2061e | 756 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 757 | { |
6aca9b3d JA |
758 | unsigned int is_random; |
759 | ||
b4c5e1ac JA |
760 | if (td->io_ops->flags & FIO_NOIO) |
761 | goto out; | |
762 | ||
1ef2b6be | 763 | set_rw_ddir(td, io_u); |
5f9099ea | 764 | |
87dc1ab1 | 765 | /* |
ff58fced | 766 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 767 | */ |
ff58fced | 768 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 769 | goto out; |
a00735e6 | 770 | |
48f5abd3 JA |
771 | /* |
772 | * See if it's time to switch to a new zone | |
773 | */ | |
13af05ae | 774 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
775 | struct fio_file *f = io_u->file; |
776 | ||
48f5abd3 | 777 | td->zone_bytes = 0; |
4c8be5b1 JA |
778 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
779 | ||
780 | /* | |
781 | * Wrap from the beginning, if we exceed the file size | |
782 | */ | |
783 | if (f->file_offset >= f->real_file_size) | |
784 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 785 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
786 | td->io_skip_bytes += td->o.zone_skip; |
787 | } | |
788 | ||
10ba535a | 789 | /* |
c685b5b2 JA |
790 | * No log, let the seq/rand engine retrieve the next buflen and |
791 | * position. | |
10ba535a | 792 | */ |
6aca9b3d | 793 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 794 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 795 | return 1; |
2ba1c290 | 796 | } |
10ba535a | 797 | |
6aca9b3d | 798 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
799 | if (!io_u->buflen) { |
800 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 801 | return 1; |
2ba1c290 | 802 | } |
bca4ed4d | 803 | |
2ba1c290 JA |
804 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
805 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
806 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
807 | (unsigned long long) io_u->offset, io_u->buflen, | |
808 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 809 | return 1; |
2ba1c290 | 810 | } |
6a5e6884 | 811 | |
bca4ed4d JA |
812 | /* |
813 | * mark entry before potentially trimming io_u | |
814 | */ | |
303032ae | 815 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 816 | mark_random_map(td, io_u); |
bca4ed4d | 817 | |
c38e9468 | 818 | out: |
2ba1c290 | 819 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 820 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 821 | return 0; |
10ba535a JA |
822 | } |
823 | ||
838bc709 JA |
824 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
825 | { | |
2b13e716 | 826 | int idx = 0; |
838bc709 JA |
827 | |
828 | switch (nr) { | |
829 | default: | |
2b13e716 | 830 | idx = 6; |
838bc709 JA |
831 | break; |
832 | case 33 ... 64: | |
2b13e716 | 833 | idx = 5; |
838bc709 JA |
834 | break; |
835 | case 17 ... 32: | |
2b13e716 | 836 | idx = 4; |
838bc709 JA |
837 | break; |
838 | case 9 ... 16: | |
2b13e716 | 839 | idx = 3; |
838bc709 JA |
840 | break; |
841 | case 5 ... 8: | |
2b13e716 | 842 | idx = 2; |
838bc709 JA |
843 | break; |
844 | case 1 ... 4: | |
2b13e716 | 845 | idx = 1; |
838bc709 JA |
846 | case 0: |
847 | break; | |
848 | } | |
849 | ||
2b13e716 | 850 | map[idx]++; |
838bc709 JA |
851 | } |
852 | ||
853 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
854 | { | |
855 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
856 | td->ts.total_submit++; | |
857 | } | |
858 | ||
859 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
860 | { | |
861 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
862 | td->ts.total_complete++; | |
863 | } | |
864 | ||
d8005759 | 865 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 866 | { |
2b13e716 | 867 | int idx = 0; |
71619dc2 JA |
868 | |
869 | switch (td->cur_depth) { | |
870 | default: | |
2b13e716 | 871 | idx = 6; |
a783e61a | 872 | break; |
71619dc2 | 873 | case 32 ... 63: |
2b13e716 | 874 | idx = 5; |
a783e61a | 875 | break; |
71619dc2 | 876 | case 16 ... 31: |
2b13e716 | 877 | idx = 4; |
a783e61a | 878 | break; |
71619dc2 | 879 | case 8 ... 15: |
2b13e716 | 880 | idx = 3; |
a783e61a | 881 | break; |
71619dc2 | 882 | case 4 ... 7: |
2b13e716 | 883 | idx = 2; |
a783e61a | 884 | break; |
71619dc2 | 885 | case 2 ... 3: |
2b13e716 | 886 | idx = 1; |
71619dc2 JA |
887 | case 1: |
888 | break; | |
889 | } | |
890 | ||
2b13e716 | 891 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
892 | } |
893 | ||
04a0feae JA |
894 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
895 | { | |
2b13e716 | 896 | int idx = 0; |
04a0feae JA |
897 | |
898 | assert(usec < 1000); | |
899 | ||
900 | switch (usec) { | |
901 | case 750 ... 999: | |
2b13e716 | 902 | idx = 9; |
04a0feae JA |
903 | break; |
904 | case 500 ... 749: | |
2b13e716 | 905 | idx = 8; |
04a0feae JA |
906 | break; |
907 | case 250 ... 499: | |
2b13e716 | 908 | idx = 7; |
04a0feae JA |
909 | break; |
910 | case 100 ... 249: | |
2b13e716 | 911 | idx = 6; |
04a0feae JA |
912 | break; |
913 | case 50 ... 99: | |
2b13e716 | 914 | idx = 5; |
04a0feae JA |
915 | break; |
916 | case 20 ... 49: | |
2b13e716 | 917 | idx = 4; |
04a0feae JA |
918 | break; |
919 | case 10 ... 19: | |
2b13e716 | 920 | idx = 3; |
04a0feae JA |
921 | break; |
922 | case 4 ... 9: | |
2b13e716 | 923 | idx = 2; |
04a0feae JA |
924 | break; |
925 | case 2 ... 3: | |
2b13e716 | 926 | idx = 1; |
04a0feae JA |
927 | case 0 ... 1: |
928 | break; | |
929 | } | |
930 | ||
2b13e716 JA |
931 | assert(idx < FIO_IO_U_LAT_U_NR); |
932 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
933 | } |
934 | ||
935 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 936 | { |
2b13e716 | 937 | int idx = 0; |
ec118304 JA |
938 | |
939 | switch (msec) { | |
940 | default: | |
2b13e716 | 941 | idx = 11; |
04a0feae | 942 | break; |
8abdce66 | 943 | case 1000 ... 1999: |
2b13e716 | 944 | idx = 10; |
04a0feae | 945 | break; |
8abdce66 | 946 | case 750 ... 999: |
2b13e716 | 947 | idx = 9; |
04a0feae | 948 | break; |
8abdce66 | 949 | case 500 ... 749: |
2b13e716 | 950 | idx = 8; |
04a0feae | 951 | break; |
8abdce66 | 952 | case 250 ... 499: |
2b13e716 | 953 | idx = 7; |
04a0feae | 954 | break; |
8abdce66 | 955 | case 100 ... 249: |
2b13e716 | 956 | idx = 6; |
04a0feae | 957 | break; |
8abdce66 | 958 | case 50 ... 99: |
2b13e716 | 959 | idx = 5; |
04a0feae | 960 | break; |
8abdce66 | 961 | case 20 ... 49: |
2b13e716 | 962 | idx = 4; |
04a0feae | 963 | break; |
8abdce66 | 964 | case 10 ... 19: |
2b13e716 | 965 | idx = 3; |
04a0feae | 966 | break; |
8abdce66 | 967 | case 4 ... 9: |
2b13e716 | 968 | idx = 2; |
04a0feae | 969 | break; |
ec118304 | 970 | case 2 ... 3: |
2b13e716 | 971 | idx = 1; |
ec118304 JA |
972 | case 0 ... 1: |
973 | break; | |
974 | } | |
975 | ||
2b13e716 JA |
976 | assert(idx < FIO_IO_U_LAT_M_NR); |
977 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
978 | } |
979 | ||
980 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
981 | { | |
982 | if (usec < 1000) | |
983 | io_u_mark_lat_usec(td, usec); | |
984 | else | |
985 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
986 | } |
987 | ||
0aabe160 JA |
988 | /* |
989 | * Get next file to service by choosing one at random | |
990 | */ | |
2cc52930 JA |
991 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
992 | enum fio_file_flags goodf, | |
d6aed795 | 993 | enum fio_file_flags badf) |
0aabe160 | 994 | { |
0aabe160 | 995 | struct fio_file *f; |
1c178180 | 996 | int fno; |
0aabe160 JA |
997 | |
998 | do { | |
87b10676 | 999 | int opened = 0; |
1294c3ec | 1000 | unsigned long r; |
4c07ad86 | 1001 | |
d6b72507 | 1002 | r = __rand(&td->next_file_state); |
54a21917 | 1003 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 1004 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 1005 | |
126d65c6 | 1006 | f = td->files[fno]; |
d6aed795 | 1007 | if (fio_file_done(f)) |
059e63c0 | 1008 | continue; |
1c178180 | 1009 | |
d6aed795 | 1010 | if (!fio_file_open(f)) { |
87b10676 JA |
1011 | int err; |
1012 | ||
002fe734 JA |
1013 | if (td->nr_open_files >= td->o.open_files) |
1014 | return ERR_PTR(-EBUSY); | |
1015 | ||
87b10676 JA |
1016 | err = td_io_open_file(td, f); |
1017 | if (err) | |
1018 | continue; | |
1019 | opened = 1; | |
1020 | } | |
1021 | ||
2ba1c290 JA |
1022 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1023 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1024 | return f; |
2ba1c290 | 1025 | } |
87b10676 JA |
1026 | if (opened) |
1027 | td_io_close_file(td, f); | |
0aabe160 JA |
1028 | } while (1); |
1029 | } | |
1030 | ||
1031 | /* | |
1032 | * Get next file to service by doing round robin between all available ones | |
1033 | */ | |
1c178180 JA |
1034 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1035 | int badf) | |
3d7c391d JA |
1036 | { |
1037 | unsigned int old_next_file = td->next_file; | |
1038 | struct fio_file *f; | |
1039 | ||
1040 | do { | |
87b10676 JA |
1041 | int opened = 0; |
1042 | ||
126d65c6 | 1043 | f = td->files[td->next_file]; |
3d7c391d JA |
1044 | |
1045 | td->next_file++; | |
2dc1bbeb | 1046 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1047 | td->next_file = 0; |
1048 | ||
87b10676 | 1049 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1050 | if (fio_file_done(f)) { |
d5ed68ea | 1051 | f = NULL; |
059e63c0 | 1052 | continue; |
d5ed68ea | 1053 | } |
059e63c0 | 1054 | |
d6aed795 | 1055 | if (!fio_file_open(f)) { |
87b10676 JA |
1056 | int err; |
1057 | ||
002fe734 JA |
1058 | if (td->nr_open_files >= td->o.open_files) |
1059 | return ERR_PTR(-EBUSY); | |
1060 | ||
87b10676 | 1061 | err = td_io_open_file(td, f); |
b5696bfc JA |
1062 | if (err) { |
1063 | dprint(FD_FILE, "error %d on open of %s\n", | |
1064 | err, f->file_name); | |
87c27b45 | 1065 | f = NULL; |
87b10676 | 1066 | continue; |
b5696bfc | 1067 | } |
87b10676 JA |
1068 | opened = 1; |
1069 | } | |
1070 | ||
0b9d69ec JA |
1071 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1072 | f->flags); | |
1c178180 | 1073 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1074 | break; |
1075 | ||
87b10676 JA |
1076 | if (opened) |
1077 | td_io_close_file(td, f); | |
1078 | ||
3d7c391d JA |
1079 | f = NULL; |
1080 | } while (td->next_file != old_next_file); | |
1081 | ||
2ba1c290 | 1082 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1083 | return f; |
1084 | } | |
1085 | ||
7eb36574 | 1086 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1087 | { |
1907dbc6 JA |
1088 | struct fio_file *f; |
1089 | ||
2dc1bbeb | 1090 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1091 | |
b5696bfc | 1092 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1093 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1094 | " nr_files=%d\n", td->nr_open_files, | |
1095 | td->nr_done_files, | |
1096 | td->o.nr_files); | |
bdb4e2e9 | 1097 | return NULL; |
2ba1c290 | 1098 | } |
bdb4e2e9 | 1099 | |
1907dbc6 | 1100 | f = td->file_service_file; |
d6aed795 | 1101 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1102 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1103 | goto out; | |
1104 | if (td->file_service_left--) | |
1105 | goto out; | |
1106 | } | |
1907dbc6 | 1107 | |
a086c257 JA |
1108 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1109 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1110 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1111 | else |
d6aed795 | 1112 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1113 | |
002fe734 JA |
1114 | if (IS_ERR(f)) |
1115 | return f; | |
1116 | ||
1907dbc6 JA |
1117 | td->file_service_file = f; |
1118 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1119 | out: |
0dac421f JA |
1120 | if (f) |
1121 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1122 | else | |
1123 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1124 | return f; |
bdb4e2e9 JA |
1125 | } |
1126 | ||
7eb36574 JA |
1127 | static struct fio_file *get_next_file(struct thread_data *td) |
1128 | { | |
372d8962 | 1129 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1130 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1131 | |
d72be545 JA |
1132 | if (ops->get_next_file) |
1133 | return ops->get_next_file(td); | |
1134 | } | |
7eb36574 JA |
1135 | |
1136 | return __get_next_file(td); | |
1137 | } | |
1138 | ||
002fe734 | 1139 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1140 | { |
1141 | struct fio_file *f; | |
1142 | ||
1143 | do { | |
1144 | f = get_next_file(td); | |
002fe734 JA |
1145 | if (IS_ERR_OR_NULL(f)) |
1146 | return PTR_ERR(f); | |
429f6675 | 1147 | |
429f6675 JA |
1148 | io_u->file = f; |
1149 | get_file(f); | |
1150 | ||
1151 | if (!fill_io_u(td, io_u)) | |
1152 | break; | |
1153 | ||
b5696bfc | 1154 | put_file_log(td, f); |
429f6675 | 1155 | td_io_close_file(td, f); |
b5696bfc | 1156 | io_u->file = NULL; |
d6aed795 | 1157 | fio_file_set_done(f); |
429f6675 | 1158 | td->nr_done_files++; |
0b9d69ec JA |
1159 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1160 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1161 | } while (1); |
1162 | ||
1163 | return 0; | |
1164 | } | |
1165 | ||
3e260a46 JA |
1166 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1167 | unsigned long tusec, unsigned long max_usec) | |
1168 | { | |
1169 | if (!td->error) | |
1170 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1171 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1172 | icd->error = ETIMEDOUT; | |
1173 | } | |
1174 | ||
1175 | static void lat_new_cycle(struct thread_data *td) | |
1176 | { | |
1177 | fio_gettime(&td->latency_ts, NULL); | |
1178 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1179 | td->latency_failed = 0; | |
1180 | } | |
1181 | ||
1182 | /* | |
1183 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1184 | * are at QD=1, then it's time to give up. | |
1185 | */ | |
1186 | static int __lat_target_failed(struct thread_data *td) | |
1187 | { | |
1188 | if (td->latency_qd == 1) | |
1189 | return 1; | |
1190 | ||
1191 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1192 | |
1193 | if (td->latency_qd == td->latency_qd_low) | |
1194 | td->latency_qd_low--; | |
1195 | ||
3e260a46 JA |
1196 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1197 | ||
1198 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1199 | ||
1200 | /* | |
1201 | * When we ramp QD down, quiesce existing IO to prevent | |
1202 | * a storm of ramp downs due to pending higher depth. | |
1203 | */ | |
1204 | io_u_quiesce(td); | |
1205 | lat_new_cycle(td); | |
1206 | return 0; | |
1207 | } | |
1208 | ||
1209 | static int lat_target_failed(struct thread_data *td) | |
1210 | { | |
1211 | if (td->o.latency_percentile.u.f == 100.0) | |
1212 | return __lat_target_failed(td); | |
1213 | ||
1214 | td->latency_failed++; | |
1215 | return 0; | |
1216 | } | |
1217 | ||
1218 | void lat_target_init(struct thread_data *td) | |
1219 | { | |
6bb58215 JA |
1220 | td->latency_end_run = 0; |
1221 | ||
3e260a46 JA |
1222 | if (td->o.latency_target) { |
1223 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1224 | fio_gettime(&td->latency_ts, NULL); | |
1225 | td->latency_qd = 1; | |
1226 | td->latency_qd_high = td->o.iodepth; | |
1227 | td->latency_qd_low = 1; | |
1228 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1229 | } else | |
1230 | td->latency_qd = td->o.iodepth; | |
1231 | } | |
1232 | ||
6bb58215 JA |
1233 | void lat_target_reset(struct thread_data *td) |
1234 | { | |
1235 | if (!td->latency_end_run) | |
1236 | lat_target_init(td); | |
1237 | } | |
1238 | ||
3e260a46 JA |
1239 | static void lat_target_success(struct thread_data *td) |
1240 | { | |
1241 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1242 | struct thread_options *o = &td->o; |
3e260a46 JA |
1243 | |
1244 | td->latency_qd_low = td->latency_qd; | |
1245 | ||
1246 | /* | |
1247 | * If we haven't failed yet, we double up to a failing value instead | |
1248 | * of bisecting from highest possible queue depth. If we have set | |
1249 | * a limit other than td->o.iodepth, bisect between that. | |
1250 | */ | |
6bb58215 | 1251 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1252 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1253 | else | |
1254 | td->latency_qd *= 2; | |
1255 | ||
6bb58215 JA |
1256 | if (td->latency_qd > o->iodepth) |
1257 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1258 | |
1259 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1260 | |
3e260a46 | 1261 | /* |
6bb58215 JA |
1262 | * Same as last one, we are done. Let it run a latency cycle, so |
1263 | * we get only the results from the targeted depth. | |
3e260a46 | 1264 | */ |
6bb58215 JA |
1265 | if (td->latency_qd == qd) { |
1266 | if (td->latency_end_run) { | |
1267 | dprint(FD_RATE, "We are done\n"); | |
1268 | td->done = 1; | |
1269 | } else { | |
1270 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1271 | io_u_quiesce(td); | |
1272 | td->latency_end_run = 1; | |
1273 | reset_all_stats(td); | |
1274 | reset_io_stats(td); | |
1275 | } | |
1276 | } | |
3e260a46 JA |
1277 | |
1278 | lat_new_cycle(td); | |
1279 | } | |
1280 | ||
1281 | /* | |
1282 | * Check if we can bump the queue depth | |
1283 | */ | |
1284 | void lat_target_check(struct thread_data *td) | |
1285 | { | |
1286 | uint64_t usec_window; | |
1287 | uint64_t ios; | |
1288 | double success_ios; | |
1289 | ||
1290 | usec_window = utime_since_now(&td->latency_ts); | |
1291 | if (usec_window < td->o.latency_window) | |
1292 | return; | |
1293 | ||
1294 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1295 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1296 | success_ios *= 100.0; | |
1297 | ||
1298 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1299 | ||
1300 | if (success_ios >= td->o.latency_percentile.u.f) | |
1301 | lat_target_success(td); | |
1302 | else | |
1303 | __lat_target_failed(td); | |
1304 | } | |
1305 | ||
1306 | /* | |
1307 | * If latency target is enabled, we might be ramping up or down and not | |
1308 | * using the full queue depth available. | |
1309 | */ | |
effd6ff0 | 1310 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1311 | { |
1312 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1313 | ||
1314 | if (qempty) | |
1315 | return 1; | |
1316 | if (!td->o.latency_target) | |
1317 | return 0; | |
1318 | ||
1319 | return td->cur_depth >= td->latency_qd; | |
1320 | } | |
429f6675 | 1321 | |
10ba535a JA |
1322 | struct io_u *__get_io_u(struct thread_data *td) |
1323 | { | |
0cae66f6 | 1324 | struct io_u *io_u = NULL; |
10ba535a | 1325 | |
ca09be4b JA |
1326 | if (td->stop_io) |
1327 | return NULL; | |
1328 | ||
e8462bd8 JA |
1329 | td_io_u_lock(td); |
1330 | ||
1331 | again: | |
2ae0b204 JA |
1332 | if (!io_u_rempty(&td->io_u_requeues)) |
1333 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1334 | else if (!queue_full(td)) { |
2ae0b204 | 1335 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1336 | |
225ba9e3 | 1337 | io_u->file = NULL; |
6040dabc | 1338 | io_u->buflen = 0; |
10ba535a | 1339 | io_u->resid = 0; |
d7762cf8 | 1340 | io_u->end_io = NULL; |
755200a3 JA |
1341 | } |
1342 | ||
1343 | if (io_u) { | |
0c6e7517 | 1344 | assert(io_u->flags & IO_U_F_FREE); |
a9da8ab2 | 1345 | io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1346 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1347 | IO_U_F_VER_LIST); | |
0c6e7517 | 1348 | |
755200a3 | 1349 | io_u->error = 0; |
bcd5abfa | 1350 | io_u->acct_ddir = -1; |
10ba535a | 1351 | td->cur_depth++; |
a9da8ab2 JA |
1352 | assert(!(td->flags & TD_F_CHILD)); |
1353 | io_u_set(io_u, IO_U_F_IN_CUR_DEPTH); | |
f9401285 | 1354 | io_u->ipo = NULL; |
a9da8ab2 | 1355 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1356 | /* |
1357 | * We ran out, wait for async verify threads to finish and | |
1358 | * return one | |
1359 | */ | |
a9da8ab2 JA |
1360 | assert(!(td->flags & TD_F_CHILD)); |
1361 | assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock)); | |
1dec3e07 | 1362 | goto again; |
10ba535a JA |
1363 | } |
1364 | ||
e8462bd8 | 1365 | td_io_u_unlock(td); |
10ba535a JA |
1366 | return io_u; |
1367 | } | |
1368 | ||
0d29de83 | 1369 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1370 | { |
d72be545 JA |
1371 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1372 | return 0; | |
1373 | ||
1374 | if (td->trim_entries) { | |
0d29de83 | 1375 | int get_trim = 0; |
10ba535a | 1376 | |
0d29de83 JA |
1377 | if (td->trim_batch) { |
1378 | td->trim_batch--; | |
1379 | get_trim = 1; | |
1380 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1381 | td->last_ddir != DDIR_READ) { | |
1382 | td->trim_batch = td->o.trim_batch; | |
1383 | if (!td->trim_batch) | |
1384 | td->trim_batch = td->o.trim_backlog; | |
1385 | get_trim = 1; | |
1386 | } | |
1387 | ||
1388 | if (get_trim && !get_next_trim(td, io_u)) | |
1389 | return 1; | |
2ba1c290 | 1390 | } |
10ba535a | 1391 | |
0d29de83 JA |
1392 | return 0; |
1393 | } | |
1394 | ||
1395 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1396 | { | |
d72be545 JA |
1397 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1398 | return 0; | |
1399 | ||
1400 | if (td->io_hist_len) { | |
9e144189 JA |
1401 | int get_verify = 0; |
1402 | ||
d1ece0c7 | 1403 | if (td->verify_batch) |
9e144189 | 1404 | get_verify = 1; |
d1ece0c7 | 1405 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1406 | td->last_ddir != DDIR_READ) { |
1407 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1408 | if (!td->verify_batch) |
1409 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1410 | get_verify = 1; |
1411 | } | |
1412 | ||
d1ece0c7 JA |
1413 | if (get_verify && !get_next_verify(td, io_u)) { |
1414 | td->verify_batch--; | |
0d29de83 | 1415 | return 1; |
d1ece0c7 | 1416 | } |
9e144189 JA |
1417 | } |
1418 | ||
0d29de83 JA |
1419 | return 0; |
1420 | } | |
1421 | ||
de789769 JA |
1422 | /* |
1423 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1424 | * easy compressible data for simple de-dupe attempts. Do this for every |
1425 | * 512b block in the range, since that should be the smallest block size | |
1426 | * we can expect from a device. | |
de789769 JA |
1427 | */ |
1428 | static void small_content_scramble(struct io_u *io_u) | |
1429 | { | |
23f394d5 | 1430 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1431 | uint64_t boffset; |
23f394d5 JA |
1432 | unsigned int offset; |
1433 | void *p, *end; | |
de789769 | 1434 | |
23f394d5 JA |
1435 | if (!nr_blocks) |
1436 | return; | |
1437 | ||
1438 | p = io_u->xfer_buf; | |
fba76ee8 | 1439 | boffset = io_u->offset; |
81f0366c | 1440 | io_u->buf_filled_len = 0; |
fad82f76 | 1441 | |
23f394d5 JA |
1442 | for (i = 0; i < nr_blocks; i++) { |
1443 | /* | |
1444 | * Fill the byte offset into a "random" start offset of | |
1445 | * the buffer, given by the product of the usec time | |
1446 | * and the actual offset. | |
1447 | */ | |
fad82f76 | 1448 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1449 | offset &= ~(sizeof(uint64_t) - 1); |
1450 | if (offset >= 512 - sizeof(uint64_t)) | |
1451 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1452 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1453 | |
1454 | end = p + 512 - sizeof(io_u->start_time); | |
1455 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1456 | p += 512; | |
fad82f76 | 1457 | boffset += 512; |
23f394d5 | 1458 | } |
de789769 JA |
1459 | } |
1460 | ||
0d29de83 JA |
1461 | /* |
1462 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1463 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1464 | */ | |
1465 | struct io_u *get_io_u(struct thread_data *td) | |
1466 | { | |
1467 | struct fio_file *f; | |
1468 | struct io_u *io_u; | |
de789769 | 1469 | int do_scramble = 0; |
002fe734 | 1470 | long ret = 0; |
0d29de83 JA |
1471 | |
1472 | io_u = __get_io_u(td); | |
1473 | if (!io_u) { | |
1474 | dprint(FD_IO, "__get_io_u failed\n"); | |
1475 | return NULL; | |
1476 | } | |
1477 | ||
1478 | if (check_get_verify(td, io_u)) | |
1479 | goto out; | |
1480 | if (check_get_trim(td, io_u)) | |
1481 | goto out; | |
1482 | ||
755200a3 JA |
1483 | /* |
1484 | * from a requeue, io_u already setup | |
1485 | */ | |
1486 | if (io_u->file) | |
77f392bf | 1487 | goto out; |
755200a3 | 1488 | |
429f6675 JA |
1489 | /* |
1490 | * If using an iolog, grab next piece if any available. | |
1491 | */ | |
d72be545 | 1492 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1493 | if (read_iolog_get(td, io_u)) |
1494 | goto err_put; | |
2ba1c290 | 1495 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1496 | ret = -EBUSY; |
2ba1c290 | 1497 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1498 | goto err_put; |
2ba1c290 | 1499 | } |
5ec10eaa | 1500 | |
429f6675 | 1501 | f = io_u->file; |
002fe734 JA |
1502 | if (!f) { |
1503 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1504 | goto err_put; | |
1505 | } | |
1506 | ||
d6aed795 | 1507 | assert(fio_file_open(f)); |
97af62ce | 1508 | |
ff58fced | 1509 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1510 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1511 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1512 | goto err_put; |
2ba1c290 | 1513 | } |
10ba535a | 1514 | |
f1dfb668 JA |
1515 | f->last_start[io_u->ddir] = io_u->offset; |
1516 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1517 | |
fd68418e | 1518 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1519 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1520 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1521 | td->o.min_bs[DDIR_WRITE], |
1522 | io_u->xfer_buflen); | |
ff441ae8 | 1523 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1524 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1525 | do_scramble = 1; |
d72be545 | 1526 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1527 | populate_verify_io_u(td, io_u); |
1528 | do_scramble = 0; | |
1529 | } | |
fd68418e | 1530 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1531 | /* |
1532 | * Reset the buf_filled parameters so next time if the | |
1533 | * buffer is used for writes it is refilled. | |
1534 | */ | |
cbe8d756 RR |
1535 | io_u->buf_filled_len = 0; |
1536 | } | |
87dc1ab1 | 1537 | } |
10ba535a | 1538 | |
165faf16 JA |
1539 | /* |
1540 | * Set io data pointers. | |
1541 | */ | |
cec6b55d JA |
1542 | io_u->xfer_buf = io_u->buf; |
1543 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1544 | |
6ac7a331 | 1545 | out: |
0d29de83 | 1546 | assert(io_u->file); |
429f6675 | 1547 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1548 | if (!td->o.disable_slat) |
1549 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1550 | if (do_scramble) |
1551 | small_content_scramble(io_u); | |
429f6675 | 1552 | return io_u; |
36167d82 | 1553 | } |
429f6675 | 1554 | err_put: |
2ba1c290 | 1555 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1556 | put_io_u(td, io_u); |
002fe734 | 1557 | return ERR_PTR(ret); |
10ba535a JA |
1558 | } |
1559 | ||
a9da8ab2 | 1560 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1561 | { |
8b28bd41 | 1562 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1563 | |
8b28bd41 DM |
1564 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1565 | return; | |
5451792e | 1566 | |
709c8313 RE |
1567 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1568 | io_u->file ? " on file " : "", | |
1569 | io_u->file ? io_u->file->file_name : "", | |
1570 | strerror(io_u->error), | |
1571 | io_ddir_name(io_u->ddir), | |
1572 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1573 | |
1574 | if (!td->error) | |
1575 | td_verror(td, io_u->error, "io_u error"); | |
1576 | } | |
1577 | ||
a9da8ab2 JA |
1578 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1579 | { | |
1580 | __io_u_log_error(td, io_u); | |
1581 | if (td->parent) | |
1582 | __io_u_log_error(td, io_u); | |
1583 | } | |
1584 | ||
aba6c951 JA |
1585 | static inline int gtod_reduce(struct thread_data *td) |
1586 | { | |
729fe3af | 1587 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1588 | && td->o.disable_bw; |
aba6c951 JA |
1589 | } |
1590 | ||
c8eeb9df JA |
1591 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1592 | struct io_completion_data *icd, | |
1593 | const enum fio_ddir idx, unsigned int bytes) | |
1594 | { | |
a9da8ab2 | 1595 | const int no_reduce = !gtod_reduce(td); |
24d23ca7 | 1596 | unsigned long lusec = 0; |
c8eeb9df | 1597 | |
a9da8ab2 | 1598 | if (no_reduce) |
c8eeb9df JA |
1599 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1600 | ||
1601 | if (!td->o.disable_lat) { | |
1602 | unsigned long tusec; | |
1603 | ||
1604 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1605 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1606 | |
d4afedfd JA |
1607 | if (td->flags & TD_F_PROFILE_OPS) { |
1608 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1609 | ||
1610 | if (ops->io_u_lat) | |
1611 | icd->error = ops->io_u_lat(td, tusec); | |
1612 | } | |
1613 | ||
3e260a46 JA |
1614 | if (td->o.max_latency && tusec > td->o.max_latency) |
1615 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1616 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1617 | if (lat_target_failed(td)) | |
1618 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1619 | } |
c8eeb9df JA |
1620 | } |
1621 | ||
1622 | if (!td->o.disable_clat) { | |
ae588852 | 1623 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1624 | io_u_mark_latency(td, lusec); |
1625 | } | |
1626 | ||
a9da8ab2 JA |
1627 | if (td->parent) |
1628 | td = td->parent; | |
1629 | ||
c8eeb9df JA |
1630 | if (!td->o.disable_bw) |
1631 | add_bw_sample(td, idx, bytes, &icd->time); | |
1632 | ||
a9da8ab2 | 1633 | if (no_reduce) |
aba6c951 | 1634 | add_iops_sample(td, idx, bytes, &icd->time); |
66347cfa DE |
1635 | |
1636 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1637 | uint32_t *info = io_u_block_info(td, io_u); | |
1638 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1639 | if (io_u->ddir == DDIR_TRIM) { | |
1640 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1641 | BLOCK_INFO_TRIMS(*info) + 1); | |
1642 | } else if (io_u->ddir == DDIR_WRITE) { | |
1643 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1644 | *info); | |
1645 | } | |
1646 | } | |
1647 | } | |
c8eeb9df JA |
1648 | } |
1649 | ||
1b8dbf25 SL |
1650 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1651 | { | |
1ae83d45 JA |
1652 | uint64_t secs, remainder, bps, bytes; |
1653 | ||
a9da8ab2 | 1654 | assert(!(td->flags & TD_F_CHILD)); |
1b8dbf25 SL |
1655 | bytes = td->this_io_bytes[ddir]; |
1656 | bps = td->rate_bps[ddir]; | |
1657 | secs = bytes / bps; | |
1658 | remainder = bytes % bps; | |
1659 | return remainder * 1000000 / bps + secs * 1000000; | |
1660 | } | |
1661 | ||
f8b0bd10 | 1662 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1663 | struct io_completion_data *icd) |
10ba535a | 1664 | { |
f8b0bd10 JA |
1665 | struct io_u *io_u = *io_u_ptr; |
1666 | enum fio_ddir ddir = io_u->ddir; | |
1667 | struct fio_file *f = io_u->file; | |
10ba535a | 1668 | |
2ba1c290 JA |
1669 | dprint_io_u(io_u, "io complete"); |
1670 | ||
0c6e7517 | 1671 | assert(io_u->flags & IO_U_F_FLIGHT); |
a9da8ab2 | 1672 | io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1673 | |
1674 | /* | |
1675 | * Mark IO ok to verify | |
1676 | */ | |
1677 | if (io_u->ipo) { | |
890b6656 JA |
1678 | /* |
1679 | * Remove errored entry from the verification list | |
1680 | */ | |
1681 | if (io_u->error) | |
1682 | unlog_io_piece(td, io_u); | |
1683 | else { | |
1684 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1685 | write_barrier(); | |
1686 | } | |
f9401285 JA |
1687 | } |
1688 | ||
f8b0bd10 | 1689 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1690 | td->last_was_sync = 1; |
44f29692 JA |
1691 | if (f) { |
1692 | f->first_write = -1ULL; | |
1693 | f->last_write = -1ULL; | |
1694 | } | |
87dc1ab1 JA |
1695 | return; |
1696 | } | |
1697 | ||
1698 | td->last_was_sync = 0; | |
f8b0bd10 | 1699 | td->last_ddir = ddir; |
87dc1ab1 | 1700 | |
f8b0bd10 | 1701 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1702 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1703 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1704 | int ret; |
10ba535a | 1705 | |
f8b0bd10 JA |
1706 | td->io_blocks[ddir]++; |
1707 | td->this_io_blocks[ddir]++; | |
1708 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1709 | |
1710 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1711 | td->this_io_bytes[ddir] += bytes; |
1712 | ||
ca09be4b JA |
1713 | if (ddir == DDIR_WRITE) { |
1714 | if (f) { | |
1715 | if (f->first_write == -1ULL || | |
1716 | io_u->offset < f->first_write) | |
1717 | f->first_write = io_u->offset; | |
1718 | if (f->last_write == -1ULL || | |
1719 | ((io_u->offset + bytes) > f->last_write)) | |
1720 | f->last_write = io_u->offset + bytes; | |
1721 | } | |
1722 | if (td->last_write_comp) { | |
1723 | int idx = td->last_write_idx++; | |
1724 | ||
1725 | td->last_write_comp[idx] = io_u->offset; | |
1726 | if (td->last_write_idx == td->o.iodepth) | |
1727 | td->last_write_idx = 0; | |
1728 | } | |
44f29692 JA |
1729 | } |
1730 | ||
6b1190fd SL |
1731 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1732 | td->runstate == TD_VERIFYING)) { | |
a9da8ab2 JA |
1733 | struct thread_data *__td = td; |
1734 | ||
f8b0bd10 | 1735 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1736 | |
a9da8ab2 JA |
1737 | if (td->parent) |
1738 | __td = td->parent; | |
1739 | ||
1740 | if (__should_check_rate(__td, ddir)) { | |
1741 | __td->rate_pending_usleep[ddir] = | |
1742 | (usec_for_io(__td, ddir) - | |
1743 | utime_since_now(&__td->start)); | |
b23b6a2f | 1744 | } |
f8b0bd10 | 1745 | if (ddir != DDIR_TRIM && |
a9da8ab2 JA |
1746 | __should_check_rate(__td, oddir)) { |
1747 | __td->rate_pending_usleep[oddir] = | |
1748 | (usec_for_io(__td, oddir) - | |
1749 | utime_since_now(&__td->start)); | |
f8b0bd10 | 1750 | } |
721938ae | 1751 | } |
10ba535a | 1752 | |
f8b0bd10 | 1753 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1754 | |
d7762cf8 | 1755 | if (io_u->end_io) { |
f8b0bd10 JA |
1756 | ret = io_u->end_io(td, io_u_ptr); |
1757 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1758 | if (ret && !icd->error) |
1759 | icd->error = ret; | |
1760 | } | |
ff58fced | 1761 | } else if (io_u->error) { |
10ba535a | 1762 | icd->error = io_u->error; |
5451792e JA |
1763 | io_u_log_error(td, io_u); |
1764 | } | |
8b28bd41 | 1765 | if (icd->error) { |
f8b0bd10 JA |
1766 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1767 | ||
8b28bd41 DM |
1768 | if (!td_non_fatal_error(td, eb, icd->error)) |
1769 | return; | |
f8b0bd10 | 1770 | |
f2bba182 RR |
1771 | /* |
1772 | * If there is a non_fatal error, then add to the error count | |
1773 | * and clear all the errors. | |
1774 | */ | |
1775 | update_error_count(td, icd->error); | |
1776 | td_clear_error(td); | |
1777 | icd->error = 0; | |
f8b0bd10 JA |
1778 | if (io_u) |
1779 | io_u->error = 0; | |
f2bba182 | 1780 | } |
10ba535a JA |
1781 | } |
1782 | ||
9520ebb9 JA |
1783 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1784 | int nr) | |
10ba535a | 1785 | { |
6eaf09d6 | 1786 | int ddir; |
aba6c951 JA |
1787 | |
1788 | if (!gtod_reduce(td)) | |
9520ebb9 | 1789 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1790 | |
3af6ef39 JA |
1791 | icd->nr = nr; |
1792 | ||
10ba535a | 1793 | icd->error = 0; |
6eaf09d6 SL |
1794 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1795 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1796 | } |
1797 | ||
97601024 JA |
1798 | static void ios_completed(struct thread_data *td, |
1799 | struct io_completion_data *icd) | |
36167d82 JA |
1800 | { |
1801 | struct io_u *io_u; | |
1802 | int i; | |
1803 | ||
10ba535a JA |
1804 | for (i = 0; i < icd->nr; i++) { |
1805 | io_u = td->io_ops->event(td, i); | |
1806 | ||
f8b0bd10 | 1807 | io_completed(td, &io_u, icd); |
e8462bd8 | 1808 | |
f8b0bd10 | 1809 | if (io_u) |
e8462bd8 | 1810 | put_io_u(td, io_u); |
10ba535a JA |
1811 | } |
1812 | } | |
97601024 | 1813 | |
e7e6cfb4 JA |
1814 | /* |
1815 | * Complete a single io_u for the sync engines. | |
1816 | */ | |
55312f9f | 1817 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1818 | { |
1819 | struct io_completion_data icd; | |
55312f9f | 1820 | int ddir; |
97601024 | 1821 | |
9520ebb9 | 1822 | init_icd(td, &icd, 1); |
f8b0bd10 | 1823 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1824 | |
f8b0bd10 | 1825 | if (io_u) |
e8462bd8 | 1826 | put_io_u(td, io_u); |
97601024 | 1827 | |
581e7141 JA |
1828 | if (icd.error) { |
1829 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1830 | return -1; | |
1831 | } | |
97601024 | 1832 | |
55312f9f JA |
1833 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1834 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1835 | |
1836 | return 0; | |
97601024 JA |
1837 | } |
1838 | ||
e7e6cfb4 JA |
1839 | /* |
1840 | * Called to complete min_events number of io for the async engines. | |
1841 | */ | |
55312f9f | 1842 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 1843 | { |
97601024 | 1844 | struct io_completion_data icd; |
00de55ef | 1845 | struct timespec *tvp = NULL; |
55312f9f | 1846 | int ret, ddir; |
4d06a338 | 1847 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1848 | |
4950421a | 1849 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1850 | |
4950421a | 1851 | if (!min_evts) |
00de55ef | 1852 | tvp = &ts; |
5fb4b366 RE |
1853 | else if (min_evts > td->cur_depth) |
1854 | min_evts = td->cur_depth; | |
97601024 | 1855 | |
4950421a | 1856 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1857 | if (ret < 0) { |
e1161c32 | 1858 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1859 | return ret; |
1860 | } else if (!ret) | |
1861 | return ret; | |
1862 | ||
9520ebb9 | 1863 | init_icd(td, &icd, ret); |
97601024 | 1864 | ios_completed(td, &icd); |
581e7141 JA |
1865 | if (icd.error) { |
1866 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1867 | return -1; | |
1868 | } | |
97601024 | 1869 | |
55312f9f JA |
1870 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1871 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1872 | |
1873 | return 0; | |
97601024 | 1874 | } |
7e77dd02 JA |
1875 | |
1876 | /* | |
1877 | * Call when io_u is really queued, to update the submission latency. | |
1878 | */ | |
1879 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1880 | { | |
9520ebb9 JA |
1881 | if (!td->o.disable_slat) { |
1882 | unsigned long slat_time; | |
7e77dd02 | 1883 | |
9520ebb9 | 1884 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1885 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1886 | io_u->offset); | |
9520ebb9 | 1887 | } |
7e77dd02 | 1888 | } |
433afcb4 | 1889 | |
5c94b008 JA |
1890 | /* |
1891 | * See if we should reuse the last seed, if dedupe is enabled | |
1892 | */ | |
9451b93e | 1893 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
1894 | { |
1895 | unsigned int v; | |
1896 | unsigned long r; | |
1897 | ||
1898 | if (!td->o.dedupe_percentage) | |
1899 | return &td->buf_state; | |
732eedd0 | 1900 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
1901 | frand_copy(&td->buf_state_prev, &td->buf_state); |
1902 | return &td->buf_state; | |
732eedd0 | 1903 | } |
5c94b008 JA |
1904 | |
1905 | r = __rand(&td->dedupe_state); | |
1906 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1907 | ||
1908 | if (v <= td->o.dedupe_percentage) | |
1909 | return &td->buf_state_prev; | |
1910 | ||
1911 | return &td->buf_state; | |
1912 | } | |
1913 | ||
9451b93e | 1914 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 1915 | { |
9451b93e JA |
1916 | if (td->o.dedupe_percentage == 100) |
1917 | frand_copy(rs, &td->buf_state_prev); | |
1918 | else if (rs == &td->buf_state) | |
5c94b008 JA |
1919 | frand_copy(&td->buf_state_prev, rs); |
1920 | } | |
1921 | ||
cc86c395 JA |
1922 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1923 | unsigned int max_bs) | |
5973cafb | 1924 | { |
d1af2894 JA |
1925 | struct thread_options *o = &td->o; |
1926 | ||
4eff3e57 | 1927 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 1928 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1929 | struct frand_state *rs; |
1066358a | 1930 | unsigned int left = max_bs; |
5c94b008 | 1931 | |
1066358a | 1932 | do { |
9451b93e | 1933 | rs = get_buf_state(td); |
9c42684e | 1934 | |
1066358a | 1935 | min_write = min(min_write, left); |
f97a43a1 | 1936 | |
1066358a JA |
1937 | if (perc) { |
1938 | unsigned int seg = min_write; | |
cc86c395 | 1939 | |
1066358a JA |
1940 | seg = min(min_write, td->o.compress_chunk); |
1941 | if (!seg) | |
1942 | seg = min_write; | |
1943 | ||
1944 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1945 | min_write, o->buffer_pattern, |
1946 | o->buffer_pattern_bytes); | |
1066358a JA |
1947 | } else |
1948 | fill_random_buf(rs, buf, min_write); | |
1949 | ||
1950 | buf += min_write; | |
1951 | left -= min_write; | |
9451b93e | 1952 | save_buf_state(td, rs); |
1066358a | 1953 | } while (left); |
d1af2894 JA |
1954 | } else if (o->buffer_pattern_bytes) |
1955 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 1956 | else if (o->zero_buffers) |
cc86c395 | 1957 | memset(buf, 0, max_bs); |
999d245e | 1958 | else |
9451b93e | 1959 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
1960 | } |
1961 | ||
1962 | /* | |
1963 | * "randomly" fill the buffer contents | |
1964 | */ | |
1965 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1966 | unsigned int min_write, unsigned int max_bs) | |
1967 | { | |
1968 | io_u->buf_filled_len = 0; | |
1969 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1970 | } |