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; | |
211c9b89 JA |
329 | io_u->flags |= IO_U_F_BUSY_OK; |
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 JA |
338 | } else { |
339 | io_u->flags |= 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 JA |
487 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
488 | buflen = (buflen + minbs - 1) & ~(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 | ||
555 | ret = io_u_queued_complete(td, 1, NULL); | |
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 | ||
002e7183 | 594 | io_u_quiesce(td); |
78c1eda5 | 595 | |
9cc80b6d | 596 | usec = usec_sleep(td, usec); |
581e7141 JA |
597 | |
598 | td->rate_pending_usleep[ddir] -= usec; | |
599 | ||
600 | odir = ddir ^ 1; | |
601 | if (td_rw(td) && __should_check_rate(td, odir)) | |
602 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 603 | |
581e7141 JA |
604 | return ddir; |
605 | } | |
606 | ||
10ba535a JA |
607 | /* |
608 | * Return the data direction for the next io_u. If the job is a | |
609 | * mixed read/write workload, check the rwmix cycle and switch if | |
610 | * necessary. | |
611 | */ | |
1e97cce9 | 612 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 613 | { |
581e7141 JA |
614 | enum fio_ddir ddir; |
615 | ||
5f9099ea JA |
616 | /* |
617 | * see if it's time to fsync | |
618 | */ | |
619 | if (td->o.fsync_blocks && | |
620 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
621 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
622 | return DDIR_SYNC; | |
623 | ||
624 | /* | |
625 | * see if it's time to fdatasync | |
626 | */ | |
627 | if (td->o.fdatasync_blocks && | |
628 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
629 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
630 | return DDIR_DATASYNC; | |
631 | ||
44f29692 JA |
632 | /* |
633 | * see if it's time to sync_file_range | |
634 | */ | |
635 | if (td->sync_file_range_nr && | |
636 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
637 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
638 | return DDIR_SYNC_FILE_RANGE; | |
639 | ||
10ba535a | 640 | if (td_rw(td)) { |
10ba535a JA |
641 | /* |
642 | * Check if it's time to seed a new data direction. | |
643 | */ | |
e4928662 | 644 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
645 | /* |
646 | * Put a top limit on how many bytes we do for | |
647 | * one data direction, to avoid overflowing the | |
648 | * ranges too much | |
649 | */ | |
650 | ddir = get_rand_ddir(td); | |
e47f799f JA |
651 | |
652 | if (ddir != td->rwmix_ddir) | |
653 | set_rwmix_bytes(td); | |
654 | ||
655 | td->rwmix_ddir = ddir; | |
10ba535a | 656 | } |
581e7141 | 657 | ddir = td->rwmix_ddir; |
10ba535a | 658 | } else if (td_read(td)) |
581e7141 | 659 | ddir = DDIR_READ; |
6eaf09d6 | 660 | else if (td_write(td)) |
581e7141 | 661 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
662 | else |
663 | ddir = DDIR_TRIM; | |
581e7141 JA |
664 | |
665 | td->rwmix_ddir = rate_ddir(td, ddir); | |
666 | return td->rwmix_ddir; | |
10ba535a JA |
667 | } |
668 | ||
1ef2b6be JA |
669 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
670 | { | |
bcd5abfa | 671 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
672 | |
673 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
674 | td->o.barrier_blocks && | |
675 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
676 | td->io_issues[DDIR_WRITE]) | |
677 | io_u->flags |= IO_U_F_BARRIER; | |
678 | } | |
679 | ||
e8462bd8 | 680 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 681 | { |
71b84caa | 682 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
683 | |
684 | if (ret) | |
685 | td_verror(td, ret, "file close"); | |
686 | } | |
687 | ||
10ba535a JA |
688 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
689 | { | |
e8462bd8 JA |
690 | td_io_u_lock(td); |
691 | ||
f8b0bd10 | 692 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 693 | put_file_log(td, io_u->file); |
f8b0bd10 | 694 | |
10ba535a | 695 | io_u->file = NULL; |
d7ee2a7d SL |
696 | io_u->flags |= IO_U_F_FREE; |
697 | ||
0c41214f RR |
698 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
699 | td->cur_depth--; | |
2ae0b204 | 700 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
701 | td_io_u_unlock(td); |
702 | td_io_u_free_notify(td); | |
10ba535a JA |
703 | } |
704 | ||
f2bba182 RR |
705 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
706 | { | |
707 | io_u->flags &= ~IO_U_F_FLIGHT; | |
708 | put_io_u(td, io_u); | |
709 | } | |
710 | ||
755200a3 JA |
711 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
712 | { | |
713 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 714 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 715 | |
465221b0 JA |
716 | dprint(FD_IO, "requeue %p\n", __io_u); |
717 | ||
e8462bd8 JA |
718 | td_io_u_lock(td); |
719 | ||
4d2e0f49 | 720 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
721 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
722 | td->io_issues[ddir]--; | |
5ec10eaa | 723 | |
4d2e0f49 | 724 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
725 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
726 | td->cur_depth--; | |
2ae0b204 JA |
727 | |
728 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 729 | td_io_u_unlock(td); |
755200a3 JA |
730 | *io_u = NULL; |
731 | } | |
732 | ||
9bf2061e | 733 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 734 | { |
6aca9b3d JA |
735 | unsigned int is_random; |
736 | ||
b4c5e1ac JA |
737 | if (td->io_ops->flags & FIO_NOIO) |
738 | goto out; | |
739 | ||
1ef2b6be | 740 | set_rw_ddir(td, io_u); |
5f9099ea | 741 | |
87dc1ab1 | 742 | /* |
ff58fced | 743 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 744 | */ |
ff58fced | 745 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 746 | goto out; |
a00735e6 | 747 | |
48f5abd3 JA |
748 | /* |
749 | * See if it's time to switch to a new zone | |
750 | */ | |
13af05ae | 751 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
752 | struct fio_file *f = io_u->file; |
753 | ||
48f5abd3 | 754 | td->zone_bytes = 0; |
4c8be5b1 JA |
755 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
756 | ||
757 | /* | |
758 | * Wrap from the beginning, if we exceed the file size | |
759 | */ | |
760 | if (f->file_offset >= f->real_file_size) | |
761 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 762 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
763 | td->io_skip_bytes += td->o.zone_skip; |
764 | } | |
765 | ||
10ba535a | 766 | /* |
c685b5b2 JA |
767 | * No log, let the seq/rand engine retrieve the next buflen and |
768 | * position. | |
10ba535a | 769 | */ |
6aca9b3d | 770 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 771 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 772 | return 1; |
2ba1c290 | 773 | } |
10ba535a | 774 | |
6aca9b3d | 775 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
776 | if (!io_u->buflen) { |
777 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 778 | return 1; |
2ba1c290 | 779 | } |
bca4ed4d | 780 | |
2ba1c290 JA |
781 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
782 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
783 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
784 | (unsigned long long) io_u->offset, io_u->buflen, | |
785 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 786 | return 1; |
2ba1c290 | 787 | } |
6a5e6884 | 788 | |
bca4ed4d JA |
789 | /* |
790 | * mark entry before potentially trimming io_u | |
791 | */ | |
303032ae | 792 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 793 | mark_random_map(td, io_u); |
bca4ed4d | 794 | |
c38e9468 | 795 | out: |
2ba1c290 | 796 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 797 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 798 | return 0; |
10ba535a JA |
799 | } |
800 | ||
838bc709 JA |
801 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
802 | { | |
2b13e716 | 803 | int idx = 0; |
838bc709 JA |
804 | |
805 | switch (nr) { | |
806 | default: | |
2b13e716 | 807 | idx = 6; |
838bc709 JA |
808 | break; |
809 | case 33 ... 64: | |
2b13e716 | 810 | idx = 5; |
838bc709 JA |
811 | break; |
812 | case 17 ... 32: | |
2b13e716 | 813 | idx = 4; |
838bc709 JA |
814 | break; |
815 | case 9 ... 16: | |
2b13e716 | 816 | idx = 3; |
838bc709 JA |
817 | break; |
818 | case 5 ... 8: | |
2b13e716 | 819 | idx = 2; |
838bc709 JA |
820 | break; |
821 | case 1 ... 4: | |
2b13e716 | 822 | idx = 1; |
838bc709 JA |
823 | case 0: |
824 | break; | |
825 | } | |
826 | ||
2b13e716 | 827 | map[idx]++; |
838bc709 JA |
828 | } |
829 | ||
830 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
831 | { | |
832 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
833 | td->ts.total_submit++; | |
834 | } | |
835 | ||
836 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
837 | { | |
838 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
839 | td->ts.total_complete++; | |
840 | } | |
841 | ||
d8005759 | 842 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 843 | { |
2b13e716 | 844 | int idx = 0; |
71619dc2 JA |
845 | |
846 | switch (td->cur_depth) { | |
847 | default: | |
2b13e716 | 848 | idx = 6; |
a783e61a | 849 | break; |
71619dc2 | 850 | case 32 ... 63: |
2b13e716 | 851 | idx = 5; |
a783e61a | 852 | break; |
71619dc2 | 853 | case 16 ... 31: |
2b13e716 | 854 | idx = 4; |
a783e61a | 855 | break; |
71619dc2 | 856 | case 8 ... 15: |
2b13e716 | 857 | idx = 3; |
a783e61a | 858 | break; |
71619dc2 | 859 | case 4 ... 7: |
2b13e716 | 860 | idx = 2; |
a783e61a | 861 | break; |
71619dc2 | 862 | case 2 ... 3: |
2b13e716 | 863 | idx = 1; |
71619dc2 JA |
864 | case 1: |
865 | break; | |
866 | } | |
867 | ||
2b13e716 | 868 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
869 | } |
870 | ||
04a0feae JA |
871 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
872 | { | |
2b13e716 | 873 | int idx = 0; |
04a0feae JA |
874 | |
875 | assert(usec < 1000); | |
876 | ||
877 | switch (usec) { | |
878 | case 750 ... 999: | |
2b13e716 | 879 | idx = 9; |
04a0feae JA |
880 | break; |
881 | case 500 ... 749: | |
2b13e716 | 882 | idx = 8; |
04a0feae JA |
883 | break; |
884 | case 250 ... 499: | |
2b13e716 | 885 | idx = 7; |
04a0feae JA |
886 | break; |
887 | case 100 ... 249: | |
2b13e716 | 888 | idx = 6; |
04a0feae JA |
889 | break; |
890 | case 50 ... 99: | |
2b13e716 | 891 | idx = 5; |
04a0feae JA |
892 | break; |
893 | case 20 ... 49: | |
2b13e716 | 894 | idx = 4; |
04a0feae JA |
895 | break; |
896 | case 10 ... 19: | |
2b13e716 | 897 | idx = 3; |
04a0feae JA |
898 | break; |
899 | case 4 ... 9: | |
2b13e716 | 900 | idx = 2; |
04a0feae JA |
901 | break; |
902 | case 2 ... 3: | |
2b13e716 | 903 | idx = 1; |
04a0feae JA |
904 | case 0 ... 1: |
905 | break; | |
906 | } | |
907 | ||
2b13e716 JA |
908 | assert(idx < FIO_IO_U_LAT_U_NR); |
909 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
910 | } |
911 | ||
912 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 913 | { |
2b13e716 | 914 | int idx = 0; |
ec118304 JA |
915 | |
916 | switch (msec) { | |
917 | default: | |
2b13e716 | 918 | idx = 11; |
04a0feae | 919 | break; |
8abdce66 | 920 | case 1000 ... 1999: |
2b13e716 | 921 | idx = 10; |
04a0feae | 922 | break; |
8abdce66 | 923 | case 750 ... 999: |
2b13e716 | 924 | idx = 9; |
04a0feae | 925 | break; |
8abdce66 | 926 | case 500 ... 749: |
2b13e716 | 927 | idx = 8; |
04a0feae | 928 | break; |
8abdce66 | 929 | case 250 ... 499: |
2b13e716 | 930 | idx = 7; |
04a0feae | 931 | break; |
8abdce66 | 932 | case 100 ... 249: |
2b13e716 | 933 | idx = 6; |
04a0feae | 934 | break; |
8abdce66 | 935 | case 50 ... 99: |
2b13e716 | 936 | idx = 5; |
04a0feae | 937 | break; |
8abdce66 | 938 | case 20 ... 49: |
2b13e716 | 939 | idx = 4; |
04a0feae | 940 | break; |
8abdce66 | 941 | case 10 ... 19: |
2b13e716 | 942 | idx = 3; |
04a0feae | 943 | break; |
8abdce66 | 944 | case 4 ... 9: |
2b13e716 | 945 | idx = 2; |
04a0feae | 946 | break; |
ec118304 | 947 | case 2 ... 3: |
2b13e716 | 948 | idx = 1; |
ec118304 JA |
949 | case 0 ... 1: |
950 | break; | |
951 | } | |
952 | ||
2b13e716 JA |
953 | assert(idx < FIO_IO_U_LAT_M_NR); |
954 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
955 | } |
956 | ||
957 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
958 | { | |
959 | if (usec < 1000) | |
960 | io_u_mark_lat_usec(td, usec); | |
961 | else | |
962 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
963 | } |
964 | ||
0aabe160 JA |
965 | /* |
966 | * Get next file to service by choosing one at random | |
967 | */ | |
2cc52930 JA |
968 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
969 | enum fio_file_flags goodf, | |
d6aed795 | 970 | enum fio_file_flags badf) |
0aabe160 | 971 | { |
0aabe160 | 972 | struct fio_file *f; |
1c178180 | 973 | int fno; |
0aabe160 JA |
974 | |
975 | do { | |
87b10676 | 976 | int opened = 0; |
1294c3ec | 977 | unsigned long r; |
4c07ad86 | 978 | |
d6b72507 | 979 | r = __rand(&td->next_file_state); |
54a21917 | 980 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 981 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 982 | |
126d65c6 | 983 | f = td->files[fno]; |
d6aed795 | 984 | if (fio_file_done(f)) |
059e63c0 | 985 | continue; |
1c178180 | 986 | |
d6aed795 | 987 | if (!fio_file_open(f)) { |
87b10676 JA |
988 | int err; |
989 | ||
002fe734 JA |
990 | if (td->nr_open_files >= td->o.open_files) |
991 | return ERR_PTR(-EBUSY); | |
992 | ||
87b10676 JA |
993 | err = td_io_open_file(td, f); |
994 | if (err) | |
995 | continue; | |
996 | opened = 1; | |
997 | } | |
998 | ||
2ba1c290 JA |
999 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1000 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1001 | return f; |
2ba1c290 | 1002 | } |
87b10676 JA |
1003 | if (opened) |
1004 | td_io_close_file(td, f); | |
0aabe160 JA |
1005 | } while (1); |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * Get next file to service by doing round robin between all available ones | |
1010 | */ | |
1c178180 JA |
1011 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1012 | int badf) | |
3d7c391d JA |
1013 | { |
1014 | unsigned int old_next_file = td->next_file; | |
1015 | struct fio_file *f; | |
1016 | ||
1017 | do { | |
87b10676 JA |
1018 | int opened = 0; |
1019 | ||
126d65c6 | 1020 | f = td->files[td->next_file]; |
3d7c391d JA |
1021 | |
1022 | td->next_file++; | |
2dc1bbeb | 1023 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1024 | td->next_file = 0; |
1025 | ||
87b10676 | 1026 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1027 | if (fio_file_done(f)) { |
d5ed68ea | 1028 | f = NULL; |
059e63c0 | 1029 | continue; |
d5ed68ea | 1030 | } |
059e63c0 | 1031 | |
d6aed795 | 1032 | if (!fio_file_open(f)) { |
87b10676 JA |
1033 | int err; |
1034 | ||
002fe734 JA |
1035 | if (td->nr_open_files >= td->o.open_files) |
1036 | return ERR_PTR(-EBUSY); | |
1037 | ||
87b10676 | 1038 | err = td_io_open_file(td, f); |
b5696bfc JA |
1039 | if (err) { |
1040 | dprint(FD_FILE, "error %d on open of %s\n", | |
1041 | err, f->file_name); | |
87c27b45 | 1042 | f = NULL; |
87b10676 | 1043 | continue; |
b5696bfc | 1044 | } |
87b10676 JA |
1045 | opened = 1; |
1046 | } | |
1047 | ||
0b9d69ec JA |
1048 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1049 | f->flags); | |
1c178180 | 1050 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1051 | break; |
1052 | ||
87b10676 JA |
1053 | if (opened) |
1054 | td_io_close_file(td, f); | |
1055 | ||
3d7c391d JA |
1056 | f = NULL; |
1057 | } while (td->next_file != old_next_file); | |
1058 | ||
2ba1c290 | 1059 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1060 | return f; |
1061 | } | |
1062 | ||
7eb36574 | 1063 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1064 | { |
1907dbc6 JA |
1065 | struct fio_file *f; |
1066 | ||
2dc1bbeb | 1067 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1068 | |
b5696bfc | 1069 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1070 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1071 | " nr_files=%d\n", td->nr_open_files, | |
1072 | td->nr_done_files, | |
1073 | td->o.nr_files); | |
bdb4e2e9 | 1074 | return NULL; |
2ba1c290 | 1075 | } |
bdb4e2e9 | 1076 | |
1907dbc6 | 1077 | f = td->file_service_file; |
d6aed795 | 1078 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1079 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1080 | goto out; | |
1081 | if (td->file_service_left--) | |
1082 | goto out; | |
1083 | } | |
1907dbc6 | 1084 | |
a086c257 JA |
1085 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1086 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1087 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1088 | else |
d6aed795 | 1089 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1090 | |
002fe734 JA |
1091 | if (IS_ERR(f)) |
1092 | return f; | |
1093 | ||
1907dbc6 JA |
1094 | td->file_service_file = f; |
1095 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1096 | out: |
0dac421f JA |
1097 | if (f) |
1098 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1099 | else | |
1100 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1101 | return f; |
bdb4e2e9 JA |
1102 | } |
1103 | ||
7eb36574 JA |
1104 | static struct fio_file *get_next_file(struct thread_data *td) |
1105 | { | |
372d8962 | 1106 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1107 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1108 | |
d72be545 JA |
1109 | if (ops->get_next_file) |
1110 | return ops->get_next_file(td); | |
1111 | } | |
7eb36574 JA |
1112 | |
1113 | return __get_next_file(td); | |
1114 | } | |
1115 | ||
002fe734 | 1116 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1117 | { |
1118 | struct fio_file *f; | |
1119 | ||
1120 | do { | |
1121 | f = get_next_file(td); | |
002fe734 JA |
1122 | if (IS_ERR_OR_NULL(f)) |
1123 | return PTR_ERR(f); | |
429f6675 | 1124 | |
429f6675 JA |
1125 | io_u->file = f; |
1126 | get_file(f); | |
1127 | ||
1128 | if (!fill_io_u(td, io_u)) | |
1129 | break; | |
1130 | ||
b5696bfc | 1131 | put_file_log(td, f); |
429f6675 | 1132 | td_io_close_file(td, f); |
b5696bfc | 1133 | io_u->file = NULL; |
d6aed795 | 1134 | fio_file_set_done(f); |
429f6675 | 1135 | td->nr_done_files++; |
0b9d69ec JA |
1136 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1137 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1138 | } while (1); |
1139 | ||
1140 | return 0; | |
1141 | } | |
1142 | ||
3e260a46 JA |
1143 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1144 | unsigned long tusec, unsigned long max_usec) | |
1145 | { | |
1146 | if (!td->error) | |
1147 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1148 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1149 | icd->error = ETIMEDOUT; | |
1150 | } | |
1151 | ||
1152 | static void lat_new_cycle(struct thread_data *td) | |
1153 | { | |
1154 | fio_gettime(&td->latency_ts, NULL); | |
1155 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1156 | td->latency_failed = 0; | |
1157 | } | |
1158 | ||
1159 | /* | |
1160 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1161 | * are at QD=1, then it's time to give up. | |
1162 | */ | |
1163 | static int __lat_target_failed(struct thread_data *td) | |
1164 | { | |
1165 | if (td->latency_qd == 1) | |
1166 | return 1; | |
1167 | ||
1168 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1169 | |
1170 | if (td->latency_qd == td->latency_qd_low) | |
1171 | td->latency_qd_low--; | |
1172 | ||
3e260a46 JA |
1173 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1174 | ||
1175 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1176 | ||
1177 | /* | |
1178 | * When we ramp QD down, quiesce existing IO to prevent | |
1179 | * a storm of ramp downs due to pending higher depth. | |
1180 | */ | |
1181 | io_u_quiesce(td); | |
1182 | lat_new_cycle(td); | |
1183 | return 0; | |
1184 | } | |
1185 | ||
1186 | static int lat_target_failed(struct thread_data *td) | |
1187 | { | |
1188 | if (td->o.latency_percentile.u.f == 100.0) | |
1189 | return __lat_target_failed(td); | |
1190 | ||
1191 | td->latency_failed++; | |
1192 | return 0; | |
1193 | } | |
1194 | ||
1195 | void lat_target_init(struct thread_data *td) | |
1196 | { | |
6bb58215 JA |
1197 | td->latency_end_run = 0; |
1198 | ||
3e260a46 JA |
1199 | if (td->o.latency_target) { |
1200 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1201 | fio_gettime(&td->latency_ts, NULL); | |
1202 | td->latency_qd = 1; | |
1203 | td->latency_qd_high = td->o.iodepth; | |
1204 | td->latency_qd_low = 1; | |
1205 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1206 | } else | |
1207 | td->latency_qd = td->o.iodepth; | |
1208 | } | |
1209 | ||
6bb58215 JA |
1210 | void lat_target_reset(struct thread_data *td) |
1211 | { | |
1212 | if (!td->latency_end_run) | |
1213 | lat_target_init(td); | |
1214 | } | |
1215 | ||
3e260a46 JA |
1216 | static void lat_target_success(struct thread_data *td) |
1217 | { | |
1218 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1219 | struct thread_options *o = &td->o; |
3e260a46 JA |
1220 | |
1221 | td->latency_qd_low = td->latency_qd; | |
1222 | ||
1223 | /* | |
1224 | * If we haven't failed yet, we double up to a failing value instead | |
1225 | * of bisecting from highest possible queue depth. If we have set | |
1226 | * a limit other than td->o.iodepth, bisect between that. | |
1227 | */ | |
6bb58215 | 1228 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1229 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1230 | else | |
1231 | td->latency_qd *= 2; | |
1232 | ||
6bb58215 JA |
1233 | if (td->latency_qd > o->iodepth) |
1234 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1235 | |
1236 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1237 | |
3e260a46 | 1238 | /* |
6bb58215 JA |
1239 | * Same as last one, we are done. Let it run a latency cycle, so |
1240 | * we get only the results from the targeted depth. | |
3e260a46 | 1241 | */ |
6bb58215 JA |
1242 | if (td->latency_qd == qd) { |
1243 | if (td->latency_end_run) { | |
1244 | dprint(FD_RATE, "We are done\n"); | |
1245 | td->done = 1; | |
1246 | } else { | |
1247 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1248 | io_u_quiesce(td); | |
1249 | td->latency_end_run = 1; | |
1250 | reset_all_stats(td); | |
1251 | reset_io_stats(td); | |
1252 | } | |
1253 | } | |
3e260a46 JA |
1254 | |
1255 | lat_new_cycle(td); | |
1256 | } | |
1257 | ||
1258 | /* | |
1259 | * Check if we can bump the queue depth | |
1260 | */ | |
1261 | void lat_target_check(struct thread_data *td) | |
1262 | { | |
1263 | uint64_t usec_window; | |
1264 | uint64_t ios; | |
1265 | double success_ios; | |
1266 | ||
1267 | usec_window = utime_since_now(&td->latency_ts); | |
1268 | if (usec_window < td->o.latency_window) | |
1269 | return; | |
1270 | ||
1271 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1272 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1273 | success_ios *= 100.0; | |
1274 | ||
1275 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1276 | ||
1277 | if (success_ios >= td->o.latency_percentile.u.f) | |
1278 | lat_target_success(td); | |
1279 | else | |
1280 | __lat_target_failed(td); | |
1281 | } | |
1282 | ||
1283 | /* | |
1284 | * If latency target is enabled, we might be ramping up or down and not | |
1285 | * using the full queue depth available. | |
1286 | */ | |
effd6ff0 | 1287 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1288 | { |
1289 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1290 | ||
1291 | if (qempty) | |
1292 | return 1; | |
1293 | if (!td->o.latency_target) | |
1294 | return 0; | |
1295 | ||
1296 | return td->cur_depth >= td->latency_qd; | |
1297 | } | |
429f6675 | 1298 | |
10ba535a JA |
1299 | struct io_u *__get_io_u(struct thread_data *td) |
1300 | { | |
0cae66f6 | 1301 | struct io_u *io_u = NULL; |
10ba535a | 1302 | |
ca09be4b JA |
1303 | if (td->stop_io) |
1304 | return NULL; | |
1305 | ||
e8462bd8 JA |
1306 | td_io_u_lock(td); |
1307 | ||
1308 | again: | |
2ae0b204 JA |
1309 | if (!io_u_rempty(&td->io_u_requeues)) |
1310 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1311 | else if (!queue_full(td)) { |
2ae0b204 | 1312 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1313 | |
225ba9e3 | 1314 | io_u->file = NULL; |
6040dabc | 1315 | io_u->buflen = 0; |
10ba535a | 1316 | io_u->resid = 0; |
d7762cf8 | 1317 | io_u->end_io = NULL; |
755200a3 JA |
1318 | } |
1319 | ||
1320 | if (io_u) { | |
0c6e7517 | 1321 | assert(io_u->flags & IO_U_F_FREE); |
f8b0bd10 JA |
1322 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
1323 | IO_U_F_TRIMMED | IO_U_F_BARRIER | | |
1324 | IO_U_F_VER_LIST); | |
0c6e7517 | 1325 | |
755200a3 | 1326 | io_u->error = 0; |
bcd5abfa | 1327 | io_u->acct_ddir = -1; |
10ba535a | 1328 | td->cur_depth++; |
0c41214f | 1329 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1330 | io_u->ipo = NULL; |
1dec3e07 JA |
1331 | } else if (td->o.verify_async) { |
1332 | /* | |
1333 | * We ran out, wait for async verify threads to finish and | |
1334 | * return one | |
1335 | */ | |
1336 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1337 | goto again; | |
10ba535a JA |
1338 | } |
1339 | ||
e8462bd8 | 1340 | td_io_u_unlock(td); |
10ba535a JA |
1341 | return io_u; |
1342 | } | |
1343 | ||
0d29de83 | 1344 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1345 | { |
d72be545 JA |
1346 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1347 | return 0; | |
1348 | ||
1349 | if (td->trim_entries) { | |
0d29de83 | 1350 | int get_trim = 0; |
10ba535a | 1351 | |
0d29de83 JA |
1352 | if (td->trim_batch) { |
1353 | td->trim_batch--; | |
1354 | get_trim = 1; | |
1355 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1356 | td->last_ddir != DDIR_READ) { | |
1357 | td->trim_batch = td->o.trim_batch; | |
1358 | if (!td->trim_batch) | |
1359 | td->trim_batch = td->o.trim_backlog; | |
1360 | get_trim = 1; | |
1361 | } | |
1362 | ||
1363 | if (get_trim && !get_next_trim(td, io_u)) | |
1364 | return 1; | |
2ba1c290 | 1365 | } |
10ba535a | 1366 | |
0d29de83 JA |
1367 | return 0; |
1368 | } | |
1369 | ||
1370 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1371 | { | |
d72be545 JA |
1372 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1373 | return 0; | |
1374 | ||
1375 | if (td->io_hist_len) { | |
9e144189 JA |
1376 | int get_verify = 0; |
1377 | ||
d1ece0c7 | 1378 | if (td->verify_batch) |
9e144189 | 1379 | get_verify = 1; |
d1ece0c7 | 1380 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1381 | td->last_ddir != DDIR_READ) { |
1382 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1383 | if (!td->verify_batch) |
1384 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1385 | get_verify = 1; |
1386 | } | |
1387 | ||
d1ece0c7 JA |
1388 | if (get_verify && !get_next_verify(td, io_u)) { |
1389 | td->verify_batch--; | |
0d29de83 | 1390 | return 1; |
d1ece0c7 | 1391 | } |
9e144189 JA |
1392 | } |
1393 | ||
0d29de83 JA |
1394 | return 0; |
1395 | } | |
1396 | ||
de789769 JA |
1397 | /* |
1398 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1399 | * easy compressible data for simple de-dupe attempts. Do this for every |
1400 | * 512b block in the range, since that should be the smallest block size | |
1401 | * we can expect from a device. | |
de789769 JA |
1402 | */ |
1403 | static void small_content_scramble(struct io_u *io_u) | |
1404 | { | |
23f394d5 | 1405 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1406 | uint64_t boffset; |
23f394d5 JA |
1407 | unsigned int offset; |
1408 | void *p, *end; | |
de789769 | 1409 | |
23f394d5 JA |
1410 | if (!nr_blocks) |
1411 | return; | |
1412 | ||
1413 | p = io_u->xfer_buf; | |
fba76ee8 | 1414 | boffset = io_u->offset; |
81f0366c | 1415 | io_u->buf_filled_len = 0; |
fad82f76 | 1416 | |
23f394d5 JA |
1417 | for (i = 0; i < nr_blocks; i++) { |
1418 | /* | |
1419 | * Fill the byte offset into a "random" start offset of | |
1420 | * the buffer, given by the product of the usec time | |
1421 | * and the actual offset. | |
1422 | */ | |
fad82f76 | 1423 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1424 | offset &= ~(sizeof(uint64_t) - 1); |
1425 | if (offset >= 512 - sizeof(uint64_t)) | |
1426 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1427 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1428 | |
1429 | end = p + 512 - sizeof(io_u->start_time); | |
1430 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1431 | p += 512; | |
fad82f76 | 1432 | boffset += 512; |
23f394d5 | 1433 | } |
de789769 JA |
1434 | } |
1435 | ||
0d29de83 JA |
1436 | /* |
1437 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1438 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1439 | */ | |
1440 | struct io_u *get_io_u(struct thread_data *td) | |
1441 | { | |
1442 | struct fio_file *f; | |
1443 | struct io_u *io_u; | |
de789769 | 1444 | int do_scramble = 0; |
002fe734 | 1445 | long ret = 0; |
0d29de83 JA |
1446 | |
1447 | io_u = __get_io_u(td); | |
1448 | if (!io_u) { | |
1449 | dprint(FD_IO, "__get_io_u failed\n"); | |
1450 | return NULL; | |
1451 | } | |
1452 | ||
1453 | if (check_get_verify(td, io_u)) | |
1454 | goto out; | |
1455 | if (check_get_trim(td, io_u)) | |
1456 | goto out; | |
1457 | ||
755200a3 JA |
1458 | /* |
1459 | * from a requeue, io_u already setup | |
1460 | */ | |
1461 | if (io_u->file) | |
77f392bf | 1462 | goto out; |
755200a3 | 1463 | |
429f6675 JA |
1464 | /* |
1465 | * If using an iolog, grab next piece if any available. | |
1466 | */ | |
d72be545 | 1467 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1468 | if (read_iolog_get(td, io_u)) |
1469 | goto err_put; | |
2ba1c290 | 1470 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1471 | ret = -EBUSY; |
2ba1c290 | 1472 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1473 | goto err_put; |
2ba1c290 | 1474 | } |
5ec10eaa | 1475 | |
429f6675 | 1476 | f = io_u->file; |
002fe734 JA |
1477 | if (!f) { |
1478 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1479 | goto err_put; | |
1480 | } | |
1481 | ||
d6aed795 | 1482 | assert(fio_file_open(f)); |
97af62ce | 1483 | |
ff58fced | 1484 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1485 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1486 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1487 | goto err_put; |
2ba1c290 | 1488 | } |
10ba535a | 1489 | |
f1dfb668 JA |
1490 | f->last_start[io_u->ddir] = io_u->offset; |
1491 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1492 | |
fd68418e | 1493 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1494 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1495 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1496 | td->o.min_bs[DDIR_WRITE], |
1497 | io_u->xfer_buflen); | |
ff441ae8 | 1498 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1499 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1500 | do_scramble = 1; |
d72be545 | 1501 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1502 | populate_verify_io_u(td, io_u); |
1503 | do_scramble = 0; | |
1504 | } | |
fd68418e | 1505 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1506 | /* |
1507 | * Reset the buf_filled parameters so next time if the | |
1508 | * buffer is used for writes it is refilled. | |
1509 | */ | |
cbe8d756 RR |
1510 | io_u->buf_filled_len = 0; |
1511 | } | |
87dc1ab1 | 1512 | } |
10ba535a | 1513 | |
165faf16 JA |
1514 | /* |
1515 | * Set io data pointers. | |
1516 | */ | |
cec6b55d JA |
1517 | io_u->xfer_buf = io_u->buf; |
1518 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1519 | |
6ac7a331 | 1520 | out: |
0d29de83 | 1521 | assert(io_u->file); |
429f6675 | 1522 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1523 | if (!td->o.disable_slat) |
1524 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1525 | if (do_scramble) |
1526 | small_content_scramble(io_u); | |
429f6675 | 1527 | return io_u; |
36167d82 | 1528 | } |
429f6675 | 1529 | err_put: |
2ba1c290 | 1530 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1531 | put_io_u(td, io_u); |
002fe734 | 1532 | return ERR_PTR(ret); |
10ba535a JA |
1533 | } |
1534 | ||
5451792e JA |
1535 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1536 | { | |
8b28bd41 | 1537 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1538 | |
8b28bd41 DM |
1539 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1540 | return; | |
5451792e | 1541 | |
709c8313 RE |
1542 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1543 | io_u->file ? " on file " : "", | |
1544 | io_u->file ? io_u->file->file_name : "", | |
1545 | strerror(io_u->error), | |
1546 | io_ddir_name(io_u->ddir), | |
1547 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1548 | |
1549 | if (!td->error) | |
1550 | td_verror(td, io_u->error, "io_u error"); | |
1551 | } | |
1552 | ||
aba6c951 JA |
1553 | static inline int gtod_reduce(struct thread_data *td) |
1554 | { | |
729fe3af | 1555 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1556 | && td->o.disable_bw; |
aba6c951 JA |
1557 | } |
1558 | ||
c8eeb9df JA |
1559 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1560 | struct io_completion_data *icd, | |
1561 | const enum fio_ddir idx, unsigned int bytes) | |
1562 | { | |
24d23ca7 | 1563 | unsigned long lusec = 0; |
c8eeb9df | 1564 | |
aba6c951 | 1565 | if (!gtod_reduce(td)) |
c8eeb9df JA |
1566 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1567 | ||
1568 | if (!td->o.disable_lat) { | |
1569 | unsigned long tusec; | |
1570 | ||
1571 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1572 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1573 | |
d4afedfd JA |
1574 | if (td->flags & TD_F_PROFILE_OPS) { |
1575 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1576 | ||
1577 | if (ops->io_u_lat) | |
1578 | icd->error = ops->io_u_lat(td, tusec); | |
1579 | } | |
1580 | ||
3e260a46 JA |
1581 | if (td->o.max_latency && tusec > td->o.max_latency) |
1582 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1583 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1584 | if (lat_target_failed(td)) | |
1585 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1586 | } |
c8eeb9df JA |
1587 | } |
1588 | ||
1589 | if (!td->o.disable_clat) { | |
ae588852 | 1590 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1591 | io_u_mark_latency(td, lusec); |
1592 | } | |
1593 | ||
1594 | if (!td->o.disable_bw) | |
1595 | add_bw_sample(td, idx, bytes, &icd->time); | |
1596 | ||
aba6c951 JA |
1597 | if (!gtod_reduce(td)) |
1598 | add_iops_sample(td, idx, bytes, &icd->time); | |
c8eeb9df JA |
1599 | } |
1600 | ||
1b8dbf25 SL |
1601 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1602 | { | |
1ae83d45 JA |
1603 | uint64_t secs, remainder, bps, bytes; |
1604 | ||
1b8dbf25 SL |
1605 | bytes = td->this_io_bytes[ddir]; |
1606 | bps = td->rate_bps[ddir]; | |
1607 | secs = bytes / bps; | |
1608 | remainder = bytes % bps; | |
1609 | return remainder * 1000000 / bps + secs * 1000000; | |
1610 | } | |
1611 | ||
f8b0bd10 | 1612 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1613 | struct io_completion_data *icd) |
10ba535a | 1614 | { |
f8b0bd10 JA |
1615 | struct io_u *io_u = *io_u_ptr; |
1616 | enum fio_ddir ddir = io_u->ddir; | |
1617 | struct fio_file *f = io_u->file; | |
10ba535a | 1618 | |
2ba1c290 JA |
1619 | dprint_io_u(io_u, "io complete"); |
1620 | ||
2ecc1b57 | 1621 | td_io_u_lock(td); |
0c6e7517 | 1622 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1623 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1624 | |
1625 | /* | |
1626 | * Mark IO ok to verify | |
1627 | */ | |
1628 | if (io_u->ipo) { | |
890b6656 JA |
1629 | /* |
1630 | * Remove errored entry from the verification list | |
1631 | */ | |
1632 | if (io_u->error) | |
1633 | unlog_io_piece(td, io_u); | |
1634 | else { | |
1635 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1636 | write_barrier(); | |
1637 | } | |
f9401285 JA |
1638 | } |
1639 | ||
2ecc1b57 | 1640 | td_io_u_unlock(td); |
0c6e7517 | 1641 | |
f8b0bd10 | 1642 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1643 | td->last_was_sync = 1; |
44f29692 JA |
1644 | if (f) { |
1645 | f->first_write = -1ULL; | |
1646 | f->last_write = -1ULL; | |
1647 | } | |
87dc1ab1 JA |
1648 | return; |
1649 | } | |
1650 | ||
1651 | td->last_was_sync = 0; | |
f8b0bd10 | 1652 | td->last_ddir = ddir; |
87dc1ab1 | 1653 | |
f8b0bd10 | 1654 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1655 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1656 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1657 | int ret; |
10ba535a | 1658 | |
f8b0bd10 JA |
1659 | td->io_blocks[ddir]++; |
1660 | td->this_io_blocks[ddir]++; | |
1661 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1662 | |
1663 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1664 | td->this_io_bytes[ddir] += bytes; |
1665 | ||
ca09be4b JA |
1666 | if (ddir == DDIR_WRITE) { |
1667 | if (f) { | |
1668 | if (f->first_write == -1ULL || | |
1669 | io_u->offset < f->first_write) | |
1670 | f->first_write = io_u->offset; | |
1671 | if (f->last_write == -1ULL || | |
1672 | ((io_u->offset + bytes) > f->last_write)) | |
1673 | f->last_write = io_u->offset + bytes; | |
1674 | } | |
1675 | if (td->last_write_comp) { | |
1676 | int idx = td->last_write_idx++; | |
1677 | ||
1678 | td->last_write_comp[idx] = io_u->offset; | |
1679 | if (td->last_write_idx == td->o.iodepth) | |
1680 | td->last_write_idx = 0; | |
1681 | } | |
44f29692 JA |
1682 | } |
1683 | ||
6b1190fd SL |
1684 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1685 | td->runstate == TD_VERIFYING)) { | |
f8b0bd10 | 1686 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1687 | |
f8b0bd10 JA |
1688 | if (__should_check_rate(td, ddir)) { |
1689 | td->rate_pending_usleep[ddir] = | |
1690 | (usec_for_io(td, ddir) - | |
ba3e4e0c | 1691 | utime_since_now(&td->start)); |
b23b6a2f | 1692 | } |
f8b0bd10 JA |
1693 | if (ddir != DDIR_TRIM && |
1694 | __should_check_rate(td, oddir)) { | |
1695 | td->rate_pending_usleep[oddir] = | |
1696 | (usec_for_io(td, oddir) - | |
ba3e4e0c | 1697 | utime_since_now(&td->start)); |
f8b0bd10 | 1698 | } |
721938ae | 1699 | } |
10ba535a | 1700 | |
f8b0bd10 | 1701 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1702 | |
d7762cf8 | 1703 | if (io_u->end_io) { |
f8b0bd10 JA |
1704 | ret = io_u->end_io(td, io_u_ptr); |
1705 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1706 | if (ret && !icd->error) |
1707 | icd->error = ret; | |
1708 | } | |
ff58fced | 1709 | } else if (io_u->error) { |
10ba535a | 1710 | icd->error = io_u->error; |
5451792e JA |
1711 | io_u_log_error(td, io_u); |
1712 | } | |
8b28bd41 | 1713 | if (icd->error) { |
f8b0bd10 JA |
1714 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1715 | ||
8b28bd41 DM |
1716 | if (!td_non_fatal_error(td, eb, icd->error)) |
1717 | return; | |
f8b0bd10 | 1718 | |
f2bba182 RR |
1719 | /* |
1720 | * If there is a non_fatal error, then add to the error count | |
1721 | * and clear all the errors. | |
1722 | */ | |
1723 | update_error_count(td, icd->error); | |
1724 | td_clear_error(td); | |
1725 | icd->error = 0; | |
f8b0bd10 JA |
1726 | if (io_u) |
1727 | io_u->error = 0; | |
f2bba182 | 1728 | } |
10ba535a JA |
1729 | } |
1730 | ||
9520ebb9 JA |
1731 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1732 | int nr) | |
10ba535a | 1733 | { |
6eaf09d6 | 1734 | int ddir; |
aba6c951 JA |
1735 | |
1736 | if (!gtod_reduce(td)) | |
9520ebb9 | 1737 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1738 | |
3af6ef39 JA |
1739 | icd->nr = nr; |
1740 | ||
10ba535a | 1741 | icd->error = 0; |
6eaf09d6 SL |
1742 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1743 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1744 | } |
1745 | ||
97601024 JA |
1746 | static void ios_completed(struct thread_data *td, |
1747 | struct io_completion_data *icd) | |
36167d82 JA |
1748 | { |
1749 | struct io_u *io_u; | |
1750 | int i; | |
1751 | ||
10ba535a JA |
1752 | for (i = 0; i < icd->nr; i++) { |
1753 | io_u = td->io_ops->event(td, i); | |
1754 | ||
f8b0bd10 | 1755 | io_completed(td, &io_u, icd); |
e8462bd8 | 1756 | |
f8b0bd10 | 1757 | if (io_u) |
e8462bd8 | 1758 | put_io_u(td, io_u); |
10ba535a JA |
1759 | } |
1760 | } | |
97601024 | 1761 | |
e7e6cfb4 JA |
1762 | /* |
1763 | * Complete a single io_u for the sync engines. | |
1764 | */ | |
581e7141 | 1765 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1766 | uint64_t *bytes) |
97601024 JA |
1767 | { |
1768 | struct io_completion_data icd; | |
1769 | ||
9520ebb9 | 1770 | init_icd(td, &icd, 1); |
f8b0bd10 | 1771 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1772 | |
f8b0bd10 | 1773 | if (io_u) |
e8462bd8 | 1774 | put_io_u(td, io_u); |
97601024 | 1775 | |
581e7141 JA |
1776 | if (icd.error) { |
1777 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1778 | return -1; | |
1779 | } | |
97601024 | 1780 | |
581e7141 | 1781 | if (bytes) { |
6eaf09d6 SL |
1782 | int ddir; |
1783 | ||
1784 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1785 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1786 | } |
1787 | ||
1788 | return 0; | |
97601024 JA |
1789 | } |
1790 | ||
e7e6cfb4 JA |
1791 | /* |
1792 | * Called to complete min_events number of io for the async engines. | |
1793 | */ | |
581e7141 | 1794 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1795 | uint64_t *bytes) |
97601024 | 1796 | { |
97601024 | 1797 | struct io_completion_data icd; |
00de55ef | 1798 | struct timespec *tvp = NULL; |
97601024 | 1799 | int ret; |
4d06a338 | 1800 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1801 | |
4950421a | 1802 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1803 | |
4950421a | 1804 | if (!min_evts) |
00de55ef | 1805 | tvp = &ts; |
5fb4b366 RE |
1806 | else if (min_evts > td->cur_depth) |
1807 | min_evts = td->cur_depth; | |
97601024 | 1808 | |
4950421a | 1809 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1810 | if (ret < 0) { |
e1161c32 | 1811 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1812 | return ret; |
1813 | } else if (!ret) | |
1814 | return ret; | |
1815 | ||
9520ebb9 | 1816 | init_icd(td, &icd, ret); |
97601024 | 1817 | ios_completed(td, &icd); |
581e7141 JA |
1818 | if (icd.error) { |
1819 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1820 | return -1; | |
1821 | } | |
97601024 | 1822 | |
581e7141 | 1823 | if (bytes) { |
6eaf09d6 SL |
1824 | int ddir; |
1825 | ||
1826 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1827 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1828 | } |
1829 | ||
1830 | return 0; | |
97601024 | 1831 | } |
7e77dd02 JA |
1832 | |
1833 | /* | |
1834 | * Call when io_u is really queued, to update the submission latency. | |
1835 | */ | |
1836 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1837 | { | |
9520ebb9 JA |
1838 | if (!td->o.disable_slat) { |
1839 | unsigned long slat_time; | |
7e77dd02 | 1840 | |
9520ebb9 | 1841 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1842 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1843 | io_u->offset); | |
9520ebb9 | 1844 | } |
7e77dd02 | 1845 | } |
433afcb4 | 1846 | |
5c94b008 JA |
1847 | /* |
1848 | * See if we should reuse the last seed, if dedupe is enabled | |
1849 | */ | |
1850 | static struct frand_state *get_buf_state(struct thread_data *td) | |
1851 | { | |
1852 | unsigned int v; | |
1853 | unsigned long r; | |
1854 | ||
1855 | if (!td->o.dedupe_percentage) | |
1856 | return &td->buf_state; | |
c0b69b92 JA |
1857 | else if (td->o.dedupe_percentage == 100) |
1858 | return &td->buf_state_prev; | |
5c94b008 JA |
1859 | |
1860 | r = __rand(&td->dedupe_state); | |
1861 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1862 | ||
1863 | if (v <= td->o.dedupe_percentage) | |
1864 | return &td->buf_state_prev; | |
1865 | ||
1866 | return &td->buf_state; | |
1867 | } | |
1868 | ||
1869 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) | |
1870 | { | |
1871 | if (rs == &td->buf_state) | |
1872 | frand_copy(&td->buf_state_prev, rs); | |
1873 | } | |
1874 | ||
cc86c395 JA |
1875 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1876 | unsigned int max_bs) | |
5973cafb | 1877 | { |
d1af2894 JA |
1878 | struct thread_options *o = &td->o; |
1879 | ||
4eff3e57 | 1880 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 1881 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1882 | struct frand_state *rs; |
1066358a | 1883 | unsigned int left = max_bs; |
5c94b008 | 1884 | |
1066358a JA |
1885 | do { |
1886 | rs = get_buf_state(td); | |
9c42684e | 1887 | |
1066358a | 1888 | min_write = min(min_write, left); |
f97a43a1 | 1889 | |
1066358a JA |
1890 | if (perc) { |
1891 | unsigned int seg = min_write; | |
cc86c395 | 1892 | |
1066358a JA |
1893 | seg = min(min_write, td->o.compress_chunk); |
1894 | if (!seg) | |
1895 | seg = min_write; | |
1896 | ||
1897 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1898 | min_write, o->buffer_pattern, |
1899 | o->buffer_pattern_bytes); | |
1066358a JA |
1900 | } else |
1901 | fill_random_buf(rs, buf, min_write); | |
1902 | ||
1903 | buf += min_write; | |
1904 | left -= min_write; | |
5c94b008 | 1905 | save_buf_state(td, rs); |
1066358a | 1906 | } while (left); |
d1af2894 JA |
1907 | } else if (o->buffer_pattern_bytes) |
1908 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 1909 | else if (o->zero_buffers) |
cc86c395 | 1910 | memset(buf, 0, max_bs); |
999d245e JA |
1911 | else |
1912 | fill_random_buf(get_buf_state(td), buf, max_bs); | |
cc86c395 JA |
1913 | } |
1914 | ||
1915 | /* | |
1916 | * "randomly" fill the buffer contents | |
1917 | */ | |
1918 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1919 | unsigned int min_write, unsigned int max_bs) | |
1920 | { | |
1921 | io_u->buf_filled_len = 0; | |
1922 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1923 | } |