Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
5973cafb | 9 | #include "hash.h" |
4f5af7b2 | 10 | #include "verify.h" |
0d29de83 | 11 | #include "trim.h" |
1fbbf72e | 12 | #include "lib/rand.h" |
7ebd796f | 13 | #include "lib/axmap.h" |
002fe734 | 14 | #include "err.h" |
10ba535a | 15 | |
97601024 JA |
16 | struct io_completion_data { |
17 | int nr; /* input */ | |
97601024 JA |
18 | |
19 | int error; /* output */ | |
100f49f1 | 20 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
97601024 JA |
21 | struct timeval time; /* output */ |
22 | }; | |
23 | ||
10ba535a | 24 | /* |
7ebd796f | 25 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
26 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
27 | */ | |
1ae83d45 | 28 | static int random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 29 | { |
7ebd796f | 30 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
31 | } |
32 | ||
df415585 JA |
33 | /* |
34 | * Mark a given offset as used in the map. | |
35 | */ | |
9bf2061e | 36 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 37 | { |
2dc1bbeb | 38 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 39 | struct fio_file *f = io_u->file; |
51ede0b1 | 40 | unsigned int nr_blocks; |
1ae83d45 | 41 | uint64_t block; |
df415585 | 42 | |
1ae83d45 | 43 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 44 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 45 | |
2ab9e98b | 46 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 47 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 48 | |
51ede0b1 JA |
49 | if ((nr_blocks * min_bs) < io_u->buflen) |
50 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
51 | } |
52 | ||
74776733 JA |
53 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
54 | enum fio_ddir ddir) | |
2ba1c290 | 55 | { |
74776733 JA |
56 | uint64_t max_blocks; |
57 | uint64_t max_size; | |
2ba1c290 | 58 | |
ff58fced JA |
59 | assert(ddir_rw(ddir)); |
60 | ||
d9dd70f7 JA |
61 | /* |
62 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
63 | */ | |
64 | max_size = f->io_size; | |
65 | if (max_size > f->real_file_size) | |
66 | max_size = f->real_file_size; | |
67 | ||
ed335855 SN |
68 | if (td->o.zone_range) |
69 | max_size = td->o.zone_range; | |
70 | ||
1ae83d45 | 71 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
72 | if (!max_blocks) |
73 | return 0; | |
74 | ||
67778e88 | 75 | return max_blocks; |
2ba1c290 JA |
76 | } |
77 | ||
1ae83d45 JA |
78 | struct rand_off { |
79 | struct flist_head list; | |
80 | uint64_t off; | |
81 | }; | |
82 | ||
e25839d4 | 83 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 84 | enum fio_ddir ddir, uint64_t *b) |
ec4015da | 85 | { |
1ae83d45 | 86 | uint64_t r, lastb; |
ec4015da | 87 | |
74776733 JA |
88 | lastb = last_block(td, f, ddir); |
89 | if (!lastb) | |
90 | return 1; | |
5e0baa7f | 91 | |
74776733 | 92 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) { |
1ae83d45 | 93 | uint64_t rmax; |
2615cc4b | 94 | |
8055e41d JA |
95 | rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX; |
96 | ||
97 | if (td->o.use_os_rand) { | |
98 | rmax = OS_RAND_MAX; | |
99 | r = os_random_long(&td->random_state); | |
100 | } else { | |
101 | rmax = FRAND_MAX; | |
102 | r = __rand(&td->__random_state); | |
103 | } | |
104 | ||
4b91ee8f | 105 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 106 | |
80e88a1d | 107 | *b = lastb * (r / ((uint64_t) rmax + 1.0)); |
51ede0b1 | 108 | } else { |
8055e41d | 109 | uint64_t off = 0; |
43c63a78 | 110 | |
74776733 | 111 | if (lfsr_next(&f->lfsr, &off, lastb)) |
8055e41d | 112 | return 1; |
ec4015da | 113 | |
8055e41d JA |
114 | *b = off; |
115 | } | |
0ce8b119 | 116 | |
ec4015da | 117 | /* |
51ede0b1 | 118 | * if we are not maintaining a random map, we are done. |
ec4015da | 119 | */ |
51ede0b1 JA |
120 | if (!file_randommap(td, f)) |
121 | goto ret; | |
43c63a78 JA |
122 | |
123 | /* | |
51ede0b1 | 124 | * calculate map offset and check if it's free |
43c63a78 | 125 | */ |
51ede0b1 JA |
126 | if (random_map_free(f, *b)) |
127 | goto ret; | |
128 | ||
4b91ee8f JA |
129 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
130 | (unsigned long long) *b); | |
51ede0b1 | 131 | |
7ebd796f | 132 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
133 | if (*b == (uint64_t) -1ULL) |
134 | return 1; | |
0ce8b119 JA |
135 | ret: |
136 | return 0; | |
ec4015da JA |
137 | } |
138 | ||
925fee33 JA |
139 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
140 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 141 | uint64_t *b) |
e25839d4 | 142 | { |
9c6f6316 | 143 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
144 | return 0; |
145 | } | |
146 | ||
925fee33 JA |
147 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
148 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 149 | uint64_t *b) |
925fee33 | 150 | { |
9c6f6316 | 151 | *b = pareto_next(&f->zipf); |
925fee33 JA |
152 | return 0; |
153 | } | |
154 | ||
1ae83d45 JA |
155 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
156 | { | |
157 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
158 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
159 | ||
160 | return r1->off - r2->off; | |
161 | } | |
162 | ||
163 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
164 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
165 | { |
166 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
167 | return __get_next_rand_offset(td, f, ddir, b); | |
168 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
169 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
170 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
171 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
e25839d4 JA |
172 | |
173 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
174 | return 1; | |
175 | } | |
176 | ||
bcd5abfa JA |
177 | /* |
178 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
179 | * specified. | |
180 | */ | |
181 | static inline int should_sort_io(struct thread_data *td) | |
182 | { | |
183 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
184 | return 0; | |
185 | if (!td_random(td)) | |
186 | return 0; | |
187 | if (td->runstate != TD_VERIFYING) | |
188 | return 0; | |
189 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
190 | return 0; | |
191 | ||
192 | return 1; | |
193 | } | |
194 | ||
d9472271 | 195 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
196 | { |
197 | unsigned int v; | |
198 | unsigned long r; | |
199 | ||
d9472271 | 200 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
201 | return 1; |
202 | ||
203 | if (td->o.use_os_rand) { | |
d9472271 | 204 | r = os_random_long(&td->seq_rand_state[ddir]); |
211c9b89 JA |
205 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); |
206 | } else { | |
d9472271 | 207 | r = __rand(&td->__seq_rand_state[ddir]); |
211c9b89 JA |
208 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
209 | } | |
210 | ||
d9472271 | 211 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
212 | } |
213 | ||
1ae83d45 JA |
214 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
215 | enum fio_ddir ddir, uint64_t *b) | |
216 | { | |
217 | struct rand_off *r; | |
218 | int i, ret = 1; | |
219 | ||
bcd5abfa | 220 | if (!should_sort_io(td)) |
1ae83d45 JA |
221 | return get_off_from_method(td, f, ddir, b); |
222 | ||
223 | if (!flist_empty(&td->next_rand_list)) { | |
224 | struct rand_off *r; | |
225 | fetch: | |
9342d5f8 | 226 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
227 | flist_del(&r->list); |
228 | *b = r->off; | |
229 | free(r); | |
230 | return 0; | |
231 | } | |
232 | ||
233 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
234 | r = malloc(sizeof(*r)); | |
235 | ||
236 | ret = get_off_from_method(td, f, ddir, &r->off); | |
237 | if (ret) { | |
238 | free(r); | |
239 | break; | |
240 | } | |
241 | ||
242 | flist_add(&r->list, &td->next_rand_list); | |
243 | } | |
244 | ||
245 | if (ret && !i) | |
246 | return ret; | |
247 | ||
248 | assert(!flist_empty(&td->next_rand_list)); | |
249 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
250 | goto fetch; | |
251 | } | |
252 | ||
38dad62d | 253 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 254 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 255 | { |
c04e4661 DE |
256 | if (!get_next_rand_offset(td, f, ddir, b)) |
257 | return 0; | |
258 | ||
259 | if (td->o.time_based) { | |
33c48814 | 260 | fio_file_reset(td, f); |
c04e4661 DE |
261 | if (!get_next_rand_offset(td, f, ddir, b)) |
262 | return 0; | |
38dad62d JA |
263 | } |
264 | ||
c04e4661 | 265 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
4b91ee8f JA |
266 | f->file_name, (unsigned long long) f->last_pos, |
267 | (unsigned long long) f->real_file_size); | |
c04e4661 | 268 | return 1; |
38dad62d JA |
269 | } |
270 | ||
37cf9e3c | 271 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 272 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 273 | { |
ac002339 JA |
274 | struct thread_options *o = &td->o; |
275 | ||
ff58fced JA |
276 | assert(ddir_rw(ddir)); |
277 | ||
ac002339 JA |
278 | if (f->last_pos >= f->io_size + get_start_offset(td, f) && |
279 | o->time_based) | |
c04e4661 DE |
280 | f->last_pos = f->last_pos - f->io_size; |
281 | ||
38dad62d | 282 | if (f->last_pos < f->real_file_size) { |
1ae83d45 | 283 | uint64_t pos; |
059b0802 | 284 | |
ac002339 | 285 | if (f->last_pos == f->file_offset && o->ddir_seq_add < 0) |
a66da7a2 JA |
286 | f->last_pos = f->real_file_size; |
287 | ||
288 | pos = f->last_pos - f->file_offset; | |
ac002339 JA |
289 | if (pos && o->ddir_seq_add) { |
290 | pos += o->ddir_seq_add; | |
291 | ||
292 | /* | |
293 | * If we reach beyond the end of the file | |
294 | * with holed IO, wrap around to the | |
295 | * beginning again. | |
296 | */ | |
297 | if (pos >= f->real_file_size) | |
298 | pos = f->file_offset; | |
299 | } | |
059b0802 | 300 | |
37cf9e3c | 301 | *offset = pos; |
38dad62d JA |
302 | return 0; |
303 | } | |
304 | ||
305 | return 1; | |
306 | } | |
307 | ||
308 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
309 | enum fio_ddir ddir, int rw_seq, |
310 | unsigned int *is_random) | |
38dad62d JA |
311 | { |
312 | struct fio_file *f = io_u->file; | |
1ae83d45 | 313 | uint64_t b, offset; |
38dad62d JA |
314 | int ret; |
315 | ||
ff58fced JA |
316 | assert(ddir_rw(ddir)); |
317 | ||
37cf9e3c JA |
318 | b = offset = -1ULL; |
319 | ||
38dad62d | 320 | if (rw_seq) { |
211c9b89 | 321 | if (td_random(td)) { |
6aca9b3d | 322 | if (should_do_random(td, ddir)) { |
211c9b89 | 323 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
324 | *is_random = 1; |
325 | } else { | |
326 | *is_random = 0; | |
211c9b89 JA |
327 | io_u->flags |= IO_U_F_BUSY_OK; |
328 | ret = get_next_seq_offset(td, f, ddir, &offset); | |
329 | if (ret) | |
330 | ret = get_next_rand_block(td, f, ddir, &b); | |
331 | } | |
6aca9b3d JA |
332 | } else { |
333 | *is_random = 0; | |
37cf9e3c | 334 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 335 | } |
38dad62d JA |
336 | } else { |
337 | io_u->flags |= IO_U_F_BUSY_OK; | |
6aca9b3d | 338 | *is_random = 0; |
38dad62d JA |
339 | |
340 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 341 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 342 | if (ret) { |
37cf9e3c | 343 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
344 | *is_random = 0; |
345 | } | |
38dad62d JA |
346 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
347 | if (f->last_start != -1ULL) | |
37cf9e3c | 348 | offset = f->last_start - f->file_offset; |
38dad62d | 349 | else |
37cf9e3c | 350 | offset = 0; |
38dad62d JA |
351 | ret = 0; |
352 | } else { | |
353 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
354 | ret = 1; | |
355 | } | |
356 | } | |
6d68b997 | 357 | |
37cf9e3c JA |
358 | if (!ret) { |
359 | if (offset != -1ULL) | |
360 | io_u->offset = offset; | |
361 | else if (b != -1ULL) | |
362 | io_u->offset = b * td->o.ba[ddir]; | |
363 | else { | |
4e0a8fa2 | 364 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
365 | ret = 1; |
366 | } | |
367 | } | |
368 | ||
38dad62d JA |
369 | return ret; |
370 | } | |
371 | ||
10ba535a JA |
372 | /* |
373 | * For random io, generate a random new block and see if it's used. Repeat | |
374 | * until we find a free one. For sequential io, just return the end of | |
375 | * the last io issued. | |
376 | */ | |
6aca9b3d JA |
377 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
378 | unsigned int *is_random) | |
10ba535a | 379 | { |
9bf2061e | 380 | struct fio_file *f = io_u->file; |
4ba66134 | 381 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 382 | int rw_seq_hit = 0; |
10ba535a | 383 | |
ff58fced JA |
384 | assert(ddir_rw(ddir)); |
385 | ||
38dad62d JA |
386 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
387 | rw_seq_hit = 1; | |
5736c10d | 388 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 389 | } |
211097b2 | 390 | |
6aca9b3d | 391 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 392 | return 1; |
10ba535a | 393 | |
009bd847 JA |
394 | if (io_u->offset >= f->io_size) { |
395 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
396 | (unsigned long long) io_u->offset, |
397 | (unsigned long long) f->io_size); | |
009bd847 JA |
398 | return 1; |
399 | } | |
400 | ||
401 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
402 | if (io_u->offset >= f->real_file_size) { |
403 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
404 | (unsigned long long) io_u->offset, |
405 | (unsigned long long) f->real_file_size); | |
10ba535a | 406 | return 1; |
2ba1c290 | 407 | } |
10ba535a JA |
408 | |
409 | return 0; | |
410 | } | |
411 | ||
6aca9b3d JA |
412 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
413 | unsigned int *is_random) | |
15dc1934 | 414 | { |
d72be545 JA |
415 | if (td->flags & TD_F_PROFILE_OPS) { |
416 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 417 | |
d72be545 | 418 | if (ops->fill_io_u_off) |
6aca9b3d | 419 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 420 | } |
15dc1934 | 421 | |
6aca9b3d | 422 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
423 | } |
424 | ||
79944128 JA |
425 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
426 | unsigned int buflen) | |
427 | { | |
428 | struct fio_file *f = io_u->file; | |
429 | ||
bedc9dc2 | 430 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
431 | } |
432 | ||
6aca9b3d JA |
433 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
434 | unsigned int is_random) | |
10ba535a | 435 | { |
6aca9b3d | 436 | int ddir = io_u->ddir; |
24d23ca7 | 437 | unsigned int buflen = 0; |
f3059de1 | 438 | unsigned int minbs, maxbs; |
1294c3ec | 439 | unsigned long r, rand_max; |
10ba535a | 440 | |
9ee1c647 | 441 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
442 | |
443 | if (td->o.bs_is_seq_rand) | |
444 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 445 | |
f3059de1 JA |
446 | minbs = td->o.min_bs[ddir]; |
447 | maxbs = td->o.max_bs[ddir]; | |
448 | ||
79944128 JA |
449 | if (minbs == maxbs) |
450 | return minbs; | |
451 | ||
52c58027 JA |
452 | /* |
453 | * If we can't satisfy the min block size from here, then fail | |
454 | */ | |
455 | if (!io_u_fits(td, io_u, minbs)) | |
456 | return 0; | |
457 | ||
4c07ad86 JA |
458 | if (td->o.use_os_rand) |
459 | rand_max = OS_RAND_MAX; | |
460 | else | |
461 | rand_max = FRAND_MAX; | |
462 | ||
79944128 | 463 | do { |
4c07ad86 JA |
464 | if (td->o.use_os_rand) |
465 | r = os_random_long(&td->bsrange_state); | |
466 | else | |
467 | r = __rand(&td->__bsrange_state); | |
468 | ||
720e84ad | 469 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 470 | buflen = 1 + (unsigned int) ((double) maxbs * |
4c07ad86 | 471 | (r / (rand_max + 1.0))); |
f3059de1 JA |
472 | if (buflen < minbs) |
473 | buflen = minbs; | |
5ec10eaa | 474 | } else { |
564ca972 JA |
475 | long perc = 0; |
476 | unsigned int i; | |
477 | ||
720e84ad JA |
478 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
479 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
480 | |
481 | buflen = bsp->bs; | |
482 | perc += bsp->perc; | |
79944128 JA |
483 | if ((r <= ((rand_max / 100L) * perc)) && |
484 | io_u_fits(td, io_u, buflen)) | |
564ca972 JA |
485 | break; |
486 | } | |
487 | } | |
79944128 | 488 | |
a9f70b1f JB |
489 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
490 | buflen = (buflen + td->o.verify_interval - 1) & | |
491 | ~(td->o.verify_interval - 1); | |
492 | ||
f3059de1 JA |
493 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
494 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 495 | |
79944128 | 496 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 497 | |
10ba535a JA |
498 | return buflen; |
499 | } | |
500 | ||
6aca9b3d JA |
501 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
502 | unsigned int is_random) | |
15dc1934 | 503 | { |
d72be545 JA |
504 | if (td->flags & TD_F_PROFILE_OPS) { |
505 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 506 | |
d72be545 | 507 | if (ops->fill_io_u_size) |
6aca9b3d | 508 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 509 | } |
15dc1934 | 510 | |
6aca9b3d | 511 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
512 | } |
513 | ||
afe24a5a JA |
514 | static void set_rwmix_bytes(struct thread_data *td) |
515 | { | |
afe24a5a JA |
516 | unsigned int diff; |
517 | ||
518 | /* | |
519 | * we do time or byte based switch. this is needed because | |
520 | * buffered writes may issue a lot quicker than they complete, | |
521 | * whereas reads do not. | |
522 | */ | |
e47f799f | 523 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 524 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
525 | } |
526 | ||
527 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
528 | { | |
529 | unsigned int v; | |
1294c3ec | 530 | unsigned long r; |
e47f799f | 531 | |
4c07ad86 JA |
532 | if (td->o.use_os_rand) { |
533 | r = os_random_long(&td->rwmix_state); | |
534 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); | |
535 | } else { | |
536 | r = __rand(&td->__rwmix_state); | |
537 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
538 | } | |
539 | ||
04c540d9 | 540 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
541 | return DDIR_READ; |
542 | ||
543 | return DDIR_WRITE; | |
afe24a5a JA |
544 | } |
545 | ||
002e7183 JA |
546 | void io_u_quiesce(struct thread_data *td) |
547 | { | |
548 | /* | |
549 | * We are going to sleep, ensure that we flush anything pending as | |
550 | * not to skew our latency numbers. | |
551 | * | |
552 | * Changed to only monitor 'in flight' requests here instead of the | |
553 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
554 | * io's that have been actually submitted to an async engine, | |
555 | * and cur_depth is meaningless for sync engines. | |
556 | */ | |
557 | while (td->io_u_in_flight) { | |
558 | int fio_unused ret; | |
559 | ||
560 | ret = io_u_queued_complete(td, 1, NULL); | |
561 | } | |
562 | } | |
563 | ||
581e7141 JA |
564 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
565 | { | |
566 | enum fio_ddir odir = ddir ^ 1; | |
567 | struct timeval t; | |
568 | long usec; | |
569 | ||
ff58fced JA |
570 | assert(ddir_rw(ddir)); |
571 | ||
315fcfec | 572 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
573 | return ddir; |
574 | ||
575 | /* | |
576 | * We have too much pending sleep in this direction. See if we | |
577 | * should switch. | |
578 | */ | |
315fcfec | 579 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
580 | /* |
581 | * Other direction does not have too much pending, switch | |
582 | */ | |
583 | if (td->rate_pending_usleep[odir] < 100000) | |
584 | return odir; | |
585 | ||
586 | /* | |
587 | * Both directions have pending sleep. Sleep the minimum time | |
588 | * and deduct from both. | |
589 | */ | |
590 | if (td->rate_pending_usleep[ddir] <= | |
591 | td->rate_pending_usleep[odir]) { | |
592 | usec = td->rate_pending_usleep[ddir]; | |
593 | } else { | |
594 | usec = td->rate_pending_usleep[odir]; | |
595 | ddir = odir; | |
596 | } | |
597 | } else | |
598 | usec = td->rate_pending_usleep[ddir]; | |
599 | ||
002e7183 | 600 | io_u_quiesce(td); |
78c1eda5 | 601 | |
581e7141 JA |
602 | fio_gettime(&t, NULL); |
603 | usec_sleep(td, usec); | |
604 | usec = utime_since_now(&t); | |
605 | ||
606 | td->rate_pending_usleep[ddir] -= usec; | |
607 | ||
608 | odir = ddir ^ 1; | |
609 | if (td_rw(td) && __should_check_rate(td, odir)) | |
610 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 611 | |
6eaf09d6 SL |
612 | if (ddir_trim(ddir)) |
613 | return ddir; | |
e0224c6b | 614 | |
581e7141 JA |
615 | return ddir; |
616 | } | |
617 | ||
10ba535a JA |
618 | /* |
619 | * Return the data direction for the next io_u. If the job is a | |
620 | * mixed read/write workload, check the rwmix cycle and switch if | |
621 | * necessary. | |
622 | */ | |
1e97cce9 | 623 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 624 | { |
581e7141 JA |
625 | enum fio_ddir ddir; |
626 | ||
5f9099ea JA |
627 | /* |
628 | * see if it's time to fsync | |
629 | */ | |
630 | if (td->o.fsync_blocks && | |
631 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
632 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
633 | return DDIR_SYNC; | |
634 | ||
635 | /* | |
636 | * see if it's time to fdatasync | |
637 | */ | |
638 | if (td->o.fdatasync_blocks && | |
639 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
640 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
641 | return DDIR_DATASYNC; | |
642 | ||
44f29692 JA |
643 | /* |
644 | * see if it's time to sync_file_range | |
645 | */ | |
646 | if (td->sync_file_range_nr && | |
647 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
648 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
649 | return DDIR_SYNC_FILE_RANGE; | |
650 | ||
10ba535a | 651 | if (td_rw(td)) { |
10ba535a JA |
652 | /* |
653 | * Check if it's time to seed a new data direction. | |
654 | */ | |
e4928662 | 655 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
656 | /* |
657 | * Put a top limit on how many bytes we do for | |
658 | * one data direction, to avoid overflowing the | |
659 | * ranges too much | |
660 | */ | |
661 | ddir = get_rand_ddir(td); | |
e47f799f JA |
662 | |
663 | if (ddir != td->rwmix_ddir) | |
664 | set_rwmix_bytes(td); | |
665 | ||
666 | td->rwmix_ddir = ddir; | |
10ba535a | 667 | } |
581e7141 | 668 | ddir = td->rwmix_ddir; |
10ba535a | 669 | } else if (td_read(td)) |
581e7141 | 670 | ddir = DDIR_READ; |
6eaf09d6 | 671 | else if (td_write(td)) |
581e7141 | 672 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
673 | else |
674 | ddir = DDIR_TRIM; | |
581e7141 JA |
675 | |
676 | td->rwmix_ddir = rate_ddir(td, ddir); | |
677 | return td->rwmix_ddir; | |
10ba535a JA |
678 | } |
679 | ||
1ef2b6be JA |
680 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
681 | { | |
bcd5abfa | 682 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
683 | |
684 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
685 | td->o.barrier_blocks && | |
686 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
687 | td->io_issues[DDIR_WRITE]) | |
688 | io_u->flags |= IO_U_F_BARRIER; | |
689 | } | |
690 | ||
e8462bd8 | 691 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 692 | { |
71b84caa | 693 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
694 | |
695 | if (ret) | |
696 | td_verror(td, ret, "file close"); | |
697 | } | |
698 | ||
10ba535a JA |
699 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
700 | { | |
e8462bd8 JA |
701 | td_io_u_lock(td); |
702 | ||
f8b0bd10 | 703 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 704 | put_file_log(td, io_u->file); |
f8b0bd10 | 705 | |
10ba535a | 706 | io_u->file = NULL; |
d7ee2a7d SL |
707 | io_u->flags |= IO_U_F_FREE; |
708 | ||
0c41214f RR |
709 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
710 | td->cur_depth--; | |
2ae0b204 | 711 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
712 | td_io_u_unlock(td); |
713 | td_io_u_free_notify(td); | |
10ba535a JA |
714 | } |
715 | ||
f2bba182 RR |
716 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
717 | { | |
718 | io_u->flags &= ~IO_U_F_FLIGHT; | |
719 | put_io_u(td, io_u); | |
720 | } | |
721 | ||
755200a3 JA |
722 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
723 | { | |
724 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 725 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 726 | |
465221b0 JA |
727 | dprint(FD_IO, "requeue %p\n", __io_u); |
728 | ||
e8462bd8 JA |
729 | td_io_u_lock(td); |
730 | ||
4d2e0f49 | 731 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
732 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
733 | td->io_issues[ddir]--; | |
5ec10eaa | 734 | |
4d2e0f49 | 735 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
736 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
737 | td->cur_depth--; | |
2ae0b204 JA |
738 | |
739 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 740 | td_io_u_unlock(td); |
755200a3 JA |
741 | *io_u = NULL; |
742 | } | |
743 | ||
9bf2061e | 744 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 745 | { |
6aca9b3d JA |
746 | unsigned int is_random; |
747 | ||
b4c5e1ac JA |
748 | if (td->io_ops->flags & FIO_NOIO) |
749 | goto out; | |
750 | ||
1ef2b6be | 751 | set_rw_ddir(td, io_u); |
5f9099ea | 752 | |
87dc1ab1 | 753 | /* |
ff58fced | 754 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 755 | */ |
ff58fced | 756 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 757 | goto out; |
a00735e6 | 758 | |
48f5abd3 JA |
759 | /* |
760 | * See if it's time to switch to a new zone | |
761 | */ | |
13af05ae | 762 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
763 | struct fio_file *f = io_u->file; |
764 | ||
48f5abd3 | 765 | td->zone_bytes = 0; |
4c8be5b1 JA |
766 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
767 | ||
768 | /* | |
769 | * Wrap from the beginning, if we exceed the file size | |
770 | */ | |
771 | if (f->file_offset >= f->real_file_size) | |
772 | f->file_offset = f->real_file_size - f->file_offset; | |
773 | f->last_pos = f->file_offset; | |
48f5abd3 JA |
774 | td->io_skip_bytes += td->o.zone_skip; |
775 | } | |
776 | ||
10ba535a | 777 | /* |
c685b5b2 JA |
778 | * No log, let the seq/rand engine retrieve the next buflen and |
779 | * position. | |
10ba535a | 780 | */ |
6aca9b3d | 781 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 782 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 783 | return 1; |
2ba1c290 | 784 | } |
10ba535a | 785 | |
6aca9b3d | 786 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
787 | if (!io_u->buflen) { |
788 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 789 | return 1; |
2ba1c290 | 790 | } |
bca4ed4d | 791 | |
2ba1c290 JA |
792 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
793 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
794 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
795 | (unsigned long long) io_u->offset, io_u->buflen, | |
796 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 797 | return 1; |
2ba1c290 | 798 | } |
6a5e6884 | 799 | |
bca4ed4d JA |
800 | /* |
801 | * mark entry before potentially trimming io_u | |
802 | */ | |
303032ae | 803 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 804 | mark_random_map(td, io_u); |
bca4ed4d | 805 | |
c38e9468 | 806 | out: |
2ba1c290 | 807 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 808 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 809 | return 0; |
10ba535a JA |
810 | } |
811 | ||
838bc709 JA |
812 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
813 | { | |
2b13e716 | 814 | int idx = 0; |
838bc709 JA |
815 | |
816 | switch (nr) { | |
817 | default: | |
2b13e716 | 818 | idx = 6; |
838bc709 JA |
819 | break; |
820 | case 33 ... 64: | |
2b13e716 | 821 | idx = 5; |
838bc709 JA |
822 | break; |
823 | case 17 ... 32: | |
2b13e716 | 824 | idx = 4; |
838bc709 JA |
825 | break; |
826 | case 9 ... 16: | |
2b13e716 | 827 | idx = 3; |
838bc709 JA |
828 | break; |
829 | case 5 ... 8: | |
2b13e716 | 830 | idx = 2; |
838bc709 JA |
831 | break; |
832 | case 1 ... 4: | |
2b13e716 | 833 | idx = 1; |
838bc709 JA |
834 | case 0: |
835 | break; | |
836 | } | |
837 | ||
2b13e716 | 838 | map[idx]++; |
838bc709 JA |
839 | } |
840 | ||
841 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
842 | { | |
843 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
844 | td->ts.total_submit++; | |
845 | } | |
846 | ||
847 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
848 | { | |
849 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
850 | td->ts.total_complete++; | |
851 | } | |
852 | ||
d8005759 | 853 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 854 | { |
2b13e716 | 855 | int idx = 0; |
71619dc2 JA |
856 | |
857 | switch (td->cur_depth) { | |
858 | default: | |
2b13e716 | 859 | idx = 6; |
a783e61a | 860 | break; |
71619dc2 | 861 | case 32 ... 63: |
2b13e716 | 862 | idx = 5; |
a783e61a | 863 | break; |
71619dc2 | 864 | case 16 ... 31: |
2b13e716 | 865 | idx = 4; |
a783e61a | 866 | break; |
71619dc2 | 867 | case 8 ... 15: |
2b13e716 | 868 | idx = 3; |
a783e61a | 869 | break; |
71619dc2 | 870 | case 4 ... 7: |
2b13e716 | 871 | idx = 2; |
a783e61a | 872 | break; |
71619dc2 | 873 | case 2 ... 3: |
2b13e716 | 874 | idx = 1; |
71619dc2 JA |
875 | case 1: |
876 | break; | |
877 | } | |
878 | ||
2b13e716 | 879 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
880 | } |
881 | ||
04a0feae JA |
882 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
883 | { | |
2b13e716 | 884 | int idx = 0; |
04a0feae JA |
885 | |
886 | assert(usec < 1000); | |
887 | ||
888 | switch (usec) { | |
889 | case 750 ... 999: | |
2b13e716 | 890 | idx = 9; |
04a0feae JA |
891 | break; |
892 | case 500 ... 749: | |
2b13e716 | 893 | idx = 8; |
04a0feae JA |
894 | break; |
895 | case 250 ... 499: | |
2b13e716 | 896 | idx = 7; |
04a0feae JA |
897 | break; |
898 | case 100 ... 249: | |
2b13e716 | 899 | idx = 6; |
04a0feae JA |
900 | break; |
901 | case 50 ... 99: | |
2b13e716 | 902 | idx = 5; |
04a0feae JA |
903 | break; |
904 | case 20 ... 49: | |
2b13e716 | 905 | idx = 4; |
04a0feae JA |
906 | break; |
907 | case 10 ... 19: | |
2b13e716 | 908 | idx = 3; |
04a0feae JA |
909 | break; |
910 | case 4 ... 9: | |
2b13e716 | 911 | idx = 2; |
04a0feae JA |
912 | break; |
913 | case 2 ... 3: | |
2b13e716 | 914 | idx = 1; |
04a0feae JA |
915 | case 0 ... 1: |
916 | break; | |
917 | } | |
918 | ||
2b13e716 JA |
919 | assert(idx < FIO_IO_U_LAT_U_NR); |
920 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
921 | } |
922 | ||
923 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 924 | { |
2b13e716 | 925 | int idx = 0; |
ec118304 JA |
926 | |
927 | switch (msec) { | |
928 | default: | |
2b13e716 | 929 | idx = 11; |
04a0feae | 930 | break; |
8abdce66 | 931 | case 1000 ... 1999: |
2b13e716 | 932 | idx = 10; |
04a0feae | 933 | break; |
8abdce66 | 934 | case 750 ... 999: |
2b13e716 | 935 | idx = 9; |
04a0feae | 936 | break; |
8abdce66 | 937 | case 500 ... 749: |
2b13e716 | 938 | idx = 8; |
04a0feae | 939 | break; |
8abdce66 | 940 | case 250 ... 499: |
2b13e716 | 941 | idx = 7; |
04a0feae | 942 | break; |
8abdce66 | 943 | case 100 ... 249: |
2b13e716 | 944 | idx = 6; |
04a0feae | 945 | break; |
8abdce66 | 946 | case 50 ... 99: |
2b13e716 | 947 | idx = 5; |
04a0feae | 948 | break; |
8abdce66 | 949 | case 20 ... 49: |
2b13e716 | 950 | idx = 4; |
04a0feae | 951 | break; |
8abdce66 | 952 | case 10 ... 19: |
2b13e716 | 953 | idx = 3; |
04a0feae | 954 | break; |
8abdce66 | 955 | case 4 ... 9: |
2b13e716 | 956 | idx = 2; |
04a0feae | 957 | break; |
ec118304 | 958 | case 2 ... 3: |
2b13e716 | 959 | idx = 1; |
ec118304 JA |
960 | case 0 ... 1: |
961 | break; | |
962 | } | |
963 | ||
2b13e716 JA |
964 | assert(idx < FIO_IO_U_LAT_M_NR); |
965 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
966 | } |
967 | ||
968 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
969 | { | |
970 | if (usec < 1000) | |
971 | io_u_mark_lat_usec(td, usec); | |
972 | else | |
973 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
974 | } |
975 | ||
0aabe160 JA |
976 | /* |
977 | * Get next file to service by choosing one at random | |
978 | */ | |
2cc52930 JA |
979 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
980 | enum fio_file_flags goodf, | |
d6aed795 | 981 | enum fio_file_flags badf) |
0aabe160 | 982 | { |
0aabe160 | 983 | struct fio_file *f; |
1c178180 | 984 | int fno; |
0aabe160 JA |
985 | |
986 | do { | |
87b10676 | 987 | int opened = 0; |
1294c3ec | 988 | unsigned long r; |
4c07ad86 JA |
989 | |
990 | if (td->o.use_os_rand) { | |
991 | r = os_random_long(&td->next_file_state); | |
992 | fno = (unsigned int) ((double) td->o.nr_files | |
993 | * (r / (OS_RAND_MAX + 1.0))); | |
994 | } else { | |
995 | r = __rand(&td->__next_file_state); | |
996 | fno = (unsigned int) ((double) td->o.nr_files | |
997 | * (r / (FRAND_MAX + 1.0))); | |
998 | } | |
7c83c089 | 999 | |
126d65c6 | 1000 | f = td->files[fno]; |
d6aed795 | 1001 | if (fio_file_done(f)) |
059e63c0 | 1002 | continue; |
1c178180 | 1003 | |
d6aed795 | 1004 | if (!fio_file_open(f)) { |
87b10676 JA |
1005 | int err; |
1006 | ||
002fe734 JA |
1007 | if (td->nr_open_files >= td->o.open_files) |
1008 | return ERR_PTR(-EBUSY); | |
1009 | ||
87b10676 JA |
1010 | err = td_io_open_file(td, f); |
1011 | if (err) | |
1012 | continue; | |
1013 | opened = 1; | |
1014 | } | |
1015 | ||
2ba1c290 JA |
1016 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1017 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1018 | return f; |
2ba1c290 | 1019 | } |
87b10676 JA |
1020 | if (opened) |
1021 | td_io_close_file(td, f); | |
0aabe160 JA |
1022 | } while (1); |
1023 | } | |
1024 | ||
1025 | /* | |
1026 | * Get next file to service by doing round robin between all available ones | |
1027 | */ | |
1c178180 JA |
1028 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1029 | int badf) | |
3d7c391d JA |
1030 | { |
1031 | unsigned int old_next_file = td->next_file; | |
1032 | struct fio_file *f; | |
1033 | ||
1034 | do { | |
87b10676 JA |
1035 | int opened = 0; |
1036 | ||
126d65c6 | 1037 | f = td->files[td->next_file]; |
3d7c391d JA |
1038 | |
1039 | td->next_file++; | |
2dc1bbeb | 1040 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1041 | td->next_file = 0; |
1042 | ||
87b10676 | 1043 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1044 | if (fio_file_done(f)) { |
d5ed68ea | 1045 | f = NULL; |
059e63c0 | 1046 | continue; |
d5ed68ea | 1047 | } |
059e63c0 | 1048 | |
d6aed795 | 1049 | if (!fio_file_open(f)) { |
87b10676 JA |
1050 | int err; |
1051 | ||
002fe734 JA |
1052 | if (td->nr_open_files >= td->o.open_files) |
1053 | return ERR_PTR(-EBUSY); | |
1054 | ||
87b10676 | 1055 | err = td_io_open_file(td, f); |
b5696bfc JA |
1056 | if (err) { |
1057 | dprint(FD_FILE, "error %d on open of %s\n", | |
1058 | err, f->file_name); | |
87c27b45 | 1059 | f = NULL; |
87b10676 | 1060 | continue; |
b5696bfc | 1061 | } |
87b10676 JA |
1062 | opened = 1; |
1063 | } | |
1064 | ||
0b9d69ec JA |
1065 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1066 | f->flags); | |
1c178180 | 1067 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1068 | break; |
1069 | ||
87b10676 JA |
1070 | if (opened) |
1071 | td_io_close_file(td, f); | |
1072 | ||
3d7c391d JA |
1073 | f = NULL; |
1074 | } while (td->next_file != old_next_file); | |
1075 | ||
2ba1c290 | 1076 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1077 | return f; |
1078 | } | |
1079 | ||
7eb36574 | 1080 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1081 | { |
1907dbc6 JA |
1082 | struct fio_file *f; |
1083 | ||
2dc1bbeb | 1084 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1085 | |
b5696bfc | 1086 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1087 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1088 | " nr_files=%d\n", td->nr_open_files, | |
1089 | td->nr_done_files, | |
1090 | td->o.nr_files); | |
bdb4e2e9 | 1091 | return NULL; |
2ba1c290 | 1092 | } |
bdb4e2e9 | 1093 | |
1907dbc6 | 1094 | f = td->file_service_file; |
d6aed795 | 1095 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1096 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1097 | goto out; | |
1098 | if (td->file_service_left--) | |
1099 | goto out; | |
1100 | } | |
1907dbc6 | 1101 | |
a086c257 JA |
1102 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1103 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1104 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1105 | else |
d6aed795 | 1106 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1107 | |
002fe734 JA |
1108 | if (IS_ERR(f)) |
1109 | return f; | |
1110 | ||
1907dbc6 JA |
1111 | td->file_service_file = f; |
1112 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1113 | out: |
0dac421f JA |
1114 | if (f) |
1115 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1116 | else | |
1117 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1118 | return f; |
bdb4e2e9 JA |
1119 | } |
1120 | ||
7eb36574 JA |
1121 | static struct fio_file *get_next_file(struct thread_data *td) |
1122 | { | |
372d8962 | 1123 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1124 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1125 | |
d72be545 JA |
1126 | if (ops->get_next_file) |
1127 | return ops->get_next_file(td); | |
1128 | } | |
7eb36574 JA |
1129 | |
1130 | return __get_next_file(td); | |
1131 | } | |
1132 | ||
002fe734 | 1133 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1134 | { |
1135 | struct fio_file *f; | |
1136 | ||
1137 | do { | |
1138 | f = get_next_file(td); | |
002fe734 JA |
1139 | if (IS_ERR_OR_NULL(f)) |
1140 | return PTR_ERR(f); | |
429f6675 | 1141 | |
429f6675 JA |
1142 | io_u->file = f; |
1143 | get_file(f); | |
1144 | ||
1145 | if (!fill_io_u(td, io_u)) | |
1146 | break; | |
1147 | ||
b5696bfc | 1148 | put_file_log(td, f); |
429f6675 | 1149 | td_io_close_file(td, f); |
b5696bfc | 1150 | io_u->file = NULL; |
d6aed795 | 1151 | fio_file_set_done(f); |
429f6675 | 1152 | td->nr_done_files++; |
0b9d69ec JA |
1153 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1154 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1155 | } while (1); |
1156 | ||
1157 | return 0; | |
1158 | } | |
1159 | ||
3e260a46 JA |
1160 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1161 | unsigned long tusec, unsigned long max_usec) | |
1162 | { | |
1163 | if (!td->error) | |
1164 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1165 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1166 | icd->error = ETIMEDOUT; | |
1167 | } | |
1168 | ||
1169 | static void lat_new_cycle(struct thread_data *td) | |
1170 | { | |
1171 | fio_gettime(&td->latency_ts, NULL); | |
1172 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1173 | td->latency_failed = 0; | |
1174 | } | |
1175 | ||
1176 | /* | |
1177 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1178 | * are at QD=1, then it's time to give up. | |
1179 | */ | |
1180 | static int __lat_target_failed(struct thread_data *td) | |
1181 | { | |
1182 | if (td->latency_qd == 1) | |
1183 | return 1; | |
1184 | ||
1185 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1186 | |
1187 | if (td->latency_qd == td->latency_qd_low) | |
1188 | td->latency_qd_low--; | |
1189 | ||
3e260a46 JA |
1190 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1191 | ||
1192 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1193 | ||
1194 | /* | |
1195 | * When we ramp QD down, quiesce existing IO to prevent | |
1196 | * a storm of ramp downs due to pending higher depth. | |
1197 | */ | |
1198 | io_u_quiesce(td); | |
1199 | lat_new_cycle(td); | |
1200 | return 0; | |
1201 | } | |
1202 | ||
1203 | static int lat_target_failed(struct thread_data *td) | |
1204 | { | |
1205 | if (td->o.latency_percentile.u.f == 100.0) | |
1206 | return __lat_target_failed(td); | |
1207 | ||
1208 | td->latency_failed++; | |
1209 | return 0; | |
1210 | } | |
1211 | ||
1212 | void lat_target_init(struct thread_data *td) | |
1213 | { | |
6bb58215 JA |
1214 | td->latency_end_run = 0; |
1215 | ||
3e260a46 JA |
1216 | if (td->o.latency_target) { |
1217 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1218 | fio_gettime(&td->latency_ts, NULL); | |
1219 | td->latency_qd = 1; | |
1220 | td->latency_qd_high = td->o.iodepth; | |
1221 | td->latency_qd_low = 1; | |
1222 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1223 | } else | |
1224 | td->latency_qd = td->o.iodepth; | |
1225 | } | |
1226 | ||
6bb58215 JA |
1227 | void lat_target_reset(struct thread_data *td) |
1228 | { | |
1229 | if (!td->latency_end_run) | |
1230 | lat_target_init(td); | |
1231 | } | |
1232 | ||
3e260a46 JA |
1233 | static void lat_target_success(struct thread_data *td) |
1234 | { | |
1235 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1236 | struct thread_options *o = &td->o; |
3e260a46 JA |
1237 | |
1238 | td->latency_qd_low = td->latency_qd; | |
1239 | ||
1240 | /* | |
1241 | * If we haven't failed yet, we double up to a failing value instead | |
1242 | * of bisecting from highest possible queue depth. If we have set | |
1243 | * a limit other than td->o.iodepth, bisect between that. | |
1244 | */ | |
6bb58215 | 1245 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1246 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1247 | else | |
1248 | td->latency_qd *= 2; | |
1249 | ||
6bb58215 JA |
1250 | if (td->latency_qd > o->iodepth) |
1251 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1252 | |
1253 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1254 | |
3e260a46 | 1255 | /* |
6bb58215 JA |
1256 | * Same as last one, we are done. Let it run a latency cycle, so |
1257 | * we get only the results from the targeted depth. | |
3e260a46 | 1258 | */ |
6bb58215 JA |
1259 | if (td->latency_qd == qd) { |
1260 | if (td->latency_end_run) { | |
1261 | dprint(FD_RATE, "We are done\n"); | |
1262 | td->done = 1; | |
1263 | } else { | |
1264 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1265 | io_u_quiesce(td); | |
1266 | td->latency_end_run = 1; | |
1267 | reset_all_stats(td); | |
1268 | reset_io_stats(td); | |
1269 | } | |
1270 | } | |
3e260a46 JA |
1271 | |
1272 | lat_new_cycle(td); | |
1273 | } | |
1274 | ||
1275 | /* | |
1276 | * Check if we can bump the queue depth | |
1277 | */ | |
1278 | void lat_target_check(struct thread_data *td) | |
1279 | { | |
1280 | uint64_t usec_window; | |
1281 | uint64_t ios; | |
1282 | double success_ios; | |
1283 | ||
1284 | usec_window = utime_since_now(&td->latency_ts); | |
1285 | if (usec_window < td->o.latency_window) | |
1286 | return; | |
1287 | ||
1288 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1289 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1290 | success_ios *= 100.0; | |
1291 | ||
1292 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1293 | ||
1294 | if (success_ios >= td->o.latency_percentile.u.f) | |
1295 | lat_target_success(td); | |
1296 | else | |
1297 | __lat_target_failed(td); | |
1298 | } | |
1299 | ||
1300 | /* | |
1301 | * If latency target is enabled, we might be ramping up or down and not | |
1302 | * using the full queue depth available. | |
1303 | */ | |
effd6ff0 | 1304 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1305 | { |
1306 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1307 | ||
1308 | if (qempty) | |
1309 | return 1; | |
1310 | if (!td->o.latency_target) | |
1311 | return 0; | |
1312 | ||
1313 | return td->cur_depth >= td->latency_qd; | |
1314 | } | |
429f6675 | 1315 | |
10ba535a JA |
1316 | struct io_u *__get_io_u(struct thread_data *td) |
1317 | { | |
0cae66f6 | 1318 | struct io_u *io_u = NULL; |
10ba535a | 1319 | |
e8462bd8 JA |
1320 | td_io_u_lock(td); |
1321 | ||
1322 | again: | |
2ae0b204 JA |
1323 | if (!io_u_rempty(&td->io_u_requeues)) |
1324 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1325 | else if (!queue_full(td)) { |
2ae0b204 | 1326 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1327 | |
225ba9e3 | 1328 | io_u->file = NULL; |
6040dabc | 1329 | io_u->buflen = 0; |
10ba535a | 1330 | io_u->resid = 0; |
d7762cf8 | 1331 | io_u->end_io = NULL; |
755200a3 JA |
1332 | } |
1333 | ||
1334 | if (io_u) { | |
0c6e7517 | 1335 | assert(io_u->flags & IO_U_F_FREE); |
f8b0bd10 JA |
1336 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
1337 | IO_U_F_TRIMMED | IO_U_F_BARRIER | | |
1338 | IO_U_F_VER_LIST); | |
0c6e7517 | 1339 | |
755200a3 | 1340 | io_u->error = 0; |
bcd5abfa | 1341 | io_u->acct_ddir = -1; |
10ba535a | 1342 | td->cur_depth++; |
0c41214f | 1343 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1344 | io_u->ipo = NULL; |
1dec3e07 JA |
1345 | } else if (td->o.verify_async) { |
1346 | /* | |
1347 | * We ran out, wait for async verify threads to finish and | |
1348 | * return one | |
1349 | */ | |
1350 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1351 | goto again; | |
10ba535a JA |
1352 | } |
1353 | ||
e8462bd8 | 1354 | td_io_u_unlock(td); |
10ba535a JA |
1355 | return io_u; |
1356 | } | |
1357 | ||
0d29de83 | 1358 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1359 | { |
d72be545 JA |
1360 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1361 | return 0; | |
1362 | ||
1363 | if (td->trim_entries) { | |
0d29de83 | 1364 | int get_trim = 0; |
10ba535a | 1365 | |
0d29de83 JA |
1366 | if (td->trim_batch) { |
1367 | td->trim_batch--; | |
1368 | get_trim = 1; | |
1369 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1370 | td->last_ddir != DDIR_READ) { | |
1371 | td->trim_batch = td->o.trim_batch; | |
1372 | if (!td->trim_batch) | |
1373 | td->trim_batch = td->o.trim_backlog; | |
1374 | get_trim = 1; | |
1375 | } | |
1376 | ||
1377 | if (get_trim && !get_next_trim(td, io_u)) | |
1378 | return 1; | |
2ba1c290 | 1379 | } |
10ba535a | 1380 | |
0d29de83 JA |
1381 | return 0; |
1382 | } | |
1383 | ||
1384 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1385 | { | |
d72be545 JA |
1386 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1387 | return 0; | |
1388 | ||
1389 | if (td->io_hist_len) { | |
9e144189 JA |
1390 | int get_verify = 0; |
1391 | ||
d1ece0c7 | 1392 | if (td->verify_batch) |
9e144189 | 1393 | get_verify = 1; |
d1ece0c7 | 1394 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1395 | td->last_ddir != DDIR_READ) { |
1396 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1397 | if (!td->verify_batch) |
1398 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1399 | get_verify = 1; |
1400 | } | |
1401 | ||
d1ece0c7 JA |
1402 | if (get_verify && !get_next_verify(td, io_u)) { |
1403 | td->verify_batch--; | |
0d29de83 | 1404 | return 1; |
d1ece0c7 | 1405 | } |
9e144189 JA |
1406 | } |
1407 | ||
0d29de83 JA |
1408 | return 0; |
1409 | } | |
1410 | ||
de789769 JA |
1411 | /* |
1412 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1413 | * easy compressible data for simple de-dupe attempts. Do this for every |
1414 | * 512b block in the range, since that should be the smallest block size | |
1415 | * we can expect from a device. | |
de789769 JA |
1416 | */ |
1417 | static void small_content_scramble(struct io_u *io_u) | |
1418 | { | |
23f394d5 | 1419 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1420 | uint64_t boffset; |
23f394d5 JA |
1421 | unsigned int offset; |
1422 | void *p, *end; | |
de789769 | 1423 | |
23f394d5 JA |
1424 | if (!nr_blocks) |
1425 | return; | |
1426 | ||
1427 | p = io_u->xfer_buf; | |
fba76ee8 | 1428 | boffset = io_u->offset; |
81f0366c | 1429 | io_u->buf_filled_len = 0; |
fad82f76 | 1430 | |
23f394d5 JA |
1431 | for (i = 0; i < nr_blocks; i++) { |
1432 | /* | |
1433 | * Fill the byte offset into a "random" start offset of | |
1434 | * the buffer, given by the product of the usec time | |
1435 | * and the actual offset. | |
1436 | */ | |
fad82f76 | 1437 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1438 | offset &= ~(sizeof(uint64_t) - 1); |
1439 | if (offset >= 512 - sizeof(uint64_t)) | |
1440 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1441 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1442 | |
1443 | end = p + 512 - sizeof(io_u->start_time); | |
1444 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1445 | p += 512; | |
fad82f76 | 1446 | boffset += 512; |
23f394d5 | 1447 | } |
de789769 JA |
1448 | } |
1449 | ||
0d29de83 JA |
1450 | /* |
1451 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1452 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1453 | */ | |
1454 | struct io_u *get_io_u(struct thread_data *td) | |
1455 | { | |
1456 | struct fio_file *f; | |
1457 | struct io_u *io_u; | |
de789769 | 1458 | int do_scramble = 0; |
002fe734 | 1459 | long ret = 0; |
0d29de83 JA |
1460 | |
1461 | io_u = __get_io_u(td); | |
1462 | if (!io_u) { | |
1463 | dprint(FD_IO, "__get_io_u failed\n"); | |
1464 | return NULL; | |
1465 | } | |
1466 | ||
1467 | if (check_get_verify(td, io_u)) | |
1468 | goto out; | |
1469 | if (check_get_trim(td, io_u)) | |
1470 | goto out; | |
1471 | ||
755200a3 JA |
1472 | /* |
1473 | * from a requeue, io_u already setup | |
1474 | */ | |
1475 | if (io_u->file) | |
77f392bf | 1476 | goto out; |
755200a3 | 1477 | |
429f6675 JA |
1478 | /* |
1479 | * If using an iolog, grab next piece if any available. | |
1480 | */ | |
d72be545 | 1481 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1482 | if (read_iolog_get(td, io_u)) |
1483 | goto err_put; | |
2ba1c290 | 1484 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1485 | ret = -EBUSY; |
2ba1c290 | 1486 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1487 | goto err_put; |
2ba1c290 | 1488 | } |
5ec10eaa | 1489 | |
429f6675 | 1490 | f = io_u->file; |
002fe734 JA |
1491 | if (!f) { |
1492 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1493 | goto err_put; | |
1494 | } | |
1495 | ||
d6aed795 | 1496 | assert(fio_file_open(f)); |
97af62ce | 1497 | |
ff58fced | 1498 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1499 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1500 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1501 | goto err_put; |
2ba1c290 | 1502 | } |
10ba535a | 1503 | |
38dad62d | 1504 | f->last_start = io_u->offset; |
36167d82 | 1505 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1506 | |
fd68418e | 1507 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1508 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1509 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1510 | td->o.min_bs[DDIR_WRITE], |
1511 | io_u->xfer_buflen); | |
bedc9dc2 JA |
1512 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
1513 | !(td->flags & TD_F_COMPRESS)) | |
fd68418e | 1514 | do_scramble = 1; |
d72be545 | 1515 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1516 | populate_verify_io_u(td, io_u); |
1517 | do_scramble = 0; | |
1518 | } | |
fd68418e | 1519 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1520 | /* |
1521 | * Reset the buf_filled parameters so next time if the | |
1522 | * buffer is used for writes it is refilled. | |
1523 | */ | |
cbe8d756 RR |
1524 | io_u->buf_filled_len = 0; |
1525 | } | |
87dc1ab1 | 1526 | } |
10ba535a | 1527 | |
165faf16 JA |
1528 | /* |
1529 | * Set io data pointers. | |
1530 | */ | |
cec6b55d JA |
1531 | io_u->xfer_buf = io_u->buf; |
1532 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1533 | |
6ac7a331 | 1534 | out: |
0d29de83 | 1535 | assert(io_u->file); |
429f6675 | 1536 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1537 | if (!td->o.disable_slat) |
1538 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1539 | if (do_scramble) |
1540 | small_content_scramble(io_u); | |
429f6675 | 1541 | return io_u; |
36167d82 | 1542 | } |
429f6675 | 1543 | err_put: |
2ba1c290 | 1544 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1545 | put_io_u(td, io_u); |
002fe734 | 1546 | return ERR_PTR(ret); |
10ba535a JA |
1547 | } |
1548 | ||
5451792e JA |
1549 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1550 | { | |
8b28bd41 | 1551 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1552 | |
8b28bd41 DM |
1553 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1554 | return; | |
5451792e | 1555 | |
709c8313 RE |
1556 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1557 | io_u->file ? " on file " : "", | |
1558 | io_u->file ? io_u->file->file_name : "", | |
1559 | strerror(io_u->error), | |
1560 | io_ddir_name(io_u->ddir), | |
1561 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1562 | |
1563 | if (!td->error) | |
1564 | td_verror(td, io_u->error, "io_u error"); | |
1565 | } | |
1566 | ||
aba6c951 JA |
1567 | static inline int gtod_reduce(struct thread_data *td) |
1568 | { | |
729fe3af | 1569 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1570 | && td->o.disable_bw; |
aba6c951 JA |
1571 | } |
1572 | ||
c8eeb9df JA |
1573 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1574 | struct io_completion_data *icd, | |
1575 | const enum fio_ddir idx, unsigned int bytes) | |
1576 | { | |
24d23ca7 | 1577 | unsigned long lusec = 0; |
c8eeb9df | 1578 | |
aba6c951 | 1579 | if (!gtod_reduce(td)) |
c8eeb9df JA |
1580 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1581 | ||
1582 | if (!td->o.disable_lat) { | |
1583 | unsigned long tusec; | |
1584 | ||
1585 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1586 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1587 | |
d4afedfd JA |
1588 | if (td->flags & TD_F_PROFILE_OPS) { |
1589 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1590 | ||
1591 | if (ops->io_u_lat) | |
1592 | icd->error = ops->io_u_lat(td, tusec); | |
1593 | } | |
1594 | ||
3e260a46 JA |
1595 | if (td->o.max_latency && tusec > td->o.max_latency) |
1596 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1597 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1598 | if (lat_target_failed(td)) | |
1599 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1600 | } |
c8eeb9df JA |
1601 | } |
1602 | ||
1603 | if (!td->o.disable_clat) { | |
ae588852 | 1604 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1605 | io_u_mark_latency(td, lusec); |
1606 | } | |
1607 | ||
1608 | if (!td->o.disable_bw) | |
1609 | add_bw_sample(td, idx, bytes, &icd->time); | |
1610 | ||
aba6c951 JA |
1611 | if (!gtod_reduce(td)) |
1612 | add_iops_sample(td, idx, bytes, &icd->time); | |
c8eeb9df JA |
1613 | } |
1614 | ||
1b8dbf25 SL |
1615 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1616 | { | |
1ae83d45 JA |
1617 | uint64_t secs, remainder, bps, bytes; |
1618 | ||
1b8dbf25 SL |
1619 | bytes = td->this_io_bytes[ddir]; |
1620 | bps = td->rate_bps[ddir]; | |
1621 | secs = bytes / bps; | |
1622 | remainder = bytes % bps; | |
1623 | return remainder * 1000000 / bps + secs * 1000000; | |
1624 | } | |
1625 | ||
f8b0bd10 | 1626 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1627 | struct io_completion_data *icd) |
10ba535a | 1628 | { |
f8b0bd10 JA |
1629 | struct io_u *io_u = *io_u_ptr; |
1630 | enum fio_ddir ddir = io_u->ddir; | |
1631 | struct fio_file *f = io_u->file; | |
10ba535a | 1632 | |
2ba1c290 JA |
1633 | dprint_io_u(io_u, "io complete"); |
1634 | ||
2ecc1b57 | 1635 | td_io_u_lock(td); |
0c6e7517 | 1636 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1637 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1638 | |
1639 | /* | |
1640 | * Mark IO ok to verify | |
1641 | */ | |
1642 | if (io_u->ipo) { | |
890b6656 JA |
1643 | /* |
1644 | * Remove errored entry from the verification list | |
1645 | */ | |
1646 | if (io_u->error) | |
1647 | unlog_io_piece(td, io_u); | |
1648 | else { | |
1649 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1650 | write_barrier(); | |
1651 | } | |
f9401285 JA |
1652 | } |
1653 | ||
2ecc1b57 | 1654 | td_io_u_unlock(td); |
0c6e7517 | 1655 | |
f8b0bd10 | 1656 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1657 | td->last_was_sync = 1; |
44f29692 JA |
1658 | if (f) { |
1659 | f->first_write = -1ULL; | |
1660 | f->last_write = -1ULL; | |
1661 | } | |
87dc1ab1 JA |
1662 | return; |
1663 | } | |
1664 | ||
1665 | td->last_was_sync = 0; | |
f8b0bd10 | 1666 | td->last_ddir = ddir; |
87dc1ab1 | 1667 | |
f8b0bd10 | 1668 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1669 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1670 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1671 | int ret; |
10ba535a | 1672 | |
f8b0bd10 JA |
1673 | td->io_blocks[ddir]++; |
1674 | td->this_io_blocks[ddir]++; | |
1675 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1676 | |
1677 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1678 | td->this_io_bytes[ddir] += bytes; |
1679 | ||
1680 | if (ddir == DDIR_WRITE && f) { | |
1681 | if (f->first_write == -1ULL || | |
1682 | io_u->offset < f->first_write) | |
1683 | f->first_write = io_u->offset; | |
1684 | if (f->last_write == -1ULL || | |
1685 | ((io_u->offset + bytes) > f->last_write)) | |
1686 | f->last_write = io_u->offset + bytes; | |
44f29692 JA |
1687 | } |
1688 | ||
6b1190fd SL |
1689 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1690 | td->runstate == TD_VERIFYING)) { | |
f8b0bd10 | 1691 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1692 | |
f8b0bd10 JA |
1693 | if (__should_check_rate(td, ddir)) { |
1694 | td->rate_pending_usleep[ddir] = | |
1695 | (usec_for_io(td, ddir) - | |
ba3e4e0c | 1696 | utime_since_now(&td->start)); |
b23b6a2f | 1697 | } |
f8b0bd10 JA |
1698 | if (ddir != DDIR_TRIM && |
1699 | __should_check_rate(td, oddir)) { | |
1700 | td->rate_pending_usleep[oddir] = | |
1701 | (usec_for_io(td, oddir) - | |
ba3e4e0c | 1702 | utime_since_now(&td->start)); |
f8b0bd10 | 1703 | } |
721938ae | 1704 | } |
10ba535a | 1705 | |
f8b0bd10 | 1706 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1707 | |
d7762cf8 | 1708 | if (io_u->end_io) { |
f8b0bd10 JA |
1709 | ret = io_u->end_io(td, io_u_ptr); |
1710 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1711 | if (ret && !icd->error) |
1712 | icd->error = ret; | |
1713 | } | |
ff58fced | 1714 | } else if (io_u->error) { |
10ba535a | 1715 | icd->error = io_u->error; |
5451792e JA |
1716 | io_u_log_error(td, io_u); |
1717 | } | |
8b28bd41 | 1718 | if (icd->error) { |
f8b0bd10 JA |
1719 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1720 | ||
8b28bd41 DM |
1721 | if (!td_non_fatal_error(td, eb, icd->error)) |
1722 | return; | |
f8b0bd10 | 1723 | |
f2bba182 RR |
1724 | /* |
1725 | * If there is a non_fatal error, then add to the error count | |
1726 | * and clear all the errors. | |
1727 | */ | |
1728 | update_error_count(td, icd->error); | |
1729 | td_clear_error(td); | |
1730 | icd->error = 0; | |
f8b0bd10 JA |
1731 | if (io_u) |
1732 | io_u->error = 0; | |
f2bba182 | 1733 | } |
10ba535a JA |
1734 | } |
1735 | ||
9520ebb9 JA |
1736 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1737 | int nr) | |
10ba535a | 1738 | { |
6eaf09d6 | 1739 | int ddir; |
aba6c951 JA |
1740 | |
1741 | if (!gtod_reduce(td)) | |
9520ebb9 | 1742 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1743 | |
3af6ef39 JA |
1744 | icd->nr = nr; |
1745 | ||
10ba535a | 1746 | icd->error = 0; |
6eaf09d6 SL |
1747 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1748 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1749 | } |
1750 | ||
97601024 JA |
1751 | static void ios_completed(struct thread_data *td, |
1752 | struct io_completion_data *icd) | |
36167d82 JA |
1753 | { |
1754 | struct io_u *io_u; | |
1755 | int i; | |
1756 | ||
10ba535a JA |
1757 | for (i = 0; i < icd->nr; i++) { |
1758 | io_u = td->io_ops->event(td, i); | |
1759 | ||
f8b0bd10 | 1760 | io_completed(td, &io_u, icd); |
e8462bd8 | 1761 | |
f8b0bd10 | 1762 | if (io_u) |
e8462bd8 | 1763 | put_io_u(td, io_u); |
10ba535a JA |
1764 | } |
1765 | } | |
97601024 | 1766 | |
e7e6cfb4 JA |
1767 | /* |
1768 | * Complete a single io_u for the sync engines. | |
1769 | */ | |
581e7141 | 1770 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1771 | uint64_t *bytes) |
97601024 JA |
1772 | { |
1773 | struct io_completion_data icd; | |
1774 | ||
9520ebb9 | 1775 | init_icd(td, &icd, 1); |
f8b0bd10 | 1776 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1777 | |
f8b0bd10 | 1778 | if (io_u) |
e8462bd8 | 1779 | put_io_u(td, io_u); |
97601024 | 1780 | |
581e7141 JA |
1781 | if (icd.error) { |
1782 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1783 | return -1; | |
1784 | } | |
97601024 | 1785 | |
581e7141 | 1786 | if (bytes) { |
6eaf09d6 SL |
1787 | int ddir; |
1788 | ||
1789 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1790 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1791 | } |
1792 | ||
1793 | return 0; | |
97601024 JA |
1794 | } |
1795 | ||
e7e6cfb4 JA |
1796 | /* |
1797 | * Called to complete min_events number of io for the async engines. | |
1798 | */ | |
581e7141 | 1799 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1800 | uint64_t *bytes) |
97601024 | 1801 | { |
97601024 | 1802 | struct io_completion_data icd; |
00de55ef | 1803 | struct timespec *tvp = NULL; |
97601024 | 1804 | int ret; |
4d06a338 | 1805 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1806 | |
4950421a | 1807 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1808 | |
4950421a | 1809 | if (!min_evts) |
00de55ef | 1810 | tvp = &ts; |
5fb4b366 RE |
1811 | else if (min_evts > td->cur_depth) |
1812 | min_evts = td->cur_depth; | |
97601024 | 1813 | |
4950421a | 1814 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1815 | if (ret < 0) { |
e1161c32 | 1816 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1817 | return ret; |
1818 | } else if (!ret) | |
1819 | return ret; | |
1820 | ||
9520ebb9 | 1821 | init_icd(td, &icd, ret); |
97601024 | 1822 | ios_completed(td, &icd); |
581e7141 JA |
1823 | if (icd.error) { |
1824 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1825 | return -1; | |
1826 | } | |
97601024 | 1827 | |
581e7141 | 1828 | if (bytes) { |
6eaf09d6 SL |
1829 | int ddir; |
1830 | ||
1831 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1832 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1833 | } |
1834 | ||
1835 | return 0; | |
97601024 | 1836 | } |
7e77dd02 JA |
1837 | |
1838 | /* | |
1839 | * Call when io_u is really queued, to update the submission latency. | |
1840 | */ | |
1841 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1842 | { | |
9520ebb9 JA |
1843 | if (!td->o.disable_slat) { |
1844 | unsigned long slat_time; | |
7e77dd02 | 1845 | |
9520ebb9 | 1846 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1847 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1848 | io_u->offset); | |
9520ebb9 | 1849 | } |
7e77dd02 | 1850 | } |
433afcb4 | 1851 | |
5c94b008 JA |
1852 | /* |
1853 | * See if we should reuse the last seed, if dedupe is enabled | |
1854 | */ | |
1855 | static struct frand_state *get_buf_state(struct thread_data *td) | |
1856 | { | |
1857 | unsigned int v; | |
1858 | unsigned long r; | |
1859 | ||
1860 | if (!td->o.dedupe_percentage) | |
1861 | return &td->buf_state; | |
c0b69b92 JA |
1862 | else if (td->o.dedupe_percentage == 100) |
1863 | return &td->buf_state_prev; | |
5c94b008 JA |
1864 | |
1865 | r = __rand(&td->dedupe_state); | |
1866 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1867 | ||
1868 | if (v <= td->o.dedupe_percentage) | |
1869 | return &td->buf_state_prev; | |
1870 | ||
1871 | return &td->buf_state; | |
1872 | } | |
1873 | ||
1874 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) | |
1875 | { | |
1876 | if (rs == &td->buf_state) | |
1877 | frand_copy(&td->buf_state_prev, rs); | |
1878 | } | |
1879 | ||
cc86c395 JA |
1880 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1881 | unsigned int max_bs) | |
5973cafb | 1882 | { |
ce35b1ec JA |
1883 | if (td->o.buffer_pattern_bytes) |
1884 | fill_buffer_pattern(td, buf, max_bs); | |
1885 | else if (!td->o.zero_buffers) { | |
9c42684e | 1886 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1887 | struct frand_state *rs; |
1066358a | 1888 | unsigned int left = max_bs; |
5c94b008 | 1889 | |
1066358a JA |
1890 | do { |
1891 | rs = get_buf_state(td); | |
9c42684e | 1892 | |
1066358a | 1893 | min_write = min(min_write, left); |
f97a43a1 | 1894 | |
1066358a JA |
1895 | if (perc) { |
1896 | unsigned int seg = min_write; | |
cc86c395 | 1897 | |
1066358a JA |
1898 | seg = min(min_write, td->o.compress_chunk); |
1899 | if (!seg) | |
1900 | seg = min_write; | |
1901 | ||
1902 | fill_random_buf_percentage(rs, buf, perc, seg, | |
1903 | min_write); | |
1904 | } else | |
1905 | fill_random_buf(rs, buf, min_write); | |
1906 | ||
1907 | buf += min_write; | |
1908 | left -= min_write; | |
5c94b008 | 1909 | save_buf_state(td, rs); |
1066358a | 1910 | } while (left); |
9c42684e | 1911 | } else |
cc86c395 JA |
1912 | memset(buf, 0, max_bs); |
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 | } |