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: | |
226 | r = flist_entry(td->next_rand_list.next, struct rand_off, list); | |
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 | { |
ff58fced JA |
274 | assert(ddir_rw(ddir)); |
275 | ||
bedc9dc2 | 276 | if (f->last_pos >= f->io_size + get_start_offset(td, f) && td->o.time_based) |
c04e4661 DE |
277 | f->last_pos = f->last_pos - f->io_size; |
278 | ||
38dad62d | 279 | if (f->last_pos < f->real_file_size) { |
1ae83d45 | 280 | uint64_t pos; |
059b0802 | 281 | |
a66da7a2 JA |
282 | if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0) |
283 | f->last_pos = f->real_file_size; | |
284 | ||
285 | pos = f->last_pos - f->file_offset; | |
059b0802 JA |
286 | if (pos) |
287 | pos += td->o.ddir_seq_add; | |
288 | ||
37cf9e3c | 289 | *offset = pos; |
38dad62d JA |
290 | return 0; |
291 | } | |
292 | ||
293 | return 1; | |
294 | } | |
295 | ||
296 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
297 | enum fio_ddir ddir, int rw_seq, |
298 | unsigned int *is_random) | |
38dad62d JA |
299 | { |
300 | struct fio_file *f = io_u->file; | |
1ae83d45 | 301 | uint64_t b, offset; |
38dad62d JA |
302 | int ret; |
303 | ||
ff58fced JA |
304 | assert(ddir_rw(ddir)); |
305 | ||
37cf9e3c JA |
306 | b = offset = -1ULL; |
307 | ||
38dad62d | 308 | if (rw_seq) { |
211c9b89 | 309 | if (td_random(td)) { |
6aca9b3d | 310 | if (should_do_random(td, ddir)) { |
211c9b89 | 311 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
312 | *is_random = 1; |
313 | } else { | |
314 | *is_random = 0; | |
211c9b89 JA |
315 | io_u->flags |= IO_U_F_BUSY_OK; |
316 | ret = get_next_seq_offset(td, f, ddir, &offset); | |
317 | if (ret) | |
318 | ret = get_next_rand_block(td, f, ddir, &b); | |
319 | } | |
6aca9b3d JA |
320 | } else { |
321 | *is_random = 0; | |
37cf9e3c | 322 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 323 | } |
38dad62d JA |
324 | } else { |
325 | io_u->flags |= IO_U_F_BUSY_OK; | |
6aca9b3d | 326 | *is_random = 0; |
38dad62d JA |
327 | |
328 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 329 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 330 | if (ret) { |
37cf9e3c | 331 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
332 | *is_random = 0; |
333 | } | |
38dad62d JA |
334 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
335 | if (f->last_start != -1ULL) | |
37cf9e3c | 336 | offset = f->last_start - f->file_offset; |
38dad62d | 337 | else |
37cf9e3c | 338 | offset = 0; |
38dad62d JA |
339 | ret = 0; |
340 | } else { | |
341 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
342 | ret = 1; | |
343 | } | |
344 | } | |
6d68b997 | 345 | |
37cf9e3c JA |
346 | if (!ret) { |
347 | if (offset != -1ULL) | |
348 | io_u->offset = offset; | |
349 | else if (b != -1ULL) | |
350 | io_u->offset = b * td->o.ba[ddir]; | |
351 | else { | |
4e0a8fa2 | 352 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
353 | ret = 1; |
354 | } | |
355 | } | |
356 | ||
38dad62d JA |
357 | return ret; |
358 | } | |
359 | ||
10ba535a JA |
360 | /* |
361 | * For random io, generate a random new block and see if it's used. Repeat | |
362 | * until we find a free one. For sequential io, just return the end of | |
363 | * the last io issued. | |
364 | */ | |
6aca9b3d JA |
365 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
366 | unsigned int *is_random) | |
10ba535a | 367 | { |
9bf2061e | 368 | struct fio_file *f = io_u->file; |
4ba66134 | 369 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 370 | int rw_seq_hit = 0; |
10ba535a | 371 | |
ff58fced JA |
372 | assert(ddir_rw(ddir)); |
373 | ||
38dad62d JA |
374 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
375 | rw_seq_hit = 1; | |
5736c10d | 376 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 377 | } |
211097b2 | 378 | |
6aca9b3d | 379 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 380 | return 1; |
10ba535a | 381 | |
009bd847 JA |
382 | if (io_u->offset >= f->io_size) { |
383 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
384 | (unsigned long long) io_u->offset, |
385 | (unsigned long long) f->io_size); | |
009bd847 JA |
386 | return 1; |
387 | } | |
388 | ||
389 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
390 | if (io_u->offset >= f->real_file_size) { |
391 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
392 | (unsigned long long) io_u->offset, |
393 | (unsigned long long) f->real_file_size); | |
10ba535a | 394 | return 1; |
2ba1c290 | 395 | } |
10ba535a JA |
396 | |
397 | return 0; | |
398 | } | |
399 | ||
6aca9b3d JA |
400 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
401 | unsigned int *is_random) | |
15dc1934 | 402 | { |
d72be545 JA |
403 | if (td->flags & TD_F_PROFILE_OPS) { |
404 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 405 | |
d72be545 | 406 | if (ops->fill_io_u_off) |
6aca9b3d | 407 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 408 | } |
15dc1934 | 409 | |
6aca9b3d | 410 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
411 | } |
412 | ||
79944128 JA |
413 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
414 | unsigned int buflen) | |
415 | { | |
416 | struct fio_file *f = io_u->file; | |
417 | ||
bedc9dc2 | 418 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
419 | } |
420 | ||
6aca9b3d JA |
421 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
422 | unsigned int is_random) | |
10ba535a | 423 | { |
6aca9b3d | 424 | int ddir = io_u->ddir; |
24d23ca7 | 425 | unsigned int buflen = 0; |
f3059de1 | 426 | unsigned int minbs, maxbs; |
1294c3ec | 427 | unsigned long r, rand_max; |
10ba535a | 428 | |
6aca9b3d JA |
429 | assert(ddir_rw(io_u->ddir)); |
430 | ||
431 | if (td->o.bs_is_seq_rand) | |
432 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
433 | else | |
434 | ddir = io_u->ddir; | |
ff58fced | 435 | |
f3059de1 JA |
436 | minbs = td->o.min_bs[ddir]; |
437 | maxbs = td->o.max_bs[ddir]; | |
438 | ||
79944128 JA |
439 | if (minbs == maxbs) |
440 | return minbs; | |
441 | ||
52c58027 JA |
442 | /* |
443 | * If we can't satisfy the min block size from here, then fail | |
444 | */ | |
445 | if (!io_u_fits(td, io_u, minbs)) | |
446 | return 0; | |
447 | ||
4c07ad86 JA |
448 | if (td->o.use_os_rand) |
449 | rand_max = OS_RAND_MAX; | |
450 | else | |
451 | rand_max = FRAND_MAX; | |
452 | ||
79944128 | 453 | do { |
4c07ad86 JA |
454 | if (td->o.use_os_rand) |
455 | r = os_random_long(&td->bsrange_state); | |
456 | else | |
457 | r = __rand(&td->__bsrange_state); | |
458 | ||
720e84ad | 459 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 460 | buflen = 1 + (unsigned int) ((double) maxbs * |
4c07ad86 | 461 | (r / (rand_max + 1.0))); |
f3059de1 JA |
462 | if (buflen < minbs) |
463 | buflen = minbs; | |
5ec10eaa | 464 | } else { |
564ca972 JA |
465 | long perc = 0; |
466 | unsigned int i; | |
467 | ||
720e84ad JA |
468 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
469 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
470 | |
471 | buflen = bsp->bs; | |
472 | perc += bsp->perc; | |
79944128 JA |
473 | if ((r <= ((rand_max / 100L) * perc)) && |
474 | io_u_fits(td, io_u, buflen)) | |
564ca972 JA |
475 | break; |
476 | } | |
477 | } | |
79944128 | 478 | |
a9f70b1f JB |
479 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
480 | buflen = (buflen + td->o.verify_interval - 1) & | |
481 | ~(td->o.verify_interval - 1); | |
482 | ||
f3059de1 JA |
483 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
484 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 485 | |
79944128 | 486 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 487 | |
10ba535a JA |
488 | return buflen; |
489 | } | |
490 | ||
6aca9b3d JA |
491 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
492 | unsigned int is_random) | |
15dc1934 | 493 | { |
d72be545 JA |
494 | if (td->flags & TD_F_PROFILE_OPS) { |
495 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 496 | |
d72be545 | 497 | if (ops->fill_io_u_size) |
6aca9b3d | 498 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 499 | } |
15dc1934 | 500 | |
6aca9b3d | 501 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
502 | } |
503 | ||
afe24a5a JA |
504 | static void set_rwmix_bytes(struct thread_data *td) |
505 | { | |
afe24a5a JA |
506 | unsigned int diff; |
507 | ||
508 | /* | |
509 | * we do time or byte based switch. this is needed because | |
510 | * buffered writes may issue a lot quicker than they complete, | |
511 | * whereas reads do not. | |
512 | */ | |
e47f799f | 513 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 514 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
515 | } |
516 | ||
517 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
518 | { | |
519 | unsigned int v; | |
1294c3ec | 520 | unsigned long r; |
e47f799f | 521 | |
4c07ad86 JA |
522 | if (td->o.use_os_rand) { |
523 | r = os_random_long(&td->rwmix_state); | |
524 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); | |
525 | } else { | |
526 | r = __rand(&td->__rwmix_state); | |
527 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
528 | } | |
529 | ||
04c540d9 | 530 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
531 | return DDIR_READ; |
532 | ||
533 | return DDIR_WRITE; | |
afe24a5a JA |
534 | } |
535 | ||
002e7183 JA |
536 | void io_u_quiesce(struct thread_data *td) |
537 | { | |
538 | /* | |
539 | * We are going to sleep, ensure that we flush anything pending as | |
540 | * not to skew our latency numbers. | |
541 | * | |
542 | * Changed to only monitor 'in flight' requests here instead of the | |
543 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
544 | * io's that have been actually submitted to an async engine, | |
545 | * and cur_depth is meaningless for sync engines. | |
546 | */ | |
547 | while (td->io_u_in_flight) { | |
548 | int fio_unused ret; | |
549 | ||
550 | ret = io_u_queued_complete(td, 1, NULL); | |
551 | } | |
552 | } | |
553 | ||
581e7141 JA |
554 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
555 | { | |
556 | enum fio_ddir odir = ddir ^ 1; | |
557 | struct timeval t; | |
558 | long usec; | |
559 | ||
ff58fced JA |
560 | assert(ddir_rw(ddir)); |
561 | ||
315fcfec | 562 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
563 | return ddir; |
564 | ||
565 | /* | |
566 | * We have too much pending sleep in this direction. See if we | |
567 | * should switch. | |
568 | */ | |
315fcfec | 569 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
570 | /* |
571 | * Other direction does not have too much pending, switch | |
572 | */ | |
573 | if (td->rate_pending_usleep[odir] < 100000) | |
574 | return odir; | |
575 | ||
576 | /* | |
577 | * Both directions have pending sleep. Sleep the minimum time | |
578 | * and deduct from both. | |
579 | */ | |
580 | if (td->rate_pending_usleep[ddir] <= | |
581 | td->rate_pending_usleep[odir]) { | |
582 | usec = td->rate_pending_usleep[ddir]; | |
583 | } else { | |
584 | usec = td->rate_pending_usleep[odir]; | |
585 | ddir = odir; | |
586 | } | |
587 | } else | |
588 | usec = td->rate_pending_usleep[ddir]; | |
589 | ||
002e7183 | 590 | io_u_quiesce(td); |
78c1eda5 | 591 | |
581e7141 JA |
592 | fio_gettime(&t, NULL); |
593 | usec_sleep(td, usec); | |
594 | usec = utime_since_now(&t); | |
595 | ||
596 | td->rate_pending_usleep[ddir] -= usec; | |
597 | ||
598 | odir = ddir ^ 1; | |
599 | if (td_rw(td) && __should_check_rate(td, odir)) | |
600 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 601 | |
6eaf09d6 SL |
602 | if (ddir_trim(ddir)) |
603 | return ddir; | |
e0224c6b | 604 | |
581e7141 JA |
605 | return ddir; |
606 | } | |
607 | ||
10ba535a JA |
608 | /* |
609 | * Return the data direction for the next io_u. If the job is a | |
610 | * mixed read/write workload, check the rwmix cycle and switch if | |
611 | * necessary. | |
612 | */ | |
1e97cce9 | 613 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 614 | { |
581e7141 JA |
615 | enum fio_ddir ddir; |
616 | ||
5f9099ea JA |
617 | /* |
618 | * see if it's time to fsync | |
619 | */ | |
620 | if (td->o.fsync_blocks && | |
621 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
622 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
623 | return DDIR_SYNC; | |
624 | ||
625 | /* | |
626 | * see if it's time to fdatasync | |
627 | */ | |
628 | if (td->o.fdatasync_blocks && | |
629 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
630 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
631 | return DDIR_DATASYNC; | |
632 | ||
44f29692 JA |
633 | /* |
634 | * see if it's time to sync_file_range | |
635 | */ | |
636 | if (td->sync_file_range_nr && | |
637 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
638 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
639 | return DDIR_SYNC_FILE_RANGE; | |
640 | ||
10ba535a | 641 | if (td_rw(td)) { |
10ba535a JA |
642 | /* |
643 | * Check if it's time to seed a new data direction. | |
644 | */ | |
e4928662 | 645 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
646 | /* |
647 | * Put a top limit on how many bytes we do for | |
648 | * one data direction, to avoid overflowing the | |
649 | * ranges too much | |
650 | */ | |
651 | ddir = get_rand_ddir(td); | |
e47f799f JA |
652 | |
653 | if (ddir != td->rwmix_ddir) | |
654 | set_rwmix_bytes(td); | |
655 | ||
656 | td->rwmix_ddir = ddir; | |
10ba535a | 657 | } |
581e7141 | 658 | ddir = td->rwmix_ddir; |
10ba535a | 659 | } else if (td_read(td)) |
581e7141 | 660 | ddir = DDIR_READ; |
6eaf09d6 | 661 | else if (td_write(td)) |
581e7141 | 662 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
663 | else |
664 | ddir = DDIR_TRIM; | |
581e7141 JA |
665 | |
666 | td->rwmix_ddir = rate_ddir(td, ddir); | |
667 | return td->rwmix_ddir; | |
10ba535a JA |
668 | } |
669 | ||
1ef2b6be JA |
670 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
671 | { | |
bcd5abfa | 672 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
673 | |
674 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
675 | td->o.barrier_blocks && | |
676 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
677 | td->io_issues[DDIR_WRITE]) | |
678 | io_u->flags |= IO_U_F_BARRIER; | |
679 | } | |
680 | ||
e8462bd8 | 681 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 JA |
682 | { |
683 | int ret = put_file(td, f); | |
684 | ||
685 | if (ret) | |
686 | td_verror(td, ret, "file close"); | |
687 | } | |
688 | ||
10ba535a JA |
689 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
690 | { | |
e8462bd8 JA |
691 | td_io_u_lock(td); |
692 | ||
d7ee2a7d | 693 | if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF)) |
60f2c658 | 694 | put_file_log(td, io_u->file); |
10ba535a | 695 | io_u->file = NULL; |
d7ee2a7d SL |
696 | io_u->flags &= ~IO_U_F_FREE_DEF; |
697 | io_u->flags |= IO_U_F_FREE; | |
698 | ||
0c41214f RR |
699 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
700 | td->cur_depth--; | |
2ae0b204 | 701 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
702 | td_io_u_unlock(td); |
703 | td_io_u_free_notify(td); | |
10ba535a JA |
704 | } |
705 | ||
f2bba182 RR |
706 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
707 | { | |
708 | io_u->flags &= ~IO_U_F_FLIGHT; | |
709 | put_io_u(td, io_u); | |
710 | } | |
711 | ||
755200a3 JA |
712 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
713 | { | |
714 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 715 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 716 | |
465221b0 JA |
717 | dprint(FD_IO, "requeue %p\n", __io_u); |
718 | ||
e8462bd8 JA |
719 | td_io_u_lock(td); |
720 | ||
4d2e0f49 | 721 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
722 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
723 | td->io_issues[ddir]--; | |
5ec10eaa | 724 | |
4d2e0f49 | 725 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
726 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
727 | td->cur_depth--; | |
2ae0b204 JA |
728 | |
729 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 730 | td_io_u_unlock(td); |
755200a3 JA |
731 | *io_u = NULL; |
732 | } | |
733 | ||
9bf2061e | 734 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 735 | { |
6aca9b3d JA |
736 | unsigned int is_random; |
737 | ||
b4c5e1ac JA |
738 | if (td->io_ops->flags & FIO_NOIO) |
739 | goto out; | |
740 | ||
1ef2b6be | 741 | set_rw_ddir(td, io_u); |
5f9099ea | 742 | |
87dc1ab1 | 743 | /* |
ff58fced | 744 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 745 | */ |
ff58fced | 746 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 747 | goto out; |
a00735e6 | 748 | |
48f5abd3 JA |
749 | /* |
750 | * See if it's time to switch to a new zone | |
751 | */ | |
13af05ae | 752 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
48f5abd3 | 753 | td->zone_bytes = 0; |
ed335855 SN |
754 | io_u->file->file_offset += td->o.zone_range + td->o.zone_skip; |
755 | io_u->file->last_pos = io_u->file->file_offset; | |
48f5abd3 JA |
756 | td->io_skip_bytes += td->o.zone_skip; |
757 | } | |
758 | ||
10ba535a | 759 | /* |
c685b5b2 JA |
760 | * No log, let the seq/rand engine retrieve the next buflen and |
761 | * position. | |
10ba535a | 762 | */ |
6aca9b3d | 763 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 764 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 765 | return 1; |
2ba1c290 | 766 | } |
10ba535a | 767 | |
6aca9b3d | 768 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
769 | if (!io_u->buflen) { |
770 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 771 | return 1; |
2ba1c290 | 772 | } |
bca4ed4d | 773 | |
2ba1c290 JA |
774 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
775 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
776 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
777 | (unsigned long long) io_u->offset, io_u->buflen, | |
778 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 779 | return 1; |
2ba1c290 | 780 | } |
6a5e6884 | 781 | |
bca4ed4d JA |
782 | /* |
783 | * mark entry before potentially trimming io_u | |
784 | */ | |
303032ae | 785 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 786 | mark_random_map(td, io_u); |
bca4ed4d | 787 | |
c38e9468 | 788 | out: |
2ba1c290 | 789 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 790 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 791 | return 0; |
10ba535a JA |
792 | } |
793 | ||
838bc709 JA |
794 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
795 | { | |
2b13e716 | 796 | int idx = 0; |
838bc709 JA |
797 | |
798 | switch (nr) { | |
799 | default: | |
2b13e716 | 800 | idx = 6; |
838bc709 JA |
801 | break; |
802 | case 33 ... 64: | |
2b13e716 | 803 | idx = 5; |
838bc709 JA |
804 | break; |
805 | case 17 ... 32: | |
2b13e716 | 806 | idx = 4; |
838bc709 JA |
807 | break; |
808 | case 9 ... 16: | |
2b13e716 | 809 | idx = 3; |
838bc709 JA |
810 | break; |
811 | case 5 ... 8: | |
2b13e716 | 812 | idx = 2; |
838bc709 JA |
813 | break; |
814 | case 1 ... 4: | |
2b13e716 | 815 | idx = 1; |
838bc709 JA |
816 | case 0: |
817 | break; | |
818 | } | |
819 | ||
2b13e716 | 820 | map[idx]++; |
838bc709 JA |
821 | } |
822 | ||
823 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
824 | { | |
825 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
826 | td->ts.total_submit++; | |
827 | } | |
828 | ||
829 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
830 | { | |
831 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
832 | td->ts.total_complete++; | |
833 | } | |
834 | ||
d8005759 | 835 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 836 | { |
2b13e716 | 837 | int idx = 0; |
71619dc2 JA |
838 | |
839 | switch (td->cur_depth) { | |
840 | default: | |
2b13e716 | 841 | idx = 6; |
a783e61a | 842 | break; |
71619dc2 | 843 | case 32 ... 63: |
2b13e716 | 844 | idx = 5; |
a783e61a | 845 | break; |
71619dc2 | 846 | case 16 ... 31: |
2b13e716 | 847 | idx = 4; |
a783e61a | 848 | break; |
71619dc2 | 849 | case 8 ... 15: |
2b13e716 | 850 | idx = 3; |
a783e61a | 851 | break; |
71619dc2 | 852 | case 4 ... 7: |
2b13e716 | 853 | idx = 2; |
a783e61a | 854 | break; |
71619dc2 | 855 | case 2 ... 3: |
2b13e716 | 856 | idx = 1; |
71619dc2 JA |
857 | case 1: |
858 | break; | |
859 | } | |
860 | ||
2b13e716 | 861 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
862 | } |
863 | ||
04a0feae JA |
864 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
865 | { | |
2b13e716 | 866 | int idx = 0; |
04a0feae JA |
867 | |
868 | assert(usec < 1000); | |
869 | ||
870 | switch (usec) { | |
871 | case 750 ... 999: | |
2b13e716 | 872 | idx = 9; |
04a0feae JA |
873 | break; |
874 | case 500 ... 749: | |
2b13e716 | 875 | idx = 8; |
04a0feae JA |
876 | break; |
877 | case 250 ... 499: | |
2b13e716 | 878 | idx = 7; |
04a0feae JA |
879 | break; |
880 | case 100 ... 249: | |
2b13e716 | 881 | idx = 6; |
04a0feae JA |
882 | break; |
883 | case 50 ... 99: | |
2b13e716 | 884 | idx = 5; |
04a0feae JA |
885 | break; |
886 | case 20 ... 49: | |
2b13e716 | 887 | idx = 4; |
04a0feae JA |
888 | break; |
889 | case 10 ... 19: | |
2b13e716 | 890 | idx = 3; |
04a0feae JA |
891 | break; |
892 | case 4 ... 9: | |
2b13e716 | 893 | idx = 2; |
04a0feae JA |
894 | break; |
895 | case 2 ... 3: | |
2b13e716 | 896 | idx = 1; |
04a0feae JA |
897 | case 0 ... 1: |
898 | break; | |
899 | } | |
900 | ||
2b13e716 JA |
901 | assert(idx < FIO_IO_U_LAT_U_NR); |
902 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
903 | } |
904 | ||
905 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 906 | { |
2b13e716 | 907 | int idx = 0; |
ec118304 JA |
908 | |
909 | switch (msec) { | |
910 | default: | |
2b13e716 | 911 | idx = 11; |
04a0feae | 912 | break; |
8abdce66 | 913 | case 1000 ... 1999: |
2b13e716 | 914 | idx = 10; |
04a0feae | 915 | break; |
8abdce66 | 916 | case 750 ... 999: |
2b13e716 | 917 | idx = 9; |
04a0feae | 918 | break; |
8abdce66 | 919 | case 500 ... 749: |
2b13e716 | 920 | idx = 8; |
04a0feae | 921 | break; |
8abdce66 | 922 | case 250 ... 499: |
2b13e716 | 923 | idx = 7; |
04a0feae | 924 | break; |
8abdce66 | 925 | case 100 ... 249: |
2b13e716 | 926 | idx = 6; |
04a0feae | 927 | break; |
8abdce66 | 928 | case 50 ... 99: |
2b13e716 | 929 | idx = 5; |
04a0feae | 930 | break; |
8abdce66 | 931 | case 20 ... 49: |
2b13e716 | 932 | idx = 4; |
04a0feae | 933 | break; |
8abdce66 | 934 | case 10 ... 19: |
2b13e716 | 935 | idx = 3; |
04a0feae | 936 | break; |
8abdce66 | 937 | case 4 ... 9: |
2b13e716 | 938 | idx = 2; |
04a0feae | 939 | break; |
ec118304 | 940 | case 2 ... 3: |
2b13e716 | 941 | idx = 1; |
ec118304 JA |
942 | case 0 ... 1: |
943 | break; | |
944 | } | |
945 | ||
2b13e716 JA |
946 | assert(idx < FIO_IO_U_LAT_M_NR); |
947 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
948 | } |
949 | ||
950 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
951 | { | |
952 | if (usec < 1000) | |
953 | io_u_mark_lat_usec(td, usec); | |
954 | else | |
955 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
956 | } |
957 | ||
0aabe160 JA |
958 | /* |
959 | * Get next file to service by choosing one at random | |
960 | */ | |
2cc52930 JA |
961 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
962 | enum fio_file_flags goodf, | |
d6aed795 | 963 | enum fio_file_flags badf) |
0aabe160 | 964 | { |
0aabe160 | 965 | struct fio_file *f; |
1c178180 | 966 | int fno; |
0aabe160 JA |
967 | |
968 | do { | |
87b10676 | 969 | int opened = 0; |
1294c3ec | 970 | unsigned long r; |
4c07ad86 JA |
971 | |
972 | if (td->o.use_os_rand) { | |
973 | r = os_random_long(&td->next_file_state); | |
974 | fno = (unsigned int) ((double) td->o.nr_files | |
975 | * (r / (OS_RAND_MAX + 1.0))); | |
976 | } else { | |
977 | r = __rand(&td->__next_file_state); | |
978 | fno = (unsigned int) ((double) td->o.nr_files | |
979 | * (r / (FRAND_MAX + 1.0))); | |
980 | } | |
7c83c089 | 981 | |
126d65c6 | 982 | f = td->files[fno]; |
d6aed795 | 983 | if (fio_file_done(f)) |
059e63c0 | 984 | continue; |
1c178180 | 985 | |
d6aed795 | 986 | if (!fio_file_open(f)) { |
87b10676 JA |
987 | int err; |
988 | ||
002fe734 JA |
989 | if (td->nr_open_files >= td->o.open_files) |
990 | return ERR_PTR(-EBUSY); | |
991 | ||
87b10676 JA |
992 | err = td_io_open_file(td, f); |
993 | if (err) | |
994 | continue; | |
995 | opened = 1; | |
996 | } | |
997 | ||
2ba1c290 JA |
998 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
999 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1000 | return f; |
2ba1c290 | 1001 | } |
87b10676 JA |
1002 | if (opened) |
1003 | td_io_close_file(td, f); | |
0aabe160 JA |
1004 | } while (1); |
1005 | } | |
1006 | ||
1007 | /* | |
1008 | * Get next file to service by doing round robin between all available ones | |
1009 | */ | |
1c178180 JA |
1010 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1011 | int badf) | |
3d7c391d JA |
1012 | { |
1013 | unsigned int old_next_file = td->next_file; | |
1014 | struct fio_file *f; | |
1015 | ||
1016 | do { | |
87b10676 JA |
1017 | int opened = 0; |
1018 | ||
126d65c6 | 1019 | f = td->files[td->next_file]; |
3d7c391d JA |
1020 | |
1021 | td->next_file++; | |
2dc1bbeb | 1022 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1023 | td->next_file = 0; |
1024 | ||
87b10676 | 1025 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1026 | if (fio_file_done(f)) { |
d5ed68ea | 1027 | f = NULL; |
059e63c0 | 1028 | continue; |
d5ed68ea | 1029 | } |
059e63c0 | 1030 | |
d6aed795 | 1031 | if (!fio_file_open(f)) { |
87b10676 JA |
1032 | int err; |
1033 | ||
002fe734 JA |
1034 | if (td->nr_open_files >= td->o.open_files) |
1035 | return ERR_PTR(-EBUSY); | |
1036 | ||
87b10676 | 1037 | err = td_io_open_file(td, f); |
b5696bfc JA |
1038 | if (err) { |
1039 | dprint(FD_FILE, "error %d on open of %s\n", | |
1040 | err, f->file_name); | |
87c27b45 | 1041 | f = NULL; |
87b10676 | 1042 | continue; |
b5696bfc | 1043 | } |
87b10676 JA |
1044 | opened = 1; |
1045 | } | |
1046 | ||
0b9d69ec JA |
1047 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1048 | f->flags); | |
1c178180 | 1049 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1050 | break; |
1051 | ||
87b10676 JA |
1052 | if (opened) |
1053 | td_io_close_file(td, f); | |
1054 | ||
3d7c391d JA |
1055 | f = NULL; |
1056 | } while (td->next_file != old_next_file); | |
1057 | ||
2ba1c290 | 1058 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1059 | return f; |
1060 | } | |
1061 | ||
7eb36574 | 1062 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1063 | { |
1907dbc6 JA |
1064 | struct fio_file *f; |
1065 | ||
2dc1bbeb | 1066 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1067 | |
b5696bfc | 1068 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1069 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1070 | " nr_files=%d\n", td->nr_open_files, | |
1071 | td->nr_done_files, | |
1072 | td->o.nr_files); | |
bdb4e2e9 | 1073 | return NULL; |
2ba1c290 | 1074 | } |
bdb4e2e9 | 1075 | |
1907dbc6 | 1076 | f = td->file_service_file; |
d6aed795 | 1077 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1078 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1079 | goto out; | |
1080 | if (td->file_service_left--) | |
1081 | goto out; | |
1082 | } | |
1907dbc6 | 1083 | |
a086c257 JA |
1084 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1085 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1086 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1087 | else |
d6aed795 | 1088 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1089 | |
002fe734 JA |
1090 | if (IS_ERR(f)) |
1091 | return f; | |
1092 | ||
1907dbc6 JA |
1093 | td->file_service_file = f; |
1094 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1095 | out: |
0dac421f JA |
1096 | if (f) |
1097 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1098 | else | |
1099 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1100 | return f; |
bdb4e2e9 JA |
1101 | } |
1102 | ||
7eb36574 JA |
1103 | static struct fio_file *get_next_file(struct thread_data *td) |
1104 | { | |
d72be545 JA |
1105 | if (!(td->flags & TD_F_PROFILE_OPS)) { |
1106 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 1107 | |
d72be545 JA |
1108 | if (ops->get_next_file) |
1109 | return ops->get_next_file(td); | |
1110 | } | |
7eb36574 JA |
1111 | |
1112 | return __get_next_file(td); | |
1113 | } | |
1114 | ||
002fe734 | 1115 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1116 | { |
1117 | struct fio_file *f; | |
1118 | ||
1119 | do { | |
1120 | f = get_next_file(td); | |
002fe734 JA |
1121 | if (IS_ERR_OR_NULL(f)) |
1122 | return PTR_ERR(f); | |
429f6675 | 1123 | |
429f6675 JA |
1124 | io_u->file = f; |
1125 | get_file(f); | |
1126 | ||
1127 | if (!fill_io_u(td, io_u)) | |
1128 | break; | |
1129 | ||
b5696bfc | 1130 | put_file_log(td, f); |
429f6675 | 1131 | td_io_close_file(td, f); |
b5696bfc | 1132 | io_u->file = NULL; |
d6aed795 | 1133 | fio_file_set_done(f); |
429f6675 | 1134 | td->nr_done_files++; |
0b9d69ec JA |
1135 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1136 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1137 | } while (1); |
1138 | ||
1139 | return 0; | |
1140 | } | |
1141 | ||
3e260a46 JA |
1142 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1143 | unsigned long tusec, unsigned long max_usec) | |
1144 | { | |
1145 | if (!td->error) | |
1146 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1147 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1148 | icd->error = ETIMEDOUT; | |
1149 | } | |
1150 | ||
1151 | static void lat_new_cycle(struct thread_data *td) | |
1152 | { | |
1153 | fio_gettime(&td->latency_ts, NULL); | |
1154 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1155 | td->latency_failed = 0; | |
1156 | } | |
1157 | ||
1158 | /* | |
1159 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1160 | * are at QD=1, then it's time to give up. | |
1161 | */ | |
1162 | static int __lat_target_failed(struct thread_data *td) | |
1163 | { | |
1164 | if (td->latency_qd == 1) | |
1165 | return 1; | |
1166 | ||
1167 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1168 | |
1169 | if (td->latency_qd == td->latency_qd_low) | |
1170 | td->latency_qd_low--; | |
1171 | ||
3e260a46 JA |
1172 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1173 | ||
1174 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1175 | ||
1176 | /* | |
1177 | * When we ramp QD down, quiesce existing IO to prevent | |
1178 | * a storm of ramp downs due to pending higher depth. | |
1179 | */ | |
1180 | io_u_quiesce(td); | |
1181 | lat_new_cycle(td); | |
1182 | return 0; | |
1183 | } | |
1184 | ||
1185 | static int lat_target_failed(struct thread_data *td) | |
1186 | { | |
1187 | if (td->o.latency_percentile.u.f == 100.0) | |
1188 | return __lat_target_failed(td); | |
1189 | ||
1190 | td->latency_failed++; | |
1191 | return 0; | |
1192 | } | |
1193 | ||
1194 | void lat_target_init(struct thread_data *td) | |
1195 | { | |
6bb58215 JA |
1196 | td->latency_end_run = 0; |
1197 | ||
3e260a46 JA |
1198 | if (td->o.latency_target) { |
1199 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1200 | fio_gettime(&td->latency_ts, NULL); | |
1201 | td->latency_qd = 1; | |
1202 | td->latency_qd_high = td->o.iodepth; | |
1203 | td->latency_qd_low = 1; | |
1204 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1205 | } else | |
1206 | td->latency_qd = td->o.iodepth; | |
1207 | } | |
1208 | ||
6bb58215 JA |
1209 | void lat_target_reset(struct thread_data *td) |
1210 | { | |
1211 | if (!td->latency_end_run) | |
1212 | lat_target_init(td); | |
1213 | } | |
1214 | ||
3e260a46 JA |
1215 | static void lat_target_success(struct thread_data *td) |
1216 | { | |
1217 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1218 | struct thread_options *o = &td->o; |
3e260a46 JA |
1219 | |
1220 | td->latency_qd_low = td->latency_qd; | |
1221 | ||
1222 | /* | |
1223 | * If we haven't failed yet, we double up to a failing value instead | |
1224 | * of bisecting from highest possible queue depth. If we have set | |
1225 | * a limit other than td->o.iodepth, bisect between that. | |
1226 | */ | |
6bb58215 | 1227 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1228 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1229 | else | |
1230 | td->latency_qd *= 2; | |
1231 | ||
6bb58215 JA |
1232 | if (td->latency_qd > o->iodepth) |
1233 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1234 | |
1235 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1236 | |
3e260a46 | 1237 | /* |
6bb58215 JA |
1238 | * Same as last one, we are done. Let it run a latency cycle, so |
1239 | * we get only the results from the targeted depth. | |
3e260a46 | 1240 | */ |
6bb58215 JA |
1241 | if (td->latency_qd == qd) { |
1242 | if (td->latency_end_run) { | |
1243 | dprint(FD_RATE, "We are done\n"); | |
1244 | td->done = 1; | |
1245 | } else { | |
1246 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1247 | io_u_quiesce(td); | |
1248 | td->latency_end_run = 1; | |
1249 | reset_all_stats(td); | |
1250 | reset_io_stats(td); | |
1251 | } | |
1252 | } | |
3e260a46 JA |
1253 | |
1254 | lat_new_cycle(td); | |
1255 | } | |
1256 | ||
1257 | /* | |
1258 | * Check if we can bump the queue depth | |
1259 | */ | |
1260 | void lat_target_check(struct thread_data *td) | |
1261 | { | |
1262 | uint64_t usec_window; | |
1263 | uint64_t ios; | |
1264 | double success_ios; | |
1265 | ||
1266 | usec_window = utime_since_now(&td->latency_ts); | |
1267 | if (usec_window < td->o.latency_window) | |
1268 | return; | |
1269 | ||
1270 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1271 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1272 | success_ios *= 100.0; | |
1273 | ||
1274 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1275 | ||
1276 | if (success_ios >= td->o.latency_percentile.u.f) | |
1277 | lat_target_success(td); | |
1278 | else | |
1279 | __lat_target_failed(td); | |
1280 | } | |
1281 | ||
1282 | /* | |
1283 | * If latency target is enabled, we might be ramping up or down and not | |
1284 | * using the full queue depth available. | |
1285 | */ | |
1286 | int queue_full(struct thread_data *td) | |
1287 | { | |
1288 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1289 | ||
1290 | if (qempty) | |
1291 | return 1; | |
1292 | if (!td->o.latency_target) | |
1293 | return 0; | |
1294 | ||
1295 | return td->cur_depth >= td->latency_qd; | |
1296 | } | |
429f6675 | 1297 | |
10ba535a JA |
1298 | struct io_u *__get_io_u(struct thread_data *td) |
1299 | { | |
0cae66f6 | 1300 | struct io_u *io_u = NULL; |
10ba535a | 1301 | |
e8462bd8 JA |
1302 | td_io_u_lock(td); |
1303 | ||
1304 | again: | |
2ae0b204 JA |
1305 | if (!io_u_rempty(&td->io_u_requeues)) |
1306 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1307 | else if (!queue_full(td)) { |
2ae0b204 | 1308 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1309 | |
225ba9e3 | 1310 | io_u->file = NULL; |
6040dabc | 1311 | io_u->buflen = 0; |
10ba535a | 1312 | io_u->resid = 0; |
d7762cf8 | 1313 | io_u->end_io = NULL; |
755200a3 JA |
1314 | } |
1315 | ||
1316 | if (io_u) { | |
0c6e7517 | 1317 | assert(io_u->flags & IO_U_F_FREE); |
2ecc1b57 | 1318 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF); |
1ef2b6be | 1319 | io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER); |
82af2a7c | 1320 | io_u->flags &= ~IO_U_F_VER_LIST; |
0c6e7517 | 1321 | |
755200a3 | 1322 | io_u->error = 0; |
bcd5abfa | 1323 | io_u->acct_ddir = -1; |
10ba535a | 1324 | td->cur_depth++; |
0c41214f | 1325 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1326 | io_u->ipo = NULL; |
1dec3e07 JA |
1327 | } else if (td->o.verify_async) { |
1328 | /* | |
1329 | * We ran out, wait for async verify threads to finish and | |
1330 | * return one | |
1331 | */ | |
1332 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1333 | goto again; | |
10ba535a JA |
1334 | } |
1335 | ||
e8462bd8 | 1336 | td_io_u_unlock(td); |
10ba535a JA |
1337 | return io_u; |
1338 | } | |
1339 | ||
0d29de83 | 1340 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1341 | { |
d72be545 JA |
1342 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1343 | return 0; | |
1344 | ||
1345 | if (td->trim_entries) { | |
0d29de83 | 1346 | int get_trim = 0; |
10ba535a | 1347 | |
0d29de83 JA |
1348 | if (td->trim_batch) { |
1349 | td->trim_batch--; | |
1350 | get_trim = 1; | |
1351 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1352 | td->last_ddir != DDIR_READ) { | |
1353 | td->trim_batch = td->o.trim_batch; | |
1354 | if (!td->trim_batch) | |
1355 | td->trim_batch = td->o.trim_backlog; | |
1356 | get_trim = 1; | |
1357 | } | |
1358 | ||
1359 | if (get_trim && !get_next_trim(td, io_u)) | |
1360 | return 1; | |
2ba1c290 | 1361 | } |
10ba535a | 1362 | |
0d29de83 JA |
1363 | return 0; |
1364 | } | |
1365 | ||
1366 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1367 | { | |
d72be545 JA |
1368 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1369 | return 0; | |
1370 | ||
1371 | if (td->io_hist_len) { | |
9e144189 JA |
1372 | int get_verify = 0; |
1373 | ||
d1ece0c7 | 1374 | if (td->verify_batch) |
9e144189 | 1375 | get_verify = 1; |
d1ece0c7 | 1376 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1377 | td->last_ddir != DDIR_READ) { |
1378 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1379 | if (!td->verify_batch) |
1380 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1381 | get_verify = 1; |
1382 | } | |
1383 | ||
d1ece0c7 JA |
1384 | if (get_verify && !get_next_verify(td, io_u)) { |
1385 | td->verify_batch--; | |
0d29de83 | 1386 | return 1; |
d1ece0c7 | 1387 | } |
9e144189 JA |
1388 | } |
1389 | ||
0d29de83 JA |
1390 | return 0; |
1391 | } | |
1392 | ||
de789769 JA |
1393 | /* |
1394 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1395 | * easy compressible data for simple de-dupe attempts. Do this for every |
1396 | * 512b block in the range, since that should be the smallest block size | |
1397 | * we can expect from a device. | |
de789769 JA |
1398 | */ |
1399 | static void small_content_scramble(struct io_u *io_u) | |
1400 | { | |
23f394d5 | 1401 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1402 | uint64_t boffset; |
23f394d5 JA |
1403 | unsigned int offset; |
1404 | void *p, *end; | |
de789769 | 1405 | |
23f394d5 JA |
1406 | if (!nr_blocks) |
1407 | return; | |
1408 | ||
1409 | p = io_u->xfer_buf; | |
fba76ee8 | 1410 | boffset = io_u->offset; |
81f0366c | 1411 | io_u->buf_filled_len = 0; |
fad82f76 | 1412 | |
23f394d5 JA |
1413 | for (i = 0; i < nr_blocks; i++) { |
1414 | /* | |
1415 | * Fill the byte offset into a "random" start offset of | |
1416 | * the buffer, given by the product of the usec time | |
1417 | * and the actual offset. | |
1418 | */ | |
fad82f76 | 1419 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1420 | offset &= ~(sizeof(uint64_t) - 1); |
1421 | if (offset >= 512 - sizeof(uint64_t)) | |
1422 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1423 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1424 | |
1425 | end = p + 512 - sizeof(io_u->start_time); | |
1426 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1427 | p += 512; | |
fad82f76 | 1428 | boffset += 512; |
23f394d5 | 1429 | } |
de789769 JA |
1430 | } |
1431 | ||
0d29de83 JA |
1432 | /* |
1433 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1434 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1435 | */ | |
1436 | struct io_u *get_io_u(struct thread_data *td) | |
1437 | { | |
1438 | struct fio_file *f; | |
1439 | struct io_u *io_u; | |
de789769 | 1440 | int do_scramble = 0; |
002fe734 | 1441 | long ret = 0; |
0d29de83 JA |
1442 | |
1443 | io_u = __get_io_u(td); | |
1444 | if (!io_u) { | |
1445 | dprint(FD_IO, "__get_io_u failed\n"); | |
1446 | return NULL; | |
1447 | } | |
1448 | ||
1449 | if (check_get_verify(td, io_u)) | |
1450 | goto out; | |
1451 | if (check_get_trim(td, io_u)) | |
1452 | goto out; | |
1453 | ||
755200a3 JA |
1454 | /* |
1455 | * from a requeue, io_u already setup | |
1456 | */ | |
1457 | if (io_u->file) | |
77f392bf | 1458 | goto out; |
755200a3 | 1459 | |
429f6675 JA |
1460 | /* |
1461 | * If using an iolog, grab next piece if any available. | |
1462 | */ | |
d72be545 | 1463 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1464 | if (read_iolog_get(td, io_u)) |
1465 | goto err_put; | |
2ba1c290 | 1466 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1467 | ret = -EBUSY; |
2ba1c290 | 1468 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1469 | goto err_put; |
2ba1c290 | 1470 | } |
5ec10eaa | 1471 | |
429f6675 | 1472 | f = io_u->file; |
002fe734 JA |
1473 | if (!f) { |
1474 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1475 | goto err_put; | |
1476 | } | |
1477 | ||
d6aed795 | 1478 | assert(fio_file_open(f)); |
97af62ce | 1479 | |
ff58fced | 1480 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1481 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1482 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1483 | goto err_put; |
2ba1c290 | 1484 | } |
10ba535a | 1485 | |
38dad62d | 1486 | f->last_start = io_u->offset; |
36167d82 | 1487 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1488 | |
fd68418e | 1489 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1490 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e JA |
1491 | io_u_fill_buffer(td, io_u, |
1492 | io_u->xfer_buflen, io_u->xfer_buflen); | |
bedc9dc2 JA |
1493 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
1494 | !(td->flags & TD_F_COMPRESS)) | |
fd68418e | 1495 | do_scramble = 1; |
d72be545 | 1496 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1497 | populate_verify_io_u(td, io_u); |
1498 | do_scramble = 0; | |
1499 | } | |
fd68418e | 1500 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1501 | /* |
1502 | * Reset the buf_filled parameters so next time if the | |
1503 | * buffer is used for writes it is refilled. | |
1504 | */ | |
cbe8d756 RR |
1505 | io_u->buf_filled_len = 0; |
1506 | } | |
87dc1ab1 | 1507 | } |
10ba535a | 1508 | |
165faf16 JA |
1509 | /* |
1510 | * Set io data pointers. | |
1511 | */ | |
cec6b55d JA |
1512 | io_u->xfer_buf = io_u->buf; |
1513 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1514 | |
6ac7a331 | 1515 | out: |
0d29de83 | 1516 | assert(io_u->file); |
429f6675 | 1517 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1518 | if (!td->o.disable_slat) |
1519 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1520 | if (do_scramble) |
1521 | small_content_scramble(io_u); | |
429f6675 | 1522 | return io_u; |
36167d82 | 1523 | } |
429f6675 | 1524 | err_put: |
2ba1c290 | 1525 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1526 | put_io_u(td, io_u); |
002fe734 | 1527 | return ERR_PTR(ret); |
10ba535a JA |
1528 | } |
1529 | ||
5451792e JA |
1530 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1531 | { | |
8b28bd41 | 1532 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e JA |
1533 | const char *msg[] = { "read", "write", "sync", "datasync", |
1534 | "sync_file_range", "wait", "trim" }; | |
1535 | ||
8b28bd41 DM |
1536 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1537 | return; | |
5451792e JA |
1538 | |
1539 | log_err("fio: io_u error"); | |
1540 | ||
1541 | if (io_u->file) | |
1542 | log_err(" on file %s", io_u->file->file_name); | |
1543 | ||
1544 | log_err(": %s\n", strerror(io_u->error)); | |
1545 | ||
5ec10eaa JA |
1546 | log_err(" %s offset=%llu, buflen=%lu\n", msg[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); | |
1572 | add_lat_sample(td, idx, tusec, bytes); | |
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) { | |
1590 | add_clat_sample(td, idx, lusec, bytes); | |
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 | ||
97601024 JA |
1612 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
1613 | struct io_completion_data *icd) | |
10ba535a | 1614 | { |
44f29692 | 1615 | struct fio_file *f; |
10ba535a | 1616 | |
2ba1c290 JA |
1617 | dprint_io_u(io_u, "io complete"); |
1618 | ||
2ecc1b57 | 1619 | td_io_u_lock(td); |
0c6e7517 | 1620 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1621 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1622 | |
1623 | /* | |
1624 | * Mark IO ok to verify | |
1625 | */ | |
1626 | if (io_u->ipo) { | |
1627 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1628 | write_barrier(); | |
1629 | } | |
1630 | ||
2ecc1b57 | 1631 | td_io_u_unlock(td); |
0c6e7517 | 1632 | |
5f9099ea | 1633 | if (ddir_sync(io_u->ddir)) { |
87dc1ab1 | 1634 | td->last_was_sync = 1; |
44f29692 JA |
1635 | f = io_u->file; |
1636 | if (f) { | |
1637 | f->first_write = -1ULL; | |
1638 | f->last_write = -1ULL; | |
1639 | } | |
87dc1ab1 JA |
1640 | return; |
1641 | } | |
1642 | ||
1643 | td->last_was_sync = 0; | |
9e144189 | 1644 | td->last_ddir = io_u->ddir; |
87dc1ab1 | 1645 | |
ff58fced | 1646 | if (!io_u->error && ddir_rw(io_u->ddir)) { |
10ba535a | 1647 | unsigned int bytes = io_u->buflen - io_u->resid; |
1e97cce9 | 1648 | const enum fio_ddir idx = io_u->ddir; |
ba3e4e0c | 1649 | const enum fio_ddir odx = io_u->ddir ^ 1; |
b29ee5b3 | 1650 | int ret; |
10ba535a | 1651 | |
b29ee5b3 | 1652 | td->io_blocks[idx]++; |
c8eeb9df | 1653 | td->this_io_blocks[idx]++; |
b29ee5b3 | 1654 | td->io_bytes[idx] += bytes; |
ae2fafc8 | 1655 | |
1656 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
1657 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 1658 | |
44f29692 JA |
1659 | if (idx == DDIR_WRITE) { |
1660 | f = io_u->file; | |
1661 | if (f) { | |
1662 | if (f->first_write == -1ULL || | |
1663 | io_u->offset < f->first_write) | |
1664 | f->first_write = io_u->offset; | |
1665 | if (f->last_write == -1ULL || | |
1666 | ((io_u->offset + bytes) > f->last_write)) | |
1667 | f->last_write = io_u->offset + bytes; | |
1668 | } | |
1669 | } | |
1670 | ||
6b1190fd SL |
1671 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1672 | td->runstate == TD_VERIFYING)) { | |
c8eeb9df | 1673 | account_io_completion(td, io_u, icd, idx, bytes); |
40e1a6f0 | 1674 | |
b23b6a2f | 1675 | if (__should_check_rate(td, idx)) { |
ba3e4e0c | 1676 | td->rate_pending_usleep[idx] = |
1b8dbf25 | 1677 | (usec_for_io(td, idx) - |
ba3e4e0c | 1678 | utime_since_now(&td->start)); |
b23b6a2f | 1679 | } |
6eaf09d6 | 1680 | if (idx != DDIR_TRIM && __should_check_rate(td, odx)) |
ba3e4e0c | 1681 | td->rate_pending_usleep[odx] = |
1b8dbf25 | 1682 | (usec_for_io(td, odx) - |
ba3e4e0c | 1683 | utime_since_now(&td->start)); |
721938ae | 1684 | } |
10ba535a | 1685 | |
b29ee5b3 | 1686 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 1687 | |
d7762cf8 | 1688 | if (io_u->end_io) { |
36690c9b | 1689 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
1690 | if (ret && !icd->error) |
1691 | icd->error = ret; | |
1692 | } | |
ff58fced | 1693 | } else if (io_u->error) { |
10ba535a | 1694 | icd->error = io_u->error; |
5451792e JA |
1695 | io_u_log_error(td, io_u); |
1696 | } | |
8b28bd41 DM |
1697 | if (icd->error) { |
1698 | enum error_type_bit eb = td_error_type(io_u->ddir, icd->error); | |
1699 | if (!td_non_fatal_error(td, eb, icd->error)) | |
1700 | return; | |
f2bba182 RR |
1701 | /* |
1702 | * If there is a non_fatal error, then add to the error count | |
1703 | * and clear all the errors. | |
1704 | */ | |
1705 | update_error_count(td, icd->error); | |
1706 | td_clear_error(td); | |
1707 | icd->error = 0; | |
1708 | io_u->error = 0; | |
1709 | } | |
10ba535a JA |
1710 | } |
1711 | ||
9520ebb9 JA |
1712 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1713 | int nr) | |
10ba535a | 1714 | { |
6eaf09d6 | 1715 | int ddir; |
aba6c951 JA |
1716 | |
1717 | if (!gtod_reduce(td)) | |
9520ebb9 | 1718 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1719 | |
3af6ef39 JA |
1720 | icd->nr = nr; |
1721 | ||
10ba535a | 1722 | icd->error = 0; |
6eaf09d6 SL |
1723 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1724 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1725 | } |
1726 | ||
97601024 JA |
1727 | static void ios_completed(struct thread_data *td, |
1728 | struct io_completion_data *icd) | |
36167d82 JA |
1729 | { |
1730 | struct io_u *io_u; | |
1731 | int i; | |
1732 | ||
10ba535a JA |
1733 | for (i = 0; i < icd->nr; i++) { |
1734 | io_u = td->io_ops->event(td, i); | |
1735 | ||
1736 | io_completed(td, io_u, icd); | |
e8462bd8 JA |
1737 | |
1738 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1739 | put_io_u(td, io_u); | |
10ba535a JA |
1740 | } |
1741 | } | |
97601024 | 1742 | |
e7e6cfb4 JA |
1743 | /* |
1744 | * Complete a single io_u for the sync engines. | |
1745 | */ | |
581e7141 | 1746 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1747 | uint64_t *bytes) |
97601024 JA |
1748 | { |
1749 | struct io_completion_data icd; | |
1750 | ||
9520ebb9 | 1751 | init_icd(td, &icd, 1); |
97601024 | 1752 | io_completed(td, io_u, &icd); |
e8462bd8 JA |
1753 | |
1754 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1755 | put_io_u(td, io_u); | |
97601024 | 1756 | |
581e7141 JA |
1757 | if (icd.error) { |
1758 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1759 | return -1; | |
1760 | } | |
97601024 | 1761 | |
581e7141 | 1762 | if (bytes) { |
6eaf09d6 SL |
1763 | int ddir; |
1764 | ||
1765 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1766 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1767 | } |
1768 | ||
1769 | return 0; | |
97601024 JA |
1770 | } |
1771 | ||
e7e6cfb4 JA |
1772 | /* |
1773 | * Called to complete min_events number of io for the async engines. | |
1774 | */ | |
581e7141 | 1775 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1776 | uint64_t *bytes) |
97601024 | 1777 | { |
97601024 | 1778 | struct io_completion_data icd; |
00de55ef | 1779 | struct timespec *tvp = NULL; |
97601024 | 1780 | int ret; |
4d06a338 | 1781 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1782 | |
4950421a | 1783 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1784 | |
4950421a | 1785 | if (!min_evts) |
00de55ef | 1786 | tvp = &ts; |
97601024 | 1787 | |
4950421a | 1788 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1789 | if (ret < 0) { |
e1161c32 | 1790 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1791 | return ret; |
1792 | } else if (!ret) | |
1793 | return ret; | |
1794 | ||
9520ebb9 | 1795 | init_icd(td, &icd, ret); |
97601024 | 1796 | ios_completed(td, &icd); |
581e7141 JA |
1797 | if (icd.error) { |
1798 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1799 | return -1; | |
1800 | } | |
97601024 | 1801 | |
581e7141 | 1802 | if (bytes) { |
6eaf09d6 SL |
1803 | int ddir; |
1804 | ||
1805 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1806 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1807 | } |
1808 | ||
1809 | return 0; | |
97601024 | 1810 | } |
7e77dd02 JA |
1811 | |
1812 | /* | |
1813 | * Call when io_u is really queued, to update the submission latency. | |
1814 | */ | |
1815 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1816 | { | |
9520ebb9 JA |
1817 | if (!td->o.disable_slat) { |
1818 | unsigned long slat_time; | |
7e77dd02 | 1819 | |
9520ebb9 | 1820 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
29a90ddb | 1821 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen); |
9520ebb9 | 1822 | } |
7e77dd02 | 1823 | } |
433afcb4 | 1824 | |
cc86c395 JA |
1825 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1826 | unsigned int max_bs) | |
5973cafb | 1827 | { |
ce35b1ec JA |
1828 | if (td->o.buffer_pattern_bytes) |
1829 | fill_buffer_pattern(td, buf, max_bs); | |
1830 | else if (!td->o.zero_buffers) { | |
9c42684e JA |
1831 | unsigned int perc = td->o.compress_percentage; |
1832 | ||
1833 | if (perc) { | |
f97a43a1 JA |
1834 | unsigned int seg = min_write; |
1835 | ||
1836 | seg = min(min_write, td->o.compress_chunk); | |
cc86c395 JA |
1837 | if (!seg) |
1838 | seg = min_write; | |
1839 | ||
1840 | fill_random_buf_percentage(&td->buf_state, buf, | |
f97a43a1 | 1841 | perc, seg, max_bs); |
9c42684e | 1842 | } else |
cc86c395 | 1843 | fill_random_buf(&td->buf_state, buf, max_bs); |
9c42684e | 1844 | } else |
cc86c395 JA |
1845 | memset(buf, 0, max_bs); |
1846 | } | |
1847 | ||
1848 | /* | |
1849 | * "randomly" fill the buffer contents | |
1850 | */ | |
1851 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1852 | unsigned int min_write, unsigned int max_bs) | |
1853 | { | |
1854 | io_u->buf_filled_len = 0; | |
1855 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1856 | } |