Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
5973cafb | 9 | #include "hash.h" |
4f5af7b2 | 10 | #include "verify.h" |
0d29de83 | 11 | #include "trim.h" |
1fbbf72e | 12 | #include "lib/rand.h" |
7ebd796f | 13 | #include "lib/axmap.h" |
002fe734 | 14 | #include "err.h" |
10ba535a | 15 | |
97601024 JA |
16 | struct io_completion_data { |
17 | int nr; /* input */ | |
97601024 JA |
18 | |
19 | int error; /* output */ | |
100f49f1 | 20 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
97601024 JA |
21 | struct timeval time; /* output */ |
22 | }; | |
23 | ||
10ba535a | 24 | /* |
7ebd796f | 25 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
26 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
27 | */ | |
1ae83d45 | 28 | static int random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 29 | { |
7ebd796f | 30 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
31 | } |
32 | ||
df415585 JA |
33 | /* |
34 | * Mark a given offset as used in the map. | |
35 | */ | |
9bf2061e | 36 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 37 | { |
2dc1bbeb | 38 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 39 | struct fio_file *f = io_u->file; |
51ede0b1 | 40 | unsigned int nr_blocks; |
1ae83d45 | 41 | uint64_t block; |
df415585 | 42 | |
1ae83d45 | 43 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 44 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 45 | |
2ab9e98b | 46 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 47 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 48 | |
51ede0b1 JA |
49 | if ((nr_blocks * min_bs) < io_u->buflen) |
50 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
51 | } |
52 | ||
74776733 JA |
53 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
54 | enum fio_ddir ddir) | |
2ba1c290 | 55 | { |
74776733 JA |
56 | uint64_t max_blocks; |
57 | uint64_t max_size; | |
2ba1c290 | 58 | |
ff58fced JA |
59 | assert(ddir_rw(ddir)); |
60 | ||
d9dd70f7 JA |
61 | /* |
62 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
63 | */ | |
64 | max_size = f->io_size; | |
65 | if (max_size > f->real_file_size) | |
66 | max_size = f->real_file_size; | |
67 | ||
ed335855 SN |
68 | if (td->o.zone_range) |
69 | max_size = td->o.zone_range; | |
70 | ||
0412d12e JE |
71 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
72 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
73 | ||
1ae83d45 | 74 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
75 | if (!max_blocks) |
76 | return 0; | |
77 | ||
67778e88 | 78 | return max_blocks; |
2ba1c290 JA |
79 | } |
80 | ||
1ae83d45 JA |
81 | struct rand_off { |
82 | struct flist_head list; | |
83 | uint64_t off; | |
84 | }; | |
85 | ||
e25839d4 | 86 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 87 | enum fio_ddir ddir, uint64_t *b) |
ec4015da | 88 | { |
6f49f8bc | 89 | uint64_t r; |
5e0baa7f | 90 | |
74776733 | 91 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) { |
6f49f8bc JA |
92 | uint64_t lastb; |
93 | ||
94 | lastb = last_block(td, f, ddir); | |
95 | if (!lastb) | |
96 | return 1; | |
97 | ||
d6b72507 | 98 | r = __rand(&td->random_state); |
8055e41d | 99 | |
4b91ee8f | 100 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 101 | |
54a21917 | 102 | *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0)); |
51ede0b1 | 103 | } else { |
8055e41d | 104 | uint64_t off = 0; |
43c63a78 | 105 | |
967d1b63 JA |
106 | assert(fio_file_lfsr(f)); |
107 | ||
6f49f8bc | 108 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 109 | return 1; |
ec4015da | 110 | |
8055e41d JA |
111 | *b = off; |
112 | } | |
0ce8b119 | 113 | |
ec4015da | 114 | /* |
51ede0b1 | 115 | * if we are not maintaining a random map, we are done. |
ec4015da | 116 | */ |
51ede0b1 JA |
117 | if (!file_randommap(td, f)) |
118 | goto ret; | |
43c63a78 JA |
119 | |
120 | /* | |
51ede0b1 | 121 | * calculate map offset and check if it's free |
43c63a78 | 122 | */ |
51ede0b1 JA |
123 | if (random_map_free(f, *b)) |
124 | goto ret; | |
125 | ||
4b91ee8f JA |
126 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
127 | (unsigned long long) *b); | |
51ede0b1 | 128 | |
7ebd796f | 129 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
130 | if (*b == (uint64_t) -1ULL) |
131 | return 1; | |
0ce8b119 JA |
132 | ret: |
133 | return 0; | |
ec4015da JA |
134 | } |
135 | ||
925fee33 JA |
136 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
137 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 138 | uint64_t *b) |
e25839d4 | 139 | { |
9c6f6316 | 140 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
141 | return 0; |
142 | } | |
143 | ||
925fee33 JA |
144 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
145 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 146 | uint64_t *b) |
925fee33 | 147 | { |
9c6f6316 | 148 | *b = pareto_next(&f->zipf); |
925fee33 JA |
149 | return 0; |
150 | } | |
151 | ||
1ae83d45 JA |
152 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
153 | { | |
154 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
155 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
156 | ||
157 | return r1->off - r2->off; | |
158 | } | |
159 | ||
160 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
161 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
162 | { |
163 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
164 | return __get_next_rand_offset(td, f, ddir, b); | |
165 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
166 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
167 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
168 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
e25839d4 JA |
169 | |
170 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
171 | return 1; | |
172 | } | |
173 | ||
bcd5abfa JA |
174 | /* |
175 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
176 | * specified. | |
177 | */ | |
178 | static inline int should_sort_io(struct thread_data *td) | |
179 | { | |
180 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
181 | return 0; | |
182 | if (!td_random(td)) | |
183 | return 0; | |
184 | if (td->runstate != TD_VERIFYING) | |
185 | return 0; | |
186 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) | |
187 | return 0; | |
188 | ||
189 | return 1; | |
190 | } | |
191 | ||
d9472271 | 192 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
193 | { |
194 | unsigned int v; | |
195 | unsigned long r; | |
196 | ||
d9472271 | 197 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
198 | return 1; |
199 | ||
d6b72507 | 200 | r = __rand(&td->seq_rand_state[ddir]); |
54a21917 | 201 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
211c9b89 | 202 | |
d9472271 | 203 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
204 | } |
205 | ||
1ae83d45 JA |
206 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
207 | enum fio_ddir ddir, uint64_t *b) | |
208 | { | |
209 | struct rand_off *r; | |
210 | int i, ret = 1; | |
211 | ||
bcd5abfa | 212 | if (!should_sort_io(td)) |
1ae83d45 JA |
213 | return get_off_from_method(td, f, ddir, b); |
214 | ||
215 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 216 | fetch: |
9342d5f8 | 217 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
218 | flist_del(&r->list); |
219 | *b = r->off; | |
220 | free(r); | |
221 | return 0; | |
222 | } | |
223 | ||
224 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
225 | r = malloc(sizeof(*r)); | |
226 | ||
227 | ret = get_off_from_method(td, f, ddir, &r->off); | |
228 | if (ret) { | |
229 | free(r); | |
230 | break; | |
231 | } | |
232 | ||
233 | flist_add(&r->list, &td->next_rand_list); | |
234 | } | |
235 | ||
236 | if (ret && !i) | |
237 | return ret; | |
238 | ||
239 | assert(!flist_empty(&td->next_rand_list)); | |
240 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
241 | goto fetch; | |
242 | } | |
243 | ||
38dad62d | 244 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 245 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 246 | { |
c04e4661 DE |
247 | if (!get_next_rand_offset(td, f, ddir, b)) |
248 | return 0; | |
249 | ||
250 | if (td->o.time_based) { | |
33c48814 | 251 | fio_file_reset(td, f); |
c04e4661 DE |
252 | if (!get_next_rand_offset(td, f, ddir, b)) |
253 | return 0; | |
38dad62d JA |
254 | } |
255 | ||
c04e4661 | 256 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 257 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 258 | (unsigned long long) f->real_file_size); |
c04e4661 | 259 | return 1; |
38dad62d JA |
260 | } |
261 | ||
37cf9e3c | 262 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 263 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 264 | { |
ac002339 JA |
265 | struct thread_options *o = &td->o; |
266 | ||
ff58fced JA |
267 | assert(ddir_rw(ddir)); |
268 | ||
f1dfb668 | 269 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
ac002339 | 270 | o->time_based) |
f1dfb668 | 271 | f->last_pos[ddir] = f->last_pos[ddir] - f->io_size; |
c04e4661 | 272 | |
f1dfb668 | 273 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 274 | uint64_t pos; |
059b0802 | 275 | |
f1dfb668 JA |
276 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) |
277 | f->last_pos[ddir] = f->real_file_size; | |
a66da7a2 | 278 | |
f1dfb668 | 279 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
280 | if (pos && o->ddir_seq_add) { |
281 | pos += o->ddir_seq_add; | |
282 | ||
283 | /* | |
284 | * If we reach beyond the end of the file | |
285 | * with holed IO, wrap around to the | |
286 | * beginning again. | |
287 | */ | |
288 | if (pos >= f->real_file_size) | |
289 | pos = f->file_offset; | |
290 | } | |
059b0802 | 291 | |
37cf9e3c | 292 | *offset = pos; |
38dad62d JA |
293 | return 0; |
294 | } | |
295 | ||
296 | return 1; | |
297 | } | |
298 | ||
299 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
300 | enum fio_ddir ddir, int rw_seq, |
301 | unsigned int *is_random) | |
38dad62d JA |
302 | { |
303 | struct fio_file *f = io_u->file; | |
1ae83d45 | 304 | uint64_t b, offset; |
38dad62d JA |
305 | int ret; |
306 | ||
ff58fced JA |
307 | assert(ddir_rw(ddir)); |
308 | ||
37cf9e3c JA |
309 | b = offset = -1ULL; |
310 | ||
38dad62d | 311 | if (rw_seq) { |
211c9b89 | 312 | if (td_random(td)) { |
6aca9b3d | 313 | if (should_do_random(td, ddir)) { |
211c9b89 | 314 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
315 | *is_random = 1; |
316 | } else { | |
317 | *is_random = 0; | |
211c9b89 JA |
318 | io_u->flags |= IO_U_F_BUSY_OK; |
319 | ret = get_next_seq_offset(td, f, ddir, &offset); | |
320 | if (ret) | |
321 | ret = get_next_rand_block(td, f, ddir, &b); | |
322 | } | |
6aca9b3d JA |
323 | } else { |
324 | *is_random = 0; | |
37cf9e3c | 325 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 326 | } |
38dad62d JA |
327 | } else { |
328 | io_u->flags |= IO_U_F_BUSY_OK; | |
6aca9b3d | 329 | *is_random = 0; |
38dad62d JA |
330 | |
331 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 332 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 333 | if (ret) { |
37cf9e3c | 334 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
335 | *is_random = 0; |
336 | } | |
38dad62d | 337 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
338 | if (f->last_start[ddir] != -1ULL) |
339 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 340 | else |
37cf9e3c | 341 | offset = 0; |
38dad62d JA |
342 | ret = 0; |
343 | } else { | |
344 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
345 | ret = 1; | |
346 | } | |
347 | } | |
6d68b997 | 348 | |
37cf9e3c JA |
349 | if (!ret) { |
350 | if (offset != -1ULL) | |
351 | io_u->offset = offset; | |
352 | else if (b != -1ULL) | |
353 | io_u->offset = b * td->o.ba[ddir]; | |
354 | else { | |
4e0a8fa2 | 355 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
356 | ret = 1; |
357 | } | |
358 | } | |
359 | ||
38dad62d JA |
360 | return ret; |
361 | } | |
362 | ||
10ba535a JA |
363 | /* |
364 | * For random io, generate a random new block and see if it's used. Repeat | |
365 | * until we find a free one. For sequential io, just return the end of | |
366 | * the last io issued. | |
367 | */ | |
6aca9b3d JA |
368 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
369 | unsigned int *is_random) | |
10ba535a | 370 | { |
9bf2061e | 371 | struct fio_file *f = io_u->file; |
4ba66134 | 372 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 373 | int rw_seq_hit = 0; |
10ba535a | 374 | |
ff58fced JA |
375 | assert(ddir_rw(ddir)); |
376 | ||
38dad62d JA |
377 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
378 | rw_seq_hit = 1; | |
5736c10d | 379 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 380 | } |
211097b2 | 381 | |
6aca9b3d | 382 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 383 | return 1; |
10ba535a | 384 | |
009bd847 JA |
385 | if (io_u->offset >= f->io_size) { |
386 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
387 | (unsigned long long) io_u->offset, |
388 | (unsigned long long) f->io_size); | |
009bd847 JA |
389 | return 1; |
390 | } | |
391 | ||
392 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
393 | if (io_u->offset >= f->real_file_size) { |
394 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
395 | (unsigned long long) io_u->offset, |
396 | (unsigned long long) f->real_file_size); | |
10ba535a | 397 | return 1; |
2ba1c290 | 398 | } |
10ba535a JA |
399 | |
400 | return 0; | |
401 | } | |
402 | ||
6aca9b3d JA |
403 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
404 | unsigned int *is_random) | |
15dc1934 | 405 | { |
d72be545 JA |
406 | if (td->flags & TD_F_PROFILE_OPS) { |
407 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 408 | |
d72be545 | 409 | if (ops->fill_io_u_off) |
6aca9b3d | 410 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 411 | } |
15dc1934 | 412 | |
6aca9b3d | 413 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
414 | } |
415 | ||
79944128 JA |
416 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
417 | unsigned int buflen) | |
418 | { | |
419 | struct fio_file *f = io_u->file; | |
420 | ||
bedc9dc2 | 421 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
422 | } |
423 | ||
6aca9b3d JA |
424 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
425 | unsigned int is_random) | |
10ba535a | 426 | { |
6aca9b3d | 427 | int ddir = io_u->ddir; |
24d23ca7 | 428 | unsigned int buflen = 0; |
f3059de1 | 429 | unsigned int minbs, maxbs; |
54a21917 | 430 | unsigned long r; |
10ba535a | 431 | |
9ee1c647 | 432 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
433 | |
434 | if (td->o.bs_is_seq_rand) | |
435 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 436 | |
f3059de1 JA |
437 | minbs = td->o.min_bs[ddir]; |
438 | maxbs = td->o.max_bs[ddir]; | |
439 | ||
79944128 JA |
440 | if (minbs == maxbs) |
441 | return minbs; | |
442 | ||
52c58027 JA |
443 | /* |
444 | * If we can't satisfy the min block size from here, then fail | |
445 | */ | |
446 | if (!io_u_fits(td, io_u, minbs)) | |
447 | return 0; | |
448 | ||
79944128 | 449 | do { |
d6b72507 | 450 | r = __rand(&td->bsrange_state); |
4c07ad86 | 451 | |
720e84ad | 452 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 453 | buflen = 1 + (unsigned int) ((double) maxbs * |
54a21917 | 454 | (r / (FRAND_MAX + 1.0))); |
f3059de1 JA |
455 | if (buflen < minbs) |
456 | buflen = minbs; | |
5ec10eaa | 457 | } else { |
564ca972 JA |
458 | long perc = 0; |
459 | unsigned int i; | |
460 | ||
720e84ad JA |
461 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
462 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
463 | |
464 | buflen = bsp->bs; | |
465 | perc += bsp->perc; | |
54a21917 | 466 | if ((r <= ((FRAND_MAX / 100L) * perc)) && |
79944128 | 467 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
468 | break; |
469 | } | |
470 | } | |
79944128 | 471 | |
a9f70b1f JB |
472 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
473 | buflen = (buflen + td->o.verify_interval - 1) & | |
474 | ~(td->o.verify_interval - 1); | |
475 | ||
f3059de1 JA |
476 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
477 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 478 | |
79944128 | 479 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 480 | |
10ba535a JA |
481 | return buflen; |
482 | } | |
483 | ||
6aca9b3d JA |
484 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
485 | unsigned int is_random) | |
15dc1934 | 486 | { |
d72be545 JA |
487 | if (td->flags & TD_F_PROFILE_OPS) { |
488 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 489 | |
d72be545 | 490 | if (ops->fill_io_u_size) |
6aca9b3d | 491 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 492 | } |
15dc1934 | 493 | |
6aca9b3d | 494 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
495 | } |
496 | ||
afe24a5a JA |
497 | static void set_rwmix_bytes(struct thread_data *td) |
498 | { | |
afe24a5a JA |
499 | unsigned int diff; |
500 | ||
501 | /* | |
502 | * we do time or byte based switch. this is needed because | |
503 | * buffered writes may issue a lot quicker than they complete, | |
504 | * whereas reads do not. | |
505 | */ | |
e47f799f | 506 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 507 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
508 | } |
509 | ||
510 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
511 | { | |
512 | unsigned int v; | |
1294c3ec | 513 | unsigned long r; |
e47f799f | 514 | |
d6b72507 | 515 | r = __rand(&td->rwmix_state); |
54a21917 | 516 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); |
4c07ad86 | 517 | |
04c540d9 | 518 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
519 | return DDIR_READ; |
520 | ||
521 | return DDIR_WRITE; | |
afe24a5a JA |
522 | } |
523 | ||
002e7183 JA |
524 | void io_u_quiesce(struct thread_data *td) |
525 | { | |
526 | /* | |
527 | * We are going to sleep, ensure that we flush anything pending as | |
528 | * not to skew our latency numbers. | |
529 | * | |
530 | * Changed to only monitor 'in flight' requests here instead of the | |
531 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
532 | * io's that have been actually submitted to an async engine, | |
533 | * and cur_depth is meaningless for sync engines. | |
534 | */ | |
9cc80b6d JA |
535 | if (td->io_u_queued || td->cur_depth) { |
536 | int fio_unused ret; | |
537 | ||
538 | ret = td_io_commit(td); | |
539 | } | |
540 | ||
002e7183 JA |
541 | while (td->io_u_in_flight) { |
542 | int fio_unused ret; | |
543 | ||
544 | ret = io_u_queued_complete(td, 1, NULL); | |
545 | } | |
546 | } | |
547 | ||
581e7141 JA |
548 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
549 | { | |
550 | enum fio_ddir odir = ddir ^ 1; | |
581e7141 JA |
551 | long usec; |
552 | ||
ff58fced JA |
553 | assert(ddir_rw(ddir)); |
554 | ||
315fcfec | 555 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
556 | return ddir; |
557 | ||
558 | /* | |
559 | * We have too much pending sleep in this direction. See if we | |
560 | * should switch. | |
561 | */ | |
315fcfec | 562 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
563 | /* |
564 | * Other direction does not have too much pending, switch | |
565 | */ | |
566 | if (td->rate_pending_usleep[odir] < 100000) | |
567 | return odir; | |
568 | ||
569 | /* | |
570 | * Both directions have pending sleep. Sleep the minimum time | |
571 | * and deduct from both. | |
572 | */ | |
573 | if (td->rate_pending_usleep[ddir] <= | |
574 | td->rate_pending_usleep[odir]) { | |
575 | usec = td->rate_pending_usleep[ddir]; | |
576 | } else { | |
577 | usec = td->rate_pending_usleep[odir]; | |
578 | ddir = odir; | |
579 | } | |
580 | } else | |
581 | usec = td->rate_pending_usleep[ddir]; | |
582 | ||
002e7183 | 583 | io_u_quiesce(td); |
78c1eda5 | 584 | |
9cc80b6d | 585 | usec = usec_sleep(td, usec); |
581e7141 JA |
586 | |
587 | td->rate_pending_usleep[ddir] -= usec; | |
588 | ||
589 | odir = ddir ^ 1; | |
590 | if (td_rw(td) && __should_check_rate(td, odir)) | |
591 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 592 | |
77b451cc JA |
593 | if (ddir == DDIR_TRIM) |
594 | return DDIR_TRIM; | |
e0224c6b | 595 | |
581e7141 JA |
596 | return ddir; |
597 | } | |
598 | ||
10ba535a JA |
599 | /* |
600 | * Return the data direction for the next io_u. If the job is a | |
601 | * mixed read/write workload, check the rwmix cycle and switch if | |
602 | * necessary. | |
603 | */ | |
1e97cce9 | 604 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 605 | { |
581e7141 JA |
606 | enum fio_ddir ddir; |
607 | ||
5f9099ea JA |
608 | /* |
609 | * see if it's time to fsync | |
610 | */ | |
611 | if (td->o.fsync_blocks && | |
612 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
613 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
614 | return DDIR_SYNC; | |
615 | ||
616 | /* | |
617 | * see if it's time to fdatasync | |
618 | */ | |
619 | if (td->o.fdatasync_blocks && | |
620 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
621 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
622 | return DDIR_DATASYNC; | |
623 | ||
44f29692 JA |
624 | /* |
625 | * see if it's time to sync_file_range | |
626 | */ | |
627 | if (td->sync_file_range_nr && | |
628 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
629 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
630 | return DDIR_SYNC_FILE_RANGE; | |
631 | ||
10ba535a | 632 | if (td_rw(td)) { |
10ba535a JA |
633 | /* |
634 | * Check if it's time to seed a new data direction. | |
635 | */ | |
e4928662 | 636 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
637 | /* |
638 | * Put a top limit on how many bytes we do for | |
639 | * one data direction, to avoid overflowing the | |
640 | * ranges too much | |
641 | */ | |
642 | ddir = get_rand_ddir(td); | |
e47f799f JA |
643 | |
644 | if (ddir != td->rwmix_ddir) | |
645 | set_rwmix_bytes(td); | |
646 | ||
647 | td->rwmix_ddir = ddir; | |
10ba535a | 648 | } |
581e7141 | 649 | ddir = td->rwmix_ddir; |
10ba535a | 650 | } else if (td_read(td)) |
581e7141 | 651 | ddir = DDIR_READ; |
6eaf09d6 | 652 | else if (td_write(td)) |
581e7141 | 653 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
654 | else |
655 | ddir = DDIR_TRIM; | |
581e7141 JA |
656 | |
657 | td->rwmix_ddir = rate_ddir(td, ddir); | |
658 | return td->rwmix_ddir; | |
10ba535a JA |
659 | } |
660 | ||
1ef2b6be JA |
661 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
662 | { | |
bcd5abfa | 663 | io_u->ddir = io_u->acct_ddir = get_rw_ddir(td); |
1ef2b6be JA |
664 | |
665 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
666 | td->o.barrier_blocks && | |
667 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
668 | td->io_issues[DDIR_WRITE]) | |
669 | io_u->flags |= IO_U_F_BARRIER; | |
670 | } | |
671 | ||
e8462bd8 | 672 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 673 | { |
71b84caa | 674 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
675 | |
676 | if (ret) | |
677 | td_verror(td, ret, "file close"); | |
678 | } | |
679 | ||
10ba535a JA |
680 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
681 | { | |
e8462bd8 JA |
682 | td_io_u_lock(td); |
683 | ||
f8b0bd10 | 684 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 685 | put_file_log(td, io_u->file); |
f8b0bd10 | 686 | |
10ba535a | 687 | io_u->file = NULL; |
d7ee2a7d SL |
688 | io_u->flags |= IO_U_F_FREE; |
689 | ||
0c41214f RR |
690 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
691 | td->cur_depth--; | |
2ae0b204 | 692 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
693 | td_io_u_unlock(td); |
694 | td_io_u_free_notify(td); | |
10ba535a JA |
695 | } |
696 | ||
f2bba182 RR |
697 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
698 | { | |
699 | io_u->flags &= ~IO_U_F_FLIGHT; | |
700 | put_io_u(td, io_u); | |
701 | } | |
702 | ||
755200a3 JA |
703 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
704 | { | |
705 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 706 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 707 | |
465221b0 JA |
708 | dprint(FD_IO, "requeue %p\n", __io_u); |
709 | ||
e8462bd8 JA |
710 | td_io_u_lock(td); |
711 | ||
4d2e0f49 | 712 | __io_u->flags |= IO_U_F_FREE; |
bcd5abfa JA |
713 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
714 | td->io_issues[ddir]--; | |
5ec10eaa | 715 | |
4d2e0f49 | 716 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
717 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
718 | td->cur_depth--; | |
2ae0b204 JA |
719 | |
720 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 721 | td_io_u_unlock(td); |
755200a3 JA |
722 | *io_u = NULL; |
723 | } | |
724 | ||
9bf2061e | 725 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 726 | { |
6aca9b3d JA |
727 | unsigned int is_random; |
728 | ||
b4c5e1ac JA |
729 | if (td->io_ops->flags & FIO_NOIO) |
730 | goto out; | |
731 | ||
1ef2b6be | 732 | set_rw_ddir(td, io_u); |
5f9099ea | 733 | |
87dc1ab1 | 734 | /* |
ff58fced | 735 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 736 | */ |
ff58fced | 737 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 738 | goto out; |
a00735e6 | 739 | |
48f5abd3 JA |
740 | /* |
741 | * See if it's time to switch to a new zone | |
742 | */ | |
13af05ae | 743 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
744 | struct fio_file *f = io_u->file; |
745 | ||
48f5abd3 | 746 | td->zone_bytes = 0; |
4c8be5b1 JA |
747 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
748 | ||
749 | /* | |
750 | * Wrap from the beginning, if we exceed the file size | |
751 | */ | |
752 | if (f->file_offset >= f->real_file_size) | |
753 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 754 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
755 | td->io_skip_bytes += td->o.zone_skip; |
756 | } | |
757 | ||
10ba535a | 758 | /* |
c685b5b2 JA |
759 | * No log, let the seq/rand engine retrieve the next buflen and |
760 | * position. | |
10ba535a | 761 | */ |
6aca9b3d | 762 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 763 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 764 | return 1; |
2ba1c290 | 765 | } |
10ba535a | 766 | |
6aca9b3d | 767 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
768 | if (!io_u->buflen) { |
769 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 770 | return 1; |
2ba1c290 | 771 | } |
bca4ed4d | 772 | |
2ba1c290 JA |
773 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
774 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
775 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
776 | (unsigned long long) io_u->offset, io_u->buflen, | |
777 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 778 | return 1; |
2ba1c290 | 779 | } |
6a5e6884 | 780 | |
bca4ed4d JA |
781 | /* |
782 | * mark entry before potentially trimming io_u | |
783 | */ | |
303032ae | 784 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 785 | mark_random_map(td, io_u); |
bca4ed4d | 786 | |
c38e9468 | 787 | out: |
2ba1c290 | 788 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 789 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 790 | return 0; |
10ba535a JA |
791 | } |
792 | ||
838bc709 JA |
793 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
794 | { | |
2b13e716 | 795 | int idx = 0; |
838bc709 JA |
796 | |
797 | switch (nr) { | |
798 | default: | |
2b13e716 | 799 | idx = 6; |
838bc709 JA |
800 | break; |
801 | case 33 ... 64: | |
2b13e716 | 802 | idx = 5; |
838bc709 JA |
803 | break; |
804 | case 17 ... 32: | |
2b13e716 | 805 | idx = 4; |
838bc709 JA |
806 | break; |
807 | case 9 ... 16: | |
2b13e716 | 808 | idx = 3; |
838bc709 JA |
809 | break; |
810 | case 5 ... 8: | |
2b13e716 | 811 | idx = 2; |
838bc709 JA |
812 | break; |
813 | case 1 ... 4: | |
2b13e716 | 814 | idx = 1; |
838bc709 JA |
815 | case 0: |
816 | break; | |
817 | } | |
818 | ||
2b13e716 | 819 | map[idx]++; |
838bc709 JA |
820 | } |
821 | ||
822 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
823 | { | |
824 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
825 | td->ts.total_submit++; | |
826 | } | |
827 | ||
828 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
829 | { | |
830 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
831 | td->ts.total_complete++; | |
832 | } | |
833 | ||
d8005759 | 834 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 835 | { |
2b13e716 | 836 | int idx = 0; |
71619dc2 JA |
837 | |
838 | switch (td->cur_depth) { | |
839 | default: | |
2b13e716 | 840 | idx = 6; |
a783e61a | 841 | break; |
71619dc2 | 842 | case 32 ... 63: |
2b13e716 | 843 | idx = 5; |
a783e61a | 844 | break; |
71619dc2 | 845 | case 16 ... 31: |
2b13e716 | 846 | idx = 4; |
a783e61a | 847 | break; |
71619dc2 | 848 | case 8 ... 15: |
2b13e716 | 849 | idx = 3; |
a783e61a | 850 | break; |
71619dc2 | 851 | case 4 ... 7: |
2b13e716 | 852 | idx = 2; |
a783e61a | 853 | break; |
71619dc2 | 854 | case 2 ... 3: |
2b13e716 | 855 | idx = 1; |
71619dc2 JA |
856 | case 1: |
857 | break; | |
858 | } | |
859 | ||
2b13e716 | 860 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
861 | } |
862 | ||
04a0feae JA |
863 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
864 | { | |
2b13e716 | 865 | int idx = 0; |
04a0feae JA |
866 | |
867 | assert(usec < 1000); | |
868 | ||
869 | switch (usec) { | |
870 | case 750 ... 999: | |
2b13e716 | 871 | idx = 9; |
04a0feae JA |
872 | break; |
873 | case 500 ... 749: | |
2b13e716 | 874 | idx = 8; |
04a0feae JA |
875 | break; |
876 | case 250 ... 499: | |
2b13e716 | 877 | idx = 7; |
04a0feae JA |
878 | break; |
879 | case 100 ... 249: | |
2b13e716 | 880 | idx = 6; |
04a0feae JA |
881 | break; |
882 | case 50 ... 99: | |
2b13e716 | 883 | idx = 5; |
04a0feae JA |
884 | break; |
885 | case 20 ... 49: | |
2b13e716 | 886 | idx = 4; |
04a0feae JA |
887 | break; |
888 | case 10 ... 19: | |
2b13e716 | 889 | idx = 3; |
04a0feae JA |
890 | break; |
891 | case 4 ... 9: | |
2b13e716 | 892 | idx = 2; |
04a0feae JA |
893 | break; |
894 | case 2 ... 3: | |
2b13e716 | 895 | idx = 1; |
04a0feae JA |
896 | case 0 ... 1: |
897 | break; | |
898 | } | |
899 | ||
2b13e716 JA |
900 | assert(idx < FIO_IO_U_LAT_U_NR); |
901 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
902 | } |
903 | ||
904 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 905 | { |
2b13e716 | 906 | int idx = 0; |
ec118304 JA |
907 | |
908 | switch (msec) { | |
909 | default: | |
2b13e716 | 910 | idx = 11; |
04a0feae | 911 | break; |
8abdce66 | 912 | case 1000 ... 1999: |
2b13e716 | 913 | idx = 10; |
04a0feae | 914 | break; |
8abdce66 | 915 | case 750 ... 999: |
2b13e716 | 916 | idx = 9; |
04a0feae | 917 | break; |
8abdce66 | 918 | case 500 ... 749: |
2b13e716 | 919 | idx = 8; |
04a0feae | 920 | break; |
8abdce66 | 921 | case 250 ... 499: |
2b13e716 | 922 | idx = 7; |
04a0feae | 923 | break; |
8abdce66 | 924 | case 100 ... 249: |
2b13e716 | 925 | idx = 6; |
04a0feae | 926 | break; |
8abdce66 | 927 | case 50 ... 99: |
2b13e716 | 928 | idx = 5; |
04a0feae | 929 | break; |
8abdce66 | 930 | case 20 ... 49: |
2b13e716 | 931 | idx = 4; |
04a0feae | 932 | break; |
8abdce66 | 933 | case 10 ... 19: |
2b13e716 | 934 | idx = 3; |
04a0feae | 935 | break; |
8abdce66 | 936 | case 4 ... 9: |
2b13e716 | 937 | idx = 2; |
04a0feae | 938 | break; |
ec118304 | 939 | case 2 ... 3: |
2b13e716 | 940 | idx = 1; |
ec118304 JA |
941 | case 0 ... 1: |
942 | break; | |
943 | } | |
944 | ||
2b13e716 JA |
945 | assert(idx < FIO_IO_U_LAT_M_NR); |
946 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
947 | } |
948 | ||
949 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
950 | { | |
951 | if (usec < 1000) | |
952 | io_u_mark_lat_usec(td, usec); | |
953 | else | |
954 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
955 | } |
956 | ||
0aabe160 JA |
957 | /* |
958 | * Get next file to service by choosing one at random | |
959 | */ | |
2cc52930 JA |
960 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
961 | enum fio_file_flags goodf, | |
d6aed795 | 962 | enum fio_file_flags badf) |
0aabe160 | 963 | { |
0aabe160 | 964 | struct fio_file *f; |
1c178180 | 965 | int fno; |
0aabe160 JA |
966 | |
967 | do { | |
87b10676 | 968 | int opened = 0; |
1294c3ec | 969 | unsigned long r; |
4c07ad86 | 970 | |
d6b72507 | 971 | r = __rand(&td->next_file_state); |
54a21917 | 972 | fno = (unsigned int) ((double) td->o.nr_files |
4c07ad86 | 973 | * (r / (FRAND_MAX + 1.0))); |
7c83c089 | 974 | |
126d65c6 | 975 | f = td->files[fno]; |
d6aed795 | 976 | if (fio_file_done(f)) |
059e63c0 | 977 | continue; |
1c178180 | 978 | |
d6aed795 | 979 | if (!fio_file_open(f)) { |
87b10676 JA |
980 | int err; |
981 | ||
002fe734 JA |
982 | if (td->nr_open_files >= td->o.open_files) |
983 | return ERR_PTR(-EBUSY); | |
984 | ||
87b10676 JA |
985 | err = td_io_open_file(td, f); |
986 | if (err) | |
987 | continue; | |
988 | opened = 1; | |
989 | } | |
990 | ||
2ba1c290 JA |
991 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
992 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 993 | return f; |
2ba1c290 | 994 | } |
87b10676 JA |
995 | if (opened) |
996 | td_io_close_file(td, f); | |
0aabe160 JA |
997 | } while (1); |
998 | } | |
999 | ||
1000 | /* | |
1001 | * Get next file to service by doing round robin between all available ones | |
1002 | */ | |
1c178180 JA |
1003 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1004 | int badf) | |
3d7c391d JA |
1005 | { |
1006 | unsigned int old_next_file = td->next_file; | |
1007 | struct fio_file *f; | |
1008 | ||
1009 | do { | |
87b10676 JA |
1010 | int opened = 0; |
1011 | ||
126d65c6 | 1012 | f = td->files[td->next_file]; |
3d7c391d JA |
1013 | |
1014 | td->next_file++; | |
2dc1bbeb | 1015 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1016 | td->next_file = 0; |
1017 | ||
87b10676 | 1018 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1019 | if (fio_file_done(f)) { |
d5ed68ea | 1020 | f = NULL; |
059e63c0 | 1021 | continue; |
d5ed68ea | 1022 | } |
059e63c0 | 1023 | |
d6aed795 | 1024 | if (!fio_file_open(f)) { |
87b10676 JA |
1025 | int err; |
1026 | ||
002fe734 JA |
1027 | if (td->nr_open_files >= td->o.open_files) |
1028 | return ERR_PTR(-EBUSY); | |
1029 | ||
87b10676 | 1030 | err = td_io_open_file(td, f); |
b5696bfc JA |
1031 | if (err) { |
1032 | dprint(FD_FILE, "error %d on open of %s\n", | |
1033 | err, f->file_name); | |
87c27b45 | 1034 | f = NULL; |
87b10676 | 1035 | continue; |
b5696bfc | 1036 | } |
87b10676 JA |
1037 | opened = 1; |
1038 | } | |
1039 | ||
0b9d69ec JA |
1040 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1041 | f->flags); | |
1c178180 | 1042 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1043 | break; |
1044 | ||
87b10676 JA |
1045 | if (opened) |
1046 | td_io_close_file(td, f); | |
1047 | ||
3d7c391d JA |
1048 | f = NULL; |
1049 | } while (td->next_file != old_next_file); | |
1050 | ||
2ba1c290 | 1051 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1052 | return f; |
1053 | } | |
1054 | ||
7eb36574 | 1055 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1056 | { |
1907dbc6 JA |
1057 | struct fio_file *f; |
1058 | ||
2dc1bbeb | 1059 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1060 | |
b5696bfc | 1061 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1062 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1063 | " nr_files=%d\n", td->nr_open_files, | |
1064 | td->nr_done_files, | |
1065 | td->o.nr_files); | |
bdb4e2e9 | 1066 | return NULL; |
2ba1c290 | 1067 | } |
bdb4e2e9 | 1068 | |
1907dbc6 | 1069 | f = td->file_service_file; |
d6aed795 | 1070 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1071 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1072 | goto out; | |
1073 | if (td->file_service_left--) | |
1074 | goto out; | |
1075 | } | |
1907dbc6 | 1076 | |
a086c257 JA |
1077 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1078 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1079 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1080 | else |
d6aed795 | 1081 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1082 | |
002fe734 JA |
1083 | if (IS_ERR(f)) |
1084 | return f; | |
1085 | ||
1907dbc6 JA |
1086 | td->file_service_file = f; |
1087 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1088 | out: |
0dac421f JA |
1089 | if (f) |
1090 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1091 | else | |
1092 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1093 | return f; |
bdb4e2e9 JA |
1094 | } |
1095 | ||
7eb36574 JA |
1096 | static struct fio_file *get_next_file(struct thread_data *td) |
1097 | { | |
372d8962 | 1098 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1099 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1100 | |
d72be545 JA |
1101 | if (ops->get_next_file) |
1102 | return ops->get_next_file(td); | |
1103 | } | |
7eb36574 JA |
1104 | |
1105 | return __get_next_file(td); | |
1106 | } | |
1107 | ||
002fe734 | 1108 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1109 | { |
1110 | struct fio_file *f; | |
1111 | ||
1112 | do { | |
1113 | f = get_next_file(td); | |
002fe734 JA |
1114 | if (IS_ERR_OR_NULL(f)) |
1115 | return PTR_ERR(f); | |
429f6675 | 1116 | |
429f6675 JA |
1117 | io_u->file = f; |
1118 | get_file(f); | |
1119 | ||
1120 | if (!fill_io_u(td, io_u)) | |
1121 | break; | |
1122 | ||
b5696bfc | 1123 | put_file_log(td, f); |
429f6675 | 1124 | td_io_close_file(td, f); |
b5696bfc | 1125 | io_u->file = NULL; |
d6aed795 | 1126 | fio_file_set_done(f); |
429f6675 | 1127 | td->nr_done_files++; |
0b9d69ec JA |
1128 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1129 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1130 | } while (1); |
1131 | ||
1132 | return 0; | |
1133 | } | |
1134 | ||
3e260a46 JA |
1135 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1136 | unsigned long tusec, unsigned long max_usec) | |
1137 | { | |
1138 | if (!td->error) | |
1139 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1140 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1141 | icd->error = ETIMEDOUT; | |
1142 | } | |
1143 | ||
1144 | static void lat_new_cycle(struct thread_data *td) | |
1145 | { | |
1146 | fio_gettime(&td->latency_ts, NULL); | |
1147 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1148 | td->latency_failed = 0; | |
1149 | } | |
1150 | ||
1151 | /* | |
1152 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1153 | * are at QD=1, then it's time to give up. | |
1154 | */ | |
1155 | static int __lat_target_failed(struct thread_data *td) | |
1156 | { | |
1157 | if (td->latency_qd == 1) | |
1158 | return 1; | |
1159 | ||
1160 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1161 | |
1162 | if (td->latency_qd == td->latency_qd_low) | |
1163 | td->latency_qd_low--; | |
1164 | ||
3e260a46 JA |
1165 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1166 | ||
1167 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1168 | ||
1169 | /* | |
1170 | * When we ramp QD down, quiesce existing IO to prevent | |
1171 | * a storm of ramp downs due to pending higher depth. | |
1172 | */ | |
1173 | io_u_quiesce(td); | |
1174 | lat_new_cycle(td); | |
1175 | return 0; | |
1176 | } | |
1177 | ||
1178 | static int lat_target_failed(struct thread_data *td) | |
1179 | { | |
1180 | if (td->o.latency_percentile.u.f == 100.0) | |
1181 | return __lat_target_failed(td); | |
1182 | ||
1183 | td->latency_failed++; | |
1184 | return 0; | |
1185 | } | |
1186 | ||
1187 | void lat_target_init(struct thread_data *td) | |
1188 | { | |
6bb58215 JA |
1189 | td->latency_end_run = 0; |
1190 | ||
3e260a46 JA |
1191 | if (td->o.latency_target) { |
1192 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1193 | fio_gettime(&td->latency_ts, NULL); | |
1194 | td->latency_qd = 1; | |
1195 | td->latency_qd_high = td->o.iodepth; | |
1196 | td->latency_qd_low = 1; | |
1197 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1198 | } else | |
1199 | td->latency_qd = td->o.iodepth; | |
1200 | } | |
1201 | ||
6bb58215 JA |
1202 | void lat_target_reset(struct thread_data *td) |
1203 | { | |
1204 | if (!td->latency_end_run) | |
1205 | lat_target_init(td); | |
1206 | } | |
1207 | ||
3e260a46 JA |
1208 | static void lat_target_success(struct thread_data *td) |
1209 | { | |
1210 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1211 | struct thread_options *o = &td->o; |
3e260a46 JA |
1212 | |
1213 | td->latency_qd_low = td->latency_qd; | |
1214 | ||
1215 | /* | |
1216 | * If we haven't failed yet, we double up to a failing value instead | |
1217 | * of bisecting from highest possible queue depth. If we have set | |
1218 | * a limit other than td->o.iodepth, bisect between that. | |
1219 | */ | |
6bb58215 | 1220 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1221 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1222 | else | |
1223 | td->latency_qd *= 2; | |
1224 | ||
6bb58215 JA |
1225 | if (td->latency_qd > o->iodepth) |
1226 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1227 | |
1228 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1229 | |
3e260a46 | 1230 | /* |
6bb58215 JA |
1231 | * Same as last one, we are done. Let it run a latency cycle, so |
1232 | * we get only the results from the targeted depth. | |
3e260a46 | 1233 | */ |
6bb58215 JA |
1234 | if (td->latency_qd == qd) { |
1235 | if (td->latency_end_run) { | |
1236 | dprint(FD_RATE, "We are done\n"); | |
1237 | td->done = 1; | |
1238 | } else { | |
1239 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1240 | io_u_quiesce(td); | |
1241 | td->latency_end_run = 1; | |
1242 | reset_all_stats(td); | |
1243 | reset_io_stats(td); | |
1244 | } | |
1245 | } | |
3e260a46 JA |
1246 | |
1247 | lat_new_cycle(td); | |
1248 | } | |
1249 | ||
1250 | /* | |
1251 | * Check if we can bump the queue depth | |
1252 | */ | |
1253 | void lat_target_check(struct thread_data *td) | |
1254 | { | |
1255 | uint64_t usec_window; | |
1256 | uint64_t ios; | |
1257 | double success_ios; | |
1258 | ||
1259 | usec_window = utime_since_now(&td->latency_ts); | |
1260 | if (usec_window < td->o.latency_window) | |
1261 | return; | |
1262 | ||
1263 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1264 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1265 | success_ios *= 100.0; | |
1266 | ||
1267 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1268 | ||
1269 | if (success_ios >= td->o.latency_percentile.u.f) | |
1270 | lat_target_success(td); | |
1271 | else | |
1272 | __lat_target_failed(td); | |
1273 | } | |
1274 | ||
1275 | /* | |
1276 | * If latency target is enabled, we might be ramping up or down and not | |
1277 | * using the full queue depth available. | |
1278 | */ | |
effd6ff0 | 1279 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1280 | { |
1281 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1282 | ||
1283 | if (qempty) | |
1284 | return 1; | |
1285 | if (!td->o.latency_target) | |
1286 | return 0; | |
1287 | ||
1288 | return td->cur_depth >= td->latency_qd; | |
1289 | } | |
429f6675 | 1290 | |
10ba535a JA |
1291 | struct io_u *__get_io_u(struct thread_data *td) |
1292 | { | |
0cae66f6 | 1293 | struct io_u *io_u = NULL; |
10ba535a | 1294 | |
ca09be4b JA |
1295 | if (td->stop_io) |
1296 | return NULL; | |
1297 | ||
e8462bd8 JA |
1298 | td_io_u_lock(td); |
1299 | ||
1300 | again: | |
2ae0b204 JA |
1301 | if (!io_u_rempty(&td->io_u_requeues)) |
1302 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1303 | else if (!queue_full(td)) { |
2ae0b204 | 1304 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1305 | |
225ba9e3 | 1306 | io_u->file = NULL; |
6040dabc | 1307 | io_u->buflen = 0; |
10ba535a | 1308 | io_u->resid = 0; |
d7762cf8 | 1309 | io_u->end_io = NULL; |
755200a3 JA |
1310 | } |
1311 | ||
1312 | if (io_u) { | |
0c6e7517 | 1313 | assert(io_u->flags & IO_U_F_FREE); |
f8b0bd10 JA |
1314 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
1315 | IO_U_F_TRIMMED | IO_U_F_BARRIER | | |
1316 | IO_U_F_VER_LIST); | |
0c6e7517 | 1317 | |
755200a3 | 1318 | io_u->error = 0; |
bcd5abfa | 1319 | io_u->acct_ddir = -1; |
10ba535a | 1320 | td->cur_depth++; |
0c41214f | 1321 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
f9401285 | 1322 | io_u->ipo = NULL; |
1dec3e07 JA |
1323 | } else if (td->o.verify_async) { |
1324 | /* | |
1325 | * We ran out, wait for async verify threads to finish and | |
1326 | * return one | |
1327 | */ | |
1328 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1329 | goto again; | |
10ba535a JA |
1330 | } |
1331 | ||
e8462bd8 | 1332 | td_io_u_unlock(td); |
10ba535a JA |
1333 | return io_u; |
1334 | } | |
1335 | ||
0d29de83 | 1336 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1337 | { |
d72be545 JA |
1338 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1339 | return 0; | |
1340 | ||
1341 | if (td->trim_entries) { | |
0d29de83 | 1342 | int get_trim = 0; |
10ba535a | 1343 | |
0d29de83 JA |
1344 | if (td->trim_batch) { |
1345 | td->trim_batch--; | |
1346 | get_trim = 1; | |
1347 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1348 | td->last_ddir != DDIR_READ) { | |
1349 | td->trim_batch = td->o.trim_batch; | |
1350 | if (!td->trim_batch) | |
1351 | td->trim_batch = td->o.trim_backlog; | |
1352 | get_trim = 1; | |
1353 | } | |
1354 | ||
1355 | if (get_trim && !get_next_trim(td, io_u)) | |
1356 | return 1; | |
2ba1c290 | 1357 | } |
10ba535a | 1358 | |
0d29de83 JA |
1359 | return 0; |
1360 | } | |
1361 | ||
1362 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1363 | { | |
d72be545 JA |
1364 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1365 | return 0; | |
1366 | ||
1367 | if (td->io_hist_len) { | |
9e144189 JA |
1368 | int get_verify = 0; |
1369 | ||
d1ece0c7 | 1370 | if (td->verify_batch) |
9e144189 | 1371 | get_verify = 1; |
d1ece0c7 | 1372 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1373 | td->last_ddir != DDIR_READ) { |
1374 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1375 | if (!td->verify_batch) |
1376 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1377 | get_verify = 1; |
1378 | } | |
1379 | ||
d1ece0c7 JA |
1380 | if (get_verify && !get_next_verify(td, io_u)) { |
1381 | td->verify_batch--; | |
0d29de83 | 1382 | return 1; |
d1ece0c7 | 1383 | } |
9e144189 JA |
1384 | } |
1385 | ||
0d29de83 JA |
1386 | return 0; |
1387 | } | |
1388 | ||
de789769 JA |
1389 | /* |
1390 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1391 | * easy compressible data for simple de-dupe attempts. Do this for every |
1392 | * 512b block in the range, since that should be the smallest block size | |
1393 | * we can expect from a device. | |
de789769 JA |
1394 | */ |
1395 | static void small_content_scramble(struct io_u *io_u) | |
1396 | { | |
23f394d5 | 1397 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1398 | uint64_t boffset; |
23f394d5 JA |
1399 | unsigned int offset; |
1400 | void *p, *end; | |
de789769 | 1401 | |
23f394d5 JA |
1402 | if (!nr_blocks) |
1403 | return; | |
1404 | ||
1405 | p = io_u->xfer_buf; | |
fba76ee8 | 1406 | boffset = io_u->offset; |
81f0366c | 1407 | io_u->buf_filled_len = 0; |
fad82f76 | 1408 | |
23f394d5 JA |
1409 | for (i = 0; i < nr_blocks; i++) { |
1410 | /* | |
1411 | * Fill the byte offset into a "random" start offset of | |
1412 | * the buffer, given by the product of the usec time | |
1413 | * and the actual offset. | |
1414 | */ | |
fad82f76 | 1415 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1416 | offset &= ~(sizeof(uint64_t) - 1); |
1417 | if (offset >= 512 - sizeof(uint64_t)) | |
1418 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1419 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1420 | |
1421 | end = p + 512 - sizeof(io_u->start_time); | |
1422 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1423 | p += 512; | |
fad82f76 | 1424 | boffset += 512; |
23f394d5 | 1425 | } |
de789769 JA |
1426 | } |
1427 | ||
0d29de83 JA |
1428 | /* |
1429 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1430 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1431 | */ | |
1432 | struct io_u *get_io_u(struct thread_data *td) | |
1433 | { | |
1434 | struct fio_file *f; | |
1435 | struct io_u *io_u; | |
de789769 | 1436 | int do_scramble = 0; |
002fe734 | 1437 | long ret = 0; |
0d29de83 JA |
1438 | |
1439 | io_u = __get_io_u(td); | |
1440 | if (!io_u) { | |
1441 | dprint(FD_IO, "__get_io_u failed\n"); | |
1442 | return NULL; | |
1443 | } | |
1444 | ||
1445 | if (check_get_verify(td, io_u)) | |
1446 | goto out; | |
1447 | if (check_get_trim(td, io_u)) | |
1448 | goto out; | |
1449 | ||
755200a3 JA |
1450 | /* |
1451 | * from a requeue, io_u already setup | |
1452 | */ | |
1453 | if (io_u->file) | |
77f392bf | 1454 | goto out; |
755200a3 | 1455 | |
429f6675 JA |
1456 | /* |
1457 | * If using an iolog, grab next piece if any available. | |
1458 | */ | |
d72be545 | 1459 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1460 | if (read_iolog_get(td, io_u)) |
1461 | goto err_put; | |
2ba1c290 | 1462 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1463 | ret = -EBUSY; |
2ba1c290 | 1464 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1465 | goto err_put; |
2ba1c290 | 1466 | } |
5ec10eaa | 1467 | |
429f6675 | 1468 | f = io_u->file; |
002fe734 JA |
1469 | if (!f) { |
1470 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1471 | goto err_put; | |
1472 | } | |
1473 | ||
d6aed795 | 1474 | assert(fio_file_open(f)); |
97af62ce | 1475 | |
ff58fced | 1476 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1477 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1478 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1479 | goto err_put; |
2ba1c290 | 1480 | } |
10ba535a | 1481 | |
f1dfb668 JA |
1482 | f->last_start[io_u->ddir] = io_u->offset; |
1483 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1484 | |
fd68418e | 1485 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1486 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1487 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1488 | td->o.min_bs[DDIR_WRITE], |
1489 | io_u->xfer_buflen); | |
bedc9dc2 JA |
1490 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
1491 | !(td->flags & TD_F_COMPRESS)) | |
fd68418e | 1492 | do_scramble = 1; |
d72be545 | 1493 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1494 | populate_verify_io_u(td, io_u); |
1495 | do_scramble = 0; | |
1496 | } | |
fd68418e | 1497 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1498 | /* |
1499 | * Reset the buf_filled parameters so next time if the | |
1500 | * buffer is used for writes it is refilled. | |
1501 | */ | |
cbe8d756 RR |
1502 | io_u->buf_filled_len = 0; |
1503 | } | |
87dc1ab1 | 1504 | } |
10ba535a | 1505 | |
165faf16 JA |
1506 | /* |
1507 | * Set io data pointers. | |
1508 | */ | |
cec6b55d JA |
1509 | io_u->xfer_buf = io_u->buf; |
1510 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1511 | |
6ac7a331 | 1512 | out: |
0d29de83 | 1513 | assert(io_u->file); |
429f6675 | 1514 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1515 | if (!td->o.disable_slat) |
1516 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1517 | if (do_scramble) |
1518 | small_content_scramble(io_u); | |
429f6675 | 1519 | return io_u; |
36167d82 | 1520 | } |
429f6675 | 1521 | err_put: |
2ba1c290 | 1522 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1523 | put_io_u(td, io_u); |
002fe734 | 1524 | return ERR_PTR(ret); |
10ba535a JA |
1525 | } |
1526 | ||
5451792e JA |
1527 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1528 | { | |
8b28bd41 | 1529 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1530 | |
8b28bd41 DM |
1531 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1532 | return; | |
5451792e | 1533 | |
709c8313 RE |
1534 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1535 | io_u->file ? " on file " : "", | |
1536 | io_u->file ? io_u->file->file_name : "", | |
1537 | strerror(io_u->error), | |
1538 | io_ddir_name(io_u->ddir), | |
1539 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1540 | |
1541 | if (!td->error) | |
1542 | td_verror(td, io_u->error, "io_u error"); | |
1543 | } | |
1544 | ||
aba6c951 JA |
1545 | static inline int gtod_reduce(struct thread_data *td) |
1546 | { | |
729fe3af | 1547 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1548 | && td->o.disable_bw; |
aba6c951 JA |
1549 | } |
1550 | ||
c8eeb9df JA |
1551 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1552 | struct io_completion_data *icd, | |
1553 | const enum fio_ddir idx, unsigned int bytes) | |
1554 | { | |
24d23ca7 | 1555 | unsigned long lusec = 0; |
c8eeb9df | 1556 | |
aba6c951 | 1557 | if (!gtod_reduce(td)) |
c8eeb9df JA |
1558 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1559 | ||
1560 | if (!td->o.disable_lat) { | |
1561 | unsigned long tusec; | |
1562 | ||
1563 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1564 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1565 | |
d4afedfd JA |
1566 | if (td->flags & TD_F_PROFILE_OPS) { |
1567 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1568 | ||
1569 | if (ops->io_u_lat) | |
1570 | icd->error = ops->io_u_lat(td, tusec); | |
1571 | } | |
1572 | ||
3e260a46 JA |
1573 | if (td->o.max_latency && tusec > td->o.max_latency) |
1574 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1575 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1576 | if (lat_target_failed(td)) | |
1577 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1578 | } |
c8eeb9df JA |
1579 | } |
1580 | ||
1581 | if (!td->o.disable_clat) { | |
ae588852 | 1582 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1583 | io_u_mark_latency(td, lusec); |
1584 | } | |
1585 | ||
1586 | if (!td->o.disable_bw) | |
1587 | add_bw_sample(td, idx, bytes, &icd->time); | |
1588 | ||
aba6c951 JA |
1589 | if (!gtod_reduce(td)) |
1590 | add_iops_sample(td, idx, bytes, &icd->time); | |
c8eeb9df JA |
1591 | } |
1592 | ||
1b8dbf25 SL |
1593 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1594 | { | |
1ae83d45 JA |
1595 | uint64_t secs, remainder, bps, bytes; |
1596 | ||
1b8dbf25 SL |
1597 | bytes = td->this_io_bytes[ddir]; |
1598 | bps = td->rate_bps[ddir]; | |
1599 | secs = bytes / bps; | |
1600 | remainder = bytes % bps; | |
1601 | return remainder * 1000000 / bps + secs * 1000000; | |
1602 | } | |
1603 | ||
f8b0bd10 | 1604 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1605 | struct io_completion_data *icd) |
10ba535a | 1606 | { |
f8b0bd10 JA |
1607 | struct io_u *io_u = *io_u_ptr; |
1608 | enum fio_ddir ddir = io_u->ddir; | |
1609 | struct fio_file *f = io_u->file; | |
10ba535a | 1610 | |
2ba1c290 JA |
1611 | dprint_io_u(io_u, "io complete"); |
1612 | ||
2ecc1b57 | 1613 | td_io_u_lock(td); |
0c6e7517 | 1614 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1615 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1616 | |
1617 | /* | |
1618 | * Mark IO ok to verify | |
1619 | */ | |
1620 | if (io_u->ipo) { | |
890b6656 JA |
1621 | /* |
1622 | * Remove errored entry from the verification list | |
1623 | */ | |
1624 | if (io_u->error) | |
1625 | unlog_io_piece(td, io_u); | |
1626 | else { | |
1627 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1628 | write_barrier(); | |
1629 | } | |
f9401285 JA |
1630 | } |
1631 | ||
2ecc1b57 | 1632 | td_io_u_unlock(td); |
0c6e7517 | 1633 | |
f8b0bd10 | 1634 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1635 | td->last_was_sync = 1; |
44f29692 JA |
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; | |
f8b0bd10 | 1644 | td->last_ddir = ddir; |
87dc1ab1 | 1645 | |
f8b0bd10 | 1646 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1647 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1648 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1649 | int ret; |
10ba535a | 1650 | |
f8b0bd10 JA |
1651 | td->io_blocks[ddir]++; |
1652 | td->this_io_blocks[ddir]++; | |
1653 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1654 | |
1655 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1656 | td->this_io_bytes[ddir] += bytes; |
1657 | ||
ca09be4b JA |
1658 | if (ddir == DDIR_WRITE) { |
1659 | if (f) { | |
1660 | if (f->first_write == -1ULL || | |
1661 | io_u->offset < f->first_write) | |
1662 | f->first_write = io_u->offset; | |
1663 | if (f->last_write == -1ULL || | |
1664 | ((io_u->offset + bytes) > f->last_write)) | |
1665 | f->last_write = io_u->offset + bytes; | |
1666 | } | |
1667 | if (td->last_write_comp) { | |
1668 | int idx = td->last_write_idx++; | |
1669 | ||
1670 | td->last_write_comp[idx] = io_u->offset; | |
1671 | if (td->last_write_idx == td->o.iodepth) | |
1672 | td->last_write_idx = 0; | |
1673 | } | |
44f29692 JA |
1674 | } |
1675 | ||
6b1190fd SL |
1676 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1677 | td->runstate == TD_VERIFYING)) { | |
f8b0bd10 | 1678 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1679 | |
f8b0bd10 JA |
1680 | if (__should_check_rate(td, ddir)) { |
1681 | td->rate_pending_usleep[ddir] = | |
1682 | (usec_for_io(td, ddir) - | |
ba3e4e0c | 1683 | utime_since_now(&td->start)); |
b23b6a2f | 1684 | } |
f8b0bd10 JA |
1685 | if (ddir != DDIR_TRIM && |
1686 | __should_check_rate(td, oddir)) { | |
1687 | td->rate_pending_usleep[oddir] = | |
1688 | (usec_for_io(td, oddir) - | |
ba3e4e0c | 1689 | utime_since_now(&td->start)); |
f8b0bd10 | 1690 | } |
721938ae | 1691 | } |
10ba535a | 1692 | |
f8b0bd10 | 1693 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1694 | |
d7762cf8 | 1695 | if (io_u->end_io) { |
f8b0bd10 JA |
1696 | ret = io_u->end_io(td, io_u_ptr); |
1697 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1698 | if (ret && !icd->error) |
1699 | icd->error = ret; | |
1700 | } | |
ff58fced | 1701 | } else if (io_u->error) { |
10ba535a | 1702 | icd->error = io_u->error; |
5451792e JA |
1703 | io_u_log_error(td, io_u); |
1704 | } | |
8b28bd41 | 1705 | if (icd->error) { |
f8b0bd10 JA |
1706 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1707 | ||
8b28bd41 DM |
1708 | if (!td_non_fatal_error(td, eb, icd->error)) |
1709 | return; | |
f8b0bd10 | 1710 | |
f2bba182 RR |
1711 | /* |
1712 | * If there is a non_fatal error, then add to the error count | |
1713 | * and clear all the errors. | |
1714 | */ | |
1715 | update_error_count(td, icd->error); | |
1716 | td_clear_error(td); | |
1717 | icd->error = 0; | |
f8b0bd10 JA |
1718 | if (io_u) |
1719 | io_u->error = 0; | |
f2bba182 | 1720 | } |
10ba535a JA |
1721 | } |
1722 | ||
9520ebb9 JA |
1723 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1724 | int nr) | |
10ba535a | 1725 | { |
6eaf09d6 | 1726 | int ddir; |
aba6c951 JA |
1727 | |
1728 | if (!gtod_reduce(td)) | |
9520ebb9 | 1729 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1730 | |
3af6ef39 JA |
1731 | icd->nr = nr; |
1732 | ||
10ba535a | 1733 | icd->error = 0; |
6eaf09d6 SL |
1734 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1735 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1736 | } |
1737 | ||
97601024 JA |
1738 | static void ios_completed(struct thread_data *td, |
1739 | struct io_completion_data *icd) | |
36167d82 JA |
1740 | { |
1741 | struct io_u *io_u; | |
1742 | int i; | |
1743 | ||
10ba535a JA |
1744 | for (i = 0; i < icd->nr; i++) { |
1745 | io_u = td->io_ops->event(td, i); | |
1746 | ||
f8b0bd10 | 1747 | io_completed(td, &io_u, icd); |
e8462bd8 | 1748 | |
f8b0bd10 | 1749 | if (io_u) |
e8462bd8 | 1750 | put_io_u(td, io_u); |
10ba535a JA |
1751 | } |
1752 | } | |
97601024 | 1753 | |
e7e6cfb4 JA |
1754 | /* |
1755 | * Complete a single io_u for the sync engines. | |
1756 | */ | |
581e7141 | 1757 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
100f49f1 | 1758 | uint64_t *bytes) |
97601024 JA |
1759 | { |
1760 | struct io_completion_data icd; | |
1761 | ||
9520ebb9 | 1762 | init_icd(td, &icd, 1); |
f8b0bd10 | 1763 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1764 | |
f8b0bd10 | 1765 | if (io_u) |
e8462bd8 | 1766 | put_io_u(td, io_u); |
97601024 | 1767 | |
581e7141 JA |
1768 | if (icd.error) { |
1769 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1770 | return -1; | |
1771 | } | |
97601024 | 1772 | |
581e7141 | 1773 | if (bytes) { |
6eaf09d6 SL |
1774 | int ddir; |
1775 | ||
1776 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1777 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1778 | } |
1779 | ||
1780 | return 0; | |
97601024 JA |
1781 | } |
1782 | ||
e7e6cfb4 JA |
1783 | /* |
1784 | * Called to complete min_events number of io for the async engines. | |
1785 | */ | |
581e7141 | 1786 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
100f49f1 | 1787 | uint64_t *bytes) |
97601024 | 1788 | { |
97601024 | 1789 | struct io_completion_data icd; |
00de55ef | 1790 | struct timespec *tvp = NULL; |
97601024 | 1791 | int ret; |
4d06a338 | 1792 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1793 | |
4950421a | 1794 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1795 | |
4950421a | 1796 | if (!min_evts) |
00de55ef | 1797 | tvp = &ts; |
5fb4b366 RE |
1798 | else if (min_evts > td->cur_depth) |
1799 | min_evts = td->cur_depth; | |
97601024 | 1800 | |
4950421a | 1801 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1802 | if (ret < 0) { |
e1161c32 | 1803 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1804 | return ret; |
1805 | } else if (!ret) | |
1806 | return ret; | |
1807 | ||
9520ebb9 | 1808 | init_icd(td, &icd, ret); |
97601024 | 1809 | ios_completed(td, &icd); |
581e7141 JA |
1810 | if (icd.error) { |
1811 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1812 | return -1; | |
1813 | } | |
97601024 | 1814 | |
581e7141 | 1815 | if (bytes) { |
6eaf09d6 SL |
1816 | int ddir; |
1817 | ||
1818 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) | |
1819 | bytes[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1820 | } |
1821 | ||
1822 | return 0; | |
97601024 | 1823 | } |
7e77dd02 JA |
1824 | |
1825 | /* | |
1826 | * Call when io_u is really queued, to update the submission latency. | |
1827 | */ | |
1828 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1829 | { | |
9520ebb9 JA |
1830 | if (!td->o.disable_slat) { |
1831 | unsigned long slat_time; | |
7e77dd02 | 1832 | |
9520ebb9 | 1833 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1834 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1835 | io_u->offset); | |
9520ebb9 | 1836 | } |
7e77dd02 | 1837 | } |
433afcb4 | 1838 | |
5c94b008 JA |
1839 | /* |
1840 | * See if we should reuse the last seed, if dedupe is enabled | |
1841 | */ | |
1842 | static struct frand_state *get_buf_state(struct thread_data *td) | |
1843 | { | |
1844 | unsigned int v; | |
1845 | unsigned long r; | |
1846 | ||
1847 | if (!td->o.dedupe_percentage) | |
1848 | return &td->buf_state; | |
c0b69b92 JA |
1849 | else if (td->o.dedupe_percentage == 100) |
1850 | return &td->buf_state_prev; | |
5c94b008 JA |
1851 | |
1852 | r = __rand(&td->dedupe_state); | |
1853 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
1854 | ||
1855 | if (v <= td->o.dedupe_percentage) | |
1856 | return &td->buf_state_prev; | |
1857 | ||
1858 | return &td->buf_state; | |
1859 | } | |
1860 | ||
1861 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) | |
1862 | { | |
1863 | if (rs == &td->buf_state) | |
1864 | frand_copy(&td->buf_state_prev, rs); | |
1865 | } | |
1866 | ||
cc86c395 JA |
1867 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1868 | unsigned int max_bs) | |
5973cafb | 1869 | { |
d1af2894 JA |
1870 | struct thread_options *o = &td->o; |
1871 | ||
4eff3e57 | 1872 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 1873 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1874 | struct frand_state *rs; |
1066358a | 1875 | unsigned int left = max_bs; |
5c94b008 | 1876 | |
1066358a JA |
1877 | do { |
1878 | rs = get_buf_state(td); | |
9c42684e | 1879 | |
1066358a | 1880 | min_write = min(min_write, left); |
f97a43a1 | 1881 | |
1066358a JA |
1882 | if (perc) { |
1883 | unsigned int seg = min_write; | |
cc86c395 | 1884 | |
1066358a JA |
1885 | seg = min(min_write, td->o.compress_chunk); |
1886 | if (!seg) | |
1887 | seg = min_write; | |
1888 | ||
1889 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1890 | min_write, o->buffer_pattern, |
1891 | o->buffer_pattern_bytes); | |
1066358a JA |
1892 | } else |
1893 | fill_random_buf(rs, buf, min_write); | |
1894 | ||
1895 | buf += min_write; | |
1896 | left -= min_write; | |
5c94b008 | 1897 | save_buf_state(td, rs); |
1066358a | 1898 | } while (left); |
d1af2894 JA |
1899 | } else if (o->buffer_pattern_bytes) |
1900 | fill_buffer_pattern(td, buf, max_bs); | |
1901 | else | |
cc86c395 JA |
1902 | memset(buf, 0, max_bs); |
1903 | } | |
1904 | ||
1905 | /* | |
1906 | * "randomly" fill the buffer contents | |
1907 | */ | |
1908 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1909 | unsigned int min_write, unsigned int max_bs) | |
1910 | { | |
1911 | io_u->buf_filled_len = 0; | |
1912 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1913 | } |