Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
5973cafb | 9 | #include "hash.h" |
4f5af7b2 | 10 | #include "verify.h" |
0d29de83 | 11 | #include "trim.h" |
1fbbf72e | 12 | #include "lib/rand.h" |
7ebd796f | 13 | #include "lib/axmap.h" |
002fe734 | 14 | #include "err.h" |
0f38bbef | 15 | #include "lib/pow2.h" |
10ba535a | 16 | |
97601024 JA |
17 | struct io_completion_data { |
18 | int nr; /* input */ | |
97601024 JA |
19 | |
20 | int error; /* output */ | |
100f49f1 | 21 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
97601024 JA |
22 | struct timeval time; /* output */ |
23 | }; | |
24 | ||
10ba535a | 25 | /* |
7ebd796f | 26 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
27 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
28 | */ | |
1ae83d45 | 29 | static int random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 30 | { |
7ebd796f | 31 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
32 | } |
33 | ||
df415585 JA |
34 | /* |
35 | * Mark a given offset as used in the map. | |
36 | */ | |
9bf2061e | 37 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 38 | { |
2dc1bbeb | 39 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 40 | struct fio_file *f = io_u->file; |
51ede0b1 | 41 | unsigned int nr_blocks; |
1ae83d45 | 42 | uint64_t block; |
df415585 | 43 | |
1ae83d45 | 44 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 45 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 46 | |
2ab9e98b | 47 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 48 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 49 | |
51ede0b1 JA |
50 | if ((nr_blocks * min_bs) < io_u->buflen) |
51 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
52 | } |
53 | ||
74776733 JA |
54 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
55 | enum fio_ddir ddir) | |
2ba1c290 | 56 | { |
74776733 JA |
57 | uint64_t max_blocks; |
58 | uint64_t max_size; | |
2ba1c290 | 59 | |
ff58fced JA |
60 | assert(ddir_rw(ddir)); |
61 | ||
d9dd70f7 JA |
62 | /* |
63 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
64 | */ | |
65 | max_size = f->io_size; | |
66 | if (max_size > f->real_file_size) | |
67 | max_size = f->real_file_size; | |
68 | ||
ed335855 SN |
69 | if (td->o.zone_range) |
70 | max_size = td->o.zone_range; | |
71 | ||
0412d12e JE |
72 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
73 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
74 | ||
1ae83d45 | 75 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
76 | if (!max_blocks) |
77 | return 0; | |
78 | ||
67778e88 | 79 | return max_blocks; |
2ba1c290 JA |
80 | } |
81 | ||
1ae83d45 JA |
82 | struct rand_off { |
83 | struct flist_head list; | |
84 | uint64_t off; | |
85 | }; | |
86 | ||
e25839d4 | 87 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 88 | enum fio_ddir ddir, uint64_t *b) |
ec4015da | 89 | { |
6f49f8bc | 90 | uint64_t r; |
5e0baa7f | 91 | |
c3546b53 JA |
92 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
93 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { | |
94 | uint64_t frand_max, lastb; | |
6f49f8bc JA |
95 | |
96 | lastb = last_block(td, f, ddir); | |
97 | if (!lastb) | |
98 | return 1; | |
99 | ||
c3546b53 | 100 | frand_max = rand_max(&td->random_state); |
d6b72507 | 101 | r = __rand(&td->random_state); |
8055e41d | 102 | |
4b91ee8f | 103 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 104 | |
c3546b53 | 105 | *b = lastb * (r / ((uint64_t) frand_max + 1.0)); |
51ede0b1 | 106 | } else { |
8055e41d | 107 | uint64_t off = 0; |
43c63a78 | 108 | |
967d1b63 JA |
109 | assert(fio_file_lfsr(f)); |
110 | ||
6f49f8bc | 111 | if (lfsr_next(&f->lfsr, &off)) |
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 | ||
56d9fa4b JA |
155 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
156 | struct fio_file *f, enum fio_ddir ddir, | |
157 | uint64_t *b) | |
158 | { | |
159 | *b = gauss_next(&f->gauss); | |
160 | return 0; | |
161 | } | |
162 | ||
163 | ||
1ae83d45 JA |
164 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
165 | { | |
166 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
167 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
168 | ||
169 | return r1->off - r2->off; | |
170 | } | |
171 | ||
172 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
173 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 JA |
174 | { |
175 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) | |
176 | return __get_next_rand_offset(td, f, ddir, b); | |
177 | else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
178 | return __get_next_rand_offset_zipf(td, f, ddir, b); | |
925fee33 JA |
179 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
180 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
181 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
182 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e25839d4 JA |
183 | |
184 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
185 | return 1; | |
186 | } | |
187 | ||
bcd5abfa JA |
188 | /* |
189 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
190 | * specified. | |
191 | */ | |
192 | static inline int should_sort_io(struct thread_data *td) | |
193 | { | |
194 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
195 | return 0; | |
196 | if (!td_random(td)) | |
197 | return 0; | |
198 | if (td->runstate != TD_VERIFYING) | |
199 | return 0; | |
c3546b53 JA |
200 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
201 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) | |
bcd5abfa JA |
202 | return 0; |
203 | ||
204 | return 1; | |
205 | } | |
206 | ||
d9472271 | 207 | static int should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 | 208 | { |
c3546b53 | 209 | uint64_t frand_max; |
211c9b89 JA |
210 | unsigned int v; |
211 | unsigned long r; | |
212 | ||
d9472271 | 213 | if (td->o.perc_rand[ddir] == 100) |
211c9b89 JA |
214 | return 1; |
215 | ||
c3546b53 | 216 | frand_max = rand_max(&td->seq_rand_state[ddir]); |
d6b72507 | 217 | r = __rand(&td->seq_rand_state[ddir]); |
c3546b53 | 218 | v = 1 + (int) (100.0 * (r / (frand_max + 1.0))); |
211c9b89 | 219 | |
d9472271 | 220 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
221 | } |
222 | ||
1ae83d45 JA |
223 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
224 | enum fio_ddir ddir, uint64_t *b) | |
225 | { | |
226 | struct rand_off *r; | |
227 | int i, ret = 1; | |
228 | ||
bcd5abfa | 229 | if (!should_sort_io(td)) |
1ae83d45 JA |
230 | return get_off_from_method(td, f, ddir, b); |
231 | ||
232 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 233 | fetch: |
9342d5f8 | 234 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
235 | flist_del(&r->list); |
236 | *b = r->off; | |
237 | free(r); | |
238 | return 0; | |
239 | } | |
240 | ||
241 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
242 | r = malloc(sizeof(*r)); | |
243 | ||
244 | ret = get_off_from_method(td, f, ddir, &r->off); | |
245 | if (ret) { | |
246 | free(r); | |
247 | break; | |
248 | } | |
249 | ||
250 | flist_add(&r->list, &td->next_rand_list); | |
251 | } | |
252 | ||
253 | if (ret && !i) | |
254 | return ret; | |
255 | ||
256 | assert(!flist_empty(&td->next_rand_list)); | |
257 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
258 | goto fetch; | |
259 | } | |
260 | ||
38dad62d | 261 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 262 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 263 | { |
c04e4661 DE |
264 | if (!get_next_rand_offset(td, f, ddir, b)) |
265 | return 0; | |
266 | ||
267 | if (td->o.time_based) { | |
33c48814 | 268 | fio_file_reset(td, f); |
c04e4661 DE |
269 | if (!get_next_rand_offset(td, f, ddir, b)) |
270 | return 0; | |
38dad62d JA |
271 | } |
272 | ||
c04e4661 | 273 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 274 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 275 | (unsigned long long) f->real_file_size); |
c04e4661 | 276 | return 1; |
38dad62d JA |
277 | } |
278 | ||
37cf9e3c | 279 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 280 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 281 | { |
ac002339 JA |
282 | struct thread_options *o = &td->o; |
283 | ||
ff58fced JA |
284 | assert(ddir_rw(ddir)); |
285 | ||
f1dfb668 | 286 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
ac002339 | 287 | o->time_based) |
f1dfb668 | 288 | f->last_pos[ddir] = f->last_pos[ddir] - f->io_size; |
c04e4661 | 289 | |
f1dfb668 | 290 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 291 | uint64_t pos; |
059b0802 | 292 | |
f1dfb668 JA |
293 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) |
294 | f->last_pos[ddir] = f->real_file_size; | |
a66da7a2 | 295 | |
f1dfb668 | 296 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
297 | if (pos && o->ddir_seq_add) { |
298 | pos += o->ddir_seq_add; | |
299 | ||
300 | /* | |
301 | * If we reach beyond the end of the file | |
302 | * with holed IO, wrap around to the | |
303 | * beginning again. | |
304 | */ | |
305 | if (pos >= f->real_file_size) | |
306 | pos = f->file_offset; | |
307 | } | |
059b0802 | 308 | |
37cf9e3c | 309 | *offset = pos; |
38dad62d JA |
310 | return 0; |
311 | } | |
312 | ||
313 | return 1; | |
314 | } | |
315 | ||
316 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
317 | enum fio_ddir ddir, int rw_seq, |
318 | unsigned int *is_random) | |
38dad62d JA |
319 | { |
320 | struct fio_file *f = io_u->file; | |
1ae83d45 | 321 | uint64_t b, offset; |
38dad62d JA |
322 | int ret; |
323 | ||
ff58fced JA |
324 | assert(ddir_rw(ddir)); |
325 | ||
37cf9e3c JA |
326 | b = offset = -1ULL; |
327 | ||
38dad62d | 328 | if (rw_seq) { |
211c9b89 | 329 | if (td_random(td)) { |
6aca9b3d | 330 | if (should_do_random(td, ddir)) { |
211c9b89 | 331 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
332 | *is_random = 1; |
333 | } else { | |
334 | *is_random = 0; | |
a9da8ab2 | 335 | io_u_set(io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
336 | ret = get_next_seq_offset(td, f, ddir, &offset); |
337 | if (ret) | |
338 | ret = get_next_rand_block(td, f, ddir, &b); | |
339 | } | |
6aca9b3d JA |
340 | } else { |
341 | *is_random = 0; | |
37cf9e3c | 342 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 343 | } |
38dad62d | 344 | } else { |
a9da8ab2 | 345 | io_u_set(io_u, IO_U_F_BUSY_OK); |
6aca9b3d | 346 | *is_random = 0; |
38dad62d JA |
347 | |
348 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 349 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 350 | if (ret) { |
37cf9e3c | 351 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
352 | *is_random = 0; |
353 | } | |
38dad62d | 354 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
355 | if (f->last_start[ddir] != -1ULL) |
356 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 357 | else |
37cf9e3c | 358 | offset = 0; |
38dad62d JA |
359 | ret = 0; |
360 | } else { | |
361 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
362 | ret = 1; | |
363 | } | |
364 | } | |
6d68b997 | 365 | |
37cf9e3c JA |
366 | if (!ret) { |
367 | if (offset != -1ULL) | |
368 | io_u->offset = offset; | |
369 | else if (b != -1ULL) | |
370 | io_u->offset = b * td->o.ba[ddir]; | |
371 | else { | |
4e0a8fa2 | 372 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
373 | ret = 1; |
374 | } | |
375 | } | |
376 | ||
38dad62d JA |
377 | return ret; |
378 | } | |
379 | ||
10ba535a JA |
380 | /* |
381 | * For random io, generate a random new block and see if it's used. Repeat | |
382 | * until we find a free one. For sequential io, just return the end of | |
383 | * the last io issued. | |
384 | */ | |
6aca9b3d JA |
385 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
386 | unsigned int *is_random) | |
10ba535a | 387 | { |
9bf2061e | 388 | struct fio_file *f = io_u->file; |
4ba66134 | 389 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 390 | int rw_seq_hit = 0; |
10ba535a | 391 | |
ff58fced JA |
392 | assert(ddir_rw(ddir)); |
393 | ||
38dad62d JA |
394 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
395 | rw_seq_hit = 1; | |
5736c10d | 396 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 397 | } |
211097b2 | 398 | |
6aca9b3d | 399 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 400 | return 1; |
10ba535a | 401 | |
009bd847 JA |
402 | if (io_u->offset >= f->io_size) { |
403 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
404 | (unsigned long long) io_u->offset, |
405 | (unsigned long long) f->io_size); | |
009bd847 JA |
406 | return 1; |
407 | } | |
408 | ||
409 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
410 | if (io_u->offset >= f->real_file_size) { |
411 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
412 | (unsigned long long) io_u->offset, |
413 | (unsigned long long) f->real_file_size); | |
10ba535a | 414 | return 1; |
2ba1c290 | 415 | } |
10ba535a JA |
416 | |
417 | return 0; | |
418 | } | |
419 | ||
6aca9b3d JA |
420 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
421 | unsigned int *is_random) | |
15dc1934 | 422 | { |
d72be545 JA |
423 | if (td->flags & TD_F_PROFILE_OPS) { |
424 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 425 | |
d72be545 | 426 | if (ops->fill_io_u_off) |
6aca9b3d | 427 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 428 | } |
15dc1934 | 429 | |
6aca9b3d | 430 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
431 | } |
432 | ||
79944128 JA |
433 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
434 | unsigned int buflen) | |
435 | { | |
436 | struct fio_file *f = io_u->file; | |
437 | ||
bedc9dc2 | 438 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
439 | } |
440 | ||
6aca9b3d JA |
441 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
442 | unsigned int is_random) | |
10ba535a | 443 | { |
6aca9b3d | 444 | int ddir = io_u->ddir; |
24d23ca7 | 445 | unsigned int buflen = 0; |
f3059de1 | 446 | unsigned int minbs, maxbs; |
c3546b53 | 447 | uint64_t frand_max; |
54a21917 | 448 | unsigned long r; |
10ba535a | 449 | |
9ee1c647 | 450 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
451 | |
452 | if (td->o.bs_is_seq_rand) | |
453 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 454 | |
f3059de1 JA |
455 | minbs = td->o.min_bs[ddir]; |
456 | maxbs = td->o.max_bs[ddir]; | |
457 | ||
79944128 JA |
458 | if (minbs == maxbs) |
459 | return minbs; | |
460 | ||
52c58027 JA |
461 | /* |
462 | * If we can't satisfy the min block size from here, then fail | |
463 | */ | |
464 | if (!io_u_fits(td, io_u, minbs)) | |
465 | return 0; | |
466 | ||
c3546b53 | 467 | frand_max = rand_max(&td->bsrange_state); |
79944128 | 468 | do { |
d6b72507 | 469 | r = __rand(&td->bsrange_state); |
4c07ad86 | 470 | |
720e84ad | 471 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 472 | buflen = 1 + (unsigned int) ((double) maxbs * |
c3546b53 | 473 | (r / (frand_max + 1.0))); |
f3059de1 JA |
474 | if (buflen < minbs) |
475 | buflen = minbs; | |
5ec10eaa | 476 | } else { |
564ca972 JA |
477 | long perc = 0; |
478 | unsigned int i; | |
479 | ||
720e84ad JA |
480 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
481 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
482 | |
483 | buflen = bsp->bs; | |
484 | perc += bsp->perc; | |
c3546b53 | 485 | if ((r <= ((frand_max / 100L) * perc)) && |
79944128 | 486 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
487 | break; |
488 | } | |
489 | } | |
79944128 | 490 | |
a9f70b1f JB |
491 | if (td->o.do_verify && td->o.verify != VERIFY_NONE) |
492 | buflen = (buflen + td->o.verify_interval - 1) & | |
493 | ~(td->o.verify_interval - 1); | |
494 | ||
f3059de1 | 495 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
7c306bb1 | 496 | buflen &= ~(minbs - 1); |
10ba535a | 497 | |
79944128 | 498 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 499 | |
10ba535a JA |
500 | return buflen; |
501 | } | |
502 | ||
6aca9b3d JA |
503 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
504 | unsigned int is_random) | |
15dc1934 | 505 | { |
d72be545 JA |
506 | if (td->flags & TD_F_PROFILE_OPS) { |
507 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 508 | |
d72be545 | 509 | if (ops->fill_io_u_size) |
6aca9b3d | 510 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 511 | } |
15dc1934 | 512 | |
6aca9b3d | 513 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
514 | } |
515 | ||
afe24a5a JA |
516 | static void set_rwmix_bytes(struct thread_data *td) |
517 | { | |
afe24a5a JA |
518 | unsigned int diff; |
519 | ||
520 | /* | |
521 | * we do time or byte based switch. this is needed because | |
522 | * buffered writes may issue a lot quicker than they complete, | |
523 | * whereas reads do not. | |
524 | */ | |
e47f799f | 525 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 526 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
527 | } |
528 | ||
529 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
530 | { | |
c3546b53 | 531 | uint64_t frand_max = rand_max(&td->rwmix_state); |
e47f799f | 532 | unsigned int v; |
1294c3ec | 533 | unsigned long r; |
e47f799f | 534 | |
d6b72507 | 535 | r = __rand(&td->rwmix_state); |
c3546b53 | 536 | v = 1 + (int) (100.0 * (r / (frand_max + 1.0))); |
4c07ad86 | 537 | |
04c540d9 | 538 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
539 | return DDIR_READ; |
540 | ||
541 | return DDIR_WRITE; | |
afe24a5a JA |
542 | } |
543 | ||
002e7183 JA |
544 | void io_u_quiesce(struct thread_data *td) |
545 | { | |
546 | /* | |
547 | * We are going to sleep, ensure that we flush anything pending as | |
548 | * not to skew our latency numbers. | |
549 | * | |
550 | * Changed to only monitor 'in flight' requests here instead of the | |
551 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
552 | * io's that have been actually submitted to an async engine, | |
553 | * and cur_depth is meaningless for sync engines. | |
554 | */ | |
9cc80b6d JA |
555 | if (td->io_u_queued || td->cur_depth) { |
556 | int fio_unused ret; | |
557 | ||
558 | ret = td_io_commit(td); | |
559 | } | |
560 | ||
002e7183 JA |
561 | while (td->io_u_in_flight) { |
562 | int fio_unused ret; | |
563 | ||
55312f9f | 564 | ret = io_u_queued_complete(td, 1); |
002e7183 JA |
565 | } |
566 | } | |
567 | ||
581e7141 JA |
568 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
569 | { | |
570 | enum fio_ddir odir = ddir ^ 1; | |
581e7141 JA |
571 | long usec; |
572 | ||
ff58fced JA |
573 | assert(ddir_rw(ddir)); |
574 | ||
315fcfec | 575 | if (td->rate_pending_usleep[ddir] <= 0) |
581e7141 JA |
576 | return ddir; |
577 | ||
578 | /* | |
579 | * We have too much pending sleep in this direction. See if we | |
580 | * should switch. | |
581 | */ | |
315fcfec | 582 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 JA |
583 | /* |
584 | * Other direction does not have too much pending, switch | |
585 | */ | |
586 | if (td->rate_pending_usleep[odir] < 100000) | |
587 | return odir; | |
588 | ||
589 | /* | |
590 | * Both directions have pending sleep. Sleep the minimum time | |
591 | * and deduct from both. | |
592 | */ | |
593 | if (td->rate_pending_usleep[ddir] <= | |
594 | td->rate_pending_usleep[odir]) { | |
595 | usec = td->rate_pending_usleep[ddir]; | |
596 | } else { | |
597 | usec = td->rate_pending_usleep[odir]; | |
598 | ddir = odir; | |
599 | } | |
600 | } else | |
601 | usec = td->rate_pending_usleep[ddir]; | |
602 | ||
a9da8ab2 JA |
603 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
604 | io_u_quiesce(td); | |
78c1eda5 | 605 | |
9cc80b6d | 606 | usec = usec_sleep(td, usec); |
581e7141 JA |
607 | |
608 | td->rate_pending_usleep[ddir] -= usec; | |
609 | ||
610 | odir = ddir ^ 1; | |
611 | if (td_rw(td) && __should_check_rate(td, odir)) | |
612 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 613 | |
581e7141 JA |
614 | return ddir; |
615 | } | |
616 | ||
10ba535a JA |
617 | /* |
618 | * Return the data direction for the next io_u. If the job is a | |
619 | * mixed read/write workload, check the rwmix cycle and switch if | |
620 | * necessary. | |
621 | */ | |
1e97cce9 | 622 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 623 | { |
581e7141 JA |
624 | enum fio_ddir ddir; |
625 | ||
5f9099ea JA |
626 | /* |
627 | * see if it's time to fsync | |
628 | */ | |
629 | if (td->o.fsync_blocks && | |
630 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
631 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
632 | return DDIR_SYNC; | |
633 | ||
634 | /* | |
635 | * see if it's time to fdatasync | |
636 | */ | |
637 | if (td->o.fdatasync_blocks && | |
638 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
639 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
640 | return DDIR_DATASYNC; | |
641 | ||
44f29692 JA |
642 | /* |
643 | * see if it's time to sync_file_range | |
644 | */ | |
645 | if (td->sync_file_range_nr && | |
646 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
647 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
648 | return DDIR_SYNC_FILE_RANGE; | |
649 | ||
10ba535a | 650 | if (td_rw(td)) { |
10ba535a JA |
651 | /* |
652 | * Check if it's time to seed a new data direction. | |
653 | */ | |
e4928662 | 654 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
655 | /* |
656 | * Put a top limit on how many bytes we do for | |
657 | * one data direction, to avoid overflowing the | |
658 | * ranges too much | |
659 | */ | |
660 | ddir = get_rand_ddir(td); | |
e47f799f JA |
661 | |
662 | if (ddir != td->rwmix_ddir) | |
663 | set_rwmix_bytes(td); | |
664 | ||
665 | td->rwmix_ddir = ddir; | |
10ba535a | 666 | } |
581e7141 | 667 | ddir = td->rwmix_ddir; |
10ba535a | 668 | } else if (td_read(td)) |
581e7141 | 669 | ddir = DDIR_READ; |
6eaf09d6 | 670 | else if (td_write(td)) |
581e7141 | 671 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
672 | else |
673 | ddir = DDIR_TRIM; | |
581e7141 JA |
674 | |
675 | td->rwmix_ddir = rate_ddir(td, ddir); | |
676 | return td->rwmix_ddir; | |
10ba535a JA |
677 | } |
678 | ||
1ef2b6be JA |
679 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
680 | { | |
0e4dd95c DE |
681 | enum fio_ddir ddir = get_rw_ddir(td); |
682 | ||
82a90686 | 683 | if (td_trimwrite(td)) { |
0e4dd95c DE |
684 | struct fio_file *f = io_u->file; |
685 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
686 | ddir = DDIR_TRIM; | |
687 | else | |
688 | ddir = DDIR_WRITE; | |
689 | } | |
690 | ||
691 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be JA |
692 | |
693 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
694 | td->o.barrier_blocks && | |
695 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
696 | td->io_issues[DDIR_WRITE]) | |
a9da8ab2 | 697 | io_u_set(io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
698 | } |
699 | ||
e8462bd8 | 700 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 701 | { |
71b84caa | 702 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
703 | |
704 | if (ret) | |
705 | td_verror(td, ret, "file close"); | |
706 | } | |
707 | ||
10ba535a JA |
708 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
709 | { | |
a9da8ab2 JA |
710 | if (td->parent) |
711 | td = td->parent; | |
712 | ||
e8462bd8 JA |
713 | td_io_u_lock(td); |
714 | ||
f8b0bd10 | 715 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 716 | put_file_log(td, io_u->file); |
f8b0bd10 | 717 | |
10ba535a | 718 | io_u->file = NULL; |
a9da8ab2 | 719 | io_u_set(io_u, IO_U_F_FREE); |
d7ee2a7d | 720 | |
a9da8ab2 | 721 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 722 | td->cur_depth--; |
a9da8ab2 JA |
723 | assert(!(td->flags & TD_F_CHILD)); |
724 | } | |
2ae0b204 | 725 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
726 | td_io_u_unlock(td); |
727 | td_io_u_free_notify(td); | |
10ba535a JA |
728 | } |
729 | ||
f2bba182 RR |
730 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
731 | { | |
a9da8ab2 | 732 | io_u_clear(io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
733 | put_io_u(td, io_u); |
734 | } | |
735 | ||
755200a3 JA |
736 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
737 | { | |
738 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 739 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 740 | |
465221b0 JA |
741 | dprint(FD_IO, "requeue %p\n", __io_u); |
742 | ||
a9da8ab2 JA |
743 | if (td->parent) |
744 | td = td->parent; | |
745 | ||
e8462bd8 JA |
746 | td_io_u_lock(td); |
747 | ||
a9da8ab2 | 748 | io_u_set(__io_u, IO_U_F_FREE); |
bcd5abfa JA |
749 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
750 | td->io_issues[ddir]--; | |
5ec10eaa | 751 | |
a9da8ab2 JA |
752 | io_u_clear(__io_u, IO_U_F_FLIGHT); |
753 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { | |
0c41214f | 754 | td->cur_depth--; |
a9da8ab2 JA |
755 | assert(!(td->flags & TD_F_CHILD)); |
756 | } | |
2ae0b204 JA |
757 | |
758 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 759 | td_io_u_unlock(td); |
a9da8ab2 | 760 | td_io_u_free_notify(td); |
755200a3 JA |
761 | *io_u = NULL; |
762 | } | |
763 | ||
9bf2061e | 764 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 765 | { |
6aca9b3d JA |
766 | unsigned int is_random; |
767 | ||
b4c5e1ac JA |
768 | if (td->io_ops->flags & FIO_NOIO) |
769 | goto out; | |
770 | ||
1ef2b6be | 771 | set_rw_ddir(td, io_u); |
5f9099ea | 772 | |
87dc1ab1 | 773 | /* |
ff58fced | 774 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 775 | */ |
ff58fced | 776 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 777 | goto out; |
a00735e6 | 778 | |
48f5abd3 JA |
779 | /* |
780 | * See if it's time to switch to a new zone | |
781 | */ | |
13af05ae | 782 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
783 | struct fio_file *f = io_u->file; |
784 | ||
48f5abd3 | 785 | td->zone_bytes = 0; |
4c8be5b1 JA |
786 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
787 | ||
788 | /* | |
789 | * Wrap from the beginning, if we exceed the file size | |
790 | */ | |
791 | if (f->file_offset >= f->real_file_size) | |
792 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 793 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
794 | td->io_skip_bytes += td->o.zone_skip; |
795 | } | |
796 | ||
10ba535a | 797 | /* |
c685b5b2 JA |
798 | * No log, let the seq/rand engine retrieve the next buflen and |
799 | * position. | |
10ba535a | 800 | */ |
6aca9b3d | 801 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 802 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 803 | return 1; |
2ba1c290 | 804 | } |
10ba535a | 805 | |
6aca9b3d | 806 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
807 | if (!io_u->buflen) { |
808 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 809 | return 1; |
2ba1c290 | 810 | } |
bca4ed4d | 811 | |
2ba1c290 JA |
812 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
813 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
814 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
815 | (unsigned long long) io_u->offset, io_u->buflen, | |
816 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 817 | return 1; |
2ba1c290 | 818 | } |
6a5e6884 | 819 | |
bca4ed4d JA |
820 | /* |
821 | * mark entry before potentially trimming io_u | |
822 | */ | |
303032ae | 823 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 824 | mark_random_map(td, io_u); |
bca4ed4d | 825 | |
c38e9468 | 826 | out: |
2ba1c290 | 827 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 828 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 829 | return 0; |
10ba535a JA |
830 | } |
831 | ||
838bc709 JA |
832 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
833 | { | |
2b13e716 | 834 | int idx = 0; |
838bc709 JA |
835 | |
836 | switch (nr) { | |
837 | default: | |
2b13e716 | 838 | idx = 6; |
838bc709 JA |
839 | break; |
840 | case 33 ... 64: | |
2b13e716 | 841 | idx = 5; |
838bc709 JA |
842 | break; |
843 | case 17 ... 32: | |
2b13e716 | 844 | idx = 4; |
838bc709 JA |
845 | break; |
846 | case 9 ... 16: | |
2b13e716 | 847 | idx = 3; |
838bc709 JA |
848 | break; |
849 | case 5 ... 8: | |
2b13e716 | 850 | idx = 2; |
838bc709 JA |
851 | break; |
852 | case 1 ... 4: | |
2b13e716 | 853 | idx = 1; |
838bc709 JA |
854 | case 0: |
855 | break; | |
856 | } | |
857 | ||
2b13e716 | 858 | map[idx]++; |
838bc709 JA |
859 | } |
860 | ||
861 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
862 | { | |
863 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
864 | td->ts.total_submit++; | |
865 | } | |
866 | ||
867 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
868 | { | |
869 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
870 | td->ts.total_complete++; | |
871 | } | |
872 | ||
d8005759 | 873 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 874 | { |
2b13e716 | 875 | int idx = 0; |
71619dc2 JA |
876 | |
877 | switch (td->cur_depth) { | |
878 | default: | |
2b13e716 | 879 | idx = 6; |
a783e61a | 880 | break; |
71619dc2 | 881 | case 32 ... 63: |
2b13e716 | 882 | idx = 5; |
a783e61a | 883 | break; |
71619dc2 | 884 | case 16 ... 31: |
2b13e716 | 885 | idx = 4; |
a783e61a | 886 | break; |
71619dc2 | 887 | case 8 ... 15: |
2b13e716 | 888 | idx = 3; |
a783e61a | 889 | break; |
71619dc2 | 890 | case 4 ... 7: |
2b13e716 | 891 | idx = 2; |
a783e61a | 892 | break; |
71619dc2 | 893 | case 2 ... 3: |
2b13e716 | 894 | idx = 1; |
71619dc2 JA |
895 | case 1: |
896 | break; | |
897 | } | |
898 | ||
2b13e716 | 899 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
900 | } |
901 | ||
04a0feae JA |
902 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
903 | { | |
2b13e716 | 904 | int idx = 0; |
04a0feae JA |
905 | |
906 | assert(usec < 1000); | |
907 | ||
908 | switch (usec) { | |
909 | case 750 ... 999: | |
2b13e716 | 910 | idx = 9; |
04a0feae JA |
911 | break; |
912 | case 500 ... 749: | |
2b13e716 | 913 | idx = 8; |
04a0feae JA |
914 | break; |
915 | case 250 ... 499: | |
2b13e716 | 916 | idx = 7; |
04a0feae JA |
917 | break; |
918 | case 100 ... 249: | |
2b13e716 | 919 | idx = 6; |
04a0feae JA |
920 | break; |
921 | case 50 ... 99: | |
2b13e716 | 922 | idx = 5; |
04a0feae JA |
923 | break; |
924 | case 20 ... 49: | |
2b13e716 | 925 | idx = 4; |
04a0feae JA |
926 | break; |
927 | case 10 ... 19: | |
2b13e716 | 928 | idx = 3; |
04a0feae JA |
929 | break; |
930 | case 4 ... 9: | |
2b13e716 | 931 | idx = 2; |
04a0feae JA |
932 | break; |
933 | case 2 ... 3: | |
2b13e716 | 934 | idx = 1; |
04a0feae JA |
935 | case 0 ... 1: |
936 | break; | |
937 | } | |
938 | ||
2b13e716 JA |
939 | assert(idx < FIO_IO_U_LAT_U_NR); |
940 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
941 | } |
942 | ||
943 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 944 | { |
2b13e716 | 945 | int idx = 0; |
ec118304 JA |
946 | |
947 | switch (msec) { | |
948 | default: | |
2b13e716 | 949 | idx = 11; |
04a0feae | 950 | break; |
8abdce66 | 951 | case 1000 ... 1999: |
2b13e716 | 952 | idx = 10; |
04a0feae | 953 | break; |
8abdce66 | 954 | case 750 ... 999: |
2b13e716 | 955 | idx = 9; |
04a0feae | 956 | break; |
8abdce66 | 957 | case 500 ... 749: |
2b13e716 | 958 | idx = 8; |
04a0feae | 959 | break; |
8abdce66 | 960 | case 250 ... 499: |
2b13e716 | 961 | idx = 7; |
04a0feae | 962 | break; |
8abdce66 | 963 | case 100 ... 249: |
2b13e716 | 964 | idx = 6; |
04a0feae | 965 | break; |
8abdce66 | 966 | case 50 ... 99: |
2b13e716 | 967 | idx = 5; |
04a0feae | 968 | break; |
8abdce66 | 969 | case 20 ... 49: |
2b13e716 | 970 | idx = 4; |
04a0feae | 971 | break; |
8abdce66 | 972 | case 10 ... 19: |
2b13e716 | 973 | idx = 3; |
04a0feae | 974 | break; |
8abdce66 | 975 | case 4 ... 9: |
2b13e716 | 976 | idx = 2; |
04a0feae | 977 | break; |
ec118304 | 978 | case 2 ... 3: |
2b13e716 | 979 | idx = 1; |
ec118304 JA |
980 | case 0 ... 1: |
981 | break; | |
982 | } | |
983 | ||
2b13e716 JA |
984 | assert(idx < FIO_IO_U_LAT_M_NR); |
985 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
986 | } |
987 | ||
988 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
989 | { | |
990 | if (usec < 1000) | |
991 | io_u_mark_lat_usec(td, usec); | |
992 | else | |
993 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
994 | } |
995 | ||
0aabe160 JA |
996 | /* |
997 | * Get next file to service by choosing one at random | |
998 | */ | |
2cc52930 JA |
999 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1000 | enum fio_file_flags goodf, | |
d6aed795 | 1001 | enum fio_file_flags badf) |
0aabe160 | 1002 | { |
c3546b53 | 1003 | uint64_t frand_max = rand_max(&td->next_file_state); |
0aabe160 | 1004 | struct fio_file *f; |
1c178180 | 1005 | int fno; |
0aabe160 JA |
1006 | |
1007 | do { | |
87b10676 | 1008 | int opened = 0; |
1294c3ec | 1009 | unsigned long r; |
4c07ad86 | 1010 | |
d6b72507 | 1011 | r = __rand(&td->next_file_state); |
54a21917 | 1012 | fno = (unsigned int) ((double) td->o.nr_files |
c3546b53 | 1013 | * (r / (frand_max + 1.0))); |
7c83c089 | 1014 | |
126d65c6 | 1015 | f = td->files[fno]; |
d6aed795 | 1016 | if (fio_file_done(f)) |
059e63c0 | 1017 | continue; |
1c178180 | 1018 | |
d6aed795 | 1019 | if (!fio_file_open(f)) { |
87b10676 JA |
1020 | int err; |
1021 | ||
002fe734 JA |
1022 | if (td->nr_open_files >= td->o.open_files) |
1023 | return ERR_PTR(-EBUSY); | |
1024 | ||
87b10676 JA |
1025 | err = td_io_open_file(td, f); |
1026 | if (err) | |
1027 | continue; | |
1028 | opened = 1; | |
1029 | } | |
1030 | ||
2ba1c290 JA |
1031 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1032 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1033 | return f; |
2ba1c290 | 1034 | } |
87b10676 JA |
1035 | if (opened) |
1036 | td_io_close_file(td, f); | |
0aabe160 JA |
1037 | } while (1); |
1038 | } | |
1039 | ||
1040 | /* | |
1041 | * Get next file to service by doing round robin between all available ones | |
1042 | */ | |
1c178180 JA |
1043 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1044 | int badf) | |
3d7c391d JA |
1045 | { |
1046 | unsigned int old_next_file = td->next_file; | |
1047 | struct fio_file *f; | |
1048 | ||
1049 | do { | |
87b10676 JA |
1050 | int opened = 0; |
1051 | ||
126d65c6 | 1052 | f = td->files[td->next_file]; |
3d7c391d JA |
1053 | |
1054 | td->next_file++; | |
2dc1bbeb | 1055 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1056 | td->next_file = 0; |
1057 | ||
87b10676 | 1058 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1059 | if (fio_file_done(f)) { |
d5ed68ea | 1060 | f = NULL; |
059e63c0 | 1061 | continue; |
d5ed68ea | 1062 | } |
059e63c0 | 1063 | |
d6aed795 | 1064 | if (!fio_file_open(f)) { |
87b10676 JA |
1065 | int err; |
1066 | ||
002fe734 JA |
1067 | if (td->nr_open_files >= td->o.open_files) |
1068 | return ERR_PTR(-EBUSY); | |
1069 | ||
87b10676 | 1070 | err = td_io_open_file(td, f); |
b5696bfc JA |
1071 | if (err) { |
1072 | dprint(FD_FILE, "error %d on open of %s\n", | |
1073 | err, f->file_name); | |
87c27b45 | 1074 | f = NULL; |
87b10676 | 1075 | continue; |
b5696bfc | 1076 | } |
87b10676 JA |
1077 | opened = 1; |
1078 | } | |
1079 | ||
0b9d69ec JA |
1080 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1081 | f->flags); | |
1c178180 | 1082 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1083 | break; |
1084 | ||
87b10676 JA |
1085 | if (opened) |
1086 | td_io_close_file(td, f); | |
1087 | ||
3d7c391d JA |
1088 | f = NULL; |
1089 | } while (td->next_file != old_next_file); | |
1090 | ||
2ba1c290 | 1091 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1092 | return f; |
1093 | } | |
1094 | ||
7eb36574 | 1095 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1096 | { |
1907dbc6 JA |
1097 | struct fio_file *f; |
1098 | ||
2dc1bbeb | 1099 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1100 | |
b5696bfc | 1101 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1102 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1103 | " nr_files=%d\n", td->nr_open_files, | |
1104 | td->nr_done_files, | |
1105 | td->o.nr_files); | |
bdb4e2e9 | 1106 | return NULL; |
2ba1c290 | 1107 | } |
bdb4e2e9 | 1108 | |
1907dbc6 | 1109 | f = td->file_service_file; |
d6aed795 | 1110 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1111 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1112 | goto out; | |
1113 | if (td->file_service_left--) | |
1114 | goto out; | |
1115 | } | |
1907dbc6 | 1116 | |
a086c257 JA |
1117 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1118 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1119 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1120 | else |
d6aed795 | 1121 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1122 | |
002fe734 JA |
1123 | if (IS_ERR(f)) |
1124 | return f; | |
1125 | ||
1907dbc6 JA |
1126 | td->file_service_file = f; |
1127 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1128 | out: |
0dac421f JA |
1129 | if (f) |
1130 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1131 | else | |
1132 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1133 | return f; |
bdb4e2e9 JA |
1134 | } |
1135 | ||
7eb36574 JA |
1136 | static struct fio_file *get_next_file(struct thread_data *td) |
1137 | { | |
372d8962 | 1138 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1139 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1140 | |
d72be545 JA |
1141 | if (ops->get_next_file) |
1142 | return ops->get_next_file(td); | |
1143 | } | |
7eb36574 JA |
1144 | |
1145 | return __get_next_file(td); | |
1146 | } | |
1147 | ||
002fe734 | 1148 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1149 | { |
1150 | struct fio_file *f; | |
1151 | ||
1152 | do { | |
1153 | f = get_next_file(td); | |
002fe734 JA |
1154 | if (IS_ERR_OR_NULL(f)) |
1155 | return PTR_ERR(f); | |
429f6675 | 1156 | |
429f6675 JA |
1157 | io_u->file = f; |
1158 | get_file(f); | |
1159 | ||
1160 | if (!fill_io_u(td, io_u)) | |
1161 | break; | |
1162 | ||
b5696bfc | 1163 | put_file_log(td, f); |
429f6675 | 1164 | td_io_close_file(td, f); |
b5696bfc | 1165 | io_u->file = NULL; |
d6aed795 | 1166 | fio_file_set_done(f); |
429f6675 | 1167 | td->nr_done_files++; |
0b9d69ec JA |
1168 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1169 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1170 | } while (1); |
1171 | ||
1172 | return 0; | |
1173 | } | |
1174 | ||
3e260a46 JA |
1175 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1176 | unsigned long tusec, unsigned long max_usec) | |
1177 | { | |
1178 | if (!td->error) | |
1179 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1180 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1181 | icd->error = ETIMEDOUT; | |
1182 | } | |
1183 | ||
1184 | static void lat_new_cycle(struct thread_data *td) | |
1185 | { | |
1186 | fio_gettime(&td->latency_ts, NULL); | |
1187 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1188 | td->latency_failed = 0; | |
1189 | } | |
1190 | ||
1191 | /* | |
1192 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1193 | * are at QD=1, then it's time to give up. | |
1194 | */ | |
1195 | static int __lat_target_failed(struct thread_data *td) | |
1196 | { | |
1197 | if (td->latency_qd == 1) | |
1198 | return 1; | |
1199 | ||
1200 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1201 | |
1202 | if (td->latency_qd == td->latency_qd_low) | |
1203 | td->latency_qd_low--; | |
1204 | ||
3e260a46 JA |
1205 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1206 | ||
1207 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1208 | ||
1209 | /* | |
1210 | * When we ramp QD down, quiesce existing IO to prevent | |
1211 | * a storm of ramp downs due to pending higher depth. | |
1212 | */ | |
1213 | io_u_quiesce(td); | |
1214 | lat_new_cycle(td); | |
1215 | return 0; | |
1216 | } | |
1217 | ||
1218 | static int lat_target_failed(struct thread_data *td) | |
1219 | { | |
1220 | if (td->o.latency_percentile.u.f == 100.0) | |
1221 | return __lat_target_failed(td); | |
1222 | ||
1223 | td->latency_failed++; | |
1224 | return 0; | |
1225 | } | |
1226 | ||
1227 | void lat_target_init(struct thread_data *td) | |
1228 | { | |
6bb58215 JA |
1229 | td->latency_end_run = 0; |
1230 | ||
3e260a46 JA |
1231 | if (td->o.latency_target) { |
1232 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1233 | fio_gettime(&td->latency_ts, NULL); | |
1234 | td->latency_qd = 1; | |
1235 | td->latency_qd_high = td->o.iodepth; | |
1236 | td->latency_qd_low = 1; | |
1237 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1238 | } else | |
1239 | td->latency_qd = td->o.iodepth; | |
1240 | } | |
1241 | ||
6bb58215 JA |
1242 | void lat_target_reset(struct thread_data *td) |
1243 | { | |
1244 | if (!td->latency_end_run) | |
1245 | lat_target_init(td); | |
1246 | } | |
1247 | ||
3e260a46 JA |
1248 | static void lat_target_success(struct thread_data *td) |
1249 | { | |
1250 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1251 | struct thread_options *o = &td->o; |
3e260a46 JA |
1252 | |
1253 | td->latency_qd_low = td->latency_qd; | |
1254 | ||
1255 | /* | |
1256 | * If we haven't failed yet, we double up to a failing value instead | |
1257 | * of bisecting from highest possible queue depth. If we have set | |
1258 | * a limit other than td->o.iodepth, bisect between that. | |
1259 | */ | |
6bb58215 | 1260 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1261 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1262 | else | |
1263 | td->latency_qd *= 2; | |
1264 | ||
6bb58215 JA |
1265 | if (td->latency_qd > o->iodepth) |
1266 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1267 | |
1268 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1269 | |
3e260a46 | 1270 | /* |
6bb58215 JA |
1271 | * Same as last one, we are done. Let it run a latency cycle, so |
1272 | * we get only the results from the targeted depth. | |
3e260a46 | 1273 | */ |
6bb58215 JA |
1274 | if (td->latency_qd == qd) { |
1275 | if (td->latency_end_run) { | |
1276 | dprint(FD_RATE, "We are done\n"); | |
1277 | td->done = 1; | |
1278 | } else { | |
1279 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1280 | io_u_quiesce(td); | |
1281 | td->latency_end_run = 1; | |
1282 | reset_all_stats(td); | |
1283 | reset_io_stats(td); | |
1284 | } | |
1285 | } | |
3e260a46 JA |
1286 | |
1287 | lat_new_cycle(td); | |
1288 | } | |
1289 | ||
1290 | /* | |
1291 | * Check if we can bump the queue depth | |
1292 | */ | |
1293 | void lat_target_check(struct thread_data *td) | |
1294 | { | |
1295 | uint64_t usec_window; | |
1296 | uint64_t ios; | |
1297 | double success_ios; | |
1298 | ||
1299 | usec_window = utime_since_now(&td->latency_ts); | |
1300 | if (usec_window < td->o.latency_window) | |
1301 | return; | |
1302 | ||
1303 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1304 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1305 | success_ios *= 100.0; | |
1306 | ||
1307 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1308 | ||
1309 | if (success_ios >= td->o.latency_percentile.u.f) | |
1310 | lat_target_success(td); | |
1311 | else | |
1312 | __lat_target_failed(td); | |
1313 | } | |
1314 | ||
1315 | /* | |
1316 | * If latency target is enabled, we might be ramping up or down and not | |
1317 | * using the full queue depth available. | |
1318 | */ | |
effd6ff0 | 1319 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1320 | { |
1321 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1322 | ||
1323 | if (qempty) | |
1324 | return 1; | |
1325 | if (!td->o.latency_target) | |
1326 | return 0; | |
1327 | ||
1328 | return td->cur_depth >= td->latency_qd; | |
1329 | } | |
429f6675 | 1330 | |
10ba535a JA |
1331 | struct io_u *__get_io_u(struct thread_data *td) |
1332 | { | |
0cae66f6 | 1333 | struct io_u *io_u = NULL; |
10ba535a | 1334 | |
ca09be4b JA |
1335 | if (td->stop_io) |
1336 | return NULL; | |
1337 | ||
e8462bd8 JA |
1338 | td_io_u_lock(td); |
1339 | ||
1340 | again: | |
2ae0b204 JA |
1341 | if (!io_u_rempty(&td->io_u_requeues)) |
1342 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1343 | else if (!queue_full(td)) { |
2ae0b204 | 1344 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1345 | |
225ba9e3 | 1346 | io_u->file = NULL; |
6040dabc | 1347 | io_u->buflen = 0; |
10ba535a | 1348 | io_u->resid = 0; |
d7762cf8 | 1349 | io_u->end_io = NULL; |
755200a3 JA |
1350 | } |
1351 | ||
1352 | if (io_u) { | |
0c6e7517 | 1353 | assert(io_u->flags & IO_U_F_FREE); |
a9da8ab2 | 1354 | io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1355 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1356 | IO_U_F_VER_LIST); | |
0c6e7517 | 1357 | |
755200a3 | 1358 | io_u->error = 0; |
bcd5abfa | 1359 | io_u->acct_ddir = -1; |
10ba535a | 1360 | td->cur_depth++; |
a9da8ab2 JA |
1361 | assert(!(td->flags & TD_F_CHILD)); |
1362 | io_u_set(io_u, IO_U_F_IN_CUR_DEPTH); | |
f9401285 | 1363 | io_u->ipo = NULL; |
a9da8ab2 | 1364 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1365 | /* |
1366 | * We ran out, wait for async verify threads to finish and | |
1367 | * return one | |
1368 | */ | |
a9da8ab2 JA |
1369 | assert(!(td->flags & TD_F_CHILD)); |
1370 | assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock)); | |
1dec3e07 | 1371 | goto again; |
10ba535a JA |
1372 | } |
1373 | ||
e8462bd8 | 1374 | td_io_u_unlock(td); |
10ba535a JA |
1375 | return io_u; |
1376 | } | |
1377 | ||
0d29de83 | 1378 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1379 | { |
d72be545 JA |
1380 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1381 | return 0; | |
1382 | ||
1383 | if (td->trim_entries) { | |
0d29de83 | 1384 | int get_trim = 0; |
10ba535a | 1385 | |
0d29de83 JA |
1386 | if (td->trim_batch) { |
1387 | td->trim_batch--; | |
1388 | get_trim = 1; | |
1389 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1390 | td->last_ddir != DDIR_READ) { | |
1391 | td->trim_batch = td->o.trim_batch; | |
1392 | if (!td->trim_batch) | |
1393 | td->trim_batch = td->o.trim_backlog; | |
1394 | get_trim = 1; | |
1395 | } | |
1396 | ||
1397 | if (get_trim && !get_next_trim(td, io_u)) | |
1398 | return 1; | |
2ba1c290 | 1399 | } |
10ba535a | 1400 | |
0d29de83 JA |
1401 | return 0; |
1402 | } | |
1403 | ||
1404 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1405 | { | |
d72be545 JA |
1406 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1407 | return 0; | |
1408 | ||
1409 | if (td->io_hist_len) { | |
9e144189 JA |
1410 | int get_verify = 0; |
1411 | ||
d1ece0c7 | 1412 | if (td->verify_batch) |
9e144189 | 1413 | get_verify = 1; |
d1ece0c7 | 1414 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1415 | td->last_ddir != DDIR_READ) { |
1416 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1417 | if (!td->verify_batch) |
1418 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1419 | get_verify = 1; |
1420 | } | |
1421 | ||
d1ece0c7 JA |
1422 | if (get_verify && !get_next_verify(td, io_u)) { |
1423 | td->verify_batch--; | |
0d29de83 | 1424 | return 1; |
d1ece0c7 | 1425 | } |
9e144189 JA |
1426 | } |
1427 | ||
0d29de83 JA |
1428 | return 0; |
1429 | } | |
1430 | ||
de789769 JA |
1431 | /* |
1432 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1433 | * easy compressible data for simple de-dupe attempts. Do this for every |
1434 | * 512b block in the range, since that should be the smallest block size | |
1435 | * we can expect from a device. | |
de789769 JA |
1436 | */ |
1437 | static void small_content_scramble(struct io_u *io_u) | |
1438 | { | |
23f394d5 | 1439 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1440 | uint64_t boffset; |
23f394d5 JA |
1441 | unsigned int offset; |
1442 | void *p, *end; | |
de789769 | 1443 | |
23f394d5 JA |
1444 | if (!nr_blocks) |
1445 | return; | |
1446 | ||
1447 | p = io_u->xfer_buf; | |
fba76ee8 | 1448 | boffset = io_u->offset; |
81f0366c | 1449 | io_u->buf_filled_len = 0; |
fad82f76 | 1450 | |
23f394d5 JA |
1451 | for (i = 0; i < nr_blocks; i++) { |
1452 | /* | |
1453 | * Fill the byte offset into a "random" start offset of | |
1454 | * the buffer, given by the product of the usec time | |
1455 | * and the actual offset. | |
1456 | */ | |
fad82f76 | 1457 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1458 | offset &= ~(sizeof(uint64_t) - 1); |
1459 | if (offset >= 512 - sizeof(uint64_t)) | |
1460 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1461 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1462 | |
1463 | end = p + 512 - sizeof(io_u->start_time); | |
1464 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1465 | p += 512; | |
fad82f76 | 1466 | boffset += 512; |
23f394d5 | 1467 | } |
de789769 JA |
1468 | } |
1469 | ||
0d29de83 JA |
1470 | /* |
1471 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1472 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1473 | */ | |
1474 | struct io_u *get_io_u(struct thread_data *td) | |
1475 | { | |
1476 | struct fio_file *f; | |
1477 | struct io_u *io_u; | |
de789769 | 1478 | int do_scramble = 0; |
002fe734 | 1479 | long ret = 0; |
0d29de83 JA |
1480 | |
1481 | io_u = __get_io_u(td); | |
1482 | if (!io_u) { | |
1483 | dprint(FD_IO, "__get_io_u failed\n"); | |
1484 | return NULL; | |
1485 | } | |
1486 | ||
1487 | if (check_get_verify(td, io_u)) | |
1488 | goto out; | |
1489 | if (check_get_trim(td, io_u)) | |
1490 | goto out; | |
1491 | ||
755200a3 JA |
1492 | /* |
1493 | * from a requeue, io_u already setup | |
1494 | */ | |
1495 | if (io_u->file) | |
77f392bf | 1496 | goto out; |
755200a3 | 1497 | |
429f6675 JA |
1498 | /* |
1499 | * If using an iolog, grab next piece if any available. | |
1500 | */ | |
d72be545 | 1501 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1502 | if (read_iolog_get(td, io_u)) |
1503 | goto err_put; | |
2ba1c290 | 1504 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1505 | ret = -EBUSY; |
2ba1c290 | 1506 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1507 | goto err_put; |
2ba1c290 | 1508 | } |
5ec10eaa | 1509 | |
429f6675 | 1510 | f = io_u->file; |
002fe734 JA |
1511 | if (!f) { |
1512 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1513 | goto err_put; | |
1514 | } | |
1515 | ||
d6aed795 | 1516 | assert(fio_file_open(f)); |
97af62ce | 1517 | |
ff58fced | 1518 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1519 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1520 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1521 | goto err_put; |
2ba1c290 | 1522 | } |
10ba535a | 1523 | |
f1dfb668 JA |
1524 | f->last_start[io_u->ddir] = io_u->offset; |
1525 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1526 | |
fd68418e | 1527 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1528 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1529 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1530 | td->o.min_bs[DDIR_WRITE], |
1531 | io_u->xfer_buflen); | |
ff441ae8 | 1532 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1533 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1534 | do_scramble = 1; |
d72be545 | 1535 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1536 | populate_verify_io_u(td, io_u); |
1537 | do_scramble = 0; | |
1538 | } | |
fd68418e | 1539 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1540 | /* |
1541 | * Reset the buf_filled parameters so next time if the | |
1542 | * buffer is used for writes it is refilled. | |
1543 | */ | |
cbe8d756 RR |
1544 | io_u->buf_filled_len = 0; |
1545 | } | |
87dc1ab1 | 1546 | } |
10ba535a | 1547 | |
165faf16 JA |
1548 | /* |
1549 | * Set io data pointers. | |
1550 | */ | |
cec6b55d JA |
1551 | io_u->xfer_buf = io_u->buf; |
1552 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1553 | |
6ac7a331 | 1554 | out: |
0d29de83 | 1555 | assert(io_u->file); |
429f6675 | 1556 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1557 | if (!td->o.disable_slat) |
1558 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1559 | if (do_scramble) |
1560 | small_content_scramble(io_u); | |
429f6675 | 1561 | return io_u; |
36167d82 | 1562 | } |
429f6675 | 1563 | err_put: |
2ba1c290 | 1564 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1565 | put_io_u(td, io_u); |
002fe734 | 1566 | return ERR_PTR(ret); |
10ba535a JA |
1567 | } |
1568 | ||
a9da8ab2 | 1569 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1570 | { |
8b28bd41 | 1571 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1572 | |
8b28bd41 DM |
1573 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1574 | return; | |
5451792e | 1575 | |
709c8313 RE |
1576 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1577 | io_u->file ? " on file " : "", | |
1578 | io_u->file ? io_u->file->file_name : "", | |
1579 | strerror(io_u->error), | |
1580 | io_ddir_name(io_u->ddir), | |
1581 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1582 | |
1583 | if (!td->error) | |
1584 | td_verror(td, io_u->error, "io_u error"); | |
1585 | } | |
1586 | ||
a9da8ab2 JA |
1587 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1588 | { | |
1589 | __io_u_log_error(td, io_u); | |
1590 | if (td->parent) | |
1591 | __io_u_log_error(td, io_u); | |
1592 | } | |
1593 | ||
aba6c951 JA |
1594 | static inline int gtod_reduce(struct thread_data *td) |
1595 | { | |
729fe3af | 1596 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1597 | && td->o.disable_bw; |
aba6c951 JA |
1598 | } |
1599 | ||
c8eeb9df JA |
1600 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1601 | struct io_completion_data *icd, | |
1602 | const enum fio_ddir idx, unsigned int bytes) | |
1603 | { | |
a9da8ab2 | 1604 | const int no_reduce = !gtod_reduce(td); |
24d23ca7 | 1605 | unsigned long lusec = 0; |
c8eeb9df | 1606 | |
a9da8ab2 | 1607 | if (no_reduce) |
c8eeb9df JA |
1608 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1609 | ||
1610 | if (!td->o.disable_lat) { | |
1611 | unsigned long tusec; | |
1612 | ||
1613 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1614 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1615 | |
d4afedfd JA |
1616 | if (td->flags & TD_F_PROFILE_OPS) { |
1617 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1618 | ||
1619 | if (ops->io_u_lat) | |
1620 | icd->error = ops->io_u_lat(td, tusec); | |
1621 | } | |
1622 | ||
3e260a46 JA |
1623 | if (td->o.max_latency && tusec > td->o.max_latency) |
1624 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1625 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1626 | if (lat_target_failed(td)) | |
1627 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1628 | } |
c8eeb9df JA |
1629 | } |
1630 | ||
1631 | if (!td->o.disable_clat) { | |
ae588852 | 1632 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1633 | io_u_mark_latency(td, lusec); |
1634 | } | |
1635 | ||
a9da8ab2 JA |
1636 | if (td->parent) |
1637 | td = td->parent; | |
1638 | ||
c8eeb9df JA |
1639 | if (!td->o.disable_bw) |
1640 | add_bw_sample(td, idx, bytes, &icd->time); | |
1641 | ||
a9da8ab2 | 1642 | if (no_reduce) |
aba6c951 | 1643 | add_iops_sample(td, idx, bytes, &icd->time); |
66347cfa DE |
1644 | |
1645 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1646 | uint32_t *info = io_u_block_info(td, io_u); | |
1647 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1648 | if (io_u->ddir == DDIR_TRIM) { | |
1649 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1650 | BLOCK_INFO_TRIMS(*info) + 1); | |
1651 | } else if (io_u->ddir == DDIR_WRITE) { | |
1652 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1653 | *info); | |
1654 | } | |
1655 | } | |
1656 | } | |
c8eeb9df JA |
1657 | } |
1658 | ||
1b8dbf25 SL |
1659 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1660 | { | |
1ae83d45 JA |
1661 | uint64_t secs, remainder, bps, bytes; |
1662 | ||
a9da8ab2 | 1663 | assert(!(td->flags & TD_F_CHILD)); |
1b8dbf25 SL |
1664 | bytes = td->this_io_bytes[ddir]; |
1665 | bps = td->rate_bps[ddir]; | |
1666 | secs = bytes / bps; | |
1667 | remainder = bytes % bps; | |
1668 | return remainder * 1000000 / bps + secs * 1000000; | |
1669 | } | |
1670 | ||
f8b0bd10 | 1671 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1672 | struct io_completion_data *icd) |
10ba535a | 1673 | { |
f8b0bd10 JA |
1674 | struct io_u *io_u = *io_u_ptr; |
1675 | enum fio_ddir ddir = io_u->ddir; | |
1676 | struct fio_file *f = io_u->file; | |
10ba535a | 1677 | |
2ba1c290 JA |
1678 | dprint_io_u(io_u, "io complete"); |
1679 | ||
0c6e7517 | 1680 | assert(io_u->flags & IO_U_F_FLIGHT); |
a9da8ab2 | 1681 | io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1682 | |
1683 | /* | |
1684 | * Mark IO ok to verify | |
1685 | */ | |
1686 | if (io_u->ipo) { | |
890b6656 JA |
1687 | /* |
1688 | * Remove errored entry from the verification list | |
1689 | */ | |
1690 | if (io_u->error) | |
1691 | unlog_io_piece(td, io_u); | |
1692 | else { | |
1693 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1694 | write_barrier(); | |
1695 | } | |
f9401285 JA |
1696 | } |
1697 | ||
f8b0bd10 | 1698 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1699 | td->last_was_sync = 1; |
44f29692 JA |
1700 | if (f) { |
1701 | f->first_write = -1ULL; | |
1702 | f->last_write = -1ULL; | |
1703 | } | |
87dc1ab1 JA |
1704 | return; |
1705 | } | |
1706 | ||
1707 | td->last_was_sync = 0; | |
f8b0bd10 | 1708 | td->last_ddir = ddir; |
87dc1ab1 | 1709 | |
f8b0bd10 | 1710 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1711 | unsigned int bytes = io_u->buflen - io_u->resid; |
f8b0bd10 | 1712 | const enum fio_ddir oddir = ddir ^ 1; |
b29ee5b3 | 1713 | int ret; |
10ba535a | 1714 | |
f8b0bd10 JA |
1715 | td->io_blocks[ddir]++; |
1716 | td->this_io_blocks[ddir]++; | |
1717 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1718 | |
1719 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1720 | td->this_io_bytes[ddir] += bytes; |
1721 | ||
ca09be4b JA |
1722 | if (ddir == DDIR_WRITE) { |
1723 | if (f) { | |
1724 | if (f->first_write == -1ULL || | |
1725 | io_u->offset < f->first_write) | |
1726 | f->first_write = io_u->offset; | |
1727 | if (f->last_write == -1ULL || | |
1728 | ((io_u->offset + bytes) > f->last_write)) | |
1729 | f->last_write = io_u->offset + bytes; | |
1730 | } | |
1731 | if (td->last_write_comp) { | |
1732 | int idx = td->last_write_idx++; | |
1733 | ||
1734 | td->last_write_comp[idx] = io_u->offset; | |
1735 | if (td->last_write_idx == td->o.iodepth) | |
1736 | td->last_write_idx = 0; | |
1737 | } | |
44f29692 JA |
1738 | } |
1739 | ||
6b1190fd SL |
1740 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
1741 | td->runstate == TD_VERIFYING)) { | |
a9da8ab2 JA |
1742 | struct thread_data *__td = td; |
1743 | ||
f8b0bd10 | 1744 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1745 | |
a9da8ab2 JA |
1746 | if (td->parent) |
1747 | __td = td->parent; | |
1748 | ||
1749 | if (__should_check_rate(__td, ddir)) { | |
1750 | __td->rate_pending_usleep[ddir] = | |
1751 | (usec_for_io(__td, ddir) - | |
1752 | utime_since_now(&__td->start)); | |
b23b6a2f | 1753 | } |
f8b0bd10 | 1754 | if (ddir != DDIR_TRIM && |
a9da8ab2 JA |
1755 | __should_check_rate(__td, oddir)) { |
1756 | __td->rate_pending_usleep[oddir] = | |
1757 | (usec_for_io(__td, oddir) - | |
1758 | utime_since_now(&__td->start)); | |
f8b0bd10 | 1759 | } |
721938ae | 1760 | } |
10ba535a | 1761 | |
f8b0bd10 | 1762 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1763 | |
d7762cf8 | 1764 | if (io_u->end_io) { |
f8b0bd10 JA |
1765 | ret = io_u->end_io(td, io_u_ptr); |
1766 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1767 | if (ret && !icd->error) |
1768 | icd->error = ret; | |
1769 | } | |
ff58fced | 1770 | } else if (io_u->error) { |
10ba535a | 1771 | icd->error = io_u->error; |
5451792e JA |
1772 | io_u_log_error(td, io_u); |
1773 | } | |
8b28bd41 | 1774 | if (icd->error) { |
f8b0bd10 JA |
1775 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1776 | ||
8b28bd41 DM |
1777 | if (!td_non_fatal_error(td, eb, icd->error)) |
1778 | return; | |
f8b0bd10 | 1779 | |
f2bba182 RR |
1780 | /* |
1781 | * If there is a non_fatal error, then add to the error count | |
1782 | * and clear all the errors. | |
1783 | */ | |
1784 | update_error_count(td, icd->error); | |
1785 | td_clear_error(td); | |
1786 | icd->error = 0; | |
f8b0bd10 JA |
1787 | if (io_u) |
1788 | io_u->error = 0; | |
f2bba182 | 1789 | } |
10ba535a JA |
1790 | } |
1791 | ||
9520ebb9 JA |
1792 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1793 | int nr) | |
10ba535a | 1794 | { |
6eaf09d6 | 1795 | int ddir; |
aba6c951 JA |
1796 | |
1797 | if (!gtod_reduce(td)) | |
9520ebb9 | 1798 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1799 | |
3af6ef39 JA |
1800 | icd->nr = nr; |
1801 | ||
10ba535a | 1802 | icd->error = 0; |
6eaf09d6 SL |
1803 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1804 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1805 | } |
1806 | ||
97601024 JA |
1807 | static void ios_completed(struct thread_data *td, |
1808 | struct io_completion_data *icd) | |
36167d82 JA |
1809 | { |
1810 | struct io_u *io_u; | |
1811 | int i; | |
1812 | ||
10ba535a JA |
1813 | for (i = 0; i < icd->nr; i++) { |
1814 | io_u = td->io_ops->event(td, i); | |
1815 | ||
f8b0bd10 | 1816 | io_completed(td, &io_u, icd); |
e8462bd8 | 1817 | |
f8b0bd10 | 1818 | if (io_u) |
e8462bd8 | 1819 | put_io_u(td, io_u); |
10ba535a JA |
1820 | } |
1821 | } | |
97601024 | 1822 | |
e7e6cfb4 JA |
1823 | /* |
1824 | * Complete a single io_u for the sync engines. | |
1825 | */ | |
55312f9f | 1826 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1827 | { |
1828 | struct io_completion_data icd; | |
55312f9f | 1829 | int ddir; |
97601024 | 1830 | |
9520ebb9 | 1831 | init_icd(td, &icd, 1); |
f8b0bd10 | 1832 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1833 | |
f8b0bd10 | 1834 | if (io_u) |
e8462bd8 | 1835 | put_io_u(td, io_u); |
97601024 | 1836 | |
581e7141 JA |
1837 | if (icd.error) { |
1838 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1839 | return -1; | |
1840 | } | |
97601024 | 1841 | |
55312f9f JA |
1842 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1843 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1844 | |
1845 | return 0; | |
97601024 JA |
1846 | } |
1847 | ||
e7e6cfb4 JA |
1848 | /* |
1849 | * Called to complete min_events number of io for the async engines. | |
1850 | */ | |
55312f9f | 1851 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 1852 | { |
97601024 | 1853 | struct io_completion_data icd; |
00de55ef | 1854 | struct timespec *tvp = NULL; |
55312f9f | 1855 | int ret, ddir; |
4d06a338 | 1856 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1857 | |
4950421a | 1858 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1859 | |
4950421a | 1860 | if (!min_evts) |
00de55ef | 1861 | tvp = &ts; |
5fb4b366 RE |
1862 | else if (min_evts > td->cur_depth) |
1863 | min_evts = td->cur_depth; | |
97601024 | 1864 | |
4950421a | 1865 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1866 | if (ret < 0) { |
e1161c32 | 1867 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1868 | return ret; |
1869 | } else if (!ret) | |
1870 | return ret; | |
1871 | ||
9520ebb9 | 1872 | init_icd(td, &icd, ret); |
97601024 | 1873 | ios_completed(td, &icd); |
581e7141 JA |
1874 | if (icd.error) { |
1875 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1876 | return -1; | |
1877 | } | |
97601024 | 1878 | |
55312f9f JA |
1879 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1880 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1881 | |
1882 | return 0; | |
97601024 | 1883 | } |
7e77dd02 JA |
1884 | |
1885 | /* | |
1886 | * Call when io_u is really queued, to update the submission latency. | |
1887 | */ | |
1888 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1889 | { | |
9520ebb9 JA |
1890 | if (!td->o.disable_slat) { |
1891 | unsigned long slat_time; | |
7e77dd02 | 1892 | |
9520ebb9 | 1893 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
ae588852 JA |
1894 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1895 | io_u->offset); | |
9520ebb9 | 1896 | } |
7e77dd02 | 1897 | } |
433afcb4 | 1898 | |
5c94b008 JA |
1899 | /* |
1900 | * See if we should reuse the last seed, if dedupe is enabled | |
1901 | */ | |
9451b93e | 1902 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 | 1903 | { |
c3546b53 | 1904 | uint64_t frand_max; |
5c94b008 JA |
1905 | unsigned int v; |
1906 | unsigned long r; | |
1907 | ||
1908 | if (!td->o.dedupe_percentage) | |
1909 | return &td->buf_state; | |
732eedd0 | 1910 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
1911 | frand_copy(&td->buf_state_prev, &td->buf_state); |
1912 | return &td->buf_state; | |
732eedd0 | 1913 | } |
5c94b008 | 1914 | |
c3546b53 | 1915 | frand_max = rand_max(&td->dedupe_state); |
5c94b008 | 1916 | r = __rand(&td->dedupe_state); |
c3546b53 | 1917 | v = 1 + (int) (100.0 * (r / (frand_max + 1.0))); |
5c94b008 JA |
1918 | |
1919 | if (v <= td->o.dedupe_percentage) | |
1920 | return &td->buf_state_prev; | |
1921 | ||
1922 | return &td->buf_state; | |
1923 | } | |
1924 | ||
9451b93e | 1925 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 1926 | { |
9451b93e JA |
1927 | if (td->o.dedupe_percentage == 100) |
1928 | frand_copy(rs, &td->buf_state_prev); | |
1929 | else if (rs == &td->buf_state) | |
5c94b008 JA |
1930 | frand_copy(&td->buf_state_prev, rs); |
1931 | } | |
1932 | ||
cc86c395 JA |
1933 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1934 | unsigned int max_bs) | |
5973cafb | 1935 | { |
d1af2894 JA |
1936 | struct thread_options *o = &td->o; |
1937 | ||
4eff3e57 | 1938 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 1939 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1940 | struct frand_state *rs; |
1066358a | 1941 | unsigned int left = max_bs; |
5c94b008 | 1942 | |
1066358a | 1943 | do { |
9451b93e | 1944 | rs = get_buf_state(td); |
9c42684e | 1945 | |
1066358a | 1946 | min_write = min(min_write, left); |
f97a43a1 | 1947 | |
1066358a JA |
1948 | if (perc) { |
1949 | unsigned int seg = min_write; | |
cc86c395 | 1950 | |
1066358a JA |
1951 | seg = min(min_write, td->o.compress_chunk); |
1952 | if (!seg) | |
1953 | seg = min_write; | |
1954 | ||
1955 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1956 | min_write, o->buffer_pattern, |
1957 | o->buffer_pattern_bytes); | |
1066358a JA |
1958 | } else |
1959 | fill_random_buf(rs, buf, min_write); | |
1960 | ||
1961 | buf += min_write; | |
1962 | left -= min_write; | |
9451b93e | 1963 | save_buf_state(td, rs); |
1066358a | 1964 | } while (left); |
d1af2894 JA |
1965 | } else if (o->buffer_pattern_bytes) |
1966 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 1967 | else if (o->zero_buffers) |
cc86c395 | 1968 | memset(buf, 0, max_bs); |
999d245e | 1969 | else |
9451b93e | 1970 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
1971 | } |
1972 | ||
1973 | /* | |
1974 | * "randomly" fill the buffer contents | |
1975 | */ | |
1976 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1977 | unsigned int min_write, unsigned int max_bs) | |
1978 | { | |
1979 | io_u->buf_filled_len = 0; | |
1980 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1981 | } |