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; | |
50a8ce86 | 571 | long usec, now; |
581e7141 | 572 | |
ff58fced | 573 | assert(ddir_rw(ddir)); |
50a8ce86 | 574 | now = utime_since_now(&td->start); |
ff58fced | 575 | |
50a8ce86 D |
576 | /* |
577 | * if rate_next_io_time is in the past, need to catch up to rate | |
578 | */ | |
579 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
580 | return ddir; |
581 | ||
582 | /* | |
50a8ce86 | 583 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
584 | * should switch. |
585 | */ | |
315fcfec | 586 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 587 | /* |
50a8ce86 | 588 | * Other direction is behind rate, switch |
581e7141 | 589 | */ |
50a8ce86 | 590 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
591 | return odir; |
592 | ||
593 | /* | |
50a8ce86 D |
594 | * Both directions are ahead of rate. sleep the min |
595 | * switch if necissary | |
581e7141 | 596 | */ |
50a8ce86 D |
597 | if (td->rate_next_io_time[ddir] <= |
598 | td->rate_next_io_time[odir]) { | |
599 | usec = td->rate_next_io_time[ddir] - now; | |
581e7141 | 600 | } else { |
50a8ce86 | 601 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
602 | ddir = odir; |
603 | } | |
604 | } else | |
50a8ce86 | 605 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 606 | |
a9da8ab2 JA |
607 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
608 | io_u_quiesce(td); | |
78c1eda5 | 609 | |
9cc80b6d | 610 | usec = usec_sleep(td, usec); |
581e7141 | 611 | |
581e7141 JA |
612 | return ddir; |
613 | } | |
614 | ||
10ba535a JA |
615 | /* |
616 | * Return the data direction for the next io_u. If the job is a | |
617 | * mixed read/write workload, check the rwmix cycle and switch if | |
618 | * necessary. | |
619 | */ | |
1e97cce9 | 620 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 621 | { |
581e7141 JA |
622 | enum fio_ddir ddir; |
623 | ||
5f9099ea JA |
624 | /* |
625 | * see if it's time to fsync | |
626 | */ | |
627 | if (td->o.fsync_blocks && | |
628 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
629 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
630 | return DDIR_SYNC; | |
631 | ||
632 | /* | |
633 | * see if it's time to fdatasync | |
634 | */ | |
635 | if (td->o.fdatasync_blocks && | |
636 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
637 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
638 | return DDIR_DATASYNC; | |
639 | ||
44f29692 JA |
640 | /* |
641 | * see if it's time to sync_file_range | |
642 | */ | |
643 | if (td->sync_file_range_nr && | |
644 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
645 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
646 | return DDIR_SYNC_FILE_RANGE; | |
647 | ||
10ba535a | 648 | if (td_rw(td)) { |
10ba535a JA |
649 | /* |
650 | * Check if it's time to seed a new data direction. | |
651 | */ | |
e4928662 | 652 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
653 | /* |
654 | * Put a top limit on how many bytes we do for | |
655 | * one data direction, to avoid overflowing the | |
656 | * ranges too much | |
657 | */ | |
658 | ddir = get_rand_ddir(td); | |
e47f799f JA |
659 | |
660 | if (ddir != td->rwmix_ddir) | |
661 | set_rwmix_bytes(td); | |
662 | ||
663 | td->rwmix_ddir = ddir; | |
10ba535a | 664 | } |
581e7141 | 665 | ddir = td->rwmix_ddir; |
10ba535a | 666 | } else if (td_read(td)) |
581e7141 | 667 | ddir = DDIR_READ; |
6eaf09d6 | 668 | else if (td_write(td)) |
581e7141 | 669 | ddir = DDIR_WRITE; |
6eaf09d6 SL |
670 | else |
671 | ddir = DDIR_TRIM; | |
581e7141 JA |
672 | |
673 | td->rwmix_ddir = rate_ddir(td, ddir); | |
674 | return td->rwmix_ddir; | |
10ba535a JA |
675 | } |
676 | ||
1ef2b6be JA |
677 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
678 | { | |
0e4dd95c DE |
679 | enum fio_ddir ddir = get_rw_ddir(td); |
680 | ||
82a90686 | 681 | if (td_trimwrite(td)) { |
0e4dd95c DE |
682 | struct fio_file *f = io_u->file; |
683 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
684 | ddir = DDIR_TRIM; | |
685 | else | |
686 | ddir = DDIR_WRITE; | |
687 | } | |
688 | ||
689 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be JA |
690 | |
691 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
692 | td->o.barrier_blocks && | |
693 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
694 | td->io_issues[DDIR_WRITE]) | |
a9da8ab2 | 695 | io_u_set(io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
696 | } |
697 | ||
e8462bd8 | 698 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 699 | { |
71b84caa | 700 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
701 | |
702 | if (ret) | |
703 | td_verror(td, ret, "file close"); | |
704 | } | |
705 | ||
10ba535a JA |
706 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
707 | { | |
a9da8ab2 JA |
708 | if (td->parent) |
709 | td = td->parent; | |
710 | ||
e8462bd8 JA |
711 | td_io_u_lock(td); |
712 | ||
f8b0bd10 | 713 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 714 | put_file_log(td, io_u->file); |
f8b0bd10 | 715 | |
10ba535a | 716 | io_u->file = NULL; |
a9da8ab2 | 717 | io_u_set(io_u, IO_U_F_FREE); |
d7ee2a7d | 718 | |
a9da8ab2 | 719 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 720 | td->cur_depth--; |
a9da8ab2 JA |
721 | assert(!(td->flags & TD_F_CHILD)); |
722 | } | |
2ae0b204 | 723 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
724 | td_io_u_unlock(td); |
725 | td_io_u_free_notify(td); | |
10ba535a JA |
726 | } |
727 | ||
f2bba182 RR |
728 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
729 | { | |
a9da8ab2 | 730 | io_u_clear(io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
731 | put_io_u(td, io_u); |
732 | } | |
733 | ||
755200a3 JA |
734 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
735 | { | |
736 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 737 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 738 | |
465221b0 JA |
739 | dprint(FD_IO, "requeue %p\n", __io_u); |
740 | ||
a9da8ab2 JA |
741 | if (td->parent) |
742 | td = td->parent; | |
743 | ||
e8462bd8 JA |
744 | td_io_u_lock(td); |
745 | ||
a9da8ab2 | 746 | io_u_set(__io_u, IO_U_F_FREE); |
bcd5abfa JA |
747 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
748 | td->io_issues[ddir]--; | |
5ec10eaa | 749 | |
a9da8ab2 JA |
750 | io_u_clear(__io_u, IO_U_F_FLIGHT); |
751 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { | |
0c41214f | 752 | td->cur_depth--; |
a9da8ab2 JA |
753 | assert(!(td->flags & TD_F_CHILD)); |
754 | } | |
2ae0b204 JA |
755 | |
756 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 757 | td_io_u_unlock(td); |
a9da8ab2 | 758 | td_io_u_free_notify(td); |
755200a3 JA |
759 | *io_u = NULL; |
760 | } | |
761 | ||
9bf2061e | 762 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 763 | { |
6aca9b3d JA |
764 | unsigned int is_random; |
765 | ||
b4c5e1ac JA |
766 | if (td->io_ops->flags & FIO_NOIO) |
767 | goto out; | |
768 | ||
1ef2b6be | 769 | set_rw_ddir(td, io_u); |
5f9099ea | 770 | |
87dc1ab1 | 771 | /* |
ff58fced | 772 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 773 | */ |
ff58fced | 774 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 775 | goto out; |
a00735e6 | 776 | |
48f5abd3 JA |
777 | /* |
778 | * See if it's time to switch to a new zone | |
779 | */ | |
13af05ae | 780 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
781 | struct fio_file *f = io_u->file; |
782 | ||
48f5abd3 | 783 | td->zone_bytes = 0; |
4c8be5b1 JA |
784 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
785 | ||
786 | /* | |
787 | * Wrap from the beginning, if we exceed the file size | |
788 | */ | |
789 | if (f->file_offset >= f->real_file_size) | |
790 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 791 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
792 | td->io_skip_bytes += td->o.zone_skip; |
793 | } | |
794 | ||
10ba535a | 795 | /* |
c685b5b2 JA |
796 | * No log, let the seq/rand engine retrieve the next buflen and |
797 | * position. | |
10ba535a | 798 | */ |
6aca9b3d | 799 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 800 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 801 | return 1; |
2ba1c290 | 802 | } |
10ba535a | 803 | |
6aca9b3d | 804 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
805 | if (!io_u->buflen) { |
806 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 807 | return 1; |
2ba1c290 | 808 | } |
bca4ed4d | 809 | |
2ba1c290 JA |
810 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
811 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4b91ee8f JA |
812 | dprint(FD_IO, " off=%llu/%lu > %llu\n", |
813 | (unsigned long long) io_u->offset, io_u->buflen, | |
814 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 815 | return 1; |
2ba1c290 | 816 | } |
6a5e6884 | 817 | |
bca4ed4d JA |
818 | /* |
819 | * mark entry before potentially trimming io_u | |
820 | */ | |
303032ae | 821 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 822 | mark_random_map(td, io_u); |
bca4ed4d | 823 | |
c38e9468 | 824 | out: |
2ba1c290 | 825 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 826 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 827 | return 0; |
10ba535a JA |
828 | } |
829 | ||
838bc709 JA |
830 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
831 | { | |
2b13e716 | 832 | int idx = 0; |
838bc709 JA |
833 | |
834 | switch (nr) { | |
835 | default: | |
2b13e716 | 836 | idx = 6; |
838bc709 JA |
837 | break; |
838 | case 33 ... 64: | |
2b13e716 | 839 | idx = 5; |
838bc709 JA |
840 | break; |
841 | case 17 ... 32: | |
2b13e716 | 842 | idx = 4; |
838bc709 JA |
843 | break; |
844 | case 9 ... 16: | |
2b13e716 | 845 | idx = 3; |
838bc709 JA |
846 | break; |
847 | case 5 ... 8: | |
2b13e716 | 848 | idx = 2; |
838bc709 JA |
849 | break; |
850 | case 1 ... 4: | |
2b13e716 | 851 | idx = 1; |
838bc709 JA |
852 | case 0: |
853 | break; | |
854 | } | |
855 | ||
2b13e716 | 856 | map[idx]++; |
838bc709 JA |
857 | } |
858 | ||
859 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
860 | { | |
861 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
862 | td->ts.total_submit++; | |
863 | } | |
864 | ||
865 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
866 | { | |
867 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
868 | td->ts.total_complete++; | |
869 | } | |
870 | ||
d8005759 | 871 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 872 | { |
2b13e716 | 873 | int idx = 0; |
71619dc2 JA |
874 | |
875 | switch (td->cur_depth) { | |
876 | default: | |
2b13e716 | 877 | idx = 6; |
a783e61a | 878 | break; |
71619dc2 | 879 | case 32 ... 63: |
2b13e716 | 880 | idx = 5; |
a783e61a | 881 | break; |
71619dc2 | 882 | case 16 ... 31: |
2b13e716 | 883 | idx = 4; |
a783e61a | 884 | break; |
71619dc2 | 885 | case 8 ... 15: |
2b13e716 | 886 | idx = 3; |
a783e61a | 887 | break; |
71619dc2 | 888 | case 4 ... 7: |
2b13e716 | 889 | idx = 2; |
a783e61a | 890 | break; |
71619dc2 | 891 | case 2 ... 3: |
2b13e716 | 892 | idx = 1; |
71619dc2 JA |
893 | case 1: |
894 | break; | |
895 | } | |
896 | ||
2b13e716 | 897 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
898 | } |
899 | ||
04a0feae JA |
900 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
901 | { | |
2b13e716 | 902 | int idx = 0; |
04a0feae JA |
903 | |
904 | assert(usec < 1000); | |
905 | ||
906 | switch (usec) { | |
907 | case 750 ... 999: | |
2b13e716 | 908 | idx = 9; |
04a0feae JA |
909 | break; |
910 | case 500 ... 749: | |
2b13e716 | 911 | idx = 8; |
04a0feae JA |
912 | break; |
913 | case 250 ... 499: | |
2b13e716 | 914 | idx = 7; |
04a0feae JA |
915 | break; |
916 | case 100 ... 249: | |
2b13e716 | 917 | idx = 6; |
04a0feae JA |
918 | break; |
919 | case 50 ... 99: | |
2b13e716 | 920 | idx = 5; |
04a0feae JA |
921 | break; |
922 | case 20 ... 49: | |
2b13e716 | 923 | idx = 4; |
04a0feae JA |
924 | break; |
925 | case 10 ... 19: | |
2b13e716 | 926 | idx = 3; |
04a0feae JA |
927 | break; |
928 | case 4 ... 9: | |
2b13e716 | 929 | idx = 2; |
04a0feae JA |
930 | break; |
931 | case 2 ... 3: | |
2b13e716 | 932 | idx = 1; |
04a0feae JA |
933 | case 0 ... 1: |
934 | break; | |
935 | } | |
936 | ||
2b13e716 JA |
937 | assert(idx < FIO_IO_U_LAT_U_NR); |
938 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
939 | } |
940 | ||
941 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 942 | { |
2b13e716 | 943 | int idx = 0; |
ec118304 JA |
944 | |
945 | switch (msec) { | |
946 | default: | |
2b13e716 | 947 | idx = 11; |
04a0feae | 948 | break; |
8abdce66 | 949 | case 1000 ... 1999: |
2b13e716 | 950 | idx = 10; |
04a0feae | 951 | break; |
8abdce66 | 952 | case 750 ... 999: |
2b13e716 | 953 | idx = 9; |
04a0feae | 954 | break; |
8abdce66 | 955 | case 500 ... 749: |
2b13e716 | 956 | idx = 8; |
04a0feae | 957 | break; |
8abdce66 | 958 | case 250 ... 499: |
2b13e716 | 959 | idx = 7; |
04a0feae | 960 | break; |
8abdce66 | 961 | case 100 ... 249: |
2b13e716 | 962 | idx = 6; |
04a0feae | 963 | break; |
8abdce66 | 964 | case 50 ... 99: |
2b13e716 | 965 | idx = 5; |
04a0feae | 966 | break; |
8abdce66 | 967 | case 20 ... 49: |
2b13e716 | 968 | idx = 4; |
04a0feae | 969 | break; |
8abdce66 | 970 | case 10 ... 19: |
2b13e716 | 971 | idx = 3; |
04a0feae | 972 | break; |
8abdce66 | 973 | case 4 ... 9: |
2b13e716 | 974 | idx = 2; |
04a0feae | 975 | break; |
ec118304 | 976 | case 2 ... 3: |
2b13e716 | 977 | idx = 1; |
ec118304 JA |
978 | case 0 ... 1: |
979 | break; | |
980 | } | |
981 | ||
2b13e716 JA |
982 | assert(idx < FIO_IO_U_LAT_M_NR); |
983 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
984 | } |
985 | ||
986 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
987 | { | |
988 | if (usec < 1000) | |
989 | io_u_mark_lat_usec(td, usec); | |
990 | else | |
991 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
992 | } |
993 | ||
0aabe160 JA |
994 | /* |
995 | * Get next file to service by choosing one at random | |
996 | */ | |
2cc52930 JA |
997 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
998 | enum fio_file_flags goodf, | |
d6aed795 | 999 | enum fio_file_flags badf) |
0aabe160 | 1000 | { |
c3546b53 | 1001 | uint64_t frand_max = rand_max(&td->next_file_state); |
0aabe160 | 1002 | struct fio_file *f; |
1c178180 | 1003 | int fno; |
0aabe160 JA |
1004 | |
1005 | do { | |
87b10676 | 1006 | int opened = 0; |
1294c3ec | 1007 | unsigned long r; |
4c07ad86 | 1008 | |
d6b72507 | 1009 | r = __rand(&td->next_file_state); |
54a21917 | 1010 | fno = (unsigned int) ((double) td->o.nr_files |
c3546b53 | 1011 | * (r / (frand_max + 1.0))); |
7c83c089 | 1012 | |
126d65c6 | 1013 | f = td->files[fno]; |
d6aed795 | 1014 | if (fio_file_done(f)) |
059e63c0 | 1015 | continue; |
1c178180 | 1016 | |
d6aed795 | 1017 | if (!fio_file_open(f)) { |
87b10676 JA |
1018 | int err; |
1019 | ||
002fe734 JA |
1020 | if (td->nr_open_files >= td->o.open_files) |
1021 | return ERR_PTR(-EBUSY); | |
1022 | ||
87b10676 JA |
1023 | err = td_io_open_file(td, f); |
1024 | if (err) | |
1025 | continue; | |
1026 | opened = 1; | |
1027 | } | |
1028 | ||
2ba1c290 JA |
1029 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1030 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1031 | return f; |
2ba1c290 | 1032 | } |
87b10676 JA |
1033 | if (opened) |
1034 | td_io_close_file(td, f); | |
0aabe160 JA |
1035 | } while (1); |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * Get next file to service by doing round robin between all available ones | |
1040 | */ | |
1c178180 JA |
1041 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1042 | int badf) | |
3d7c391d JA |
1043 | { |
1044 | unsigned int old_next_file = td->next_file; | |
1045 | struct fio_file *f; | |
1046 | ||
1047 | do { | |
87b10676 JA |
1048 | int opened = 0; |
1049 | ||
126d65c6 | 1050 | f = td->files[td->next_file]; |
3d7c391d JA |
1051 | |
1052 | td->next_file++; | |
2dc1bbeb | 1053 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1054 | td->next_file = 0; |
1055 | ||
87b10676 | 1056 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1057 | if (fio_file_done(f)) { |
d5ed68ea | 1058 | f = NULL; |
059e63c0 | 1059 | continue; |
d5ed68ea | 1060 | } |
059e63c0 | 1061 | |
d6aed795 | 1062 | if (!fio_file_open(f)) { |
87b10676 JA |
1063 | int err; |
1064 | ||
002fe734 JA |
1065 | if (td->nr_open_files >= td->o.open_files) |
1066 | return ERR_PTR(-EBUSY); | |
1067 | ||
87b10676 | 1068 | err = td_io_open_file(td, f); |
b5696bfc JA |
1069 | if (err) { |
1070 | dprint(FD_FILE, "error %d on open of %s\n", | |
1071 | err, f->file_name); | |
87c27b45 | 1072 | f = NULL; |
87b10676 | 1073 | continue; |
b5696bfc | 1074 | } |
87b10676 JA |
1075 | opened = 1; |
1076 | } | |
1077 | ||
0b9d69ec JA |
1078 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1079 | f->flags); | |
1c178180 | 1080 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1081 | break; |
1082 | ||
87b10676 JA |
1083 | if (opened) |
1084 | td_io_close_file(td, f); | |
1085 | ||
3d7c391d JA |
1086 | f = NULL; |
1087 | } while (td->next_file != old_next_file); | |
1088 | ||
2ba1c290 | 1089 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1090 | return f; |
1091 | } | |
1092 | ||
7eb36574 | 1093 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1094 | { |
1907dbc6 JA |
1095 | struct fio_file *f; |
1096 | ||
2dc1bbeb | 1097 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1098 | |
b5696bfc | 1099 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1100 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1101 | " nr_files=%d\n", td->nr_open_files, | |
1102 | td->nr_done_files, | |
1103 | td->o.nr_files); | |
bdb4e2e9 | 1104 | return NULL; |
2ba1c290 | 1105 | } |
bdb4e2e9 | 1106 | |
1907dbc6 | 1107 | f = td->file_service_file; |
d6aed795 | 1108 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1109 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1110 | goto out; | |
1111 | if (td->file_service_left--) | |
1112 | goto out; | |
1113 | } | |
1907dbc6 | 1114 | |
a086c257 JA |
1115 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1116 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1117 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1118 | else |
d6aed795 | 1119 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1120 | |
002fe734 JA |
1121 | if (IS_ERR(f)) |
1122 | return f; | |
1123 | ||
1907dbc6 JA |
1124 | td->file_service_file = f; |
1125 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1126 | out: |
0dac421f JA |
1127 | if (f) |
1128 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1129 | else | |
1130 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1131 | return f; |
bdb4e2e9 JA |
1132 | } |
1133 | ||
7eb36574 JA |
1134 | static struct fio_file *get_next_file(struct thread_data *td) |
1135 | { | |
372d8962 | 1136 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1137 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1138 | |
d72be545 JA |
1139 | if (ops->get_next_file) |
1140 | return ops->get_next_file(td); | |
1141 | } | |
7eb36574 JA |
1142 | |
1143 | return __get_next_file(td); | |
1144 | } | |
1145 | ||
002fe734 | 1146 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1147 | { |
1148 | struct fio_file *f; | |
1149 | ||
1150 | do { | |
1151 | f = get_next_file(td); | |
002fe734 JA |
1152 | if (IS_ERR_OR_NULL(f)) |
1153 | return PTR_ERR(f); | |
429f6675 | 1154 | |
429f6675 JA |
1155 | io_u->file = f; |
1156 | get_file(f); | |
1157 | ||
1158 | if (!fill_io_u(td, io_u)) | |
1159 | break; | |
1160 | ||
b5696bfc | 1161 | put_file_log(td, f); |
429f6675 | 1162 | td_io_close_file(td, f); |
b5696bfc | 1163 | io_u->file = NULL; |
d6aed795 | 1164 | fio_file_set_done(f); |
429f6675 | 1165 | td->nr_done_files++; |
0b9d69ec JA |
1166 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1167 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1168 | } while (1); |
1169 | ||
1170 | return 0; | |
1171 | } | |
1172 | ||
3e260a46 JA |
1173 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1174 | unsigned long tusec, unsigned long max_usec) | |
1175 | { | |
1176 | if (!td->error) | |
1177 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1178 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1179 | icd->error = ETIMEDOUT; | |
1180 | } | |
1181 | ||
1182 | static void lat_new_cycle(struct thread_data *td) | |
1183 | { | |
1184 | fio_gettime(&td->latency_ts, NULL); | |
1185 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1186 | td->latency_failed = 0; | |
1187 | } | |
1188 | ||
1189 | /* | |
1190 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1191 | * are at QD=1, then it's time to give up. | |
1192 | */ | |
1193 | static int __lat_target_failed(struct thread_data *td) | |
1194 | { | |
1195 | if (td->latency_qd == 1) | |
1196 | return 1; | |
1197 | ||
1198 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1199 | |
1200 | if (td->latency_qd == td->latency_qd_low) | |
1201 | td->latency_qd_low--; | |
1202 | ||
3e260a46 JA |
1203 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1204 | ||
1205 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1206 | ||
1207 | /* | |
1208 | * When we ramp QD down, quiesce existing IO to prevent | |
1209 | * a storm of ramp downs due to pending higher depth. | |
1210 | */ | |
1211 | io_u_quiesce(td); | |
1212 | lat_new_cycle(td); | |
1213 | return 0; | |
1214 | } | |
1215 | ||
1216 | static int lat_target_failed(struct thread_data *td) | |
1217 | { | |
1218 | if (td->o.latency_percentile.u.f == 100.0) | |
1219 | return __lat_target_failed(td); | |
1220 | ||
1221 | td->latency_failed++; | |
1222 | return 0; | |
1223 | } | |
1224 | ||
1225 | void lat_target_init(struct thread_data *td) | |
1226 | { | |
6bb58215 JA |
1227 | td->latency_end_run = 0; |
1228 | ||
3e260a46 JA |
1229 | if (td->o.latency_target) { |
1230 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1231 | fio_gettime(&td->latency_ts, NULL); | |
1232 | td->latency_qd = 1; | |
1233 | td->latency_qd_high = td->o.iodepth; | |
1234 | td->latency_qd_low = 1; | |
1235 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1236 | } else | |
1237 | td->latency_qd = td->o.iodepth; | |
1238 | } | |
1239 | ||
6bb58215 JA |
1240 | void lat_target_reset(struct thread_data *td) |
1241 | { | |
1242 | if (!td->latency_end_run) | |
1243 | lat_target_init(td); | |
1244 | } | |
1245 | ||
3e260a46 JA |
1246 | static void lat_target_success(struct thread_data *td) |
1247 | { | |
1248 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1249 | struct thread_options *o = &td->o; |
3e260a46 JA |
1250 | |
1251 | td->latency_qd_low = td->latency_qd; | |
1252 | ||
1253 | /* | |
1254 | * If we haven't failed yet, we double up to a failing value instead | |
1255 | * of bisecting from highest possible queue depth. If we have set | |
1256 | * a limit other than td->o.iodepth, bisect between that. | |
1257 | */ | |
6bb58215 | 1258 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1259 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1260 | else | |
1261 | td->latency_qd *= 2; | |
1262 | ||
6bb58215 JA |
1263 | if (td->latency_qd > o->iodepth) |
1264 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1265 | |
1266 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1267 | |
3e260a46 | 1268 | /* |
6bb58215 JA |
1269 | * Same as last one, we are done. Let it run a latency cycle, so |
1270 | * we get only the results from the targeted depth. | |
3e260a46 | 1271 | */ |
6bb58215 JA |
1272 | if (td->latency_qd == qd) { |
1273 | if (td->latency_end_run) { | |
1274 | dprint(FD_RATE, "We are done\n"); | |
1275 | td->done = 1; | |
1276 | } else { | |
1277 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1278 | io_u_quiesce(td); | |
1279 | td->latency_end_run = 1; | |
1280 | reset_all_stats(td); | |
1281 | reset_io_stats(td); | |
1282 | } | |
1283 | } | |
3e260a46 JA |
1284 | |
1285 | lat_new_cycle(td); | |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * Check if we can bump the queue depth | |
1290 | */ | |
1291 | void lat_target_check(struct thread_data *td) | |
1292 | { | |
1293 | uint64_t usec_window; | |
1294 | uint64_t ios; | |
1295 | double success_ios; | |
1296 | ||
1297 | usec_window = utime_since_now(&td->latency_ts); | |
1298 | if (usec_window < td->o.latency_window) | |
1299 | return; | |
1300 | ||
1301 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1302 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1303 | success_ios *= 100.0; | |
1304 | ||
1305 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1306 | ||
1307 | if (success_ios >= td->o.latency_percentile.u.f) | |
1308 | lat_target_success(td); | |
1309 | else | |
1310 | __lat_target_failed(td); | |
1311 | } | |
1312 | ||
1313 | /* | |
1314 | * If latency target is enabled, we might be ramping up or down and not | |
1315 | * using the full queue depth available. | |
1316 | */ | |
effd6ff0 | 1317 | int queue_full(const struct thread_data *td) |
3e260a46 JA |
1318 | { |
1319 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1320 | ||
1321 | if (qempty) | |
1322 | return 1; | |
1323 | if (!td->o.latency_target) | |
1324 | return 0; | |
1325 | ||
1326 | return td->cur_depth >= td->latency_qd; | |
1327 | } | |
429f6675 | 1328 | |
10ba535a JA |
1329 | struct io_u *__get_io_u(struct thread_data *td) |
1330 | { | |
0cae66f6 | 1331 | struct io_u *io_u = NULL; |
10ba535a | 1332 | |
ca09be4b JA |
1333 | if (td->stop_io) |
1334 | return NULL; | |
1335 | ||
e8462bd8 JA |
1336 | td_io_u_lock(td); |
1337 | ||
1338 | again: | |
2ae0b204 JA |
1339 | if (!io_u_rempty(&td->io_u_requeues)) |
1340 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1341 | else if (!queue_full(td)) { |
2ae0b204 | 1342 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1343 | |
225ba9e3 | 1344 | io_u->file = NULL; |
6040dabc | 1345 | io_u->buflen = 0; |
10ba535a | 1346 | io_u->resid = 0; |
d7762cf8 | 1347 | io_u->end_io = NULL; |
755200a3 JA |
1348 | } |
1349 | ||
1350 | if (io_u) { | |
0c6e7517 | 1351 | assert(io_u->flags & IO_U_F_FREE); |
a9da8ab2 | 1352 | io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1353 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1354 | IO_U_F_VER_LIST); | |
0c6e7517 | 1355 | |
755200a3 | 1356 | io_u->error = 0; |
bcd5abfa | 1357 | io_u->acct_ddir = -1; |
10ba535a | 1358 | td->cur_depth++; |
a9da8ab2 JA |
1359 | assert(!(td->flags & TD_F_CHILD)); |
1360 | io_u_set(io_u, IO_U_F_IN_CUR_DEPTH); | |
f9401285 | 1361 | io_u->ipo = NULL; |
a9da8ab2 | 1362 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1363 | /* |
1364 | * We ran out, wait for async verify threads to finish and | |
1365 | * return one | |
1366 | */ | |
a9da8ab2 JA |
1367 | assert(!(td->flags & TD_F_CHILD)); |
1368 | assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock)); | |
1dec3e07 | 1369 | goto again; |
10ba535a JA |
1370 | } |
1371 | ||
e8462bd8 | 1372 | td_io_u_unlock(td); |
10ba535a JA |
1373 | return io_u; |
1374 | } | |
1375 | ||
0d29de83 | 1376 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1377 | { |
d72be545 JA |
1378 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
1379 | return 0; | |
1380 | ||
1381 | if (td->trim_entries) { | |
0d29de83 | 1382 | int get_trim = 0; |
10ba535a | 1383 | |
0d29de83 JA |
1384 | if (td->trim_batch) { |
1385 | td->trim_batch--; | |
1386 | get_trim = 1; | |
1387 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1388 | td->last_ddir != DDIR_READ) { | |
1389 | td->trim_batch = td->o.trim_batch; | |
1390 | if (!td->trim_batch) | |
1391 | td->trim_batch = td->o.trim_backlog; | |
1392 | get_trim = 1; | |
1393 | } | |
1394 | ||
1395 | if (get_trim && !get_next_trim(td, io_u)) | |
1396 | return 1; | |
2ba1c290 | 1397 | } |
10ba535a | 1398 | |
0d29de83 JA |
1399 | return 0; |
1400 | } | |
1401 | ||
1402 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1403 | { | |
d72be545 JA |
1404 | if (!(td->flags & TD_F_VER_BACKLOG)) |
1405 | return 0; | |
1406 | ||
1407 | if (td->io_hist_len) { | |
9e144189 JA |
1408 | int get_verify = 0; |
1409 | ||
d1ece0c7 | 1410 | if (td->verify_batch) |
9e144189 | 1411 | get_verify = 1; |
d1ece0c7 | 1412 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1413 | td->last_ddir != DDIR_READ) { |
1414 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1415 | if (!td->verify_batch) |
1416 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1417 | get_verify = 1; |
1418 | } | |
1419 | ||
d1ece0c7 JA |
1420 | if (get_verify && !get_next_verify(td, io_u)) { |
1421 | td->verify_batch--; | |
0d29de83 | 1422 | return 1; |
d1ece0c7 | 1423 | } |
9e144189 JA |
1424 | } |
1425 | ||
0d29de83 JA |
1426 | return 0; |
1427 | } | |
1428 | ||
de789769 JA |
1429 | /* |
1430 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1431 | * easy compressible data for simple de-dupe attempts. Do this for every |
1432 | * 512b block in the range, since that should be the smallest block size | |
1433 | * we can expect from a device. | |
de789769 JA |
1434 | */ |
1435 | static void small_content_scramble(struct io_u *io_u) | |
1436 | { | |
23f394d5 | 1437 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1438 | uint64_t boffset; |
23f394d5 JA |
1439 | unsigned int offset; |
1440 | void *p, *end; | |
de789769 | 1441 | |
23f394d5 JA |
1442 | if (!nr_blocks) |
1443 | return; | |
1444 | ||
1445 | p = io_u->xfer_buf; | |
fba76ee8 | 1446 | boffset = io_u->offset; |
81f0366c | 1447 | io_u->buf_filled_len = 0; |
fad82f76 | 1448 | |
23f394d5 JA |
1449 | for (i = 0; i < nr_blocks; i++) { |
1450 | /* | |
1451 | * Fill the byte offset into a "random" start offset of | |
1452 | * the buffer, given by the product of the usec time | |
1453 | * and the actual offset. | |
1454 | */ | |
fad82f76 | 1455 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
1ae83d45 JA |
1456 | offset &= ~(sizeof(uint64_t) - 1); |
1457 | if (offset >= 512 - sizeof(uint64_t)) | |
1458 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1459 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1460 | |
1461 | end = p + 512 - sizeof(io_u->start_time); | |
1462 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1463 | p += 512; | |
fad82f76 | 1464 | boffset += 512; |
23f394d5 | 1465 | } |
de789769 JA |
1466 | } |
1467 | ||
0d29de83 JA |
1468 | /* |
1469 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1470 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1471 | */ | |
1472 | struct io_u *get_io_u(struct thread_data *td) | |
1473 | { | |
1474 | struct fio_file *f; | |
1475 | struct io_u *io_u; | |
de789769 | 1476 | int do_scramble = 0; |
002fe734 | 1477 | long ret = 0; |
0d29de83 JA |
1478 | |
1479 | io_u = __get_io_u(td); | |
1480 | if (!io_u) { | |
1481 | dprint(FD_IO, "__get_io_u failed\n"); | |
1482 | return NULL; | |
1483 | } | |
1484 | ||
1485 | if (check_get_verify(td, io_u)) | |
1486 | goto out; | |
1487 | if (check_get_trim(td, io_u)) | |
1488 | goto out; | |
1489 | ||
755200a3 JA |
1490 | /* |
1491 | * from a requeue, io_u already setup | |
1492 | */ | |
1493 | if (io_u->file) | |
77f392bf | 1494 | goto out; |
755200a3 | 1495 | |
429f6675 JA |
1496 | /* |
1497 | * If using an iolog, grab next piece if any available. | |
1498 | */ | |
d72be545 | 1499 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1500 | if (read_iolog_get(td, io_u)) |
1501 | goto err_put; | |
2ba1c290 | 1502 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1503 | ret = -EBUSY; |
2ba1c290 | 1504 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1505 | goto err_put; |
2ba1c290 | 1506 | } |
5ec10eaa | 1507 | |
429f6675 | 1508 | f = io_u->file; |
002fe734 JA |
1509 | if (!f) { |
1510 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1511 | goto err_put; | |
1512 | } | |
1513 | ||
d6aed795 | 1514 | assert(fio_file_open(f)); |
97af62ce | 1515 | |
ff58fced | 1516 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1517 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1518 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1519 | goto err_put; |
2ba1c290 | 1520 | } |
10ba535a | 1521 | |
f1dfb668 JA |
1522 | f->last_start[io_u->ddir] = io_u->offset; |
1523 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1524 | |
fd68418e | 1525 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1526 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1527 | io_u_fill_buffer(td, io_u, |
1066358a JA |
1528 | td->o.min_bs[DDIR_WRITE], |
1529 | io_u->xfer_buflen); | |
ff441ae8 | 1530 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1531 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1532 | do_scramble = 1; |
d72be545 | 1533 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1534 | populate_verify_io_u(td, io_u); |
1535 | do_scramble = 0; | |
1536 | } | |
fd68418e | 1537 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1538 | /* |
1539 | * Reset the buf_filled parameters so next time if the | |
1540 | * buffer is used for writes it is refilled. | |
1541 | */ | |
cbe8d756 RR |
1542 | io_u->buf_filled_len = 0; |
1543 | } | |
87dc1ab1 | 1544 | } |
10ba535a | 1545 | |
165faf16 JA |
1546 | /* |
1547 | * Set io data pointers. | |
1548 | */ | |
cec6b55d JA |
1549 | io_u->xfer_buf = io_u->buf; |
1550 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1551 | |
6ac7a331 | 1552 | out: |
0d29de83 | 1553 | assert(io_u->file); |
429f6675 | 1554 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1555 | if (!td->o.disable_slat) |
1556 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1557 | if (do_scramble) |
1558 | small_content_scramble(io_u); | |
429f6675 | 1559 | return io_u; |
36167d82 | 1560 | } |
429f6675 | 1561 | err_put: |
2ba1c290 | 1562 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1563 | put_io_u(td, io_u); |
002fe734 | 1564 | return ERR_PTR(ret); |
10ba535a JA |
1565 | } |
1566 | ||
a9da8ab2 | 1567 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1568 | { |
8b28bd41 | 1569 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1570 | |
8b28bd41 DM |
1571 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1572 | return; | |
5451792e | 1573 | |
709c8313 RE |
1574 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1575 | io_u->file ? " on file " : "", | |
1576 | io_u->file ? io_u->file->file_name : "", | |
1577 | strerror(io_u->error), | |
1578 | io_ddir_name(io_u->ddir), | |
1579 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1580 | |
5ad7be56 KD |
1581 | if (td->io_ops->errdetails) { |
1582 | char *err = td->io_ops->errdetails(io_u); | |
1583 | ||
1584 | log_err("fio: %s\n", err); | |
1585 | free(err); | |
1586 | } | |
1587 | ||
5451792e JA |
1588 | if (!td->error) |
1589 | td_verror(td, io_u->error, "io_u error"); | |
1590 | } | |
1591 | ||
a9da8ab2 JA |
1592 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1593 | { | |
1594 | __io_u_log_error(td, io_u); | |
1595 | if (td->parent) | |
1596 | __io_u_log_error(td, io_u); | |
1597 | } | |
1598 | ||
aba6c951 JA |
1599 | static inline int gtod_reduce(struct thread_data *td) |
1600 | { | |
729fe3af | 1601 | return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat |
b74b8208 | 1602 | && td->o.disable_bw; |
aba6c951 JA |
1603 | } |
1604 | ||
c8eeb9df JA |
1605 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1606 | struct io_completion_data *icd, | |
1607 | const enum fio_ddir idx, unsigned int bytes) | |
1608 | { | |
a9da8ab2 | 1609 | const int no_reduce = !gtod_reduce(td); |
24d23ca7 | 1610 | unsigned long lusec = 0; |
c8eeb9df | 1611 | |
75dc383e JA |
1612 | if (td->parent) |
1613 | td = td->parent; | |
1614 | ||
a9da8ab2 | 1615 | if (no_reduce) |
c8eeb9df JA |
1616 | lusec = utime_since(&io_u->issue_time, &icd->time); |
1617 | ||
1618 | if (!td->o.disable_lat) { | |
1619 | unsigned long tusec; | |
1620 | ||
1621 | tusec = utime_since(&io_u->start_time, &icd->time); | |
ae588852 | 1622 | add_lat_sample(td, idx, tusec, bytes, io_u->offset); |
15501535 | 1623 | |
d4afedfd JA |
1624 | if (td->flags & TD_F_PROFILE_OPS) { |
1625 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1626 | ||
1627 | if (ops->io_u_lat) | |
1628 | icd->error = ops->io_u_lat(td, tusec); | |
1629 | } | |
1630 | ||
3e260a46 JA |
1631 | if (td->o.max_latency && tusec > td->o.max_latency) |
1632 | lat_fatal(td, icd, tusec, td->o.max_latency); | |
1633 | if (td->o.latency_target && tusec > td->o.latency_target) { | |
1634 | if (lat_target_failed(td)) | |
1635 | lat_fatal(td, icd, tusec, td->o.latency_target); | |
15501535 | 1636 | } |
c8eeb9df JA |
1637 | } |
1638 | ||
1639 | if (!td->o.disable_clat) { | |
ae588852 | 1640 | add_clat_sample(td, idx, lusec, bytes, io_u->offset); |
c8eeb9df JA |
1641 | io_u_mark_latency(td, lusec); |
1642 | } | |
1643 | ||
1644 | if (!td->o.disable_bw) | |
1645 | add_bw_sample(td, idx, bytes, &icd->time); | |
1646 | ||
a9da8ab2 | 1647 | if (no_reduce) |
aba6c951 | 1648 | add_iops_sample(td, idx, bytes, &icd->time); |
66347cfa DE |
1649 | |
1650 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1651 | uint32_t *info = io_u_block_info(td, io_u); | |
1652 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1653 | if (io_u->ddir == DDIR_TRIM) { | |
1654 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1655 | BLOCK_INFO_TRIMS(*info) + 1); | |
1656 | } else if (io_u->ddir == DDIR_WRITE) { | |
1657 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1658 | *info); | |
1659 | } | |
1660 | } | |
1661 | } | |
c8eeb9df JA |
1662 | } |
1663 | ||
f8b0bd10 | 1664 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1665 | struct io_completion_data *icd) |
10ba535a | 1666 | { |
f8b0bd10 JA |
1667 | struct io_u *io_u = *io_u_ptr; |
1668 | enum fio_ddir ddir = io_u->ddir; | |
1669 | struct fio_file *f = io_u->file; | |
10ba535a | 1670 | |
2ba1c290 JA |
1671 | dprint_io_u(io_u, "io complete"); |
1672 | ||
0c6e7517 | 1673 | assert(io_u->flags & IO_U_F_FLIGHT); |
a9da8ab2 | 1674 | io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1675 | |
1676 | /* | |
1677 | * Mark IO ok to verify | |
1678 | */ | |
1679 | if (io_u->ipo) { | |
890b6656 JA |
1680 | /* |
1681 | * Remove errored entry from the verification list | |
1682 | */ | |
1683 | if (io_u->error) | |
1684 | unlog_io_piece(td, io_u); | |
1685 | else { | |
1686 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1687 | write_barrier(); | |
1688 | } | |
f9401285 JA |
1689 | } |
1690 | ||
f8b0bd10 | 1691 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1692 | td->last_was_sync = 1; |
44f29692 JA |
1693 | if (f) { |
1694 | f->first_write = -1ULL; | |
1695 | f->last_write = -1ULL; | |
1696 | } | |
87dc1ab1 JA |
1697 | return; |
1698 | } | |
1699 | ||
1700 | td->last_was_sync = 0; | |
f8b0bd10 | 1701 | td->last_ddir = ddir; |
87dc1ab1 | 1702 | |
f8b0bd10 | 1703 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1704 | unsigned int bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1705 | int ret; |
10ba535a | 1706 | |
f8b0bd10 JA |
1707 | td->io_blocks[ddir]++; |
1708 | td->this_io_blocks[ddir]++; | |
1709 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1710 | |
1711 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1712 | td->this_io_bytes[ddir] += bytes; |
1713 | ||
ca09be4b JA |
1714 | if (ddir == DDIR_WRITE) { |
1715 | if (f) { | |
1716 | if (f->first_write == -1ULL || | |
1717 | io_u->offset < f->first_write) | |
1718 | f->first_write = io_u->offset; | |
1719 | if (f->last_write == -1ULL || | |
1720 | ((io_u->offset + bytes) > f->last_write)) | |
1721 | f->last_write = io_u->offset + bytes; | |
1722 | } | |
1723 | if (td->last_write_comp) { | |
1724 | int idx = td->last_write_idx++; | |
1725 | ||
1726 | td->last_write_comp[idx] = io_u->offset; | |
1727 | if (td->last_write_idx == td->o.iodepth) | |
1728 | td->last_write_idx = 0; | |
1729 | } | |
44f29692 JA |
1730 | } |
1731 | ||
6b1190fd | 1732 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
50a8ce86 | 1733 | td->runstate == TD_VERIFYING)) |
f8b0bd10 | 1734 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1735 | |
f8b0bd10 | 1736 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1737 | |
d7762cf8 | 1738 | if (io_u->end_io) { |
f8b0bd10 JA |
1739 | ret = io_u->end_io(td, io_u_ptr); |
1740 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1741 | if (ret && !icd->error) |
1742 | icd->error = ret; | |
1743 | } | |
ff58fced | 1744 | } else if (io_u->error) { |
10ba535a | 1745 | icd->error = io_u->error; |
5451792e JA |
1746 | io_u_log_error(td, io_u); |
1747 | } | |
8b28bd41 | 1748 | if (icd->error) { |
f8b0bd10 JA |
1749 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1750 | ||
8b28bd41 DM |
1751 | if (!td_non_fatal_error(td, eb, icd->error)) |
1752 | return; | |
f8b0bd10 | 1753 | |
f2bba182 RR |
1754 | /* |
1755 | * If there is a non_fatal error, then add to the error count | |
1756 | * and clear all the errors. | |
1757 | */ | |
1758 | update_error_count(td, icd->error); | |
1759 | td_clear_error(td); | |
1760 | icd->error = 0; | |
f8b0bd10 JA |
1761 | if (io_u) |
1762 | io_u->error = 0; | |
f2bba182 | 1763 | } |
10ba535a JA |
1764 | } |
1765 | ||
9520ebb9 JA |
1766 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1767 | int nr) | |
10ba535a | 1768 | { |
6eaf09d6 | 1769 | int ddir; |
aba6c951 JA |
1770 | |
1771 | if (!gtod_reduce(td)) | |
9520ebb9 | 1772 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1773 | |
3af6ef39 JA |
1774 | icd->nr = nr; |
1775 | ||
10ba535a | 1776 | icd->error = 0; |
6eaf09d6 SL |
1777 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1778 | icd->bytes_done[ddir] = 0; | |
36167d82 JA |
1779 | } |
1780 | ||
97601024 JA |
1781 | static void ios_completed(struct thread_data *td, |
1782 | struct io_completion_data *icd) | |
36167d82 JA |
1783 | { |
1784 | struct io_u *io_u; | |
1785 | int i; | |
1786 | ||
10ba535a JA |
1787 | for (i = 0; i < icd->nr; i++) { |
1788 | io_u = td->io_ops->event(td, i); | |
1789 | ||
f8b0bd10 | 1790 | io_completed(td, &io_u, icd); |
e8462bd8 | 1791 | |
f8b0bd10 | 1792 | if (io_u) |
e8462bd8 | 1793 | put_io_u(td, io_u); |
10ba535a JA |
1794 | } |
1795 | } | |
97601024 | 1796 | |
e7e6cfb4 JA |
1797 | /* |
1798 | * Complete a single io_u for the sync engines. | |
1799 | */ | |
55312f9f | 1800 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
1801 | { |
1802 | struct io_completion_data icd; | |
55312f9f | 1803 | int ddir; |
97601024 | 1804 | |
9520ebb9 | 1805 | init_icd(td, &icd, 1); |
f8b0bd10 | 1806 | io_completed(td, &io_u, &icd); |
e8462bd8 | 1807 | |
f8b0bd10 | 1808 | if (io_u) |
e8462bd8 | 1809 | put_io_u(td, io_u); |
97601024 | 1810 | |
581e7141 JA |
1811 | if (icd.error) { |
1812 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1813 | return -1; | |
1814 | } | |
97601024 | 1815 | |
55312f9f JA |
1816 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1817 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1818 | |
1819 | return 0; | |
97601024 JA |
1820 | } |
1821 | ||
e7e6cfb4 JA |
1822 | /* |
1823 | * Called to complete min_events number of io for the async engines. | |
1824 | */ | |
55312f9f | 1825 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 1826 | { |
97601024 | 1827 | struct io_completion_data icd; |
00de55ef | 1828 | struct timespec *tvp = NULL; |
55312f9f | 1829 | int ret, ddir; |
4d06a338 | 1830 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1831 | |
4950421a | 1832 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1833 | |
4950421a | 1834 | if (!min_evts) |
00de55ef | 1835 | tvp = &ts; |
5fb4b366 RE |
1836 | else if (min_evts > td->cur_depth) |
1837 | min_evts = td->cur_depth; | |
97601024 | 1838 | |
82407585 RP |
1839 | /* No worries, td_io_getevents fixes min and max if they are |
1840 | * set incorrectly */ | |
1841 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 1842 | if (ret < 0) { |
e1161c32 | 1843 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1844 | return ret; |
1845 | } else if (!ret) | |
1846 | return ret; | |
1847 | ||
9520ebb9 | 1848 | init_icd(td, &icd, ret); |
97601024 | 1849 | ios_completed(td, &icd); |
581e7141 JA |
1850 | if (icd.error) { |
1851 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1852 | return -1; | |
1853 | } | |
97601024 | 1854 | |
55312f9f JA |
1855 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) |
1856 | td->bytes_done[ddir] += icd.bytes_done[ddir]; | |
581e7141 JA |
1857 | |
1858 | return 0; | |
97601024 | 1859 | } |
7e77dd02 JA |
1860 | |
1861 | /* | |
1862 | * Call when io_u is really queued, to update the submission latency. | |
1863 | */ | |
1864 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1865 | { | |
9520ebb9 JA |
1866 | if (!td->o.disable_slat) { |
1867 | unsigned long slat_time; | |
7e77dd02 | 1868 | |
9520ebb9 | 1869 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
1870 | |
1871 | if (td->parent) | |
1872 | td = td->parent; | |
1873 | ||
ae588852 JA |
1874 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
1875 | io_u->offset); | |
9520ebb9 | 1876 | } |
7e77dd02 | 1877 | } |
433afcb4 | 1878 | |
5c94b008 JA |
1879 | /* |
1880 | * See if we should reuse the last seed, if dedupe is enabled | |
1881 | */ | |
9451b93e | 1882 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 | 1883 | { |
c3546b53 | 1884 | uint64_t frand_max; |
5c94b008 JA |
1885 | unsigned int v; |
1886 | unsigned long r; | |
1887 | ||
1888 | if (!td->o.dedupe_percentage) | |
1889 | return &td->buf_state; | |
732eedd0 | 1890 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
1891 | frand_copy(&td->buf_state_prev, &td->buf_state); |
1892 | return &td->buf_state; | |
732eedd0 | 1893 | } |
5c94b008 | 1894 | |
c3546b53 | 1895 | frand_max = rand_max(&td->dedupe_state); |
5c94b008 | 1896 | r = __rand(&td->dedupe_state); |
c3546b53 | 1897 | v = 1 + (int) (100.0 * (r / (frand_max + 1.0))); |
5c94b008 JA |
1898 | |
1899 | if (v <= td->o.dedupe_percentage) | |
1900 | return &td->buf_state_prev; | |
1901 | ||
1902 | return &td->buf_state; | |
1903 | } | |
1904 | ||
9451b93e | 1905 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 1906 | { |
9451b93e JA |
1907 | if (td->o.dedupe_percentage == 100) |
1908 | frand_copy(rs, &td->buf_state_prev); | |
1909 | else if (rs == &td->buf_state) | |
5c94b008 JA |
1910 | frand_copy(&td->buf_state_prev, rs); |
1911 | } | |
1912 | ||
cc86c395 JA |
1913 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
1914 | unsigned int max_bs) | |
5973cafb | 1915 | { |
d1af2894 JA |
1916 | struct thread_options *o = &td->o; |
1917 | ||
4eff3e57 | 1918 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 1919 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 1920 | struct frand_state *rs; |
1066358a | 1921 | unsigned int left = max_bs; |
5c94b008 | 1922 | |
1066358a | 1923 | do { |
9451b93e | 1924 | rs = get_buf_state(td); |
9c42684e | 1925 | |
1066358a | 1926 | min_write = min(min_write, left); |
f97a43a1 | 1927 | |
1066358a JA |
1928 | if (perc) { |
1929 | unsigned int seg = min_write; | |
cc86c395 | 1930 | |
1066358a JA |
1931 | seg = min(min_write, td->o.compress_chunk); |
1932 | if (!seg) | |
1933 | seg = min_write; | |
1934 | ||
1935 | fill_random_buf_percentage(rs, buf, perc, seg, | |
d1af2894 JA |
1936 | min_write, o->buffer_pattern, |
1937 | o->buffer_pattern_bytes); | |
1066358a JA |
1938 | } else |
1939 | fill_random_buf(rs, buf, min_write); | |
1940 | ||
1941 | buf += min_write; | |
1942 | left -= min_write; | |
9451b93e | 1943 | save_buf_state(td, rs); |
1066358a | 1944 | } while (left); |
d1af2894 JA |
1945 | } else if (o->buffer_pattern_bytes) |
1946 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 1947 | else if (o->zero_buffers) |
cc86c395 | 1948 | memset(buf, 0, max_bs); |
999d245e | 1949 | else |
9451b93e | 1950 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
1951 | } |
1952 | ||
1953 | /* | |
1954 | * "randomly" fill the buffer contents | |
1955 | */ | |
1956 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1957 | unsigned int min_write, unsigned int max_bs) | |
1958 | { | |
1959 | io_u->buf_filled_len = 0; | |
1960 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 1961 | } |