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" |
10ba535a | 13 | |
97601024 JA |
14 | struct io_completion_data { |
15 | int nr; /* input */ | |
c8eeb9df | 16 | int account; /* input */ |
97601024 JA |
17 | |
18 | int error; /* output */ | |
19 | unsigned long bytes_done[2]; /* output */ | |
20 | struct timeval time; /* output */ | |
21 | }; | |
22 | ||
10ba535a JA |
23 | /* |
24 | * The ->file_map[] contains a map of blocks we have or have not done io | |
25 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
26 | */ | |
aec2de20 | 27 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 28 | { |
aec2de20 JA |
29 | unsigned int idx = RAND_MAP_IDX(f, block); |
30 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 31 | |
84422acd JA |
32 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
33 | ||
0ce8b119 | 34 | return (f->file_map[idx] & (1UL << bit)) == 0; |
10ba535a JA |
35 | } |
36 | ||
df415585 JA |
37 | /* |
38 | * Mark a given offset as used in the map. | |
39 | */ | |
9bf2061e | 40 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 41 | { |
2dc1bbeb | 42 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 43 | struct fio_file *f = io_u->file; |
a00735e6 | 44 | unsigned long long block; |
3e3357b1 | 45 | unsigned int blocks, nr_blocks; |
38dad62d | 46 | int busy_check; |
df415585 | 47 | |
b9c5b644 | 48 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
c685b5b2 | 49 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
3e3357b1 | 50 | blocks = 0; |
38dad62d | 51 | busy_check = !(io_u->flags & IO_U_F_BUSY_OK); |
c685b5b2 | 52 | |
3e3357b1 | 53 | while (nr_blocks) { |
df415585 | 54 | unsigned int idx, bit; |
0ce8b119 | 55 | unsigned long mask, this_blocks; |
df415585 | 56 | |
1e3d53ac JA |
57 | /* |
58 | * If we have a mixed random workload, we may | |
59 | * encounter blocks we already did IO to. | |
60 | */ | |
38dad62d JA |
61 | if (!busy_check) { |
62 | blocks = nr_blocks; | |
63 | break; | |
64 | } | |
5736c10d | 65 | if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block)) |
df415585 JA |
66 | break; |
67 | ||
aec2de20 JA |
68 | idx = RAND_MAP_IDX(f, block); |
69 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 70 | |
0032bf9f | 71 | fio_assert(td, idx < f->num_maps); |
df415585 | 72 | |
3e3357b1 JA |
73 | this_blocks = nr_blocks; |
74 | if (this_blocks + bit > BLOCKS_PER_MAP) | |
75 | this_blocks = BLOCKS_PER_MAP - bit; | |
76 | ||
da3758bf JA |
77 | do { |
78 | if (this_blocks == BLOCKS_PER_MAP) | |
0ce8b119 | 79 | mask = -1UL; |
da3758bf | 80 | else |
0ce8b119 | 81 | mask = ((1UL << this_blocks) - 1) << bit; |
da3758bf JA |
82 | |
83 | if (!(f->file_map[idx] & mask)) | |
84 | break; | |
85 | ||
86 | this_blocks--; | |
87 | } while (this_blocks); | |
88 | ||
89 | if (!this_blocks) | |
90 | break; | |
3e3357b1 | 91 | |
3e3357b1 JA |
92 | f->file_map[idx] |= mask; |
93 | nr_blocks -= this_blocks; | |
94 | blocks += this_blocks; | |
c9dd34b2 | 95 | block += this_blocks; |
df415585 JA |
96 | } |
97 | ||
a00735e6 JA |
98 | if ((blocks * min_bs) < io_u->buflen) |
99 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
100 | } |
101 | ||
3e3357b1 JA |
102 | static unsigned long long last_block(struct thread_data *td, struct fio_file *f, |
103 | enum fio_ddir ddir) | |
2ba1c290 JA |
104 | { |
105 | unsigned long long max_blocks; | |
d9dd70f7 | 106 | unsigned long long max_size; |
2ba1c290 | 107 | |
ff58fced JA |
108 | assert(ddir_rw(ddir)); |
109 | ||
d9dd70f7 JA |
110 | /* |
111 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
112 | */ | |
113 | max_size = f->io_size; | |
114 | if (max_size > f->real_file_size) | |
115 | max_size = f->real_file_size; | |
116 | ||
2b7a01d0 | 117 | max_blocks = max_size / (unsigned long long) td->o.ba[ddir]; |
2ba1c290 JA |
118 | if (!max_blocks) |
119 | return 0; | |
120 | ||
67778e88 | 121 | return max_blocks; |
2ba1c290 JA |
122 | } |
123 | ||
10ba535a JA |
124 | /* |
125 | * Return the next free block in the map. | |
126 | */ | |
127 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 128 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 129 | { |
0ce8b119 | 130 | unsigned long long block, min_bs = td->o.rw_min_bs, lastb; |
10ba535a JA |
131 | int i; |
132 | ||
5e0baa7f JA |
133 | lastb = last_block(td, f, ddir); |
134 | if (!lastb) | |
135 | return 1; | |
136 | ||
c685b5b2 | 137 | i = f->last_free_lookup; |
0ce8b119 JA |
138 | block = i * BLOCKS_PER_MAP; |
139 | while (block * min_bs < f->real_file_size && | |
140 | block * min_bs < f->io_size) { | |
141 | if (f->file_map[i] != -1UL) { | |
142 | block += ffz(f->file_map[i]); | |
143 | if (block > lastb) | |
2ba1c290 | 144 | break; |
c685b5b2 | 145 | f->last_free_lookup = i; |
0ce8b119 | 146 | *b = block; |
10ba535a JA |
147 | return 0; |
148 | } | |
149 | ||
0ce8b119 | 150 | block += BLOCKS_PER_MAP; |
10ba535a JA |
151 | i++; |
152 | } | |
153 | ||
2ba1c290 | 154 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
155 | return 1; |
156 | } | |
157 | ||
ec4015da | 158 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 159 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 160 | { |
5e0baa7f | 161 | unsigned long long r, lastb; |
ec4015da JA |
162 | int loops = 5; |
163 | ||
5e0baa7f JA |
164 | lastb = last_block(td, f, ddir); |
165 | if (!lastb) | |
166 | return 1; | |
167 | ||
0ce8b119 JA |
168 | if (f->failed_rands >= 200) |
169 | goto ffz; | |
170 | ||
ec4015da | 171 | do { |
2615cc4b JA |
172 | if (td->o.use_os_rand) { |
173 | r = os_random_long(&td->random_state); | |
174 | *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0)); | |
175 | } else { | |
176 | r = __rand(&td->__random_state); | |
177 | *b = (lastb - 1) * (r / ((unsigned long long) FRAND_MAX + 1.0)); | |
178 | } | |
179 | ||
84422acd | 180 | dprint(FD_RANDOM, "off rand %llu\n", r); |
2615cc4b | 181 | |
2ba1c290 | 182 | |
43c63a78 JA |
183 | /* |
184 | * if we are not maintaining a random map, we are done. | |
185 | */ | |
303032ae | 186 | if (!file_randommap(td, f)) |
0ce8b119 | 187 | goto ret_good; |
43c63a78 JA |
188 | |
189 | /* | |
84422acd | 190 | * calculate map offset and check if it's free |
43c63a78 | 191 | */ |
aec2de20 | 192 | if (random_map_free(f, *b)) |
0ce8b119 | 193 | goto ret_good; |
43c63a78 | 194 | |
84422acd JA |
195 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
196 | *b); | |
43c63a78 | 197 | } while (--loops); |
ec4015da | 198 | |
0ce8b119 JA |
199 | if (!f->failed_rands++) |
200 | f->last_free_lookup = 0; | |
201 | ||
ec4015da | 202 | /* |
43c63a78 JA |
203 | * we get here, if we didn't suceed in looking up a block. generate |
204 | * a random start offset into the filemap, and find the first free | |
205 | * block from there. | |
ec4015da | 206 | */ |
43c63a78 JA |
207 | loops = 10; |
208 | do { | |
dc873b6f JA |
209 | f->last_free_lookup = (f->num_maps - 1) * |
210 | (r / (OS_RAND_MAX + 1.0)); | |
4ba66134 | 211 | if (!get_next_free_block(td, f, ddir, b)) |
0ce8b119 | 212 | goto ret; |
ec4015da | 213 | |
d1cea521 JA |
214 | if (td->o.use_os_rand) |
215 | r = os_random_long(&td->random_state); | |
216 | else | |
217 | r = __rand(&td->__random_state); | |
43c63a78 JA |
218 | } while (--loops); |
219 | ||
220 | /* | |
221 | * that didn't work either, try exhaustive search from the start | |
222 | */ | |
223 | f->last_free_lookup = 0; | |
0ce8b119 JA |
224 | ffz: |
225 | if (!get_next_free_block(td, f, ddir, b)) | |
226 | return 0; | |
227 | f->last_free_lookup = 0; | |
4ba66134 | 228 | return get_next_free_block(td, f, ddir, b); |
0ce8b119 JA |
229 | ret_good: |
230 | f->failed_rands = 0; | |
231 | ret: | |
232 | return 0; | |
ec4015da JA |
233 | } |
234 | ||
38dad62d JA |
235 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
236 | enum fio_ddir ddir, unsigned long long *b) | |
237 | { | |
238 | if (get_next_rand_offset(td, f, ddir, b)) { | |
239 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", | |
240 | f->file_name, f->last_pos, f->real_file_size); | |
241 | return 1; | |
242 | } | |
243 | ||
244 | return 0; | |
245 | } | |
246 | ||
247 | static int get_next_seq_block(struct thread_data *td, struct fio_file *f, | |
248 | enum fio_ddir ddir, unsigned long long *b) | |
249 | { | |
ff58fced JA |
250 | assert(ddir_rw(ddir)); |
251 | ||
38dad62d | 252 | if (f->last_pos < f->real_file_size) { |
a66da7a2 | 253 | unsigned long long pos; |
059b0802 | 254 | |
a66da7a2 JA |
255 | if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0) |
256 | f->last_pos = f->real_file_size; | |
257 | ||
258 | pos = f->last_pos - f->file_offset; | |
059b0802 JA |
259 | if (pos) |
260 | pos += td->o.ddir_seq_add; | |
261 | ||
262 | *b = pos / td->o.min_bs[ddir]; | |
38dad62d JA |
263 | return 0; |
264 | } | |
265 | ||
266 | return 1; | |
267 | } | |
268 | ||
269 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
270 | enum fio_ddir ddir, int rw_seq, unsigned long long *b) | |
271 | { | |
272 | struct fio_file *f = io_u->file; | |
273 | int ret; | |
274 | ||
ff58fced JA |
275 | assert(ddir_rw(ddir)); |
276 | ||
38dad62d JA |
277 | if (rw_seq) { |
278 | if (td_random(td)) | |
279 | ret = get_next_rand_block(td, f, ddir, b); | |
280 | else | |
281 | ret = get_next_seq_block(td, f, ddir, b); | |
282 | } else { | |
283 | io_u->flags |= IO_U_F_BUSY_OK; | |
284 | ||
285 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
286 | ret = get_next_seq_block(td, f, ddir, b); | |
287 | if (ret) | |
288 | ret = get_next_rand_block(td, f, ddir, b); | |
289 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { | |
290 | if (f->last_start != -1ULL) | |
0508320e DN |
291 | *b = (f->last_start - f->file_offset) |
292 | / td->o.min_bs[ddir]; | |
38dad62d JA |
293 | else |
294 | *b = 0; | |
295 | ret = 0; | |
296 | } else { | |
297 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
298 | ret = 1; | |
299 | } | |
300 | } | |
301 | ||
302 | return ret; | |
303 | } | |
304 | ||
10ba535a JA |
305 | /* |
306 | * For random io, generate a random new block and see if it's used. Repeat | |
307 | * until we find a free one. For sequential io, just return the end of | |
308 | * the last io issued. | |
309 | */ | |
15dc1934 | 310 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 311 | { |
9bf2061e | 312 | struct fio_file *f = io_u->file; |
ec4015da | 313 | unsigned long long b; |
4ba66134 | 314 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 315 | int rw_seq_hit = 0; |
10ba535a | 316 | |
ff58fced JA |
317 | assert(ddir_rw(ddir)); |
318 | ||
38dad62d JA |
319 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
320 | rw_seq_hit = 1; | |
5736c10d | 321 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 322 | } |
211097b2 | 323 | |
0508320e | 324 | if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) |
38dad62d | 325 | return 1; |
10ba535a | 326 | |
2b7a01d0 | 327 | io_u->offset = b * td->o.ba[ddir]; |
009bd847 JA |
328 | if (io_u->offset >= f->io_size) { |
329 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
330 | io_u->offset, f->io_size); | |
331 | return 1; | |
332 | } | |
333 | ||
334 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
335 | if (io_u->offset >= f->real_file_size) { |
336 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
337 | io_u->offset, f->real_file_size); | |
10ba535a | 338 | return 1; |
2ba1c290 | 339 | } |
10ba535a JA |
340 | |
341 | return 0; | |
342 | } | |
343 | ||
15dc1934 JA |
344 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
345 | { | |
7eb36574 JA |
346 | struct prof_io_ops *ops = &td->prof_io_ops; |
347 | ||
348 | if (ops->fill_io_u_off) | |
349 | return ops->fill_io_u_off(td, io_u); | |
15dc1934 JA |
350 | |
351 | return __get_next_offset(td, io_u); | |
352 | } | |
353 | ||
79944128 JA |
354 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
355 | unsigned int buflen) | |
356 | { | |
357 | struct fio_file *f = io_u->file; | |
358 | ||
359 | return io_u->offset + buflen <= f->io_size + td->o.start_offset; | |
360 | } | |
361 | ||
15dc1934 | 362 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 363 | { |
bca4ed4d | 364 | const int ddir = io_u->ddir; |
f3f552b9 | 365 | unsigned int uninitialized_var(buflen); |
f3059de1 | 366 | unsigned int minbs, maxbs; |
1294c3ec | 367 | unsigned long r, rand_max; |
10ba535a | 368 | |
ff58fced JA |
369 | assert(ddir_rw(ddir)); |
370 | ||
f3059de1 JA |
371 | minbs = td->o.min_bs[ddir]; |
372 | maxbs = td->o.max_bs[ddir]; | |
373 | ||
79944128 JA |
374 | if (minbs == maxbs) |
375 | return minbs; | |
376 | ||
4c07ad86 JA |
377 | if (td->o.use_os_rand) |
378 | rand_max = OS_RAND_MAX; | |
379 | else | |
380 | rand_max = FRAND_MAX; | |
381 | ||
79944128 | 382 | do { |
4c07ad86 JA |
383 | if (td->o.use_os_rand) |
384 | r = os_random_long(&td->bsrange_state); | |
385 | else | |
386 | r = __rand(&td->__bsrange_state); | |
387 | ||
720e84ad | 388 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 389 | buflen = 1 + (unsigned int) ((double) maxbs * |
4c07ad86 | 390 | (r / (rand_max + 1.0))); |
f3059de1 JA |
391 | if (buflen < minbs) |
392 | buflen = minbs; | |
5ec10eaa | 393 | } else { |
564ca972 JA |
394 | long perc = 0; |
395 | unsigned int i; | |
396 | ||
720e84ad JA |
397 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
398 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
399 | |
400 | buflen = bsp->bs; | |
401 | perc += bsp->perc; | |
79944128 JA |
402 | if ((r <= ((rand_max / 100L) * perc)) && |
403 | io_u_fits(td, io_u, buflen)) | |
564ca972 JA |
404 | break; |
405 | } | |
406 | } | |
79944128 | 407 | |
f3059de1 JA |
408 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
409 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 410 | |
79944128 | 411 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 412 | |
10ba535a JA |
413 | return buflen; |
414 | } | |
415 | ||
15dc1934 JA |
416 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
417 | { | |
7eb36574 JA |
418 | struct prof_io_ops *ops = &td->prof_io_ops; |
419 | ||
420 | if (ops->fill_io_u_size) | |
421 | return ops->fill_io_u_size(td, io_u); | |
15dc1934 JA |
422 | |
423 | return __get_next_buflen(td, io_u); | |
424 | } | |
425 | ||
afe24a5a JA |
426 | static void set_rwmix_bytes(struct thread_data *td) |
427 | { | |
afe24a5a JA |
428 | unsigned int diff; |
429 | ||
430 | /* | |
431 | * we do time or byte based switch. this is needed because | |
432 | * buffered writes may issue a lot quicker than they complete, | |
433 | * whereas reads do not. | |
434 | */ | |
e47f799f | 435 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 436 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
437 | } |
438 | ||
439 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
440 | { | |
441 | unsigned int v; | |
1294c3ec | 442 | unsigned long r; |
e47f799f | 443 | |
4c07ad86 JA |
444 | if (td->o.use_os_rand) { |
445 | r = os_random_long(&td->rwmix_state); | |
446 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); | |
447 | } else { | |
448 | r = __rand(&td->__rwmix_state); | |
449 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
450 | } | |
451 | ||
04c540d9 | 452 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
453 | return DDIR_READ; |
454 | ||
455 | return DDIR_WRITE; | |
afe24a5a JA |
456 | } |
457 | ||
581e7141 JA |
458 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
459 | { | |
460 | enum fio_ddir odir = ddir ^ 1; | |
461 | struct timeval t; | |
462 | long usec; | |
463 | ||
ff58fced JA |
464 | assert(ddir_rw(ddir)); |
465 | ||
581e7141 JA |
466 | if (td->rate_pending_usleep[ddir] <= 0) |
467 | return ddir; | |
468 | ||
469 | /* | |
470 | * We have too much pending sleep in this direction. See if we | |
471 | * should switch. | |
472 | */ | |
473 | if (td_rw(td)) { | |
474 | /* | |
475 | * Other direction does not have too much pending, switch | |
476 | */ | |
477 | if (td->rate_pending_usleep[odir] < 100000) | |
478 | return odir; | |
479 | ||
480 | /* | |
481 | * Both directions have pending sleep. Sleep the minimum time | |
482 | * and deduct from both. | |
483 | */ | |
484 | if (td->rate_pending_usleep[ddir] <= | |
485 | td->rate_pending_usleep[odir]) { | |
486 | usec = td->rate_pending_usleep[ddir]; | |
487 | } else { | |
488 | usec = td->rate_pending_usleep[odir]; | |
489 | ddir = odir; | |
490 | } | |
491 | } else | |
492 | usec = td->rate_pending_usleep[ddir]; | |
493 | ||
78c1eda5 JA |
494 | /* |
495 | * We are going to sleep, ensure that we flush anything pending as | |
496 | * not to skew our latency numbers | |
497 | */ | |
498 | if (td->cur_depth) { | |
499 | int fio_unused ret; | |
500 | ||
501 | ret = io_u_queued_complete(td, td->cur_depth, NULL); | |
502 | } | |
503 | ||
581e7141 JA |
504 | fio_gettime(&t, NULL); |
505 | usec_sleep(td, usec); | |
506 | usec = utime_since_now(&t); | |
507 | ||
508 | td->rate_pending_usleep[ddir] -= usec; | |
509 | ||
510 | odir = ddir ^ 1; | |
511 | if (td_rw(td) && __should_check_rate(td, odir)) | |
512 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 513 | |
581e7141 JA |
514 | return ddir; |
515 | } | |
516 | ||
10ba535a JA |
517 | /* |
518 | * Return the data direction for the next io_u. If the job is a | |
519 | * mixed read/write workload, check the rwmix cycle and switch if | |
520 | * necessary. | |
521 | */ | |
1e97cce9 | 522 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 523 | { |
581e7141 JA |
524 | enum fio_ddir ddir; |
525 | ||
5f9099ea JA |
526 | /* |
527 | * see if it's time to fsync | |
528 | */ | |
529 | if (td->o.fsync_blocks && | |
530 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
531 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
532 | return DDIR_SYNC; | |
533 | ||
534 | /* | |
535 | * see if it's time to fdatasync | |
536 | */ | |
537 | if (td->o.fdatasync_blocks && | |
538 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
539 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
540 | return DDIR_DATASYNC; | |
541 | ||
44f29692 JA |
542 | /* |
543 | * see if it's time to sync_file_range | |
544 | */ | |
545 | if (td->sync_file_range_nr && | |
546 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
547 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
548 | return DDIR_SYNC_FILE_RANGE; | |
549 | ||
10ba535a | 550 | if (td_rw(td)) { |
10ba535a JA |
551 | /* |
552 | * Check if it's time to seed a new data direction. | |
553 | */ | |
e4928662 | 554 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
555 | /* |
556 | * Put a top limit on how many bytes we do for | |
557 | * one data direction, to avoid overflowing the | |
558 | * ranges too much | |
559 | */ | |
560 | ddir = get_rand_ddir(td); | |
e47f799f JA |
561 | |
562 | if (ddir != td->rwmix_ddir) | |
563 | set_rwmix_bytes(td); | |
564 | ||
565 | td->rwmix_ddir = ddir; | |
10ba535a | 566 | } |
581e7141 | 567 | ddir = td->rwmix_ddir; |
10ba535a | 568 | } else if (td_read(td)) |
581e7141 | 569 | ddir = DDIR_READ; |
10ba535a | 570 | else |
581e7141 JA |
571 | ddir = DDIR_WRITE; |
572 | ||
573 | td->rwmix_ddir = rate_ddir(td, ddir); | |
574 | return td->rwmix_ddir; | |
10ba535a JA |
575 | } |
576 | ||
1ef2b6be JA |
577 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
578 | { | |
579 | io_u->ddir = get_rw_ddir(td); | |
580 | ||
581 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
582 | td->o.barrier_blocks && | |
583 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
584 | td->io_issues[DDIR_WRITE]) | |
585 | io_u->flags |= IO_U_F_BARRIER; | |
586 | } | |
587 | ||
e8462bd8 | 588 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 JA |
589 | { |
590 | int ret = put_file(td, f); | |
591 | ||
592 | if (ret) | |
593 | td_verror(td, ret, "file close"); | |
594 | } | |
595 | ||
10ba535a JA |
596 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
597 | { | |
e8462bd8 JA |
598 | td_io_u_lock(td); |
599 | ||
0c6e7517 | 600 | io_u->flags |= IO_U_F_FREE; |
e8462bd8 | 601 | io_u->flags &= ~IO_U_F_FREE_DEF; |
0c6e7517 | 602 | |
60f2c658 JA |
603 | if (io_u->file) |
604 | put_file_log(td, io_u->file); | |
2dbdab7e | 605 | |
10ba535a | 606 | io_u->file = NULL; |
0c41214f RR |
607 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
608 | td->cur_depth--; | |
e8462bd8 | 609 | flist_del_init(&io_u->list); |
01743ee1 | 610 | flist_add(&io_u->list, &td->io_u_freelist); |
e8462bd8 JA |
611 | td_io_u_unlock(td); |
612 | td_io_u_free_notify(td); | |
10ba535a JA |
613 | } |
614 | ||
f2bba182 RR |
615 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
616 | { | |
617 | io_u->flags &= ~IO_U_F_FLIGHT; | |
618 | put_io_u(td, io_u); | |
619 | } | |
620 | ||
755200a3 JA |
621 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
622 | { | |
623 | struct io_u *__io_u = *io_u; | |
624 | ||
465221b0 JA |
625 | dprint(FD_IO, "requeue %p\n", __io_u); |
626 | ||
e8462bd8 JA |
627 | td_io_u_lock(td); |
628 | ||
4d2e0f49 | 629 | __io_u->flags |= IO_U_F_FREE; |
ff58fced | 630 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir)) |
e4f54adb | 631 | td->io_issues[__io_u->ddir]--; |
5ec10eaa | 632 | |
4d2e0f49 | 633 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
634 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
635 | td->cur_depth--; | |
01743ee1 JA |
636 | flist_del(&__io_u->list); |
637 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
e8462bd8 | 638 | td_io_u_unlock(td); |
755200a3 JA |
639 | *io_u = NULL; |
640 | } | |
641 | ||
9bf2061e | 642 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 643 | { |
b4c5e1ac JA |
644 | if (td->io_ops->flags & FIO_NOIO) |
645 | goto out; | |
646 | ||
1ef2b6be | 647 | set_rw_ddir(td, io_u); |
5f9099ea | 648 | |
87dc1ab1 | 649 | /* |
ff58fced | 650 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 651 | */ |
ff58fced | 652 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 653 | goto out; |
a00735e6 | 654 | |
48f5abd3 JA |
655 | /* |
656 | * See if it's time to switch to a new zone | |
657 | */ | |
658 | if (td->zone_bytes >= td->o.zone_size) { | |
659 | td->zone_bytes = 0; | |
660 | io_u->file->last_pos += td->o.zone_skip; | |
661 | td->io_skip_bytes += td->o.zone_skip; | |
662 | } | |
663 | ||
10ba535a | 664 | /* |
c685b5b2 JA |
665 | * No log, let the seq/rand engine retrieve the next buflen and |
666 | * position. | |
10ba535a | 667 | */ |
2ba1c290 JA |
668 | if (get_next_offset(td, io_u)) { |
669 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 670 | return 1; |
2ba1c290 | 671 | } |
10ba535a | 672 | |
9bf2061e | 673 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
674 | if (!io_u->buflen) { |
675 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 676 | return 1; |
2ba1c290 | 677 | } |
bca4ed4d | 678 | |
2ba1c290 JA |
679 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
680 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
681 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
682 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 683 | return 1; |
2ba1c290 | 684 | } |
6a5e6884 | 685 | |
bca4ed4d JA |
686 | /* |
687 | * mark entry before potentially trimming io_u | |
688 | */ | |
303032ae | 689 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 690 | mark_random_map(td, io_u); |
bca4ed4d JA |
691 | |
692 | /* | |
693 | * If using a write iolog, store this entry. | |
694 | */ | |
c38e9468 | 695 | out: |
2ba1c290 | 696 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 697 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 698 | log_io_u(td, io_u); |
bca4ed4d | 699 | return 0; |
10ba535a JA |
700 | } |
701 | ||
838bc709 JA |
702 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
703 | { | |
2b13e716 | 704 | int idx = 0; |
838bc709 JA |
705 | |
706 | switch (nr) { | |
707 | default: | |
2b13e716 | 708 | idx = 6; |
838bc709 JA |
709 | break; |
710 | case 33 ... 64: | |
2b13e716 | 711 | idx = 5; |
838bc709 JA |
712 | break; |
713 | case 17 ... 32: | |
2b13e716 | 714 | idx = 4; |
838bc709 JA |
715 | break; |
716 | case 9 ... 16: | |
2b13e716 | 717 | idx = 3; |
838bc709 JA |
718 | break; |
719 | case 5 ... 8: | |
2b13e716 | 720 | idx = 2; |
838bc709 JA |
721 | break; |
722 | case 1 ... 4: | |
2b13e716 | 723 | idx = 1; |
838bc709 JA |
724 | case 0: |
725 | break; | |
726 | } | |
727 | ||
2b13e716 | 728 | map[idx]++; |
838bc709 JA |
729 | } |
730 | ||
731 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
732 | { | |
733 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
734 | td->ts.total_submit++; | |
735 | } | |
736 | ||
737 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
738 | { | |
739 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
740 | td->ts.total_complete++; | |
741 | } | |
742 | ||
d8005759 | 743 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 744 | { |
2b13e716 | 745 | int idx = 0; |
71619dc2 JA |
746 | |
747 | switch (td->cur_depth) { | |
748 | default: | |
2b13e716 | 749 | idx = 6; |
a783e61a | 750 | break; |
71619dc2 | 751 | case 32 ... 63: |
2b13e716 | 752 | idx = 5; |
a783e61a | 753 | break; |
71619dc2 | 754 | case 16 ... 31: |
2b13e716 | 755 | idx = 4; |
a783e61a | 756 | break; |
71619dc2 | 757 | case 8 ... 15: |
2b13e716 | 758 | idx = 3; |
a783e61a | 759 | break; |
71619dc2 | 760 | case 4 ... 7: |
2b13e716 | 761 | idx = 2; |
a783e61a | 762 | break; |
71619dc2 | 763 | case 2 ... 3: |
2b13e716 | 764 | idx = 1; |
71619dc2 JA |
765 | case 1: |
766 | break; | |
767 | } | |
768 | ||
2b13e716 | 769 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
770 | } |
771 | ||
04a0feae JA |
772 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
773 | { | |
2b13e716 | 774 | int idx = 0; |
04a0feae JA |
775 | |
776 | assert(usec < 1000); | |
777 | ||
778 | switch (usec) { | |
779 | case 750 ... 999: | |
2b13e716 | 780 | idx = 9; |
04a0feae JA |
781 | break; |
782 | case 500 ... 749: | |
2b13e716 | 783 | idx = 8; |
04a0feae JA |
784 | break; |
785 | case 250 ... 499: | |
2b13e716 | 786 | idx = 7; |
04a0feae JA |
787 | break; |
788 | case 100 ... 249: | |
2b13e716 | 789 | idx = 6; |
04a0feae JA |
790 | break; |
791 | case 50 ... 99: | |
2b13e716 | 792 | idx = 5; |
04a0feae JA |
793 | break; |
794 | case 20 ... 49: | |
2b13e716 | 795 | idx = 4; |
04a0feae JA |
796 | break; |
797 | case 10 ... 19: | |
2b13e716 | 798 | idx = 3; |
04a0feae JA |
799 | break; |
800 | case 4 ... 9: | |
2b13e716 | 801 | idx = 2; |
04a0feae JA |
802 | break; |
803 | case 2 ... 3: | |
2b13e716 | 804 | idx = 1; |
04a0feae JA |
805 | case 0 ... 1: |
806 | break; | |
807 | } | |
808 | ||
2b13e716 JA |
809 | assert(idx < FIO_IO_U_LAT_U_NR); |
810 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
811 | } |
812 | ||
813 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 814 | { |
2b13e716 | 815 | int idx = 0; |
ec118304 JA |
816 | |
817 | switch (msec) { | |
818 | default: | |
2b13e716 | 819 | idx = 11; |
04a0feae | 820 | break; |
8abdce66 | 821 | case 1000 ... 1999: |
2b13e716 | 822 | idx = 10; |
04a0feae | 823 | break; |
8abdce66 | 824 | case 750 ... 999: |
2b13e716 | 825 | idx = 9; |
04a0feae | 826 | break; |
8abdce66 | 827 | case 500 ... 749: |
2b13e716 | 828 | idx = 8; |
04a0feae | 829 | break; |
8abdce66 | 830 | case 250 ... 499: |
2b13e716 | 831 | idx = 7; |
04a0feae | 832 | break; |
8abdce66 | 833 | case 100 ... 249: |
2b13e716 | 834 | idx = 6; |
04a0feae | 835 | break; |
8abdce66 | 836 | case 50 ... 99: |
2b13e716 | 837 | idx = 5; |
04a0feae | 838 | break; |
8abdce66 | 839 | case 20 ... 49: |
2b13e716 | 840 | idx = 4; |
04a0feae | 841 | break; |
8abdce66 | 842 | case 10 ... 19: |
2b13e716 | 843 | idx = 3; |
04a0feae | 844 | break; |
8abdce66 | 845 | case 4 ... 9: |
2b13e716 | 846 | idx = 2; |
04a0feae | 847 | break; |
ec118304 | 848 | case 2 ... 3: |
2b13e716 | 849 | idx = 1; |
ec118304 JA |
850 | case 0 ... 1: |
851 | break; | |
852 | } | |
853 | ||
2b13e716 JA |
854 | assert(idx < FIO_IO_U_LAT_M_NR); |
855 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
856 | } |
857 | ||
858 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
859 | { | |
860 | if (usec < 1000) | |
861 | io_u_mark_lat_usec(td, usec); | |
862 | else | |
863 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
864 | } |
865 | ||
0aabe160 JA |
866 | /* |
867 | * Get next file to service by choosing one at random | |
868 | */ | |
2cc52930 JA |
869 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
870 | enum fio_file_flags goodf, | |
d6aed795 | 871 | enum fio_file_flags badf) |
0aabe160 | 872 | { |
0aabe160 | 873 | struct fio_file *f; |
1c178180 | 874 | int fno; |
0aabe160 JA |
875 | |
876 | do { | |
87b10676 | 877 | int opened = 0; |
1294c3ec | 878 | unsigned long r; |
4c07ad86 JA |
879 | |
880 | if (td->o.use_os_rand) { | |
881 | r = os_random_long(&td->next_file_state); | |
882 | fno = (unsigned int) ((double) td->o.nr_files | |
883 | * (r / (OS_RAND_MAX + 1.0))); | |
884 | } else { | |
885 | r = __rand(&td->__next_file_state); | |
886 | fno = (unsigned int) ((double) td->o.nr_files | |
887 | * (r / (FRAND_MAX + 1.0))); | |
888 | } | |
7c83c089 | 889 | |
126d65c6 | 890 | f = td->files[fno]; |
d6aed795 | 891 | if (fio_file_done(f)) |
059e63c0 | 892 | continue; |
1c178180 | 893 | |
d6aed795 | 894 | if (!fio_file_open(f)) { |
87b10676 JA |
895 | int err; |
896 | ||
897 | err = td_io_open_file(td, f); | |
898 | if (err) | |
899 | continue; | |
900 | opened = 1; | |
901 | } | |
902 | ||
2ba1c290 JA |
903 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
904 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 905 | return f; |
2ba1c290 | 906 | } |
87b10676 JA |
907 | if (opened) |
908 | td_io_close_file(td, f); | |
0aabe160 JA |
909 | } while (1); |
910 | } | |
911 | ||
912 | /* | |
913 | * Get next file to service by doing round robin between all available ones | |
914 | */ | |
1c178180 JA |
915 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
916 | int badf) | |
3d7c391d JA |
917 | { |
918 | unsigned int old_next_file = td->next_file; | |
919 | struct fio_file *f; | |
920 | ||
921 | do { | |
87b10676 JA |
922 | int opened = 0; |
923 | ||
126d65c6 | 924 | f = td->files[td->next_file]; |
3d7c391d JA |
925 | |
926 | td->next_file++; | |
2dc1bbeb | 927 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
928 | td->next_file = 0; |
929 | ||
87b10676 | 930 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 931 | if (fio_file_done(f)) { |
d5ed68ea | 932 | f = NULL; |
059e63c0 | 933 | continue; |
d5ed68ea | 934 | } |
059e63c0 | 935 | |
d6aed795 | 936 | if (!fio_file_open(f)) { |
87b10676 JA |
937 | int err; |
938 | ||
939 | err = td_io_open_file(td, f); | |
b5696bfc JA |
940 | if (err) { |
941 | dprint(FD_FILE, "error %d on open of %s\n", | |
942 | err, f->file_name); | |
87c27b45 | 943 | f = NULL; |
87b10676 | 944 | continue; |
b5696bfc | 945 | } |
87b10676 JA |
946 | opened = 1; |
947 | } | |
948 | ||
0b9d69ec JA |
949 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
950 | f->flags); | |
1c178180 | 951 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
952 | break; |
953 | ||
87b10676 JA |
954 | if (opened) |
955 | td_io_close_file(td, f); | |
956 | ||
3d7c391d JA |
957 | f = NULL; |
958 | } while (td->next_file != old_next_file); | |
959 | ||
2ba1c290 | 960 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
961 | return f; |
962 | } | |
963 | ||
7eb36574 | 964 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 965 | { |
1907dbc6 JA |
966 | struct fio_file *f; |
967 | ||
2dc1bbeb | 968 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 969 | |
b5696bfc | 970 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
971 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
972 | " nr_files=%d\n", td->nr_open_files, | |
973 | td->nr_done_files, | |
974 | td->o.nr_files); | |
bdb4e2e9 | 975 | return NULL; |
2ba1c290 | 976 | } |
bdb4e2e9 | 977 | |
1907dbc6 | 978 | f = td->file_service_file; |
d6aed795 | 979 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
980 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
981 | goto out; | |
982 | if (td->file_service_left--) | |
983 | goto out; | |
984 | } | |
1907dbc6 | 985 | |
a086c257 JA |
986 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
987 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 988 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 989 | else |
d6aed795 | 990 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 JA |
991 | |
992 | td->file_service_file = f; | |
993 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 994 | out: |
683023e8 | 995 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 996 | return f; |
bdb4e2e9 JA |
997 | } |
998 | ||
7eb36574 JA |
999 | static struct fio_file *get_next_file(struct thread_data *td) |
1000 | { | |
1001 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1002 | ||
1003 | if (ops->get_next_file) | |
1004 | return ops->get_next_file(td); | |
1005 | ||
1006 | return __get_next_file(td); | |
1007 | } | |
1008 | ||
429f6675 JA |
1009 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
1010 | { | |
1011 | struct fio_file *f; | |
1012 | ||
1013 | do { | |
1014 | f = get_next_file(td); | |
1015 | if (!f) | |
1016 | return 1; | |
1017 | ||
429f6675 JA |
1018 | io_u->file = f; |
1019 | get_file(f); | |
1020 | ||
1021 | if (!fill_io_u(td, io_u)) | |
1022 | break; | |
1023 | ||
b5696bfc | 1024 | put_file_log(td, f); |
429f6675 | 1025 | td_io_close_file(td, f); |
b5696bfc | 1026 | io_u->file = NULL; |
d6aed795 | 1027 | fio_file_set_done(f); |
429f6675 | 1028 | td->nr_done_files++; |
0b9d69ec JA |
1029 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1030 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1031 | } while (1); |
1032 | ||
1033 | return 0; | |
1034 | } | |
1035 | ||
1036 | ||
10ba535a JA |
1037 | struct io_u *__get_io_u(struct thread_data *td) |
1038 | { | |
1039 | struct io_u *io_u = NULL; | |
1040 | ||
e8462bd8 JA |
1041 | td_io_u_lock(td); |
1042 | ||
1043 | again: | |
01743ee1 JA |
1044 | if (!flist_empty(&td->io_u_requeues)) |
1045 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 1046 | else if (!queue_full(td)) { |
01743ee1 | 1047 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 1048 | |
6040dabc | 1049 | io_u->buflen = 0; |
10ba535a | 1050 | io_u->resid = 0; |
755200a3 | 1051 | io_u->file = NULL; |
d7762cf8 | 1052 | io_u->end_io = NULL; |
755200a3 JA |
1053 | } |
1054 | ||
1055 | if (io_u) { | |
0c6e7517 | 1056 | assert(io_u->flags & IO_U_F_FREE); |
2ecc1b57 | 1057 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF); |
1ef2b6be | 1058 | io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER); |
0c6e7517 | 1059 | |
755200a3 | 1060 | io_u->error = 0; |
01743ee1 JA |
1061 | flist_del(&io_u->list); |
1062 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a | 1063 | td->cur_depth++; |
0c41214f | 1064 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
1dec3e07 JA |
1065 | } else if (td->o.verify_async) { |
1066 | /* | |
1067 | * We ran out, wait for async verify threads to finish and | |
1068 | * return one | |
1069 | */ | |
1070 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1071 | goto again; | |
10ba535a JA |
1072 | } |
1073 | ||
e8462bd8 | 1074 | td_io_u_unlock(td); |
10ba535a JA |
1075 | return io_u; |
1076 | } | |
1077 | ||
0d29de83 | 1078 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1079 | { |
0d29de83 JA |
1080 | if (td->o.trim_backlog && td->trim_entries) { |
1081 | int get_trim = 0; | |
10ba535a | 1082 | |
0d29de83 JA |
1083 | if (td->trim_batch) { |
1084 | td->trim_batch--; | |
1085 | get_trim = 1; | |
1086 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1087 | td->last_ddir != DDIR_READ) { | |
1088 | td->trim_batch = td->o.trim_batch; | |
1089 | if (!td->trim_batch) | |
1090 | td->trim_batch = td->o.trim_backlog; | |
1091 | get_trim = 1; | |
1092 | } | |
1093 | ||
1094 | if (get_trim && !get_next_trim(td, io_u)) | |
1095 | return 1; | |
2ba1c290 | 1096 | } |
10ba535a | 1097 | |
0d29de83 JA |
1098 | return 0; |
1099 | } | |
1100 | ||
1101 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1102 | { | |
9e144189 JA |
1103 | if (td->o.verify_backlog && td->io_hist_len) { |
1104 | int get_verify = 0; | |
1105 | ||
1106 | if (td->verify_batch) { | |
1107 | td->verify_batch--; | |
1108 | get_verify = 1; | |
1109 | } else if (!(td->io_hist_len % td->o.verify_backlog) && | |
1110 | td->last_ddir != DDIR_READ) { | |
1111 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1112 | if (!td->verify_batch) |
1113 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1114 | get_verify = 1; |
1115 | } | |
1116 | ||
1117 | if (get_verify && !get_next_verify(td, io_u)) | |
0d29de83 | 1118 | return 1; |
9e144189 JA |
1119 | } |
1120 | ||
0d29de83 JA |
1121 | return 0; |
1122 | } | |
1123 | ||
de789769 JA |
1124 | /* |
1125 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1126 | * easy compressible data for simple de-dupe attempts. Do this for every |
1127 | * 512b block in the range, since that should be the smallest block size | |
1128 | * we can expect from a device. | |
de789769 JA |
1129 | */ |
1130 | static void small_content_scramble(struct io_u *io_u) | |
1131 | { | |
23f394d5 | 1132 | unsigned int i, nr_blocks = io_u->buflen / 512; |
fad82f76 | 1133 | unsigned long long boffset; |
23f394d5 JA |
1134 | unsigned int offset; |
1135 | void *p, *end; | |
de789769 | 1136 | |
23f394d5 JA |
1137 | if (!nr_blocks) |
1138 | return; | |
1139 | ||
1140 | p = io_u->xfer_buf; | |
fba76ee8 | 1141 | boffset = io_u->offset; |
fad82f76 | 1142 | |
23f394d5 JA |
1143 | for (i = 0; i < nr_blocks; i++) { |
1144 | /* | |
1145 | * Fill the byte offset into a "random" start offset of | |
1146 | * the buffer, given by the product of the usec time | |
1147 | * and the actual offset. | |
1148 | */ | |
fad82f76 | 1149 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
fba76ee8 | 1150 | offset &= ~(sizeof(unsigned long long) - 1); |
23f394d5 JA |
1151 | if (offset >= 512 - sizeof(unsigned long long)) |
1152 | offset -= sizeof(unsigned long long); | |
fba76ee8 | 1153 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1154 | |
1155 | end = p + 512 - sizeof(io_u->start_time); | |
1156 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1157 | p += 512; | |
fad82f76 | 1158 | boffset += 512; |
23f394d5 | 1159 | } |
de789769 JA |
1160 | } |
1161 | ||
0d29de83 JA |
1162 | /* |
1163 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1164 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1165 | */ | |
1166 | struct io_u *get_io_u(struct thread_data *td) | |
1167 | { | |
1168 | struct fio_file *f; | |
1169 | struct io_u *io_u; | |
de789769 | 1170 | int do_scramble = 0; |
0d29de83 JA |
1171 | |
1172 | io_u = __get_io_u(td); | |
1173 | if (!io_u) { | |
1174 | dprint(FD_IO, "__get_io_u failed\n"); | |
1175 | return NULL; | |
1176 | } | |
1177 | ||
1178 | if (check_get_verify(td, io_u)) | |
1179 | goto out; | |
1180 | if (check_get_trim(td, io_u)) | |
1181 | goto out; | |
1182 | ||
755200a3 JA |
1183 | /* |
1184 | * from a requeue, io_u already setup | |
1185 | */ | |
1186 | if (io_u->file) | |
77f392bf | 1187 | goto out; |
755200a3 | 1188 | |
429f6675 JA |
1189 | /* |
1190 | * If using an iolog, grab next piece if any available. | |
1191 | */ | |
1192 | if (td->o.read_iolog_file) { | |
1193 | if (read_iolog_get(td, io_u)) | |
1194 | goto err_put; | |
2ba1c290 JA |
1195 | } else if (set_io_u_file(td, io_u)) { |
1196 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 1197 | goto err_put; |
2ba1c290 | 1198 | } |
5ec10eaa | 1199 | |
429f6675 | 1200 | f = io_u->file; |
d6aed795 | 1201 | assert(fio_file_open(f)); |
97af62ce | 1202 | |
ff58fced | 1203 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1204 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1205 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1206 | goto err_put; |
2ba1c290 | 1207 | } |
10ba535a | 1208 | |
38dad62d | 1209 | f->last_start = io_u->offset; |
36167d82 | 1210 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1211 | |
fd68418e JA |
1212 | if (io_u->ddir == DDIR_WRITE) { |
1213 | if (td->o.verify != VERIFY_NONE) | |
1214 | populate_verify_io_u(td, io_u); | |
1215 | else if (td->o.refill_buffers) | |
1216 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
1217 | else if (td->o.scramble_buffers) | |
1218 | do_scramble = 1; | |
1219 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1220 | /* |
1221 | * Reset the buf_filled parameters so next time if the | |
1222 | * buffer is used for writes it is refilled. | |
1223 | */ | |
cbe8d756 RR |
1224 | io_u->buf_filled_len = 0; |
1225 | } | |
87dc1ab1 | 1226 | } |
10ba535a | 1227 | |
165faf16 JA |
1228 | /* |
1229 | * Set io data pointers. | |
1230 | */ | |
cec6b55d JA |
1231 | io_u->xfer_buf = io_u->buf; |
1232 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1233 | |
6ac7a331 | 1234 | out: |
0d29de83 | 1235 | assert(io_u->file); |
429f6675 | 1236 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1237 | if (!td->o.disable_slat) |
1238 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1239 | if (do_scramble) |
1240 | small_content_scramble(io_u); | |
429f6675 | 1241 | return io_u; |
36167d82 | 1242 | } |
429f6675 | 1243 | err_put: |
2ba1c290 | 1244 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
1245 | put_io_u(td, io_u); |
1246 | return NULL; | |
10ba535a JA |
1247 | } |
1248 | ||
5451792e JA |
1249 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1250 | { | |
825f818e JA |
1251 | const char *msg[] = { "read", "write", "sync", "datasync", |
1252 | "sync_file_range", "wait", "trim" }; | |
1253 | ||
1254 | ||
5451792e JA |
1255 | |
1256 | log_err("fio: io_u error"); | |
1257 | ||
1258 | if (io_u->file) | |
1259 | log_err(" on file %s", io_u->file->file_name); | |
1260 | ||
1261 | log_err(": %s\n", strerror(io_u->error)); | |
1262 | ||
5ec10eaa JA |
1263 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
1264 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1265 | |
1266 | if (!td->error) | |
1267 | td_verror(td, io_u->error, "io_u error"); | |
1268 | } | |
1269 | ||
c8eeb9df JA |
1270 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1271 | struct io_completion_data *icd, | |
1272 | const enum fio_ddir idx, unsigned int bytes) | |
1273 | { | |
1274 | unsigned long uninitialized_var(lusec); | |
1275 | ||
1276 | if (!icd->account) | |
1277 | return; | |
1278 | ||
1279 | if (!td->o.disable_clat || !td->o.disable_bw) | |
1280 | lusec = utime_since(&io_u->issue_time, &icd->time); | |
1281 | ||
1282 | if (!td->o.disable_lat) { | |
1283 | unsigned long tusec; | |
1284 | ||
1285 | tusec = utime_since(&io_u->start_time, &icd->time); | |
1286 | add_lat_sample(td, idx, tusec, bytes); | |
1287 | } | |
1288 | ||
1289 | if (!td->o.disable_clat) { | |
1290 | add_clat_sample(td, idx, lusec, bytes); | |
1291 | io_u_mark_latency(td, lusec); | |
1292 | } | |
1293 | ||
1294 | if (!td->o.disable_bw) | |
1295 | add_bw_sample(td, idx, bytes, &icd->time); | |
1296 | ||
1297 | add_iops_sample(td, idx, &icd->time); | |
1298 | } | |
1299 | ||
97601024 JA |
1300 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
1301 | struct io_completion_data *icd) | |
10ba535a | 1302 | { |
dbad30b0 JA |
1303 | /* |
1304 | * Older gcc's are too dumb to realize that usec is always used | |
1305 | * initialized, silence that warning. | |
1306 | */ | |
1307 | unsigned long uninitialized_var(usec); | |
44f29692 | 1308 | struct fio_file *f; |
10ba535a | 1309 | |
2ba1c290 JA |
1310 | dprint_io_u(io_u, "io complete"); |
1311 | ||
2ecc1b57 | 1312 | td_io_u_lock(td); |
0c6e7517 | 1313 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1314 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
2ecc1b57 | 1315 | td_io_u_unlock(td); |
0c6e7517 | 1316 | |
5f9099ea | 1317 | if (ddir_sync(io_u->ddir)) { |
87dc1ab1 | 1318 | td->last_was_sync = 1; |
44f29692 JA |
1319 | f = io_u->file; |
1320 | if (f) { | |
1321 | f->first_write = -1ULL; | |
1322 | f->last_write = -1ULL; | |
1323 | } | |
87dc1ab1 JA |
1324 | return; |
1325 | } | |
1326 | ||
1327 | td->last_was_sync = 0; | |
9e144189 | 1328 | td->last_ddir = io_u->ddir; |
87dc1ab1 | 1329 | |
ff58fced | 1330 | if (!io_u->error && ddir_rw(io_u->ddir)) { |
10ba535a | 1331 | unsigned int bytes = io_u->buflen - io_u->resid; |
1e97cce9 | 1332 | const enum fio_ddir idx = io_u->ddir; |
ba3e4e0c | 1333 | const enum fio_ddir odx = io_u->ddir ^ 1; |
b29ee5b3 | 1334 | int ret; |
10ba535a | 1335 | |
b29ee5b3 | 1336 | td->io_blocks[idx]++; |
c8eeb9df | 1337 | td->this_io_blocks[idx]++; |
b29ee5b3 JA |
1338 | td->io_bytes[idx] += bytes; |
1339 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 1340 | |
44f29692 JA |
1341 | if (idx == DDIR_WRITE) { |
1342 | f = io_u->file; | |
1343 | if (f) { | |
1344 | if (f->first_write == -1ULL || | |
1345 | io_u->offset < f->first_write) | |
1346 | f->first_write = io_u->offset; | |
1347 | if (f->last_write == -1ULL || | |
1348 | ((io_u->offset + bytes) > f->last_write)) | |
1349 | f->last_write = io_u->offset + bytes; | |
1350 | } | |
1351 | } | |
1352 | ||
c8eeb9df JA |
1353 | if (ramp_time_over(td) && td->runstate == TD_RUNNING) { |
1354 | account_io_completion(td, io_u, icd, idx, bytes); | |
40e1a6f0 | 1355 | |
b23b6a2f | 1356 | if (__should_check_rate(td, idx)) { |
ba3e4e0c RR |
1357 | td->rate_pending_usleep[idx] = |
1358 | ((td->this_io_bytes[idx] * | |
1359 | td->rate_nsec_cycle[idx]) / 1000 - | |
1360 | utime_since_now(&td->start)); | |
b23b6a2f | 1361 | } |
581e7141 | 1362 | if (__should_check_rate(td, idx ^ 1)) |
ba3e4e0c RR |
1363 | td->rate_pending_usleep[odx] = |
1364 | ((td->this_io_bytes[odx] * | |
1365 | td->rate_nsec_cycle[odx]) / 1000 - | |
1366 | utime_since_now(&td->start)); | |
721938ae | 1367 | } |
10ba535a | 1368 | |
660a1cb5 | 1369 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 1370 | td->o.do_verify && |
41128405 | 1371 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
1372 | log_io_piece(td, io_u); |
1373 | ||
b29ee5b3 | 1374 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 1375 | |
d7762cf8 | 1376 | if (io_u->end_io) { |
36690c9b | 1377 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
1378 | if (ret && !icd->error) |
1379 | icd->error = ret; | |
1380 | } | |
ff58fced | 1381 | } else if (io_u->error) { |
10ba535a | 1382 | icd->error = io_u->error; |
5451792e JA |
1383 | io_u_log_error(td, io_u); |
1384 | } | |
f2bba182 RR |
1385 | if (td->o.continue_on_error && icd->error && |
1386 | td_non_fatal_error(icd->error)) { | |
1387 | /* | |
1388 | * If there is a non_fatal error, then add to the error count | |
1389 | * and clear all the errors. | |
1390 | */ | |
1391 | update_error_count(td, icd->error); | |
1392 | td_clear_error(td); | |
1393 | icd->error = 0; | |
1394 | io_u->error = 0; | |
1395 | } | |
10ba535a JA |
1396 | } |
1397 | ||
9520ebb9 JA |
1398 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1399 | int nr) | |
10ba535a | 1400 | { |
9520ebb9 JA |
1401 | if (!td->o.disable_clat || !td->o.disable_bw) |
1402 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 1403 | |
3af6ef39 | 1404 | icd->nr = nr; |
c8eeb9df | 1405 | icd->account = 1; |
3af6ef39 | 1406 | |
10ba535a JA |
1407 | icd->error = 0; |
1408 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
1409 | } |
1410 | ||
97601024 JA |
1411 | static void ios_completed(struct thread_data *td, |
1412 | struct io_completion_data *icd) | |
36167d82 JA |
1413 | { |
1414 | struct io_u *io_u; | |
1415 | int i; | |
1416 | ||
10ba535a JA |
1417 | for (i = 0; i < icd->nr; i++) { |
1418 | io_u = td->io_ops->event(td, i); | |
1419 | ||
1420 | io_completed(td, io_u, icd); | |
e8462bd8 JA |
1421 | |
1422 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1423 | put_io_u(td, io_u); | |
c8eeb9df JA |
1424 | |
1425 | icd->account = 0; | |
10ba535a JA |
1426 | } |
1427 | } | |
97601024 | 1428 | |
e7e6cfb4 JA |
1429 | /* |
1430 | * Complete a single io_u for the sync engines. | |
1431 | */ | |
581e7141 JA |
1432 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
1433 | unsigned long *bytes) | |
97601024 JA |
1434 | { |
1435 | struct io_completion_data icd; | |
1436 | ||
9520ebb9 | 1437 | init_icd(td, &icd, 1); |
97601024 | 1438 | io_completed(td, io_u, &icd); |
e8462bd8 JA |
1439 | |
1440 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1441 | put_io_u(td, io_u); | |
97601024 | 1442 | |
581e7141 JA |
1443 | if (icd.error) { |
1444 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1445 | return -1; | |
1446 | } | |
97601024 | 1447 | |
581e7141 JA |
1448 | if (bytes) { |
1449 | bytes[0] += icd.bytes_done[0]; | |
1450 | bytes[1] += icd.bytes_done[1]; | |
1451 | } | |
1452 | ||
1453 | return 0; | |
97601024 JA |
1454 | } |
1455 | ||
e7e6cfb4 JA |
1456 | /* |
1457 | * Called to complete min_events number of io for the async engines. | |
1458 | */ | |
581e7141 JA |
1459 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
1460 | unsigned long *bytes) | |
97601024 | 1461 | { |
97601024 | 1462 | struct io_completion_data icd; |
00de55ef | 1463 | struct timespec *tvp = NULL; |
97601024 | 1464 | int ret; |
4d06a338 | 1465 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1466 | |
4950421a | 1467 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1468 | |
4950421a | 1469 | if (!min_evts) |
00de55ef | 1470 | tvp = &ts; |
97601024 | 1471 | |
4950421a | 1472 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1473 | if (ret < 0) { |
e1161c32 | 1474 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1475 | return ret; |
1476 | } else if (!ret) | |
1477 | return ret; | |
1478 | ||
9520ebb9 | 1479 | init_icd(td, &icd, ret); |
97601024 | 1480 | ios_completed(td, &icd); |
581e7141 JA |
1481 | if (icd.error) { |
1482 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1483 | return -1; | |
1484 | } | |
97601024 | 1485 | |
581e7141 JA |
1486 | if (bytes) { |
1487 | bytes[0] += icd.bytes_done[0]; | |
1488 | bytes[1] += icd.bytes_done[1]; | |
1489 | } | |
1490 | ||
1491 | return 0; | |
97601024 | 1492 | } |
7e77dd02 JA |
1493 | |
1494 | /* | |
1495 | * Call when io_u is really queued, to update the submission latency. | |
1496 | */ | |
1497 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1498 | { | |
9520ebb9 JA |
1499 | if (!td->o.disable_slat) { |
1500 | unsigned long slat_time; | |
7e77dd02 | 1501 | |
9520ebb9 | 1502 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
29a90ddb | 1503 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen); |
9520ebb9 | 1504 | } |
7e77dd02 | 1505 | } |
433afcb4 | 1506 | |
5973cafb JA |
1507 | /* |
1508 | * "randomly" fill the buffer contents | |
1509 | */ | |
1510 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1511 | unsigned int max_bs) | |
1512 | { | |
e84b371e JA |
1513 | io_u->buf_filled_len = 0; |
1514 | ||
637ef8d9 | 1515 | if (!td->o.zero_buffers) |
3545a109 | 1516 | fill_random_buf(&td->buf_state, io_u->buf, max_bs); |
637ef8d9 JA |
1517 | else |
1518 | memset(io_u->buf, 0, max_bs); | |
5973cafb | 1519 | } |