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 */ | |
97601024 JA |
16 | |
17 | int error; /* output */ | |
18 | unsigned long bytes_done[2]; /* output */ | |
19 | struct timeval time; /* output */ | |
20 | }; | |
21 | ||
10ba535a JA |
22 | /* |
23 | * The ->file_map[] contains a map of blocks we have or have not done io | |
24 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
25 | */ | |
aec2de20 | 26 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 27 | { |
aec2de20 JA |
28 | unsigned int idx = RAND_MAP_IDX(f, block); |
29 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 30 | |
84422acd JA |
31 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
32 | ||
0ce8b119 | 33 | return (f->file_map[idx] & (1UL << bit)) == 0; |
10ba535a JA |
34 | } |
35 | ||
df415585 JA |
36 | /* |
37 | * Mark a given offset as used in the map. | |
38 | */ | |
9bf2061e | 39 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 40 | { |
2dc1bbeb | 41 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 42 | struct fio_file *f = io_u->file; |
a00735e6 | 43 | unsigned long long block; |
3e3357b1 | 44 | unsigned int blocks, nr_blocks; |
38dad62d | 45 | int busy_check; |
df415585 | 46 | |
b9c5b644 | 47 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
c685b5b2 | 48 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
3e3357b1 | 49 | blocks = 0; |
38dad62d | 50 | busy_check = !(io_u->flags & IO_U_F_BUSY_OK); |
c685b5b2 | 51 | |
3e3357b1 | 52 | while (nr_blocks) { |
df415585 | 53 | unsigned int idx, bit; |
0ce8b119 | 54 | unsigned long mask, this_blocks; |
df415585 | 55 | |
1e3d53ac JA |
56 | /* |
57 | * If we have a mixed random workload, we may | |
58 | * encounter blocks we already did IO to. | |
59 | */ | |
38dad62d JA |
60 | if (!busy_check) { |
61 | blocks = nr_blocks; | |
62 | break; | |
63 | } | |
5736c10d | 64 | if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block)) |
df415585 JA |
65 | break; |
66 | ||
aec2de20 JA |
67 | idx = RAND_MAP_IDX(f, block); |
68 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 69 | |
0032bf9f | 70 | fio_assert(td, idx < f->num_maps); |
df415585 | 71 | |
3e3357b1 JA |
72 | this_blocks = nr_blocks; |
73 | if (this_blocks + bit > BLOCKS_PER_MAP) | |
74 | this_blocks = BLOCKS_PER_MAP - bit; | |
75 | ||
da3758bf JA |
76 | do { |
77 | if (this_blocks == BLOCKS_PER_MAP) | |
0ce8b119 | 78 | mask = -1UL; |
da3758bf | 79 | else |
0ce8b119 | 80 | mask = ((1UL << this_blocks) - 1) << bit; |
da3758bf JA |
81 | |
82 | if (!(f->file_map[idx] & mask)) | |
83 | break; | |
84 | ||
85 | this_blocks--; | |
86 | } while (this_blocks); | |
87 | ||
88 | if (!this_blocks) | |
89 | break; | |
3e3357b1 | 90 | |
3e3357b1 JA |
91 | f->file_map[idx] |= mask; |
92 | nr_blocks -= this_blocks; | |
93 | blocks += this_blocks; | |
c9dd34b2 | 94 | block += this_blocks; |
df415585 JA |
95 | } |
96 | ||
a00735e6 JA |
97 | if ((blocks * min_bs) < io_u->buflen) |
98 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
99 | } |
100 | ||
3e3357b1 JA |
101 | static unsigned long long last_block(struct thread_data *td, struct fio_file *f, |
102 | enum fio_ddir ddir) | |
2ba1c290 JA |
103 | { |
104 | unsigned long long max_blocks; | |
d9dd70f7 | 105 | unsigned long long max_size; |
2ba1c290 | 106 | |
ff58fced JA |
107 | assert(ddir_rw(ddir)); |
108 | ||
d9dd70f7 JA |
109 | /* |
110 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
111 | */ | |
112 | max_size = f->io_size; | |
113 | if (max_size > f->real_file_size) | |
114 | max_size = f->real_file_size; | |
115 | ||
2b7a01d0 | 116 | max_blocks = max_size / (unsigned long long) td->o.ba[ddir]; |
2ba1c290 JA |
117 | if (!max_blocks) |
118 | return 0; | |
119 | ||
67778e88 | 120 | return max_blocks; |
2ba1c290 JA |
121 | } |
122 | ||
10ba535a JA |
123 | /* |
124 | * Return the next free block in the map. | |
125 | */ | |
126 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 127 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 128 | { |
0ce8b119 | 129 | unsigned long long block, min_bs = td->o.rw_min_bs, lastb; |
10ba535a JA |
130 | int i; |
131 | ||
5e0baa7f JA |
132 | lastb = last_block(td, f, ddir); |
133 | if (!lastb) | |
134 | return 1; | |
135 | ||
c685b5b2 | 136 | i = f->last_free_lookup; |
0ce8b119 JA |
137 | block = i * BLOCKS_PER_MAP; |
138 | while (block * min_bs < f->real_file_size && | |
139 | block * min_bs < f->io_size) { | |
140 | if (f->file_map[i] != -1UL) { | |
141 | block += ffz(f->file_map[i]); | |
142 | if (block > lastb) | |
2ba1c290 | 143 | break; |
c685b5b2 | 144 | f->last_free_lookup = i; |
0ce8b119 | 145 | *b = block; |
10ba535a JA |
146 | return 0; |
147 | } | |
148 | ||
0ce8b119 | 149 | block += BLOCKS_PER_MAP; |
10ba535a JA |
150 | i++; |
151 | } | |
152 | ||
2ba1c290 | 153 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
154 | return 1; |
155 | } | |
156 | ||
ec4015da | 157 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 158 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 159 | { |
5e0baa7f | 160 | unsigned long long r, lastb; |
ec4015da JA |
161 | int loops = 5; |
162 | ||
5e0baa7f JA |
163 | lastb = last_block(td, f, ddir); |
164 | if (!lastb) | |
165 | return 1; | |
166 | ||
0ce8b119 JA |
167 | if (f->failed_rands >= 200) |
168 | goto ffz; | |
169 | ||
ec4015da | 170 | do { |
2615cc4b JA |
171 | if (td->o.use_os_rand) { |
172 | r = os_random_long(&td->random_state); | |
173 | *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0)); | |
174 | } else { | |
175 | r = __rand(&td->__random_state); | |
176 | *b = (lastb - 1) * (r / ((unsigned long long) FRAND_MAX + 1.0)); | |
177 | } | |
178 | ||
84422acd | 179 | dprint(FD_RANDOM, "off rand %llu\n", r); |
2615cc4b | 180 | |
2ba1c290 | 181 | |
43c63a78 JA |
182 | /* |
183 | * if we are not maintaining a random map, we are done. | |
184 | */ | |
303032ae | 185 | if (!file_randommap(td, f)) |
0ce8b119 | 186 | goto ret_good; |
43c63a78 JA |
187 | |
188 | /* | |
84422acd | 189 | * calculate map offset and check if it's free |
43c63a78 | 190 | */ |
aec2de20 | 191 | if (random_map_free(f, *b)) |
0ce8b119 | 192 | goto ret_good; |
43c63a78 | 193 | |
84422acd JA |
194 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
195 | *b); | |
43c63a78 | 196 | } while (--loops); |
ec4015da | 197 | |
0ce8b119 JA |
198 | if (!f->failed_rands++) |
199 | f->last_free_lookup = 0; | |
200 | ||
ec4015da | 201 | /* |
43c63a78 JA |
202 | * we get here, if we didn't suceed in looking up a block. generate |
203 | * a random start offset into the filemap, and find the first free | |
204 | * block from there. | |
ec4015da | 205 | */ |
43c63a78 JA |
206 | loops = 10; |
207 | do { | |
dc873b6f JA |
208 | f->last_free_lookup = (f->num_maps - 1) * |
209 | (r / (OS_RAND_MAX + 1.0)); | |
4ba66134 | 210 | if (!get_next_free_block(td, f, ddir, b)) |
0ce8b119 | 211 | goto ret; |
ec4015da | 212 | |
d1cea521 JA |
213 | if (td->o.use_os_rand) |
214 | r = os_random_long(&td->random_state); | |
215 | else | |
216 | r = __rand(&td->__random_state); | |
43c63a78 JA |
217 | } while (--loops); |
218 | ||
219 | /* | |
220 | * that didn't work either, try exhaustive search from the start | |
221 | */ | |
222 | f->last_free_lookup = 0; | |
0ce8b119 JA |
223 | ffz: |
224 | if (!get_next_free_block(td, f, ddir, b)) | |
225 | return 0; | |
226 | f->last_free_lookup = 0; | |
4ba66134 | 227 | return get_next_free_block(td, f, ddir, b); |
0ce8b119 JA |
228 | ret_good: |
229 | f->failed_rands = 0; | |
230 | ret: | |
231 | return 0; | |
ec4015da JA |
232 | } |
233 | ||
38dad62d JA |
234 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
235 | enum fio_ddir ddir, unsigned long long *b) | |
236 | { | |
237 | if (get_next_rand_offset(td, f, ddir, b)) { | |
238 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", | |
239 | f->file_name, f->last_pos, f->real_file_size); | |
240 | return 1; | |
241 | } | |
242 | ||
243 | return 0; | |
244 | } | |
245 | ||
246 | static int get_next_seq_block(struct thread_data *td, struct fio_file *f, | |
247 | enum fio_ddir ddir, unsigned long long *b) | |
248 | { | |
ff58fced JA |
249 | assert(ddir_rw(ddir)); |
250 | ||
38dad62d | 251 | if (f->last_pos < f->real_file_size) { |
a66da7a2 | 252 | unsigned long long pos; |
059b0802 | 253 | |
a66da7a2 JA |
254 | if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0) |
255 | f->last_pos = f->real_file_size; | |
256 | ||
257 | pos = f->last_pos - f->file_offset; | |
059b0802 JA |
258 | if (pos) |
259 | pos += td->o.ddir_seq_add; | |
260 | ||
261 | *b = pos / td->o.min_bs[ddir]; | |
38dad62d JA |
262 | return 0; |
263 | } | |
264 | ||
265 | return 1; | |
266 | } | |
267 | ||
268 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
269 | enum fio_ddir ddir, int rw_seq, unsigned long long *b) | |
270 | { | |
271 | struct fio_file *f = io_u->file; | |
272 | int ret; | |
273 | ||
ff58fced JA |
274 | assert(ddir_rw(ddir)); |
275 | ||
38dad62d JA |
276 | if (rw_seq) { |
277 | if (td_random(td)) | |
278 | ret = get_next_rand_block(td, f, ddir, b); | |
279 | else | |
280 | ret = get_next_seq_block(td, f, ddir, b); | |
281 | } else { | |
282 | io_u->flags |= IO_U_F_BUSY_OK; | |
283 | ||
284 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
285 | ret = get_next_seq_block(td, f, ddir, b); | |
286 | if (ret) | |
287 | ret = get_next_rand_block(td, f, ddir, b); | |
288 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { | |
289 | if (f->last_start != -1ULL) | |
0508320e DN |
290 | *b = (f->last_start - f->file_offset) |
291 | / td->o.min_bs[ddir]; | |
38dad62d JA |
292 | else |
293 | *b = 0; | |
294 | ret = 0; | |
295 | } else { | |
296 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
297 | ret = 1; | |
298 | } | |
299 | } | |
300 | ||
301 | return ret; | |
302 | } | |
303 | ||
10ba535a JA |
304 | /* |
305 | * For random io, generate a random new block and see if it's used. Repeat | |
306 | * until we find a free one. For sequential io, just return the end of | |
307 | * the last io issued. | |
308 | */ | |
15dc1934 | 309 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 310 | { |
9bf2061e | 311 | struct fio_file *f = io_u->file; |
ec4015da | 312 | unsigned long long b; |
4ba66134 | 313 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 314 | int rw_seq_hit = 0; |
10ba535a | 315 | |
ff58fced JA |
316 | assert(ddir_rw(ddir)); |
317 | ||
38dad62d JA |
318 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
319 | rw_seq_hit = 1; | |
5736c10d | 320 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 321 | } |
211097b2 | 322 | |
0508320e | 323 | if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) |
38dad62d | 324 | return 1; |
10ba535a | 325 | |
2b7a01d0 | 326 | io_u->offset = b * td->o.ba[ddir]; |
009bd847 JA |
327 | if (io_u->offset >= f->io_size) { |
328 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
329 | io_u->offset, f->io_size); | |
330 | return 1; | |
331 | } | |
332 | ||
333 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
334 | if (io_u->offset >= f->real_file_size) { |
335 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
336 | io_u->offset, f->real_file_size); | |
10ba535a | 337 | return 1; |
2ba1c290 | 338 | } |
10ba535a JA |
339 | |
340 | return 0; | |
341 | } | |
342 | ||
15dc1934 JA |
343 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
344 | { | |
7eb36574 JA |
345 | struct prof_io_ops *ops = &td->prof_io_ops; |
346 | ||
347 | if (ops->fill_io_u_off) | |
348 | return ops->fill_io_u_off(td, io_u); | |
15dc1934 JA |
349 | |
350 | return __get_next_offset(td, io_u); | |
351 | } | |
352 | ||
79944128 JA |
353 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
354 | unsigned int buflen) | |
355 | { | |
356 | struct fio_file *f = io_u->file; | |
357 | ||
358 | return io_u->offset + buflen <= f->io_size + td->o.start_offset; | |
359 | } | |
360 | ||
15dc1934 | 361 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 362 | { |
bca4ed4d | 363 | const int ddir = io_u->ddir; |
f3f552b9 | 364 | unsigned int uninitialized_var(buflen); |
f3059de1 | 365 | unsigned int minbs, maxbs; |
1294c3ec | 366 | unsigned long r, rand_max; |
10ba535a | 367 | |
ff58fced JA |
368 | assert(ddir_rw(ddir)); |
369 | ||
f3059de1 JA |
370 | minbs = td->o.min_bs[ddir]; |
371 | maxbs = td->o.max_bs[ddir]; | |
372 | ||
79944128 JA |
373 | if (minbs == maxbs) |
374 | return minbs; | |
375 | ||
4c07ad86 JA |
376 | if (td->o.use_os_rand) |
377 | rand_max = OS_RAND_MAX; | |
378 | else | |
379 | rand_max = FRAND_MAX; | |
380 | ||
79944128 | 381 | do { |
4c07ad86 JA |
382 | if (td->o.use_os_rand) |
383 | r = os_random_long(&td->bsrange_state); | |
384 | else | |
385 | r = __rand(&td->__bsrange_state); | |
386 | ||
720e84ad | 387 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 388 | buflen = 1 + (unsigned int) ((double) maxbs * |
4c07ad86 | 389 | (r / (rand_max + 1.0))); |
f3059de1 JA |
390 | if (buflen < minbs) |
391 | buflen = minbs; | |
5ec10eaa | 392 | } else { |
564ca972 JA |
393 | long perc = 0; |
394 | unsigned int i; | |
395 | ||
720e84ad JA |
396 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
397 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
398 | |
399 | buflen = bsp->bs; | |
400 | perc += bsp->perc; | |
79944128 JA |
401 | if ((r <= ((rand_max / 100L) * perc)) && |
402 | io_u_fits(td, io_u, buflen)) | |
564ca972 JA |
403 | break; |
404 | } | |
405 | } | |
79944128 | 406 | |
f3059de1 JA |
407 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
408 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 409 | |
79944128 | 410 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 411 | |
10ba535a JA |
412 | return buflen; |
413 | } | |
414 | ||
15dc1934 JA |
415 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
416 | { | |
7eb36574 JA |
417 | struct prof_io_ops *ops = &td->prof_io_ops; |
418 | ||
419 | if (ops->fill_io_u_size) | |
420 | return ops->fill_io_u_size(td, io_u); | |
15dc1934 JA |
421 | |
422 | return __get_next_buflen(td, io_u); | |
423 | } | |
424 | ||
afe24a5a JA |
425 | static void set_rwmix_bytes(struct thread_data *td) |
426 | { | |
afe24a5a JA |
427 | unsigned int diff; |
428 | ||
429 | /* | |
430 | * we do time or byte based switch. this is needed because | |
431 | * buffered writes may issue a lot quicker than they complete, | |
432 | * whereas reads do not. | |
433 | */ | |
e47f799f | 434 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 435 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
436 | } |
437 | ||
438 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
439 | { | |
440 | unsigned int v; | |
1294c3ec | 441 | unsigned long r; |
e47f799f | 442 | |
4c07ad86 JA |
443 | if (td->o.use_os_rand) { |
444 | r = os_random_long(&td->rwmix_state); | |
445 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); | |
446 | } else { | |
447 | r = __rand(&td->__rwmix_state); | |
448 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
449 | } | |
450 | ||
04c540d9 | 451 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
452 | return DDIR_READ; |
453 | ||
454 | return DDIR_WRITE; | |
afe24a5a JA |
455 | } |
456 | ||
581e7141 JA |
457 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
458 | { | |
459 | enum fio_ddir odir = ddir ^ 1; | |
460 | struct timeval t; | |
461 | long usec; | |
462 | ||
ff58fced JA |
463 | assert(ddir_rw(ddir)); |
464 | ||
581e7141 JA |
465 | if (td->rate_pending_usleep[ddir] <= 0) |
466 | return ddir; | |
467 | ||
468 | /* | |
469 | * We have too much pending sleep in this direction. See if we | |
470 | * should switch. | |
471 | */ | |
472 | if (td_rw(td)) { | |
473 | /* | |
474 | * Other direction does not have too much pending, switch | |
475 | */ | |
476 | if (td->rate_pending_usleep[odir] < 100000) | |
477 | return odir; | |
478 | ||
479 | /* | |
480 | * Both directions have pending sleep. Sleep the minimum time | |
481 | * and deduct from both. | |
482 | */ | |
483 | if (td->rate_pending_usleep[ddir] <= | |
484 | td->rate_pending_usleep[odir]) { | |
485 | usec = td->rate_pending_usleep[ddir]; | |
486 | } else { | |
487 | usec = td->rate_pending_usleep[odir]; | |
488 | ddir = odir; | |
489 | } | |
490 | } else | |
491 | usec = td->rate_pending_usleep[ddir]; | |
492 | ||
78c1eda5 JA |
493 | /* |
494 | * We are going to sleep, ensure that we flush anything pending as | |
495 | * not to skew our latency numbers | |
496 | */ | |
497 | if (td->cur_depth) { | |
498 | int fio_unused ret; | |
499 | ||
500 | ret = io_u_queued_complete(td, td->cur_depth, NULL); | |
501 | } | |
502 | ||
581e7141 JA |
503 | fio_gettime(&t, NULL); |
504 | usec_sleep(td, usec); | |
505 | usec = utime_since_now(&t); | |
506 | ||
507 | td->rate_pending_usleep[ddir] -= usec; | |
508 | ||
509 | odir = ddir ^ 1; | |
510 | if (td_rw(td) && __should_check_rate(td, odir)) | |
511 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 512 | |
581e7141 JA |
513 | return ddir; |
514 | } | |
515 | ||
10ba535a JA |
516 | /* |
517 | * Return the data direction for the next io_u. If the job is a | |
518 | * mixed read/write workload, check the rwmix cycle and switch if | |
519 | * necessary. | |
520 | */ | |
1e97cce9 | 521 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 522 | { |
581e7141 JA |
523 | enum fio_ddir ddir; |
524 | ||
5f9099ea JA |
525 | /* |
526 | * see if it's time to fsync | |
527 | */ | |
528 | if (td->o.fsync_blocks && | |
529 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
530 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
531 | return DDIR_SYNC; | |
532 | ||
533 | /* | |
534 | * see if it's time to fdatasync | |
535 | */ | |
536 | if (td->o.fdatasync_blocks && | |
537 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
538 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
539 | return DDIR_DATASYNC; | |
540 | ||
44f29692 JA |
541 | /* |
542 | * see if it's time to sync_file_range | |
543 | */ | |
544 | if (td->sync_file_range_nr && | |
545 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
546 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
547 | return DDIR_SYNC_FILE_RANGE; | |
548 | ||
10ba535a | 549 | if (td_rw(td)) { |
10ba535a JA |
550 | /* |
551 | * Check if it's time to seed a new data direction. | |
552 | */ | |
e4928662 | 553 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
554 | /* |
555 | * Put a top limit on how many bytes we do for | |
556 | * one data direction, to avoid overflowing the | |
557 | * ranges too much | |
558 | */ | |
559 | ddir = get_rand_ddir(td); | |
e47f799f JA |
560 | |
561 | if (ddir != td->rwmix_ddir) | |
562 | set_rwmix_bytes(td); | |
563 | ||
564 | td->rwmix_ddir = ddir; | |
10ba535a | 565 | } |
581e7141 | 566 | ddir = td->rwmix_ddir; |
10ba535a | 567 | } else if (td_read(td)) |
581e7141 | 568 | ddir = DDIR_READ; |
10ba535a | 569 | else |
581e7141 JA |
570 | ddir = DDIR_WRITE; |
571 | ||
572 | td->rwmix_ddir = rate_ddir(td, ddir); | |
573 | return td->rwmix_ddir; | |
10ba535a JA |
574 | } |
575 | ||
1ef2b6be JA |
576 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
577 | { | |
578 | io_u->ddir = get_rw_ddir(td); | |
579 | ||
580 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
581 | td->o.barrier_blocks && | |
582 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
583 | td->io_issues[DDIR_WRITE]) | |
584 | io_u->flags |= IO_U_F_BARRIER; | |
585 | } | |
586 | ||
e8462bd8 | 587 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 JA |
588 | { |
589 | int ret = put_file(td, f); | |
590 | ||
591 | if (ret) | |
592 | td_verror(td, ret, "file close"); | |
593 | } | |
594 | ||
10ba535a JA |
595 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
596 | { | |
e8462bd8 JA |
597 | td_io_u_lock(td); |
598 | ||
0c6e7517 | 599 | io_u->flags |= IO_U_F_FREE; |
e8462bd8 | 600 | io_u->flags &= ~IO_U_F_FREE_DEF; |
0c6e7517 | 601 | |
60f2c658 JA |
602 | if (io_u->file) |
603 | put_file_log(td, io_u->file); | |
2dbdab7e | 604 | |
10ba535a | 605 | io_u->file = NULL; |
0c41214f RR |
606 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
607 | td->cur_depth--; | |
e8462bd8 | 608 | flist_del_init(&io_u->list); |
01743ee1 | 609 | flist_add(&io_u->list, &td->io_u_freelist); |
e8462bd8 JA |
610 | td_io_u_unlock(td); |
611 | td_io_u_free_notify(td); | |
10ba535a JA |
612 | } |
613 | ||
f2bba182 RR |
614 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
615 | { | |
616 | io_u->flags &= ~IO_U_F_FLIGHT; | |
617 | put_io_u(td, io_u); | |
618 | } | |
619 | ||
755200a3 JA |
620 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
621 | { | |
622 | struct io_u *__io_u = *io_u; | |
623 | ||
465221b0 JA |
624 | dprint(FD_IO, "requeue %p\n", __io_u); |
625 | ||
e8462bd8 JA |
626 | td_io_u_lock(td); |
627 | ||
4d2e0f49 | 628 | __io_u->flags |= IO_U_F_FREE; |
ff58fced | 629 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir)) |
e4f54adb | 630 | td->io_issues[__io_u->ddir]--; |
5ec10eaa | 631 | |
4d2e0f49 | 632 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
633 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
634 | td->cur_depth--; | |
01743ee1 JA |
635 | flist_del(&__io_u->list); |
636 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
e8462bd8 | 637 | td_io_u_unlock(td); |
755200a3 JA |
638 | *io_u = NULL; |
639 | } | |
640 | ||
9bf2061e | 641 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 642 | { |
b4c5e1ac JA |
643 | if (td->io_ops->flags & FIO_NOIO) |
644 | goto out; | |
645 | ||
1ef2b6be | 646 | set_rw_ddir(td, io_u); |
5f9099ea | 647 | |
87dc1ab1 | 648 | /* |
ff58fced | 649 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 650 | */ |
ff58fced | 651 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 652 | goto out; |
a00735e6 | 653 | |
48f5abd3 JA |
654 | /* |
655 | * See if it's time to switch to a new zone | |
656 | */ | |
657 | if (td->zone_bytes >= td->o.zone_size) { | |
658 | td->zone_bytes = 0; | |
659 | io_u->file->last_pos += td->o.zone_skip; | |
660 | td->io_skip_bytes += td->o.zone_skip; | |
661 | } | |
662 | ||
10ba535a | 663 | /* |
c685b5b2 JA |
664 | * No log, let the seq/rand engine retrieve the next buflen and |
665 | * position. | |
10ba535a | 666 | */ |
2ba1c290 JA |
667 | if (get_next_offset(td, io_u)) { |
668 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 669 | return 1; |
2ba1c290 | 670 | } |
10ba535a | 671 | |
9bf2061e | 672 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
673 | if (!io_u->buflen) { |
674 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 675 | return 1; |
2ba1c290 | 676 | } |
bca4ed4d | 677 | |
2ba1c290 JA |
678 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
679 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
680 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
681 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 682 | return 1; |
2ba1c290 | 683 | } |
6a5e6884 | 684 | |
bca4ed4d JA |
685 | /* |
686 | * mark entry before potentially trimming io_u | |
687 | */ | |
303032ae | 688 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 689 | mark_random_map(td, io_u); |
bca4ed4d JA |
690 | |
691 | /* | |
692 | * If using a write iolog, store this entry. | |
693 | */ | |
c38e9468 | 694 | out: |
2ba1c290 | 695 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 696 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 697 | log_io_u(td, io_u); |
bca4ed4d | 698 | return 0; |
10ba535a JA |
699 | } |
700 | ||
838bc709 JA |
701 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
702 | { | |
2b13e716 | 703 | int idx = 0; |
838bc709 JA |
704 | |
705 | switch (nr) { | |
706 | default: | |
2b13e716 | 707 | idx = 6; |
838bc709 JA |
708 | break; |
709 | case 33 ... 64: | |
2b13e716 | 710 | idx = 5; |
838bc709 JA |
711 | break; |
712 | case 17 ... 32: | |
2b13e716 | 713 | idx = 4; |
838bc709 JA |
714 | break; |
715 | case 9 ... 16: | |
2b13e716 | 716 | idx = 3; |
838bc709 JA |
717 | break; |
718 | case 5 ... 8: | |
2b13e716 | 719 | idx = 2; |
838bc709 JA |
720 | break; |
721 | case 1 ... 4: | |
2b13e716 | 722 | idx = 1; |
838bc709 JA |
723 | case 0: |
724 | break; | |
725 | } | |
726 | ||
2b13e716 | 727 | map[idx]++; |
838bc709 JA |
728 | } |
729 | ||
730 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
731 | { | |
732 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
733 | td->ts.total_submit++; | |
734 | } | |
735 | ||
736 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
737 | { | |
738 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
739 | td->ts.total_complete++; | |
740 | } | |
741 | ||
d8005759 | 742 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 743 | { |
2b13e716 | 744 | int idx = 0; |
71619dc2 JA |
745 | |
746 | switch (td->cur_depth) { | |
747 | default: | |
2b13e716 | 748 | idx = 6; |
a783e61a | 749 | break; |
71619dc2 | 750 | case 32 ... 63: |
2b13e716 | 751 | idx = 5; |
a783e61a | 752 | break; |
71619dc2 | 753 | case 16 ... 31: |
2b13e716 | 754 | idx = 4; |
a783e61a | 755 | break; |
71619dc2 | 756 | case 8 ... 15: |
2b13e716 | 757 | idx = 3; |
a783e61a | 758 | break; |
71619dc2 | 759 | case 4 ... 7: |
2b13e716 | 760 | idx = 2; |
a783e61a | 761 | break; |
71619dc2 | 762 | case 2 ... 3: |
2b13e716 | 763 | idx = 1; |
71619dc2 JA |
764 | case 1: |
765 | break; | |
766 | } | |
767 | ||
2b13e716 | 768 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
769 | } |
770 | ||
04a0feae JA |
771 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
772 | { | |
2b13e716 | 773 | int idx = 0; |
04a0feae JA |
774 | |
775 | assert(usec < 1000); | |
776 | ||
777 | switch (usec) { | |
778 | case 750 ... 999: | |
2b13e716 | 779 | idx = 9; |
04a0feae JA |
780 | break; |
781 | case 500 ... 749: | |
2b13e716 | 782 | idx = 8; |
04a0feae JA |
783 | break; |
784 | case 250 ... 499: | |
2b13e716 | 785 | idx = 7; |
04a0feae JA |
786 | break; |
787 | case 100 ... 249: | |
2b13e716 | 788 | idx = 6; |
04a0feae JA |
789 | break; |
790 | case 50 ... 99: | |
2b13e716 | 791 | idx = 5; |
04a0feae JA |
792 | break; |
793 | case 20 ... 49: | |
2b13e716 | 794 | idx = 4; |
04a0feae JA |
795 | break; |
796 | case 10 ... 19: | |
2b13e716 | 797 | idx = 3; |
04a0feae JA |
798 | break; |
799 | case 4 ... 9: | |
2b13e716 | 800 | idx = 2; |
04a0feae JA |
801 | break; |
802 | case 2 ... 3: | |
2b13e716 | 803 | idx = 1; |
04a0feae JA |
804 | case 0 ... 1: |
805 | break; | |
806 | } | |
807 | ||
2b13e716 JA |
808 | assert(idx < FIO_IO_U_LAT_U_NR); |
809 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
810 | } |
811 | ||
812 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 813 | { |
2b13e716 | 814 | int idx = 0; |
ec118304 JA |
815 | |
816 | switch (msec) { | |
817 | default: | |
2b13e716 | 818 | idx = 11; |
04a0feae | 819 | break; |
8abdce66 | 820 | case 1000 ... 1999: |
2b13e716 | 821 | idx = 10; |
04a0feae | 822 | break; |
8abdce66 | 823 | case 750 ... 999: |
2b13e716 | 824 | idx = 9; |
04a0feae | 825 | break; |
8abdce66 | 826 | case 500 ... 749: |
2b13e716 | 827 | idx = 8; |
04a0feae | 828 | break; |
8abdce66 | 829 | case 250 ... 499: |
2b13e716 | 830 | idx = 7; |
04a0feae | 831 | break; |
8abdce66 | 832 | case 100 ... 249: |
2b13e716 | 833 | idx = 6; |
04a0feae | 834 | break; |
8abdce66 | 835 | case 50 ... 99: |
2b13e716 | 836 | idx = 5; |
04a0feae | 837 | break; |
8abdce66 | 838 | case 20 ... 49: |
2b13e716 | 839 | idx = 4; |
04a0feae | 840 | break; |
8abdce66 | 841 | case 10 ... 19: |
2b13e716 | 842 | idx = 3; |
04a0feae | 843 | break; |
8abdce66 | 844 | case 4 ... 9: |
2b13e716 | 845 | idx = 2; |
04a0feae | 846 | break; |
ec118304 | 847 | case 2 ... 3: |
2b13e716 | 848 | idx = 1; |
ec118304 JA |
849 | case 0 ... 1: |
850 | break; | |
851 | } | |
852 | ||
2b13e716 JA |
853 | assert(idx < FIO_IO_U_LAT_M_NR); |
854 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
855 | } |
856 | ||
857 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
858 | { | |
859 | if (usec < 1000) | |
860 | io_u_mark_lat_usec(td, usec); | |
861 | else | |
862 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
863 | } |
864 | ||
0aabe160 JA |
865 | /* |
866 | * Get next file to service by choosing one at random | |
867 | */ | |
2cc52930 JA |
868 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
869 | enum fio_file_flags goodf, | |
d6aed795 | 870 | enum fio_file_flags badf) |
0aabe160 | 871 | { |
0aabe160 | 872 | struct fio_file *f; |
1c178180 | 873 | int fno; |
0aabe160 JA |
874 | |
875 | do { | |
87b10676 | 876 | int opened = 0; |
1294c3ec | 877 | unsigned long r; |
4c07ad86 JA |
878 | |
879 | if (td->o.use_os_rand) { | |
880 | r = os_random_long(&td->next_file_state); | |
881 | fno = (unsigned int) ((double) td->o.nr_files | |
882 | * (r / (OS_RAND_MAX + 1.0))); | |
883 | } else { | |
884 | r = __rand(&td->__next_file_state); | |
885 | fno = (unsigned int) ((double) td->o.nr_files | |
886 | * (r / (FRAND_MAX + 1.0))); | |
887 | } | |
7c83c089 | 888 | |
126d65c6 | 889 | f = td->files[fno]; |
d6aed795 | 890 | if (fio_file_done(f)) |
059e63c0 | 891 | continue; |
1c178180 | 892 | |
d6aed795 | 893 | if (!fio_file_open(f)) { |
87b10676 JA |
894 | int err; |
895 | ||
896 | err = td_io_open_file(td, f); | |
897 | if (err) | |
898 | continue; | |
899 | opened = 1; | |
900 | } | |
901 | ||
2ba1c290 JA |
902 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
903 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 904 | return f; |
2ba1c290 | 905 | } |
87b10676 JA |
906 | if (opened) |
907 | td_io_close_file(td, f); | |
0aabe160 JA |
908 | } while (1); |
909 | } | |
910 | ||
911 | /* | |
912 | * Get next file to service by doing round robin between all available ones | |
913 | */ | |
1c178180 JA |
914 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
915 | int badf) | |
3d7c391d JA |
916 | { |
917 | unsigned int old_next_file = td->next_file; | |
918 | struct fio_file *f; | |
919 | ||
920 | do { | |
87b10676 JA |
921 | int opened = 0; |
922 | ||
126d65c6 | 923 | f = td->files[td->next_file]; |
3d7c391d JA |
924 | |
925 | td->next_file++; | |
2dc1bbeb | 926 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
927 | td->next_file = 0; |
928 | ||
87b10676 | 929 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 930 | if (fio_file_done(f)) { |
d5ed68ea | 931 | f = NULL; |
059e63c0 | 932 | continue; |
d5ed68ea | 933 | } |
059e63c0 | 934 | |
d6aed795 | 935 | if (!fio_file_open(f)) { |
87b10676 JA |
936 | int err; |
937 | ||
938 | err = td_io_open_file(td, f); | |
b5696bfc JA |
939 | if (err) { |
940 | dprint(FD_FILE, "error %d on open of %s\n", | |
941 | err, f->file_name); | |
87c27b45 | 942 | f = NULL; |
87b10676 | 943 | continue; |
b5696bfc | 944 | } |
87b10676 JA |
945 | opened = 1; |
946 | } | |
947 | ||
0b9d69ec JA |
948 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
949 | f->flags); | |
1c178180 | 950 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
951 | break; |
952 | ||
87b10676 JA |
953 | if (opened) |
954 | td_io_close_file(td, f); | |
955 | ||
3d7c391d JA |
956 | f = NULL; |
957 | } while (td->next_file != old_next_file); | |
958 | ||
2ba1c290 | 959 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
960 | return f; |
961 | } | |
962 | ||
7eb36574 | 963 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 964 | { |
1907dbc6 JA |
965 | struct fio_file *f; |
966 | ||
2dc1bbeb | 967 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 968 | |
b5696bfc | 969 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
970 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
971 | " nr_files=%d\n", td->nr_open_files, | |
972 | td->nr_done_files, | |
973 | td->o.nr_files); | |
bdb4e2e9 | 974 | return NULL; |
2ba1c290 | 975 | } |
bdb4e2e9 | 976 | |
1907dbc6 | 977 | f = td->file_service_file; |
d6aed795 | 978 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
979 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
980 | goto out; | |
981 | if (td->file_service_left--) | |
982 | goto out; | |
983 | } | |
1907dbc6 | 984 | |
a086c257 JA |
985 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
986 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 987 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 988 | else |
d6aed795 | 989 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 JA |
990 | |
991 | td->file_service_file = f; | |
992 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 993 | out: |
683023e8 | 994 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 995 | return f; |
bdb4e2e9 JA |
996 | } |
997 | ||
7eb36574 JA |
998 | static struct fio_file *get_next_file(struct thread_data *td) |
999 | { | |
1000 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1001 | ||
1002 | if (ops->get_next_file) | |
1003 | return ops->get_next_file(td); | |
1004 | ||
1005 | return __get_next_file(td); | |
1006 | } | |
1007 | ||
429f6675 JA |
1008 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
1009 | { | |
1010 | struct fio_file *f; | |
1011 | ||
1012 | do { | |
1013 | f = get_next_file(td); | |
1014 | if (!f) | |
1015 | return 1; | |
1016 | ||
429f6675 JA |
1017 | io_u->file = f; |
1018 | get_file(f); | |
1019 | ||
1020 | if (!fill_io_u(td, io_u)) | |
1021 | break; | |
1022 | ||
b5696bfc | 1023 | put_file_log(td, f); |
429f6675 | 1024 | td_io_close_file(td, f); |
b5696bfc | 1025 | io_u->file = NULL; |
d6aed795 | 1026 | fio_file_set_done(f); |
429f6675 | 1027 | td->nr_done_files++; |
0b9d69ec JA |
1028 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1029 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1030 | } while (1); |
1031 | ||
1032 | return 0; | |
1033 | } | |
1034 | ||
1035 | ||
10ba535a JA |
1036 | struct io_u *__get_io_u(struct thread_data *td) |
1037 | { | |
1038 | struct io_u *io_u = NULL; | |
1039 | ||
e8462bd8 JA |
1040 | td_io_u_lock(td); |
1041 | ||
1042 | again: | |
01743ee1 JA |
1043 | if (!flist_empty(&td->io_u_requeues)) |
1044 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 1045 | else if (!queue_full(td)) { |
01743ee1 | 1046 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 1047 | |
6040dabc | 1048 | io_u->buflen = 0; |
10ba535a | 1049 | io_u->resid = 0; |
755200a3 | 1050 | io_u->file = NULL; |
d7762cf8 | 1051 | io_u->end_io = NULL; |
755200a3 JA |
1052 | } |
1053 | ||
1054 | if (io_u) { | |
0c6e7517 | 1055 | assert(io_u->flags & IO_U_F_FREE); |
2ecc1b57 | 1056 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF); |
1ef2b6be | 1057 | io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER); |
0c6e7517 | 1058 | |
755200a3 | 1059 | io_u->error = 0; |
01743ee1 JA |
1060 | flist_del(&io_u->list); |
1061 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a | 1062 | td->cur_depth++; |
0c41214f | 1063 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
1dec3e07 JA |
1064 | } else if (td->o.verify_async) { |
1065 | /* | |
1066 | * We ran out, wait for async verify threads to finish and | |
1067 | * return one | |
1068 | */ | |
1069 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1070 | goto again; | |
10ba535a JA |
1071 | } |
1072 | ||
e8462bd8 | 1073 | td_io_u_unlock(td); |
10ba535a JA |
1074 | return io_u; |
1075 | } | |
1076 | ||
0d29de83 | 1077 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1078 | { |
0d29de83 JA |
1079 | if (td->o.trim_backlog && td->trim_entries) { |
1080 | int get_trim = 0; | |
10ba535a | 1081 | |
0d29de83 JA |
1082 | if (td->trim_batch) { |
1083 | td->trim_batch--; | |
1084 | get_trim = 1; | |
1085 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1086 | td->last_ddir != DDIR_READ) { | |
1087 | td->trim_batch = td->o.trim_batch; | |
1088 | if (!td->trim_batch) | |
1089 | td->trim_batch = td->o.trim_backlog; | |
1090 | get_trim = 1; | |
1091 | } | |
1092 | ||
1093 | if (get_trim && !get_next_trim(td, io_u)) | |
1094 | return 1; | |
2ba1c290 | 1095 | } |
10ba535a | 1096 | |
0d29de83 JA |
1097 | return 0; |
1098 | } | |
1099 | ||
1100 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1101 | { | |
9e144189 JA |
1102 | if (td->o.verify_backlog && td->io_hist_len) { |
1103 | int get_verify = 0; | |
1104 | ||
1105 | if (td->verify_batch) { | |
1106 | td->verify_batch--; | |
1107 | get_verify = 1; | |
1108 | } else if (!(td->io_hist_len % td->o.verify_backlog) && | |
1109 | td->last_ddir != DDIR_READ) { | |
1110 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1111 | if (!td->verify_batch) |
1112 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1113 | get_verify = 1; |
1114 | } | |
1115 | ||
1116 | if (get_verify && !get_next_verify(td, io_u)) | |
0d29de83 | 1117 | return 1; |
9e144189 JA |
1118 | } |
1119 | ||
0d29de83 JA |
1120 | return 0; |
1121 | } | |
1122 | ||
de789769 JA |
1123 | /* |
1124 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1125 | * easy compressible data for simple de-dupe attempts. Do this for every |
1126 | * 512b block in the range, since that should be the smallest block size | |
1127 | * we can expect from a device. | |
de789769 JA |
1128 | */ |
1129 | static void small_content_scramble(struct io_u *io_u) | |
1130 | { | |
23f394d5 | 1131 | unsigned int i, nr_blocks = io_u->buflen / 512; |
fad82f76 | 1132 | unsigned long long boffset; |
23f394d5 JA |
1133 | unsigned int offset; |
1134 | void *p, *end; | |
de789769 | 1135 | |
23f394d5 JA |
1136 | if (!nr_blocks) |
1137 | return; | |
1138 | ||
1139 | p = io_u->xfer_buf; | |
fba76ee8 | 1140 | boffset = io_u->offset; |
fad82f76 | 1141 | |
23f394d5 JA |
1142 | for (i = 0; i < nr_blocks; i++) { |
1143 | /* | |
1144 | * Fill the byte offset into a "random" start offset of | |
1145 | * the buffer, given by the product of the usec time | |
1146 | * and the actual offset. | |
1147 | */ | |
fad82f76 | 1148 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
fba76ee8 | 1149 | offset &= ~(sizeof(unsigned long long) - 1); |
23f394d5 JA |
1150 | if (offset >= 512 - sizeof(unsigned long long)) |
1151 | offset -= sizeof(unsigned long long); | |
fba76ee8 | 1152 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1153 | |
1154 | end = p + 512 - sizeof(io_u->start_time); | |
1155 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1156 | p += 512; | |
fad82f76 | 1157 | boffset += 512; |
23f394d5 | 1158 | } |
de789769 JA |
1159 | } |
1160 | ||
0d29de83 JA |
1161 | /* |
1162 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1163 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1164 | */ | |
1165 | struct io_u *get_io_u(struct thread_data *td) | |
1166 | { | |
1167 | struct fio_file *f; | |
1168 | struct io_u *io_u; | |
de789769 | 1169 | int do_scramble = 0; |
0d29de83 JA |
1170 | |
1171 | io_u = __get_io_u(td); | |
1172 | if (!io_u) { | |
1173 | dprint(FD_IO, "__get_io_u failed\n"); | |
1174 | return NULL; | |
1175 | } | |
1176 | ||
1177 | if (check_get_verify(td, io_u)) | |
1178 | goto out; | |
1179 | if (check_get_trim(td, io_u)) | |
1180 | goto out; | |
1181 | ||
755200a3 JA |
1182 | /* |
1183 | * from a requeue, io_u already setup | |
1184 | */ | |
1185 | if (io_u->file) | |
77f392bf | 1186 | goto out; |
755200a3 | 1187 | |
429f6675 JA |
1188 | /* |
1189 | * If using an iolog, grab next piece if any available. | |
1190 | */ | |
1191 | if (td->o.read_iolog_file) { | |
1192 | if (read_iolog_get(td, io_u)) | |
1193 | goto err_put; | |
2ba1c290 JA |
1194 | } else if (set_io_u_file(td, io_u)) { |
1195 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 1196 | goto err_put; |
2ba1c290 | 1197 | } |
5ec10eaa | 1198 | |
429f6675 | 1199 | f = io_u->file; |
d6aed795 | 1200 | assert(fio_file_open(f)); |
97af62ce | 1201 | |
ff58fced | 1202 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1203 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1204 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1205 | goto err_put; |
2ba1c290 | 1206 | } |
10ba535a | 1207 | |
38dad62d | 1208 | f->last_start = io_u->offset; |
36167d82 | 1209 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1210 | |
fd68418e JA |
1211 | if (io_u->ddir == DDIR_WRITE) { |
1212 | if (td->o.verify != VERIFY_NONE) | |
1213 | populate_verify_io_u(td, io_u); | |
1214 | else if (td->o.refill_buffers) | |
1215 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
1216 | else if (td->o.scramble_buffers) | |
1217 | do_scramble = 1; | |
1218 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1219 | /* |
1220 | * Reset the buf_filled parameters so next time if the | |
1221 | * buffer is used for writes it is refilled. | |
1222 | */ | |
cbe8d756 RR |
1223 | io_u->buf_filled_len = 0; |
1224 | } | |
87dc1ab1 | 1225 | } |
10ba535a | 1226 | |
165faf16 JA |
1227 | /* |
1228 | * Set io data pointers. | |
1229 | */ | |
cec6b55d JA |
1230 | io_u->xfer_buf = io_u->buf; |
1231 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1232 | |
6ac7a331 | 1233 | out: |
0d29de83 | 1234 | assert(io_u->file); |
429f6675 | 1235 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1236 | if (!td->o.disable_slat) |
1237 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1238 | if (do_scramble) |
1239 | small_content_scramble(io_u); | |
429f6675 | 1240 | return io_u; |
36167d82 | 1241 | } |
429f6675 | 1242 | err_put: |
2ba1c290 | 1243 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
1244 | put_io_u(td, io_u); |
1245 | return NULL; | |
10ba535a JA |
1246 | } |
1247 | ||
5451792e JA |
1248 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1249 | { | |
825f818e JA |
1250 | const char *msg[] = { "read", "write", "sync", "datasync", |
1251 | "sync_file_range", "wait", "trim" }; | |
1252 | ||
1253 | ||
5451792e JA |
1254 | |
1255 | log_err("fio: io_u error"); | |
1256 | ||
1257 | if (io_u->file) | |
1258 | log_err(" on file %s", io_u->file->file_name); | |
1259 | ||
1260 | log_err(": %s\n", strerror(io_u->error)); | |
1261 | ||
5ec10eaa JA |
1262 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
1263 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1264 | |
1265 | if (!td->error) | |
1266 | td_verror(td, io_u->error, "io_u error"); | |
1267 | } | |
1268 | ||
97601024 JA |
1269 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
1270 | struct io_completion_data *icd) | |
10ba535a | 1271 | { |
dbad30b0 JA |
1272 | /* |
1273 | * Older gcc's are too dumb to realize that usec is always used | |
1274 | * initialized, silence that warning. | |
1275 | */ | |
1276 | unsigned long uninitialized_var(usec); | |
44f29692 | 1277 | struct fio_file *f; |
10ba535a | 1278 | |
2ba1c290 JA |
1279 | dprint_io_u(io_u, "io complete"); |
1280 | ||
2ecc1b57 | 1281 | td_io_u_lock(td); |
0c6e7517 | 1282 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1283 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
2ecc1b57 | 1284 | td_io_u_unlock(td); |
0c6e7517 | 1285 | |
5f9099ea | 1286 | if (ddir_sync(io_u->ddir)) { |
87dc1ab1 | 1287 | td->last_was_sync = 1; |
44f29692 JA |
1288 | f = io_u->file; |
1289 | if (f) { | |
1290 | f->first_write = -1ULL; | |
1291 | f->last_write = -1ULL; | |
1292 | } | |
87dc1ab1 JA |
1293 | return; |
1294 | } | |
1295 | ||
1296 | td->last_was_sync = 0; | |
9e144189 | 1297 | td->last_ddir = io_u->ddir; |
87dc1ab1 | 1298 | |
ff58fced | 1299 | if (!io_u->error && ddir_rw(io_u->ddir)) { |
10ba535a | 1300 | unsigned int bytes = io_u->buflen - io_u->resid; |
1e97cce9 | 1301 | const enum fio_ddir idx = io_u->ddir; |
ba3e4e0c | 1302 | const enum fio_ddir odx = io_u->ddir ^ 1; |
b29ee5b3 | 1303 | int ret; |
10ba535a | 1304 | |
b29ee5b3 JA |
1305 | td->io_blocks[idx]++; |
1306 | td->io_bytes[idx] += bytes; | |
1307 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 1308 | |
44f29692 JA |
1309 | if (idx == DDIR_WRITE) { |
1310 | f = io_u->file; | |
1311 | if (f) { | |
1312 | if (f->first_write == -1ULL || | |
1313 | io_u->offset < f->first_write) | |
1314 | f->first_write = io_u->offset; | |
1315 | if (f->last_write == -1ULL || | |
1316 | ((io_u->offset + bytes) > f->last_write)) | |
1317 | f->last_write = io_u->offset + bytes; | |
1318 | } | |
1319 | } | |
1320 | ||
b29ee5b3 | 1321 | if (ramp_time_over(td)) { |
40e1a6f0 | 1322 | unsigned long uninitialized_var(lusec); |
40e1a6f0 JA |
1323 | |
1324 | if (!td->o.disable_clat || !td->o.disable_bw) | |
1325 | lusec = utime_since(&io_u->issue_time, | |
1326 | &icd->time); | |
02af0988 JA |
1327 | if (!td->o.disable_lat) { |
1328 | unsigned long tusec; | |
721938ae | 1329 | |
02af0988 JA |
1330 | tusec = utime_since(&io_u->start_time, |
1331 | &icd->time); | |
1332 | add_lat_sample(td, idx, tusec, bytes); | |
1333 | } | |
9520ebb9 | 1334 | if (!td->o.disable_clat) { |
01a1caa5 | 1335 | add_clat_sample(td, idx, lusec, bytes); |
40e1a6f0 | 1336 | io_u_mark_latency(td, lusec); |
9520ebb9 JA |
1337 | } |
1338 | if (!td->o.disable_bw) | |
306ddc97 | 1339 | add_bw_sample(td, idx, bytes, &icd->time); |
b23b6a2f | 1340 | if (__should_check_rate(td, idx)) { |
ba3e4e0c RR |
1341 | td->rate_pending_usleep[idx] = |
1342 | ((td->this_io_bytes[idx] * | |
1343 | td->rate_nsec_cycle[idx]) / 1000 - | |
1344 | utime_since_now(&td->start)); | |
b23b6a2f | 1345 | } |
581e7141 | 1346 | if (__should_check_rate(td, idx ^ 1)) |
ba3e4e0c RR |
1347 | td->rate_pending_usleep[odx] = |
1348 | ((td->this_io_bytes[odx] * | |
1349 | td->rate_nsec_cycle[odx]) / 1000 - | |
1350 | utime_since_now(&td->start)); | |
721938ae | 1351 | } |
10ba535a | 1352 | |
660a1cb5 | 1353 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 1354 | td->o.do_verify && |
41128405 | 1355 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
1356 | log_io_piece(td, io_u); |
1357 | ||
b29ee5b3 | 1358 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 1359 | |
d7762cf8 | 1360 | if (io_u->end_io) { |
36690c9b | 1361 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
1362 | if (ret && !icd->error) |
1363 | icd->error = ret; | |
1364 | } | |
ff58fced | 1365 | } else if (io_u->error) { |
10ba535a | 1366 | icd->error = io_u->error; |
5451792e JA |
1367 | io_u_log_error(td, io_u); |
1368 | } | |
f2bba182 RR |
1369 | if (td->o.continue_on_error && icd->error && |
1370 | td_non_fatal_error(icd->error)) { | |
1371 | /* | |
1372 | * If there is a non_fatal error, then add to the error count | |
1373 | * and clear all the errors. | |
1374 | */ | |
1375 | update_error_count(td, icd->error); | |
1376 | td_clear_error(td); | |
1377 | icd->error = 0; | |
1378 | io_u->error = 0; | |
1379 | } | |
10ba535a JA |
1380 | } |
1381 | ||
9520ebb9 JA |
1382 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1383 | int nr) | |
10ba535a | 1384 | { |
9520ebb9 JA |
1385 | if (!td->o.disable_clat || !td->o.disable_bw) |
1386 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 1387 | |
3af6ef39 JA |
1388 | icd->nr = nr; |
1389 | ||
10ba535a JA |
1390 | icd->error = 0; |
1391 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
1392 | } |
1393 | ||
97601024 JA |
1394 | static void ios_completed(struct thread_data *td, |
1395 | struct io_completion_data *icd) | |
36167d82 JA |
1396 | { |
1397 | struct io_u *io_u; | |
1398 | int i; | |
1399 | ||
10ba535a JA |
1400 | for (i = 0; i < icd->nr; i++) { |
1401 | io_u = td->io_ops->event(td, i); | |
1402 | ||
1403 | io_completed(td, io_u, icd); | |
e8462bd8 JA |
1404 | |
1405 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1406 | put_io_u(td, io_u); | |
10ba535a JA |
1407 | } |
1408 | } | |
97601024 | 1409 | |
e7e6cfb4 JA |
1410 | /* |
1411 | * Complete a single io_u for the sync engines. | |
1412 | */ | |
581e7141 JA |
1413 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
1414 | unsigned long *bytes) | |
97601024 JA |
1415 | { |
1416 | struct io_completion_data icd; | |
1417 | ||
9520ebb9 | 1418 | init_icd(td, &icd, 1); |
97601024 | 1419 | io_completed(td, io_u, &icd); |
e8462bd8 JA |
1420 | |
1421 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1422 | put_io_u(td, io_u); | |
97601024 | 1423 | |
581e7141 JA |
1424 | if (icd.error) { |
1425 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1426 | return -1; | |
1427 | } | |
97601024 | 1428 | |
581e7141 JA |
1429 | if (bytes) { |
1430 | bytes[0] += icd.bytes_done[0]; | |
1431 | bytes[1] += icd.bytes_done[1]; | |
1432 | } | |
1433 | ||
1434 | return 0; | |
97601024 JA |
1435 | } |
1436 | ||
e7e6cfb4 JA |
1437 | /* |
1438 | * Called to complete min_events number of io for the async engines. | |
1439 | */ | |
581e7141 JA |
1440 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
1441 | unsigned long *bytes) | |
97601024 | 1442 | { |
97601024 | 1443 | struct io_completion_data icd; |
00de55ef | 1444 | struct timespec *tvp = NULL; |
97601024 | 1445 | int ret; |
4d06a338 | 1446 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1447 | |
4950421a | 1448 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1449 | |
4950421a | 1450 | if (!min_evts) |
00de55ef | 1451 | tvp = &ts; |
97601024 | 1452 | |
4950421a | 1453 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1454 | if (ret < 0) { |
e1161c32 | 1455 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1456 | return ret; |
1457 | } else if (!ret) | |
1458 | return ret; | |
1459 | ||
9520ebb9 | 1460 | init_icd(td, &icd, ret); |
97601024 | 1461 | ios_completed(td, &icd); |
581e7141 JA |
1462 | if (icd.error) { |
1463 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1464 | return -1; | |
1465 | } | |
97601024 | 1466 | |
581e7141 JA |
1467 | if (bytes) { |
1468 | bytes[0] += icd.bytes_done[0]; | |
1469 | bytes[1] += icd.bytes_done[1]; | |
1470 | } | |
1471 | ||
1472 | return 0; | |
97601024 | 1473 | } |
7e77dd02 JA |
1474 | |
1475 | /* | |
1476 | * Call when io_u is really queued, to update the submission latency. | |
1477 | */ | |
1478 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1479 | { | |
9520ebb9 JA |
1480 | if (!td->o.disable_slat) { |
1481 | unsigned long slat_time; | |
7e77dd02 | 1482 | |
9520ebb9 | 1483 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
29a90ddb | 1484 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen); |
9520ebb9 | 1485 | } |
7e77dd02 | 1486 | } |
433afcb4 | 1487 | |
5973cafb JA |
1488 | /* |
1489 | * "randomly" fill the buffer contents | |
1490 | */ | |
1491 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1492 | unsigned int max_bs) | |
1493 | { | |
e84b371e JA |
1494 | io_u->buf_filled_len = 0; |
1495 | ||
637ef8d9 | 1496 | if (!td->o.zero_buffers) |
3545a109 | 1497 | fill_random_buf(&td->buf_state, io_u->buf, max_bs); |
637ef8d9 JA |
1498 | else |
1499 | memset(io_u->buf, 0, max_bs); | |
5973cafb | 1500 | } |