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) { |
059b0802 JA |
252 | unsigned long long pos = f->last_pos - f->file_offset; |
253 | ||
254 | if (pos) | |
255 | pos += td->o.ddir_seq_add; | |
256 | ||
257 | *b = pos / td->o.min_bs[ddir]; | |
38dad62d JA |
258 | return 0; |
259 | } | |
260 | ||
261 | return 1; | |
262 | } | |
263 | ||
264 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
265 | enum fio_ddir ddir, int rw_seq, unsigned long long *b) | |
266 | { | |
267 | struct fio_file *f = io_u->file; | |
268 | int ret; | |
269 | ||
ff58fced JA |
270 | assert(ddir_rw(ddir)); |
271 | ||
38dad62d JA |
272 | if (rw_seq) { |
273 | if (td_random(td)) | |
274 | ret = get_next_rand_block(td, f, ddir, b); | |
275 | else | |
276 | ret = get_next_seq_block(td, f, ddir, b); | |
277 | } else { | |
278 | io_u->flags |= IO_U_F_BUSY_OK; | |
279 | ||
280 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
281 | ret = get_next_seq_block(td, f, ddir, b); | |
282 | if (ret) | |
283 | ret = get_next_rand_block(td, f, ddir, b); | |
284 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { | |
285 | if (f->last_start != -1ULL) | |
0508320e DN |
286 | *b = (f->last_start - f->file_offset) |
287 | / td->o.min_bs[ddir]; | |
38dad62d JA |
288 | else |
289 | *b = 0; | |
290 | ret = 0; | |
291 | } else { | |
292 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
293 | ret = 1; | |
294 | } | |
295 | } | |
296 | ||
297 | return ret; | |
298 | } | |
299 | ||
10ba535a JA |
300 | /* |
301 | * For random io, generate a random new block and see if it's used. Repeat | |
302 | * until we find a free one. For sequential io, just return the end of | |
303 | * the last io issued. | |
304 | */ | |
15dc1934 | 305 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 306 | { |
9bf2061e | 307 | struct fio_file *f = io_u->file; |
ec4015da | 308 | unsigned long long b; |
4ba66134 | 309 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 310 | int rw_seq_hit = 0; |
10ba535a | 311 | |
ff58fced JA |
312 | assert(ddir_rw(ddir)); |
313 | ||
38dad62d JA |
314 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
315 | rw_seq_hit = 1; | |
5736c10d | 316 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 317 | } |
211097b2 | 318 | |
0508320e | 319 | if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) |
38dad62d | 320 | return 1; |
10ba535a | 321 | |
2b7a01d0 | 322 | io_u->offset = b * td->o.ba[ddir]; |
009bd847 JA |
323 | if (io_u->offset >= f->io_size) { |
324 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
325 | io_u->offset, f->io_size); | |
326 | return 1; | |
327 | } | |
328 | ||
329 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
330 | if (io_u->offset >= f->real_file_size) { |
331 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
332 | io_u->offset, f->real_file_size); | |
10ba535a | 333 | return 1; |
2ba1c290 | 334 | } |
10ba535a JA |
335 | |
336 | return 0; | |
337 | } | |
338 | ||
15dc1934 JA |
339 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
340 | { | |
7eb36574 JA |
341 | struct prof_io_ops *ops = &td->prof_io_ops; |
342 | ||
343 | if (ops->fill_io_u_off) | |
344 | return ops->fill_io_u_off(td, io_u); | |
15dc1934 JA |
345 | |
346 | return __get_next_offset(td, io_u); | |
347 | } | |
348 | ||
79944128 JA |
349 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
350 | unsigned int buflen) | |
351 | { | |
352 | struct fio_file *f = io_u->file; | |
353 | ||
354 | return io_u->offset + buflen <= f->io_size + td->o.start_offset; | |
355 | } | |
356 | ||
15dc1934 | 357 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 358 | { |
bca4ed4d | 359 | const int ddir = io_u->ddir; |
f3f552b9 | 360 | unsigned int uninitialized_var(buflen); |
f3059de1 | 361 | unsigned int minbs, maxbs; |
1294c3ec | 362 | unsigned long r, rand_max; |
10ba535a | 363 | |
ff58fced JA |
364 | assert(ddir_rw(ddir)); |
365 | ||
f3059de1 JA |
366 | minbs = td->o.min_bs[ddir]; |
367 | maxbs = td->o.max_bs[ddir]; | |
368 | ||
79944128 JA |
369 | if (minbs == maxbs) |
370 | return minbs; | |
371 | ||
4c07ad86 JA |
372 | if (td->o.use_os_rand) |
373 | rand_max = OS_RAND_MAX; | |
374 | else | |
375 | rand_max = FRAND_MAX; | |
376 | ||
79944128 | 377 | do { |
4c07ad86 JA |
378 | if (td->o.use_os_rand) |
379 | r = os_random_long(&td->bsrange_state); | |
380 | else | |
381 | r = __rand(&td->__bsrange_state); | |
382 | ||
720e84ad | 383 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 384 | buflen = 1 + (unsigned int) ((double) maxbs * |
4c07ad86 | 385 | (r / (rand_max + 1.0))); |
f3059de1 JA |
386 | if (buflen < minbs) |
387 | buflen = minbs; | |
5ec10eaa | 388 | } else { |
564ca972 JA |
389 | long perc = 0; |
390 | unsigned int i; | |
391 | ||
720e84ad JA |
392 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
393 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
394 | |
395 | buflen = bsp->bs; | |
396 | perc += bsp->perc; | |
79944128 JA |
397 | if ((r <= ((rand_max / 100L) * perc)) && |
398 | io_u_fits(td, io_u, buflen)) | |
564ca972 JA |
399 | break; |
400 | } | |
401 | } | |
79944128 | 402 | |
f3059de1 JA |
403 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
404 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 405 | |
79944128 | 406 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 407 | |
10ba535a JA |
408 | return buflen; |
409 | } | |
410 | ||
15dc1934 JA |
411 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
412 | { | |
7eb36574 JA |
413 | struct prof_io_ops *ops = &td->prof_io_ops; |
414 | ||
415 | if (ops->fill_io_u_size) | |
416 | return ops->fill_io_u_size(td, io_u); | |
15dc1934 JA |
417 | |
418 | return __get_next_buflen(td, io_u); | |
419 | } | |
420 | ||
afe24a5a JA |
421 | static void set_rwmix_bytes(struct thread_data *td) |
422 | { | |
afe24a5a JA |
423 | unsigned int diff; |
424 | ||
425 | /* | |
426 | * we do time or byte based switch. this is needed because | |
427 | * buffered writes may issue a lot quicker than they complete, | |
428 | * whereas reads do not. | |
429 | */ | |
e47f799f | 430 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 431 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
432 | } |
433 | ||
434 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
435 | { | |
436 | unsigned int v; | |
1294c3ec | 437 | unsigned long r; |
e47f799f | 438 | |
4c07ad86 JA |
439 | if (td->o.use_os_rand) { |
440 | r = os_random_long(&td->rwmix_state); | |
441 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); | |
442 | } else { | |
443 | r = __rand(&td->__rwmix_state); | |
444 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
445 | } | |
446 | ||
04c540d9 | 447 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
448 | return DDIR_READ; |
449 | ||
450 | return DDIR_WRITE; | |
afe24a5a JA |
451 | } |
452 | ||
581e7141 JA |
453 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
454 | { | |
455 | enum fio_ddir odir = ddir ^ 1; | |
456 | struct timeval t; | |
457 | long usec; | |
458 | ||
ff58fced JA |
459 | assert(ddir_rw(ddir)); |
460 | ||
581e7141 JA |
461 | if (td->rate_pending_usleep[ddir] <= 0) |
462 | return ddir; | |
463 | ||
464 | /* | |
465 | * We have too much pending sleep in this direction. See if we | |
466 | * should switch. | |
467 | */ | |
468 | if (td_rw(td)) { | |
469 | /* | |
470 | * Other direction does not have too much pending, switch | |
471 | */ | |
472 | if (td->rate_pending_usleep[odir] < 100000) | |
473 | return odir; | |
474 | ||
475 | /* | |
476 | * Both directions have pending sleep. Sleep the minimum time | |
477 | * and deduct from both. | |
478 | */ | |
479 | if (td->rate_pending_usleep[ddir] <= | |
480 | td->rate_pending_usleep[odir]) { | |
481 | usec = td->rate_pending_usleep[ddir]; | |
482 | } else { | |
483 | usec = td->rate_pending_usleep[odir]; | |
484 | ddir = odir; | |
485 | } | |
486 | } else | |
487 | usec = td->rate_pending_usleep[ddir]; | |
488 | ||
78c1eda5 JA |
489 | /* |
490 | * We are going to sleep, ensure that we flush anything pending as | |
491 | * not to skew our latency numbers | |
492 | */ | |
493 | if (td->cur_depth) { | |
494 | int fio_unused ret; | |
495 | ||
496 | ret = io_u_queued_complete(td, td->cur_depth, NULL); | |
497 | } | |
498 | ||
581e7141 JA |
499 | fio_gettime(&t, NULL); |
500 | usec_sleep(td, usec); | |
501 | usec = utime_since_now(&t); | |
502 | ||
503 | td->rate_pending_usleep[ddir] -= usec; | |
504 | ||
505 | odir = ddir ^ 1; | |
506 | if (td_rw(td) && __should_check_rate(td, odir)) | |
507 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 508 | |
581e7141 JA |
509 | return ddir; |
510 | } | |
511 | ||
10ba535a JA |
512 | /* |
513 | * Return the data direction for the next io_u. If the job is a | |
514 | * mixed read/write workload, check the rwmix cycle and switch if | |
515 | * necessary. | |
516 | */ | |
1e97cce9 | 517 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 518 | { |
581e7141 JA |
519 | enum fio_ddir ddir; |
520 | ||
5f9099ea JA |
521 | /* |
522 | * see if it's time to fsync | |
523 | */ | |
524 | if (td->o.fsync_blocks && | |
525 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
526 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
527 | return DDIR_SYNC; | |
528 | ||
529 | /* | |
530 | * see if it's time to fdatasync | |
531 | */ | |
532 | if (td->o.fdatasync_blocks && | |
533 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
534 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
535 | return DDIR_DATASYNC; | |
536 | ||
44f29692 JA |
537 | /* |
538 | * see if it's time to sync_file_range | |
539 | */ | |
540 | if (td->sync_file_range_nr && | |
541 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
542 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
543 | return DDIR_SYNC_FILE_RANGE; | |
544 | ||
10ba535a | 545 | if (td_rw(td)) { |
10ba535a JA |
546 | /* |
547 | * Check if it's time to seed a new data direction. | |
548 | */ | |
e4928662 | 549 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
550 | /* |
551 | * Put a top limit on how many bytes we do for | |
552 | * one data direction, to avoid overflowing the | |
553 | * ranges too much | |
554 | */ | |
555 | ddir = get_rand_ddir(td); | |
e47f799f JA |
556 | |
557 | if (ddir != td->rwmix_ddir) | |
558 | set_rwmix_bytes(td); | |
559 | ||
560 | td->rwmix_ddir = ddir; | |
10ba535a | 561 | } |
581e7141 | 562 | ddir = td->rwmix_ddir; |
10ba535a | 563 | } else if (td_read(td)) |
581e7141 | 564 | ddir = DDIR_READ; |
10ba535a | 565 | else |
581e7141 JA |
566 | ddir = DDIR_WRITE; |
567 | ||
568 | td->rwmix_ddir = rate_ddir(td, ddir); | |
569 | return td->rwmix_ddir; | |
10ba535a JA |
570 | } |
571 | ||
1ef2b6be JA |
572 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
573 | { | |
574 | io_u->ddir = get_rw_ddir(td); | |
575 | ||
576 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
577 | td->o.barrier_blocks && | |
578 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
579 | td->io_issues[DDIR_WRITE]) | |
580 | io_u->flags |= IO_U_F_BARRIER; | |
581 | } | |
582 | ||
e8462bd8 | 583 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 JA |
584 | { |
585 | int ret = put_file(td, f); | |
586 | ||
587 | if (ret) | |
588 | td_verror(td, ret, "file close"); | |
589 | } | |
590 | ||
10ba535a JA |
591 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
592 | { | |
e8462bd8 JA |
593 | td_io_u_lock(td); |
594 | ||
0c6e7517 | 595 | io_u->flags |= IO_U_F_FREE; |
e8462bd8 | 596 | io_u->flags &= ~IO_U_F_FREE_DEF; |
0c6e7517 | 597 | |
60f2c658 JA |
598 | if (io_u->file) |
599 | put_file_log(td, io_u->file); | |
2dbdab7e | 600 | |
10ba535a | 601 | io_u->file = NULL; |
0c41214f RR |
602 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
603 | td->cur_depth--; | |
e8462bd8 | 604 | flist_del_init(&io_u->list); |
01743ee1 | 605 | flist_add(&io_u->list, &td->io_u_freelist); |
e8462bd8 JA |
606 | td_io_u_unlock(td); |
607 | td_io_u_free_notify(td); | |
10ba535a JA |
608 | } |
609 | ||
f2bba182 RR |
610 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
611 | { | |
612 | io_u->flags &= ~IO_U_F_FLIGHT; | |
613 | put_io_u(td, io_u); | |
614 | } | |
615 | ||
755200a3 JA |
616 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
617 | { | |
618 | struct io_u *__io_u = *io_u; | |
619 | ||
465221b0 JA |
620 | dprint(FD_IO, "requeue %p\n", __io_u); |
621 | ||
e8462bd8 JA |
622 | td_io_u_lock(td); |
623 | ||
4d2e0f49 | 624 | __io_u->flags |= IO_U_F_FREE; |
ff58fced | 625 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir)) |
e4f54adb | 626 | td->io_issues[__io_u->ddir]--; |
5ec10eaa | 627 | |
4d2e0f49 | 628 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
629 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
630 | td->cur_depth--; | |
01743ee1 JA |
631 | flist_del(&__io_u->list); |
632 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
e8462bd8 | 633 | td_io_u_unlock(td); |
755200a3 JA |
634 | *io_u = NULL; |
635 | } | |
636 | ||
9bf2061e | 637 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 638 | { |
b4c5e1ac JA |
639 | if (td->io_ops->flags & FIO_NOIO) |
640 | goto out; | |
641 | ||
1ef2b6be | 642 | set_rw_ddir(td, io_u); |
5f9099ea | 643 | |
87dc1ab1 | 644 | /* |
ff58fced | 645 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 646 | */ |
ff58fced | 647 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 648 | goto out; |
a00735e6 | 649 | |
48f5abd3 JA |
650 | /* |
651 | * See if it's time to switch to a new zone | |
652 | */ | |
653 | if (td->zone_bytes >= td->o.zone_size) { | |
654 | td->zone_bytes = 0; | |
655 | io_u->file->last_pos += td->o.zone_skip; | |
656 | td->io_skip_bytes += td->o.zone_skip; | |
657 | } | |
658 | ||
10ba535a | 659 | /* |
c685b5b2 JA |
660 | * No log, let the seq/rand engine retrieve the next buflen and |
661 | * position. | |
10ba535a | 662 | */ |
2ba1c290 JA |
663 | if (get_next_offset(td, io_u)) { |
664 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 665 | return 1; |
2ba1c290 | 666 | } |
10ba535a | 667 | |
9bf2061e | 668 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
669 | if (!io_u->buflen) { |
670 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 671 | return 1; |
2ba1c290 | 672 | } |
bca4ed4d | 673 | |
2ba1c290 JA |
674 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
675 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
676 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
677 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 678 | return 1; |
2ba1c290 | 679 | } |
6a5e6884 | 680 | |
bca4ed4d JA |
681 | /* |
682 | * mark entry before potentially trimming io_u | |
683 | */ | |
303032ae | 684 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 685 | mark_random_map(td, io_u); |
bca4ed4d JA |
686 | |
687 | /* | |
688 | * If using a write iolog, store this entry. | |
689 | */ | |
c38e9468 | 690 | out: |
2ba1c290 | 691 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 692 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 693 | log_io_u(td, io_u); |
bca4ed4d | 694 | return 0; |
10ba535a JA |
695 | } |
696 | ||
838bc709 JA |
697 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
698 | { | |
2b13e716 | 699 | int idx = 0; |
838bc709 JA |
700 | |
701 | switch (nr) { | |
702 | default: | |
2b13e716 | 703 | idx = 6; |
838bc709 JA |
704 | break; |
705 | case 33 ... 64: | |
2b13e716 | 706 | idx = 5; |
838bc709 JA |
707 | break; |
708 | case 17 ... 32: | |
2b13e716 | 709 | idx = 4; |
838bc709 JA |
710 | break; |
711 | case 9 ... 16: | |
2b13e716 | 712 | idx = 3; |
838bc709 JA |
713 | break; |
714 | case 5 ... 8: | |
2b13e716 | 715 | idx = 2; |
838bc709 JA |
716 | break; |
717 | case 1 ... 4: | |
2b13e716 | 718 | idx = 1; |
838bc709 JA |
719 | case 0: |
720 | break; | |
721 | } | |
722 | ||
2b13e716 | 723 | map[idx]++; |
838bc709 JA |
724 | } |
725 | ||
726 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
727 | { | |
728 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
729 | td->ts.total_submit++; | |
730 | } | |
731 | ||
732 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
733 | { | |
734 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
735 | td->ts.total_complete++; | |
736 | } | |
737 | ||
d8005759 | 738 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 739 | { |
2b13e716 | 740 | int idx = 0; |
71619dc2 JA |
741 | |
742 | switch (td->cur_depth) { | |
743 | default: | |
2b13e716 | 744 | idx = 6; |
a783e61a | 745 | break; |
71619dc2 | 746 | case 32 ... 63: |
2b13e716 | 747 | idx = 5; |
a783e61a | 748 | break; |
71619dc2 | 749 | case 16 ... 31: |
2b13e716 | 750 | idx = 4; |
a783e61a | 751 | break; |
71619dc2 | 752 | case 8 ... 15: |
2b13e716 | 753 | idx = 3; |
a783e61a | 754 | break; |
71619dc2 | 755 | case 4 ... 7: |
2b13e716 | 756 | idx = 2; |
a783e61a | 757 | break; |
71619dc2 | 758 | case 2 ... 3: |
2b13e716 | 759 | idx = 1; |
71619dc2 JA |
760 | case 1: |
761 | break; | |
762 | } | |
763 | ||
2b13e716 | 764 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
765 | } |
766 | ||
04a0feae JA |
767 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
768 | { | |
2b13e716 | 769 | int idx = 0; |
04a0feae JA |
770 | |
771 | assert(usec < 1000); | |
772 | ||
773 | switch (usec) { | |
774 | case 750 ... 999: | |
2b13e716 | 775 | idx = 9; |
04a0feae JA |
776 | break; |
777 | case 500 ... 749: | |
2b13e716 | 778 | idx = 8; |
04a0feae JA |
779 | break; |
780 | case 250 ... 499: | |
2b13e716 | 781 | idx = 7; |
04a0feae JA |
782 | break; |
783 | case 100 ... 249: | |
2b13e716 | 784 | idx = 6; |
04a0feae JA |
785 | break; |
786 | case 50 ... 99: | |
2b13e716 | 787 | idx = 5; |
04a0feae JA |
788 | break; |
789 | case 20 ... 49: | |
2b13e716 | 790 | idx = 4; |
04a0feae JA |
791 | break; |
792 | case 10 ... 19: | |
2b13e716 | 793 | idx = 3; |
04a0feae JA |
794 | break; |
795 | case 4 ... 9: | |
2b13e716 | 796 | idx = 2; |
04a0feae JA |
797 | break; |
798 | case 2 ... 3: | |
2b13e716 | 799 | idx = 1; |
04a0feae JA |
800 | case 0 ... 1: |
801 | break; | |
802 | } | |
803 | ||
2b13e716 JA |
804 | assert(idx < FIO_IO_U_LAT_U_NR); |
805 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
806 | } |
807 | ||
808 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 809 | { |
2b13e716 | 810 | int idx = 0; |
ec118304 JA |
811 | |
812 | switch (msec) { | |
813 | default: | |
2b13e716 | 814 | idx = 11; |
04a0feae | 815 | break; |
8abdce66 | 816 | case 1000 ... 1999: |
2b13e716 | 817 | idx = 10; |
04a0feae | 818 | break; |
8abdce66 | 819 | case 750 ... 999: |
2b13e716 | 820 | idx = 9; |
04a0feae | 821 | break; |
8abdce66 | 822 | case 500 ... 749: |
2b13e716 | 823 | idx = 8; |
04a0feae | 824 | break; |
8abdce66 | 825 | case 250 ... 499: |
2b13e716 | 826 | idx = 7; |
04a0feae | 827 | break; |
8abdce66 | 828 | case 100 ... 249: |
2b13e716 | 829 | idx = 6; |
04a0feae | 830 | break; |
8abdce66 | 831 | case 50 ... 99: |
2b13e716 | 832 | idx = 5; |
04a0feae | 833 | break; |
8abdce66 | 834 | case 20 ... 49: |
2b13e716 | 835 | idx = 4; |
04a0feae | 836 | break; |
8abdce66 | 837 | case 10 ... 19: |
2b13e716 | 838 | idx = 3; |
04a0feae | 839 | break; |
8abdce66 | 840 | case 4 ... 9: |
2b13e716 | 841 | idx = 2; |
04a0feae | 842 | break; |
ec118304 | 843 | case 2 ... 3: |
2b13e716 | 844 | idx = 1; |
ec118304 JA |
845 | case 0 ... 1: |
846 | break; | |
847 | } | |
848 | ||
2b13e716 JA |
849 | assert(idx < FIO_IO_U_LAT_M_NR); |
850 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
851 | } |
852 | ||
853 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
854 | { | |
855 | if (usec < 1000) | |
856 | io_u_mark_lat_usec(td, usec); | |
857 | else | |
858 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
859 | } |
860 | ||
0aabe160 JA |
861 | /* |
862 | * Get next file to service by choosing one at random | |
863 | */ | |
2cc52930 JA |
864 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
865 | enum fio_file_flags goodf, | |
d6aed795 | 866 | enum fio_file_flags badf) |
0aabe160 | 867 | { |
0aabe160 | 868 | struct fio_file *f; |
1c178180 | 869 | int fno; |
0aabe160 JA |
870 | |
871 | do { | |
87b10676 | 872 | int opened = 0; |
1294c3ec | 873 | unsigned long r; |
4c07ad86 JA |
874 | |
875 | if (td->o.use_os_rand) { | |
876 | r = os_random_long(&td->next_file_state); | |
877 | fno = (unsigned int) ((double) td->o.nr_files | |
878 | * (r / (OS_RAND_MAX + 1.0))); | |
879 | } else { | |
880 | r = __rand(&td->__next_file_state); | |
881 | fno = (unsigned int) ((double) td->o.nr_files | |
882 | * (r / (FRAND_MAX + 1.0))); | |
883 | } | |
7c83c089 | 884 | |
126d65c6 | 885 | f = td->files[fno]; |
d6aed795 | 886 | if (fio_file_done(f)) |
059e63c0 | 887 | continue; |
1c178180 | 888 | |
d6aed795 | 889 | if (!fio_file_open(f)) { |
87b10676 JA |
890 | int err; |
891 | ||
892 | err = td_io_open_file(td, f); | |
893 | if (err) | |
894 | continue; | |
895 | opened = 1; | |
896 | } | |
897 | ||
2ba1c290 JA |
898 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
899 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 900 | return f; |
2ba1c290 | 901 | } |
87b10676 JA |
902 | if (opened) |
903 | td_io_close_file(td, f); | |
0aabe160 JA |
904 | } while (1); |
905 | } | |
906 | ||
907 | /* | |
908 | * Get next file to service by doing round robin between all available ones | |
909 | */ | |
1c178180 JA |
910 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
911 | int badf) | |
3d7c391d JA |
912 | { |
913 | unsigned int old_next_file = td->next_file; | |
914 | struct fio_file *f; | |
915 | ||
916 | do { | |
87b10676 JA |
917 | int opened = 0; |
918 | ||
126d65c6 | 919 | f = td->files[td->next_file]; |
3d7c391d JA |
920 | |
921 | td->next_file++; | |
2dc1bbeb | 922 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
923 | td->next_file = 0; |
924 | ||
87b10676 | 925 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 926 | if (fio_file_done(f)) { |
d5ed68ea | 927 | f = NULL; |
059e63c0 | 928 | continue; |
d5ed68ea | 929 | } |
059e63c0 | 930 | |
d6aed795 | 931 | if (!fio_file_open(f)) { |
87b10676 JA |
932 | int err; |
933 | ||
934 | err = td_io_open_file(td, f); | |
b5696bfc JA |
935 | if (err) { |
936 | dprint(FD_FILE, "error %d on open of %s\n", | |
937 | err, f->file_name); | |
87c27b45 | 938 | f = NULL; |
87b10676 | 939 | continue; |
b5696bfc | 940 | } |
87b10676 JA |
941 | opened = 1; |
942 | } | |
943 | ||
0b9d69ec JA |
944 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
945 | f->flags); | |
1c178180 | 946 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
947 | break; |
948 | ||
87b10676 JA |
949 | if (opened) |
950 | td_io_close_file(td, f); | |
951 | ||
3d7c391d JA |
952 | f = NULL; |
953 | } while (td->next_file != old_next_file); | |
954 | ||
2ba1c290 | 955 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
956 | return f; |
957 | } | |
958 | ||
7eb36574 | 959 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 960 | { |
1907dbc6 JA |
961 | struct fio_file *f; |
962 | ||
2dc1bbeb | 963 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 964 | |
b5696bfc | 965 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
966 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
967 | " nr_files=%d\n", td->nr_open_files, | |
968 | td->nr_done_files, | |
969 | td->o.nr_files); | |
bdb4e2e9 | 970 | return NULL; |
2ba1c290 | 971 | } |
bdb4e2e9 | 972 | |
1907dbc6 | 973 | f = td->file_service_file; |
d6aed795 | 974 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
975 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
976 | goto out; | |
977 | if (td->file_service_left--) | |
978 | goto out; | |
979 | } | |
1907dbc6 | 980 | |
a086c257 JA |
981 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
982 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 983 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 984 | else |
d6aed795 | 985 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 JA |
986 | |
987 | td->file_service_file = f; | |
988 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 989 | out: |
683023e8 | 990 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 991 | return f; |
bdb4e2e9 JA |
992 | } |
993 | ||
7eb36574 JA |
994 | static struct fio_file *get_next_file(struct thread_data *td) |
995 | { | |
996 | struct prof_io_ops *ops = &td->prof_io_ops; | |
997 | ||
998 | if (ops->get_next_file) | |
999 | return ops->get_next_file(td); | |
1000 | ||
1001 | return __get_next_file(td); | |
1002 | } | |
1003 | ||
429f6675 JA |
1004 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
1005 | { | |
1006 | struct fio_file *f; | |
1007 | ||
1008 | do { | |
1009 | f = get_next_file(td); | |
1010 | if (!f) | |
1011 | return 1; | |
1012 | ||
429f6675 JA |
1013 | io_u->file = f; |
1014 | get_file(f); | |
1015 | ||
1016 | if (!fill_io_u(td, io_u)) | |
1017 | break; | |
1018 | ||
b5696bfc | 1019 | put_file_log(td, f); |
429f6675 | 1020 | td_io_close_file(td, f); |
b5696bfc | 1021 | io_u->file = NULL; |
d6aed795 | 1022 | fio_file_set_done(f); |
429f6675 | 1023 | td->nr_done_files++; |
0b9d69ec JA |
1024 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1025 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1026 | } while (1); |
1027 | ||
1028 | return 0; | |
1029 | } | |
1030 | ||
1031 | ||
10ba535a JA |
1032 | struct io_u *__get_io_u(struct thread_data *td) |
1033 | { | |
1034 | struct io_u *io_u = NULL; | |
1035 | ||
e8462bd8 JA |
1036 | td_io_u_lock(td); |
1037 | ||
1038 | again: | |
01743ee1 JA |
1039 | if (!flist_empty(&td->io_u_requeues)) |
1040 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 1041 | else if (!queue_full(td)) { |
01743ee1 | 1042 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 1043 | |
6040dabc | 1044 | io_u->buflen = 0; |
10ba535a | 1045 | io_u->resid = 0; |
755200a3 | 1046 | io_u->file = NULL; |
d7762cf8 | 1047 | io_u->end_io = NULL; |
755200a3 JA |
1048 | } |
1049 | ||
1050 | if (io_u) { | |
0c6e7517 | 1051 | assert(io_u->flags & IO_U_F_FREE); |
2ecc1b57 | 1052 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF); |
1ef2b6be | 1053 | io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER); |
0c6e7517 | 1054 | |
755200a3 | 1055 | io_u->error = 0; |
01743ee1 JA |
1056 | flist_del(&io_u->list); |
1057 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a | 1058 | td->cur_depth++; |
0c41214f | 1059 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
1dec3e07 JA |
1060 | } else if (td->o.verify_async) { |
1061 | /* | |
1062 | * We ran out, wait for async verify threads to finish and | |
1063 | * return one | |
1064 | */ | |
1065 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1066 | goto again; | |
10ba535a JA |
1067 | } |
1068 | ||
e8462bd8 | 1069 | td_io_u_unlock(td); |
10ba535a JA |
1070 | return io_u; |
1071 | } | |
1072 | ||
0d29de83 | 1073 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1074 | { |
0d29de83 JA |
1075 | if (td->o.trim_backlog && td->trim_entries) { |
1076 | int get_trim = 0; | |
10ba535a | 1077 | |
0d29de83 JA |
1078 | if (td->trim_batch) { |
1079 | td->trim_batch--; | |
1080 | get_trim = 1; | |
1081 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1082 | td->last_ddir != DDIR_READ) { | |
1083 | td->trim_batch = td->o.trim_batch; | |
1084 | if (!td->trim_batch) | |
1085 | td->trim_batch = td->o.trim_backlog; | |
1086 | get_trim = 1; | |
1087 | } | |
1088 | ||
1089 | if (get_trim && !get_next_trim(td, io_u)) | |
1090 | return 1; | |
2ba1c290 | 1091 | } |
10ba535a | 1092 | |
0d29de83 JA |
1093 | return 0; |
1094 | } | |
1095 | ||
1096 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1097 | { | |
9e144189 JA |
1098 | if (td->o.verify_backlog && td->io_hist_len) { |
1099 | int get_verify = 0; | |
1100 | ||
1101 | if (td->verify_batch) { | |
1102 | td->verify_batch--; | |
1103 | get_verify = 1; | |
1104 | } else if (!(td->io_hist_len % td->o.verify_backlog) && | |
1105 | td->last_ddir != DDIR_READ) { | |
1106 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1107 | if (!td->verify_batch) |
1108 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1109 | get_verify = 1; |
1110 | } | |
1111 | ||
1112 | if (get_verify && !get_next_verify(td, io_u)) | |
0d29de83 | 1113 | return 1; |
9e144189 JA |
1114 | } |
1115 | ||
0d29de83 JA |
1116 | return 0; |
1117 | } | |
1118 | ||
1119 | /* | |
1120 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1121 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1122 | */ | |
1123 | struct io_u *get_io_u(struct thread_data *td) | |
1124 | { | |
1125 | struct fio_file *f; | |
1126 | struct io_u *io_u; | |
1127 | ||
1128 | io_u = __get_io_u(td); | |
1129 | if (!io_u) { | |
1130 | dprint(FD_IO, "__get_io_u failed\n"); | |
1131 | return NULL; | |
1132 | } | |
1133 | ||
1134 | if (check_get_verify(td, io_u)) | |
1135 | goto out; | |
1136 | if (check_get_trim(td, io_u)) | |
1137 | goto out; | |
1138 | ||
755200a3 JA |
1139 | /* |
1140 | * from a requeue, io_u already setup | |
1141 | */ | |
1142 | if (io_u->file) | |
77f392bf | 1143 | goto out; |
755200a3 | 1144 | |
429f6675 JA |
1145 | /* |
1146 | * If using an iolog, grab next piece if any available. | |
1147 | */ | |
1148 | if (td->o.read_iolog_file) { | |
1149 | if (read_iolog_get(td, io_u)) | |
1150 | goto err_put; | |
2ba1c290 JA |
1151 | } else if (set_io_u_file(td, io_u)) { |
1152 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 1153 | goto err_put; |
2ba1c290 | 1154 | } |
5ec10eaa | 1155 | |
429f6675 | 1156 | f = io_u->file; |
d6aed795 | 1157 | assert(fio_file_open(f)); |
97af62ce | 1158 | |
ff58fced | 1159 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1160 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1161 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1162 | goto err_put; |
2ba1c290 | 1163 | } |
10ba535a | 1164 | |
38dad62d | 1165 | f->last_start = io_u->offset; |
36167d82 | 1166 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1167 | |
c311cd2a | 1168 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE) |
87dc1ab1 | 1169 | populate_verify_io_u(td, io_u); |
e4dad9c6 JA |
1170 | else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) |
1171 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
cbe8d756 RR |
1172 | else if (io_u->ddir == DDIR_READ) { |
1173 | /* | |
1174 | * Reset the buf_filled parameters so next time if the | |
1175 | * buffer is used for writes it is refilled. | |
1176 | */ | |
cbe8d756 RR |
1177 | io_u->buf_filled_len = 0; |
1178 | } | |
87dc1ab1 | 1179 | } |
10ba535a | 1180 | |
165faf16 JA |
1181 | /* |
1182 | * Set io data pointers. | |
1183 | */ | |
cec6b55d JA |
1184 | io_u->xfer_buf = io_u->buf; |
1185 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1186 | |
6ac7a331 | 1187 | out: |
0d29de83 | 1188 | assert(io_u->file); |
429f6675 | 1189 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1190 | if (!td->o.disable_slat) |
1191 | fio_gettime(&io_u->start_time, NULL); | |
429f6675 | 1192 | return io_u; |
36167d82 | 1193 | } |
429f6675 | 1194 | err_put: |
2ba1c290 | 1195 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
1196 | put_io_u(td, io_u); |
1197 | return NULL; | |
10ba535a JA |
1198 | } |
1199 | ||
5451792e JA |
1200 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1201 | { | |
825f818e JA |
1202 | const char *msg[] = { "read", "write", "sync", "datasync", |
1203 | "sync_file_range", "wait", "trim" }; | |
1204 | ||
1205 | ||
5451792e JA |
1206 | |
1207 | log_err("fio: io_u error"); | |
1208 | ||
1209 | if (io_u->file) | |
1210 | log_err(" on file %s", io_u->file->file_name); | |
1211 | ||
1212 | log_err(": %s\n", strerror(io_u->error)); | |
1213 | ||
5ec10eaa JA |
1214 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
1215 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1216 | |
1217 | if (!td->error) | |
1218 | td_verror(td, io_u->error, "io_u error"); | |
1219 | } | |
1220 | ||
97601024 JA |
1221 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
1222 | struct io_completion_data *icd) | |
10ba535a | 1223 | { |
dbad30b0 JA |
1224 | /* |
1225 | * Older gcc's are too dumb to realize that usec is always used | |
1226 | * initialized, silence that warning. | |
1227 | */ | |
1228 | unsigned long uninitialized_var(usec); | |
44f29692 | 1229 | struct fio_file *f; |
10ba535a | 1230 | |
2ba1c290 JA |
1231 | dprint_io_u(io_u, "io complete"); |
1232 | ||
2ecc1b57 | 1233 | td_io_u_lock(td); |
0c6e7517 | 1234 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1235 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
2ecc1b57 | 1236 | td_io_u_unlock(td); |
0c6e7517 | 1237 | |
5f9099ea | 1238 | if (ddir_sync(io_u->ddir)) { |
87dc1ab1 | 1239 | td->last_was_sync = 1; |
44f29692 JA |
1240 | f = io_u->file; |
1241 | if (f) { | |
1242 | f->first_write = -1ULL; | |
1243 | f->last_write = -1ULL; | |
1244 | } | |
87dc1ab1 JA |
1245 | return; |
1246 | } | |
1247 | ||
1248 | td->last_was_sync = 0; | |
9e144189 | 1249 | td->last_ddir = io_u->ddir; |
87dc1ab1 | 1250 | |
ff58fced | 1251 | if (!io_u->error && ddir_rw(io_u->ddir)) { |
10ba535a | 1252 | unsigned int bytes = io_u->buflen - io_u->resid; |
1e97cce9 | 1253 | const enum fio_ddir idx = io_u->ddir; |
ba3e4e0c | 1254 | const enum fio_ddir odx = io_u->ddir ^ 1; |
b29ee5b3 | 1255 | int ret; |
10ba535a | 1256 | |
b29ee5b3 JA |
1257 | td->io_blocks[idx]++; |
1258 | td->io_bytes[idx] += bytes; | |
1259 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 1260 | |
44f29692 JA |
1261 | if (idx == DDIR_WRITE) { |
1262 | f = io_u->file; | |
1263 | if (f) { | |
1264 | if (f->first_write == -1ULL || | |
1265 | io_u->offset < f->first_write) | |
1266 | f->first_write = io_u->offset; | |
1267 | if (f->last_write == -1ULL || | |
1268 | ((io_u->offset + bytes) > f->last_write)) | |
1269 | f->last_write = io_u->offset + bytes; | |
1270 | } | |
1271 | } | |
1272 | ||
b29ee5b3 | 1273 | if (ramp_time_over(td)) { |
40e1a6f0 | 1274 | unsigned long uninitialized_var(lusec); |
40e1a6f0 JA |
1275 | |
1276 | if (!td->o.disable_clat || !td->o.disable_bw) | |
1277 | lusec = utime_since(&io_u->issue_time, | |
1278 | &icd->time); | |
02af0988 JA |
1279 | if (!td->o.disable_lat) { |
1280 | unsigned long tusec; | |
721938ae | 1281 | |
02af0988 JA |
1282 | tusec = utime_since(&io_u->start_time, |
1283 | &icd->time); | |
1284 | add_lat_sample(td, idx, tusec, bytes); | |
1285 | } | |
9520ebb9 | 1286 | if (!td->o.disable_clat) { |
01a1caa5 | 1287 | add_clat_sample(td, idx, lusec, bytes); |
40e1a6f0 | 1288 | io_u_mark_latency(td, lusec); |
9520ebb9 JA |
1289 | } |
1290 | if (!td->o.disable_bw) | |
306ddc97 | 1291 | add_bw_sample(td, idx, bytes, &icd->time); |
b23b6a2f | 1292 | if (__should_check_rate(td, idx)) { |
ba3e4e0c RR |
1293 | td->rate_pending_usleep[idx] = |
1294 | ((td->this_io_bytes[idx] * | |
1295 | td->rate_nsec_cycle[idx]) / 1000 - | |
1296 | utime_since_now(&td->start)); | |
b23b6a2f | 1297 | } |
581e7141 | 1298 | if (__should_check_rate(td, idx ^ 1)) |
ba3e4e0c RR |
1299 | td->rate_pending_usleep[odx] = |
1300 | ((td->this_io_bytes[odx] * | |
1301 | td->rate_nsec_cycle[odx]) / 1000 - | |
1302 | utime_since_now(&td->start)); | |
721938ae | 1303 | } |
10ba535a | 1304 | |
660a1cb5 | 1305 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 1306 | td->o.do_verify && |
41128405 | 1307 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
1308 | log_io_piece(td, io_u); |
1309 | ||
b29ee5b3 | 1310 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 1311 | |
d7762cf8 | 1312 | if (io_u->end_io) { |
36690c9b | 1313 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
1314 | if (ret && !icd->error) |
1315 | icd->error = ret; | |
1316 | } | |
ff58fced | 1317 | } else if (io_u->error) { |
10ba535a | 1318 | icd->error = io_u->error; |
5451792e JA |
1319 | io_u_log_error(td, io_u); |
1320 | } | |
f2bba182 RR |
1321 | if (td->o.continue_on_error && icd->error && |
1322 | td_non_fatal_error(icd->error)) { | |
1323 | /* | |
1324 | * If there is a non_fatal error, then add to the error count | |
1325 | * and clear all the errors. | |
1326 | */ | |
1327 | update_error_count(td, icd->error); | |
1328 | td_clear_error(td); | |
1329 | icd->error = 0; | |
1330 | io_u->error = 0; | |
1331 | } | |
10ba535a JA |
1332 | } |
1333 | ||
9520ebb9 JA |
1334 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1335 | int nr) | |
10ba535a | 1336 | { |
9520ebb9 JA |
1337 | if (!td->o.disable_clat || !td->o.disable_bw) |
1338 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 1339 | |
3af6ef39 JA |
1340 | icd->nr = nr; |
1341 | ||
10ba535a JA |
1342 | icd->error = 0; |
1343 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
1344 | } |
1345 | ||
97601024 JA |
1346 | static void ios_completed(struct thread_data *td, |
1347 | struct io_completion_data *icd) | |
36167d82 JA |
1348 | { |
1349 | struct io_u *io_u; | |
1350 | int i; | |
1351 | ||
10ba535a JA |
1352 | for (i = 0; i < icd->nr; i++) { |
1353 | io_u = td->io_ops->event(td, i); | |
1354 | ||
1355 | io_completed(td, io_u, icd); | |
e8462bd8 JA |
1356 | |
1357 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1358 | put_io_u(td, io_u); | |
10ba535a JA |
1359 | } |
1360 | } | |
97601024 | 1361 | |
e7e6cfb4 JA |
1362 | /* |
1363 | * Complete a single io_u for the sync engines. | |
1364 | */ | |
581e7141 JA |
1365 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
1366 | unsigned long *bytes) | |
97601024 JA |
1367 | { |
1368 | struct io_completion_data icd; | |
1369 | ||
9520ebb9 | 1370 | init_icd(td, &icd, 1); |
97601024 | 1371 | io_completed(td, io_u, &icd); |
e8462bd8 JA |
1372 | |
1373 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1374 | put_io_u(td, io_u); | |
97601024 | 1375 | |
581e7141 JA |
1376 | if (icd.error) { |
1377 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1378 | return -1; | |
1379 | } | |
97601024 | 1380 | |
581e7141 JA |
1381 | if (bytes) { |
1382 | bytes[0] += icd.bytes_done[0]; | |
1383 | bytes[1] += icd.bytes_done[1]; | |
1384 | } | |
1385 | ||
1386 | return 0; | |
97601024 JA |
1387 | } |
1388 | ||
e7e6cfb4 JA |
1389 | /* |
1390 | * Called to complete min_events number of io for the async engines. | |
1391 | */ | |
581e7141 JA |
1392 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
1393 | unsigned long *bytes) | |
97601024 | 1394 | { |
97601024 | 1395 | struct io_completion_data icd; |
00de55ef | 1396 | struct timespec *tvp = NULL; |
97601024 | 1397 | int ret; |
4d06a338 | 1398 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1399 | |
4950421a | 1400 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1401 | |
4950421a | 1402 | if (!min_evts) |
00de55ef | 1403 | tvp = &ts; |
97601024 | 1404 | |
4950421a | 1405 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1406 | if (ret < 0) { |
e1161c32 | 1407 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1408 | return ret; |
1409 | } else if (!ret) | |
1410 | return ret; | |
1411 | ||
9520ebb9 | 1412 | init_icd(td, &icd, ret); |
97601024 | 1413 | ios_completed(td, &icd); |
581e7141 JA |
1414 | if (icd.error) { |
1415 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1416 | return -1; | |
1417 | } | |
97601024 | 1418 | |
581e7141 JA |
1419 | if (bytes) { |
1420 | bytes[0] += icd.bytes_done[0]; | |
1421 | bytes[1] += icd.bytes_done[1]; | |
1422 | } | |
1423 | ||
1424 | return 0; | |
97601024 | 1425 | } |
7e77dd02 JA |
1426 | |
1427 | /* | |
1428 | * Call when io_u is really queued, to update the submission latency. | |
1429 | */ | |
1430 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1431 | { | |
9520ebb9 JA |
1432 | if (!td->o.disable_slat) { |
1433 | unsigned long slat_time; | |
7e77dd02 | 1434 | |
9520ebb9 | 1435 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
29a90ddb | 1436 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen); |
9520ebb9 | 1437 | } |
7e77dd02 | 1438 | } |
433afcb4 | 1439 | |
5973cafb JA |
1440 | /* |
1441 | * "randomly" fill the buffer contents | |
1442 | */ | |
1443 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1444 | unsigned int max_bs) | |
1445 | { | |
e84b371e JA |
1446 | io_u->buf_filled_len = 0; |
1447 | ||
637ef8d9 JA |
1448 | if (!td->o.zero_buffers) |
1449 | fill_random_buf(io_u->buf, max_bs); | |
1450 | else | |
1451 | memset(io_u->buf, 0, max_bs); | |
5973cafb | 1452 | } |