Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
5973cafb | 9 | #include "hash.h" |
4f5af7b2 | 10 | #include "verify.h" |
0d29de83 | 11 | #include "trim.h" |
1fbbf72e | 12 | #include "lib/rand.h" |
10ba535a | 13 | |
97601024 JA |
14 | struct io_completion_data { |
15 | int nr; /* input */ | |
c8eeb9df | 16 | int account; /* input */ |
97601024 JA |
17 | |
18 | int error; /* output */ | |
19 | unsigned long bytes_done[2]; /* output */ | |
20 | struct timeval time; /* output */ | |
21 | }; | |
22 | ||
10ba535a JA |
23 | /* |
24 | * The ->file_map[] contains a map of blocks we have or have not done io | |
25 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
26 | */ | |
aec2de20 | 27 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
10ba535a | 28 | { |
aec2de20 JA |
29 | unsigned int idx = RAND_MAP_IDX(f, block); |
30 | unsigned int bit = RAND_MAP_BIT(f, block); | |
10ba535a | 31 | |
84422acd JA |
32 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
33 | ||
0ce8b119 | 34 | return (f->file_map[idx] & (1UL << bit)) == 0; |
10ba535a JA |
35 | } |
36 | ||
df415585 JA |
37 | /* |
38 | * Mark a given offset as used in the map. | |
39 | */ | |
9bf2061e | 40 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 41 | { |
2dc1bbeb | 42 | unsigned int min_bs = td->o.rw_min_bs; |
9bf2061e | 43 | struct fio_file *f = io_u->file; |
a00735e6 | 44 | unsigned long long block; |
3e3357b1 | 45 | unsigned int blocks, nr_blocks; |
38dad62d | 46 | int busy_check; |
df415585 | 47 | |
b9c5b644 | 48 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
c685b5b2 | 49 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
3e3357b1 | 50 | blocks = 0; |
38dad62d | 51 | busy_check = !(io_u->flags & IO_U_F_BUSY_OK); |
c685b5b2 | 52 | |
3e3357b1 | 53 | while (nr_blocks) { |
df415585 | 54 | unsigned int idx, bit; |
0ce8b119 | 55 | unsigned long mask, this_blocks; |
df415585 | 56 | |
1e3d53ac JA |
57 | /* |
58 | * If we have a mixed random workload, we may | |
59 | * encounter blocks we already did IO to. | |
60 | */ | |
38dad62d JA |
61 | if (!busy_check) { |
62 | blocks = nr_blocks; | |
63 | break; | |
64 | } | |
5736c10d | 65 | if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block)) |
df415585 JA |
66 | break; |
67 | ||
aec2de20 JA |
68 | idx = RAND_MAP_IDX(f, block); |
69 | bit = RAND_MAP_BIT(f, block); | |
df415585 | 70 | |
0032bf9f | 71 | fio_assert(td, idx < f->num_maps); |
df415585 | 72 | |
3e3357b1 JA |
73 | this_blocks = nr_blocks; |
74 | if (this_blocks + bit > BLOCKS_PER_MAP) | |
75 | this_blocks = BLOCKS_PER_MAP - bit; | |
76 | ||
da3758bf JA |
77 | do { |
78 | if (this_blocks == BLOCKS_PER_MAP) | |
0ce8b119 | 79 | mask = -1UL; |
da3758bf | 80 | else |
0ce8b119 | 81 | mask = ((1UL << this_blocks) - 1) << bit; |
da3758bf JA |
82 | |
83 | if (!(f->file_map[idx] & mask)) | |
84 | break; | |
85 | ||
86 | this_blocks--; | |
87 | } while (this_blocks); | |
88 | ||
89 | if (!this_blocks) | |
90 | break; | |
3e3357b1 | 91 | |
3e3357b1 JA |
92 | f->file_map[idx] |= mask; |
93 | nr_blocks -= this_blocks; | |
94 | blocks += this_blocks; | |
c9dd34b2 | 95 | block += this_blocks; |
df415585 JA |
96 | } |
97 | ||
a00735e6 JA |
98 | if ((blocks * min_bs) < io_u->buflen) |
99 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
100 | } |
101 | ||
3e3357b1 JA |
102 | static unsigned long long last_block(struct thread_data *td, struct fio_file *f, |
103 | enum fio_ddir ddir) | |
2ba1c290 JA |
104 | { |
105 | unsigned long long max_blocks; | |
d9dd70f7 | 106 | unsigned long long max_size; |
2ba1c290 | 107 | |
ff58fced JA |
108 | assert(ddir_rw(ddir)); |
109 | ||
d9dd70f7 JA |
110 | /* |
111 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
112 | */ | |
113 | max_size = f->io_size; | |
114 | if (max_size > f->real_file_size) | |
115 | max_size = f->real_file_size; | |
116 | ||
ed335855 SN |
117 | if (td->o.zone_range) |
118 | max_size = td->o.zone_range; | |
119 | ||
2b7a01d0 | 120 | max_blocks = max_size / (unsigned long long) td->o.ba[ddir]; |
2ba1c290 JA |
121 | if (!max_blocks) |
122 | return 0; | |
123 | ||
67778e88 | 124 | return max_blocks; |
2ba1c290 JA |
125 | } |
126 | ||
10ba535a JA |
127 | /* |
128 | * Return the next free block in the map. | |
129 | */ | |
130 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
4ba66134 | 131 | enum fio_ddir ddir, unsigned long long *b) |
10ba535a | 132 | { |
0ce8b119 | 133 | unsigned long long block, min_bs = td->o.rw_min_bs, lastb; |
10ba535a JA |
134 | int i; |
135 | ||
5e0baa7f JA |
136 | lastb = last_block(td, f, ddir); |
137 | if (!lastb) | |
138 | return 1; | |
139 | ||
c685b5b2 | 140 | i = f->last_free_lookup; |
0ce8b119 JA |
141 | block = i * BLOCKS_PER_MAP; |
142 | while (block * min_bs < f->real_file_size && | |
143 | block * min_bs < f->io_size) { | |
144 | if (f->file_map[i] != -1UL) { | |
145 | block += ffz(f->file_map[i]); | |
146 | if (block > lastb) | |
2ba1c290 | 147 | break; |
c685b5b2 | 148 | f->last_free_lookup = i; |
0ce8b119 | 149 | *b = block; |
10ba535a JA |
150 | return 0; |
151 | } | |
152 | ||
0ce8b119 | 153 | block += BLOCKS_PER_MAP; |
10ba535a JA |
154 | i++; |
155 | } | |
156 | ||
2ba1c290 | 157 | dprint(FD_IO, "failed finding a free block\n"); |
10ba535a JA |
158 | return 1; |
159 | } | |
160 | ||
ec4015da | 161 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
4ba66134 | 162 | enum fio_ddir ddir, unsigned long long *b) |
ec4015da | 163 | { |
15b87723 | 164 | unsigned long long rmax, r, lastb; |
ec4015da JA |
165 | int loops = 5; |
166 | ||
5e0baa7f JA |
167 | lastb = last_block(td, f, ddir); |
168 | if (!lastb) | |
169 | return 1; | |
170 | ||
0ce8b119 JA |
171 | if (f->failed_rands >= 200) |
172 | goto ffz; | |
173 | ||
15b87723 | 174 | rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX; |
ec4015da | 175 | do { |
15b87723 | 176 | if (td->o.use_os_rand) |
2615cc4b | 177 | r = os_random_long(&td->random_state); |
15b87723 | 178 | else |
2615cc4b | 179 | r = __rand(&td->__random_state); |
15b87723 JA |
180 | |
181 | *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0)); | |
2615cc4b | 182 | |
84422acd | 183 | dprint(FD_RANDOM, "off rand %llu\n", r); |
2615cc4b | 184 | |
2ba1c290 | 185 | |
43c63a78 JA |
186 | /* |
187 | * if we are not maintaining a random map, we are done. | |
188 | */ | |
303032ae | 189 | if (!file_randommap(td, f)) |
0ce8b119 | 190 | goto ret_good; |
43c63a78 JA |
191 | |
192 | /* | |
84422acd | 193 | * calculate map offset and check if it's free |
43c63a78 | 194 | */ |
aec2de20 | 195 | if (random_map_free(f, *b)) |
0ce8b119 | 196 | goto ret_good; |
43c63a78 | 197 | |
84422acd JA |
198 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
199 | *b); | |
43c63a78 | 200 | } while (--loops); |
ec4015da | 201 | |
0ce8b119 JA |
202 | if (!f->failed_rands++) |
203 | f->last_free_lookup = 0; | |
204 | ||
ec4015da | 205 | /* |
43c63a78 JA |
206 | * we get here, if we didn't suceed in looking up a block. generate |
207 | * a random start offset into the filemap, and find the first free | |
208 | * block from there. | |
ec4015da | 209 | */ |
43c63a78 JA |
210 | loops = 10; |
211 | do { | |
dc873b6f | 212 | f->last_free_lookup = (f->num_maps - 1) * |
15b87723 | 213 | (r / ((unsigned long long) rmax + 1.0)); |
4ba66134 | 214 | if (!get_next_free_block(td, f, ddir, b)) |
0ce8b119 | 215 | goto ret; |
ec4015da | 216 | |
d1cea521 JA |
217 | if (td->o.use_os_rand) |
218 | r = os_random_long(&td->random_state); | |
219 | else | |
220 | r = __rand(&td->__random_state); | |
43c63a78 JA |
221 | } while (--loops); |
222 | ||
223 | /* | |
224 | * that didn't work either, try exhaustive search from the start | |
225 | */ | |
226 | f->last_free_lookup = 0; | |
0ce8b119 JA |
227 | ffz: |
228 | if (!get_next_free_block(td, f, ddir, b)) | |
229 | return 0; | |
230 | f->last_free_lookup = 0; | |
4ba66134 | 231 | return get_next_free_block(td, f, ddir, b); |
0ce8b119 JA |
232 | ret_good: |
233 | f->failed_rands = 0; | |
234 | ret: | |
235 | return 0; | |
ec4015da JA |
236 | } |
237 | ||
38dad62d JA |
238 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
239 | enum fio_ddir ddir, unsigned long long *b) | |
240 | { | |
241 | if (get_next_rand_offset(td, f, ddir, b)) { | |
242 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", | |
243 | f->file_name, f->last_pos, f->real_file_size); | |
244 | return 1; | |
245 | } | |
246 | ||
247 | return 0; | |
248 | } | |
249 | ||
250 | static int get_next_seq_block(struct thread_data *td, struct fio_file *f, | |
251 | enum fio_ddir ddir, unsigned long long *b) | |
252 | { | |
ff58fced JA |
253 | assert(ddir_rw(ddir)); |
254 | ||
38dad62d | 255 | if (f->last_pos < f->real_file_size) { |
a66da7a2 | 256 | unsigned long long pos; |
059b0802 | 257 | |
a66da7a2 JA |
258 | if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0) |
259 | f->last_pos = f->real_file_size; | |
260 | ||
261 | pos = f->last_pos - f->file_offset; | |
059b0802 JA |
262 | if (pos) |
263 | pos += td->o.ddir_seq_add; | |
264 | ||
265 | *b = pos / td->o.min_bs[ddir]; | |
38dad62d JA |
266 | return 0; |
267 | } | |
268 | ||
269 | return 1; | |
270 | } | |
271 | ||
272 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
273 | enum fio_ddir ddir, int rw_seq, unsigned long long *b) | |
274 | { | |
275 | struct fio_file *f = io_u->file; | |
276 | int ret; | |
277 | ||
ff58fced JA |
278 | assert(ddir_rw(ddir)); |
279 | ||
38dad62d JA |
280 | if (rw_seq) { |
281 | if (td_random(td)) | |
282 | ret = get_next_rand_block(td, f, ddir, b); | |
283 | else | |
284 | ret = get_next_seq_block(td, f, ddir, b); | |
285 | } else { | |
286 | io_u->flags |= IO_U_F_BUSY_OK; | |
287 | ||
288 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
289 | ret = get_next_seq_block(td, f, ddir, b); | |
290 | if (ret) | |
291 | ret = get_next_rand_block(td, f, ddir, b); | |
292 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { | |
293 | if (f->last_start != -1ULL) | |
0508320e DN |
294 | *b = (f->last_start - f->file_offset) |
295 | / td->o.min_bs[ddir]; | |
38dad62d JA |
296 | else |
297 | *b = 0; | |
298 | ret = 0; | |
299 | } else { | |
300 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
301 | ret = 1; | |
302 | } | |
303 | } | |
304 | ||
305 | return ret; | |
306 | } | |
307 | ||
10ba535a JA |
308 | /* |
309 | * For random io, generate a random new block and see if it's used. Repeat | |
310 | * until we find a free one. For sequential io, just return the end of | |
311 | * the last io issued. | |
312 | */ | |
15dc1934 | 313 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u) |
10ba535a | 314 | { |
9bf2061e | 315 | struct fio_file *f = io_u->file; |
ec4015da | 316 | unsigned long long b; |
4ba66134 | 317 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 318 | int rw_seq_hit = 0; |
10ba535a | 319 | |
ff58fced JA |
320 | assert(ddir_rw(ddir)); |
321 | ||
38dad62d JA |
322 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
323 | rw_seq_hit = 1; | |
5736c10d | 324 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 325 | } |
211097b2 | 326 | |
0508320e | 327 | if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) |
38dad62d | 328 | return 1; |
10ba535a | 329 | |
2b7a01d0 | 330 | io_u->offset = b * td->o.ba[ddir]; |
009bd847 JA |
331 | if (io_u->offset >= f->io_size) { |
332 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
333 | io_u->offset, f->io_size); | |
334 | return 1; | |
335 | } | |
336 | ||
337 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
338 | if (io_u->offset >= f->real_file_size) { |
339 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
340 | io_u->offset, f->real_file_size); | |
10ba535a | 341 | return 1; |
2ba1c290 | 342 | } |
10ba535a JA |
343 | |
344 | return 0; | |
345 | } | |
346 | ||
15dc1934 JA |
347 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
348 | { | |
7eb36574 JA |
349 | struct prof_io_ops *ops = &td->prof_io_ops; |
350 | ||
351 | if (ops->fill_io_u_off) | |
352 | return ops->fill_io_u_off(td, io_u); | |
15dc1934 JA |
353 | |
354 | return __get_next_offset(td, io_u); | |
355 | } | |
356 | ||
79944128 JA |
357 | static inline int io_u_fits(struct thread_data *td, struct io_u *io_u, |
358 | unsigned int buflen) | |
359 | { | |
360 | struct fio_file *f = io_u->file; | |
361 | ||
362 | return io_u->offset + buflen <= f->io_size + td->o.start_offset; | |
363 | } | |
364 | ||
15dc1934 | 365 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u) |
10ba535a | 366 | { |
bca4ed4d | 367 | const int ddir = io_u->ddir; |
f3f552b9 | 368 | unsigned int uninitialized_var(buflen); |
f3059de1 | 369 | unsigned int minbs, maxbs; |
1294c3ec | 370 | unsigned long r, rand_max; |
10ba535a | 371 | |
ff58fced JA |
372 | assert(ddir_rw(ddir)); |
373 | ||
f3059de1 JA |
374 | minbs = td->o.min_bs[ddir]; |
375 | maxbs = td->o.max_bs[ddir]; | |
376 | ||
79944128 JA |
377 | if (minbs == maxbs) |
378 | return minbs; | |
379 | ||
4c07ad86 JA |
380 | if (td->o.use_os_rand) |
381 | rand_max = OS_RAND_MAX; | |
382 | else | |
383 | rand_max = FRAND_MAX; | |
384 | ||
79944128 | 385 | do { |
4c07ad86 JA |
386 | if (td->o.use_os_rand) |
387 | r = os_random_long(&td->bsrange_state); | |
388 | else | |
389 | r = __rand(&td->__bsrange_state); | |
390 | ||
720e84ad | 391 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 392 | buflen = 1 + (unsigned int) ((double) maxbs * |
4c07ad86 | 393 | (r / (rand_max + 1.0))); |
f3059de1 JA |
394 | if (buflen < minbs) |
395 | buflen = minbs; | |
5ec10eaa | 396 | } else { |
564ca972 JA |
397 | long perc = 0; |
398 | unsigned int i; | |
399 | ||
720e84ad JA |
400 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
401 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
402 | |
403 | buflen = bsp->bs; | |
404 | perc += bsp->perc; | |
79944128 JA |
405 | if ((r <= ((rand_max / 100L) * perc)) && |
406 | io_u_fits(td, io_u, buflen)) | |
564ca972 JA |
407 | break; |
408 | } | |
409 | } | |
79944128 | 410 | |
f3059de1 JA |
411 | if (!td->o.bs_unaligned && is_power_of_2(minbs)) |
412 | buflen = (buflen + minbs - 1) & ~(minbs - 1); | |
10ba535a | 413 | |
79944128 | 414 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 415 | |
10ba535a JA |
416 | return buflen; |
417 | } | |
418 | ||
15dc1934 JA |
419 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
420 | { | |
7eb36574 JA |
421 | struct prof_io_ops *ops = &td->prof_io_ops; |
422 | ||
423 | if (ops->fill_io_u_size) | |
424 | return ops->fill_io_u_size(td, io_u); | |
15dc1934 JA |
425 | |
426 | return __get_next_buflen(td, io_u); | |
427 | } | |
428 | ||
afe24a5a JA |
429 | static void set_rwmix_bytes(struct thread_data *td) |
430 | { | |
afe24a5a JA |
431 | unsigned int diff; |
432 | ||
433 | /* | |
434 | * we do time or byte based switch. this is needed because | |
435 | * buffered writes may issue a lot quicker than they complete, | |
436 | * whereas reads do not. | |
437 | */ | |
e47f799f | 438 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 439 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
440 | } |
441 | ||
442 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
443 | { | |
444 | unsigned int v; | |
1294c3ec | 445 | unsigned long r; |
e47f799f | 446 | |
4c07ad86 JA |
447 | if (td->o.use_os_rand) { |
448 | r = os_random_long(&td->rwmix_state); | |
449 | v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0))); | |
450 | } else { | |
451 | r = __rand(&td->__rwmix_state); | |
452 | v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0))); | |
453 | } | |
454 | ||
04c540d9 | 455 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
456 | return DDIR_READ; |
457 | ||
458 | return DDIR_WRITE; | |
afe24a5a JA |
459 | } |
460 | ||
581e7141 JA |
461 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
462 | { | |
463 | enum fio_ddir odir = ddir ^ 1; | |
464 | struct timeval t; | |
465 | long usec; | |
466 | ||
ff58fced JA |
467 | assert(ddir_rw(ddir)); |
468 | ||
581e7141 JA |
469 | if (td->rate_pending_usleep[ddir] <= 0) |
470 | return ddir; | |
471 | ||
472 | /* | |
473 | * We have too much pending sleep in this direction. See if we | |
474 | * should switch. | |
475 | */ | |
476 | if (td_rw(td)) { | |
477 | /* | |
478 | * Other direction does not have too much pending, switch | |
479 | */ | |
480 | if (td->rate_pending_usleep[odir] < 100000) | |
481 | return odir; | |
482 | ||
483 | /* | |
484 | * Both directions have pending sleep. Sleep the minimum time | |
485 | * and deduct from both. | |
486 | */ | |
487 | if (td->rate_pending_usleep[ddir] <= | |
488 | td->rate_pending_usleep[odir]) { | |
489 | usec = td->rate_pending_usleep[ddir]; | |
490 | } else { | |
491 | usec = td->rate_pending_usleep[odir]; | |
492 | ddir = odir; | |
493 | } | |
494 | } else | |
495 | usec = td->rate_pending_usleep[ddir]; | |
496 | ||
78c1eda5 JA |
497 | /* |
498 | * We are going to sleep, ensure that we flush anything pending as | |
422f9e4b RM |
499 | * not to skew our latency numbers. |
500 | * | |
501 | * Changed to only monitor 'in flight' requests here instead of the | |
502 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
503 | * io's that have been actually submitted to an async engine, | |
504 | * and cur_depth is meaningless for sync engines. | |
78c1eda5 | 505 | */ |
422f9e4b | 506 | if (td->io_u_in_flight) { |
78c1eda5 JA |
507 | int fio_unused ret; |
508 | ||
422f9e4b | 509 | ret = io_u_queued_complete(td, td->io_u_in_flight, NULL); |
78c1eda5 JA |
510 | } |
511 | ||
581e7141 JA |
512 | fio_gettime(&t, NULL); |
513 | usec_sleep(td, usec); | |
514 | usec = utime_since_now(&t); | |
515 | ||
516 | td->rate_pending_usleep[ddir] -= usec; | |
517 | ||
518 | odir = ddir ^ 1; | |
519 | if (td_rw(td) && __should_check_rate(td, odir)) | |
520 | td->rate_pending_usleep[odir] -= usec; | |
0b9d69ec | 521 | |
581e7141 JA |
522 | return ddir; |
523 | } | |
524 | ||
10ba535a JA |
525 | /* |
526 | * Return the data direction for the next io_u. If the job is a | |
527 | * mixed read/write workload, check the rwmix cycle and switch if | |
528 | * necessary. | |
529 | */ | |
1e97cce9 | 530 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 531 | { |
581e7141 JA |
532 | enum fio_ddir ddir; |
533 | ||
5f9099ea JA |
534 | /* |
535 | * see if it's time to fsync | |
536 | */ | |
537 | if (td->o.fsync_blocks && | |
538 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && | |
539 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
540 | return DDIR_SYNC; | |
541 | ||
542 | /* | |
543 | * see if it's time to fdatasync | |
544 | */ | |
545 | if (td->o.fdatasync_blocks && | |
546 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) && | |
547 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
548 | return DDIR_DATASYNC; | |
549 | ||
44f29692 JA |
550 | /* |
551 | * see if it's time to sync_file_range | |
552 | */ | |
553 | if (td->sync_file_range_nr && | |
554 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) && | |
555 | td->io_issues[DDIR_WRITE] && should_fsync(td)) | |
556 | return DDIR_SYNC_FILE_RANGE; | |
557 | ||
10ba535a | 558 | if (td_rw(td)) { |
10ba535a JA |
559 | /* |
560 | * Check if it's time to seed a new data direction. | |
561 | */ | |
e4928662 | 562 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
563 | /* |
564 | * Put a top limit on how many bytes we do for | |
565 | * one data direction, to avoid overflowing the | |
566 | * ranges too much | |
567 | */ | |
568 | ddir = get_rand_ddir(td); | |
e47f799f JA |
569 | |
570 | if (ddir != td->rwmix_ddir) | |
571 | set_rwmix_bytes(td); | |
572 | ||
573 | td->rwmix_ddir = ddir; | |
10ba535a | 574 | } |
581e7141 | 575 | ddir = td->rwmix_ddir; |
10ba535a | 576 | } else if (td_read(td)) |
581e7141 | 577 | ddir = DDIR_READ; |
10ba535a | 578 | else |
581e7141 JA |
579 | ddir = DDIR_WRITE; |
580 | ||
581 | td->rwmix_ddir = rate_ddir(td, ddir); | |
582 | return td->rwmix_ddir; | |
10ba535a JA |
583 | } |
584 | ||
1ef2b6be JA |
585 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
586 | { | |
587 | io_u->ddir = get_rw_ddir(td); | |
588 | ||
589 | if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && | |
590 | td->o.barrier_blocks && | |
591 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
592 | td->io_issues[DDIR_WRITE]) | |
593 | io_u->flags |= IO_U_F_BARRIER; | |
594 | } | |
595 | ||
e8462bd8 | 596 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 JA |
597 | { |
598 | int ret = put_file(td, f); | |
599 | ||
600 | if (ret) | |
601 | td_verror(td, ret, "file close"); | |
602 | } | |
603 | ||
10ba535a JA |
604 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
605 | { | |
e8462bd8 JA |
606 | td_io_u_lock(td); |
607 | ||
d7ee2a7d | 608 | if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF)) |
60f2c658 | 609 | put_file_log(td, io_u->file); |
10ba535a | 610 | io_u->file = NULL; |
d7ee2a7d SL |
611 | io_u->flags &= ~IO_U_F_FREE_DEF; |
612 | io_u->flags |= IO_U_F_FREE; | |
613 | ||
0c41214f RR |
614 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) |
615 | td->cur_depth--; | |
e8462bd8 | 616 | flist_del_init(&io_u->list); |
01743ee1 | 617 | flist_add(&io_u->list, &td->io_u_freelist); |
e8462bd8 JA |
618 | td_io_u_unlock(td); |
619 | td_io_u_free_notify(td); | |
10ba535a JA |
620 | } |
621 | ||
f2bba182 RR |
622 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
623 | { | |
624 | io_u->flags &= ~IO_U_F_FLIGHT; | |
625 | put_io_u(td, io_u); | |
626 | } | |
627 | ||
755200a3 JA |
628 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
629 | { | |
630 | struct io_u *__io_u = *io_u; | |
631 | ||
465221b0 JA |
632 | dprint(FD_IO, "requeue %p\n", __io_u); |
633 | ||
e8462bd8 JA |
634 | td_io_u_lock(td); |
635 | ||
4d2e0f49 | 636 | __io_u->flags |= IO_U_F_FREE; |
ff58fced | 637 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir)) |
e4f54adb | 638 | td->io_issues[__io_u->ddir]--; |
5ec10eaa | 639 | |
4d2e0f49 | 640 | __io_u->flags &= ~IO_U_F_FLIGHT; |
0c41214f RR |
641 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) |
642 | td->cur_depth--; | |
01743ee1 JA |
643 | flist_del(&__io_u->list); |
644 | flist_add_tail(&__io_u->list, &td->io_u_requeues); | |
e8462bd8 | 645 | td_io_u_unlock(td); |
755200a3 JA |
646 | *io_u = NULL; |
647 | } | |
648 | ||
9bf2061e | 649 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 650 | { |
b4c5e1ac JA |
651 | if (td->io_ops->flags & FIO_NOIO) |
652 | goto out; | |
653 | ||
1ef2b6be | 654 | set_rw_ddir(td, io_u); |
5f9099ea | 655 | |
87dc1ab1 | 656 | /* |
ff58fced | 657 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 658 | */ |
ff58fced | 659 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 660 | goto out; |
a00735e6 | 661 | |
48f5abd3 JA |
662 | /* |
663 | * See if it's time to switch to a new zone | |
664 | */ | |
665 | if (td->zone_bytes >= td->o.zone_size) { | |
666 | td->zone_bytes = 0; | |
ed335855 SN |
667 | io_u->file->file_offset += td->o.zone_range + td->o.zone_skip; |
668 | io_u->file->last_pos = io_u->file->file_offset; | |
48f5abd3 JA |
669 | td->io_skip_bytes += td->o.zone_skip; |
670 | } | |
671 | ||
10ba535a | 672 | /* |
c685b5b2 JA |
673 | * No log, let the seq/rand engine retrieve the next buflen and |
674 | * position. | |
10ba535a | 675 | */ |
2ba1c290 JA |
676 | if (get_next_offset(td, io_u)) { |
677 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); | |
bca4ed4d | 678 | return 1; |
2ba1c290 | 679 | } |
10ba535a | 680 | |
9bf2061e | 681 | io_u->buflen = get_next_buflen(td, io_u); |
2ba1c290 JA |
682 | if (!io_u->buflen) { |
683 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 684 | return 1; |
2ba1c290 | 685 | } |
bca4ed4d | 686 | |
2ba1c290 JA |
687 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
688 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); | |
4ba66134 JA |
689 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
690 | io_u->buflen, io_u->file->real_file_size); | |
6a5e6884 | 691 | return 1; |
2ba1c290 | 692 | } |
6a5e6884 | 693 | |
bca4ed4d JA |
694 | /* |
695 | * mark entry before potentially trimming io_u | |
696 | */ | |
303032ae | 697 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 698 | mark_random_map(td, io_u); |
bca4ed4d JA |
699 | |
700 | /* | |
701 | * If using a write iolog, store this entry. | |
702 | */ | |
c38e9468 | 703 | out: |
2ba1c290 | 704 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 705 | td->zone_bytes += io_u->buflen; |
f29b25a3 | 706 | log_io_u(td, io_u); |
bca4ed4d | 707 | return 0; |
10ba535a JA |
708 | } |
709 | ||
838bc709 JA |
710 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
711 | { | |
2b13e716 | 712 | int idx = 0; |
838bc709 JA |
713 | |
714 | switch (nr) { | |
715 | default: | |
2b13e716 | 716 | idx = 6; |
838bc709 JA |
717 | break; |
718 | case 33 ... 64: | |
2b13e716 | 719 | idx = 5; |
838bc709 JA |
720 | break; |
721 | case 17 ... 32: | |
2b13e716 | 722 | idx = 4; |
838bc709 JA |
723 | break; |
724 | case 9 ... 16: | |
2b13e716 | 725 | idx = 3; |
838bc709 JA |
726 | break; |
727 | case 5 ... 8: | |
2b13e716 | 728 | idx = 2; |
838bc709 JA |
729 | break; |
730 | case 1 ... 4: | |
2b13e716 | 731 | idx = 1; |
838bc709 JA |
732 | case 0: |
733 | break; | |
734 | } | |
735 | ||
2b13e716 | 736 | map[idx]++; |
838bc709 JA |
737 | } |
738 | ||
739 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
740 | { | |
741 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
742 | td->ts.total_submit++; | |
743 | } | |
744 | ||
745 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
746 | { | |
747 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
748 | td->ts.total_complete++; | |
749 | } | |
750 | ||
d8005759 | 751 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 752 | { |
2b13e716 | 753 | int idx = 0; |
71619dc2 JA |
754 | |
755 | switch (td->cur_depth) { | |
756 | default: | |
2b13e716 | 757 | idx = 6; |
a783e61a | 758 | break; |
71619dc2 | 759 | case 32 ... 63: |
2b13e716 | 760 | idx = 5; |
a783e61a | 761 | break; |
71619dc2 | 762 | case 16 ... 31: |
2b13e716 | 763 | idx = 4; |
a783e61a | 764 | break; |
71619dc2 | 765 | case 8 ... 15: |
2b13e716 | 766 | idx = 3; |
a783e61a | 767 | break; |
71619dc2 | 768 | case 4 ... 7: |
2b13e716 | 769 | idx = 2; |
a783e61a | 770 | break; |
71619dc2 | 771 | case 2 ... 3: |
2b13e716 | 772 | idx = 1; |
71619dc2 JA |
773 | case 1: |
774 | break; | |
775 | } | |
776 | ||
2b13e716 | 777 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
778 | } |
779 | ||
04a0feae JA |
780 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
781 | { | |
2b13e716 | 782 | int idx = 0; |
04a0feae JA |
783 | |
784 | assert(usec < 1000); | |
785 | ||
786 | switch (usec) { | |
787 | case 750 ... 999: | |
2b13e716 | 788 | idx = 9; |
04a0feae JA |
789 | break; |
790 | case 500 ... 749: | |
2b13e716 | 791 | idx = 8; |
04a0feae JA |
792 | break; |
793 | case 250 ... 499: | |
2b13e716 | 794 | idx = 7; |
04a0feae JA |
795 | break; |
796 | case 100 ... 249: | |
2b13e716 | 797 | idx = 6; |
04a0feae JA |
798 | break; |
799 | case 50 ... 99: | |
2b13e716 | 800 | idx = 5; |
04a0feae JA |
801 | break; |
802 | case 20 ... 49: | |
2b13e716 | 803 | idx = 4; |
04a0feae JA |
804 | break; |
805 | case 10 ... 19: | |
2b13e716 | 806 | idx = 3; |
04a0feae JA |
807 | break; |
808 | case 4 ... 9: | |
2b13e716 | 809 | idx = 2; |
04a0feae JA |
810 | break; |
811 | case 2 ... 3: | |
2b13e716 | 812 | idx = 1; |
04a0feae JA |
813 | case 0 ... 1: |
814 | break; | |
815 | } | |
816 | ||
2b13e716 JA |
817 | assert(idx < FIO_IO_U_LAT_U_NR); |
818 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
819 | } |
820 | ||
821 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) | |
ec118304 | 822 | { |
2b13e716 | 823 | int idx = 0; |
ec118304 JA |
824 | |
825 | switch (msec) { | |
826 | default: | |
2b13e716 | 827 | idx = 11; |
04a0feae | 828 | break; |
8abdce66 | 829 | case 1000 ... 1999: |
2b13e716 | 830 | idx = 10; |
04a0feae | 831 | break; |
8abdce66 | 832 | case 750 ... 999: |
2b13e716 | 833 | idx = 9; |
04a0feae | 834 | break; |
8abdce66 | 835 | case 500 ... 749: |
2b13e716 | 836 | idx = 8; |
04a0feae | 837 | break; |
8abdce66 | 838 | case 250 ... 499: |
2b13e716 | 839 | idx = 7; |
04a0feae | 840 | break; |
8abdce66 | 841 | case 100 ... 249: |
2b13e716 | 842 | idx = 6; |
04a0feae | 843 | break; |
8abdce66 | 844 | case 50 ... 99: |
2b13e716 | 845 | idx = 5; |
04a0feae | 846 | break; |
8abdce66 | 847 | case 20 ... 49: |
2b13e716 | 848 | idx = 4; |
04a0feae | 849 | break; |
8abdce66 | 850 | case 10 ... 19: |
2b13e716 | 851 | idx = 3; |
04a0feae | 852 | break; |
8abdce66 | 853 | case 4 ... 9: |
2b13e716 | 854 | idx = 2; |
04a0feae | 855 | break; |
ec118304 | 856 | case 2 ... 3: |
2b13e716 | 857 | idx = 1; |
ec118304 JA |
858 | case 0 ... 1: |
859 | break; | |
860 | } | |
861 | ||
2b13e716 JA |
862 | assert(idx < FIO_IO_U_LAT_M_NR); |
863 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
864 | } |
865 | ||
866 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) | |
867 | { | |
868 | if (usec < 1000) | |
869 | io_u_mark_lat_usec(td, usec); | |
870 | else | |
871 | io_u_mark_lat_msec(td, usec / 1000); | |
ec118304 JA |
872 | } |
873 | ||
0aabe160 JA |
874 | /* |
875 | * Get next file to service by choosing one at random | |
876 | */ | |
2cc52930 JA |
877 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
878 | enum fio_file_flags goodf, | |
d6aed795 | 879 | enum fio_file_flags badf) |
0aabe160 | 880 | { |
0aabe160 | 881 | struct fio_file *f; |
1c178180 | 882 | int fno; |
0aabe160 JA |
883 | |
884 | do { | |
87b10676 | 885 | int opened = 0; |
1294c3ec | 886 | unsigned long r; |
4c07ad86 JA |
887 | |
888 | if (td->o.use_os_rand) { | |
889 | r = os_random_long(&td->next_file_state); | |
890 | fno = (unsigned int) ((double) td->o.nr_files | |
891 | * (r / (OS_RAND_MAX + 1.0))); | |
892 | } else { | |
893 | r = __rand(&td->__next_file_state); | |
894 | fno = (unsigned int) ((double) td->o.nr_files | |
895 | * (r / (FRAND_MAX + 1.0))); | |
896 | } | |
7c83c089 | 897 | |
126d65c6 | 898 | f = td->files[fno]; |
d6aed795 | 899 | if (fio_file_done(f)) |
059e63c0 | 900 | continue; |
1c178180 | 901 | |
d6aed795 | 902 | if (!fio_file_open(f)) { |
87b10676 JA |
903 | int err; |
904 | ||
905 | err = td_io_open_file(td, f); | |
906 | if (err) | |
907 | continue; | |
908 | opened = 1; | |
909 | } | |
910 | ||
2ba1c290 JA |
911 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
912 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 913 | return f; |
2ba1c290 | 914 | } |
87b10676 JA |
915 | if (opened) |
916 | td_io_close_file(td, f); | |
0aabe160 JA |
917 | } while (1); |
918 | } | |
919 | ||
920 | /* | |
921 | * Get next file to service by doing round robin between all available ones | |
922 | */ | |
1c178180 JA |
923 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
924 | int badf) | |
3d7c391d JA |
925 | { |
926 | unsigned int old_next_file = td->next_file; | |
927 | struct fio_file *f; | |
928 | ||
929 | do { | |
87b10676 JA |
930 | int opened = 0; |
931 | ||
126d65c6 | 932 | f = td->files[td->next_file]; |
3d7c391d JA |
933 | |
934 | td->next_file++; | |
2dc1bbeb | 935 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
936 | td->next_file = 0; |
937 | ||
87b10676 | 938 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 939 | if (fio_file_done(f)) { |
d5ed68ea | 940 | f = NULL; |
059e63c0 | 941 | continue; |
d5ed68ea | 942 | } |
059e63c0 | 943 | |
d6aed795 | 944 | if (!fio_file_open(f)) { |
87b10676 JA |
945 | int err; |
946 | ||
947 | err = td_io_open_file(td, f); | |
b5696bfc JA |
948 | if (err) { |
949 | dprint(FD_FILE, "error %d on open of %s\n", | |
950 | err, f->file_name); | |
87c27b45 | 951 | f = NULL; |
87b10676 | 952 | continue; |
b5696bfc | 953 | } |
87b10676 JA |
954 | opened = 1; |
955 | } | |
956 | ||
0b9d69ec JA |
957 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
958 | f->flags); | |
1c178180 | 959 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
960 | break; |
961 | ||
87b10676 JA |
962 | if (opened) |
963 | td_io_close_file(td, f); | |
964 | ||
3d7c391d JA |
965 | f = NULL; |
966 | } while (td->next_file != old_next_file); | |
967 | ||
2ba1c290 | 968 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
969 | return f; |
970 | } | |
971 | ||
7eb36574 | 972 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 973 | { |
1907dbc6 JA |
974 | struct fio_file *f; |
975 | ||
2dc1bbeb | 976 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 977 | |
b5696bfc | 978 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
979 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
980 | " nr_files=%d\n", td->nr_open_files, | |
981 | td->nr_done_files, | |
982 | td->o.nr_files); | |
bdb4e2e9 | 983 | return NULL; |
2ba1c290 | 984 | } |
bdb4e2e9 | 985 | |
1907dbc6 | 986 | f = td->file_service_file; |
d6aed795 | 987 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
988 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
989 | goto out; | |
990 | if (td->file_service_left--) | |
991 | goto out; | |
992 | } | |
1907dbc6 | 993 | |
a086c257 JA |
994 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
995 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 996 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 997 | else |
d6aed795 | 998 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 JA |
999 | |
1000 | td->file_service_file = f; | |
1001 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1002 | out: |
683023e8 | 1003 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
1907dbc6 | 1004 | return f; |
bdb4e2e9 JA |
1005 | } |
1006 | ||
7eb36574 JA |
1007 | static struct fio_file *get_next_file(struct thread_data *td) |
1008 | { | |
1009 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1010 | ||
1011 | if (ops->get_next_file) | |
1012 | return ops->get_next_file(td); | |
1013 | ||
1014 | return __get_next_file(td); | |
1015 | } | |
1016 | ||
429f6675 JA |
1017 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
1018 | { | |
1019 | struct fio_file *f; | |
1020 | ||
1021 | do { | |
1022 | f = get_next_file(td); | |
1023 | if (!f) | |
1024 | return 1; | |
1025 | ||
429f6675 JA |
1026 | io_u->file = f; |
1027 | get_file(f); | |
1028 | ||
1029 | if (!fill_io_u(td, io_u)) | |
1030 | break; | |
1031 | ||
b5696bfc | 1032 | put_file_log(td, f); |
429f6675 | 1033 | td_io_close_file(td, f); |
b5696bfc | 1034 | io_u->file = NULL; |
d6aed795 | 1035 | fio_file_set_done(f); |
429f6675 | 1036 | td->nr_done_files++; |
0b9d69ec JA |
1037 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
1038 | td->nr_done_files, td->o.nr_files); | |
429f6675 JA |
1039 | } while (1); |
1040 | ||
1041 | return 0; | |
1042 | } | |
1043 | ||
1044 | ||
10ba535a JA |
1045 | struct io_u *__get_io_u(struct thread_data *td) |
1046 | { | |
1047 | struct io_u *io_u = NULL; | |
1048 | ||
e8462bd8 JA |
1049 | td_io_u_lock(td); |
1050 | ||
1051 | again: | |
01743ee1 JA |
1052 | if (!flist_empty(&td->io_u_requeues)) |
1053 | io_u = flist_entry(td->io_u_requeues.next, struct io_u, list); | |
755200a3 | 1054 | else if (!queue_full(td)) { |
01743ee1 | 1055 | io_u = flist_entry(td->io_u_freelist.next, struct io_u, list); |
10ba535a | 1056 | |
6040dabc | 1057 | io_u->buflen = 0; |
10ba535a | 1058 | io_u->resid = 0; |
755200a3 | 1059 | io_u->file = NULL; |
d7762cf8 | 1060 | io_u->end_io = NULL; |
755200a3 JA |
1061 | } |
1062 | ||
1063 | if (io_u) { | |
0c6e7517 | 1064 | assert(io_u->flags & IO_U_F_FREE); |
2ecc1b57 | 1065 | io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF); |
1ef2b6be | 1066 | io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER); |
0c6e7517 | 1067 | |
755200a3 | 1068 | io_u->error = 0; |
01743ee1 JA |
1069 | flist_del(&io_u->list); |
1070 | flist_add(&io_u->list, &td->io_u_busylist); | |
10ba535a | 1071 | td->cur_depth++; |
0c41214f | 1072 | io_u->flags |= IO_U_F_IN_CUR_DEPTH; |
1dec3e07 JA |
1073 | } else if (td->o.verify_async) { |
1074 | /* | |
1075 | * We ran out, wait for async verify threads to finish and | |
1076 | * return one | |
1077 | */ | |
1078 | pthread_cond_wait(&td->free_cond, &td->io_u_lock); | |
1079 | goto again; | |
10ba535a JA |
1080 | } |
1081 | ||
e8462bd8 | 1082 | td_io_u_unlock(td); |
10ba535a JA |
1083 | return io_u; |
1084 | } | |
1085 | ||
0d29de83 | 1086 | static int check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1087 | { |
0d29de83 JA |
1088 | if (td->o.trim_backlog && td->trim_entries) { |
1089 | int get_trim = 0; | |
10ba535a | 1090 | |
0d29de83 JA |
1091 | if (td->trim_batch) { |
1092 | td->trim_batch--; | |
1093 | get_trim = 1; | |
1094 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1095 | td->last_ddir != DDIR_READ) { | |
1096 | td->trim_batch = td->o.trim_batch; | |
1097 | if (!td->trim_batch) | |
1098 | td->trim_batch = td->o.trim_backlog; | |
1099 | get_trim = 1; | |
1100 | } | |
1101 | ||
1102 | if (get_trim && !get_next_trim(td, io_u)) | |
1103 | return 1; | |
2ba1c290 | 1104 | } |
10ba535a | 1105 | |
0d29de83 JA |
1106 | return 0; |
1107 | } | |
1108 | ||
1109 | static int check_get_verify(struct thread_data *td, struct io_u *io_u) | |
1110 | { | |
9e144189 JA |
1111 | if (td->o.verify_backlog && td->io_hist_len) { |
1112 | int get_verify = 0; | |
1113 | ||
1114 | if (td->verify_batch) { | |
1115 | td->verify_batch--; | |
1116 | get_verify = 1; | |
1117 | } else if (!(td->io_hist_len % td->o.verify_backlog) && | |
1118 | td->last_ddir != DDIR_READ) { | |
1119 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1120 | if (!td->verify_batch) |
1121 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1122 | get_verify = 1; |
1123 | } | |
1124 | ||
1125 | if (get_verify && !get_next_verify(td, io_u)) | |
0d29de83 | 1126 | return 1; |
9e144189 JA |
1127 | } |
1128 | ||
0d29de83 JA |
1129 | return 0; |
1130 | } | |
1131 | ||
de789769 JA |
1132 | /* |
1133 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1134 | * easy compressible data for simple de-dupe attempts. Do this for every |
1135 | * 512b block in the range, since that should be the smallest block size | |
1136 | * we can expect from a device. | |
de789769 JA |
1137 | */ |
1138 | static void small_content_scramble(struct io_u *io_u) | |
1139 | { | |
23f394d5 | 1140 | unsigned int i, nr_blocks = io_u->buflen / 512; |
fad82f76 | 1141 | unsigned long long boffset; |
23f394d5 JA |
1142 | unsigned int offset; |
1143 | void *p, *end; | |
de789769 | 1144 | |
23f394d5 JA |
1145 | if (!nr_blocks) |
1146 | return; | |
1147 | ||
1148 | p = io_u->xfer_buf; | |
fba76ee8 | 1149 | boffset = io_u->offset; |
fad82f76 | 1150 | |
23f394d5 JA |
1151 | for (i = 0; i < nr_blocks; i++) { |
1152 | /* | |
1153 | * Fill the byte offset into a "random" start offset of | |
1154 | * the buffer, given by the product of the usec time | |
1155 | * and the actual offset. | |
1156 | */ | |
fad82f76 | 1157 | offset = (io_u->start_time.tv_usec ^ boffset) & 511; |
fba76ee8 | 1158 | offset &= ~(sizeof(unsigned long long) - 1); |
23f394d5 JA |
1159 | if (offset >= 512 - sizeof(unsigned long long)) |
1160 | offset -= sizeof(unsigned long long); | |
fba76ee8 | 1161 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1162 | |
1163 | end = p + 512 - sizeof(io_u->start_time); | |
1164 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1165 | p += 512; | |
fad82f76 | 1166 | boffset += 512; |
23f394d5 | 1167 | } |
de789769 JA |
1168 | } |
1169 | ||
0d29de83 JA |
1170 | /* |
1171 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1172 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1173 | */ | |
1174 | struct io_u *get_io_u(struct thread_data *td) | |
1175 | { | |
1176 | struct fio_file *f; | |
1177 | struct io_u *io_u; | |
de789769 | 1178 | int do_scramble = 0; |
0d29de83 JA |
1179 | |
1180 | io_u = __get_io_u(td); | |
1181 | if (!io_u) { | |
1182 | dprint(FD_IO, "__get_io_u failed\n"); | |
1183 | return NULL; | |
1184 | } | |
1185 | ||
1186 | if (check_get_verify(td, io_u)) | |
1187 | goto out; | |
1188 | if (check_get_trim(td, io_u)) | |
1189 | goto out; | |
1190 | ||
755200a3 JA |
1191 | /* |
1192 | * from a requeue, io_u already setup | |
1193 | */ | |
1194 | if (io_u->file) | |
77f392bf | 1195 | goto out; |
755200a3 | 1196 | |
429f6675 JA |
1197 | /* |
1198 | * If using an iolog, grab next piece if any available. | |
1199 | */ | |
1200 | if (td->o.read_iolog_file) { | |
1201 | if (read_iolog_get(td, io_u)) | |
1202 | goto err_put; | |
2ba1c290 JA |
1203 | } else if (set_io_u_file(td, io_u)) { |
1204 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
429f6675 | 1205 | goto err_put; |
2ba1c290 | 1206 | } |
5ec10eaa | 1207 | |
429f6675 | 1208 | f = io_u->file; |
d6aed795 | 1209 | assert(fio_file_open(f)); |
97af62ce | 1210 | |
ff58fced | 1211 | if (ddir_rw(io_u->ddir)) { |
d0656a93 | 1212 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
2ba1c290 | 1213 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1214 | goto err_put; |
2ba1c290 | 1215 | } |
10ba535a | 1216 | |
38dad62d | 1217 | f->last_start = io_u->offset; |
36167d82 | 1218 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 1219 | |
fd68418e JA |
1220 | if (io_u->ddir == DDIR_WRITE) { |
1221 | if (td->o.verify != VERIFY_NONE) | |
1222 | populate_verify_io_u(td, io_u); | |
1223 | else if (td->o.refill_buffers) | |
1224 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); | |
1225 | else if (td->o.scramble_buffers) | |
1226 | do_scramble = 1; | |
1227 | } else if (io_u->ddir == DDIR_READ) { | |
cbe8d756 RR |
1228 | /* |
1229 | * Reset the buf_filled parameters so next time if the | |
1230 | * buffer is used for writes it is refilled. | |
1231 | */ | |
cbe8d756 RR |
1232 | io_u->buf_filled_len = 0; |
1233 | } | |
87dc1ab1 | 1234 | } |
10ba535a | 1235 | |
165faf16 JA |
1236 | /* |
1237 | * Set io data pointers. | |
1238 | */ | |
cec6b55d JA |
1239 | io_u->xfer_buf = io_u->buf; |
1240 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1241 | |
6ac7a331 | 1242 | out: |
0d29de83 | 1243 | assert(io_u->file); |
429f6675 | 1244 | if (!td_io_prep(td, io_u)) { |
993bf48b JA |
1245 | if (!td->o.disable_slat) |
1246 | fio_gettime(&io_u->start_time, NULL); | |
de789769 JA |
1247 | if (do_scramble) |
1248 | small_content_scramble(io_u); | |
429f6675 | 1249 | return io_u; |
36167d82 | 1250 | } |
429f6675 | 1251 | err_put: |
2ba1c290 | 1252 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 JA |
1253 | put_io_u(td, io_u); |
1254 | return NULL; | |
10ba535a JA |
1255 | } |
1256 | ||
5451792e JA |
1257 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1258 | { | |
825f818e JA |
1259 | const char *msg[] = { "read", "write", "sync", "datasync", |
1260 | "sync_file_range", "wait", "trim" }; | |
1261 | ||
1262 | ||
5451792e JA |
1263 | |
1264 | log_err("fio: io_u error"); | |
1265 | ||
1266 | if (io_u->file) | |
1267 | log_err(" on file %s", io_u->file->file_name); | |
1268 | ||
1269 | log_err(": %s\n", strerror(io_u->error)); | |
1270 | ||
5ec10eaa JA |
1271 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
1272 | io_u->offset, io_u->xfer_buflen); | |
5451792e JA |
1273 | |
1274 | if (!td->error) | |
1275 | td_verror(td, io_u->error, "io_u error"); | |
1276 | } | |
1277 | ||
c8eeb9df JA |
1278 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1279 | struct io_completion_data *icd, | |
1280 | const enum fio_ddir idx, unsigned int bytes) | |
1281 | { | |
1282 | unsigned long uninitialized_var(lusec); | |
1283 | ||
1284 | if (!icd->account) | |
1285 | return; | |
1286 | ||
1287 | if (!td->o.disable_clat || !td->o.disable_bw) | |
1288 | lusec = utime_since(&io_u->issue_time, &icd->time); | |
1289 | ||
1290 | if (!td->o.disable_lat) { | |
1291 | unsigned long tusec; | |
1292 | ||
1293 | tusec = utime_since(&io_u->start_time, &icd->time); | |
1294 | add_lat_sample(td, idx, tusec, bytes); | |
1295 | } | |
1296 | ||
1297 | if (!td->o.disable_clat) { | |
1298 | add_clat_sample(td, idx, lusec, bytes); | |
1299 | io_u_mark_latency(td, lusec); | |
1300 | } | |
1301 | ||
1302 | if (!td->o.disable_bw) | |
1303 | add_bw_sample(td, idx, bytes, &icd->time); | |
1304 | ||
1305 | add_iops_sample(td, idx, &icd->time); | |
1306 | } | |
1307 | ||
1b8dbf25 SL |
1308 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
1309 | { | |
1310 | unsigned long long secs, remainder, bps, bytes; | |
1311 | bytes = td->this_io_bytes[ddir]; | |
1312 | bps = td->rate_bps[ddir]; | |
1313 | secs = bytes / bps; | |
1314 | remainder = bytes % bps; | |
1315 | return remainder * 1000000 / bps + secs * 1000000; | |
1316 | } | |
1317 | ||
97601024 JA |
1318 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
1319 | struct io_completion_data *icd) | |
10ba535a | 1320 | { |
dbad30b0 JA |
1321 | /* |
1322 | * Older gcc's are too dumb to realize that usec is always used | |
1323 | * initialized, silence that warning. | |
1324 | */ | |
1325 | unsigned long uninitialized_var(usec); | |
44f29692 | 1326 | struct fio_file *f; |
10ba535a | 1327 | |
2ba1c290 JA |
1328 | dprint_io_u(io_u, "io complete"); |
1329 | ||
2ecc1b57 | 1330 | td_io_u_lock(td); |
0c6e7517 | 1331 | assert(io_u->flags & IO_U_F_FLIGHT); |
38dad62d | 1332 | io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
2ecc1b57 | 1333 | td_io_u_unlock(td); |
0c6e7517 | 1334 | |
5f9099ea | 1335 | if (ddir_sync(io_u->ddir)) { |
87dc1ab1 | 1336 | td->last_was_sync = 1; |
44f29692 JA |
1337 | f = io_u->file; |
1338 | if (f) { | |
1339 | f->first_write = -1ULL; | |
1340 | f->last_write = -1ULL; | |
1341 | } | |
87dc1ab1 JA |
1342 | return; |
1343 | } | |
1344 | ||
1345 | td->last_was_sync = 0; | |
9e144189 | 1346 | td->last_ddir = io_u->ddir; |
87dc1ab1 | 1347 | |
ff58fced | 1348 | if (!io_u->error && ddir_rw(io_u->ddir)) { |
10ba535a | 1349 | unsigned int bytes = io_u->buflen - io_u->resid; |
1e97cce9 | 1350 | const enum fio_ddir idx = io_u->ddir; |
ba3e4e0c | 1351 | const enum fio_ddir odx = io_u->ddir ^ 1; |
b29ee5b3 | 1352 | int ret; |
10ba535a | 1353 | |
b29ee5b3 | 1354 | td->io_blocks[idx]++; |
c8eeb9df | 1355 | td->this_io_blocks[idx]++; |
b29ee5b3 JA |
1356 | td->io_bytes[idx] += bytes; |
1357 | td->this_io_bytes[idx] += bytes; | |
10ba535a | 1358 | |
44f29692 JA |
1359 | if (idx == DDIR_WRITE) { |
1360 | f = io_u->file; | |
1361 | if (f) { | |
1362 | if (f->first_write == -1ULL || | |
1363 | io_u->offset < f->first_write) | |
1364 | f->first_write = io_u->offset; | |
1365 | if (f->last_write == -1ULL || | |
1366 | ((io_u->offset + bytes) > f->last_write)) | |
1367 | f->last_write = io_u->offset + bytes; | |
1368 | } | |
1369 | } | |
1370 | ||
c8eeb9df JA |
1371 | if (ramp_time_over(td) && td->runstate == TD_RUNNING) { |
1372 | account_io_completion(td, io_u, icd, idx, bytes); | |
40e1a6f0 | 1373 | |
b23b6a2f | 1374 | if (__should_check_rate(td, idx)) { |
ba3e4e0c | 1375 | td->rate_pending_usleep[idx] = |
1b8dbf25 | 1376 | (usec_for_io(td, idx) - |
ba3e4e0c | 1377 | utime_since_now(&td->start)); |
b23b6a2f | 1378 | } |
1b8dbf25 | 1379 | if (__should_check_rate(td, odx)) |
ba3e4e0c | 1380 | td->rate_pending_usleep[odx] = |
1b8dbf25 | 1381 | (usec_for_io(td, odx) - |
ba3e4e0c | 1382 | utime_since_now(&td->start)); |
721938ae | 1383 | } |
10ba535a | 1384 | |
660a1cb5 | 1385 | if (td_write(td) && idx == DDIR_WRITE && |
8670579e | 1386 | td->o.do_verify && |
41128405 | 1387 | td->o.verify != VERIFY_NONE) |
10ba535a JA |
1388 | log_io_piece(td, io_u); |
1389 | ||
b29ee5b3 | 1390 | icd->bytes_done[idx] += bytes; |
3af6ef39 | 1391 | |
d7762cf8 | 1392 | if (io_u->end_io) { |
36690c9b | 1393 | ret = io_u->end_io(td, io_u); |
3af6ef39 JA |
1394 | if (ret && !icd->error) |
1395 | icd->error = ret; | |
1396 | } | |
ff58fced | 1397 | } else if (io_u->error) { |
10ba535a | 1398 | icd->error = io_u->error; |
5451792e JA |
1399 | io_u_log_error(td, io_u); |
1400 | } | |
06842027 SL |
1401 | if (icd->error && td_non_fatal_error(icd->error) && |
1402 | (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) { | |
f2bba182 RR |
1403 | /* |
1404 | * If there is a non_fatal error, then add to the error count | |
1405 | * and clear all the errors. | |
1406 | */ | |
1407 | update_error_count(td, icd->error); | |
1408 | td_clear_error(td); | |
1409 | icd->error = 0; | |
1410 | io_u->error = 0; | |
1411 | } | |
10ba535a JA |
1412 | } |
1413 | ||
9520ebb9 JA |
1414 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1415 | int nr) | |
10ba535a | 1416 | { |
9520ebb9 JA |
1417 | if (!td->o.disable_clat || !td->o.disable_bw) |
1418 | fio_gettime(&icd->time, NULL); | |
02bcaa8c | 1419 | |
3af6ef39 | 1420 | icd->nr = nr; |
c8eeb9df | 1421 | icd->account = 1; |
3af6ef39 | 1422 | |
10ba535a JA |
1423 | icd->error = 0; |
1424 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
1425 | } |
1426 | ||
97601024 JA |
1427 | static void ios_completed(struct thread_data *td, |
1428 | struct io_completion_data *icd) | |
36167d82 JA |
1429 | { |
1430 | struct io_u *io_u; | |
1431 | int i; | |
1432 | ||
10ba535a JA |
1433 | for (i = 0; i < icd->nr; i++) { |
1434 | io_u = td->io_ops->event(td, i); | |
1435 | ||
1436 | io_completed(td, io_u, icd); | |
e8462bd8 JA |
1437 | |
1438 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1439 | put_io_u(td, io_u); | |
c8eeb9df JA |
1440 | |
1441 | icd->account = 0; | |
10ba535a JA |
1442 | } |
1443 | } | |
97601024 | 1444 | |
e7e6cfb4 JA |
1445 | /* |
1446 | * Complete a single io_u for the sync engines. | |
1447 | */ | |
581e7141 JA |
1448 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
1449 | unsigned long *bytes) | |
97601024 JA |
1450 | { |
1451 | struct io_completion_data icd; | |
1452 | ||
9520ebb9 | 1453 | init_icd(td, &icd, 1); |
97601024 | 1454 | io_completed(td, io_u, &icd); |
e8462bd8 JA |
1455 | |
1456 | if (!(io_u->flags & IO_U_F_FREE_DEF)) | |
1457 | put_io_u(td, io_u); | |
97601024 | 1458 | |
581e7141 JA |
1459 | if (icd.error) { |
1460 | td_verror(td, icd.error, "io_u_sync_complete"); | |
1461 | return -1; | |
1462 | } | |
97601024 | 1463 | |
581e7141 JA |
1464 | if (bytes) { |
1465 | bytes[0] += icd.bytes_done[0]; | |
1466 | bytes[1] += icd.bytes_done[1]; | |
1467 | } | |
1468 | ||
1469 | return 0; | |
97601024 JA |
1470 | } |
1471 | ||
e7e6cfb4 JA |
1472 | /* |
1473 | * Called to complete min_events number of io for the async engines. | |
1474 | */ | |
581e7141 JA |
1475 | int io_u_queued_complete(struct thread_data *td, int min_evts, |
1476 | unsigned long *bytes) | |
97601024 | 1477 | { |
97601024 | 1478 | struct io_completion_data icd; |
00de55ef | 1479 | struct timespec *tvp = NULL; |
97601024 | 1480 | int ret; |
4d06a338 | 1481 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 1482 | |
4950421a | 1483 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts); |
b271fe62 | 1484 | |
4950421a | 1485 | if (!min_evts) |
00de55ef | 1486 | tvp = &ts; |
97601024 | 1487 | |
4950421a | 1488 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp); |
97601024 | 1489 | if (ret < 0) { |
e1161c32 | 1490 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
1491 | return ret; |
1492 | } else if (!ret) | |
1493 | return ret; | |
1494 | ||
9520ebb9 | 1495 | init_icd(td, &icd, ret); |
97601024 | 1496 | ios_completed(td, &icd); |
581e7141 JA |
1497 | if (icd.error) { |
1498 | td_verror(td, icd.error, "io_u_queued_complete"); | |
1499 | return -1; | |
1500 | } | |
97601024 | 1501 | |
581e7141 JA |
1502 | if (bytes) { |
1503 | bytes[0] += icd.bytes_done[0]; | |
1504 | bytes[1] += icd.bytes_done[1]; | |
1505 | } | |
1506 | ||
1507 | return 0; | |
97601024 | 1508 | } |
7e77dd02 JA |
1509 | |
1510 | /* | |
1511 | * Call when io_u is really queued, to update the submission latency. | |
1512 | */ | |
1513 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
1514 | { | |
9520ebb9 JA |
1515 | if (!td->o.disable_slat) { |
1516 | unsigned long slat_time; | |
7e77dd02 | 1517 | |
9520ebb9 | 1518 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
29a90ddb | 1519 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen); |
9520ebb9 | 1520 | } |
7e77dd02 | 1521 | } |
433afcb4 | 1522 | |
5973cafb JA |
1523 | /* |
1524 | * "randomly" fill the buffer contents | |
1525 | */ | |
1526 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
1527 | unsigned int max_bs) | |
1528 | { | |
e84b371e JA |
1529 | io_u->buf_filled_len = 0; |
1530 | ||
637ef8d9 | 1531 | if (!td->o.zero_buffers) |
3545a109 | 1532 | fill_random_buf(&td->buf_state, io_u->buf, max_bs); |
637ef8d9 JA |
1533 | else |
1534 | memset(io_u->buf, 0, max_bs); | |
5973cafb | 1535 | } |