Commit | Line | Data |
---|---|---|
abe4da87 JA |
1 | /* |
2 | * fio - the flexible io tester | |
3 | * | |
4 | * Copyright (C) 2005 Jens Axboe <axboe@suse.de> | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License as published by | |
8 | * the Free Software Foundation; either version 2 of the License, or | |
9 | * (at your option) any later version. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * along with this program; if not, write to the Free Software | |
18 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
19 | * | |
20 | */ | |
892199bd JA |
21 | #include <stdio.h> |
22 | #include <stdlib.h> | |
23 | #include <unistd.h> | |
24 | #include <fcntl.h> | |
25 | #include <string.h> | |
26 | #include <errno.h> | |
27 | #include <signal.h> | |
28 | #include <time.h> | |
e128065d | 29 | #include <math.h> |
49d2caab | 30 | #include <assert.h> |
189873de | 31 | #include <pthread.h> |
debf703a JA |
32 | #include <dirent.h> |
33 | #include <libgen.h> | |
892199bd JA |
34 | #include <sys/types.h> |
35 | #include <sys/stat.h> | |
36 | #include <sys/wait.h> | |
892199bd JA |
37 | #include <sys/ipc.h> |
38 | #include <sys/shm.h> | |
c94deb1c | 39 | #include <sys/ioctl.h> |
6e2c38cc | 40 | #include <sys/mman.h> |
892199bd JA |
41 | #include <asm/unistd.h> |
42 | ||
27c32a38 | 43 | #include "fio.h" |
892199bd | 44 | |
892199bd JA |
45 | #define MASK (4095) |
46 | ||
4240cfa1 | 47 | #define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK)) |
892199bd | 48 | |
27c32a38 JA |
49 | int groupid = 0; |
50 | int thread_number = 0; | |
51 | char run_str[MAX_JOBS + 1]; | |
52 | int shm_id = 0; | |
6f75d67c | 53 | static LIST_HEAD(disk_list); |
892199bd | 54 | |
8882f4e9 JA |
55 | static void update_io_ticks(void); |
56 | static void disk_util_timer_arm(void); | |
57 | ||
02bdd9ba JA |
58 | /* |
59 | * thread life cycle | |
60 | */ | |
61 | enum { | |
62 | TD_NOT_CREATED = 0, | |
63 | TD_CREATED, | |
e8457004 JA |
64 | TD_RUNNING, |
65 | TD_VERIFYING, | |
02bdd9ba JA |
66 | TD_EXITED, |
67 | TD_REAPED, | |
68 | }; | |
69 | ||
2c83567e JA |
70 | /* |
71 | * The io unit | |
72 | */ | |
73 | struct io_u { | |
74 | struct iocb iocb; | |
57d753e3 | 75 | struct timeval start_time; |
2c83567e JA |
76 | struct timeval issue_time; |
77 | ||
2c83567e JA |
78 | char *buf; |
79 | unsigned int buflen; | |
4ac89145 | 80 | unsigned long long offset; |
2c83567e JA |
81 | |
82 | struct list_head list; | |
83 | }; | |
84 | ||
62bb4285 | 85 | #define should_fsync(td) (td_write(td) && !(td)->odirect) |
02983297 | 86 | |
892199bd JA |
87 | static sem_t startup_sem; |
88 | ||
27c32a38 JA |
89 | #define TERMINATE_ALL (-1) |
90 | ||
91 | static void terminate_threads(int groupid) | |
892199bd JA |
92 | { |
93 | int i; | |
94 | ||
213b446c JA |
95 | for (i = 0; i < thread_number; i++) { |
96 | struct thread_data *td = &threads[i]; | |
97 | ||
27c32a38 JA |
98 | if (groupid == TERMINATE_ALL || groupid == td->groupid) { |
99 | td->terminate = 1; | |
100 | td->start_delay = 0; | |
101 | } | |
213b446c | 102 | } |
02bdd9ba JA |
103 | } |
104 | ||
27c32a38 | 105 | static void sig_handler(int sig) |
946d8870 | 106 | { |
8882f4e9 JA |
107 | switch (sig) { |
108 | case SIGALRM: | |
109 | update_io_ticks(); | |
110 | disk_util_timer_arm(); | |
111 | break; | |
112 | default: | |
eaa7586c JA |
113 | printf("\nfio: terminating on signal\n"); |
114 | fflush(stdout); | |
8882f4e9 JA |
115 | terminate_threads(TERMINATE_ALL); |
116 | break; | |
117 | } | |
892199bd JA |
118 | } |
119 | ||
5c24b2c4 | 120 | static unsigned long utime_since(struct timeval *s, struct timeval *e) |
892199bd JA |
121 | { |
122 | double sec, usec; | |
123 | ||
124 | sec = e->tv_sec - s->tv_sec; | |
125 | usec = e->tv_usec - s->tv_usec; | |
126 | if (sec > 0 && usec < 0) { | |
127 | sec--; | |
128 | usec += 1000000; | |
129 | } | |
130 | ||
131 | sec *= (double) 1000000; | |
132 | ||
133 | return sec + usec; | |
134 | } | |
135 | ||
fd11d7af JA |
136 | static unsigned long utime_since_now(struct timeval *s) |
137 | { | |
138 | struct timeval t; | |
139 | ||
140 | gettimeofday(&t, NULL); | |
141 | return utime_since(s, &t); | |
142 | } | |
143 | ||
5c24b2c4 | 144 | static unsigned long mtime_since(struct timeval *s, struct timeval *e) |
892199bd JA |
145 | { |
146 | double sec, usec; | |
147 | ||
148 | sec = e->tv_sec - s->tv_sec; | |
149 | usec = e->tv_usec - s->tv_usec; | |
150 | if (sec > 0 && usec < 0) { | |
151 | sec--; | |
152 | usec += 1000000; | |
153 | } | |
154 | ||
155 | sec *= (double) 1000; | |
156 | usec /= (double) 1000; | |
157 | ||
158 | return sec + usec; | |
159 | } | |
160 | ||
be33abe4 JA |
161 | static unsigned long mtime_since_now(struct timeval *s) |
162 | { | |
163 | struct timeval t; | |
164 | ||
165 | gettimeofday(&t, NULL); | |
166 | return mtime_since(s, &t); | |
167 | } | |
168 | ||
98168d55 JA |
169 | static inline unsigned long msec_now(struct timeval *s) |
170 | { | |
171 | return s->tv_sec * 1000 + s->tv_usec / 1000; | |
172 | } | |
173 | ||
49d2caab JA |
174 | static int random_map_free(struct thread_data *td, unsigned long long block) |
175 | { | |
75b2ab2c JA |
176 | unsigned int idx = RAND_MAP_IDX(td, block); |
177 | unsigned int bit = RAND_MAP_BIT(td, block); | |
49d2caab JA |
178 | |
179 | return (td->file_map[idx] & (1UL << bit)) == 0; | |
180 | } | |
181 | ||
182 | static int get_next_free_block(struct thread_data *td, unsigned long long *b) | |
892199bd | 183 | { |
49d2caab JA |
184 | int i; |
185 | ||
186 | *b = 0; | |
187 | i = 0; | |
188 | while ((*b) * td->min_bs < td->io_size) { | |
189 | if (td->file_map[i] != -1UL) { | |
190 | *b += ffz(td->file_map[i]); | |
191 | return 0; | |
192 | } | |
193 | ||
194 | *b += BLOCKS_PER_MAP; | |
195 | i++; | |
196 | } | |
197 | ||
198 | return 1; | |
199 | } | |
200 | ||
201 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) | |
202 | { | |
203 | unsigned long block = io_u->offset / td->min_bs; | |
204 | unsigned int blocks = 0; | |
205 | ||
206 | while (blocks < (io_u->buflen / td->min_bs)) { | |
207 | int idx, bit; | |
208 | ||
209 | if (!random_map_free(td, block)) | |
210 | break; | |
211 | ||
75b2ab2c JA |
212 | idx = RAND_MAP_IDX(td, block); |
213 | bit = RAND_MAP_BIT(td, block); | |
49d2caab JA |
214 | |
215 | assert(idx < td->num_maps); | |
216 | ||
217 | td->file_map[idx] |= (1UL << bit); | |
218 | block++; | |
219 | blocks++; | |
220 | } | |
221 | ||
222 | if ((blocks * td->min_bs) < io_u->buflen) | |
223 | io_u->buflen = blocks * td->min_bs; | |
224 | } | |
225 | ||
226 | static int get_next_offset(struct thread_data *td, unsigned long long *offset) | |
227 | { | |
75b2ab2c | 228 | unsigned long long b, rb; |
d32d9284 | 229 | long r; |
892199bd JA |
230 | |
231 | if (!td->sequential) { | |
49d2caab JA |
232 | unsigned long max_blocks = td->io_size / td->min_bs; |
233 | int loops = 50; | |
234 | ||
235 | do { | |
236 | lrand48_r(&td->random_state, &r); | |
237 | b = ((max_blocks - 1) * r / (RAND_MAX+1.0)); | |
75b2ab2c | 238 | rb = b + (td->file_offset / td->min_bs); |
49d2caab | 239 | loops--; |
75b2ab2c | 240 | } while (!random_map_free(td, rb) && loops); |
49d2caab JA |
241 | |
242 | if (!loops) { | |
243 | if (get_next_free_block(td, &b)) | |
244 | return 1; | |
245 | } | |
7889f07b | 246 | } else |
49d2caab | 247 | b = td->last_bytes / td->min_bs; |
7889f07b | 248 | |
49d2caab | 249 | *offset = (b * td->min_bs) + td->file_offset; |
75b2ab2c JA |
250 | if (*offset > td->file_size) |
251 | return 1; | |
252 | ||
49d2caab | 253 | return 0; |
7889f07b JA |
254 | } |
255 | ||
256 | static unsigned int get_next_buflen(struct thread_data *td) | |
257 | { | |
258 | unsigned int buflen; | |
d32d9284 | 259 | long r; |
7889f07b JA |
260 | |
261 | if (td->min_bs == td->max_bs) | |
262 | buflen = td->min_bs; | |
263 | else { | |
d32d9284 | 264 | lrand48_r(&td->bsrange_state, &r); |
7889f07b JA |
265 | buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0)); |
266 | buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1); | |
892199bd JA |
267 | } |
268 | ||
49d2caab JA |
269 | if (buflen > td->io_size - td->this_io_bytes) |
270 | buflen = td->io_size - td->this_io_bytes; | |
7889f07b | 271 | |
7889f07b | 272 | return buflen; |
892199bd JA |
273 | } |
274 | ||
57d753e3 JA |
275 | static inline void add_stat_sample(struct thread_data *td, struct io_stat *is, |
276 | unsigned long val) | |
892199bd | 277 | { |
57d753e3 JA |
278 | if (val > is->max_val) |
279 | is->max_val = val; | |
280 | if (val < is->min_val) | |
281 | is->min_val = val; | |
282 | ||
283 | is->val += val; | |
284 | is->val_sq += val * val; | |
285 | is->samples++; | |
286 | } | |
fd1ae4c9 | 287 | |
a0a9b35b JA |
288 | static void add_log_sample(struct thread_data *td, struct io_log *log, |
289 | unsigned long val) | |
290 | { | |
291 | if (log->nr_samples == log->max_samples) { | |
292 | int new_size = sizeof(struct io_sample) * log->max_samples * 2; | |
293 | ||
294 | log->log = realloc(log->log, new_size); | |
295 | log->max_samples <<= 1; | |
296 | } | |
297 | ||
298 | log->log[log->nr_samples].val = val; | |
299 | log->log[log->nr_samples].time = mtime_since_now(&td->start); | |
300 | log->nr_samples++; | |
301 | } | |
302 | ||
57d753e3 JA |
303 | static void add_clat_sample(struct thread_data *td, unsigned long msec) |
304 | { | |
305 | add_stat_sample(td, &td->clat_stat, msec); | |
a0a9b35b JA |
306 | |
307 | if (td->lat_log) | |
308 | add_log_sample(td, td->lat_log, msec); | |
57d753e3 | 309 | } |
fd1ae4c9 | 310 | |
57d753e3 JA |
311 | static void add_slat_sample(struct thread_data *td, unsigned long msec) |
312 | { | |
313 | add_stat_sample(td, &td->slat_stat, msec); | |
314 | } | |
fd1ae4c9 | 315 | |
645785e5 | 316 | static void add_bw_sample(struct thread_data *td) |
57d753e3 JA |
317 | { |
318 | unsigned long spent = mtime_since_now(&td->stat_sample_time); | |
319 | unsigned long rate; | |
320 | ||
1d035750 | 321 | if (spent < td->bw_avg_time) |
57d753e3 JA |
322 | return; |
323 | ||
49d2caab | 324 | rate = (td->this_io_bytes - td->stat_io_bytes) / spent; |
57d753e3 JA |
325 | add_stat_sample(td, &td->bw_stat, rate); |
326 | ||
a0a9b35b JA |
327 | if (td->bw_log) |
328 | add_log_sample(td, td->bw_log, rate); | |
329 | ||
57d753e3 | 330 | gettimeofday(&td->stat_sample_time, NULL); |
49d2caab | 331 | td->stat_io_bytes = td->this_io_bytes; |
892199bd JA |
332 | } |
333 | ||
fd11d7af JA |
334 | /* |
335 | * busy looping version for the last few usec | |
336 | */ | |
337 | static void __usec_sleep(int usec) | |
338 | { | |
339 | struct timeval start; | |
340 | ||
341 | gettimeofday(&start, NULL); | |
342 | while (utime_since_now(&start) < usec) | |
3782a8cd | 343 | nop; |
fd11d7af JA |
344 | } |
345 | ||
d15c5195 | 346 | static void usec_sleep(struct thread_data *td, unsigned long usec) |
892199bd | 347 | { |
d15c5195 JA |
348 | struct timespec req, rem; |
349 | ||
350 | req.tv_sec = usec / 1000000; | |
351 | req.tv_nsec = usec * 1000 - req.tv_sec * 1000000; | |
892199bd JA |
352 | |
353 | do { | |
fd11d7af JA |
354 | if (usec < 5000) { |
355 | __usec_sleep(usec); | |
356 | break; | |
357 | } | |
d15c5195 | 358 | |
86184d14 | 359 | rem.tv_sec = rem.tv_nsec = 0; |
d15c5195 JA |
360 | if (nanosleep(&req, &rem) < 0) |
361 | break; | |
362 | ||
363 | if ((rem.tv_sec + rem.tv_nsec) == 0) | |
892199bd | 364 | break; |
86184d14 JA |
365 | |
366 | req.tv_nsec = rem.tv_nsec; | |
d15c5195 JA |
367 | req.tv_sec = rem.tv_sec; |
368 | ||
369 | usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000; | |
370 | } while (!td->terminate); | |
892199bd JA |
371 | } |
372 | ||
9e850933 JA |
373 | static void rate_throttle(struct thread_data *td, unsigned long time_spent, |
374 | unsigned int bytes) | |
86184d14 | 375 | { |
9e850933 JA |
376 | unsigned long usec_cycle; |
377 | ||
4240cfa1 JA |
378 | if (!td->rate) |
379 | return; | |
380 | ||
9e850933 JA |
381 | usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs); |
382 | ||
383 | if (time_spent < usec_cycle) { | |
384 | unsigned long s = usec_cycle - time_spent; | |
86184d14 JA |
385 | |
386 | td->rate_pending_usleep += s; | |
fad86e6a | 387 | if (td->rate_pending_usleep >= 100000) { |
d15c5195 | 388 | usec_sleep(td, td->rate_pending_usleep); |
86184d14 JA |
389 | td->rate_pending_usleep = 0; |
390 | } | |
4240cfa1 | 391 | } else { |
9e850933 | 392 | long overtime = time_spent - usec_cycle; |
42b2b9fe | 393 | |
4240cfa1 JA |
394 | td->rate_pending_usleep -= overtime; |
395 | } | |
396 | } | |
397 | ||
5c24b2c4 | 398 | static int check_min_rate(struct thread_data *td, struct timeval *now) |
4240cfa1 | 399 | { |
7607bc6b | 400 | unsigned long spent; |
4240cfa1 JA |
401 | unsigned long rate; |
402 | ||
403 | /* | |
404 | * allow a 2 second settle period in the beginning | |
405 | */ | |
7607bc6b | 406 | if (mtime_since(&td->start, now) < 2000) |
4240cfa1 JA |
407 | return 0; |
408 | ||
409 | /* | |
410 | * if rate blocks is set, sample is running | |
411 | */ | |
49d2caab | 412 | if (td->rate_bytes) { |
4240cfa1 JA |
413 | spent = mtime_since(&td->lastrate, now); |
414 | if (spent < td->ratecycle) | |
415 | return 0; | |
416 | ||
49d2caab | 417 | rate = (td->this_io_bytes - td->rate_bytes) / spent; |
4240cfa1 JA |
418 | if (rate < td->ratemin) { |
419 | printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate); | |
02bdd9ba | 420 | if (rate_quit) |
27c32a38 | 421 | terminate_threads(td->groupid); |
4240cfa1 JA |
422 | return 1; |
423 | } | |
86184d14 | 424 | } |
4240cfa1 | 425 | |
49d2caab | 426 | td->rate_bytes = td->this_io_bytes; |
4240cfa1 JA |
427 | memcpy(&td->lastrate, now, sizeof(*now)); |
428 | return 0; | |
86184d14 JA |
429 | } |
430 | ||
67903a2e JA |
431 | static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) |
432 | { | |
01f79976 JA |
433 | if (!td->timeout) |
434 | return 0; | |
67903a2e JA |
435 | if (mtime_since(&td->start, t) >= td->timeout * 1000) |
436 | return 1; | |
437 | ||
438 | return 0; | |
439 | } | |
440 | ||
e8457004 JA |
441 | static void fill_random_bytes(struct thread_data *td, |
442 | unsigned char *p, unsigned int len) | |
443 | { | |
645785e5 | 444 | unsigned int todo; |
40ef7f64 | 445 | double r; |
e8457004 JA |
446 | |
447 | while (len) { | |
40ef7f64 | 448 | drand48_r(&td->verify_state, &r); |
e8457004 | 449 | |
40ef7f64 JA |
450 | /* |
451 | * lrand48_r seems to be broken and only fill the bottom | |
452 | * 32-bits, even on 64-bit archs with 64-bit longs | |
453 | */ | |
454 | todo = sizeof(r); | |
e8457004 JA |
455 | if (todo > len) |
456 | todo = len; | |
457 | ||
458 | memcpy(p, &r, todo); | |
459 | ||
460 | len -= todo; | |
461 | p += todo; | |
462 | } | |
463 | } | |
464 | ||
9d0c6ca2 JA |
465 | static void hexdump(void *buffer, int len) |
466 | { | |
467 | unsigned char *p = buffer; | |
468 | int i; | |
469 | ||
470 | for (i = 0; i < len; i++) | |
471 | printf("%02x", p[i]); | |
472 | printf("\n"); | |
473 | } | |
474 | ||
7f46ef08 JA |
475 | static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u) |
476 | { | |
477 | unsigned char *p = (unsigned char *) io_u->buf; | |
478 | unsigned long c; | |
479 | ||
480 | p += sizeof(*hdr); | |
481 | c = crc32(p, hdr->len - sizeof(*hdr)); | |
482 | ||
483 | return c != hdr->crc32; | |
484 | } | |
485 | ||
486 | static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u) | |
e8457004 | 487 | { |
e8457004 JA |
488 | unsigned char *p = (unsigned char *) io_u->buf; |
489 | struct md5_ctx md5_ctx; | |
9d0c6ca2 | 490 | int ret; |
e8457004 | 491 | |
e8457004 JA |
492 | memset(&md5_ctx, 0, sizeof(md5_ctx)); |
493 | p += sizeof(*hdr); | |
494 | md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr)); | |
495 | ||
9d0c6ca2 JA |
496 | ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash)); |
497 | if (ret) { | |
498 | hexdump(hdr->md5_digest, sizeof(hdr->md5_digest)); | |
499 | hexdump(md5_ctx.hash, sizeof(md5_ctx.hash)); | |
500 | } | |
501 | ||
502 | return ret; | |
e8457004 JA |
503 | } |
504 | ||
7f46ef08 JA |
505 | static int verify_io_u(struct io_u *io_u) |
506 | { | |
507 | struct verify_header *hdr = (struct verify_header *) io_u->buf; | |
508 | int ret; | |
509 | ||
510 | if (hdr->fio_magic != FIO_HDR_MAGIC) | |
511 | return 1; | |
512 | ||
513 | if (hdr->verify_type == VERIFY_MD5) | |
514 | ret = verify_io_u_md5(hdr, io_u); | |
515 | else if (hdr->verify_type == VERIFY_CRC32) | |
516 | ret = verify_io_u_crc32(hdr, io_u); | |
517 | else { | |
518 | fprintf(stderr, "Bad verify type %d\n", hdr->verify_type); | |
519 | ret = 1; | |
520 | } | |
521 | ||
522 | return ret; | |
523 | } | |
524 | ||
525 | static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len) | |
526 | { | |
527 | hdr->crc32 = crc32(p, len); | |
528 | } | |
529 | ||
530 | static void fill_md5(struct verify_header *hdr, void *p, unsigned int len) | |
531 | { | |
532 | struct md5_ctx md5_ctx; | |
533 | ||
534 | memset(&md5_ctx, 0, sizeof(md5_ctx)); | |
535 | md5_update(&md5_ctx, p, len); | |
536 | memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash)); | |
537 | } | |
538 | ||
cfc702bd JA |
539 | /* |
540 | * fill body of io_u->buf with random data and add a header with the | |
541 | * (eg) sha1sum of that data. | |
542 | */ | |
e8457004 | 543 | static void populate_io_u(struct thread_data *td, struct io_u *io_u) |
cfc702bd | 544 | { |
e8457004 | 545 | unsigned char *p = (unsigned char *) io_u->buf; |
7f46ef08 | 546 | struct verify_header hdr; |
e8457004 JA |
547 | |
548 | hdr.fio_magic = FIO_HDR_MAGIC; | |
549 | hdr.len = io_u->buflen; | |
550 | p += sizeof(hdr); | |
551 | fill_random_bytes(td, p, io_u->buflen - sizeof(hdr)); | |
552 | ||
7f46ef08 JA |
553 | if (td->verify == VERIFY_MD5) { |
554 | fill_md5(&hdr, p, io_u->buflen - sizeof(hdr)); | |
555 | hdr.verify_type = VERIFY_MD5; | |
556 | } else { | |
557 | fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr)); | |
558 | hdr.verify_type = VERIFY_CRC32; | |
559 | } | |
560 | ||
e8457004 | 561 | memcpy(io_u->buf, &hdr, sizeof(hdr)); |
cfc702bd JA |
562 | } |
563 | ||
2c83567e JA |
564 | static void put_io_u(struct thread_data *td, struct io_u *io_u) |
565 | { | |
566 | list_del(&io_u->list); | |
567 | list_add(&io_u->list, &td->io_u_freelist); | |
568 | td->cur_depth--; | |
569 | } | |
570 | ||
f0f3411b JA |
571 | #define queue_full(td) (list_empty(&(td)->io_u_freelist)) |
572 | ||
e8457004 JA |
573 | static struct io_u *__get_io_u(struct thread_data *td) |
574 | { | |
575 | struct io_u *io_u; | |
576 | ||
f0f3411b | 577 | if (queue_full(td)) |
e8457004 JA |
578 | return NULL; |
579 | ||
580 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); | |
581 | list_del(&io_u->list); | |
582 | list_add(&io_u->list, &td->io_u_busylist); | |
f4bb2243 | 583 | td->cur_depth++; |
e8457004 JA |
584 | return io_u; |
585 | } | |
586 | ||
2c83567e JA |
587 | static struct io_u *get_io_u(struct thread_data *td) |
588 | { | |
589 | struct io_u *io_u; | |
590 | ||
e8457004 JA |
591 | io_u = __get_io_u(td); |
592 | if (!io_u) | |
2c83567e JA |
593 | return NULL; |
594 | ||
406e7b7c JA |
595 | if (get_next_offset(td, &io_u->offset)) { |
596 | put_io_u(td, io_u); | |
49d2caab | 597 | return NULL; |
406e7b7c | 598 | } |
49d2caab | 599 | |
b2a369fb JA |
600 | io_u->buflen = get_next_buflen(td); |
601 | if (!io_u->buflen) { | |
e8457004 | 602 | put_io_u(td, io_u); |
7889f07b | 603 | return NULL; |
e8457004 | 604 | } |
2c83567e | 605 | |
75b2ab2c JA |
606 | if (io_u->buflen + io_u->offset > td->file_size) |
607 | io_u->buflen = td->file_size - io_u->offset; | |
49d2caab JA |
608 | |
609 | if (!td->sequential) | |
610 | mark_random_map(td, io_u); | |
611 | ||
612 | td->last_bytes += io_u->buflen; | |
613 | ||
7f46ef08 | 614 | if (td->verify != VERIFY_NONE) |
e8457004 | 615 | populate_io_u(td, io_u); |
cfc702bd | 616 | |
2c83567e JA |
617 | if (td->use_aio) { |
618 | if (td_read(td)) | |
619 | io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset); | |
620 | else | |
621 | io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset); | |
622 | } | |
623 | ||
57d753e3 | 624 | gettimeofday(&io_u->start_time, NULL); |
2c83567e JA |
625 | return io_u; |
626 | } | |
627 | ||
40ef7f64 JA |
628 | static inline void td_set_runstate(struct thread_data *td, int runstate) |
629 | { | |
630 | td->old_runstate = td->runstate; | |
631 | td->runstate = runstate; | |
632 | } | |
633 | ||
645785e5 JA |
634 | static int get_next_verify(struct thread_data *td, |
635 | unsigned long long *offset, unsigned int *len) | |
636 | { | |
637 | struct io_piece *ipo; | |
638 | ||
639 | if (list_empty(&td->io_hist_list)) | |
640 | return 1; | |
641 | ||
642 | ipo = list_entry(td->io_hist_list.next, struct io_piece, list); | |
643 | list_del(&ipo->list); | |
644 | ||
645 | *offset = ipo->offset; | |
646 | *len = ipo->len; | |
647 | free(ipo); | |
648 | return 0; | |
649 | } | |
650 | ||
9d0c6ca2 JA |
651 | static void prune_io_piece_log(struct thread_data *td) |
652 | { | |
653 | struct io_piece *ipo; | |
654 | ||
655 | while (!list_empty(&td->io_hist_list)) { | |
656 | ipo = list_entry(td->io_hist_list.next, struct io_piece, list); | |
657 | ||
658 | list_del(&ipo->list); | |
659 | free(ipo); | |
660 | } | |
661 | } | |
662 | ||
9d0c6ca2 JA |
663 | /* |
664 | * log a succesful write, so we can unwind the log for verify | |
665 | */ | |
666 | static void log_io_piece(struct thread_data *td, struct io_u *io_u) | |
667 | { | |
49d2caab | 668 | struct io_piece *ipo = malloc(sizeof(struct io_piece)); |
9d0c6ca2 JA |
669 | struct list_head *entry; |
670 | ||
671 | INIT_LIST_HEAD(&ipo->list); | |
672 | ipo->offset = io_u->offset; | |
673 | ipo->len = io_u->buflen; | |
674 | ||
49d2caab JA |
675 | /* |
676 | * for random io where the writes extend the file, it will typically | |
677 | * be laid out with the block scattered as written. it's faster to | |
678 | * read them in in that order again, so don't sort | |
679 | */ | |
680 | if (td->sequential || !td->overwrite) { | |
9d0c6ca2 JA |
681 | list_add_tail(&ipo->list, &td->io_hist_list); |
682 | return; | |
683 | } | |
684 | ||
685 | /* | |
686 | * for random io, sort the list so verify will run faster | |
687 | */ | |
688 | entry = &td->io_hist_list; | |
689 | while ((entry = entry->prev) != &td->io_hist_list) { | |
690 | struct io_piece *__ipo = list_entry(entry, struct io_piece, list); | |
691 | ||
9d0c6ca2 JA |
692 | if (__ipo->offset < ipo->offset) |
693 | break; | |
694 | } | |
695 | ||
696 | list_add(&ipo->list, entry); | |
697 | } | |
698 | ||
91fc5dc9 | 699 | static void do_sync_verify(struct thread_data *td) |
cfc702bd | 700 | { |
40ef7f64 | 701 | struct timeval t; |
e8457004 | 702 | struct io_u *io_u = NULL; |
645785e5 | 703 | int ret; |
e8457004 | 704 | |
40ef7f64 | 705 | td_set_runstate(td, TD_VERIFYING); |
e8457004 JA |
706 | |
707 | io_u = __get_io_u(td); | |
708 | ||
40ef7f64 | 709 | if (!td->odirect) { |
6e2c38cc JA |
710 | if (!td->use_mmap) { |
711 | if (fadvise(td->fd, td->file_offset, td->io_size, POSIX_FADV_DONTNEED) < 0) { | |
712 | td->error = errno; | |
713 | goto out; | |
714 | } | |
715 | } else { | |
716 | if (madvise(td->mmap, td->io_size, MADV_DONTNEED)) { | |
717 | td->error = errno; | |
718 | goto out; | |
719 | } | |
40ef7f64 JA |
720 | } |
721 | } | |
722 | ||
e8457004 JA |
723 | do { |
724 | if (td->terminate) | |
725 | break; | |
40ef7f64 JA |
726 | |
727 | gettimeofday(&t, NULL); | |
728 | if (runtime_exceeded(td, &t)) | |
729 | break; | |
730 | ||
645785e5 JA |
731 | if (get_next_verify(td, &io_u->offset, &io_u->buflen)) |
732 | break; | |
733 | ||
734 | if (td->cur_off != io_u->offset) { | |
735 | if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) { | |
736 | td->error = errno; | |
737 | break; | |
738 | } | |
739 | } | |
e8457004 JA |
740 | |
741 | ret = read(td->fd, io_u->buf, io_u->buflen); | |
742 | if (ret < (int) io_u->buflen) { | |
743 | if (ret == -1) { | |
744 | td->error = errno; | |
745 | break; | |
746 | } else if (!ret) | |
747 | break; | |
748 | else | |
749 | io_u->buflen = ret; | |
750 | } | |
751 | ||
645785e5 | 752 | if (verify_io_u(io_u)) |
e8457004 JA |
753 | break; |
754 | ||
645785e5 | 755 | td->cur_off = io_u->offset + io_u->buflen; |
e8457004 JA |
756 | } while (1); |
757 | ||
758 | out: | |
40ef7f64 | 759 | td_set_runstate(td, TD_RUNNING); |
e8457004 | 760 | put_io_u(td, io_u); |
cfc702bd JA |
761 | } |
762 | ||
6e2c38cc JA |
763 | static int __do_sync_mmap(struct thread_data *td, struct io_u *io_u) |
764 | { | |
765 | unsigned long long real_off = io_u->offset - td->file_offset; | |
766 | ||
767 | if (td_read(td)) | |
768 | memcpy(io_u->buf, td->mmap + real_off, io_u->buflen); | |
769 | else | |
770 | memcpy(td->mmap + real_off, io_u->buf, io_u->buflen); | |
c009cdac JA |
771 | |
772 | /* | |
773 | * not really direct, but should drop the pages from the cache | |
774 | */ | |
775 | if (td->odirect) { | |
776 | msync(td->mmap + real_off, io_u->buflen, MS_SYNC); | |
777 | madvise(td->mmap + real_off, io_u->buflen, MADV_DONTNEED); | |
778 | } | |
6e2c38cc JA |
779 | |
780 | return io_u->buflen; | |
781 | } | |
782 | ||
783 | static int __do_sync_rw(struct thread_data *td, struct io_u *io_u) | |
784 | { | |
785 | if (td->cur_off != io_u->offset) { | |
786 | if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) { | |
787 | td->error = errno; | |
788 | return 1; | |
789 | } | |
790 | } | |
791 | ||
792 | if (td_read(td)) | |
793 | return read(td->fd, io_u->buf, io_u->buflen); | |
794 | else | |
795 | return write(td->fd, io_u->buf, io_u->buflen); | |
796 | } | |
797 | ||
798 | static void sync_td(struct thread_data *td) | |
799 | { | |
800 | if (!td->use_mmap) | |
801 | fsync(td->fd); | |
802 | else | |
803 | msync(td->mmap, td->file_size, MS_SYNC); | |
804 | } | |
805 | ||
43000118 | 806 | static void do_sync_io(struct thread_data *td) |
892199bd | 807 | { |
7889f07b | 808 | unsigned long msec, usec; |
e8457004 | 809 | struct io_u *io_u = NULL; |
2c83567e | 810 | struct timeval e; |
892199bd | 811 | |
49d2caab | 812 | while (td->this_io_bytes < td->io_size) { |
892199bd JA |
813 | int ret; |
814 | ||
815 | if (td->terminate) | |
816 | break; | |
817 | ||
2c83567e | 818 | io_u = get_io_u(td); |
7889f07b JA |
819 | if (!io_u) |
820 | break; | |
2c83567e | 821 | |
6e2c38cc JA |
822 | if (!td->use_mmap) |
823 | ret = __do_sync_rw(td, io_u); | |
892199bd | 824 | else |
6e2c38cc | 825 | ret = __do_sync_mmap(td, io_u); |
892199bd | 826 | |
2c83567e | 827 | if (ret < (int) io_u->buflen) { |
892199bd JA |
828 | if (ret == -1) |
829 | td->error = errno; | |
830 | break; | |
831 | } | |
832 | ||
62bb4285 | 833 | if (td_write(td)) |
645785e5 JA |
834 | log_io_piece(td, io_u); |
835 | ||
4240cfa1 | 836 | td->io_blocks++; |
49d2caab JA |
837 | td->io_bytes += io_u->buflen; |
838 | td->this_io_bytes += io_u->buflen; | |
63a09e51 | 839 | td->cur_off = io_u->offset + io_u->buflen; |
4240cfa1 | 840 | |
86184d14 JA |
841 | gettimeofday(&e, NULL); |
842 | ||
57d753e3 | 843 | usec = utime_since(&io_u->start_time, &e); |
86184d14 | 844 | |
9e850933 | 845 | rate_throttle(td, usec, io_u->buflen); |
892199bd | 846 | |
4240cfa1 JA |
847 | if (check_min_rate(td, &e)) { |
848 | td->error = ENODATA; | |
849 | break; | |
850 | } | |
892199bd | 851 | |
4240cfa1 | 852 | msec = usec / 1000; |
57d753e3 | 853 | add_clat_sample(td, msec); |
645785e5 | 854 | add_bw_sample(td); |
67903a2e JA |
855 | |
856 | if (runtime_exceeded(td, &e)) | |
857 | break; | |
2c83567e | 858 | |
cdf92433 | 859 | put_io_u(td, io_u); |
e8457004 | 860 | io_u = NULL; |
cdf92433 | 861 | |
e97712ed | 862 | if (td->thinktime) |
d15c5195 | 863 | usec_sleep(td, td->thinktime); |
e97712ed | 864 | |
cdf92433 JA |
865 | if (should_fsync(td) && td->fsync_blocks && |
866 | (td->io_blocks % td->fsync_blocks) == 0) | |
6e2c38cc | 867 | sync_td(td); |
892199bd JA |
868 | } |
869 | ||
e8457004 JA |
870 | if (io_u) |
871 | put_io_u(td, io_u); | |
872 | ||
4240cfa1 | 873 | if (should_fsync(td)) |
6e2c38cc | 874 | sync_td(td); |
892199bd | 875 | } |
43000118 | 876 | |
1ad72b11 JA |
877 | static int io_u_getevents(struct thread_data *td, int min, int max, |
878 | struct timespec *t) | |
879 | { | |
880 | int r; | |
881 | ||
882 | do { | |
883 | r = io_getevents(td->aio_ctx, min, max, td->aio_events, t); | |
884 | if (r != -EAGAIN && r != -EINTR) | |
885 | break; | |
886 | } while (1); | |
887 | ||
888 | return r; | |
889 | } | |
890 | ||
2c83567e | 891 | static int io_u_queue(struct thread_data *td, struct io_u *io_u) |
56b0eff0 | 892 | { |
2c83567e | 893 | struct iocb *iocb = &io_u->iocb; |
56b0eff0 JA |
894 | int ret; |
895 | ||
896 | do { | |
254605cd | 897 | ret = io_submit(td->aio_ctx, 1, &iocb); |
56b0eff0 JA |
898 | if (ret == 1) |
899 | return 0; | |
406e7b7c | 900 | else if (ret == -EAGAIN) |
56b0eff0 | 901 | usleep(100); |
406e7b7c | 902 | else if (ret == -EINTR) |
a592bd33 | 903 | continue; |
56b0eff0 JA |
904 | else |
905 | break; | |
906 | } while (1); | |
907 | ||
a592bd33 | 908 | return ret; |
56b0eff0 JA |
909 | } |
910 | ||
98168d55 | 911 | #define iocb_time(iocb) ((unsigned long) (iocb)->data) |
2c83567e JA |
912 | #define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj) |
913 | ||
f0f3411b | 914 | static int ios_completed(struct thread_data *td, int nr) |
2c83567e JA |
915 | { |
916 | unsigned long msec; | |
917 | struct io_u *io_u; | |
918 | struct timeval e; | |
9e850933 | 919 | int i, bytes_done; |
2c83567e | 920 | |
f0f3411b | 921 | gettimeofday(&e, NULL); |
2c83567e | 922 | |
9e850933 | 923 | for (i = 0, bytes_done = 0; i < nr; i++) { |
2c83567e JA |
924 | io_u = ev_to_iou(td->aio_events + i); |
925 | ||
f0f3411b | 926 | td->io_blocks++; |
49d2caab JA |
927 | td->io_bytes += io_u->buflen; |
928 | td->this_io_bytes += io_u->buflen; | |
8c033f93 | 929 | |
f0f3411b | 930 | msec = mtime_since(&io_u->issue_time, &e); |
2c83567e | 931 | |
f0f3411b | 932 | add_clat_sample(td, msec); |
645785e5 JA |
933 | add_bw_sample(td); |
934 | ||
62bb4285 | 935 | if (td_write(td)) |
645785e5 | 936 | log_io_piece(td, io_u); |
2c83567e | 937 | |
f4bb2243 | 938 | bytes_done += io_u->buflen; |
2c83567e JA |
939 | put_io_u(td, io_u); |
940 | } | |
9e850933 JA |
941 | |
942 | return bytes_done; | |
2c83567e JA |
943 | } |
944 | ||
945 | static void cleanup_pending_aio(struct thread_data *td) | |
946 | { | |
947 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; | |
948 | struct list_head *entry, *n; | |
949 | struct io_u *io_u; | |
950 | int r; | |
951 | ||
952 | /* | |
953 | * get immediately available events, if any | |
954 | */ | |
1ad72b11 | 955 | r = io_u_getevents(td, 0, td->cur_depth, &ts); |
2c83567e | 956 | if (r > 0) |
f0f3411b | 957 | ios_completed(td, r); |
2c83567e JA |
958 | |
959 | /* | |
960 | * now cancel remaining active events | |
961 | */ | |
962 | list_for_each_safe(entry, n, &td->io_u_busylist) { | |
963 | io_u = list_entry(entry, struct io_u, list); | |
964 | ||
965 | r = io_cancel(td->aio_ctx, &io_u->iocb, td->aio_events); | |
966 | if (!r) | |
967 | put_io_u(td, io_u); | |
968 | } | |
969 | ||
970 | if (td->cur_depth) { | |
1ad72b11 | 971 | r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL); |
2c83567e | 972 | if (r > 0) |
f0f3411b | 973 | ios_completed(td, r); |
2c83567e JA |
974 | } |
975 | } | |
98168d55 | 976 | |
d32d9284 JA |
977 | static int async_do_verify(struct thread_data *td, struct io_u **io_u) |
978 | { | |
979 | struct io_u *v_io_u = *io_u; | |
980 | int ret = 0; | |
981 | ||
982 | if (v_io_u) { | |
645785e5 | 983 | ret = verify_io_u(v_io_u); |
d32d9284 JA |
984 | put_io_u(td, v_io_u); |
985 | *io_u = NULL; | |
986 | } | |
987 | ||
988 | return ret; | |
989 | } | |
990 | ||
91fc5dc9 | 991 | static void do_async_verify(struct thread_data *td) |
cfc702bd | 992 | { |
f4bb2243 | 993 | struct timeval t; |
d32d9284 | 994 | struct io_u *io_u, *v_io_u = NULL; |
645785e5 | 995 | int ret; |
f4bb2243 JA |
996 | |
997 | td_set_runstate(td, TD_VERIFYING); | |
998 | ||
f4bb2243 JA |
999 | do { |
1000 | if (td->terminate) | |
1001 | break; | |
1002 | ||
1003 | gettimeofday(&t, NULL); | |
1004 | if (runtime_exceeded(td, &t)) | |
1005 | break; | |
1006 | ||
1007 | io_u = __get_io_u(td); | |
1008 | if (!io_u) | |
1009 | break; | |
1010 | ||
645785e5 JA |
1011 | if (get_next_verify(td, &io_u->offset, &io_u->buflen)) { |
1012 | put_io_u(td, io_u); | |
1013 | break; | |
f4bb2243 JA |
1014 | } |
1015 | ||
1016 | io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset); | |
1017 | ret = io_u_queue(td, io_u); | |
1018 | if (ret) { | |
1019 | put_io_u(td, io_u); | |
1020 | td->error = ret; | |
1021 | break; | |
1022 | } | |
1023 | ||
f0f3411b JA |
1024 | /* |
1025 | * we have one pending to verify, do that while the next | |
1026 | * we are doing io on the next one | |
1027 | */ | |
d32d9284 JA |
1028 | if (async_do_verify(td, &v_io_u)) |
1029 | break; | |
f0f3411b | 1030 | |
1ad72b11 | 1031 | ret = io_u_getevents(td, 1, 1, NULL); |
f0f3411b JA |
1032 | if (ret != 1) { |
1033 | if (ret < 0) | |
1034 | td->error = ret; | |
f4bb2243 JA |
1035 | break; |
1036 | } | |
1037 | ||
f0f3411b | 1038 | v_io_u = ev_to_iou(td->aio_events); |
f4bb2243 | 1039 | |
645785e5 | 1040 | td->cur_off = v_io_u->offset + v_io_u->buflen; |
f0f3411b JA |
1041 | |
1042 | /* | |
d32d9284 | 1043 | * if we can't submit more io, we need to verify now |
f0f3411b | 1044 | */ |
d32d9284 JA |
1045 | if (queue_full(td) && async_do_verify(td, &v_io_u)) |
1046 | break; | |
1047 | ||
f4bb2243 JA |
1048 | } while (1); |
1049 | ||
d32d9284 | 1050 | async_do_verify(td, &v_io_u); |
f0f3411b | 1051 | |
f4bb2243 JA |
1052 | if (td->cur_depth) |
1053 | cleanup_pending_aio(td); | |
1054 | ||
1055 | td_set_runstate(td, TD_RUNNING); | |
cfc702bd JA |
1056 | } |
1057 | ||
43000118 JA |
1058 | static void do_async_io(struct thread_data *td) |
1059 | { | |
1060 | struct timeval s, e; | |
7889f07b | 1061 | unsigned long usec; |
43000118 | 1062 | |
49d2caab | 1063 | while (td->this_io_bytes < td->io_size) { |
43000118 JA |
1064 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; |
1065 | struct timespec *timeout; | |
2c83567e JA |
1066 | int ret, min_evts = 0; |
1067 | struct io_u *io_u; | |
9e850933 | 1068 | unsigned int bytes_done; |
43000118 JA |
1069 | |
1070 | if (td->terminate) | |
1071 | break; | |
1072 | ||
2c83567e | 1073 | io_u = get_io_u(td); |
7889f07b JA |
1074 | if (!io_u) |
1075 | break; | |
43000118 | 1076 | |
57d753e3 | 1077 | memcpy(&s, &io_u->start_time, sizeof(s)); |
8baf1bcc | 1078 | |
2c83567e | 1079 | ret = io_u_queue(td, io_u); |
56b0eff0 | 1080 | if (ret) { |
a3fdb993 | 1081 | put_io_u(td, io_u); |
a592bd33 | 1082 | td->error = ret; |
43000118 JA |
1083 | break; |
1084 | } | |
1085 | ||
57d753e3 JA |
1086 | gettimeofday(&io_u->issue_time, NULL); |
1087 | add_slat_sample(td, mtime_since(&io_u->start_time, &io_u->issue_time)); | |
2c83567e | 1088 | if (td->cur_depth < td->aio_depth) { |
43000118 JA |
1089 | timeout = &ts; |
1090 | min_evts = 0; | |
1091 | } else { | |
1092 | timeout = NULL; | |
1093 | min_evts = 1; | |
1094 | } | |
1095 | ||
1ad72b11 | 1096 | ret = io_u_getevents(td, min_evts, td->cur_depth, timeout); |
43000118 | 1097 | if (ret < 0) { |
406e7b7c | 1098 | td->error = ret; |
43000118 JA |
1099 | break; |
1100 | } else if (!ret) | |
1101 | continue; | |
1102 | ||
f0f3411b | 1103 | bytes_done = ios_completed(td, ret); |
43000118 | 1104 | |
98168d55 JA |
1105 | /* |
1106 | * the rate is batched for now, it should work for batches | |
1107 | * of completions except the very first one which may look | |
1108 | * a little bursty | |
1109 | */ | |
2c83567e | 1110 | gettimeofday(&e, NULL); |
43000118 JA |
1111 | usec = utime_since(&s, &e); |
1112 | ||
9e850933 | 1113 | rate_throttle(td, usec, bytes_done); |
43000118 JA |
1114 | |
1115 | if (check_min_rate(td, &e)) { | |
1116 | td->error = ENODATA; | |
1117 | break; | |
1118 | } | |
67903a2e JA |
1119 | |
1120 | if (runtime_exceeded(td, &e)) | |
1121 | break; | |
765d9223 JA |
1122 | |
1123 | if (td->thinktime) | |
d15c5195 | 1124 | usec_sleep(td, td->thinktime); |
cdf92433 JA |
1125 | |
1126 | if (should_fsync(td) && td->fsync_blocks && | |
1127 | (td->io_blocks % td->fsync_blocks) == 0) | |
1128 | fsync(td->fd); | |
43000118 | 1129 | } |
56b0eff0 | 1130 | |
2c83567e JA |
1131 | if (td->cur_depth) |
1132 | cleanup_pending_aio(td); | |
4ac89145 JA |
1133 | |
1134 | if (should_fsync(td)) | |
1135 | fsync(td->fd); | |
56b0eff0 JA |
1136 | } |
1137 | ||
1138 | static void cleanup_aio(struct thread_data *td) | |
1139 | { | |
254605cd JA |
1140 | io_destroy(td->aio_ctx); |
1141 | ||
43000118 JA |
1142 | if (td->aio_events) |
1143 | free(td->aio_events); | |
43000118 JA |
1144 | } |
1145 | ||
1146 | static int init_aio(struct thread_data *td) | |
1147 | { | |
254605cd | 1148 | if (io_queue_init(td->aio_depth, &td->aio_ctx)) { |
43000118 JA |
1149 | td->error = errno; |
1150 | return 1; | |
1151 | } | |
1152 | ||
43000118 | 1153 | td->aio_events = malloc(td->aio_depth * sizeof(struct io_event)); |
43000118 JA |
1154 | return 0; |
1155 | } | |
1156 | ||
2c83567e JA |
1157 | static void cleanup_io_u(struct thread_data *td) |
1158 | { | |
1159 | struct list_head *entry, *n; | |
1160 | struct io_u *io_u; | |
1161 | ||
1162 | list_for_each_safe(entry, n, &td->io_u_freelist) { | |
1163 | io_u = list_entry(entry, struct io_u, list); | |
1164 | ||
1165 | list_del(&io_u->list); | |
2c83567e JA |
1166 | free(io_u); |
1167 | } | |
6b71c826 | 1168 | |
99c6704f JA |
1169 | if (td->mem_type == MEM_MALLOC) |
1170 | free(td->orig_buffer); | |
1171 | else if (td->mem_type == MEM_SHM) { | |
1172 | struct shmid_ds sbuf; | |
1173 | ||
1174 | shmdt(td->orig_buffer); | |
1175 | shmctl(td->shm_id, IPC_RMID, &sbuf); | |
891e70f8 JA |
1176 | } else if (td->mem_type == MEM_MMAP) |
1177 | munmap(td->orig_buffer, td->orig_buffer_size); | |
1178 | else | |
1179 | fprintf(stderr, "Bad memory type %d\n", td->mem_type); | |
1180 | ||
1181 | td->orig_buffer = NULL; | |
2c83567e JA |
1182 | } |
1183 | ||
99c6704f | 1184 | static int init_io_u(struct thread_data *td) |
2c83567e JA |
1185 | { |
1186 | struct io_u *io_u; | |
891e70f8 | 1187 | int i, max_units; |
6b71c826 | 1188 | char *p; |
2c83567e JA |
1189 | |
1190 | if (!td->use_aio) | |
1191 | max_units = 1; | |
1192 | else | |
1193 | max_units = td->aio_depth; | |
1194 | ||
891e70f8 | 1195 | td->orig_buffer_size = td->max_bs * max_units + MASK; |
99c6704f JA |
1196 | |
1197 | if (td->mem_type == MEM_MALLOC) | |
891e70f8 | 1198 | td->orig_buffer = malloc(td->orig_buffer_size); |
99c6704f | 1199 | else if (td->mem_type == MEM_SHM) { |
891e70f8 | 1200 | td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600); |
99c6704f JA |
1201 | if (td->shm_id < 0) { |
1202 | td->error = errno; | |
1203 | perror("shmget"); | |
1204 | return 1; | |
1205 | } | |
1206 | ||
1207 | td->orig_buffer = shmat(td->shm_id, NULL, 0); | |
1208 | if (td->orig_buffer == (void *) -1) { | |
1209 | td->error = errno; | |
1210 | perror("shmat"); | |
891e70f8 JA |
1211 | td->orig_buffer = NULL; |
1212 | return 1; | |
1213 | } | |
1214 | } else if (td->mem_type == MEM_MMAP) { | |
1215 | td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, 0, 0); | |
1216 | if (td->orig_buffer == MAP_FAILED) { | |
1217 | td->error = errno; | |
1218 | perror("mmap"); | |
1219 | td->orig_buffer = NULL; | |
99c6704f JA |
1220 | return 1; |
1221 | } | |
1222 | } | |
6b71c826 | 1223 | |
2c83567e JA |
1224 | INIT_LIST_HEAD(&td->io_u_freelist); |
1225 | INIT_LIST_HEAD(&td->io_u_busylist); | |
645785e5 | 1226 | INIT_LIST_HEAD(&td->io_hist_list); |
2c83567e | 1227 | |
99c6704f | 1228 | p = ALIGN(td->orig_buffer); |
2c83567e JA |
1229 | for (i = 0; i < max_units; i++) { |
1230 | io_u = malloc(sizeof(*io_u)); | |
1231 | memset(io_u, 0, sizeof(*io_u)); | |
1232 | INIT_LIST_HEAD(&io_u->list); | |
1233 | ||
7889f07b | 1234 | io_u->buf = p + td->max_bs * i; |
2c83567e JA |
1235 | list_add(&io_u->list, &td->io_u_freelist); |
1236 | } | |
99c6704f JA |
1237 | |
1238 | return 0; | |
2c83567e JA |
1239 | } |
1240 | ||
ea6b64b1 JA |
1241 | static int create_file(struct thread_data *td, unsigned long long size, |
1242 | int extend) | |
02983297 | 1243 | { |
7889f07b | 1244 | unsigned long long left; |
645785e5 | 1245 | unsigned int bs; |
ea6b64b1 | 1246 | int r, oflags; |
02983297 JA |
1247 | char *b; |
1248 | ||
02983297 JA |
1249 | /* |
1250 | * unless specifically asked for overwrite, let normal io extend it | |
1251 | */ | |
62bb4285 | 1252 | if (td_write(td) && !td->overwrite) |
02983297 JA |
1253 | return 0; |
1254 | ||
ea6b64b1 | 1255 | if (!size) { |
57d753e3 JA |
1256 | fprintf(stderr, "Need size for create\n"); |
1257 | td->error = EINVAL; | |
1258 | return 1; | |
1259 | } | |
1260 | ||
ea6b64b1 JA |
1261 | if (!extend) { |
1262 | oflags = O_CREAT | O_TRUNC; | |
1263 | printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20); | |
1264 | } else { | |
1265 | oflags = O_APPEND; | |
1266 | printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20); | |
1267 | } | |
42fd89a7 | 1268 | |
ea6b64b1 | 1269 | td->fd = open(td->file_name, O_WRONLY | oflags, 0644); |
02983297 JA |
1270 | if (td->fd < 0) { |
1271 | td->error = errno; | |
1272 | return 1; | |
1273 | } | |
1274 | ||
ea6b64b1 | 1275 | if (!extend && ftruncate(td->fd, td->file_size) == -1) { |
c94deb1c JA |
1276 | td->error = errno; |
1277 | return 1; | |
1278 | } | |
1279 | ||
49d2caab | 1280 | td->io_size = td->file_size; |
7889f07b JA |
1281 | b = malloc(td->max_bs); |
1282 | memset(b, 0, td->max_bs); | |
1283 | ||
ea6b64b1 | 1284 | left = size; |
b31d826c | 1285 | while (left && !td->terminate) { |
7889f07b JA |
1286 | bs = td->max_bs; |
1287 | if (bs > left) | |
1288 | bs = left; | |
02983297 | 1289 | |
7889f07b | 1290 | r = write(td->fd, b, bs); |
02983297 | 1291 | |
645785e5 | 1292 | if (r == (int) bs) { |
7889f07b | 1293 | left -= bs; |
02983297 | 1294 | continue; |
7889f07b | 1295 | } else { |
02983297 JA |
1296 | if (r < 0) |
1297 | td->error = errno; | |
1298 | else | |
1299 | td->error = EIO; | |
1300 | ||
1301 | break; | |
1302 | } | |
1303 | } | |
1304 | ||
b31d826c JA |
1305 | if (td->terminate) |
1306 | unlink(td->file_name); | |
1307 | else if (td->create_fsync) | |
fc097bfe JA |
1308 | fsync(td->fd); |
1309 | ||
02983297 JA |
1310 | close(td->fd); |
1311 | td->fd = -1; | |
1312 | free(b); | |
1313 | return 0; | |
1314 | } | |
1315 | ||
c4c8f7b3 | 1316 | static int file_size(struct thread_data *td) |
02983297 JA |
1317 | { |
1318 | struct stat st; | |
c94deb1c JA |
1319 | |
1320 | if (fstat(td->fd, &st) == -1) { | |
1321 | td->error = errno; | |
1322 | return 1; | |
1323 | } | |
1324 | ||
c6ea46a7 JA |
1325 | if (!td->file_size) |
1326 | td->file_size = st.st_size; | |
c4c8f7b3 JA |
1327 | |
1328 | return 0; | |
1329 | } | |
1330 | ||
1331 | static int bdev_size(struct thread_data *td) | |
1332 | { | |
1333 | size_t bytes; | |
c94deb1c | 1334 | |
c4c8f7b3 JA |
1335 | if (ioctl(td->fd, BLKGETSIZE64, &bytes) < 0) { |
1336 | td->error = errno; | |
1337 | return 1; | |
c94deb1c JA |
1338 | } |
1339 | ||
c6ea46a7 JA |
1340 | /* |
1341 | * no extend possibilities, so limit size to device size if too large | |
1342 | */ | |
1343 | if (!td->file_size || td->file_size > bytes) | |
c4c8f7b3 JA |
1344 | td->file_size = bytes; |
1345 | ||
1346 | return 0; | |
1347 | } | |
1348 | ||
1349 | static int get_file_size(struct thread_data *td) | |
1350 | { | |
1351 | int ret; | |
1352 | ||
1353 | if (td->filetype == FIO_TYPE_FILE) | |
1354 | ret = file_size(td); | |
1355 | else | |
1356 | ret = bdev_size(td); | |
1357 | ||
1358 | if (ret) | |
1359 | return ret; | |
1360 | ||
1361 | if (td->file_offset > td->file_size) { | |
091d845a | 1362 | fprintf(stderr, "Client%d: offset larger than length (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->file_size); |
c94deb1c JA |
1363 | return 1; |
1364 | } | |
1365 | ||
c4c8f7b3 | 1366 | td->io_size = td->file_size - td->file_offset; |
c94deb1c JA |
1367 | if (td->io_size == 0) { |
1368 | fprintf(stderr, "Client%d: no io blocks\n", td->thread_number); | |
1369 | td->error = EINVAL; | |
1370 | return 1; | |
1371 | } | |
1372 | ||
1373 | return 0; | |
1374 | } | |
1375 | ||
6e2c38cc JA |
1376 | static int setup_file_mmap(struct thread_data *td) |
1377 | { | |
1378 | int flags; | |
1379 | ||
1380 | if (td_read(td)) | |
1381 | flags = PROT_READ; | |
1382 | else { | |
1383 | flags = PROT_WRITE; | |
1384 | ||
7f46ef08 | 1385 | if (td->verify != VERIFY_NONE) |
6e2c38cc JA |
1386 | flags |= PROT_READ; |
1387 | } | |
1388 | ||
1389 | td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset); | |
1390 | if (td->mmap == MAP_FAILED) { | |
1391 | td->mmap = NULL; | |
1392 | td->error = errno; | |
1393 | return 1; | |
1394 | } | |
1395 | ||
1396 | if (td->invalidate_cache) { | |
1397 | if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) { | |
1398 | td->error = errno; | |
1399 | return 1; | |
1400 | } | |
1401 | } | |
1402 | ||
1403 | if (td->sequential) { | |
1404 | if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) { | |
1405 | td->error = errno; | |
1406 | return 1; | |
1407 | } | |
1408 | } else { | |
1409 | if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) { | |
1410 | td->error = errno; | |
1411 | return 1; | |
1412 | } | |
1413 | } | |
1414 | ||
1415 | return 0; | |
1416 | } | |
1417 | ||
1418 | static int setup_file_plain(struct thread_data *td) | |
1419 | { | |
1420 | if (td->invalidate_cache) { | |
1421 | if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) { | |
1422 | td->error = errno; | |
1423 | return 1; | |
1424 | } | |
1425 | } | |
1426 | ||
1427 | if (td->sequential) { | |
1428 | if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) { | |
1429 | td->error = errno; | |
1430 | return 1; | |
1431 | } | |
1432 | } else { | |
1433 | if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) { | |
1434 | td->error = errno; | |
1435 | return 1; | |
1436 | } | |
1437 | } | |
1438 | ||
1439 | return 0; | |
1440 | } | |
1441 | ||
c94deb1c JA |
1442 | static int setup_file(struct thread_data *td) |
1443 | { | |
ea6b64b1 | 1444 | struct stat st; |
02983297 JA |
1445 | int flags = 0; |
1446 | ||
ea6b64b1 JA |
1447 | if (stat(td->file_name, &st) == -1) { |
1448 | if (errno != ENOENT) { | |
1449 | td->error = errno; | |
1450 | return 1; | |
1451 | } | |
02983297 JA |
1452 | if (!td->create_file) { |
1453 | td->error = ENOENT; | |
1454 | return 1; | |
1455 | } | |
ea6b64b1 JA |
1456 | if (create_file(td, td->file_size, 0)) |
1457 | return 1; | |
c6ea46a7 JA |
1458 | } else if (td->filetype == FIO_TYPE_FILE) { |
1459 | if (st.st_size < td->file_size) { | |
1460 | if (create_file(td, td->file_size - st.st_size, 1)) | |
1461 | return 1; | |
1462 | } | |
02983297 JA |
1463 | } |
1464 | ||
1465 | if (td->odirect) | |
1466 | flags |= O_DIRECT; | |
1467 | ||
1468 | if (td_read(td)) | |
1469 | td->fd = open(td->file_name, flags | O_RDONLY); | |
1470 | else { | |
c6ea46a7 JA |
1471 | if (td->filetype == FIO_TYPE_FILE) { |
1472 | if (!td->overwrite) | |
1473 | flags |= O_TRUNC; | |
1474 | ||
1475 | flags |= O_CREAT; | |
1476 | } | |
74b4b5fb JA |
1477 | if (td->sync_io) |
1478 | flags |= O_SYNC; | |
6e2c38cc JA |
1479 | |
1480 | flags |= O_RDWR; | |
02983297 | 1481 | |
c6ea46a7 | 1482 | td->fd = open(td->file_name, flags, 0600); |
02983297 JA |
1483 | } |
1484 | ||
1485 | if (td->fd == -1) { | |
1486 | td->error = errno; | |
1487 | return 1; | |
1488 | } | |
1489 | ||
c94deb1c | 1490 | if (get_file_size(td)) |
49d2caab | 1491 | return 1; |
49d2caab | 1492 | |
6e2c38cc JA |
1493 | if (!td->use_mmap) |
1494 | return setup_file_plain(td); | |
1495 | else | |
1496 | return setup_file_mmap(td); | |
02983297 JA |
1497 | } |
1498 | ||
debf703a JA |
1499 | static int check_dev_match(dev_t dev, char *path) |
1500 | { | |
1501 | int major, minor; | |
1502 | char line[256], *p; | |
1503 | FILE *f; | |
1504 | ||
1505 | f = fopen(path, "r"); | |
1506 | if (!f) { | |
1507 | perror("open path"); | |
1508 | return 1; | |
1509 | } | |
1510 | ||
1511 | p = fgets(line, sizeof(line), f); | |
1512 | if (!p) { | |
1513 | fclose(f); | |
1514 | return 1; | |
1515 | } | |
1516 | ||
1517 | if (sscanf(p, "%u:%u", &major, &minor) != 2) { | |
1518 | fclose(f); | |
1519 | return 1; | |
1520 | } | |
1521 | ||
1522 | if (((major << 8) | minor) == dev) { | |
1523 | fclose(f); | |
1524 | return 0; | |
1525 | } | |
1526 | ||
1527 | fclose(f); | |
1528 | return 1; | |
1529 | } | |
1530 | ||
1531 | static char *find_block_dir(dev_t dev, char *path) | |
1532 | { | |
1533 | struct dirent *dir; | |
1534 | char *found = NULL; | |
1535 | struct stat st; | |
1536 | DIR *D; | |
1537 | ||
1538 | D = opendir(path); | |
1539 | if (!D) | |
1540 | return NULL; | |
1541 | ||
1542 | while ((dir = readdir(D)) != NULL) { | |
1543 | char full_path[256]; | |
1544 | ||
1545 | if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, "..")) | |
1546 | continue; | |
1547 | if (!strcmp(dir->d_name, "device")) | |
1548 | continue; | |
1549 | ||
1550 | sprintf(full_path, "%s/%s", path, dir->d_name); | |
1551 | ||
1552 | if (!strcmp(dir->d_name, "dev")) { | |
1553 | if (!check_dev_match(dev, full_path)) { | |
1554 | found = path; | |
1555 | break; | |
1556 | } | |
1557 | } | |
1558 | ||
1559 | if (stat(full_path, &st) == -1) { | |
1560 | perror("stat"); | |
1561 | break; | |
1562 | } | |
1563 | ||
1564 | if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode)) | |
1565 | continue; | |
1566 | ||
1567 | if ((found = find_block_dir(dev, full_path)) != NULL) | |
1568 | break; | |
1569 | } | |
1570 | ||
1571 | closedir(D); | |
1572 | return found; | |
1573 | } | |
1574 | ||
6f75d67c | 1575 | static int get_io_ticks(struct disk_util *du) |
debf703a JA |
1576 | { |
1577 | int i1, i2, i3, i4, i5, i6, i7, i8, i9; | |
1578 | unsigned long long ull1, ull2; | |
1579 | char line[256]; | |
1580 | FILE *f; | |
1581 | char *p; | |
1582 | ||
6f75d67c | 1583 | f = fopen(du->path, "r"); |
debf703a JA |
1584 | if (!f) |
1585 | return 0; | |
1586 | ||
1587 | p = fgets(line, sizeof(line), f); | |
1588 | if (!p) { | |
1589 | fclose(f); | |
1590 | return 0; | |
1591 | } | |
1592 | ||
1593 | if (sscanf(p, "%8u %8u %8llu %8u %8u %8u %8llu %8u %8u %8u %8u", &i1, &i2, &ull1, &i3, &i4, &i5, &ull2, &i6, &i7, &i8, &i9) != 11) { | |
1594 | fclose(f); | |
1595 | return 0; | |
1596 | } | |
1597 | ||
1598 | fclose(f); | |
1599 | return i8; | |
1600 | } | |
1601 | ||
6f75d67c | 1602 | static void update_io_tick_disk(struct disk_util *du) |
debf703a | 1603 | { |
6f75d67c JA |
1604 | unsigned long ticks = get_io_ticks(du); |
1605 | struct timeval t; | |
1606 | ||
1607 | du->io_ticks += (ticks - du->last_io_ticks); | |
1608 | du->last_io_ticks = ticks; | |
1609 | gettimeofday(&t, NULL); | |
1610 | du->msec += mtime_since(&du->time, &t); | |
1611 | memcpy(&du->time, &t, sizeof(t)); | |
1612 | } | |
1613 | ||
1614 | static void update_io_ticks(void) | |
1615 | { | |
1616 | struct list_head *entry; | |
1617 | struct disk_util *du; | |
1618 | ||
1619 | list_for_each(entry, &disk_list) { | |
1620 | du = list_entry(entry, struct disk_util, list); | |
1621 | update_io_tick_disk(du); | |
1622 | } | |
1623 | } | |
1624 | ||
1625 | static int disk_util_exists(dev_t dev) | |
1626 | { | |
1627 | struct list_head *entry; | |
1628 | struct disk_util *du; | |
1629 | ||
1630 | list_for_each(entry, &disk_list) { | |
1631 | du = list_entry(entry, struct disk_util, list); | |
1632 | ||
1633 | if (du->dev == dev) | |
1634 | return 1; | |
1635 | } | |
debf703a | 1636 | |
6f75d67c | 1637 | return 0; |
debf703a JA |
1638 | } |
1639 | ||
6f75d67c JA |
1640 | static void disk_util_add(dev_t dev, char *path) |
1641 | { | |
1642 | struct disk_util *du = malloc(sizeof(*du)); | |
1643 | ||
1644 | memset(du, 0, sizeof(*du)); | |
1645 | INIT_LIST_HEAD(&du->list); | |
1646 | sprintf(du->path, "%s/stat", path); | |
1647 | du->name = basename(path); | |
1648 | du->dev = dev; | |
1649 | ||
1650 | gettimeofday(&du->time, NULL); | |
1651 | du->last_io_ticks = get_io_ticks(du); | |
1652 | ||
1653 | list_add_tail(&du->list, &disk_list); | |
1654 | } | |
1655 | ||
1656 | static void init_disk_util(struct thread_data *td) | |
debf703a JA |
1657 | { |
1658 | struct stat st; | |
1659 | char foo[256], tmp[256]; | |
1660 | dev_t dev; | |
1661 | char *p, *dir; | |
1662 | ||
6f75d67c JA |
1663 | if (!stat(td->file_name, &st)) { |
1664 | if (S_ISBLK(st.st_mode)) | |
1665 | dev = st.st_rdev; | |
1666 | else | |
1667 | dev = st.st_dev; | |
1668 | } else { | |
1669 | /* | |
1670 | * must be a file, open "." in that path | |
1671 | */ | |
8cf48a05 JA |
1672 | strcpy(foo, td->file_name); |
1673 | p = dirname(foo); | |
6f75d67c JA |
1674 | if (stat(p, &st)) { |
1675 | perror("disk util stat"); | |
1676 | return; | |
1677 | } | |
debf703a | 1678 | |
debf703a | 1679 | dev = st.st_dev; |
6f75d67c JA |
1680 | } |
1681 | ||
1682 | if (disk_util_exists(dev)) | |
1683 | return; | |
debf703a JA |
1684 | |
1685 | sprintf(foo, "/sys/block"); | |
1686 | dir = find_block_dir(dev, foo); | |
1687 | if (!dir) | |
6f75d67c | 1688 | return; |
debf703a JA |
1689 | |
1690 | /* | |
1691 | * if this is inside a partition dir, jump back to parent | |
1692 | */ | |
1693 | sprintf(tmp, "%s/queue", dir); | |
1694 | if (stat(tmp, &st)) { | |
1695 | p = dirname(dir); | |
1696 | sprintf(tmp, "%s/queue", p); | |
1697 | if (stat(tmp, &st)) { | |
1698 | fprintf(stderr, "unknown sysfs layout\n"); | |
6f75d67c | 1699 | return; |
debf703a | 1700 | } |
6f75d67c | 1701 | sprintf(foo, "%s", p); |
debf703a JA |
1702 | } |
1703 | ||
6f75d67c JA |
1704 | disk_util_add(dev, foo); |
1705 | } | |
1706 | ||
1707 | static void disk_util_timer_arm(void) | |
1708 | { | |
1709 | struct itimerval itimer; | |
1710 | ||
1711 | itimer.it_value.tv_sec = 0; | |
1712 | itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000; | |
1713 | setitimer(ITIMER_REAL, &itimer, NULL); | |
1714 | } | |
1715 | ||
d32d9284 JA |
1716 | static void clear_io_state(struct thread_data *td) |
1717 | { | |
9d0c6ca2 JA |
1718 | if (!td->use_aio) |
1719 | lseek(td->fd, SEEK_SET, 0); | |
1720 | ||
d32d9284 | 1721 | td->cur_off = 0; |
49d2caab JA |
1722 | td->last_bytes = 0; |
1723 | td->stat_io_bytes = 0; | |
1724 | td->this_io_bytes = 0; | |
1725 | ||
1726 | if (td->file_map) | |
1727 | memset(td->file_map, 0, td->num_maps * sizeof(long)); | |
d32d9284 JA |
1728 | } |
1729 | ||
f6dcd824 JA |
1730 | static void update_rusage_stat(struct thread_data *td) |
1731 | { | |
1732 | if (!td->runtime) | |
1733 | return; | |
1734 | ||
1735 | getrusage(RUSAGE_SELF, &td->ru_end); | |
1736 | ||
1737 | td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime); | |
1738 | td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime); | |
1739 | td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw); | |
1740 | ||
1741 | ||
1742 | memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end)); | |
1743 | } | |
1744 | ||
189873de | 1745 | static void *thread_main(void *data) |
892199bd | 1746 | { |
189873de | 1747 | struct thread_data *td = data; |
02983297 | 1748 | int ret = 1; |
892199bd | 1749 | |
7292613b | 1750 | setsid(); |
892199bd JA |
1751 | td->pid = getpid(); |
1752 | ||
99c6704f JA |
1753 | if (init_io_u(td)) |
1754 | goto err; | |
2c83567e | 1755 | |
18e0b78c JA |
1756 | if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) { |
1757 | td->error = errno; | |
1758 | goto err; | |
1759 | } | |
1760 | ||
43000118 JA |
1761 | if (td->use_aio && init_aio(td)) |
1762 | goto err; | |
1763 | ||
f737299d | 1764 | if (td->ioprio) { |
892199bd JA |
1765 | if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) { |
1766 | td->error = errno; | |
599002b3 | 1767 | goto err; |
892199bd JA |
1768 | } |
1769 | } | |
1770 | ||
1771 | sem_post(&startup_sem); | |
1772 | sem_wait(&td->mutex); | |
43000118 | 1773 | |
fc097bfe JA |
1774 | if (!td->create_serialize && setup_file(td)) |
1775 | goto err; | |
1776 | ||
49d2caab JA |
1777 | if (init_random_state(td)) |
1778 | goto err; | |
1779 | ||
293753bb | 1780 | while (td->loops--) { |
f6dcd824 JA |
1781 | getrusage(RUSAGE_SELF, &td->ru_start); |
1782 | gettimeofday(&td->start, NULL); | |
1783 | memcpy(&td->stat_sample_time, &td->start, sizeof(td->start)); | |
293753bb JA |
1784 | |
1785 | if (td->ratemin) | |
1786 | memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate)); | |
7292613b | 1787 | |
d32d9284 | 1788 | clear_io_state(td); |
9d0c6ca2 | 1789 | prune_io_piece_log(td); |
fd1ae4c9 | 1790 | |
b2de0ed2 | 1791 | if (!td->use_aio) |
b6794fbf | 1792 | do_sync_io(td); |
b2de0ed2 JA |
1793 | else |
1794 | do_async_io(td); | |
1795 | ||
91fc5dc9 JA |
1796 | if (td->error) |
1797 | break; | |
1798 | ||
f6dcd824 JA |
1799 | td->runtime += mtime_since_now(&td->start); |
1800 | update_rusage_stat(td); | |
1801 | ||
7f46ef08 | 1802 | if (td->verify == VERIFY_NONE) |
b2de0ed2 | 1803 | continue; |
cfc702bd | 1804 | |
b2de0ed2 | 1805 | clear_io_state(td); |
d32d9284 | 1806 | |
91fc5dc9 JA |
1807 | if (!td->use_aio) |
1808 | do_sync_verify(td); | |
1809 | else | |
1810 | do_async_verify(td); | |
1811 | ||
1812 | if (td->error) | |
1813 | break; | |
b6794fbf | 1814 | } |
7292613b | 1815 | |
892199bd | 1816 | ret = 0; |
a0a9b35b JA |
1817 | |
1818 | if (td->bw_log) | |
1819 | finish_log(td, td->bw_log, "bw"); | |
1820 | if (td->lat_log) | |
1821 | finish_log(td, td->lat_log, "lat"); | |
4ac89145 | 1822 | |
98dd52d6 | 1823 | if (exitall_on_terminate) |
27c32a38 | 1824 | terminate_threads(td->groupid); |
98dd52d6 | 1825 | |
892199bd | 1826 | err: |
7292613b JA |
1827 | if (td->fd != -1) { |
1828 | close(td->fd); | |
1829 | td->fd = -1; | |
1830 | } | |
6e2c38cc JA |
1831 | if (td->mmap) |
1832 | munmap(td->mmap, td->file_size); | |
4ac89145 JA |
1833 | if (td->use_aio) |
1834 | cleanup_aio(td); | |
2c83567e | 1835 | cleanup_io_u(td); |
599002b3 | 1836 | if (ret) { |
892199bd | 1837 | sem_post(&startup_sem); |
599002b3 JA |
1838 | sem_wait(&td->mutex); |
1839 | } | |
40ef7f64 | 1840 | td_set_runstate(td, TD_EXITED); |
189873de JA |
1841 | return NULL; |
1842 | ||
1843 | } | |
1844 | ||
1845 | static void *fork_main(int shm_id, int offset) | |
1846 | { | |
1847 | struct thread_data *td; | |
1848 | void *data; | |
1849 | ||
1850 | data = shmat(shm_id, NULL, 0); | |
1851 | if (data == (void *) -1) { | |
1852 | perror("shmat"); | |
1853 | return NULL; | |
1854 | } | |
1855 | ||
1856 | td = data + offset * sizeof(struct thread_data); | |
1857 | thread_main(td); | |
4240cfa1 | 1858 | shmdt(data); |
892199bd JA |
1859 | return NULL; |
1860 | } | |
1861 | ||
57d753e3 JA |
1862 | static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, |
1863 | double *mean, double *dev) | |
1864 | { | |
1865 | double n; | |
1866 | ||
1867 | if (is->samples == 0) | |
1868 | return 0; | |
1869 | ||
1870 | *min = is->min_val; | |
1871 | *max = is->max_val; | |
1872 | ||
1873 | n = (double) is->samples; | |
1874 | *mean = (double) is->val / n; | |
1875 | *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1)); | |
1876 | return 1; | |
1877 | } | |
1878 | ||
557e4102 JA |
1879 | static void show_thread_status(struct thread_data *td, |
1880 | struct group_run_stats *rs) | |
892199bd JA |
1881 | { |
1882 | int prio, prio_class; | |
f6dcd824 | 1883 | unsigned long min, max, bw = 0; |
92b229ed | 1884 | double mean, dev, usr_cpu, sys_cpu; |
892199bd | 1885 | |
49d2caab | 1886 | if (!td->io_bytes && !td->error) |
213b446c JA |
1887 | return; |
1888 | ||
892199bd | 1889 | if (td->runtime) |
49d2caab | 1890 | bw = td->io_bytes / td->runtime; |
892199bd JA |
1891 | |
1892 | prio = td->ioprio & 0xff; | |
1893 | prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT; | |
1894 | ||
f6dcd824 | 1895 | printf("Client%d (g=%d): err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->groupid, td->error, td->io_bytes >> 20, bw, td->runtime); |
fd1ae4c9 | 1896 | |
57d753e3 JA |
1897 | if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev)) |
1898 | printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev); | |
1899 | if (calc_lat(&td->clat_stat, &min, &max, &mean, &dev)) | |
1900 | printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev); | |
557e4102 JA |
1901 | if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev)) { |
1902 | double p_of_agg; | |
1903 | ||
1904 | p_of_agg = mean * 100 / (double) rs->agg[td->ddir]; | |
1905 | printf(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev); | |
1906 | } | |
92b229ed JA |
1907 | |
1908 | if (td->runtime) { | |
f6dcd824 JA |
1909 | usr_cpu = (double) td->usr_time * 100 / (double) td->runtime; |
1910 | sys_cpu = (double) td->sys_time * 100 / (double) td->runtime; | |
92b229ed JA |
1911 | } else { |
1912 | usr_cpu = 0; | |
1913 | sys_cpu = 0; | |
1914 | } | |
1915 | ||
f6dcd824 | 1916 | printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx); |
892199bd JA |
1917 | } |
1918 | ||
3f39453a | 1919 | static void print_thread_status(struct thread_data *td, int nr_running, |
8dbff0b1 | 1920 | int t_rate, int m_rate) |
3f39453a | 1921 | { |
3f39453a JA |
1922 | printf("Threads now running: %d", nr_running); |
1923 | if (m_rate || t_rate) | |
1924 | printf(", commitrate %d/%dKiB/sec", t_rate, m_rate); | |
8dbff0b1 JA |
1925 | printf(" : [%s]\r", run_str); |
1926 | fflush(stdout); | |
3f39453a JA |
1927 | } |
1928 | ||
40ef7f64 JA |
1929 | static void check_str_update(struct thread_data *td, int n, int t, int m) |
1930 | { | |
1931 | char c = run_str[td->thread_number - 1]; | |
1932 | ||
1933 | if (td->runstate == td->old_runstate) | |
1934 | return; | |
1935 | ||
1936 | switch (td->runstate) { | |
1937 | case TD_REAPED: | |
1938 | c = '_'; | |
1939 | break; | |
f4bb2243 JA |
1940 | case TD_EXITED: |
1941 | c = 'E'; | |
1942 | break; | |
40ef7f64 | 1943 | case TD_RUNNING: |
af678352 JA |
1944 | if (td_read(td)) { |
1945 | if (td->sequential) | |
1946 | c = 'R'; | |
1947 | else | |
1948 | c = 'r'; | |
1949 | } else { | |
1950 | if (td->sequential) | |
1951 | c = 'W'; | |
1952 | else | |
1953 | c = 'w'; | |
1954 | } | |
40ef7f64 JA |
1955 | break; |
1956 | case TD_VERIFYING: | |
1957 | c = 'V'; | |
1958 | break; | |
1959 | case TD_CREATED: | |
1960 | c = 'C'; | |
1961 | break; | |
1962 | case TD_NOT_CREATED: | |
1963 | c = 'P'; | |
1964 | break; | |
1965 | default: | |
1966 | printf("state %d\n", td->runstate); | |
1967 | } | |
1968 | ||
1969 | run_str[td->thread_number - 1] = c; | |
1970 | print_thread_status(td, n, t, m); | |
1971 | td->old_runstate = td->runstate; | |
1972 | } | |
1973 | ||
213b446c | 1974 | static void reap_threads(int *nr_running, int *t_rate, int *m_rate) |
02bdd9ba | 1975 | { |
213b446c | 1976 | int i; |
02bdd9ba | 1977 | |
3f39453a JA |
1978 | /* |
1979 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
1980 | */ | |
02bdd9ba JA |
1981 | for (i = 0; i < thread_number; i++) { |
1982 | struct thread_data *td = &threads[i]; | |
1983 | ||
40ef7f64 JA |
1984 | check_str_update(td, *nr_running, *t_rate, *m_rate); |
1985 | ||
213b446c JA |
1986 | if (td->runstate != TD_EXITED) |
1987 | continue; | |
02bdd9ba | 1988 | |
40ef7f64 | 1989 | td_set_runstate(td, TD_REAPED); |
189873de JA |
1990 | |
1991 | if (td->use_thread) { | |
1992 | long ret; | |
1993 | ||
1994 | if (pthread_join(td->thread, (void *) &ret)) | |
1995 | perror("thread_join"); | |
1996 | } else | |
1997 | waitpid(td->pid, NULL, 0); | |
1998 | ||
213b446c JA |
1999 | (*nr_running)--; |
2000 | (*m_rate) -= td->ratemin; | |
2001 | (*t_rate) -= td->rate; | |
40ef7f64 | 2002 | check_str_update(td, *nr_running, *t_rate, *m_rate); |
213b446c | 2003 | } |
02bdd9ba JA |
2004 | } |
2005 | ||
fc24389f JA |
2006 | static void run_threads(char *argv[]) |
2007 | { | |
be33abe4 | 2008 | struct timeval genesis; |
fc24389f JA |
2009 | struct thread_data *td; |
2010 | unsigned long spent; | |
2a81240d | 2011 | int i, todo, nr_running, m_rate, t_rate, nr_started; |
fc24389f | 2012 | |
fc24389f JA |
2013 | printf("Starting %d threads\n", thread_number); |
2014 | fflush(stdout); | |
2015 | ||
7292613b | 2016 | signal(SIGINT, sig_handler); |
8882f4e9 | 2017 | signal(SIGALRM, sig_handler); |
7292613b | 2018 | |
fc24389f | 2019 | todo = thread_number; |
02bdd9ba | 2020 | nr_running = 0; |
2a81240d | 2021 | nr_started = 0; |
213b446c | 2022 | m_rate = t_rate = 0; |
fc24389f | 2023 | |
8bdcfab5 JA |
2024 | for (i = 0; i < thread_number; i++) { |
2025 | td = &threads[i]; | |
2026 | ||
6f75d67c JA |
2027 | init_disk_util(td); |
2028 | ||
fc097bfe JA |
2029 | if (!td->create_serialize) |
2030 | continue; | |
2031 | ||
8bdcfab5 JA |
2032 | /* |
2033 | * do file setup here so it happens sequentially, | |
2034 | * we don't want X number of threads getting their | |
2035 | * client data interspersed on disk | |
2036 | */ | |
2037 | if (setup_file(td)) { | |
40ef7f64 | 2038 | td_set_runstate(td, TD_REAPED); |
8bdcfab5 JA |
2039 | todo--; |
2040 | } | |
2041 | } | |
2042 | ||
2043 | gettimeofday(&genesis, NULL); | |
2044 | ||
213b446c | 2045 | while (todo) { |
3f39453a JA |
2046 | /* |
2047 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
2048 | */ | |
fc24389f JA |
2049 | for (i = 0; i < thread_number; i++) { |
2050 | td = &threads[i]; | |
2051 | ||
02bdd9ba | 2052 | if (td->runstate != TD_NOT_CREATED) |
fc24389f JA |
2053 | continue; |
2054 | ||
213b446c JA |
2055 | /* |
2056 | * never got a chance to start, killed by other | |
2057 | * thread for some reason | |
2058 | */ | |
2059 | if (td->terminate) { | |
2060 | todo--; | |
2061 | continue; | |
2062 | } | |
2063 | ||
fc24389f | 2064 | if (td->start_delay) { |
be33abe4 | 2065 | spent = mtime_since_now(&genesis); |
fc24389f JA |
2066 | |
2067 | if (td->start_delay * 1000 > spent) | |
2068 | continue; | |
2069 | } | |
2070 | ||
2a81240d | 2071 | if (td->stonewall && (nr_started || nr_running)) |
ea6f96a2 | 2072 | break; |
2a81240d | 2073 | |
40ef7f64 JA |
2074 | td_set_runstate(td, TD_CREATED); |
2075 | check_str_update(td, nr_running, t_rate, m_rate); | |
fc24389f JA |
2076 | sem_init(&startup_sem, 1, 1); |
2077 | todo--; | |
2a81240d | 2078 | nr_started++; |
fc24389f | 2079 | |
189873de JA |
2080 | if (td->use_thread) { |
2081 | if (pthread_create(&td->thread, NULL, thread_main, td)) { | |
2082 | perror("thread_create"); | |
2083 | nr_started--; | |
2084 | } | |
2085 | } else { | |
2086 | if (fork()) | |
2087 | sem_wait(&startup_sem); | |
2088 | else { | |
2089 | fork_main(shm_id, i); | |
2090 | exit(0); | |
2091 | } | |
fc24389f JA |
2092 | } |
2093 | } | |
2094 | ||
3f39453a | 2095 | /* |
e8457004 | 2096 | * start created threads (TD_CREATED -> TD_RUNNING) |
3f39453a | 2097 | */ |
fc24389f JA |
2098 | for (i = 0; i < thread_number; i++) { |
2099 | struct thread_data *td = &threads[i]; | |
2100 | ||
3f39453a JA |
2101 | if (td->runstate != TD_CREATED) |
2102 | continue; | |
2103 | ||
40ef7f64 | 2104 | td_set_runstate(td, TD_RUNNING); |
3f39453a | 2105 | nr_running++; |
2a81240d | 2106 | nr_started--; |
3f39453a JA |
2107 | m_rate += td->ratemin; |
2108 | t_rate += td->rate; | |
40ef7f64 | 2109 | check_str_update(td, nr_running, t_rate, m_rate); |
3f39453a | 2110 | sem_post(&td->mutex); |
fc24389f JA |
2111 | } |
2112 | ||
e8457004 JA |
2113 | for (i = 0; i < thread_number; i++) { |
2114 | struct thread_data *td = &threads[i]; | |
2115 | ||
b48889bb JA |
2116 | if (td->runstate != TD_RUNNING && |
2117 | td->runstate != TD_VERIFYING) | |
e8457004 JA |
2118 | continue; |
2119 | ||
40ef7f64 | 2120 | check_str_update(td, nr_running, t_rate, m_rate); |
e8457004 JA |
2121 | } |
2122 | ||
213b446c | 2123 | reap_threads(&nr_running, &t_rate, &m_rate); |
02bdd9ba | 2124 | |
fc24389f JA |
2125 | if (todo) |
2126 | usleep(100000); | |
2127 | } | |
02bdd9ba JA |
2128 | |
2129 | while (nr_running) { | |
213b446c | 2130 | reap_threads(&nr_running, &t_rate, &m_rate); |
02bdd9ba JA |
2131 | usleep(10000); |
2132 | } | |
6f75d67c JA |
2133 | |
2134 | update_io_ticks(); | |
fc24389f JA |
2135 | } |
2136 | ||
0d80f40d | 2137 | static void show_group_stats(struct group_run_stats *rs, int id) |
8867c0a8 | 2138 | { |
0d80f40d JA |
2139 | printf("\nRun status group %d:\n", id); |
2140 | ||
2141 | if (rs->max_run[DDIR_READ]) | |
2142 | printf(" READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]); | |
2143 | if (rs->max_run[DDIR_WRITE]) | |
2144 | printf(" WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]); | |
2145 | } | |
2146 | ||
6f75d67c JA |
2147 | static void show_disk_util(void) |
2148 | { | |
2149 | struct list_head *entry; | |
2150 | struct disk_util *du; | |
2151 | double util; | |
2152 | ||
2153 | printf("\nDisk utilization:\n"); | |
2154 | ||
2155 | list_for_each(entry, &disk_list) { | |
2156 | du = list_entry(entry, struct disk_util, list); | |
2157 | ||
2158 | util = (double) 100 * du->io_ticks / (double) du->msec; | |
2159 | if (util > 100.0) | |
2160 | util = 100.0; | |
2161 | ||
2162 | printf(" %s: %3.2f%%\n", du->name, util); | |
2163 | } | |
2164 | } | |
2165 | ||
0d80f40d JA |
2166 | static void show_run_stats(void) |
2167 | { | |
2168 | struct group_run_stats *runstats, *rs; | |
557e4102 | 2169 | struct thread_data *td; |
8867c0a8 JA |
2170 | int i; |
2171 | ||
0d80f40d JA |
2172 | runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); |
2173 | ||
2174 | for (i = 0; i < groupid + 1; i++) { | |
2175 | rs = &runstats[i]; | |
2176 | ||
f6dcd824 | 2177 | memset(rs, 0, sizeof(*rs)); |
0d80f40d JA |
2178 | rs->min_bw[0] = rs->min_run[0] = ~0UL; |
2179 | rs->min_bw[1] = rs->min_run[1] = ~0UL; | |
0d80f40d JA |
2180 | } |
2181 | ||
2182 | for (i = 0; i < thread_number; i++) { | |
0d80f40d JA |
2183 | unsigned long bw = 0; |
2184 | ||
557e4102 JA |
2185 | td = &threads[i]; |
2186 | ||
2187 | if (td->error) | |
2188 | continue; | |
2189 | ||
0d80f40d JA |
2190 | rs = &runstats[td->groupid]; |
2191 | ||
557e4102 JA |
2192 | if (td->runtime < rs->min_run[td->ddir]) |
2193 | rs->min_run[td->ddir] = td->runtime; | |
2194 | if (td->runtime > rs->max_run[td->ddir]) | |
2195 | rs->max_run[td->ddir] = td->runtime; | |
0d80f40d | 2196 | |
557e4102 JA |
2197 | if (td->runtime) |
2198 | bw = td->io_bytes / td->runtime; | |
2199 | if (bw < rs->min_bw[td->ddir]) | |
2200 | rs->min_bw[td->ddir] = bw; | |
2201 | if (bw > rs->max_bw[td->ddir]) | |
2202 | rs->max_bw[td->ddir] = bw; | |
0d80f40d | 2203 | |
557e4102 | 2204 | rs->io_mb[td->ddir] += td->io_bytes >> 20; |
0d80f40d | 2205 | } |
9d489c62 | 2206 | |
0d80f40d JA |
2207 | for (i = 0; i < groupid + 1; i++) { |
2208 | rs = &runstats[i]; | |
2209 | ||
2210 | if (rs->max_run[0]) | |
2211 | rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0]; | |
2212 | if (rs->max_run[1]) | |
2213 | rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1]; | |
0d80f40d | 2214 | } |
557e4102 JA |
2215 | |
2216 | for (i = 0; i < thread_number; i++) { | |
2217 | td = &threads[i]; | |
2218 | rs = &runstats[td->groupid]; | |
2219 | ||
c4c8f7b3 | 2220 | show_thread_status(td, rs); |
557e4102 | 2221 | } |
9d489c62 JA |
2222 | |
2223 | for (i = 0; i < groupid + 1; i++) | |
2224 | show_group_stats(&runstats[i], i); | |
6f75d67c JA |
2225 | |
2226 | show_disk_util(); | |
0d80f40d JA |
2227 | } |
2228 | ||
2229 | int main(int argc, char *argv[]) | |
2230 | { | |
27c32a38 | 2231 | memset(run_str, 0, sizeof(run_str)); |
5961d92c | 2232 | |
27c32a38 | 2233 | if (parse_options(argc, argv)) |
5961d92c | 2234 | return 1; |
7dd1389e | 2235 | |
4240cfa1 JA |
2236 | if (!thread_number) { |
2237 | printf("Nothing to do\n"); | |
2238 | return 1; | |
2239 | } | |
7dd1389e | 2240 | |
8882f4e9 | 2241 | disk_util_timer_arm(); |
6f75d67c | 2242 | |
fc24389f | 2243 | run_threads(argv); |
0d80f40d | 2244 | show_run_stats(); |
fc24389f | 2245 | |
892199bd JA |
2246 | return 0; |
2247 | } |