Commit | Line | Data |
---|---|---|
ebac4655 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 | */ | |
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> | |
29 | #include <math.h> | |
30 | #include <assert.h> | |
31 | #include <dirent.h> | |
32 | #include <libgen.h> | |
33 | #include <sys/types.h> | |
34 | #include <sys/stat.h> | |
35 | #include <sys/wait.h> | |
36 | #include <sys/ipc.h> | |
37 | #include <sys/shm.h> | |
38 | #include <sys/ioctl.h> | |
39 | #include <sys/mman.h> | |
40 | ||
41 | #include "fio.h" | |
42 | #include "os.h" | |
43 | ||
44 | #define MASK (4095) | |
45 | ||
46 | #define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK)) | |
47 | ||
48 | int groupid = 0; | |
49 | int thread_number = 0; | |
50 | static char run_str[MAX_JOBS + 1]; | |
51 | int shm_id = 0; | |
52 | static LIST_HEAD(disk_list); | |
53 | static struct itimerval itimer; | |
5289b847 | 54 | static struct timeval genesis; |
ebac4655 JA |
55 | |
56 | static void update_io_ticks(void); | |
57 | static void disk_util_timer_arm(void); | |
58 | static void print_thread_status(void); | |
59 | ||
60 | /* | |
61 | * thread life cycle | |
62 | */ | |
63 | enum { | |
64 | TD_NOT_CREATED = 0, | |
65 | TD_CREATED, | |
66 | TD_RUNNING, | |
67 | TD_VERIFYING, | |
68 | TD_EXITED, | |
69 | TD_REAPED, | |
70 | }; | |
71 | ||
3d60d1ed | 72 | #define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync)) |
ebac4655 JA |
73 | |
74 | static sem_t startup_sem; | |
75 | ||
76 | #define TERMINATE_ALL (-1) | |
77 | ||
78 | static void terminate_threads(int group_id) | |
79 | { | |
80 | int i; | |
81 | ||
82 | for (i = 0; i < thread_number; i++) { | |
83 | struct thread_data *td = &threads[i]; | |
84 | ||
85 | if (group_id == TERMINATE_ALL || groupid == td->groupid) { | |
86 | td->terminate = 1; | |
87 | td->start_delay = 0; | |
88 | } | |
89 | } | |
90 | } | |
91 | ||
92 | static void sig_handler(int sig) | |
93 | { | |
94 | switch (sig) { | |
95 | case SIGALRM: | |
96 | update_io_ticks(); | |
97 | disk_util_timer_arm(); | |
98 | print_thread_status(); | |
99 | break; | |
100 | default: | |
101 | printf("\nfio: terminating on signal\n"); | |
102 | fflush(stdout); | |
103 | terminate_threads(TERMINATE_ALL); | |
104 | break; | |
105 | } | |
106 | } | |
107 | ||
108 | static unsigned long utime_since(struct timeval *s, struct timeval *e) | |
109 | { | |
110 | double sec, usec; | |
111 | ||
112 | sec = e->tv_sec - s->tv_sec; | |
113 | usec = e->tv_usec - s->tv_usec; | |
114 | if (sec > 0 && usec < 0) { | |
115 | sec--; | |
116 | usec += 1000000; | |
117 | } | |
118 | ||
119 | sec *= (double) 1000000; | |
120 | ||
121 | return sec + usec; | |
122 | } | |
123 | ||
124 | static unsigned long utime_since_now(struct timeval *s) | |
125 | { | |
126 | struct timeval t; | |
127 | ||
128 | gettimeofday(&t, NULL); | |
129 | return utime_since(s, &t); | |
130 | } | |
131 | ||
132 | static unsigned long mtime_since(struct timeval *s, struct timeval *e) | |
133 | { | |
134 | double sec, usec; | |
135 | ||
136 | sec = e->tv_sec - s->tv_sec; | |
137 | usec = e->tv_usec - s->tv_usec; | |
138 | if (sec > 0 && usec < 0) { | |
139 | sec--; | |
140 | usec += 1000000; | |
141 | } | |
142 | ||
143 | sec *= (double) 1000; | |
144 | usec /= (double) 1000; | |
145 | ||
146 | return sec + usec; | |
147 | } | |
148 | ||
149 | static unsigned long mtime_since_now(struct timeval *s) | |
150 | { | |
151 | struct timeval t; | |
152 | ||
153 | gettimeofday(&t, NULL); | |
154 | return mtime_since(s, &t); | |
155 | } | |
156 | ||
157 | static inline unsigned long msec_now(struct timeval *s) | |
158 | { | |
159 | return s->tv_sec * 1000 + s->tv_usec / 1000; | |
160 | } | |
161 | ||
5289b847 JA |
162 | static unsigned long time_since_now(struct timeval *s) |
163 | { | |
164 | return mtime_since_now(s) / 1000; | |
165 | } | |
166 | ||
ebac4655 JA |
167 | static int random_map_free(struct thread_data *td, unsigned long long block) |
168 | { | |
169 | unsigned int idx = RAND_MAP_IDX(td, block); | |
170 | unsigned int bit = RAND_MAP_BIT(td, block); | |
171 | ||
172 | return (td->file_map[idx] & (1UL << bit)) == 0; | |
173 | } | |
174 | ||
175 | static int get_next_free_block(struct thread_data *td, unsigned long long *b) | |
176 | { | |
177 | int i; | |
178 | ||
179 | *b = 0; | |
180 | i = 0; | |
181 | while ((*b) * td->min_bs < td->io_size) { | |
182 | if (td->file_map[i] != -1UL) { | |
183 | *b += ffz(td->file_map[i]); | |
184 | return 0; | |
185 | } | |
186 | ||
187 | *b += BLOCKS_PER_MAP; | |
188 | i++; | |
189 | } | |
190 | ||
191 | return 1; | |
192 | } | |
193 | ||
194 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) | |
195 | { | |
196 | unsigned long block = io_u->offset / td->min_bs; | |
197 | unsigned int blocks = 0; | |
198 | ||
199 | while (blocks < (io_u->buflen / td->min_bs)) { | |
200 | unsigned int idx, bit; | |
201 | ||
202 | if (!random_map_free(td, block)) | |
203 | break; | |
204 | ||
205 | idx = RAND_MAP_IDX(td, block); | |
206 | bit = RAND_MAP_BIT(td, block); | |
207 | ||
208 | assert(idx < td->num_maps); | |
209 | ||
210 | td->file_map[idx] |= (1UL << bit); | |
211 | block++; | |
212 | blocks++; | |
213 | } | |
214 | ||
215 | if ((blocks * td->min_bs) < io_u->buflen) | |
216 | io_u->buflen = blocks * td->min_bs; | |
217 | } | |
218 | ||
ebac4655 JA |
219 | static inline void add_stat_sample(struct io_stat *is, unsigned long val) |
220 | { | |
221 | if (val > is->max_val) | |
222 | is->max_val = val; | |
223 | if (val < is->min_val) | |
224 | is->min_val = val; | |
225 | ||
226 | is->val += val; | |
227 | is->val_sq += val * val; | |
228 | is->samples++; | |
229 | } | |
230 | ||
231 | static void add_log_sample(struct thread_data *td, struct io_log *iolog, | |
232 | unsigned long val, int ddir) | |
233 | { | |
234 | if (iolog->nr_samples == iolog->max_samples) { | |
235 | int new_size = sizeof(struct io_sample) * iolog->max_samples*2; | |
236 | ||
237 | iolog->log = realloc(iolog->log, new_size); | |
238 | iolog->max_samples <<= 1; | |
239 | } | |
240 | ||
241 | iolog->log[iolog->nr_samples].val = val; | |
242 | iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch); | |
243 | iolog->log[iolog->nr_samples].ddir = ddir; | |
244 | iolog->nr_samples++; | |
245 | } | |
246 | ||
247 | static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec) | |
248 | { | |
249 | add_stat_sample(&td->clat_stat[ddir], msec); | |
250 | ||
251 | if (td->clat_log) | |
252 | add_log_sample(td, td->clat_log, msec, ddir); | |
253 | } | |
254 | ||
255 | static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec) | |
256 | { | |
257 | add_stat_sample(&td->slat_stat[ddir], msec); | |
258 | ||
259 | if (td->slat_log) | |
260 | add_log_sample(td, td->slat_log, msec, ddir); | |
261 | } | |
262 | ||
263 | static void add_bw_sample(struct thread_data *td, int ddir) | |
264 | { | |
265 | unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]); | |
266 | unsigned long rate; | |
267 | ||
268 | if (spent < td->bw_avg_time) | |
269 | return; | |
270 | ||
271 | rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent; | |
272 | add_stat_sample(&td->bw_stat[ddir], rate); | |
273 | ||
274 | if (td->bw_log) | |
275 | add_log_sample(td, td->bw_log, rate, ddir); | |
276 | ||
277 | gettimeofday(&td->stat_sample_time[ddir], NULL); | |
278 | td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; | |
279 | } | |
280 | ||
20dc95c4 JA |
281 | static int get_next_offset(struct thread_data *td, unsigned long long *offset) |
282 | { | |
283 | unsigned long long b, rb; | |
284 | long r; | |
285 | ||
286 | if (!td->sequential) { | |
287 | unsigned long max_blocks = td->io_size / td->min_bs; | |
288 | int loops = 50; | |
289 | ||
290 | do { | |
291 | lrand48_r(&td->random_state, &r); | |
292 | b = ((max_blocks - 1) * r / (RAND_MAX+1.0)); | |
293 | rb = b + (td->file_offset / td->min_bs); | |
294 | loops--; | |
295 | } while (!random_map_free(td, rb) && loops); | |
296 | ||
297 | if (!loops) { | |
298 | if (get_next_free_block(td, &b)) | |
299 | return 1; | |
300 | } | |
301 | } else | |
302 | b = td->last_pos / td->min_bs; | |
303 | ||
304 | *offset = (b * td->min_bs) + td->file_offset; | |
838a3cd3 | 305 | if (*offset > td->real_file_size) |
20dc95c4 JA |
306 | return 1; |
307 | ||
308 | return 0; | |
309 | } | |
310 | ||
311 | static unsigned int get_next_buflen(struct thread_data *td) | |
312 | { | |
313 | unsigned int buflen; | |
314 | long r; | |
315 | ||
316 | if (td->min_bs == td->max_bs) | |
317 | buflen = td->min_bs; | |
318 | else { | |
319 | lrand48_r(&td->bsrange_state, &r); | |
320 | buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0)); | |
321 | buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1); | |
322 | } | |
323 | ||
324 | if (buflen > td->io_size - td->this_io_bytes[td->ddir]) | |
325 | buflen = td->io_size - td->this_io_bytes[td->ddir]; | |
326 | ||
327 | return buflen; | |
328 | } | |
329 | ||
ebac4655 JA |
330 | /* |
331 | * busy looping version for the last few usec | |
332 | */ | |
333 | static void __usec_sleep(unsigned int usec) | |
334 | { | |
335 | struct timeval start; | |
336 | ||
337 | gettimeofday(&start, NULL); | |
338 | while (utime_since_now(&start) < usec) | |
339 | nop; | |
340 | } | |
341 | ||
342 | static void usec_sleep(struct thread_data *td, unsigned long usec) | |
343 | { | |
344 | struct timespec req, rem; | |
345 | ||
346 | req.tv_sec = usec / 1000000; | |
347 | req.tv_nsec = usec * 1000 - req.tv_sec * 1000000; | |
348 | ||
349 | do { | |
350 | if (usec < 5000) { | |
351 | __usec_sleep(usec); | |
352 | break; | |
353 | } | |
354 | ||
355 | rem.tv_sec = rem.tv_nsec = 0; | |
356 | if (nanosleep(&req, &rem) < 0) | |
357 | break; | |
358 | ||
359 | if ((rem.tv_sec + rem.tv_nsec) == 0) | |
360 | break; | |
361 | ||
362 | req.tv_nsec = rem.tv_nsec; | |
363 | req.tv_sec = rem.tv_sec; | |
364 | ||
365 | usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000; | |
366 | } while (!td->terminate); | |
367 | } | |
368 | ||
369 | static void rate_throttle(struct thread_data *td, unsigned long time_spent, | |
370 | unsigned int bytes) | |
371 | { | |
372 | unsigned long usec_cycle; | |
373 | ||
374 | if (!td->rate) | |
375 | return; | |
376 | ||
377 | usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs); | |
378 | ||
379 | if (time_spent < usec_cycle) { | |
380 | unsigned long s = usec_cycle - time_spent; | |
381 | ||
382 | td->rate_pending_usleep += s; | |
383 | if (td->rate_pending_usleep >= 100000) { | |
384 | usec_sleep(td, td->rate_pending_usleep); | |
385 | td->rate_pending_usleep = 0; | |
386 | } | |
387 | } else { | |
388 | long overtime = time_spent - usec_cycle; | |
389 | ||
390 | td->rate_pending_usleep -= overtime; | |
391 | } | |
392 | } | |
393 | ||
394 | static int check_min_rate(struct thread_data *td, struct timeval *now) | |
395 | { | |
396 | unsigned long spent; | |
397 | unsigned long rate; | |
398 | int ddir = td->ddir; | |
399 | ||
400 | /* | |
401 | * allow a 2 second settle period in the beginning | |
402 | */ | |
403 | if (mtime_since(&td->start, now) < 2000) | |
404 | return 0; | |
405 | ||
406 | /* | |
407 | * if rate blocks is set, sample is running | |
408 | */ | |
409 | if (td->rate_bytes) { | |
410 | spent = mtime_since(&td->lastrate, now); | |
411 | if (spent < td->ratecycle) | |
412 | return 0; | |
413 | ||
414 | rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent; | |
415 | if (rate < td->ratemin) { | |
416 | printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate); | |
417 | if (rate_quit) | |
418 | terminate_threads(td->groupid); | |
419 | return 1; | |
420 | } | |
421 | } | |
422 | ||
423 | td->rate_bytes = td->this_io_bytes[ddir]; | |
424 | memcpy(&td->lastrate, now, sizeof(*now)); | |
425 | return 0; | |
426 | } | |
427 | ||
428 | static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) | |
429 | { | |
430 | if (!td->timeout) | |
431 | return 0; | |
432 | if (mtime_since(&td->epoch, t) >= td->timeout * 1000) | |
433 | return 1; | |
434 | ||
435 | return 0; | |
436 | } | |
437 | ||
438 | static void fill_random_bytes(struct thread_data *td, | |
439 | unsigned char *p, unsigned int len) | |
440 | { | |
441 | unsigned int todo; | |
442 | double r; | |
443 | ||
444 | while (len) { | |
445 | drand48_r(&td->verify_state, &r); | |
446 | ||
447 | /* | |
448 | * lrand48_r seems to be broken and only fill the bottom | |
449 | * 32-bits, even on 64-bit archs with 64-bit longs | |
450 | */ | |
451 | todo = sizeof(r); | |
452 | if (todo > len) | |
453 | todo = len; | |
454 | ||
455 | memcpy(p, &r, todo); | |
456 | ||
457 | len -= todo; | |
458 | p += todo; | |
459 | } | |
460 | } | |
461 | ||
462 | static void hexdump(void *buffer, int len) | |
463 | { | |
464 | unsigned char *p = buffer; | |
465 | int i; | |
466 | ||
467 | for (i = 0; i < len; i++) | |
468 | printf("%02x", p[i]); | |
469 | printf("\n"); | |
470 | } | |
471 | ||
472 | static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u) | |
473 | { | |
474 | unsigned char *p = (unsigned char *) io_u->buf; | |
475 | unsigned long c; | |
476 | int ret; | |
477 | ||
478 | p += sizeof(*hdr); | |
479 | c = crc32(p, hdr->len - sizeof(*hdr)); | |
480 | ret = c != hdr->crc32; | |
481 | ||
482 | if (ret) { | |
483 | fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen); | |
484 | fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c); | |
485 | } | |
486 | ||
487 | return ret; | |
488 | } | |
489 | ||
490 | static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u) | |
491 | { | |
492 | unsigned char *p = (unsigned char *) io_u->buf; | |
493 | struct md5_ctx md5_ctx; | |
494 | int ret; | |
495 | ||
496 | memset(&md5_ctx, 0, sizeof(md5_ctx)); | |
497 | p += sizeof(*hdr); | |
498 | md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr)); | |
499 | ||
500 | ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash)); | |
501 | if (ret) { | |
502 | fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen); | |
503 | hexdump(hdr->md5_digest, sizeof(hdr->md5_digest)); | |
504 | hexdump(md5_ctx.hash, sizeof(md5_ctx.hash)); | |
505 | } | |
506 | ||
507 | return ret; | |
508 | } | |
509 | ||
510 | static int verify_io_u(struct io_u *io_u) | |
511 | { | |
512 | struct verify_header *hdr = (struct verify_header *) io_u->buf; | |
513 | int ret; | |
514 | ||
515 | if (hdr->fio_magic != FIO_HDR_MAGIC) | |
516 | return 1; | |
517 | ||
518 | if (hdr->verify_type == VERIFY_MD5) | |
519 | ret = verify_io_u_md5(hdr, io_u); | |
520 | else if (hdr->verify_type == VERIFY_CRC32) | |
521 | ret = verify_io_u_crc32(hdr, io_u); | |
522 | else { | |
523 | fprintf(stderr, "Bad verify type %d\n", hdr->verify_type); | |
524 | ret = 1; | |
525 | } | |
526 | ||
527 | return ret; | |
528 | } | |
529 | ||
530 | static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len) | |
531 | { | |
532 | hdr->crc32 = crc32(p, len); | |
533 | } | |
534 | ||
535 | static void fill_md5(struct verify_header *hdr, void *p, unsigned int len) | |
536 | { | |
537 | struct md5_ctx md5_ctx; | |
538 | ||
539 | memset(&md5_ctx, 0, sizeof(md5_ctx)); | |
540 | md5_update(&md5_ctx, p, len); | |
541 | memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash)); | |
542 | } | |
543 | ||
3d60d1ed JA |
544 | unsigned int hweight32(unsigned int w) |
545 | { | |
546 | unsigned int res = w - ((w >> 1) & 0x55555555); | |
547 | ||
548 | res = (res & 0x33333333) + ((res >> 2) & 0x33333333); | |
549 | res = (res + (res >> 4)) & 0x0F0F0F0F; | |
550 | res = res + (res >> 8); | |
551 | ||
552 | return (res + (res >> 16)) & 0x000000FF; | |
553 | } | |
554 | ||
555 | unsigned long hweight64(unsigned long long w) | |
556 | { | |
557 | #if __WORDSIZE == 32 | |
558 | return hweight32((unsigned int)(w >> 32)) + hweight32((unsigned int)w); | |
559 | #elif __WORDSIZE == 64 | |
560 | unsigned long long v = w - ((w >> 1) & 0x5555555555555555ul); | |
561 | ||
562 | v = (v & 0x3333333333333333ul) + ((v >> 2) & 0x3333333333333333ul); | |
563 | v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0Ful; | |
564 | v = v + (v >> 8); | |
565 | v = v + (v >> 16); | |
566 | ||
567 | return (v + (v >> 32)) & 0x00000000000000FFul; | |
568 | #else | |
569 | #error __WORDSIZE not defined | |
570 | #endif | |
571 | } | |
572 | ||
573 | static int get_rw_ddir(struct thread_data *td) | |
574 | { | |
575 | /* | |
576 | * perhaps cheasy, but use the hamming weight of the position | |
577 | * as a randomizer for data direction. | |
578 | */ | |
579 | if (td_rw(td)) | |
580 | return hweight64(td->last_pos) & 1; | |
581 | else if (td_read(td)) | |
582 | return DDIR_READ; | |
583 | else | |
584 | return DDIR_WRITE; | |
585 | } | |
586 | ||
ebac4655 JA |
587 | /* |
588 | * fill body of io_u->buf with random data and add a header with the | |
589 | * (eg) sha1sum of that data. | |
590 | */ | |
591 | static void populate_io_u(struct thread_data *td, struct io_u *io_u) | |
592 | { | |
593 | unsigned char *p = (unsigned char *) io_u->buf; | |
594 | struct verify_header hdr; | |
595 | ||
596 | hdr.fio_magic = FIO_HDR_MAGIC; | |
597 | hdr.len = io_u->buflen; | |
598 | p += sizeof(hdr); | |
599 | fill_random_bytes(td, p, io_u->buflen - sizeof(hdr)); | |
600 | ||
601 | if (td->verify == VERIFY_MD5) { | |
602 | fill_md5(&hdr, p, io_u->buflen - sizeof(hdr)); | |
603 | hdr.verify_type = VERIFY_MD5; | |
604 | } else { | |
605 | fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr)); | |
606 | hdr.verify_type = VERIFY_CRC32; | |
607 | } | |
608 | ||
609 | memcpy(io_u->buf, &hdr, sizeof(hdr)); | |
610 | } | |
611 | ||
20dc95c4 JA |
612 | static int td_io_prep(struct thread_data *td, struct io_u *io_u, int read) |
613 | { | |
614 | if (read) | |
615 | io_u->ddir = DDIR_READ; | |
616 | else | |
617 | io_u->ddir = DDIR_WRITE; | |
618 | ||
619 | if (td->io_prep && td->io_prep(td, io_u)) | |
620 | return 1; | |
621 | ||
622 | return 0; | |
623 | } | |
624 | ||
b1ff3403 | 625 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
ebac4655 JA |
626 | { |
627 | list_del(&io_u->list); | |
628 | list_add(&io_u->list, &td->io_u_freelist); | |
629 | td->cur_depth--; | |
630 | } | |
631 | ||
632 | #define queue_full(td) (list_empty(&(td)->io_u_freelist)) | |
633 | ||
b1ff3403 | 634 | struct io_u *__get_io_u(struct thread_data *td) |
ebac4655 JA |
635 | { |
636 | struct io_u *io_u; | |
637 | ||
638 | if (queue_full(td)) | |
639 | return NULL; | |
640 | ||
641 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); | |
642 | io_u->error = 0; | |
643 | io_u->resid = 0; | |
644 | list_del(&io_u->list); | |
645 | list_add(&io_u->list, &td->io_u_busylist); | |
646 | td->cur_depth++; | |
647 | return io_u; | |
648 | } | |
649 | ||
ebac4655 JA |
650 | static struct io_u *get_io_u(struct thread_data *td) |
651 | { | |
652 | struct io_u *io_u; | |
653 | ||
654 | io_u = __get_io_u(td); | |
655 | if (!io_u) | |
656 | return NULL; | |
657 | ||
20dc95c4 JA |
658 | if (td->zone_bytes >= td->zone_size) { |
659 | td->zone_bytes = 0; | |
660 | td->last_pos += td->zone_skip; | |
661 | } | |
662 | ||
ebac4655 JA |
663 | if (get_next_offset(td, &io_u->offset)) { |
664 | put_io_u(td, io_u); | |
665 | return NULL; | |
666 | } | |
667 | ||
668 | io_u->buflen = get_next_buflen(td); | |
669 | if (!io_u->buflen) { | |
670 | put_io_u(td, io_u); | |
671 | return NULL; | |
672 | } | |
673 | ||
838a3cd3 JA |
674 | if (io_u->buflen + io_u->offset > td->real_file_size) |
675 | io_u->buflen = td->real_file_size - io_u->offset; | |
ebac4655 JA |
676 | |
677 | if (!io_u->buflen) { | |
678 | put_io_u(td, io_u); | |
679 | return NULL; | |
680 | } | |
681 | ||
682 | if (!td->sequential) | |
683 | mark_random_map(td, io_u); | |
684 | ||
20dc95c4 | 685 | td->last_pos += io_u->buflen; |
ebac4655 JA |
686 | |
687 | if (td->verify != VERIFY_NONE) | |
688 | populate_io_u(td, io_u); | |
689 | ||
3d60d1ed | 690 | if (td_io_prep(td, io_u, get_rw_ddir(td))) { |
ebac4655 JA |
691 | put_io_u(td, io_u); |
692 | return NULL; | |
693 | } | |
694 | ||
695 | gettimeofday(&io_u->start_time, NULL); | |
696 | return io_u; | |
697 | } | |
698 | ||
699 | static inline void td_set_runstate(struct thread_data *td, int runstate) | |
700 | { | |
701 | td->old_runstate = td->runstate; | |
702 | td->runstate = runstate; | |
703 | } | |
704 | ||
705 | static int get_next_verify(struct thread_data *td, | |
706 | unsigned long long *offset, unsigned int *len) | |
707 | { | |
708 | struct io_piece *ipo; | |
709 | ||
710 | if (list_empty(&td->io_hist_list)) | |
711 | return 1; | |
712 | ||
713 | ipo = list_entry(td->io_hist_list.next, struct io_piece, list); | |
714 | list_del(&ipo->list); | |
715 | ||
716 | *offset = ipo->offset; | |
717 | *len = ipo->len; | |
718 | free(ipo); | |
719 | return 0; | |
720 | } | |
721 | ||
722 | static void prune_io_piece_log(struct thread_data *td) | |
723 | { | |
724 | struct io_piece *ipo; | |
725 | ||
726 | while (!list_empty(&td->io_hist_list)) { | |
727 | ipo = list_entry(td->io_hist_list.next, struct io_piece, list); | |
728 | ||
729 | list_del(&ipo->list); | |
730 | free(ipo); | |
731 | } | |
732 | } | |
733 | ||
734 | /* | |
735 | * log a succesful write, so we can unwind the log for verify | |
736 | */ | |
737 | static void log_io_piece(struct thread_data *td, struct io_u *io_u) | |
738 | { | |
739 | struct io_piece *ipo = malloc(sizeof(struct io_piece)); | |
740 | struct list_head *entry; | |
741 | ||
742 | INIT_LIST_HEAD(&ipo->list); | |
743 | ipo->offset = io_u->offset; | |
744 | ipo->len = io_u->buflen; | |
745 | ||
746 | /* | |
747 | * for random io where the writes extend the file, it will typically | |
748 | * be laid out with the block scattered as written. it's faster to | |
749 | * read them in in that order again, so don't sort | |
750 | */ | |
751 | if (td->sequential || !td->overwrite) { | |
752 | list_add_tail(&ipo->list, &td->io_hist_list); | |
753 | return; | |
754 | } | |
755 | ||
756 | /* | |
757 | * for random io, sort the list so verify will run faster | |
758 | */ | |
759 | entry = &td->io_hist_list; | |
760 | while ((entry = entry->prev) != &td->io_hist_list) { | |
761 | struct io_piece *__ipo = list_entry(entry, struct io_piece, list); | |
762 | ||
763 | if (__ipo->offset < ipo->offset) | |
764 | break; | |
765 | } | |
766 | ||
767 | list_add(&ipo->list, entry); | |
768 | } | |
769 | ||
770 | static int sync_td(struct thread_data *td) | |
771 | { | |
772 | if (td->io_sync) | |
773 | return td->io_sync(td); | |
774 | ||
775 | return 0; | |
776 | } | |
777 | ||
778 | static int io_u_getevents(struct thread_data *td, int min, int max, | |
779 | struct timespec *t) | |
780 | { | |
781 | return td->io_getevents(td, min, max, t); | |
782 | } | |
783 | ||
784 | static int io_u_queue(struct thread_data *td, struct io_u *io_u) | |
785 | { | |
786 | gettimeofday(&io_u->issue_time, NULL); | |
787 | ||
788 | return td->io_queue(td, io_u); | |
789 | } | |
790 | ||
791 | #define iocb_time(iocb) ((unsigned long) (iocb)->data) | |
792 | ||
793 | static void io_completed(struct thread_data *td, struct io_u *io_u, | |
794 | struct io_completion_data *icd) | |
795 | { | |
796 | struct timeval e; | |
797 | unsigned long msec; | |
798 | ||
799 | gettimeofday(&e, NULL); | |
800 | ||
801 | if (!io_u->error) { | |
20dc95c4 | 802 | unsigned int bytes = io_u->buflen - io_u->resid; |
3d60d1ed | 803 | const int idx = io_u->ddir; |
ebac4655 JA |
804 | |
805 | td->io_blocks[idx]++; | |
20dc95c4 JA |
806 | td->io_bytes[idx] += bytes; |
807 | td->zone_bytes += bytes; | |
808 | td->this_io_bytes[idx] += bytes; | |
ebac4655 JA |
809 | |
810 | msec = mtime_since(&io_u->issue_time, &e); | |
811 | ||
3d60d1ed JA |
812 | add_clat_sample(td, idx, msec); |
813 | add_bw_sample(td, idx); | |
ebac4655 | 814 | |
3d60d1ed | 815 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) |
ebac4655 JA |
816 | log_io_piece(td, io_u); |
817 | ||
20dc95c4 | 818 | icd->bytes_done[idx] += bytes; |
ebac4655 JA |
819 | } else |
820 | icd->error = io_u->error; | |
821 | } | |
822 | ||
823 | static void ios_completed(struct thread_data *td,struct io_completion_data *icd) | |
824 | { | |
825 | struct io_u *io_u; | |
826 | int i; | |
827 | ||
828 | icd->error = 0; | |
829 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
830 | ||
831 | for (i = 0; i < icd->nr; i++) { | |
832 | io_u = td->io_event(td, i); | |
833 | ||
834 | io_completed(td, io_u, icd); | |
835 | put_io_u(td, io_u); | |
836 | } | |
837 | } | |
838 | ||
839 | static void cleanup_pending_aio(struct thread_data *td) | |
840 | { | |
841 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; | |
842 | struct list_head *entry, *n; | |
843 | struct io_completion_data icd; | |
844 | struct io_u *io_u; | |
845 | int r; | |
846 | ||
847 | /* | |
848 | * get immediately available events, if any | |
849 | */ | |
850 | r = io_u_getevents(td, 0, td->cur_depth, &ts); | |
851 | if (r > 0) { | |
852 | icd.nr = r; | |
853 | ios_completed(td, &icd); | |
854 | } | |
855 | ||
856 | /* | |
857 | * now cancel remaining active events | |
858 | */ | |
859 | if (td->io_cancel) { | |
860 | list_for_each_safe(entry, n, &td->io_u_busylist) { | |
861 | io_u = list_entry(entry, struct io_u, list); | |
862 | ||
863 | r = td->io_cancel(td, io_u); | |
864 | if (!r) | |
865 | put_io_u(td, io_u); | |
866 | } | |
867 | } | |
868 | ||
869 | if (td->cur_depth) { | |
870 | r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL); | |
871 | if (r > 0) { | |
872 | icd.nr = r; | |
873 | ios_completed(td, &icd); | |
874 | } | |
875 | } | |
876 | } | |
877 | ||
878 | static int do_io_u_verify(struct thread_data *td, struct io_u **io_u) | |
879 | { | |
880 | struct io_u *v_io_u = *io_u; | |
881 | int ret = 0; | |
882 | ||
883 | if (v_io_u) { | |
884 | ret = verify_io_u(v_io_u); | |
885 | put_io_u(td, v_io_u); | |
886 | *io_u = NULL; | |
887 | } | |
888 | ||
889 | return ret; | |
890 | } | |
891 | ||
892 | static void do_verify(struct thread_data *td) | |
893 | { | |
894 | struct timeval t; | |
895 | struct io_u *io_u, *v_io_u = NULL; | |
896 | struct io_completion_data icd; | |
897 | int ret; | |
898 | ||
899 | td_set_runstate(td, TD_VERIFYING); | |
900 | ||
901 | do { | |
902 | if (td->terminate) | |
903 | break; | |
904 | ||
905 | gettimeofday(&t, NULL); | |
906 | if (runtime_exceeded(td, &t)) | |
907 | break; | |
908 | ||
909 | io_u = __get_io_u(td); | |
910 | if (!io_u) | |
911 | break; | |
912 | ||
913 | if (get_next_verify(td, &io_u->offset, &io_u->buflen)) { | |
914 | put_io_u(td, io_u); | |
915 | break; | |
916 | } | |
917 | ||
918 | if (td_io_prep(td, io_u, 1)) { | |
919 | put_io_u(td, io_u); | |
920 | break; | |
921 | } | |
922 | ||
923 | ret = io_u_queue(td, io_u); | |
924 | if (ret) { | |
925 | put_io_u(td, io_u); | |
926 | td_verror(td, ret); | |
927 | break; | |
928 | } | |
929 | ||
930 | /* | |
931 | * we have one pending to verify, do that while | |
932 | * we are doing io on the next one | |
933 | */ | |
934 | if (do_io_u_verify(td, &v_io_u)) | |
935 | break; | |
936 | ||
937 | ret = io_u_getevents(td, 1, 1, NULL); | |
938 | if (ret != 1) { | |
939 | if (ret < 0) | |
940 | td_verror(td, ret); | |
941 | break; | |
942 | } | |
943 | ||
944 | v_io_u = td->io_event(td, 0); | |
945 | icd.nr = 1; | |
946 | icd.error = 0; | |
947 | io_completed(td, v_io_u, &icd); | |
948 | ||
949 | if (icd.error) { | |
950 | td_verror(td, icd.error); | |
951 | put_io_u(td, v_io_u); | |
952 | v_io_u = NULL; | |
953 | break; | |
954 | } | |
955 | ||
ebac4655 JA |
956 | /* |
957 | * if we can't submit more io, we need to verify now | |
958 | */ | |
959 | if (queue_full(td) && do_io_u_verify(td, &v_io_u)) | |
960 | break; | |
961 | ||
962 | } while (1); | |
963 | ||
964 | do_io_u_verify(td, &v_io_u); | |
965 | ||
966 | if (td->cur_depth) | |
967 | cleanup_pending_aio(td); | |
968 | ||
969 | td_set_runstate(td, TD_RUNNING); | |
970 | } | |
971 | ||
972 | static void do_io(struct thread_data *td) | |
973 | { | |
974 | struct io_completion_data icd; | |
975 | struct timeval s, e; | |
976 | unsigned long usec; | |
977 | ||
978 | while (td->this_io_bytes[td->ddir] < td->io_size) { | |
979 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; | |
980 | struct timespec *timeout; | |
981 | int ret, min_evts = 0; | |
982 | struct io_u *io_u; | |
983 | ||
984 | if (td->terminate) | |
985 | break; | |
986 | ||
987 | io_u = get_io_u(td); | |
988 | if (!io_u) | |
989 | break; | |
990 | ||
991 | memcpy(&s, &io_u->start_time, sizeof(s)); | |
992 | ||
993 | ret = io_u_queue(td, io_u); | |
994 | if (ret) { | |
995 | put_io_u(td, io_u); | |
996 | td_verror(td, ret); | |
997 | break; | |
998 | } | |
999 | ||
1000 | add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time)); | |
1001 | ||
1002 | if (td->cur_depth < td->iodepth) { | |
1003 | timeout = &ts; | |
1004 | min_evts = 0; | |
1005 | } else { | |
1006 | timeout = NULL; | |
1007 | min_evts = 1; | |
1008 | } | |
1009 | ||
1010 | ret = io_u_getevents(td, min_evts, td->cur_depth, timeout); | |
1011 | if (ret < 0) { | |
1012 | td_verror(td, ret); | |
1013 | break; | |
1014 | } else if (!ret) | |
1015 | continue; | |
1016 | ||
1017 | icd.nr = ret; | |
1018 | ios_completed(td, &icd); | |
1019 | if (icd.error) { | |
1020 | td_verror(td, icd.error); | |
1021 | break; | |
1022 | } | |
1023 | ||
1024 | /* | |
1025 | * the rate is batched for now, it should work for batches | |
1026 | * of completions except the very first one which may look | |
1027 | * a little bursty | |
1028 | */ | |
1029 | gettimeofday(&e, NULL); | |
1030 | usec = utime_since(&s, &e); | |
1031 | ||
1032 | rate_throttle(td, usec, icd.bytes_done[td->ddir]); | |
1033 | ||
1034 | if (check_min_rate(td, &e)) { | |
1035 | td_verror(td, ENOMEM); | |
1036 | break; | |
1037 | } | |
1038 | ||
1039 | if (runtime_exceeded(td, &e)) | |
1040 | break; | |
1041 | ||
1042 | if (td->thinktime) | |
1043 | usec_sleep(td, td->thinktime); | |
1044 | ||
1045 | if (should_fsync(td) && td->fsync_blocks && | |
1046 | (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0) | |
1047 | sync_td(td); | |
1048 | } | |
1049 | ||
1050 | if (td->cur_depth) | |
1051 | cleanup_pending_aio(td); | |
1052 | ||
2f844d30 | 1053 | if (should_fsync(td) && td->fsync_blocks) |
ebac4655 JA |
1054 | sync_td(td); |
1055 | } | |
1056 | ||
1057 | static void cleanup_io(struct thread_data *td) | |
1058 | { | |
1059 | if (td->io_cleanup) | |
1060 | td->io_cleanup(td); | |
1061 | } | |
1062 | ||
1063 | static int init_io(struct thread_data *td) | |
1064 | { | |
1065 | if (td->io_engine == FIO_SYNCIO) | |
1066 | return fio_syncio_init(td); | |
1067 | else if (td->io_engine == FIO_MMAPIO) | |
1068 | return fio_mmapio_init(td); | |
1069 | else if (td->io_engine == FIO_LIBAIO) | |
1070 | return fio_libaio_init(td); | |
1071 | else if (td->io_engine == FIO_POSIXAIO) | |
1072 | return fio_posixaio_init(td); | |
1073 | else if (td->io_engine == FIO_SGIO) | |
1074 | return fio_sgio_init(td); | |
1075 | else { | |
1076 | fprintf(stderr, "bad io_engine %d\n", td->io_engine); | |
1077 | return 1; | |
1078 | } | |
1079 | } | |
1080 | ||
1081 | static void cleanup_io_u(struct thread_data *td) | |
1082 | { | |
1083 | struct list_head *entry, *n; | |
1084 | struct io_u *io_u; | |
1085 | ||
1086 | list_for_each_safe(entry, n, &td->io_u_freelist) { | |
1087 | io_u = list_entry(entry, struct io_u, list); | |
1088 | ||
1089 | list_del(&io_u->list); | |
1090 | free(io_u); | |
1091 | } | |
1092 | ||
1093 | if (td->mem_type == MEM_MALLOC) | |
1094 | free(td->orig_buffer); | |
1095 | else if (td->mem_type == MEM_SHM) { | |
1096 | struct shmid_ds sbuf; | |
1097 | ||
1098 | shmdt(td->orig_buffer); | |
1099 | shmctl(td->shm_id, IPC_RMID, &sbuf); | |
1100 | } else if (td->mem_type == MEM_MMAP) | |
1101 | munmap(td->orig_buffer, td->orig_buffer_size); | |
1102 | else | |
1103 | fprintf(stderr, "Bad memory type %d\n", td->mem_type); | |
1104 | ||
1105 | td->orig_buffer = NULL; | |
1106 | } | |
1107 | ||
1108 | static int init_io_u(struct thread_data *td) | |
1109 | { | |
1110 | struct io_u *io_u; | |
1111 | int i, max_units; | |
1112 | char *p; | |
1113 | ||
1114 | if (td->io_engine & FIO_SYNCIO) | |
1115 | max_units = 1; | |
1116 | else | |
1117 | max_units = td->iodepth; | |
1118 | ||
1119 | td->orig_buffer_size = td->max_bs * max_units + MASK; | |
1120 | ||
1121 | if (td->mem_type == MEM_MALLOC) | |
1122 | td->orig_buffer = malloc(td->orig_buffer_size); | |
1123 | else if (td->mem_type == MEM_SHM) { | |
1124 | td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600); | |
1125 | if (td->shm_id < 0) { | |
1126 | td_verror(td, errno); | |
1127 | perror("shmget"); | |
1128 | return 1; | |
1129 | } | |
1130 | ||
1131 | td->orig_buffer = shmat(td->shm_id, NULL, 0); | |
1132 | if (td->orig_buffer == (void *) -1) { | |
1133 | td_verror(td, errno); | |
1134 | perror("shmat"); | |
1135 | td->orig_buffer = NULL; | |
1136 | return 1; | |
1137 | } | |
1138 | } else if (td->mem_type == MEM_MMAP) { | |
1139 | td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0); | |
1140 | if (td->orig_buffer == MAP_FAILED) { | |
1141 | td_verror(td, errno); | |
1142 | perror("mmap"); | |
1143 | td->orig_buffer = NULL; | |
1144 | return 1; | |
1145 | } | |
1146 | } | |
1147 | ||
1148 | INIT_LIST_HEAD(&td->io_u_freelist); | |
1149 | INIT_LIST_HEAD(&td->io_u_busylist); | |
1150 | INIT_LIST_HEAD(&td->io_hist_list); | |
1151 | ||
1152 | p = ALIGN(td->orig_buffer); | |
1153 | for (i = 0; i < max_units; i++) { | |
1154 | io_u = malloc(sizeof(*io_u)); | |
1155 | memset(io_u, 0, sizeof(*io_u)); | |
1156 | INIT_LIST_HEAD(&io_u->list); | |
1157 | ||
1158 | io_u->buf = p + td->max_bs * i; | |
b1ff3403 | 1159 | io_u->index = i; |
ebac4655 JA |
1160 | list_add(&io_u->list, &td->io_u_freelist); |
1161 | } | |
1162 | ||
1163 | return 0; | |
1164 | } | |
1165 | ||
1166 | static int create_file(struct thread_data *td, unsigned long long size, | |
1167 | int extend) | |
1168 | { | |
1169 | unsigned long long left; | |
1170 | unsigned int bs; | |
1171 | int r, oflags; | |
1172 | char *b; | |
1173 | ||
1174 | /* | |
1175 | * unless specifically asked for overwrite, let normal io extend it | |
1176 | */ | |
1177 | if (td_write(td) && !td->overwrite) | |
1178 | return 0; | |
1179 | ||
1180 | if (!size) { | |
1181 | fprintf(stderr, "Need size for create\n"); | |
1182 | td_verror(td, EINVAL); | |
1183 | return 1; | |
1184 | } | |
1185 | ||
1186 | if (!extend) { | |
1187 | oflags = O_CREAT | O_TRUNC; | |
1188 | printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20); | |
1189 | } else { | |
1190 | oflags = O_APPEND; | |
1191 | printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20); | |
1192 | } | |
1193 | ||
1194 | td->fd = open(td->file_name, O_WRONLY | oflags, 0644); | |
1195 | if (td->fd < 0) { | |
1196 | td_verror(td, errno); | |
1197 | return 1; | |
1198 | } | |
1199 | ||
1200 | if (!extend && ftruncate(td->fd, td->file_size) == -1) { | |
1201 | td_verror(td, errno); | |
1202 | return 1; | |
1203 | } | |
1204 | ||
1205 | td->io_size = td->file_size; | |
1206 | b = malloc(td->max_bs); | |
1207 | memset(b, 0, td->max_bs); | |
1208 | ||
1209 | left = size; | |
1210 | while (left && !td->terminate) { | |
1211 | bs = td->max_bs; | |
1212 | if (bs > left) | |
1213 | bs = left; | |
1214 | ||
1215 | r = write(td->fd, b, bs); | |
1216 | ||
1217 | if (r == (int) bs) { | |
1218 | left -= bs; | |
1219 | continue; | |
1220 | } else { | |
1221 | if (r < 0) | |
1222 | td_verror(td, errno); | |
1223 | else | |
1224 | td_verror(td, EIO); | |
1225 | ||
1226 | break; | |
1227 | } | |
1228 | } | |
1229 | ||
1230 | if (td->terminate) | |
1231 | unlink(td->file_name); | |
1232 | else if (td->create_fsync) | |
1233 | fsync(td->fd); | |
1234 | ||
1235 | close(td->fd); | |
1236 | td->fd = -1; | |
1237 | free(b); | |
1238 | return 0; | |
1239 | } | |
1240 | ||
1241 | static int file_size(struct thread_data *td) | |
1242 | { | |
1243 | struct stat st; | |
1244 | ||
1245 | if (fstat(td->fd, &st) == -1) { | |
1246 | td_verror(td, errno); | |
1247 | return 1; | |
1248 | } | |
1249 | ||
838a3cd3 JA |
1250 | td->real_file_size = st.st_size; |
1251 | ||
1252 | if (!td->file_size || td->file_size > td->real_file_size) | |
1253 | td->file_size = td->real_file_size; | |
ebac4655 | 1254 | |
e8e387c1 | 1255 | td->file_size -= td->file_offset; |
ebac4655 JA |
1256 | return 0; |
1257 | } | |
1258 | ||
1259 | static int bdev_size(struct thread_data *td) | |
1260 | { | |
9104f874 | 1261 | unsigned long long bytes; |
ebac4655 JA |
1262 | int r; |
1263 | ||
1264 | r = blockdev_size(td->fd, &bytes); | |
1265 | if (r) { | |
1266 | td_verror(td, r); | |
1267 | return 1; | |
1268 | } | |
1269 | ||
838a3cd3 JA |
1270 | td->real_file_size = bytes; |
1271 | ||
ebac4655 JA |
1272 | /* |
1273 | * no extend possibilities, so limit size to device size if too large | |
1274 | */ | |
838a3cd3 JA |
1275 | if (!td->file_size || td->file_size > td->real_file_size) |
1276 | td->file_size = td->real_file_size; | |
ebac4655 | 1277 | |
e8e387c1 | 1278 | td->file_size -= td->file_offset; |
ebac4655 JA |
1279 | return 0; |
1280 | } | |
1281 | ||
1282 | static int get_file_size(struct thread_data *td) | |
1283 | { | |
0af7b542 | 1284 | int ret = 0; |
ebac4655 JA |
1285 | |
1286 | if (td->filetype == FIO_TYPE_FILE) | |
1287 | ret = file_size(td); | |
0af7b542 | 1288 | else if (td->filetype == FIO_TYPE_BD) |
ebac4655 | 1289 | ret = bdev_size(td); |
0af7b542 JA |
1290 | else |
1291 | td->real_file_size = -1; | |
ebac4655 JA |
1292 | |
1293 | if (ret) | |
1294 | return ret; | |
1295 | ||
e8e387c1 JA |
1296 | if (td->file_offset > td->real_file_size) { |
1297 | fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size); | |
ebac4655 JA |
1298 | return 1; |
1299 | } | |
1300 | ||
838a3cd3 | 1301 | td->io_size = td->file_size; |
ebac4655 JA |
1302 | if (td->io_size == 0) { |
1303 | fprintf(stderr, "Client%d: no io blocks\n", td->thread_number); | |
1304 | td_verror(td, EINVAL); | |
1305 | return 1; | |
1306 | } | |
1307 | ||
20dc95c4 JA |
1308 | if (!td->zone_size) |
1309 | td->zone_size = td->io_size; | |
1310 | ||
ebac4655 JA |
1311 | td->total_io_size = td->io_size * td->loops; |
1312 | return 0; | |
1313 | } | |
1314 | ||
1315 | static int setup_file_mmap(struct thread_data *td) | |
1316 | { | |
1317 | int flags; | |
1318 | ||
3d60d1ed JA |
1319 | if (td_rw(td)) |
1320 | flags = PROT_READ | PROT_WRITE; | |
1321 | else if (td_write(td)) { | |
ebac4655 JA |
1322 | flags = PROT_WRITE; |
1323 | ||
1324 | if (td->verify != VERIFY_NONE) | |
1325 | flags |= PROT_READ; | |
3d60d1ed JA |
1326 | } else |
1327 | flags = PROT_READ; | |
ebac4655 JA |
1328 | |
1329 | td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset); | |
1330 | if (td->mmap == MAP_FAILED) { | |
1331 | td->mmap = NULL; | |
1332 | td_verror(td, errno); | |
1333 | return 1; | |
1334 | } | |
1335 | ||
1336 | if (td->invalidate_cache) { | |
1337 | if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) { | |
1338 | td_verror(td, errno); | |
1339 | return 1; | |
1340 | } | |
1341 | } | |
1342 | ||
1343 | if (td->sequential) { | |
1344 | if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) { | |
1345 | td_verror(td, errno); | |
1346 | return 1; | |
1347 | } | |
1348 | } else { | |
1349 | if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) { | |
1350 | td_verror(td, errno); | |
1351 | return 1; | |
1352 | } | |
1353 | } | |
1354 | ||
1355 | return 0; | |
1356 | } | |
1357 | ||
1358 | static int setup_file_plain(struct thread_data *td) | |
1359 | { | |
1360 | if (td->invalidate_cache) { | |
1361 | if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) { | |
1362 | td_verror(td, errno); | |
1363 | return 1; | |
1364 | } | |
1365 | } | |
1366 | ||
1367 | if (td->sequential) { | |
1368 | if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) { | |
1369 | td_verror(td, errno); | |
1370 | return 1; | |
1371 | } | |
1372 | } else { | |
1373 | if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) { | |
1374 | td_verror(td, errno); | |
1375 | return 1; | |
1376 | } | |
1377 | } | |
1378 | ||
1379 | return 0; | |
1380 | } | |
1381 | ||
1382 | static int setup_file(struct thread_data *td) | |
1383 | { | |
1384 | struct stat st; | |
1385 | int flags = 0; | |
1386 | ||
1387 | if (stat(td->file_name, &st) == -1) { | |
1388 | if (errno != ENOENT) { | |
1389 | td_verror(td, errno); | |
1390 | return 1; | |
1391 | } | |
1392 | if (!td->create_file) { | |
1393 | td_verror(td, ENOENT); | |
1394 | return 1; | |
1395 | } | |
1396 | if (create_file(td, td->file_size, 0)) | |
1397 | return 1; | |
1398 | } else if (td->filetype == FIO_TYPE_FILE) { | |
6a0106a0 | 1399 | if (st.st_size < (off_t) td->file_size) { |
ebac4655 JA |
1400 | if (create_file(td, td->file_size - st.st_size, 1)) |
1401 | return 1; | |
1402 | } | |
1403 | } | |
1404 | ||
1405 | if (td->odirect) | |
1406 | flags |= O_DIRECT; | |
1407 | ||
3d60d1ed | 1408 | if (td_write(td) || td_rw(td)) { |
ebac4655 JA |
1409 | if (td->filetype == FIO_TYPE_FILE) { |
1410 | if (!td->overwrite) | |
1411 | flags |= O_TRUNC; | |
1412 | ||
1413 | flags |= O_CREAT; | |
1414 | } | |
1415 | if (td->sync_io) | |
1416 | flags |= O_SYNC; | |
1417 | ||
1418 | flags |= O_RDWR; | |
1419 | ||
1420 | td->fd = open(td->file_name, flags, 0600); | |
3d60d1ed JA |
1421 | } else { |
1422 | if (td->filetype == FIO_TYPE_CHAR) | |
1423 | flags |= O_RDWR; | |
1424 | else | |
1425 | flags |= O_RDONLY; | |
1426 | ||
1427 | td->fd = open(td->file_name, flags); | |
ebac4655 JA |
1428 | } |
1429 | ||
1430 | if (td->fd == -1) { | |
1431 | td_verror(td, errno); | |
1432 | return 1; | |
1433 | } | |
1434 | ||
1435 | if (get_file_size(td)) | |
1436 | return 1; | |
1437 | ||
1438 | if (td->io_engine != FIO_MMAPIO) | |
1439 | return setup_file_plain(td); | |
1440 | else | |
1441 | return setup_file_mmap(td); | |
1442 | } | |
1443 | ||
1444 | static int check_dev_match(dev_t dev, char *path) | |
1445 | { | |
1446 | unsigned int major, minor; | |
1447 | char line[256], *p; | |
1448 | FILE *f; | |
1449 | ||
1450 | f = fopen(path, "r"); | |
1451 | if (!f) { | |
1452 | perror("open path"); | |
1453 | return 1; | |
1454 | } | |
1455 | ||
1456 | p = fgets(line, sizeof(line), f); | |
1457 | if (!p) { | |
1458 | fclose(f); | |
1459 | return 1; | |
1460 | } | |
1461 | ||
1462 | if (sscanf(p, "%u:%u", &major, &minor) != 2) { | |
1463 | fclose(f); | |
1464 | return 1; | |
1465 | } | |
1466 | ||
1467 | if (((major << 8) | minor) == dev) { | |
1468 | fclose(f); | |
1469 | return 0; | |
1470 | } | |
1471 | ||
1472 | fclose(f); | |
1473 | return 1; | |
1474 | } | |
1475 | ||
1476 | static int find_block_dir(dev_t dev, char *path) | |
1477 | { | |
1478 | struct dirent *dir; | |
1479 | struct stat st; | |
1480 | int found = 0; | |
1481 | DIR *D; | |
1482 | ||
1483 | D = opendir(path); | |
1484 | if (!D) | |
1485 | return 0; | |
1486 | ||
1487 | while ((dir = readdir(D)) != NULL) { | |
1488 | char full_path[256]; | |
1489 | ||
1490 | if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, "..")) | |
1491 | continue; | |
1492 | if (!strcmp(dir->d_name, "device")) | |
1493 | continue; | |
1494 | ||
1495 | sprintf(full_path, "%s/%s", path, dir->d_name); | |
1496 | ||
1497 | if (!strcmp(dir->d_name, "dev")) { | |
1498 | if (!check_dev_match(dev, full_path)) { | |
1499 | found = 1; | |
1500 | break; | |
1501 | } | |
1502 | } | |
1503 | ||
1504 | if (stat(full_path, &st) == -1) { | |
1505 | perror("stat"); | |
1506 | break; | |
1507 | } | |
1508 | ||
1509 | if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode)) | |
1510 | continue; | |
1511 | ||
1512 | found = find_block_dir(dev, full_path); | |
1513 | if (found) { | |
1514 | strcpy(path, full_path); | |
1515 | break; | |
1516 | } | |
1517 | } | |
1518 | ||
1519 | closedir(D); | |
1520 | return found; | |
1521 | } | |
1522 | ||
1523 | static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus) | |
1524 | { | |
1525 | unsigned in_flight; | |
1526 | char line[256]; | |
1527 | FILE *f; | |
1528 | char *p; | |
1529 | ||
1530 | f = fopen(du->path, "r"); | |
1531 | if (!f) | |
1532 | return 1; | |
1533 | ||
1534 | p = fgets(line, sizeof(line), f); | |
1535 | if (!p) { | |
1536 | fclose(f); | |
1537 | return 1; | |
1538 | } | |
1539 | ||
1540 | if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) { | |
1541 | fclose(f); | |
1542 | return 1; | |
1543 | } | |
1544 | ||
1545 | fclose(f); | |
1546 | return 0; | |
1547 | } | |
1548 | ||
1549 | static void update_io_tick_disk(struct disk_util *du) | |
1550 | { | |
1551 | struct disk_util_stat __dus, *dus, *ldus; | |
1552 | struct timeval t; | |
1553 | ||
1554 | if (get_io_ticks(du, &__dus)) | |
1555 | return; | |
1556 | ||
1557 | dus = &du->dus; | |
1558 | ldus = &du->last_dus; | |
1559 | ||
1560 | dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]); | |
1561 | dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]); | |
1562 | dus->ios[0] += (__dus.ios[0] - ldus->ios[0]); | |
1563 | dus->ios[1] += (__dus.ios[1] - ldus->ios[1]); | |
1564 | dus->merges[0] += (__dus.merges[0] - ldus->merges[0]); | |
1565 | dus->merges[1] += (__dus.merges[1] - ldus->merges[1]); | |
1566 | dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]); | |
1567 | dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]); | |
1568 | dus->io_ticks += (__dus.io_ticks - ldus->io_ticks); | |
1569 | dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue); | |
1570 | ||
1571 | gettimeofday(&t, NULL); | |
1572 | du->msec += mtime_since(&du->time, &t); | |
1573 | memcpy(&du->time, &t, sizeof(t)); | |
1574 | memcpy(ldus, &__dus, sizeof(__dus)); | |
1575 | } | |
1576 | ||
1577 | static void update_io_ticks(void) | |
1578 | { | |
1579 | struct list_head *entry; | |
1580 | struct disk_util *du; | |
1581 | ||
1582 | list_for_each(entry, &disk_list) { | |
1583 | du = list_entry(entry, struct disk_util, list); | |
1584 | update_io_tick_disk(du); | |
1585 | } | |
1586 | } | |
1587 | ||
1588 | static int disk_util_exists(dev_t dev) | |
1589 | { | |
1590 | struct list_head *entry; | |
1591 | struct disk_util *du; | |
1592 | ||
1593 | list_for_each(entry, &disk_list) { | |
1594 | du = list_entry(entry, struct disk_util, list); | |
1595 | ||
1596 | if (du->dev == dev) | |
1597 | return 1; | |
1598 | } | |
1599 | ||
1600 | return 0; | |
1601 | } | |
1602 | ||
1603 | static void disk_util_add(dev_t dev, char *path) | |
1604 | { | |
1605 | struct disk_util *du = malloc(sizeof(*du)); | |
1606 | ||
1607 | memset(du, 0, sizeof(*du)); | |
1608 | INIT_LIST_HEAD(&du->list); | |
1609 | sprintf(du->path, "%s/stat", path); | |
1610 | du->name = strdup(basename(path)); | |
1611 | du->dev = dev; | |
1612 | ||
1613 | gettimeofday(&du->time, NULL); | |
1614 | get_io_ticks(du, &du->last_dus); | |
1615 | ||
1616 | list_add_tail(&du->list, &disk_list); | |
1617 | } | |
1618 | ||
1619 | static void init_disk_util(struct thread_data *td) | |
1620 | { | |
1621 | struct stat st; | |
1622 | char foo[256], tmp[256]; | |
1623 | dev_t dev; | |
1624 | char *p; | |
1625 | ||
1626 | if (!td->do_disk_util) | |
1627 | return; | |
1628 | ||
1629 | if (!stat(td->file_name, &st)) { | |
1630 | if (S_ISBLK(st.st_mode)) | |
1631 | dev = st.st_rdev; | |
1632 | else | |
1633 | dev = st.st_dev; | |
1634 | } else { | |
1635 | /* | |
1636 | * must be a file, open "." in that path | |
1637 | */ | |
1638 | strcpy(foo, td->file_name); | |
1639 | p = dirname(foo); | |
1640 | if (stat(p, &st)) { | |
1641 | perror("disk util stat"); | |
1642 | return; | |
1643 | } | |
1644 | ||
1645 | dev = st.st_dev; | |
1646 | } | |
1647 | ||
1648 | if (disk_util_exists(dev)) | |
1649 | return; | |
1650 | ||
1651 | sprintf(foo, "/sys/block"); | |
1652 | if (!find_block_dir(dev, foo)) | |
1653 | return; | |
1654 | ||
1655 | /* | |
64cb3e10 | 1656 | * for md/dm, there's no queue dir. we already have the right place |
ebac4655 | 1657 | */ |
64cb3e10 | 1658 | sprintf(tmp, "%s/stat", foo); |
ebac4655 | 1659 | if (stat(tmp, &st)) { |
64cb3e10 JA |
1660 | /* |
1661 | * if this is inside a partition dir, jump back to parent | |
1662 | */ | |
1663 | sprintf(tmp, "%s/queue", foo); | |
ebac4655 | 1664 | if (stat(tmp, &st)) { |
64cb3e10 JA |
1665 | p = dirname(foo); |
1666 | sprintf(tmp, "%s/queue", p); | |
1667 | if (stat(tmp, &st)) { | |
1668 | fprintf(stderr, "unknown sysfs layout\n"); | |
1669 | return; | |
1670 | } | |
1671 | sprintf(foo, "%s", p); | |
ebac4655 | 1672 | } |
ebac4655 JA |
1673 | } |
1674 | ||
1675 | disk_util_add(dev, foo); | |
1676 | } | |
1677 | ||
1678 | static void disk_util_timer_arm(void) | |
1679 | { | |
1680 | itimer.it_value.tv_sec = 0; | |
1681 | itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000; | |
1682 | setitimer(ITIMER_REAL, &itimer, NULL); | |
1683 | } | |
1684 | ||
1685 | static void clear_io_state(struct thread_data *td) | |
1686 | { | |
1687 | if (td->io_engine == FIO_SYNCIO) | |
1688 | lseek(td->fd, SEEK_SET, 0); | |
1689 | ||
20dc95c4 | 1690 | td->last_pos = 0; |
ebac4655 JA |
1691 | td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0; |
1692 | td->this_io_bytes[0] = td->this_io_bytes[1] = 0; | |
20dc95c4 | 1693 | td->zone_bytes = 0; |
ebac4655 JA |
1694 | |
1695 | if (td->file_map) | |
1696 | memset(td->file_map, 0, td->num_maps * sizeof(long)); | |
1697 | } | |
1698 | ||
1699 | static void update_rusage_stat(struct thread_data *td) | |
1700 | { | |
1701 | if (!(td->runtime[0] + td->runtime[1])) | |
1702 | return; | |
1703 | ||
1704 | getrusage(RUSAGE_SELF, &td->ru_end); | |
1705 | ||
1706 | td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime); | |
1707 | td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime); | |
1708 | td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw); | |
1709 | ||
1710 | ||
1711 | memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end)); | |
1712 | } | |
1713 | ||
1714 | static void *thread_main(void *data) | |
1715 | { | |
1716 | struct thread_data *td = data; | |
1717 | int ret = 1; | |
1718 | ||
1719 | if (!td->use_thread) | |
1720 | setsid(); | |
1721 | ||
1722 | td->pid = getpid(); | |
1723 | ||
1724 | if (init_io_u(td)) | |
1725 | goto err; | |
1726 | ||
1727 | if (fio_setaffinity(td) == -1) { | |
1728 | td_verror(td, errno); | |
1729 | goto err; | |
1730 | } | |
1731 | ||
1732 | if (init_io(td)) | |
1733 | goto err; | |
1734 | ||
1735 | if (td->ioprio) { | |
1736 | if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) { | |
1737 | td_verror(td, errno); | |
1738 | goto err; | |
1739 | } | |
1740 | } | |
1741 | ||
1742 | sem_post(&startup_sem); | |
1743 | sem_wait(&td->mutex); | |
1744 | ||
1745 | if (!td->create_serialize && setup_file(td)) | |
1746 | goto err; | |
1747 | ||
1748 | if (init_random_state(td)) | |
1749 | goto err; | |
1750 | ||
1751 | gettimeofday(&td->epoch, NULL); | |
1752 | ||
1753 | while (td->loops--) { | |
1754 | getrusage(RUSAGE_SELF, &td->ru_start); | |
1755 | gettimeofday(&td->start, NULL); | |
1756 | memcpy(&td->stat_sample_time, &td->start, sizeof(td->start)); | |
1757 | ||
1758 | if (td->ratemin) | |
1759 | memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate)); | |
1760 | ||
1761 | clear_io_state(td); | |
1762 | prune_io_piece_log(td); | |
1763 | ||
1764 | do_io(td); | |
1765 | ||
1766 | td->runtime[td->ddir] += mtime_since_now(&td->start); | |
3d60d1ed JA |
1767 | if (td_rw(td)) |
1768 | td->runtime[td->ddir ^ 1] = td->runtime[td->ddir]; | |
1769 | ||
ebac4655 JA |
1770 | update_rusage_stat(td); |
1771 | ||
1772 | if (td->error || td->terminate) | |
1773 | break; | |
1774 | ||
1775 | if (td->verify == VERIFY_NONE) | |
1776 | continue; | |
1777 | ||
1778 | clear_io_state(td); | |
1779 | gettimeofday(&td->start, NULL); | |
1780 | ||
1781 | do_verify(td); | |
1782 | ||
1783 | td->runtime[DDIR_READ] += mtime_since_now(&td->start); | |
1784 | ||
1785 | if (td->error || td->terminate) | |
1786 | break; | |
1787 | } | |
1788 | ||
1789 | ret = 0; | |
1790 | ||
1791 | if (td->bw_log) | |
1792 | finish_log(td, td->bw_log, "bw"); | |
1793 | if (td->slat_log) | |
1794 | finish_log(td, td->slat_log, "slat"); | |
1795 | if (td->clat_log) | |
1796 | finish_log(td, td->clat_log, "clat"); | |
1797 | ||
1798 | if (exitall_on_terminate) | |
1799 | terminate_threads(td->groupid); | |
1800 | ||
1801 | err: | |
1802 | if (td->fd != -1) { | |
1803 | close(td->fd); | |
1804 | td->fd = -1; | |
1805 | } | |
1806 | if (td->mmap) | |
1807 | munmap(td->mmap, td->file_size); | |
1808 | cleanup_io(td); | |
1809 | cleanup_io_u(td); | |
1810 | if (ret) { | |
1811 | sem_post(&startup_sem); | |
1812 | sem_wait(&td->mutex); | |
1813 | } | |
1814 | td_set_runstate(td, TD_EXITED); | |
1815 | return NULL; | |
1816 | ||
1817 | } | |
1818 | ||
1819 | static void *fork_main(int shmid, int offset) | |
1820 | { | |
1821 | struct thread_data *td; | |
1822 | void *data; | |
1823 | ||
1824 | data = shmat(shmid, NULL, 0); | |
1825 | if (data == (void *) -1) { | |
1826 | perror("shmat"); | |
1827 | return NULL; | |
1828 | } | |
1829 | ||
1830 | td = data + offset * sizeof(struct thread_data); | |
1831 | thread_main(td); | |
1832 | shmdt(data); | |
1833 | return NULL; | |
1834 | } | |
1835 | ||
1836 | static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, | |
1837 | double *mean, double *dev) | |
1838 | { | |
1839 | double n; | |
1840 | ||
1841 | if (is->samples == 0) | |
1842 | return 0; | |
1843 | ||
1844 | *min = is->min_val; | |
1845 | *max = is->max_val; | |
1846 | ||
1847 | n = (double) is->samples; | |
1848 | *mean = (double) is->val / n; | |
1849 | *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1)); | |
1850 | if (!(*min + *max) && !(*mean + *dev)) | |
1851 | return 0; | |
1852 | ||
1853 | return 1; | |
1854 | } | |
1855 | ||
1856 | static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs, | |
1857 | int ddir) | |
1858 | { | |
1859 | char *ddir_str[] = { "read ", "write" }; | |
1860 | unsigned long min, max, bw; | |
1861 | double mean, dev; | |
1862 | ||
1863 | if (!td->runtime[ddir]) | |
1864 | return; | |
1865 | ||
1866 | bw = td->io_bytes[ddir] / td->runtime[ddir]; | |
9104f874 | 1867 | printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]); |
ebac4655 JA |
1868 | |
1869 | if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev)) | |
1870 | printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev); | |
1871 | ||
1872 | if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev)) | |
1873 | printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev); | |
1874 | ||
1875 | if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) { | |
1876 | double p_of_agg; | |
1877 | ||
1878 | p_of_agg = mean * 100 / (double) rs->agg[ddir]; | |
1879 | 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); | |
1880 | } | |
1881 | } | |
1882 | ||
1883 | static void show_thread_status(struct thread_data *td, | |
1884 | struct group_run_stats *rs) | |
1885 | { | |
1886 | double usr_cpu, sys_cpu; | |
1887 | ||
1888 | if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error) | |
1889 | return; | |
1890 | ||
1891 | printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error); | |
1892 | ||
1893 | show_ddir_status(td, rs, td->ddir); | |
1894 | show_ddir_status(td, rs, td->ddir ^ 1); | |
1895 | ||
1896 | if (td->runtime[0] + td->runtime[1]) { | |
1897 | double runt = td->runtime[0] + td->runtime[1]; | |
1898 | ||
1899 | usr_cpu = (double) td->usr_time * 100 / runt; | |
1900 | sys_cpu = (double) td->sys_time * 100 / runt; | |
1901 | } else { | |
1902 | usr_cpu = 0; | |
1903 | sys_cpu = 0; | |
1904 | } | |
1905 | ||
1906 | printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx); | |
1907 | } | |
1908 | ||
1909 | static void check_str_update(struct thread_data *td) | |
1910 | { | |
1911 | char c = run_str[td->thread_number - 1]; | |
1912 | ||
1913 | if (td->runstate == td->old_runstate) | |
1914 | return; | |
1915 | ||
1916 | switch (td->runstate) { | |
1917 | case TD_REAPED: | |
1918 | c = '_'; | |
1919 | break; | |
1920 | case TD_EXITED: | |
1921 | c = 'E'; | |
1922 | break; | |
1923 | case TD_RUNNING: | |
3d60d1ed JA |
1924 | if (td_rw(td)) { |
1925 | if (td->sequential) | |
1926 | c = 'M'; | |
1927 | else | |
1928 | c = 'm'; | |
1929 | } else if (td_read(td)) { | |
ebac4655 JA |
1930 | if (td->sequential) |
1931 | c = 'R'; | |
1932 | else | |
1933 | c = 'r'; | |
1934 | } else { | |
1935 | if (td->sequential) | |
1936 | c = 'W'; | |
1937 | else | |
1938 | c = 'w'; | |
1939 | } | |
1940 | break; | |
1941 | case TD_VERIFYING: | |
1942 | c = 'V'; | |
1943 | break; | |
1944 | case TD_CREATED: | |
1945 | c = 'C'; | |
1946 | break; | |
1947 | case TD_NOT_CREATED: | |
1948 | c = 'P'; | |
1949 | break; | |
1950 | default: | |
1951 | printf("state %d\n", td->runstate); | |
1952 | } | |
1953 | ||
1954 | run_str[td->thread_number - 1] = c; | |
1955 | td->old_runstate = td->runstate; | |
1956 | } | |
1957 | ||
5289b847 JA |
1958 | static void eta_to_str(char *str, int eta_sec) |
1959 | { | |
1960 | unsigned int d, h, m, s; | |
1961 | static int always_d, always_h; | |
1962 | ||
1963 | d = h = m = s = 0; | |
1964 | ||
1965 | s = eta_sec % 60; | |
1966 | eta_sec /= 60; | |
1967 | m = eta_sec % 60; | |
1968 | eta_sec /= 60; | |
1969 | h = eta_sec % 24; | |
1970 | eta_sec /= 24; | |
1971 | d = eta_sec; | |
1972 | ||
1973 | if (d || always_d) { | |
1974 | always_d = 1; | |
1975 | str += sprintf(str, "%02dd:", d); | |
1976 | } | |
1977 | if (h || always_h) { | |
1978 | always_h = 1; | |
1979 | str += sprintf(str, "%02dh:", h); | |
1980 | } | |
1981 | ||
1982 | str += sprintf(str, "%02dm:", m); | |
1983 | str += sprintf(str, "%02ds", s); | |
1984 | } | |
1985 | ||
6a0106a0 JA |
1986 | static int thread_eta(struct thread_data *td, unsigned long elapsed) |
1987 | { | |
1988 | unsigned long long bytes_total, bytes_done; | |
1989 | unsigned int eta_sec = 0; | |
1990 | ||
1991 | bytes_total = td->total_io_size; | |
3d60d1ed JA |
1992 | |
1993 | /* | |
1994 | * if writing, bytes_total will be twice the size. If mixing, | |
1995 | * assume a 50/50 split and thus bytes_total will be 50% larger. | |
1996 | */ | |
1997 | if (td->verify) { | |
1998 | if (td_rw(td)) | |
1999 | bytes_total = bytes_total * 3 / 2; | |
2000 | else | |
2001 | bytes_total <<= 1; | |
2002 | } | |
6a0106a0 JA |
2003 | if (td->zone_size && td->zone_skip) |
2004 | bytes_total /= (td->zone_skip / td->zone_size); | |
2005 | ||
2006 | if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) { | |
2007 | double perc; | |
8b611c34 | 2008 | |
6a0106a0 JA |
2009 | bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]; |
2010 | perc = (double) bytes_done / (double) bytes_total; | |
2011 | if (perc > 1.0) | |
2012 | perc = 1.0; | |
2013 | ||
2014 | eta_sec = (elapsed * (1.0 / perc)) - elapsed; | |
2015 | ||
8b611c34 | 2016 | if (td->timeout && eta_sec > (td->timeout - elapsed)) |
6a0106a0 JA |
2017 | eta_sec = td->timeout - elapsed; |
2018 | } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) { | |
2019 | int t_eta = 0, r_eta = 0; | |
2020 | ||
2021 | /* | |
2022 | * We can only guess - assume it'll run the full timeout | |
2023 | * if given, otherwise assume it'll run at the specified rate. | |
2024 | */ | |
2025 | if (td->timeout) | |
2026 | t_eta = td->timeout + td->start_delay - elapsed; | |
2027 | if (td->rate) { | |
2028 | r_eta = (bytes_total / 1024) / td->rate; | |
2029 | r_eta += td->start_delay - elapsed; | |
2030 | } | |
2031 | ||
2032 | if (r_eta && t_eta) | |
2033 | eta_sec = min(r_eta, t_eta); | |
2034 | else if (r_eta) | |
2035 | eta_sec = r_eta; | |
2036 | else if (t_eta) | |
2037 | eta_sec = t_eta; | |
2038 | else | |
2039 | eta_sec = INT_MAX; | |
2040 | } else { | |
2041 | /* | |
2042 | * thread is already done | |
2043 | */ | |
2044 | eta_sec = 0; | |
2045 | } | |
2046 | ||
2047 | return eta_sec; | |
2048 | } | |
2049 | ||
ebac4655 JA |
2050 | static void print_thread_status(void) |
2051 | { | |
6a0106a0 JA |
2052 | unsigned long elapsed = time_since_now(&genesis); |
2053 | int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec; | |
5289b847 | 2054 | char eta_str[32]; |
6a0106a0 JA |
2055 | double perc = 0.0; |
2056 | ||
2057 | eta_secs = malloc(thread_number * sizeof(int)); | |
2058 | memset(eta_secs, 0, thread_number * sizeof(int)); | |
ebac4655 | 2059 | |
ebac4655 JA |
2060 | nr_running = t_rate = m_rate = 0; |
2061 | for (i = 0; i < thread_number; i++) { | |
2062 | struct thread_data *td = &threads[i]; | |
2063 | ||
2064 | if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){ | |
2065 | nr_running++; | |
2066 | t_rate += td->rate; | |
2067 | m_rate += td->ratemin; | |
2068 | } | |
2069 | ||
6a0106a0 JA |
2070 | if (elapsed >= 3) |
2071 | eta_secs[i] = thread_eta(td, elapsed); | |
8b611c34 JA |
2072 | else |
2073 | eta_secs[i] = INT_MAX; | |
ebac4655 JA |
2074 | |
2075 | check_str_update(td); | |
2076 | } | |
2077 | ||
6a0106a0 JA |
2078 | if (exitall_on_terminate) |
2079 | eta_sec = INT_MAX; | |
2080 | else | |
2081 | eta_sec = 0; | |
5289b847 | 2082 | |
6a0106a0 JA |
2083 | for (i = 0; i < thread_number; i++) { |
2084 | if (exitall_on_terminate) { | |
2085 | if (eta_secs[i] < eta_sec) | |
2086 | eta_sec = eta_secs[i]; | |
2087 | } else { | |
2088 | if (eta_secs[i] > eta_sec) | |
2089 | eta_sec = eta_secs[i]; | |
5289b847 | 2090 | } |
6a0106a0 | 2091 | } |
5289b847 | 2092 | |
8b611c34 | 2093 | if (eta_sec != INT_MAX && elapsed) { |
6a0106a0 JA |
2094 | perc = (double) elapsed / (double) (elapsed + eta_sec); |
2095 | eta_to_str(eta_str, eta_sec); | |
ebac4655 JA |
2096 | } |
2097 | ||
5289b847 | 2098 | printf("Threads now running (%d)", nr_running); |
ebac4655 JA |
2099 | if (m_rate || t_rate) |
2100 | printf(", commitrate %d/%dKiB/sec", t_rate, m_rate); | |
8b611c34 | 2101 | if (eta_sec != INT_MAX) { |
6a0106a0 JA |
2102 | perc *= 100.0; |
2103 | printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str); | |
2104 | } | |
5289b847 | 2105 | printf("\r"); |
ebac4655 | 2106 | fflush(stdout); |
6a0106a0 | 2107 | free(eta_secs); |
ebac4655 JA |
2108 | } |
2109 | ||
2110 | static void reap_threads(int *nr_running, int *t_rate, int *m_rate) | |
2111 | { | |
2112 | int i; | |
2113 | ||
2114 | /* | |
2115 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
2116 | */ | |
2117 | for (i = 0; i < thread_number; i++) { | |
2118 | struct thread_data *td = &threads[i]; | |
2119 | ||
2120 | if (td->runstate != TD_EXITED) | |
2121 | continue; | |
2122 | ||
2123 | td_set_runstate(td, TD_REAPED); | |
2124 | ||
2125 | if (td->use_thread) { | |
2126 | long ret; | |
2127 | ||
2128 | if (pthread_join(td->thread, (void *) &ret)) | |
2129 | perror("thread_join"); | |
2130 | } else | |
2131 | waitpid(td->pid, NULL, 0); | |
2132 | ||
2133 | (*nr_running)--; | |
2134 | (*m_rate) -= td->ratemin; | |
2135 | (*t_rate) -= td->rate; | |
2136 | } | |
2137 | } | |
2138 | ||
2139 | static void run_threads(void) | |
2140 | { | |
ebac4655 JA |
2141 | struct thread_data *td; |
2142 | unsigned long spent; | |
2143 | int i, todo, nr_running, m_rate, t_rate, nr_started; | |
2144 | ||
2145 | printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : ""); | |
2146 | fflush(stdout); | |
2147 | ||
2148 | signal(SIGINT, sig_handler); | |
2149 | signal(SIGALRM, sig_handler); | |
2150 | ||
2151 | todo = thread_number; | |
2152 | nr_running = 0; | |
2153 | nr_started = 0; | |
2154 | m_rate = t_rate = 0; | |
2155 | ||
2156 | for (i = 0; i < thread_number; i++) { | |
2157 | td = &threads[i]; | |
2158 | ||
2159 | run_str[td->thread_number - 1] = 'P'; | |
2160 | ||
2161 | init_disk_util(td); | |
2162 | ||
2163 | if (!td->create_serialize) | |
2164 | continue; | |
2165 | ||
2166 | /* | |
2167 | * do file setup here so it happens sequentially, | |
2168 | * we don't want X number of threads getting their | |
2169 | * client data interspersed on disk | |
2170 | */ | |
2171 | if (setup_file(td)) { | |
2172 | td_set_runstate(td, TD_REAPED); | |
2173 | todo--; | |
2174 | } | |
2175 | } | |
2176 | ||
2177 | gettimeofday(&genesis, NULL); | |
2178 | ||
2179 | while (todo) { | |
2180 | /* | |
2181 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
2182 | */ | |
2183 | for (i = 0; i < thread_number; i++) { | |
2184 | td = &threads[i]; | |
2185 | ||
2186 | if (td->runstate != TD_NOT_CREATED) | |
2187 | continue; | |
2188 | ||
2189 | /* | |
2190 | * never got a chance to start, killed by other | |
2191 | * thread for some reason | |
2192 | */ | |
2193 | if (td->terminate) { | |
2194 | todo--; | |
2195 | continue; | |
2196 | } | |
2197 | ||
2198 | if (td->start_delay) { | |
2199 | spent = mtime_since_now(&genesis); | |
2200 | ||
2201 | if (td->start_delay * 1000 > spent) | |
2202 | continue; | |
2203 | } | |
2204 | ||
2205 | if (td->stonewall && (nr_started || nr_running)) | |
2206 | break; | |
2207 | ||
2208 | td_set_runstate(td, TD_CREATED); | |
2209 | sem_init(&startup_sem, 0, 1); | |
2210 | todo--; | |
2211 | nr_started++; | |
2212 | ||
2213 | if (td->use_thread) { | |
2214 | if (pthread_create(&td->thread, NULL, thread_main, td)) { | |
2215 | perror("thread_create"); | |
2216 | nr_started--; | |
2217 | } | |
2218 | } else { | |
2219 | if (fork()) | |
2220 | sem_wait(&startup_sem); | |
2221 | else { | |
2222 | fork_main(shm_id, i); | |
2223 | exit(0); | |
2224 | } | |
2225 | } | |
2226 | } | |
2227 | ||
2228 | /* | |
2229 | * start created threads (TD_CREATED -> TD_RUNNING) | |
2230 | */ | |
2231 | for (i = 0; i < thread_number; i++) { | |
2232 | td = &threads[i]; | |
2233 | ||
2234 | if (td->runstate != TD_CREATED) | |
2235 | continue; | |
2236 | ||
2237 | td_set_runstate(td, TD_RUNNING); | |
2238 | nr_running++; | |
2239 | nr_started--; | |
2240 | m_rate += td->ratemin; | |
2241 | t_rate += td->rate; | |
2242 | sem_post(&td->mutex); | |
2243 | } | |
2244 | ||
2245 | reap_threads(&nr_running, &t_rate, &m_rate); | |
2246 | ||
2247 | if (todo) | |
2248 | usleep(100000); | |
2249 | } | |
2250 | ||
2251 | while (nr_running) { | |
2252 | reap_threads(&nr_running, &t_rate, &m_rate); | |
2253 | usleep(10000); | |
2254 | } | |
2255 | ||
2256 | update_io_ticks(); | |
2257 | } | |
2258 | ||
2259 | static void show_group_stats(struct group_run_stats *rs, int id) | |
2260 | { | |
2261 | printf("\nRun status group %d (all jobs):\n", id); | |
2262 | ||
2263 | if (rs->max_run[DDIR_READ]) | |
9104f874 | 2264 | printf(" READ: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]); |
ebac4655 | 2265 | if (rs->max_run[DDIR_WRITE]) |
9104f874 | 2266 | printf(" WRITE: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]); |
ebac4655 JA |
2267 | } |
2268 | ||
2269 | static void show_disk_util(void) | |
2270 | { | |
2271 | struct disk_util_stat *dus; | |
2272 | struct list_head *entry; | |
2273 | struct disk_util *du; | |
2274 | double util; | |
2275 | ||
2276 | printf("\nDisk stats (read/write):\n"); | |
2277 | ||
2278 | list_for_each(entry, &disk_list) { | |
2279 | du = list_entry(entry, struct disk_util, list); | |
2280 | dus = &du->dus; | |
2281 | ||
2282 | util = (double) 100 * du->dus.io_ticks / (double) du->msec; | |
2283 | if (util > 100.0) | |
2284 | util = 100.0; | |
2285 | ||
2286 | printf(" %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util); | |
2287 | } | |
2288 | } | |
2289 | ||
2290 | static void show_run_stats(void) | |
2291 | { | |
2292 | struct group_run_stats *runstats, *rs; | |
2293 | struct thread_data *td; | |
2294 | int i; | |
2295 | ||
2296 | runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); | |
2297 | ||
2298 | for (i = 0; i < groupid + 1; i++) { | |
2299 | rs = &runstats[i]; | |
2300 | ||
2301 | memset(rs, 0, sizeof(*rs)); | |
2302 | rs->min_bw[0] = rs->min_run[0] = ~0UL; | |
2303 | rs->min_bw[1] = rs->min_run[1] = ~0UL; | |
2304 | } | |
2305 | ||
2306 | for (i = 0; i < thread_number; i++) { | |
2307 | unsigned long rbw, wbw; | |
2308 | ||
2309 | td = &threads[i]; | |
2310 | ||
2311 | if (td->error) { | |
2312 | printf("Client%d: %s\n", td->thread_number, td->verror); | |
2313 | continue; | |
2314 | } | |
2315 | ||
2316 | rs = &runstats[td->groupid]; | |
2317 | ||
2318 | if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0]) | |
2319 | rs->min_run[0] = td->runtime[0]; | |
2320 | if (td->runtime[0] > rs->max_run[0]) | |
2321 | rs->max_run[0] = td->runtime[0]; | |
2322 | if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1]) | |
2323 | rs->min_run[1] = td->runtime[1]; | |
2324 | if (td->runtime[1] > rs->max_run[1]) | |
2325 | rs->max_run[1] = td->runtime[1]; | |
2326 | ||
2327 | rbw = wbw = 0; | |
2328 | if (td->runtime[0]) | |
2329 | rbw = td->io_bytes[0] / td->runtime[0]; | |
2330 | if (td->runtime[1]) | |
2331 | wbw = td->io_bytes[1] / td->runtime[1]; | |
2332 | ||
2333 | if (rbw < rs->min_bw[0]) | |
2334 | rs->min_bw[0] = rbw; | |
2335 | if (wbw < rs->min_bw[1]) | |
2336 | rs->min_bw[1] = wbw; | |
2337 | if (rbw > rs->max_bw[0]) | |
2338 | rs->max_bw[0] = rbw; | |
2339 | if (wbw > rs->max_bw[1]) | |
2340 | rs->max_bw[1] = wbw; | |
2341 | ||
2342 | rs->io_mb[0] += td->io_bytes[0] >> 20; | |
2343 | rs->io_mb[1] += td->io_bytes[1] >> 20; | |
2344 | } | |
2345 | ||
2346 | for (i = 0; i < groupid + 1; i++) { | |
2347 | rs = &runstats[i]; | |
2348 | ||
2349 | if (rs->max_run[0]) | |
2350 | rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0]; | |
2351 | if (rs->max_run[1]) | |
2352 | rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1]; | |
2353 | } | |
2354 | ||
2355 | /* | |
2356 | * don't overwrite last signal output | |
2357 | */ | |
2358 | printf("\n"); | |
2359 | ||
2360 | for (i = 0; i < thread_number; i++) { | |
2361 | td = &threads[i]; | |
2362 | rs = &runstats[td->groupid]; | |
2363 | ||
2364 | show_thread_status(td, rs); | |
2365 | } | |
2366 | ||
2367 | for (i = 0; i < groupid + 1; i++) | |
2368 | show_group_stats(&runstats[i], i); | |
2369 | ||
2370 | show_disk_util(); | |
2371 | } | |
2372 | ||
2373 | int main(int argc, char *argv[]) | |
2374 | { | |
2375 | if (parse_options(argc, argv)) | |
2376 | return 1; | |
2377 | ||
2378 | if (!thread_number) { | |
2379 | printf("Nothing to do\n"); | |
2380 | return 1; | |
2381 | } | |
2382 | ||
2383 | disk_util_timer_arm(); | |
2384 | ||
2385 | run_threads(); | |
2386 | show_run_stats(); | |
2387 | ||
2388 | return 0; | |
2389 | } |