1 /*****************************************************************************
2 * RRDtool 1.3.2 Copyright by Tobi Oetiker, 1997-2008
3 *****************************************************************************
4 * rrd_open.c Open an RRD File
5 *****************************************************************************
6 * $Id$
7 *****************************************************************************/
9 #include "rrd_tool.h"
10 #include "unused.h"
11 #define MEMBLK 8192
13 /* DEBUG 2 prints information obtained via mincore(2) */
14 #define DEBUG 1
15 /* do not calculate exact madvise hints but assume 1 page for headers and
16 * set DONTNEED for the rest, which is assumed to be data */
17 /* Avoid calling madvise on areas that were already hinted. May be benefical if
18 * your syscalls are very slow */
20 #ifdef HAVE_MMAP
21 /* the cast to void* is there to avoid this warning seen on ia64 with certain
22 versions of gcc: 'cast increases required alignment of target type'
23 */
24 #define __rrd_read(dst, dst_t, cnt) { \
25 size_t wanted = sizeof(dst_t)*(cnt); \
26 if (offset + wanted > rrd_file->file_len) { \
27 rrd_set_error("reached EOF while loading header " #dst); \
28 goto out_nullify_head; \
29 } \
30 (dst) = (dst_t*)(void*) (data + offset); \
31 offset += wanted; \
32 }
33 #else
34 #define __rrd_read(dst, dst_t, cnt) { \
35 size_t wanted = sizeof(dst_t)*(cnt); \
36 size_t got; \
37 if ((dst = malloc(wanted)) == NULL) { \
38 rrd_set_error(#dst " malloc"); \
39 goto out_nullify_head; \
40 } \
41 got = read (rrd_file->fd, dst, wanted); \
42 if (got != wanted) { \
43 rrd_set_error("short read while reading header " #dst); \
44 goto out_nullify_head; \
45 } \
46 offset += got; \
47 }
48 #endif
50 /* get the address of the start of this page */
51 #if defined USE_MADVISE || defined HAVE_POSIX_FADVISE
52 #ifndef PAGE_START
53 #define PAGE_START(addr) ((addr)&(~(_page_size-1)))
54 #endif
55 #endif
57 /* Open a database file, return its header and an open filehandle,
58 * positioned to the first cdp in the first rra.
59 * In the error path of rrd_open, only rrd_free(&rrd) has to be called
60 * before returning an error. Do not call rrd_close upon failure of rrd_open.
61 */
63 rrd_file_t *rrd_open(
64 const char *const file_name,
65 rrd_t *rrd,
66 unsigned rdwr)
67 {
68 int flags = 0;
69 mode_t mode = S_IRUSR;
70 int version;
72 #ifdef HAVE_MMAP
73 ssize_t _page_size = sysconf(_SC_PAGESIZE);
74 int mm_prot = PROT_READ, mm_flags = 0;
75 char *data = MAP_FAILED;
76 #endif
77 off_t offset = 0;
78 struct stat statb;
79 rrd_file_t *rrd_file = NULL;
80 off_t newfile_size = 0;
82 if (rdwr & RRD_CREAT) {
83 /* yes bad inline signaling alert, we are using the
84 floatcookie to pass the size in ... only used in resize */
85 newfile_size = (off_t) rrd->stat_head->float_cookie;
86 free(rrd->stat_head);
87 }
88 rrd_init(rrd);
89 rrd_file = malloc(sizeof(rrd_file_t));
90 if (rrd_file == NULL) {
91 rrd_set_error("allocating rrd_file descriptor for '%s'", file_name);
92 return NULL;
93 }
94 memset(rrd_file, 0, sizeof(rrd_file_t));
96 #ifdef DEBUG
97 if ((rdwr & (RRD_READONLY | RRD_READWRITE)) ==
98 (RRD_READONLY | RRD_READWRITE)) {
99 /* Both READONLY and READWRITE were given, which is invalid. */
100 rrd_set_error("in read/write request mask");
101 exit(-1);
102 }
103 #endif
104 if (rdwr & RRD_READONLY) {
105 flags |= O_RDONLY;
106 #ifdef HAVE_MMAP
107 mm_flags = MAP_PRIVATE;
108 # ifdef MAP_NORESERVE
109 mm_flags |= MAP_NORESERVE; /* readonly, so no swap backing needed */
110 # endif
111 #endif
112 } else {
113 if (rdwr & RRD_READWRITE) {
114 mode |= S_IWUSR;
115 flags |= O_RDWR;
116 #ifdef HAVE_MMAP
117 mm_flags = MAP_SHARED;
118 mm_prot |= PROT_WRITE;
119 #endif
120 }
121 if (rdwr & RRD_CREAT) {
122 flags |= (O_CREAT | O_TRUNC);
123 }
124 }
125 if (rdwr & RRD_READAHEAD) {
126 #ifdef MAP_POPULATE
127 mm_flags |= MAP_POPULATE; /* populate ptes and data */
128 #endif
129 #if defined MAP_NONBLOCK
130 mm_flags |= MAP_NONBLOCK; /* just populate ptes */
131 #endif
132 }
133 #if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__CYGWIN32__)
134 flags |= O_BINARY;
135 #endif
137 if ((rrd_file->fd = open(file_name, flags, mode)) < 0) {
138 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
139 goto out_free;
140 }
142 /* Better try to avoid seeks as much as possible. stat may be heavy but
143 * many concurrent seeks are even worse. */
144 if (newfile_size == 0 && ((fstat(rrd_file->fd, &statb)) < 0)) {
145 rrd_set_error("fstat '%s': %s", file_name, rrd_strerror(errno));
146 goto out_close;
147 }
148 if (newfile_size == 0) {
149 rrd_file->file_len = statb.st_size;
150 } else {
151 rrd_file->file_len = newfile_size;
152 lseek(rrd_file->fd, newfile_size - 1, SEEK_SET);
153 write(rrd_file->fd, "\0", 1); /* poke */
154 lseek(rrd_file->fd, 0, SEEK_SET);
155 }
156 #ifdef HAVE_POSIX_FADVISE
157 /* In general we need no read-ahead when dealing with rrd_files.
158 When we stop reading, it is highly unlikely that we start up again.
159 In this manner we actually save time and diskaccess (and buffer cache).
160 Thanks to Dave Plonka for the Idea of using POSIX_FADV_RANDOM here. */
161 posix_fadvise(rrd_file->fd, 0, 0, POSIX_FADV_RANDOM);
162 #endif
164 /*
165 if (rdwr & RRD_READWRITE)
166 {
167 if (setvbuf((rrd_file->fd),NULL,_IONBF,2)) {
168 rrd_set_error("failed to disable the stream buffer\n");
169 return (-1);
170 }
171 }
172 */
173 #ifdef HAVE_MMAP
174 data = mmap(0, rrd_file->file_len, mm_prot, mm_flags,
175 rrd_file->fd, offset);
177 /* lets see if the first read worked */
178 if (data == MAP_FAILED) {
179 rrd_set_error("mmaping file '%s': %s", file_name,
180 rrd_strerror(errno));
181 goto out_close;
182 }
183 rrd_file->file_start = data;
184 if (rdwr & RRD_CREAT) {
185 memset(data, DNAN, newfile_size - 1);
186 goto out_done;
187 }
188 #endif
189 if (rdwr & RRD_CREAT)
190 goto out_done;
191 #ifdef USE_MADVISE
192 if (rdwr & RRD_COPY) {
193 /* We will read everything in a moment (copying) */
194 madvise(data, rrd_file->file_len, MADV_WILLNEED | MADV_SEQUENTIAL);
195 } else {
196 /* We do not need to read anything in for the moment */
197 madvise(data, rrd_file->file_len, MADV_RANDOM);
198 /* the stat_head will be needed soonish, so hint accordingly */
199 madvise(data, sizeof(stat_head_t), MADV_WILLNEED | MADV_RANDOM);
200 }
201 #endif
203 __rrd_read(rrd->stat_head, stat_head_t,
204 1);
206 /* lets do some test if we are on track ... */
207 if (memcmp(rrd->stat_head->cookie, RRD_COOKIE, sizeof(RRD_COOKIE)) != 0) {
208 rrd_set_error("'%s' is not an RRD file", file_name);
209 goto out_nullify_head;
210 }
212 if (rrd->stat_head->float_cookie != FLOAT_COOKIE) {
213 rrd_set_error("This RRD was created on another architecture");
214 goto out_nullify_head;
215 }
217 version = atoi(rrd->stat_head->version);
219 if (version > atoi(RRD_VERSION)) {
220 rrd_set_error("can't handle RRD file version %s",
221 rrd->stat_head->version);
222 goto out_nullify_head;
223 }
224 #if defined USE_MADVISE
225 /* the ds_def will be needed soonish, so hint accordingly */
226 madvise(data + PAGE_START(offset),
227 sizeof(ds_def_t) * rrd->stat_head->ds_cnt, MADV_WILLNEED);
228 #endif
229 __rrd_read(rrd->ds_def, ds_def_t,
230 rrd->stat_head->ds_cnt);
232 #if defined USE_MADVISE
233 /* the rra_def will be needed soonish, so hint accordingly */
234 madvise(data + PAGE_START(offset),
235 sizeof(rra_def_t) * rrd->stat_head->rra_cnt, MADV_WILLNEED);
236 #endif
237 __rrd_read(rrd->rra_def, rra_def_t,
238 rrd->stat_head->rra_cnt);
240 /* handle different format for the live_head */
241 if (version < 3) {
242 rrd->live_head = (live_head_t *) malloc(sizeof(live_head_t));
243 if (rrd->live_head == NULL) {
244 rrd_set_error("live_head_t malloc");
245 goto out_close;
246 }
247 #if defined USE_MADVISE
248 /* the live_head will be needed soonish, so hint accordingly */
249 madvise(data + PAGE_START(offset), sizeof(time_t), MADV_WILLNEED);
250 #endif
251 __rrd_read(rrd->legacy_last_up, time_t,
252 1);
254 rrd->live_head->last_up = *rrd->legacy_last_up;
255 rrd->live_head->last_up_usec = 0;
256 } else {
257 #if defined USE_MADVISE
258 /* the live_head will be needed soonish, so hint accordingly */
259 madvise(data + PAGE_START(offset),
260 sizeof(live_head_t), MADV_WILLNEED);
261 #endif
262 __rrd_read(rrd->live_head, live_head_t,
263 1);
264 }
265 __rrd_read(rrd->pdp_prep, pdp_prep_t,
266 rrd->stat_head->ds_cnt);
267 __rrd_read(rrd->cdp_prep, cdp_prep_t,
268 rrd->stat_head->rra_cnt * rrd->stat_head->ds_cnt);
269 __rrd_read(rrd->rra_ptr, rra_ptr_t,
270 rrd->stat_head->rra_cnt);
272 rrd_file->header_len = offset;
273 rrd_file->pos = offset;
275 {
276 unsigned long row_cnt = 0;
277 unsigned long i;
279 for (i=0; i<rrd->stat_head->rra_cnt; i++)
280 row_cnt += rrd->rra_def[i].row_cnt;
282 off_t correct_len = rrd_file->header_len +
283 sizeof(rrd_value_t) * row_cnt * rrd->stat_head->ds_cnt;
285 if (correct_len > rrd_file->file_len)
286 {
287 rrd_set_error("'%s' is too small (should be %ld bytes)",
288 file_name, (long long) correct_len);
289 goto out_nullify_head;
290 }
291 }
293 out_done:
294 return (rrd_file);
295 out_nullify_head:
296 rrd->stat_head = NULL;
297 out_close:
298 #ifdef HAVE_MMAP
299 if (data != MAP_FAILED)
300 munmap(data, rrd_file->file_len);
301 #endif
302 close(rrd_file->fd);
303 out_free:
304 free(rrd_file);
305 return NULL;
306 }
309 #if defined DEBUG && DEBUG > 1
310 /* Print list of in-core pages of a the current rrd_file. */
311 static
312 void mincore_print(
313 rrd_file_t *rrd_file,
314 char *mark)
315 {
316 #ifdef HAVE_MMAP
317 /* pretty print blocks in core */
318 off_t off;
319 unsigned char *vec;
320 ssize_t _page_size = sysconf(_SC_PAGESIZE);
322 off = rrd_file->file_len +
323 ((rrd_file->file_len + _page_size - 1) / _page_size);
324 vec = malloc(off);
325 if (vec != NULL) {
326 memset(vec, 0, off);
327 if (mincore(rrd_file->file_start, rrd_file->file_len, vec) == 0) {
328 int prev;
329 unsigned is_in = 0, was_in = 0;
331 for (off = 0, prev = 0; off < rrd_file->file_len; ++off) {
332 is_in = vec[off] & 1; /* if lsb set then is core resident */
333 if (off == 0)
334 was_in = is_in;
335 if (was_in != is_in) {
336 fprintf(stderr, "%s: %sin core: %p len %ld\n", mark,
337 was_in ? "" : "not ", vec + prev, off - prev);
338 was_in = is_in;
339 prev = off;
340 }
341 }
342 fprintf(stderr,
343 "%s: %sin core: %p len %ld\n", mark,
344 was_in ? "" : "not ", vec + prev, off - prev);
345 } else
346 fprintf(stderr, "mincore: %s", rrd_strerror(errno));
347 }
348 #else
349 fprintf(stderr, "sorry mincore only works with mmap");
350 #endif
351 }
352 #endif /* defined DEBUG && DEBUG > 1 */
355 /* drop cache except for the header and the active pages */
356 void rrd_dontneed(
357 rrd_file_t *rrd_file,
358 rrd_t *rrd)
359 {
360 #if defined USE_MADVISE || defined HAVE_POSIX_FADVISE
361 off_t dontneed_start;
362 off_t rra_start;
363 off_t active_block;
364 unsigned long i;
365 ssize_t _page_size = sysconf(_SC_PAGESIZE);
367 #if defined DEBUG && DEBUG > 1
368 mincore_print(rrd_file, "before");
369 #endif
371 /* ignoring errors from RRDs that are smaller then the file_len+rounding */
372 rra_start = rrd_file->header_len;
373 dontneed_start = PAGE_START(rra_start) + _page_size;
374 for (i = 0; i < rrd->stat_head->rra_cnt; ++i) {
375 active_block =
376 PAGE_START(rra_start
377 + rrd->rra_ptr[i].cur_row
378 * rrd->stat_head->ds_cnt * sizeof(rrd_value_t));
379 if (active_block > dontneed_start) {
380 #ifdef USE_MADVISE
381 madvise(rrd_file->file_start + dontneed_start,
382 active_block - dontneed_start - 1, MADV_DONTNEED);
383 #endif
384 /* in linux at least only fadvise DONTNEED seems to purge pages from cache */
385 #ifdef HAVE_POSIX_FADVISE
386 posix_fadvise(rrd_file->fd, dontneed_start,
387 active_block - dontneed_start - 1,
388 POSIX_FADV_DONTNEED);
389 #endif
390 }
391 dontneed_start = active_block;
392 /* do not release 'hot' block if update for this RAA will occur
393 * within 10 minutes */
394 if (rrd->stat_head->pdp_step * rrd->rra_def[i].pdp_cnt -
395 rrd->live_head->last_up % (rrd->stat_head->pdp_step *
396 rrd->rra_def[i].pdp_cnt) < 10 * 60) {
397 dontneed_start += _page_size;
398 }
399 rra_start +=
400 rrd->rra_def[i].row_cnt * rrd->stat_head->ds_cnt *
401 sizeof(rrd_value_t);
402 }
404 if (dontneed_start < rrd_file->file_len) {
405 #ifdef USE_MADVISE
406 madvise(rrd_file->file_start + dontneed_start,
407 rrd_file->file_len - dontneed_start, MADV_DONTNEED);
408 #endif
409 #ifdef HAVE_POSIX_FADVISE
410 posix_fadvise(rrd_file->fd, dontneed_start,
411 rrd_file->file_len - dontneed_start,
412 POSIX_FADV_DONTNEED);
413 #endif
414 }
416 #if defined DEBUG && DEBUG > 1
417 mincore_print(rrd_file, "after");
418 #endif
419 #endif /* without madvise and posix_fadvise ist does not make much sense todo anything */
420 }
426 int rrd_close(
427 rrd_file_t *rrd_file)
428 {
429 int ret;
431 #ifdef HAVE_MMAP
432 ret = msync(rrd_file->file_start, rrd_file->file_len, MS_ASYNC);
433 if (ret != 0)
434 rrd_set_error("msync rrd_file: %s", rrd_strerror(errno));
435 ret = munmap(rrd_file->file_start, rrd_file->file_len);
436 if (ret != 0)
437 rrd_set_error("munmap rrd_file: %s", rrd_strerror(errno));
438 #endif
439 ret = close(rrd_file->fd);
440 if (ret != 0)
441 rrd_set_error("closing file: %s", rrd_strerror(errno));
442 free(rrd_file);
443 rrd_file = NULL;
444 return ret;
445 }
448 /* Set position of rrd_file. */
450 off_t rrd_seek(
451 rrd_file_t *rrd_file,
452 off_t off,
453 int whence)
454 {
455 off_t ret = 0;
457 #ifdef HAVE_MMAP
458 if (whence == SEEK_SET)
459 rrd_file->pos = off;
460 else if (whence == SEEK_CUR)
461 rrd_file->pos += off;
462 else if (whence == SEEK_END)
463 rrd_file->pos = rrd_file->file_len + off;
464 #else
465 ret = lseek(rrd_file->fd, off, whence);
466 if (ret < 0)
467 rrd_set_error("lseek: %s", rrd_strerror(errno));
468 rrd_file->pos = ret;
469 #endif
470 /* mimic fseek, which returns 0 upon success */
471 return ret < 0; /*XXX: or just ret to mimic lseek */
472 }
475 /* Get current position in rrd_file. */
477 off_t rrd_tell(
478 rrd_file_t *rrd_file)
479 {
480 return rrd_file->pos;
481 }
484 /* Read count bytes into buffer buf, starting at rrd_file->pos.
485 * Returns the number of bytes read or <0 on error. */
487 ssize_t rrd_read(
488 rrd_file_t *rrd_file,
489 void *buf,
490 size_t count)
491 {
492 #ifdef HAVE_MMAP
493 size_t _cnt = count;
494 ssize_t _surplus;
496 if (rrd_file->pos > rrd_file->file_len || _cnt == 0) /* EOF */
497 return 0;
498 if (buf == NULL)
499 return -1; /* EINVAL */
500 _surplus = rrd_file->pos + _cnt - rrd_file->file_len;
501 if (_surplus > 0) { /* short read */
502 _cnt -= _surplus;
503 }
504 if (_cnt == 0)
505 return 0; /* EOF */
506 buf = memcpy(buf, rrd_file->file_start + rrd_file->pos, _cnt);
508 rrd_file->pos += _cnt; /* mimmic read() semantics */
509 return _cnt;
510 #else
511 ssize_t ret;
513 ret = read(rrd_file->fd, buf, count);
514 if (ret > 0)
515 rrd_file->pos += ret; /* mimmic read() semantics */
516 return ret;
517 #endif
518 }
521 /* Write count bytes from buffer buf to the current position
522 * rrd_file->pos of rrd_file->fd.
523 * Returns the number of bytes written or <0 on error. */
525 ssize_t rrd_write(
526 rrd_file_t *rrd_file,
527 const void *buf,
528 size_t count)
529 {
530 #ifdef HAVE_MMAP
531 if (count == 0)
532 return 0;
533 if (buf == NULL)
534 return -1; /* EINVAL */
535 memcpy(rrd_file->file_start + rrd_file->pos, buf, count);
536 rrd_file->pos += count;
537 return count; /* mimmic write() semantics */
538 #else
539 ssize_t _sz = write(rrd_file->fd, buf, count);
541 if (_sz > 0)
542 rrd_file->pos += _sz;
543 return _sz;
544 #endif
545 }
548 /* flush all data pending to be written to FD. */
550 void rrd_flush(
551 rrd_file_t *rrd_file)
552 {
553 if (fdatasync(rrd_file->fd) != 0) {
554 rrd_set_error("flushing fd %d: %s", rrd_file->fd,
555 rrd_strerror(errno));
556 }
557 }
560 /* Initialize RRD header. */
562 void rrd_init(
563 rrd_t *rrd)
564 {
565 rrd->stat_head = NULL;
566 rrd->ds_def = NULL;
567 rrd->rra_def = NULL;
568 rrd->live_head = NULL;
569 rrd->legacy_last_up = NULL;
570 rrd->rra_ptr = NULL;
571 rrd->pdp_prep = NULL;
572 rrd->cdp_prep = NULL;
573 rrd->rrd_value = NULL;
574 }
577 /* free RRD header data. */
579 #ifdef HAVE_MMAP
580 void rrd_free(
581 rrd_t *rrd)
582 {
583 if (rrd->legacy_last_up) { /* this gets set for version < 3 only */
584 free(rrd->live_head);
585 }
586 }
587 #else
588 void rrd_free(
589 rrd_t *rrd)
590 {
591 free(rrd->live_head);
592 free(rrd->stat_head);
593 free(rrd->ds_def);
594 free(rrd->rra_def);
595 free(rrd->rra_ptr);
596 free(rrd->pdp_prep);
597 free(rrd->cdp_prep);
598 free(rrd->rrd_value);
599 }
600 #endif
603 /* routine used by external libraries to free memory allocated by
604 * rrd library */
606 void rrd_freemem(
607 void *mem)
608 {
609 free(mem);
610 }