1 /*****************************************************************************
2 * RRDtool 1.3.2 Copyright by Tobi Oetiker, 1997-2008
3 *****************************************************************************
4 * rrd_open.c Open an RRD File
5 *****************************************************************************
7 *****************************************************************************/
15 #ifdef HAVE_BROKEN_MS_ASYNC
16 #include <sys/types.h>
25 #define _LK_UNLCK 0 /* Unlock */
26 #define _LK_LOCK 1 /* Lock */
27 #define _LK_NBLCK 2 /* Non-blocking lock */
28 #define _LK_RLCK 3 /* Lock for read only */
29 #define _LK_NBRLCK 4 /* Non-blocking lock for read only */
32 #define LK_UNLCK _LK_UNLCK
33 #define LK_LOCK _LK_LOCK
34 #define LK_NBLCK _LK_NBLCK
35 #define LK_RLCK _LK_RLCK
36 #define LK_NBRLCK _LK_NBRLCK
39 /* DEBUG 2 prints information obtained via mincore(2) */
41 /* do not calculate exact madvise hints but assume 1 page for headers and
42 * set DONTNEED for the rest, which is assumed to be data */
43 /* Avoid calling madvise on areas that were already hinted. May be benefical if
44 * your syscalls are very slow */
47 /* the cast to void* is there to avoid this warning seen on ia64 with certain
48 versions of gcc: 'cast increases required alignment of target type'
50 #define __rrd_read(dst, dst_t, cnt) { \
51 size_t wanted = sizeof(dst_t)*(cnt); \
52 if (offset + wanted > rrd_file->file_len) { \
53 rrd_set_error("reached EOF while loading header " #dst); \
54 goto out_nullify_head; \
56 (dst) = (dst_t*)(void*) (data + offset); \
60 #define __rrd_read(dst, dst_t, cnt) { \
61 size_t wanted = sizeof(dst_t)*(cnt); \
63 if ((dst = (dst_t*)malloc(wanted)) == NULL) { \
64 rrd_set_error(#dst " malloc"); \
65 goto out_nullify_head; \
67 got = read (rrd_simple_file->fd, dst, wanted); \
68 if (got != wanted) { \
69 rrd_set_error("short read while reading header " #dst); \
70 goto out_nullify_head; \
76 /* get the address of the start of this page */
77 #if defined USE_MADVISE || defined HAVE_POSIX_FADVISE
79 #define PAGE_START(addr) ((addr)&(~(_page_size-1)))
83 /* Open a database file, return its header and an open filehandle,
84 * positioned to the first cdp in the first rra.
85 * In the error path of rrd_open, only rrd_free(&rrd) has to be called
86 * before returning an error. Do not call rrd_close upon failure of rrd_open.
87 * If creating a new file, the parameter rrd must be initialised with
88 * details of the file content.
89 * If opening an existing file, then use rrd must be initialised by
90 * rrd_init(rrd) prior to invoking rrd_open
94 const char *const file_name,
103 ssize_t _page_size = sysconf(_SC_PAGESIZE);
104 char *data = MAP_FAILED;
108 rrd_file_t *rrd_file = NULL;
109 rrd_simple_file_t *rrd_simple_file = NULL;
110 size_t newfile_size = 0;
111 size_t header_len, value_cnt, data_len;
113 /* Are we creating a new file? */
114 if((rdwr & RRD_CREAT) && (rrd->stat_head != NULL))
116 header_len = rrd_get_header_size(rrd);
119 for (ui = 0; ui < rrd->stat_head->rra_cnt; ui++)
120 value_cnt += rrd->stat_head->ds_cnt * rrd->rra_def[ui].row_cnt;
122 data_len = sizeof(rrd_value_t) * value_cnt;
124 newfile_size = header_len + data_len;
127 rrd_file = (rrd_file_t*)malloc(sizeof(rrd_file_t));
128 if (rrd_file == NULL) {
129 rrd_set_error("allocating rrd_file descriptor for '%s'", file_name);
132 memset(rrd_file, 0, sizeof(rrd_file_t));
134 rrd_file->pvt = malloc(sizeof(rrd_simple_file_t));
135 if(rrd_file->pvt == NULL) {
136 rrd_set_error("allocating rrd_simple_file for '%s'", file_name);
139 memset(rrd_file->pvt, 0, sizeof(rrd_simple_file_t));
140 rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
143 if ((rdwr & (RRD_READONLY | RRD_READWRITE)) ==
144 (RRD_READONLY | RRD_READWRITE)) {
145 /* Both READONLY and READWRITE were given, which is invalid. */
146 rrd_set_error("in read/write request mask");
152 rrd_simple_file->mm_prot = PROT_READ;
153 rrd_simple_file->mm_flags = 0;
156 if (rdwr & RRD_READONLY) {
159 rrd_simple_file->mm_flags = MAP_PRIVATE;
160 # ifdef MAP_NORESERVE
161 rrd_simple_file->mm_flags |= MAP_NORESERVE; /* readonly, so no swap backing needed */
165 if (rdwr & RRD_READWRITE) {
168 rrd_simple_file->mm_flags = MAP_SHARED;
169 rrd_simple_file->mm_prot |= PROT_WRITE;
172 if (rdwr & RRD_CREAT) {
173 flags |= (O_CREAT | O_TRUNC);
176 if (rdwr & RRD_READAHEAD) {
178 rrd_simple_file->mm_flags |= MAP_POPULATE; /* populate ptes and data */
180 #if defined MAP_NONBLOCK
181 rrd_simple_file->mm_flags |= MAP_NONBLOCK; /* just populate ptes */
184 #if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__CYGWIN32__)
188 if ((rrd_simple_file->fd = open(file_name, flags, 0666)) < 0) {
189 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
194 #ifdef HAVE_BROKEN_MS_ASYNC
195 if (rdwr & RRD_READWRITE) {
196 /* some unices, the files mtime does not get update
197 on msync MS_ASYNC, in order to help them,
198 we update the the timestamp at this point.
199 The thing happens pretty 'close' to the open
200 call so the chances of a race should be minimal.
202 Maybe ask your vendor to fix your OS ... */
203 utime(file_name,NULL);
208 /* Better try to avoid seeks as much as possible. stat may be heavy but
209 * many concurrent seeks are even worse. */
210 if (newfile_size == 0 && ((fstat(rrd_simple_file->fd, &statb)) < 0)) {
211 rrd_set_error("fstat '%s': %s", file_name, rrd_strerror(errno));
214 if (newfile_size == 0) {
215 rrd_file->file_len = statb.st_size;
217 rrd_file->file_len = newfile_size;
218 lseek(rrd_simple_file->fd, newfile_size - 1, SEEK_SET);
219 write(rrd_simple_file->fd, "\0", 1); /* poke */
220 lseek(rrd_simple_file->fd, 0, SEEK_SET);
222 #ifdef HAVE_POSIX_FADVISE
223 /* In general we need no read-ahead when dealing with rrd_files.
224 When we stop reading, it is highly unlikely that we start up again.
225 In this manner we actually save time and diskaccess (and buffer cache).
226 Thanks to Dave Plonka for the Idea of using POSIX_FADV_RANDOM here. */
227 posix_fadvise(rrd_simple_file->fd, 0, 0, POSIX_FADV_RANDOM);
231 if (rdwr & RRD_READWRITE)
233 if (setvbuf((rrd_simple_file->fd),NULL,_IONBF,2)) {
234 rrd_set_error("failed to disable the stream buffer\n");
241 data = mmap(0, rrd_file->file_len,
242 rrd_simple_file->mm_prot, rrd_simple_file->mm_flags,
243 rrd_simple_file->fd, offset);
245 /* lets see if the first read worked */
246 if (data == MAP_FAILED) {
247 rrd_set_error("mmaping file '%s': %s", file_name,
248 rrd_strerror(errno));
251 rrd_simple_file->file_start = data;
252 if (rdwr & RRD_CREAT) {
253 memset(data, DNAN, newfile_size - 1);
257 if (rdwr & RRD_CREAT)
260 if (rdwr & RRD_COPY) {
261 /* We will read everything in a moment (copying) */
262 madvise(data, rrd_file->file_len, MADV_WILLNEED | MADV_SEQUENTIAL);
264 /* We do not need to read anything in for the moment */
265 madvise(data, rrd_file->file_len, MADV_RANDOM);
266 /* the stat_head will be needed soonish, so hint accordingly */
267 madvise(data, sizeof(stat_head_t), MADV_WILLNEED | MADV_RANDOM);
271 __rrd_read(rrd->stat_head, stat_head_t,
274 /* lets do some test if we are on track ... */
275 if (memcmp(rrd->stat_head->cookie, RRD_COOKIE, sizeof(RRD_COOKIE)) != 0) {
276 rrd_set_error("'%s' is not an RRD file", file_name);
277 goto out_nullify_head;
280 if (rrd->stat_head->float_cookie != FLOAT_COOKIE) {
281 rrd_set_error("This RRD was created on another architecture");
282 goto out_nullify_head;
285 version = atoi(rrd->stat_head->version);
287 if (version > atoi(RRD_VERSION)) {
288 rrd_set_error("can't handle RRD file version %s",
289 rrd->stat_head->version);
290 goto out_nullify_head;
292 #if defined USE_MADVISE
293 /* the ds_def will be needed soonish, so hint accordingly */
294 madvise(data + PAGE_START(offset),
295 sizeof(ds_def_t) * rrd->stat_head->ds_cnt, MADV_WILLNEED);
297 __rrd_read(rrd->ds_def, ds_def_t,
298 rrd->stat_head->ds_cnt);
300 #if defined USE_MADVISE
301 /* the rra_def will be needed soonish, so hint accordingly */
302 madvise(data + PAGE_START(offset),
303 sizeof(rra_def_t) * rrd->stat_head->rra_cnt, MADV_WILLNEED);
305 __rrd_read(rrd->rra_def, rra_def_t,
306 rrd->stat_head->rra_cnt);
308 /* handle different format for the live_head */
310 rrd->live_head = (live_head_t *) malloc(sizeof(live_head_t));
311 if (rrd->live_head == NULL) {
312 rrd_set_error("live_head_t malloc");
315 #if defined USE_MADVISE
316 /* the live_head will be needed soonish, so hint accordingly */
317 madvise(data + PAGE_START(offset), sizeof(time_t), MADV_WILLNEED);
319 __rrd_read(rrd->legacy_last_up, time_t,
322 rrd->live_head->last_up = *rrd->legacy_last_up;
323 rrd->live_head->last_up_usec = 0;
325 #if defined USE_MADVISE
326 /* the live_head will be needed soonish, so hint accordingly */
327 madvise(data + PAGE_START(offset),
328 sizeof(live_head_t), MADV_WILLNEED);
330 __rrd_read(rrd->live_head, live_head_t,
333 __rrd_read(rrd->pdp_prep, pdp_prep_t,
334 rrd->stat_head->ds_cnt);
335 __rrd_read(rrd->cdp_prep, cdp_prep_t,
336 rrd->stat_head->rra_cnt * rrd->stat_head->ds_cnt);
337 __rrd_read(rrd->rra_ptr, rra_ptr_t,
338 rrd->stat_head->rra_cnt);
340 rrd_file->header_len = offset;
341 rrd_file->pos = offset;
344 unsigned long row_cnt = 0;
346 for (ui=0; ui<rrd->stat_head->rra_cnt; ui++)
347 row_cnt += rrd->rra_def[ui].row_cnt;
349 size_t correct_len = rrd_file->header_len +
350 sizeof(rrd_value_t) * row_cnt * rrd->stat_head->ds_cnt;
352 if (correct_len > rrd_file->file_len)
354 rrd_set_error("'%s' is too small (should be %ld bytes)",
355 file_name, (long long) correct_len);
356 goto out_nullify_head;
363 rrd->stat_head = NULL;
366 if (data != MAP_FAILED)
367 munmap(data, rrd_file->file_len);
370 close(rrd_simple_file->fd);
378 #if defined DEBUG && DEBUG > 1
379 /* Print list of in-core pages of a the current rrd_file. */
382 rrd_file_t *rrd_file,
385 rrd_simple_file_t *rrd_simple_file;
386 rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
388 /* pretty print blocks in core */
391 ssize_t _page_size = sysconf(_SC_PAGESIZE);
393 off = rrd_file->file_len +
394 ((rrd_file->file_len + _page_size - 1) / _page_size);
398 if (mincore(rrd_simple_file->file_start, rrd_file->file_len, vec) == 0) {
400 unsigned is_in = 0, was_in = 0;
402 for (off = 0, prev = 0; off < rrd_file->file_len; ++off) {
403 is_in = vec[off] & 1; /* if lsb set then is core resident */
406 if (was_in != is_in) {
407 fprintf(stderr, "%s: %sin core: %p len %ld\n", mark,
408 was_in ? "" : "not ", vec + prev, off - prev);
414 "%s: %sin core: %p len %ld\n", mark,
415 was_in ? "" : "not ", vec + prev, off - prev);
417 fprintf(stderr, "mincore: %s", rrd_strerror(errno));
420 fprintf(stderr, "sorry mincore only works with mmap");
423 #endif /* defined DEBUG && DEBUG > 1 */
426 * get exclusive lock to whole file.
427 * lock gets removed when we close the file
429 * returns 0 on success
432 rrd_file_t *rrd_file)
435 rrd_simple_file_t *rrd_simple_file;
436 rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
439 #if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__CYGWIN32__)
442 if (_fstat(rrd_simple_file->fd, &st) == 0) {
443 rcstat = _locking(rrd_simple_file->fd, _LK_NBLCK, st.st_size);
450 lock.l_type = F_WRLCK; /* exclusive write lock */
451 lock.l_len = 0; /* whole file */
452 lock.l_start = 0; /* start of file */
453 lock.l_whence = SEEK_SET; /* end of file */
455 rcstat = fcntl(rrd_simple_file->fd, F_SETLK, &lock);
463 /* drop cache except for the header and the active pages */
465 rrd_file_t *rrd_file,
468 rrd_simple_file_t *rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
469 #if defined USE_MADVISE || defined HAVE_POSIX_FADVISE
470 size_t dontneed_start;
474 ssize_t _page_size = sysconf(_SC_PAGESIZE);
476 if (rrd_file == NULL) {
477 #if defined DEBUG && DEBUG
478 fprintf (stderr, "rrd_dontneed: Argument 'rrd_file' is NULL.\n");
483 #if defined DEBUG && DEBUG > 1
484 mincore_print(rrd_file, "before");
487 /* ignoring errors from RRDs that are smaller then the file_len+rounding */
488 rra_start = rrd_file->header_len;
489 dontneed_start = PAGE_START(rra_start) + _page_size;
490 for (i = 0; i < rrd->stat_head->rra_cnt; ++i) {
493 + rrd->rra_ptr[i].cur_row
494 * rrd->stat_head->ds_cnt * sizeof(rrd_value_t));
495 if (active_block > dontneed_start) {
497 madvise(rrd_simple_file->file_start + dontneed_start,
498 active_block - dontneed_start - 1, MADV_DONTNEED);
500 /* in linux at least only fadvise DONTNEED seems to purge pages from cache */
501 #ifdef HAVE_POSIX_FADVISE
502 posix_fadvise(rrd_simple_file->fd, dontneed_start,
503 active_block - dontneed_start - 1,
504 POSIX_FADV_DONTNEED);
507 dontneed_start = active_block;
508 /* do not release 'hot' block if update for this RAA will occur
509 * within 10 minutes */
510 if (rrd->stat_head->pdp_step * rrd->rra_def[i].pdp_cnt -
511 rrd->live_head->last_up % (rrd->stat_head->pdp_step *
512 rrd->rra_def[i].pdp_cnt) < 10 * 60) {
513 dontneed_start += _page_size;
516 rrd->rra_def[i].row_cnt * rrd->stat_head->ds_cnt *
520 if (dontneed_start < rrd_file->file_len) {
522 madvise(rrd_simple_file->file_start + dontneed_start,
523 rrd_file->file_len - dontneed_start, MADV_DONTNEED);
525 #ifdef HAVE_POSIX_FADVISE
526 posix_fadvise(rrd_simple_file->fd, dontneed_start,
527 rrd_file->file_len - dontneed_start,
528 POSIX_FADV_DONTNEED);
532 #if defined DEBUG && DEBUG > 1
533 mincore_print(rrd_file, "after");
535 #endif /* without madvise and posix_fadvise it does not make much sense todo anything */
543 rrd_file_t *rrd_file)
545 rrd_simple_file_t *rrd_simple_file;
546 rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
550 ret = msync(rrd_simple_file->file_start, rrd_file->file_len, MS_ASYNC);
552 rrd_set_error("msync rrd_file: %s", rrd_strerror(errno));
553 ret = munmap(rrd_simple_file->file_start, rrd_file->file_len);
555 rrd_set_error("munmap rrd_file: %s", rrd_strerror(errno));
557 ret = close(rrd_simple_file->fd);
559 rrd_set_error("closing file: %s", rrd_strerror(errno));
567 /* Set position of rrd_file. */
570 rrd_file_t *rrd_file,
575 rrd_simple_file_t *rrd_simple_file;
576 rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
579 if (whence == SEEK_SET)
581 else if (whence == SEEK_CUR)
582 rrd_file->pos += off;
583 else if (whence == SEEK_END)
584 rrd_file->pos = rrd_file->file_len + off;
586 ret = lseek(rrd_simple_file->fd, off, whence);
588 rrd_set_error("lseek: %s", rrd_strerror(errno));
591 /* mimic fseek, which returns 0 upon success */
592 return ret < 0; /*XXX: or just ret to mimic lseek */
596 /* Get current position in rrd_file. */
599 rrd_file_t *rrd_file)
601 return rrd_file->pos;
605 /* Read count bytes into buffer buf, starting at rrd_file->pos.
606 * Returns the number of bytes read or <0 on error. */
609 rrd_file_t *rrd_file,
613 rrd_simple_file_t *rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
618 if (rrd_file->pos > rrd_file->file_len || _cnt == 0) /* EOF */
621 return -1; /* EINVAL */
622 _surplus = rrd_file->pos + _cnt - rrd_file->file_len;
623 if (_surplus > 0) { /* short read */
628 buf = memcpy(buf, rrd_simple_file->file_start + rrd_file->pos, _cnt);
630 rrd_file->pos += _cnt; /* mimmic read() semantics */
635 ret = read(rrd_simple_file->fd, buf, count);
637 rrd_file->pos += ret; /* mimmic read() semantics */
643 /* Write count bytes from buffer buf to the current position
644 * rrd_file->pos of rrd_simple_file->fd.
645 * Returns the number of bytes written or <0 on error. */
648 rrd_file_t *rrd_file,
652 rrd_simple_file_t *rrd_simple_file = (rrd_simple_file_t *)rrd_file->pvt;
654 size_t old_size = rrd_file->file_len;
658 return -1; /* EINVAL */
660 if((rrd_file->pos + count) > old_size)
662 rrd_set_error("attempting to write beyond end of file");
665 memcpy(rrd_simple_file->file_start + rrd_file->pos, buf, count);
666 rrd_file->pos += count;
667 return count; /* mimmic write() semantics */
669 ssize_t _sz = write(rrd_simple_file->fd, buf, count);
672 rrd_file->pos += _sz;
678 /* this is a leftover from the old days, it serves no purpose
679 and is therefore turned into a no-op */
681 rrd_file_t *rrd_file __attribute__((unused)))
685 /* Initialize RRD header. */
690 rrd->stat_head = NULL;
693 rrd->live_head = NULL;
694 rrd->legacy_last_up = NULL;
696 rrd->pdp_prep = NULL;
697 rrd->cdp_prep = NULL;
698 rrd->rrd_value = NULL;
702 /* free RRD header data. */
708 if (rrd->legacy_last_up) { /* this gets set for version < 3 only */
709 free(rrd->live_head);
716 free(rrd->live_head);
717 free(rrd->stat_head);
723 free(rrd->rrd_value);
728 /* routine used by external libraries to free memory allocated by
738 * rra_update informs us about the RRAs being updated
739 * The low level storage API may use this information for
740 * aligning RRAs within stripes, or other performance enhancements
743 rrd_file_t *rrd_file __attribute__((unused)),
744 int rra_idx __attribute__((unused)),
745 unsigned long rra_row __attribute__((unused)),
746 time_t rra_time __attribute__((unused)))
751 * This function is called when creating a new RRD
752 * The storage implementation can use this opportunity to select
753 * a sensible starting row within the file.
754 * The default implementation is random, to ensure that all RRAs
755 * don't change to a new disk block at the same time
757 unsigned long rrd_select_initial_row(
758 rrd_file_t *rrd_file __attribute__((unused)),
759 int rra_idx __attribute__((unused)),
763 return rrd_random() % rra->row_cnt;