1 /*****************************************************************************
2 * RRDtool 1.2.23 Copyright by Tobi Oetiker, 1997-2007
3 *****************************************************************************
4 * rrd_open.c Open an RRD File
5 *****************************************************************************
8 * Revision 1.10 2004/05/26 22:11:12 oetiker
9 * reduce compiler warnings. Many small fixes. -- Mike Slifcak <slif@bellsouth.net>
11 * Revision 1.9 2003/04/29 21:56:49 oetiker
12 * readline in rrd_open.c reads the file in 8 KB blocks, and calls realloc for
13 * each block. realloc is very slow in Mac OS X for huge blocks, e.g. when
14 * restoring databases from huge xml files. This patch finds the size of the
15 * file, and starts out with malloc'ing the full size.
16 * -- Peter Speck <speck@ruc.dk>
18 * Revision 1.8 2003/04/11 19:43:44 oetiker
19 * New special value COUNT which allows calculations based on the position of a
20 * value within a data set. Bug fix in rrd_rpncalc.c. PREV returned erroneus
21 * value for the second value. Bug fix in rrd_restore.c. Bug causing seek error
22 * when accesing an RRD restored from an xml that holds an RRD version <3.
23 * -- Ruben Justo <ruben@ainek.com>
25 * Revision 1.7 2003/03/31 21:22:12 oetiker
26 * enables RRDtool updates with microsecond or in case of windows millisecond
27 * precision. This is needed to reduce time measurement error when archive step
28 * is small. (<30s) -- Sasha Mikheev <sasha@avalon-net.co.il>
30 * Revision 1.6 2003/02/13 07:05:27 oetiker
31 * Find attached the patch I promised to send to you. Please note that there
32 * are three new source files (src/rrd_is_thread_safe.h, src/rrd_thread_safe.c
33 * and src/rrd_not_thread_safe.c) and the introduction of librrd_th. This
34 * library is identical to librrd, but it contains support code for per-thread
35 * global variables currently used for error information only. This is similar
36 * to how errno per-thread variables are implemented. librrd_th must be linked
37 * alongside of libpthred
39 * There is also a new file "THREADS", holding some documentation.
41 * -- Peter Stamfest <peter@stamfest.at>
43 * Revision 1.5 2002/06/20 00:21:03 jake
44 * More Win32 build changes; thanks to Kerry Calvert.
46 * Revision 1.4 2002/02/01 20:34:49 oetiker
47 * fixed version number and date/time
49 * Revision 1.3 2001/03/04 13:01:55 oetiker
50 * Aberrant Behavior Detection support. A brief overview added to rrdtool.pod.
51 * Major updates to rrd_update.c, rrd_create.c. Minor update to other core files.
52 * This is backwards compatible! But new files using the Aberrant stuff are not readable
53 * by old rrdtool versions. See http://cricket.sourceforge.net/aberrant/rrd_hw.htm
54 * -- Jake Brutlag <jakeb@corp.webtv.net>
56 * Revision 1.2 2001/03/04 10:29:20 oetiker
57 * fixed filedescriptor leak
58 * -- Mike Franusich <mike@franusich.com>
60 * Revision 1.1.1.1 2001/02/25 22:25:05 oetiker
63 *****************************************************************************/
69 /* DEBUG 2 prints information obtained via mincore(2) */
71 /* do not calculate exact madvise hints but assume 1 page for headers and
72 * set DONTNEED for the rest, which is assumed to be data */
74 /* Avoid calling madvise on areas that were already hinted. May be benefical if
75 * your syscalls are very slow */
76 #define CHECK_MADVISE_OVERLAPS 1
79 /* the cast to void* is there to avoid this warning seen on ia64 with certain
80 versions of gcc: 'cast increases required alignment of target type'
82 #define __rrd_read(dst, dst_t, cnt) \
83 (dst) = (dst_t*)(void*) (data + offset); \
84 offset += sizeof(dst_t) * (cnt)
86 #define __rrd_read(dst, dst_t, cnt) \
87 if ((dst = malloc(sizeof(dst_t)*(cnt))) == NULL) { \
88 rrd_set_error(#dst " malloc"); \
89 goto out_nullify_head; \
91 offset += read (rrd_file->fd, dst, sizeof(dst_t)*(cnt))
94 /* next page-aligned (i.e. page-align up) */
96 #define PAGE_ALIGN(addr) (((addr)+_page_size-1)&(~(_page_size-1)))
98 /* previous page-aligned (i.e. page-align down) */
99 #ifndef PAGE_ALIGN_DOWN
100 #define PAGE_ALIGN_DOWN(addr) (((addr)+_page_size-1)&(~(_page_size-1)))
104 /* vector of last madvise hint */
105 typedef struct _madvise_vec_t {
109 _madvise_vec_t _madv_vec = { NULL, 0 };
112 #if defined CHECK_MADVISE_OVERLAPS
113 #define _madvise(_start, _off, _hint) \
114 if ((_start) != _madv_vec.start && (ssize_t)(_off) != _madv_vec.length) { \
115 _madv_vec.start = (_start) ; _madv_vec.length = (_off); \
116 madvise((_start), (_off), (_hint)); \
119 #define _madvise(_start, _off, _hint) \
120 madvise((_start), (_off), (_hint))
123 /* Open a database file, return its header and an open filehandle,
124 * positioned to the first cdp in the first rra.
125 * In the error path of rrd_open, only rrd_free(&rrd) has to be called
126 * before returning an error. Do not call rrd_close upon failure of rrd_open.
129 rrd_file_t *rrd_open(
130 const char *const file_name,
135 mode_t mode = S_IRUSR;
139 ssize_t _page_size = sysconf(_SC_PAGESIZE);
140 int mm_prot = PROT_READ, mm_flags = 0;
145 rrd_file_t *rrd_file = NULL;
146 off_t newfile_size = 0;
148 if (rdwr & RRD_CREAT) {
149 newfile_size = (off_t) rrd->stat_head->float_cookie;
150 free(rrd->stat_head);
153 rrd_file = malloc(sizeof(rrd_file_t));
154 if (rrd_file == NULL) {
155 rrd_set_error("allocating rrd_file descriptor for '%s'", file_name);
158 memset(rrd_file, 0, sizeof(rrd_file_t));
161 if ((rdwr & (RRD_READONLY | RRD_READWRITE)) ==
162 (RRD_READONLY | RRD_READWRITE)) {
163 /* Both READONLY and READWRITE were given, which is invalid. */
164 rrd_set_error("in read/write request mask");
168 if (rdwr & RRD_READONLY) {
171 mm_flags = MAP_PRIVATE;
172 # ifdef MAP_NORESERVE
173 mm_flags |= MAP_NORESERVE; /* readonly, so no swap backing needed */
177 if (rdwr & RRD_READWRITE) {
181 mm_flags = MAP_SHARED;
182 mm_prot |= PROT_WRITE;
185 if (rdwr & RRD_CREAT) {
186 flags |= (O_CREAT | O_TRUNC);
189 if (rdwr & RRD_READAHEAD) {
191 mm_flags |= MAP_POPULATE; /* populate ptes and data */
193 #if defined MAP_NONBLOCK
194 mm_flags |= MAP_NONBLOCK; /* just populate ptes */
205 if ((rrd_file->fd = open(file_name, flags, mode)) < 0) {
206 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
210 /* Better try to avoid seeks as much as possible. stat may be heavy but
211 * many concurrent seeks are even worse. */
212 if (newfile_size == 0 && ((fstat(rrd_file->fd, &statb)) < 0)) {
213 rrd_set_error("fstat '%s': %s", file_name, rrd_strerror(errno));
216 if (newfile_size == 0) {
217 rrd_file->file_len = statb.st_size;
219 rrd_file->file_len = newfile_size;
220 lseek(rrd_file->fd, newfile_size - 1, SEEK_SET);
221 write(rrd_file->fd, "\0", 1); /* poke */
222 lseek(rrd_file->fd, 0, SEEK_SET);
224 #ifdef HAVE_POSIX_FADVISE
225 /* In general we need no read-ahead when dealing with rrd_files.
226 When we stop reading, it is highly unlikely that we start up again.
227 In this manner we actually save time and diskaccess (and buffer cache).
228 Thanks to Dave Plonka for the Idea of using POSIX_FADV_RANDOM here. */
229 if (0 != posix_fadvise(rrd_file->fd, 0, 0, POSIX_FADV_RANDOM)) {
230 rrd_set_error("setting POSIX_FADV_RANDOM on '%s': %s", file_name,
231 rrd_strerror(errno));
237 if (rdwr & RRD_READWRITE)
239 if (setvbuf((rrd_file->fd),NULL,_IONBF,2)) {
240 rrd_set_error("failed to disable the stream buffer\n");
246 data = mmap(0, rrd_file->file_len, mm_prot, mm_flags,
247 rrd_file->fd, offset);
249 /* lets see if the first read worked */
250 if (data == MAP_FAILED) {
251 rrd_set_error("mmaping file '%s': %s", file_name,
252 rrd_strerror(errno));
255 rrd_file->file_start = data;
256 if (rdwr & RRD_CREAT) {
257 memset(data, DNAN, newfile_size - 1);
261 if (rdwr & RRD_CREAT)
264 if (rdwr & RRD_COPY) {
265 /* We will read everything in a moment (copying) */
266 _madvise(data, rrd_file->file_len, MADV_WILLNEED | MADV_SEQUENTIAL);
269 /* We do not need to read anything in for the moment */
270 _madvise(data, rrd_file->file_len, MADV_DONTNEED);
271 /* the stat_head will be needed soonish, so hint accordingly */
272 _madvise(data + PAGE_ALIGN_DOWN(offset),
273 PAGE_ALIGN(sizeof(stat_head_t)),
274 MADV_WILLNEED | MADV_RANDOM);
277 /* alternatively: keep 1 page worth of data, likely headers,
278 * don't need the rest. */
279 _madvise(data, _page_size, MADV_WILLNEED | MADV_SEQUENTIAL);
280 _madvise(data + _page_size, (rrd_file->file_len >= _page_size)
281 ? rrd_file->file_len - _page_size : 0, MADV_DONTNEED);
286 __rrd_read(rrd->stat_head, stat_head_t,
289 /* lets do some test if we are on track ... */
290 if (memcmp(rrd->stat_head->cookie, RRD_COOKIE, sizeof(RRD_COOKIE)) != 0) {
291 rrd_set_error("'%s' is not an RRD file", file_name);
292 goto out_nullify_head;
295 if (rrd->stat_head->float_cookie != FLOAT_COOKIE) {
296 rrd_set_error("This RRD was created on another architecture");
297 goto out_nullify_head;
300 version = atoi(rrd->stat_head->version);
302 if (version > atoi(RRD_VERSION)) {
303 rrd_set_error("can't handle RRD file version %s",
304 rrd->stat_head->version);
305 goto out_nullify_head;
307 #if defined USE_MADVISE && !defined ONE_PAGE
308 /* the ds_def will be needed soonish, so hint accordingly */
309 _madvise(data + PAGE_ALIGN_DOWN(offset),
310 PAGE_ALIGN(sizeof(ds_def_t) * rrd->stat_head->ds_cnt),
313 __rrd_read(rrd->ds_def, ds_def_t,
314 rrd->stat_head->ds_cnt);
316 #if defined USE_MADVISE && !defined ONE_PAGE
317 /* the rra_def will be needed soonish, so hint accordingly */
318 _madvise(data + PAGE_ALIGN_DOWN(offset),
319 PAGE_ALIGN(sizeof(rra_def_t) * rrd->stat_head->rra_cnt),
322 __rrd_read(rrd->rra_def, rra_def_t,
323 rrd->stat_head->rra_cnt);
325 /* handle different format for the live_head */
327 rrd->live_head = (live_head_t *) malloc(sizeof(live_head_t));
328 if (rrd->live_head == NULL) {
329 rrd_set_error("live_head_t malloc");
333 memmove(&rrd->live_head->last_up, data + offset, sizeof(long));
334 offset += sizeof(long);
336 offset += read(rrd_file->fd, &rrd->live_head->last_up, sizeof(long));
338 rrd->live_head->last_up_usec = 0;
340 #if defined USE_MADVISE && !defined ONE_PAGE
341 /* the live_head will be needed soonish, so hint accordingly */
342 _madvise(data + PAGE_ALIGN_DOWN(offset),
343 PAGE_ALIGN(sizeof(live_head_t)), MADV_WILLNEED);
345 __rrd_read(rrd->live_head, live_head_t,
348 //XXX: This doesn't look like it needs madvise
349 __rrd_read(rrd->pdp_prep, pdp_prep_t,
350 rrd->stat_head->ds_cnt);
352 //XXX: This could benefit from madvise()ing
353 __rrd_read(rrd->cdp_prep, cdp_prep_t,
354 rrd->stat_head->rra_cnt * rrd->stat_head->ds_cnt);
356 //XXX: This could benefit from madvise()ing
357 __rrd_read(rrd->rra_ptr, rra_ptr_t,
358 rrd->stat_head->rra_cnt);
360 rrd_file->header_len = offset;
361 rrd_file->pos = offset;
365 rrd->stat_head = NULL;
374 /* Close a reference to an rrd_file. */
377 rrd_file_t *rrd_file)
381 #if defined HAVE_MMAP || defined DEBUG
382 ssize_t _page_size = sysconf(_SC_PAGESIZE);
384 #if defined DEBUG && DEBUG > 1
385 /* pretty print blocks in core */
389 off = rrd_file->file_len +
390 ((rrd_file->file_len + _page_size - 1) / _page_size);
394 if (mincore(rrd_file->file_start, rrd_file->file_len, vec) == 0) {
396 unsigned is_in = 0, was_in = 0;
398 for (off = 0, prev = 0; off < rrd_file->file_len; ++off) {
399 is_in = vec[off] & 1; /* if lsb set then is core resident */
402 if (was_in != is_in) {
403 fprintf(stderr, "%sin core: %p len %ld\n",
404 was_in ? "" : "not ", vec + prev, off - prev);
410 "%sin core: %p len %ld\n",
411 was_in ? "" : "not ", vec + prev, off - prev);
413 fprintf(stderr, "mincore: %s", rrd_strerror(errno));
419 /* Keep headers around, round up to next page boundary. */
421 PAGE_ALIGN(rrd_file->header_len % _page_size + rrd_file->header_len);
422 if (rrd_file->file_len > ret)
423 _madvise(rrd_file->file_start + ret,
424 rrd_file->file_len - ret, MADV_DONTNEED);
426 /* ignoring errors from RRDs that are smaller then the file_len+rounding */
427 _madvise(rrd_file->file_start + PAGE_ALIGN_DOWN(rrd_file->header_len),
428 rrd_file->file_len - PAGE_ALIGN(rrd_file->header_len),
433 ret = munmap(rrd_file->file_start, rrd_file->file_len);
435 rrd_set_error("munmap rrd_file: %s", rrd_strerror(errno));
437 ret = close(rrd_file->fd);
439 rrd_set_error("closing file: %s", rrd_strerror(errno));
446 /* Set position of rrd_file. */
449 rrd_file_t *rrd_file,
456 if (whence == SEEK_SET)
458 else if (whence == SEEK_CUR)
459 rrd_file->pos += off;
460 else if (whence == SEEK_END)
461 rrd_file->pos = rrd_file->file_len + off;
463 ret = lseek(rrd_file->fd, off, whence);
465 rrd_set_error("lseek: %s", rrd_strerror(errno));
468 //XXX: mimic fseek, which returns 0 upon success
469 return ret == -1; //XXX: or just ret to mimic lseek
473 /* Get current position in rrd_file. */
475 inline off_t rrd_tell(
476 rrd_file_t *rrd_file)
478 return rrd_file->pos;
482 /* read count bytes into buffer buf, starting at rrd_file->pos.
483 * Returns the number of bytes read or <0 on error. */
485 inline ssize_t rrd_read(
486 rrd_file_t *rrd_file,
492 ssize_t _surplus = rrd_file->pos + _cnt - rrd_file->file_len;
494 if (_surplus > 0) { /* short read */
499 buf = memcpy(buf, rrd_file->file_start + rrd_file->pos, _cnt);
501 rrd_file->pos += _cnt; /* mimmic read() semantics */
506 ret = read(rrd_file->fd, buf, count);
508 rrd_file->pos += ret; /* mimmic read() semantics */
514 /* write count bytes from buffer buf to the current position
515 * rrd_file->pos of rrd_file->fd.
516 * Returns the number of bytes written. */
518 inline ssize_t rrd_write(
519 rrd_file_t *rrd_file,
524 memcpy(rrd_file->file_start + rrd_file->pos, buf, count);
525 rrd_file->pos += count;
526 return count; /* mimmic write() semantics */
528 ssize_t _sz = write(rrd_file->fd, buf, count);
531 rrd_file->pos += _sz;
537 /* flush all data pending to be written to FD. */
539 inline void rrd_flush(
540 rrd_file_t *rrd_file)
542 if (fdatasync(rrd_file->fd) != 0) {
543 rrd_set_error("flushing fd %d: %s", rrd_file->fd,
544 rrd_strerror(errno));
549 /* Initialize RRD header. */
554 rrd->stat_head = NULL;
557 rrd->live_head = NULL;
559 rrd->pdp_prep = NULL;
560 rrd->cdp_prep = NULL;
561 rrd->rrd_value = NULL;
565 /* free RRD header data. */
568 inline void rrd_free(
576 free(rrd->live_head);
577 free(rrd->stat_head);
583 free(rrd->rrd_value);
588 /* routine used by external libraries to free memory allocated by
598 /* XXX: FIXME: missing documentation. */
599 /*XXX: FIXME should be renamed to rrd_readfile or _rrd_readfile */
601 int /*_rrd_*/ readfile(
602 const char *file_name,
606 long writecnt = 0, totalcnt = MEMBLK;
611 if ((strcmp("-", file_name) == 0)) {
614 if ((input = fopen(file_name, "rb")) == NULL) {
615 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
623 } while (c != '\n' && !feof(input));
625 if (strcmp("-", file_name)) {
626 fseek(input, 0, SEEK_END);
627 /* have extra space for detecting EOF without realloc */
628 totalcnt = (ftell(input) + 1) / sizeof(char) - offset;
629 if (totalcnt < MEMBLK)
630 totalcnt = MEMBLK; /* sanitize */
631 fseek(input, offset * sizeof(char), SEEK_SET);
633 if (((*buffer) = (char *) malloc((totalcnt + 4) * sizeof(char))) == NULL) {
634 perror("Allocate Buffer:");
639 fread((*buffer) + writecnt, 1,
640 (totalcnt - writecnt) * sizeof(char), input);
641 if (writecnt >= totalcnt) {
644 rrd_realloc((*buffer),
645 (totalcnt + 4) * sizeof(char))) == NULL) {
646 perror("Realloc Buffer:");
650 } while (!feof(input));
651 (*buffer)[writecnt] = '\0';
652 if (strcmp("-", file_name) != 0) {