1 /*****************************************************************************
2 * RRDtool 1.3.2 Copyright by Tobi Oetiker, 1997-2008
3 *****************************************************************************
4 * rrd_open.c Open an RRD File
5 *****************************************************************************
7 *****************************************************************************/
13 /* DEBUG 2 prints information obtained via mincore(2) */
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 */
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'
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; \
30 (dst) = (dst_t*)(void*) (data + offset); \
34 #define __rrd_read(dst, dst_t, cnt) { \
35 size_t wanted = sizeof(dst_t)*(cnt); \
37 if ((dst = malloc(wanted)) == NULL) { \
38 rrd_set_error(#dst " malloc"); \
39 goto out_nullify_head; \
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; \
50 /* get the address of the start of this page */
51 #if defined USE_MADVISE || defined HAVE_POSIX_FADVISE
53 #define PAGE_START(addr) ((addr)&(~(_page_size-1)))
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.
64 const char *const file_name,
69 mode_t mode = S_IRUSR;
73 ssize_t _page_size = sysconf(_SC_PAGESIZE);
74 int mm_prot = PROT_READ, mm_flags = 0;
75 char *data = MAP_FAILED;
79 rrd_file_t *rrd_file = NULL;
80 off_t newfile_size = 0;
82 if ((rdwr & RRD_CREAT) && (rdwr & RRD_CREAT_SETSIZE)) {
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;
88 if(!(rdwr & RRD_CREAT))
90 rrd_file = malloc(sizeof(rrd_file_t));
91 if (rrd_file == NULL) {
92 rrd_set_error("allocating rrd_file descriptor for '%s'", file_name);
95 memset(rrd_file, 0, sizeof(rrd_file_t));
98 if ((rdwr & (RRD_READONLY | RRD_READWRITE)) ==
99 (RRD_READONLY | RRD_READWRITE)) {
100 /* Both READONLY and READWRITE were given, which is invalid. */
101 rrd_set_error("in read/write request mask");
105 if (rdwr & RRD_READONLY) {
108 mm_flags = MAP_PRIVATE;
109 # ifdef MAP_NORESERVE
110 mm_flags |= MAP_NORESERVE; /* readonly, so no swap backing needed */
114 if (rdwr & RRD_READWRITE) {
118 mm_flags = MAP_SHARED;
119 mm_prot |= PROT_WRITE;
122 if (rdwr & RRD_CREAT) {
123 flags |= (O_CREAT | O_TRUNC);
126 if (rdwr & RRD_READAHEAD) {
128 mm_flags |= MAP_POPULATE; /* populate ptes and data */
130 #if defined MAP_NONBLOCK
131 mm_flags |= MAP_NONBLOCK; /* just populate ptes */
134 #if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__CYGWIN32__)
138 if ((rrd_file->fd = open(file_name, flags, mode)) < 0) {
139 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
143 /* Better try to avoid seeks as much as possible. stat may be heavy but
144 * many concurrent seeks are even worse. */
145 if (newfile_size == 0 && ((fstat(rrd_file->fd, &statb)) < 0)) {
146 rrd_set_error("fstat '%s': %s", file_name, rrd_strerror(errno));
149 if (newfile_size == 0) {
150 rrd_file->file_len = statb.st_size;
152 rrd_file->file_len = newfile_size;
153 lseek(rrd_file->fd, newfile_size - 1, SEEK_SET);
154 write(rrd_file->fd, "\0", 1); /* poke */
155 lseek(rrd_file->fd, 0, SEEK_SET);
157 #ifdef HAVE_POSIX_FADVISE
158 /* In general we need no read-ahead when dealing with rrd_files.
159 When we stop reading, it is highly unlikely that we start up again.
160 In this manner we actually save time and diskaccess (and buffer cache).
161 Thanks to Dave Plonka for the Idea of using POSIX_FADV_RANDOM here. */
162 posix_fadvise(rrd_file->fd, 0, 0, POSIX_FADV_RANDOM);
166 if (rdwr & RRD_READWRITE)
168 if (setvbuf((rrd_file->fd),NULL,_IONBF,2)) {
169 rrd_set_error("failed to disable the stream buffer\n");
176 if(rrd_file->file_len == 0 && (rdwr & RRD_CREAT))
178 rrd_file->file_start = NULL;
181 data = mmap(0, rrd_file->file_len, mm_prot, mm_flags,
182 rrd_file->fd, offset);
184 /* lets see if the first read worked */
185 if (data == MAP_FAILED) {
186 rrd_set_error("mmaping file '%s': %s", file_name,
187 rrd_strerror(errno));
190 rrd_file->file_start = data;
191 if (rdwr & RRD_CREAT) {
192 memset(data, DNAN, newfile_size - 1);
196 if (rdwr & RRD_CREAT)
199 if (rdwr & RRD_COPY) {
200 /* We will read everything in a moment (copying) */
201 madvise(data, rrd_file->file_len, MADV_WILLNEED | MADV_SEQUENTIAL);
203 /* We do not need to read anything in for the moment */
204 madvise(data, rrd_file->file_len, MADV_RANDOM);
205 /* the stat_head will be needed soonish, so hint accordingly */
206 madvise(data, sizeof(stat_head_t), MADV_WILLNEED | MADV_RANDOM);
210 __rrd_read(rrd->stat_head, stat_head_t,
213 /* lets do some test if we are on track ... */
214 if (memcmp(rrd->stat_head->cookie, RRD_COOKIE, sizeof(RRD_COOKIE)) != 0) {
215 rrd_set_error("'%s' is not an RRD file", file_name);
216 goto out_nullify_head;
219 if (rrd->stat_head->float_cookie != FLOAT_COOKIE) {
220 rrd_set_error("This RRD was created on another architecture");
221 goto out_nullify_head;
224 version = atoi(rrd->stat_head->version);
226 if (version > atoi(RRD_VERSION)) {
227 rrd_set_error("can't handle RRD file version %s",
228 rrd->stat_head->version);
229 goto out_nullify_head;
231 #if defined USE_MADVISE
232 /* the ds_def will be needed soonish, so hint accordingly */
233 madvise(data + PAGE_START(offset),
234 sizeof(ds_def_t) * rrd->stat_head->ds_cnt, MADV_WILLNEED);
236 __rrd_read(rrd->ds_def, ds_def_t,
237 rrd->stat_head->ds_cnt);
239 #if defined USE_MADVISE
240 /* the rra_def will be needed soonish, so hint accordingly */
241 madvise(data + PAGE_START(offset),
242 sizeof(rra_def_t) * rrd->stat_head->rra_cnt, MADV_WILLNEED);
244 __rrd_read(rrd->rra_def, rra_def_t,
245 rrd->stat_head->rra_cnt);
247 /* handle different format for the live_head */
249 rrd->live_head = (live_head_t *) malloc(sizeof(live_head_t));
250 if (rrd->live_head == NULL) {
251 rrd_set_error("live_head_t malloc");
254 #if defined USE_MADVISE
255 /* the live_head will be needed soonish, so hint accordingly */
256 madvise(data + PAGE_START(offset), sizeof(time_t), MADV_WILLNEED);
258 __rrd_read(rrd->legacy_last_up, time_t,
261 rrd->live_head->last_up = *rrd->legacy_last_up;
262 rrd->live_head->last_up_usec = 0;
264 #if defined USE_MADVISE
265 /* the live_head will be needed soonish, so hint accordingly */
266 madvise(data + PAGE_START(offset),
267 sizeof(live_head_t), MADV_WILLNEED);
269 __rrd_read(rrd->live_head, live_head_t,
272 __rrd_read(rrd->pdp_prep, pdp_prep_t,
273 rrd->stat_head->ds_cnt);
274 __rrd_read(rrd->cdp_prep, cdp_prep_t,
275 rrd->stat_head->rra_cnt * rrd->stat_head->ds_cnt);
276 __rrd_read(rrd->rra_ptr, rra_ptr_t,
277 rrd->stat_head->rra_cnt);
279 rrd_file->header_len = offset;
280 rrd_file->pos = offset;
283 unsigned long row_cnt = 0;
286 for (i=0; i<rrd->stat_head->rra_cnt; i++)
287 row_cnt += rrd->rra_def[i].row_cnt;
289 off_t correct_len = rrd_file->header_len +
290 sizeof(rrd_value_t) * row_cnt * rrd->stat_head->ds_cnt;
292 if (correct_len > rrd_file->file_len)
294 rrd_set_error("'%s' is too small (should be %ld bytes)",
295 file_name, (long long) correct_len);
296 goto out_nullify_head;
303 rrd->stat_head = NULL;
306 if (data != MAP_FAILED)
307 munmap(data, rrd_file->file_len);
316 #if defined DEBUG && DEBUG > 1
317 /* Print list of in-core pages of a the current rrd_file. */
320 rrd_file_t *rrd_file,
324 /* pretty print blocks in core */
327 ssize_t _page_size = sysconf(_SC_PAGESIZE);
329 off = rrd_file->file_len +
330 ((rrd_file->file_len + _page_size - 1) / _page_size);
334 if (mincore(rrd_file->file_start, rrd_file->file_len, vec) == 0) {
336 unsigned is_in = 0, was_in = 0;
338 for (off = 0, prev = 0; off < rrd_file->file_len; ++off) {
339 is_in = vec[off] & 1; /* if lsb set then is core resident */
342 if (was_in != is_in) {
343 fprintf(stderr, "%s: %sin core: %p len %ld\n", mark,
344 was_in ? "" : "not ", vec + prev, off - prev);
350 "%s: %sin core: %p len %ld\n", mark,
351 was_in ? "" : "not ", vec + prev, off - prev);
353 fprintf(stderr, "mincore: %s", rrd_strerror(errno));
356 fprintf(stderr, "sorry mincore only works with mmap");
359 #endif /* defined DEBUG && DEBUG > 1 */
362 /* drop cache except for the header and the active pages */
364 rrd_file_t *rrd_file,
367 #if defined USE_MADVISE || defined HAVE_POSIX_FADVISE
368 off_t dontneed_start;
372 ssize_t _page_size = sysconf(_SC_PAGESIZE);
374 if (rrd_file == NULL) {
375 #if defined DEBUG && DEBUG
376 fprintf (stderr, "rrd_dontneed: Argument 'rrd_file' is NULL.\n");
381 #if defined DEBUG && DEBUG > 1
382 mincore_print(rrd_file, "before");
385 /* ignoring errors from RRDs that are smaller then the file_len+rounding */
386 rra_start = rrd_file->header_len;
387 dontneed_start = PAGE_START(rra_start) + _page_size;
388 for (i = 0; i < rrd->stat_head->rra_cnt; ++i) {
391 + rrd->rra_ptr[i].cur_row
392 * rrd->stat_head->ds_cnt * sizeof(rrd_value_t));
393 if (active_block > dontneed_start) {
395 madvise(rrd_file->file_start + dontneed_start,
396 active_block - dontneed_start - 1, MADV_DONTNEED);
398 /* in linux at least only fadvise DONTNEED seems to purge pages from cache */
399 #ifdef HAVE_POSIX_FADVISE
400 posix_fadvise(rrd_file->fd, dontneed_start,
401 active_block - dontneed_start - 1,
402 POSIX_FADV_DONTNEED);
405 dontneed_start = active_block;
406 /* do not release 'hot' block if update for this RAA will occur
407 * within 10 minutes */
408 if (rrd->stat_head->pdp_step * rrd->rra_def[i].pdp_cnt -
409 rrd->live_head->last_up % (rrd->stat_head->pdp_step *
410 rrd->rra_def[i].pdp_cnt) < 10 * 60) {
411 dontneed_start += _page_size;
414 rrd->rra_def[i].row_cnt * rrd->stat_head->ds_cnt *
418 if (dontneed_start < rrd_file->file_len) {
420 madvise(rrd_file->file_start + dontneed_start,
421 rrd_file->file_len - dontneed_start, MADV_DONTNEED);
423 #ifdef HAVE_POSIX_FADVISE
424 posix_fadvise(rrd_file->fd, dontneed_start,
425 rrd_file->file_len - dontneed_start,
426 POSIX_FADV_DONTNEED);
430 #if defined DEBUG && DEBUG > 1
431 mincore_print(rrd_file, "after");
433 #endif /* without madvise and posix_fadvise ist does not make much sense todo anything */
441 rrd_file_t *rrd_file)
446 ret = msync(rrd_file->file_start, rrd_file->file_len, MS_ASYNC);
448 rrd_set_error("msync rrd_file: %s", rrd_strerror(errno));
449 ret = munmap(rrd_file->file_start, rrd_file->file_len);
451 rrd_set_error("munmap rrd_file: %s", rrd_strerror(errno));
453 ret = close(rrd_file->fd);
455 rrd_set_error("closing file: %s", rrd_strerror(errno));
462 /* Set position of rrd_file. */
465 rrd_file_t *rrd_file,
472 if (whence == SEEK_SET)
474 else if (whence == SEEK_CUR)
475 rrd_file->pos += off;
476 else if (whence == SEEK_END)
477 rrd_file->pos = rrd_file->file_len + off;
479 ret = lseek(rrd_file->fd, off, whence);
481 rrd_set_error("lseek: %s", rrd_strerror(errno));
484 /* mimic fseek, which returns 0 upon success */
485 return ret < 0; /*XXX: or just ret to mimic lseek */
489 /* Get current position in rrd_file. */
492 rrd_file_t *rrd_file)
494 return rrd_file->pos;
498 /* Read count bytes into buffer buf, starting at rrd_file->pos.
499 * Returns the number of bytes read or <0 on error. */
502 rrd_file_t *rrd_file,
510 if (rrd_file->pos > rrd_file->file_len || _cnt == 0) /* EOF */
513 return -1; /* EINVAL */
514 _surplus = rrd_file->pos + _cnt - rrd_file->file_len;
515 if (_surplus > 0) { /* short read */
520 buf = memcpy(buf, rrd_file->file_start + rrd_file->pos, _cnt);
522 rrd_file->pos += _cnt; /* mimmic read() semantics */
527 ret = read(rrd_file->fd, buf, count);
529 rrd_file->pos += ret; /* mimmic read() semantics */
535 /* Write count bytes from buffer buf to the current position
536 * rrd_file->pos of rrd_file->fd.
537 * Returns the number of bytes written or <0 on error. */
540 rrd_file_t *rrd_file,
545 /* These flags are used if creating a new RRD */
546 int mm_prot = PROT_READ | PROT_WRITE, mm_flags = MAP_SHARED;
547 int old_size = rrd_file->file_len;
548 int new_size = rrd_file->file_len;
552 return -1; /* EINVAL */
554 if((rrd_file->pos + count) > old_size)
556 new_size = rrd_file->pos + count;
557 rrd_file->file_len = new_size;
558 lseek(rrd_file->fd, new_size - 1, SEEK_SET);
559 write(rrd_file->fd, "\0", 1); /* poke */
560 lseek(rrd_file->fd, 0, SEEK_SET);
561 if(rrd_file->file_start == NULL)
563 rrd_file->file_start = mmap(0, new_size, mm_prot, mm_flags,
567 rrd_file->file_start = mremap(rrd_file->file_start, old_size, new_size, MREMAP_MAYMOVE);
569 if (rrd_file->file_start == MAP_FAILED) {
570 rrd_set_error("m(re)maping file : %s",
571 rrd_strerror(errno));
575 memcpy(rrd_file->file_start + rrd_file->pos, buf, count);
576 rrd_file->pos += count;
577 return count; /* mimmic write() semantics */
579 ssize_t _sz = write(rrd_file->fd, buf, count);
582 rrd_file->pos += _sz;
588 /* flush all data pending to be written to FD. */
591 rrd_file_t *rrd_file)
593 if (fdatasync(rrd_file->fd) != 0) {
594 rrd_set_error("flushing fd %d: %s", rrd_file->fd,
595 rrd_strerror(errno));
600 /* Initialize RRD header. */
605 rrd->stat_head = NULL;
608 rrd->live_head = NULL;
609 rrd->legacy_last_up = NULL;
611 rrd->pdp_prep = NULL;
612 rrd->cdp_prep = NULL;
613 rrd->rrd_value = NULL;
617 /* free RRD header data. */
623 if (rrd->legacy_last_up) { /* this gets set for version < 3 only */
624 free(rrd->live_head);
631 free(rrd->live_head);
632 free(rrd->stat_head);
638 free(rrd->rrd_value);
643 /* routine used by external libraries to free memory allocated by