1 /*****************************************************************************
2 * RRDtool 1.3rc4 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 (dst) = (dst_t*)(void*) (data + offset); \
26 offset += sizeof(dst_t) * (cnt)
28 #define __rrd_read(dst, dst_t, cnt) \
29 if ((dst = malloc(sizeof(dst_t)*(cnt))) == NULL) { \
30 rrd_set_error(#dst " malloc"); \
31 goto out_nullify_head; \
33 offset += read (rrd_file->fd, dst, sizeof(dst_t)*(cnt))
36 /* get the address of the start of this page */
38 #define PAGE_START(addr) ((addr)&(~(_page_size-1)))
42 /* Open a database file, return its header and an open filehandle,
43 * positioned to the first cdp in the first rra.
44 * In the error path of rrd_open, only rrd_free(&rrd) has to be called
45 * before returning an error. Do not call rrd_close upon failure of rrd_open.
49 const char *const file_name,
54 mode_t mode = S_IRUSR;
58 ssize_t _page_size = sysconf(_SC_PAGESIZE);
59 int mm_prot = PROT_READ, mm_flags = 0;
64 rrd_file_t *rrd_file = NULL;
65 off_t newfile_size = 0;
67 if (rdwr & RRD_CREAT) {
68 /* yes bad inline signaling alert, we are using the
69 floatcookie to pass the size in ... only used in resize */
70 newfile_size = (off_t) rrd->stat_head->float_cookie;
74 rrd_file = malloc(sizeof(rrd_file_t));
75 if (rrd_file == NULL) {
76 rrd_set_error("allocating rrd_file descriptor for '%s'", file_name);
79 memset(rrd_file, 0, sizeof(rrd_file_t));
82 if ((rdwr & (RRD_READONLY | RRD_READWRITE)) ==
83 (RRD_READONLY | RRD_READWRITE)) {
84 /* Both READONLY and READWRITE were given, which is invalid. */
85 rrd_set_error("in read/write request mask");
89 if (rdwr & RRD_READONLY) {
92 mm_flags = MAP_PRIVATE;
94 mm_flags |= MAP_NORESERVE; /* readonly, so no swap backing needed */
98 if (rdwr & RRD_READWRITE) {
102 mm_flags = MAP_SHARED;
103 mm_prot |= PROT_WRITE;
106 if (rdwr & RRD_CREAT) {
107 flags |= (O_CREAT | O_TRUNC);
110 if (rdwr & RRD_READAHEAD) {
112 mm_flags |= MAP_POPULATE; /* populate ptes and data */
114 #if defined MAP_NONBLOCK
115 mm_flags |= MAP_NONBLOCK; /* just populate ptes */
119 if ((rrd_file->fd = open(file_name, flags, mode)) < 0) {
120 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
124 /* Better try to avoid seeks as much as possible. stat may be heavy but
125 * many concurrent seeks are even worse. */
126 if (newfile_size == 0 && ((fstat(rrd_file->fd, &statb)) < 0)) {
127 rrd_set_error("fstat '%s': %s", file_name, rrd_strerror(errno));
130 if (newfile_size == 0) {
131 rrd_file->file_len = statb.st_size;
133 rrd_file->file_len = newfile_size;
134 lseek(rrd_file->fd, newfile_size - 1, SEEK_SET);
135 write(rrd_file->fd, "\0", 1); /* poke */
136 lseek(rrd_file->fd, 0, SEEK_SET);
138 #ifdef HAVE_POSIX_FADVISE
139 /* In general we need no read-ahead when dealing with rrd_files.
140 When we stop reading, it is highly unlikely that we start up again.
141 In this manner we actually save time and diskaccess (and buffer cache).
142 Thanks to Dave Plonka for the Idea of using POSIX_FADV_RANDOM here. */
143 posix_fadvise(rrd_file->fd, 0, 0, POSIX_FADV_RANDOM);
147 if (rdwr & RRD_READWRITE)
149 if (setvbuf((rrd_file->fd),NULL,_IONBF,2)) {
150 rrd_set_error("failed to disable the stream buffer\n");
156 data = mmap(0, rrd_file->file_len, mm_prot, mm_flags,
157 rrd_file->fd, offset);
159 /* lets see if the first read worked */
160 if (data == MAP_FAILED) {
161 rrd_set_error("mmaping file '%s': %s", file_name,
162 rrd_strerror(errno));
165 rrd_file->file_start = data;
166 if (rdwr & RRD_CREAT) {
167 memset(data, DNAN, newfile_size - 1);
171 if (rdwr & RRD_CREAT)
174 if (rdwr & RRD_COPY) {
175 /* We will read everything in a moment (copying) */
176 madvise(data, rrd_file->file_len, MADV_WILLNEED | MADV_SEQUENTIAL);
178 /* We do not need to read anything in for the moment */
179 madvise(data, rrd_file->file_len, MADV_RANDOM);
180 /* the stat_head will be needed soonish, so hint accordingly */
181 madvise(data, sizeof(stat_head_t), MADV_WILLNEED | MADV_RANDOM);
185 __rrd_read(rrd->stat_head, stat_head_t,
188 /* lets do some test if we are on track ... */
189 if (memcmp(rrd->stat_head->cookie, RRD_COOKIE, sizeof(RRD_COOKIE)) != 0) {
190 rrd_set_error("'%s' is not an RRD file", file_name);
191 goto out_nullify_head;
194 if (rrd->stat_head->float_cookie != FLOAT_COOKIE) {
195 rrd_set_error("This RRD was created on another architecture");
196 goto out_nullify_head;
199 version = atoi(rrd->stat_head->version);
201 if (version > atoi(RRD_VERSION)) {
202 rrd_set_error("can't handle RRD file version %s",
203 rrd->stat_head->version);
204 goto out_nullify_head;
206 #if defined USE_MADVISE
207 /* the ds_def will be needed soonish, so hint accordingly */
208 madvise(data + PAGE_START(offset),
209 sizeof(ds_def_t) * rrd->stat_head->ds_cnt, MADV_WILLNEED);
211 __rrd_read(rrd->ds_def, ds_def_t,
212 rrd->stat_head->ds_cnt);
214 #if defined USE_MADVISE
215 /* the rra_def will be needed soonish, so hint accordingly */
216 madvise(data + PAGE_START(offset),
217 sizeof(rra_def_t) * rrd->stat_head->rra_cnt, MADV_WILLNEED);
219 __rrd_read(rrd->rra_def, rra_def_t,
220 rrd->stat_head->rra_cnt);
222 /* handle different format for the live_head */
224 rrd->live_head = (live_head_t *) malloc(sizeof(live_head_t));
225 if (rrd->live_head == NULL) {
226 rrd_set_error("live_head_t malloc");
230 memmove(&rrd->live_head->last_up, data + offset, sizeof(long));
231 offset += sizeof(long);
233 offset += read(rrd_file->fd, &rrd->live_head->last_up, sizeof(long));
235 rrd->live_head->last_up_usec = 0;
237 #if defined USE_MADVISE
238 /* the live_head will be needed soonish, so hint accordingly */
239 madvise(data + PAGE_START(offset),
240 sizeof(live_head_t), MADV_WILLNEED);
242 __rrd_read(rrd->live_head, live_head_t,
245 __rrd_read(rrd->pdp_prep, pdp_prep_t,
246 rrd->stat_head->ds_cnt);
247 __rrd_read(rrd->cdp_prep, cdp_prep_t,
248 rrd->stat_head->rra_cnt * rrd->stat_head->ds_cnt);
249 __rrd_read(rrd->rra_ptr, rra_ptr_t,
250 rrd->stat_head->rra_cnt);
252 rrd_file->header_len = offset;
253 rrd_file->pos = offset;
257 rrd->stat_head = NULL;
266 #if defined DEBUG && DEBUG > 1
267 /* Print list of in-core pages of a the current rrd_file. */
270 rrd_file_t *rrd_file,
274 /* pretty print blocks in core */
277 ssize_t _page_size = sysconf(_SC_PAGESIZE);
279 off = rrd_file->file_len +
280 ((rrd_file->file_len + _page_size - 1) / _page_size);
284 if (mincore(rrd_file->file_start, rrd_file->file_len, vec) == 0) {
286 unsigned is_in = 0, was_in = 0;
288 for (off = 0, prev = 0; off < rrd_file->file_len; ++off) {
289 is_in = vec[off] & 1; /* if lsb set then is core resident */
292 if (was_in != is_in) {
293 fprintf(stderr, "%s: %sin core: %p len %ld\n", mark,
294 was_in ? "" : "not ", vec + prev, off - prev);
300 "%s: %sin core: %p len %ld\n", mark,
301 was_in ? "" : "not ", vec + prev, off - prev);
303 fprintf(stderr, "mincore: %s", rrd_strerror(errno));
306 fprintf(stderr, "sorry mincore only works with mmap");
309 #endif /* defined DEBUG && DEBUG > 1 */
312 /* drop cache except for the header and the active pages */
314 rrd_file_t *rrd_file,
317 unsigned long dontneed_start;
318 unsigned long rra_start;
319 unsigned long active_block;
321 ssize_t _page_size = sysconf(_SC_PAGESIZE);
323 #if defined DEBUG && DEBUG > 1
324 mincore_print(rrd_file, "before");
327 /* ignoring errors from RRDs that are smaller then the file_len+rounding */
328 rra_start = rrd_file->header_len;
329 dontneed_start = PAGE_START(rra_start) + _page_size;
330 for (i = 0; i < rrd->stat_head->rra_cnt; ++i) {
333 + rrd->rra_ptr[i].cur_row
334 * rrd->stat_head->ds_cnt * sizeof(rrd_value_t));
335 if (active_block > dontneed_start) {
337 madvise(rrd_file->file_start + dontneed_start,
338 active_block - dontneed_start - 1, MADV_DONTNEED);
340 /* in linux at least only fadvise DONTNEED seems to purge pages from cache */
341 #ifdef HAVE_POSIX_FADVISE
342 posix_fadvise(rrd_file->fd, dontneed_start,
343 active_block - dontneed_start - 1,
344 POSIX_FADV_DONTNEED);
347 dontneed_start = active_block;
348 /* do not release 'hot' block if update for this RAA will occur
349 * within 10 minutes */
350 if (rrd->stat_head->pdp_step * rrd->rra_def[i].pdp_cnt -
351 rrd->live_head->last_up % (rrd->stat_head->pdp_step *
352 rrd->rra_def[i].pdp_cnt) < 10 * 60) {
353 dontneed_start += _page_size;
356 rrd->rra_def[i].row_cnt * rrd->stat_head->ds_cnt *
360 madvise(rrd_file->file_start + dontneed_start,
361 rrd_file->file_len - dontneed_start, MADV_DONTNEED);
363 #ifdef HAVE_POSIX_FADVISE
364 posix_fadvise(rrd_file->fd, dontneed_start,
365 rrd_file->file_len - dontneed_start, POSIX_FADV_DONTNEED);
367 #if defined DEBUG && DEBUG > 1
368 mincore_print(rrd_file, "after");
373 rrd_file_t *rrd_file)
378 ret = msync(rrd_file->file_start, rrd_file->file_len, MS_ASYNC);
380 rrd_set_error("msync rrd_file: %s", rrd_strerror(errno));
381 ret = munmap(rrd_file->file_start, rrd_file->file_len);
383 rrd_set_error("munmap rrd_file: %s", rrd_strerror(errno));
385 ret = close(rrd_file->fd);
387 rrd_set_error("closing file: %s", rrd_strerror(errno));
394 /* Set position of rrd_file. */
397 rrd_file_t *rrd_file,
404 if (whence == SEEK_SET)
406 else if (whence == SEEK_CUR)
407 rrd_file->pos += off;
408 else if (whence == SEEK_END)
409 rrd_file->pos = rrd_file->file_len + off;
411 ret = lseek(rrd_file->fd, off, whence);
413 rrd_set_error("lseek: %s", rrd_strerror(errno));
416 /* mimic fseek, which returns 0 upon success */
417 return ret < 0; /*XXX: or just ret to mimic lseek */
421 /* Get current position in rrd_file. */
423 inline off_t rrd_tell(
424 rrd_file_t *rrd_file)
426 return rrd_file->pos;
430 /* Read count bytes into buffer buf, starting at rrd_file->pos.
431 * Returns the number of bytes read or <0 on error. */
433 inline ssize_t rrd_read(
434 rrd_file_t *rrd_file,
442 if (rrd_file->pos > rrd_file->file_len || _cnt == 0) /* EOF */
445 return -1; /* EINVAL */
446 _surplus = rrd_file->pos + _cnt - rrd_file->file_len;
447 if (_surplus > 0) { /* short read */
452 buf = memcpy(buf, rrd_file->file_start + rrd_file->pos, _cnt);
454 rrd_file->pos += _cnt; /* mimmic read() semantics */
459 ret = read(rrd_file->fd, buf, count);
461 rrd_file->pos += ret; /* mimmic read() semantics */
467 /* Write count bytes from buffer buf to the current position
468 * rrd_file->pos of rrd_file->fd.
469 * Returns the number of bytes written or <0 on error. */
471 inline ssize_t rrd_write(
472 rrd_file_t *rrd_file,
480 return -1; /* EINVAL */
481 memcpy(rrd_file->file_start + rrd_file->pos, buf, count);
482 rrd_file->pos += count;
483 return count; /* mimmic write() semantics */
485 ssize_t _sz = write(rrd_file->fd, buf, count);
488 rrd_file->pos += _sz;
494 /* flush all data pending to be written to FD. */
496 inline void rrd_flush(
497 rrd_file_t *rrd_file)
499 if (fdatasync(rrd_file->fd) != 0) {
500 rrd_set_error("flushing fd %d: %s", rrd_file->fd,
501 rrd_strerror(errno));
506 /* Initialize RRD header. */
511 rrd->stat_head = NULL;
514 rrd->live_head = NULL;
516 rrd->pdp_prep = NULL;
517 rrd->cdp_prep = NULL;
518 rrd->rrd_value = NULL;
522 /* free RRD header data. */
525 inline void rrd_free(
533 free(rrd->live_head);
534 free(rrd->stat_head);
540 free(rrd->rrd_value);
545 /* routine used by external libraries to free memory allocated by
555 /* XXX: FIXME: missing documentation. */
556 /*XXX: FIXME should be renamed to rrd_readfile or _rrd_readfile */
558 int /*_rrd_*/ readfile(
559 const char *file_name,
563 long writecnt = 0, totalcnt = MEMBLK;
568 if ((strcmp("-", file_name) == 0)) {
571 if ((input = fopen(file_name, "rb")) == NULL) {
572 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
580 } while (c != '\n' && !feof(input));
582 if (strcmp("-", file_name)) {
583 fseek(input, 0, SEEK_END);
584 /* have extra space for detecting EOF without realloc */
585 totalcnt = (ftell(input) + 1) / sizeof(char) - offset;
586 if (totalcnt < MEMBLK)
587 totalcnt = MEMBLK; /* sanitize */
588 fseek(input, offset * sizeof(char), SEEK_SET);
590 if (((*buffer) = (char *) malloc((totalcnt + 4) * sizeof(char))) == NULL) {
591 perror("Allocate Buffer:");
596 fread((*buffer) + writecnt, 1,
597 (totalcnt - writecnt) * sizeof(char), input);
598 if (writecnt >= totalcnt) {
601 rrd_realloc((*buffer),
602 (totalcnt + 4) * sizeof(char))) == NULL) {
603 perror("Realloc Buffer:");
607 } while (!feof(input));
608 (*buffer)[writecnt] = '\0';
609 if (strcmp("-", file_name) != 0) {