1 /*****************************************************************************
2 * RRDtool 1.2.99907080300 Copyright by Tobi Oetiker, 1997-2007
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 (dst) = (dst_t*)(void*) (data + offset); \
26 offset += sizeof(dst_t) * (cnt)
27 #else
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; \
32 } \
33 offset += read (rrd_file->fd, dst, sizeof(dst_t)*(cnt))
34 #endif
36 /* get the address of the start of this page */
37 #ifndef PAGE_START
38 #define PAGE_START(addr) ((addr)&(~(_page_size-1)))
39 #endif
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.
46 */
48 rrd_file_t *rrd_open(
49 const char *const file_name,
50 rrd_t *rrd,
51 unsigned rdwr)
52 {
53 int flags = 0;
54 mode_t mode = S_IRUSR;
55 int version;
57 #ifdef HAVE_MMAP
58 ssize_t _page_size = sysconf(_SC_PAGESIZE);
59 int mm_prot = PROT_READ, mm_flags = 0;
60 char *data;
61 #endif
62 off_t offset = 0;
63 struct stat statb;
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;
71 free(rrd->stat_head);
72 }
73 rrd_init(rrd);
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);
77 return NULL;
78 }
79 memset(rrd_file, 0, sizeof(rrd_file_t));
81 #ifdef DEBUG
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");
86 exit(-1);
87 }
88 #endif
89 if (rdwr & RRD_READONLY) {
90 flags |= O_RDONLY;
91 #ifdef HAVE_MMAP
92 mm_flags = MAP_PRIVATE;
93 # ifdef MAP_NORESERVE
94 mm_flags |= MAP_NORESERVE; /* readonly, so no swap backing needed */
95 # endif
96 #endif
97 } else {
98 if (rdwr & RRD_READWRITE) {
99 mode |= S_IWUSR;
100 flags |= O_RDWR;
101 #ifdef HAVE_MMAP
102 mm_flags = MAP_SHARED;
103 mm_prot |= PROT_WRITE;
104 #endif
105 }
106 if (rdwr & RRD_CREAT) {
107 flags |= (O_CREAT | O_TRUNC);
108 }
109 }
110 if (rdwr & RRD_READAHEAD) {
111 #ifdef MAP_POPULATE
112 mm_flags |= MAP_POPULATE; /* populate ptes and data */
113 #endif
114 #if defined MAP_NONBLOCK
115 mm_flags |= MAP_NONBLOCK; /* just populate ptes */
116 #endif
117 }
119 if ((rrd_file->fd = open(file_name, flags, mode)) < 0) {
120 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
121 goto out_free;
122 }
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));
128 goto out_close;
129 }
130 if (newfile_size == 0) {
131 rrd_file->file_len = statb.st_size;
132 } else {
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);
137 }
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);
144 #endif
146 /*
147 if (rdwr & RRD_READWRITE)
148 {
149 if (setvbuf((rrd_file->fd),NULL,_IONBF,2)) {
150 rrd_set_error("failed to disable the stream buffer\n");
151 return (-1);
152 }
153 }
154 */
155 #ifdef HAVE_MMAP
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));
163 goto out_close;
164 }
165 rrd_file->file_start = data;
166 if (rdwr & RRD_CREAT) {
167 memset(data, DNAN, newfile_size - 1);
168 goto out_done;
169 }
170 #endif
171 if (rdwr & RRD_CREAT)
172 goto out_done;
173 #ifdef USE_MADVISE
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);
177 } else {
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);
182 }
183 #endif
185 __rrd_read(rrd->stat_head, stat_head_t, 1);
187 /* lets do some test if we are on track ... */
188 if (memcmp(rrd->stat_head->cookie, RRD_COOKIE, sizeof(RRD_COOKIE)) != 0) {
189 rrd_set_error("'%s' is not an RRD file", file_name);
190 goto out_nullify_head;
191 }
193 if (rrd->stat_head->float_cookie != FLOAT_COOKIE) {
194 rrd_set_error("This RRD was created on another architecture");
195 goto out_nullify_head;
196 }
198 version = atoi(rrd->stat_head->version);
200 if (version > atoi(RRD_VERSION)) {
201 rrd_set_error("can't handle RRD file version %s",
202 rrd->stat_head->version);
203 goto out_nullify_head;
204 }
205 #if defined USE_MADVISE
206 /* the ds_def will be needed soonish, so hint accordingly */
207 madvise(data + PAGE_START(offset),
208 sizeof(ds_def_t) * rrd->stat_head->ds_cnt, MADV_WILLNEED);
209 #endif
210 __rrd_read(rrd->ds_def, ds_def_t,
211 rrd->stat_head->ds_cnt);
213 #if defined USE_MADVISE
214 /* the rra_def will be needed soonish, so hint accordingly */
215 madvise(data + PAGE_START(offset),
216 sizeof(rra_def_t) * rrd->stat_head->rra_cnt, MADV_WILLNEED);
217 #endif
218 __rrd_read(rrd->rra_def, rra_def_t,
219 rrd->stat_head->rra_cnt);
221 /* handle different format for the live_head */
222 if (version < 3) {
223 rrd->live_head = (live_head_t *) malloc(sizeof(live_head_t));
224 if (rrd->live_head == NULL) {
225 rrd_set_error("live_head_t malloc");
226 goto out_close;
227 }
228 #ifdef HAVE_MMAP
229 memmove(&rrd->live_head->last_up, data + offset, sizeof(long));
230 offset += sizeof(long);
231 #else
232 offset += read(rrd_file->fd, &rrd->live_head->last_up, sizeof(long));
233 #endif
234 rrd->live_head->last_up_usec = 0;
235 } else {
236 #if defined USE_MADVISE
237 /* the live_head will be needed soonish, so hint accordingly */
238 madvise(data + PAGE_START(offset),
239 sizeof(live_head_t), MADV_WILLNEED);
240 #endif
241 __rrd_read(rrd->live_head, live_head_t,
242 1);
243 }
244 //XXX: This doesn't look like it needs madvise
245 __rrd_read(rrd->pdp_prep, pdp_prep_t,
246 rrd->stat_head->ds_cnt);
248 //XXX: This could benefit from madvise()ing
249 __rrd_read(rrd->cdp_prep, cdp_prep_t,
250 rrd->stat_head->rra_cnt * rrd->stat_head->ds_cnt);
252 //XXX: This could benefit from madvise()ing
253 __rrd_read(rrd->rra_ptr, rra_ptr_t,
254 rrd->stat_head->rra_cnt);
256 rrd_file->header_len = offset;
257 rrd_file->pos = offset;
258 out_done:
259 return (rrd_file);
260 out_nullify_head:
261 rrd->stat_head = NULL;
262 out_close:
263 close(rrd_file->fd);
264 out_free:
265 free(rrd_file);
266 return NULL;
267 }
270 /* Print list of in-core pages of a the current rrd_file. */
271 static
272 void mincore_print(
273 rrd_file_t *rrd_file,
274 char *mark)
275 {
276 #ifdef HAVE_MMAP
277 /* pretty print blocks in core */
278 off_t off;
279 unsigned char *vec;
280 ssize_t _page_size = sysconf(_SC_PAGESIZE);
282 off = rrd_file->file_len +
283 ((rrd_file->file_len + _page_size - 1) / _page_size);
284 vec = malloc(off);
285 if (vec != NULL) {
286 memset(vec, 0, off);
287 if (mincore(rrd_file->file_start, rrd_file->file_len, vec) == 0) {
288 int prev;
289 unsigned is_in = 0, was_in = 0;
291 for (off = 0, prev = 0; off < rrd_file->file_len; ++off) {
292 is_in = vec[off] & 1; /* if lsb set then is core resident */
293 if (off == 0)
294 was_in = is_in;
295 if (was_in != is_in) {
296 fprintf(stderr, "%s: %sin core: %p len %ld\n", mark,
297 was_in ? "" : "not ", vec + prev, off - prev);
298 was_in = is_in;
299 prev = off;
300 }
301 }
302 fprintf(stderr,
303 "%s: %sin core: %p len %ld\n", mark,
304 was_in ? "" : "not ", vec + prev, off - prev);
305 } else
306 fprintf(stderr, "mincore: %s", rrd_strerror(errno));
307 }
308 #else
309 fprintf(stderr, "sorry mincore only works with mmap");
310 #endif
311 }
314 /* drop cache except for the header and the active pages */
315 void rrd_dontneed(
316 rrd_file_t *rrd_file,
317 rrd_t *rrd)
318 {
319 unsigned long dontneed_start;
320 unsigned long rra_start;
321 unsigned long active_block;
322 unsigned long i;
323 ssize_t _page_size = sysconf(_SC_PAGESIZE);
325 #if defined DEBUG && DEBUG > 1
326 mincore_print(rrd_file, "before");
327 #endif
329 /* ignoring errors from RRDs that are smaller then the file_len+rounding */
330 rra_start = rrd_file->header_len;
331 dontneed_start = PAGE_START(rra_start) + _page_size;
332 for (i = 0; i < rrd->stat_head->rra_cnt; ++i) {
333 active_block =
334 PAGE_START(rra_start
335 + rrd->rra_ptr[i].cur_row
336 * rrd->stat_head->ds_cnt * sizeof(rrd_value_t));
337 if (active_block > dontneed_start) {
338 #ifdef USE_MADVISE
339 madvise(rrd_file->file_start + dontneed_start,
340 active_block - dontneed_start - 1, MADV_DONTNEED);
341 #endif
342 /* in linux at least only fadvise DONTNEED seems to purge pages from cache */
343 #ifdef HAVE_POSIX_FADVISE
344 posix_fadvise(rrd_file->fd, dontneed_start,
345 active_block - dontneed_start - 1,
346 POSIX_FADV_DONTNEED);
347 #endif
348 }
349 dontneed_start = active_block;
350 /* do not release 'hot' block if update for this RAA will occur
351 * within 10 minutes */
352 if (rrd->stat_head->pdp_step * rrd->rra_def[i].pdp_cnt -
353 rrd->live_head->last_up % (rrd->stat_head->pdp_step *
354 rrd->rra_def[i].pdp_cnt) < 10 * 60) {
355 dontneed_start += _page_size;
356 }
357 rra_start +=
358 rrd->rra_def[i].row_cnt * rrd->stat_head->ds_cnt *
359 sizeof(rrd_value_t);
360 }
361 #ifdef USE_MADVISE
362 madvise(rrd_file->file_start + dontneed_start,
363 rrd_file->file_len - dontneed_start, MADV_DONTNEED);
364 #endif
365 #ifdef HAVE_POSIX_FADVISE
366 posix_fadvise(rrd_file->fd, dontneed_start,
367 rrd_file->file_len - dontneed_start, POSIX_FADV_DONTNEED);
368 #endif
369 #if defined DEBUG && DEBUG > 1
370 mincore_print(rrd_file, "after");
371 #endif
372 }
374 int rrd_close(
375 rrd_file_t *rrd_file)
376 {
377 int ret;
379 #ifdef HAVE_MMAP
380 ret = msync(rrd_file->file_start, rrd_file->file_len,MS_ASYNC);
381 if (ret != 0)
382 rrd_set_error("msync rrd_file: %s", rrd_strerror(errno));
383 ret = munmap(rrd_file->file_start, rrd_file->file_len);
384 if (ret != 0)
385 rrd_set_error("munmap rrd_file: %s", rrd_strerror(errno));
386 #endif
387 ret = close(rrd_file->fd);
388 if (ret != 0)
389 rrd_set_error("closing file: %s", rrd_strerror(errno));
390 free(rrd_file);
391 rrd_file = NULL;
392 return ret;
393 }
396 /* Set position of rrd_file. */
398 off_t rrd_seek(
399 rrd_file_t *rrd_file,
400 off_t off,
401 int whence)
402 {
403 off_t ret = 0;
405 #ifdef HAVE_MMAP
406 if (whence == SEEK_SET)
407 rrd_file->pos = off;
408 else if (whence == SEEK_CUR)
409 rrd_file->pos += off;
410 else if (whence == SEEK_END)
411 rrd_file->pos = rrd_file->file_len + off;
412 #else
413 ret = lseek(rrd_file->fd, off, whence);
414 if (ret < 0)
415 rrd_set_error("lseek: %s", rrd_strerror(errno));
416 rrd_file->pos = ret;
417 #endif
418 /* mimic fseek, which returns 0 upon success */
419 return ret < 0; //XXX: or just ret to mimic lseek
420 }
423 /* Get current position in rrd_file. */
425 inline off_t rrd_tell(
426 rrd_file_t *rrd_file)
427 {
428 return rrd_file->pos;
429 }
432 /* Read count bytes into buffer buf, starting at rrd_file->pos.
433 * Returns the number of bytes read or <0 on error. */
435 inline ssize_t rrd_read(
436 rrd_file_t *rrd_file,
437 void *buf,
438 size_t count)
439 {
440 #ifdef HAVE_MMAP
441 size_t _cnt = count;
442 ssize_t _surplus;
444 if (rrd_file->pos > rrd_file->file_len || _cnt == 0) /* EOF */
445 return 0;
446 if (buf == NULL)
447 return -1; /* EINVAL */
448 _surplus = rrd_file->pos + _cnt - rrd_file->file_len;
449 if (_surplus > 0) { /* short read */
450 _cnt -= _surplus;
451 }
452 if (_cnt == 0)
453 return 0; /* EOF */
454 buf = memcpy(buf, rrd_file->file_start + rrd_file->pos, _cnt);
456 rrd_file->pos += _cnt; /* mimmic read() semantics */
457 return _cnt;
458 #else
459 ssize_t ret;
461 ret = read(rrd_file->fd, buf, count);
462 if (ret > 0)
463 rrd_file->pos += ret; /* mimmic read() semantics */
464 return ret;
465 #endif
466 }
469 /* Write count bytes from buffer buf to the current position
470 * rrd_file->pos of rrd_file->fd.
471 * Returns the number of bytes written or <0 on error. */
473 inline ssize_t rrd_write(
474 rrd_file_t *rrd_file,
475 const void *buf,
476 size_t count)
477 {
478 #ifdef HAVE_MMAP
479 if (count == 0)
480 return 0;
481 if (buf == NULL)
482 return -1; /* EINVAL */
483 memcpy(rrd_file->file_start + rrd_file->pos, buf, count);
484 rrd_file->pos += count;
485 return count; /* mimmic write() semantics */
486 #else
487 ssize_t _sz = write(rrd_file->fd, buf, count);
489 if (_sz > 0)
490 rrd_file->pos += _sz;
491 return _sz;
492 #endif
493 }
496 /* flush all data pending to be written to FD. */
498 inline void rrd_flush(
499 rrd_file_t *rrd_file)
500 {
501 if (fdatasync(rrd_file->fd) != 0) {
502 rrd_set_error("flushing fd %d: %s", rrd_file->fd,
503 rrd_strerror(errno));
504 }
505 }
508 /* Initialize RRD header. */
510 void rrd_init(
511 rrd_t *rrd)
512 {
513 rrd->stat_head = NULL;
514 rrd->ds_def = NULL;
515 rrd->rra_def = NULL;
516 rrd->live_head = NULL;
517 rrd->rra_ptr = NULL;
518 rrd->pdp_prep = NULL;
519 rrd->cdp_prep = NULL;
520 rrd->rrd_value = NULL;
521 }
524 /* free RRD header data. */
526 #ifdef HAVE_MMAP
527 inline void rrd_free(
528 rrd_t UNUSED(*rrd))
529 {
530 }
531 #else
532 void rrd_free(
533 rrd_t *rrd)
534 {
535 free(rrd->live_head);
536 free(rrd->stat_head);
537 free(rrd->ds_def);
538 free(rrd->rra_def);
539 free(rrd->rra_ptr);
540 free(rrd->pdp_prep);
541 free(rrd->cdp_prep);
542 free(rrd->rrd_value);
543 }
544 #endif
547 /* routine used by external libraries to free memory allocated by
548 * rrd library */
550 void rrd_freemem(
551 void *mem)
552 {
553 free(mem);
554 }
557 /* XXX: FIXME: missing documentation. */
558 /*XXX: FIXME should be renamed to rrd_readfile or _rrd_readfile */
560 int /*_rrd_*/ readfile(
561 const char *file_name,
562 char **buffer,
563 int skipfirst)
564 {
565 long writecnt = 0, totalcnt = MEMBLK;
566 long offset = 0;
567 FILE *input = NULL;
568 char c;
570 if ((strcmp("-", file_name) == 0)) {
571 input = stdin;
572 } else {
573 if ((input = fopen(file_name, "rb")) == NULL) {
574 rrd_set_error("opening '%s': %s", file_name, rrd_strerror(errno));
575 return (-1);
576 }
577 }
578 if (skipfirst) {
579 do {
580 c = getc(input);
581 offset++;
582 } while (c != '\n' && !feof(input));
583 }
584 if (strcmp("-", file_name)) {
585 fseek(input, 0, SEEK_END);
586 /* have extra space for detecting EOF without realloc */
587 totalcnt = (ftell(input) + 1) / sizeof(char) - offset;
588 if (totalcnt < MEMBLK)
589 totalcnt = MEMBLK; /* sanitize */
590 fseek(input, offset * sizeof(char), SEEK_SET);
591 }
592 if (((*buffer) = (char *) malloc((totalcnt + 4) * sizeof(char))) == NULL) {
593 perror("Allocate Buffer:");
594 exit(1);
595 };
596 do {
597 writecnt +=
598 fread((*buffer) + writecnt, 1,
599 (totalcnt - writecnt) * sizeof(char), input);
600 if (writecnt >= totalcnt) {
601 totalcnt += MEMBLK;
602 if (((*buffer) =
603 rrd_realloc((*buffer),
604 (totalcnt + 4) * sizeof(char))) == NULL) {
605 perror("Realloc Buffer:");
606 exit(1);
607 };
608 }
609 } while (!feof(input));
610 (*buffer)[writecnt] = '\0';
611 if (strcmp("-", file_name) != 0) {
612 fclose(input);
613 };
614 return writecnt;
615 }