1 /* Code to load locale data from the locale archive file.
2 Copyright (C) 2002-2023 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <https://www.gnu.org/licenses/>. */
18
19 #include <locale.h>
20 #include <stddef.h>
21 #include <stdlib.h>
22 #include <stdbool.h>
23 #include <errno.h>
24 #include <assert.h>
25 #include <string.h>
26 #include <fcntl.h>
27 #include <unistd.h>
28 #include <stdint.h>
29 #include <sys/mman.h>
30 #include <sys/stat.h>
31 #include <sys/param.h>
32
33 #include "localeinfo.h"
34 #include "locarchive.h"
35 #include <not-cancel.h>
36
37 /* Define the hash function. We define the function as static inline. */
38 #define compute_hashval static inline compute_hashval
39 #define hashval_t uint32_t
40 #include "hashval.h"
41 #undef compute_hashval
42
43
44 /* Name of the locale archive file. */
45 static const char archfname[] = COMPLOCALEDIR "/locale-archive";
46
47 /* Size of initial mapping window, optimal if large enough to
48 cover the header plus the initial locale. */
49 #define ARCHIVE_MAPPING_WINDOW (2 * 1024 * 1024)
50
51 #ifndef MAP_COPY
52 /* This is not quite as good as MAP_COPY since unexamined pages
53 can change out from under us and give us inconsistent data.
54 But we rely on the user not to diddle the system's live archive.
55 Even though we only ever use PROT_READ, using MAP_SHARED would
56 not give the system sufficient freedom to e.g. let the on disk
57 file go away because it doesn't know we won't call mprotect later. */
58 # define MAP_COPY MAP_PRIVATE
59 #endif
60 #ifndef MAP_FILE
61 /* Some systems do not have this flag; it is superfluous. */
62 # define MAP_FILE 0
63 #endif
64
65 /* Record of contiguous pages already mapped from the locale archive. */
66 struct archmapped
67 {
68 void *ptr;
69 uint32_t from;
70 uint32_t len;
71 struct archmapped *next;
72 };
73 static struct archmapped *archmapped;
74
75 /* This describes the mapping at the beginning of the file that contains
76 the header data. There could be data in the following partial page,
77 so this is searched like any other. Once the archive has been used,
78 ARCHMAPPED points to this; if mapping the archive header failed,
79 then headmap.ptr is null. */
80 static struct archmapped headmap;
81 static struct __stat64_t64 archive_stat; /* stat of archive when header mapped. */
82
83 /* Record of locales that we have already loaded from the archive. */
84 struct locale_in_archive
85 {
86 struct locale_in_archive *next;
87 char *name;
88 struct __locale_data *data[__LC_LAST];
89 };
90 static struct locale_in_archive *archloaded;
91
92
93 /* Local structure and subroutine of _nl_load_archive, see below. */
94 struct range
95 {
96 uint32_t from;
97 uint32_t len;
98 int category;
99 void *result;
100 };
101
102 static int
103 rangecmp (const void *p1, const void *p2)
104 {
105 return ((struct range *) p1)->from - ((struct range *) p2)->from;
106 }
107
108
109 /* Calculate the amount of space needed for all the tables described
110 by the given header. Note we do not include the empty table space
111 that has been preallocated in the file, so our mapping may not be
112 large enough if localedef adds data to the file in place. However,
113 doing that would permute the header fields while we are accessing
114 them and thus not be safe anyway, so we don't allow for that. */
115 static inline off_t
116 calculate_head_size (const struct locarhead *h)
117 {
118 off_t namehash_end = (h->namehash_offset
119 + h->namehash_size * sizeof (struct namehashent));
120 off_t string_end = h->string_offset + h->string_used;
121 off_t locrectab_end = (h->locrectab_offset
122 + h->locrectab_used * sizeof (struct locrecent));
123 return MAX (namehash_end, MAX (string_end, locrectab_end));
124 }
125
126
127 /* Find the locale *NAMEP in the locale archive, and return the
128 internalized data structure for its CATEGORY data. If this locale has
129 already been loaded from the archive, just returns the existing data
130 structure. If successful, sets *NAMEP to point directly into the mapped
131 archive string table; that way, the next call can short-circuit strcmp. */
132 struct __locale_data *
133 _nl_load_locale_from_archive (int category, const char **namep)
134 {
135 const char *name = *namep;
136 struct
137 {
138 void *addr;
139 size_t len;
140 } results[__LC_LAST];
141 struct locale_in_archive *lia;
142 struct locarhead *head;
143 struct namehashent *namehashtab;
144 struct locrecent *locrec;
145 struct archmapped *mapped;
146 struct archmapped *last;
147 unsigned long int hval;
148 size_t idx;
149 size_t incr;
150 struct range ranges[__LC_LAST - 1];
151 int nranges;
152 int cnt;
153 size_t ps = __sysconf (_SC_PAGE_SIZE);
154 int fd = -1;
155
156 /* Check if we have already loaded this locale from the archive.
157 If we previously loaded the locale but found bogons in the data,
158 then we will have stored a null pointer to return here. */
159 for (lia = archloaded; lia != NULL; lia = lia->next)
160 if (name == lia->name || !strcmp (name, lia->name))
161 {
162 *namep = lia->name;
163 return lia->data[category];
164 }
165
166 {
167 /* If the name contains a codeset, then we normalize the name before
168 doing the lookup. */
169 const char *p = strchr (name, '.');
170 if (p != NULL && p[1] != '@' && p[1] != '\0')
171 {
172 const char *rest = __strchrnul (++p, '@');
173 const char *normalized_codeset = _nl_normalize_codeset (p, rest - p);
174 if (normalized_codeset == NULL) /* malloc failure */
175 return NULL;
176 if (strncmp (normalized_codeset, p, rest - p) != 0
177 || normalized_codeset[rest - p] != '\0')
178 {
179 /* There is a normalized codeset name that is different from
180 what was specified; reconstruct a new locale name using it. */
181 size_t normlen = strlen (normalized_codeset);
182 size_t restlen = strlen (rest) + 1;
183 char *newname = alloca (p - name + normlen + restlen);
184 memcpy (__mempcpy (__mempcpy (newname, name, p - name),
185 normalized_codeset, normlen),
186 rest, restlen);
187 name = newname;
188 }
189 free ((char *) normalized_codeset);
190 }
191 }
192
193 /* Make sure the archive is loaded. */
194 if (archmapped == NULL)
195 {
196 void *result;
197 size_t headsize, mapsize;
198
199 /* We do this early as a sign that we have tried to open the archive.
200 If headmap.ptr remains null, that's an indication that we tried
201 and failed, so we won't try again. */
202 archmapped = &headmap;
203
204 /* The archive has never been opened. */
205 fd = __open_nocancel (archfname, O_RDONLY|O_LARGEFILE|O_CLOEXEC);
206 if (fd < 0)
207 /* Cannot open the archive, for whatever reason. */
208 return NULL;
209
210 if (__fstat64_time64 (fd, &archive_stat) == -1)
211 {
212 /* stat failed, very strange. */
213 close_and_out:
214 if (fd >= 0)
215 __close_nocancel_nostatus (fd);
216 return NULL;
217 }
218
219
220 /* Map an initial window probably large enough to cover the header
221 and the first locale's data. With a large address space, we can
222 just map the whole file and be sure everything is covered. */
223
224 mapsize = (sizeof (void *) > 4 ? archive_stat.st_size
225 : MIN (archive_stat.st_size, ARCHIVE_MAPPING_WINDOW));
226
227 result = __mmap64 (NULL, mapsize, PROT_READ, MAP_FILE|MAP_COPY, fd, 0);
228 if (result == MAP_FAILED)
229 goto close_and_out;
230
231 /* Check whether the file is large enough for the sizes given in
232 the header. Theoretically an archive could be so large that
233 just the header fails to fit in our initial mapping window. */
234 headsize = calculate_head_size ((const struct locarhead *) result);
235 if (headsize > mapsize)
236 {
237 (void) __munmap (result, mapsize);
238 if (sizeof (void *) > 4 || headsize > archive_stat.st_size)
239 /* The file is not big enough for the header. Bogus. */
240 goto close_and_out;
241
242 /* Freakishly long header. */
243 /* XXX could use mremap when available */
244 mapsize = (headsize + ps - 1) & ~(ps - 1);
245 result = __mmap64 (NULL, mapsize, PROT_READ, MAP_FILE|MAP_COPY,
246 fd, 0);
247 if (result == MAP_FAILED)
248 goto close_and_out;
249 }
250
251 if (sizeof (void *) > 4 || mapsize >= archive_stat.st_size)
252 {
253 /* We've mapped the whole file already, so we can be
254 sure we won't need this file descriptor later. */
255 __close_nocancel_nostatus (fd);
256 fd = -1;
257 }
258
259 headmap.ptr = result;
260 /* headmap.from already initialized to zero. */
261 headmap.len = mapsize;
262 }
263
264 /* If there is no archive or it cannot be loaded for some reason fail. */
265 if (__glibc_unlikely (headmap.ptr == NULL))
266 goto close_and_out;
267
268 /* We have the archive available. To find the name we first have to
269 determine its hash value. */
270 hval = compute_hashval (name, strlen (name));
271
272 head = headmap.ptr;
273 namehashtab = (struct namehashent *) ((char *) head
274 + head->namehash_offset);
275
276 /* Avoid division by 0 if the file is corrupted. */
277 if (__glibc_unlikely (head->namehash_size <= 2))
278 goto close_and_out;
279
280 idx = hval % head->namehash_size;
281 incr = 1 + hval % (head->namehash_size - 2);
282
283 /* If the name_offset field is zero this means this is a
284 deleted entry and therefore no entry can be found. */
285 while (1)
286 {
287 if (namehashtab[idx].name_offset == 0)
288 /* Not found. */
289 goto close_and_out;
290
291 if (namehashtab[idx].hashval == hval
292 && strcmp (name, headmap.ptr + namehashtab[idx].name_offset) == 0)
293 /* Found the entry. */
294 break;
295
296 idx += incr;
297 if (idx >= head->namehash_size)
298 idx -= head->namehash_size;
299 }
300
301 /* We found an entry. It might be a placeholder for a removed one. */
302 if (namehashtab[idx].locrec_offset == 0)
303 goto close_and_out;
304
305 locrec = (struct locrecent *) (headmap.ptr + namehashtab[idx].locrec_offset);
306
307 if (sizeof (void *) > 4 /* || headmap.len == archive_stat.st_size */)
308 {
309 /* We already have the whole locale archive mapped in. */
310 assert (headmap.len == archive_stat.st_size);
311 for (cnt = 0; cnt < __LC_LAST; ++cnt)
312 if (cnt != LC_ALL)
313 {
314 if (locrec->record[cnt].offset + locrec->record[cnt].len
315 > headmap.len)
316 /* The archive locrectab contains bogus offsets. */
317 goto close_and_out;
318 results[cnt].addr = headmap.ptr + locrec->record[cnt].offset;
319 results[cnt].len = locrec->record[cnt].len;
320 }
321 }
322 else
323 {
324 /* Get the offsets of the data files and sort them. */
325 for (cnt = nranges = 0; cnt < __LC_LAST; ++cnt)
326 if (cnt != LC_ALL)
327 {
328 ranges[nranges].from = locrec->record[cnt].offset;
329 ranges[nranges].len = locrec->record[cnt].len;
330 ranges[nranges].category = cnt;
331 ranges[nranges].result = NULL;
332
333 ++nranges;
334 }
335
336 qsort (ranges, nranges, sizeof (ranges[0]), rangecmp);
337
338 /* The information about mmap'd blocks is kept in a list.
339 Skip over the blocks which are before the data we need. */
340 last = mapped = archmapped;
341 for (cnt = 0; cnt < nranges; ++cnt)
342 {
343 int upper;
344 size_t from;
345 size_t to;
346 void *addr;
347 struct archmapped *newp;
348
349 /* Determine whether the appropriate page is already mapped. */
350 while (mapped != NULL
351 && (mapped->from + mapped->len
352 <= ranges[cnt].from + ranges[cnt].len))
353 {
354 last = mapped;
355 mapped = mapped->next;
356 }
357
358 /* Do we have a match? */
359 if (mapped != NULL
360 && mapped->from <= ranges[cnt].from
361 && (ranges[cnt].from + ranges[cnt].len
362 <= mapped->from + mapped->len))
363 {
364 /* Yep, already loaded. */
365 results[ranges[cnt].category].addr = ((char *) mapped->ptr
366 + ranges[cnt].from
367 - mapped->from);
368 results[ranges[cnt].category].len = ranges[cnt].len;
369 continue;
370 }
371
372 /* Map the range with the locale data from the file. We will
373 try to cover as much of the locale as possible. I.e., if the
374 next category (next as in "next offset") is on the current or
375 immediately following page we use it as well. */
376 assert (powerof2 (ps));
377 from = ranges[cnt].from & ~(ps - 1);
378 upper = cnt;
379 do
380 {
381 to = ranges[upper].from + ranges[upper].len;
382 if (to > (size_t) archive_stat.st_size)
383 /* The archive locrectab contains bogus offsets. */
384 goto close_and_out;
385 to = (to + ps - 1) & ~(ps - 1);
386
387 /* If a range is already mmaped in, stop. */
388 if (mapped != NULL && ranges[upper].from >= mapped->from)
389 break;
390
391 ++upper;
392 }
393 /* Loop while still in contiguous pages. */
394 while (upper < nranges && ranges[upper].from < to + ps);
395
396 /* Open the file if it hasn't happened yet. */
397 if (fd == -1)
398 {
399 struct __stat64_t64 st;
400 fd = __open_nocancel (archfname,
401 O_RDONLY|O_LARGEFILE|O_CLOEXEC);
402 if (fd == -1)
403 /* Cannot open the archive, for whatever reason. */
404 return NULL;
405 /* Now verify we think this is really the same archive file
406 we opened before. If it has been changed we cannot trust
407 the header we read previously. */
408 if (__fstat64_time64 (fd, &st) < 0
409 || st.st_size != archive_stat.st_size
410 || st.st_mtime != archive_stat.st_mtime
411 || st.st_dev != archive_stat.st_dev
412 || st.st_ino != archive_stat.st_ino)
413 goto close_and_out;
414 }
415
416 /* Map the range from the archive. */
417 addr = __mmap64 (NULL, to - from, PROT_READ, MAP_FILE|MAP_COPY,
418 fd, from);
419 if (addr == MAP_FAILED)
420 goto close_and_out;
421
422 /* Allocate a record for this mapping. */
423 newp = (struct archmapped *) malloc (sizeof (struct archmapped));
424 if (newp == NULL)
425 {
426 (void) __munmap (addr, to - from);
427 goto close_and_out;
428 }
429
430 /* And queue it. */
431 newp->ptr = addr;
432 newp->from = from;
433 newp->len = to - from;
434 assert (last->next == mapped);
435 newp->next = mapped;
436 last->next = newp;
437 last = newp;
438
439 /* Determine the load addresses for the category data. */
440 do
441 {
442 assert (ranges[cnt].from >= from);
443 results[ranges[cnt].category].addr = ((char *) addr
444 + ranges[cnt].from - from);
445 results[ranges[cnt].category].len = ranges[cnt].len;
446 }
447 while (++cnt < upper);
448 --cnt; /* The 'for' will increase 'cnt' again. */
449 }
450 }
451
452 /* We don't need the file descriptor any longer. */
453 if (fd >= 0)
454 __close_nocancel_nostatus (fd);
455 fd = -1;
456
457 /* We succeeded in mapping all the necessary regions of the archive.
458 Now we need the expected data structures to point into the data. */
459
460 lia = malloc (sizeof *lia);
461 if (__glibc_unlikely (lia == NULL))
462 return NULL;
463
464 lia->name = __strdup (*namep);
465 if (__glibc_unlikely (lia->name == NULL))
466 {
467 free (lia);
468 return NULL;
469 }
470
471 lia->next = archloaded;
472 archloaded = lia;
473
474 for (cnt = 0; cnt < __LC_LAST; ++cnt)
475 if (cnt != LC_ALL)
476 {
477 lia->data[cnt] = _nl_intern_locale_data (cnt,
478 results[cnt].addr,
479 results[cnt].len);
480 if (__glibc_likely (lia->data[cnt] != NULL))
481 {
482 /* _nl_intern_locale_data leaves us these fields to initialize. */
483 lia->data[cnt]->alloc = ld_archive;
484 lia->data[cnt]->name = lia->name;
485
486 /* We do this instead of bumping the count each time we return
487 this data because the mappings stay around forever anyway
488 and we might as well hold on to a little more memory and not
489 have to rebuild it on the next lookup of the same thing.
490 If we were to maintain the usage_count normally and let the
491 structures be freed, we would have to remove the elements
492 from archloaded too. */
493 lia->data[cnt]->usage_count = UNDELETABLE;
494 }
495 }
496
497 *namep = lia->name;
498 return lia->data[category];
499 }
500
501 void
502 _nl_archive_subfreeres (void)
503 {
504 struct locale_in_archive *lia;
505 struct archmapped *am;
506
507 /* Toss out our cached locales. */
508 lia = archloaded;
509 while (lia != NULL)
510 {
511 int category;
512 struct locale_in_archive *dead = lia;
513 lia = lia->next;
514
515 free (dead->name);
516 for (category = 0; category < __LC_LAST; ++category)
517 if (category != LC_ALL && dead->data[category] != NULL)
518 _nl_unload_locale (category, dead->data[category]);
519 free (dead);
520 }
521 archloaded = NULL;
522
523 if (archmapped != NULL)
524 {
525 /* Now toss all the mapping windows, which we know nothing is using any
526 more because we just tossed all the locales that point into them. */
527
528 assert (archmapped == &headmap);
529 archmapped = NULL;
530 (void) __munmap (headmap.ptr, headmap.len);
531 am = headmap.next;
532 while (am != NULL)
533 {
534 struct archmapped *dead = am;
535 am = am->next;
536 (void) __munmap (dead->ptr, dead->len);
537 free (dead);
538 }
539 }
540 }