xref: /illumos-gate/usr/src/lib/libc/port/gen/getenv.c (revision b6805bf78d2bbbeeaea8909a05623587b42d58b3)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 1989, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 
26 /*	Copyright (c) 1988 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 #pragma	weak _putenv = putenv
30 
31 #include "lint.h"
32 #include <mtlib.h>
33 #include <sys/types.h>
34 #include <thread.h>
35 #include <synch.h>
36 #include <stdlib.h>
37 #include <errno.h>
38 #include <string.h>
39 #include <atomic.h>
40 
41 #define	MIN_ENV_SIZE		128
42 
43 extern const char		**_environ;
44 extern void			clean_env();
45 
46 /*
47  * For performance and consistency reasons we expand the _environ list using
48  * the trusted "power of two, drop it on the floor" method. This allows for
49  * a lockless, single pass implementation of getenv(), yet the memory leak
50  * is bounded - in normal circumstances total wastage is never greater than
51  * 3x the space needed to hold any _environ list.
52  *
53  * The only abnormal circumstance is if an application modifies the _environ
54  * list pointer directly. Such an application does not conform to POSIX.1
55  * 2001. However, we also care about standards which did not foresee this
56  * issue. For this reason we keep a working copy of our notion of _environ in
57  * my_environ. If, when we are called upon to modify _environ, we ever detect
58  * a mismatch between _environ and my_environ we discard all our assumptions
59  * concerning the location and size of the _environ list. As an additional
60  * precaution we only ever update _environ once we have finished manipulating
61  * our working copy.
62  *
63  * The setenv() API is inherently leaky but we are completely at the mercy
64  * of the application.
65  *
66  * To pacify leak detectors we chain all allocations which are at risk of
67  * being leaked in either of the above two scenarios. chunk_list must only
68  * be updated under the protection of update_lock.
69  *
70  * Although we don't allocate the original _environ list it is likely that
71  * we will leak this too. Accordingly, we create a reference in initenv().
72  * However, we can't be held responsible for such leaks in abnormal (see
73  * above) circumstances.
74  */
75 
76 typedef struct chunk {
77 	struct chunk		*next;
78 } chunk_t;
79 
80 static mutex_t			update_lock = DEFAULTMUTEX;
81 static const char		**orig_environ = NULL;
82 static const char		**my_environ = NULL;
83 static const char		**environ_base = NULL;
84 static int			environ_size = 0;
85 static int			environ_gen = 0;
86 static int			initenv_done = 0;
87 static chunk_t			*chunk_list = NULL;
88 
89 /*
90  * Compute the size an _environ list including the terminating NULL entry.
91  * This is the only way we have to determine the size of an _environ list
92  * we didn't allocate.
93  */
94 static int
95 envsize(const char **e)
96 {
97 	int			size;
98 
99 	if (e == NULL)
100 		return (0);
101 
102 	for (size = 1; *e != NULL; e++)
103 		size++;
104 
105 	return (size);
106 }
107 
108 /*
109  * Initialization for the following scenarios:
110  * 1. The very first time we reference the _environ list we must call in the
111  *    NLSPATH janitor, make a reference to the original _environ list to keep
112  *    leak detectors happy, initialize my_environ and environ_base, and then
113  *    compute environ_size.
114  * 2. Whenever we detect that someone else has hijacked _environ (something
115  *    very abnormal) we need to reinitialize my_environ and environ_base,
116  *    and then recompute environ_size.
117  *
118  * The local globals my_environ, environ_base and environ_size may be used
119  * by others only if initenv_done is true and only under the protection of
120  * update_lock. However, our callers, who must NOT be holding update_lock,
121  * may safely test initenv_done or my_environ against _environ just prior to
122  * calling us because we test these again whilst holding update_lock.
123  */
124 static void
125 initenv()
126 {
127 	if ((my_environ != _environ) || !initenv_done) {
128 		lmutex_lock(&update_lock);
129 		if ((my_environ != _environ) || !initenv_done) {
130 			if (!initenv_done) {
131 				/* Call the NLSPATH janitor in. */
132 				clean_env();
133 
134 				/* Pacify leak detectors in normal operation. */
135 				orig_environ = _environ;
136 #ifdef __lint
137 				my_environ = orig_environ;
138 #endif
139 			}
140 
141 			my_environ = _environ;
142 			environ_base = my_environ;
143 			environ_size = envsize(environ_base);
144 			membar_producer();
145 			initenv_done = 1;
146 		}
147 		lmutex_unlock(&update_lock);
148 	}
149 	membar_consumer();
150 }
151 
152 /*
153  * Search an _environ list for a particular entry. If name_only is set, then
154  * string must be the entry name only, and we return the value of the first
155  * match. Otherwise, string must be of the form "name=value", and we return
156  * the address of the first matching entry.
157  */
158 static const char **
159 findenv(const char **e, const char *string, int name_only, char **value)
160 {
161 	char			target;
162 	const char		*s1;
163 	const char		*s2;
164 
165 	*value = NULL;
166 
167 	if (e == NULL)
168 		return (NULL);
169 
170 	target = name_only ? '\0' : '=';
171 
172 	for (; (s2 = *e) != NULL; e++) {
173 		s1 =  string;
174 
175 		/* Fast comparison for first char. */
176 		if (*s1 != *s2)
177 			continue;
178 
179 		/* Slow comparison for rest of string. */
180 		while (*s1 == *s2 && *s2 != '=') {
181 			s1++;
182 			s2++;
183 		}
184 
185 		if (*s1 == target && *s2 == '=') {
186 			*value = (char *)s2 + 1;
187 			return (e);
188 		}
189 	}
190 	return (NULL);
191 }
192 
193 /*
194  * Common code for putenv() and setenv(). We support the lockless getenv()
195  * by inserting new entries at the bottom of the list, and by growing the
196  * list using the trusted "power of two, drop it on the floor" method. We
197  * use a lock (update_lock) to protect all updates to the _environ list, but
198  * we are obliged to release this lock whenever we call malloc() or free().
199  * A generation number (environ_gen) is bumped whenever names are added to,
200  * or removed from, the _environ list so that we can detect collisions with
201  * other updaters.
202  *
203  * Return values
204  *   0 : success
205  *  -1 : with errno set
206  *  -2 : an entry already existed and overwrite was zero
207  */
208 static int
209 addtoenv(char *string, int overwrite)
210 {
211 	char			*value;
212 	const char		**p;
213 	chunk_t			*new_chunk;
214 	const char		**new_environ;
215 	const char		**new_base;
216 	int			new_size;
217 	int			old_gen;
218 
219 	initenv();
220 
221 	lmutex_lock(&update_lock);
222 
223 	for (;;) {
224 		/*
225 		 * If the name already exists just overwrite the existing
226 		 * entry -- except when we were called by setenv() without
227 		 * the overwrite flag.
228 		 */
229 		if ((p = findenv(my_environ, string, 0, &value)) != NULL) {
230 			if (overwrite) {
231 				/*
232 				 * Replace the value in situ. No name was
233 				 * added, so there is no need to bump the
234 				 * generation number.
235 				 */
236 				*p = string;
237 				lmutex_unlock(&update_lock);
238 				return (0);
239 			} else {
240 				/* No change. */
241 				lmutex_unlock(&update_lock);
242 				return (-2);
243 			}
244 		}
245 
246 		/* Try to insert the new entry at the bottom of the list. */
247 		if (environ_base < my_environ) {
248 			/*
249 			 * The new value must be visible before we decrement
250 			 * the _environ list pointer.
251 			 */
252 			my_environ[-1] = string;
253 			membar_producer();
254 			my_environ--;
255 			_environ = my_environ;
256 
257 			/*
258 			 * We've added a name, so bump the generation number.
259 			 */
260 			environ_gen++;
261 
262 			lmutex_unlock(&update_lock);
263 			return (0);
264 		}
265 
266 		/*
267 		 * There is no room. Attempt to allocate a new _environ list
268 		 * which is at least double the size of the current one. See
269 		 * comment above concerning locking and malloc() etc.
270 		 */
271 		new_size = environ_size * 2;
272 		if (new_size < MIN_ENV_SIZE)
273 			new_size = MIN_ENV_SIZE;
274 
275 		old_gen = environ_gen;
276 		lmutex_unlock(&update_lock);
277 
278 		new_chunk = malloc(sizeof (chunk_t) +
279 		    new_size * sizeof (char *));
280 		if (new_chunk == NULL) {
281 			errno = ENOMEM;
282 			return (-1);
283 		}
284 
285 		lmutex_lock(&update_lock);
286 
287 		/*
288 		 * If no other thread added or removed names while the lock
289 		 * was dropped, it is time to break out of this loop.
290 		 */
291 		if (environ_gen == old_gen)
292 			break;
293 
294 		/*
295 		 * At least one name has been added or removed, so we need to
296 		 * try again. It is very likely that we will find sufficient
297 		 * space the next time around.
298 		 */
299 		lmutex_unlock(&update_lock);
300 		free(new_chunk);
301 		lmutex_lock(&update_lock);
302 	}
303 
304 	/* Add the new chunk to chunk_list to hide potential future leak. */
305 	new_chunk->next = chunk_list;
306 	chunk_list = new_chunk;
307 
308 	/* Copy the old _environ list into the top of the new _environ list. */
309 	new_base = (const char **)(new_chunk + 1);
310 	new_environ = &new_base[(new_size - 1) - environ_size];
311 	(void) memcpy(new_environ, my_environ, environ_size * sizeof (char *));
312 
313 	/* Insert the new entry at the bottom of the new _environ list. */
314 	new_environ[-1] = string;
315 	new_environ--;
316 
317 	/* Ensure that the new _environ list is visible to all. */
318 	membar_producer();
319 
320 	/* Make the switch (dropping the old _environ list on the floor). */
321 	environ_base = new_base;
322 	my_environ = new_environ;
323 	_environ = my_environ;
324 	environ_size = new_size;
325 
326 	/* We've added a name, so bump the generation number. */
327 	environ_gen++;
328 
329 	lmutex_unlock(&update_lock);
330 	return (0);
331 }
332 
333 /*
334  * All the work for putenv() is done in addtoenv().
335  */
336 int
337 putenv(char *string)
338 {
339 	return (addtoenv(string, 1));
340 }
341 
342 /*
343  * setenv() is a little more complex than putenv() because we have to allocate
344  * and construct an _environ entry on behalf of the caller. The bulk of the
345  * work is still done in addtoenv().
346  */
347 
348 int
349 setenv(const char *envname, const char *envval, int overwrite)
350 {
351 	chunk_t			*new_chunk;
352 	char			*new_string;
353 	size_t			name_len;
354 	size_t			val_len;
355 	int			res;
356 
357 	if (envname == NULL || *envname == 0 || strchr(envname, '=') != NULL) {
358 		errno = EINVAL;
359 		return (-1);
360 	}
361 
362 	name_len = strlen(envname);
363 	val_len = strlen(envval);
364 
365 	new_chunk = malloc(sizeof (chunk_t) + name_len + val_len + 2);
366 	if (new_chunk == NULL) {
367 		errno = ENOMEM;
368 		return (-1);
369 	}
370 	new_string = (char *)(new_chunk + 1);
371 
372 	(void) memcpy(new_string, envname, name_len);
373 	new_string[name_len] = '=';
374 	(void) memcpy(new_string + name_len + 1, envval, val_len);
375 	new_string[name_len + 1 + val_len] = 0;
376 
377 	if ((res = addtoenv(new_string, overwrite)) < 0) {
378 		free(new_chunk);
379 		if (res == -2) {
380 			/* The name already existed, but not an error. */
381 			return (0);
382 		} else {
383 			/* i.e. res == -1 which means only one thing. */
384 			errno = ENOMEM;
385 			return (-1);
386 		}
387 	}
388 
389 	/* Hide potential leak of new_string. */
390 	lmutex_lock(&update_lock);
391 	new_chunk->next = chunk_list;
392 	chunk_list = new_chunk;
393 	lmutex_unlock(&update_lock);
394 
395 	return (0);
396 }
397 
398 /*
399  * unsetenv() is tricky because we need to compress the _environ list in a way
400  * which supports a lockless getenv(). The approach here is to move the first
401  * entry from the enrivon list into the space occupied by the entry to be
402  * deleted, and then to increment _environ. This has the added advantage of
403  * making _any_ incremental linear search of the _environ list consistent (i.e.
404  * we will not break any naughty apps which read the list without our help).
405  */
406 int
407 unsetenv(const char *name)
408 {
409 	const char		**p;
410 	char			*value;
411 
412 	if (name == NULL || *name == 0 || strchr(name, '=') != NULL) {
413 		errno = EINVAL;
414 		return (-1);
415 	}
416 
417 	initenv();
418 
419 	lmutex_lock(&update_lock);
420 
421 	/*
422 	 * Find the target, overwrite it with the first entry, increment the
423 	 * _environ pointer.
424 	 */
425 	if ((p = findenv(my_environ, name, 1, &value)) != NULL) {
426 		/* Overwrite target with the first entry. */
427 		*p = my_environ[0];
428 
429 		/* Ensure that the moved entry is visible to all.  */
430 		membar_producer();
431 
432 		/* Shrink the _environ list. */
433 		my_environ++;
434 		_environ = my_environ;
435 
436 		/* Make sure addtoenv() knows that we've removed a name. */
437 		environ_gen++;
438 	}
439 
440 	lmutex_unlock(&update_lock);
441 	return (0);
442 }
443 
444 /*
445  * Dump entire environment.
446  */
447 int
448 clearenv(void)
449 {
450 	/*
451 	 * Just drop the entire environment list on the floor, as it
452 	 * would be non-trivial to try and free the used memory.
453 	 */
454 	static const char *nullp = NULL;
455 
456 	lmutex_lock(&update_lock);
457 	_environ = &nullp;
458 	my_environ = NULL;
459 	environ_base = NULL;
460 	environ_size = 0;
461 	environ_gen++;
462 	membar_producer();
463 	lmutex_unlock(&update_lock);
464 
465 	return (0);
466 }
467 
468 /*
469  * At last, a lockless implementation of getenv()!
470  */
471 char *
472 getenv(const char *name)
473 {
474 	char			*value;
475 
476 	initenv();
477 
478 	if (findenv(_environ, name, 1, &value) != NULL)
479 		return (value);
480 
481 	return (NULL);
482 }
483