xref: /illumos-gate/usr/src/lib/libc/port/gen/getenv.c (revision 59f081ed)
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 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1988 AT&T	*/
28 /*	  All Rights Reserved  	*/
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 #pragma	weak _putenv = putenv
33 
34 #include "lint.h"
35 #include <mtlib.h>
36 #include <sys/types.h>
37 #include <thread.h>
38 #include <synch.h>
39 #include <stdlib.h>
40 #include <errno.h>
41 #include <string.h>
42 #include <atomic.h>
43 
44 #define	MIN_ENV_SIZE		128
45 
46 extern const char		**_environ;
47 extern void			clean_env();
48 
49 /*
50  * For performance and consistency reasons we expand the _environ list using
51  * the trusted "power of two, drop it on the floor" method. This allows for
52  * a lockless, single pass implementation of getenv(), yet the memory leak
53  * is bounded - in normal circumstances total wastage is never greater than
54  * 3x the space needed to hold any _environ list.
55  *
56  * The only abnormal circumstance is if an application modifies the _environ
57  * list pointer directly. Such an application does not conform to POSIX.1
58  * 2001. However, we also care about standards which did not foresee this
59  * issue. For this reason we keep a working copy of our notion of _environ in
60  * my_environ. If, when we are called upon to modify _environ, we ever detect
61  * a mismatch between _environ and my_environ we discard all our assumptions
62  * concerning the location and size of the _environ list. As an additional
63  * precaution we only ever update _environ once we have finished manipulating
64  * our working copy.
65  *
66  * The setenv() API is inherently leaky but we are completely at the mercy
67  * of the application.
68  *
69  * To pacify leak detectors we chain all allocations which are at risk of
70  * being leaked in either of the above two scenarios. chunk_list must only
71  * be updated under the protection of update_lock.
72  *
73  * Although we don't allocate the original _environ list it is likely that
74  * we will leak this too. Accordingly, we create a reference in initenv().
75  * However, we can't be held responsible for such leaks in abnormal (see
76  * above) circumstances.
77  */
78 
79 typedef struct chunk {
80 	struct chunk		*next;
81 } chunk_t;
82 
83 static mutex_t			update_lock = DEFAULTMUTEX;
84 static const char		**orig_environ = NULL;
85 static const char		**my_environ = NULL;
86 static const char		**environ_base = NULL;
87 static int			environ_size = 0;
88 static int			environ_gen = 0;
89 static int			initenv_done = 0;
90 static chunk_t			*chunk_list = NULL;
91 
92 /*
93  * Compute the size an _environ list including the terminating NULL entry.
94  * This is the only way we have to determine the size of an _environ list
95  * we didn't allocate.
96  */
97 static int
98 envsize(const char **e)
99 {
100 	int			size;
101 
102 	if (e == NULL)
103 		return (0);
104 
105 	for (size = 1; *e != NULL; e++)
106 		size++;
107 
108 	return (size);
109 }
110 
111 /*
112  * Initialization for the following scenarios:
113  * 1. The very first time we reference the _environ list we must call in the
114  *    NLSPATH janitor, make a reference to the original _environ list to keep
115  *    leak detectors happy, initialize my_environ and environ_base, and then
116  *    compute environ_size.
117  * 2. Whenever we detect that someone else has hijacked _environ (something
118  *    very abnormal) we need to reinitialize my_environ and environ_base,
119  *    and then recompute environ_size.
120  *
121  * The local globals my_environ, environ_base and environ_size may be used
122  * by others only if initenv_done is true and only under the protection of
123  * update_lock. However, our callers, who must NOT be holding update_lock,
124  * may safely test initenv_done or my_environ against _environ just prior to
125  * calling us because we test these again whilst holding update_lock.
126  */
127 static void
128 initenv()
129 {
130 	if ((my_environ != _environ) || !initenv_done) {
131 		lmutex_lock(&update_lock);
132 		if ((my_environ != _environ) || !initenv_done) {
133 			if (!initenv_done) {
134 				/* Call the NLSPATH janitor in. */
135 				clean_env();
136 
137 				/* Pacify leak detectors in normal operation. */
138 				orig_environ = _environ;
139 #ifdef __lint
140 				my_environ = orig_environ;
141 #endif
142 			}
143 
144 			my_environ = _environ;
145 			environ_base = my_environ;
146 			environ_size = envsize(environ_base);
147 			membar_producer();
148 			initenv_done = 1;
149 		}
150 		lmutex_unlock(&update_lock);
151 	}
152 	membar_consumer();
153 }
154 
155 /*
156  * Search an _environ list for a particular entry. If name_only is set, then
157  * string must be the entry name only, and we return the value of the first
158  * match. Otherwise, string must be of the form "name=value", and we return
159  * the address of the first matching entry.
160  */
161 static const char **
162 findenv(const char **e, const char *string, int name_only, char **value)
163 {
164 	char			target;
165 	const char		*s1;
166 	const char		*s2;
167 
168 	*value = NULL;
169 
170 	if (e == NULL)
171 		return (NULL);
172 
173 	target = name_only ? '\0' : '=';
174 
175 	for (; (s2 = *e) != NULL; e++) {
176 		s1 =  string;
177 
178 		/* Fast comparison for first char. */
179 		if (*s1 != *s2)
180 			continue;
181 
182 		/* Slow comparison for rest of string. */
183 		while (*s1 == *s2 && *s2 != '=') {
184 			s1++;
185 			s2++;
186 		}
187 
188 		if (*s1 == target && *s2 == '=') {
189 			*value = (char *)s2 + 1;
190 			return (e);
191 		}
192 	}
193 	return (NULL);
194 }
195 
196 /*
197  * Common code for putenv() and setenv(). We support the lockless getenv()
198  * by inserting new entries at the bottom of the list, and by growing the
199  * list using the trusted "power of two, drop it on the floor" method. We
200  * use a lock (update_lock) to protect all updates to the _environ list, but
201  * we are obliged to release this lock whenever we call malloc() or free().
202  * A generation number (environ_gen) is bumped whenever names are added to,
203  * or removed from, the _environ list so that we can detect collisions with
204  * other updaters.
205  *
206  * Return values
207  *   0 : success
208  *  -1 : with errno set
209  *  -2 : an entry already existed and overwrite was zero
210  */
211 static int
212 addtoenv(char *string, int overwrite)
213 {
214 	char			*value;
215 	const char		**p;
216 	chunk_t			*new_chunk;
217 	const char		**new_environ;
218 	const char		**new_base;
219 	int			new_size;
220 	int			old_gen;
221 
222 	initenv();
223 
224 	lmutex_lock(&update_lock);
225 
226 	for (;;) {
227 		/*
228 		 * If the name already exists just overwrite the existing
229 		 * entry -- except when we were called by setenv() without
230 		 * the overwrite flag.
231 		 */
232 		if ((p = findenv(my_environ, string, 0, &value)) != NULL) {
233 			if (overwrite) {
234 				/*
235 				 * Replace the value in situ. No name was
236 				 * added, so there is no need to bump the
237 				 * generation number.
238 				 */
239 				*p = string;
240 				lmutex_unlock(&update_lock);
241 				return (0);
242 			} else {
243 				/* No change. */
244 				lmutex_unlock(&update_lock);
245 				return (-2);
246 			}
247 		}
248 
249 		/* Try to insert the new entry at the bottom of the list. */
250 		if (environ_base < my_environ) {
251 			/*
252 			 * The new value must be visible before we decrement
253 			 * the _environ list pointer.
254 			 */
255 			my_environ[-1] = string;
256 			membar_producer();
257 			my_environ--;
258 			_environ = my_environ;
259 
260 			/*
261 			 * We've added a name, so bump the generation number.
262 			 */
263 			environ_gen++;
264 
265 			lmutex_unlock(&update_lock);
266 			return (0);
267 		}
268 
269 		/*
270 		 * There is no room. Attempt to allocate a new _environ list
271 		 * which is at least double the size of the current one. See
272 		 * comment above concerning locking and malloc() etc.
273 		 */
274 		new_size = environ_size * 2;
275 		if (new_size < MIN_ENV_SIZE)
276 			new_size = MIN_ENV_SIZE;
277 
278 		old_gen = environ_gen;
279 		lmutex_unlock(&update_lock);
280 
281 		new_chunk = malloc(sizeof (chunk_t) +
282 		    new_size * sizeof (char *));
283 		if (new_chunk == NULL) {
284 			errno = ENOMEM;
285 			return (-1);
286 		}
287 
288 		lmutex_lock(&update_lock);
289 
290 		/*
291 		 * If no other thread added or removed names while the lock
292 		 * was dropped, it is time to break out of this loop.
293 		 */
294 		if (environ_gen == old_gen)
295 			break;
296 
297 		/*
298 		 * At least one name has been added or removed, so we need to
299 		 * try again. It is very likely that we will find sufficient
300 		 * space the next time around.
301 		 */
302 		lmutex_unlock(&update_lock);
303 		free(new_chunk);
304 		lmutex_lock(&update_lock);
305 	}
306 
307 	/* Add the new chunk to chunk_list to hide potential future leak. */
308 	new_chunk->next = chunk_list;
309 	chunk_list = new_chunk;
310 
311 	/* Copy the old _environ list into the top of the new _environ list. */
312 	new_base = (const char **)(new_chunk + 1);
313 	new_environ = &new_base[(new_size - 1) - environ_size];
314 	(void) memcpy(new_environ, my_environ, environ_size * sizeof (char *));
315 
316 	/* Insert the new entry at the bottom of the new _environ list. */
317 	new_environ[-1] = string;
318 	new_environ--;
319 
320 	/* Ensure that the new _environ list is visible to all. */
321 	membar_producer();
322 
323 	/* Make the switch (dropping the old _environ list on the floor). */
324 	environ_base = new_base;
325 	my_environ = new_environ;
326 	_environ = my_environ;
327 	environ_size = new_size;
328 
329 	/* We've added a name, so bump the generation number. */
330 	environ_gen++;
331 
332 	lmutex_unlock(&update_lock);
333 	return (0);
334 }
335 
336 /*
337  * All the work for putenv() is done in addtoenv().
338  */
339 int
340 putenv(char *string)
341 {
342 	return (addtoenv(string, 1));
343 }
344 
345 /*
346  * setenv() is a little more complex than putenv() because we have to allocate
347  * and construct an _environ entry on behalf of the caller. The bulk of the
348  * work is still done in addtoenv().
349  */
350 
351 int
352 setenv(const char *envname, const char *envval, int overwrite)
353 {
354 	chunk_t			*new_chunk;
355 	char			*new_string;
356 	size_t			name_len;
357 	size_t			val_len;
358 	int			res;
359 
360 	if (envname == NULL || *envname == 0 || strchr(envname, '=') != NULL) {
361 		errno = EINVAL;
362 		return (-1);
363 	}
364 
365 	name_len = strlen(envname);
366 	val_len = strlen(envval);
367 
368 	new_chunk = malloc(sizeof (chunk_t) + name_len + val_len + 2);
369 	if (new_chunk == NULL) {
370 		errno = ENOMEM;
371 		return (-1);
372 	}
373 	new_string = (char *)(new_chunk + 1);
374 
375 	(void) memcpy(new_string, envname, name_len);
376 	new_string[name_len] = '=';
377 	(void) memcpy(new_string + name_len + 1, envval, val_len);
378 	new_string[name_len + 1 + val_len] = 0;
379 
380 	if ((res = addtoenv(new_string, overwrite)) < 0) {
381 		free(new_chunk);
382 		if (res == -2) {
383 			/* The name already existed, but not an error. */
384 			return (0);
385 		} else {
386 			/* i.e. res == -1 which means only one thing. */
387 			errno = ENOMEM;
388 			return (-1);
389 		}
390 	}
391 
392 	/* Hide potential leak of new_string. */
393 	lmutex_lock(&update_lock);
394 	new_chunk->next = chunk_list;
395 	chunk_list = new_chunk;
396 	lmutex_unlock(&update_lock);
397 
398 	return (0);
399 }
400 
401 /*
402  * unsetenv() is tricky because we need to compress the _environ list in a way
403  * which supports a lockless getenv(). The approach here is to move the first
404  * entry from the enrivon list into the space occupied by the entry to be
405  * deleted, and then to increment _environ. This has the added advantage of
406  * making _any_ incremental linear search of the _environ list consistent (i.e.
407  * we will not break any naughty apps which read the list without our help).
408  */
409 int
410 unsetenv(const char *name)
411 {
412 	const char		**p;
413 	char			*value;
414 
415 	if (name == NULL || *name == 0 || strchr(name, '=') != NULL) {
416 		errno = EINVAL;
417 		return (-1);
418 	}
419 
420 	initenv();
421 
422 	lmutex_lock(&update_lock);
423 
424 	/*
425 	 * Find the target, overwrite it with the first entry, increment the
426 	 * _environ pointer.
427 	 */
428 	if ((p = findenv(my_environ, name, 1, &value)) != NULL) {
429 		/* Overwrite target with the first entry. */
430 		*p = my_environ[0];
431 
432 		/* Ensure that the moved entry is visible to all.  */
433 		membar_producer();
434 
435 		/* Shrink the _environ list. */
436 		my_environ++;
437 		_environ = my_environ;
438 
439 		/* Make sure addtoenv() knows that we've removed a name. */
440 		environ_gen++;
441 	}
442 
443 	lmutex_unlock(&update_lock);
444 	return (0);
445 }
446 
447 /*
448  * At last, a lockless implementation of getenv()!
449  */
450 char *
451 getenv(const char *name)
452 {
453 	char			*value;
454 
455 	initenv();
456 
457 	if (findenv(_environ, name, 1, &value) != NULL)
458 		return (value);
459 
460 	return (NULL);
461 }
462