xref: /linux/net/unix/garbage.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 /*
2  * NET3:	Garbage Collector For AF_UNIX sockets
3  *
4  * Garbage Collector:
5  *	Copyright (C) Barak A. Pearlmutter.
6  *	Released under the GPL version 2 or later.
7  *
8  * Chopped about by Alan Cox 22/3/96 to make it fit the AF_UNIX socket problem.
9  * If it doesn't work blame me, it worked when Barak sent it.
10  *
11  * Assumptions:
12  *
13  *  - object w/ a bit
14  *  - free list
15  *
16  * Current optimizations:
17  *
18  *  - explicit stack instead of recursion
19  *  - tail recurse on first born instead of immediate push/pop
20  *  - we gather the stuff that should not be killed into tree
21  *    and stack is just a path from root to the current pointer.
22  *
23  *  Future optimizations:
24  *
25  *  - don't just push entire root set; process in place
26  *
27  *	This program is free software; you can redistribute it and/or
28  *	modify it under the terms of the GNU General Public License
29  *	as published by the Free Software Foundation; either version
30  *	2 of the License, or (at your option) any later version.
31  *
32  *  Fixes:
33  *	Alan Cox	07 Sept	1997	Vmalloc internal stack as needed.
34  *					Cope with changing max_files.
35  *	Al Viro		11 Oct 1998
36  *		Graph may have cycles. That is, we can send the descriptor
37  *		of foo to bar and vice versa. Current code chokes on that.
38  *		Fix: move SCM_RIGHTS ones into the separate list and then
39  *		skb_free() them all instead of doing explicit fput's.
40  *		Another problem: since fput() may block somebody may
41  *		create a new unix_socket when we are in the middle of sweep
42  *		phase. Fix: revert the logic wrt MARKED. Mark everything
43  *		upon the beginning and unmark non-junk ones.
44  *
45  *		[12 Oct 1998] AAARGH! New code purges all SCM_RIGHTS
46  *		sent to connect()'ed but still not accept()'ed sockets.
47  *		Fixed. Old code had slightly different problem here:
48  *		extra fput() in situation when we passed the descriptor via
49  *		such socket and closed it (descriptor). That would happen on
50  *		each unix_gc() until the accept(). Since the struct file in
51  *		question would go to the free list and might be reused...
52  *		That might be the reason of random oopses on filp_close()
53  *		in unrelated processes.
54  *
55  *	AV		28 Feb 1999
56  *		Kill the explicit allocation of stack. Now we keep the tree
57  *		with root in dummy + pointer (gc_current) to one of the nodes.
58  *		Stack is represented as path from gc_current to dummy. Unmark
59  *		now means "add to tree". Push == "make it a son of gc_current".
60  *		Pop == "move gc_current to parent". We keep only pointers to
61  *		parents (->gc_tree).
62  *	AV		1 Mar 1999
63  *		Damn. Added missing check for ->dead in listen queues scanning.
64  *
65  *	Miklos Szeredi 25 Jun 2007
66  *		Reimplement with a cycle collecting algorithm. This should
67  *		solve several problems with the previous code, like being racy
68  *		wrt receive and holding up unrelated socket operations.
69  */
70 
71 #include <linux/kernel.h>
72 #include <linux/string.h>
73 #include <linux/socket.h>
74 #include <linux/un.h>
75 #include <linux/net.h>
76 #include <linux/fs.h>
77 #include <linux/skbuff.h>
78 #include <linux/netdevice.h>
79 #include <linux/file.h>
80 #include <linux/proc_fs.h>
81 #include <linux/mutex.h>
82 #include <linux/wait.h>
83 
84 #include <net/sock.h>
85 #include <net/af_unix.h>
86 #include <net/scm.h>
87 #include <net/tcp_states.h>
88 
89 /* Internal data structures and random procedures: */
90 
91 static LIST_HEAD(gc_inflight_list);
92 static LIST_HEAD(gc_candidates);
93 static DEFINE_SPINLOCK(unix_gc_lock);
94 static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
95 
96 unsigned int unix_tot_inflight;
97 
98 struct sock *unix_get_socket(struct file *filp)
99 {
100 	struct sock *u_sock = NULL;
101 	struct inode *inode = file_inode(filp);
102 
103 	/* Socket ? */
104 	if (S_ISSOCK(inode->i_mode) && !(filp->f_mode & FMODE_PATH)) {
105 		struct socket *sock = SOCKET_I(inode);
106 		struct sock *s = sock->sk;
107 
108 		/* PF_UNIX ? */
109 		if (s && sock->ops && sock->ops->family == PF_UNIX)
110 			u_sock = s;
111 	}
112 	return u_sock;
113 }
114 
115 /* Keep the number of times in flight count for the file
116  * descriptor if it is for an AF_UNIX socket.
117  */
118 
119 void unix_inflight(struct user_struct *user, struct file *fp)
120 {
121 	struct sock *s = unix_get_socket(fp);
122 
123 	spin_lock(&unix_gc_lock);
124 
125 	if (s) {
126 		struct unix_sock *u = unix_sk(s);
127 
128 		if (atomic_long_inc_return(&u->inflight) == 1) {
129 			BUG_ON(!list_empty(&u->link));
130 			list_add_tail(&u->link, &gc_inflight_list);
131 		} else {
132 			BUG_ON(list_empty(&u->link));
133 		}
134 		unix_tot_inflight++;
135 	}
136 	user->unix_inflight++;
137 	spin_unlock(&unix_gc_lock);
138 }
139 
140 void unix_notinflight(struct user_struct *user, struct file *fp)
141 {
142 	struct sock *s = unix_get_socket(fp);
143 
144 	spin_lock(&unix_gc_lock);
145 
146 	if (s) {
147 		struct unix_sock *u = unix_sk(s);
148 
149 		BUG_ON(!atomic_long_read(&u->inflight));
150 		BUG_ON(list_empty(&u->link));
151 
152 		if (atomic_long_dec_and_test(&u->inflight))
153 			list_del_init(&u->link);
154 		unix_tot_inflight--;
155 	}
156 	user->unix_inflight--;
157 	spin_unlock(&unix_gc_lock);
158 }
159 
160 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
161 			  struct sk_buff_head *hitlist)
162 {
163 	struct sk_buff *skb;
164 	struct sk_buff *next;
165 
166 	spin_lock(&x->sk_receive_queue.lock);
167 	skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
168 		/* Do we have file descriptors ? */
169 		if (UNIXCB(skb).fp) {
170 			bool hit = false;
171 			/* Process the descriptors of this socket */
172 			int nfd = UNIXCB(skb).fp->count;
173 			struct file **fp = UNIXCB(skb).fp->fp;
174 
175 			while (nfd--) {
176 				/* Get the socket the fd matches if it indeed does so */
177 				struct sock *sk = unix_get_socket(*fp++);
178 
179 				if (sk) {
180 					struct unix_sock *u = unix_sk(sk);
181 
182 					/* Ignore non-candidates, they could
183 					 * have been added to the queues after
184 					 * starting the garbage collection
185 					 */
186 					if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
187 						hit = true;
188 
189 						func(u);
190 					}
191 				}
192 			}
193 			if (hit && hitlist != NULL) {
194 				__skb_unlink(skb, &x->sk_receive_queue);
195 				__skb_queue_tail(hitlist, skb);
196 			}
197 		}
198 	}
199 	spin_unlock(&x->sk_receive_queue.lock);
200 }
201 
202 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
203 			  struct sk_buff_head *hitlist)
204 {
205 	if (x->sk_state != TCP_LISTEN) {
206 		scan_inflight(x, func, hitlist);
207 	} else {
208 		struct sk_buff *skb;
209 		struct sk_buff *next;
210 		struct unix_sock *u;
211 		LIST_HEAD(embryos);
212 
213 		/* For a listening socket collect the queued embryos
214 		 * and perform a scan on them as well.
215 		 */
216 		spin_lock(&x->sk_receive_queue.lock);
217 		skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
218 			u = unix_sk(skb->sk);
219 
220 			/* An embryo cannot be in-flight, so it's safe
221 			 * to use the list link.
222 			 */
223 			BUG_ON(!list_empty(&u->link));
224 			list_add_tail(&u->link, &embryos);
225 		}
226 		spin_unlock(&x->sk_receive_queue.lock);
227 
228 		while (!list_empty(&embryos)) {
229 			u = list_entry(embryos.next, struct unix_sock, link);
230 			scan_inflight(&u->sk, func, hitlist);
231 			list_del_init(&u->link);
232 		}
233 	}
234 }
235 
236 static void dec_inflight(struct unix_sock *usk)
237 {
238 	atomic_long_dec(&usk->inflight);
239 }
240 
241 static void inc_inflight(struct unix_sock *usk)
242 {
243 	atomic_long_inc(&usk->inflight);
244 }
245 
246 static void inc_inflight_move_tail(struct unix_sock *u)
247 {
248 	atomic_long_inc(&u->inflight);
249 	/* If this still might be part of a cycle, move it to the end
250 	 * of the list, so that it's checked even if it was already
251 	 * passed over
252 	 */
253 	if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
254 		list_move_tail(&u->link, &gc_candidates);
255 }
256 
257 static bool gc_in_progress;
258 #define UNIX_INFLIGHT_TRIGGER_GC 16000
259 
260 void wait_for_unix_gc(void)
261 {
262 	/* If number of inflight sockets is insane,
263 	 * force a garbage collect right now.
264 	 */
265 	if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
266 		unix_gc();
267 	wait_event(unix_gc_wait, gc_in_progress == false);
268 }
269 
270 /* The external entry point: unix_gc() */
271 void unix_gc(void)
272 {
273 	struct unix_sock *u;
274 	struct unix_sock *next;
275 	struct sk_buff_head hitlist;
276 	struct list_head cursor;
277 	LIST_HEAD(not_cycle_list);
278 
279 	spin_lock(&unix_gc_lock);
280 
281 	/* Avoid a recursive GC. */
282 	if (gc_in_progress)
283 		goto out;
284 
285 	gc_in_progress = true;
286 	/* First, select candidates for garbage collection.  Only
287 	 * in-flight sockets are considered, and from those only ones
288 	 * which don't have any external reference.
289 	 *
290 	 * Holding unix_gc_lock will protect these candidates from
291 	 * being detached, and hence from gaining an external
292 	 * reference.  Since there are no possible receivers, all
293 	 * buffers currently on the candidates' queues stay there
294 	 * during the garbage collection.
295 	 *
296 	 * We also know that no new candidate can be added onto the
297 	 * receive queues.  Other, non candidate sockets _can_ be
298 	 * added to queue, so we must make sure only to touch
299 	 * candidates.
300 	 */
301 	list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
302 		long total_refs;
303 		long inflight_refs;
304 
305 		total_refs = file_count(u->sk.sk_socket->file);
306 		inflight_refs = atomic_long_read(&u->inflight);
307 
308 		BUG_ON(inflight_refs < 1);
309 		BUG_ON(total_refs < inflight_refs);
310 		if (total_refs == inflight_refs) {
311 			list_move_tail(&u->link, &gc_candidates);
312 			__set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
313 			__set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
314 		}
315 	}
316 
317 	/* Now remove all internal in-flight reference to children of
318 	 * the candidates.
319 	 */
320 	list_for_each_entry(u, &gc_candidates, link)
321 		scan_children(&u->sk, dec_inflight, NULL);
322 
323 	/* Restore the references for children of all candidates,
324 	 * which have remaining references.  Do this recursively, so
325 	 * only those remain, which form cyclic references.
326 	 *
327 	 * Use a "cursor" link, to make the list traversal safe, even
328 	 * though elements might be moved about.
329 	 */
330 	list_add(&cursor, &gc_candidates);
331 	while (cursor.next != &gc_candidates) {
332 		u = list_entry(cursor.next, struct unix_sock, link);
333 
334 		/* Move cursor to after the current position. */
335 		list_move(&cursor, &u->link);
336 
337 		if (atomic_long_read(&u->inflight) > 0) {
338 			list_move_tail(&u->link, &not_cycle_list);
339 			__clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
340 			scan_children(&u->sk, inc_inflight_move_tail, NULL);
341 		}
342 	}
343 	list_del(&cursor);
344 
345 	/* Now gc_candidates contains only garbage.  Restore original
346 	 * inflight counters for these as well, and remove the skbuffs
347 	 * which are creating the cycle(s).
348 	 */
349 	skb_queue_head_init(&hitlist);
350 	list_for_each_entry(u, &gc_candidates, link)
351 		scan_children(&u->sk, inc_inflight, &hitlist);
352 
353 	/* not_cycle_list contains those sockets which do not make up a
354 	 * cycle.  Restore these to the inflight list.
355 	 */
356 	while (!list_empty(&not_cycle_list)) {
357 		u = list_entry(not_cycle_list.next, struct unix_sock, link);
358 		__clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
359 		list_move_tail(&u->link, &gc_inflight_list);
360 	}
361 
362 	spin_unlock(&unix_gc_lock);
363 
364 	/* Here we are. Hitlist is filled. Die. */
365 	__skb_queue_purge(&hitlist);
366 
367 	spin_lock(&unix_gc_lock);
368 
369 	/* All candidates should have been detached by now. */
370 	BUG_ON(!list_empty(&gc_candidates));
371 	gc_in_progress = false;
372 	wake_up(&unix_gc_wait);
373 
374  out:
375 	spin_unlock(&unix_gc_lock);
376 }
377