xref: /openbmc/linux/net/unix/garbage.c (revision 827634ad)
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 file *fp)
120 {
121 	struct sock *s = unix_get_socket(fp);
122 
123 	if (s) {
124 		struct unix_sock *u = unix_sk(s);
125 
126 		spin_lock(&unix_gc_lock);
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 		spin_unlock(&unix_gc_lock);
136 	}
137 }
138 
139 void unix_notinflight(struct file *fp)
140 {
141 	struct sock *s = unix_get_socket(fp);
142 
143 	if (s) {
144 		struct unix_sock *u = unix_sk(s);
145 
146 		spin_lock(&unix_gc_lock);
147 		BUG_ON(list_empty(&u->link));
148 
149 		if (atomic_long_dec_and_test(&u->inflight))
150 			list_del_init(&u->link);
151 		unix_tot_inflight--;
152 		spin_unlock(&unix_gc_lock);
153 	}
154 }
155 
156 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
157 			  struct sk_buff_head *hitlist)
158 {
159 	struct sk_buff *skb;
160 	struct sk_buff *next;
161 
162 	spin_lock(&x->sk_receive_queue.lock);
163 	skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
164 		/* Do we have file descriptors ? */
165 		if (UNIXCB(skb).fp) {
166 			bool hit = false;
167 			/* Process the descriptors of this socket */
168 			int nfd = UNIXCB(skb).fp->count;
169 			struct file **fp = UNIXCB(skb).fp->fp;
170 
171 			while (nfd--) {
172 				/* Get the socket the fd matches if it indeed does so */
173 				struct sock *sk = unix_get_socket(*fp++);
174 
175 				if (sk) {
176 					struct unix_sock *u = unix_sk(sk);
177 
178 					/* Ignore non-candidates, they could
179 					 * have been added to the queues after
180 					 * starting the garbage collection
181 					 */
182 					if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
183 						hit = true;
184 
185 						func(u);
186 					}
187 				}
188 			}
189 			if (hit && hitlist != NULL) {
190 				__skb_unlink(skb, &x->sk_receive_queue);
191 				__skb_queue_tail(hitlist, skb);
192 			}
193 		}
194 	}
195 	spin_unlock(&x->sk_receive_queue.lock);
196 }
197 
198 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
199 			  struct sk_buff_head *hitlist)
200 {
201 	if (x->sk_state != TCP_LISTEN) {
202 		scan_inflight(x, func, hitlist);
203 	} else {
204 		struct sk_buff *skb;
205 		struct sk_buff *next;
206 		struct unix_sock *u;
207 		LIST_HEAD(embryos);
208 
209 		/* For a listening socket collect the queued embryos
210 		 * and perform a scan on them as well.
211 		 */
212 		spin_lock(&x->sk_receive_queue.lock);
213 		skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
214 			u = unix_sk(skb->sk);
215 
216 			/* An embryo cannot be in-flight, so it's safe
217 			 * to use the list link.
218 			 */
219 			BUG_ON(!list_empty(&u->link));
220 			list_add_tail(&u->link, &embryos);
221 		}
222 		spin_unlock(&x->sk_receive_queue.lock);
223 
224 		while (!list_empty(&embryos)) {
225 			u = list_entry(embryos.next, struct unix_sock, link);
226 			scan_inflight(&u->sk, func, hitlist);
227 			list_del_init(&u->link);
228 		}
229 	}
230 }
231 
232 static void dec_inflight(struct unix_sock *usk)
233 {
234 	atomic_long_dec(&usk->inflight);
235 }
236 
237 static void inc_inflight(struct unix_sock *usk)
238 {
239 	atomic_long_inc(&usk->inflight);
240 }
241 
242 static void inc_inflight_move_tail(struct unix_sock *u)
243 {
244 	atomic_long_inc(&u->inflight);
245 	/* If this still might be part of a cycle, move it to the end
246 	 * of the list, so that it's checked even if it was already
247 	 * passed over
248 	 */
249 	if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
250 		list_move_tail(&u->link, &gc_candidates);
251 }
252 
253 static bool gc_in_progress;
254 #define UNIX_INFLIGHT_TRIGGER_GC 16000
255 
256 void wait_for_unix_gc(void)
257 {
258 	/* If number of inflight sockets is insane,
259 	 * force a garbage collect right now.
260 	 */
261 	if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
262 		unix_gc();
263 	wait_event(unix_gc_wait, gc_in_progress == false);
264 }
265 
266 /* The external entry point: unix_gc() */
267 void unix_gc(void)
268 {
269 	struct unix_sock *u;
270 	struct unix_sock *next;
271 	struct sk_buff_head hitlist;
272 	struct list_head cursor;
273 	LIST_HEAD(not_cycle_list);
274 
275 	spin_lock(&unix_gc_lock);
276 
277 	/* Avoid a recursive GC. */
278 	if (gc_in_progress)
279 		goto out;
280 
281 	gc_in_progress = true;
282 	/* First, select candidates for garbage collection.  Only
283 	 * in-flight sockets are considered, and from those only ones
284 	 * which don't have any external reference.
285 	 *
286 	 * Holding unix_gc_lock will protect these candidates from
287 	 * being detached, and hence from gaining an external
288 	 * reference.  Since there are no possible receivers, all
289 	 * buffers currently on the candidates' queues stay there
290 	 * during the garbage collection.
291 	 *
292 	 * We also know that no new candidate can be added onto the
293 	 * receive queues.  Other, non candidate sockets _can_ be
294 	 * added to queue, so we must make sure only to touch
295 	 * candidates.
296 	 */
297 	list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
298 		long total_refs;
299 		long inflight_refs;
300 
301 		total_refs = file_count(u->sk.sk_socket->file);
302 		inflight_refs = atomic_long_read(&u->inflight);
303 
304 		BUG_ON(inflight_refs < 1);
305 		BUG_ON(total_refs < inflight_refs);
306 		if (total_refs == inflight_refs) {
307 			list_move_tail(&u->link, &gc_candidates);
308 			__set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
309 			__set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
310 		}
311 	}
312 
313 	/* Now remove all internal in-flight reference to children of
314 	 * the candidates.
315 	 */
316 	list_for_each_entry(u, &gc_candidates, link)
317 		scan_children(&u->sk, dec_inflight, NULL);
318 
319 	/* Restore the references for children of all candidates,
320 	 * which have remaining references.  Do this recursively, so
321 	 * only those remain, which form cyclic references.
322 	 *
323 	 * Use a "cursor" link, to make the list traversal safe, even
324 	 * though elements might be moved about.
325 	 */
326 	list_add(&cursor, &gc_candidates);
327 	while (cursor.next != &gc_candidates) {
328 		u = list_entry(cursor.next, struct unix_sock, link);
329 
330 		/* Move cursor to after the current position. */
331 		list_move(&cursor, &u->link);
332 
333 		if (atomic_long_read(&u->inflight) > 0) {
334 			list_move_tail(&u->link, &not_cycle_list);
335 			__clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
336 			scan_children(&u->sk, inc_inflight_move_tail, NULL);
337 		}
338 	}
339 	list_del(&cursor);
340 
341 	/* not_cycle_list contains those sockets which do not make up a
342 	 * cycle.  Restore these to the inflight list.
343 	 */
344 	while (!list_empty(&not_cycle_list)) {
345 		u = list_entry(not_cycle_list.next, struct unix_sock, link);
346 		__clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
347 		list_move_tail(&u->link, &gc_inflight_list);
348 	}
349 
350 	/* Now gc_candidates contains only garbage.  Restore original
351 	 * inflight counters for these as well, and remove the skbuffs
352 	 * which are creating the cycle(s).
353 	 */
354 	skb_queue_head_init(&hitlist);
355 	list_for_each_entry(u, &gc_candidates, link)
356 	scan_children(&u->sk, inc_inflight, &hitlist);
357 
358 	spin_unlock(&unix_gc_lock);
359 
360 	/* Here we are. Hitlist is filled. Die. */
361 	__skb_queue_purge(&hitlist);
362 
363 	spin_lock(&unix_gc_lock);
364 
365 	/* All candidates should have been detached by now. */
366 	BUG_ON(!list_empty(&gc_candidates));
367 	gc_in_progress = false;
368 	wake_up(&unix_gc_wait);
369 
370  out:
371 	spin_unlock(&unix_gc_lock);
372 }
373