xref: /openbmc/linux/net/unix/garbage.c (revision 89e33ea7)
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 #include "scm.h"
90 
91 /* Internal data structures and random procedures: */
92 
93 static LIST_HEAD(gc_candidates);
94 static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
95 
96 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
97 			  struct sk_buff_head *hitlist)
98 {
99 	struct sk_buff *skb;
100 	struct sk_buff *next;
101 
102 	spin_lock(&x->sk_receive_queue.lock);
103 	skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
104 		/* Do we have file descriptors ? */
105 		if (UNIXCB(skb).fp) {
106 			bool hit = false;
107 			/* Process the descriptors of this socket */
108 			int nfd = UNIXCB(skb).fp->count;
109 			struct file **fp = UNIXCB(skb).fp->fp;
110 
111 			while (nfd--) {
112 				/* Get the socket the fd matches if it indeed does so */
113 				struct sock *sk = unix_get_socket(*fp++);
114 
115 				if (sk) {
116 					struct unix_sock *u = unix_sk(sk);
117 
118 					/* Ignore non-candidates, they could
119 					 * have been added to the queues after
120 					 * starting the garbage collection
121 					 */
122 					if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
123 						hit = true;
124 
125 						func(u);
126 					}
127 				}
128 			}
129 			if (hit && hitlist != NULL) {
130 				__skb_unlink(skb, &x->sk_receive_queue);
131 				__skb_queue_tail(hitlist, skb);
132 			}
133 		}
134 	}
135 	spin_unlock(&x->sk_receive_queue.lock);
136 }
137 
138 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
139 			  struct sk_buff_head *hitlist)
140 {
141 	if (x->sk_state != TCP_LISTEN) {
142 		scan_inflight(x, func, hitlist);
143 	} else {
144 		struct sk_buff *skb;
145 		struct sk_buff *next;
146 		struct unix_sock *u;
147 		LIST_HEAD(embryos);
148 
149 		/* For a listening socket collect the queued embryos
150 		 * and perform a scan on them as well.
151 		 */
152 		spin_lock(&x->sk_receive_queue.lock);
153 		skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
154 			u = unix_sk(skb->sk);
155 
156 			/* An embryo cannot be in-flight, so it's safe
157 			 * to use the list link.
158 			 */
159 			BUG_ON(!list_empty(&u->link));
160 			list_add_tail(&u->link, &embryos);
161 		}
162 		spin_unlock(&x->sk_receive_queue.lock);
163 
164 		while (!list_empty(&embryos)) {
165 			u = list_entry(embryos.next, struct unix_sock, link);
166 			scan_inflight(&u->sk, func, hitlist);
167 			list_del_init(&u->link);
168 		}
169 	}
170 }
171 
172 static void dec_inflight(struct unix_sock *usk)
173 {
174 	atomic_long_dec(&usk->inflight);
175 }
176 
177 static void inc_inflight(struct unix_sock *usk)
178 {
179 	atomic_long_inc(&usk->inflight);
180 }
181 
182 static void inc_inflight_move_tail(struct unix_sock *u)
183 {
184 	atomic_long_inc(&u->inflight);
185 	/* If this still might be part of a cycle, move it to the end
186 	 * of the list, so that it's checked even if it was already
187 	 * passed over
188 	 */
189 	if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
190 		list_move_tail(&u->link, &gc_candidates);
191 }
192 
193 static bool gc_in_progress;
194 #define UNIX_INFLIGHT_TRIGGER_GC 16000
195 
196 void wait_for_unix_gc(void)
197 {
198 	/* If number of inflight sockets is insane,
199 	 * force a garbage collect right now.
200 	 */
201 	if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
202 		unix_gc();
203 	wait_event(unix_gc_wait, gc_in_progress == false);
204 }
205 
206 /* The external entry point: unix_gc() */
207 void unix_gc(void)
208 {
209 	struct unix_sock *u;
210 	struct unix_sock *next;
211 	struct sk_buff_head hitlist;
212 	struct list_head cursor;
213 	LIST_HEAD(not_cycle_list);
214 
215 	spin_lock(&unix_gc_lock);
216 
217 	/* Avoid a recursive GC. */
218 	if (gc_in_progress)
219 		goto out;
220 
221 	gc_in_progress = true;
222 	/* First, select candidates for garbage collection.  Only
223 	 * in-flight sockets are considered, and from those only ones
224 	 * which don't have any external reference.
225 	 *
226 	 * Holding unix_gc_lock will protect these candidates from
227 	 * being detached, and hence from gaining an external
228 	 * reference.  Since there are no possible receivers, all
229 	 * buffers currently on the candidates' queues stay there
230 	 * during the garbage collection.
231 	 *
232 	 * We also know that no new candidate can be added onto the
233 	 * receive queues.  Other, non candidate sockets _can_ be
234 	 * added to queue, so we must make sure only to touch
235 	 * candidates.
236 	 */
237 	list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
238 		long total_refs;
239 		long inflight_refs;
240 
241 		total_refs = file_count(u->sk.sk_socket->file);
242 		inflight_refs = atomic_long_read(&u->inflight);
243 
244 		BUG_ON(inflight_refs < 1);
245 		BUG_ON(total_refs < inflight_refs);
246 		if (total_refs == inflight_refs) {
247 			list_move_tail(&u->link, &gc_candidates);
248 			__set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
249 			__set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
250 		}
251 	}
252 
253 	/* Now remove all internal in-flight reference to children of
254 	 * the candidates.
255 	 */
256 	list_for_each_entry(u, &gc_candidates, link)
257 		scan_children(&u->sk, dec_inflight, NULL);
258 
259 	/* Restore the references for children of all candidates,
260 	 * which have remaining references.  Do this recursively, so
261 	 * only those remain, which form cyclic references.
262 	 *
263 	 * Use a "cursor" link, to make the list traversal safe, even
264 	 * though elements might be moved about.
265 	 */
266 	list_add(&cursor, &gc_candidates);
267 	while (cursor.next != &gc_candidates) {
268 		u = list_entry(cursor.next, struct unix_sock, link);
269 
270 		/* Move cursor to after the current position. */
271 		list_move(&cursor, &u->link);
272 
273 		if (atomic_long_read(&u->inflight) > 0) {
274 			list_move_tail(&u->link, &not_cycle_list);
275 			__clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
276 			scan_children(&u->sk, inc_inflight_move_tail, NULL);
277 		}
278 	}
279 	list_del(&cursor);
280 
281 	/* Now gc_candidates contains only garbage.  Restore original
282 	 * inflight counters for these as well, and remove the skbuffs
283 	 * which are creating the cycle(s).
284 	 */
285 	skb_queue_head_init(&hitlist);
286 	list_for_each_entry(u, &gc_candidates, link)
287 		scan_children(&u->sk, inc_inflight, &hitlist);
288 
289 	/* not_cycle_list contains those sockets which do not make up a
290 	 * cycle.  Restore these to the inflight list.
291 	 */
292 	while (!list_empty(&not_cycle_list)) {
293 		u = list_entry(not_cycle_list.next, struct unix_sock, link);
294 		__clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
295 		list_move_tail(&u->link, &gc_inflight_list);
296 	}
297 
298 	spin_unlock(&unix_gc_lock);
299 
300 	/* Here we are. Hitlist is filled. Die. */
301 	__skb_queue_purge(&hitlist);
302 
303 	spin_lock(&unix_gc_lock);
304 
305 	/* All candidates should have been detached by now. */
306 	BUG_ON(!list_empty(&gc_candidates));
307 	gc_in_progress = false;
308 	wake_up(&unix_gc_wait);
309 
310  out:
311 	spin_unlock(&unix_gc_lock);
312 }
313