xref: /linux/net/unix/garbage.c (revision 2da68a77)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * NET3:	Garbage Collector For AF_UNIX sockets
4  *
5  * Garbage Collector:
6  *	Copyright (C) Barak A. Pearlmutter.
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  *  Fixes:
28  *	Alan Cox	07 Sept	1997	Vmalloc internal stack as needed.
29  *					Cope with changing max_files.
30  *	Al Viro		11 Oct 1998
31  *		Graph may have cycles. That is, we can send the descriptor
32  *		of foo to bar and vice versa. Current code chokes on that.
33  *		Fix: move SCM_RIGHTS ones into the separate list and then
34  *		skb_free() them all instead of doing explicit fput's.
35  *		Another problem: since fput() may block somebody may
36  *		create a new unix_socket when we are in the middle of sweep
37  *		phase. Fix: revert the logic wrt MARKED. Mark everything
38  *		upon the beginning and unmark non-junk ones.
39  *
40  *		[12 Oct 1998] AAARGH! New code purges all SCM_RIGHTS
41  *		sent to connect()'ed but still not accept()'ed sockets.
42  *		Fixed. Old code had slightly different problem here:
43  *		extra fput() in situation when we passed the descriptor via
44  *		such socket and closed it (descriptor). That would happen on
45  *		each unix_gc() until the accept(). Since the struct file in
46  *		question would go to the free list and might be reused...
47  *		That might be the reason of random oopses on filp_close()
48  *		in unrelated processes.
49  *
50  *	AV		28 Feb 1999
51  *		Kill the explicit allocation of stack. Now we keep the tree
52  *		with root in dummy + pointer (gc_current) to one of the nodes.
53  *		Stack is represented as path from gc_current to dummy. Unmark
54  *		now means "add to tree". Push == "make it a son of gc_current".
55  *		Pop == "move gc_current to parent". We keep only pointers to
56  *		parents (->gc_tree).
57  *	AV		1 Mar 1999
58  *		Damn. Added missing check for ->dead in listen queues scanning.
59  *
60  *	Miklos Szeredi 25 Jun 2007
61  *		Reimplement with a cycle collecting algorithm. This should
62  *		solve several problems with the previous code, like being racy
63  *		wrt receive and holding up unrelated socket operations.
64  */
65 
66 #include <linux/kernel.h>
67 #include <linux/string.h>
68 #include <linux/socket.h>
69 #include <linux/un.h>
70 #include <linux/net.h>
71 #include <linux/fs.h>
72 #include <linux/skbuff.h>
73 #include <linux/netdevice.h>
74 #include <linux/file.h>
75 #include <linux/proc_fs.h>
76 #include <linux/mutex.h>
77 #include <linux/wait.h>
78 
79 #include <net/sock.h>
80 #include <net/af_unix.h>
81 #include <net/scm.h>
82 #include <net/tcp_states.h>
83 
84 #include "scm.h"
85 
86 /* Internal data structures and random procedures: */
87 
88 static LIST_HEAD(gc_candidates);
89 static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
90 
91 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
92 			  struct sk_buff_head *hitlist)
93 {
94 	struct sk_buff *skb;
95 	struct sk_buff *next;
96 
97 	spin_lock(&x->sk_receive_queue.lock);
98 	skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
99 		/* Do we have file descriptors ? */
100 		if (UNIXCB(skb).fp) {
101 			bool hit = false;
102 			/* Process the descriptors of this socket */
103 			int nfd = UNIXCB(skb).fp->count;
104 			struct file **fp = UNIXCB(skb).fp->fp;
105 
106 			while (nfd--) {
107 				/* Get the socket the fd matches if it indeed does so */
108 				struct sock *sk = unix_get_socket(*fp++);
109 
110 				if (sk) {
111 					struct unix_sock *u = unix_sk(sk);
112 
113 					/* Ignore non-candidates, they could
114 					 * have been added to the queues after
115 					 * starting the garbage collection
116 					 */
117 					if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
118 						hit = true;
119 
120 						func(u);
121 					}
122 				}
123 			}
124 			if (hit && hitlist != NULL) {
125 				__skb_unlink(skb, &x->sk_receive_queue);
126 				__skb_queue_tail(hitlist, skb);
127 			}
128 		}
129 	}
130 	spin_unlock(&x->sk_receive_queue.lock);
131 }
132 
133 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
134 			  struct sk_buff_head *hitlist)
135 {
136 	if (x->sk_state != TCP_LISTEN) {
137 		scan_inflight(x, func, hitlist);
138 	} else {
139 		struct sk_buff *skb;
140 		struct sk_buff *next;
141 		struct unix_sock *u;
142 		LIST_HEAD(embryos);
143 
144 		/* For a listening socket collect the queued embryos
145 		 * and perform a scan on them as well.
146 		 */
147 		spin_lock(&x->sk_receive_queue.lock);
148 		skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
149 			u = unix_sk(skb->sk);
150 
151 			/* An embryo cannot be in-flight, so it's safe
152 			 * to use the list link.
153 			 */
154 			BUG_ON(!list_empty(&u->link));
155 			list_add_tail(&u->link, &embryos);
156 		}
157 		spin_unlock(&x->sk_receive_queue.lock);
158 
159 		while (!list_empty(&embryos)) {
160 			u = list_entry(embryos.next, struct unix_sock, link);
161 			scan_inflight(&u->sk, func, hitlist);
162 			list_del_init(&u->link);
163 		}
164 	}
165 }
166 
167 static void dec_inflight(struct unix_sock *usk)
168 {
169 	atomic_long_dec(&usk->inflight);
170 }
171 
172 static void inc_inflight(struct unix_sock *usk)
173 {
174 	atomic_long_inc(&usk->inflight);
175 }
176 
177 static void inc_inflight_move_tail(struct unix_sock *u)
178 {
179 	atomic_long_inc(&u->inflight);
180 	/* If this still might be part of a cycle, move it to the end
181 	 * of the list, so that it's checked even if it was already
182 	 * passed over
183 	 */
184 	if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
185 		list_move_tail(&u->link, &gc_candidates);
186 }
187 
188 static bool gc_in_progress;
189 #define UNIX_INFLIGHT_TRIGGER_GC 16000
190 
191 void wait_for_unix_gc(void)
192 {
193 	/* If number of inflight sockets is insane,
194 	 * force a garbage collect right now.
195 	 * Paired with the WRITE_ONCE() in unix_inflight(),
196 	 * unix_notinflight() and gc_in_progress().
197 	 */
198 	if (READ_ONCE(unix_tot_inflight) > UNIX_INFLIGHT_TRIGGER_GC &&
199 	    !READ_ONCE(gc_in_progress))
200 		unix_gc();
201 	wait_event(unix_gc_wait, gc_in_progress == false);
202 }
203 
204 /* The external entry point: unix_gc() */
205 void unix_gc(void)
206 {
207 	struct sk_buff *next_skb, *skb;
208 	struct unix_sock *u;
209 	struct unix_sock *next;
210 	struct sk_buff_head hitlist;
211 	struct list_head cursor;
212 	LIST_HEAD(not_cycle_list);
213 
214 	spin_lock(&unix_gc_lock);
215 
216 	/* Avoid a recursive GC. */
217 	if (gc_in_progress)
218 		goto out;
219 
220 	/* Paired with READ_ONCE() in wait_for_unix_gc(). */
221 	WRITE_ONCE(gc_in_progress, true);
222 
223 	/* First, select candidates for garbage collection.  Only
224 	 * in-flight sockets are considered, and from those only ones
225 	 * which don't have any external reference.
226 	 *
227 	 * Holding unix_gc_lock will protect these candidates from
228 	 * being detached, and hence from gaining an external
229 	 * reference.  Since there are no possible receivers, all
230 	 * buffers currently on the candidates' queues stay there
231 	 * during the garbage collection.
232 	 *
233 	 * We also know that no new candidate can be added onto the
234 	 * receive queues.  Other, non candidate sockets _can_ be
235 	 * added to queue, so we must make sure only to touch
236 	 * candidates.
237 	 */
238 	list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
239 		long total_refs;
240 		long inflight_refs;
241 
242 		total_refs = file_count(u->sk.sk_socket->file);
243 		inflight_refs = atomic_long_read(&u->inflight);
244 
245 		BUG_ON(inflight_refs < 1);
246 		BUG_ON(total_refs < inflight_refs);
247 		if (total_refs == inflight_refs) {
248 			list_move_tail(&u->link, &gc_candidates);
249 			__set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
250 			__set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
251 		}
252 	}
253 
254 	/* Now remove all internal in-flight reference to children of
255 	 * the candidates.
256 	 */
257 	list_for_each_entry(u, &gc_candidates, link)
258 		scan_children(&u->sk, dec_inflight, NULL);
259 
260 	/* Restore the references for children of all candidates,
261 	 * which have remaining references.  Do this recursively, so
262 	 * only those remain, which form cyclic references.
263 	 *
264 	 * Use a "cursor" link, to make the list traversal safe, even
265 	 * though elements might be moved about.
266 	 */
267 	list_add(&cursor, &gc_candidates);
268 	while (cursor.next != &gc_candidates) {
269 		u = list_entry(cursor.next, struct unix_sock, link);
270 
271 		/* Move cursor to after the current position. */
272 		list_move(&cursor, &u->link);
273 
274 		if (atomic_long_read(&u->inflight) > 0) {
275 			list_move_tail(&u->link, &not_cycle_list);
276 			__clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
277 			scan_children(&u->sk, inc_inflight_move_tail, NULL);
278 		}
279 	}
280 	list_del(&cursor);
281 
282 	/* Now gc_candidates contains only garbage.  Restore original
283 	 * inflight counters for these as well, and remove the skbuffs
284 	 * which are creating the cycle(s).
285 	 */
286 	skb_queue_head_init(&hitlist);
287 	list_for_each_entry(u, &gc_candidates, link)
288 		scan_children(&u->sk, inc_inflight, &hitlist);
289 
290 	/* not_cycle_list contains those sockets which do not make up a
291 	 * cycle.  Restore these to the inflight list.
292 	 */
293 	while (!list_empty(&not_cycle_list)) {
294 		u = list_entry(not_cycle_list.next, struct unix_sock, link);
295 		__clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
296 		list_move_tail(&u->link, &gc_inflight_list);
297 	}
298 
299 	spin_unlock(&unix_gc_lock);
300 
301 	/* We need io_uring to clean its registered files, ignore all io_uring
302 	 * originated skbs. It's fine as io_uring doesn't keep references to
303 	 * other io_uring instances and so killing all other files in the cycle
304 	 * will put all io_uring references forcing it to go through normal
305 	 * release.path eventually putting registered files.
306 	 */
307 	skb_queue_walk_safe(&hitlist, skb, next_skb) {
308 		if (skb->scm_io_uring) {
309 			__skb_unlink(skb, &hitlist);
310 			skb_queue_tail(&skb->sk->sk_receive_queue, skb);
311 		}
312 	}
313 
314 	/* Here we are. Hitlist is filled. Die. */
315 	__skb_queue_purge(&hitlist);
316 
317 	spin_lock(&unix_gc_lock);
318 
319 	/* There could be io_uring registered files, just push them back to
320 	 * the inflight list
321 	 */
322 	list_for_each_entry_safe(u, next, &gc_candidates, link)
323 		list_move_tail(&u->link, &gc_inflight_list);
324 
325 	/* All candidates should have been detached by now. */
326 	BUG_ON(!list_empty(&gc_candidates));
327 
328 	/* Paired with READ_ONCE() in wait_for_unix_gc(). */
329 	WRITE_ONCE(gc_in_progress, false);
330 
331 	wake_up(&unix_gc_wait);
332 
333  out:
334 	spin_unlock(&unix_gc_lock);
335 }
336