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
unix_get_socket(struct file * filp)99 struct sock *unix_get_socket(struct file *filp)
100 {
101 struct sock *u_sock = NULL;
102 struct inode *inode = file_inode(filp);
103
104 /*
105 * Socket ?
106 */
107 if (S_ISSOCK(inode->i_mode) && !(filp->f_mode & FMODE_PATH)) {
108 struct socket *sock = SOCKET_I(inode);
109 struct sock *s = sock->sk;
110
111 /*
112 * PF_UNIX ?
113 */
114 if (s && sock->ops && sock->ops->family == PF_UNIX)
115 u_sock = s;
116 }
117 return u_sock;
118 }
119
120 /*
121 * Keep the number of times in flight count for the file
122 * descriptor if it is for an AF_UNIX socket.
123 */
124
unix_inflight(struct file * fp)125 void unix_inflight(struct file *fp)
126 {
127 struct sock *s = unix_get_socket(fp);
128
129 spin_lock(&unix_gc_lock);
130
131 if (s) {
132 struct unix_sock *u = unix_sk(s);
133
134 if (atomic_long_inc_return(&u->inflight) == 1) {
135 BUG_ON(!list_empty(&u->link));
136 list_add_tail(&u->link, &gc_inflight_list);
137 } else {
138 BUG_ON(list_empty(&u->link));
139 }
140 unix_tot_inflight++;
141 }
142 fp->f_cred->user->unix_inflight++;
143 spin_unlock(&unix_gc_lock);
144 }
145
unix_notinflight(struct file * fp)146 void unix_notinflight(struct file *fp)
147 {
148 struct sock *s = unix_get_socket(fp);
149
150 spin_lock(&unix_gc_lock);
151
152 if (s) {
153 struct unix_sock *u = unix_sk(s);
154
155 BUG_ON(!atomic_long_read(&u->inflight));
156 BUG_ON(list_empty(&u->link));
157 if (atomic_long_dec_and_test(&u->inflight))
158 list_del_init(&u->link);
159 unix_tot_inflight--;
160 }
161 fp->f_cred->user->unix_inflight--;
162 spin_unlock(&unix_gc_lock);
163 }
164
scan_inflight(struct sock * x,void (* func)(struct unix_sock *),struct sk_buff_head * hitlist)165 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
166 struct sk_buff_head *hitlist)
167 {
168 struct sk_buff *skb;
169 struct sk_buff *next;
170
171 spin_lock(&x->sk_receive_queue.lock);
172 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
173 /*
174 * Do we have file descriptors ?
175 */
176 if (UNIXCB(skb).fp) {
177 bool hit = false;
178 /*
179 * Process the descriptors of this socket
180 */
181 int nfd = UNIXCB(skb).fp->count;
182 struct file **fp = UNIXCB(skb).fp->fp;
183 while (nfd--) {
184 /*
185 * Get the socket the fd matches
186 * if it indeed does so
187 */
188 struct sock *sk = unix_get_socket(*fp++);
189 if (sk) {
190 struct unix_sock *u = unix_sk(sk);
191
192 /*
193 * Ignore non-candidates, they could
194 * have been added to the queues after
195 * starting the garbage collection
196 */
197 if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
198 hit = true;
199 func(u);
200 }
201 }
202 }
203 if (hit && hitlist != NULL) {
204 __skb_unlink(skb, &x->sk_receive_queue);
205 __skb_queue_tail(hitlist, skb);
206 }
207 }
208 }
209 spin_unlock(&x->sk_receive_queue.lock);
210 }
211
scan_children(struct sock * x,void (* func)(struct unix_sock *),struct sk_buff_head * hitlist)212 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
213 struct sk_buff_head *hitlist)
214 {
215 if (x->sk_state != TCP_LISTEN)
216 scan_inflight(x, func, hitlist);
217 else {
218 struct sk_buff *skb;
219 struct sk_buff *next;
220 struct unix_sock *u;
221 LIST_HEAD(embryos);
222
223 /*
224 * For a listening socket collect the queued embryos
225 * and perform a scan on them as well.
226 */
227 spin_lock(&x->sk_receive_queue.lock);
228 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
229 u = unix_sk(skb->sk);
230
231 /*
232 * An embryo cannot be in-flight, so it's safe
233 * to use the list link.
234 */
235 BUG_ON(!list_empty(&u->link));
236 list_add_tail(&u->link, &embryos);
237 }
238 spin_unlock(&x->sk_receive_queue.lock);
239
240 while (!list_empty(&embryos)) {
241 u = list_entry(embryos.next, struct unix_sock, link);
242 scan_inflight(&u->sk, func, hitlist);
243 list_del_init(&u->link);
244 }
245 }
246 }
247
dec_inflight(struct unix_sock * usk)248 static void dec_inflight(struct unix_sock *usk)
249 {
250 atomic_long_dec(&usk->inflight);
251 }
252
inc_inflight(struct unix_sock * usk)253 static void inc_inflight(struct unix_sock *usk)
254 {
255 atomic_long_inc(&usk->inflight);
256 }
257
inc_inflight_move_tail(struct unix_sock * u)258 static void inc_inflight_move_tail(struct unix_sock *u)
259 {
260 atomic_long_inc(&u->inflight);
261 /*
262 * If this still might be part of a cycle, move it to the end
263 * of the list, so that it's checked even if it was already
264 * passed over
265 */
266 if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
267 list_move_tail(&u->link, &gc_candidates);
268 }
269
270 static bool gc_in_progress;
271 #define UNIX_INFLIGHT_TRIGGER_GC 16000
272
wait_for_unix_gc(void)273 void wait_for_unix_gc(void)
274 {
275 /*
276 * If number of inflight sockets is insane,
277 * force a garbage collect right now.
278 */
279 if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
280 unix_gc();
281 wait_event(unix_gc_wait, gc_in_progress == false);
282 }
283
284 /* The external entry point: unix_gc() */
unix_gc(void)285 void unix_gc(void)
286 {
287 struct unix_sock *u;
288 struct unix_sock *next;
289 struct sk_buff_head hitlist;
290 struct list_head cursor;
291 LIST_HEAD(not_cycle_list);
292
293 spin_lock(&unix_gc_lock);
294
295 /* Avoid a recursive GC. */
296 if (gc_in_progress)
297 goto out;
298
299 gc_in_progress = true;
300 /*
301 * First, select candidates for garbage collection. Only
302 * in-flight sockets are considered, and from those only ones
303 * which don't have any external reference.
304 *
305 * Holding unix_gc_lock will protect these candidates from
306 * being detached, and hence from gaining an external
307 * reference. Since there are no possible receivers, all
308 * buffers currently on the candidates' queues stay there
309 * during the garbage collection.
310 *
311 * We also know that no new candidate can be added onto the
312 * receive queues. Other, non candidate sockets _can_ be
313 * added to queue, so we must make sure only to touch
314 * candidates.
315 */
316 list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
317 long total_refs;
318 long inflight_refs;
319
320 total_refs = file_count(u->sk.sk_socket->file);
321 inflight_refs = atomic_long_read(&u->inflight);
322
323 BUG_ON(inflight_refs < 1);
324 BUG_ON(total_refs < inflight_refs);
325 if (total_refs == inflight_refs) {
326 list_move_tail(&u->link, &gc_candidates);
327 __set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
328 __set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
329 }
330 }
331
332 /*
333 * Now remove all internal in-flight reference to children of
334 * the candidates.
335 */
336 list_for_each_entry(u, &gc_candidates, link)
337 scan_children(&u->sk, dec_inflight, NULL);
338
339 /*
340 * Restore the references for children of all candidates,
341 * which have remaining references. Do this recursively, so
342 * only those remain, which form cyclic references.
343 *
344 * Use a "cursor" link, to make the list traversal safe, even
345 * though elements might be moved about.
346 */
347 list_add(&cursor, &gc_candidates);
348 while (cursor.next != &gc_candidates) {
349 u = list_entry(cursor.next, struct unix_sock, link);
350
351 /* Move cursor to after the current position. */
352 list_move(&cursor, &u->link);
353
354 if (atomic_long_read(&u->inflight) > 0) {
355 list_move_tail(&u->link, ¬_cycle_list);
356 __clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
357 scan_children(&u->sk, inc_inflight_move_tail, NULL);
358 }
359 }
360 list_del(&cursor);
361
362 /* Now gc_candidates contains only garbage. Restore original
363 * inflight counters for these as well, and remove the skbuffs
364 * which are creating the cycle(s).
365 */
366 skb_queue_head_init(&hitlist);
367 list_for_each_entry(u, &gc_candidates, link)
368 scan_children(&u->sk, inc_inflight, &hitlist);
369
370 /*
371 * not_cycle_list contains those sockets which do not make up a
372 * cycle. Restore these to the inflight list.
373 */
374 while (!list_empty(¬_cycle_list)) {
375 u = list_entry(not_cycle_list.next, struct unix_sock, link);
376 __clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
377 list_move_tail(&u->link, &gc_inflight_list);
378 }
379
380 spin_unlock(&unix_gc_lock);
381
382 /* Here we are. Hitlist is filled. Die. */
383 __skb_queue_purge(&hitlist);
384
385 spin_lock(&unix_gc_lock);
386
387 /* All candidates should have been detached by now. */
388 BUG_ON(!list_empty(&gc_candidates));
389 gc_in_progress = false;
390 wake_up(&unix_gc_wait);
391
392 out:
393 spin_unlock(&unix_gc_lock);
394 }
395