2 * NET3: Garbage Collector For AF_UNIX sockets
5 * Copyright (C) Barak A. Pearlmutter.
6 * Released under the GPL version 2 or later.
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.
16 * Current optimizations:
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.
23 * Future optimizations:
25 * - don't just push entire root set; process in place
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.
33 * Alan Cox 07 Sept 1997 Vmalloc internal stack as needed.
34 * Cope with changing max_files.
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.
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.
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).
63 * Damn. Added missing check for ->dead in listen queues scanning.
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.
71 #include <linux/kernel.h>
72 #include <linux/string.h>
73 #include <linux/socket.h>
75 #include <linux/net.h>
77 #include <linux/slab.h>
78 #include <linux/skbuff.h>
79 #include <linux/netdevice.h>
80 #include <linux/file.h>
81 #include <linux/proc_fs.h>
82 #include <linux/mutex.h>
85 #include <net/af_unix.h>
87 #include <net/tcp_states.h>
89 /* Internal data structures and random procedures: */
91 static LIST_HEAD(gc_inflight_list);
92 static LIST_HEAD(gc_candidates);
93 static DEFINE_SPINLOCK(unix_gc_lock);
95 unsigned int unix_tot_inflight;
98 static struct sock *unix_get_socket(struct file *filp)
100 struct sock *u_sock = NULL;
101 struct inode *inode = filp->f_path.dentry->d_inode;
106 if (S_ISSOCK(inode->i_mode)) {
107 struct socket * sock = SOCKET_I(inode);
108 struct sock * s = sock->sk;
113 if (s && sock->ops && sock->ops->family == PF_UNIX)
120 * Keep the number of times in flight count for the file
121 * descriptor if it is for an AF_UNIX socket.
124 void unix_inflight(struct file *fp)
126 struct sock *s = unix_get_socket(fp);
128 struct unix_sock *u = unix_sk(s);
129 spin_lock(&unix_gc_lock);
130 if (atomic_long_inc_return(&u->inflight) == 1) {
131 BUG_ON(!list_empty(&u->link));
132 list_add_tail(&u->link, &gc_inflight_list);
134 BUG_ON(list_empty(&u->link));
137 spin_unlock(&unix_gc_lock);
141 void unix_notinflight(struct file *fp)
143 struct sock *s = unix_get_socket(fp);
145 struct unix_sock *u = unix_sk(s);
146 spin_lock(&unix_gc_lock);
147 BUG_ON(list_empty(&u->link));
148 if (atomic_long_dec_and_test(&u->inflight))
149 list_del_init(&u->link);
151 spin_unlock(&unix_gc_lock);
155 static inline struct sk_buff *sock_queue_head(struct sock *sk)
157 return (struct sk_buff *) &sk->sk_receive_queue;
160 #define receive_queue_for_each_skb(sk, next, skb) \
161 for (skb = sock_queue_head(sk)->next, next = skb->next; \
162 skb != sock_queue_head(sk); skb = next, next = skb->next)
164 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
165 struct sk_buff_head *hitlist)
168 struct sk_buff *next;
170 spin_lock(&x->sk_receive_queue.lock);
171 receive_queue_for_each_skb(x, next, skb) {
173 * Do we have file descriptors ?
175 if (UNIXCB(skb).fp) {
178 * Process the descriptors of this socket
180 int nfd = UNIXCB(skb).fp->count;
181 struct file **fp = UNIXCB(skb).fp->fp;
184 * Get the socket the fd matches
185 * if it indeed does so
187 struct sock *sk = unix_get_socket(*fp++);
189 struct unix_sock *u = unix_sk(sk);
192 * Ignore non-candidates, they could
193 * have been added to the queues after
194 * starting the garbage collection
196 if (u->gc_candidate) {
202 if (hit && hitlist != NULL) {
203 __skb_unlink(skb, &x->sk_receive_queue);
204 __skb_queue_tail(hitlist, skb);
208 spin_unlock(&x->sk_receive_queue.lock);
211 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
212 struct sk_buff_head *hitlist)
214 if (x->sk_state != TCP_LISTEN)
215 scan_inflight(x, func, hitlist);
218 struct sk_buff *next;
223 * For a listening socket collect the queued embryos
224 * and perform a scan on them as well.
226 spin_lock(&x->sk_receive_queue.lock);
227 receive_queue_for_each_skb(x, next, skb) {
228 u = unix_sk(skb->sk);
231 * An embryo cannot be in-flight, so it's safe
232 * to use the list link.
234 BUG_ON(!list_empty(&u->link));
235 list_add_tail(&u->link, &embryos);
237 spin_unlock(&x->sk_receive_queue.lock);
239 while (!list_empty(&embryos)) {
240 u = list_entry(embryos.next, struct unix_sock, link);
241 scan_inflight(&u->sk, func, hitlist);
242 list_del_init(&u->link);
247 static void dec_inflight(struct unix_sock *usk)
249 atomic_long_dec(&usk->inflight);
252 static void inc_inflight(struct unix_sock *usk)
254 atomic_long_inc(&usk->inflight);
257 static void inc_inflight_move_tail(struct unix_sock *u)
259 atomic_long_inc(&u->inflight);
261 * If this still might be part of a cycle, move it to the end
262 * of the list, so that it's checked even if it was already
265 if (u->gc_maybe_cycle)
266 list_move_tail(&u->link, &gc_candidates);
269 /* The external entry point: unix_gc() */
273 static bool gc_in_progress = false;
276 struct unix_sock *next;
277 struct sk_buff_head hitlist;
278 struct list_head cursor;
279 LIST_HEAD(not_cycle_list);
281 spin_lock(&unix_gc_lock);
283 /* Avoid a recursive GC. */
287 gc_in_progress = true;
289 * First, select candidates for garbage collection. Only
290 * in-flight sockets are considered, and from those only ones
291 * which don't have any external reference.
293 * Holding unix_gc_lock will protect these candidates from
294 * being detached, and hence from gaining an external
295 * reference. Since there are no possible receivers, all
296 * buffers currently on the candidates' queues stay there
297 * during the garbage collection.
299 * We also know that no new candidate can be added onto the
300 * receive queues. Other, non candidate sockets _can_ be
301 * added to queue, so we must make sure only to touch
304 list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
308 total_refs = file_count(u->sk.sk_socket->file);
309 inflight_refs = atomic_long_read(&u->inflight);
311 BUG_ON(inflight_refs < 1);
312 BUG_ON(total_refs < inflight_refs);
313 if (total_refs == inflight_refs) {
314 list_move_tail(&u->link, &gc_candidates);
316 u->gc_maybe_cycle = 1;
321 * Now remove all internal in-flight reference to children of
324 list_for_each_entry(u, &gc_candidates, link)
325 scan_children(&u->sk, dec_inflight, NULL);
328 * Restore the references for children of all candidates,
329 * which have remaining references. Do this recursively, so
330 * only those remain, which form cyclic references.
332 * Use a "cursor" link, to make the list traversal safe, even
333 * though elements might be moved about.
335 list_add(&cursor, &gc_candidates);
336 while (cursor.next != &gc_candidates) {
337 u = list_entry(cursor.next, struct unix_sock, link);
339 /* Move cursor to after the current position. */
340 list_move(&cursor, &u->link);
342 if (atomic_long_read(&u->inflight) > 0) {
343 list_move_tail(&u->link, ¬_cycle_list);
344 u->gc_maybe_cycle = 0;
345 scan_children(&u->sk, inc_inflight_move_tail, NULL);
351 * not_cycle_list contains those sockets which do not make up a
352 * cycle. Restore these to the inflight list.
354 while (!list_empty(¬_cycle_list)) {
355 u = list_entry(not_cycle_list.next, struct unix_sock, link);
357 list_move_tail(&u->link, &gc_inflight_list);
361 * Now gc_candidates contains only garbage. Restore original
362 * inflight counters for these as well, and remove the skbuffs
363 * which are creating the cycle(s).
365 skb_queue_head_init(&hitlist);
366 list_for_each_entry(u, &gc_candidates, link)
367 scan_children(&u->sk, inc_inflight, &hitlist);
369 spin_unlock(&unix_gc_lock);
371 /* Here we are. Hitlist is filled. Die. */
372 __skb_queue_purge(&hitlist);
374 spin_lock(&unix_gc_lock);
376 /* All candidates should have been detached by now. */
377 BUG_ON(!list_empty(&gc_candidates));
378 gc_in_progress = false;
381 spin_unlock(&unix_gc_lock);