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.
67 #include <linux/kernel.h>
68 #include <linux/string.h>
69 #include <linux/socket.h>
71 #include <linux/net.h>
73 #include <linux/slab.h>
74 #include <linux/skbuff.h>
75 #include <linux/netdevice.h>
76 #include <linux/file.h>
77 #include <linux/proc_fs.h>
78 #include <linux/mutex.h>
81 #include <net/af_unix.h>
83 #include <net/tcp_states.h>
85 /* Internal data structures and random procedures: */
87 #define GC_HEAD ((struct sock *)(-1))
88 #define GC_ORPHAN ((struct sock *)(-3))
90 static struct sock *gc_current = GC_HEAD; /* stack of objects to mark */
92 atomic_t unix_tot_inflight = ATOMIC_INIT(0);
95 static struct sock *unix_get_socket(struct file *filp)
97 struct sock *u_sock = NULL;
98 struct inode *inode = filp->f_path.dentry->d_inode;
103 if (S_ISSOCK(inode->i_mode)) {
104 struct socket * sock = SOCKET_I(inode);
105 struct sock * s = sock->sk;
110 if (s && sock->ops && sock->ops->family == PF_UNIX)
117 * Keep the number of times in flight count for the file
118 * descriptor if it is for an AF_UNIX socket.
121 void unix_inflight(struct file *fp)
123 struct sock *s = unix_get_socket(fp);
125 atomic_inc(&unix_sk(s)->inflight);
126 atomic_inc(&unix_tot_inflight);
130 void unix_notinflight(struct file *fp)
132 struct sock *s = unix_get_socket(fp);
134 atomic_dec(&unix_sk(s)->inflight);
135 atomic_dec(&unix_tot_inflight);
141 * Garbage Collector Support Functions
144 static inline struct sock *pop_stack(void)
146 struct sock *p = gc_current;
147 gc_current = unix_sk(p)->gc_tree;
151 static inline int empty_stack(void)
153 return gc_current == GC_HEAD;
156 static void maybe_unmark_and_push(struct sock *x)
158 struct unix_sock *u = unix_sk(x);
160 if (u->gc_tree != GC_ORPHAN)
163 u->gc_tree = gc_current;
168 /* The external entry point: unix_gc() */
172 static DEFINE_MUTEX(unix_gc_sem);
175 struct sk_buff_head hitlist;
179 * Avoid a recursive GC.
182 if (!mutex_trylock(&unix_gc_sem))
185 spin_lock(&unix_table_lock);
187 forall_unix_sockets(i, s)
189 unix_sk(s)->gc_tree = GC_ORPHAN;
192 * Everything is now marked
195 /* Invariant to be maintained:
196 - everything unmarked is either:
197 -- (a) on the stack, or
198 -- (b) has all of its children unmarked
199 - everything on the stack is always unmarked
200 - nothing is ever pushed onto the stack twice, because:
201 -- nothing previously unmarked is ever pushed on the stack
208 forall_unix_sockets(i, s)
213 * If all instances of the descriptor are not
214 * in flight we are in use.
216 * Special case: when socket s is embrion, it may be
217 * hashed but still not in queue of listening socket.
218 * In this case (see unix_create1()) we set artificial
219 * negative inflight counter to close race window.
220 * It is trick of course and dirty one.
222 if (s->sk_socket && s->sk_socket->file)
223 open_count = file_count(s->sk_socket->file);
224 if (open_count > atomic_read(&unix_sk(s)->inflight))
225 maybe_unmark_and_push(s);
232 while (!empty_stack())
234 struct sock *x = pop_stack();
237 spin_lock(&x->sk_receive_queue.lock);
238 skb = skb_peek(&x->sk_receive_queue);
241 * Loop through all but first born
244 while (skb && skb != (struct sk_buff *)&x->sk_receive_queue) {
246 * Do we have file descriptors ?
251 * Process the descriptors of this socket
253 int nfd=UNIXCB(skb).fp->count;
254 struct file **fp = UNIXCB(skb).fp->fp;
258 * Get the socket the fd matches if
261 if((sk=unix_get_socket(*fp++))!=NULL)
263 maybe_unmark_and_push(sk);
267 /* We have to scan not-yet-accepted ones too */
268 if (x->sk_state == TCP_LISTEN)
269 maybe_unmark_and_push(skb->sk);
272 spin_unlock(&x->sk_receive_queue.lock);
276 skb_queue_head_init(&hitlist);
278 forall_unix_sockets(i, s)
280 struct unix_sock *u = unix_sk(s);
282 if (u->gc_tree == GC_ORPHAN) {
283 struct sk_buff *nextsk;
285 spin_lock(&s->sk_receive_queue.lock);
286 skb = skb_peek(&s->sk_receive_queue);
288 skb != (struct sk_buff *)&s->sk_receive_queue) {
291 * Do we have file descriptors ?
293 if (UNIXCB(skb).fp) {
295 &s->sk_receive_queue);
296 __skb_queue_tail(&hitlist, skb);
300 spin_unlock(&s->sk_receive_queue.lock);
302 u->gc_tree = GC_ORPHAN;
304 spin_unlock(&unix_table_lock);
307 * Here we are. Hitlist is filled. Die.
310 __skb_queue_purge(&hitlist);
311 mutex_unlock(&unix_gc_sem);