blob: 746c2f4a5fa62a290418a0b7780b165678ae7f96 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 */
66
67#include <linux/kernel.h>
68#include <linux/sched.h>
69#include <linux/string.h>
70#include <linux/socket.h>
71#include <linux/un.h>
72#include <linux/net.h>
73#include <linux/fs.h>
74#include <linux/slab.h>
75#include <linux/skbuff.h>
76#include <linux/netdevice.h>
77#include <linux/file.h>
78#include <linux/proc_fs.h>
Arjan van de Ven4a3e2f72006-03-20 22:33:17 -080079#include <linux/mutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070080
81#include <net/sock.h>
82#include <net/af_unix.h>
83#include <net/scm.h>
Arnaldo Carvalho de Meloc752f072005-08-09 20:08:28 -070084#include <net/tcp_states.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
86/* Internal data structures and random procedures: */
87
88#define GC_HEAD ((struct sock *)(-1))
89#define GC_ORPHAN ((struct sock *)(-3))
90
91static struct sock *gc_current = GC_HEAD; /* stack of objects to mark */
92
93atomic_t unix_tot_inflight = ATOMIC_INIT(0);
94
95
96static struct sock *unix_get_socket(struct file *filp)
97{
98 struct sock *u_sock = NULL;
99 struct inode *inode = filp->f_dentry->d_inode;
100
101 /*
102 * Socket ?
103 */
104 if (S_ISSOCK(inode->i_mode)) {
105 struct socket * sock = SOCKET_I(inode);
106 struct sock * s = sock->sk;
107
108 /*
109 * PF_UNIX ?
110 */
111 if (s && sock->ops && sock->ops->family == PF_UNIX)
112 u_sock = s;
113 }
114 return u_sock;
115}
116
117/*
118 * Keep the number of times in flight count for the file
119 * descriptor if it is for an AF_UNIX socket.
120 */
121
122void unix_inflight(struct file *fp)
123{
124 struct sock *s = unix_get_socket(fp);
125 if(s) {
126 atomic_inc(&unix_sk(s)->inflight);
127 atomic_inc(&unix_tot_inflight);
128 }
129}
130
131void unix_notinflight(struct file *fp)
132{
133 struct sock *s = unix_get_socket(fp);
134 if(s) {
135 atomic_dec(&unix_sk(s)->inflight);
136 atomic_dec(&unix_tot_inflight);
137 }
138}
139
140
141/*
142 * Garbage Collector Support Functions
143 */
144
145static inline struct sock *pop_stack(void)
146{
147 struct sock *p = gc_current;
148 gc_current = unix_sk(p)->gc_tree;
149 return p;
150}
151
152static inline int empty_stack(void)
153{
154 return gc_current == GC_HEAD;
155}
156
157static void maybe_unmark_and_push(struct sock *x)
158{
159 struct unix_sock *u = unix_sk(x);
160
161 if (u->gc_tree != GC_ORPHAN)
162 return;
163 sock_hold(x);
164 u->gc_tree = gc_current;
165 gc_current = x;
166}
167
168
169/* The external entry point: unix_gc() */
170
171void unix_gc(void)
172{
Arjan van de Ven4a3e2f72006-03-20 22:33:17 -0800173 static DEFINE_MUTEX(unix_gc_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 int i;
175 struct sock *s;
176 struct sk_buff_head hitlist;
177 struct sk_buff *skb;
178
179 /*
180 * Avoid a recursive GC.
181 */
182
Arjan van de Ven4a3e2f72006-03-20 22:33:17 -0800183 if (!mutex_trylock(&unix_gc_sem))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 return;
185
David S. Millerfbe9cc42005-12-13 23:26:29 -0800186 spin_lock(&unix_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187
188 forall_unix_sockets(i, s)
189 {
190 unix_sk(s)->gc_tree = GC_ORPHAN;
191 }
192 /*
193 * Everything is now marked
194 */
195
196 /* Invariant to be maintained:
197 - everything unmarked is either:
198 -- (a) on the stack, or
199 -- (b) has all of its children unmarked
200 - everything on the stack is always unmarked
201 - nothing is ever pushed onto the stack twice, because:
202 -- nothing previously unmarked is ever pushed on the stack
203 */
204
205 /*
206 * Push root set
207 */
208
209 forall_unix_sockets(i, s)
210 {
211 int open_count = 0;
212
213 /*
214 * If all instances of the descriptor are not
215 * in flight we are in use.
216 *
217 * Special case: when socket s is embrion, it may be
218 * hashed but still not in queue of listening socket.
219 * In this case (see unix_create1()) we set artificial
220 * negative inflight counter to close race window.
221 * It is trick of course and dirty one.
222 */
223 if (s->sk_socket && s->sk_socket->file)
224 open_count = file_count(s->sk_socket->file);
225 if (open_count > atomic_read(&unix_sk(s)->inflight))
226 maybe_unmark_and_push(s);
227 }
228
229 /*
230 * Mark phase
231 */
232
233 while (!empty_stack())
234 {
235 struct sock *x = pop_stack();
236 struct sock *sk;
237
238 spin_lock(&x->sk_receive_queue.lock);
239 skb = skb_peek(&x->sk_receive_queue);
240
241 /*
242 * Loop through all but first born
243 */
244
245 while (skb && skb != (struct sk_buff *)&x->sk_receive_queue) {
246 /*
247 * Do we have file descriptors ?
248 */
249 if(UNIXCB(skb).fp)
250 {
251 /*
252 * Process the descriptors of this socket
253 */
254 int nfd=UNIXCB(skb).fp->count;
255 struct file **fp = UNIXCB(skb).fp->fp;
256 while(nfd--)
257 {
258 /*
259 * Get the socket the fd matches if
260 * it indeed does so
261 */
262 if((sk=unix_get_socket(*fp++))!=NULL)
263 {
264 maybe_unmark_and_push(sk);
265 }
266 }
267 }
268 /* We have to scan not-yet-accepted ones too */
269 if (x->sk_state == TCP_LISTEN)
270 maybe_unmark_and_push(skb->sk);
271 skb=skb->next;
272 }
273 spin_unlock(&x->sk_receive_queue.lock);
274 sock_put(x);
275 }
276
277 skb_queue_head_init(&hitlist);
278
279 forall_unix_sockets(i, s)
280 {
281 struct unix_sock *u = unix_sk(s);
282
283 if (u->gc_tree == GC_ORPHAN) {
284 struct sk_buff *nextsk;
285
286 spin_lock(&s->sk_receive_queue.lock);
287 skb = skb_peek(&s->sk_receive_queue);
288 while (skb &&
289 skb != (struct sk_buff *)&s->sk_receive_queue) {
David S. Miller8728b832005-08-09 19:25:21 -0700290 nextsk = skb->next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 /*
292 * Do we have file descriptors ?
293 */
David S. Miller8728b832005-08-09 19:25:21 -0700294 if (UNIXCB(skb).fp) {
295 __skb_unlink(skb,
296 &s->sk_receive_queue);
297 __skb_queue_tail(&hitlist, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 }
David S. Miller8728b832005-08-09 19:25:21 -0700299 skb = nextsk;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 }
301 spin_unlock(&s->sk_receive_queue.lock);
302 }
303 u->gc_tree = GC_ORPHAN;
304 }
David S. Millerfbe9cc42005-12-13 23:26:29 -0800305 spin_unlock(&unix_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306
307 /*
308 * Here we are. Hitlist is filled. Die.
309 */
310
311 __skb_queue_purge(&hitlist);
Arjan van de Ven4a3e2f72006-03-20 22:33:17 -0800312 mutex_unlock(&unix_gc_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313}