Commit | Line | Data |
---|---|---|
1da177e4 LT |
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 | * | |
1fd05ba5 MS |
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. | |
1da177e4 | 69 | */ |
ac7bfa62 | 70 | |
1da177e4 | 71 | #include <linux/kernel.h> |
1da177e4 LT |
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/slab.h> | |
78 | #include <linux/skbuff.h> | |
79 | #include <linux/netdevice.h> | |
80 | #include <linux/file.h> | |
81 | #include <linux/proc_fs.h> | |
4a3e2f71 | 82 | #include <linux/mutex.h> |
5f23b734 | 83 | #include <linux/wait.h> |
1da177e4 LT |
84 | |
85 | #include <net/sock.h> | |
86 | #include <net/af_unix.h> | |
87 | #include <net/scm.h> | |
c752f073 | 88 | #include <net/tcp_states.h> |
1da177e4 LT |
89 | |
90 | /* Internal data structures and random procedures: */ | |
91 | ||
1fd05ba5 MS |
92 | static LIST_HEAD(gc_inflight_list); |
93 | static LIST_HEAD(gc_candidates); | |
94 | static DEFINE_SPINLOCK(unix_gc_lock); | |
5f23b734 | 95 | static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait); |
1da177e4 | 96 | |
9305cfa4 | 97 | unsigned int unix_tot_inflight; |
1da177e4 LT |
98 | |
99 | ||
100 | static struct sock *unix_get_socket(struct file *filp) | |
101 | { | |
102 | struct sock *u_sock = NULL; | |
592ccbf9 | 103 | struct inode *inode = filp->f_path.dentry->d_inode; |
1da177e4 LT |
104 | |
105 | /* | |
106 | * Socket ? | |
107 | */ | |
108 | if (S_ISSOCK(inode->i_mode)) { | |
e27dfcea JK |
109 | struct socket *sock = SOCKET_I(inode); |
110 | struct sock *s = sock->sk; | |
1da177e4 LT |
111 | |
112 | /* | |
113 | * PF_UNIX ? | |
114 | */ | |
115 | if (s && sock->ops && sock->ops->family == PF_UNIX) | |
116 | u_sock = s; | |
117 | } | |
118 | return u_sock; | |
119 | } | |
120 | ||
121 | /* | |
122 | * Keep the number of times in flight count for the file | |
123 | * descriptor if it is for an AF_UNIX socket. | |
124 | */ | |
ac7bfa62 | 125 | |
1da177e4 LT |
126 | void unix_inflight(struct file *fp) |
127 | { | |
128 | struct sock *s = unix_get_socket(fp); | |
e27dfcea | 129 | if (s) { |
1fd05ba5 MS |
130 | struct unix_sock *u = unix_sk(s); |
131 | spin_lock(&unix_gc_lock); | |
516e0cc5 | 132 | if (atomic_long_inc_return(&u->inflight) == 1) { |
1fd05ba5 MS |
133 | BUG_ON(!list_empty(&u->link)); |
134 | list_add_tail(&u->link, &gc_inflight_list); | |
135 | } else { | |
136 | BUG_ON(list_empty(&u->link)); | |
137 | } | |
9305cfa4 | 138 | unix_tot_inflight++; |
1fd05ba5 | 139 | spin_unlock(&unix_gc_lock); |
1da177e4 LT |
140 | } |
141 | } | |
142 | ||
143 | void unix_notinflight(struct file *fp) | |
144 | { | |
145 | struct sock *s = unix_get_socket(fp); | |
e27dfcea | 146 | if (s) { |
1fd05ba5 MS |
147 | struct unix_sock *u = unix_sk(s); |
148 | spin_lock(&unix_gc_lock); | |
149 | BUG_ON(list_empty(&u->link)); | |
516e0cc5 | 150 | if (atomic_long_dec_and_test(&u->inflight)) |
1fd05ba5 | 151 | list_del_init(&u->link); |
9305cfa4 | 152 | unix_tot_inflight--; |
1fd05ba5 | 153 | spin_unlock(&unix_gc_lock); |
1da177e4 LT |
154 | } |
155 | } | |
156 | ||
1fd05ba5 MS |
157 | static inline struct sk_buff *sock_queue_head(struct sock *sk) |
158 | { | |
e27dfcea | 159 | return (struct sk_buff *)&sk->sk_receive_queue; |
1fd05ba5 | 160 | } |
1da177e4 | 161 | |
1fd05ba5 MS |
162 | #define receive_queue_for_each_skb(sk, next, skb) \ |
163 | for (skb = sock_queue_head(sk)->next, next = skb->next; \ | |
164 | skb != sock_queue_head(sk); skb = next, next = skb->next) | |
1da177e4 | 165 | |
5c80f1ae | 166 | static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *), |
1fd05ba5 | 167 | struct sk_buff_head *hitlist) |
1da177e4 | 168 | { |
1fd05ba5 MS |
169 | struct sk_buff *skb; |
170 | struct sk_buff *next; | |
171 | ||
172 | spin_lock(&x->sk_receive_queue.lock); | |
173 | receive_queue_for_each_skb(x, next, skb) { | |
174 | /* | |
175 | * Do we have file descriptors ? | |
176 | */ | |
177 | if (UNIXCB(skb).fp) { | |
178 | bool hit = false; | |
179 | /* | |
180 | * Process the descriptors of this socket | |
181 | */ | |
182 | int nfd = UNIXCB(skb).fp->count; | |
183 | struct file **fp = UNIXCB(skb).fp->fp; | |
184 | while (nfd--) { | |
185 | /* | |
186 | * Get the socket the fd matches | |
187 | * if it indeed does so | |
188 | */ | |
189 | struct sock *sk = unix_get_socket(*fp++); | |
5c80f1ae | 190 | if (sk) { |
6209344f MS |
191 | struct unix_sock *u = unix_sk(sk); |
192 | ||
193 | /* | |
194 | * Ignore non-candidates, they could | |
195 | * have been added to the queues after | |
196 | * starting the garbage collection | |
197 | */ | |
198 | if (u->gc_candidate) { | |
199 | hit = true; | |
200 | func(u); | |
201 | } | |
1fd05ba5 MS |
202 | } |
203 | } | |
204 | if (hit && hitlist != NULL) { | |
205 | __skb_unlink(skb, &x->sk_receive_queue); | |
206 | __skb_queue_tail(hitlist, skb); | |
207 | } | |
208 | } | |
209 | } | |
210 | spin_unlock(&x->sk_receive_queue.lock); | |
1da177e4 LT |
211 | } |
212 | ||
5c80f1ae | 213 | static void scan_children(struct sock *x, void (*func)(struct unix_sock *), |
1fd05ba5 | 214 | struct sk_buff_head *hitlist) |
1da177e4 | 215 | { |
1fd05ba5 MS |
216 | if (x->sk_state != TCP_LISTEN) |
217 | scan_inflight(x, func, hitlist); | |
218 | else { | |
219 | struct sk_buff *skb; | |
220 | struct sk_buff *next; | |
221 | struct unix_sock *u; | |
222 | LIST_HEAD(embryos); | |
223 | ||
224 | /* | |
225 | * For a listening socket collect the queued embryos | |
226 | * and perform a scan on them as well. | |
227 | */ | |
228 | spin_lock(&x->sk_receive_queue.lock); | |
229 | receive_queue_for_each_skb(x, next, skb) { | |
230 | u = unix_sk(skb->sk); | |
231 | ||
232 | /* | |
233 | * An embryo cannot be in-flight, so it's safe | |
234 | * to use the list link. | |
235 | */ | |
236 | BUG_ON(!list_empty(&u->link)); | |
237 | list_add_tail(&u->link, &embryos); | |
238 | } | |
239 | spin_unlock(&x->sk_receive_queue.lock); | |
240 | ||
241 | while (!list_empty(&embryos)) { | |
242 | u = list_entry(embryos.next, struct unix_sock, link); | |
243 | scan_inflight(&u->sk, func, hitlist); | |
244 | list_del_init(&u->link); | |
245 | } | |
246 | } | |
1da177e4 LT |
247 | } |
248 | ||
5c80f1ae | 249 | static void dec_inflight(struct unix_sock *usk) |
1da177e4 | 250 | { |
516e0cc5 | 251 | atomic_long_dec(&usk->inflight); |
1fd05ba5 | 252 | } |
1da177e4 | 253 | |
5c80f1ae | 254 | static void inc_inflight(struct unix_sock *usk) |
1fd05ba5 | 255 | { |
516e0cc5 | 256 | atomic_long_inc(&usk->inflight); |
1da177e4 LT |
257 | } |
258 | ||
5c80f1ae | 259 | static void inc_inflight_move_tail(struct unix_sock *u) |
1fd05ba5 | 260 | { |
516e0cc5 | 261 | atomic_long_inc(&u->inflight); |
1fd05ba5 | 262 | /* |
6209344f MS |
263 | * If this still might be part of a cycle, move it to the end |
264 | * of the list, so that it's checked even if it was already | |
265 | * passed over | |
1fd05ba5 | 266 | */ |
6209344f | 267 | if (u->gc_maybe_cycle) |
1fd05ba5 MS |
268 | list_move_tail(&u->link, &gc_candidates); |
269 | } | |
1da177e4 | 270 | |
5f23b734 | 271 | static bool gc_in_progress = false; |
1da177e4 | 272 | |
5f23b734 | 273 | void wait_for_unix_gc(void) |
1da177e4 | 274 | { |
5f23b734 | 275 | wait_event(unix_gc_wait, gc_in_progress == false); |
276 | } | |
1da177e4 | 277 | |
5f23b734 | 278 | /* The external entry point: unix_gc() */ |
279 | void unix_gc(void) | |
280 | { | |
1fd05ba5 MS |
281 | struct unix_sock *u; |
282 | struct unix_sock *next; | |
283 | struct sk_buff_head hitlist; | |
284 | struct list_head cursor; | |
6209344f | 285 | LIST_HEAD(not_cycle_list); |
1da177e4 | 286 | |
1fd05ba5 | 287 | spin_lock(&unix_gc_lock); |
1da177e4 | 288 | |
1fd05ba5 MS |
289 | /* Avoid a recursive GC. */ |
290 | if (gc_in_progress) | |
291 | goto out; | |
1da177e4 | 292 | |
1fd05ba5 | 293 | gc_in_progress = true; |
1da177e4 | 294 | /* |
1fd05ba5 MS |
295 | * First, select candidates for garbage collection. Only |
296 | * in-flight sockets are considered, and from those only ones | |
297 | * which don't have any external reference. | |
298 | * | |
299 | * Holding unix_gc_lock will protect these candidates from | |
300 | * being detached, and hence from gaining an external | |
6209344f MS |
301 | * reference. Since there are no possible receivers, all |
302 | * buffers currently on the candidates' queues stay there | |
303 | * during the garbage collection. | |
304 | * | |
305 | * We also know that no new candidate can be added onto the | |
306 | * receive queues. Other, non candidate sockets _can_ be | |
307 | * added to queue, so we must make sure only to touch | |
308 | * candidates. | |
1da177e4 | 309 | */ |
1fd05ba5 | 310 | list_for_each_entry_safe(u, next, &gc_inflight_list, link) { |
516e0cc5 AV |
311 | long total_refs; |
312 | long inflight_refs; | |
1fd05ba5 MS |
313 | |
314 | total_refs = file_count(u->sk.sk_socket->file); | |
516e0cc5 | 315 | inflight_refs = atomic_long_read(&u->inflight); |
1fd05ba5 MS |
316 | |
317 | BUG_ON(inflight_refs < 1); | |
318 | BUG_ON(total_refs < inflight_refs); | |
319 | if (total_refs == inflight_refs) { | |
320 | list_move_tail(&u->link, &gc_candidates); | |
321 | u->gc_candidate = 1; | |
6209344f | 322 | u->gc_maybe_cycle = 1; |
1fd05ba5 MS |
323 | } |
324 | } | |
1da177e4 LT |
325 | |
326 | /* | |
1fd05ba5 MS |
327 | * Now remove all internal in-flight reference to children of |
328 | * the candidates. | |
1da177e4 | 329 | */ |
1fd05ba5 MS |
330 | list_for_each_entry(u, &gc_candidates, link) |
331 | scan_children(&u->sk, dec_inflight, NULL); | |
1da177e4 LT |
332 | |
333 | /* | |
1fd05ba5 MS |
334 | * Restore the references for children of all candidates, |
335 | * which have remaining references. Do this recursively, so | |
336 | * only those remain, which form cyclic references. | |
337 | * | |
338 | * Use a "cursor" link, to make the list traversal safe, even | |
339 | * though elements might be moved about. | |
1da177e4 | 340 | */ |
1fd05ba5 MS |
341 | list_add(&cursor, &gc_candidates); |
342 | while (cursor.next != &gc_candidates) { | |
343 | u = list_entry(cursor.next, struct unix_sock, link); | |
1da177e4 | 344 | |
1fd05ba5 MS |
345 | /* Move cursor to after the current position. */ |
346 | list_move(&cursor, &u->link); | |
ac7bfa62 | 347 | |
516e0cc5 | 348 | if (atomic_long_read(&u->inflight) > 0) { |
6209344f MS |
349 | list_move_tail(&u->link, ¬_cycle_list); |
350 | u->gc_maybe_cycle = 0; | |
1fd05ba5 | 351 | scan_children(&u->sk, inc_inflight_move_tail, NULL); |
1da177e4 | 352 | } |
1da177e4 | 353 | } |
1fd05ba5 | 354 | list_del(&cursor); |
1da177e4 | 355 | |
6209344f MS |
356 | /* |
357 | * not_cycle_list contains those sockets which do not make up a | |
358 | * cycle. Restore these to the inflight list. | |
359 | */ | |
360 | while (!list_empty(¬_cycle_list)) { | |
361 | u = list_entry(not_cycle_list.next, struct unix_sock, link); | |
362 | u->gc_candidate = 0; | |
363 | list_move_tail(&u->link, &gc_inflight_list); | |
364 | } | |
365 | ||
1fd05ba5 MS |
366 | /* |
367 | * Now gc_candidates contains only garbage. Restore original | |
368 | * inflight counters for these as well, and remove the skbuffs | |
369 | * which are creating the cycle(s). | |
370 | */ | |
1da177e4 | 371 | skb_queue_head_init(&hitlist); |
1fd05ba5 | 372 | list_for_each_entry(u, &gc_candidates, link) |
e27dfcea | 373 | scan_children(&u->sk, inc_inflight, &hitlist); |
1da177e4 | 374 | |
1fd05ba5 | 375 | spin_unlock(&unix_gc_lock); |
1da177e4 | 376 | |
1fd05ba5 MS |
377 | /* Here we are. Hitlist is filled. Die. */ |
378 | __skb_queue_purge(&hitlist); | |
1da177e4 | 379 | |
1fd05ba5 | 380 | spin_lock(&unix_gc_lock); |
1da177e4 | 381 | |
1fd05ba5 MS |
382 | /* All candidates should have been detached by now. */ |
383 | BUG_ON(!list_empty(&gc_candidates)); | |
384 | gc_in_progress = false; | |
5f23b734 | 385 | wake_up(&unix_gc_wait); |
1da177e4 | 386 | |
1fd05ba5 MS |
387 | out: |
388 | spin_unlock(&unix_gc_lock); | |
1da177e4 | 389 | } |