blob: b6f4b994eb356a0cd9c05851d49fff16bece3eb0 [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 *
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -070065 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 */
YOSHIFUJI Hideakiac7bfa62007-02-09 23:25:23 +090070
Linus Torvalds1da177e2005-04-16 15:20:36 -070071#include <linux/kernel.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070072#include <linux/string.h>
73#include <linux/socket.h>
74#include <linux/un.h>
75#include <linux/net.h>
76#include <linux/fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077#include <linux/skbuff.h>
78#include <linux/netdevice.h>
79#include <linux/file.h>
80#include <linux/proc_fs.h>
Arjan van de Ven4a3e2f72006-03-20 22:33:17 -080081#include <linux/mutex.h>
dann frazier5f23b732008-11-26 15:32:27 -080082#include <linux/wait.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070083
84#include <net/sock.h>
85#include <net/af_unix.h>
86#include <net/scm.h>
Arnaldo Carvalho de Meloc752f072005-08-09 20:08:28 -070087#include <net/tcp_states.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89/* Internal data structures and random procedures: */
90
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -070091static LIST_HEAD(gc_inflight_list);
92static LIST_HEAD(gc_candidates);
93static DEFINE_SPINLOCK(unix_gc_lock);
dann frazier5f23b732008-11-26 15:32:27 -080094static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
Pavel Emelyanov9305cfa2007-11-10 22:06:01 -080096unsigned int unix_tot_inflight;
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
98
Eric Dumazet25888e32010-11-25 04:11:39 +000099struct sock *unix_get_socket(struct file *filp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100{
101 struct sock *u_sock = NULL;
Josef Sipek592ccbf2006-12-08 02:37:45 -0800102 struct inode *inode = filp->f_path.dentry->d_inode;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
104 /*
105 * Socket ?
106 */
Al Viro326be7b2011-03-13 17:08:22 -0400107 if (S_ISSOCK(inode->i_mode) && !(filp->f_mode & FMODE_PATH)) {
Jianjun Konge27dfce2008-11-01 21:38:31 -0700108 struct socket *sock = SOCKET_I(inode);
109 struct sock *s = sock->sk;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
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 */
YOSHIFUJI Hideakiac7bfa62007-02-09 23:25:23 +0900124
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125void unix_inflight(struct file *fp)
126{
127 struct sock *s = unix_get_socket(fp);
Jianjun Konge27dfce2008-11-01 21:38:31 -0700128 if (s) {
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700129 struct unix_sock *u = unix_sk(s);
130 spin_lock(&unix_gc_lock);
Al Viro516e0cc2008-07-26 00:39:17 -0400131 if (atomic_long_inc_return(&u->inflight) == 1) {
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700132 BUG_ON(!list_empty(&u->link));
133 list_add_tail(&u->link, &gc_inflight_list);
134 } else {
135 BUG_ON(list_empty(&u->link));
136 }
Pavel Emelyanov9305cfa2007-11-10 22:06:01 -0800137 unix_tot_inflight++;
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700138 spin_unlock(&unix_gc_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 }
140}
141
142void unix_notinflight(struct file *fp)
143{
144 struct sock *s = unix_get_socket(fp);
Jianjun Konge27dfce2008-11-01 21:38:31 -0700145 if (s) {
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700146 struct unix_sock *u = unix_sk(s);
147 spin_lock(&unix_gc_lock);
148 BUG_ON(list_empty(&u->link));
Al Viro516e0cc2008-07-26 00:39:17 -0400149 if (atomic_long_dec_and_test(&u->inflight))
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700150 list_del_init(&u->link);
Pavel Emelyanov9305cfa2007-11-10 22:06:01 -0800151 unix_tot_inflight--;
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700152 spin_unlock(&unix_gc_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 }
154}
155
Pavel Emelyanov5c80f1a2007-11-10 22:07:13 -0800156static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700157 struct sk_buff_head *hitlist)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158{
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700159 struct sk_buff *skb;
160 struct sk_buff *next;
161
162 spin_lock(&x->sk_receive_queue.lock);
Ilpo Järvinena2f3be12010-05-03 03:22:18 +0000163 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700164 /*
165 * Do we have file descriptors ?
166 */
167 if (UNIXCB(skb).fp) {
168 bool hit = false;
169 /*
170 * Process the descriptors of this socket
171 */
172 int nfd = UNIXCB(skb).fp->count;
173 struct file **fp = UNIXCB(skb).fp->fp;
174 while (nfd--) {
175 /*
176 * Get the socket the fd matches
177 * if it indeed does so
178 */
179 struct sock *sk = unix_get_socket(*fp++);
Pavel Emelyanov5c80f1a2007-11-10 22:07:13 -0800180 if (sk) {
Miklos Szeredi62093442008-11-09 15:23:57 +0100181 struct unix_sock *u = unix_sk(sk);
182
183 /*
184 * Ignore non-candidates, they could
185 * have been added to the queues after
186 * starting the garbage collection
187 */
188 if (u->gc_candidate) {
189 hit = true;
190 func(u);
191 }
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700192 }
193 }
194 if (hit && hitlist != NULL) {
195 __skb_unlink(skb, &x->sk_receive_queue);
196 __skb_queue_tail(hitlist, skb);
197 }
198 }
199 }
200 spin_unlock(&x->sk_receive_queue.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201}
202
Pavel Emelyanov5c80f1a2007-11-10 22:07:13 -0800203static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700204 struct sk_buff_head *hitlist)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205{
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700206 if (x->sk_state != TCP_LISTEN)
207 scan_inflight(x, func, hitlist);
208 else {
209 struct sk_buff *skb;
210 struct sk_buff *next;
211 struct unix_sock *u;
212 LIST_HEAD(embryos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700214 /*
215 * For a listening socket collect the queued embryos
216 * and perform a scan on them as well.
217 */
218 spin_lock(&x->sk_receive_queue.lock);
Ilpo Järvinena2f3be12010-05-03 03:22:18 +0000219 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700220 u = unix_sk(skb->sk);
221
222 /*
223 * An embryo cannot be in-flight, so it's safe
224 * to use the list link.
225 */
226 BUG_ON(!list_empty(&u->link));
227 list_add_tail(&u->link, &embryos);
228 }
229 spin_unlock(&x->sk_receive_queue.lock);
230
231 while (!list_empty(&embryos)) {
232 u = list_entry(embryos.next, struct unix_sock, link);
233 scan_inflight(&u->sk, func, hitlist);
234 list_del_init(&u->link);
235 }
236 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237}
238
Pavel Emelyanov5c80f1a2007-11-10 22:07:13 -0800239static void dec_inflight(struct unix_sock *usk)
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700240{
Al Viro516e0cc2008-07-26 00:39:17 -0400241 atomic_long_dec(&usk->inflight);
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700242}
243
Pavel Emelyanov5c80f1a2007-11-10 22:07:13 -0800244static void inc_inflight(struct unix_sock *usk)
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700245{
Al Viro516e0cc2008-07-26 00:39:17 -0400246 atomic_long_inc(&usk->inflight);
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700247}
248
Pavel Emelyanov5c80f1a2007-11-10 22:07:13 -0800249static void inc_inflight_move_tail(struct unix_sock *u)
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700250{
Al Viro516e0cc2008-07-26 00:39:17 -0400251 atomic_long_inc(&u->inflight);
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700252 /*
Miklos Szeredi62093442008-11-09 15:23:57 +0100253 * If this still might be part of a cycle, move it to the end
254 * of the list, so that it's checked even if it was already
255 * passed over
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700256 */
Miklos Szeredi62093442008-11-09 15:23:57 +0100257 if (u->gc_maybe_cycle)
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700258 list_move_tail(&u->link, &gc_candidates);
259}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
dann frazier5f23b732008-11-26 15:32:27 -0800261static bool gc_in_progress = false;
Eric Dumazet99156722010-11-24 09:15:27 -0800262#define UNIX_INFLIGHT_TRIGGER_GC 16000
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
dann frazier5f23b732008-11-26 15:32:27 -0800264void wait_for_unix_gc(void)
265{
Eric Dumazet99156722010-11-24 09:15:27 -0800266 /*
267 * If number of inflight sockets is insane,
268 * force a garbage collect right now.
269 */
270 if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
271 unix_gc();
dann frazier5f23b732008-11-26 15:32:27 -0800272 wait_event(unix_gc_wait, gc_in_progress == false);
273}
274
275/* The external entry point: unix_gc() */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276void unix_gc(void)
277{
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700278 struct unix_sock *u;
279 struct unix_sock *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 struct sk_buff_head hitlist;
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700281 struct list_head cursor;
Miklos Szeredi62093442008-11-09 15:23:57 +0100282 LIST_HEAD(not_cycle_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700284 spin_lock(&unix_gc_lock);
285
286 /* Avoid a recursive GC. */
287 if (gc_in_progress)
288 goto out;
289
290 gc_in_progress = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 /*
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700292 * First, select candidates for garbage collection. Only
293 * in-flight sockets are considered, and from those only ones
294 * which don't have any external reference.
295 *
296 * Holding unix_gc_lock will protect these candidates from
297 * being detached, and hence from gaining an external
Miklos Szeredi62093442008-11-09 15:23:57 +0100298 * reference. Since there are no possible receivers, all
299 * buffers currently on the candidates' queues stay there
300 * during the garbage collection.
301 *
302 * We also know that no new candidate can be added onto the
303 * receive queues. Other, non candidate sockets _can_ be
304 * added to queue, so we must make sure only to touch
305 * candidates.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 */
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700307 list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
Al Viro516e0cc2008-07-26 00:39:17 -0400308 long total_refs;
309 long inflight_refs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700311 total_refs = file_count(u->sk.sk_socket->file);
Al Viro516e0cc2008-07-26 00:39:17 -0400312 inflight_refs = atomic_long_read(&u->inflight);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700314 BUG_ON(inflight_refs < 1);
315 BUG_ON(total_refs < inflight_refs);
316 if (total_refs == inflight_refs) {
317 list_move_tail(&u->link, &gc_candidates);
318 u->gc_candidate = 1;
Miklos Szeredi62093442008-11-09 15:23:57 +0100319 u->gc_maybe_cycle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 }
322
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700323 /*
324 * Now remove all internal in-flight reference to children of
325 * the candidates.
326 */
327 list_for_each_entry(u, &gc_candidates, link)
328 scan_children(&u->sk, dec_inflight, NULL);
329
330 /*
331 * Restore the references for children of all candidates,
332 * which have remaining references. Do this recursively, so
333 * only those remain, which form cyclic references.
334 *
335 * Use a "cursor" link, to make the list traversal safe, even
336 * though elements might be moved about.
337 */
338 list_add(&cursor, &gc_candidates);
339 while (cursor.next != &gc_candidates) {
340 u = list_entry(cursor.next, struct unix_sock, link);
341
342 /* Move cursor to after the current position. */
343 list_move(&cursor, &u->link);
344
Al Viro516e0cc2008-07-26 00:39:17 -0400345 if (atomic_long_read(&u->inflight) > 0) {
Miklos Szeredi62093442008-11-09 15:23:57 +0100346 list_move_tail(&u->link, &not_cycle_list);
347 u->gc_maybe_cycle = 0;
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700348 scan_children(&u->sk, inc_inflight_move_tail, NULL);
349 }
350 }
351 list_del(&cursor);
352
353 /*
Miklos Szeredi62093442008-11-09 15:23:57 +0100354 * not_cycle_list contains those sockets which do not make up a
355 * cycle. Restore these to the inflight list.
356 */
357 while (!list_empty(&not_cycle_list)) {
358 u = list_entry(not_cycle_list.next, struct unix_sock, link);
359 u->gc_candidate = 0;
360 list_move_tail(&u->link, &gc_inflight_list);
361 }
362
363 /*
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700364 * Now gc_candidates contains only garbage. Restore original
365 * inflight counters for these as well, and remove the skbuffs
366 * which are creating the cycle(s).
367 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 skb_queue_head_init(&hitlist);
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700369 list_for_each_entry(u, &gc_candidates, link)
Jianjun Konge27dfce2008-11-01 21:38:31 -0700370 scan_children(&u->sk, inc_inflight, &hitlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700372 spin_unlock(&unix_gc_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700374 /* Here we are. Hitlist is filled. Die. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 __skb_queue_purge(&hitlist);
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700376
377 spin_lock(&unix_gc_lock);
378
379 /* All candidates should have been detached by now. */
380 BUG_ON(!list_empty(&gc_candidates));
381 gc_in_progress = false;
dann frazier5f23b732008-11-26 15:32:27 -0800382 wake_up(&unix_gc_wait);
Miklos Szeredi1fd05ba2007-07-11 14:22:39 -0700383
384 out:
385 spin_unlock(&unix_gc_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}