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/skbuff.h>
78 #include <linux/netdevice.h>
79 #include <linux/file.h>
80 #include <linux/proc_fs.h>
81 #include <linux/mutex.h>
82 #include <linux/wait.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);
94 static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
96 unsigned int unix_tot_inflight;
99 struct sock *unix_get_socket(struct file *filp)
101 struct sock *u_sock = NULL;
102 struct inode *inode = filp->f_path.dentry->d_inode;
107 if (S_ISSOCK(inode->i_mode) && !(filp->f_mode & FMODE_PATH)) {
108 struct socket *sock = SOCKET_I(inode);
109 struct sock *s = sock->sk;
114 if (s && sock->ops && sock->ops->family == PF_UNIX)
121 * Keep the number of times in flight count for the file
122 * descriptor if it is for an AF_UNIX socket.
125 void unix_inflight(struct user_struct *user, struct file *fp)
127 struct sock *s = unix_get_socket(fp);
129 spin_lock(&unix_gc_lock);
132 struct unix_sock *u = unix_sk(s);
133 if (atomic_long_inc_return(&u->inflight) == 1) {
134 BUG_ON(!list_empty(&u->link));
135 list_add_tail(&u->link, &gc_inflight_list);
137 BUG_ON(list_empty(&u->link));
141 user->unix_inflight++;
142 spin_unlock(&unix_gc_lock);
145 void unix_notinflight(struct user_struct *user, struct file *fp)
147 struct sock *s = unix_get_socket(fp);
149 spin_lock(&unix_gc_lock);
152 struct unix_sock *u = unix_sk(s);
153 BUG_ON(list_empty(&u->link));
154 if (atomic_long_dec_and_test(&u->inflight))
155 list_del_init(&u->link);
158 user->unix_inflight--;
159 spin_unlock(&unix_gc_lock);
162 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
163 struct sk_buff_head *hitlist)
166 struct sk_buff *next;
168 spin_lock(&x->sk_receive_queue.lock);
169 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
171 * Do we have file descriptors ?
173 if (UNIXCB(skb).fp) {
176 * Process the descriptors of this socket
178 int nfd = UNIXCB(skb).fp->count;
179 struct file **fp = UNIXCB(skb).fp->fp;
182 * Get the socket the fd matches
183 * if it indeed does so
185 struct sock *sk = unix_get_socket(*fp++);
187 struct unix_sock *u = unix_sk(sk);
190 * Ignore non-candidates, they could
191 * have been added to the queues after
192 * starting the garbage collection
194 if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
200 if (hit && hitlist != NULL) {
201 __skb_unlink(skb, &x->sk_receive_queue);
202 __skb_queue_tail(hitlist, skb);
206 spin_unlock(&x->sk_receive_queue.lock);
209 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
210 struct sk_buff_head *hitlist)
212 if (x->sk_state != TCP_LISTEN)
213 scan_inflight(x, func, hitlist);
216 struct sk_buff *next;
221 * For a listening socket collect the queued embryos
222 * and perform a scan on them as well.
224 spin_lock(&x->sk_receive_queue.lock);
225 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
226 u = unix_sk(skb->sk);
229 * An embryo cannot be in-flight, so it's safe
230 * to use the list link.
232 BUG_ON(!list_empty(&u->link));
233 list_add_tail(&u->link, &embryos);
235 spin_unlock(&x->sk_receive_queue.lock);
237 while (!list_empty(&embryos)) {
238 u = list_entry(embryos.next, struct unix_sock, link);
239 scan_inflight(&u->sk, func, hitlist);
240 list_del_init(&u->link);
245 static void dec_inflight(struct unix_sock *usk)
247 atomic_long_dec(&usk->inflight);
250 static void inc_inflight(struct unix_sock *usk)
252 atomic_long_inc(&usk->inflight);
255 static void inc_inflight_move_tail(struct unix_sock *u)
257 atomic_long_inc(&u->inflight);
259 * If this still might be part of a cycle, move it to the end
260 * of the list, so that it's checked even if it was already
263 if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
264 list_move_tail(&u->link, &gc_candidates);
267 static bool gc_in_progress = false;
268 #define UNIX_INFLIGHT_TRIGGER_GC 16000
270 void wait_for_unix_gc(void)
273 * If number of inflight sockets is insane,
274 * force a garbage collect right now.
276 if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
278 wait_event(unix_gc_wait, gc_in_progress == false);
281 /* The external entry point: unix_gc() */
285 struct unix_sock *next;
286 struct sk_buff_head hitlist;
287 struct list_head cursor;
288 LIST_HEAD(not_cycle_list);
290 spin_lock(&unix_gc_lock);
292 /* Avoid a recursive GC. */
296 gc_in_progress = true;
298 * First, select candidates for garbage collection. Only
299 * in-flight sockets are considered, and from those only ones
300 * which don't have any external reference.
302 * Holding unix_gc_lock will protect these candidates from
303 * being detached, and hence from gaining an external
304 * reference. Since there are no possible receivers, all
305 * buffers currently on the candidates' queues stay there
306 * during the garbage collection.
308 * We also know that no new candidate can be added onto the
309 * receive queues. Other, non candidate sockets _can_ be
310 * added to queue, so we must make sure only to touch
313 list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
317 total_refs = file_count(u->sk.sk_socket->file);
318 inflight_refs = atomic_long_read(&u->inflight);
320 BUG_ON(inflight_refs < 1);
321 BUG_ON(total_refs < inflight_refs);
322 if (total_refs == inflight_refs) {
323 list_move_tail(&u->link, &gc_candidates);
324 __set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
325 __set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
330 * Now remove all internal in-flight reference to children of
333 list_for_each_entry(u, &gc_candidates, link)
334 scan_children(&u->sk, dec_inflight, NULL);
337 * Restore the references for children of all candidates,
338 * which have remaining references. Do this recursively, so
339 * only those remain, which form cyclic references.
341 * Use a "cursor" link, to make the list traversal safe, even
342 * though elements might be moved about.
344 list_add(&cursor, &gc_candidates);
345 while (cursor.next != &gc_candidates) {
346 u = list_entry(cursor.next, struct unix_sock, link);
348 /* Move cursor to after the current position. */
349 list_move(&cursor, &u->link);
351 if (atomic_long_read(&u->inflight) > 0) {
352 list_move_tail(&u->link, ¬_cycle_list);
353 __clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
354 scan_children(&u->sk, inc_inflight_move_tail, NULL);
360 * not_cycle_list contains those sockets which do not make up a
361 * cycle. Restore these to the inflight list.
363 while (!list_empty(¬_cycle_list)) {
364 u = list_entry(not_cycle_list.next, struct unix_sock, link);
365 __clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
366 list_move_tail(&u->link, &gc_inflight_list);
370 * Now gc_candidates contains only garbage. Restore original
371 * inflight counters for these as well, and remove the skbuffs
372 * which are creating the cycle(s).
374 skb_queue_head_init(&hitlist);
375 list_for_each_entry(u, &gc_candidates, link)
376 scan_children(&u->sk, inc_inflight, &hitlist);
378 spin_unlock(&unix_gc_lock);
380 /* Here we are. Hitlist is filled. Die. */
381 __skb_queue_purge(&hitlist);
383 spin_lock(&unix_gc_lock);
385 /* All candidates should have been detached by now. */
386 BUG_ON(!list_empty(&gc_candidates));
387 gc_in_progress = false;
388 wake_up(&unix_gc_wait);
391 spin_unlock(&unix_gc_lock);