1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * NET3: Garbage Collector For AF_UNIX sockets
6 * Copyright (C) Barak A. Pearlmutter.
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
28 * Alan Cox 07 Sept 1997 Vmalloc internal stack as needed.
29 * Cope with changing max_files.
31 * Graph may have cycles. That is, we can send the descriptor
32 * of foo to bar and vice versa. Current code chokes on that.
33 * Fix: move SCM_RIGHTS ones into the separate list and then
34 * skb_free() them all instead of doing explicit fput's.
35 * Another problem: since fput() may block somebody may
36 * create a new unix_socket when we are in the middle of sweep
37 * phase. Fix: revert the logic wrt MARKED. Mark everything
38 * upon the beginning and unmark non-junk ones.
40 * [12 Oct 1998] AAARGH! New code purges all SCM_RIGHTS
41 * sent to connect()'ed but still not accept()'ed sockets.
42 * Fixed. Old code had slightly different problem here:
43 * extra fput() in situation when we passed the descriptor via
44 * such socket and closed it (descriptor). That would happen on
45 * each unix_gc() until the accept(). Since the struct file in
46 * question would go to the free list and might be reused...
47 * That might be the reason of random oopses on filp_close()
48 * in unrelated processes.
51 * Kill the explicit allocation of stack. Now we keep the tree
52 * with root in dummy + pointer (gc_current) to one of the nodes.
53 * Stack is represented as path from gc_current to dummy. Unmark
54 * now means "add to tree". Push == "make it a son of gc_current".
55 * Pop == "move gc_current to parent". We keep only pointers to
56 * parents (->gc_tree).
58 * Damn. Added missing check for ->dead in listen queues scanning.
60 * Miklos Szeredi 25 Jun 2007
61 * Reimplement with a cycle collecting algorithm. This should
62 * solve several problems with the previous code, like being racy
63 * wrt receive and holding up unrelated socket operations.
66 #include <linux/kernel.h>
67 #include <linux/string.h>
68 #include <linux/socket.h>
70 #include <linux/net.h>
72 #include <linux/skbuff.h>
73 #include <linux/netdevice.h>
74 #include <linux/file.h>
75 #include <linux/proc_fs.h>
76 #include <linux/mutex.h>
77 #include <linux/wait.h>
80 #include <net/af_unix.h>
82 #include <net/tcp_states.h>
86 /* Internal data structures and random procedures: */
88 static LIST_HEAD(gc_candidates);
89 static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
91 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
92 struct sk_buff_head *hitlist)
97 spin_lock(&x->sk_receive_queue.lock);
98 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
99 /* Do we have file descriptors ? */
100 if (UNIXCB(skb).fp) {
102 /* Process the descriptors of this socket */
103 int nfd = UNIXCB(skb).fp->count;
104 struct file **fp = UNIXCB(skb).fp->fp;
107 /* Get the socket the fd matches if it indeed does so */
108 struct sock *sk = unix_get_socket(*fp++);
111 struct unix_sock *u = unix_sk(sk);
113 /* Ignore non-candidates, they could
114 * have been added to the queues after
115 * starting the garbage collection
117 if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
124 if (hit && hitlist != NULL) {
125 __skb_unlink(skb, &x->sk_receive_queue);
126 __skb_queue_tail(hitlist, skb);
130 spin_unlock(&x->sk_receive_queue.lock);
133 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
134 struct sk_buff_head *hitlist)
136 if (x->sk_state != TCP_LISTEN) {
137 scan_inflight(x, func, hitlist);
140 struct sk_buff *next;
144 /* For a listening socket collect the queued embryos
145 * and perform a scan on them as well.
147 spin_lock(&x->sk_receive_queue.lock);
148 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
149 u = unix_sk(skb->sk);
151 /* An embryo cannot be in-flight, so it's safe
152 * to use the list link.
154 BUG_ON(!list_empty(&u->link));
155 list_add_tail(&u->link, &embryos);
157 spin_unlock(&x->sk_receive_queue.lock);
159 while (!list_empty(&embryos)) {
160 u = list_entry(embryos.next, struct unix_sock, link);
161 scan_inflight(&u->sk, func, hitlist);
162 list_del_init(&u->link);
167 static void dec_inflight(struct unix_sock *usk)
172 static void inc_inflight(struct unix_sock *usk)
177 static void inc_inflight_move_tail(struct unix_sock *u)
181 /* If this still might be part of a cycle, move it to the end
182 * of the list, so that it's checked even if it was already
185 if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
186 list_move_tail(&u->link, &gc_candidates);
189 static bool gc_in_progress;
190 #define UNIX_INFLIGHT_TRIGGER_GC 16000
192 void wait_for_unix_gc(void)
194 /* If number of inflight sockets is insane,
195 * force a garbage collect right now.
196 * Paired with the WRITE_ONCE() in unix_inflight(),
197 * unix_notinflight() and gc_in_progress().
199 if (READ_ONCE(unix_tot_inflight) > UNIX_INFLIGHT_TRIGGER_GC &&
200 !READ_ONCE(gc_in_progress))
202 wait_event(unix_gc_wait, !READ_ONCE(gc_in_progress));
205 /* The external entry point: unix_gc() */
208 struct sk_buff *next_skb, *skb;
210 struct unix_sock *next;
211 struct sk_buff_head hitlist;
212 struct list_head cursor;
213 LIST_HEAD(not_cycle_list);
215 spin_lock(&unix_gc_lock);
217 /* Avoid a recursive GC. */
221 /* Paired with READ_ONCE() in wait_for_unix_gc(). */
222 WRITE_ONCE(gc_in_progress, true);
224 /* First, select candidates for garbage collection. Only
225 * in-flight sockets are considered, and from those only ones
226 * which don't have any external reference.
228 * Holding unix_gc_lock will protect these candidates from
229 * being detached, and hence from gaining an external
230 * reference. Since there are no possible receivers, all
231 * buffers currently on the candidates' queues stay there
232 * during the garbage collection.
234 * We also know that no new candidate can be added onto the
235 * receive queues. Other, non candidate sockets _can_ be
236 * added to queue, so we must make sure only to touch
239 * Embryos, though never candidates themselves, affect which
240 * candidates are reachable by the garbage collector. Before
241 * being added to a listener's queue, an embryo may already
242 * receive data carrying SCM_RIGHTS, potentially making the
243 * passed socket a candidate that is not yet reachable by the
244 * collector. It becomes reachable once the embryo is
245 * enqueued. Therefore, we must ensure that no SCM-laden
246 * embryo appears in a (candidate) listener's queue between
247 * consecutive scan_children() calls.
249 list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
250 struct sock *sk = &u->sk;
253 total_refs = file_count(sk->sk_socket->file);
255 BUG_ON(!u->inflight);
256 BUG_ON(total_refs < u->inflight);
257 if (total_refs == u->inflight) {
258 list_move_tail(&u->link, &gc_candidates);
259 __set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
260 __set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
262 if (sk->sk_state == TCP_LISTEN) {
263 unix_state_lock_nested(sk, U_LOCK_GC_LISTENER);
264 unix_state_unlock(sk);
269 /* Now remove all internal in-flight reference to children of
272 list_for_each_entry(u, &gc_candidates, link)
273 scan_children(&u->sk, dec_inflight, NULL);
275 /* Restore the references for children of all candidates,
276 * which have remaining references. Do this recursively, so
277 * only those remain, which form cyclic references.
279 * Use a "cursor" link, to make the list traversal safe, even
280 * though elements might be moved about.
282 list_add(&cursor, &gc_candidates);
283 while (cursor.next != &gc_candidates) {
284 u = list_entry(cursor.next, struct unix_sock, link);
286 /* Move cursor to after the current position. */
287 list_move(&cursor, &u->link);
290 list_move_tail(&u->link, ¬_cycle_list);
291 __clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
292 scan_children(&u->sk, inc_inflight_move_tail, NULL);
297 /* Now gc_candidates contains only garbage. Restore original
298 * inflight counters for these as well, and remove the skbuffs
299 * which are creating the cycle(s).
301 skb_queue_head_init(&hitlist);
302 list_for_each_entry(u, &gc_candidates, link) {
303 scan_children(&u->sk, inc_inflight, &hitlist);
305 #if IS_ENABLED(CONFIG_AF_UNIX_OOB)
307 kfree_skb(u->oob_skb);
313 /* not_cycle_list contains those sockets which do not make up a
314 * cycle. Restore these to the inflight list.
316 while (!list_empty(¬_cycle_list)) {
317 u = list_entry(not_cycle_list.next, struct unix_sock, link);
318 __clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
319 list_move_tail(&u->link, &gc_inflight_list);
322 spin_unlock(&unix_gc_lock);
324 /* We need io_uring to clean its registered files, ignore all io_uring
325 * originated skbs. It's fine as io_uring doesn't keep references to
326 * other io_uring instances and so killing all other files in the cycle
327 * will put all io_uring references forcing it to go through normal
328 * release.path eventually putting registered files.
330 skb_queue_walk_safe(&hitlist, skb, next_skb) {
331 if (skb->destructor == io_uring_destruct_scm) {
332 __skb_unlink(skb, &hitlist);
333 skb_queue_tail(&skb->sk->sk_receive_queue, skb);
337 /* Here we are. Hitlist is filled. Die. */
338 __skb_queue_purge(&hitlist);
340 spin_lock(&unix_gc_lock);
342 /* There could be io_uring registered files, just push them back to
345 list_for_each_entry_safe(u, next, &gc_candidates, link)
346 list_move_tail(&u->link, &gc_inflight_list);
348 /* All candidates should have been detached by now. */
349 BUG_ON(!list_empty(&gc_candidates));
351 /* Paired with READ_ONCE() in wait_for_unix_gc(). */
352 WRITE_ONCE(gc_in_progress, false);
354 wake_up(&unix_gc_wait);
357 spin_unlock(&unix_gc_lock);