Merge branch 'work.dcache2' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux-2.6-block.git] / net / unix / garbage.c
CommitLineData
a85036f6 1// SPDX-License-Identifier: GPL-2.0-or-later
1da177e4
LT
2/*
3 * NET3: Garbage Collector For AF_UNIX sockets
4 *
5 * Garbage Collector:
6 * Copyright (C) Barak A. Pearlmutter.
1da177e4
LT
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 *
1da177e4
LT
27 * Fixes:
28 * Alan Cox 07 Sept 1997 Vmalloc internal stack as needed.
29 * Cope with changing max_files.
30 * Al Viro 11 Oct 1998
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.
39 *
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.
49 *
50 * AV 28 Feb 1999
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).
57 * AV 1 Mar 1999
58 * Damn. Added missing check for ->dead in listen queues scanning.
59 *
1fd05ba5
MS
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.
1da177e4 64 */
ac7bfa62 65
1da177e4 66#include <linux/kernel.h>
1da177e4
LT
67#include <linux/string.h>
68#include <linux/socket.h>
69#include <linux/un.h>
70#include <linux/net.h>
71#include <linux/fs.h>
1da177e4
LT
72#include <linux/skbuff.h>
73#include <linux/netdevice.h>
74#include <linux/file.h>
75#include <linux/proc_fs.h>
4a3e2f71 76#include <linux/mutex.h>
5f23b734 77#include <linux/wait.h>
1da177e4
LT
78
79#include <net/sock.h>
80#include <net/af_unix.h>
81#include <net/scm.h>
c752f073 82#include <net/tcp_states.h>
1da177e4 83
f4e65870
JA
84#include "scm.h"
85
1da177e4
LT
86/* Internal data structures and random procedures: */
87
1fd05ba5 88static LIST_HEAD(gc_candidates);
5f23b734 89static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
1da177e4 90
5c80f1ae 91static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
1fd05ba5 92 struct sk_buff_head *hitlist)
1da177e4 93{
1fd05ba5
MS
94 struct sk_buff *skb;
95 struct sk_buff *next;
96
97 spin_lock(&x->sk_receive_queue.lock);
a2f3be17 98 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
d1ab39f1 99 /* Do we have file descriptors ? */
1fd05ba5
MS
100 if (UNIXCB(skb).fp) {
101 bool hit = false;
d1ab39f1 102 /* Process the descriptors of this socket */
1fd05ba5
MS
103 int nfd = UNIXCB(skb).fp->count;
104 struct file **fp = UNIXCB(skb).fp->fp;
d1ab39f1 105
1fd05ba5 106 while (nfd--) {
d1ab39f1 107 /* Get the socket the fd matches if it indeed does so */
1fd05ba5 108 struct sock *sk = unix_get_socket(*fp++);
d1ab39f1 109
5c80f1ae 110 if (sk) {
6209344f
MS
111 struct unix_sock *u = unix_sk(sk);
112
d1ab39f1 113 /* Ignore non-candidates, they could
6209344f
MS
114 * have been added to the queues after
115 * starting the garbage collection
116 */
60bc851a 117 if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
6209344f 118 hit = true;
d1ab39f1 119
6209344f
MS
120 func(u);
121 }
1fd05ba5
MS
122 }
123 }
124 if (hit && hitlist != NULL) {
125 __skb_unlink(skb, &x->sk_receive_queue);
126 __skb_queue_tail(hitlist, skb);
127 }
128 }
129 }
130 spin_unlock(&x->sk_receive_queue.lock);
1da177e4
LT
131}
132
5c80f1ae 133static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
1fd05ba5 134 struct sk_buff_head *hitlist)
1da177e4 135{
d1ab39f1 136 if (x->sk_state != TCP_LISTEN) {
1fd05ba5 137 scan_inflight(x, func, hitlist);
d1ab39f1 138 } else {
1fd05ba5
MS
139 struct sk_buff *skb;
140 struct sk_buff *next;
141 struct unix_sock *u;
142 LIST_HEAD(embryos);
143
d1ab39f1 144 /* For a listening socket collect the queued embryos
1fd05ba5
MS
145 * and perform a scan on them as well.
146 */
147 spin_lock(&x->sk_receive_queue.lock);
a2f3be17 148 skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
1fd05ba5
MS
149 u = unix_sk(skb->sk);
150
d1ab39f1 151 /* An embryo cannot be in-flight, so it's safe
1fd05ba5
MS
152 * to use the list link.
153 */
154 BUG_ON(!list_empty(&u->link));
155 list_add_tail(&u->link, &embryos);
156 }
157 spin_unlock(&x->sk_receive_queue.lock);
158
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);
163 }
164 }
1da177e4
LT
165}
166
5c80f1ae 167static void dec_inflight(struct unix_sock *usk)
1da177e4 168{
516e0cc5 169 atomic_long_dec(&usk->inflight);
1fd05ba5 170}
1da177e4 171
5c80f1ae 172static void inc_inflight(struct unix_sock *usk)
1fd05ba5 173{
516e0cc5 174 atomic_long_inc(&usk->inflight);
1da177e4
LT
175}
176
5c80f1ae 177static void inc_inflight_move_tail(struct unix_sock *u)
1fd05ba5 178{
516e0cc5 179 atomic_long_inc(&u->inflight);
d1ab39f1 180 /* If this still might be part of a cycle, move it to the end
6209344f
MS
181 * of the list, so that it's checked even if it was already
182 * passed over
1fd05ba5 183 */
60bc851a 184 if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
1fd05ba5
MS
185 list_move_tail(&u->link, &gc_candidates);
186}
1da177e4 187
505e907d 188static bool gc_in_progress;
9915672d 189#define UNIX_INFLIGHT_TRIGGER_GC 16000
1da177e4 190
5f23b734 191void wait_for_unix_gc(void)
1da177e4 192{
d1ab39f1 193 /* If number of inflight sockets is insane,
9915672d
ED
194 * force a garbage collect right now.
195 */
196 if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
197 unix_gc();
5f23b734 198 wait_event(unix_gc_wait, gc_in_progress == false);
199}
1da177e4 200
5f23b734 201/* The external entry point: unix_gc() */
202void unix_gc(void)
203{
1fd05ba5
MS
204 struct unix_sock *u;
205 struct unix_sock *next;
206 struct sk_buff_head hitlist;
207 struct list_head cursor;
6209344f 208 LIST_HEAD(not_cycle_list);
1da177e4 209
1fd05ba5 210 spin_lock(&unix_gc_lock);
1da177e4 211
1fd05ba5
MS
212 /* Avoid a recursive GC. */
213 if (gc_in_progress)
214 goto out;
1da177e4 215
1fd05ba5 216 gc_in_progress = true;
d1ab39f1 217 /* First, select candidates for garbage collection. Only
1fd05ba5
MS
218 * in-flight sockets are considered, and from those only ones
219 * which don't have any external reference.
220 *
221 * Holding unix_gc_lock will protect these candidates from
222 * being detached, and hence from gaining an external
6209344f
MS
223 * reference. Since there are no possible receivers, all
224 * buffers currently on the candidates' queues stay there
225 * during the garbage collection.
226 *
227 * We also know that no new candidate can be added onto the
228 * receive queues. Other, non candidate sockets _can_ be
229 * added to queue, so we must make sure only to touch
230 * candidates.
1da177e4 231 */
1fd05ba5 232 list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
516e0cc5
AV
233 long total_refs;
234 long inflight_refs;
1fd05ba5
MS
235
236 total_refs = file_count(u->sk.sk_socket->file);
516e0cc5 237 inflight_refs = atomic_long_read(&u->inflight);
1fd05ba5
MS
238
239 BUG_ON(inflight_refs < 1);
240 BUG_ON(total_refs < inflight_refs);
241 if (total_refs == inflight_refs) {
242 list_move_tail(&u->link, &gc_candidates);
60bc851a
ED
243 __set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
244 __set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
1fd05ba5
MS
245 }
246 }
1da177e4 247
d1ab39f1 248 /* Now remove all internal in-flight reference to children of
1fd05ba5 249 * the candidates.
1da177e4 250 */
1fd05ba5
MS
251 list_for_each_entry(u, &gc_candidates, link)
252 scan_children(&u->sk, dec_inflight, NULL);
1da177e4 253
d1ab39f1 254 /* Restore the references for children of all candidates,
1fd05ba5
MS
255 * which have remaining references. Do this recursively, so
256 * only those remain, which form cyclic references.
257 *
258 * Use a "cursor" link, to make the list traversal safe, even
259 * though elements might be moved about.
1da177e4 260 */
1fd05ba5
MS
261 list_add(&cursor, &gc_candidates);
262 while (cursor.next != &gc_candidates) {
263 u = list_entry(cursor.next, struct unix_sock, link);
1da177e4 264
1fd05ba5
MS
265 /* Move cursor to after the current position. */
266 list_move(&cursor, &u->link);
ac7bfa62 267
516e0cc5 268 if (atomic_long_read(&u->inflight) > 0) {
6209344f 269 list_move_tail(&u->link, &not_cycle_list);
60bc851a 270 __clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
1fd05ba5 271 scan_children(&u->sk, inc_inflight_move_tail, NULL);
1da177e4 272 }
1da177e4 273 }
1fd05ba5 274 list_del(&cursor);
1da177e4 275
7df9c246
AU
276 /* Now gc_candidates contains only garbage. Restore original
277 * inflight counters for these as well, and remove the skbuffs
278 * which are creating the cycle(s).
279 */
280 skb_queue_head_init(&hitlist);
281 list_for_each_entry(u, &gc_candidates, link)
282 scan_children(&u->sk, inc_inflight, &hitlist);
283
d1ab39f1 284 /* not_cycle_list contains those sockets which do not make up a
6209344f
MS
285 * cycle. Restore these to the inflight list.
286 */
287 while (!list_empty(&not_cycle_list)) {
288 u = list_entry(not_cycle_list.next, struct unix_sock, link);
60bc851a 289 __clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
6209344f
MS
290 list_move_tail(&u->link, &gc_inflight_list);
291 }
292
1fd05ba5 293 spin_unlock(&unix_gc_lock);
1da177e4 294
1fd05ba5
MS
295 /* Here we are. Hitlist is filled. Die. */
296 __skb_queue_purge(&hitlist);
1da177e4 297
1fd05ba5 298 spin_lock(&unix_gc_lock);
1da177e4 299
1fd05ba5
MS
300 /* All candidates should have been detached by now. */
301 BUG_ON(!list_empty(&gc_candidates));
302 gc_in_progress = false;
5f23b734 303 wake_up(&unix_gc_wait);
1da177e4 304
1fd05ba5
MS
305 out:
306 spin_unlock(&unix_gc_lock);
1da177e4 307}