1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * NET3: Garbage Collector For AF_UNIX sockets 4 * 5 * Garbage Collector: 6 * Copyright (C) Barak A. Pearlmutter. 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 * 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 * 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. 64 */ 65 66#include <linux/kernel.h> 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> 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> 78 79#include <net/sock.h> 80#include <net/af_unix.h> 81#include <net/scm.h> 82#include <net/tcp_states.h> 83 84#include "scm.h" 85 86/* Internal data structures and random procedures: */ 87 88static LIST_HEAD(gc_candidates); 89static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait); 90 91static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *), 92 struct sk_buff_head *hitlist) 93{ 94 struct sk_buff *skb; 95 struct sk_buff *next; 96 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) { 101 bool hit = false; 102 /* Process the descriptors of this socket */ 103 int nfd = UNIXCB(skb).fp->count; 104 struct file **fp = UNIXCB(skb).fp->fp; 105 106 while (nfd--) { 107 /* Get the socket the fd matches if it indeed does so */ 108 struct sock *sk = unix_get_socket(*fp++); 109 110 if (sk) { 111 struct unix_sock *u = unix_sk(sk); 112 113 /* Ignore non-candidates, they could 114 * have been added to the queues after 115 * starting the garbage collection 116 */ 117 if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) { 118 hit = true; 119 120 func(u); 121 } 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); 131} 132 133static void scan_children(struct sock *x, void (*func)(struct unix_sock *), 134 struct sk_buff_head *hitlist) 135{ 136 if (x->sk_state != TCP_LISTEN) { 137 scan_inflight(x, func, hitlist); 138 } else { 139 struct sk_buff *skb; 140 struct sk_buff *next; 141 struct unix_sock *u; 142 LIST_HEAD(embryos); 143 144 /* For a listening socket collect the queued embryos 145 * and perform a scan on them as well. 146 */ 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); 150 151 /* An embryo cannot be in-flight, so it's safe 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 } 165} 166 167static void dec_inflight(struct unix_sock *usk) 168{ 169 atomic_long_dec(&usk->inflight); 170} 171 172static void inc_inflight(struct unix_sock *usk) 173{ 174 atomic_long_inc(&usk->inflight); 175} 176 177static void inc_inflight_move_tail(struct unix_sock *u) 178{ 179 atomic_long_inc(&u->inflight); 180 /* If this still might be part of a cycle, move it to the end 181 * of the list, so that it's checked even if it was already 182 * passed over 183 */ 184 if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags)) 185 list_move_tail(&u->link, &gc_candidates); 186} 187 188static bool gc_in_progress; 189#define UNIX_INFLIGHT_TRIGGER_GC 16000 190 191void wait_for_unix_gc(void) 192{ 193 /* If number of inflight sockets is insane, 194 * force a garbage collect right now. 195 * Paired with the WRITE_ONCE() in unix_inflight(), 196 * unix_notinflight() and gc_in_progress(). 197 */ 198 if (READ_ONCE(unix_tot_inflight) > UNIX_INFLIGHT_TRIGGER_GC && 199 !READ_ONCE(gc_in_progress)) 200 unix_gc(); 201 wait_event(unix_gc_wait, gc_in_progress == false); 202} 203 204/* The external entry point: unix_gc() */ 205void unix_gc(void) 206{ 207 struct sk_buff *next_skb, *skb; 208 struct unix_sock *u; 209 struct unix_sock *next; 210 struct sk_buff_head hitlist; 211 struct list_head cursor; 212 LIST_HEAD(not_cycle_list); 213 214 spin_lock(&unix_gc_lock); 215 216 /* Avoid a recursive GC. */ 217 if (gc_in_progress) 218 goto out; 219 220 /* Paired with READ_ONCE() in wait_for_unix_gc(). */ 221 WRITE_ONCE(gc_in_progress, true); 222 223 /* First, select candidates for garbage collection. Only 224 * in-flight sockets are considered, and from those only ones 225 * which don't have any external reference. 226 * 227 * Holding unix_gc_lock will protect these candidates from 228 * being detached, and hence from gaining an external 229 * reference. Since there are no possible receivers, all 230 * buffers currently on the candidates' queues stay there 231 * during the garbage collection. 232 * 233 * We also know that no new candidate can be added onto the 234 * receive queues. Other, non candidate sockets _can_ be 235 * added to queue, so we must make sure only to touch 236 * candidates. 237 * 238 * Embryos, though never candidates themselves, affect which 239 * candidates are reachable by the garbage collector. Before 240 * being added to a listener's queue, an embryo may already 241 * receive data carrying SCM_RIGHTS, potentially making the 242 * passed socket a candidate that is not yet reachable by the 243 * collector. It becomes reachable once the embryo is 244 * enqueued. Therefore, we must ensure that no SCM-laden 245 * embryo appears in a (candidate) listener's queue between 246 * consecutive scan_children() calls. 247 */ 248 list_for_each_entry_safe(u, next, &gc_inflight_list, link) { 249 struct sock *sk = &u->sk; 250 long total_refs; 251 long inflight_refs; 252 253 total_refs = file_count(sk->sk_socket->file); 254 inflight_refs = atomic_long_read(&u->inflight); 255 256 BUG_ON(inflight_refs < 1); 257 BUG_ON(total_refs < inflight_refs); 258 if (total_refs == inflight_refs) { 259 list_move_tail(&u->link, &gc_candidates); 260 __set_bit(UNIX_GC_CANDIDATE, &u->gc_flags); 261 __set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags); 262 263 if (sk->sk_state == TCP_LISTEN) { 264 unix_state_lock(sk); 265 unix_state_unlock(sk); 266 } 267 } 268 } 269 270 /* Now remove all internal in-flight reference to children of 271 * the candidates. 272 */ 273 list_for_each_entry(u, &gc_candidates, link) 274 scan_children(&u->sk, dec_inflight, NULL); 275 276 /* Restore the references for children of all candidates, 277 * which have remaining references. Do this recursively, so 278 * only those remain, which form cyclic references. 279 * 280 * Use a "cursor" link, to make the list traversal safe, even 281 * though elements might be moved about. 282 */ 283 list_add(&cursor, &gc_candidates); 284 while (cursor.next != &gc_candidates) { 285 u = list_entry(cursor.next, struct unix_sock, link); 286 287 /* Move cursor to after the current position. */ 288 list_move(&cursor, &u->link); 289 290 if (atomic_long_read(&u->inflight) > 0) { 291 list_move_tail(&u->link, ¬_cycle_list); 292 __clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags); 293 scan_children(&u->sk, inc_inflight_move_tail, NULL); 294 } 295 } 296 list_del(&cursor); 297 298 /* Now gc_candidates contains only garbage. Restore original 299 * inflight counters for these as well, and remove the skbuffs 300 * which are creating the cycle(s). 301 */ 302 skb_queue_head_init(&hitlist); 303 list_for_each_entry(u, &gc_candidates, link) 304 scan_children(&u->sk, inc_inflight, &hitlist); 305 306 /* not_cycle_list contains those sockets which do not make up a 307 * cycle. Restore these to the inflight list. 308 */ 309 while (!list_empty(¬_cycle_list)) { 310 u = list_entry(not_cycle_list.next, struct unix_sock, link); 311 __clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags); 312 list_move_tail(&u->link, &gc_inflight_list); 313 } 314 315 spin_unlock(&unix_gc_lock); 316 317 /* We need io_uring to clean its registered files, ignore all io_uring 318 * originated skbs. It's fine as io_uring doesn't keep references to 319 * other io_uring instances and so killing all other files in the cycle 320 * will put all io_uring references forcing it to go through normal 321 * release.path eventually putting registered files. 322 */ 323 skb_queue_walk_safe(&hitlist, skb, next_skb) { 324 if (skb->scm_io_uring) { 325 __skb_unlink(skb, &hitlist); 326 skb_queue_tail(&skb->sk->sk_receive_queue, skb); 327 } 328 } 329 330 /* Here we are. Hitlist is filled. Die. */ 331 __skb_queue_purge(&hitlist); 332 333 spin_lock(&unix_gc_lock); 334 335 /* There could be io_uring registered files, just push them back to 336 * the inflight list 337 */ 338 list_for_each_entry_safe(u, next, &gc_candidates, link) 339 list_move_tail(&u->link, &gc_inflight_list); 340 341 /* All candidates should have been detached by now. */ 342 BUG_ON(!list_empty(&gc_candidates)); 343 344 /* Paired with READ_ONCE() in wait_for_unix_gc(). */ 345 WRITE_ONCE(gc_in_progress, false); 346 347 wake_up(&unix_gc_wait); 348 349 out: 350 spin_unlock(&unix_gc_lock); 351} 352