net/unix/garbage.c

   1 /*
   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  *
  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.
  69  */
  70
  71 #include <linux/kernel.h>
  72 #include <linux/string.h>
  73 #include <linux/socket.h>
  74 #include <linux/un.h>
  75 #include <linux/net.h>
  76 #include <linux/fs.h>
  77 #include <linux/slab.h>
  78 #include <linux/skbuff.h>
  79 #include <linux/netdevice.h>
  80 #include <linux/file.h>
  81 #include <linux/proc_fs.h>
  82 #include <linux/mutex.h>
  83
  84 #include <net/sock.h>
  85 #include <net/af_unix.h>
  86 #include <net/scm.h>
  87 #include <net/tcp_states.h>
  88
  89 /* Internal data structures and random procedures: */
  90
  91 static LIST_HEAD(gc_inflight_list);
  92 static LIST_HEAD(gc_candidates);
  93 static DEFINE_SPINLOCK(unix_gc_lock);
  94
  95 unsigned int unix_tot_inflight;
  96
  97
  98 static struct sock *unix_get_socket(struct file *filp)
  99 {
 100         struct sock *u_sock = NULL;
 101         struct inode *inode = filp->f_path.dentry->d_inode;
 102
 103         /*
 104          *      Socket ?
 105          */
 106         if (S_ISSOCK(inode->i_mode)) {
 107                 struct socket * sock = SOCKET_I(inode);
 108                 struct sock * s = sock->sk;
 109
 110                 /*
 111                  *      PF_UNIX ?
 112                  */
 113                 if (s && sock->ops && sock->ops->family == PF_UNIX)
 114                         u_sock = s;
 115         }
 116         return u_sock;
 117 }
 118
 119 /*
 120  *      Keep the number of times in flight count for the file
 121  *      descriptor if it is for an AF_UNIX socket.
 122  */
 123
 124 void unix_inflight(struct file *fp)
 125 {
 126         struct sock *s = unix_get_socket(fp);
 127         if(s) {
 128                 struct unix_sock *u = unix_sk(s);
 129                 spin_lock(&unix_gc_lock);
 130                 if (atomic_inc_return(&u->inflight) == 1) {
 131                         BUG_ON(!list_empty(&u->link));
 132                         list_add_tail(&u->link, &gc_inflight_list);
 133                 } else {
 134                         BUG_ON(list_empty(&u->link));
 135                 }
 136                 unix_tot_inflight++;
 137                 spin_unlock(&unix_gc_lock);
 138         }
 139 }
 140
 141 void unix_notinflight(struct file *fp)
 142 {
 143         struct sock *s = unix_get_socket(fp);
 144         if(s) {
 145                 struct unix_sock *u = unix_sk(s);
 146                 spin_lock(&unix_gc_lock);
 147                 BUG_ON(list_empty(&u->link));
 148                 if (atomic_dec_and_test(&u->inflight))
 149                         list_del_init(&u->link);
 150                 unix_tot_inflight--;
 151                 spin_unlock(&unix_gc_lock);
 152         }
 153 }
 154
 155 static inline struct sk_buff *sock_queue_head(struct sock *sk)
 156 {
 157         return (struct sk_buff *) &sk->sk_receive_queue;
 158 }
 159
 160 #define receive_queue_for_each_skb(sk, next, skb) \
 161         for (skb = sock_queue_head(sk)->next, next = skb->next; \
 162              skb != sock_queue_head(sk); skb = next, next = skb->next)
 163
 164 static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
 165                           struct sk_buff_head *hitlist)
 166 {
 167         struct sk_buff *skb;
 168         struct sk_buff *next;
 169
 170         spin_lock(&x->sk_receive_queue.lock);
 171         receive_queue_for_each_skb(x, next, skb) {
 172                 /*
 173                  *      Do we have file descriptors ?
 174                  */
 175                 if (UNIXCB(skb).fp) {
 176                         bool hit = false;
 177                         /*
 178                          *      Process the descriptors of this socket
 179                          */
 180                         int nfd = UNIXCB(skb).fp->count;
 181                         struct file **fp = UNIXCB(skb).fp->fp;
 182                         while (nfd--) {
 183                                 /*
 184                                  *      Get the socket the fd matches
 185                                  *      if it indeed does so
 186                                  */
 187                                 struct sock *sk = unix_get_socket(*fp++);
 188                                 if (sk) {
 189                                         hit = true;
 190                                         func(unix_sk(sk));
 191                                 }
 192                         }
 193                         if (hit && hitlist != NULL) {
 194                                 __skb_unlink(skb, &x->sk_receive_queue);
 195                                 __skb_queue_tail(hitlist, skb);
 196                         }
 197                 }
 198         }
 199         spin_unlock(&x->sk_receive_queue.lock);
 200 }
 201
 202 static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
 203                           struct sk_buff_head *hitlist)
 204 {
 205         if (x->sk_state != TCP_LISTEN)
 206                 scan_inflight(x, func, hitlist);
 207         else {
 208                 struct sk_buff *skb;
 209                 struct sk_buff *next;
 210                 struct unix_sock *u;
 211                 LIST_HEAD(embryos);
 212
 213                 /*
 214                  * For a listening socket collect the queued embryos
 215                  * and perform a scan on them as well.
 216                  */
 217                 spin_lock(&x->sk_receive_queue.lock);
 218                 receive_queue_for_each_skb(x, next, skb) {
 219                         u = unix_sk(skb->sk);
 220
 221                         /*
 222                          * An embryo cannot be in-flight, so it's safe
 223                          * to use the list link.
 224                          */
 225                         BUG_ON(!list_empty(&u->link));
 226                         list_add_tail(&u->link, &embryos);
 227                 }
 228                 spin_unlock(&x->sk_receive_queue.lock);
 229
 230                 while (!list_empty(&embryos)) {
 231                         u = list_entry(embryos.next, struct unix_sock, link);
 232                         scan_inflight(&u->sk, func, hitlist);
 233                         list_del_init(&u->link);
 234                 }
 235         }
 236 }
 237
 238 static void dec_inflight(struct unix_sock *usk)
 239 {
 240         atomic_dec(&usk->inflight);
 241 }
 242
 243 static void inc_inflight(struct unix_sock *usk)
 244 {
 245         atomic_inc(&usk->inflight);
 246 }
 247
 248 static void inc_inflight_move_tail(struct unix_sock *u)
 249 {
 250         atomic_inc(&u->inflight);
 251         /*
 252          * If this is still a candidate, move it to the end of the
 253          * list, so that it's checked even if it was already passed
 254          * over
 255          */
 256         if (u->gc_candidate)
 257                 list_move_tail(&u->link, &gc_candidates);
 258 }
 259
 260 /* The external entry point: unix_gc() */
 261
 262 void unix_gc(void)
 263 {
 264         static bool gc_in_progress = false;
 265
 266         struct unix_sock *u;
 267         struct unix_sock *next;
 268         struct sk_buff_head hitlist;
 269         struct list_head cursor;
 270
 271         spin_lock(&unix_gc_lock);
 272
 273         /* Avoid a recursive GC. */
 274         if (gc_in_progress)
 275                 goto out;
 276
 277         gc_in_progress = true;
 278         /*
 279          * First, select candidates for garbage collection.  Only
 280          * in-flight sockets are considered, and from those only ones
 281          * which don't have any external reference.
 282          *
 283          * Holding unix_gc_lock will protect these candidates from
 284          * being detached, and hence from gaining an external
 285          * reference.  This also means, that since there are no
 286          * possible receivers, the receive queues of these sockets are
 287          * static during the GC, even though the dequeue is done
 288          * before the detach without atomicity guarantees.
 289          */
 290         list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
 291                 int total_refs;
 292                 int inflight_refs;
 293
 294                 total_refs = file_count(u->sk.sk_socket->file);
 295                 inflight_refs = atomic_read(&u->inflight);
 296
 297                 BUG_ON(inflight_refs < 1);
 298                 BUG_ON(total_refs < inflight_refs);
 299                 if (total_refs == inflight_refs) {
 300                         list_move_tail(&u->link, &gc_candidates);
 301                         u->gc_candidate = 1;
 302                 }
 303         }
 304
 305         /*
 306          * Now remove all internal in-flight reference to children of
 307          * the candidates.
 308          */
 309         list_for_each_entry(u, &gc_candidates, link)
 310                 scan_children(&u->sk, dec_inflight, NULL);
 311
 312         /*
 313          * Restore the references for children of all candidates,
 314          * which have remaining references.  Do this recursively, so
 315          * only those remain, which form cyclic references.
 316          *
 317          * Use a "cursor" link, to make the list traversal safe, even
 318          * though elements might be moved about.
 319          */
 320         list_add(&cursor, &gc_candidates);
 321         while (cursor.next != &gc_candidates) {
 322                 u = list_entry(cursor.next, struct unix_sock, link);
 323
 324                 /* Move cursor to after the current position. */
 325                 list_move(&cursor, &u->link);
 326
 327                 if (atomic_read(&u->inflight) > 0) {
 328                         list_move_tail(&u->link, &gc_inflight_list);
 329                         u->gc_candidate = 0;
 330                         scan_children(&u->sk, inc_inflight_move_tail, NULL);
 331                 }
 332         }
 333         list_del(&cursor);
 334
 335         /*
 336          * Now gc_candidates contains only garbage.  Restore original
 337          * inflight counters for these as well, and remove the skbuffs
 338          * which are creating the cycle(s).
 339          */
 340         skb_queue_head_init(&hitlist);
 341         list_for_each_entry(u, &gc_candidates, link)
 342                 scan_children(&u->sk, inc_inflight, &hitlist);
 343
 344         spin_unlock(&unix_gc_lock);
 345
 346         /* Here we are. Hitlist is filled. Die. */
 347         __skb_queue_purge(&hitlist);
 348
 349         spin_lock(&unix_gc_lock);
 350
 351         /* All candidates should have been detached by now. */
 352         BUG_ON(!list_empty(&gc_candidates));
 353         gc_in_progress = false;
 354
 355  out:
 356         spin_unlock(&unix_gc_lock);
 357 }