net/ipv4/inetpeer.c

   1 /*
   2  *              INETPEER - A storage for permanent information about peers
   3  *
   4  *  This source is covered by the GNU GPL, the same as all kernel sources.
   5  *
   6  *  Authors:    Andrey V. Savochkin <saw@msu.ru>
   7  */
   8
   9 #include <linux/module.h>
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/random.h>
  15 #include <linux/timer.h>
  16 #include <linux/time.h>
  17 #include <linux/kernel.h>
  18 #include <linux/mm.h>
  19 #include <linux/net.h>
  20 #include <net/ip.h>
  21 #include <net/inetpeer.h>
  22
  23 /*
  24  *  Theory of operations.
  25  *  We keep one entry for each peer IP address.  The nodes contains long-living
  26  *  information about the peer which doesn't depend on routes.
  27  *  At this moment this information consists only of ID field for the next
  28  *  outgoing IP packet.  This field is incremented with each packet as encoded
  29  *  in inet_getid() function (include/net/inetpeer.h).
  30  *  At the moment of writing this notes identifier of IP packets is generated
  31  *  to be unpredictable using this code only for packets subjected
  32  *  (actually or potentially) to defragmentation.  I.e. DF packets less than
  33  *  PMTU in size uses a constant ID and do not use this code (see
  34  *  ip_select_ident() in include/net/ip.h).
  35  *
  36  *  Route cache entries hold references to our nodes.
  37  *  New cache entries get references via lookup by destination IP address in
  38  *  the avl tree.  The reference is grabbed only when it's needed i.e. only
  39  *  when we try to output IP packet which needs an unpredictable ID (see
  40  *  __ip_select_ident() in net/ipv4/route.c).
  41  *  Nodes are removed only when reference counter goes to 0.
  42  *  When it's happened the node may be removed when a sufficient amount of
  43  *  time has been passed since its last use.  The less-recently-used entry can
  44  *  also be removed if the pool is overloaded i.e. if the total amount of
  45  *  entries is greater-or-equal than the threshold.
  46  *
  47  *  Node pool is organised as an AVL tree.
  48  *  Such an implementation has been chosen not just for fun.  It's a way to
  49  *  prevent easy and efficient DoS attacks by creating hash collisions.  A huge
  50  *  amount of long living nodes in a single hash slot would significantly delay
  51  *  lookups performed with disabled BHs.
  52  *
  53  *  Serialisation issues.
  54  *  1.  Nodes may appear in the tree only with the pool lock held.
  55  *  2.  Nodes may disappear from the tree only with the pool lock held
  56  *      AND reference count being 0.
  57  *  3.  Nodes appears and disappears from unused node list only under
  58  *      "inet_peer_unused_lock".
  59  *  4.  Global variable peer_total is modified under the pool lock.
  60  *  5.  struct inet_peer fields modification:
  61  *              avl_left, avl_right, avl_parent, avl_height: pool lock
  62  *              unused: unused node list lock
  63  *              refcnt: atomically against modifications on other CPU;
  64  *                 usually under some other lock to prevent node disappearing
  65  *              dtime: unused node list lock
  66  *              v4daddr: unchangeable
  67  *              ip_id_count: atomic value (no lock needed)
  68  */
  69
  70 static struct kmem_cache *peer_cachep __read_mostly;
  71
  72 #define node_height(x) x->avl_height
  73
  74 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
  75 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
  76 static const struct inet_peer peer_fake_node = {
  77         .avl_left       = peer_avl_empty_rcu,
  78         .avl_right      = peer_avl_empty_rcu,
  79         .avl_height     = 0
  80 };
  81
  82 static struct {
  83         struct inet_peer __rcu *root;
  84         spinlock_t      lock;
  85         int             total;
  86 } peers = {
  87         .root           = peer_avl_empty_rcu,
  88         .lock           = __SPIN_LOCK_UNLOCKED(peers.lock),
  89         .total          = 0,
  90 };
  91 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
  92
  93 /* Exported for sysctl_net_ipv4.  */
  94 int inet_peer_threshold __read_mostly = 65536 + 128;    /* start to throw entries more
  95                                          * aggressively at this stage */
  96 int inet_peer_minttl __read_mostly = 120 * HZ;  /* TTL under high load: 120 sec */
  97 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;      /* usual time to live: 10 min */
  98 int inet_peer_gc_mintime __read_mostly = 10 * HZ;
  99 int inet_peer_gc_maxtime __read_mostly = 120 * HZ;
 100
 101 static struct {
 102         struct list_head        list;
 103         spinlock_t              lock;
 104 } unused_peers = {
 105         .list                   = LIST_HEAD_INIT(unused_peers.list),
 106         .lock                   = __SPIN_LOCK_UNLOCKED(unused_peers.lock),
 107 };
 108
 109 static void peer_check_expire(unsigned long dummy);
 110 static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0);
 111
 112
 113 /* Called from ip_output.c:ip_init  */
 114 void __init inet_initpeers(void)
 115 {
 116         struct sysinfo si;
 117
 118         /* Use the straight interface to information about memory. */
 119         si_meminfo(&si);
 120         /* The values below were suggested by Alexey Kuznetsov
 121          * <kuznet@ms2.inr.ac.ru>.  I don't have any opinion about the values
 122          * myself.  --SAW
 123          */
 124         if (si.totalram <= (32768*1024)/PAGE_SIZE)
 125                 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
 126         if (si.totalram <= (16384*1024)/PAGE_SIZE)
 127                 inet_peer_threshold >>= 1; /* about 512KB */
 128         if (si.totalram <= (8192*1024)/PAGE_SIZE)
 129                 inet_peer_threshold >>= 2; /* about 128KB */
 130
 131         peer_cachep = kmem_cache_create("inet_peer_cache",
 132                         sizeof(struct inet_peer),
 133                         0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
 134                         NULL);
 135
 136         /* All the timers, started at system startup tend
 137            to synchronize. Perturb it a bit.
 138          */
 139         peer_periodic_timer.expires = jiffies
 140                 + net_random() % inet_peer_gc_maxtime
 141                 + inet_peer_gc_maxtime;
 142         add_timer(&peer_periodic_timer);
 143 }
 144
 145 /* Called with or without local BH being disabled. */
 146 static void unlink_from_unused(struct inet_peer *p)
 147 {
 148         if (!list_empty(&p->unused)) {
 149                 spin_lock_bh(&unused_peers.lock);
 150                 list_del_init(&p->unused);
 151                 spin_unlock_bh(&unused_peers.lock);
 152         }
 153 }
 154
 155 /*
 156  * Called with local BH disabled and the pool lock held.
 157  */
 158 #define lookup(_daddr, _stack)                                  \
 159 ({                                                              \
 160         struct inet_peer *u;                                    \
 161         struct inet_peer __rcu **v;                             \
 162                                                                 \
 163         stackptr = _stack;                                      \
 164         *stackptr++ = &peers.root;                              \
 165         for (u = rcu_dereference_protected(peers.root,          \
 166                         lockdep_is_held(&peers.lock));          \
 167              u != peer_avl_empty; ) {                           \
 168                 if (_daddr == u->v4daddr)                       \
 169                         break;                                  \
 170                 if ((__force __u32)_daddr < (__force __u32)u->v4daddr)  \
 171                         v = &u->avl_left;                       \
 172                 else                                            \
 173                         v = &u->avl_right;                      \
 174                 *stackptr++ = v;                                \
 175                 u = rcu_dereference_protected(*v,               \
 176                         lockdep_is_held(&peers.lock));          \
 177         }                                                       \
 178         u;                                                      \
 179 })
 180
 181 /*
 182  * Called with rcu_read_lock_bh()
 183  * Because we hold no lock against a writer, its quite possible we fall
 184  * in an endless loop.
 185  * But every pointer we follow is guaranteed to be valid thanks to RCU.
 186  * We exit from this function if number of links exceeds PEER_MAXDEPTH
 187  */
 188 static struct inet_peer *lookup_rcu_bh(__be32 daddr)
 189 {
 190         struct inet_peer *u = rcu_dereference_bh(peers.root);
 191         int count = 0;
 192
 193         while (u != peer_avl_empty) {
 194                 if (daddr == u->v4daddr) {
 195                         /* Before taking a reference, check if this entry was
 196                          * deleted, unlink_from_pool() sets refcnt=-1 to make
 197                          * distinction between an unused entry (refcnt=0) and
 198                          * a freed one.
 199                          */
 200                         if (unlikely(!atomic_add_unless(&u->refcnt, 1, -1)))
 201                                 u = NULL;
 202                         return u;
 203                 }
 204                 if ((__force __u32)daddr < (__force __u32)u->v4daddr)
 205                         u = rcu_dereference_bh(u->avl_left);
 206                 else
 207                         u = rcu_dereference_bh(u->avl_right);
 208                 if (unlikely(++count == PEER_MAXDEPTH))
 209                         break;
 210         }
 211         return NULL;
 212 }
 213
 214 /* Called with local BH disabled and the pool lock held. */
 215 #define lookup_rightempty(start)                                \
 216 ({                                                              \
 217         struct inet_peer *u;                                    \
 218         struct inet_peer __rcu **v;                             \
 219         *stackptr++ = &start->avl_left;                         \
 220         v = &start->avl_left;                                   \
 221         for (u = rcu_dereference_protected(*v,                  \
 222                         lockdep_is_held(&peers.lock));          \
 223              u->avl_right != peer_avl_empty_rcu; ) {            \
 224                 v = &u->avl_right;                              \
 225                 *stackptr++ = v;                                \
 226                 u = rcu_dereference_protected(*v,               \
 227                         lockdep_is_held(&peers.lock));          \
 228         }                                                       \
 229         u;                                                      \
 230 })
 231
 232 /* Called with local BH disabled and the pool lock held.
 233  * Variable names are the proof of operation correctness.
 234  * Look into mm/map_avl.c for more detail description of the ideas.
 235  */
 236 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
 237                 struct inet_peer __rcu ***stackend)
 238 {
 239         struct inet_peer __rcu **nodep;
 240         struct inet_peer *node, *l, *r;
 241         int lh, rh;
 242
 243         while (stackend > stack) {
 244                 nodep = *--stackend;
 245                 node = rcu_dereference_protected(*nodep,
 246                                 lockdep_is_held(&peers.lock));
 247                 l = rcu_dereference_protected(node->avl_left,
 248                                 lockdep_is_held(&peers.lock));
 249                 r = rcu_dereference_protected(node->avl_right,
 250                                 lockdep_is_held(&peers.lock));
 251                 lh = node_height(l);
 252                 rh = node_height(r);
 253                 if (lh > rh + 1) { /* l: RH+2 */
 254                         struct inet_peer *ll, *lr, *lrl, *lrr;
 255                         int lrh;
 256                         ll = rcu_dereference_protected(l->avl_left,
 257                                 lockdep_is_held(&peers.lock));
 258                         lr = rcu_dereference_protected(l->avl_right,
 259                                 lockdep_is_held(&peers.lock));
 260                         lrh = node_height(lr);
 261                         if (lrh <= node_height(ll)) {   /* ll: RH+1 */
 262                                 RCU_INIT_POINTER(node->avl_left, lr);   /* lr: RH or RH+1 */
 263                                 RCU_INIT_POINTER(node->avl_right, r);   /* r: RH */
 264                                 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
 265                                 RCU_INIT_POINTER(l->avl_left, ll);       /* ll: RH+1 */
 266                                 RCU_INIT_POINTER(l->avl_right, node);   /* node: RH+1 or RH+2 */
 267                                 l->avl_height = node->avl_height + 1;
 268                                 RCU_INIT_POINTER(*nodep, l);
 269                         } else { /* ll: RH, lr: RH+1 */
 270                                 lrl = rcu_dereference_protected(lr->avl_left,
 271                                         lockdep_is_held(&peers.lock));  /* lrl: RH or RH-1 */
 272                                 lrr = rcu_dereference_protected(lr->avl_right,
 273                                         lockdep_is_held(&peers.lock));  /* lrr: RH or RH-1 */
 274                                 RCU_INIT_POINTER(node->avl_left, lrr);  /* lrr: RH or RH-1 */
 275                                 RCU_INIT_POINTER(node->avl_right, r);   /* r: RH */
 276                                 node->avl_height = rh + 1; /* node: RH+1 */
 277                                 RCU_INIT_POINTER(l->avl_left, ll);      /* ll: RH */
 278                                 RCU_INIT_POINTER(l->avl_right, lrl);    /* lrl: RH or RH-1 */
 279                                 l->avl_height = rh + 1; /* l: RH+1 */
 280                                 RCU_INIT_POINTER(lr->avl_left, l);      /* l: RH+1 */
 281                                 RCU_INIT_POINTER(lr->avl_right, node);  /* node: RH+1 */
 282                                 lr->avl_height = rh + 2;
 283                                 RCU_INIT_POINTER(*nodep, lr);
 284                         }
 285                 } else if (rh > lh + 1) { /* r: LH+2 */
 286                         struct inet_peer *rr, *rl, *rlr, *rll;
 287                         int rlh;
 288                         rr = rcu_dereference_protected(r->avl_right,
 289                                 lockdep_is_held(&peers.lock));
 290                         rl = rcu_dereference_protected(r->avl_left,
 291                                 lockdep_is_held(&peers.lock));
 292                         rlh = node_height(rl);
 293                         if (rlh <= node_height(rr)) {   /* rr: LH+1 */
 294                                 RCU_INIT_POINTER(node->avl_right, rl);  /* rl: LH or LH+1 */
 295                                 RCU_INIT_POINTER(node->avl_left, l);    /* l: LH */
 296                                 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
 297                                 RCU_INIT_POINTER(r->avl_right, rr);     /* rr: LH+1 */
 298                                 RCU_INIT_POINTER(r->avl_left, node);    /* node: LH+1 or LH+2 */
 299                                 r->avl_height = node->avl_height + 1;
 300                                 RCU_INIT_POINTER(*nodep, r);
 301                         } else { /* rr: RH, rl: RH+1 */
 302                                 rlr = rcu_dereference_protected(rl->avl_right,
 303                                         lockdep_is_held(&peers.lock));  /* rlr: LH or LH-1 */
 304                                 rll = rcu_dereference_protected(rl->avl_left,
 305                                         lockdep_is_held(&peers.lock));  /* rll: LH or LH-1 */
 306                                 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
 307                                 RCU_INIT_POINTER(node->avl_left, l);    /* l: LH */
 308                                 node->avl_height = lh + 1; /* node: LH+1 */
 309                                 RCU_INIT_POINTER(r->avl_right, rr);     /* rr: LH */
 310                                 RCU_INIT_POINTER(r->avl_left, rlr);     /* rlr: LH or LH-1 */
 311                                 r->avl_height = lh + 1; /* r: LH+1 */
 312                                 RCU_INIT_POINTER(rl->avl_right, r);     /* r: LH+1 */
 313                                 RCU_INIT_POINTER(rl->avl_left, node);   /* node: LH+1 */
 314                                 rl->avl_height = lh + 2;
 315                                 RCU_INIT_POINTER(*nodep, rl);
 316                         }
 317                 } else {
 318                         node->avl_height = (lh > rh ? lh : rh) + 1;
 319                 }
 320         }
 321 }
 322
 323 /* Called with local BH disabled and the pool lock held. */
 324 #define link_to_pool(n)                                         \
 325 do {                                                            \
 326         n->avl_height = 1;                                      \
 327         n->avl_left = peer_avl_empty_rcu;                       \
 328         n->avl_right = peer_avl_empty_rcu;                      \
 329         /* lockless readers can catch us now */                 \
 330         rcu_assign_pointer(**--stackptr, n);                    \
 331         peer_avl_rebalance(stack, stackptr);                    \
 332 } while (0)
 333
 334 static void inetpeer_free_rcu(struct rcu_head *head)
 335 {
 336         kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
 337 }
 338
 339 /* May be called with local BH enabled. */
 340 static void unlink_from_pool(struct inet_peer *p)
 341 {
 342         int do_free;
 343
 344         do_free = 0;
 345
 346         spin_lock_bh(&peers.lock);
 347         /* Check the reference counter.  It was artificially incremented by 1
 348          * in cleanup() function to prevent sudden disappearing.  If we can
 349          * atomically (because of lockless readers) take this last reference,
 350          * it's safe to remove the node and free it later.
 351          * We use refcnt=-1 to alert lockless readers this entry is deleted.
 352          */
 353         if (atomic_cmpxchg(&p->refcnt, 1, -1) == 1) {
 354                 struct inet_peer __rcu **stack[PEER_MAXDEPTH];
 355                 struct inet_peer __rcu ***stackptr, ***delp;
 356                 if (lookup(p->v4daddr, stack) != p)
 357                         BUG();
 358                 delp = stackptr - 1; /* *delp[0] == p */
 359                 if (p->avl_left == peer_avl_empty_rcu) {
 360                         *delp[0] = p->avl_right;
 361                         --stackptr;
 362                 } else {
 363                         /* look for a node to insert instead of p */
 364                         struct inet_peer *t;
 365                         t = lookup_rightempty(p);
 366                         BUG_ON(rcu_dereference_protected(*stackptr[-1],
 367                                         lockdep_is_held(&peers.lock)) != t);
 368                         **--stackptr = t->avl_left;
 369                         /* t is removed, t->v4daddr > x->v4daddr for any
 370                          * x in p->avl_left subtree.
 371                          * Put t in the old place of p. */
 372                         RCU_INIT_POINTER(*delp[0], t);
 373                         t->avl_left = p->avl_left;
 374                         t->avl_right = p->avl_right;
 375                         t->avl_height = p->avl_height;
 376                         BUG_ON(delp[1] != &p->avl_left);
 377                         delp[1] = &t->avl_left; /* was &p->avl_left */
 378                 }
 379                 peer_avl_rebalance(stack, stackptr);
 380                 peers.total--;
 381                 do_free = 1;
 382         }
 383         spin_unlock_bh(&peers.lock);
 384
 385         if (do_free)
 386                 call_rcu_bh(&p->rcu, inetpeer_free_rcu);
 387         else
 388                 /* The node is used again.  Decrease the reference counter
 389                  * back.  The loop "cleanup -> unlink_from_unused
 390                  *   -> unlink_from_pool -> putpeer -> link_to_unused
 391                  *   -> cleanup (for the same node)"
 392                  * doesn't really exist because the entry will have a
 393                  * recent deletion time and will not be cleaned again soon.
 394                  */
 395                 inet_putpeer(p);
 396 }
 397
 398 /* May be called with local BH enabled. */
 399 static int cleanup_once(unsigned long ttl)
 400 {
 401         struct inet_peer *p = NULL;
 402
 403         /* Remove the first entry from the list of unused nodes. */
 404         spin_lock_bh(&unused_peers.lock);
 405         if (!list_empty(&unused_peers.list)) {
 406                 __u32 delta;
 407
 408                 p = list_first_entry(&unused_peers.list, struct inet_peer, unused);
 409                 delta = (__u32)jiffies - p->dtime;
 410
 411                 if (delta < ttl) {
 412                         /* Do not prune fresh entries. */
 413                         spin_unlock_bh(&unused_peers.lock);
 414                         return -1;
 415                 }
 416
 417                 list_del_init(&p->unused);
 418
 419                 /* Grab an extra reference to prevent node disappearing
 420                  * before unlink_from_pool() call. */
 421                 atomic_inc(&p->refcnt);
 422         }
 423         spin_unlock_bh(&unused_peers.lock);
 424
 425         if (p == NULL)
 426                 /* It means that the total number of USED entries has
 427                  * grown over inet_peer_threshold.  It shouldn't really
 428                  * happen because of entry limits in route cache. */
 429                 return -1;
 430
 431         unlink_from_pool(p);
 432         return 0;
 433 }
 434
 435 /* Called with or without local BH being disabled. */
 436 struct inet_peer *inet_getpeer(__be32 daddr, int create)
 437 {
 438         struct inet_peer *p;
 439         struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
 440
 441         /* Look up for the address quickly, lockless.
 442          * Because of a concurrent writer, we might not find an existing entry.
 443          */
 444         rcu_read_lock_bh();
 445         p = lookup_rcu_bh(daddr);
 446         rcu_read_unlock_bh();
 447
 448         if (p) {
 449                 /* The existing node has been found.
 450                  * Remove the entry from unused list if it was there.
 451                  */
 452                 unlink_from_unused(p);
 453                 return p;
 454         }
 455
 456         /* retry an exact lookup, taking the lock before.
 457          * At least, nodes should be hot in our cache.
 458          */
 459         spin_lock_bh(&peers.lock);
 460         p = lookup(daddr, stack);
 461         if (p != peer_avl_empty) {
 462                 atomic_inc(&p->refcnt);
 463                 spin_unlock_bh(&peers.lock);
 464                 /* Remove the entry from unused list if it was there. */
 465                 unlink_from_unused(p);
 466                 return p;
 467         }
 468         p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
 469         if (p) {
 470                 p->v4daddr = daddr;
 471                 atomic_set(&p->refcnt, 1);
 472                 atomic_set(&p->rid, 0);
 473                 atomic_set(&p->ip_id_count, secure_ip_id(daddr));
 474                 p->tcp_ts_stamp = 0;
 475                 INIT_LIST_HEAD(&p->unused);
 476
 477
 478                 /* Link the node. */
 479                 link_to_pool(p);
 480                 peers.total++;
 481         }
 482         spin_unlock_bh(&peers.lock);
 483
 484         if (peers.total >= inet_peer_threshold)
 485                 /* Remove one less-recently-used entry. */
 486                 cleanup_once(0);
 487
 488         return p;
 489 }
 490
 491 /* Called with local BH disabled. */
 492 static void peer_check_expire(unsigned long dummy)
 493 {
 494         unsigned long now = jiffies;
 495         int ttl;
 496
 497         if (peers.total >= inet_peer_threshold)
 498                 ttl = inet_peer_minttl;
 499         else
 500                 ttl = inet_peer_maxttl
 501                                 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
 502                                         peers.total / inet_peer_threshold * HZ;
 503         while (!cleanup_once(ttl)) {
 504                 if (jiffies != now)
 505                         break;
 506         }
 507
 508         /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
 509          * interval depending on the total number of entries (more entries,
 510          * less interval). */
 511         if (peers.total >= inet_peer_threshold)
 512                 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
 513         else
 514                 peer_periodic_timer.expires = jiffies
 515                         + inet_peer_gc_maxtime
 516                         - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
 517                                 peers.total / inet_peer_threshold * HZ;
 518         add_timer(&peer_periodic_timer);
 519 }
 520
 521 void inet_putpeer(struct inet_peer *p)
 522 {
 523         local_bh_disable();
 524
 525         if (atomic_dec_and_lock(&p->refcnt, &unused_peers.lock)) {
 526                 list_add_tail(&p->unused, &unused_peers.list);
 527                 p->dtime = (__u32)jiffies;
 528                 spin_unlock(&unused_peers.lock);
 529         }
 530
 531         local_bh_enable();
 532 }