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.  Global variable peer_total is modified under the pool lock.
  58  *  4.  struct inet_peer fields modification:
  59  *              avl_left, avl_right, avl_parent, avl_height: pool lock
  60  *              refcnt: atomically against modifications on other CPU;
  61  *                 usually under some other lock to prevent node disappearing
  62  *              daddr: unchangeable
  63  *              ip_id_count: atomic value (no lock needed)
  64  */
  65
  66 static struct kmem_cache *peer_cachep __read_mostly;
  67
  68 #define node_height(x) x->avl_height
  69
  70 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
  71 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
  72 static const struct inet_peer peer_fake_node = {
  73         .avl_left       = peer_avl_empty_rcu,
  74         .avl_right      = peer_avl_empty_rcu,
  75         .avl_height     = 0
  76 };
  77
  78 struct inet_peer_base {
  79         struct inet_peer __rcu *root;
  80         seqlock_t       lock;
  81         int             total;
  82 };
  83
  84 static struct inet_peer_base v4_peers = {
  85         .root           = peer_avl_empty_rcu,
  86         .lock           = __SEQLOCK_UNLOCKED(v4_peers.lock),
  87         .total          = 0,
  88 };
  89
  90 static struct inet_peer_base v6_peers = {
  91         .root           = peer_avl_empty_rcu,
  92         .lock           = __SEQLOCK_UNLOCKED(v6_peers.lock),
  93         .total          = 0,
  94 };
  95
  96 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
  97
  98 /* Exported for sysctl_net_ipv4.  */
  99 int inet_peer_threshold __read_mostly = 65536 + 128;    /* start to throw entries more
 100                                          * aggressively at this stage */
 101 int inet_peer_minttl __read_mostly = 120 * HZ;  /* TTL under high load: 120 sec */
 102 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;      /* usual time to live: 10 min */
 103
 104
 105 /* Called from ip_output.c:ip_init  */
 106 void __init inet_initpeers(void)
 107 {
 108         struct sysinfo si;
 109
 110         /* Use the straight interface to information about memory. */
 111         si_meminfo(&si);
 112         /* The values below were suggested by Alexey Kuznetsov
 113          * <kuznet@ms2.inr.ac.ru>.  I don't have any opinion about the values
 114          * myself.  --SAW
 115          */
 116         if (si.totalram <= (32768*1024)/PAGE_SIZE)
 117                 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
 118         if (si.totalram <= (16384*1024)/PAGE_SIZE)
 119                 inet_peer_threshold >>= 1; /* about 512KB */
 120         if (si.totalram <= (8192*1024)/PAGE_SIZE)
 121                 inet_peer_threshold >>= 2; /* about 128KB */
 122
 123         peer_cachep = kmem_cache_create("inet_peer_cache",
 124                         sizeof(struct inet_peer),
 125                         0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
 126                         NULL);
 127
 128 }
 129
 130 static int addr_compare(const struct inetpeer_addr *a,
 131                         const struct inetpeer_addr *b)
 132 {
 133         int i, n = (a->family == AF_INET ? 1 : 4);
 134
 135         for (i = 0; i < n; i++) {
 136                 if (a->addr.a6[i] == b->addr.a6[i])
 137                         continue;
 138                 if (a->addr.a6[i] < b->addr.a6[i])
 139                         return -1;
 140                 return 1;
 141         }
 142
 143         return 0;
 144 }
 145
 146 #define rcu_deref_locked(X, BASE)                               \
 147         rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
 148
 149 /*
 150  * Called with local BH disabled and the pool lock held.
 151  */
 152 #define lookup(_daddr, _stack, _base)                           \
 153 ({                                                              \
 154         struct inet_peer *u;                                    \
 155         struct inet_peer __rcu **v;                             \
 156                                                                 \
 157         stackptr = _stack;                                      \
 158         *stackptr++ = &_base->root;                             \
 159         for (u = rcu_deref_locked(_base->root, _base);          \
 160              u != peer_avl_empty; ) {                           \
 161                 int cmp = addr_compare(_daddr, &u->daddr);      \
 162                 if (cmp == 0)                                   \
 163                         break;                                  \
 164                 if (cmp == -1)                                  \
 165                         v = &u->avl_left;                       \
 166                 else                                            \
 167                         v = &u->avl_right;                      \
 168                 *stackptr++ = v;                                \
 169                 u = rcu_deref_locked(*v, _base);                \
 170         }                                                       \
 171         u;                                                      \
 172 })
 173
 174 /*
 175  * Called with rcu_read_lock()
 176  * Because we hold no lock against a writer, its quite possible we fall
 177  * in an endless loop.
 178  * But every pointer we follow is guaranteed to be valid thanks to RCU.
 179  * We exit from this function if number of links exceeds PEER_MAXDEPTH
 180  */
 181 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
 182                                     struct inet_peer_base *base)
 183 {
 184         struct inet_peer *u = rcu_dereference(base->root);
 185         int count = 0;
 186
 187         while (u != peer_avl_empty) {
 188                 int cmp = addr_compare(daddr, &u->daddr);
 189                 if (cmp == 0) {
 190                         /* Before taking a reference, check if this entry was
 191                          * deleted (refcnt=-1)
 192                          */
 193                         if (!atomic_add_unless(&u->refcnt, 1, -1))
 194                                 u = NULL;
 195                         return u;
 196                 }
 197                 if (cmp == -1)
 198                         u = rcu_dereference(u->avl_left);
 199                 else
 200                         u = rcu_dereference(u->avl_right);
 201                 if (unlikely(++count == PEER_MAXDEPTH))
 202                         break;
 203         }
 204         return NULL;
 205 }
 206
 207 /* Called with local BH disabled and the pool lock held. */
 208 #define lookup_rightempty(start, base)                          \
 209 ({                                                              \
 210         struct inet_peer *u;                                    \
 211         struct inet_peer __rcu **v;                             \
 212         *stackptr++ = &start->avl_left;                         \
 213         v = &start->avl_left;                                   \
 214         for (u = rcu_deref_locked(*v, base);                    \
 215              u->avl_right != peer_avl_empty_rcu; ) {            \
 216                 v = &u->avl_right;                              \
 217                 *stackptr++ = v;                                \
 218                 u = rcu_deref_locked(*v, base);                 \
 219         }                                                       \
 220         u;                                                      \
 221 })
 222
 223 /* Called with local BH disabled and the pool lock held.
 224  * Variable names are the proof of operation correctness.
 225  * Look into mm/map_avl.c for more detail description of the ideas.
 226  */
 227 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
 228                                struct inet_peer __rcu ***stackend,
 229                                struct inet_peer_base *base)
 230 {
 231         struct inet_peer __rcu **nodep;
 232         struct inet_peer *node, *l, *r;
 233         int lh, rh;
 234
 235         while (stackend > stack) {
 236                 nodep = *--stackend;
 237                 node = rcu_deref_locked(*nodep, base);
 238                 l = rcu_deref_locked(node->avl_left, base);
 239                 r = rcu_deref_locked(node->avl_right, base);
 240                 lh = node_height(l);
 241                 rh = node_height(r);
 242                 if (lh > rh + 1) { /* l: RH+2 */
 243                         struct inet_peer *ll, *lr, *lrl, *lrr;
 244                         int lrh;
 245                         ll = rcu_deref_locked(l->avl_left, base);
 246                         lr = rcu_deref_locked(l->avl_right, base);
 247                         lrh = node_height(lr);
 248                         if (lrh <= node_height(ll)) {   /* ll: RH+1 */
 249                                 RCU_INIT_POINTER(node->avl_left, lr);   /* lr: RH or RH+1 */
 250                                 RCU_INIT_POINTER(node->avl_right, r);   /* r: RH */
 251                                 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
 252                                 RCU_INIT_POINTER(l->avl_left, ll);       /* ll: RH+1 */
 253                                 RCU_INIT_POINTER(l->avl_right, node);   /* node: RH+1 or RH+2 */
 254                                 l->avl_height = node->avl_height + 1;
 255                                 RCU_INIT_POINTER(*nodep, l);
 256                         } else { /* ll: RH, lr: RH+1 */
 257                                 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
 258                                 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
 259                                 RCU_INIT_POINTER(node->avl_left, lrr);  /* lrr: RH or RH-1 */
 260                                 RCU_INIT_POINTER(node->avl_right, r);   /* r: RH */
 261                                 node->avl_height = rh + 1; /* node: RH+1 */
 262                                 RCU_INIT_POINTER(l->avl_left, ll);      /* ll: RH */
 263                                 RCU_INIT_POINTER(l->avl_right, lrl);    /* lrl: RH or RH-1 */
 264                                 l->avl_height = rh + 1; /* l: RH+1 */
 265                                 RCU_INIT_POINTER(lr->avl_left, l);      /* l: RH+1 */
 266                                 RCU_INIT_POINTER(lr->avl_right, node);  /* node: RH+1 */
 267                                 lr->avl_height = rh + 2;
 268                                 RCU_INIT_POINTER(*nodep, lr);
 269                         }
 270                 } else if (rh > lh + 1) { /* r: LH+2 */
 271                         struct inet_peer *rr, *rl, *rlr, *rll;
 272                         int rlh;
 273                         rr = rcu_deref_locked(r->avl_right, base);
 274                         rl = rcu_deref_locked(r->avl_left, base);
 275                         rlh = node_height(rl);
 276                         if (rlh <= node_height(rr)) {   /* rr: LH+1 */
 277                                 RCU_INIT_POINTER(node->avl_right, rl);  /* rl: LH or LH+1 */
 278                                 RCU_INIT_POINTER(node->avl_left, l);    /* l: LH */
 279                                 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
 280                                 RCU_INIT_POINTER(r->avl_right, rr);     /* rr: LH+1 */
 281                                 RCU_INIT_POINTER(r->avl_left, node);    /* node: LH+1 or LH+2 */
 282                                 r->avl_height = node->avl_height + 1;
 283                                 RCU_INIT_POINTER(*nodep, r);
 284                         } else { /* rr: RH, rl: RH+1 */
 285                                 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
 286                                 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
 287                                 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
 288                                 RCU_INIT_POINTER(node->avl_left, l);    /* l: LH */
 289                                 node->avl_height = lh + 1; /* node: LH+1 */
 290                                 RCU_INIT_POINTER(r->avl_right, rr);     /* rr: LH */
 291                                 RCU_INIT_POINTER(r->avl_left, rlr);     /* rlr: LH or LH-1 */
 292                                 r->avl_height = lh + 1; /* r: LH+1 */
 293                                 RCU_INIT_POINTER(rl->avl_right, r);     /* r: LH+1 */
 294                                 RCU_INIT_POINTER(rl->avl_left, node);   /* node: LH+1 */
 295                                 rl->avl_height = lh + 2;
 296                                 RCU_INIT_POINTER(*nodep, rl);
 297                         }
 298                 } else {
 299                         node->avl_height = (lh > rh ? lh : rh) + 1;
 300                 }
 301         }
 302 }
 303
 304 /* Called with local BH disabled and the pool lock held. */
 305 #define link_to_pool(n, base)                                   \
 306 do {                                                            \
 307         n->avl_height = 1;                                      \
 308         n->avl_left = peer_avl_empty_rcu;                       \
 309         n->avl_right = peer_avl_empty_rcu;                      \
 310         /* lockless readers can catch us now */                 \
 311         rcu_assign_pointer(**--stackptr, n);                    \
 312         peer_avl_rebalance(stack, stackptr, base);              \
 313 } while (0)
 314
 315 static void inetpeer_free_rcu(struct rcu_head *head)
 316 {
 317         kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
 318 }
 319
 320 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
 321                              struct inet_peer __rcu **stack[PEER_MAXDEPTH])
 322 {
 323         struct inet_peer __rcu ***stackptr, ***delp;
 324
 325         if (lookup(&p->daddr, stack, base) != p)
 326                 BUG();
 327         delp = stackptr - 1; /* *delp[0] == p */
 328         if (p->avl_left == peer_avl_empty_rcu) {
 329                 *delp[0] = p->avl_right;
 330                 --stackptr;
 331         } else {
 332                 /* look for a node to insert instead of p */
 333                 struct inet_peer *t;
 334                 t = lookup_rightempty(p, base);
 335                 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
 336                 **--stackptr = t->avl_left;
 337                 /* t is removed, t->daddr > x->daddr for any
 338                  * x in p->avl_left subtree.
 339                  * Put t in the old place of p. */
 340                 RCU_INIT_POINTER(*delp[0], t);
 341                 t->avl_left = p->avl_left;
 342                 t->avl_right = p->avl_right;
 343                 t->avl_height = p->avl_height;
 344                 BUG_ON(delp[1] != &p->avl_left);
 345                 delp[1] = &t->avl_left; /* was &p->avl_left */
 346         }
 347         peer_avl_rebalance(stack, stackptr, base);
 348         base->total--;
 349         call_rcu(&p->rcu, inetpeer_free_rcu);
 350 }
 351
 352 static struct inet_peer_base *family_to_base(int family)
 353 {
 354         return family == AF_INET ? &v4_peers : &v6_peers;
 355 }
 356
 357 /* perform garbage collect on all items stacked during a lookup */
 358 static int inet_peer_gc(struct inet_peer_base *base,
 359                         struct inet_peer __rcu **stack[PEER_MAXDEPTH],
 360                         struct inet_peer __rcu ***stackptr)
 361 {
 362         struct inet_peer *p, *gchead = NULL;
 363         __u32 delta, ttl;
 364         int cnt = 0;
 365
 366         if (base->total >= inet_peer_threshold)
 367                 ttl = 0; /* be aggressive */
 368         else
 369                 ttl = inet_peer_maxttl
 370                                 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
 371                                         base->total / inet_peer_threshold * HZ;
 372         stackptr--; /* last stack slot is peer_avl_empty */
 373         while (stackptr > stack) {
 374                 stackptr--;
 375                 p = rcu_deref_locked(**stackptr, base);
 376                 delta = (__u32)jiffies - p->dtime;
 377                 if (atomic_read(&p->refcnt) == 0 && delta >= ttl &&
 378                     atomic_cmpxchg(&p->refcnt, 0, -1) == 0) {
 379                         p->gc_next = gchead;
 380                         gchead = p;
 381                 }
 382         }
 383         while ((p = gchead) != NULL) {
 384                 gchead = p->gc_next;
 385                 cnt++;
 386                 unlink_from_pool(p, base, stack);
 387         }
 388         return cnt;
 389 }
 390
 391 struct inet_peer *inet_getpeer(struct inetpeer_addr *daddr, int create)
 392 {
 393         struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
 394         struct inet_peer_base *base = family_to_base(daddr->family);
 395         struct inet_peer *p;
 396         unsigned int sequence;
 397         int invalidated, gccnt = 0;
 398
 399         /* Attempt a lockless lookup first.
 400          * Because of a concurrent writer, we might not find an existing entry.
 401          */
 402         rcu_read_lock();
 403         sequence = read_seqbegin(&base->lock);
 404         p = lookup_rcu(daddr, base);
 405         invalidated = read_seqretry(&base->lock, sequence);
 406         rcu_read_unlock();
 407
 408         if (p)
 409                 return p;
 410
 411         /* If no writer did a change during our lookup, we can return early. */
 412         if (!create && !invalidated)
 413                 return NULL;
 414
 415         /* retry an exact lookup, taking the lock before.
 416          * At least, nodes should be hot in our cache.
 417          */
 418         write_seqlock_bh(&base->lock);
 419 relookup:
 420         p = lookup(daddr, stack, base);
 421         if (p != peer_avl_empty) {
 422                 atomic_inc(&p->refcnt);
 423                 write_sequnlock_bh(&base->lock);
 424                 return p;
 425         }
 426         if (!gccnt) {
 427                 gccnt = inet_peer_gc(base, stack, stackptr);
 428                 if (gccnt && create)
 429                         goto relookup;
 430         }
 431         p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
 432         if (p) {
 433                 p->daddr = *daddr;
 434                 atomic_set(&p->refcnt, 1);
 435                 atomic_set(&p->rid, 0);
 436                 atomic_set(&p->ip_id_count, secure_ip_id(daddr->addr.a4));
 437                 p->tcp_ts_stamp = 0;
 438                 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
 439                 p->rate_tokens = 0;
 440                 p->rate_last = 0;
 441                 p->pmtu_expires = 0;
 442                 p->pmtu_orig = 0;
 443                 memset(&p->redirect_learned, 0, sizeof(p->redirect_learned));
 444
 445
 446                 /* Link the node. */
 447                 link_to_pool(p, base);
 448                 base->total++;
 449         }
 450         write_sequnlock_bh(&base->lock);
 451
 452         return p;
 453 }
 454 EXPORT_SYMBOL_GPL(inet_getpeer);
 455
 456 void inet_putpeer(struct inet_peer *p)
 457 {
 458         p->dtime = (__u32)jiffies;
 459         atomic_dec(&p->refcnt);
 460 }
 461 EXPORT_SYMBOL_GPL(inet_putpeer);
 462
 463 /*
 464  *      Check transmit rate limitation for given message.
 465  *      The rate information is held in the inet_peer entries now.
 466  *      This function is generic and could be used for other purposes
 467  *      too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
 468  *
 469  *      Note that the same inet_peer fields are modified by functions in
 470  *      route.c too, but these work for packet destinations while xrlim_allow
 471  *      works for icmp destinations. This means the rate limiting information
 472  *      for one "ip object" is shared - and these ICMPs are twice limited:
 473  *      by source and by destination.
 474  *
 475  *      RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
 476  *                        SHOULD allow setting of rate limits
 477  *
 478  *      Shared between ICMPv4 and ICMPv6.
 479  */
 480 #define XRLIM_BURST_FACTOR 6
 481 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
 482 {
 483         unsigned long now, token;
 484         bool rc = false;
 485
 486         if (!peer)
 487                 return true;
 488
 489         token = peer->rate_tokens;
 490         now = jiffies;
 491         token += now - peer->rate_last;
 492         peer->rate_last = now;
 493         if (token > XRLIM_BURST_FACTOR * timeout)
 494                 token = XRLIM_BURST_FACTOR * timeout;
 495         if (token >= timeout) {
 496                 token -= timeout;
 497                 rc = true;
 498         }
 499         peer->rate_tokens = token;
 500         return rc;
 501 }
 502 EXPORT_SYMBOL(inet_peer_xrlim_allow);