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 #include <net/secure_seq.h> 23 24 /* 25 * Theory of operations. 26 * We keep one entry for each peer IP address. The nodes contains long-living 27 * information about the peer which doesn't depend on routes. 28 * At this moment this information consists only of ID field for the next 29 * outgoing IP packet. This field is incremented with each packet as encoded 30 * in inet_getid() function (include/net/inetpeer.h). 31 * At the moment of writing this notes identifier of IP packets is generated 32 * to be unpredictable using this code only for packets subjected 33 * (actually or potentially) to defragmentation. I.e. DF packets less than 34 * PMTU in size uses a constant ID and do not use this code (see 35 * ip_select_ident() in include/net/ip.h). 36 * 37 * Route cache entries hold references to our nodes. 38 * New cache entries get references via lookup by destination IP address in 39 * the avl tree. The reference is grabbed only when it's needed i.e. only 40 * when we try to output IP packet which needs an unpredictable ID (see 41 * __ip_select_ident() in net/ipv4/route.c). 42 * Nodes are removed only when reference counter goes to 0. 43 * When it's happened the node may be removed when a sufficient amount of 44 * time has been passed since its last use. The less-recently-used entry can 45 * also be removed if the pool is overloaded i.e. if the total amount of 46 * entries is greater-or-equal than the threshold. 47 * 48 * Node pool is organised as an AVL tree. 49 * Such an implementation has been chosen not just for fun. It's a way to 50 * prevent easy and efficient DoS attacks by creating hash collisions. A huge 51 * amount of long living nodes in a single hash slot would significantly delay 52 * lookups performed with disabled BHs. 53 * 54 * Serialisation issues. 55 * 1. Nodes may appear in the tree only with the pool lock held. 56 * 2. Nodes may disappear from the tree only with the pool lock held 57 * AND reference count being 0. 58 * 3. Global variable peer_total is modified under the pool lock. 59 * 4. struct inet_peer fields modification: 60 * avl_left, avl_right, avl_parent, avl_height: pool lock 61 * refcnt: atomically against modifications on other CPU; 62 * usually under some other lock to prevent node disappearing 63 * daddr: unchangeable 64 * ip_id_count: atomic value (no lock needed) 65 */ 66 67 static struct kmem_cache *peer_cachep __read_mostly; 68 69 #define node_height(x) x->avl_height 70 71 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node) 72 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node) 73 static const struct inet_peer peer_fake_node = { 74 .avl_left = peer_avl_empty_rcu, 75 .avl_right = peer_avl_empty_rcu, 76 .avl_height = 0 77 }; 78 79 struct inet_peer_base { 80 struct inet_peer __rcu *root; 81 seqlock_t lock; 82 int total; 83 }; 84 85 static struct inet_peer_base v4_peers = { 86 .root = peer_avl_empty_rcu, 87 .lock = __SEQLOCK_UNLOCKED(v4_peers.lock), 88 .total = 0, 89 }; 90 91 static struct inet_peer_base v6_peers = { 92 .root = peer_avl_empty_rcu, 93 .lock = __SEQLOCK_UNLOCKED(v6_peers.lock), 94 .total = 0, 95 }; 96 97 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */ 98 99 /* Exported for sysctl_net_ipv4. */ 100 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more 101 * aggressively at this stage */ 102 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */ 103 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */ 104 105 106 /* Called from ip_output.c:ip_init */ 107 void __init inet_initpeers(void) 108 { 109 struct sysinfo si; 110 111 /* Use the straight interface to information about memory. */ 112 si_meminfo(&si); 113 /* The values below were suggested by Alexey Kuznetsov 114 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values 115 * myself. --SAW 116 */ 117 if (si.totalram <= (32768*1024)/PAGE_SIZE) 118 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */ 119 if (si.totalram <= (16384*1024)/PAGE_SIZE) 120 inet_peer_threshold >>= 1; /* about 512KB */ 121 if (si.totalram <= (8192*1024)/PAGE_SIZE) 122 inet_peer_threshold >>= 2; /* about 128KB */ 123 124 peer_cachep = kmem_cache_create("inet_peer_cache", 125 sizeof(struct inet_peer), 126 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, 127 NULL); 128 129 } 130 131 static int addr_compare(const struct inetpeer_addr *a, 132 const struct inetpeer_addr *b) 133 { 134 int i, n = (a->family == AF_INET ? 1 : 4); 135 136 for (i = 0; i < n; i++) { 137 if (a->addr.a6[i] == b->addr.a6[i]) 138 continue; 139 if ((__force u32)a->addr.a6[i] < (__force u32)b->addr.a6[i]) 140 return -1; 141 return 1; 142 } 143 144 return 0; 145 } 146 147 #define rcu_deref_locked(X, BASE) \ 148 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock)) 149 150 /* 151 * Called with local BH disabled and the pool lock held. 152 */ 153 #define lookup(_daddr, _stack, _base) \ 154 ({ \ 155 struct inet_peer *u; \ 156 struct inet_peer __rcu **v; \ 157 \ 158 stackptr = _stack; \ 159 *stackptr++ = &_base->root; \ 160 for (u = rcu_deref_locked(_base->root, _base); \ 161 u != peer_avl_empty; ) { \ 162 int cmp = addr_compare(_daddr, &u->daddr); \ 163 if (cmp == 0) \ 164 break; \ 165 if (cmp == -1) \ 166 v = &u->avl_left; \ 167 else \ 168 v = &u->avl_right; \ 169 *stackptr++ = v; \ 170 u = rcu_deref_locked(*v, _base); \ 171 } \ 172 u; \ 173 }) 174 175 /* 176 * Called with rcu_read_lock() 177 * Because we hold no lock against a writer, its quite possible we fall 178 * in an endless loop. 179 * But every pointer we follow is guaranteed to be valid thanks to RCU. 180 * We exit from this function if number of links exceeds PEER_MAXDEPTH 181 */ 182 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr, 183 struct inet_peer_base *base) 184 { 185 struct inet_peer *u = rcu_dereference(base->root); 186 int count = 0; 187 188 while (u != peer_avl_empty) { 189 int cmp = addr_compare(daddr, &u->daddr); 190 if (cmp == 0) { 191 /* Before taking a reference, check if this entry was 192 * deleted (refcnt=-1) 193 */ 194 if (!atomic_add_unless(&u->refcnt, 1, -1)) 195 u = NULL; 196 return u; 197 } 198 if (cmp == -1) 199 u = rcu_dereference(u->avl_left); 200 else 201 u = rcu_dereference(u->avl_right); 202 if (unlikely(++count == PEER_MAXDEPTH)) 203 break; 204 } 205 return NULL; 206 } 207 208 /* Called with local BH disabled and the pool lock held. */ 209 #define lookup_rightempty(start, base) \ 210 ({ \ 211 struct inet_peer *u; \ 212 struct inet_peer __rcu **v; \ 213 *stackptr++ = &start->avl_left; \ 214 v = &start->avl_left; \ 215 for (u = rcu_deref_locked(*v, base); \ 216 u->avl_right != peer_avl_empty_rcu; ) { \ 217 v = &u->avl_right; \ 218 *stackptr++ = v; \ 219 u = rcu_deref_locked(*v, base); \ 220 } \ 221 u; \ 222 }) 223 224 /* Called with local BH disabled and the pool lock held. 225 * Variable names are the proof of operation correctness. 226 * Look into mm/map_avl.c for more detail description of the ideas. 227 */ 228 static void peer_avl_rebalance(struct inet_peer __rcu **stack[], 229 struct inet_peer __rcu ***stackend, 230 struct inet_peer_base *base) 231 { 232 struct inet_peer __rcu **nodep; 233 struct inet_peer *node, *l, *r; 234 int lh, rh; 235 236 while (stackend > stack) { 237 nodep = *--stackend; 238 node = rcu_deref_locked(*nodep, base); 239 l = rcu_deref_locked(node->avl_left, base); 240 r = rcu_deref_locked(node->avl_right, base); 241 lh = node_height(l); 242 rh = node_height(r); 243 if (lh > rh + 1) { /* l: RH+2 */ 244 struct inet_peer *ll, *lr, *lrl, *lrr; 245 int lrh; 246 ll = rcu_deref_locked(l->avl_left, base); 247 lr = rcu_deref_locked(l->avl_right, base); 248 lrh = node_height(lr); 249 if (lrh <= node_height(ll)) { /* ll: RH+1 */ 250 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */ 251 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */ 252 node->avl_height = lrh + 1; /* RH+1 or RH+2 */ 253 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */ 254 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */ 255 l->avl_height = node->avl_height + 1; 256 RCU_INIT_POINTER(*nodep, l); 257 } else { /* ll: RH, lr: RH+1 */ 258 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */ 259 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */ 260 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */ 261 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */ 262 node->avl_height = rh + 1; /* node: RH+1 */ 263 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */ 264 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */ 265 l->avl_height = rh + 1; /* l: RH+1 */ 266 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */ 267 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */ 268 lr->avl_height = rh + 2; 269 RCU_INIT_POINTER(*nodep, lr); 270 } 271 } else if (rh > lh + 1) { /* r: LH+2 */ 272 struct inet_peer *rr, *rl, *rlr, *rll; 273 int rlh; 274 rr = rcu_deref_locked(r->avl_right, base); 275 rl = rcu_deref_locked(r->avl_left, base); 276 rlh = node_height(rl); 277 if (rlh <= node_height(rr)) { /* rr: LH+1 */ 278 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */ 279 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */ 280 node->avl_height = rlh + 1; /* LH+1 or LH+2 */ 281 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */ 282 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */ 283 r->avl_height = node->avl_height + 1; 284 RCU_INIT_POINTER(*nodep, r); 285 } else { /* rr: RH, rl: RH+1 */ 286 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */ 287 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */ 288 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */ 289 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */ 290 node->avl_height = lh + 1; /* node: LH+1 */ 291 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */ 292 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */ 293 r->avl_height = lh + 1; /* r: LH+1 */ 294 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */ 295 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */ 296 rl->avl_height = lh + 2; 297 RCU_INIT_POINTER(*nodep, rl); 298 } 299 } else { 300 node->avl_height = (lh > rh ? lh : rh) + 1; 301 } 302 } 303 } 304 305 /* Called with local BH disabled and the pool lock held. */ 306 #define link_to_pool(n, base) \ 307 do { \ 308 n->avl_height = 1; \ 309 n->avl_left = peer_avl_empty_rcu; \ 310 n->avl_right = peer_avl_empty_rcu; \ 311 /* lockless readers can catch us now */ \ 312 rcu_assign_pointer(**--stackptr, n); \ 313 peer_avl_rebalance(stack, stackptr, base); \ 314 } while (0) 315 316 static void inetpeer_free_rcu(struct rcu_head *head) 317 { 318 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu)); 319 } 320 321 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base, 322 struct inet_peer __rcu **stack[PEER_MAXDEPTH]) 323 { 324 struct inet_peer __rcu ***stackptr, ***delp; 325 326 if (lookup(&p->daddr, stack, base) != p) 327 BUG(); 328 delp = stackptr - 1; /* *delp[0] == p */ 329 if (p->avl_left == peer_avl_empty_rcu) { 330 *delp[0] = p->avl_right; 331 --stackptr; 332 } else { 333 /* look for a node to insert instead of p */ 334 struct inet_peer *t; 335 t = lookup_rightempty(p, base); 336 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t); 337 **--stackptr = t->avl_left; 338 /* t is removed, t->daddr > x->daddr for any 339 * x in p->avl_left subtree. 340 * Put t in the old place of p. */ 341 RCU_INIT_POINTER(*delp[0], t); 342 t->avl_left = p->avl_left; 343 t->avl_right = p->avl_right; 344 t->avl_height = p->avl_height; 345 BUG_ON(delp[1] != &p->avl_left); 346 delp[1] = &t->avl_left; /* was &p->avl_left */ 347 } 348 peer_avl_rebalance(stack, stackptr, base); 349 base->total--; 350 call_rcu(&p->rcu, inetpeer_free_rcu); 351 } 352 353 static struct inet_peer_base *family_to_base(int family) 354 { 355 return family == AF_INET ? &v4_peers : &v6_peers; 356 } 357 358 /* perform garbage collect on all items stacked during a lookup */ 359 static int inet_peer_gc(struct inet_peer_base *base, 360 struct inet_peer __rcu **stack[PEER_MAXDEPTH], 361 struct inet_peer __rcu ***stackptr) 362 { 363 struct inet_peer *p, *gchead = NULL; 364 __u32 delta, ttl; 365 int cnt = 0; 366 367 if (base->total >= inet_peer_threshold) 368 ttl = 0; /* be aggressive */ 369 else 370 ttl = inet_peer_maxttl 371 - (inet_peer_maxttl - inet_peer_minttl) / HZ * 372 base->total / inet_peer_threshold * HZ; 373 stackptr--; /* last stack slot is peer_avl_empty */ 374 while (stackptr > stack) { 375 stackptr--; 376 p = rcu_deref_locked(**stackptr, base); 377 if (atomic_read(&p->refcnt) == 0) { 378 smp_rmb(); 379 delta = (__u32)jiffies - p->dtime; 380 if (delta >= ttl && 381 atomic_cmpxchg(&p->refcnt, 0, -1) == 0) { 382 p->gc_next = gchead; 383 gchead = p; 384 } 385 } 386 } 387 while ((p = gchead) != NULL) { 388 gchead = p->gc_next; 389 cnt++; 390 unlink_from_pool(p, base, stack); 391 } 392 return cnt; 393 } 394 395 struct inet_peer *inet_getpeer(const struct inetpeer_addr *daddr, int create) 396 { 397 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr; 398 struct inet_peer_base *base = family_to_base(daddr->family); 399 struct inet_peer *p; 400 unsigned int sequence; 401 int invalidated, gccnt = 0; 402 403 /* Attempt a lockless lookup first. 404 * Because of a concurrent writer, we might not find an existing entry. 405 */ 406 rcu_read_lock(); 407 sequence = read_seqbegin(&base->lock); 408 p = lookup_rcu(daddr, base); 409 invalidated = read_seqretry(&base->lock, sequence); 410 rcu_read_unlock(); 411 412 if (p) 413 return p; 414 415 /* If no writer did a change during our lookup, we can return early. */ 416 if (!create && !invalidated) 417 return NULL; 418 419 /* retry an exact lookup, taking the lock before. 420 * At least, nodes should be hot in our cache. 421 */ 422 write_seqlock_bh(&base->lock); 423 relookup: 424 p = lookup(daddr, stack, base); 425 if (p != peer_avl_empty) { 426 atomic_inc(&p->refcnt); 427 write_sequnlock_bh(&base->lock); 428 return p; 429 } 430 if (!gccnt) { 431 gccnt = inet_peer_gc(base, stack, stackptr); 432 if (gccnt && create) 433 goto relookup; 434 } 435 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL; 436 if (p) { 437 p->daddr = *daddr; 438 atomic_set(&p->refcnt, 1); 439 atomic_set(&p->rid, 0); 440 atomic_set(&p->ip_id_count, 441 (daddr->family == AF_INET) ? 442 secure_ip_id(daddr->addr.a4) : 443 secure_ipv6_id(daddr->addr.a6)); 444 p->tcp_ts_stamp = 0; 445 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW; 446 p->rate_tokens = 0; 447 p->rate_last = 0; 448 p->pmtu_expires = 0; 449 p->pmtu_orig = 0; 450 p->redirect_genid = 0; 451 memset(&p->redirect_learned, 0, sizeof(p->redirect_learned)); 452 453 454 /* Link the node. */ 455 link_to_pool(p, base); 456 base->total++; 457 } 458 write_sequnlock_bh(&base->lock); 459 460 return p; 461 } 462 EXPORT_SYMBOL_GPL(inet_getpeer); 463 464 void inet_putpeer(struct inet_peer *p) 465 { 466 p->dtime = (__u32)jiffies; 467 smp_mb__before_atomic_dec(); 468 atomic_dec(&p->refcnt); 469 } 470 EXPORT_SYMBOL_GPL(inet_putpeer); 471 472 /* 473 * Check transmit rate limitation for given message. 474 * The rate information is held in the inet_peer entries now. 475 * This function is generic and could be used for other purposes 476 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov. 477 * 478 * Note that the same inet_peer fields are modified by functions in 479 * route.c too, but these work for packet destinations while xrlim_allow 480 * works for icmp destinations. This means the rate limiting information 481 * for one "ip object" is shared - and these ICMPs are twice limited: 482 * by source and by destination. 483 * 484 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate 485 * SHOULD allow setting of rate limits 486 * 487 * Shared between ICMPv4 and ICMPv6. 488 */ 489 #define XRLIM_BURST_FACTOR 6 490 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout) 491 { 492 unsigned long now, token; 493 bool rc = false; 494 495 if (!peer) 496 return true; 497 498 token = peer->rate_tokens; 499 now = jiffies; 500 token += now - peer->rate_last; 501 peer->rate_last = now; 502 if (token > XRLIM_BURST_FACTOR * timeout) 503 token = XRLIM_BURST_FACTOR * timeout; 504 if (token >= timeout) { 505 token -= timeout; 506 rc = true; 507 } 508 peer->rate_tokens = token; 509 return rc; 510 } 511 EXPORT_SYMBOL(inet_peer_xrlim_allow); 512