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 <linux/workqueue.h> 21 #include <net/ip.h> 22 #include <net/inetpeer.h> 23 #include <net/secure_seq.h> 24 25 /* 26 * Theory of operations. 27 * We keep one entry for each peer IP address. The nodes contains long-living 28 * information about the peer which doesn't depend on routes. 29 * At this moment this information consists only of ID field for the next 30 * outgoing IP packet. This field is incremented with each packet as encoded 31 * in inet_getid() function (include/net/inetpeer.h). 32 * At the moment of writing this notes identifier of IP packets is generated 33 * to be unpredictable using this code only for packets subjected 34 * (actually or potentially) to defragmentation. I.e. DF packets less than 35 * PMTU in size when local fragmentation is disabled use a constant ID and do 36 * not use this code (see ip_select_ident() in include/net/ip.h). 37 * 38 * Route cache entries hold references to our nodes. 39 * New cache entries get references via lookup by destination IP address in 40 * the avl tree. The reference is grabbed only when it's needed i.e. only 41 * when we try to output IP packet which needs an unpredictable ID (see 42 * __ip_select_ident() in net/ipv4/route.c). 43 * Nodes are removed only when reference counter goes to 0. 44 * When it's happened the node may be removed when a sufficient amount of 45 * time has been passed since its last use. The less-recently-used entry can 46 * also be removed if the pool is overloaded i.e. if the total amount of 47 * entries is greater-or-equal than the threshold. 48 * 49 * Node pool is organised as an AVL tree. 50 * Such an implementation has been chosen not just for fun. It's a way to 51 * prevent easy and efficient DoS attacks by creating hash collisions. A huge 52 * amount of long living nodes in a single hash slot would significantly delay 53 * lookups performed with disabled BHs. 54 * 55 * Serialisation issues. 56 * 1. Nodes may appear in the tree only with the pool lock held. 57 * 2. Nodes may disappear from the tree only with the pool lock held 58 * AND reference count being 0. 59 * 3. Global variable peer_total is modified under the pool lock. 60 * 4. struct inet_peer fields modification: 61 * avl_left, avl_right, avl_parent, avl_height: pool lock 62 * refcnt: atomically against modifications on other CPU; 63 * usually under some other lock to prevent node disappearing 64 * daddr: unchangeable 65 * ip_id_count: atomic value (no lock needed) 66 */ 67 68 static struct kmem_cache *peer_cachep __read_mostly; 69 70 static LIST_HEAD(gc_list); 71 static const int gc_delay = 60 * HZ; 72 static struct delayed_work gc_work; 73 static DEFINE_SPINLOCK(gc_lock); 74 75 #define node_height(x) x->avl_height 76 77 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node) 78 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node) 79 static const struct inet_peer peer_fake_node = { 80 .avl_left = peer_avl_empty_rcu, 81 .avl_right = peer_avl_empty_rcu, 82 .avl_height = 0 83 }; 84 85 void inet_peer_base_init(struct inet_peer_base *bp) 86 { 87 bp->root = peer_avl_empty_rcu; 88 seqlock_init(&bp->lock); 89 bp->flush_seq = ~0U; 90 bp->total = 0; 91 } 92 EXPORT_SYMBOL_GPL(inet_peer_base_init); 93 94 static atomic_t v4_seq = ATOMIC_INIT(0); 95 static atomic_t v6_seq = ATOMIC_INIT(0); 96 97 static atomic_t *inetpeer_seq_ptr(int family) 98 { 99 return (family == AF_INET ? &v4_seq : &v6_seq); 100 } 101 102 static inline void flush_check(struct inet_peer_base *base, int family) 103 { 104 atomic_t *fp = inetpeer_seq_ptr(family); 105 106 if (unlikely(base->flush_seq != atomic_read(fp))) { 107 inetpeer_invalidate_tree(base); 108 base->flush_seq = atomic_read(fp); 109 } 110 } 111 112 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */ 113 114 /* Exported for sysctl_net_ipv4. */ 115 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more 116 * aggressively at this stage */ 117 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */ 118 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */ 119 120 static void inetpeer_gc_worker(struct work_struct *work) 121 { 122 struct inet_peer *p, *n, *c; 123 LIST_HEAD(list); 124 125 spin_lock_bh(&gc_lock); 126 list_replace_init(&gc_list, &list); 127 spin_unlock_bh(&gc_lock); 128 129 if (list_empty(&list)) 130 return; 131 132 list_for_each_entry_safe(p, n, &list, gc_list) { 133 134 if (need_resched()) 135 cond_resched(); 136 137 c = rcu_dereference_protected(p->avl_left, 1); 138 if (c != peer_avl_empty) { 139 list_add_tail(&c->gc_list, &list); 140 p->avl_left = peer_avl_empty_rcu; 141 } 142 143 c = rcu_dereference_protected(p->avl_right, 1); 144 if (c != peer_avl_empty) { 145 list_add_tail(&c->gc_list, &list); 146 p->avl_right = peer_avl_empty_rcu; 147 } 148 149 n = list_entry(p->gc_list.next, struct inet_peer, gc_list); 150 151 if (!atomic_read(&p->refcnt)) { 152 list_del(&p->gc_list); 153 kmem_cache_free(peer_cachep, p); 154 } 155 } 156 157 if (list_empty(&list)) 158 return; 159 160 spin_lock_bh(&gc_lock); 161 list_splice(&list, &gc_list); 162 spin_unlock_bh(&gc_lock); 163 164 schedule_delayed_work(&gc_work, gc_delay); 165 } 166 167 /* Called from ip_output.c:ip_init */ 168 void __init inet_initpeers(void) 169 { 170 struct sysinfo si; 171 172 /* Use the straight interface to information about memory. */ 173 si_meminfo(&si); 174 /* The values below were suggested by Alexey Kuznetsov 175 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values 176 * myself. --SAW 177 */ 178 if (si.totalram <= (32768*1024)/PAGE_SIZE) 179 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */ 180 if (si.totalram <= (16384*1024)/PAGE_SIZE) 181 inet_peer_threshold >>= 1; /* about 512KB */ 182 if (si.totalram <= (8192*1024)/PAGE_SIZE) 183 inet_peer_threshold >>= 2; /* about 128KB */ 184 185 peer_cachep = kmem_cache_create("inet_peer_cache", 186 sizeof(struct inet_peer), 187 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, 188 NULL); 189 190 INIT_DEFERRABLE_WORK(&gc_work, inetpeer_gc_worker); 191 } 192 193 static int addr_compare(const struct inetpeer_addr *a, 194 const struct inetpeer_addr *b) 195 { 196 int i, n = (a->family == AF_INET ? 1 : 4); 197 198 for (i = 0; i < n; i++) { 199 if (a->addr.a6[i] == b->addr.a6[i]) 200 continue; 201 if ((__force u32)a->addr.a6[i] < (__force u32)b->addr.a6[i]) 202 return -1; 203 return 1; 204 } 205 206 return 0; 207 } 208 209 #define rcu_deref_locked(X, BASE) \ 210 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock)) 211 212 /* 213 * Called with local BH disabled and the pool lock held. 214 */ 215 #define lookup(_daddr, _stack, _base) \ 216 ({ \ 217 struct inet_peer *u; \ 218 struct inet_peer __rcu **v; \ 219 \ 220 stackptr = _stack; \ 221 *stackptr++ = &_base->root; \ 222 for (u = rcu_deref_locked(_base->root, _base); \ 223 u != peer_avl_empty;) { \ 224 int cmp = addr_compare(_daddr, &u->daddr); \ 225 if (cmp == 0) \ 226 break; \ 227 if (cmp == -1) \ 228 v = &u->avl_left; \ 229 else \ 230 v = &u->avl_right; \ 231 *stackptr++ = v; \ 232 u = rcu_deref_locked(*v, _base); \ 233 } \ 234 u; \ 235 }) 236 237 /* 238 * Called with rcu_read_lock() 239 * Because we hold no lock against a writer, its quite possible we fall 240 * in an endless loop. 241 * But every pointer we follow is guaranteed to be valid thanks to RCU. 242 * We exit from this function if number of links exceeds PEER_MAXDEPTH 243 */ 244 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr, 245 struct inet_peer_base *base) 246 { 247 struct inet_peer *u = rcu_dereference(base->root); 248 int count = 0; 249 250 while (u != peer_avl_empty) { 251 int cmp = addr_compare(daddr, &u->daddr); 252 if (cmp == 0) { 253 /* Before taking a reference, check if this entry was 254 * deleted (refcnt=-1) 255 */ 256 if (!atomic_add_unless(&u->refcnt, 1, -1)) 257 u = NULL; 258 return u; 259 } 260 if (cmp == -1) 261 u = rcu_dereference(u->avl_left); 262 else 263 u = rcu_dereference(u->avl_right); 264 if (unlikely(++count == PEER_MAXDEPTH)) 265 break; 266 } 267 return NULL; 268 } 269 270 /* Called with local BH disabled and the pool lock held. */ 271 #define lookup_rightempty(start, base) \ 272 ({ \ 273 struct inet_peer *u; \ 274 struct inet_peer __rcu **v; \ 275 *stackptr++ = &start->avl_left; \ 276 v = &start->avl_left; \ 277 for (u = rcu_deref_locked(*v, base); \ 278 u->avl_right != peer_avl_empty_rcu;) { \ 279 v = &u->avl_right; \ 280 *stackptr++ = v; \ 281 u = rcu_deref_locked(*v, base); \ 282 } \ 283 u; \ 284 }) 285 286 /* Called with local BH disabled and the pool lock held. 287 * Variable names are the proof of operation correctness. 288 * Look into mm/map_avl.c for more detail description of the ideas. 289 */ 290 static void peer_avl_rebalance(struct inet_peer __rcu **stack[], 291 struct inet_peer __rcu ***stackend, 292 struct inet_peer_base *base) 293 { 294 struct inet_peer __rcu **nodep; 295 struct inet_peer *node, *l, *r; 296 int lh, rh; 297 298 while (stackend > stack) { 299 nodep = *--stackend; 300 node = rcu_deref_locked(*nodep, base); 301 l = rcu_deref_locked(node->avl_left, base); 302 r = rcu_deref_locked(node->avl_right, base); 303 lh = node_height(l); 304 rh = node_height(r); 305 if (lh > rh + 1) { /* l: RH+2 */ 306 struct inet_peer *ll, *lr, *lrl, *lrr; 307 int lrh; 308 ll = rcu_deref_locked(l->avl_left, base); 309 lr = rcu_deref_locked(l->avl_right, base); 310 lrh = node_height(lr); 311 if (lrh <= node_height(ll)) { /* ll: RH+1 */ 312 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */ 313 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */ 314 node->avl_height = lrh + 1; /* RH+1 or RH+2 */ 315 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */ 316 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */ 317 l->avl_height = node->avl_height + 1; 318 RCU_INIT_POINTER(*nodep, l); 319 } else { /* ll: RH, lr: RH+1 */ 320 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */ 321 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */ 322 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */ 323 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */ 324 node->avl_height = rh + 1; /* node: RH+1 */ 325 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */ 326 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */ 327 l->avl_height = rh + 1; /* l: RH+1 */ 328 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */ 329 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */ 330 lr->avl_height = rh + 2; 331 RCU_INIT_POINTER(*nodep, lr); 332 } 333 } else if (rh > lh + 1) { /* r: LH+2 */ 334 struct inet_peer *rr, *rl, *rlr, *rll; 335 int rlh; 336 rr = rcu_deref_locked(r->avl_right, base); 337 rl = rcu_deref_locked(r->avl_left, base); 338 rlh = node_height(rl); 339 if (rlh <= node_height(rr)) { /* rr: LH+1 */ 340 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */ 341 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */ 342 node->avl_height = rlh + 1; /* LH+1 or LH+2 */ 343 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */ 344 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */ 345 r->avl_height = node->avl_height + 1; 346 RCU_INIT_POINTER(*nodep, r); 347 } else { /* rr: RH, rl: RH+1 */ 348 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */ 349 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */ 350 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */ 351 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */ 352 node->avl_height = lh + 1; /* node: LH+1 */ 353 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */ 354 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */ 355 r->avl_height = lh + 1; /* r: LH+1 */ 356 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */ 357 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */ 358 rl->avl_height = lh + 2; 359 RCU_INIT_POINTER(*nodep, rl); 360 } 361 } else { 362 node->avl_height = (lh > rh ? lh : rh) + 1; 363 } 364 } 365 } 366 367 /* Called with local BH disabled and the pool lock held. */ 368 #define link_to_pool(n, base) \ 369 do { \ 370 n->avl_height = 1; \ 371 n->avl_left = peer_avl_empty_rcu; \ 372 n->avl_right = peer_avl_empty_rcu; \ 373 /* lockless readers can catch us now */ \ 374 rcu_assign_pointer(**--stackptr, n); \ 375 peer_avl_rebalance(stack, stackptr, base); \ 376 } while (0) 377 378 static void inetpeer_free_rcu(struct rcu_head *head) 379 { 380 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu)); 381 } 382 383 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base, 384 struct inet_peer __rcu **stack[PEER_MAXDEPTH]) 385 { 386 struct inet_peer __rcu ***stackptr, ***delp; 387 388 if (lookup(&p->daddr, stack, base) != p) 389 BUG(); 390 delp = stackptr - 1; /* *delp[0] == p */ 391 if (p->avl_left == peer_avl_empty_rcu) { 392 *delp[0] = p->avl_right; 393 --stackptr; 394 } else { 395 /* look for a node to insert instead of p */ 396 struct inet_peer *t; 397 t = lookup_rightempty(p, base); 398 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t); 399 **--stackptr = t->avl_left; 400 /* t is removed, t->daddr > x->daddr for any 401 * x in p->avl_left subtree. 402 * Put t in the old place of p. */ 403 RCU_INIT_POINTER(*delp[0], t); 404 t->avl_left = p->avl_left; 405 t->avl_right = p->avl_right; 406 t->avl_height = p->avl_height; 407 BUG_ON(delp[1] != &p->avl_left); 408 delp[1] = &t->avl_left; /* was &p->avl_left */ 409 } 410 peer_avl_rebalance(stack, stackptr, base); 411 base->total--; 412 call_rcu(&p->rcu, inetpeer_free_rcu); 413 } 414 415 /* perform garbage collect on all items stacked during a lookup */ 416 static int inet_peer_gc(struct inet_peer_base *base, 417 struct inet_peer __rcu **stack[PEER_MAXDEPTH], 418 struct inet_peer __rcu ***stackptr) 419 { 420 struct inet_peer *p, *gchead = NULL; 421 __u32 delta, ttl; 422 int cnt = 0; 423 424 if (base->total >= inet_peer_threshold) 425 ttl = 0; /* be aggressive */ 426 else 427 ttl = inet_peer_maxttl 428 - (inet_peer_maxttl - inet_peer_minttl) / HZ * 429 base->total / inet_peer_threshold * HZ; 430 stackptr--; /* last stack slot is peer_avl_empty */ 431 while (stackptr > stack) { 432 stackptr--; 433 p = rcu_deref_locked(**stackptr, base); 434 if (atomic_read(&p->refcnt) == 0) { 435 smp_rmb(); 436 delta = (__u32)jiffies - p->dtime; 437 if (delta >= ttl && 438 atomic_cmpxchg(&p->refcnt, 0, -1) == 0) { 439 p->gc_next = gchead; 440 gchead = p; 441 } 442 } 443 } 444 while ((p = gchead) != NULL) { 445 gchead = p->gc_next; 446 cnt++; 447 unlink_from_pool(p, base, stack); 448 } 449 return cnt; 450 } 451 452 struct inet_peer *inet_getpeer(struct inet_peer_base *base, 453 const struct inetpeer_addr *daddr, 454 int create) 455 { 456 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr; 457 struct inet_peer *p; 458 unsigned int sequence; 459 int invalidated, gccnt = 0; 460 461 flush_check(base, daddr->family); 462 463 /* Attempt a lockless lookup first. 464 * Because of a concurrent writer, we might not find an existing entry. 465 */ 466 rcu_read_lock(); 467 sequence = read_seqbegin(&base->lock); 468 p = lookup_rcu(daddr, base); 469 invalidated = read_seqretry(&base->lock, sequence); 470 rcu_read_unlock(); 471 472 if (p) 473 return p; 474 475 /* If no writer did a change during our lookup, we can return early. */ 476 if (!create && !invalidated) 477 return NULL; 478 479 /* retry an exact lookup, taking the lock before. 480 * At least, nodes should be hot in our cache. 481 */ 482 write_seqlock_bh(&base->lock); 483 relookup: 484 p = lookup(daddr, stack, base); 485 if (p != peer_avl_empty) { 486 atomic_inc(&p->refcnt); 487 write_sequnlock_bh(&base->lock); 488 return p; 489 } 490 if (!gccnt) { 491 gccnt = inet_peer_gc(base, stack, stackptr); 492 if (gccnt && create) 493 goto relookup; 494 } 495 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL; 496 if (p) { 497 p->daddr = *daddr; 498 atomic_set(&p->refcnt, 1); 499 atomic_set(&p->rid, 0); 500 atomic_set(&p->ip_id_count, 501 (daddr->family == AF_INET) ? 502 secure_ip_id(daddr->addr.a4) : 503 secure_ipv6_id(daddr->addr.a6)); 504 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW; 505 p->rate_tokens = 0; 506 /* 60*HZ is arbitrary, but chosen enough high so that the first 507 * calculation of tokens is at its maximum. 508 */ 509 p->rate_last = jiffies - 60*HZ; 510 INIT_LIST_HEAD(&p->gc_list); 511 512 /* Link the node. */ 513 link_to_pool(p, base); 514 base->total++; 515 } 516 write_sequnlock_bh(&base->lock); 517 518 return p; 519 } 520 EXPORT_SYMBOL_GPL(inet_getpeer); 521 522 void inet_putpeer(struct inet_peer *p) 523 { 524 p->dtime = (__u32)jiffies; 525 smp_mb__before_atomic_dec(); 526 atomic_dec(&p->refcnt); 527 } 528 EXPORT_SYMBOL_GPL(inet_putpeer); 529 530 /* 531 * Check transmit rate limitation for given message. 532 * The rate information is held in the inet_peer entries now. 533 * This function is generic and could be used for other purposes 534 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov. 535 * 536 * Note that the same inet_peer fields are modified by functions in 537 * route.c too, but these work for packet destinations while xrlim_allow 538 * works for icmp destinations. This means the rate limiting information 539 * for one "ip object" is shared - and these ICMPs are twice limited: 540 * by source and by destination. 541 * 542 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate 543 * SHOULD allow setting of rate limits 544 * 545 * Shared between ICMPv4 and ICMPv6. 546 */ 547 #define XRLIM_BURST_FACTOR 6 548 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout) 549 { 550 unsigned long now, token; 551 bool rc = false; 552 553 if (!peer) 554 return true; 555 556 token = peer->rate_tokens; 557 now = jiffies; 558 token += now - peer->rate_last; 559 peer->rate_last = now; 560 if (token > XRLIM_BURST_FACTOR * timeout) 561 token = XRLIM_BURST_FACTOR * timeout; 562 if (token >= timeout) { 563 token -= timeout; 564 rc = true; 565 } 566 peer->rate_tokens = token; 567 return rc; 568 } 569 EXPORT_SYMBOL(inet_peer_xrlim_allow); 570 571 static void inetpeer_inval_rcu(struct rcu_head *head) 572 { 573 struct inet_peer *p = container_of(head, struct inet_peer, gc_rcu); 574 575 spin_lock_bh(&gc_lock); 576 list_add_tail(&p->gc_list, &gc_list); 577 spin_unlock_bh(&gc_lock); 578 579 schedule_delayed_work(&gc_work, gc_delay); 580 } 581 582 void inetpeer_invalidate_tree(struct inet_peer_base *base) 583 { 584 struct inet_peer *root; 585 586 write_seqlock_bh(&base->lock); 587 588 root = rcu_deref_locked(base->root, base); 589 if (root != peer_avl_empty) { 590 base->root = peer_avl_empty_rcu; 591 base->total = 0; 592 call_rcu(&root->gc_rcu, inetpeer_inval_rcu); 593 } 594 595 write_sequnlock_bh(&base->lock); 596 } 597 EXPORT_SYMBOL(inetpeer_invalidate_tree); 598