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 * Version: $Id: inetpeer.c,v 1.1.1.1 2007/08/03 18:53:51 Exp $ 7 * 8 * Authors: Andrey V. Savochkin <saw@msu.ru> 9 */ 10 11#include <linux/module.h> 12#include <linux/types.h> 13#include <linux/slab.h> 14#include <linux/interrupt.h> 15#include <linux/spinlock.h> 16#include <linux/random.h> 17#include <linux/timer.h> 18#include <linux/time.h> 19#include <linux/kernel.h> 20#include <linux/mm.h> 21#include <linux/net.h> 22#include <net/ip.h> 23#include <net/inetpeer.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 uses a constant ID and do not use this code (see 36 * 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 write lock held. 57 * 2. Nodes may disappear from the tree only with the pool write lock held 58 * AND reference count being 0. 59 * 3. Nodes appears and disappears from unused node list only under 60 * "inet_peer_unused_lock". 61 * 4. Global variable peer_total is modified under the pool lock. 62 * 5. struct inet_peer fields modification: 63 * avl_left, avl_right, avl_parent, avl_height: pool lock 64 * unused_next, unused_prevp: unused node list lock 65 * refcnt: atomically against modifications on other CPU; 66 * usually under some other lock to prevent node disappearing 67 * dtime: unused node list lock 68 * v4daddr: unchangeable 69 * ip_id_count: idlock 70 */ 71 72/* Exported for inet_getid inline function. */ 73DEFINE_SPINLOCK(inet_peer_idlock); 74 75static struct kmem_cache *peer_cachep __read_mostly; 76 77#define node_height(x) x->avl_height 78static struct inet_peer peer_fake_node = { 79 .avl_left = &peer_fake_node, 80 .avl_right = &peer_fake_node, 81 .avl_height = 0 82}; 83#define peer_avl_empty (&peer_fake_node) 84static struct inet_peer *peer_root = peer_avl_empty; 85static DEFINE_RWLOCK(peer_pool_lock); 86#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */ 87 88static int peer_total; 89/* Exported for sysctl_net_ipv4. */ 90int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more 91 * aggressively at this stage */ 92int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */ 93int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */ 94int inet_peer_gc_mintime __read_mostly = 10 * HZ; 95int inet_peer_gc_maxtime __read_mostly = 120 * HZ; 96 97static struct inet_peer *inet_peer_unused_head; 98static struct inet_peer **inet_peer_unused_tailp = &inet_peer_unused_head; 99static DEFINE_SPINLOCK(inet_peer_unused_lock); 100 101static void peer_check_expire(unsigned long dummy); 102static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0); 103 104 105/* Called from ip_output.c:ip_init */ 106void __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, NULL); 127 128 /* All the timers, started at system startup tend 129 to synchronize. Perturb it a bit. 130 */ 131 peer_periodic_timer.expires = jiffies 132 + net_random() % inet_peer_gc_maxtime 133 + inet_peer_gc_maxtime; 134 add_timer(&peer_periodic_timer); 135} 136 137/* Called with or without local BH being disabled. */ 138static void unlink_from_unused(struct inet_peer *p) 139{ 140 spin_lock_bh(&inet_peer_unused_lock); 141 if (p->unused_prevp != NULL) { 142 /* On unused list. */ 143 *p->unused_prevp = p->unused_next; 144 if (p->unused_next != NULL) 145 p->unused_next->unused_prevp = p->unused_prevp; 146 else 147 inet_peer_unused_tailp = p->unused_prevp; 148 p->unused_prevp = NULL; /* mark it as removed */ 149 } 150 spin_unlock_bh(&inet_peer_unused_lock); 151} 152 153/* 154 * Called with local BH disabled and the pool lock held. 155 * _stack is known to be NULL or not at compile time, 156 * so compiler will optimize the if (_stack) tests. 157 */ 158#define lookup(_daddr,_stack) \ 159({ \ 160 struct inet_peer *u, **v; \ 161 if (_stack) { \ 162 stackptr = _stack; \ 163 *stackptr++ = &peer_root; \ 164 } \ 165 for (u = peer_root; u != peer_avl_empty; ) { \ 166 if (_daddr == u->v4daddr) \ 167 break; \ 168 if ((__force __u32)_daddr < (__force __u32)u->v4daddr) \ 169 v = &u->avl_left; \ 170 else \ 171 v = &u->avl_right; \ 172 if (_stack) \ 173 *stackptr++ = v; \ 174 u = *v; \ 175 } \ 176 u; \ 177}) 178 179/* Called with local BH disabled and the pool write lock held. */ 180#define lookup_rightempty(start) \ 181({ \ 182 struct inet_peer *u, **v; \ 183 *stackptr++ = &start->avl_left; \ 184 v = &start->avl_left; \ 185 for (u = *v; u->avl_right != peer_avl_empty; ) { \ 186 v = &u->avl_right; \ 187 *stackptr++ = v; \ 188 u = *v; \ 189 } \ 190 u; \ 191}) 192 193/* Called with local BH disabled and the pool write lock held. 194 * Variable names are the proof of operation correctness. 195 * Look into mm/map_avl.c for more detail description of the ideas. */ 196static void peer_avl_rebalance(struct inet_peer **stack[], 197 struct inet_peer ***stackend) 198{ 199 struct inet_peer **nodep, *node, *l, *r; 200 int lh, rh; 201 202 while (stackend > stack) { 203 nodep = *--stackend; 204 node = *nodep; 205 l = node->avl_left; 206 r = node->avl_right; 207 lh = node_height(l); 208 rh = node_height(r); 209 if (lh > rh + 1) { /* l: RH+2 */ 210 struct inet_peer *ll, *lr, *lrl, *lrr; 211 int lrh; 212 ll = l->avl_left; 213 lr = l->avl_right; 214 lrh = node_height(lr); 215 if (lrh <= node_height(ll)) { /* ll: RH+1 */ 216 node->avl_left = lr; /* lr: RH or RH+1 */ 217 node->avl_right = r; /* r: RH */ 218 node->avl_height = lrh + 1; /* RH+1 or RH+2 */ 219 l->avl_left = ll; /* ll: RH+1 */ 220 l->avl_right = node; /* node: RH+1 or RH+2 */ 221 l->avl_height = node->avl_height + 1; 222 *nodep = l; 223 } else { /* ll: RH, lr: RH+1 */ 224 lrl = lr->avl_left; /* lrl: RH or RH-1 */ 225 lrr = lr->avl_right; /* lrr: RH or RH-1 */ 226 node->avl_left = lrr; /* lrr: RH or RH-1 */ 227 node->avl_right = r; /* r: RH */ 228 node->avl_height = rh + 1; /* node: RH+1 */ 229 l->avl_left = ll; /* ll: RH */ 230 l->avl_right = lrl; /* lrl: RH or RH-1 */ 231 l->avl_height = rh + 1; /* l: RH+1 */ 232 lr->avl_left = l; /* l: RH+1 */ 233 lr->avl_right = node; /* node: RH+1 */ 234 lr->avl_height = rh + 2; 235 *nodep = lr; 236 } 237 } else if (rh > lh + 1) { /* r: LH+2 */ 238 struct inet_peer *rr, *rl, *rlr, *rll; 239 int rlh; 240 rr = r->avl_right; 241 rl = r->avl_left; 242 rlh = node_height(rl); 243 if (rlh <= node_height(rr)) { /* rr: LH+1 */ 244 node->avl_right = rl; /* rl: LH or LH+1 */ 245 node->avl_left = l; /* l: LH */ 246 node->avl_height = rlh + 1; /* LH+1 or LH+2 */ 247 r->avl_right = rr; /* rr: LH+1 */ 248 r->avl_left = node; /* node: LH+1 or LH+2 */ 249 r->avl_height = node->avl_height + 1; 250 *nodep = r; 251 } else { /* rr: RH, rl: RH+1 */ 252 rlr = rl->avl_right; /* rlr: LH or LH-1 */ 253 rll = rl->avl_left; /* rll: LH or LH-1 */ 254 node->avl_right = rll; /* rll: LH or LH-1 */ 255 node->avl_left = l; /* l: LH */ 256 node->avl_height = lh + 1; /* node: LH+1 */ 257 r->avl_right = rr; /* rr: LH */ 258 r->avl_left = rlr; /* rlr: LH or LH-1 */ 259 r->avl_height = lh + 1; /* r: LH+1 */ 260 rl->avl_right = r; /* r: LH+1 */ 261 rl->avl_left = node; /* node: LH+1 */ 262 rl->avl_height = lh + 2; 263 *nodep = rl; 264 } 265 } else { 266 node->avl_height = (lh > rh ? lh : rh) + 1; 267 } 268 } 269} 270 271/* Called with local BH disabled and the pool write lock held. */ 272#define link_to_pool(n) \ 273do { \ 274 n->avl_height = 1; \ 275 n->avl_left = peer_avl_empty; \ 276 n->avl_right = peer_avl_empty; \ 277 **--stackptr = n; \ 278 peer_avl_rebalance(stack, stackptr); \ 279} while(0) 280 281/* May be called with local BH enabled. */ 282static void unlink_from_pool(struct inet_peer *p) 283{ 284 int do_free; 285 286 do_free = 0; 287 288 write_lock_bh(&peer_pool_lock); 289 /* Check the reference counter. It was artificially incremented by 1 290 * in cleanup() function to prevent sudden disappearing. If the 291 * reference count is still 1 then the node is referenced only as `p' 292 * here and from the pool. So under the exclusive pool lock it's safe 293 * to remove the node and free it later. */ 294 if (atomic_read(&p->refcnt) == 1) { 295 struct inet_peer **stack[PEER_MAXDEPTH]; 296 struct inet_peer ***stackptr, ***delp; 297 if (lookup(p->v4daddr, stack) != p) 298 BUG(); 299 delp = stackptr - 1; /* *delp[0] == p */ 300 if (p->avl_left == peer_avl_empty) { 301 *delp[0] = p->avl_right; 302 --stackptr; 303 } else { 304 /* look for a node to insert instead of p */ 305 struct inet_peer *t; 306 t = lookup_rightempty(p); 307 BUG_ON(*stackptr[-1] != t); 308 **--stackptr = t->avl_left; 309 /* t is removed, t->v4daddr > x->v4daddr for any 310 * x in p->avl_left subtree. 311 * Put t in the old place of p. */ 312 *delp[0] = t; 313 t->avl_left = p->avl_left; 314 t->avl_right = p->avl_right; 315 t->avl_height = p->avl_height; 316 BUG_ON(delp[1] != &p->avl_left); 317 delp[1] = &t->avl_left; /* was &p->avl_left */ 318 } 319 peer_avl_rebalance(stack, stackptr); 320 peer_total--; 321 do_free = 1; 322 } 323 write_unlock_bh(&peer_pool_lock); 324 325 if (do_free) 326 kmem_cache_free(peer_cachep, p); 327 else 328 /* The node is used again. Decrease the reference counter 329 * back. The loop "cleanup -> unlink_from_unused 330 * -> unlink_from_pool -> putpeer -> link_to_unused 331 * -> cleanup (for the same node)" 332 * doesn't really exist because the entry will have a 333 * recent deletion time and will not be cleaned again soon. */ 334 inet_putpeer(p); 335} 336 337/* May be called with local BH enabled. */ 338static int cleanup_once(unsigned long ttl) 339{ 340 struct inet_peer *p; 341 342 /* Remove the first entry from the list of unused nodes. */ 343 spin_lock_bh(&inet_peer_unused_lock); 344 p = inet_peer_unused_head; 345 if (p != NULL) { 346 __u32 delta = (__u32)jiffies - p->dtime; 347 if (delta < ttl) { 348 /* Do not prune fresh entries. */ 349 spin_unlock_bh(&inet_peer_unused_lock); 350 return -1; 351 } 352 inet_peer_unused_head = p->unused_next; 353 if (p->unused_next != NULL) 354 p->unused_next->unused_prevp = p->unused_prevp; 355 else 356 inet_peer_unused_tailp = p->unused_prevp; 357 p->unused_prevp = NULL; /* mark as not on the list */ 358 /* Grab an extra reference to prevent node disappearing 359 * before unlink_from_pool() call. */ 360 atomic_inc(&p->refcnt); 361 } 362 spin_unlock_bh(&inet_peer_unused_lock); 363 364 if (p == NULL) 365 /* It means that the total number of USED entries has 366 * grown over inet_peer_threshold. It shouldn't really 367 * happen because of entry limits in route cache. */ 368 return -1; 369 370 unlink_from_pool(p); 371 return 0; 372} 373 374/* Called with or without local BH being disabled. */ 375struct inet_peer *inet_getpeer(__be32 daddr, int create) 376{ 377 struct inet_peer *p, *n; 378 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr; 379 380 /* Look up for the address quickly. */ 381 read_lock_bh(&peer_pool_lock); 382 p = lookup(daddr, NULL); 383 if (p != peer_avl_empty) 384 atomic_inc(&p->refcnt); 385 read_unlock_bh(&peer_pool_lock); 386 387 if (p != peer_avl_empty) { 388 /* The existing node has been found. */ 389 /* Remove the entry from unused list if it was there. */ 390 unlink_from_unused(p); 391 return p; 392 } 393 394 if (!create) 395 return NULL; 396 397 /* Allocate the space outside the locked region. */ 398 n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC); 399 if (n == NULL) 400 return NULL; 401 n->v4daddr = daddr; 402 atomic_set(&n->refcnt, 1); 403 atomic_set(&n->rid, 0); 404 n->ip_id_count = secure_ip_id(daddr); 405 n->tcp_ts_stamp = 0; 406 407 write_lock_bh(&peer_pool_lock); 408 /* Check if an entry has suddenly appeared. */ 409 p = lookup(daddr, stack); 410 if (p != peer_avl_empty) 411 goto out_free; 412 413 /* Link the node. */ 414 link_to_pool(n); 415 n->unused_prevp = NULL; /* not on the list */ 416 peer_total++; 417 write_unlock_bh(&peer_pool_lock); 418 419 if (peer_total >= inet_peer_threshold) 420 /* Remove one less-recently-used entry. */ 421 cleanup_once(0); 422 423 return n; 424 425out_free: 426 /* The appropriate node is already in the pool. */ 427 atomic_inc(&p->refcnt); 428 write_unlock_bh(&peer_pool_lock); 429 /* Remove the entry from unused list if it was there. */ 430 unlink_from_unused(p); 431 /* Free preallocated the preallocated node. */ 432 kmem_cache_free(peer_cachep, n); 433 return p; 434} 435 436/* Called with local BH disabled. */ 437static void peer_check_expire(unsigned long dummy) 438{ 439 unsigned long now = jiffies; 440 int ttl; 441 442 if (peer_total >= inet_peer_threshold) 443 ttl = inet_peer_minttl; 444 else 445 ttl = inet_peer_maxttl 446 - (inet_peer_maxttl - inet_peer_minttl) / HZ * 447 peer_total / inet_peer_threshold * HZ; 448 while (!cleanup_once(ttl)) { 449 if (jiffies != now) 450 break; 451 } 452 453 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime 454 * interval depending on the total number of entries (more entries, 455 * less interval). */ 456 if (peer_total >= inet_peer_threshold) 457 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime; 458 else 459 peer_periodic_timer.expires = jiffies 460 + inet_peer_gc_maxtime 461 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ * 462 peer_total / inet_peer_threshold * HZ; 463 add_timer(&peer_periodic_timer); 464} 465 466void inet_putpeer(struct inet_peer *p) 467{ 468 spin_lock_bh(&inet_peer_unused_lock); 469 if (atomic_dec_and_test(&p->refcnt)) { 470 p->unused_prevp = inet_peer_unused_tailp; 471 p->unused_next = NULL; 472 *inet_peer_unused_tailp = p; 473 inet_peer_unused_tailp = &p->unused_next; 474 p->dtime = (__u32)jiffies; 475 } 476 spin_unlock_bh(&inet_peer_unused_lock); 477} 478