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 2008/10/15 03:27:33 james26_jang Exp $ 7 * 8 * Authors: Andrey V. Savochkin <saw@msu.ru> 9 */ 10 11#include <linux/types.h> 12#include <linux/slab.h> 13#include <linux/interrupt.h> 14#include <linux/spinlock.h> 15#include <linux/random.h> 16#include <linux/sched.h> 17#include <linux/timer.h> 18#include <linux/time.h> 19#include <linux/kernel.h> 20#include <linux/mm.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 write lock held. 55 * 2. Nodes may disappear from the tree only with the pool write 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_next, unused_prevp: 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: idlock 68 */ 69 70/* Exported for inet_getid inline function. */ 71spinlock_t inet_peer_idlock = SPIN_LOCK_UNLOCKED; 72 73static kmem_cache_t *peer_cachep; 74 75#define node_height(x) x->avl_height 76static struct inet_peer peer_fake_node = { 77 avl_left : &peer_fake_node, 78 avl_right : &peer_fake_node, 79 avl_height : 0 80}; 81#define peer_avl_empty (&peer_fake_node) 82static struct inet_peer *peer_root = peer_avl_empty; 83static rwlock_t peer_pool_lock = RW_LOCK_UNLOCKED; 84#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */ 85 86static volatile int peer_total; 87/* Exported for sysctl_net_ipv4. */ 88int inet_peer_threshold = 65536 + 128; /* start to throw entries more 89 * aggressively at this stage */ 90int inet_peer_minttl = 120 * HZ; /* TTL under high load: 120 sec */ 91int inet_peer_maxttl = 10 * 60 * HZ; /* usual time to live: 10 min */ 92 93/* Exported for inet_putpeer inline function. */ 94struct inet_peer *inet_peer_unused_head, 95 **inet_peer_unused_tailp = &inet_peer_unused_head; 96spinlock_t inet_peer_unused_lock = SPIN_LOCK_UNLOCKED; 97#define PEER_MAX_CLEANUP_WORK 30 98 99static void peer_check_expire(unsigned long dummy); 100static struct timer_list peer_periodic_timer = 101 { { NULL, NULL }, 0, 0, &peer_check_expire }; 102 103/* Exported for sysctl_net_ipv4. */ 104int inet_peer_gc_mintime = 10 * HZ, 105 inet_peer_gc_maxtime = 120 * HZ; 106 107/* Called from ip_output.c:ip_init */ 108void __init inet_initpeers(void) 109{ 110 struct sysinfo si; 111 112 /* Use the straight interface to information about memory. */ 113 si_meminfo(&si); 114 /* The values below were suggested by Alexey Kuznetsov 115 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values 116 * myself. --SAW 117 */ 118 if (si.totalram <= (32768*1024)/PAGE_SIZE) 119 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */ 120 if (si.totalram <= (16384*1024)/PAGE_SIZE) 121 inet_peer_threshold >>= 1; /* about 512KB */ 122 if (si.totalram <= (8192*1024)/PAGE_SIZE) 123 inet_peer_threshold >>= 2; /* about 128KB */ 124 125 peer_cachep = kmem_cache_create("inet_peer_cache", 126 sizeof(struct inet_peer), 127 0, SLAB_HWCACHE_ALIGN, 128 NULL, NULL); 129 130 /* All the timers, started at system startup tend 131 to synchronize. Perturb it a bit. 132 */ 133 peer_periodic_timer.expires = jiffies 134 + net_random() % inet_peer_gc_maxtime 135 + inet_peer_gc_maxtime; 136 add_timer(&peer_periodic_timer); 137} 138 139/* Called with or without local BH being disabled. */ 140static void unlink_from_unused(struct inet_peer *p) 141{ 142 spin_lock_bh(&inet_peer_unused_lock); 143 if (p->unused_prevp != NULL) { 144 /* On unused list. */ 145 *p->unused_prevp = p->unused_next; 146 if (p->unused_next != NULL) 147 p->unused_next->unused_prevp = p->unused_prevp; 148 else 149 inet_peer_unused_tailp = p->unused_prevp; 150 p->unused_prevp = NULL; /* mark it as removed */ 151 } 152 spin_unlock_bh(&inet_peer_unused_lock); 153} 154 155/* Called with local BH disabled and the pool lock held. */ 156#define lookup(daddr) \ 157({ \ 158 struct inet_peer *u, **v; \ 159 stackptr = stack; \ 160 *stackptr++ = &peer_root; \ 161 for (u = peer_root; u != peer_avl_empty; ) { \ 162 if (daddr == u->v4daddr) \ 163 break; \ 164 if (daddr < u->v4daddr) \ 165 v = &u->avl_left; \ 166 else \ 167 v = &u->avl_right; \ 168 *stackptr++ = v; \ 169 u = *v; \ 170 } \ 171 u; \ 172}) 173 174/* Called with local BH disabled and the pool write lock held. */ 175#define lookup_rightempty(start) \ 176({ \ 177 struct inet_peer *u, **v; \ 178 *stackptr++ = &start->avl_left; \ 179 v = &start->avl_left; \ 180 for (u = *v; u->avl_right != peer_avl_empty; ) { \ 181 v = &u->avl_right; \ 182 *stackptr++ = v; \ 183 u = *v; \ 184 } \ 185 u; \ 186}) 187 188/* Called with local BH disabled and the pool write lock held. 189 * Variable names are the proof of operation correctness. 190 * Look into mm/map_avl.c for more detail description of the ideas. */ 191static void peer_avl_rebalance(struct inet_peer **stack[], 192 struct inet_peer ***stackend) 193{ 194 struct inet_peer **nodep, *node, *l, *r; 195 int lh, rh; 196 197 while (stackend > stack) { 198 nodep = *--stackend; 199 node = *nodep; 200 l = node->avl_left; 201 r = node->avl_right; 202 lh = node_height(l); 203 rh = node_height(r); 204 if (lh > rh + 1) { /* l: RH+2 */ 205 struct inet_peer *ll, *lr, *lrl, *lrr; 206 int lrh; 207 ll = l->avl_left; 208 lr = l->avl_right; 209 lrh = node_height(lr); 210 if (lrh <= node_height(ll)) { /* ll: RH+1 */ 211 node->avl_left = lr; /* lr: RH or RH+1 */ 212 node->avl_right = r; /* r: RH */ 213 node->avl_height = lrh + 1; /* RH+1 or RH+2 */ 214 l->avl_left = ll; /* ll: RH+1 */ 215 l->avl_right = node; /* node: RH+1 or RH+2 */ 216 l->avl_height = node->avl_height + 1; 217 *nodep = l; 218 } else { /* ll: RH, lr: RH+1 */ 219 lrl = lr->avl_left; /* lrl: RH or RH-1 */ 220 lrr = lr->avl_right; /* lrr: RH or RH-1 */ 221 node->avl_left = lrr; /* lrr: RH or RH-1 */ 222 node->avl_right = r; /* r: RH */ 223 node->avl_height = rh + 1; /* node: RH+1 */ 224 l->avl_left = ll; /* ll: RH */ 225 l->avl_right = lrl; /* lrl: RH or RH-1 */ 226 l->avl_height = rh + 1; /* l: RH+1 */ 227 lr->avl_left = l; /* l: RH+1 */ 228 lr->avl_right = node; /* node: RH+1 */ 229 lr->avl_height = rh + 2; 230 *nodep = lr; 231 } 232 } else if (rh > lh + 1) { /* r: LH+2 */ 233 struct inet_peer *rr, *rl, *rlr, *rll; 234 int rlh; 235 rr = r->avl_right; 236 rl = r->avl_left; 237 rlh = node_height(rl); 238 if (rlh <= node_height(rr)) { /* rr: LH+1 */ 239 node->avl_right = rl; /* rl: LH or LH+1 */ 240 node->avl_left = l; /* l: LH */ 241 node->avl_height = rlh + 1; /* LH+1 or LH+2 */ 242 r->avl_right = rr; /* rr: LH+1 */ 243 r->avl_left = node; /* node: LH+1 or LH+2 */ 244 r->avl_height = node->avl_height + 1; 245 *nodep = r; 246 } else { /* rr: RH, rl: RH+1 */ 247 rlr = rl->avl_right; /* rlr: LH or LH-1 */ 248 rll = rl->avl_left; /* rll: LH or LH-1 */ 249 node->avl_right = rll; /* rll: LH or LH-1 */ 250 node->avl_left = l; /* l: LH */ 251 node->avl_height = lh + 1; /* node: LH+1 */ 252 r->avl_right = rr; /* rr: LH */ 253 r->avl_left = rlr; /* rlr: LH or LH-1 */ 254 r->avl_height = lh + 1; /* r: LH+1 */ 255 rl->avl_right = r; /* r: LH+1 */ 256 rl->avl_left = node; /* node: LH+1 */ 257 rl->avl_height = lh + 2; 258 *nodep = rl; 259 } 260 } else { 261 node->avl_height = (lh > rh ? lh : rh) + 1; 262 } 263 } 264} 265 266/* Called with local BH disabled and the pool write lock held. */ 267#define link_to_pool(n) \ 268do { \ 269 n->avl_height = 1; \ 270 n->avl_left = peer_avl_empty; \ 271 n->avl_right = peer_avl_empty; \ 272 **--stackptr = n; \ 273 peer_avl_rebalance(stack, stackptr); \ 274} while(0) 275 276/* May be called with local BH enabled. */ 277static void unlink_from_pool(struct inet_peer *p) 278{ 279 int do_free; 280 281 do_free = 0; 282 283 write_lock_bh(&peer_pool_lock); 284 /* Check the reference counter. It was artificially incremented by 1 285 * in cleanup() function to prevent sudden disappearing. If the 286 * reference count is still 1 then the node is referenced only as `p' 287 * here and from the pool. So under the exclusive pool lock it's safe 288 * to remove the node and free it later. */ 289 if (atomic_read(&p->refcnt) == 1) { 290 struct inet_peer **stack[PEER_MAXDEPTH]; 291 struct inet_peer ***stackptr, ***delp; 292 if (lookup(p->v4daddr) != p) 293 BUG(); 294 delp = stackptr - 1; /* *delp[0] == p */ 295 if (p->avl_left == peer_avl_empty) { 296 *delp[0] = p->avl_right; 297 --stackptr; 298 } else { 299 /* look for a node to insert instead of p */ 300 struct inet_peer *t; 301 t = lookup_rightempty(p); 302 if (*stackptr[-1] != t) 303 BUG(); 304 **--stackptr = t->avl_left; 305 /* t is removed, t->v4daddr > x->v4daddr for any 306 * x in p->avl_left subtree. 307 * Put t in the old place of p. */ 308 *delp[0] = t; 309 t->avl_left = p->avl_left; 310 t->avl_right = p->avl_right; 311 t->avl_height = p->avl_height; 312 if (delp[1] != &p->avl_left) 313 BUG(); 314 delp[1] = &t->avl_left; /* was &p->avl_left */ 315 } 316 peer_avl_rebalance(stack, stackptr); 317 peer_total--; 318 do_free = 1; 319 } 320 write_unlock_bh(&peer_pool_lock); 321 322 if (do_free) 323 kmem_cache_free(peer_cachep, p); 324 else 325 /* The node is used again. Decrease the reference counter 326 * back. The loop "cleanup -> unlink_from_unused 327 * -> unlink_from_pool -> putpeer -> link_to_unused 328 * -> cleanup (for the same node)" 329 * doesn't really exist because the entry will have a 330 * recent deletion time and will not be cleaned again soon. */ 331 inet_putpeer(p); 332} 333 334/* May be called with local BH enabled. */ 335static int cleanup_once(unsigned long ttl) 336{ 337 struct inet_peer *p; 338 339 /* Remove the first entry from the list of unused nodes. */ 340 spin_lock_bh(&inet_peer_unused_lock); 341 p = inet_peer_unused_head; 342 if (p != NULL) { 343 if (time_after(p->dtime + ttl, jiffies)) { 344 /* Do not prune fresh entries. */ 345 spin_unlock_bh(&inet_peer_unused_lock); 346 return -1; 347 } 348 inet_peer_unused_head = p->unused_next; 349 if (p->unused_next != NULL) 350 p->unused_next->unused_prevp = p->unused_prevp; 351 else 352 inet_peer_unused_tailp = p->unused_prevp; 353 p->unused_prevp = NULL; /* mark as not on the list */ 354 /* Grab an extra reference to prevent node disappearing 355 * before unlink_from_pool() call. */ 356 atomic_inc(&p->refcnt); 357 } 358 spin_unlock_bh(&inet_peer_unused_lock); 359 360 if (p == NULL) 361 /* It means that the total number of USED entries has 362 * grown over inet_peer_threshold. It shouldn't really 363 * happen because of entry limits in route cache. */ 364 return -1; 365 366 unlink_from_pool(p); 367 return 0; 368} 369 370/* Called with or without local BH being disabled. */ 371struct inet_peer *inet_getpeer(__u32 daddr, int create) 372{ 373 struct inet_peer *p, *n; 374 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr; 375 376 /* Look up for the address quickly. */ 377 read_lock_bh(&peer_pool_lock); 378 p = lookup(daddr); 379 if (p != peer_avl_empty) 380 atomic_inc(&p->refcnt); 381 read_unlock_bh(&peer_pool_lock); 382 383 if (p != peer_avl_empty) { 384 /* The existing node has been found. */ 385 /* Remove the entry from unused list if it was there. */ 386 unlink_from_unused(p); 387 return p; 388 } 389 390 if (!create) 391 return NULL; 392 393 /* Allocate the space outside the locked region. */ 394 n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC); 395 if (n == NULL) 396 return NULL; 397 n->v4daddr = daddr; 398 atomic_set(&n->refcnt, 1); 399 n->ip_id_count = secure_ip_id(daddr); 400 n->tcp_ts_stamp = 0; 401 402 write_lock_bh(&peer_pool_lock); 403 /* Check if an entry has suddenly appeared. */ 404 p = lookup(daddr); 405 if (p != peer_avl_empty) 406 goto out_free; 407 408 /* Link the node. */ 409 link_to_pool(n); 410 n->unused_prevp = NULL; /* not on the list */ 411 peer_total++; 412 write_unlock_bh(&peer_pool_lock); 413 414 if (peer_total >= inet_peer_threshold) 415 /* Remove one less-recently-used entry. */ 416 cleanup_once(0); 417 418 return n; 419 420out_free: 421 /* The appropriate node is already in the pool. */ 422 atomic_inc(&p->refcnt); 423 write_unlock_bh(&peer_pool_lock); 424 /* Remove the entry from unused list if it was there. */ 425 unlink_from_unused(p); 426 /* Free preallocated the preallocated node. */ 427 kmem_cache_free(peer_cachep, n); 428 return p; 429} 430 431/* Called with local BH disabled. */ 432static void peer_check_expire(unsigned long dummy) 433{ 434 int i; 435 int ttl; 436 437 if (peer_total >= inet_peer_threshold) 438 ttl = inet_peer_minttl; 439 else 440 ttl = inet_peer_maxttl 441 - (inet_peer_maxttl - inet_peer_minttl) / HZ * 442 peer_total / inet_peer_threshold * HZ; 443 for (i = 0; i < PEER_MAX_CLEANUP_WORK && !cleanup_once(ttl); i++); 444 445 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime 446 * interval depending on the total number of entries (more entries, 447 * less interval). */ 448 peer_periodic_timer.expires = jiffies 449 + inet_peer_gc_maxtime 450 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ * 451 peer_total / inet_peer_threshold * HZ; 452 add_timer(&peer_periodic_timer); 453} 454