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