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/cache.h>
  10 #include <linux/module.h>
  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/timer.h>
  17 #include <linux/time.h>
  18 #include <linux/kernel.h>
  19 #include <linux/mm.h>
  20 #include <linux/net.h>
  21 #include <linux/workqueue.h>
  22 #include <net/ip.h>
  23 #include <net/inetpeer.h>
  24 #include <net/secure_seq.h>
  25
  26 /*
  27  *  Theory of operations.
  28  *  We keep one entry for each peer IP address.  The nodes contains long-living
  29  *  information about the peer which doesn't depend on routes.
  30  *
  31  *  Nodes are removed only when reference counter goes to 0.
  32  *  When it's happened the node may be removed when a sufficient amount of
  33  *  time has been passed since its last use.  The less-recently-used entry can
  34  *  also be removed if the pool is overloaded i.e. if the total amount of
  35  *  entries is greater-or-equal than the threshold.
  36  *
  37  *  Node pool is organised as an RB tree.
  38  *  Such an implementation has been chosen not just for fun.  It's a way to
  39  *  prevent easy and efficient DoS attacks by creating hash collisions.  A huge
  40  *  amount of long living nodes in a single hash slot would significantly delay
  41  *  lookups performed with disabled BHs.
  42  *
  43  *  Serialisation issues.
  44  *  1.  Nodes may appear in the tree only with the pool lock held.
  45  *  2.  Nodes may disappear from the tree only with the pool lock held
  46  *      AND reference count being 0.
  47  *  3.  Global variable peer_total is modified under the pool lock.
  48  *  4.  struct inet_peer fields modification:
  49  *              rb_node: pool lock
  50  *              refcnt: atomically against modifications on other CPU;
  51  *                 usually under some other lock to prevent node disappearing
  52  *              daddr: unchangeable
  53  */
  54
  55 static struct kmem_cache *peer_cachep __ro_after_init;
  56
  57 void inet_peer_base_init(struct inet_peer_base *bp)
  58 {
  59         bp->rb_root = RB_ROOT;
  60         seqlock_init(&bp->lock);
  61         bp->total = 0;
  62 }
  63 EXPORT_SYMBOL_GPL(inet_peer_base_init);
  64
  65 #define PEER_MAX_GC 32
  66
  67 /* Exported for sysctl_net_ipv4.  */
  68 int inet_peer_threshold __read_mostly;  /* start to throw entries more
  69                                          * aggressively at this stage */
  70 int inet_peer_minttl __read_mostly = 120 * HZ;  /* TTL under high load: 120 sec */
  71 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;      /* usual time to live: 10 min */
  72
  73 /* Called from ip_output.c:ip_init  */
  74 void __init inet_initpeers(void)
  75 {
  76         u64 nr_entries;
  77
  78          /* 1% of physical memory */
  79         nr_entries = div64_ul((u64)totalram_pages() << PAGE_SHIFT,
  80                               100 * L1_CACHE_ALIGN(sizeof(struct inet_peer)));
  81
  82         inet_peer_threshold = clamp_val(nr_entries, 4096, 65536 + 128);
  83
  84         peer_cachep = kmem_cache_create("inet_peer_cache",
  85                         sizeof(struct inet_peer),
  86                         0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
  87                         NULL);
  88 }
  89
  90 /* Called with rcu_read_lock() or base->lock held */
  91 static struct inet_peer *lookup(const struct inetpeer_addr *daddr,
  92                                 struct inet_peer_base *base,
  93                                 unsigned int seq,
  94                                 struct inet_peer *gc_stack[],
  95                                 unsigned int *gc_cnt,
  96                                 struct rb_node **parent_p,
  97                                 struct rb_node ***pp_p)
  98 {
  99         struct rb_node **pp, *parent, *next;
 100         struct inet_peer *p;
 101
 102         pp = &base->rb_root.rb_node;
 103         parent = NULL;
 104         while (1) {
 105                 int cmp;
 106
 107                 next = rcu_dereference_raw(*pp);
 108                 if (!next)
 109                         break;
 110                 parent = next;
 111                 p = rb_entry(parent, struct inet_peer, rb_node);
 112                 cmp = inetpeer_addr_cmp(daddr, &p->daddr);
 113                 if (cmp == 0) {
 114                         if (!refcount_inc_not_zero(&p->refcnt))
 115                                 break;
 116                         return p;
 117                 }
 118                 if (gc_stack) {
 119                         if (*gc_cnt < PEER_MAX_GC)
 120                                 gc_stack[(*gc_cnt)++] = p;
 121                 } else if (unlikely(read_seqretry(&base->lock, seq))) {
 122                         break;
 123                 }
 124                 if (cmp == -1)
 125                         pp = &next->rb_left;
 126                 else
 127                         pp = &next->rb_right;
 128         }
 129         *parent_p = parent;
 130         *pp_p = pp;
 131         return NULL;
 132 }
 133
 134 static void inetpeer_free_rcu(struct rcu_head *head)
 135 {
 136         kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
 137 }
 138
 139 /* perform garbage collect on all items stacked during a lookup */
 140 static void inet_peer_gc(struct inet_peer_base *base,
 141                          struct inet_peer *gc_stack[],
 142                          unsigned int gc_cnt)
 143 {
 144         struct inet_peer *p;
 145         __u32 delta, ttl;
 146         int i;
 147
 148         if (base->total >= inet_peer_threshold)
 149                 ttl = 0; /* be aggressive */
 150         else
 151                 ttl = inet_peer_maxttl
 152                                 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
 153                                         base->total / inet_peer_threshold * HZ;
 154         for (i = 0; i < gc_cnt; i++) {
 155                 p = gc_stack[i];
 156
 157                 /* The READ_ONCE() pairs with the WRITE_ONCE()
 158                  * in inet_putpeer()
 159                  */
 160                 delta = (__u32)jiffies - READ_ONCE(p->dtime);
 161
 162                 if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
 163                         gc_stack[i] = NULL;
 164         }
 165         for (i = 0; i < gc_cnt; i++) {
 166                 p = gc_stack[i];
 167                 if (p) {
 168                         rb_erase(&p->rb_node, &base->rb_root);
 169                         base->total--;
 170                         call_rcu(&p->rcu, inetpeer_free_rcu);
 171                 }
 172         }
 173 }
 174
 175 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
 176                                const struct inetpeer_addr *daddr,
 177                                int create)
 178 {
 179         struct inet_peer *p, *gc_stack[PEER_MAX_GC];
 180         struct rb_node **pp, *parent;
 181         unsigned int gc_cnt, seq;
 182         int invalidated;
 183
 184         /* Attempt a lockless lookup first.
 185          * Because of a concurrent writer, we might not find an existing entry.
 186          */
 187         rcu_read_lock();
 188         seq = read_seqbegin(&base->lock);
 189         p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
 190         invalidated = read_seqretry(&base->lock, seq);
 191         rcu_read_unlock();
 192
 193         if (p)
 194                 return p;
 195
 196         /* If no writer did a change during our lookup, we can return early. */
 197         if (!create && !invalidated)
 198                 return NULL;
 199
 200         /* retry an exact lookup, taking the lock before.
 201          * At least, nodes should be hot in our cache.
 202          */
 203         parent = NULL;
 204         write_seqlock_bh(&base->lock);
 205
 206         gc_cnt = 0;
 207         p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
 208         if (!p && create) {
 209                 p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
 210                 if (p) {
 211                         p->daddr = *daddr;
 212                         p->dtime = (__u32)jiffies;
 213                         refcount_set(&p->refcnt, 2);
 214                         atomic_set(&p->rid, 0);
 215                         p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
 216                         p->rate_tokens = 0;
 217                         p->n_redirects = 0;
 218                         /* 60*HZ is arbitrary, but chosen enough high so that the first
 219                          * calculation of tokens is at its maximum.
 220                          */
 221                         p->rate_last = jiffies - 60*HZ;
 222
 223                         rb_link_node(&p->rb_node, parent, pp);
 224                         rb_insert_color(&p->rb_node, &base->rb_root);
 225                         base->total++;
 226                 }
 227         }
 228         if (gc_cnt)
 229                 inet_peer_gc(base, gc_stack, gc_cnt);
 230         write_sequnlock_bh(&base->lock);
 231
 232         return p;
 233 }
 234 EXPORT_SYMBOL_GPL(inet_getpeer);
 235
 236 void inet_putpeer(struct inet_peer *p)
 237 {
 238         /* The WRITE_ONCE() pairs with itself (we run lockless)
 239          * and the READ_ONCE() in inet_peer_gc()
 240          */
 241         WRITE_ONCE(p->dtime, (__u32)jiffies);
 242
 243         if (refcount_dec_and_test(&p->refcnt))
 244                 call_rcu(&p->rcu, inetpeer_free_rcu);
 245 }
 246 EXPORT_SYMBOL_GPL(inet_putpeer);
 247
 248 /*
 249  *      Check transmit rate limitation for given message.
 250  *      The rate information is held in the inet_peer entries now.
 251  *      This function is generic and could be used for other purposes
 252  *      too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
 253  *
 254  *      Note that the same inet_peer fields are modified by functions in
 255  *      route.c too, but these work for packet destinations while xrlim_allow
 256  *      works for icmp destinations. This means the rate limiting information
 257  *      for one "ip object" is shared - and these ICMPs are twice limited:
 258  *      by source and by destination.
 259  *
 260  *      RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
 261  *                        SHOULD allow setting of rate limits
 262  *
 263  *      Shared between ICMPv4 and ICMPv6.
 264  */
 265 #define XRLIM_BURST_FACTOR 6
 266 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
 267 {
 268         unsigned long now, token;
 269         bool rc = false;
 270
 271         if (!peer)
 272                 return true;
 273
 274         token = peer->rate_tokens;
 275         now = jiffies;
 276         token += now - peer->rate_last;
 277         peer->rate_last = now;
 278         if (token > XRLIM_BURST_FACTOR * timeout)
 279                 token = XRLIM_BURST_FACTOR * timeout;
 280         if (token >= timeout) {
 281                 token -= timeout;
 282                 rc = true;
 283         }
 284         peer->rate_tokens = token;
 285         return rc;
 286 }
 287 EXPORT_SYMBOL(inet_peer_xrlim_allow);
 288
 289 void inetpeer_invalidate_tree(struct inet_peer_base *base)
 290 {
 291         struct rb_node *p = rb_first(&base->rb_root);
 292
 293         while (p) {
 294                 struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
 295
 296                 p = rb_next(p);
 297                 rb_erase(&peer->rb_node, &base->rb_root);
 298                 inet_putpeer(peer);
 299                 cond_resched();
 300         }
 301
 302         base->total = 0;
 303 }
 304 EXPORT_SYMBOL(inetpeer_invalidate_tree);