linux/net/ipv4/inetpeer.c
<<
>>
Prefs
   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 RB 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 *              rb_node: 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
  54static struct kmem_cache *peer_cachep __read_mostly;
  55
  56void inet_peer_base_init(struct inet_peer_base *bp)
  57{
  58        bp->rb_root = RB_ROOT;
  59        seqlock_init(&bp->lock);
  60        bp->total = 0;
  61}
  62EXPORT_SYMBOL_GPL(inet_peer_base_init);
  63
  64#define PEER_MAX_GC 32
  65
  66/* Exported for sysctl_net_ipv4.  */
  67int inet_peer_threshold __read_mostly = 65536 + 128;    /* start to throw entries more
  68                                         * aggressively at this stage */
  69int inet_peer_minttl __read_mostly = 120 * HZ;  /* TTL under high load: 120 sec */
  70int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;      /* usual time to live: 10 min */
  71
  72/* Called from ip_output.c:ip_init  */
  73void __init inet_initpeers(void)
  74{
  75        struct sysinfo si;
  76
  77        /* Use the straight interface to information about memory. */
  78        si_meminfo(&si);
  79        /* The values below were suggested by Alexey Kuznetsov
  80         * <kuznet@ms2.inr.ac.ru>.  I don't have any opinion about the values
  81         * myself.  --SAW
  82         */
  83        if (si.totalram <= (32768*1024)/PAGE_SIZE)
  84                inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
  85        if (si.totalram <= (16384*1024)/PAGE_SIZE)
  86                inet_peer_threshold >>= 1; /* about 512KB */
  87        if (si.totalram <= (8192*1024)/PAGE_SIZE)
  88                inet_peer_threshold >>= 2; /* about 128KB */
  89
  90        peer_cachep = kmem_cache_create("inet_peer_cache",
  91                        sizeof(struct inet_peer),
  92                        0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
  93                        NULL);
  94}
  95
  96/* Called with rcu_read_lock() or base->lock held */
  97static struct inet_peer *lookup(const struct inetpeer_addr *daddr,
  98                                struct inet_peer_base *base,
  99                                unsigned int seq,
 100                                struct inet_peer *gc_stack[],
 101                                unsigned int *gc_cnt,
 102                                struct rb_node **parent_p,
 103                                struct rb_node ***pp_p)
 104{
 105        struct rb_node **pp, *parent, *next;
 106        struct inet_peer *p;
 107
 108        pp = &base->rb_root.rb_node;
 109        parent = NULL;
 110        while (1) {
 111                int cmp;
 112
 113                next = rcu_dereference_raw(*pp);
 114                if (!next)
 115                        break;
 116                parent = next;
 117                p = rb_entry(parent, struct inet_peer, rb_node);
 118                cmp = inetpeer_addr_cmp(daddr, &p->daddr);
 119                if (cmp == 0) {
 120                        if (!refcount_inc_not_zero(&p->refcnt))
 121                                break;
 122                        return p;
 123                }
 124                if (gc_stack) {
 125                        if (*gc_cnt < PEER_MAX_GC)
 126                                gc_stack[(*gc_cnt)++] = p;
 127                } else if (unlikely(read_seqretry(&base->lock, seq))) {
 128                        break;
 129                }
 130                if (cmp == -1)
 131                        pp = &next->rb_left;
 132                else
 133                        pp = &next->rb_right;
 134        }
 135        *parent_p = parent;
 136        *pp_p = pp;
 137        return NULL;
 138}
 139
 140static void inetpeer_free_rcu(struct rcu_head *head)
 141{
 142        kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
 143}
 144
 145/* perform garbage collect on all items stacked during a lookup */
 146static void inet_peer_gc(struct inet_peer_base *base,
 147                         struct inet_peer *gc_stack[],
 148                         unsigned int gc_cnt)
 149{
 150        struct inet_peer *p;
 151        __u32 delta, ttl;
 152        int i;
 153
 154        if (base->total >= inet_peer_threshold)
 155                ttl = 0; /* be aggressive */
 156        else
 157                ttl = inet_peer_maxttl
 158                                - (inet_peer_maxttl - inet_peer_minttl) / HZ *
 159                                        base->total / inet_peer_threshold * HZ;
 160        for (i = 0; i < gc_cnt; i++) {
 161                p = gc_stack[i];
 162                delta = (__u32)jiffies - p->dtime;
 163                if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
 164                        gc_stack[i] = NULL;
 165        }
 166        for (i = 0; i < gc_cnt; i++) {
 167                p = gc_stack[i];
 168                if (p) {
 169                        rb_erase(&p->rb_node, &base->rb_root);
 170                        base->total--;
 171                        call_rcu(&p->rcu, inetpeer_free_rcu);
 172                }
 173        }
 174}
 175
 176struct inet_peer *inet_getpeer(struct inet_peer_base *base,
 177                               const struct inetpeer_addr *daddr,
 178                               int create)
 179{
 180        struct inet_peer *p, *gc_stack[PEER_MAX_GC];
 181        struct rb_node **pp, *parent;
 182        unsigned int gc_cnt, seq;
 183        int invalidated;
 184
 185        /* Attempt a lockless lookup first.
 186         * Because of a concurrent writer, we might not find an existing entry.
 187         */
 188        rcu_read_lock();
 189        seq = read_seqbegin(&base->lock);
 190        p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
 191        invalidated = read_seqretry(&base->lock, seq);
 192        rcu_read_unlock();
 193
 194        if (p)
 195                return p;
 196
 197        /* If no writer did a change during our lookup, we can return early. */
 198        if (!create && !invalidated)
 199                return NULL;
 200
 201        /* retry an exact lookup, taking the lock before.
 202         * At least, nodes should be hot in our cache.
 203         */
 204        parent = NULL;
 205        write_seqlock_bh(&base->lock);
 206
 207        gc_cnt = 0;
 208        p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
 209        if (!p && create) {
 210                p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
 211                if (p) {
 212                        p->daddr = *daddr;
 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                        /* 60*HZ is arbitrary, but chosen enough high so that the first
 218                         * calculation of tokens is at its maximum.
 219                         */
 220                        p->rate_last = jiffies - 60*HZ;
 221
 222                        rb_link_node(&p->rb_node, parent, pp);
 223                        rb_insert_color(&p->rb_node, &base->rb_root);
 224                        base->total++;
 225                }
 226        }
 227        if (gc_cnt)
 228                inet_peer_gc(base, gc_stack, gc_cnt);
 229        write_sequnlock_bh(&base->lock);
 230
 231        return p;
 232}
 233EXPORT_SYMBOL_GPL(inet_getpeer);
 234
 235void inet_putpeer(struct inet_peer *p)
 236{
 237        p->dtime = (__u32)jiffies;
 238
 239        if (refcount_dec_and_test(&p->refcnt))
 240                call_rcu(&p->rcu, inetpeer_free_rcu);
 241}
 242EXPORT_SYMBOL_GPL(inet_putpeer);
 243
 244/*
 245 *      Check transmit rate limitation for given message.
 246 *      The rate information is held in the inet_peer entries now.
 247 *      This function is generic and could be used for other purposes
 248 *      too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
 249 *
 250 *      Note that the same inet_peer fields are modified by functions in
 251 *      route.c too, but these work for packet destinations while xrlim_allow
 252 *      works for icmp destinations. This means the rate limiting information
 253 *      for one "ip object" is shared - and these ICMPs are twice limited:
 254 *      by source and by destination.
 255 *
 256 *      RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
 257 *                        SHOULD allow setting of rate limits
 258 *
 259 *      Shared between ICMPv4 and ICMPv6.
 260 */
 261#define XRLIM_BURST_FACTOR 6
 262bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
 263{
 264        unsigned long now, token;
 265        bool rc = false;
 266
 267        if (!peer)
 268                return true;
 269
 270        token = peer->rate_tokens;
 271        now = jiffies;
 272        token += now - peer->rate_last;
 273        peer->rate_last = now;
 274        if (token > XRLIM_BURST_FACTOR * timeout)
 275                token = XRLIM_BURST_FACTOR * timeout;
 276        if (token >= timeout) {
 277                token -= timeout;
 278                rc = true;
 279        }
 280        peer->rate_tokens = token;
 281        return rc;
 282}
 283EXPORT_SYMBOL(inet_peer_xrlim_allow);
 284
 285void inetpeer_invalidate_tree(struct inet_peer_base *base)
 286{
 287        struct rb_node *p = rb_first(&base->rb_root);
 288
 289        while (p) {
 290                struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
 291
 292                p = rb_next(p);
 293                rb_erase(&peer->rb_node, &base->rb_root);
 294                inet_putpeer(peer);
 295                cond_resched();
 296        }
 297
 298        base->total = 0;
 299}
 300EXPORT_SYMBOL(inetpeer_invalidate_tree);
 301