linux-2.4.20-nethashfix2.patch

   1 # This is a BitKeeper generated patch for the following project:
   2 # Project Name: Linux kernel tree
   3 # This patch format is intended for GNU patch command version 2.5 or higher.
   4 # This patch includes the following deltas:
   5 #                  ChangeSet    1.1226  -> 1.1229
   6 #       include/linux/sysctl.h  1.23    -> 1.24
   7 #        net/ipv6/tcp_ipv6.c    1.23    -> 1.24
   8 #        net/ipv6/af_inet6.c    1.9     -> 1.10
   9 #         net/ipv4/af_inet.c    1.10    -> 1.11
  10 #       net/ipv4/ip_fragment.c  1.6     -> 1.8
  11 #       net/ipv6/reassembly.c   1.6     -> 1.7
  12 #       include/net/transp_v6.h 1.1     -> 1.2
  13 #       net/ipv4/sysctl_net_ipv4.c      1.8     -> 1.9
  14 #
  15 # The following is the BitKeeper ChangeSet Log
  16 # --------------------------------------------
  17 # 03/05/27      davem@nuts.ninka.net    1.1227
  18 # [IPV4/IPV6]: Use Jenkins hash for fragment reassembly handling.
  19 # --------------------------------------------
  20 # 03/05/27      davem@nuts.ninka.net    1.1228
  21 # [IPV6]: Input full addresses into TCP_SYNQ hash function.
  22 # --------------------------------------------
  23 # 03/05/28      davem@nuts.ninka.net    1.1229
  24 # [IPV4]: Add sysctl to control ipfrag_secret_interval.
  25 # --------------------------------------------
  26 #
  27 diff -Nru a/include/linux/sysctl.h b/include/linux/sysctl.h
  28 --- a/include/linux/sysctl.h    Wed May 28 01:21:40 2003
  29 +++ b/include/linux/sysctl.h    Wed May 28 01:21:40 2003
  30 @@ -295,7 +295,8 @@
  31         NET_IPV4_NONLOCAL_BIND=88,
  32         NET_IPV4_ICMP_RATELIMIT=89,
  33         NET_IPV4_ICMP_RATEMASK=90,
  34 -       NET_TCP_TW_REUSE=91
  35 +       NET_TCP_TW_REUSE=91,
  36 +       NET_IPV4_IPFRAG_SECRET_INTERVAL=94
  37  };
  38
  39  enum {
  40 diff -Nru a/include/net/transp_v6.h b/include/net/transp_v6.h
  41 --- a/include/net/transp_v6.h   Wed May 28 01:21:40 2003
  42 +++ b/include/net/transp_v6.h   Wed May 28 01:21:40 2003
  43 @@ -15,6 +15,8 @@
  44
  45  struct flowi;
  46
  47 +extern void                            ipv6_frag_init(void);
  48 +
  49  extern void                            rawv6_init(void);
  50  extern void                            udpv6_init(void);
  51  extern void                            tcpv6_init(void);
  52 diff -Nru a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
  53 --- a/net/ipv4/af_inet.c        Wed May 28 01:21:40 2003
  54 +++ b/net/ipv4/af_inet.c        Wed May 28 01:21:40 2003
  55 @@ -1100,6 +1100,7 @@
  56         }
  57  }
  58
  59 +extern void ipfrag_init(void);
  60
  61  /*
  62   *     Called by socket.c on kernel startup.
  63 @@ -1196,6 +1197,9 @@
  64         proc_net_create ("tcp", 0, tcp_get_info);
  65         proc_net_create ("udp", 0, udp_get_info);
  66  #endif         /* CONFIG_PROC_FS */
  67 +
  68 +       ipfrag_init();
  69 +
  70         return 0;
  71  }
  72  module_init(inet_init);
  73 diff -Nru a/net/ipv4/ip_fragment.c b/net/ipv4/ip_fragment.c
  74 --- a/net/ipv4/ip_fragment.c    Wed May 28 01:21:40 2003
  75 +++ b/net/ipv4/ip_fragment.c    Wed May 28 01:21:40 2003
  76 @@ -31,6 +31,8 @@
  77  #include <linux/ip.h>
  78  #include <linux/icmp.h>
  79  #include <linux/netdevice.h>
  80 +#include <linux/jhash.h>
  81 +#include <linux/random.h>
  82  #include <net/sock.h>
  83  #include <net/ip.h>
  84  #include <net/icmp.h>
  85 @@ -97,6 +99,7 @@
  86  /* Per-bucket lock is easy to add now. */
  87  static struct ipq *ipq_hash[IPQ_HASHSZ];
  88  static rwlock_t ipfrag_lock = RW_LOCK_UNLOCKED;
  89 +static u32 ipfrag_hash_rnd;
  90  int ip_frag_nqueues = 0;
  91
  92  static __inline__ void __ipq_unlink(struct ipq *qp)
  93 @@ -116,21 +119,51 @@
  94         write_unlock(&ipfrag_lock);
  95  }
  96
  97 -/*
  98 - * Was:        ((((id) >> 1) ^ (saddr) ^ (daddr) ^ (prot)) & (IPQ_HASHSZ - 1))
  99 - *
 100 - * I see, I see evil hand of bigendian mafia. On Intel all the packets hit
 101 - * one hash bucket with this hash function. 8)
 102 - */
 103 -static __inline__ unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
 104 +static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
 105  {
 106 -       unsigned int h = saddr ^ daddr;
 107 -
 108 -       h ^= (h>>16)^id;
 109 -       h ^= (h>>8)^prot;
 110 -       return h & (IPQ_HASHSZ - 1);
 111 +       return jhash_3words((u32)id << 16 | prot, saddr, daddr,
 112 +                           ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
 113  }
 114
 115 +static struct timer_list ipfrag_secret_timer;
 116 +int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
 117 +
 118 +static void ipfrag_secret_rebuild(unsigned long dummy)
 119 +{
 120 +       unsigned long now = jiffies;
 121 +       int i;
 122 +
 123 +       write_lock(&ipfrag_lock);
 124 +       get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
 125 +       for (i = 0; i < IPQ_HASHSZ; i++) {
 126 +               struct ipq *q;
 127 +
 128 +               q = ipq_hash[i];
 129 +               while (q) {
 130 +                       struct ipq *next = q->next;
 131 +                       unsigned int hval = ipqhashfn(q->id, q->saddr,
 132 +                                                     q->daddr, q->protocol);
 133 +
 134 +                       if (hval != i) {
 135 +                               /* Unlink. */
 136 +                               if (q->next)
 137 +                                       q->next->pprev = q->pprev;
 138 +                               *q->pprev = q->next;
 139 +
 140 +                               /* Relink to new hash chain. */
 141 +                               if ((q->next = ipq_hash[hval]) != NULL)
 142 +                                       q->next->pprev = &q->next;
 143 +                               ipq_hash[hval] = q;
 144 +                               q->pprev = &ipq_hash[hval];
 145 +                       }
 146 +
 147 +                       q = next;
 148 +               }
 149 +       }
 150 +       write_unlock(&ipfrag_lock);
 151 +
 152 +       mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
 153 +}
 154
 155  atomic_t ip_frag_mem = ATOMIC_INIT(0); /* Memory used for fragments */
 156
 157 @@ -630,4 +663,15 @@
 158         IP_INC_STATS_BH(IpReasmFails);
 159         kfree_skb(skb);
 160         return NULL;
 161 +}
 162 +
 163 +void ipfrag_init(void)
 164 +{
 165 +       ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
 166 +                                (jiffies ^ (jiffies >> 6)));
 167 +
 168 +       init_timer(&ipfrag_secret_timer);
 169 +       ipfrag_secret_timer.function = ipfrag_secret_rebuild;
 170 +       ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
 171 +       add_timer(&ipfrag_secret_timer);
 172  }
 173 diff -Nru a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
 174 --- a/net/ipv4/sysctl_net_ipv4.c        Wed May 28 01:21:40 2003
 175 +++ b/net/ipv4/sysctl_net_ipv4.c        Wed May 28 01:21:40 2003
 176 @@ -27,6 +27,7 @@
 177  extern int sysctl_ipfrag_low_thresh;
 178  extern int sysctl_ipfrag_high_thresh;
 179  extern int sysctl_ipfrag_time;
 180 +extern int sysctl_ipfrag_secret_interval;
 181
 182  /* From ip_output.c */
 183  extern int sysctl_ip_dynaddr;
 184 @@ -225,6 +226,9 @@
 185          &sysctl_icmp_ratemask, sizeof(int), 0644, NULL, &proc_dointvec},
 186         {NET_TCP_TW_REUSE, "tcp_tw_reuse",
 187          &sysctl_tcp_tw_reuse, sizeof(int), 0644, NULL, &proc_dointvec},
 188 +       {NET_IPV4_IPFRAG_SECRET_INTERVAL, "ipfrag_secret_interval",
 189 +        &sysctl_ipfrag_secret_interval, sizeof(int), 0644, NULL, &proc_dointvec_jiffies,
 190 +        &sysctl_jiffies},
 191         {0}
 192  };
 193
 194 diff -Nru a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
 195 --- a/net/ipv6/af_inet6.c       Wed May 28 01:21:40 2003
 196 +++ b/net/ipv6/af_inet6.c       Wed May 28 01:21:40 2003
 197 @@ -671,6 +671,7 @@
 198         ip6_flowlabel_init();
 199         addrconf_init();
 200         sit_init();
 201 +       ipv6_frag_init();
 202
 203         /* Init v6 transport protocols. */
 204         udpv6_init();
 205 diff -Nru a/net/ipv6/reassembly.c b/net/ipv6/reassembly.c
 206 --- a/net/ipv6/reassembly.c     Wed May 28 01:21:40 2003
 207 +++ b/net/ipv6/reassembly.c     Wed May 28 01:21:40 2003
 208 @@ -37,6 +37,8 @@
 209  #include <linux/in6.h>
 210  #include <linux/ipv6.h>
 211  #include <linux/icmpv6.h>
 212 +#include <linux/random.h>
 213 +#include <linux/jhash.h>
 214
 215  #include <net/sock.h>
 216  #include <net/snmp.h>
 217 @@ -98,6 +100,7 @@
 218
 219  static struct frag_queue *ip6_frag_hash[IP6Q_HASHSZ];
 220  static rwlock_t ip6_frag_lock = RW_LOCK_UNLOCKED;
 221 +static u32 ip6_frag_hash_rnd;
 222  int ip6_frag_nqueues = 0;
 223
 224  static __inline__ void __fq_unlink(struct frag_queue *fq)
 225 @@ -117,16 +120,73 @@
 226         write_unlock(&ip6_frag_lock);
 227  }
 228
 229 -static __inline__ unsigned int ip6qhashfn(u32 id, struct in6_addr *saddr,
 230 -                                         struct in6_addr *daddr)
 231 +static unsigned int ip6qhashfn(u32 id, struct in6_addr *saddr,
 232 +                              struct in6_addr *daddr)
 233  {
 234 -       unsigned int h = saddr->s6_addr32[3] ^ daddr->s6_addr32[3] ^ id;
 235 +       u32 a, b, c;
 236
 237 -       h ^= (h>>16);
 238 -       h ^= (h>>8);
 239 -       return h & (IP6Q_HASHSZ - 1);
 240 +       a = saddr->s6_addr32[0];
 241 +       b = saddr->s6_addr32[1];
 242 +       c = saddr->s6_addr32[2];
 243 +
 244 +       a += JHASH_GOLDEN_RATIO;
 245 +       b += JHASH_GOLDEN_RATIO;
 246 +       c += ip6_frag_hash_rnd;
 247 +       __jhash_mix(a, b, c);
 248 +
 249 +       a += saddr->s6_addr32[3];
 250 +       b += daddr->s6_addr32[0];
 251 +       c += daddr->s6_addr32[1];
 252 +       __jhash_mix(a, b, c);
 253 +
 254 +       a += daddr->s6_addr32[2];
 255 +       b += daddr->s6_addr32[3];
 256 +       c += id;
 257 +       __jhash_mix(a, b, c);
 258 +
 259 +       return c & (IP6Q_HASHSZ - 1);
 260  }
 261
 262 +static struct timer_list ip6_frag_secret_timer;
 263 +static int ip6_frag_secret_interval = 10 * 60 * HZ;
 264 +
 265 +static void ip6_frag_secret_rebuild(unsigned long dummy)
 266 +{
 267 +       unsigned long now = jiffies;
 268 +       int i;
 269 +
 270 +       write_lock(&ip6_frag_lock);
 271 +       get_random_bytes(&ip6_frag_hash_rnd, sizeof(u32));
 272 +       for (i = 0; i < IP6Q_HASHSZ; i++) {
 273 +               struct frag_queue *q;
 274 +
 275 +               q = ip6_frag_hash[i];
 276 +               while (q) {
 277 +                       struct frag_queue *next = q->next;
 278 +                       unsigned int hval = ip6qhashfn(q->id,
 279 +                                                      &q->saddr,
 280 +                                                      &q->daddr);
 281 +
 282 +                       if (hval != i) {
 283 +                               /* Unlink. */
 284 +                               if (q->next)
 285 +                                       q->next->pprev = q->pprev;
 286 +                               *q->pprev = q->next;
 287 +
 288 +                               /* Relink to new hash chain. */
 289 +                               if ((q->next = ip6_frag_hash[hval]) != NULL)
 290 +                                       q->next->pprev = &q->next;
 291 +                               ip6_frag_hash[hval] = q;
 292 +                               q->pprev = &ip6_frag_hash[hval];
 293 +                       }
 294 +
 295 +                       q = next;
 296 +               }
 297 +       }
 298 +       write_unlock(&ip6_frag_lock);
 299 +
 300 +       mod_timer(&ip6_frag_secret_timer, now + ip6_frag_secret_interval);
 301 +}
 302
 303  atomic_t ip6_frag_mem = ATOMIC_INIT(0);
 304
 305 @@ -680,4 +740,15 @@
 306         IP6_INC_STATS_BH(Ip6ReasmFails);
 307         kfree_skb(skb);
 308         return -1;
 309 +}
 310 +
 311 +void __init ipv6_frag_init(void)
 312 +{
 313 +       ip6_frag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
 314 +                                  (jiffies ^ (jiffies >> 6)));
 315 +
 316 +       init_timer(&ip6_frag_secret_timer);
 317 +       ip6_frag_secret_timer.function = ip6_frag_secret_rebuild;
 318 +       ip6_frag_secret_timer.expires = jiffies + ip6_frag_secret_interval;
 319 +       add_timer(&ip6_frag_secret_timer);
 320  }
 321 diff -Nru a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c
 322 --- a/net/ipv6/tcp_ipv6.c       Wed May 28 01:21:40 2003
 323 +++ b/net/ipv6/tcp_ipv6.c       Wed May 28 01:21:40 2003
 324 @@ -371,9 +371,22 @@
 325
 326  static u32 tcp_v6_synq_hash(struct in6_addr *raddr, u16 rport, u32 rnd)
 327  {
 328 -       return (jhash_3words(raddr->s6_addr32[0] ^ raddr->s6_addr32[1],
 329 -                            raddr->s6_addr32[2] ^ raddr->s6_addr32[3],
 330 -                            (u32) rport, rnd) & (TCP_SYNQ_HSIZE - 1));
 331 +       u32 a, b, c;
 332 +
 333 +       a = raddr->s6_addr32[0];
 334 +       b = raddr->s6_addr32[1];
 335 +       c = raddr->s6_addr32[2];
 336 +
 337 +       a += JHASH_GOLDEN_RATIO;
 338 +       b += JHASH_GOLDEN_RATIO;
 339 +       c += rnd;
 340 +       __jhash_mix(a, b, c);
 341 +
 342 +       a += raddr->s6_addr32[3];
 343 +       b += (u32) rport;
 344 +       __jhash_mix(a, b, c);
 345 +
 346 +       return c & (TCP_SYNQ_HSIZE - 1);
 347  }
 348
 349  static struct open_request *tcp_v6_search_req(struct tcp_opt *tp,