pintos-os.org Git - openvswitch/blob - datapath/flow.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008 The Board of Trustees of The Leland
   4  * Stanford Junior University
   5  */
   6
   7 #include "flow.h"
   8 #include <linux/netdevice.h>
   9 #include <linux/etherdevice.h>
  10 #include <linux/if_ether.h>
  11 #include <linux/if_vlan.h>
  12 #include <linux/if_arp.h>
  13 #include <net/llc_pdu.h>
  14 #include <linux/ip.h>
  15 #include <linux/kernel.h>
  16 #include <linux/tcp.h>
  17 #include <linux/udp.h>
  18 #include <linux/in.h>
  19 #include <linux/rcupdate.h>
  20
  21 #include "openflow.h"
  22 #include "compat.h"
  23 #include "snap.h"
  24
  25 struct kmem_cache *flow_cache;
  26
  27 /* Internal function used to compare fields in flow. */
  28 static inline
  29 int flow_fields_match(const struct sw_flow_key *a, const struct sw_flow_key *b,
  30                 uint16_t w)
  31 {
  32         return ((w & OFPFW_IN_PORT || a->in_port == b->in_port)
  33                 && (w & OFPFW_DL_VLAN || a->dl_vlan == b->dl_vlan)
  34                 && (w & OFPFW_DL_SRC || !memcmp(a->dl_src, b->dl_src, ETH_ALEN))
  35                 && (w & OFPFW_DL_DST || !memcmp(a->dl_dst, b->dl_dst, ETH_ALEN))
  36                 && (w & OFPFW_DL_TYPE || a->dl_type == b->dl_type)
  37                 && (w & OFPFW_NW_SRC || a->nw_src == b->nw_src)
  38                 && (w & OFPFW_NW_DST || a->nw_dst == b->nw_dst)
  39                 && (w & OFPFW_NW_PROTO || a->nw_proto == b->nw_proto)
  40                 && (w & OFPFW_TP_SRC || a->tp_src == b->tp_src)
  41                 && (w & OFPFW_TP_DST || a->tp_dst == b->tp_dst));
  42 }
  43
  44 /* Returns nonzero if 'a' and 'b' match, that is, if their fields are equal
  45  * modulo wildcards, zero otherwise. */
  46 inline
  47 int flow_matches(const struct sw_flow_key *a, const struct sw_flow_key *b)
  48 {
  49         return flow_fields_match(a, b, (a->wildcards | b->wildcards));
  50 }
  51
  52 /* Returns nonzero if 't' (the table entry's key) and 'd' (the key
  53  * describing the deletion) match, that is, if their fields are
  54  * equal modulo wildcards, zero otherwise.  If 'strict' is nonzero, the
  55  * wildcards must match in both 't_key' and 'd_key'.  Note that the
  56  * table's wildcards are ignored unless 'strict' is set. */
  57 inline
  58 int flow_del_matches(const struct sw_flow_key *t, const struct sw_flow_key *d, int strict)
  59 {
  60         if (strict && (t->wildcards != d->wildcards))
  61                 return 0;
  62
  63         return flow_fields_match(t, d, d->wildcards);
  64 }
  65
  66 void flow_extract_match(struct sw_flow_key* to, const struct ofp_match* from)
  67 {
  68         to->wildcards = ntohs(from->wildcards) & OFPFW_ALL;
  69         memset(to->pad, '\0', sizeof(to->pad));
  70         to->in_port = from->in_port;
  71         to->dl_vlan = from->dl_vlan;
  72         memcpy(to->dl_src, from->dl_src, ETH_ALEN);
  73         memcpy(to->dl_dst, from->dl_dst, ETH_ALEN);
  74         to->dl_type = from->dl_type;
  75
  76         if (likely(from->dl_type == htons(ETH_P_IP))) {
  77                 to->nw_src   = from->nw_src;
  78                 to->nw_dst   = from->nw_dst;
  79                 to->nw_proto = from->nw_proto;
  80
  81                 if ((from->nw_proto != IPPROTO_TCP && from->nw_proto != IPPROTO_UDP)) {
  82                         goto no_th;
  83                 }
  84                 to->tp_src = from->tp_src;
  85                 to->tp_dst = from->tp_dst;
  86                 return;
  87         } else if (from->dl_type == htons(ETH_P_ARP)) {
  88                 to->nw_src = from->nw_src;
  89                 to->nw_dst = from->nw_dst;
  90                 to->nw_proto = 0;
  91                 goto no_th;
  92         }
  93
  94         to->nw_src = 0;
  95         to->nw_dst = 0;
  96         to->nw_proto = 0;
  97
  98 no_th:
  99         to->tp_src = 0;
 100         to->tp_dst = 0;
 101 }
 102
 103 void flow_fill_match(struct ofp_match* to, const struct sw_flow_key* from)
 104 {
 105         to->wildcards = htons(from->wildcards);
 106         to->in_port   = from->in_port;
 107         to->dl_vlan   = from->dl_vlan;
 108         memcpy(to->dl_src, from->dl_src, ETH_ALEN);
 109         memcpy(to->dl_dst, from->dl_dst, ETH_ALEN);
 110         to->dl_type   = from->dl_type;
 111         to->nw_src        = from->nw_src;
 112         to->nw_dst        = from->nw_dst;
 113         to->nw_proto  = from->nw_proto;
 114         to->tp_src        = from->tp_src;
 115         to->tp_dst        = from->tp_dst;
 116         memset(to->pad, '\0', sizeof(to->pad));
 117 }
 118
 119 /* Returns true if 'flow' can be deleted and set up for a deferred free, false
 120  * if deletion has already been scheduled (by another thread).
 121  *
 122  * Caller must hold rcu_read_lock. */
 123 int flow_del(struct sw_flow *flow)
 124 {
 125         return !atomic_cmpxchg(&flow->deleted, 0, 1);
 126 }
 127
 128 /* Allocates and returns a new flow with 'n_actions' action, using allocation
 129  * flags 'flags'.  Returns the new flow or a null pointer on failure. */
 130 struct sw_flow *flow_alloc(int n_actions, gfp_t flags)
 131 {
 132         struct sw_flow *flow = kmem_cache_alloc(flow_cache, flags);
 133         if (unlikely(!flow))
 134                 return NULL;
 135
 136         flow->n_actions = n_actions;
 137         flow->actions = kmalloc(n_actions * sizeof *flow->actions,
 138                                 flags);
 139         if (unlikely(!flow->actions) && n_actions > 0) {
 140                 kmem_cache_free(flow_cache, flow);
 141                 return NULL;
 142         }
 143         return flow;
 144 }
 145
 146 /* Frees 'flow' immediately. */
 147 void flow_free(struct sw_flow *flow)
 148 {
 149         if (unlikely(!flow))
 150                 return;
 151         kfree(flow->actions);
 152         kmem_cache_free(flow_cache, flow);
 153 }
 154
 155 /* RCU callback used by flow_deferred_free. */
 156 static void rcu_callback(struct rcu_head *rcu)
 157 {
 158         struct sw_flow *flow = container_of(rcu, struct sw_flow, rcu);
 159         flow_free(flow);
 160 }
 161
 162 /* Schedules 'flow' to be freed after the next RCU grace period.
 163  * The caller must hold rcu_read_lock for this to be sensible. */
 164 void flow_deferred_free(struct sw_flow *flow)
 165 {
 166         call_rcu(&flow->rcu, rcu_callback);
 167 }
 168
 169 /* Prints a representation of 'key' to the kernel log. */
 170 void print_flow(const struct sw_flow_key *key)
 171 {
 172         printk("wild%04x port%04x:vlan%04x mac%02x:%02x:%02x:%02x:%02x:%02x"
 173                         "->%02x:%02x:%02x:%02x:%02x:%02x "
 174                         "proto%04x ip%u.%u.%u.%u->%u.%u.%u.%u port%d->%d\n",
 175                         key->wildcards, ntohs(key->in_port), ntohs(key->dl_vlan),
 176                         key->dl_src[0], key->dl_src[1], key->dl_src[2],
 177                         key->dl_src[3], key->dl_src[4], key->dl_src[5],
 178                         key->dl_dst[0], key->dl_dst[1], key->dl_dst[2],
 179                         key->dl_dst[3], key->dl_dst[4], key->dl_dst[5],
 180                         ntohs(key->dl_type),
 181                         ((unsigned char *)&key->nw_src)[0],
 182                         ((unsigned char *)&key->nw_src)[1],
 183                         ((unsigned char *)&key->nw_src)[2],
 184                         ((unsigned char *)&key->nw_src)[3],
 185                         ((unsigned char *)&key->nw_dst)[0],
 186                         ((unsigned char *)&key->nw_dst)[1],
 187                         ((unsigned char *)&key->nw_dst)[2],
 188                         ((unsigned char *)&key->nw_dst)[3],
 189                         ntohs(key->tp_src), ntohs(key->tp_dst));
 190 }
 191
 192 uint32_t hash_in6(const struct in6_addr *in)
 193 {
 194         return (in->s6_addr32[0] ^ in->s6_addr32[1]
 195                         ^ in->s6_addr32[2] ^ in->s6_addr32[3]);
 196 }
 197
 198 // with inspiration from linux/if_arp.h
 199 struct arp_eth_hdr {
 200         uint16_t  ar_hrd;  /* format of hardware address    */
 201         uint16_t  ar_pro;  /* format of protocol address    */
 202         uint8_t   ar_hln;  /* length of hardware address    */
 203         uint8_t   ar_pln;  /* length of protocol address    */
 204         uint16_t  ar_op;   /* ARP opcode (command)          */
 205
 206         uint8_t   ar_sha[ETH_ALEN]; /* source hardware addr */
 207         uint32_t  ar_sip;           /* source protocol addr */
 208         uint8_t   ar_tha[ETH_ALEN]; /* dest hardware addr   */
 209         uint32_t  ar_tip;           /* dest protocol addr   */
 210 } __attribute__((packed));
 211
 212 /* Parses the Ethernet frame in 'skb', which was received on 'in_port',
 213  * and initializes 'key' to match. */
 214 void flow_extract(struct sk_buff *skb, uint16_t in_port,
 215                   struct sw_flow_key *key)
 216 {
 217         struct ethhdr *mac;
 218         struct udphdr *th;
 219         int nh_ofs, th_ofs;
 220
 221         key->in_port = htons(in_port);
 222         key->wildcards = 0;
 223         memset(key->pad, '\0', sizeof(key->pad));
 224
 225         /* This code doesn't check that skb->len is long enough to contain the
 226          * MAC or network header.  With a 46-byte minimum length frame this
 227          * assumption is always correct. */
 228
 229         /* Doesn't verify checksums.  Should it? */
 230
 231         /* Data link layer.  We only support Ethernet. */
 232         mac = eth_hdr(skb);
 233         nh_ofs = sizeof(struct ethhdr);
 234         if (likely(ntohs(mac->h_proto) >= OFP_DL_TYPE_ETH2_CUTOFF)) {
 235                 /* This is an Ethernet II frame */
 236                 key->dl_type = mac->h_proto;
 237         } else {
 238                 /* This is an 802.2 frame */
 239                 if (snap_get_ethertype(skb, &key->dl_type) != -EINVAL) {
 240                         nh_ofs += sizeof(struct snap_hdr);
 241                 } else {
 242                         key->dl_type = OFP_DL_TYPE_NOT_ETH_TYPE;
 243                         nh_ofs += sizeof(struct llc_pdu_un);
 244                 }
 245         }
 246
 247         /* Check for a VLAN tag */
 248         if (likely(key->dl_type != htons(ETH_P_8021Q))) {
 249                 key->dl_vlan = htons(OFP_VLAN_NONE);
 250         } else {
 251                 struct vlan_hdr *vh = (struct vlan_hdr *)(skb_mac_header(skb) + nh_ofs);
 252                 key->dl_type = vh->h_vlan_encapsulated_proto;
 253                 key->dl_vlan = vh->h_vlan_TCI & htons(VLAN_VID_MASK);
 254                 nh_ofs += sizeof(*vh);
 255         }
 256         memcpy(key->dl_src, mac->h_source, ETH_ALEN);
 257         memcpy(key->dl_dst, mac->h_dest, ETH_ALEN);
 258         skb_set_network_header(skb, nh_ofs);
 259
 260         /* Network layer. */
 261         if (likely(key->dl_type == htons(ETH_P_IP))) {
 262                 struct iphdr *nh = ip_hdr(skb);
 263                 key->nw_src = nh->saddr;
 264                 key->nw_dst = nh->daddr;
 265                 key->nw_proto = nh->protocol;
 266                 th_ofs = nh_ofs + nh->ihl * 4;
 267                 skb_set_transport_header(skb, th_ofs);
 268
 269                 /* Transport layer. */
 270                 if ((key->nw_proto != IPPROTO_TCP && key->nw_proto != IPPROTO_UDP)
 271                                 || skb->len < th_ofs + sizeof(struct udphdr)) {
 272                         goto no_th;
 273                 }
 274                 th = udp_hdr(skb);
 275                 key->tp_src = th->source;
 276                 key->tp_dst = th->dest;
 277
 278                 return;
 279         } else if (key->dl_type == htons(ETH_P_ARP)) {
 280                 /* just barely within 46-byte minimum packet */
 281                 struct arp_eth_hdr *ah = (struct arp_eth_hdr *)skb_network_header(skb);
 282                 if (ah->ar_hrd == htons(ARPHRD_ETHER)
 283                     && ah->ar_pro == htons(ETH_P_IP)
 284                     && ah->ar_hln == ETH_ALEN
 285                     && ah->ar_pln == sizeof(key->nw_src))
 286                 {
 287                         /* check if sha/tha match dl_src/dl_dst? */
 288                         key->nw_src = ah->ar_sip;
 289                         key->nw_dst = ah->ar_tip;
 290                         key->nw_proto = 0;
 291                         goto no_th;
 292                 }
 293         } else {
 294                 /* Fall through. */
 295         }
 296
 297         key->nw_src = 0;
 298         key->nw_dst = 0;
 299         key->nw_proto = 0;
 300
 301 no_th:
 302         key->tp_src = 0;
 303         key->tp_dst = 0;
 304 }
 305
 306 /* Initializes the flow module.
 307  * Returns zero if successful or a negative error code. */
 308 int flow_init(void)
 309 {
 310         flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow), 0,
 311                                         0, NULL);
 312         if (flow_cache == NULL)
 313                 return -ENOMEM;
 314
 315         return 0;
 316 }
 317
 318 /* Uninitializes the flow module. */
 319 void flow_exit(void)
 320 {
 321         kmem_cache_destroy(flow_cache);
 322 }
 323