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
7 /* Functions for managing the dp interface/device. */
9 #include <linux/init.h>
10 #include <linux/module.h>
11 #include <linux/if_arp.h>
12 #include <linux/if_bridge.h>
13 #include <linux/if_vlan.h>
15 #include <net/genetlink.h>
17 #include <linux/delay.h>
18 #include <linux/etherdevice.h>
19 #include <linux/kernel.h>
20 #include <linux/kthread.h>
21 #include <linux/mutex.h>
22 #include <linux/rtnetlink.h>
23 #include <linux/rcupdate.h>
24 #include <linux/version.h>
25 #include <linux/ethtool.h>
26 #include <linux/random.h>
27 #include <asm/system.h>
28 #include <linux/netfilter_bridge.h>
29 #include <linux/inetdevice.h>
30 #include <linux/list.h>
32 #include "openflow-netlink.h"
38 #include "datapath_t.h"
43 /* Number of milliseconds between runs of the maintenance thread. */
44 #define MAINT_SLEEP_MSECS 1000
46 #define BRIDGE_PORT_NO_FLOOD 0x00000001
48 #define UINT32_MAX 4294967295U
49 #define UINT16_MAX 65535
50 #define MAX(X, Y) ((X) > (Y) ? (X) : (Y))
52 struct net_bridge_port {
56 struct net_device *dev;
57 struct list_head node; /* Element in datapath.ports. */
60 static struct genl_family dp_genl_family;
61 static struct genl_multicast_group mc_group;
63 int dp_dev_setup(struct net_device *dev);
65 /* It's hard to imagine wanting more than one datapath, but... */
68 /* datapaths. Protected on the read side by rcu_read_lock, on the write side
71 * It is safe to access the datapath and net_bridge_port structures with just
72 * the dp_mutex, but to access the chain you need to take the rcu_read_lock
73 * also (because dp_mutex doesn't prevent flows from being destroyed).
75 static struct datapath *dps[DP_MAX];
76 static DEFINE_MUTEX(dp_mutex);
78 static int dp_maint_func(void *data);
79 static int send_port_status(struct net_bridge_port *p, uint8_t status);
80 static int dp_genl_openflow_done(struct netlink_callback *);
82 /* nla_shrink - reduce amount of space reserved by nla_reserve
83 * @skb: socket buffer from which to recover room
84 * @nla: netlink attribute to adjust
85 * @len: new length of attribute payload
87 * Reduces amount of space reserved by a call to nla_reserve.
89 * No other attributes may be added between calling nla_reserve and this
90 * function, since it will create a hole in the message.
92 void nla_shrink(struct sk_buff *skb, struct nlattr *nla, int len)
94 int delta = nla_total_size(len) - nla_total_size(nla_len(nla));
98 nla->nla_len = nla_attr_size(len);
101 /* Puts a set of openflow headers for a message of the given 'type' into 'skb'.
102 * If 'sender' is nonnull, then it is used as the message's destination. 'dp'
103 * must specify the datapath to use.
105 * '*max_openflow_len' receives the maximum number of bytes that are available
106 * for the embedded OpenFlow message. The caller must call
107 * resize_openflow_skb() to set the actual size of the message to this number
110 * Returns the openflow header if successful, otherwise (if 'skb' is too small)
113 put_openflow_headers(struct datapath *dp, struct sk_buff *skb, uint8_t type,
114 const struct sender *sender, int *max_openflow_len)
116 struct ofp_header *oh;
120 /* Assemble the Generic Netlink wrapper. */
121 if (!genlmsg_put(skb,
122 sender ? sender->pid : 0,
123 sender ? sender->seq : 0,
124 &dp_genl_family, 0, DP_GENL_C_OPENFLOW))
125 return ERR_PTR(-ENOBUFS);
126 if (nla_put_u32(skb, DP_GENL_A_DP_IDX, dp->dp_idx) < 0)
127 return ERR_PTR(-ENOBUFS);
128 openflow_len = (skb_tailroom(skb) - NLA_HDRLEN) & ~(NLA_ALIGNTO - 1);
129 if (openflow_len < sizeof *oh)
130 return ERR_PTR(-ENOBUFS);
131 *max_openflow_len = openflow_len;
132 attr = nla_reserve(skb, DP_GENL_A_OPENFLOW, openflow_len);
135 /* Fill in the header. The caller is responsible for the length. */
137 oh->version = OFP_VERSION;
139 oh->xid = sender ? sender->xid : 0;
144 /* Resizes OpenFlow header 'oh', which must be at the tail end of 'skb', to new
145 * length 'new_length' (in bytes), adjusting pointers and size values as
148 resize_openflow_skb(struct sk_buff *skb,
149 struct ofp_header *oh, size_t new_length)
151 struct nlattr *attr = ((void *) oh) - NLA_HDRLEN;
152 nla_shrink(skb, attr, new_length);
153 oh->length = htons(new_length);
154 nlmsg_end(skb, (struct nlmsghdr *) skb->data);
157 /* Allocates a new skb to contain an OpenFlow message 'openflow_len' bytes in
158 * length. Returns a null pointer if memory is unavailable, otherwise returns
159 * the OpenFlow header and stores a pointer to the skb in '*pskb'.
161 * 'type' is the OpenFlow message type. If 'sender' is nonnull, then it is
162 * used as the message's destination. 'dp' must specify the datapath to
165 alloc_openflow_skb(struct datapath *dp, size_t openflow_len, uint8_t type,
166 const struct sender *sender, struct sk_buff **pskb)
168 struct ofp_header *oh;
171 int max_openflow_len;
173 if ((openflow_len + sizeof(struct ofp_header)) > UINT16_MAX) {
175 printk("alloc_openflow_skb: openflow message too large: %zu\n",
180 genl_len = nlmsg_total_size(GENL_HDRLEN + dp_genl_family.hdrsize);
181 genl_len += nla_total_size(sizeof(uint32_t)); /* DP_GENL_A_DP_IDX */
182 genl_len += nla_total_size(openflow_len); /* DP_GENL_A_OPENFLOW */
183 skb = *pskb = genlmsg_new(genl_len, GFP_ATOMIC);
186 printk("alloc_openflow_skb: genlmsg_new failed\n");
190 oh = put_openflow_headers(dp, skb, type, sender, &max_openflow_len);
191 BUG_ON(!oh || IS_ERR(oh));
192 resize_openflow_skb(skb, oh, openflow_len);
197 /* Sends 'skb' to 'sender' if it is nonnull, otherwise multicasts 'skb' to all
200 send_openflow_skb(struct sk_buff *skb, const struct sender *sender)
203 ? genlmsg_unicast(skb, sender->pid)
204 : genlmsg_multicast(skb, 0, mc_group.id, GFP_ATOMIC));
205 if (err && net_ratelimit())
206 printk(KERN_WARNING "send_openflow_skb: send failed: %d\n",
211 /* Generates a unique datapath id. It incorporates the datapath index
212 * and a hardware address, if available. If not, it generates a random
216 uint64_t gen_datapath_id(uint16_t dp_idx)
220 struct net_device *dev;
222 /* The top 16 bits are used to identify the datapath. The lower 48 bits
223 * use an interface address. */
224 id = (uint64_t)dp_idx << 48;
225 if ((dev = dev_get_by_name(&init_net, "ctl0"))
226 || (dev = dev_get_by_name(&init_net, "eth0"))) {
227 for (i=0; i<ETH_ALEN; i++) {
228 id |= (uint64_t)dev->dev_addr[i] << (8*(ETH_ALEN-1 - i));
232 /* Randomly choose the lower 48 bits if we cannot find an
233 * address and mark the most significant bit to indicate that
234 * this was randomly generated. */
235 uint8_t rand[ETH_ALEN];
236 get_random_bytes(rand, ETH_ALEN);
237 id |= (uint64_t)1 << 63;
238 for (i=0; i<ETH_ALEN; i++) {
239 id |= (uint64_t)rand[i] << (8*(ETH_ALEN-1 - i));
246 /* Creates a new datapath numbered 'dp_idx'. Returns 0 for success or a
247 * negative error code.
249 * Not called with any locks. */
250 static int new_dp(int dp_idx)
255 if (dp_idx < 0 || dp_idx >= DP_MAX)
258 if (!try_module_get(THIS_MODULE))
261 mutex_lock(&dp_mutex);
262 dp = rcu_dereference(dps[dp_idx]);
269 dp = kzalloc(sizeof *dp, GFP_KERNEL);
274 dp->id = gen_datapath_id(dp_idx);
275 dp->chain = chain_create(dp);
276 if (dp->chain == NULL)
278 INIT_LIST_HEAD(&dp->port_list);
281 /* Setup our "of" device */
284 err = dp_dev_setup(&dp->dev);
287 printk("datapath: problem setting up 'of' device\n");
291 dp->miss_send_len = OFP_DEFAULT_MISS_SEND_LEN;
293 dp->dp_task = kthread_run(dp_maint_func, dp, "dp%d", dp_idx);
294 if (IS_ERR(dp->dp_task))
297 rcu_assign_pointer(dps[dp_idx], dp);
298 mutex_unlock(&dp_mutex);
305 mutex_unlock(&dp_mutex);
306 module_put(THIS_MODULE);
310 /* Find and return a free port number under 'dp'. Called under dp_mutex. */
311 static int find_portno(struct datapath *dp)
314 for (i = 0; i < OFPP_MAX; i++)
315 if (dp->ports[i] == NULL)
320 static struct net_bridge_port *new_nbp(struct datapath *dp,
321 struct net_device *dev)
323 struct net_bridge_port *p;
326 port_no = find_portno(dp);
328 return ERR_PTR(port_no);
330 p = kzalloc(sizeof(*p), GFP_KERNEL);
332 return ERR_PTR(-ENOMEM);
337 p->port_no = port_no;
342 /* Called with dp_mutex. */
343 int add_switch_port(struct datapath *dp, struct net_device *dev)
345 struct net_bridge_port *p;
347 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER)
350 if (dev->br_port != NULL)
353 p = new_nbp(dp, dev);
358 rcu_assign_pointer(dev->br_port, p);
360 dev_set_promiscuity(dev, 1);
363 rcu_assign_pointer(dp->ports[p->port_no], p);
364 list_add_rcu(&p->node, &dp->port_list);
366 /* Notify the ctlpath that this port has been added */
367 send_port_status(p, OFPPR_ADD);
372 /* Delete 'p' from switch.
373 * Called with dp_mutex. */
374 static int del_switch_port(struct net_bridge_port *p)
376 /* First drop references to device. */
378 dev_set_promiscuity(p->dev, -1);
380 list_del_rcu(&p->node);
381 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
382 rcu_assign_pointer(p->dev->br_port, NULL);
384 /* Then wait until no one is still using it, and destroy it. */
387 /* Notify the ctlpath that this port no longer exists */
388 send_port_status(p, OFPPR_DELETE);
396 /* Called with dp_mutex. */
397 static void del_dp(struct datapath *dp)
399 struct net_bridge_port *p, *n;
402 /* Unregister the "of" device of this dp */
404 unregister_netdevice(&dp->dev);
408 kthread_stop(dp->dp_task);
410 /* Drop references to DP. */
411 list_for_each_entry_safe (p, n, &dp->port_list, node)
413 rcu_assign_pointer(dps[dp->dp_idx], NULL);
415 /* Wait until no longer in use, then destroy it. */
417 chain_destroy(dp->chain);
419 module_put(THIS_MODULE);
422 static int dp_maint_func(void *data)
424 struct datapath *dp = (struct datapath *) data;
426 while (!kthread_should_stop()) {
428 chain_timeout(dp->chain);
430 int count = chain_timeout(dp->chain);
431 chain_print_stats(dp->chain);
433 printk("%d flows timed out\n", count);
435 msleep_interruptible(MAINT_SLEEP_MSECS);
442 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
443 * different set of devices!) Returns 0 if *pskb should be processed further,
444 * 1 if *pskb is handled. */
445 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
446 /* Called with rcu_read_lock. */
447 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
450 struct ethhdr *eh = eth_hdr(skb);
451 struct sk_buff *skb_local = NULL;
454 if (compare_ether_addr(eh->h_dest, skb->dev->dev_addr) == 0)
457 if (is_broadcast_ether_addr(eh->h_dest)
458 || is_multicast_ether_addr(eh->h_dest)
459 || is_local_ether_addr(eh->h_dest))
460 skb_local = skb_clone(skb, GFP_ATOMIC);
462 /* Push the Ethernet header back on. */
463 if (skb->protocol == htons(ETH_P_8021Q))
464 skb_push(skb, VLAN_ETH_HLEN);
466 skb_push(skb, ETH_HLEN);
468 fwd_port_input(p->dp->chain, skb, p->port_no);
472 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
473 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
475 /* Push the Ethernet header back on. */
476 if ((*pskb)->protocol == htons(ETH_P_8021Q))
477 skb_push(*pskb, VLAN_ETH_HLEN);
479 skb_push(*pskb, ETH_HLEN);
481 fwd_port_input(p->dp->chain, *pskb, p->port_no);
485 /* NB: This has only been tested on 2.4.35 */
487 /* Called without any locks (?) */
488 static void dp_frame_hook(struct sk_buff *skb)
490 struct net_bridge_port *p = skb->dev->br_port;
492 /* Push the Ethernet header back on. */
493 if (skb->protocol == htons(ETH_P_8021Q))
494 skb_push(skb, VLAN_ETH_HLEN);
496 skb_push(skb, ETH_HLEN);
500 fwd_port_input(p->dp->chain, skb, p->port_no);
507 /* Forwarding output path.
508 * Based on net/bridge/br_forward.c. */
510 /* Don't forward packets to originating port. If we're flooding,
511 * then don't send out ports with flooding disabled.
513 static inline int should_deliver(const struct net_bridge_port *p,
514 const struct sk_buff *skb, int flood)
516 if (skb->dev == p->dev)
519 if (flood && (p->flags & BRIDGE_PORT_NO_FLOOD))
525 static inline unsigned packet_length(const struct sk_buff *skb)
527 int length = skb->len - ETH_HLEN;
528 if (skb->protocol == htons(ETH_P_8021Q))
533 /* Send packets out all the ports except the originating one. If the
534 * "flood" argument is set, only send along the minimum spanning tree.
537 output_all(struct datapath *dp, struct sk_buff *skb, int flood)
539 struct net_bridge_port *p;
543 list_for_each_entry_rcu (p, &dp->port_list, node) {
544 if (!should_deliver(p, skb, flood))
546 if (prev_port != -1) {
547 struct sk_buff *clone = skb_clone(skb, GFP_ATOMIC);
552 dp_output_port(dp, clone, prev_port);
554 prev_port = p->port_no;
557 dp_output_port(dp, skb, prev_port);
564 /* Marks 'skb' as having originated from 'in_port' in 'dp'.
565 FIXME: how are devices reference counted? */
566 int dp_set_origin(struct datapath *dp, uint16_t in_port,
569 if (in_port < OFPP_MAX && dp->ports[in_port]) {
570 skb->dev = dp->ports[in_port]->dev;
576 /* Takes ownership of 'skb' and transmits it to 'out_port' on 'dp'.
578 int dp_output_port(struct datapath *dp, struct sk_buff *skb, int out_port)
580 struct net_bridge_port *p;
584 if (out_port == OFPP_FLOOD)
585 return output_all(dp, skb, 1);
586 else if (out_port == OFPP_ALL)
587 return output_all(dp, skb, 0);
588 else if (out_port == OFPP_CONTROLLER)
589 return dp_output_control(dp, skb, fwd_save_skb(skb), 0,
591 else if (out_port == OFPP_TABLE) {
592 struct sw_flow_key key;
593 struct sw_flow *flow;
595 flow_extract(skb, skb->dev->br_port->port_no, &key);
596 flow = chain_lookup(dp->chain, &key);
597 if (likely(flow != NULL)) {
598 flow_used(flow, skb);
599 execute_actions(dp, skb, &key, flow->actions, flow->n_actions);
603 } else if (out_port >= OFPP_MAX)
606 p = dp->ports[out_port];
611 if (packet_length(skb) > skb->dev->mtu) {
612 printk("dropped over-mtu packet: %d > %d\n",
613 packet_length(skb), skb->dev->mtu);
625 printk("can't forward to bad port %d\n", out_port);
629 /* Takes ownership of 'skb' and transmits it to 'dp''s control path. If
630 * 'buffer_id' != -1, then only the first 64 bytes of 'skb' are sent;
631 * otherwise, all of 'skb' is sent. 'reason' indicates why 'skb' is being
632 * sent. 'max_len' sets the maximum number of bytes that the caller
633 * wants to be sent; a value of 0 indicates the entire packet should be
636 dp_output_control(struct datapath *dp, struct sk_buff *skb,
637 uint32_t buffer_id, size_t max_len, int reason)
639 /* FIXME? Can we avoid creating a new skbuff in the case where we
640 * forward the whole packet? */
641 struct sk_buff *f_skb;
642 struct ofp_packet_in *opi;
643 size_t fwd_len, opi_len;
647 if ((buffer_id != (uint32_t) -1) && max_len)
648 fwd_len = min(fwd_len, max_len);
650 opi_len = offsetof(struct ofp_packet_in, data) + fwd_len;
651 opi = alloc_openflow_skb(dp, opi_len, OFPT_PACKET_IN, NULL, &f_skb);
656 opi->buffer_id = htonl(buffer_id);
657 opi->total_len = htons(skb->len);
658 opi->in_port = htons(skb->dev->br_port->port_no);
659 opi->reason = reason;
661 memcpy(opi->data, skb_mac_header(skb), fwd_len);
662 err = send_openflow_skb(f_skb, NULL);
669 static void fill_port_desc(struct net_bridge_port *p, struct ofp_phy_port *desc)
671 desc->port_no = htons(p->port_no);
672 strncpy(desc->name, p->dev->name, OFP_MAX_PORT_NAME_LEN);
673 desc->name[OFP_MAX_PORT_NAME_LEN-1] = '\0';
674 memcpy(desc->hw_addr, p->dev->dev_addr, ETH_ALEN);
675 desc->flags = htonl(p->flags);
679 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,24)
680 if (p->dev->ethtool_ops && p->dev->ethtool_ops->get_settings) {
681 struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
683 if (!p->dev->ethtool_ops->get_settings(p->dev, &ecmd)) {
684 if (ecmd.supported & SUPPORTED_10baseT_Half)
685 desc->features |= OFPPF_10MB_HD;
686 if (ecmd.supported & SUPPORTED_10baseT_Full)
687 desc->features |= OFPPF_10MB_FD;
688 if (ecmd.supported & SUPPORTED_100baseT_Half)
689 desc->features |= OFPPF_100MB_HD;
690 if (ecmd.supported & SUPPORTED_100baseT_Full)
691 desc->features |= OFPPF_100MB_FD;
692 if (ecmd.supported & SUPPORTED_1000baseT_Half)
693 desc->features |= OFPPF_1GB_HD;
694 if (ecmd.supported & SUPPORTED_1000baseT_Full)
695 desc->features |= OFPPF_1GB_FD;
696 /* 10Gbps half-duplex doesn't exist... */
697 if (ecmd.supported & SUPPORTED_10000baseT_Full)
698 desc->features |= OFPPF_10GB_FD;
700 desc->features = htonl(desc->features);
701 desc->speed = htonl(ecmd.speed);
708 fill_features_reply(struct datapath *dp, struct ofp_switch_features *ofr)
710 struct net_bridge_port *p;
713 ofr->datapath_id = cpu_to_be64(dp->id);
715 ofr->n_exact = htonl(2 * TABLE_HASH_MAX_FLOWS);
716 ofr->n_compression = 0; /* Not supported */
717 ofr->n_general = htonl(TABLE_LINEAR_MAX_FLOWS);
718 ofr->buffer_mb = htonl(UINT32_MAX);
719 ofr->n_buffers = htonl(N_PKT_BUFFERS);
720 ofr->capabilities = htonl(OFP_SUPPORTED_CAPABILITIES);
721 ofr->actions = htonl(OFP_SUPPORTED_ACTIONS);
723 list_for_each_entry_rcu (p, &dp->port_list, node) {
724 fill_port_desc(p, &ofr->ports[port_count]);
732 dp_send_features_reply(struct datapath *dp, const struct sender *sender)
735 struct ofp_switch_features *ofr;
736 size_t ofr_len, port_max_len;
740 port_max_len = sizeof(struct ofp_phy_port) * OFPP_MAX;
741 ofr = alloc_openflow_skb(dp, sizeof(*ofr) + port_max_len,
742 OFPT_FEATURES_REPLY, sender, &skb);
747 port_count = fill_features_reply(dp, ofr);
750 ofr_len = sizeof(*ofr) + (sizeof(struct ofp_phy_port) * port_count);
751 resize_openflow_skb(skb, &ofr->header, ofr_len);
752 return send_openflow_skb(skb, sender);
756 dp_send_config_reply(struct datapath *dp, const struct sender *sender)
759 struct ofp_switch_config *osc;
761 osc = alloc_openflow_skb(dp, sizeof *osc, OFPT_GET_CONFIG_REPLY, sender,
766 osc->flags = htons(dp->flags);
767 osc->miss_send_len = htons(dp->miss_send_len);
769 return send_openflow_skb(skb, sender);
773 dp_update_port_flags(struct datapath *dp, const struct ofp_phy_port *opp)
775 struct net_bridge_port *p;
777 p = dp->ports[htons(opp->port_no)];
779 /* Make sure the port id hasn't changed since this was sent */
780 if (!p || memcmp(opp->hw_addr, p->dev->dev_addr, ETH_ALEN) != 0)
783 p->flags = htonl(opp->flags);
790 send_port_status(struct net_bridge_port *p, uint8_t status)
793 struct ofp_port_status *ops;
795 ops = alloc_openflow_skb(p->dp, sizeof *ops, OFPT_PORT_STATUS, NULL,
799 ops->reason = status;
800 memset(ops->pad, 0, sizeof ops->pad);
801 fill_port_desc(p, &ops->desc);
803 return send_openflow_skb(skb, NULL);
807 dp_send_flow_expired(struct datapath *dp, struct sw_flow *flow)
810 struct ofp_flow_expired *ofe;
811 unsigned long duration_j;
813 ofe = alloc_openflow_skb(dp, sizeof *ofe, OFPT_FLOW_EXPIRED, 0, &skb);
817 flow_fill_match(&ofe->match, &flow->key);
819 memset(ofe->pad, 0, sizeof ofe->pad);
820 ofe->priority = htons(flow->priority);
822 duration_j = (flow->timeout - HZ * flow->max_idle) - flow->init_time;
823 ofe->duration = htonl(duration_j / HZ);
824 ofe->packet_count = cpu_to_be64(flow->packet_count);
825 ofe->byte_count = cpu_to_be64(flow->byte_count);
827 return send_openflow_skb(skb, NULL);
831 dp_send_error_msg(struct datapath *dp, const struct sender *sender,
832 uint16_t type, uint16_t code, const uint8_t *data, size_t len)
835 struct ofp_error_msg *oem;
838 oem = alloc_openflow_skb(dp, sizeof(*oem)+len, OFPT_ERROR_MSG,
843 oem->type = htons(type);
844 oem->code = htons(code);
845 memcpy(oem->data, data, len);
847 return send_openflow_skb(skb, sender);
850 /* Generic Netlink interface.
852 * See netlink(7) for an introduction to netlink. See
853 * http://linux-net.osdl.org/index.php/Netlink for more information and
854 * pointers on how to work with netlink and Generic Netlink in the kernel and
857 static struct genl_family dp_genl_family = {
858 .id = GENL_ID_GENERATE,
860 .name = DP_GENL_FAMILY_NAME,
862 .maxattr = DP_GENL_A_MAX,
865 /* Attribute policy: what each attribute may contain. */
866 static struct nla_policy dp_genl_policy[DP_GENL_A_MAX + 1] = {
867 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
868 [DP_GENL_A_MC_GROUP] = { .type = NLA_U32 },
869 [DP_GENL_A_PORTNAME] = { .type = NLA_STRING }
872 static int dp_genl_add(struct sk_buff *skb, struct genl_info *info)
874 if (!info->attrs[DP_GENL_A_DP_IDX])
877 return new_dp(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
880 static struct genl_ops dp_genl_ops_add_dp = {
881 .cmd = DP_GENL_C_ADD_DP,
882 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
883 .policy = dp_genl_policy,
888 struct datapath *dp_get(int dp_idx)
890 if (dp_idx < 0 || dp_idx > DP_MAX)
892 return rcu_dereference(dps[dp_idx]);
895 static int dp_genl_del(struct sk_buff *skb, struct genl_info *info)
900 if (!info->attrs[DP_GENL_A_DP_IDX])
903 mutex_lock(&dp_mutex);
904 dp = dp_get(nla_get_u32((info->attrs[DP_GENL_A_DP_IDX])));
911 mutex_unlock(&dp_mutex);
915 static struct genl_ops dp_genl_ops_del_dp = {
916 .cmd = DP_GENL_C_DEL_DP,
917 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
918 .policy = dp_genl_policy,
923 /* Queries a datapath for related information. Currently the only relevant
924 * information is the datapath's multicast group ID. Really we want one
925 * multicast group per datapath, but because of locking issues[*] we can't
926 * easily get one. Thus, every datapath will currently return the same
927 * global multicast group ID, but in the future it would be nice to fix that.
929 * [*] dp_genl_add, to add a new datapath, is called under the genl_lock
930 * mutex, and genl_register_mc_group, called to acquire a new multicast
931 * group ID, also acquires genl_lock, thus deadlock.
933 static int dp_genl_query(struct sk_buff *skb, struct genl_info *info)
936 struct sk_buff *ans_skb = NULL;
940 if (!info->attrs[DP_GENL_A_DP_IDX])
944 dp_idx = nla_get_u32((info->attrs[DP_GENL_A_DP_IDX]));
950 ans_skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
955 data = genlmsg_put_reply(ans_skb, info, &dp_genl_family,
956 0, DP_GENL_C_QUERY_DP);
961 NLA_PUT_U32(ans_skb, DP_GENL_A_DP_IDX, dp_idx);
962 NLA_PUT_U32(ans_skb, DP_GENL_A_MC_GROUP, mc_group.id);
964 genlmsg_end(ans_skb, data);
965 err = genlmsg_reply(ans_skb, info);
977 static struct genl_ops dp_genl_ops_query_dp = {
978 .cmd = DP_GENL_C_QUERY_DP,
979 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
980 .policy = dp_genl_policy,
981 .doit = dp_genl_query,
985 static int dp_genl_add_del_port(struct sk_buff *skb, struct genl_info *info)
988 struct net_device *port;
991 if (!info->attrs[DP_GENL_A_DP_IDX] || !info->attrs[DP_GENL_A_PORTNAME])
995 mutex_lock(&dp_mutex);
996 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1002 /* Get interface to add/remove. */
1003 port = dev_get_by_name(&init_net,
1004 nla_data(info->attrs[DP_GENL_A_PORTNAME]));
1010 /* Execute operation. */
1011 if (info->genlhdr->cmd == DP_GENL_C_ADD_PORT)
1012 err = add_switch_port(dp, port);
1014 if (port->br_port == NULL || port->br_port->dp != dp) {
1018 err = del_switch_port(port->br_port);
1024 mutex_unlock(&dp_mutex);
1028 static struct genl_ops dp_genl_ops_add_port = {
1029 .cmd = DP_GENL_C_ADD_PORT,
1030 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1031 .policy = dp_genl_policy,
1032 .doit = dp_genl_add_del_port,
1036 static struct genl_ops dp_genl_ops_del_port = {
1037 .cmd = DP_GENL_C_DEL_PORT,
1038 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1039 .policy = dp_genl_policy,
1040 .doit = dp_genl_add_del_port,
1044 static int dp_genl_openflow(struct sk_buff *skb, struct genl_info *info)
1046 struct nlattr *va = info->attrs[DP_GENL_A_OPENFLOW];
1047 struct datapath *dp;
1048 struct ofp_header *oh;
1049 struct sender sender;
1052 if (!info->attrs[DP_GENL_A_DP_IDX] || !va)
1056 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1062 if (nla_len(va) < sizeof(struct ofp_header)) {
1068 sender.xid = oh->xid;
1069 sender.pid = info->snd_pid;
1070 sender.seq = info->snd_seq;
1071 err = fwd_control_input(dp->chain, &sender, nla_data(va), nla_len(va));
1078 static struct nla_policy dp_genl_openflow_policy[DP_GENL_A_MAX + 1] = {
1079 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1082 struct flow_stats_state {
1084 struct sw_table_position position;
1085 const struct ofp_flow_stats_request *rq;
1088 int bytes_used, bytes_allocated;
1091 static int flow_stats_init(struct datapath *dp, const void *body, int body_len,
1094 const struct ofp_flow_stats_request *fsr = body;
1095 struct flow_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1098 s->table_idx = fsr->table_id == 0xff ? 0 : fsr->table_id;
1099 memset(&s->position, 0, sizeof s->position);
1105 static int flow_stats_dump_callback(struct sw_flow *flow, void *private)
1107 struct flow_stats_state *s = private;
1108 struct ofp_flow_stats *ofs;
1112 actions_length = sizeof *ofs->actions * flow->n_actions;
1113 length = sizeof *ofs + sizeof *ofs->actions * flow->n_actions;
1114 if (length + s->bytes_used > s->bytes_allocated)
1117 ofs = s->body + s->bytes_used;
1118 ofs->length = htons(length);
1119 ofs->table_id = s->table_idx;
1121 ofs->match.wildcards = htons(flow->key.wildcards);
1122 ofs->match.in_port = flow->key.in_port;
1123 memcpy(ofs->match.dl_src, flow->key.dl_src, ETH_ALEN);
1124 memcpy(ofs->match.dl_dst, flow->key.dl_dst, ETH_ALEN);
1125 ofs->match.dl_vlan = flow->key.dl_vlan;
1126 ofs->match.dl_type = flow->key.dl_type;
1127 ofs->match.nw_src = flow->key.nw_src;
1128 ofs->match.nw_dst = flow->key.nw_dst;
1129 ofs->match.nw_proto = flow->key.nw_proto;
1130 memset(ofs->match.pad, 0, sizeof ofs->match.pad);
1131 ofs->match.tp_src = flow->key.tp_src;
1132 ofs->match.tp_dst = flow->key.tp_dst;
1133 ofs->duration = htonl((jiffies - flow->init_time) / HZ);
1134 ofs->packet_count = cpu_to_be64(flow->packet_count);
1135 ofs->byte_count = cpu_to_be64(flow->byte_count);
1136 ofs->priority = htons(flow->priority);
1137 ofs->max_idle = htons(flow->max_idle);
1138 memcpy(ofs->actions, flow->actions, actions_length);
1140 s->bytes_used += length;
1144 static int flow_stats_dump(struct datapath *dp, void *state,
1145 void *body, int *body_len)
1147 struct flow_stats_state *s = state;
1148 struct sw_flow_key match_key;
1152 s->bytes_allocated = *body_len;
1155 flow_extract_match(&match_key, &s->rq->match);
1156 while (s->table_idx < dp->chain->n_tables
1157 && (s->rq->table_id == 0xff || s->rq->table_id == s->table_idx))
1159 struct sw_table *table = dp->chain->tables[s->table_idx];
1161 error = table->iterate(table, &match_key, &s->position,
1162 flow_stats_dump_callback, s);
1167 memset(&s->position, 0, sizeof s->position);
1169 *body_len = s->bytes_used;
1171 /* If error is 0, we're done.
1172 * Otherwise, if some bytes were used, there are more flows to come.
1173 * Otherwise, we were not able to fit even a single flow in the body,
1174 * which indicates that we have a single flow with too many actions to
1175 * fit. We won't ever make any progress at that rate, so give up. */
1176 return !error ? 0 : s->bytes_used ? 1 : -ENOMEM;
1179 static void flow_stats_done(void *state)
1184 static int aggregate_stats_init(struct datapath *dp,
1185 const void *body, int body_len,
1188 *state = (void *)body;
1192 static int aggregate_stats_dump_callback(struct sw_flow *flow, void *private)
1194 struct ofp_aggregate_stats_reply *rpy = private;
1195 rpy->packet_count += flow->packet_count;
1196 rpy->byte_count += flow->byte_count;
1201 static int aggregate_stats_dump(struct datapath *dp, void *state,
1202 void *body, int *body_len)
1204 struct ofp_aggregate_stats_request *rq = state;
1205 struct ofp_aggregate_stats_reply *rpy;
1206 struct sw_table_position position;
1207 struct sw_flow_key match_key;
1210 if (*body_len < sizeof *rpy)
1213 *body_len = sizeof *rpy;
1215 memset(rpy, 0, sizeof *rpy);
1217 flow_extract_match(&match_key, &rq->match);
1218 table_idx = rq->table_id == 0xff ? 0 : rq->table_id;
1219 memset(&position, 0, sizeof position);
1220 while (table_idx < dp->chain->n_tables
1221 && (rq->table_id == 0xff || rq->table_id == table_idx))
1223 struct sw_table *table = dp->chain->tables[table_idx];
1226 error = table->iterate(table, &match_key, &position,
1227 aggregate_stats_dump_callback, rpy);
1232 memset(&position, 0, sizeof position);
1235 rpy->packet_count = cpu_to_be64(rpy->packet_count);
1236 rpy->byte_count = cpu_to_be64(rpy->byte_count);
1237 rpy->flow_count = htonl(rpy->flow_count);
1241 static int table_stats_dump(struct datapath *dp, void *state,
1242 void *body, int *body_len)
1244 struct ofp_table_stats *ots;
1245 int nbytes = dp->chain->n_tables * sizeof *ots;
1247 if (nbytes > *body_len)
1250 for (i = 0, ots = body; i < dp->chain->n_tables; i++, ots++) {
1251 struct sw_table_stats stats;
1252 dp->chain->tables[i]->stats(dp->chain->tables[i], &stats);
1253 strncpy(ots->name, stats.name, sizeof ots->name);
1255 memset(ots->pad, 0, sizeof ots->pad);
1256 ots->max_entries = htonl(stats.max_flows);
1257 ots->active_count = htonl(stats.n_flows);
1258 ots->matched_count = cpu_to_be64(0); /* FIXME */
1263 struct port_stats_state {
1267 static int port_stats_init(struct datapath *dp, const void *body, int body_len,
1270 struct port_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1278 static int port_stats_dump(struct datapath *dp, void *state,
1279 void *body, int *body_len)
1281 struct port_stats_state *s = state;
1282 struct ofp_port_stats *ops;
1283 int n_ports, max_ports;
1286 max_ports = *body_len / sizeof *ops;
1292 for (i = s->port; i < OFPP_MAX && n_ports < max_ports; i++) {
1293 struct net_bridge_port *p = dp->ports[i];
1294 struct net_device_stats *stats;
1297 stats = p->dev->get_stats(p->dev);
1298 ops->port_no = htons(p->port_no);
1299 memset(ops->pad, 0, sizeof ops->pad);
1300 ops->rx_count = cpu_to_be64(stats->rx_packets);
1301 ops->tx_count = cpu_to_be64(stats->tx_packets);
1302 ops->drop_count = cpu_to_be64(stats->rx_dropped
1303 + stats->tx_dropped);
1308 *body_len = n_ports * sizeof *ops;
1309 return n_ports >= max_ports;
1312 static void port_stats_done(void *state)
1318 /* Minimum and maximum acceptable number of bytes in body member of
1319 * struct ofp_stats_request. */
1320 size_t min_body, max_body;
1322 /* Prepares to dump some kind of statistics on 'dp'. 'body' and
1323 * 'body_len' are the 'body' member of the struct ofp_stats_request.
1324 * Returns zero if successful, otherwise a negative error code.
1325 * May initialize '*state' to state information. May be null if no
1326 * initialization is required.*/
1327 int (*init)(struct datapath *dp, const void *body, int body_len,
1330 /* Dumps statistics for 'dp' into the '*body_len' bytes at 'body', and
1331 * modifies '*body_len' to reflect the number of bytes actually used.
1332 * ('body' will be transmitted as the 'body' member of struct
1333 * ofp_stats_reply.) */
1334 int (*dump)(struct datapath *dp, void *state,
1335 void *body, int *body_len);
1337 /* Cleans any state created by the init or dump functions. May be null
1338 * if no cleanup is required. */
1339 void (*done)(void *state);
1342 static const struct stats_type stats[] = {
1344 sizeof(struct ofp_flow_stats_request),
1345 sizeof(struct ofp_flow_stats_request),
1350 [OFPST_AGGREGATE] = {
1351 sizeof(struct ofp_aggregate_stats_request),
1352 sizeof(struct ofp_aggregate_stats_request),
1353 aggregate_stats_init,
1354 aggregate_stats_dump,
1374 dp_genl_openflow_dumpit(struct sk_buff *skb, struct netlink_callback *cb)
1376 struct datapath *dp;
1377 struct sender sender;
1378 const struct stats_type *s;
1379 struct ofp_stats_reply *osr;
1381 int max_openflow_len, body_len;
1385 /* Set up the cleanup function for this dump. Linux 2.6.20 and later
1386 * support setting up cleanup functions via the .doneit member of
1387 * struct genl_ops. This kluge supports earlier versions also. */
1388 cb->done = dp_genl_openflow_done;
1392 struct nlattr *attrs[DP_GENL_A_MAX + 1];
1393 struct ofp_stats_request *rq;
1395 size_t len, body_len;
1398 err = nlmsg_parse(cb->nlh, GENL_HDRLEN, attrs, DP_GENL_A_MAX,
1399 dp_genl_openflow_policy);
1405 if (!attrs[DP_GENL_A_DP_IDX])
1407 dp_idx = nla_get_u16(attrs[DP_GENL_A_DP_IDX]);
1408 dp = dp_get(dp_idx);
1414 va = attrs[DP_GENL_A_OPENFLOW];
1416 if (!va || len < sizeof *rq)
1420 type = ntohs(rq->type);
1421 if (rq->header.version != OFP_VERSION
1422 || rq->header.type != OFPT_STATS_REQUEST
1423 || ntohs(rq->header.length) != len
1424 || type >= ARRAY_SIZE(stats)
1425 || !stats[type].dump)
1429 body_len = len - offsetof(struct ofp_stats_request, body);
1430 if (body_len < s->min_body || body_len > s->max_body)
1434 cb->args[1] = dp_idx;
1436 cb->args[3] = rq->header.xid;
1439 err = s->init(dp, rq->body, body_len, &state);
1442 cb->args[4] = (long) state;
1444 } else if (cb->args[0] == 1) {
1445 dp_idx = cb->args[1];
1446 s = &stats[cb->args[2]];
1448 dp = dp_get(dp_idx);
1458 sender.xid = cb->args[3];
1459 sender.pid = NETLINK_CB(cb->skb).pid;
1460 sender.seq = cb->nlh->nlmsg_seq;
1462 osr = put_openflow_headers(dp, skb, OFPT_STATS_REPLY, &sender,
1468 osr->type = htons(s - stats);
1470 resize_openflow_skb(skb, &osr->header, max_openflow_len);
1472 body_len = max_openflow_len - offsetof(struct ofp_stats_reply, body);
1474 err = s->dump(dp, (void *) cb->args[4], body, &body_len);
1479 osr->flags = ntohs(OFPSF_REPLY_MORE);
1480 resize_openflow_skb(skb, &osr->header,
1481 (offsetof(struct ofp_stats_reply, body)
1492 dp_genl_openflow_done(struct netlink_callback *cb)
1495 const struct stats_type *s = &stats[cb->args[2]];
1497 s->done((void *) cb->args[4]);
1502 static struct genl_ops dp_genl_ops_openflow = {
1503 .cmd = DP_GENL_C_OPENFLOW,
1504 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1505 .policy = dp_genl_openflow_policy,
1506 .doit = dp_genl_openflow,
1507 .dumpit = dp_genl_openflow_dumpit,
1510 static struct nla_policy dp_genl_benchmark_policy[DP_GENL_A_MAX + 1] = {
1511 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1512 [DP_GENL_A_NPACKETS] = { .type = NLA_U32 },
1513 [DP_GENL_A_PSIZE] = { .type = NLA_U32 },
1516 static struct genl_ops dp_genl_ops_benchmark_nl = {
1517 .cmd = DP_GENL_C_BENCHMARK_NL,
1518 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1519 .policy = dp_genl_benchmark_policy,
1520 .doit = dp_genl_benchmark_nl,
1524 static struct genl_ops *dp_genl_all_ops[] = {
1525 /* Keep this operation first. Generic Netlink dispatching
1526 * looks up operations with linear search, so we want it at the
1528 &dp_genl_ops_openflow,
1530 &dp_genl_ops_add_dp,
1531 &dp_genl_ops_del_dp,
1532 &dp_genl_ops_query_dp,
1533 &dp_genl_ops_add_port,
1534 &dp_genl_ops_del_port,
1535 &dp_genl_ops_benchmark_nl,
1538 static int dp_init_netlink(void)
1543 err = genl_register_family(&dp_genl_family);
1547 for (i = 0; i < ARRAY_SIZE(dp_genl_all_ops); i++) {
1548 err = genl_register_ops(&dp_genl_family, dp_genl_all_ops[i]);
1550 goto err_unregister;
1553 strcpy(mc_group.name, "openflow");
1554 err = genl_register_mc_group(&dp_genl_family, &mc_group);
1556 goto err_unregister;
1561 genl_unregister_family(&dp_genl_family);
1565 static void dp_uninit_netlink(void)
1567 genl_unregister_family(&dp_genl_family);
1570 #define DRV_NAME "openflow"
1571 #define DRV_VERSION VERSION
1572 #define DRV_DESCRIPTION "OpenFlow switching datapath implementation"
1573 #define DRV_COPYRIGHT "Copyright (c) 2007, 2008 The Board of Trustees of The Leland Stanford Junior University"
1576 static int __init dp_init(void)
1580 printk(KERN_INFO DRV_NAME ": " DRV_DESCRIPTION "\n");
1581 printk(KERN_INFO DRV_NAME ": " VERSION" built on "__DATE__" "__TIME__"\n");
1582 printk(KERN_INFO DRV_NAME ": " DRV_COPYRIGHT "\n");
1588 err = dp_init_netlink();
1590 goto error_flow_exit;
1592 /* Hook into callback used by the bridge to intercept packets.
1593 * Parasites we are. */
1594 if (br_handle_frame_hook)
1595 printk("openflow: hijacking bridge hook\n");
1596 br_handle_frame_hook = dp_frame_hook;
1603 printk(KERN_EMERG "openflow: failed to install!");
1607 static void dp_cleanup(void)
1610 dp_uninit_netlink();
1612 br_handle_frame_hook = NULL;
1615 module_init(dp_init);
1616 module_exit(dp_cleanup);
1618 MODULE_DESCRIPTION(DRV_DESCRIPTION);
1619 MODULE_AUTHOR(DRV_COPYRIGHT);
1620 MODULE_LICENSE("GPL");