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 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
444 /* Push the Ethernet header back on. */
445 if (skb->protocol == htons(ETH_P_8021Q))
446 skb_push(skb, VLAN_ETH_HLEN);
448 skb_push(skb, ETH_HLEN);
449 fwd_port_input(p->dp->chain, skb, p->port_no);
453 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
454 * different set of devices!)
456 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
457 /* Called with rcu_read_lock. */
458 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
461 do_port_input(p, skb);
464 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
465 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
467 do_port_input(p, *pskb);
471 /* NB: This has only been tested on 2.4.35 */
473 /* Called without any locks (?) */
474 static void dp_frame_hook(struct sk_buff *skb)
476 struct net_bridge_port *p = skb->dev->br_port;
479 do_port_input(p, skb);
486 /* Forwarding output path.
487 * Based on net/bridge/br_forward.c. */
489 /* Don't forward packets to originating port. If we're flooding,
490 * then don't send out ports with flooding disabled.
492 static inline int should_deliver(const struct net_bridge_port *p,
493 const struct sk_buff *skb, int flood)
495 if (skb->dev == p->dev)
498 if (flood && (p->flags & BRIDGE_PORT_NO_FLOOD))
504 static inline unsigned packet_length(const struct sk_buff *skb)
506 int length = skb->len - ETH_HLEN;
507 if (skb->protocol == htons(ETH_P_8021Q))
512 /* Send packets out all the ports except the originating one. If the
513 * "flood" argument is set, only send along the minimum spanning tree.
516 output_all(struct datapath *dp, struct sk_buff *skb, int flood)
518 struct net_bridge_port *p;
522 list_for_each_entry_rcu (p, &dp->port_list, node) {
523 if (!should_deliver(p, skb, flood))
525 if (prev_port != -1) {
526 struct sk_buff *clone = skb_clone(skb, GFP_ATOMIC);
531 dp_output_port(dp, clone, prev_port);
533 prev_port = p->port_no;
536 dp_output_port(dp, skb, prev_port);
543 /* Marks 'skb' as having originated from 'in_port' in 'dp'.
544 FIXME: how are devices reference counted? */
545 int dp_set_origin(struct datapath *dp, uint16_t in_port,
548 if (in_port < OFPP_MAX && dp->ports[in_port]) {
549 skb->dev = dp->ports[in_port]->dev;
555 /* Takes ownership of 'skb' and transmits it to 'out_port' on 'dp'.
557 int dp_output_port(struct datapath *dp, struct sk_buff *skb, int out_port)
559 struct net_bridge_port *p;
563 if (out_port == OFPP_FLOOD)
564 return output_all(dp, skb, 1);
565 else if (out_port == OFPP_ALL)
566 return output_all(dp, skb, 0);
567 else if (out_port == OFPP_CONTROLLER)
568 return dp_output_control(dp, skb, fwd_save_skb(skb), 0,
570 else if (out_port == OFPP_TABLE) {
571 struct sw_flow_key key;
572 struct sw_flow *flow;
574 flow_extract(skb, skb->dev->br_port->port_no, &key);
575 flow = chain_lookup(dp->chain, &key);
576 if (likely(flow != NULL)) {
577 flow_used(flow, skb);
578 execute_actions(dp, skb, &key, flow->actions, flow->n_actions);
582 } else if (out_port >= OFPP_MAX)
585 p = dp->ports[out_port];
590 if (packet_length(skb) > skb->dev->mtu) {
591 printk("dropped over-mtu packet: %d > %d\n",
592 packet_length(skb), skb->dev->mtu);
604 printk("can't forward to bad port %d\n", out_port);
608 /* Takes ownership of 'skb' and transmits it to 'dp''s control path. If
609 * 'buffer_id' != -1, then only the first 64 bytes of 'skb' are sent;
610 * otherwise, all of 'skb' is sent. 'reason' indicates why 'skb' is being
611 * sent. 'max_len' sets the maximum number of bytes that the caller
612 * wants to be sent; a value of 0 indicates the entire packet should be
615 dp_output_control(struct datapath *dp, struct sk_buff *skb,
616 uint32_t buffer_id, size_t max_len, int reason)
618 /* FIXME? Can we avoid creating a new skbuff in the case where we
619 * forward the whole packet? */
620 struct sk_buff *f_skb;
621 struct ofp_packet_in *opi;
622 size_t fwd_len, opi_len;
626 if ((buffer_id != (uint32_t) -1) && max_len)
627 fwd_len = min(fwd_len, max_len);
629 opi_len = offsetof(struct ofp_packet_in, data) + fwd_len;
630 opi = alloc_openflow_skb(dp, opi_len, OFPT_PACKET_IN, NULL, &f_skb);
635 opi->buffer_id = htonl(buffer_id);
636 opi->total_len = htons(skb->len);
637 opi->in_port = htons(skb->dev->br_port->port_no);
638 opi->reason = reason;
640 memcpy(opi->data, skb_mac_header(skb), fwd_len);
641 err = send_openflow_skb(f_skb, NULL);
648 static void fill_port_desc(struct net_bridge_port *p, struct ofp_phy_port *desc)
650 desc->port_no = htons(p->port_no);
651 strncpy(desc->name, p->dev->name, OFP_MAX_PORT_NAME_LEN);
652 desc->name[OFP_MAX_PORT_NAME_LEN-1] = '\0';
653 memcpy(desc->hw_addr, p->dev->dev_addr, ETH_ALEN);
654 desc->flags = htonl(p->flags);
658 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,24)
659 if (p->dev->ethtool_ops && p->dev->ethtool_ops->get_settings) {
660 struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
662 if (!p->dev->ethtool_ops->get_settings(p->dev, &ecmd)) {
663 if (ecmd.supported & SUPPORTED_10baseT_Half)
664 desc->features |= OFPPF_10MB_HD;
665 if (ecmd.supported & SUPPORTED_10baseT_Full)
666 desc->features |= OFPPF_10MB_FD;
667 if (ecmd.supported & SUPPORTED_100baseT_Half)
668 desc->features |= OFPPF_100MB_HD;
669 if (ecmd.supported & SUPPORTED_100baseT_Full)
670 desc->features |= OFPPF_100MB_FD;
671 if (ecmd.supported & SUPPORTED_1000baseT_Half)
672 desc->features |= OFPPF_1GB_HD;
673 if (ecmd.supported & SUPPORTED_1000baseT_Full)
674 desc->features |= OFPPF_1GB_FD;
675 /* 10Gbps half-duplex doesn't exist... */
676 if (ecmd.supported & SUPPORTED_10000baseT_Full)
677 desc->features |= OFPPF_10GB_FD;
679 desc->features = htonl(desc->features);
680 desc->speed = htonl(ecmd.speed);
687 fill_features_reply(struct datapath *dp, struct ofp_switch_features *ofr)
689 struct net_bridge_port *p;
692 ofr->datapath_id = cpu_to_be64(dp->id);
694 ofr->n_exact = htonl(2 * TABLE_HASH_MAX_FLOWS);
695 ofr->n_compression = 0; /* Not supported */
696 ofr->n_general = htonl(TABLE_LINEAR_MAX_FLOWS);
697 ofr->buffer_mb = htonl(UINT32_MAX);
698 ofr->n_buffers = htonl(N_PKT_BUFFERS);
699 ofr->capabilities = htonl(OFP_SUPPORTED_CAPABILITIES);
700 ofr->actions = htonl(OFP_SUPPORTED_ACTIONS);
702 list_for_each_entry_rcu (p, &dp->port_list, node) {
703 fill_port_desc(p, &ofr->ports[port_count]);
711 dp_send_features_reply(struct datapath *dp, const struct sender *sender)
714 struct ofp_switch_features *ofr;
715 size_t ofr_len, port_max_len;
719 port_max_len = sizeof(struct ofp_phy_port) * OFPP_MAX;
720 ofr = alloc_openflow_skb(dp, sizeof(*ofr) + port_max_len,
721 OFPT_FEATURES_REPLY, sender, &skb);
726 port_count = fill_features_reply(dp, ofr);
729 ofr_len = sizeof(*ofr) + (sizeof(struct ofp_phy_port) * port_count);
730 resize_openflow_skb(skb, &ofr->header, ofr_len);
731 return send_openflow_skb(skb, sender);
735 dp_send_config_reply(struct datapath *dp, const struct sender *sender)
738 struct ofp_switch_config *osc;
740 osc = alloc_openflow_skb(dp, sizeof *osc, OFPT_GET_CONFIG_REPLY, sender,
745 osc->flags = htons(dp->flags);
746 osc->miss_send_len = htons(dp->miss_send_len);
748 return send_openflow_skb(skb, sender);
752 dp_update_port_flags(struct datapath *dp, const struct ofp_phy_port *opp)
754 struct net_bridge_port *p;
756 p = dp->ports[htons(opp->port_no)];
758 /* Make sure the port id hasn't changed since this was sent */
759 if (!p || memcmp(opp->hw_addr, p->dev->dev_addr, ETH_ALEN) != 0)
762 p->flags = htonl(opp->flags);
769 send_port_status(struct net_bridge_port *p, uint8_t status)
772 struct ofp_port_status *ops;
774 ops = alloc_openflow_skb(p->dp, sizeof *ops, OFPT_PORT_STATUS, NULL,
778 ops->reason = status;
779 memset(ops->pad, 0, sizeof ops->pad);
780 fill_port_desc(p, &ops->desc);
782 return send_openflow_skb(skb, NULL);
786 dp_send_flow_expired(struct datapath *dp, struct sw_flow *flow)
789 struct ofp_flow_expired *ofe;
790 unsigned long duration_j;
792 ofe = alloc_openflow_skb(dp, sizeof *ofe, OFPT_FLOW_EXPIRED, 0, &skb);
796 flow_fill_match(&ofe->match, &flow->key);
798 memset(ofe->pad, 0, sizeof ofe->pad);
799 ofe->priority = htons(flow->priority);
801 duration_j = (flow->timeout - HZ * flow->max_idle) - flow->init_time;
802 ofe->duration = htonl(duration_j / HZ);
803 ofe->packet_count = cpu_to_be64(flow->packet_count);
804 ofe->byte_count = cpu_to_be64(flow->byte_count);
806 return send_openflow_skb(skb, NULL);
808 EXPORT_SYMBOL(dp_send_flow_expired);
811 dp_send_error_msg(struct datapath *dp, const struct sender *sender,
812 uint16_t type, uint16_t code, const uint8_t *data, size_t len)
815 struct ofp_error_msg *oem;
818 oem = alloc_openflow_skb(dp, sizeof(*oem)+len, OFPT_ERROR_MSG,
823 oem->type = htons(type);
824 oem->code = htons(code);
825 memcpy(oem->data, data, len);
827 return send_openflow_skb(skb, sender);
830 /* Generic Netlink interface.
832 * See netlink(7) for an introduction to netlink. See
833 * http://linux-net.osdl.org/index.php/Netlink for more information and
834 * pointers on how to work with netlink and Generic Netlink in the kernel and
837 static struct genl_family dp_genl_family = {
838 .id = GENL_ID_GENERATE,
840 .name = DP_GENL_FAMILY_NAME,
842 .maxattr = DP_GENL_A_MAX,
845 /* Attribute policy: what each attribute may contain. */
846 static struct nla_policy dp_genl_policy[DP_GENL_A_MAX + 1] = {
847 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
848 [DP_GENL_A_MC_GROUP] = { .type = NLA_U32 },
849 [DP_GENL_A_PORTNAME] = { .type = NLA_STRING }
852 static int dp_genl_add(struct sk_buff *skb, struct genl_info *info)
854 if (!info->attrs[DP_GENL_A_DP_IDX])
857 return new_dp(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
860 static struct genl_ops dp_genl_ops_add_dp = {
861 .cmd = DP_GENL_C_ADD_DP,
862 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
863 .policy = dp_genl_policy,
868 struct datapath *dp_get(int dp_idx)
870 if (dp_idx < 0 || dp_idx > DP_MAX)
872 return rcu_dereference(dps[dp_idx]);
875 static int dp_genl_del(struct sk_buff *skb, struct genl_info *info)
880 if (!info->attrs[DP_GENL_A_DP_IDX])
883 mutex_lock(&dp_mutex);
884 dp = dp_get(nla_get_u32((info->attrs[DP_GENL_A_DP_IDX])));
891 mutex_unlock(&dp_mutex);
895 static struct genl_ops dp_genl_ops_del_dp = {
896 .cmd = DP_GENL_C_DEL_DP,
897 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
898 .policy = dp_genl_policy,
903 /* Queries a datapath for related information. Currently the only relevant
904 * information is the datapath's multicast group ID. Really we want one
905 * multicast group per datapath, but because of locking issues[*] we can't
906 * easily get one. Thus, every datapath will currently return the same
907 * global multicast group ID, but in the future it would be nice to fix that.
909 * [*] dp_genl_add, to add a new datapath, is called under the genl_lock
910 * mutex, and genl_register_mc_group, called to acquire a new multicast
911 * group ID, also acquires genl_lock, thus deadlock.
913 static int dp_genl_query(struct sk_buff *skb, struct genl_info *info)
916 struct sk_buff *ans_skb = NULL;
920 if (!info->attrs[DP_GENL_A_DP_IDX])
924 dp_idx = nla_get_u32((info->attrs[DP_GENL_A_DP_IDX]));
930 ans_skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
935 data = genlmsg_put_reply(ans_skb, info, &dp_genl_family,
936 0, DP_GENL_C_QUERY_DP);
941 NLA_PUT_U32(ans_skb, DP_GENL_A_DP_IDX, dp_idx);
942 NLA_PUT_U32(ans_skb, DP_GENL_A_MC_GROUP, mc_group.id);
944 genlmsg_end(ans_skb, data);
945 err = genlmsg_reply(ans_skb, info);
957 static struct genl_ops dp_genl_ops_query_dp = {
958 .cmd = DP_GENL_C_QUERY_DP,
959 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
960 .policy = dp_genl_policy,
961 .doit = dp_genl_query,
965 static int dp_genl_add_del_port(struct sk_buff *skb, struct genl_info *info)
968 struct net_device *port;
971 if (!info->attrs[DP_GENL_A_DP_IDX] || !info->attrs[DP_GENL_A_PORTNAME])
975 mutex_lock(&dp_mutex);
976 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
982 /* Get interface to add/remove. */
983 port = dev_get_by_name(&init_net,
984 nla_data(info->attrs[DP_GENL_A_PORTNAME]));
990 /* Execute operation. */
991 if (info->genlhdr->cmd == DP_GENL_C_ADD_PORT)
992 err = add_switch_port(dp, port);
994 if (port->br_port == NULL || port->br_port->dp != dp) {
998 err = del_switch_port(port->br_port);
1004 mutex_unlock(&dp_mutex);
1008 static struct genl_ops dp_genl_ops_add_port = {
1009 .cmd = DP_GENL_C_ADD_PORT,
1010 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1011 .policy = dp_genl_policy,
1012 .doit = dp_genl_add_del_port,
1016 static struct genl_ops dp_genl_ops_del_port = {
1017 .cmd = DP_GENL_C_DEL_PORT,
1018 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1019 .policy = dp_genl_policy,
1020 .doit = dp_genl_add_del_port,
1024 static int dp_genl_openflow(struct sk_buff *skb, struct genl_info *info)
1026 struct nlattr *va = info->attrs[DP_GENL_A_OPENFLOW];
1027 struct datapath *dp;
1028 struct ofp_header *oh;
1029 struct sender sender;
1032 if (!info->attrs[DP_GENL_A_DP_IDX] || !va)
1036 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1042 if (nla_len(va) < sizeof(struct ofp_header)) {
1048 sender.xid = oh->xid;
1049 sender.pid = info->snd_pid;
1050 sender.seq = info->snd_seq;
1051 err = fwd_control_input(dp->chain, &sender, nla_data(va), nla_len(va));
1058 static struct nla_policy dp_genl_openflow_policy[DP_GENL_A_MAX + 1] = {
1059 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1062 struct flow_stats_state {
1064 struct sw_table_position position;
1065 const struct ofp_flow_stats_request *rq;
1068 int bytes_used, bytes_allocated;
1071 static int flow_stats_init(struct datapath *dp, const void *body, int body_len,
1074 const struct ofp_flow_stats_request *fsr = body;
1075 struct flow_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1078 s->table_idx = fsr->table_id == 0xff ? 0 : fsr->table_id;
1079 memset(&s->position, 0, sizeof s->position);
1085 static int flow_stats_dump_callback(struct sw_flow *flow, void *private)
1087 struct flow_stats_state *s = private;
1088 struct ofp_flow_stats *ofs;
1092 actions_length = sizeof *ofs->actions * flow->n_actions;
1093 length = sizeof *ofs + sizeof *ofs->actions * flow->n_actions;
1094 if (length + s->bytes_used > s->bytes_allocated)
1097 ofs = s->body + s->bytes_used;
1098 ofs->length = htons(length);
1099 ofs->table_id = s->table_idx;
1101 ofs->match.wildcards = htons(flow->key.wildcards);
1102 ofs->match.in_port = flow->key.in_port;
1103 memcpy(ofs->match.dl_src, flow->key.dl_src, ETH_ALEN);
1104 memcpy(ofs->match.dl_dst, flow->key.dl_dst, ETH_ALEN);
1105 ofs->match.dl_vlan = flow->key.dl_vlan;
1106 ofs->match.dl_type = flow->key.dl_type;
1107 ofs->match.nw_src = flow->key.nw_src;
1108 ofs->match.nw_dst = flow->key.nw_dst;
1109 ofs->match.nw_proto = flow->key.nw_proto;
1110 memset(ofs->match.pad, 0, sizeof ofs->match.pad);
1111 ofs->match.tp_src = flow->key.tp_src;
1112 ofs->match.tp_dst = flow->key.tp_dst;
1113 ofs->duration = htonl((jiffies - flow->init_time) / HZ);
1114 ofs->packet_count = cpu_to_be64(flow->packet_count);
1115 ofs->byte_count = cpu_to_be64(flow->byte_count);
1116 ofs->priority = htons(flow->priority);
1117 ofs->max_idle = htons(flow->max_idle);
1118 memcpy(ofs->actions, flow->actions, actions_length);
1120 s->bytes_used += length;
1124 static int flow_stats_dump(struct datapath *dp, void *state,
1125 void *body, int *body_len)
1127 struct flow_stats_state *s = state;
1128 struct sw_flow_key match_key;
1132 s->bytes_allocated = *body_len;
1135 flow_extract_match(&match_key, &s->rq->match);
1136 while (s->table_idx < dp->chain->n_tables
1137 && (s->rq->table_id == 0xff || s->rq->table_id == s->table_idx))
1139 struct sw_table *table = dp->chain->tables[s->table_idx];
1141 error = table->iterate(table, &match_key, &s->position,
1142 flow_stats_dump_callback, s);
1147 memset(&s->position, 0, sizeof s->position);
1149 *body_len = s->bytes_used;
1151 /* If error is 0, we're done.
1152 * Otherwise, if some bytes were used, there are more flows to come.
1153 * Otherwise, we were not able to fit even a single flow in the body,
1154 * which indicates that we have a single flow with too many actions to
1155 * fit. We won't ever make any progress at that rate, so give up. */
1156 return !error ? 0 : s->bytes_used ? 1 : -ENOMEM;
1159 static void flow_stats_done(void *state)
1164 static int aggregate_stats_init(struct datapath *dp,
1165 const void *body, int body_len,
1168 *state = (void *)body;
1172 static int aggregate_stats_dump_callback(struct sw_flow *flow, void *private)
1174 struct ofp_aggregate_stats_reply *rpy = private;
1175 rpy->packet_count += flow->packet_count;
1176 rpy->byte_count += flow->byte_count;
1181 static int aggregate_stats_dump(struct datapath *dp, void *state,
1182 void *body, int *body_len)
1184 struct ofp_aggregate_stats_request *rq = state;
1185 struct ofp_aggregate_stats_reply *rpy;
1186 struct sw_table_position position;
1187 struct sw_flow_key match_key;
1190 if (*body_len < sizeof *rpy)
1193 *body_len = sizeof *rpy;
1195 memset(rpy, 0, sizeof *rpy);
1197 flow_extract_match(&match_key, &rq->match);
1198 table_idx = rq->table_id == 0xff ? 0 : rq->table_id;
1199 memset(&position, 0, sizeof position);
1200 while (table_idx < dp->chain->n_tables
1201 && (rq->table_id == 0xff || rq->table_id == table_idx))
1203 struct sw_table *table = dp->chain->tables[table_idx];
1206 error = table->iterate(table, &match_key, &position,
1207 aggregate_stats_dump_callback, rpy);
1212 memset(&position, 0, sizeof position);
1215 rpy->packet_count = cpu_to_be64(rpy->packet_count);
1216 rpy->byte_count = cpu_to_be64(rpy->byte_count);
1217 rpy->flow_count = htonl(rpy->flow_count);
1221 static int table_stats_dump(struct datapath *dp, void *state,
1222 void *body, int *body_len)
1224 struct ofp_table_stats *ots;
1225 int nbytes = dp->chain->n_tables * sizeof *ots;
1227 if (nbytes > *body_len)
1230 for (i = 0, ots = body; i < dp->chain->n_tables; i++, ots++) {
1231 struct sw_table_stats stats;
1232 dp->chain->tables[i]->stats(dp->chain->tables[i], &stats);
1233 strncpy(ots->name, stats.name, sizeof ots->name);
1235 memset(ots->pad, 0, sizeof ots->pad);
1236 ots->max_entries = htonl(stats.max_flows);
1237 ots->active_count = htonl(stats.n_flows);
1238 ots->matched_count = cpu_to_be64(0); /* FIXME */
1243 struct port_stats_state {
1247 static int port_stats_init(struct datapath *dp, const void *body, int body_len,
1250 struct port_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1258 static int port_stats_dump(struct datapath *dp, void *state,
1259 void *body, int *body_len)
1261 struct port_stats_state *s = state;
1262 struct ofp_port_stats *ops;
1263 int n_ports, max_ports;
1266 max_ports = *body_len / sizeof *ops;
1272 for (i = s->port; i < OFPP_MAX && n_ports < max_ports; i++) {
1273 struct net_bridge_port *p = dp->ports[i];
1274 struct net_device_stats *stats;
1277 stats = p->dev->get_stats(p->dev);
1278 ops->port_no = htons(p->port_no);
1279 memset(ops->pad, 0, sizeof ops->pad);
1280 ops->rx_count = cpu_to_be64(stats->rx_packets);
1281 ops->tx_count = cpu_to_be64(stats->tx_packets);
1282 ops->drop_count = cpu_to_be64(stats->rx_dropped
1283 + stats->tx_dropped);
1288 *body_len = n_ports * sizeof *ops;
1289 return n_ports >= max_ports;
1292 static void port_stats_done(void *state)
1298 /* Minimum and maximum acceptable number of bytes in body member of
1299 * struct ofp_stats_request. */
1300 size_t min_body, max_body;
1302 /* Prepares to dump some kind of statistics on 'dp'. 'body' and
1303 * 'body_len' are the 'body' member of the struct ofp_stats_request.
1304 * Returns zero if successful, otherwise a negative error code.
1305 * May initialize '*state' to state information. May be null if no
1306 * initialization is required.*/
1307 int (*init)(struct datapath *dp, const void *body, int body_len,
1310 /* Dumps statistics for 'dp' into the '*body_len' bytes at 'body', and
1311 * modifies '*body_len' to reflect the number of bytes actually used.
1312 * ('body' will be transmitted as the 'body' member of struct
1313 * ofp_stats_reply.) */
1314 int (*dump)(struct datapath *dp, void *state,
1315 void *body, int *body_len);
1317 /* Cleans any state created by the init or dump functions. May be null
1318 * if no cleanup is required. */
1319 void (*done)(void *state);
1322 static const struct stats_type stats[] = {
1324 sizeof(struct ofp_flow_stats_request),
1325 sizeof(struct ofp_flow_stats_request),
1330 [OFPST_AGGREGATE] = {
1331 sizeof(struct ofp_aggregate_stats_request),
1332 sizeof(struct ofp_aggregate_stats_request),
1333 aggregate_stats_init,
1334 aggregate_stats_dump,
1354 dp_genl_openflow_dumpit(struct sk_buff *skb, struct netlink_callback *cb)
1356 struct datapath *dp;
1357 struct sender sender;
1358 const struct stats_type *s;
1359 struct ofp_stats_reply *osr;
1361 int max_openflow_len, body_len;
1365 /* Set up the cleanup function for this dump. Linux 2.6.20 and later
1366 * support setting up cleanup functions via the .doneit member of
1367 * struct genl_ops. This kluge supports earlier versions also. */
1368 cb->done = dp_genl_openflow_done;
1372 struct nlattr *attrs[DP_GENL_A_MAX + 1];
1373 struct ofp_stats_request *rq;
1375 size_t len, body_len;
1378 err = nlmsg_parse(cb->nlh, GENL_HDRLEN, attrs, DP_GENL_A_MAX,
1379 dp_genl_openflow_policy);
1385 if (!attrs[DP_GENL_A_DP_IDX])
1387 dp_idx = nla_get_u16(attrs[DP_GENL_A_DP_IDX]);
1388 dp = dp_get(dp_idx);
1394 va = attrs[DP_GENL_A_OPENFLOW];
1396 if (!va || len < sizeof *rq)
1400 type = ntohs(rq->type);
1401 if (rq->header.version != OFP_VERSION
1402 || rq->header.type != OFPT_STATS_REQUEST
1403 || ntohs(rq->header.length) != len
1404 || type >= ARRAY_SIZE(stats)
1405 || !stats[type].dump)
1409 body_len = len - offsetof(struct ofp_stats_request, body);
1410 if (body_len < s->min_body || body_len > s->max_body)
1414 cb->args[1] = dp_idx;
1416 cb->args[3] = rq->header.xid;
1419 err = s->init(dp, rq->body, body_len, &state);
1422 cb->args[4] = (long) state;
1424 } else if (cb->args[0] == 1) {
1425 dp_idx = cb->args[1];
1426 s = &stats[cb->args[2]];
1428 dp = dp_get(dp_idx);
1438 sender.xid = cb->args[3];
1439 sender.pid = NETLINK_CB(cb->skb).pid;
1440 sender.seq = cb->nlh->nlmsg_seq;
1442 osr = put_openflow_headers(dp, skb, OFPT_STATS_REPLY, &sender,
1448 osr->type = htons(s - stats);
1450 resize_openflow_skb(skb, &osr->header, max_openflow_len);
1452 body_len = max_openflow_len - offsetof(struct ofp_stats_reply, body);
1454 err = s->dump(dp, (void *) cb->args[4], body, &body_len);
1459 osr->flags = ntohs(OFPSF_REPLY_MORE);
1460 resize_openflow_skb(skb, &osr->header,
1461 (offsetof(struct ofp_stats_reply, body)
1472 dp_genl_openflow_done(struct netlink_callback *cb)
1475 const struct stats_type *s = &stats[cb->args[2]];
1477 s->done((void *) cb->args[4]);
1482 static struct genl_ops dp_genl_ops_openflow = {
1483 .cmd = DP_GENL_C_OPENFLOW,
1484 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1485 .policy = dp_genl_openflow_policy,
1486 .doit = dp_genl_openflow,
1487 .dumpit = dp_genl_openflow_dumpit,
1490 static struct nla_policy dp_genl_benchmark_policy[DP_GENL_A_MAX + 1] = {
1491 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1492 [DP_GENL_A_NPACKETS] = { .type = NLA_U32 },
1493 [DP_GENL_A_PSIZE] = { .type = NLA_U32 },
1496 static struct genl_ops dp_genl_ops_benchmark_nl = {
1497 .cmd = DP_GENL_C_BENCHMARK_NL,
1498 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1499 .policy = dp_genl_benchmark_policy,
1500 .doit = dp_genl_benchmark_nl,
1504 static struct genl_ops *dp_genl_all_ops[] = {
1505 /* Keep this operation first. Generic Netlink dispatching
1506 * looks up operations with linear search, so we want it at the
1508 &dp_genl_ops_openflow,
1510 &dp_genl_ops_add_dp,
1511 &dp_genl_ops_del_dp,
1512 &dp_genl_ops_query_dp,
1513 &dp_genl_ops_add_port,
1514 &dp_genl_ops_del_port,
1515 &dp_genl_ops_benchmark_nl,
1518 static int dp_init_netlink(void)
1523 err = genl_register_family(&dp_genl_family);
1527 for (i = 0; i < ARRAY_SIZE(dp_genl_all_ops); i++) {
1528 err = genl_register_ops(&dp_genl_family, dp_genl_all_ops[i]);
1530 goto err_unregister;
1533 strcpy(mc_group.name, "openflow");
1534 err = genl_register_mc_group(&dp_genl_family, &mc_group);
1536 goto err_unregister;
1541 genl_unregister_family(&dp_genl_family);
1545 static void dp_uninit_netlink(void)
1547 genl_unregister_family(&dp_genl_family);
1550 #define DRV_NAME "openflow"
1551 #define DRV_VERSION VERSION
1552 #define DRV_DESCRIPTION "OpenFlow switching datapath implementation"
1553 #define DRV_COPYRIGHT "Copyright (c) 2007, 2008 The Board of Trustees of The Leland Stanford Junior University"
1556 static int __init dp_init(void)
1560 printk(KERN_INFO DRV_NAME ": " DRV_DESCRIPTION "\n");
1561 printk(KERN_INFO DRV_NAME ": " VERSION" built on "__DATE__" "__TIME__"\n");
1562 printk(KERN_INFO DRV_NAME ": " DRV_COPYRIGHT "\n");
1568 err = dp_init_netlink();
1570 goto error_flow_exit;
1572 /* Hook into callback used by the bridge to intercept packets.
1573 * Parasites we are. */
1574 if (br_handle_frame_hook)
1575 printk("openflow: hijacking bridge hook\n");
1576 br_handle_frame_hook = dp_frame_hook;
1583 printk(KERN_EMERG "openflow: failed to install!");
1587 static void dp_cleanup(void)
1590 dp_uninit_netlink();
1592 br_handle_frame_hook = NULL;
1595 module_init(dp_init);
1596 module_exit(dp_cleanup);
1598 MODULE_DESCRIPTION(DRV_DESCRIPTION);
1599 MODULE_AUTHOR(DRV_COPYRIGHT);
1600 MODULE_LICENSE("GPL");