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"
39 #include "datapath_t.h"
44 /* Number of milliseconds between runs of the maintenance thread. */
45 #define MAINT_SLEEP_MSECS 1000
47 #define BRIDGE_PORT_NO_FLOOD 0x00000001
49 #define UINT32_MAX 4294967295U
50 #define UINT16_MAX 65535
51 #define MAX(X, Y) ((X) > (Y) ? (X) : (Y))
53 struct net_bridge_port {
57 struct net_device *dev;
58 struct list_head node; /* Element in datapath.ports. */
61 static struct genl_family dp_genl_family;
62 static struct genl_multicast_group mc_group;
64 /* It's hard to imagine wanting more than one datapath, but... */
67 /* datapaths. Protected on the read side by rcu_read_lock, on the write side
70 * It is safe to access the datapath and net_bridge_port structures with just
71 * the dp_mutex, but to access the chain you need to take the rcu_read_lock
72 * also (because dp_mutex doesn't prevent flows from being destroyed).
74 static struct datapath *dps[DP_MAX];
75 static DEFINE_MUTEX(dp_mutex);
77 static int dp_maint_func(void *data);
78 static int send_port_status(struct net_bridge_port *p, uint8_t status);
79 static int dp_genl_openflow_done(struct netlink_callback *);
80 static struct net_bridge_port *new_nbp(struct datapath *,
81 struct net_device *, int port_no);
82 static int del_switch_port(struct net_bridge_port *);
84 /* nla_shrink - reduce amount of space reserved by nla_reserve
85 * @skb: socket buffer from which to recover room
86 * @nla: netlink attribute to adjust
87 * @len: new length of attribute payload
89 * Reduces amount of space reserved by a call to nla_reserve.
91 * No other attributes may be added between calling nla_reserve and this
92 * function, since it will create a hole in the message.
94 void nla_shrink(struct sk_buff *skb, struct nlattr *nla, int len)
96 int delta = nla_total_size(len) - nla_total_size(nla_len(nla));
100 nla->nla_len = nla_attr_size(len);
103 /* Puts a set of openflow headers for a message of the given 'type' into 'skb'.
104 * If 'sender' is nonnull, then it is used as the message's destination. 'dp'
105 * must specify the datapath to use.
107 * '*max_openflow_len' receives the maximum number of bytes that are available
108 * for the embedded OpenFlow message. The caller must call
109 * resize_openflow_skb() to set the actual size of the message to this number
112 * Returns the openflow header if successful, otherwise (if 'skb' is too small)
115 put_openflow_headers(struct datapath *dp, struct sk_buff *skb, uint8_t type,
116 const struct sender *sender, int *max_openflow_len)
118 struct ofp_header *oh;
122 /* Assemble the Generic Netlink wrapper. */
123 if (!genlmsg_put(skb,
124 sender ? sender->pid : 0,
125 sender ? sender->seq : 0,
126 &dp_genl_family, 0, DP_GENL_C_OPENFLOW))
127 return ERR_PTR(-ENOBUFS);
128 if (nla_put_u32(skb, DP_GENL_A_DP_IDX, dp->dp_idx) < 0)
129 return ERR_PTR(-ENOBUFS);
130 openflow_len = (skb_tailroom(skb) - NLA_HDRLEN) & ~(NLA_ALIGNTO - 1);
131 if (openflow_len < sizeof *oh)
132 return ERR_PTR(-ENOBUFS);
133 *max_openflow_len = openflow_len;
134 attr = nla_reserve(skb, DP_GENL_A_OPENFLOW, openflow_len);
137 /* Fill in the header. The caller is responsible for the length. */
139 oh->version = OFP_VERSION;
141 oh->xid = sender ? sender->xid : 0;
146 /* Resizes OpenFlow header 'oh', which must be at the tail end of 'skb', to new
147 * length 'new_length' (in bytes), adjusting pointers and size values as
150 resize_openflow_skb(struct sk_buff *skb,
151 struct ofp_header *oh, size_t new_length)
153 struct nlattr *attr = ((void *) oh) - NLA_HDRLEN;
154 nla_shrink(skb, attr, new_length);
155 oh->length = htons(new_length);
156 nlmsg_end(skb, (struct nlmsghdr *) skb->data);
159 /* Allocates a new skb to contain an OpenFlow message 'openflow_len' bytes in
160 * length. Returns a null pointer if memory is unavailable, otherwise returns
161 * the OpenFlow header and stores a pointer to the skb in '*pskb'.
163 * 'type' is the OpenFlow message type. If 'sender' is nonnull, then it is
164 * used as the message's destination. 'dp' must specify the datapath to
167 alloc_openflow_skb(struct datapath *dp, size_t openflow_len, uint8_t type,
168 const struct sender *sender, struct sk_buff **pskb)
170 struct ofp_header *oh;
173 int max_openflow_len;
175 if ((openflow_len + sizeof(struct ofp_header)) > UINT16_MAX) {
177 printk("alloc_openflow_skb: openflow message too large: %zu\n",
182 genl_len = nlmsg_total_size(GENL_HDRLEN + dp_genl_family.hdrsize);
183 genl_len += nla_total_size(sizeof(uint32_t)); /* DP_GENL_A_DP_IDX */
184 genl_len += nla_total_size(openflow_len); /* DP_GENL_A_OPENFLOW */
185 skb = *pskb = genlmsg_new(genl_len, GFP_ATOMIC);
188 printk("alloc_openflow_skb: genlmsg_new failed\n");
192 oh = put_openflow_headers(dp, skb, type, sender, &max_openflow_len);
193 BUG_ON(!oh || IS_ERR(oh));
194 resize_openflow_skb(skb, oh, openflow_len);
199 /* Sends 'skb' to 'sender' if it is nonnull, otherwise multicasts 'skb' to all
202 send_openflow_skb(struct sk_buff *skb, const struct sender *sender)
205 ? genlmsg_unicast(skb, sender->pid)
206 : genlmsg_multicast(skb, 0, mc_group.id, GFP_ATOMIC));
207 if (err && net_ratelimit())
208 printk(KERN_WARNING "send_openflow_skb: send failed: %d\n",
213 /* Generates a unique datapath id. It incorporates the datapath index
214 * and a hardware address, if available. If not, it generates a random
218 uint64_t gen_datapath_id(uint16_t dp_idx)
222 struct net_device *dev;
224 /* The top 16 bits are used to identify the datapath. The lower 48 bits
225 * use an interface address. */
226 id = (uint64_t)dp_idx << 48;
227 if ((dev = dev_get_by_name(&init_net, "ctl0"))
228 || (dev = dev_get_by_name(&init_net, "eth0"))) {
229 for (i=0; i<ETH_ALEN; i++) {
230 id |= (uint64_t)dev->dev_addr[i] << (8*(ETH_ALEN-1 - i));
234 /* Randomly choose the lower 48 bits if we cannot find an
235 * address and mark the most significant bit to indicate that
236 * this was randomly generated. */
237 uint8_t rand[ETH_ALEN];
238 get_random_bytes(rand, ETH_ALEN);
239 id |= (uint64_t)1 << 63;
240 for (i=0; i<ETH_ALEN; i++) {
241 id |= (uint64_t)rand[i] << (8*(ETH_ALEN-1 - i));
248 /* Creates a new datapath numbered 'dp_idx'. Returns 0 for success or a
249 * negative error code.
251 * Not called with any locks. */
252 static int new_dp(int dp_idx)
257 if (dp_idx < 0 || dp_idx >= DP_MAX)
260 if (!try_module_get(THIS_MODULE))
263 mutex_lock(&dp_mutex);
264 dp = rcu_dereference(dps[dp_idx]);
271 dp = kzalloc(sizeof *dp, GFP_KERNEL);
275 /* Setup our "of" device */
276 err = dp_dev_setup(dp);
281 dp->id = gen_datapath_id(dp_idx);
282 dp->chain = chain_create(dp);
283 if (dp->chain == NULL)
284 goto err_destroy_dp_dev;
285 INIT_LIST_HEAD(&dp->port_list);
287 dp->local_port = new_nbp(dp, dp->netdev, OFPP_LOCAL);
288 if (IS_ERR(dp->local_port)) {
289 err = PTR_ERR(dp->local_port);
290 goto err_destroy_local_port;
294 dp->miss_send_len = OFP_DEFAULT_MISS_SEND_LEN;
296 dp->dp_task = kthread_run(dp_maint_func, dp, "dp%d", dp_idx);
297 if (IS_ERR(dp->dp_task))
298 goto err_destroy_chain;
300 rcu_assign_pointer(dps[dp_idx], dp);
301 mutex_unlock(&dp_mutex);
305 err_destroy_local_port:
306 del_switch_port(dp->local_port);
308 chain_destroy(dp->chain);
314 mutex_unlock(&dp_mutex);
315 module_put(THIS_MODULE);
319 /* Find and return a free port number under 'dp'. Called under dp_mutex. */
320 static int find_portno(struct datapath *dp)
323 for (i = 0; i < OFPP_MAX; i++)
324 if (dp->ports[i] == NULL)
329 static struct net_bridge_port *new_nbp(struct datapath *dp,
330 struct net_device *dev, int port_no)
332 struct net_bridge_port *p;
334 if (dev->br_port != NULL)
335 return ERR_PTR(-EBUSY);
337 p = kzalloc(sizeof(*p), GFP_KERNEL);
339 return ERR_PTR(-ENOMEM);
342 dev_set_promiscuity(dev, 1);
347 p->port_no = port_no;
348 if (port_no != OFPP_LOCAL)
349 rcu_assign_pointer(dev->br_port, p);
350 if (port_no < OFPP_MAX)
351 rcu_assign_pointer(dp->ports[port_no], p);
352 list_add_rcu(&p->node, &dp->port_list);
357 /* Called with dp_mutex. */
358 int add_switch_port(struct datapath *dp, struct net_device *dev)
360 struct net_bridge_port *p;
363 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER
367 port_no = find_portno(dp);
371 p = new_nbp(dp, dev, port_no);
375 /* Notify the ctlpath that this port has been added */
376 send_port_status(p, OFPPR_ADD);
381 /* Delete 'p' from switch.
382 * Called with dp_mutex. */
383 static int del_switch_port(struct net_bridge_port *p)
385 /* First drop references to device. */
387 dev_set_promiscuity(p->dev, -1);
389 list_del_rcu(&p->node);
390 if (p->port_no != OFPP_LOCAL)
391 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
392 rcu_assign_pointer(p->dev->br_port, NULL);
394 /* Then wait until no one is still using it, and destroy it. */
397 /* Notify the ctlpath that this port no longer exists */
398 send_port_status(p, OFPPR_DELETE);
406 /* Called with dp_mutex. */
407 static void del_dp(struct datapath *dp)
409 struct net_bridge_port *p, *n;
411 kthread_stop(dp->dp_task);
413 /* Drop references to DP. */
414 list_for_each_entry_safe (p, n, &dp->port_list, node)
416 rcu_assign_pointer(dps[dp->dp_idx], NULL);
418 /* Kill off local_port dev references from buffered packets that have
419 * associated dst entries. */
423 /* Destroy dp->netdev. (Must follow deleting switch ports since
424 * dp->local_port has a reference to it.) */
427 /* Wait until no longer in use, then destroy it. */
429 chain_destroy(dp->chain);
431 module_put(THIS_MODULE);
434 static int dp_maint_func(void *data)
436 struct datapath *dp = (struct datapath *) data;
438 while (!kthread_should_stop()) {
439 chain_timeout(dp->chain);
440 msleep_interruptible(MAINT_SLEEP_MSECS);
447 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
449 /* Push the Ethernet header back on. */
450 if (skb->protocol == htons(ETH_P_8021Q))
451 skb_push(skb, VLAN_ETH_HLEN);
453 skb_push(skb, ETH_HLEN);
454 fwd_port_input(p->dp->chain, skb, p->port_no);
458 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
459 * different set of devices!)
461 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
462 /* Called with rcu_read_lock. */
463 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
466 do_port_input(p, skb);
469 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
470 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
472 do_port_input(p, *pskb);
476 /* NB: This has only been tested on 2.4.35 */
478 /* Called without any locks (?) */
479 static void dp_frame_hook(struct sk_buff *skb)
481 struct net_bridge_port *p = skb->dev->br_port;
484 do_port_input(p, skb);
491 /* Forwarding output path.
492 * Based on net/bridge/br_forward.c. */
494 static inline unsigned packet_length(const struct sk_buff *skb)
496 int length = skb->len - ETH_HLEN;
497 if (skb->protocol == htons(ETH_P_8021Q))
502 /* Send packets out all the ports except the originating one. If the
503 * "flood" argument is set, only send along the minimum spanning tree.
506 output_all(struct datapath *dp, struct sk_buff *skb, int flood)
508 u32 disable = flood ? BRIDGE_PORT_NO_FLOOD : 0;
509 struct net_bridge_port *p;
512 list_for_each_entry_rcu (p, &dp->port_list, node) {
513 if (skb->dev == p->dev || p->flags & disable)
515 if (prev_port != -1) {
516 struct sk_buff *clone = skb_clone(skb, GFP_ATOMIC);
521 dp_output_port(dp, clone, prev_port);
523 prev_port = p->port_no;
526 dp_output_port(dp, skb, prev_port);
533 /* Marks 'skb' as having originated from 'in_port' in 'dp'.
534 FIXME: how are devices reference counted? */
535 int dp_set_origin(struct datapath *dp, uint16_t in_port,
538 struct net_bridge_port *p = (in_port < OFPP_MAX ? dp->ports[in_port]
539 : in_port == OFPP_LOCAL ? dp->local_port
548 /* Takes ownership of 'skb' and transmits it to 'out_port' on 'dp'.
550 int dp_output_port(struct datapath *dp, struct sk_buff *skb, int out_port)
553 if (out_port == OFPP_FLOOD)
554 return output_all(dp, skb, 1);
555 else if (out_port == OFPP_ALL)
556 return output_all(dp, skb, 0);
557 else if (out_port == OFPP_CONTROLLER)
558 return dp_output_control(dp, skb, fwd_save_skb(skb), 0,
560 else if (out_port == OFPP_TABLE) {
561 struct net_bridge_port *p = skb->dev->br_port;
562 struct sw_flow_key key;
563 struct sw_flow *flow;
565 flow_extract(skb, p ? p->port_no : OFPP_LOCAL, &key);
566 flow = chain_lookup(dp->chain, &key);
567 if (likely(flow != NULL)) {
568 flow_used(flow, skb);
569 execute_actions(dp, skb, &key, flow->actions, flow->n_actions);
573 } else if (out_port == OFPP_LOCAL) {
574 struct net_device *dev = dp->netdev;
575 return dev ? dp_dev_recv(dev, skb) : -ESRCH;
576 } else if (out_port >= 0 && out_port < OFPP_MAX) {
577 struct net_bridge_port *p = dp->ports[out_port];
582 if (packet_length(skb) > skb->dev->mtu) {
583 printk("dropped over-mtu packet: %d > %d\n",
584 packet_length(skb), skb->dev->mtu);
597 printk("can't forward to bad port %d\n", out_port);
601 /* Takes ownership of 'skb' and transmits it to 'dp''s control path. If
602 * 'buffer_id' != -1, then only the first 64 bytes of 'skb' are sent;
603 * otherwise, all of 'skb' is sent. 'reason' indicates why 'skb' is being
604 * sent. 'max_len' sets the maximum number of bytes that the caller
605 * wants to be sent; a value of 0 indicates the entire packet should be
608 dp_output_control(struct datapath *dp, struct sk_buff *skb,
609 uint32_t buffer_id, size_t max_len, int reason)
611 /* FIXME? Can we avoid creating a new skbuff in the case where we
612 * forward the whole packet? */
613 struct sk_buff *f_skb;
614 struct ofp_packet_in *opi;
615 struct net_bridge_port *p;
616 size_t fwd_len, opi_len;
620 if ((buffer_id != (uint32_t) -1) && max_len)
621 fwd_len = min(fwd_len, max_len);
623 opi_len = offsetof(struct ofp_packet_in, data) + fwd_len;
624 opi = alloc_openflow_skb(dp, opi_len, OFPT_PACKET_IN, NULL, &f_skb);
629 opi->buffer_id = htonl(buffer_id);
630 opi->total_len = htons(skb->len);
631 p = skb->dev->br_port;
632 opi->in_port = htons(p ? p->port_no : OFPP_LOCAL);
633 opi->reason = reason;
635 memcpy(opi->data, skb_mac_header(skb), fwd_len);
636 err = send_openflow_skb(f_skb, NULL);
643 static void fill_port_desc(struct net_bridge_port *p, struct ofp_phy_port *desc)
645 desc->port_no = htons(p->port_no);
646 strncpy(desc->name, p->dev->name, OFP_MAX_PORT_NAME_LEN);
647 desc->name[OFP_MAX_PORT_NAME_LEN-1] = '\0';
648 memcpy(desc->hw_addr, p->dev->dev_addr, ETH_ALEN);
649 desc->flags = htonl(p->flags);
653 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,24)
654 if (p->dev->ethtool_ops && p->dev->ethtool_ops->get_settings) {
655 struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
657 if (!p->dev->ethtool_ops->get_settings(p->dev, &ecmd)) {
658 if (ecmd.supported & SUPPORTED_10baseT_Half)
659 desc->features |= OFPPF_10MB_HD;
660 if (ecmd.supported & SUPPORTED_10baseT_Full)
661 desc->features |= OFPPF_10MB_FD;
662 if (ecmd.supported & SUPPORTED_100baseT_Half)
663 desc->features |= OFPPF_100MB_HD;
664 if (ecmd.supported & SUPPORTED_100baseT_Full)
665 desc->features |= OFPPF_100MB_FD;
666 if (ecmd.supported & SUPPORTED_1000baseT_Half)
667 desc->features |= OFPPF_1GB_HD;
668 if (ecmd.supported & SUPPORTED_1000baseT_Full)
669 desc->features |= OFPPF_1GB_FD;
670 /* 10Gbps half-duplex doesn't exist... */
671 if (ecmd.supported & SUPPORTED_10000baseT_Full)
672 desc->features |= OFPPF_10GB_FD;
674 desc->features = htonl(desc->features);
675 desc->speed = htonl(ecmd.speed);
682 fill_features_reply(struct datapath *dp, struct ofp_switch_features *ofr)
684 struct net_bridge_port *p;
687 ofr->datapath_id = cpu_to_be64(dp->id);
689 ofr->n_exact = htonl(2 * TABLE_HASH_MAX_FLOWS);
690 ofr->n_compression = 0; /* Not supported */
691 ofr->n_general = htonl(TABLE_LINEAR_MAX_FLOWS);
692 ofr->buffer_mb = htonl(UINT32_MAX);
693 ofr->n_buffers = htonl(N_PKT_BUFFERS);
694 ofr->capabilities = htonl(OFP_SUPPORTED_CAPABILITIES);
695 ofr->actions = htonl(OFP_SUPPORTED_ACTIONS);
697 list_for_each_entry_rcu (p, &dp->port_list, node) {
698 fill_port_desc(p, &ofr->ports[port_count]);
706 dp_send_features_reply(struct datapath *dp, const struct sender *sender)
709 struct ofp_switch_features *ofr;
710 size_t ofr_len, port_max_len;
714 port_max_len = sizeof(struct ofp_phy_port) * OFPP_MAX;
715 ofr = alloc_openflow_skb(dp, sizeof(*ofr) + port_max_len,
716 OFPT_FEATURES_REPLY, sender, &skb);
721 port_count = fill_features_reply(dp, ofr);
724 ofr_len = sizeof(*ofr) + (sizeof(struct ofp_phy_port) * port_count);
725 resize_openflow_skb(skb, &ofr->header, ofr_len);
726 return send_openflow_skb(skb, sender);
730 dp_send_config_reply(struct datapath *dp, const struct sender *sender)
733 struct ofp_switch_config *osc;
735 osc = alloc_openflow_skb(dp, sizeof *osc, OFPT_GET_CONFIG_REPLY, sender,
740 osc->flags = htons(dp->flags);
741 osc->miss_send_len = htons(dp->miss_send_len);
743 return send_openflow_skb(skb, sender);
747 dp_update_port_flags(struct datapath *dp, const struct ofp_phy_port *opp)
749 int port_no = ntohs(opp->port_no);
750 struct net_bridge_port *p = (port_no < OFPP_MAX ? dp->ports[port_no]
751 : port_no == OFPP_LOCAL ? dp->local_port
753 /* Make sure the port id hasn't changed since this was sent */
754 if (!p || memcmp(opp->hw_addr, p->dev->dev_addr, ETH_ALEN))
756 p->flags = htonl(opp->flags);
762 send_port_status(struct net_bridge_port *p, uint8_t status)
765 struct ofp_port_status *ops;
767 ops = alloc_openflow_skb(p->dp, sizeof *ops, OFPT_PORT_STATUS, NULL,
771 ops->reason = status;
772 memset(ops->pad, 0, sizeof ops->pad);
773 fill_port_desc(p, &ops->desc);
775 return send_openflow_skb(skb, NULL);
779 dp_send_flow_expired(struct datapath *dp, struct sw_flow *flow)
782 struct ofp_flow_expired *ofe;
783 unsigned long duration_j;
785 ofe = alloc_openflow_skb(dp, sizeof *ofe, OFPT_FLOW_EXPIRED, 0, &skb);
789 flow_fill_match(&ofe->match, &flow->key);
791 memset(ofe->pad, 0, sizeof ofe->pad);
792 ofe->priority = htons(flow->priority);
794 duration_j = (flow->timeout - HZ * flow->max_idle) - flow->init_time;
795 ofe->duration = htonl(duration_j / HZ);
796 ofe->packet_count = cpu_to_be64(flow->packet_count);
797 ofe->byte_count = cpu_to_be64(flow->byte_count);
799 return send_openflow_skb(skb, NULL);
801 EXPORT_SYMBOL(dp_send_flow_expired);
804 dp_send_error_msg(struct datapath *dp, const struct sender *sender,
805 uint16_t type, uint16_t code, const uint8_t *data, size_t len)
808 struct ofp_error_msg *oem;
811 oem = alloc_openflow_skb(dp, sizeof(*oem)+len, OFPT_ERROR_MSG,
816 oem->type = htons(type);
817 oem->code = htons(code);
818 memcpy(oem->data, data, len);
820 return send_openflow_skb(skb, sender);
823 /* Generic Netlink interface.
825 * See netlink(7) for an introduction to netlink. See
826 * http://linux-net.osdl.org/index.php/Netlink for more information and
827 * pointers on how to work with netlink and Generic Netlink in the kernel and
830 static struct genl_family dp_genl_family = {
831 .id = GENL_ID_GENERATE,
833 .name = DP_GENL_FAMILY_NAME,
835 .maxattr = DP_GENL_A_MAX,
838 /* Attribute policy: what each attribute may contain. */
839 static struct nla_policy dp_genl_policy[DP_GENL_A_MAX + 1] = {
840 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
841 [DP_GENL_A_MC_GROUP] = { .type = NLA_U32 },
842 [DP_GENL_A_PORTNAME] = { .type = NLA_STRING }
845 static int dp_genl_add(struct sk_buff *skb, struct genl_info *info)
847 if (!info->attrs[DP_GENL_A_DP_IDX])
850 return new_dp(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
853 static struct genl_ops dp_genl_ops_add_dp = {
854 .cmd = DP_GENL_C_ADD_DP,
855 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
856 .policy = dp_genl_policy,
861 struct datapath *dp_get(int dp_idx)
863 if (dp_idx < 0 || dp_idx > DP_MAX)
865 return rcu_dereference(dps[dp_idx]);
868 static int dp_genl_del(struct sk_buff *skb, struct genl_info *info)
873 if (!info->attrs[DP_GENL_A_DP_IDX])
876 mutex_lock(&dp_mutex);
877 dp = dp_get(nla_get_u32((info->attrs[DP_GENL_A_DP_IDX])));
884 mutex_unlock(&dp_mutex);
888 static struct genl_ops dp_genl_ops_del_dp = {
889 .cmd = DP_GENL_C_DEL_DP,
890 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
891 .policy = dp_genl_policy,
896 /* Queries a datapath for related information. Currently the only relevant
897 * information is the datapath's multicast group ID. Really we want one
898 * multicast group per datapath, but because of locking issues[*] we can't
899 * easily get one. Thus, every datapath will currently return the same
900 * global multicast group ID, but in the future it would be nice to fix that.
902 * [*] dp_genl_add, to add a new datapath, is called under the genl_lock
903 * mutex, and genl_register_mc_group, called to acquire a new multicast
904 * group ID, also acquires genl_lock, thus deadlock.
906 static int dp_genl_query(struct sk_buff *skb, struct genl_info *info)
909 struct sk_buff *ans_skb = NULL;
913 if (!info->attrs[DP_GENL_A_DP_IDX])
917 dp_idx = nla_get_u32((info->attrs[DP_GENL_A_DP_IDX]));
923 ans_skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
928 data = genlmsg_put_reply(ans_skb, info, &dp_genl_family,
929 0, DP_GENL_C_QUERY_DP);
934 NLA_PUT_U32(ans_skb, DP_GENL_A_DP_IDX, dp_idx);
935 NLA_PUT_U32(ans_skb, DP_GENL_A_MC_GROUP, mc_group.id);
937 genlmsg_end(ans_skb, data);
938 err = genlmsg_reply(ans_skb, info);
950 static struct genl_ops dp_genl_ops_query_dp = {
951 .cmd = DP_GENL_C_QUERY_DP,
952 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
953 .policy = dp_genl_policy,
954 .doit = dp_genl_query,
958 static int dp_genl_add_del_port(struct sk_buff *skb, struct genl_info *info)
961 struct net_device *port;
964 if (!info->attrs[DP_GENL_A_DP_IDX] || !info->attrs[DP_GENL_A_PORTNAME])
968 mutex_lock(&dp_mutex);
969 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
975 /* Get interface to add/remove. */
976 port = dev_get_by_name(&init_net,
977 nla_data(info->attrs[DP_GENL_A_PORTNAME]));
983 /* Execute operation. */
984 if (info->genlhdr->cmd == DP_GENL_C_ADD_PORT)
985 err = add_switch_port(dp, port);
987 if (port->br_port == NULL || port->br_port->dp != dp) {
991 err = del_switch_port(port->br_port);
997 mutex_unlock(&dp_mutex);
1001 static struct genl_ops dp_genl_ops_add_port = {
1002 .cmd = DP_GENL_C_ADD_PORT,
1003 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1004 .policy = dp_genl_policy,
1005 .doit = dp_genl_add_del_port,
1009 static struct genl_ops dp_genl_ops_del_port = {
1010 .cmd = DP_GENL_C_DEL_PORT,
1011 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1012 .policy = dp_genl_policy,
1013 .doit = dp_genl_add_del_port,
1017 static int dp_genl_openflow(struct sk_buff *skb, struct genl_info *info)
1019 struct nlattr *va = info->attrs[DP_GENL_A_OPENFLOW];
1020 struct datapath *dp;
1021 struct ofp_header *oh;
1022 struct sender sender;
1025 if (!info->attrs[DP_GENL_A_DP_IDX] || !va)
1029 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1035 if (nla_len(va) < sizeof(struct ofp_header)) {
1041 sender.xid = oh->xid;
1042 sender.pid = info->snd_pid;
1043 sender.seq = info->snd_seq;
1044 err = fwd_control_input(dp->chain, &sender, nla_data(va), nla_len(va));
1051 static struct nla_policy dp_genl_openflow_policy[DP_GENL_A_MAX + 1] = {
1052 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1055 struct flow_stats_state {
1057 struct sw_table_position position;
1058 const struct ofp_flow_stats_request *rq;
1061 int bytes_used, bytes_allocated;
1064 static int flow_stats_init(struct datapath *dp, const void *body, int body_len,
1067 const struct ofp_flow_stats_request *fsr = body;
1068 struct flow_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1071 s->table_idx = fsr->table_id == 0xff ? 0 : fsr->table_id;
1072 memset(&s->position, 0, sizeof s->position);
1078 static int flow_stats_dump_callback(struct sw_flow *flow, void *private)
1080 struct flow_stats_state *s = private;
1081 struct ofp_flow_stats *ofs;
1085 actions_length = sizeof *ofs->actions * flow->n_actions;
1086 length = sizeof *ofs + sizeof *ofs->actions * flow->n_actions;
1087 if (length + s->bytes_used > s->bytes_allocated)
1090 ofs = s->body + s->bytes_used;
1091 ofs->length = htons(length);
1092 ofs->table_id = s->table_idx;
1094 ofs->match.wildcards = htons(flow->key.wildcards);
1095 ofs->match.in_port = flow->key.in_port;
1096 memcpy(ofs->match.dl_src, flow->key.dl_src, ETH_ALEN);
1097 memcpy(ofs->match.dl_dst, flow->key.dl_dst, ETH_ALEN);
1098 ofs->match.dl_vlan = flow->key.dl_vlan;
1099 ofs->match.dl_type = flow->key.dl_type;
1100 ofs->match.nw_src = flow->key.nw_src;
1101 ofs->match.nw_dst = flow->key.nw_dst;
1102 ofs->match.nw_proto = flow->key.nw_proto;
1103 memset(ofs->match.pad, 0, sizeof ofs->match.pad);
1104 ofs->match.tp_src = flow->key.tp_src;
1105 ofs->match.tp_dst = flow->key.tp_dst;
1106 ofs->duration = htonl((jiffies - flow->init_time) / HZ);
1107 ofs->packet_count = cpu_to_be64(flow->packet_count);
1108 ofs->byte_count = cpu_to_be64(flow->byte_count);
1109 ofs->priority = htons(flow->priority);
1110 ofs->max_idle = htons(flow->max_idle);
1111 memcpy(ofs->actions, flow->actions, actions_length);
1113 s->bytes_used += length;
1117 static int flow_stats_dump(struct datapath *dp, void *state,
1118 void *body, int *body_len)
1120 struct flow_stats_state *s = state;
1121 struct sw_flow_key match_key;
1125 s->bytes_allocated = *body_len;
1128 flow_extract_match(&match_key, &s->rq->match);
1129 while (s->table_idx < dp->chain->n_tables
1130 && (s->rq->table_id == 0xff || s->rq->table_id == s->table_idx))
1132 struct sw_table *table = dp->chain->tables[s->table_idx];
1134 error = table->iterate(table, &match_key, &s->position,
1135 flow_stats_dump_callback, s);
1140 memset(&s->position, 0, sizeof s->position);
1142 *body_len = s->bytes_used;
1144 /* If error is 0, we're done.
1145 * Otherwise, if some bytes were used, there are more flows to come.
1146 * Otherwise, we were not able to fit even a single flow in the body,
1147 * which indicates that we have a single flow with too many actions to
1148 * fit. We won't ever make any progress at that rate, so give up. */
1149 return !error ? 0 : s->bytes_used ? 1 : -ENOMEM;
1152 static void flow_stats_done(void *state)
1157 static int aggregate_stats_init(struct datapath *dp,
1158 const void *body, int body_len,
1161 *state = (void *)body;
1165 static int aggregate_stats_dump_callback(struct sw_flow *flow, void *private)
1167 struct ofp_aggregate_stats_reply *rpy = private;
1168 rpy->packet_count += flow->packet_count;
1169 rpy->byte_count += flow->byte_count;
1174 static int aggregate_stats_dump(struct datapath *dp, void *state,
1175 void *body, int *body_len)
1177 struct ofp_aggregate_stats_request *rq = state;
1178 struct ofp_aggregate_stats_reply *rpy;
1179 struct sw_table_position position;
1180 struct sw_flow_key match_key;
1183 if (*body_len < sizeof *rpy)
1186 *body_len = sizeof *rpy;
1188 memset(rpy, 0, sizeof *rpy);
1190 flow_extract_match(&match_key, &rq->match);
1191 table_idx = rq->table_id == 0xff ? 0 : rq->table_id;
1192 memset(&position, 0, sizeof position);
1193 while (table_idx < dp->chain->n_tables
1194 && (rq->table_id == 0xff || rq->table_id == table_idx))
1196 struct sw_table *table = dp->chain->tables[table_idx];
1199 error = table->iterate(table, &match_key, &position,
1200 aggregate_stats_dump_callback, rpy);
1205 memset(&position, 0, sizeof position);
1208 rpy->packet_count = cpu_to_be64(rpy->packet_count);
1209 rpy->byte_count = cpu_to_be64(rpy->byte_count);
1210 rpy->flow_count = htonl(rpy->flow_count);
1214 static int table_stats_dump(struct datapath *dp, void *state,
1215 void *body, int *body_len)
1217 struct ofp_table_stats *ots;
1218 int nbytes = dp->chain->n_tables * sizeof *ots;
1220 if (nbytes > *body_len)
1223 for (i = 0, ots = body; i < dp->chain->n_tables; i++, ots++) {
1224 struct sw_table_stats stats;
1225 dp->chain->tables[i]->stats(dp->chain->tables[i], &stats);
1226 strncpy(ots->name, stats.name, sizeof ots->name);
1228 memset(ots->pad, 0, sizeof ots->pad);
1229 ots->max_entries = htonl(stats.max_flows);
1230 ots->active_count = htonl(stats.n_flows);
1231 ots->matched_count = cpu_to_be64(0); /* FIXME */
1236 struct port_stats_state {
1240 static int port_stats_init(struct datapath *dp, const void *body, int body_len,
1243 struct port_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1251 static int port_stats_dump(struct datapath *dp, void *state,
1252 void *body, int *body_len)
1254 struct port_stats_state *s = state;
1255 struct ofp_port_stats *ops;
1256 int n_ports, max_ports;
1259 max_ports = *body_len / sizeof *ops;
1265 for (i = s->port; i < OFPP_MAX && n_ports < max_ports; i++) {
1266 struct net_bridge_port *p = dp->ports[i];
1267 struct net_device_stats *stats;
1270 stats = p->dev->get_stats(p->dev);
1271 ops->port_no = htons(p->port_no);
1272 memset(ops->pad, 0, sizeof ops->pad);
1273 ops->rx_count = cpu_to_be64(stats->rx_packets);
1274 ops->tx_count = cpu_to_be64(stats->tx_packets);
1275 ops->drop_count = cpu_to_be64(stats->rx_dropped
1276 + stats->tx_dropped);
1281 *body_len = n_ports * sizeof *ops;
1282 return n_ports >= max_ports;
1285 static void port_stats_done(void *state)
1291 /* Minimum and maximum acceptable number of bytes in body member of
1292 * struct ofp_stats_request. */
1293 size_t min_body, max_body;
1295 /* Prepares to dump some kind of statistics on 'dp'. 'body' and
1296 * 'body_len' are the 'body' member of the struct ofp_stats_request.
1297 * Returns zero if successful, otherwise a negative error code.
1298 * May initialize '*state' to state information. May be null if no
1299 * initialization is required.*/
1300 int (*init)(struct datapath *dp, const void *body, int body_len,
1303 /* Dumps statistics for 'dp' into the '*body_len' bytes at 'body', and
1304 * modifies '*body_len' to reflect the number of bytes actually used.
1305 * ('body' will be transmitted as the 'body' member of struct
1306 * ofp_stats_reply.) */
1307 int (*dump)(struct datapath *dp, void *state,
1308 void *body, int *body_len);
1310 /* Cleans any state created by the init or dump functions. May be null
1311 * if no cleanup is required. */
1312 void (*done)(void *state);
1315 static const struct stats_type stats[] = {
1317 sizeof(struct ofp_flow_stats_request),
1318 sizeof(struct ofp_flow_stats_request),
1323 [OFPST_AGGREGATE] = {
1324 sizeof(struct ofp_aggregate_stats_request),
1325 sizeof(struct ofp_aggregate_stats_request),
1326 aggregate_stats_init,
1327 aggregate_stats_dump,
1347 dp_genl_openflow_dumpit(struct sk_buff *skb, struct netlink_callback *cb)
1349 struct datapath *dp;
1350 struct sender sender;
1351 const struct stats_type *s;
1352 struct ofp_stats_reply *osr;
1354 int max_openflow_len, body_len;
1358 /* Set up the cleanup function for this dump. Linux 2.6.20 and later
1359 * support setting up cleanup functions via the .doneit member of
1360 * struct genl_ops. This kluge supports earlier versions also. */
1361 cb->done = dp_genl_openflow_done;
1365 struct nlattr *attrs[DP_GENL_A_MAX + 1];
1366 struct ofp_stats_request *rq;
1368 size_t len, body_len;
1371 err = nlmsg_parse(cb->nlh, GENL_HDRLEN, attrs, DP_GENL_A_MAX,
1372 dp_genl_openflow_policy);
1378 if (!attrs[DP_GENL_A_DP_IDX])
1380 dp_idx = nla_get_u16(attrs[DP_GENL_A_DP_IDX]);
1381 dp = dp_get(dp_idx);
1387 va = attrs[DP_GENL_A_OPENFLOW];
1389 if (!va || len < sizeof *rq)
1393 type = ntohs(rq->type);
1394 if (rq->header.version != OFP_VERSION
1395 || rq->header.type != OFPT_STATS_REQUEST
1396 || ntohs(rq->header.length) != len
1397 || type >= ARRAY_SIZE(stats)
1398 || !stats[type].dump)
1402 body_len = len - offsetof(struct ofp_stats_request, body);
1403 if (body_len < s->min_body || body_len > s->max_body)
1407 cb->args[1] = dp_idx;
1409 cb->args[3] = rq->header.xid;
1412 err = s->init(dp, rq->body, body_len, &state);
1415 cb->args[4] = (long) state;
1417 } else if (cb->args[0] == 1) {
1418 dp_idx = cb->args[1];
1419 s = &stats[cb->args[2]];
1421 dp = dp_get(dp_idx);
1431 sender.xid = cb->args[3];
1432 sender.pid = NETLINK_CB(cb->skb).pid;
1433 sender.seq = cb->nlh->nlmsg_seq;
1435 osr = put_openflow_headers(dp, skb, OFPT_STATS_REPLY, &sender,
1441 osr->type = htons(s - stats);
1443 resize_openflow_skb(skb, &osr->header, max_openflow_len);
1445 body_len = max_openflow_len - offsetof(struct ofp_stats_reply, body);
1447 err = s->dump(dp, (void *) cb->args[4], body, &body_len);
1452 osr->flags = ntohs(OFPSF_REPLY_MORE);
1453 resize_openflow_skb(skb, &osr->header,
1454 (offsetof(struct ofp_stats_reply, body)
1465 dp_genl_openflow_done(struct netlink_callback *cb)
1468 const struct stats_type *s = &stats[cb->args[2]];
1470 s->done((void *) cb->args[4]);
1475 static struct genl_ops dp_genl_ops_openflow = {
1476 .cmd = DP_GENL_C_OPENFLOW,
1477 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1478 .policy = dp_genl_openflow_policy,
1479 .doit = dp_genl_openflow,
1480 .dumpit = dp_genl_openflow_dumpit,
1483 static struct nla_policy dp_genl_benchmark_policy[DP_GENL_A_MAX + 1] = {
1484 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1485 [DP_GENL_A_NPACKETS] = { .type = NLA_U32 },
1486 [DP_GENL_A_PSIZE] = { .type = NLA_U32 },
1489 static struct genl_ops dp_genl_ops_benchmark_nl = {
1490 .cmd = DP_GENL_C_BENCHMARK_NL,
1491 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1492 .policy = dp_genl_benchmark_policy,
1493 .doit = dp_genl_benchmark_nl,
1497 static struct genl_ops *dp_genl_all_ops[] = {
1498 /* Keep this operation first. Generic Netlink dispatching
1499 * looks up operations with linear search, so we want it at the
1501 &dp_genl_ops_openflow,
1503 &dp_genl_ops_add_dp,
1504 &dp_genl_ops_del_dp,
1505 &dp_genl_ops_query_dp,
1506 &dp_genl_ops_add_port,
1507 &dp_genl_ops_del_port,
1508 &dp_genl_ops_benchmark_nl,
1511 static int dp_init_netlink(void)
1516 err = genl_register_family(&dp_genl_family);
1520 for (i = 0; i < ARRAY_SIZE(dp_genl_all_ops); i++) {
1521 err = genl_register_ops(&dp_genl_family, dp_genl_all_ops[i]);
1523 goto err_unregister;
1526 strcpy(mc_group.name, "openflow");
1527 err = genl_register_mc_group(&dp_genl_family, &mc_group);
1529 goto err_unregister;
1534 genl_unregister_family(&dp_genl_family);
1538 static void dp_uninit_netlink(void)
1540 genl_unregister_family(&dp_genl_family);
1543 #define DRV_NAME "openflow"
1544 #define DRV_VERSION VERSION
1545 #define DRV_DESCRIPTION "OpenFlow switching datapath implementation"
1546 #define DRV_COPYRIGHT "Copyright (c) 2007, 2008 The Board of Trustees of The Leland Stanford Junior University"
1549 static int __init dp_init(void)
1553 printk(KERN_INFO DRV_NAME ": " DRV_DESCRIPTION "\n");
1554 printk(KERN_INFO DRV_NAME ": " VERSION" built on "__DATE__" "__TIME__"\n");
1555 printk(KERN_INFO DRV_NAME ": " DRV_COPYRIGHT "\n");
1561 err = dp_init_netlink();
1563 goto error_flow_exit;
1565 /* Hook into callback used by the bridge to intercept packets.
1566 * Parasites we are. */
1567 if (br_handle_frame_hook)
1568 printk("openflow: hijacking bridge hook\n");
1569 br_handle_frame_hook = dp_frame_hook;
1576 printk(KERN_EMERG "openflow: failed to install!");
1580 static void dp_cleanup(void)
1583 dp_uninit_netlink();
1585 br_handle_frame_hook = NULL;
1588 module_init(dp_init);
1589 module_exit(dp_cleanup);
1591 MODULE_DESCRIPTION(DRV_DESCRIPTION);
1592 MODULE_AUTHOR(DRV_COPYRIGHT);
1593 MODULE_LICENSE("GPL");