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"
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 /* It's hard to imagine wanting more than one datapath, but... */
66 /* datapaths. Protected on the read side by rcu_read_lock, on the write side
69 * It is safe to access the datapath and net_bridge_port structures with just
70 * the dp_mutex, but to access the chain you need to take the rcu_read_lock
71 * also (because dp_mutex doesn't prevent flows from being destroyed).
73 static struct datapath *dps[DP_MAX];
74 static DEFINE_MUTEX(dp_mutex);
76 static int dp_maint_func(void *data);
77 static int send_port_status(struct net_bridge_port *p, uint8_t status);
78 static int dp_genl_openflow_done(struct netlink_callback *);
79 static struct net_bridge_port *new_nbp(struct datapath *,
80 struct net_device *, int port_no);
81 static int del_switch_port(struct net_bridge_port *);
83 /* nla_shrink - reduce amount of space reserved by nla_reserve
84 * @skb: socket buffer from which to recover room
85 * @nla: netlink attribute to adjust
86 * @len: new length of attribute payload
88 * Reduces amount of space reserved by a call to nla_reserve.
90 * No other attributes may be added between calling nla_reserve and this
91 * function, since it will create a hole in the message.
93 void nla_shrink(struct sk_buff *skb, struct nlattr *nla, int len)
95 int delta = nla_total_size(len) - nla_total_size(nla_len(nla));
99 nla->nla_len = nla_attr_size(len);
102 /* Puts a set of openflow headers for a message of the given 'type' into 'skb'.
103 * If 'sender' is nonnull, then it is used as the message's destination. 'dp'
104 * must specify the datapath to use.
106 * '*max_openflow_len' receives the maximum number of bytes that are available
107 * for the embedded OpenFlow message. The caller must call
108 * resize_openflow_skb() to set the actual size of the message to this number
111 * Returns the openflow header if successful, otherwise (if 'skb' is too small)
114 put_openflow_headers(struct datapath *dp, struct sk_buff *skb, uint8_t type,
115 const struct sender *sender, int *max_openflow_len)
117 struct ofp_header *oh;
121 /* Assemble the Generic Netlink wrapper. */
122 if (!genlmsg_put(skb,
123 sender ? sender->pid : 0,
124 sender ? sender->seq : 0,
125 &dp_genl_family, 0, DP_GENL_C_OPENFLOW))
126 return ERR_PTR(-ENOBUFS);
127 if (nla_put_u32(skb, DP_GENL_A_DP_IDX, dp->dp_idx) < 0)
128 return ERR_PTR(-ENOBUFS);
129 openflow_len = (skb_tailroom(skb) - NLA_HDRLEN) & ~(NLA_ALIGNTO - 1);
130 if (openflow_len < sizeof *oh)
131 return ERR_PTR(-ENOBUFS);
132 *max_openflow_len = openflow_len;
133 attr = nla_reserve(skb, DP_GENL_A_OPENFLOW, openflow_len);
136 /* Fill in the header. The caller is responsible for the length. */
138 oh->version = OFP_VERSION;
140 oh->xid = sender ? sender->xid : 0;
145 /* Resizes OpenFlow header 'oh', which must be at the tail end of 'skb', to new
146 * length 'new_length' (in bytes), adjusting pointers and size values as
149 resize_openflow_skb(struct sk_buff *skb,
150 struct ofp_header *oh, size_t new_length)
152 struct nlattr *attr = ((void *) oh) - NLA_HDRLEN;
153 nla_shrink(skb, attr, new_length);
154 oh->length = htons(new_length);
155 nlmsg_end(skb, (struct nlmsghdr *) skb->data);
158 /* Allocates a new skb to contain an OpenFlow message 'openflow_len' bytes in
159 * length. Returns a null pointer if memory is unavailable, otherwise returns
160 * the OpenFlow header and stores a pointer to the skb in '*pskb'.
162 * 'type' is the OpenFlow message type. If 'sender' is nonnull, then it is
163 * used as the message's destination. 'dp' must specify the datapath to
166 alloc_openflow_skb(struct datapath *dp, size_t openflow_len, uint8_t type,
167 const struct sender *sender, struct sk_buff **pskb)
169 struct ofp_header *oh;
172 int max_openflow_len;
174 if ((openflow_len + sizeof(struct ofp_header)) > UINT16_MAX) {
176 printk("alloc_openflow_skb: openflow message too large: %zu\n",
181 genl_len = nlmsg_total_size(GENL_HDRLEN + dp_genl_family.hdrsize);
182 genl_len += nla_total_size(sizeof(uint32_t)); /* DP_GENL_A_DP_IDX */
183 genl_len += nla_total_size(openflow_len); /* DP_GENL_A_OPENFLOW */
184 skb = *pskb = genlmsg_new(genl_len, GFP_ATOMIC);
187 printk("alloc_openflow_skb: genlmsg_new failed\n");
191 oh = put_openflow_headers(dp, skb, type, sender, &max_openflow_len);
192 BUG_ON(!oh || IS_ERR(oh));
193 resize_openflow_skb(skb, oh, openflow_len);
198 /* Sends 'skb' to 'sender' if it is nonnull, otherwise multicasts 'skb' to all
201 send_openflow_skb(struct sk_buff *skb, const struct sender *sender)
204 ? genlmsg_unicast(skb, sender->pid)
205 : genlmsg_multicast(skb, 0, mc_group.id, GFP_ATOMIC));
208 /* Generates a unique datapath id. It incorporates the datapath index
209 * and a hardware address, if available. If not, it generates a random
213 uint64_t gen_datapath_id(uint16_t dp_idx)
217 struct net_device *dev;
219 /* The top 16 bits are used to identify the datapath. The lower 48 bits
220 * use an interface address. */
221 id = (uint64_t)dp_idx << 48;
222 if ((dev = dev_get_by_name(&init_net, "ctl0"))
223 || (dev = dev_get_by_name(&init_net, "eth0"))) {
224 for (i=0; i<ETH_ALEN; i++) {
225 id |= (uint64_t)dev->dev_addr[i] << (8*(ETH_ALEN-1 - i));
229 /* Randomly choose the lower 48 bits if we cannot find an
230 * address and mark the most significant bit to indicate that
231 * this was randomly generated. */
232 uint8_t rand[ETH_ALEN];
233 get_random_bytes(rand, ETH_ALEN);
234 id |= (uint64_t)1 << 63;
235 for (i=0; i<ETH_ALEN; i++) {
236 id |= (uint64_t)rand[i] << (8*(ETH_ALEN-1 - i));
243 /* Creates a new datapath numbered 'dp_idx'. Returns 0 for success or a
244 * negative error code.
246 * Not called with any locks. */
247 static int new_dp(int dp_idx)
252 if (dp_idx < 0 || dp_idx >= DP_MAX)
255 if (!try_module_get(THIS_MODULE))
258 mutex_lock(&dp_mutex);
259 dp = rcu_dereference(dps[dp_idx]);
266 dp = kzalloc(sizeof *dp, GFP_KERNEL);
270 /* Setup our "of" device */
271 err = dp_dev_setup(dp);
276 dp->id = gen_datapath_id(dp_idx);
277 dp->chain = chain_create(dp);
278 if (dp->chain == NULL)
279 goto err_destroy_dp_dev;
280 INIT_LIST_HEAD(&dp->port_list);
282 dp->local_port = new_nbp(dp, dp->netdev, OFPP_LOCAL);
283 if (IS_ERR(dp->local_port)) {
284 err = PTR_ERR(dp->local_port);
285 goto err_destroy_local_port;
289 dp->miss_send_len = OFP_DEFAULT_MISS_SEND_LEN;
291 dp->dp_task = kthread_run(dp_maint_func, dp, "dp%d", dp_idx);
292 if (IS_ERR(dp->dp_task))
293 goto err_destroy_chain;
295 rcu_assign_pointer(dps[dp_idx], dp);
296 mutex_unlock(&dp_mutex);
300 err_destroy_local_port:
301 del_switch_port(dp->local_port);
303 chain_destroy(dp->chain);
309 mutex_unlock(&dp_mutex);
310 module_put(THIS_MODULE);
314 /* Find and return a free port number under 'dp'. Called under dp_mutex. */
315 static int find_portno(struct datapath *dp)
318 for (i = 0; i < OFPP_MAX; i++)
319 if (dp->ports[i] == NULL)
324 static struct net_bridge_port *new_nbp(struct datapath *dp,
325 struct net_device *dev, int port_no)
327 struct net_bridge_port *p;
329 if (dev->br_port != NULL)
330 return ERR_PTR(-EBUSY);
332 p = kzalloc(sizeof(*p), GFP_KERNEL);
334 return ERR_PTR(-ENOMEM);
337 dev_set_promiscuity(dev, 1);
342 p->port_no = port_no;
343 if (port_no != OFPP_LOCAL)
344 rcu_assign_pointer(dev->br_port, p);
345 if (port_no < OFPP_MAX)
346 rcu_assign_pointer(dp->ports[port_no], p);
347 list_add_rcu(&p->node, &dp->port_list);
352 /* Called with dp_mutex. */
353 int add_switch_port(struct datapath *dp, struct net_device *dev)
355 struct net_bridge_port *p;
358 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER
362 port_no = find_portno(dp);
366 p = new_nbp(dp, dev, port_no);
370 /* Notify the ctlpath that this port has been added */
371 send_port_status(p, OFPPR_ADD);
376 /* Delete 'p' from switch.
377 * Called with dp_mutex. */
378 static int del_switch_port(struct net_bridge_port *p)
380 /* First drop references to device. */
382 dev_set_promiscuity(p->dev, -1);
384 list_del_rcu(&p->node);
385 if (p->port_no != OFPP_LOCAL)
386 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
387 rcu_assign_pointer(p->dev->br_port, NULL);
389 /* Then wait until no one is still using it, and destroy it. */
392 /* Notify the ctlpath that this port no longer exists */
393 send_port_status(p, OFPPR_DELETE);
401 /* Called with dp_mutex. */
402 static void del_dp(struct datapath *dp)
404 struct net_bridge_port *p, *n;
406 kthread_stop(dp->dp_task);
408 /* Drop references to DP. */
409 list_for_each_entry_safe (p, n, &dp->port_list, node)
411 rcu_assign_pointer(dps[dp->dp_idx], NULL);
413 /* Kill off local_port dev references from buffered packets that have
414 * associated dst entries. */
418 /* Destroy dp->netdev. (Must follow deleting switch ports since
419 * dp->local_port has a reference to it.) */
422 /* Wait until no longer in use, then destroy it. */
424 chain_destroy(dp->chain);
426 module_put(THIS_MODULE);
429 static int dp_maint_func(void *data)
431 struct datapath *dp = (struct datapath *) data;
433 while (!kthread_should_stop()) {
434 chain_timeout(dp->chain);
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 skb_push(skb, ETH_HLEN);
446 fwd_port_input(p->dp->chain, skb, p->port_no);
450 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
451 * different set of devices!)
453 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
454 /* Called with rcu_read_lock. */
455 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
458 do_port_input(p, skb);
461 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
462 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
464 do_port_input(p, *pskb);
468 /* NB: This has only been tested on 2.4.35 */
470 /* Called without any locks (?) */
471 static void dp_frame_hook(struct sk_buff *skb)
473 struct net_bridge_port *p = skb->dev->br_port;
476 do_port_input(p, skb);
483 /* Forwarding output path.
484 * Based on net/bridge/br_forward.c. */
486 static inline unsigned packet_length(const struct sk_buff *skb)
488 int length = skb->len - ETH_HLEN;
489 if (skb->protocol == htons(ETH_P_8021Q))
494 /* Send packets out all the ports except the originating one. If the
495 * "flood" argument is set, only send along the minimum spanning tree.
498 output_all(struct datapath *dp, struct sk_buff *skb, int flood)
500 u32 disable = flood ? BRIDGE_PORT_NO_FLOOD : 0;
501 struct net_bridge_port *p;
504 list_for_each_entry_rcu (p, &dp->port_list, node) {
505 if (skb->dev == p->dev || p->flags & disable)
507 if (prev_port != -1) {
508 struct sk_buff *clone = skb_clone(skb, GFP_ATOMIC);
513 dp_output_port(dp, clone, prev_port);
515 prev_port = p->port_no;
518 dp_output_port(dp, skb, prev_port);
525 /* Marks 'skb' as having originated from 'in_port' in 'dp'.
526 FIXME: how are devices reference counted? */
527 int dp_set_origin(struct datapath *dp, uint16_t in_port,
530 struct net_bridge_port *p = (in_port < OFPP_MAX ? dp->ports[in_port]
531 : in_port == OFPP_LOCAL ? dp->local_port
540 /* Takes ownership of 'skb' and transmits it to 'out_port' on 'dp'.
542 int dp_output_port(struct datapath *dp, struct sk_buff *skb, int out_port)
545 if (out_port == OFPP_FLOOD)
546 return output_all(dp, skb, 1);
547 else if (out_port == OFPP_ALL)
548 return output_all(dp, skb, 0);
549 else if (out_port == OFPP_CONTROLLER)
550 return dp_output_control(dp, skb, fwd_save_skb(skb), 0,
552 else if (out_port == OFPP_TABLE) {
553 struct net_bridge_port *p = skb->dev->br_port;
554 struct sw_flow_key key;
555 struct sw_flow *flow;
557 flow_extract(skb, p ? p->port_no : OFPP_LOCAL, &key);
558 flow = chain_lookup(dp->chain, &key);
559 if (likely(flow != NULL)) {
560 flow_used(flow, skb);
561 execute_actions(dp, skb, &key, flow->actions, flow->n_actions);
565 } else if (out_port == OFPP_LOCAL) {
566 struct net_device *dev = dp->netdev;
567 return dev ? dp_dev_recv(dev, skb) : -ESRCH;
568 } else if (out_port >= 0 && out_port < OFPP_MAX) {
569 struct net_bridge_port *p = dp->ports[out_port];
574 if (packet_length(skb) > skb->dev->mtu) {
575 printk("dropped over-mtu packet: %d > %d\n",
576 packet_length(skb), skb->dev->mtu);
589 printk("can't forward to bad port %d\n", out_port);
593 /* Takes ownership of 'skb' and transmits it to 'dp''s control path. If
594 * 'buffer_id' != -1, then only the first 64 bytes of 'skb' are sent;
595 * otherwise, all of 'skb' is sent. 'reason' indicates why 'skb' is being
596 * sent. 'max_len' sets the maximum number of bytes that the caller
597 * wants to be sent; a value of 0 indicates the entire packet should be
600 dp_output_control(struct datapath *dp, struct sk_buff *skb,
601 uint32_t buffer_id, size_t max_len, int reason)
603 /* FIXME? Can we avoid creating a new skbuff in the case where we
604 * forward the whole packet? */
605 struct sk_buff *f_skb;
606 struct ofp_packet_in *opi;
607 struct net_bridge_port *p;
608 size_t fwd_len, opi_len;
612 if ((buffer_id != (uint32_t) -1) && max_len)
613 fwd_len = min(fwd_len, max_len);
615 opi_len = offsetof(struct ofp_packet_in, data) + fwd_len;
616 opi = alloc_openflow_skb(dp, opi_len, OFPT_PACKET_IN, NULL, &f_skb);
621 opi->buffer_id = htonl(buffer_id);
622 opi->total_len = htons(skb->len);
623 p = skb->dev->br_port;
624 opi->in_port = htons(p ? p->port_no : OFPP_LOCAL);
625 opi->reason = reason;
627 memcpy(opi->data, skb_mac_header(skb), fwd_len);
628 err = send_openflow_skb(f_skb, NULL);
635 static void fill_port_desc(struct net_bridge_port *p, struct ofp_phy_port *desc)
637 desc->port_no = htons(p->port_no);
638 strncpy(desc->name, p->dev->name, OFP_MAX_PORT_NAME_LEN);
639 desc->name[OFP_MAX_PORT_NAME_LEN-1] = '\0';
640 memcpy(desc->hw_addr, p->dev->dev_addr, ETH_ALEN);
641 desc->flags = htonl(p->flags);
645 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,24)
646 if (p->dev->ethtool_ops && p->dev->ethtool_ops->get_settings) {
647 struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
649 if (!p->dev->ethtool_ops->get_settings(p->dev, &ecmd)) {
650 if (ecmd.supported & SUPPORTED_10baseT_Half)
651 desc->features |= OFPPF_10MB_HD;
652 if (ecmd.supported & SUPPORTED_10baseT_Full)
653 desc->features |= OFPPF_10MB_FD;
654 if (ecmd.supported & SUPPORTED_100baseT_Half)
655 desc->features |= OFPPF_100MB_HD;
656 if (ecmd.supported & SUPPORTED_100baseT_Full)
657 desc->features |= OFPPF_100MB_FD;
658 if (ecmd.supported & SUPPORTED_1000baseT_Half)
659 desc->features |= OFPPF_1GB_HD;
660 if (ecmd.supported & SUPPORTED_1000baseT_Full)
661 desc->features |= OFPPF_1GB_FD;
662 /* 10Gbps half-duplex doesn't exist... */
663 if (ecmd.supported & SUPPORTED_10000baseT_Full)
664 desc->features |= OFPPF_10GB_FD;
666 desc->features = htonl(desc->features);
667 desc->speed = htonl(ecmd.speed);
674 fill_features_reply(struct datapath *dp, struct ofp_switch_features *ofr)
676 struct net_bridge_port *p;
679 ofr->datapath_id = cpu_to_be64(dp->id);
681 ofr->n_exact = htonl(2 * TABLE_HASH_MAX_FLOWS);
682 ofr->n_compression = 0; /* Not supported */
683 ofr->n_general = htonl(TABLE_LINEAR_MAX_FLOWS);
684 ofr->buffer_mb = htonl(UINT32_MAX);
685 ofr->n_buffers = htonl(N_PKT_BUFFERS);
686 ofr->capabilities = htonl(OFP_SUPPORTED_CAPABILITIES);
687 ofr->actions = htonl(OFP_SUPPORTED_ACTIONS);
689 list_for_each_entry_rcu (p, &dp->port_list, node) {
690 fill_port_desc(p, &ofr->ports[port_count]);
698 dp_send_features_reply(struct datapath *dp, const struct sender *sender)
701 struct ofp_switch_features *ofr;
702 size_t ofr_len, port_max_len;
706 port_max_len = sizeof(struct ofp_phy_port) * OFPP_MAX;
707 ofr = alloc_openflow_skb(dp, sizeof(*ofr) + port_max_len,
708 OFPT_FEATURES_REPLY, sender, &skb);
713 port_count = fill_features_reply(dp, ofr);
716 ofr_len = sizeof(*ofr) + (sizeof(struct ofp_phy_port) * port_count);
717 resize_openflow_skb(skb, &ofr->header, ofr_len);
718 return send_openflow_skb(skb, sender);
722 dp_send_config_reply(struct datapath *dp, const struct sender *sender)
725 struct ofp_switch_config *osc;
727 osc = alloc_openflow_skb(dp, sizeof *osc, OFPT_GET_CONFIG_REPLY, sender,
732 osc->flags = htons(dp->flags);
733 osc->miss_send_len = htons(dp->miss_send_len);
735 return send_openflow_skb(skb, sender);
739 dp_update_port_flags(struct datapath *dp, const struct ofp_phy_port *opp)
741 int port_no = ntohs(opp->port_no);
742 struct net_bridge_port *p = (port_no < OFPP_MAX ? dp->ports[port_no]
743 : port_no == OFPP_LOCAL ? dp->local_port
745 /* Make sure the port id hasn't changed since this was sent */
746 if (!p || memcmp(opp->hw_addr, p->dev->dev_addr, ETH_ALEN))
748 p->flags = htonl(opp->flags);
754 send_port_status(struct net_bridge_port *p, uint8_t status)
757 struct ofp_port_status *ops;
759 ops = alloc_openflow_skb(p->dp, sizeof *ops, OFPT_PORT_STATUS, NULL,
763 ops->reason = status;
764 memset(ops->pad, 0, sizeof ops->pad);
765 fill_port_desc(p, &ops->desc);
767 return send_openflow_skb(skb, NULL);
771 dp_send_flow_expired(struct datapath *dp, struct sw_flow *flow)
774 struct ofp_flow_expired *ofe;
775 unsigned long duration_j;
777 ofe = alloc_openflow_skb(dp, sizeof *ofe, OFPT_FLOW_EXPIRED, 0, &skb);
781 flow_fill_match(&ofe->match, &flow->key);
783 memset(ofe->pad, 0, sizeof ofe->pad);
784 ofe->priority = htons(flow->priority);
786 duration_j = (flow->timeout - HZ * flow->max_idle) - flow->init_time;
787 ofe->duration = htonl(duration_j / HZ);
788 ofe->packet_count = cpu_to_be64(flow->packet_count);
789 ofe->byte_count = cpu_to_be64(flow->byte_count);
791 return send_openflow_skb(skb, NULL);
793 EXPORT_SYMBOL(dp_send_flow_expired);
796 dp_send_error_msg(struct datapath *dp, const struct sender *sender,
797 uint16_t type, uint16_t code, const uint8_t *data, size_t len)
800 struct ofp_error_msg *oem;
803 oem = alloc_openflow_skb(dp, sizeof(*oem)+len, OFPT_ERROR_MSG,
808 oem->type = htons(type);
809 oem->code = htons(code);
810 memcpy(oem->data, data, len);
812 return send_openflow_skb(skb, sender);
816 dp_send_echo_reply(struct datapath *dp, const struct sender *sender,
817 const struct ofp_header *rq)
820 struct ofp_header *reply;
822 reply = alloc_openflow_skb(dp, ntohs(rq->length), OFPT_ECHO_REPLY,
827 memcpy(reply + 1, rq + 1, ntohs(rq->length) - sizeof *rq);
828 return send_openflow_skb(skb, sender);
831 /* Generic Netlink interface.
833 * See netlink(7) for an introduction to netlink. See
834 * http://linux-net.osdl.org/index.php/Netlink for more information and
835 * pointers on how to work with netlink and Generic Netlink in the kernel and
838 static struct genl_family dp_genl_family = {
839 .id = GENL_ID_GENERATE,
841 .name = DP_GENL_FAMILY_NAME,
843 .maxattr = DP_GENL_A_MAX,
846 /* Attribute policy: what each attribute may contain. */
847 static struct nla_policy dp_genl_policy[DP_GENL_A_MAX + 1] = {
848 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
849 [DP_GENL_A_MC_GROUP] = { .type = NLA_U32 },
850 [DP_GENL_A_PORTNAME] = { .type = NLA_STRING }
853 static int dp_genl_add(struct sk_buff *skb, struct genl_info *info)
855 if (!info->attrs[DP_GENL_A_DP_IDX])
858 return new_dp(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
861 static struct genl_ops dp_genl_ops_add_dp = {
862 .cmd = DP_GENL_C_ADD_DP,
863 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
864 .policy = dp_genl_policy,
869 struct datapath *dp_get(int dp_idx)
871 if (dp_idx < 0 || dp_idx > DP_MAX)
873 return rcu_dereference(dps[dp_idx]);
876 static int dp_genl_del(struct sk_buff *skb, struct genl_info *info)
881 if (!info->attrs[DP_GENL_A_DP_IDX])
884 mutex_lock(&dp_mutex);
885 dp = dp_get(nla_get_u32((info->attrs[DP_GENL_A_DP_IDX])));
892 mutex_unlock(&dp_mutex);
896 static struct genl_ops dp_genl_ops_del_dp = {
897 .cmd = DP_GENL_C_DEL_DP,
898 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
899 .policy = dp_genl_policy,
904 /* Queries a datapath for related information. Currently the only relevant
905 * information is the datapath's multicast group ID. Really we want one
906 * multicast group per datapath, but because of locking issues[*] we can't
907 * easily get one. Thus, every datapath will currently return the same
908 * global multicast group ID, but in the future it would be nice to fix that.
910 * [*] dp_genl_add, to add a new datapath, is called under the genl_lock
911 * mutex, and genl_register_mc_group, called to acquire a new multicast
912 * group ID, also acquires genl_lock, thus deadlock.
914 static int dp_genl_query(struct sk_buff *skb, struct genl_info *info)
917 struct sk_buff *ans_skb = NULL;
921 if (!info->attrs[DP_GENL_A_DP_IDX])
925 dp_idx = nla_get_u32((info->attrs[DP_GENL_A_DP_IDX]));
931 ans_skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
936 data = genlmsg_put_reply(ans_skb, info, &dp_genl_family,
937 0, DP_GENL_C_QUERY_DP);
942 NLA_PUT_U32(ans_skb, DP_GENL_A_DP_IDX, dp_idx);
943 NLA_PUT_U32(ans_skb, DP_GENL_A_MC_GROUP, mc_group.id);
945 genlmsg_end(ans_skb, data);
946 err = genlmsg_reply(ans_skb, info);
958 static struct genl_ops dp_genl_ops_query_dp = {
959 .cmd = DP_GENL_C_QUERY_DP,
960 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
961 .policy = dp_genl_policy,
962 .doit = dp_genl_query,
966 static int dp_genl_add_del_port(struct sk_buff *skb, struct genl_info *info)
969 struct net_device *port;
972 if (!info->attrs[DP_GENL_A_DP_IDX] || !info->attrs[DP_GENL_A_PORTNAME])
976 mutex_lock(&dp_mutex);
977 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
983 /* Get interface to add/remove. */
984 port = dev_get_by_name(&init_net,
985 nla_data(info->attrs[DP_GENL_A_PORTNAME]));
991 /* Execute operation. */
992 if (info->genlhdr->cmd == DP_GENL_C_ADD_PORT)
993 err = add_switch_port(dp, port);
995 if (port->br_port == NULL || port->br_port->dp != dp) {
999 err = del_switch_port(port->br_port);
1005 mutex_unlock(&dp_mutex);
1009 static struct genl_ops dp_genl_ops_add_port = {
1010 .cmd = DP_GENL_C_ADD_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 struct genl_ops dp_genl_ops_del_port = {
1018 .cmd = DP_GENL_C_DEL_PORT,
1019 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1020 .policy = dp_genl_policy,
1021 .doit = dp_genl_add_del_port,
1025 static int dp_genl_openflow(struct sk_buff *skb, struct genl_info *info)
1027 struct nlattr *va = info->attrs[DP_GENL_A_OPENFLOW];
1028 struct datapath *dp;
1029 struct ofp_header *oh;
1030 struct sender sender;
1032 if (!info->attrs[DP_GENL_A_DP_IDX] || !va)
1035 dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1039 if (nla_len(va) < sizeof(struct ofp_header))
1043 sender.xid = oh->xid;
1044 sender.pid = info->snd_pid;
1045 sender.seq = info->snd_seq;
1046 return fwd_control_input(dp->chain, &sender,
1047 nla_data(va), nla_len(va));
1050 static struct nla_policy dp_genl_openflow_policy[DP_GENL_A_MAX + 1] = {
1051 [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1054 struct flow_stats_state {
1056 struct sw_table_position position;
1057 const struct ofp_flow_stats_request *rq;
1060 int bytes_used, bytes_allocated;
1063 static int flow_stats_init(struct datapath *dp, const void *body, int body_len,
1066 const struct ofp_flow_stats_request *fsr = body;
1067 struct flow_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1070 s->table_idx = fsr->table_id == 0xff ? 0 : fsr->table_id;
1071 memset(&s->position, 0, sizeof s->position);
1077 static int flow_stats_dump_callback(struct sw_flow *flow, void *private)
1079 struct flow_stats_state *s = private;
1080 struct ofp_flow_stats *ofs;
1084 actions_length = sizeof *ofs->actions * flow->n_actions;
1085 length = sizeof *ofs + sizeof *ofs->actions * flow->n_actions;
1086 if (length + s->bytes_used > s->bytes_allocated)
1089 ofs = s->body + s->bytes_used;
1090 ofs->length = htons(length);
1091 ofs->table_id = s->table_idx;
1093 ofs->match.wildcards = htons(flow->key.wildcards);
1094 ofs->match.in_port = flow->key.in_port;
1095 memcpy(ofs->match.dl_src, flow->key.dl_src, ETH_ALEN);
1096 memcpy(ofs->match.dl_dst, flow->key.dl_dst, ETH_ALEN);
1097 ofs->match.dl_vlan = flow->key.dl_vlan;
1098 ofs->match.dl_type = flow->key.dl_type;
1099 ofs->match.nw_src = flow->key.nw_src;
1100 ofs->match.nw_dst = flow->key.nw_dst;
1101 ofs->match.nw_proto = flow->key.nw_proto;
1102 memset(ofs->match.pad, 0, sizeof ofs->match.pad);
1103 ofs->match.tp_src = flow->key.tp_src;
1104 ofs->match.tp_dst = flow->key.tp_dst;
1105 ofs->duration = htonl((jiffies - flow->init_time) / HZ);
1106 ofs->packet_count = cpu_to_be64(flow->packet_count);
1107 ofs->byte_count = cpu_to_be64(flow->byte_count);
1108 ofs->priority = htons(flow->priority);
1109 ofs->max_idle = htons(flow->max_idle);
1110 memcpy(ofs->actions, flow->actions, actions_length);
1112 s->bytes_used += length;
1116 static int flow_stats_dump(struct datapath *dp, void *state,
1117 void *body, int *body_len)
1119 struct flow_stats_state *s = state;
1120 struct sw_flow_key match_key;
1124 s->bytes_allocated = *body_len;
1127 flow_extract_match(&match_key, &s->rq->match);
1128 while (s->table_idx < dp->chain->n_tables
1129 && (s->rq->table_id == 0xff || s->rq->table_id == s->table_idx))
1131 struct sw_table *table = dp->chain->tables[s->table_idx];
1133 error = table->iterate(table, &match_key, &s->position,
1134 flow_stats_dump_callback, s);
1139 memset(&s->position, 0, sizeof s->position);
1141 *body_len = s->bytes_used;
1143 /* If error is 0, we're done.
1144 * Otherwise, if some bytes were used, there are more flows to come.
1145 * Otherwise, we were not able to fit even a single flow in the body,
1146 * which indicates that we have a single flow with too many actions to
1147 * fit. We won't ever make any progress at that rate, so give up. */
1148 return !error ? 0 : s->bytes_used ? 1 : -ENOMEM;
1151 static void flow_stats_done(void *state)
1156 static int aggregate_stats_init(struct datapath *dp,
1157 const void *body, int body_len,
1160 *state = (void *)body;
1164 static int aggregate_stats_dump_callback(struct sw_flow *flow, void *private)
1166 struct ofp_aggregate_stats_reply *rpy = private;
1167 rpy->packet_count += flow->packet_count;
1168 rpy->byte_count += flow->byte_count;
1173 static int aggregate_stats_dump(struct datapath *dp, void *state,
1174 void *body, int *body_len)
1176 struct ofp_aggregate_stats_request *rq = state;
1177 struct ofp_aggregate_stats_reply *rpy;
1178 struct sw_table_position position;
1179 struct sw_flow_key match_key;
1182 if (*body_len < sizeof *rpy)
1185 *body_len = sizeof *rpy;
1187 memset(rpy, 0, sizeof *rpy);
1189 flow_extract_match(&match_key, &rq->match);
1190 table_idx = rq->table_id == 0xff ? 0 : rq->table_id;
1191 memset(&position, 0, sizeof position);
1192 while (table_idx < dp->chain->n_tables
1193 && (rq->table_id == 0xff || rq->table_id == table_idx))
1195 struct sw_table *table = dp->chain->tables[table_idx];
1198 error = table->iterate(table, &match_key, &position,
1199 aggregate_stats_dump_callback, rpy);
1204 memset(&position, 0, sizeof position);
1207 rpy->packet_count = cpu_to_be64(rpy->packet_count);
1208 rpy->byte_count = cpu_to_be64(rpy->byte_count);
1209 rpy->flow_count = htonl(rpy->flow_count);
1213 static int table_stats_dump(struct datapath *dp, void *state,
1214 void *body, int *body_len)
1216 struct ofp_table_stats *ots;
1217 int nbytes = dp->chain->n_tables * sizeof *ots;
1219 if (nbytes > *body_len)
1222 for (i = 0, ots = body; i < dp->chain->n_tables; i++, ots++) {
1223 struct sw_table_stats stats;
1224 dp->chain->tables[i]->stats(dp->chain->tables[i], &stats);
1225 strncpy(ots->name, stats.name, sizeof ots->name);
1227 memset(ots->pad, 0, sizeof ots->pad);
1228 ots->max_entries = htonl(stats.max_flows);
1229 ots->active_count = htonl(stats.n_flows);
1230 ots->matched_count = cpu_to_be64(0); /* FIXME */
1235 struct port_stats_state {
1239 static int port_stats_init(struct datapath *dp, const void *body, int body_len,
1242 struct port_stats_state *s = kmalloc(sizeof *s, GFP_ATOMIC);
1250 static int port_stats_dump(struct datapath *dp, void *state,
1251 void *body, int *body_len)
1253 struct port_stats_state *s = state;
1254 struct ofp_port_stats *ops;
1255 int n_ports, max_ports;
1258 max_ports = *body_len / sizeof *ops;
1264 for (i = s->port; i < OFPP_MAX && n_ports < max_ports; i++) {
1265 struct net_bridge_port *p = dp->ports[i];
1266 struct net_device_stats *stats;
1269 stats = p->dev->get_stats(p->dev);
1270 ops->port_no = htons(p->port_no);
1271 memset(ops->pad, 0, sizeof ops->pad);
1272 ops->rx_count = cpu_to_be64(stats->rx_packets);
1273 ops->tx_count = cpu_to_be64(stats->tx_packets);
1274 ops->drop_count = cpu_to_be64(stats->rx_dropped
1275 + stats->tx_dropped);
1280 *body_len = n_ports * sizeof *ops;
1281 return n_ports >= max_ports;
1284 static void port_stats_done(void *state)
1290 /* Minimum and maximum acceptable number of bytes in body member of
1291 * struct ofp_stats_request. */
1292 size_t min_body, max_body;
1294 /* Prepares to dump some kind of statistics on 'dp'. 'body' and
1295 * 'body_len' are the 'body' member of the struct ofp_stats_request.
1296 * Returns zero if successful, otherwise a negative error code.
1297 * May initialize '*state' to state information. May be null if no
1298 * initialization is required.*/
1299 int (*init)(struct datapath *dp, const void *body, int body_len,
1302 /* Dumps statistics for 'dp' into the '*body_len' bytes at 'body', and
1303 * modifies '*body_len' to reflect the number of bytes actually used.
1304 * ('body' will be transmitted as the 'body' member of struct
1305 * ofp_stats_reply.) */
1306 int (*dump)(struct datapath *dp, void *state,
1307 void *body, int *body_len);
1309 /* Cleans any state created by the init or dump functions. May be null
1310 * if no cleanup is required. */
1311 void (*done)(void *state);
1314 static const struct stats_type stats[] = {
1316 sizeof(struct ofp_flow_stats_request),
1317 sizeof(struct ofp_flow_stats_request),
1322 [OFPST_AGGREGATE] = {
1323 sizeof(struct ofp_aggregate_stats_request),
1324 sizeof(struct ofp_aggregate_stats_request),
1325 aggregate_stats_init,
1326 aggregate_stats_dump,
1346 dp_genl_openflow_dumpit(struct sk_buff *skb, struct netlink_callback *cb)
1348 struct datapath *dp;
1349 struct sender sender;
1350 const struct stats_type *s;
1351 struct ofp_stats_reply *osr;
1353 int max_openflow_len, body_len;
1357 /* Set up the cleanup function for this dump. Linux 2.6.20 and later
1358 * support setting up cleanup functions via the .doneit member of
1359 * struct genl_ops. This kluge supports earlier versions also. */
1360 cb->done = dp_genl_openflow_done;
1363 struct nlattr *attrs[DP_GENL_A_MAX + 1];
1364 struct ofp_stats_request *rq;
1366 size_t len, body_len;
1369 err = nlmsg_parse(cb->nlh, GENL_HDRLEN, attrs, DP_GENL_A_MAX,
1370 dp_genl_openflow_policy);
1374 if (!attrs[DP_GENL_A_DP_IDX])
1376 dp_idx = nla_get_u16(attrs[DP_GENL_A_DP_IDX]);
1377 dp = dp_get(dp_idx);
1381 va = attrs[DP_GENL_A_OPENFLOW];
1383 if (!va || len < sizeof *rq)
1387 type = ntohs(rq->type);
1388 if (rq->header.version != OFP_VERSION
1389 || rq->header.type != OFPT_STATS_REQUEST
1390 || ntohs(rq->header.length) != len
1391 || type >= ARRAY_SIZE(stats)
1392 || !stats[type].dump)
1396 body_len = len - offsetof(struct ofp_stats_request, body);
1397 if (body_len < s->min_body || body_len > s->max_body)
1401 cb->args[1] = dp_idx;
1403 cb->args[3] = rq->header.xid;
1406 err = s->init(dp, rq->body, body_len, &state);
1409 cb->args[4] = (long) state;
1411 } else if (cb->args[0] == 1) {
1412 dp_idx = cb->args[1];
1413 s = &stats[cb->args[2]];
1415 dp = dp_get(dp_idx);
1422 sender.xid = cb->args[3];
1423 sender.pid = NETLINK_CB(cb->skb).pid;
1424 sender.seq = cb->nlh->nlmsg_seq;
1426 osr = put_openflow_headers(dp, skb, OFPT_STATS_REPLY, &sender,
1429 return PTR_ERR(osr);
1430 osr->type = htons(s - stats);
1432 resize_openflow_skb(skb, &osr->header, max_openflow_len);
1434 body_len = max_openflow_len - offsetof(struct ofp_stats_reply, body);
1436 err = s->dump(dp, (void *) cb->args[4], body, &body_len);
1441 osr->flags = ntohs(OFPSF_REPLY_MORE);
1442 resize_openflow_skb(skb, &osr->header,
1443 (offsetof(struct ofp_stats_reply, body)
1452 dp_genl_openflow_done(struct netlink_callback *cb)
1455 const struct stats_type *s = &stats[cb->args[2]];
1457 s->done((void *) cb->args[4]);
1462 static struct genl_ops dp_genl_ops_openflow = {
1463 .cmd = DP_GENL_C_OPENFLOW,
1464 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1465 .policy = dp_genl_openflow_policy,
1466 .doit = dp_genl_openflow,
1467 .dumpit = dp_genl_openflow_dumpit,
1470 static struct genl_ops *dp_genl_all_ops[] = {
1471 /* Keep this operation first. Generic Netlink dispatching
1472 * looks up operations with linear search, so we want it at the
1474 &dp_genl_ops_openflow,
1476 &dp_genl_ops_add_dp,
1477 &dp_genl_ops_del_dp,
1478 &dp_genl_ops_query_dp,
1479 &dp_genl_ops_add_port,
1480 &dp_genl_ops_del_port,
1483 static int dp_init_netlink(void)
1488 err = genl_register_family(&dp_genl_family);
1492 for (i = 0; i < ARRAY_SIZE(dp_genl_all_ops); i++) {
1493 err = genl_register_ops(&dp_genl_family, dp_genl_all_ops[i]);
1495 goto err_unregister;
1498 strcpy(mc_group.name, "openflow");
1499 err = genl_register_mc_group(&dp_genl_family, &mc_group);
1501 goto err_unregister;
1506 genl_unregister_family(&dp_genl_family);
1510 static void dp_uninit_netlink(void)
1512 genl_unregister_family(&dp_genl_family);
1515 #define DRV_NAME "openflow"
1516 #define DRV_VERSION VERSION
1517 #define DRV_DESCRIPTION "OpenFlow switching datapath implementation"
1518 #define DRV_COPYRIGHT "Copyright (c) 2007, 2008 The Board of Trustees of The Leland Stanford Junior University"
1521 static int __init dp_init(void)
1525 printk(KERN_INFO DRV_NAME ": " DRV_DESCRIPTION "\n");
1526 printk(KERN_INFO DRV_NAME ": " VERSION" built on "__DATE__" "__TIME__"\n");
1527 printk(KERN_INFO DRV_NAME ": " DRV_COPYRIGHT "\n");
1533 err = dp_init_netlink();
1535 goto error_flow_exit;
1537 /* Hook into callback used by the bridge to intercept packets.
1538 * Parasites we are. */
1539 if (br_handle_frame_hook)
1540 printk("openflow: hijacking bridge hook\n");
1541 br_handle_frame_hook = dp_frame_hook;
1548 printk(KERN_EMERG "openflow: failed to install!");
1552 static void dp_cleanup(void)
1555 dp_uninit_netlink();
1557 br_handle_frame_hook = NULL;
1560 module_init(dp_init);
1561 module_exit(dp_cleanup);
1563 MODULE_DESCRIPTION(DRV_DESCRIPTION);
1564 MODULE_AUTHOR(DRV_COPYRIGHT);
1565 MODULE_LICENSE("GPL");