2 * Copyright (c) 2007, 2008, 2009, 2010 Nicira Networks.
3 * Distributed under the terms of the GNU GPL version 2.
5 * Significant portions of this file may be copied from parts of the Linux
6 * kernel, by Linus Torvalds and others.
9 /* Functions for managing the dp interface/device. */
11 #include <linux/init.h>
12 #include <linux/module.h>
14 #include <linux/if_arp.h>
15 #include <linux/if_bridge.h>
16 #include <linux/if_vlan.h>
19 #include <linux/delay.h>
20 #include <linux/time.h>
21 #include <linux/etherdevice.h>
22 #include <linux/kernel.h>
23 #include <linux/kthread.h>
24 #include <linux/llc.h>
25 #include <linux/mutex.h>
26 #include <linux/percpu.h>
27 #include <linux/rcupdate.h>
28 #include <linux/tcp.h>
29 #include <linux/udp.h>
30 #include <linux/version.h>
31 #include <linux/ethtool.h>
32 #include <linux/random.h>
33 #include <linux/wait.h>
34 #include <asm/system.h>
35 #include <asm/div64.h>
37 #include <linux/netfilter_bridge.h>
38 #include <linux/netfilter_ipv4.h>
39 #include <linux/inetdevice.h>
40 #include <linux/list.h>
41 #include <linux/rculist.h>
42 #include <linux/workqueue.h>
43 #include <linux/dmi.h>
46 #include "openvswitch/datapath-protocol.h"
55 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
56 EXPORT_SYMBOL(dp_ioctl_hook);
58 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
61 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
64 * It is safe to access the datapath and net_bridge_port structures with just
67 static struct datapath *dps[ODP_MAX];
68 static DEFINE_MUTEX(dp_mutex);
70 /* Number of milliseconds between runs of the maintenance thread. */
71 #define MAINT_SLEEP_MSECS 1000
73 static int new_nbp(struct datapath *, struct net_device *, int port_no);
75 /* Must be called with rcu_read_lock or dp_mutex. */
76 struct datapath *get_dp(int dp_idx)
78 if (dp_idx < 0 || dp_idx >= ODP_MAX)
80 return rcu_dereference(dps[dp_idx]);
82 EXPORT_SYMBOL_GPL(get_dp);
84 struct datapath *get_dp_locked(int dp_idx)
88 mutex_lock(&dp_mutex);
91 mutex_lock(&dp->mutex);
92 mutex_unlock(&dp_mutex);
96 static inline size_t br_nlmsg_size(void)
98 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
99 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
100 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
101 + nla_total_size(4) /* IFLA_MASTER */
102 + nla_total_size(4) /* IFLA_MTU */
103 + nla_total_size(4) /* IFLA_LINK */
104 + nla_total_size(1); /* IFLA_OPERSTATE */
107 static int dp_fill_ifinfo(struct sk_buff *skb,
108 const struct net_bridge_port *port,
109 int event, unsigned int flags)
111 const struct datapath *dp = port->dp;
112 const struct net_device *dev = port->dev;
113 struct ifinfomsg *hdr;
114 struct nlmsghdr *nlh;
116 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
120 hdr = nlmsg_data(nlh);
121 hdr->ifi_family = AF_BRIDGE;
123 hdr->ifi_type = dev->type;
124 hdr->ifi_index = dev->ifindex;
125 hdr->ifi_flags = dev_get_flags(dev);
128 NLA_PUT_STRING(skb, IFLA_IFNAME, dev->name);
129 NLA_PUT_U32(skb, IFLA_MASTER, dp->ports[ODPP_LOCAL]->dev->ifindex);
130 NLA_PUT_U32(skb, IFLA_MTU, dev->mtu);
131 #ifdef IFLA_OPERSTATE
132 NLA_PUT_U8(skb, IFLA_OPERSTATE,
133 netif_running(dev) ? dev->operstate : IF_OPER_DOWN);
137 NLA_PUT(skb, IFLA_ADDRESS, dev->addr_len, dev->dev_addr);
139 if (dev->ifindex != dev->iflink)
140 NLA_PUT_U32(skb, IFLA_LINK, dev->iflink);
142 return nlmsg_end(skb, nlh);
145 nlmsg_cancel(skb, nlh);
149 static void dp_ifinfo_notify(int event, struct net_bridge_port *port)
151 struct net *net = dev_net(port->dev);
155 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
159 err = dp_fill_ifinfo(skb, port, event, 0);
161 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
162 WARN_ON(err == -EMSGSIZE);
166 rtnl_notify(skb, net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
170 rtnl_set_sk_err(net, RTNLGRP_LINK, err);
173 static void release_dp(struct kobject *kobj)
175 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
179 struct kobj_type dp_ktype = {
180 .release = release_dp
183 static int create_dp(int dp_idx, const char __user *devnamep)
185 struct net_device *dp_dev;
186 char devname[IFNAMSIZ];
193 if (strncpy_from_user(devname, devnamep, IFNAMSIZ - 1) < 0)
195 devname[IFNAMSIZ - 1] = '\0';
197 snprintf(devname, sizeof devname, "of%d", dp_idx);
201 mutex_lock(&dp_mutex);
203 if (!try_module_get(THIS_MODULE))
206 /* Exit early if a datapath with that number already exists.
207 * (We don't use -EEXIST because that's ambiguous with 'devname'
208 * conflicting with an existing network device name.) */
214 dp = kzalloc(sizeof *dp, GFP_KERNEL);
217 INIT_LIST_HEAD(&dp->port_list);
218 mutex_init(&dp->mutex);
220 for (i = 0; i < DP_N_QUEUES; i++)
221 skb_queue_head_init(&dp->queues[i]);
222 init_waitqueue_head(&dp->waitqueue);
224 /* Initialize kobject for bridge. This will be added as
225 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
226 dp->ifobj.kset = NULL;
227 kobject_init(&dp->ifobj, &dp_ktype);
229 /* Allocate table. */
231 rcu_assign_pointer(dp->table, dp_table_create(DP_L1_SIZE));
235 /* Set up our datapath device. */
236 dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL);
237 err = PTR_ERR(dp_dev);
239 goto err_destroy_table;
241 err = new_nbp(dp, dp_dev, ODPP_LOCAL);
243 dp_dev_destroy(dp_dev);
244 goto err_destroy_table;
248 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
249 if (!dp->stats_percpu)
250 goto err_destroy_local_port;
252 rcu_assign_pointer(dps[dp_idx], dp);
253 mutex_unlock(&dp_mutex);
260 err_destroy_local_port:
261 dp_del_port(dp->ports[ODPP_LOCAL]);
263 dp_table_destroy(dp->table, 0);
267 module_put(THIS_MODULE);
269 mutex_unlock(&dp_mutex);
275 static void do_destroy_dp(struct datapath *dp)
277 struct net_bridge_port *p, *n;
280 list_for_each_entry_safe (p, n, &dp->port_list, node)
281 if (p->port_no != ODPP_LOCAL)
286 rcu_assign_pointer(dps[dp->dp_idx], NULL);
288 dp_del_port(dp->ports[ODPP_LOCAL]);
290 dp_table_destroy(dp->table, 1);
292 for (i = 0; i < DP_N_QUEUES; i++)
293 skb_queue_purge(&dp->queues[i]);
294 for (i = 0; i < DP_MAX_GROUPS; i++)
295 kfree(dp->groups[i]);
296 free_percpu(dp->stats_percpu);
297 kobject_put(&dp->ifobj);
298 module_put(THIS_MODULE);
301 static int destroy_dp(int dp_idx)
307 mutex_lock(&dp_mutex);
317 mutex_unlock(&dp_mutex);
322 static void release_nbp(struct kobject *kobj)
324 struct net_bridge_port *p = container_of(kobj, struct net_bridge_port, kobj);
328 struct kobj_type brport_ktype = {
330 .sysfs_ops = &brport_sysfs_ops,
332 .release = release_nbp
335 /* Called with RTNL lock and dp_mutex. */
336 static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no)
338 struct net_bridge_port *p;
340 if (dev->br_port != NULL)
343 p = kzalloc(sizeof(*p), GFP_KERNEL);
347 dev_set_promiscuity(dev, 1);
349 p->port_no = port_no;
353 rcu_assign_pointer(dev->br_port, p);
355 /* It would make sense to assign dev->br_port here too, but
356 * that causes packets received on internal ports to get caught
357 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
358 * back to network stack, but that's a waste of time. */
360 rcu_assign_pointer(dp->ports[port_no], p);
361 list_add_rcu(&p->node, &dp->port_list);
364 /* Initialize kobject for bridge. This will be added as
365 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
367 kobject_init(&p->kobj, &brport_ktype);
369 dp_ifinfo_notify(RTM_NEWLINK, p);
374 static int add_port(int dp_idx, struct odp_port __user *portp)
376 struct net_device *dev;
378 struct odp_port port;
383 if (copy_from_user(&port, portp, sizeof port))
385 port.devname[IFNAMSIZ - 1] = '\0';
388 dp = get_dp_locked(dp_idx);
391 goto out_unlock_rtnl;
393 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
394 if (!dp->ports[port_no])
400 if (!(port.flags & ODP_PORT_INTERNAL)) {
402 dev = dev_get_by_name(&init_net, port.devname);
407 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
411 dev = dp_dev_create(dp, port.devname, port_no);
418 err = new_nbp(dp, dev, port_no);
422 set_dp_devs_mtu(dp, dev);
423 dp_sysfs_add_if(dp->ports[port_no]);
425 err = __put_user(port_no, &port.port);
430 mutex_unlock(&dp->mutex);
437 int dp_del_port(struct net_bridge_port *p)
441 if (p->port_no != ODPP_LOCAL)
443 dp_ifinfo_notify(RTM_DELLINK, p);
447 if (is_dp_dev(p->dev)) {
448 /* Make sure that no packets arrive from now on, since
449 * dp_dev_xmit() will try to find itself through
450 * p->dp->ports[], and we're about to set that to null. */
451 netif_tx_disable(p->dev);
454 /* First drop references to device. */
455 dev_set_promiscuity(p->dev, -1);
456 list_del_rcu(&p->node);
457 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
458 rcu_assign_pointer(p->dev->br_port, NULL);
460 /* Then wait until no one is still using it, and destroy it. */
463 if (is_dp_dev(p->dev))
464 dp_dev_destroy(p->dev);
466 kobject_put(&p->kobj);
471 static int del_port(int dp_idx, int port_no)
473 struct net_bridge_port *p;
479 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
483 dp = get_dp_locked(dp_idx);
486 goto out_unlock_rtnl;
488 p = dp->ports[port_no];
493 err = dp_del_port(p);
496 mutex_unlock(&dp->mutex);
503 /* Must be called with rcu_read_lock. */
505 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
507 /* Make our own copy of the packet. Otherwise we will mangle the
508 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
509 * (No one comes after us, since we tell handle_bridge() that we took
511 skb = skb_share_check(skb, GFP_ATOMIC);
515 /* Push the Ethernet header back on. */
516 skb_push(skb, ETH_HLEN);
517 skb_reset_mac_header(skb);
518 dp_process_received_packet(skb, p);
521 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
522 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
524 struct datapath *dp = p->dp;
525 struct dp_stats_percpu *stats;
526 struct odp_flow_key key;
527 struct sw_flow *flow;
529 WARN_ON_ONCE(skb_shared(skb));
531 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
532 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
534 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
535 if (dp->drop_frags) {
542 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
544 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
545 flow_used(flow, skb);
546 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
551 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
556 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
557 * different set of devices!)
559 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
560 /* Called with rcu_read_lock and bottom-halves disabled. */
561 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
564 do_port_input(p, skb);
567 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
568 /* Called with rcu_read_lock and bottom-halves disabled. */
569 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
571 do_port_input(p, *pskb);
578 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
579 /* This code is based on a skb_checksum_setup from net/dev/core.c from a
580 * combination of Lenny's 2.6.26 Xen kernel and Xen's
581 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call this function
582 * directly because it isn't exported in all versions. */
583 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
585 if (ptr < (void *)skb->tail)
587 if (__pskb_pull_tail(skb,
588 ptr - (void *)skb->data - skb_headlen(skb))) {
595 int vswitch_skb_checksum_setup(struct sk_buff *skb)
600 __u16 csum_start, csum_offset;
602 if (!skb->proto_csum_blank)
605 if (skb->protocol != htons(ETH_P_IP))
608 if (!skb_pull_up_to(skb, skb_network_header(skb) + 1))
612 th = skb_network_header(skb) + 4 * iph->ihl;
614 csum_start = th - skb->head;
615 switch (iph->protocol) {
617 csum_offset = offsetof(struct tcphdr, check);
620 csum_offset = offsetof(struct udphdr, check);
624 printk(KERN_ERR "Attempting to checksum a non-"
625 "TCP/UDP packet, dropping a protocol"
626 " %d packet", iph->protocol);
630 if (!skb_pull_up_to(skb, th + csum_offset + 2))
633 skb->ip_summed = CHECKSUM_PARTIAL;
634 skb->proto_csum_blank = 0;
636 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
637 skb->csum_start = csum_start;
638 skb->csum_offset = csum_offset;
640 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
641 skb->csum = csum_offset;
650 int vswitch_skb_checksum_setup(struct sk_buff *skb) { return 0; }
651 #endif /* CONFIG_XEN && linux == 2.6.18 */
653 /* Types of checksums that we can receive (these all refer to L4 checksums):
654 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
655 * (though not verified) checksum in packet but not in skb->csum. Packets
656 * from the bridge local port will also have this type.
657 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
658 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
659 * a valid skb->csum. Importantly, both contain a full checksum (not
660 * verified) in the packet itself. The only difference is that if the
661 * packet gets to L4 processing on this machine (not in DomU) we won't
662 * have to recompute the checksum to verify. Most hardware devices do not
663 * produce packets with this type, even if they support receive checksum
664 * offloading (they produce type #5).
665 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
666 * be computed if it is sent off box. Unfortunately on earlier kernels,
667 * this case is impossible to distinguish from #2, despite having opposite
668 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
669 * to distinguish the different states. The only real user of this type
670 * with bridging is Xen (on later kernels).
671 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
672 * generated locally by a Xen DomU and has a partial checksum. If it is
673 * handled on this machine (Dom0 or DomU), then the checksum will not be
674 * computed. If it goes off box, the checksum in the packet needs to
675 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
676 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
677 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
678 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
679 * full checksum or using a protocol without a checksum. skb->csum is
680 * undefined. This is common from devices with receive checksum
681 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
682 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
684 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
685 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
686 * based on whether it is on the transmit or receive path. After the datapath
687 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
688 * checksum, we will panic. Since we can receive packets with checksums, we
689 * assume that all CHECKSUM_HW packets have checksums and map them to
690 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
691 * packet is processed by the local IP stack, in which case it will need to
692 * be reverified). If we receive a packet with CHECKSUM_HW that really means
693 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
694 * shouldn't be any devices that do this with bridging.
696 * The bridge has similar behavior and this function closely resembles
697 * skb_forward_csum(). It is slightly different because we are only concerned
698 * with bridging and not other types of forwarding and can get away with
699 * slightly more optimal behavior.*/
701 forward_ip_summed(struct sk_buff *skb)
704 if (skb->ip_summed == CHECKSUM_HW)
705 skb->ip_summed = CHECKSUM_NONE;
709 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
710 * unless we broke up a GSO packet. */
712 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
713 int queue_no, u32 arg)
715 struct sk_buff *nskb;
719 port_no = ODPP_LOCAL;
721 if (skb->dev->br_port)
722 port_no = skb->dev->br_port->port_no;
723 else if (is_dp_dev(skb->dev))
724 port_no = dp_dev_priv(skb->dev)->port_no;
728 struct odp_msg *header;
733 /* If a checksum-deferred packet is forwarded to the
734 * controller, correct the pointers and checksum. This happens
735 * on a regular basis only on Xen, on which VMs can pass up
736 * packets that do not have their checksum computed.
738 err = vswitch_skb_checksum_setup(skb);
742 if (skb->ip_summed == CHECKSUM_PARTIAL) {
743 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
744 /* Until 2.6.22, the start of the transport header was
745 * also the start of data to be checksummed. Linux
746 * 2.6.22 introduced the csum_start field for this
747 * purpose, but we should point the transport header to
748 * it anyway for backward compatibility, as
749 * dev_queue_xmit() does even in 2.6.28. */
750 skb_set_transport_header(skb, skb->csum_start -
753 err = skb_checksum_help(skb);
758 if (skb->ip_summed == CHECKSUM_HW) {
759 err = skb_checksum_help(skb, 0);
765 err = skb_cow(skb, sizeof *header);
769 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
770 header->type = queue_no;
771 header->length = skb->len;
772 header->port = port_no;
773 header->reserved = 0;
775 skb_queue_tail(queue, skb);
783 while ((skb = nskb) != NULL) {
791 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
794 struct dp_stats_percpu *stats;
795 struct sk_buff_head *queue;
798 WARN_ON_ONCE(skb_shared(skb));
799 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR);
801 queue = &dp->queues[queue_no];
803 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
806 forward_ip_summed(skb);
808 /* Break apart GSO packets into their component pieces. Otherwise
809 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
810 if (skb_is_gso(skb)) {
811 struct sk_buff *nskb = skb_gso_segment(skb, 0);
815 if (unlikely(IS_ERR(skb))) {
820 /* XXX This case might not be possible. It's hard to
821 * tell from the skb_gso_segment() code and comment. */
825 err = queue_control_packets(skb, queue, queue_no, arg);
826 wake_up_interruptible(&dp->waitqueue);
832 stats = percpu_ptr(dp->stats_percpu, get_cpu());
839 static int flush_flows(struct datapath *dp)
842 return dp_table_flush(dp);
845 static int validate_actions(const struct sw_flow_actions *actions)
849 for (i = 0; i < actions->n_actions; i++) {
850 const union odp_action *a = &actions->actions[i];
853 if (a->output.port >= DP_MAX_PORTS)
857 case ODPAT_OUTPUT_GROUP:
858 if (a->output_group.group >= DP_MAX_GROUPS)
862 case ODPAT_SET_VLAN_VID:
863 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
867 case ODPAT_SET_VLAN_PCP:
868 if (a->vlan_pcp.vlan_pcp
869 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
874 if (a->type >= ODPAT_N_ACTIONS)
883 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
885 struct sw_flow_actions *actions;
888 actions = flow_actions_alloc(flow->n_actions);
889 error = PTR_ERR(actions);
894 if (copy_from_user(actions->actions, flow->actions,
895 flow->n_actions * sizeof(union odp_action)))
896 goto error_free_actions;
897 error = validate_actions(actions);
899 goto error_free_actions;
906 return ERR_PTR(error);
909 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
911 if (flow->used.tv_sec) {
912 stats->used_sec = flow->used.tv_sec;
913 stats->used_nsec = flow->used.tv_nsec;
916 stats->used_nsec = 0;
918 stats->n_packets = flow->packet_count;
919 stats->n_bytes = flow->byte_count;
920 stats->ip_tos = flow->ip_tos;
921 stats->tcp_flags = flow->tcp_flags;
925 static void clear_stats(struct sw_flow *flow)
927 flow->used.tv_sec = flow->used.tv_nsec = 0;
930 flow->packet_count = 0;
931 flow->byte_count = 0;
934 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
936 struct odp_flow_put uf;
937 struct sw_flow *flow;
938 struct dp_table *table;
939 struct odp_flow_stats stats;
943 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
945 uf.flow.key.reserved = 0;
947 table = rcu_dereference(dp->table);
948 flow = dp_table_lookup(table, &uf.flow.key);
951 struct sw_flow_actions *acts;
954 if (!(uf.flags & ODPPF_CREATE))
957 /* Expand table, if necessary, to make room. */
958 if (dp->n_flows >= table->n_buckets) {
960 if (table->n_buckets >= DP_MAX_BUCKETS)
963 error = dp_table_expand(dp);
966 table = rcu_dereference(dp->table);
971 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
974 flow->key = uf.flow.key;
975 spin_lock_init(&flow->lock);
978 /* Obtain actions. */
979 acts = get_actions(&uf.flow);
980 error = PTR_ERR(acts);
982 goto error_free_flow;
983 rcu_assign_pointer(flow->sf_acts, acts);
985 /* Put flow in bucket. */
986 error = dp_table_insert(table, flow);
988 goto error_free_flow_acts;
990 memset(&stats, 0, sizeof(struct odp_flow_stats));
992 /* We found a matching flow. */
993 struct sw_flow_actions *old_acts, *new_acts;
994 unsigned long int flags;
996 /* Bail out if we're not allowed to modify an existing flow. */
998 if (!(uf.flags & ODPPF_MODIFY))
1002 new_acts = get_actions(&uf.flow);
1003 error = PTR_ERR(new_acts);
1004 if (IS_ERR(new_acts))
1006 old_acts = rcu_dereference(flow->sf_acts);
1007 if (old_acts->n_actions != new_acts->n_actions ||
1008 memcmp(old_acts->actions, new_acts->actions,
1009 sizeof(union odp_action) * old_acts->n_actions)) {
1010 rcu_assign_pointer(flow->sf_acts, new_acts);
1011 flow_deferred_free_acts(old_acts);
1016 /* Fetch stats, then clear them if necessary. */
1017 spin_lock_irqsave(&flow->lock, flags);
1018 get_stats(flow, &stats);
1019 if (uf.flags & ODPPF_ZERO_STATS)
1021 spin_unlock_irqrestore(&flow->lock, flags);
1024 /* Copy stats to userspace. */
1025 if (__copy_to_user(&ufp->flow.stats, &stats,
1026 sizeof(struct odp_flow_stats)))
1030 error_free_flow_acts:
1031 kfree(flow->sf_acts);
1033 kmem_cache_free(flow_cache, flow);
1038 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1040 union odp_action __user *actions;
1041 struct sw_flow_actions *sf_acts;
1044 if (__get_user(actions, &ufp->actions) ||
1045 __get_user(n_actions, &ufp->n_actions))
1051 sf_acts = rcu_dereference(flow->sf_acts);
1052 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1053 (actions && copy_to_user(actions, sf_acts->actions,
1054 sizeof(union odp_action) *
1055 min(sf_acts->n_actions, n_actions))))
1061 static int answer_query(struct sw_flow *flow, u32 query_flags,
1062 struct odp_flow __user *ufp)
1064 struct odp_flow_stats stats;
1065 unsigned long int flags;
1067 spin_lock_irqsave(&flow->lock, flags);
1068 get_stats(flow, &stats);
1070 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1071 flow->tcp_flags = 0;
1073 spin_unlock_irqrestore(&flow->lock, flags);
1075 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1077 return put_actions(flow, ufp);
1080 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1082 struct dp_table *table = rcu_dereference(dp->table);
1084 struct sw_flow *flow;
1088 if (copy_from_user(&uf, ufp, sizeof uf))
1090 uf.key.reserved = 0;
1092 flow = dp_table_lookup(table, &uf.key);
1097 /* XXX redundant lookup */
1098 error = dp_table_delete(table, flow);
1102 /* XXX These statistics might lose a few packets, since other CPUs can
1103 * be using this flow. We used to synchronize_rcu() to make sure that
1104 * we get completely accurate stats, but that blows our performance,
1107 error = answer_query(flow, 0, ufp);
1108 flow_deferred_free(flow);
1114 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1116 struct dp_table *table = rcu_dereference(dp->table);
1118 for (i = 0; i < flowvec->n_flows; i++) {
1119 struct __user odp_flow *ufp = &flowvec->flows[i];
1121 struct sw_flow *flow;
1124 if (__copy_from_user(&uf, ufp, sizeof uf))
1126 uf.key.reserved = 0;
1128 flow = dp_table_lookup(table, &uf.key);
1130 error = __put_user(ENOENT, &ufp->stats.error);
1132 error = answer_query(flow, uf.flags, ufp);
1136 return flowvec->n_flows;
1139 struct list_flows_cbdata {
1140 struct odp_flow __user *uflows;
1145 static int list_flow(struct sw_flow *flow, void *cbdata_)
1147 struct list_flows_cbdata *cbdata = cbdata_;
1148 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1151 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1153 error = answer_query(flow, 0, ufp);
1157 if (cbdata->listed_flows >= cbdata->n_flows)
1158 return cbdata->listed_flows;
1162 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1164 struct list_flows_cbdata cbdata;
1167 if (!flowvec->n_flows)
1170 cbdata.uflows = flowvec->flows;
1171 cbdata.n_flows = flowvec->n_flows;
1172 cbdata.listed_flows = 0;
1173 error = dp_table_foreach(rcu_dereference(dp->table),
1174 list_flow, &cbdata);
1175 return error ? error : cbdata.listed_flows;
1178 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1179 int (*function)(struct datapath *,
1180 const struct odp_flowvec *))
1182 struct odp_flowvec __user *uflowvec;
1183 struct odp_flowvec flowvec;
1186 uflowvec = (struct odp_flowvec __user *)argp;
1187 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1188 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1191 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1194 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1195 flowvec.n_flows * sizeof(struct odp_flow)))
1198 retval = function(dp, &flowvec);
1199 return (retval < 0 ? retval
1200 : retval == flowvec.n_flows ? 0
1201 : __put_user(retval, &uflowvec->n_flows));
1204 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1206 struct odp_execute execute;
1207 struct odp_flow_key key;
1208 struct sk_buff *skb;
1209 struct sw_flow_actions *actions;
1214 if (copy_from_user(&execute, executep, sizeof execute))
1218 if (execute.length < ETH_HLEN || execute.length > 65535)
1222 actions = flow_actions_alloc(execute.n_actions);
1227 if (copy_from_user(actions->actions, execute.actions,
1228 execute.n_actions * sizeof *execute.actions))
1229 goto error_free_actions;
1231 err = validate_actions(actions);
1233 goto error_free_actions;
1236 skb = alloc_skb(execute.length, GFP_KERNEL);
1238 goto error_free_actions;
1239 if (execute.in_port < DP_MAX_PORTS) {
1240 struct net_bridge_port *p = dp->ports[execute.in_port];
1246 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1248 goto error_free_skb;
1250 skb_reset_mac_header(skb);
1253 /* Normally, setting the skb 'protocol' field would be handled by a
1254 * call to eth_type_trans(), but it assumes there's a sending
1255 * device, which we may not have. */
1256 if (ntohs(eth->h_proto) >= 1536)
1257 skb->protocol = eth->h_proto;
1259 skb->protocol = htons(ETH_P_802_2);
1261 flow_extract(skb, execute.in_port, &key);
1262 err = execute_actions(dp, skb, &key, actions->actions,
1263 actions->n_actions, GFP_KERNEL);
1275 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1277 struct odp_stats stats;
1280 stats.n_flows = dp->n_flows;
1281 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets;
1282 stats.max_capacity = DP_MAX_BUCKETS;
1283 stats.n_ports = dp->n_ports;
1284 stats.max_ports = DP_MAX_PORTS;
1285 stats.max_groups = DP_MAX_GROUPS;
1286 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1287 for_each_possible_cpu(i) {
1288 const struct dp_stats_percpu *s;
1289 s = percpu_ptr(dp->stats_percpu, i);
1290 stats.n_frags += s->n_frags;
1291 stats.n_hit += s->n_hit;
1292 stats.n_missed += s->n_missed;
1293 stats.n_lost += s->n_lost;
1295 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1296 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1297 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1300 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1301 int dp_min_mtu(const struct datapath *dp)
1303 struct net_bridge_port *p;
1308 list_for_each_entry_rcu (p, &dp->port_list, node) {
1309 struct net_device *dev = p->dev;
1311 /* Skip any internal ports, since that's what we're trying to
1316 if (!mtu || dev->mtu < mtu)
1320 return mtu ? mtu : ETH_DATA_LEN;
1323 /* Sets the MTU of all datapath devices to the minimum of the ports. 'dev'
1324 * is the device whose MTU may have changed. Must be called with RTNL lock
1326 void set_dp_devs_mtu(const struct datapath *dp, struct net_device *dev)
1328 struct net_bridge_port *p;
1336 mtu = dp_min_mtu(dp);
1338 list_for_each_entry_rcu (p, &dp->port_list, node) {
1339 struct net_device *br_dev = p->dev;
1341 if (is_dp_dev(br_dev))
1342 dev_set_mtu(br_dev, mtu);
1347 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1350 memset(&op, 0, sizeof op);
1351 strncpy(op.devname, p->dev->name, sizeof op.devname);
1352 op.port = p->port_no;
1353 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1354 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1358 query_port(struct datapath *dp, struct odp_port __user *uport)
1360 struct odp_port port;
1362 if (copy_from_user(&port, uport, sizeof port))
1364 if (port.devname[0]) {
1365 struct net_bridge_port *p;
1366 struct net_device *dev;
1369 port.devname[IFNAMSIZ - 1] = '\0';
1371 dev = dev_get_by_name(&init_net, port.devname);
1376 if (!p && is_dp_dev(dev)) {
1377 struct dp_dev *dp_dev = dp_dev_priv(dev);
1378 if (dp_dev->dp == dp)
1379 p = dp->ports[dp_dev->port_no];
1381 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1386 if (port.port >= DP_MAX_PORTS)
1388 if (!dp->ports[port.port])
1390 return put_port(dp->ports[port.port], uport);
1395 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1397 struct odp_portvec pv;
1398 struct net_bridge_port *p;
1401 if (copy_from_user(&pv, pvp, sizeof pv))
1406 list_for_each_entry_rcu (p, &dp->port_list, node) {
1407 if (put_port(p, &pv.ports[idx]))
1409 if (idx++ >= pv.n_ports)
1413 return put_user(dp->n_ports, &pvp->n_ports);
1416 /* RCU callback for freeing a dp_port_group */
1417 static void free_port_group(struct rcu_head *rcu)
1419 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1424 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1426 struct odp_port_group pg;
1427 struct dp_port_group *new_group, *old_group;
1431 if (copy_from_user(&pg, upg, sizeof pg))
1435 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1439 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1444 new_group->n_ports = pg.n_ports;
1446 if (copy_from_user(new_group->ports, pg.ports,
1447 sizeof(u16) * pg.n_ports))
1450 old_group = rcu_dereference(dp->groups[pg.group]);
1451 rcu_assign_pointer(dp->groups[pg.group], new_group);
1453 call_rcu(&old_group->rcu, free_port_group);
1463 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1465 struct odp_port_group pg;
1466 struct dp_port_group *g;
1469 if (copy_from_user(&pg, upg, sizeof pg))
1472 if (pg.group >= DP_MAX_GROUPS)
1475 g = dp->groups[pg.group];
1476 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1477 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1480 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1486 static int get_listen_mask(const struct file *f)
1488 return (long)f->private_data;
1491 static void set_listen_mask(struct file *f, int listen_mask)
1493 f->private_data = (void*)(long)listen_mask;
1496 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1499 int dp_idx = iminor(f->f_dentry->d_inode);
1500 struct datapath *dp;
1501 int drop_frags, listeners, port_no;
1504 /* Handle commands with special locking requirements up front. */
1507 err = create_dp(dp_idx, (char __user *)argp);
1510 case ODP_DP_DESTROY:
1511 err = destroy_dp(dp_idx);
1515 err = add_port(dp_idx, (struct odp_port __user *)argp);
1519 err = get_user(port_no, (int __user *)argp);
1521 err = del_port(dp_idx, port_no);
1525 dp = get_dp_locked(dp_idx);
1532 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1535 case ODP_GET_DROP_FRAGS:
1536 err = put_user(dp->drop_frags, (int __user *)argp);
1539 case ODP_SET_DROP_FRAGS:
1540 err = get_user(drop_frags, (int __user *)argp);
1544 if (drop_frags != 0 && drop_frags != 1)
1546 dp->drop_frags = drop_frags;
1550 case ODP_GET_LISTEN_MASK:
1551 err = put_user(get_listen_mask(f), (int __user *)argp);
1554 case ODP_SET_LISTEN_MASK:
1555 err = get_user(listeners, (int __user *)argp);
1559 if (listeners & ~ODPL_ALL)
1562 set_listen_mask(f, listeners);
1565 case ODP_PORT_QUERY:
1566 err = query_port(dp, (struct odp_port __user *)argp);
1570 err = list_ports(dp, (struct odp_portvec __user *)argp);
1573 case ODP_PORT_GROUP_SET:
1574 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1577 case ODP_PORT_GROUP_GET:
1578 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1581 case ODP_FLOW_FLUSH:
1582 err = flush_flows(dp);
1586 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1590 err = del_flow(dp, (struct odp_flow __user *)argp);
1594 err = do_flowvec_ioctl(dp, argp, query_flows);
1598 err = do_flowvec_ioctl(dp, argp, list_flows);
1602 err = do_execute(dp, (struct odp_execute __user *)argp);
1609 mutex_unlock(&dp->mutex);
1614 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1617 for (i = 0; i < DP_N_QUEUES; i++) {
1618 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1624 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1627 /* XXX is there sufficient synchronization here? */
1628 int listeners = get_listen_mask(f);
1629 int dp_idx = iminor(f->f_dentry->d_inode);
1630 struct datapath *dp = get_dp(dp_idx);
1631 struct sk_buff *skb;
1632 struct iovec __user iov;
1639 if (nbytes == 0 || !listeners)
1645 for (i = 0; i < DP_N_QUEUES; i++) {
1646 if (listeners & (1 << i)) {
1647 skb = skb_dequeue(&dp->queues[i]);
1653 if (f->f_flags & O_NONBLOCK) {
1658 wait_event_interruptible(dp->waitqueue,
1659 dp_has_packet_of_interest(dp,
1662 if (signal_pending(current)) {
1663 retval = -ERESTARTSYS;
1668 copy_bytes = min_t(size_t, skb->len, nbytes);
1670 iov.iov_len = copy_bytes;
1671 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1673 retval = copy_bytes;
1680 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1682 /* XXX is there sufficient synchronization here? */
1683 int dp_idx = iminor(file->f_dentry->d_inode);
1684 struct datapath *dp = get_dp(dp_idx);
1689 poll_wait(file, &dp->waitqueue, wait);
1690 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1691 mask |= POLLIN | POLLRDNORM;
1693 mask = POLLIN | POLLRDNORM | POLLHUP;
1698 struct file_operations openvswitch_fops = {
1699 /* XXX .aio_read = openvswitch_aio_read, */
1700 .read = openvswitch_read,
1701 .poll = openvswitch_poll,
1702 .unlocked_ioctl = openvswitch_ioctl,
1703 /* XXX .fasync = openvswitch_fasync, */
1708 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)
1709 static struct llc_sap *dp_stp_sap;
1711 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1712 struct packet_type *pt, struct net_device *orig_dev)
1714 /* We don't really care about STP packets, we just listen for them for
1715 * mutual exclusion with the bridge module, so this just discards
1721 static int dp_avoid_bridge_init(void)
1723 /* Register to receive STP packets because the bridge module also
1724 * attempts to do so. Since there can only be a single listener for a
1725 * given protocol, this provides mutual exclusion against the bridge
1726 * module, preventing both of them from being loaded at the same
1728 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1730 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1736 static void dp_avoid_bridge_exit(void)
1738 llc_sap_put(dp_stp_sap);
1740 #else /* Linux 2.6.27 or later. */
1741 static int dp_avoid_bridge_init(void)
1743 /* Linux 2.6.27 introduces a way for multiple clients to register for
1744 * STP packets, which interferes with what we try to do above.
1745 * Instead, just check whether there's a bridge hook defined. This is
1746 * not as safe--the bridge module is willing to load over the top of
1747 * us--but it provides a little bit of protection. */
1748 if (br_handle_frame_hook) {
1749 printk(KERN_ERR "openvswitch: bridge module is loaded, cannot load over it\n");
1755 static void dp_avoid_bridge_exit(void)
1757 /* Nothing to do. */
1759 #endif /* Linux 2.6.27 or later */
1761 static int __init dp_init(void)
1765 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1767 err = dp_avoid_bridge_init();
1775 err = register_netdevice_notifier(&dp_device_notifier);
1777 goto error_flow_exit;
1779 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1781 goto error_unreg_notifier;
1783 /* Hook into callback used by the bridge to intercept packets.
1784 * Parasites we are. */
1785 br_handle_frame_hook = dp_frame_hook;
1789 error_unreg_notifier:
1790 unregister_netdevice_notifier(&dp_device_notifier);
1797 static void dp_cleanup(void)
1800 unregister_chrdev(major, "openvswitch");
1801 unregister_netdevice_notifier(&dp_device_notifier);
1803 br_handle_frame_hook = NULL;
1804 dp_avoid_bridge_exit();
1807 module_init(dp_init);
1808 module_exit(dp_cleanup);
1810 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1811 MODULE_LICENSE("GPL");