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 dp_sysfs_add_if(dp->ports[port_no]);
424 err = __put_user(port_no, &port.port);
429 mutex_unlock(&dp->mutex);
436 int dp_del_port(struct net_bridge_port *p)
440 if (p->port_no != ODPP_LOCAL)
442 dp_ifinfo_notify(RTM_DELLINK, p);
446 if (is_dp_dev(p->dev)) {
447 /* Make sure that no packets arrive from now on, since
448 * dp_dev_xmit() will try to find itself through
449 * p->dp->ports[], and we're about to set that to null. */
450 netif_tx_disable(p->dev);
453 /* First drop references to device. */
454 dev_set_promiscuity(p->dev, -1);
455 list_del_rcu(&p->node);
456 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
457 rcu_assign_pointer(p->dev->br_port, NULL);
459 /* Then wait until no one is still using it, and destroy it. */
462 if (is_dp_dev(p->dev))
463 dp_dev_destroy(p->dev);
465 kobject_put(&p->kobj);
470 static int del_port(int dp_idx, int port_no)
472 struct net_bridge_port *p;
478 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
482 dp = get_dp_locked(dp_idx);
485 goto out_unlock_rtnl;
487 p = dp->ports[port_no];
492 err = dp_del_port(p);
495 mutex_unlock(&dp->mutex);
502 /* Must be called with rcu_read_lock. */
504 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
506 /* Make our own copy of the packet. Otherwise we will mangle the
507 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
508 * (No one comes after us, since we tell handle_bridge() that we took
510 skb = skb_share_check(skb, GFP_ATOMIC);
514 /* Push the Ethernet header back on. */
515 skb_push(skb, ETH_HLEN);
516 skb_reset_mac_header(skb);
517 dp_process_received_packet(skb, p);
520 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
521 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
523 struct datapath *dp = p->dp;
524 struct dp_stats_percpu *stats;
525 struct odp_flow_key key;
526 struct sw_flow *flow;
528 WARN_ON_ONCE(skb_shared(skb));
530 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
531 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
533 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
534 if (dp->drop_frags) {
541 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
543 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
544 flow_used(flow, skb);
545 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
550 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
555 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
556 * different set of devices!)
558 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
559 /* Called with rcu_read_lock and bottom-halves disabled. */
560 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
563 do_port_input(p, skb);
566 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
567 /* Called with rcu_read_lock and bottom-halves disabled. */
568 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
570 do_port_input(p, *pskb);
577 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
578 /* This code is copied verbatim from net/dev/core.c in Xen's
579 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
580 * directly because they aren't exported. */
581 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
583 if (ptr < (void *)skb->tail)
585 if (__pskb_pull_tail(skb,
586 ptr - (void *)skb->data - skb_headlen(skb))) {
593 int vswitch_skb_checksum_setup(struct sk_buff *skb)
595 if (skb->proto_csum_blank) {
596 if (skb->protocol != htons(ETH_P_IP))
598 if (!skb_pull_up_to(skb, skb->nh.iph + 1))
600 skb->h.raw = (unsigned char *)skb->nh.iph + 4*skb->nh.iph->ihl;
601 switch (skb->nh.iph->protocol) {
603 skb->csum = offsetof(struct tcphdr, check);
606 skb->csum = offsetof(struct udphdr, check);
610 printk(KERN_ERR "Attempting to checksum a non-"
611 "TCP/UDP packet, dropping a protocol"
612 " %d packet", skb->nh.iph->protocol);
615 if (!skb_pull_up_to(skb, skb->h.raw + skb->csum + 2))
617 skb->ip_summed = CHECKSUM_HW;
618 skb->proto_csum_blank = 0;
625 int vswitch_skb_checksum_setup(struct sk_buff *skb) { return 0; }
626 #endif /* CONFIG_XEN && linux == 2.6.18 */
628 /* Types of checksums that we can receive (these all refer to L4 checksums):
629 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
630 * (though not verified) checksum in packet but not in skb->csum. Packets
631 * from the bridge local port will also have this type.
632 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
633 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
634 * a valid skb->csum. Importantly, both contain a full checksum (not
635 * verified) in the packet itself. The only difference is that if the
636 * packet gets to L4 processing on this machine (not in DomU) we won't
637 * have to recompute the checksum to verify. Most hardware devices do not
638 * produce packets with this type, even if they support receive checksum
639 * offloading (they produce type #5).
640 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
641 * be computed if it is sent off box. Unfortunately on earlier kernels,
642 * this case is impossible to distinguish from #2, despite having opposite
643 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
644 * to distinguish the different states. The only real user of this type
645 * with bridging is Xen (on later kernels).
646 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
647 * generated locally by a Xen DomU and has a partial checksum. If it is
648 * handled on this machine (Dom0 or DomU), then the checksum will not be
649 * computed. If it goes off box, the checksum in the packet needs to
650 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
651 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
652 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
653 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
654 * full checksum or using a protocol without a checksum. skb->csum is
655 * undefined. This is common from devices with receive checksum
656 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
657 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
659 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
660 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
661 * based on whether it is on the transmit or receive path. After the datapath
662 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
663 * checksum, we will panic. Since we can receive packets with checksums, we
664 * assume that all CHECKSUM_HW packets have checksums and map them to
665 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
666 * packet is processed by the local IP stack, in which case it will need to
667 * be reverified). If we receive a packet with CHECKSUM_HW that really means
668 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
669 * shouldn't be any devices that do this with bridging.
671 * The bridge has similar behavior and this function closely resembles
672 * skb_forward_csum(). It is slightly different because we are only concerned
673 * with bridging and not other types of forwarding and can get away with
674 * slightly more optimal behavior.*/
676 forward_ip_summed(struct sk_buff *skb)
679 if (skb->ip_summed == CHECKSUM_HW)
680 skb->ip_summed = CHECKSUM_NONE;
684 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
685 * unless we broke up a GSO packet. */
687 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
688 int queue_no, u32 arg)
690 struct sk_buff *nskb;
694 port_no = ODPP_LOCAL;
696 if (skb->dev->br_port)
697 port_no = skb->dev->br_port->port_no;
698 else if (is_dp_dev(skb->dev))
699 port_no = dp_dev_priv(skb->dev)->port_no;
703 struct odp_msg *header;
708 /* If a checksum-deferred packet is forwarded to the
709 * controller, correct the pointers and checksum. This happens
710 * on a regular basis only on Xen, on which VMs can pass up
711 * packets that do not have their checksum computed.
713 err = vswitch_skb_checksum_setup(skb);
717 if (skb->ip_summed == CHECKSUM_PARTIAL) {
718 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
719 /* Until 2.6.22, the start of the transport header was
720 * also the start of data to be checksummed. Linux
721 * 2.6.22 introduced the csum_start field for this
722 * purpose, but we should point the transport header to
723 * it anyway for backward compatibility, as
724 * dev_queue_xmit() does even in 2.6.28. */
725 skb_set_transport_header(skb, skb->csum_start -
728 err = skb_checksum_help(skb);
733 if (skb->ip_summed == CHECKSUM_HW) {
734 err = skb_checksum_help(skb, 0);
740 err = skb_cow(skb, sizeof *header);
744 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
745 header->type = queue_no;
746 header->length = skb->len;
747 header->port = port_no;
748 header->reserved = 0;
750 skb_queue_tail(queue, skb);
758 while ((skb = nskb) != NULL) {
766 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
769 struct dp_stats_percpu *stats;
770 struct sk_buff_head *queue;
773 WARN_ON_ONCE(skb_shared(skb));
774 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR);
776 queue = &dp->queues[queue_no];
778 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
781 forward_ip_summed(skb);
783 /* Break apart GSO packets into their component pieces. Otherwise
784 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
785 if (skb_is_gso(skb)) {
786 struct sk_buff *nskb = skb_gso_segment(skb, 0);
790 if (unlikely(IS_ERR(skb))) {
795 /* XXX This case might not be possible. It's hard to
796 * tell from the skb_gso_segment() code and comment. */
800 err = queue_control_packets(skb, queue, queue_no, arg);
801 wake_up_interruptible(&dp->waitqueue);
807 stats = percpu_ptr(dp->stats_percpu, get_cpu());
814 static int flush_flows(struct datapath *dp)
817 return dp_table_flush(dp);
820 static int validate_actions(const struct sw_flow_actions *actions)
824 for (i = 0; i < actions->n_actions; i++) {
825 const union odp_action *a = &actions->actions[i];
828 if (a->output.port >= DP_MAX_PORTS)
832 case ODPAT_OUTPUT_GROUP:
833 if (a->output_group.group >= DP_MAX_GROUPS)
837 case ODPAT_SET_VLAN_VID:
838 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
842 case ODPAT_SET_VLAN_PCP:
843 if (a->vlan_pcp.vlan_pcp
844 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
849 if (a->type >= ODPAT_N_ACTIONS)
858 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
860 struct sw_flow_actions *actions;
863 actions = flow_actions_alloc(flow->n_actions);
864 error = PTR_ERR(actions);
869 if (copy_from_user(actions->actions, flow->actions,
870 flow->n_actions * sizeof(union odp_action)))
871 goto error_free_actions;
872 error = validate_actions(actions);
874 goto error_free_actions;
881 return ERR_PTR(error);
884 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
886 if (flow->used.tv_sec) {
887 stats->used_sec = flow->used.tv_sec;
888 stats->used_nsec = flow->used.tv_nsec;
891 stats->used_nsec = 0;
893 stats->n_packets = flow->packet_count;
894 stats->n_bytes = flow->byte_count;
895 stats->ip_tos = flow->ip_tos;
896 stats->tcp_flags = flow->tcp_flags;
900 static void clear_stats(struct sw_flow *flow)
902 flow->used.tv_sec = flow->used.tv_nsec = 0;
905 flow->packet_count = 0;
906 flow->byte_count = 0;
909 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
911 struct odp_flow_put uf;
912 struct sw_flow *flow;
913 struct dp_table *table;
914 struct odp_flow_stats stats;
918 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
920 uf.flow.key.reserved = 0;
922 table = rcu_dereference(dp->table);
923 flow = dp_table_lookup(table, &uf.flow.key);
926 struct sw_flow_actions *acts;
929 if (!(uf.flags & ODPPF_CREATE))
932 /* Expand table, if necessary, to make room. */
933 if (dp->n_flows >= table->n_buckets) {
935 if (table->n_buckets >= DP_MAX_BUCKETS)
938 error = dp_table_expand(dp);
941 table = rcu_dereference(dp->table);
946 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
949 flow->key = uf.flow.key;
950 spin_lock_init(&flow->lock);
953 /* Obtain actions. */
954 acts = get_actions(&uf.flow);
955 error = PTR_ERR(acts);
957 goto error_free_flow;
958 rcu_assign_pointer(flow->sf_acts, acts);
960 /* Put flow in bucket. */
961 error = dp_table_insert(table, flow);
963 goto error_free_flow_acts;
965 memset(&stats, 0, sizeof(struct odp_flow_stats));
967 /* We found a matching flow. */
968 struct sw_flow_actions *old_acts, *new_acts;
969 unsigned long int flags;
971 /* Bail out if we're not allowed to modify an existing flow. */
973 if (!(uf.flags & ODPPF_MODIFY))
977 new_acts = get_actions(&uf.flow);
978 error = PTR_ERR(new_acts);
979 if (IS_ERR(new_acts))
981 old_acts = rcu_dereference(flow->sf_acts);
982 if (old_acts->n_actions != new_acts->n_actions ||
983 memcmp(old_acts->actions, new_acts->actions,
984 sizeof(union odp_action) * old_acts->n_actions)) {
985 rcu_assign_pointer(flow->sf_acts, new_acts);
986 flow_deferred_free_acts(old_acts);
991 /* Fetch stats, then clear them if necessary. */
992 spin_lock_irqsave(&flow->lock, flags);
993 get_stats(flow, &stats);
994 if (uf.flags & ODPPF_ZERO_STATS)
996 spin_unlock_irqrestore(&flow->lock, flags);
999 /* Copy stats to userspace. */
1000 if (__copy_to_user(&ufp->flow.stats, &stats,
1001 sizeof(struct odp_flow_stats)))
1005 error_free_flow_acts:
1006 kfree(flow->sf_acts);
1008 kmem_cache_free(flow_cache, flow);
1013 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1015 union odp_action __user *actions;
1016 struct sw_flow_actions *sf_acts;
1019 if (__get_user(actions, &ufp->actions) ||
1020 __get_user(n_actions, &ufp->n_actions))
1026 sf_acts = rcu_dereference(flow->sf_acts);
1027 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1028 (actions && copy_to_user(actions, sf_acts->actions,
1029 sizeof(union odp_action) *
1030 min(sf_acts->n_actions, n_actions))))
1036 static int answer_query(struct sw_flow *flow, u32 query_flags,
1037 struct odp_flow __user *ufp)
1039 struct odp_flow_stats stats;
1040 unsigned long int flags;
1042 spin_lock_irqsave(&flow->lock, flags);
1043 get_stats(flow, &stats);
1045 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1046 flow->tcp_flags = 0;
1048 spin_unlock_irqrestore(&flow->lock, flags);
1050 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1052 return put_actions(flow, ufp);
1055 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1057 struct dp_table *table = rcu_dereference(dp->table);
1059 struct sw_flow *flow;
1063 if (copy_from_user(&uf, ufp, sizeof uf))
1065 uf.key.reserved = 0;
1067 flow = dp_table_lookup(table, &uf.key);
1072 /* XXX redundant lookup */
1073 error = dp_table_delete(table, flow);
1077 /* XXX These statistics might lose a few packets, since other CPUs can
1078 * be using this flow. We used to synchronize_rcu() to make sure that
1079 * we get completely accurate stats, but that blows our performance,
1082 error = answer_query(flow, 0, ufp);
1083 flow_deferred_free(flow);
1089 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1091 struct dp_table *table = rcu_dereference(dp->table);
1093 for (i = 0; i < flowvec->n_flows; i++) {
1094 struct __user odp_flow *ufp = &flowvec->flows[i];
1096 struct sw_flow *flow;
1099 if (__copy_from_user(&uf, ufp, sizeof uf))
1101 uf.key.reserved = 0;
1103 flow = dp_table_lookup(table, &uf.key);
1105 error = __put_user(ENOENT, &ufp->stats.error);
1107 error = answer_query(flow, uf.flags, ufp);
1111 return flowvec->n_flows;
1114 struct list_flows_cbdata {
1115 struct odp_flow __user *uflows;
1120 static int list_flow(struct sw_flow *flow, void *cbdata_)
1122 struct list_flows_cbdata *cbdata = cbdata_;
1123 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1126 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1128 error = answer_query(flow, 0, ufp);
1132 if (cbdata->listed_flows >= cbdata->n_flows)
1133 return cbdata->listed_flows;
1137 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1139 struct list_flows_cbdata cbdata;
1142 if (!flowvec->n_flows)
1145 cbdata.uflows = flowvec->flows;
1146 cbdata.n_flows = flowvec->n_flows;
1147 cbdata.listed_flows = 0;
1148 error = dp_table_foreach(rcu_dereference(dp->table),
1149 list_flow, &cbdata);
1150 return error ? error : cbdata.listed_flows;
1153 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1154 int (*function)(struct datapath *,
1155 const struct odp_flowvec *))
1157 struct odp_flowvec __user *uflowvec;
1158 struct odp_flowvec flowvec;
1161 uflowvec = (struct odp_flowvec __user *)argp;
1162 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1163 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1166 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1169 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1170 flowvec.n_flows * sizeof(struct odp_flow)))
1173 retval = function(dp, &flowvec);
1174 return (retval < 0 ? retval
1175 : retval == flowvec.n_flows ? 0
1176 : __put_user(retval, &uflowvec->n_flows));
1179 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1181 struct odp_execute execute;
1182 struct odp_flow_key key;
1183 struct sk_buff *skb;
1184 struct sw_flow_actions *actions;
1189 if (copy_from_user(&execute, executep, sizeof execute))
1193 if (execute.length < ETH_HLEN || execute.length > 65535)
1197 actions = flow_actions_alloc(execute.n_actions);
1202 if (copy_from_user(actions->actions, execute.actions,
1203 execute.n_actions * sizeof *execute.actions))
1204 goto error_free_actions;
1206 err = validate_actions(actions);
1208 goto error_free_actions;
1211 skb = alloc_skb(execute.length, GFP_KERNEL);
1213 goto error_free_actions;
1214 if (execute.in_port < DP_MAX_PORTS) {
1215 struct net_bridge_port *p = dp->ports[execute.in_port];
1221 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1223 goto error_free_skb;
1225 skb_reset_mac_header(skb);
1228 /* Normally, setting the skb 'protocol' field would be handled by a
1229 * call to eth_type_trans(), but it assumes there's a sending
1230 * device, which we may not have. */
1231 if (ntohs(eth->h_proto) >= 1536)
1232 skb->protocol = eth->h_proto;
1234 skb->protocol = htons(ETH_P_802_2);
1236 flow_extract(skb, execute.in_port, &key);
1237 err = execute_actions(dp, skb, &key, actions->actions,
1238 actions->n_actions, GFP_KERNEL);
1250 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1252 struct odp_stats stats;
1255 stats.n_flows = dp->n_flows;
1256 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets;
1257 stats.max_capacity = DP_MAX_BUCKETS;
1258 stats.n_ports = dp->n_ports;
1259 stats.max_ports = DP_MAX_PORTS;
1260 stats.max_groups = DP_MAX_GROUPS;
1261 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1262 for_each_possible_cpu(i) {
1263 const struct dp_stats_percpu *s;
1264 s = percpu_ptr(dp->stats_percpu, i);
1265 stats.n_frags += s->n_frags;
1266 stats.n_hit += s->n_hit;
1267 stats.n_missed += s->n_missed;
1268 stats.n_lost += s->n_lost;
1270 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1271 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1272 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1275 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1276 int dp_min_mtu(const struct datapath *dp)
1278 struct net_bridge_port *p;
1283 list_for_each_entry_rcu (p, &dp->port_list, node) {
1284 struct net_device *dev = p->dev;
1286 /* Skip any internal ports, since that's what we're trying to
1291 if (!mtu || dev->mtu < mtu)
1295 return mtu ? mtu : ETH_DATA_LEN;
1299 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1302 memset(&op, 0, sizeof op);
1303 strncpy(op.devname, p->dev->name, sizeof op.devname);
1304 op.port = p->port_no;
1305 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1306 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1310 query_port(struct datapath *dp, struct odp_port __user *uport)
1312 struct odp_port port;
1314 if (copy_from_user(&port, uport, sizeof port))
1316 if (port.devname[0]) {
1317 struct net_bridge_port *p;
1318 struct net_device *dev;
1321 port.devname[IFNAMSIZ - 1] = '\0';
1323 dev = dev_get_by_name(&init_net, port.devname);
1328 if (!p && is_dp_dev(dev)) {
1329 struct dp_dev *dp_dev = dp_dev_priv(dev);
1330 if (dp_dev->dp == dp)
1331 p = dp->ports[dp_dev->port_no];
1333 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1338 if (port.port >= DP_MAX_PORTS)
1340 if (!dp->ports[port.port])
1342 return put_port(dp->ports[port.port], uport);
1347 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1349 struct odp_portvec pv;
1350 struct net_bridge_port *p;
1353 if (copy_from_user(&pv, pvp, sizeof pv))
1358 list_for_each_entry_rcu (p, &dp->port_list, node) {
1359 if (put_port(p, &pv.ports[idx]))
1361 if (idx++ >= pv.n_ports)
1365 return put_user(dp->n_ports, &pvp->n_ports);
1368 /* RCU callback for freeing a dp_port_group */
1369 static void free_port_group(struct rcu_head *rcu)
1371 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1376 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1378 struct odp_port_group pg;
1379 struct dp_port_group *new_group, *old_group;
1383 if (copy_from_user(&pg, upg, sizeof pg))
1387 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1391 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1396 new_group->n_ports = pg.n_ports;
1398 if (copy_from_user(new_group->ports, pg.ports,
1399 sizeof(u16) * pg.n_ports))
1402 old_group = rcu_dereference(dp->groups[pg.group]);
1403 rcu_assign_pointer(dp->groups[pg.group], new_group);
1405 call_rcu(&old_group->rcu, free_port_group);
1415 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1417 struct odp_port_group pg;
1418 struct dp_port_group *g;
1421 if (copy_from_user(&pg, upg, sizeof pg))
1424 if (pg.group >= DP_MAX_GROUPS)
1427 g = dp->groups[pg.group];
1428 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1429 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1432 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1438 static int get_listen_mask(const struct file *f)
1440 return (long)f->private_data;
1443 static void set_listen_mask(struct file *f, int listen_mask)
1445 f->private_data = (void*)(long)listen_mask;
1448 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1451 int dp_idx = iminor(f->f_dentry->d_inode);
1452 struct datapath *dp;
1453 int drop_frags, listeners, port_no;
1456 /* Handle commands with special locking requirements up front. */
1459 err = create_dp(dp_idx, (char __user *)argp);
1462 case ODP_DP_DESTROY:
1463 err = destroy_dp(dp_idx);
1467 err = add_port(dp_idx, (struct odp_port __user *)argp);
1471 err = get_user(port_no, (int __user *)argp);
1473 err = del_port(dp_idx, port_no);
1477 dp = get_dp_locked(dp_idx);
1484 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1487 case ODP_GET_DROP_FRAGS:
1488 err = put_user(dp->drop_frags, (int __user *)argp);
1491 case ODP_SET_DROP_FRAGS:
1492 err = get_user(drop_frags, (int __user *)argp);
1496 if (drop_frags != 0 && drop_frags != 1)
1498 dp->drop_frags = drop_frags;
1502 case ODP_GET_LISTEN_MASK:
1503 err = put_user(get_listen_mask(f), (int __user *)argp);
1506 case ODP_SET_LISTEN_MASK:
1507 err = get_user(listeners, (int __user *)argp);
1511 if (listeners & ~ODPL_ALL)
1514 set_listen_mask(f, listeners);
1517 case ODP_PORT_QUERY:
1518 err = query_port(dp, (struct odp_port __user *)argp);
1522 err = list_ports(dp, (struct odp_portvec __user *)argp);
1525 case ODP_PORT_GROUP_SET:
1526 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1529 case ODP_PORT_GROUP_GET:
1530 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1533 case ODP_FLOW_FLUSH:
1534 err = flush_flows(dp);
1538 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1542 err = del_flow(dp, (struct odp_flow __user *)argp);
1546 err = do_flowvec_ioctl(dp, argp, query_flows);
1550 err = do_flowvec_ioctl(dp, argp, list_flows);
1554 err = do_execute(dp, (struct odp_execute __user *)argp);
1561 mutex_unlock(&dp->mutex);
1566 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1569 for (i = 0; i < DP_N_QUEUES; i++) {
1570 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1576 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1579 /* XXX is there sufficient synchronization here? */
1580 int listeners = get_listen_mask(f);
1581 int dp_idx = iminor(f->f_dentry->d_inode);
1582 struct datapath *dp = get_dp(dp_idx);
1583 struct sk_buff *skb;
1584 struct iovec __user iov;
1591 if (nbytes == 0 || !listeners)
1597 for (i = 0; i < DP_N_QUEUES; i++) {
1598 if (listeners & (1 << i)) {
1599 skb = skb_dequeue(&dp->queues[i]);
1605 if (f->f_flags & O_NONBLOCK) {
1610 wait_event_interruptible(dp->waitqueue,
1611 dp_has_packet_of_interest(dp,
1614 if (signal_pending(current)) {
1615 retval = -ERESTARTSYS;
1620 copy_bytes = min_t(size_t, skb->len, nbytes);
1622 iov.iov_len = copy_bytes;
1623 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1625 retval = copy_bytes;
1632 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1634 /* XXX is there sufficient synchronization here? */
1635 int dp_idx = iminor(file->f_dentry->d_inode);
1636 struct datapath *dp = get_dp(dp_idx);
1641 poll_wait(file, &dp->waitqueue, wait);
1642 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1643 mask |= POLLIN | POLLRDNORM;
1645 mask = POLLIN | POLLRDNORM | POLLHUP;
1650 struct file_operations openvswitch_fops = {
1651 /* XXX .aio_read = openvswitch_aio_read, */
1652 .read = openvswitch_read,
1653 .poll = openvswitch_poll,
1654 .unlocked_ioctl = openvswitch_ioctl,
1655 /* XXX .fasync = openvswitch_fasync, */
1660 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)
1661 static struct llc_sap *dp_stp_sap;
1663 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1664 struct packet_type *pt, struct net_device *orig_dev)
1666 /* We don't really care about STP packets, we just listen for them for
1667 * mutual exclusion with the bridge module, so this just discards
1673 static int dp_avoid_bridge_init(void)
1675 /* Register to receive STP packets because the bridge module also
1676 * attempts to do so. Since there can only be a single listener for a
1677 * given protocol, this provides mutual exclusion against the bridge
1678 * module, preventing both of them from being loaded at the same
1680 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1682 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1688 static void dp_avoid_bridge_exit(void)
1690 llc_sap_put(dp_stp_sap);
1692 #else /* Linux 2.6.27 or later. */
1693 static int dp_avoid_bridge_init(void)
1695 /* Linux 2.6.27 introduces a way for multiple clients to register for
1696 * STP packets, which interferes with what we try to do above.
1697 * Instead, just check whether there's a bridge hook defined. This is
1698 * not as safe--the bridge module is willing to load over the top of
1699 * us--but it provides a little bit of protection. */
1700 if (br_handle_frame_hook) {
1701 printk(KERN_ERR "openvswitch: bridge module is loaded, cannot load over it\n");
1707 static void dp_avoid_bridge_exit(void)
1709 /* Nothing to do. */
1711 #endif /* Linux 2.6.27 or later */
1713 static int __init dp_init(void)
1717 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1719 err = dp_avoid_bridge_init();
1727 err = register_netdevice_notifier(&dp_device_notifier);
1729 goto error_flow_exit;
1731 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1733 goto error_unreg_notifier;
1735 /* Hook into callback used by the bridge to intercept packets.
1736 * Parasites we are. */
1737 br_handle_frame_hook = dp_frame_hook;
1741 error_unreg_notifier:
1742 unregister_netdevice_notifier(&dp_device_notifier);
1749 static void dp_cleanup(void)
1752 unregister_chrdev(major, "openvswitch");
1753 unregister_netdevice_notifier(&dp_device_notifier);
1755 br_handle_frame_hook = NULL;
1756 dp_avoid_bridge_exit();
1759 module_init(dp_init);
1760 module_exit(dp_cleanup);
1762 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1763 MODULE_LICENSE("GPL");