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);
74 static void compute_ip_summed(struct sk_buff *skb);
76 /* Must be called with rcu_read_lock or dp_mutex. */
77 struct datapath *get_dp(int dp_idx)
79 if (dp_idx < 0 || dp_idx >= ODP_MAX)
81 return rcu_dereference(dps[dp_idx]);
83 EXPORT_SYMBOL_GPL(get_dp);
85 static struct datapath *get_dp_locked(int dp_idx)
89 mutex_lock(&dp_mutex);
92 mutex_lock(&dp->mutex);
93 mutex_unlock(&dp_mutex);
97 static inline size_t br_nlmsg_size(void)
99 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
100 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
101 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
102 + nla_total_size(4) /* IFLA_MASTER */
103 + nla_total_size(4) /* IFLA_MTU */
104 + nla_total_size(4) /* IFLA_LINK */
105 + nla_total_size(1); /* IFLA_OPERSTATE */
108 static int dp_fill_ifinfo(struct sk_buff *skb,
109 const struct net_bridge_port *port,
110 int event, unsigned int flags)
112 const struct datapath *dp = port->dp;
113 const struct net_device *dev = port->dev;
114 struct ifinfomsg *hdr;
115 struct nlmsghdr *nlh;
117 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
121 hdr = nlmsg_data(nlh);
122 hdr->ifi_family = AF_BRIDGE;
124 hdr->ifi_type = dev->type;
125 hdr->ifi_index = dev->ifindex;
126 hdr->ifi_flags = dev_get_flags(dev);
129 NLA_PUT_STRING(skb, IFLA_IFNAME, dev->name);
130 NLA_PUT_U32(skb, IFLA_MASTER, dp->ports[ODPP_LOCAL]->dev->ifindex);
131 NLA_PUT_U32(skb, IFLA_MTU, dev->mtu);
132 #ifdef IFLA_OPERSTATE
133 NLA_PUT_U8(skb, IFLA_OPERSTATE,
134 netif_running(dev) ? dev->operstate : IF_OPER_DOWN);
138 NLA_PUT(skb, IFLA_ADDRESS, dev->addr_len, dev->dev_addr);
140 if (dev->ifindex != dev->iflink)
141 NLA_PUT_U32(skb, IFLA_LINK, dev->iflink);
143 return nlmsg_end(skb, nlh);
146 nlmsg_cancel(skb, nlh);
150 static void dp_ifinfo_notify(int event, struct net_bridge_port *port)
152 struct net *net = dev_net(port->dev);
156 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
160 err = dp_fill_ifinfo(skb, port, event, 0);
162 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
163 WARN_ON(err == -EMSGSIZE);
167 rtnl_notify(skb, net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
171 rtnl_set_sk_err(net, RTNLGRP_LINK, err);
174 static void release_dp(struct kobject *kobj)
176 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
180 static struct kobj_type dp_ktype = {
181 .release = release_dp
184 static int create_dp(int dp_idx, const char __user *devnamep)
186 struct net_device *dp_dev;
187 char devname[IFNAMSIZ];
194 if (strncpy_from_user(devname, devnamep, IFNAMSIZ - 1) < 0)
196 devname[IFNAMSIZ - 1] = '\0';
198 snprintf(devname, sizeof devname, "of%d", dp_idx);
202 mutex_lock(&dp_mutex);
204 if (!try_module_get(THIS_MODULE))
207 /* Exit early if a datapath with that number already exists.
208 * (We don't use -EEXIST because that's ambiguous with 'devname'
209 * conflicting with an existing network device name.) */
215 dp = kzalloc(sizeof *dp, GFP_KERNEL);
218 INIT_LIST_HEAD(&dp->port_list);
219 mutex_init(&dp->mutex);
221 for (i = 0; i < DP_N_QUEUES; i++)
222 skb_queue_head_init(&dp->queues[i]);
223 init_waitqueue_head(&dp->waitqueue);
225 /* Initialize kobject for bridge. This will be added as
226 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
227 dp->ifobj.kset = NULL;
228 kobject_init(&dp->ifobj, &dp_ktype);
230 /* Allocate table. */
232 rcu_assign_pointer(dp->table, dp_table_create(DP_L1_SIZE));
236 /* Set up our datapath device. */
237 dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL);
238 err = PTR_ERR(dp_dev);
240 goto err_destroy_table;
242 err = new_nbp(dp, dp_dev, ODPP_LOCAL);
244 dp_dev_destroy(dp_dev);
245 goto err_destroy_table;
249 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
250 if (!dp->stats_percpu)
251 goto err_destroy_local_port;
253 rcu_assign_pointer(dps[dp_idx], dp);
254 mutex_unlock(&dp_mutex);
261 err_destroy_local_port:
262 dp_del_port(dp->ports[ODPP_LOCAL]);
264 dp_table_destroy(dp->table, 0);
268 module_put(THIS_MODULE);
270 mutex_unlock(&dp_mutex);
276 static void do_destroy_dp(struct datapath *dp)
278 struct net_bridge_port *p, *n;
281 list_for_each_entry_safe (p, n, &dp->port_list, node)
282 if (p->port_no != ODPP_LOCAL)
287 rcu_assign_pointer(dps[dp->dp_idx], NULL);
289 dp_del_port(dp->ports[ODPP_LOCAL]);
291 dp_table_destroy(dp->table, 1);
293 for (i = 0; i < DP_N_QUEUES; i++)
294 skb_queue_purge(&dp->queues[i]);
295 for (i = 0; i < DP_MAX_GROUPS; i++)
296 kfree(dp->groups[i]);
297 free_percpu(dp->stats_percpu);
298 kobject_put(&dp->ifobj);
299 module_put(THIS_MODULE);
302 static int destroy_dp(int dp_idx)
308 mutex_lock(&dp_mutex);
318 mutex_unlock(&dp_mutex);
323 static void release_nbp(struct kobject *kobj)
325 struct net_bridge_port *p = container_of(kobj, struct net_bridge_port, kobj);
329 static struct kobj_type brport_ktype = {
331 .sysfs_ops = &brport_sysfs_ops,
333 .release = release_nbp
336 /* Called with RTNL lock and dp_mutex. */
337 static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no)
339 struct net_bridge_port *p;
341 if (dev->br_port != NULL)
344 p = kzalloc(sizeof(*p), GFP_KERNEL);
348 dev_set_promiscuity(dev, 1);
350 p->port_no = port_no;
353 atomic_set(&p->sflow_pool, 0);
355 rcu_assign_pointer(dev->br_port, p);
357 /* It would make sense to assign dev->br_port here too, but
358 * that causes packets received on internal ports to get caught
359 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
360 * back to network stack, but that's a waste of time. */
362 rcu_assign_pointer(dp->ports[port_no], p);
363 list_add_rcu(&p->node, &dp->port_list);
366 /* Initialize kobject for bridge. This will be added as
367 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
369 kobject_init(&p->kobj, &brport_ktype);
371 dp_ifinfo_notify(RTM_NEWLINK, p);
376 static int add_port(int dp_idx, struct odp_port __user *portp)
378 struct net_device *dev;
380 struct odp_port port;
385 if (copy_from_user(&port, portp, sizeof port))
387 port.devname[IFNAMSIZ - 1] = '\0';
390 dp = get_dp_locked(dp_idx);
393 goto out_unlock_rtnl;
395 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
396 if (!dp->ports[port_no])
402 if (!(port.flags & ODP_PORT_INTERNAL)) {
404 dev = dev_get_by_name(&init_net, port.devname);
409 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
413 dev = dp_dev_create(dp, port.devname, port_no);
420 err = new_nbp(dp, dev, port_no);
424 set_dp_devs_mtu(dp, dev);
425 dp_sysfs_add_if(dp->ports[port_no]);
427 err = __put_user(port_no, &portp->port);
432 mutex_unlock(&dp->mutex);
439 int dp_del_port(struct net_bridge_port *p)
443 if (p->port_no != ODPP_LOCAL)
445 dp_ifinfo_notify(RTM_DELLINK, p);
449 if (is_dp_dev(p->dev)) {
450 /* Make sure that no packets arrive from now on, since
451 * dp_dev_xmit() will try to find itself through
452 * p->dp->ports[], and we're about to set that to null. */
453 netif_tx_disable(p->dev);
456 /* First drop references to device. */
457 dev_set_promiscuity(p->dev, -1);
458 list_del_rcu(&p->node);
459 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
460 rcu_assign_pointer(p->dev->br_port, NULL);
462 /* Then wait until no one is still using it, and destroy it. */
465 if (is_dp_dev(p->dev))
466 dp_dev_destroy(p->dev);
468 kobject_put(&p->kobj);
473 static int del_port(int dp_idx, int port_no)
475 struct net_bridge_port *p;
481 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
485 dp = get_dp_locked(dp_idx);
488 goto out_unlock_rtnl;
490 p = dp->ports[port_no];
495 err = dp_del_port(p);
498 mutex_unlock(&dp->mutex);
505 /* Must be called with rcu_read_lock. */
507 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
509 /* Make our own copy of the packet. Otherwise we will mangle the
510 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
511 * (No one comes after us, since we tell handle_bridge() that we took
513 skb = skb_share_check(skb, GFP_ATOMIC);
517 /* Push the Ethernet header back on. */
518 skb_push(skb, ETH_HLEN);
519 skb_reset_mac_header(skb);
520 dp_process_received_packet(skb, p);
523 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
524 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
526 struct datapath *dp = p->dp;
527 struct dp_stats_percpu *stats;
528 struct odp_flow_key key;
529 struct sw_flow *flow;
531 WARN_ON_ONCE(skb_shared(skb));
533 compute_ip_summed(skb);
535 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
536 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
538 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
539 if (dp->drop_frags) {
546 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
548 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
549 flow_used(flow, skb);
550 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
555 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
560 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
561 * different set of devices!)
563 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
564 /* Called with rcu_read_lock and bottom-halves disabled. */
565 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
568 do_port_input(p, skb);
571 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
572 /* Called with rcu_read_lock and bottom-halves disabled. */
573 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
575 do_port_input(p, *pskb);
582 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
583 /* This code is based on a skb_checksum_setup from net/dev/core.c from a
584 * combination of Lenny's 2.6.26 Xen kernel and Xen's
585 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call this function
586 * directly because it isn't exported in all versions. */
587 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
589 if (ptr < (void *)skb->tail)
591 if (__pskb_pull_tail(skb,
592 ptr - (void *)skb->data - skb_headlen(skb))) {
599 int vswitch_skb_checksum_setup(struct sk_buff *skb)
604 __u16 csum_start, csum_offset;
606 if (!skb->proto_csum_blank)
609 if (skb->protocol != htons(ETH_P_IP))
612 if (!skb_pull_up_to(skb, skb_network_header(skb) + sizeof(struct iphdr)))
616 th = skb_network_header(skb) + 4 * iph->ihl;
618 csum_start = th - skb->head;
619 switch (iph->protocol) {
621 csum_offset = offsetof(struct tcphdr, check);
624 csum_offset = offsetof(struct udphdr, check);
628 printk(KERN_ERR "Attempting to checksum a non-"
629 "TCP/UDP packet, dropping a protocol"
630 " %d packet", iph->protocol);
634 if (!skb_pull_up_to(skb, th + csum_offset + 2))
637 skb->ip_summed = CHECKSUM_PARTIAL;
638 skb->proto_csum_blank = 0;
640 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
641 skb->csum_start = csum_start;
642 skb->csum_offset = csum_offset;
644 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
645 skb->csum = csum_offset;
653 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
655 /* Types of checksums that we can receive (these all refer to L4 checksums):
656 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
657 * (though not verified) checksum in packet but not in skb->csum. Packets
658 * from the bridge local port will also have this type.
659 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
660 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
661 * a valid skb->csum. Importantly, both contain a full checksum (not
662 * verified) in the packet itself. The only difference is that if the
663 * packet gets to L4 processing on this machine (not in DomU) we won't
664 * have to recompute the checksum to verify. Most hardware devices do not
665 * produce packets with this type, even if they support receive checksum
666 * offloading (they produce type #5).
667 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
668 * be computed if it is sent off box. Unfortunately on earlier kernels,
669 * this case is impossible to distinguish from #2, despite having opposite
670 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
671 * to distinguish the different states. The only real user of this type
672 * with bridging is Xen (on later kernels).
673 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
674 * generated locally by a Xen DomU and has a partial checksum. If it is
675 * handled on this machine (Dom0 or DomU), then the checksum will not be
676 * computed. If it goes off box, the checksum in the packet needs to be
677 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
678 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
679 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
680 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
681 * full checksum or using a protocol without a checksum. skb->csum is
682 * undefined. This is common from devices with receive checksum
683 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
684 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
686 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
687 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
688 * based on whether it is on the transmit or receive path. After the datapath
689 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
690 * checksum, we will panic. Since we can receive packets with checksums, we
691 * assume that all CHECKSUM_HW packets have checksums and map them to
692 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
693 * packet is processed by the local IP stack, in which case it will need to
694 * be reverified). If we receive a packet with CHECKSUM_HW that really means
695 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
696 * shouldn't be any devices that do this with bridging.
698 * The bridge has similar behavior and this function closely resembles
699 * skb_forward_csum(). It is slightly different because we are only concerned
700 * with bridging and not other types of forwarding and can get away with
701 * slightly more optimal behavior.*/
703 compute_ip_summed(struct sk_buff *skb)
705 OVS_CB(skb)->ip_summed = skb->ip_summed;
708 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
709 * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
710 * uses some special fields to represent this (see below). Since we
711 * can only make one type work, pick the one that actually happens in
713 if (skb->ip_summed == CHECKSUM_HW)
714 OVS_CB(skb)->ip_summed = CSUM_COMPLETE;
716 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
717 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
719 if (skb->proto_csum_blank)
720 OVS_CB(skb)->ip_summed = CSUM_PARTIAL;
725 forward_ip_summed(struct sk_buff *skb)
728 if (OVS_CB(skb)->ip_summed == CSUM_COMPLETE)
729 skb->ip_summed = CHECKSUM_NONE;
733 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
734 * unless we broke up a GSO packet. */
736 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
737 int queue_no, u32 arg)
739 struct sk_buff *nskb;
743 port_no = ODPP_LOCAL;
745 if (skb->dev->br_port)
746 port_no = skb->dev->br_port->port_no;
747 else if (is_dp_dev(skb->dev))
748 port_no = dp_dev_priv(skb->dev)->port_no;
752 struct odp_msg *header;
757 /* If a checksum-deferred packet is forwarded to the
758 * controller, correct the pointers and checksum. This happens
759 * on a regular basis only on Xen, on which VMs can pass up
760 * packets that do not have their checksum computed.
762 err = vswitch_skb_checksum_setup(skb);
766 if (skb->ip_summed == CHECKSUM_PARTIAL) {
767 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
768 /* Until 2.6.22, the start of the transport header was
769 * also the start of data to be checksummed. Linux
770 * 2.6.22 introduced the csum_start field for this
771 * purpose, but we should point the transport header to
772 * it anyway for backward compatibility, as
773 * dev_queue_xmit() does even in 2.6.28. */
774 skb_set_transport_header(skb, skb->csum_start -
777 err = skb_checksum_help(skb);
782 if (skb->ip_summed == CHECKSUM_HW) {
783 err = skb_checksum_help(skb, 0);
789 err = skb_cow(skb, sizeof *header);
793 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
794 header->type = queue_no;
795 header->length = skb->len;
796 header->port = port_no;
797 header->reserved = 0;
799 skb_queue_tail(queue, skb);
807 while ((skb = nskb) != NULL) {
815 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
818 struct dp_stats_percpu *stats;
819 struct sk_buff_head *queue;
822 WARN_ON_ONCE(skb_shared(skb));
823 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
824 queue = &dp->queues[queue_no];
826 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
829 forward_ip_summed(skb);
831 /* Break apart GSO packets into their component pieces. Otherwise
832 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
833 if (skb_is_gso(skb)) {
834 struct sk_buff *nskb = skb_gso_segment(skb, 0);
838 if (unlikely(IS_ERR(skb))) {
843 /* XXX This case might not be possible. It's hard to
844 * tell from the skb_gso_segment() code and comment. */
848 err = queue_control_packets(skb, queue, queue_no, arg);
849 wake_up_interruptible(&dp->waitqueue);
855 stats = percpu_ptr(dp->stats_percpu, get_cpu());
862 static int flush_flows(struct datapath *dp)
865 return dp_table_flush(dp);
868 static int validate_actions(const struct sw_flow_actions *actions)
872 for (i = 0; i < actions->n_actions; i++) {
873 const union odp_action *a = &actions->actions[i];
876 if (a->output.port >= DP_MAX_PORTS)
880 case ODPAT_OUTPUT_GROUP:
881 if (a->output_group.group >= DP_MAX_GROUPS)
885 case ODPAT_SET_VLAN_VID:
886 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
890 case ODPAT_SET_VLAN_PCP:
891 if (a->vlan_pcp.vlan_pcp
892 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
897 if (a->type >= ODPAT_N_ACTIONS)
906 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
908 struct sw_flow_actions *actions;
911 actions = flow_actions_alloc(flow->n_actions);
912 error = PTR_ERR(actions);
917 if (copy_from_user(actions->actions, flow->actions,
918 flow->n_actions * sizeof(union odp_action)))
919 goto error_free_actions;
920 error = validate_actions(actions);
922 goto error_free_actions;
929 return ERR_PTR(error);
932 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
934 if (flow->used.tv_sec) {
935 stats->used_sec = flow->used.tv_sec;
936 stats->used_nsec = flow->used.tv_nsec;
939 stats->used_nsec = 0;
941 stats->n_packets = flow->packet_count;
942 stats->n_bytes = flow->byte_count;
943 stats->ip_tos = flow->ip_tos;
944 stats->tcp_flags = flow->tcp_flags;
948 static void clear_stats(struct sw_flow *flow)
950 flow->used.tv_sec = flow->used.tv_nsec = 0;
953 flow->packet_count = 0;
954 flow->byte_count = 0;
957 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
959 struct odp_flow_put uf;
960 struct sw_flow *flow;
961 struct dp_table *table;
962 struct odp_flow_stats stats;
966 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
968 memset(uf.flow.key.reserved, 0, sizeof uf.flow.key.reserved);
970 table = rcu_dereference(dp->table);
971 flow = dp_table_lookup(table, &uf.flow.key);
974 struct sw_flow_actions *acts;
977 if (!(uf.flags & ODPPF_CREATE))
980 /* Expand table, if necessary, to make room. */
981 if (dp->n_flows >= table->n_buckets) {
983 if (table->n_buckets >= DP_MAX_BUCKETS)
986 error = dp_table_expand(dp);
989 table = rcu_dereference(dp->table);
994 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
997 flow->key = uf.flow.key;
998 spin_lock_init(&flow->lock);
1001 /* Obtain actions. */
1002 acts = get_actions(&uf.flow);
1003 error = PTR_ERR(acts);
1005 goto error_free_flow;
1006 rcu_assign_pointer(flow->sf_acts, acts);
1008 /* Put flow in bucket. */
1009 error = dp_table_insert(table, flow);
1011 goto error_free_flow_acts;
1013 memset(&stats, 0, sizeof(struct odp_flow_stats));
1015 /* We found a matching flow. */
1016 struct sw_flow_actions *old_acts, *new_acts;
1017 unsigned long int flags;
1019 /* Bail out if we're not allowed to modify an existing flow. */
1021 if (!(uf.flags & ODPPF_MODIFY))
1025 new_acts = get_actions(&uf.flow);
1026 error = PTR_ERR(new_acts);
1027 if (IS_ERR(new_acts))
1029 old_acts = rcu_dereference(flow->sf_acts);
1030 if (old_acts->n_actions != new_acts->n_actions ||
1031 memcmp(old_acts->actions, new_acts->actions,
1032 sizeof(union odp_action) * old_acts->n_actions)) {
1033 rcu_assign_pointer(flow->sf_acts, new_acts);
1034 flow_deferred_free_acts(old_acts);
1039 /* Fetch stats, then clear them if necessary. */
1040 spin_lock_irqsave(&flow->lock, flags);
1041 get_stats(flow, &stats);
1042 if (uf.flags & ODPPF_ZERO_STATS)
1044 spin_unlock_irqrestore(&flow->lock, flags);
1047 /* Copy stats to userspace. */
1048 if (__copy_to_user(&ufp->flow.stats, &stats,
1049 sizeof(struct odp_flow_stats)))
1053 error_free_flow_acts:
1054 kfree(flow->sf_acts);
1056 kmem_cache_free(flow_cache, flow);
1061 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1063 union odp_action __user *actions;
1064 struct sw_flow_actions *sf_acts;
1067 if (__get_user(actions, &ufp->actions) ||
1068 __get_user(n_actions, &ufp->n_actions))
1074 sf_acts = rcu_dereference(flow->sf_acts);
1075 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1076 (actions && copy_to_user(actions, sf_acts->actions,
1077 sizeof(union odp_action) *
1078 min(sf_acts->n_actions, n_actions))))
1084 static int answer_query(struct sw_flow *flow, u32 query_flags,
1085 struct odp_flow __user *ufp)
1087 struct odp_flow_stats stats;
1088 unsigned long int flags;
1090 spin_lock_irqsave(&flow->lock, flags);
1091 get_stats(flow, &stats);
1093 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1094 flow->tcp_flags = 0;
1096 spin_unlock_irqrestore(&flow->lock, flags);
1098 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1100 return put_actions(flow, ufp);
1103 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1105 struct dp_table *table = rcu_dereference(dp->table);
1107 struct sw_flow *flow;
1111 if (copy_from_user(&uf, ufp, sizeof uf))
1113 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1115 flow = dp_table_lookup(table, &uf.key);
1120 /* XXX redundant lookup */
1121 error = dp_table_delete(table, flow);
1125 /* XXX These statistics might lose a few packets, since other CPUs can
1126 * be using this flow. We used to synchronize_rcu() to make sure that
1127 * we get completely accurate stats, but that blows our performance,
1130 error = answer_query(flow, 0, ufp);
1131 flow_deferred_free(flow);
1137 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1139 struct dp_table *table = rcu_dereference(dp->table);
1141 for (i = 0; i < flowvec->n_flows; i++) {
1142 struct __user odp_flow *ufp = &flowvec->flows[i];
1144 struct sw_flow *flow;
1147 if (__copy_from_user(&uf, ufp, sizeof uf))
1149 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1151 flow = dp_table_lookup(table, &uf.key);
1153 error = __put_user(ENOENT, &ufp->stats.error);
1155 error = answer_query(flow, uf.flags, ufp);
1159 return flowvec->n_flows;
1162 struct list_flows_cbdata {
1163 struct odp_flow __user *uflows;
1168 static int list_flow(struct sw_flow *flow, void *cbdata_)
1170 struct list_flows_cbdata *cbdata = cbdata_;
1171 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1174 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1176 error = answer_query(flow, 0, ufp);
1180 if (cbdata->listed_flows >= cbdata->n_flows)
1181 return cbdata->listed_flows;
1185 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1187 struct list_flows_cbdata cbdata;
1190 if (!flowvec->n_flows)
1193 cbdata.uflows = flowvec->flows;
1194 cbdata.n_flows = flowvec->n_flows;
1195 cbdata.listed_flows = 0;
1196 error = dp_table_foreach(rcu_dereference(dp->table),
1197 list_flow, &cbdata);
1198 return error ? error : cbdata.listed_flows;
1201 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1202 int (*function)(struct datapath *,
1203 const struct odp_flowvec *))
1205 struct odp_flowvec __user *uflowvec;
1206 struct odp_flowvec flowvec;
1209 uflowvec = (struct odp_flowvec __user *)argp;
1210 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1211 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1214 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1217 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1218 flowvec.n_flows * sizeof(struct odp_flow)))
1221 retval = function(dp, &flowvec);
1222 return (retval < 0 ? retval
1223 : retval == flowvec.n_flows ? 0
1224 : __put_user(retval, &uflowvec->n_flows));
1227 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1229 struct odp_execute execute;
1230 struct odp_flow_key key;
1231 struct sk_buff *skb;
1232 struct sw_flow_actions *actions;
1237 if (copy_from_user(&execute, executep, sizeof execute))
1241 if (execute.length < ETH_HLEN || execute.length > 65535)
1245 actions = flow_actions_alloc(execute.n_actions);
1250 if (copy_from_user(actions->actions, execute.actions,
1251 execute.n_actions * sizeof *execute.actions))
1252 goto error_free_actions;
1254 err = validate_actions(actions);
1256 goto error_free_actions;
1259 skb = alloc_skb(execute.length, GFP_KERNEL);
1261 goto error_free_actions;
1262 if (execute.in_port < DP_MAX_PORTS) {
1263 struct net_bridge_port *p = dp->ports[execute.in_port];
1269 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1271 goto error_free_skb;
1273 skb_reset_mac_header(skb);
1276 /* Normally, setting the skb 'protocol' field would be handled by a
1277 * call to eth_type_trans(), but it assumes there's a sending
1278 * device, which we may not have. */
1279 if (ntohs(eth->h_proto) >= 1536)
1280 skb->protocol = eth->h_proto;
1282 skb->protocol = htons(ETH_P_802_2);
1284 flow_extract(skb, execute.in_port, &key);
1285 err = execute_actions(dp, skb, &key, actions->actions,
1286 actions->n_actions, GFP_KERNEL);
1298 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1300 struct odp_stats stats;
1303 stats.n_flows = dp->n_flows;
1304 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets;
1305 stats.max_capacity = DP_MAX_BUCKETS;
1306 stats.n_ports = dp->n_ports;
1307 stats.max_ports = DP_MAX_PORTS;
1308 stats.max_groups = DP_MAX_GROUPS;
1309 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1310 for_each_possible_cpu(i) {
1311 const struct dp_stats_percpu *s;
1312 s = percpu_ptr(dp->stats_percpu, i);
1313 stats.n_frags += s->n_frags;
1314 stats.n_hit += s->n_hit;
1315 stats.n_missed += s->n_missed;
1316 stats.n_lost += s->n_lost;
1318 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1319 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1320 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1323 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1324 int dp_min_mtu(const struct datapath *dp)
1326 struct net_bridge_port *p;
1331 list_for_each_entry_rcu (p, &dp->port_list, node) {
1332 struct net_device *dev = p->dev;
1334 /* Skip any internal ports, since that's what we're trying to
1339 if (!mtu || dev->mtu < mtu)
1343 return mtu ? mtu : ETH_DATA_LEN;
1346 /* Sets the MTU of all datapath devices to the minimum of the ports. 'dev'
1347 * is the device whose MTU may have changed. Must be called with RTNL lock
1349 void set_dp_devs_mtu(const struct datapath *dp, struct net_device *dev)
1351 struct net_bridge_port *p;
1359 mtu = dp_min_mtu(dp);
1361 list_for_each_entry_rcu (p, &dp->port_list, node) {
1362 struct net_device *br_dev = p->dev;
1364 if (is_dp_dev(br_dev))
1365 dev_set_mtu(br_dev, mtu);
1370 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1373 memset(&op, 0, sizeof op);
1374 strncpy(op.devname, p->dev->name, sizeof op.devname);
1375 op.port = p->port_no;
1376 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1377 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1381 query_port(struct datapath *dp, struct odp_port __user *uport)
1383 struct odp_port port;
1385 if (copy_from_user(&port, uport, sizeof port))
1387 if (port.devname[0]) {
1388 struct net_bridge_port *p;
1389 struct net_device *dev;
1392 port.devname[IFNAMSIZ - 1] = '\0';
1394 dev = dev_get_by_name(&init_net, port.devname);
1399 if (!p && is_dp_dev(dev)) {
1400 struct dp_dev *dp_dev = dp_dev_priv(dev);
1401 if (dp_dev->dp == dp)
1402 p = dp->ports[dp_dev->port_no];
1404 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1409 if (port.port >= DP_MAX_PORTS)
1411 if (!dp->ports[port.port])
1413 return put_port(dp->ports[port.port], uport);
1418 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1420 struct odp_portvec pv;
1421 struct net_bridge_port *p;
1424 if (copy_from_user(&pv, pvp, sizeof pv))
1429 list_for_each_entry_rcu (p, &dp->port_list, node) {
1430 if (put_port(p, &pv.ports[idx]))
1432 if (idx++ >= pv.n_ports)
1436 return put_user(dp->n_ports, &pvp->n_ports);
1439 /* RCU callback for freeing a dp_port_group */
1440 static void free_port_group(struct rcu_head *rcu)
1442 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1447 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1449 struct odp_port_group pg;
1450 struct dp_port_group *new_group, *old_group;
1454 if (copy_from_user(&pg, upg, sizeof pg))
1458 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1462 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1467 new_group->n_ports = pg.n_ports;
1469 if (copy_from_user(new_group->ports, pg.ports,
1470 sizeof(u16) * pg.n_ports))
1473 old_group = rcu_dereference(dp->groups[pg.group]);
1474 rcu_assign_pointer(dp->groups[pg.group], new_group);
1476 call_rcu(&old_group->rcu, free_port_group);
1486 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1488 struct odp_port_group pg;
1489 struct dp_port_group *g;
1492 if (copy_from_user(&pg, upg, sizeof pg))
1495 if (pg.group >= DP_MAX_GROUPS)
1498 g = dp->groups[pg.group];
1499 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1500 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1503 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1509 static int get_listen_mask(const struct file *f)
1511 return (long)f->private_data;
1514 static void set_listen_mask(struct file *f, int listen_mask)
1516 f->private_data = (void*)(long)listen_mask;
1519 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1522 int dp_idx = iminor(f->f_dentry->d_inode);
1523 struct datapath *dp;
1524 int drop_frags, listeners, port_no;
1525 unsigned int sflow_probability;
1528 /* Handle commands with special locking requirements up front. */
1531 err = create_dp(dp_idx, (char __user *)argp);
1534 case ODP_DP_DESTROY:
1535 err = destroy_dp(dp_idx);
1539 err = add_port(dp_idx, (struct odp_port __user *)argp);
1543 err = get_user(port_no, (int __user *)argp);
1545 err = del_port(dp_idx, port_no);
1549 dp = get_dp_locked(dp_idx);
1556 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1559 case ODP_GET_DROP_FRAGS:
1560 err = put_user(dp->drop_frags, (int __user *)argp);
1563 case ODP_SET_DROP_FRAGS:
1564 err = get_user(drop_frags, (int __user *)argp);
1568 if (drop_frags != 0 && drop_frags != 1)
1570 dp->drop_frags = drop_frags;
1574 case ODP_GET_LISTEN_MASK:
1575 err = put_user(get_listen_mask(f), (int __user *)argp);
1578 case ODP_SET_LISTEN_MASK:
1579 err = get_user(listeners, (int __user *)argp);
1583 if (listeners & ~ODPL_ALL)
1586 set_listen_mask(f, listeners);
1589 case ODP_GET_SFLOW_PROBABILITY:
1590 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1593 case ODP_SET_SFLOW_PROBABILITY:
1594 err = get_user(sflow_probability, (unsigned int __user *)argp);
1596 dp->sflow_probability = sflow_probability;
1599 case ODP_PORT_QUERY:
1600 err = query_port(dp, (struct odp_port __user *)argp);
1604 err = list_ports(dp, (struct odp_portvec __user *)argp);
1607 case ODP_PORT_GROUP_SET:
1608 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1611 case ODP_PORT_GROUP_GET:
1612 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1615 case ODP_FLOW_FLUSH:
1616 err = flush_flows(dp);
1620 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1624 err = del_flow(dp, (struct odp_flow __user *)argp);
1628 err = do_flowvec_ioctl(dp, argp, query_flows);
1632 err = do_flowvec_ioctl(dp, argp, list_flows);
1636 err = do_execute(dp, (struct odp_execute __user *)argp);
1643 mutex_unlock(&dp->mutex);
1648 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1651 for (i = 0; i < DP_N_QUEUES; i++) {
1652 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1658 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1661 /* XXX is there sufficient synchronization here? */
1662 int listeners = get_listen_mask(f);
1663 int dp_idx = iminor(f->f_dentry->d_inode);
1664 struct datapath *dp = get_dp(dp_idx);
1665 struct sk_buff *skb;
1666 struct iovec __user iov;
1673 if (nbytes == 0 || !listeners)
1679 for (i = 0; i < DP_N_QUEUES; i++) {
1680 if (listeners & (1 << i)) {
1681 skb = skb_dequeue(&dp->queues[i]);
1687 if (f->f_flags & O_NONBLOCK) {
1692 wait_event_interruptible(dp->waitqueue,
1693 dp_has_packet_of_interest(dp,
1696 if (signal_pending(current)) {
1697 retval = -ERESTARTSYS;
1702 copy_bytes = min_t(size_t, skb->len, nbytes);
1704 iov.iov_len = copy_bytes;
1705 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1707 retval = copy_bytes;
1714 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1716 /* XXX is there sufficient synchronization here? */
1717 int dp_idx = iminor(file->f_dentry->d_inode);
1718 struct datapath *dp = get_dp(dp_idx);
1723 poll_wait(file, &dp->waitqueue, wait);
1724 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1725 mask |= POLLIN | POLLRDNORM;
1727 mask = POLLIN | POLLRDNORM | POLLHUP;
1732 struct file_operations openvswitch_fops = {
1733 /* XXX .aio_read = openvswitch_aio_read, */
1734 .read = openvswitch_read,
1735 .poll = openvswitch_poll,
1736 .unlocked_ioctl = openvswitch_ioctl,
1737 /* XXX .fasync = openvswitch_fasync, */
1742 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)
1743 static struct llc_sap *dp_stp_sap;
1745 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1746 struct packet_type *pt, struct net_device *orig_dev)
1748 /* We don't really care about STP packets, we just listen for them for
1749 * mutual exclusion with the bridge module, so this just discards
1755 static int dp_avoid_bridge_init(void)
1757 /* Register to receive STP packets because the bridge module also
1758 * attempts to do so. Since there can only be a single listener for a
1759 * given protocol, this provides mutual exclusion against the bridge
1760 * module, preventing both of them from being loaded at the same
1762 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1764 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1770 static void dp_avoid_bridge_exit(void)
1772 llc_sap_put(dp_stp_sap);
1774 #else /* Linux 2.6.27 or later. */
1775 static int dp_avoid_bridge_init(void)
1777 /* Linux 2.6.27 introduces a way for multiple clients to register for
1778 * STP packets, which interferes with what we try to do above.
1779 * Instead, just check whether there's a bridge hook defined. This is
1780 * not as safe--the bridge module is willing to load over the top of
1781 * us--but it provides a little bit of protection. */
1782 if (br_handle_frame_hook) {
1783 printk(KERN_ERR "openvswitch: bridge module is loaded, cannot load over it\n");
1789 static void dp_avoid_bridge_exit(void)
1791 /* Nothing to do. */
1793 #endif /* Linux 2.6.27 or later */
1795 static int __init dp_init(void)
1799 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1801 err = dp_avoid_bridge_init();
1809 err = register_netdevice_notifier(&dp_device_notifier);
1811 goto error_flow_exit;
1813 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1815 goto error_unreg_notifier;
1817 /* Hook into callback used by the bridge to intercept packets.
1818 * Parasites we are. */
1819 br_handle_frame_hook = dp_frame_hook;
1823 error_unreg_notifier:
1824 unregister_netdevice_notifier(&dp_device_notifier);
1831 static void dp_cleanup(void)
1834 unregister_chrdev(major, "openvswitch");
1835 unregister_netdevice_notifier(&dp_device_notifier);
1837 br_handle_frame_hook = NULL;
1838 dp_avoid_bridge_exit();
1841 module_init(dp_init);
1842 module_exit(dp_cleanup);
1844 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1845 MODULE_LICENSE("GPL");