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>
44 #include <net/inet_ecn.h>
47 #include "openvswitch/datapath-protocol.h"
56 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
57 EXPORT_SYMBOL(dp_ioctl_hook);
59 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
62 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
65 * It is safe to access the datapath and net_bridge_port structures with just
68 static struct datapath *dps[ODP_MAX];
69 static DEFINE_MUTEX(dp_mutex);
71 /* Number of milliseconds between runs of the maintenance thread. */
72 #define MAINT_SLEEP_MSECS 1000
74 static int new_nbp(struct datapath *, struct net_device *, int port_no);
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 dev_disable_lro(dev);
363 rcu_assign_pointer(dp->ports[port_no], p);
364 list_add_rcu(&p->node, &dp->port_list);
367 /* Initialize kobject for bridge. This will be added as
368 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
370 kobject_init(&p->kobj, &brport_ktype);
372 dp_ifinfo_notify(RTM_NEWLINK, p);
377 static int add_port(int dp_idx, struct odp_port __user *portp)
379 struct net_device *dev;
381 struct odp_port port;
386 if (copy_from_user(&port, portp, sizeof port))
388 port.devname[IFNAMSIZ - 1] = '\0';
391 dp = get_dp_locked(dp_idx);
394 goto out_unlock_rtnl;
396 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
397 if (!dp->ports[port_no])
403 if (!(port.flags & ODP_PORT_INTERNAL)) {
405 dev = dev_get_by_name(&init_net, port.devname);
410 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
414 dev = dp_dev_create(dp, port.devname, port_no);
421 err = new_nbp(dp, dev, port_no);
425 set_dp_devs_mtu(dp, dev);
426 dp_sysfs_add_if(dp->ports[port_no]);
428 err = __put_user(port_no, &portp->port);
433 mutex_unlock(&dp->mutex);
440 int dp_del_port(struct net_bridge_port *p)
444 if (p->port_no != ODPP_LOCAL)
446 dp_ifinfo_notify(RTM_DELLINK, p);
450 if (is_dp_dev(p->dev)) {
451 /* Make sure that no packets arrive from now on, since
452 * dp_dev_xmit() will try to find itself through
453 * p->dp->ports[], and we're about to set that to null. */
454 netif_tx_disable(p->dev);
457 /* First drop references to device. */
458 dev_set_promiscuity(p->dev, -1);
459 list_del_rcu(&p->node);
460 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
461 rcu_assign_pointer(p->dev->br_port, NULL);
463 /* Then wait until no one is still using it, and destroy it. */
466 if (is_dp_dev(p->dev))
467 dp_dev_destroy(p->dev);
469 kobject_put(&p->kobj);
474 static int del_port(int dp_idx, int port_no)
476 struct net_bridge_port *p;
482 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
486 dp = get_dp_locked(dp_idx);
489 goto out_unlock_rtnl;
491 p = dp->ports[port_no];
496 err = dp_del_port(p);
499 mutex_unlock(&dp->mutex);
506 /* Must be called with rcu_read_lock. */
508 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
510 /* LRO isn't suitable for bridging. We turn it off but make sure
511 * that it wasn't reactivated. */
512 if (skb_warn_if_lro(skb))
515 /* Make our own copy of the packet. Otherwise we will mangle the
516 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
517 * (No one comes after us, since we tell handle_bridge() that we took
519 skb = skb_share_check(skb, GFP_ATOMIC);
523 /* Push the Ethernet header back on. */
524 skb_push(skb, ETH_HLEN);
525 skb_reset_mac_header(skb);
526 dp_process_received_packet(skb, p);
529 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
530 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
532 struct datapath *dp = p->dp;
533 struct dp_stats_percpu *stats;
534 struct odp_flow_key key;
535 struct sw_flow *flow;
537 WARN_ON_ONCE(skb_shared(skb));
539 compute_ip_summed(skb, false);
541 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
542 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
544 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
545 if (dp->drop_frags) {
552 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
554 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
555 flow_used(flow, skb);
556 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
561 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
566 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
567 * different set of devices!)
569 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
570 /* Called with rcu_read_lock and bottom-halves disabled. */
571 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
574 do_port_input(p, skb);
577 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
578 /* Called with rcu_read_lock and bottom-halves disabled. */
579 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
581 do_port_input(p, *pskb);
588 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
589 /* This code is based on a skb_checksum_setup from net/dev/core.c from a
590 * combination of Lenny's 2.6.26 Xen kernel and Xen's
591 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call this function
592 * directly because it isn't exported in all versions. */
593 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
595 if (ptr < (void *)skb->tail)
597 if (__pskb_pull_tail(skb,
598 ptr - (void *)skb->data - skb_headlen(skb))) {
605 int vswitch_skb_checksum_setup(struct sk_buff *skb)
610 __u16 csum_start, csum_offset;
612 if (!skb->proto_csum_blank)
615 if (skb->protocol != htons(ETH_P_IP))
618 if (!skb_pull_up_to(skb, skb_network_header(skb) + sizeof(struct iphdr)))
622 th = skb_network_header(skb) + 4 * iph->ihl;
624 csum_start = th - skb->head;
625 switch (iph->protocol) {
627 csum_offset = offsetof(struct tcphdr, check);
630 csum_offset = offsetof(struct udphdr, check);
634 printk(KERN_ERR "Attempting to checksum a non-"
635 "TCP/UDP packet, dropping a protocol"
636 " %d packet", iph->protocol);
640 if (!skb_pull_up_to(skb, th + csum_offset + 2))
643 skb->ip_summed = CHECKSUM_PARTIAL;
644 skb->proto_csum_blank = 0;
646 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
647 skb->csum_start = csum_start;
648 skb->csum_offset = csum_offset;
650 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
651 skb->csum = csum_offset;
659 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
661 /* Types of checksums that we can receive (these all refer to L4 checksums):
662 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
663 * (though not verified) checksum in packet but not in skb->csum. Packets
664 * from the bridge local port will also have this type.
665 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
666 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
667 * a valid skb->csum. Importantly, both contain a full checksum (not
668 * verified) in the packet itself. The only difference is that if the
669 * packet gets to L4 processing on this machine (not in DomU) we won't
670 * have to recompute the checksum to verify. Most hardware devices do not
671 * produce packets with this type, even if they support receive checksum
672 * offloading (they produce type #5).
673 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
674 * be computed if it is sent off box. Unfortunately on earlier kernels,
675 * this case is impossible to distinguish from #2, despite having opposite
676 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
677 * to distinguish the different states. The only real user of this type
678 * with bridging is Xen (on later kernels).
679 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
680 * generated locally by a Xen DomU and has a partial checksum. If it is
681 * handled on this machine (Dom0 or DomU), then the checksum will not be
682 * computed. If it goes off box, the checksum in the packet needs to be
683 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
684 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
685 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
686 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
687 * full checksum or using a protocol without a checksum. skb->csum is
688 * undefined. This is common from devices with receive checksum
689 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
690 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
692 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
693 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
694 * based on whether it is on the transmit or receive path. After the datapath
695 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
696 * checksum, we will panic. Since we can receive packets with checksums, we
697 * assume that all CHECKSUM_HW packets have checksums and map them to
698 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
699 * packet is processed by the local IP stack, in which case it will need to
700 * be reverified). If we receive a packet with CHECKSUM_HW that really means
701 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
702 * shouldn't be any devices that do this with bridging.
704 * The bridge has similar behavior and this function closely resembles
705 * skb_forward_csum(). It is slightly different because we are only concerned
706 * with bridging and not other types of forwarding and can get away with
707 * slightly more optimal behavior.*/
709 compute_ip_summed(struct sk_buff *skb, bool xmit)
711 /* For our convenience these defines change repeatedly between kernel
712 * versions, so we can't just copy them over... */
713 switch (skb->ip_summed) {
715 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
717 case CHECKSUM_UNNECESSARY:
718 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
721 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
722 * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
723 * uses some special fields to represent this (see below). Since we
724 * can only make one type work, pick the one that actually happens in
727 * The one exception to this is if we are on the transmit path
728 * (basically after skb_checksum_setup() has been run) the type has
729 * already been converted, so we should stay with that. */
732 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
734 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
738 case CHECKSUM_COMPLETE:
739 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
741 case CHECKSUM_PARTIAL:
742 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
746 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
748 /* None seems the safest... */
749 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
752 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
753 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
754 * kernels. It should not be set on the transmit path though. */
755 if (skb->proto_csum_blank)
756 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
758 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
763 forward_ip_summed(struct sk_buff *skb)
766 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
767 skb->ip_summed = CHECKSUM_NONE;
771 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
772 * unless we broke up a GSO packet. */
774 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
775 int queue_no, u32 arg)
777 struct sk_buff *nskb;
781 port_no = ODPP_LOCAL;
783 if (skb->dev->br_port)
784 port_no = skb->dev->br_port->port_no;
785 else if (is_dp_dev(skb->dev))
786 port_no = dp_dev_priv(skb->dev)->port_no;
790 struct odp_msg *header;
795 /* If a checksum-deferred packet is forwarded to the
796 * controller, correct the pointers and checksum. This happens
797 * on a regular basis only on Xen, on which VMs can pass up
798 * packets that do not have their checksum computed.
800 err = vswitch_skb_checksum_setup(skb);
804 if (skb->ip_summed == CHECKSUM_PARTIAL) {
805 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
806 /* Until 2.6.22, the start of the transport header was
807 * also the start of data to be checksummed. Linux
808 * 2.6.22 introduced the csum_start field for this
809 * purpose, but we should point the transport header to
810 * it anyway for backward compatibility, as
811 * dev_queue_xmit() does even in 2.6.28. */
812 skb_set_transport_header(skb, skb->csum_start -
815 err = skb_checksum_help(skb);
820 if (skb->ip_summed == CHECKSUM_HW) {
821 err = skb_checksum_help(skb, 0);
827 err = skb_cow(skb, sizeof *header);
831 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
832 header->type = queue_no;
833 header->length = skb->len;
834 header->port = port_no;
835 header->reserved = 0;
837 skb_queue_tail(queue, skb);
845 while ((skb = nskb) != NULL) {
853 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
856 struct dp_stats_percpu *stats;
857 struct sk_buff_head *queue;
860 WARN_ON_ONCE(skb_shared(skb));
861 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
862 queue = &dp->queues[queue_no];
864 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
867 forward_ip_summed(skb);
869 /* Break apart GSO packets into their component pieces. Otherwise
870 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
871 if (skb_is_gso(skb)) {
872 struct sk_buff *nskb = skb_gso_segment(skb, 0);
876 if (unlikely(IS_ERR(skb))) {
881 /* XXX This case might not be possible. It's hard to
882 * tell from the skb_gso_segment() code and comment. */
886 err = queue_control_packets(skb, queue, queue_no, arg);
887 wake_up_interruptible(&dp->waitqueue);
893 stats = percpu_ptr(dp->stats_percpu, get_cpu());
900 static int flush_flows(struct datapath *dp)
903 return dp_table_flush(dp);
906 static int validate_actions(const struct sw_flow_actions *actions)
910 for (i = 0; i < actions->n_actions; i++) {
911 const union odp_action *a = &actions->actions[i];
914 if (a->output.port >= DP_MAX_PORTS)
918 case ODPAT_OUTPUT_GROUP:
919 if (a->output_group.group >= DP_MAX_GROUPS)
923 case ODPAT_SET_VLAN_VID:
924 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
928 case ODPAT_SET_VLAN_PCP:
929 if (a->vlan_pcp.vlan_pcp
930 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
934 case ODPAT_SET_NW_TOS:
935 if (a->nw_tos.nw_tos & INET_ECN_MASK)
940 if (a->type >= ODPAT_N_ACTIONS)
949 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
951 struct sw_flow_actions *actions;
954 actions = flow_actions_alloc(flow->n_actions);
955 error = PTR_ERR(actions);
960 if (copy_from_user(actions->actions, flow->actions,
961 flow->n_actions * sizeof(union odp_action)))
962 goto error_free_actions;
963 error = validate_actions(actions);
965 goto error_free_actions;
972 return ERR_PTR(error);
975 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
977 if (flow->used.tv_sec) {
978 stats->used_sec = flow->used.tv_sec;
979 stats->used_nsec = flow->used.tv_nsec;
982 stats->used_nsec = 0;
984 stats->n_packets = flow->packet_count;
985 stats->n_bytes = flow->byte_count;
986 stats->ip_tos = flow->ip_tos;
987 stats->tcp_flags = flow->tcp_flags;
991 static void clear_stats(struct sw_flow *flow)
993 flow->used.tv_sec = flow->used.tv_nsec = 0;
996 flow->packet_count = 0;
997 flow->byte_count = 0;
1000 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
1002 struct odp_flow_put uf;
1003 struct sw_flow *flow;
1004 struct dp_table *table;
1005 struct odp_flow_stats stats;
1009 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
1011 memset(uf.flow.key.reserved, 0, sizeof uf.flow.key.reserved);
1013 table = rcu_dereference(dp->table);
1014 flow = dp_table_lookup(table, &uf.flow.key);
1017 struct sw_flow_actions *acts;
1020 if (!(uf.flags & ODPPF_CREATE))
1023 /* Expand table, if necessary, to make room. */
1024 if (dp->n_flows >= table->n_buckets) {
1026 if (table->n_buckets >= DP_MAX_BUCKETS)
1029 error = dp_table_expand(dp);
1032 table = rcu_dereference(dp->table);
1035 /* Allocate flow. */
1037 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
1040 flow->key = uf.flow.key;
1041 spin_lock_init(&flow->lock);
1044 /* Obtain actions. */
1045 acts = get_actions(&uf.flow);
1046 error = PTR_ERR(acts);
1048 goto error_free_flow;
1049 rcu_assign_pointer(flow->sf_acts, acts);
1051 /* Put flow in bucket. */
1052 error = dp_table_insert(table, flow);
1054 goto error_free_flow_acts;
1056 memset(&stats, 0, sizeof(struct odp_flow_stats));
1058 /* We found a matching flow. */
1059 struct sw_flow_actions *old_acts, *new_acts;
1060 unsigned long int flags;
1062 /* Bail out if we're not allowed to modify an existing flow. */
1064 if (!(uf.flags & ODPPF_MODIFY))
1068 new_acts = get_actions(&uf.flow);
1069 error = PTR_ERR(new_acts);
1070 if (IS_ERR(new_acts))
1072 old_acts = rcu_dereference(flow->sf_acts);
1073 if (old_acts->n_actions != new_acts->n_actions ||
1074 memcmp(old_acts->actions, new_acts->actions,
1075 sizeof(union odp_action) * old_acts->n_actions)) {
1076 rcu_assign_pointer(flow->sf_acts, new_acts);
1077 flow_deferred_free_acts(old_acts);
1082 /* Fetch stats, then clear them if necessary. */
1083 spin_lock_irqsave(&flow->lock, flags);
1084 get_stats(flow, &stats);
1085 if (uf.flags & ODPPF_ZERO_STATS)
1087 spin_unlock_irqrestore(&flow->lock, flags);
1090 /* Copy stats to userspace. */
1091 if (__copy_to_user(&ufp->flow.stats, &stats,
1092 sizeof(struct odp_flow_stats)))
1096 error_free_flow_acts:
1097 kfree(flow->sf_acts);
1099 kmem_cache_free(flow_cache, flow);
1104 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1106 union odp_action __user *actions;
1107 struct sw_flow_actions *sf_acts;
1110 if (__get_user(actions, &ufp->actions) ||
1111 __get_user(n_actions, &ufp->n_actions))
1117 sf_acts = rcu_dereference(flow->sf_acts);
1118 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1119 (actions && copy_to_user(actions, sf_acts->actions,
1120 sizeof(union odp_action) *
1121 min(sf_acts->n_actions, n_actions))))
1127 static int answer_query(struct sw_flow *flow, u32 query_flags,
1128 struct odp_flow __user *ufp)
1130 struct odp_flow_stats stats;
1131 unsigned long int flags;
1133 spin_lock_irqsave(&flow->lock, flags);
1134 get_stats(flow, &stats);
1136 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1137 flow->tcp_flags = 0;
1139 spin_unlock_irqrestore(&flow->lock, flags);
1141 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1143 return put_actions(flow, ufp);
1146 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1148 struct dp_table *table = rcu_dereference(dp->table);
1150 struct sw_flow *flow;
1154 if (copy_from_user(&uf, ufp, sizeof uf))
1156 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1158 flow = dp_table_lookup(table, &uf.key);
1163 /* XXX redundant lookup */
1164 error = dp_table_delete(table, flow);
1168 /* XXX These statistics might lose a few packets, since other CPUs can
1169 * be using this flow. We used to synchronize_rcu() to make sure that
1170 * we get completely accurate stats, but that blows our performance,
1173 error = answer_query(flow, 0, ufp);
1174 flow_deferred_free(flow);
1180 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1182 struct dp_table *table = rcu_dereference(dp->table);
1184 for (i = 0; i < flowvec->n_flows; i++) {
1185 struct __user odp_flow *ufp = &flowvec->flows[i];
1187 struct sw_flow *flow;
1190 if (__copy_from_user(&uf, ufp, sizeof uf))
1192 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1194 flow = dp_table_lookup(table, &uf.key);
1196 error = __put_user(ENOENT, &ufp->stats.error);
1198 error = answer_query(flow, uf.flags, ufp);
1202 return flowvec->n_flows;
1205 struct list_flows_cbdata {
1206 struct odp_flow __user *uflows;
1211 static int list_flow(struct sw_flow *flow, void *cbdata_)
1213 struct list_flows_cbdata *cbdata = cbdata_;
1214 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1217 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1219 error = answer_query(flow, 0, ufp);
1223 if (cbdata->listed_flows >= cbdata->n_flows)
1224 return cbdata->listed_flows;
1228 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1230 struct list_flows_cbdata cbdata;
1233 if (!flowvec->n_flows)
1236 cbdata.uflows = flowvec->flows;
1237 cbdata.n_flows = flowvec->n_flows;
1238 cbdata.listed_flows = 0;
1239 error = dp_table_foreach(rcu_dereference(dp->table),
1240 list_flow, &cbdata);
1241 return error ? error : cbdata.listed_flows;
1244 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1245 int (*function)(struct datapath *,
1246 const struct odp_flowvec *))
1248 struct odp_flowvec __user *uflowvec;
1249 struct odp_flowvec flowvec;
1252 uflowvec = (struct odp_flowvec __user *)argp;
1253 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1254 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1257 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1260 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1261 flowvec.n_flows * sizeof(struct odp_flow)))
1264 retval = function(dp, &flowvec);
1265 return (retval < 0 ? retval
1266 : retval == flowvec.n_flows ? 0
1267 : __put_user(retval, &uflowvec->n_flows));
1270 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1272 struct odp_execute execute;
1273 struct odp_flow_key key;
1274 struct sk_buff *skb;
1275 struct sw_flow_actions *actions;
1280 if (copy_from_user(&execute, executep, sizeof execute))
1284 if (execute.length < ETH_HLEN || execute.length > 65535)
1288 actions = flow_actions_alloc(execute.n_actions);
1293 if (copy_from_user(actions->actions, execute.actions,
1294 execute.n_actions * sizeof *execute.actions))
1295 goto error_free_actions;
1297 err = validate_actions(actions);
1299 goto error_free_actions;
1302 skb = alloc_skb(execute.length, GFP_KERNEL);
1304 goto error_free_actions;
1305 if (execute.in_port < DP_MAX_PORTS) {
1306 struct net_bridge_port *p = dp->ports[execute.in_port];
1312 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1314 goto error_free_skb;
1316 skb_reset_mac_header(skb);
1319 /* Normally, setting the skb 'protocol' field would be handled by a
1320 * call to eth_type_trans(), but it assumes there's a sending
1321 * device, which we may not have. */
1322 if (ntohs(eth->h_proto) >= 1536)
1323 skb->protocol = eth->h_proto;
1325 skb->protocol = htons(ETH_P_802_2);
1327 flow_extract(skb, execute.in_port, &key);
1328 err = execute_actions(dp, skb, &key, actions->actions,
1329 actions->n_actions, GFP_KERNEL);
1341 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1343 struct odp_stats stats;
1346 stats.n_flows = dp->n_flows;
1347 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets;
1348 stats.max_capacity = DP_MAX_BUCKETS;
1349 stats.n_ports = dp->n_ports;
1350 stats.max_ports = DP_MAX_PORTS;
1351 stats.max_groups = DP_MAX_GROUPS;
1352 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1353 for_each_possible_cpu(i) {
1354 const struct dp_stats_percpu *s;
1355 s = percpu_ptr(dp->stats_percpu, i);
1356 stats.n_frags += s->n_frags;
1357 stats.n_hit += s->n_hit;
1358 stats.n_missed += s->n_missed;
1359 stats.n_lost += s->n_lost;
1361 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1362 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1363 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1366 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1367 int dp_min_mtu(const struct datapath *dp)
1369 struct net_bridge_port *p;
1374 list_for_each_entry_rcu (p, &dp->port_list, node) {
1375 struct net_device *dev = p->dev;
1377 /* Skip any internal ports, since that's what we're trying to
1382 if (!mtu || dev->mtu < mtu)
1386 return mtu ? mtu : ETH_DATA_LEN;
1389 /* Sets the MTU of all datapath devices to the minimum of the ports. 'dev'
1390 * is the device whose MTU may have changed. Must be called with RTNL lock
1392 void set_dp_devs_mtu(const struct datapath *dp, struct net_device *dev)
1394 struct net_bridge_port *p;
1402 mtu = dp_min_mtu(dp);
1404 list_for_each_entry_rcu (p, &dp->port_list, node) {
1405 struct net_device *br_dev = p->dev;
1407 if (is_dp_dev(br_dev))
1408 dev_set_mtu(br_dev, mtu);
1413 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1416 memset(&op, 0, sizeof op);
1417 strncpy(op.devname, p->dev->name, sizeof op.devname);
1418 op.port = p->port_no;
1419 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1420 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1424 query_port(struct datapath *dp, struct odp_port __user *uport)
1426 struct odp_port port;
1428 if (copy_from_user(&port, uport, sizeof port))
1430 if (port.devname[0]) {
1431 struct net_bridge_port *p;
1432 struct net_device *dev;
1435 port.devname[IFNAMSIZ - 1] = '\0';
1437 dev = dev_get_by_name(&init_net, port.devname);
1442 if (!p && is_dp_dev(dev)) {
1443 struct dp_dev *dp_dev = dp_dev_priv(dev);
1444 if (dp_dev->dp == dp)
1445 p = dp->ports[dp_dev->port_no];
1447 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1452 if (port.port >= DP_MAX_PORTS)
1454 if (!dp->ports[port.port])
1456 return put_port(dp->ports[port.port], uport);
1461 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1463 struct odp_portvec pv;
1464 struct net_bridge_port *p;
1467 if (copy_from_user(&pv, pvp, sizeof pv))
1472 list_for_each_entry_rcu (p, &dp->port_list, node) {
1473 if (put_port(p, &pv.ports[idx]))
1475 if (idx++ >= pv.n_ports)
1479 return put_user(dp->n_ports, &pvp->n_ports);
1482 /* RCU callback for freeing a dp_port_group */
1483 static void free_port_group(struct rcu_head *rcu)
1485 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1490 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1492 struct odp_port_group pg;
1493 struct dp_port_group *new_group, *old_group;
1497 if (copy_from_user(&pg, upg, sizeof pg))
1501 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1505 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1510 new_group->n_ports = pg.n_ports;
1512 if (copy_from_user(new_group->ports, pg.ports,
1513 sizeof(u16) * pg.n_ports))
1516 old_group = rcu_dereference(dp->groups[pg.group]);
1517 rcu_assign_pointer(dp->groups[pg.group], new_group);
1519 call_rcu(&old_group->rcu, free_port_group);
1529 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1531 struct odp_port_group pg;
1532 struct dp_port_group *g;
1535 if (copy_from_user(&pg, upg, sizeof pg))
1538 if (pg.group >= DP_MAX_GROUPS)
1541 g = dp->groups[pg.group];
1542 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1543 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1546 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1552 static int get_listen_mask(const struct file *f)
1554 return (long)f->private_data;
1557 static void set_listen_mask(struct file *f, int listen_mask)
1559 f->private_data = (void*)(long)listen_mask;
1562 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1565 int dp_idx = iminor(f->f_dentry->d_inode);
1566 struct datapath *dp;
1567 int drop_frags, listeners, port_no;
1568 unsigned int sflow_probability;
1571 /* Handle commands with special locking requirements up front. */
1574 err = create_dp(dp_idx, (char __user *)argp);
1577 case ODP_DP_DESTROY:
1578 err = destroy_dp(dp_idx);
1582 err = add_port(dp_idx, (struct odp_port __user *)argp);
1586 err = get_user(port_no, (int __user *)argp);
1588 err = del_port(dp_idx, port_no);
1592 dp = get_dp_locked(dp_idx);
1599 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1602 case ODP_GET_DROP_FRAGS:
1603 err = put_user(dp->drop_frags, (int __user *)argp);
1606 case ODP_SET_DROP_FRAGS:
1607 err = get_user(drop_frags, (int __user *)argp);
1611 if (drop_frags != 0 && drop_frags != 1)
1613 dp->drop_frags = drop_frags;
1617 case ODP_GET_LISTEN_MASK:
1618 err = put_user(get_listen_mask(f), (int __user *)argp);
1621 case ODP_SET_LISTEN_MASK:
1622 err = get_user(listeners, (int __user *)argp);
1626 if (listeners & ~ODPL_ALL)
1629 set_listen_mask(f, listeners);
1632 case ODP_GET_SFLOW_PROBABILITY:
1633 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1636 case ODP_SET_SFLOW_PROBABILITY:
1637 err = get_user(sflow_probability, (unsigned int __user *)argp);
1639 dp->sflow_probability = sflow_probability;
1642 case ODP_PORT_QUERY:
1643 err = query_port(dp, (struct odp_port __user *)argp);
1647 err = list_ports(dp, (struct odp_portvec __user *)argp);
1650 case ODP_PORT_GROUP_SET:
1651 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1654 case ODP_PORT_GROUP_GET:
1655 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1658 case ODP_FLOW_FLUSH:
1659 err = flush_flows(dp);
1663 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1667 err = del_flow(dp, (struct odp_flow __user *)argp);
1671 err = do_flowvec_ioctl(dp, argp, query_flows);
1675 err = do_flowvec_ioctl(dp, argp, list_flows);
1679 err = do_execute(dp, (struct odp_execute __user *)argp);
1686 mutex_unlock(&dp->mutex);
1691 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1694 for (i = 0; i < DP_N_QUEUES; i++) {
1695 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1701 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1704 /* XXX is there sufficient synchronization here? */
1705 int listeners = get_listen_mask(f);
1706 int dp_idx = iminor(f->f_dentry->d_inode);
1707 struct datapath *dp = get_dp(dp_idx);
1708 struct sk_buff *skb;
1709 struct iovec __user iov;
1716 if (nbytes == 0 || !listeners)
1722 for (i = 0; i < DP_N_QUEUES; i++) {
1723 if (listeners & (1 << i)) {
1724 skb = skb_dequeue(&dp->queues[i]);
1730 if (f->f_flags & O_NONBLOCK) {
1735 wait_event_interruptible(dp->waitqueue,
1736 dp_has_packet_of_interest(dp,
1739 if (signal_pending(current)) {
1740 retval = -ERESTARTSYS;
1745 copy_bytes = min_t(size_t, skb->len, nbytes);
1747 iov.iov_len = copy_bytes;
1748 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1750 retval = copy_bytes;
1757 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1759 /* XXX is there sufficient synchronization here? */
1760 int dp_idx = iminor(file->f_dentry->d_inode);
1761 struct datapath *dp = get_dp(dp_idx);
1766 poll_wait(file, &dp->waitqueue, wait);
1767 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1768 mask |= POLLIN | POLLRDNORM;
1770 mask = POLLIN | POLLRDNORM | POLLHUP;
1775 struct file_operations openvswitch_fops = {
1776 /* XXX .aio_read = openvswitch_aio_read, */
1777 .read = openvswitch_read,
1778 .poll = openvswitch_poll,
1779 .unlocked_ioctl = openvswitch_ioctl,
1780 /* XXX .fasync = openvswitch_fasync, */
1785 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)
1786 static struct llc_sap *dp_stp_sap;
1788 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1789 struct packet_type *pt, struct net_device *orig_dev)
1791 /* We don't really care about STP packets, we just listen for them for
1792 * mutual exclusion with the bridge module, so this just discards
1798 static int dp_avoid_bridge_init(void)
1800 /* Register to receive STP packets because the bridge module also
1801 * attempts to do so. Since there can only be a single listener for a
1802 * given protocol, this provides mutual exclusion against the bridge
1803 * module, preventing both of them from being loaded at the same
1805 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1807 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1813 static void dp_avoid_bridge_exit(void)
1815 llc_sap_put(dp_stp_sap);
1817 #else /* Linux 2.6.27 or later. */
1818 static int dp_avoid_bridge_init(void)
1820 /* Linux 2.6.27 introduces a way for multiple clients to register for
1821 * STP packets, which interferes with what we try to do above.
1822 * Instead, just check whether there's a bridge hook defined. This is
1823 * not as safe--the bridge module is willing to load over the top of
1824 * us--but it provides a little bit of protection. */
1825 if (br_handle_frame_hook) {
1826 printk(KERN_ERR "openvswitch: bridge module is loaded, cannot load over it\n");
1832 static void dp_avoid_bridge_exit(void)
1834 /* Nothing to do. */
1836 #endif /* Linux 2.6.27 or later */
1838 static int __init dp_init(void)
1842 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1844 err = dp_avoid_bridge_init();
1852 err = register_netdevice_notifier(&dp_device_notifier);
1854 goto error_flow_exit;
1856 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1858 goto error_unreg_notifier;
1860 /* Hook into callback used by the bridge to intercept packets.
1861 * Parasites we are. */
1862 br_handle_frame_hook = dp_frame_hook;
1866 error_unreg_notifier:
1867 unregister_netdevice_notifier(&dp_device_notifier);
1874 static void dp_cleanup(void)
1877 unregister_chrdev(major, "openvswitch");
1878 unregister_netdevice_notifier(&dp_device_notifier);
1880 br_handle_frame_hook = NULL;
1881 dp_avoid_bridge_exit();
1884 module_init(dp_init);
1885 module_exit(dp_cleanup);
1887 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1888 MODULE_LICENSE("GPL");