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;
352 atomic_set(&p->sflow_pool, 0);
354 rcu_assign_pointer(dev->br_port, p);
356 /* It would make sense to assign dev->br_port here too, but
357 * that causes packets received on internal ports to get caught
358 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
359 * back to network stack, but that's a waste of time. */
361 rcu_assign_pointer(dp->ports[port_no], p);
362 list_add_rcu(&p->node, &dp->port_list);
365 /* Initialize kobject for bridge. This will be added as
366 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
368 kobject_init(&p->kobj, &brport_ktype);
370 dp_ifinfo_notify(RTM_NEWLINK, p);
375 static int add_port(int dp_idx, struct odp_port __user *portp)
377 struct net_device *dev;
379 struct odp_port port;
384 if (copy_from_user(&port, portp, sizeof port))
386 port.devname[IFNAMSIZ - 1] = '\0';
389 dp = get_dp_locked(dp_idx);
392 goto out_unlock_rtnl;
394 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
395 if (!dp->ports[port_no])
401 if (!(port.flags & ODP_PORT_INTERNAL)) {
403 dev = dev_get_by_name(&init_net, port.devname);
408 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
412 dev = dp_dev_create(dp, port.devname, port_no);
419 err = new_nbp(dp, dev, port_no);
423 set_dp_devs_mtu(dp, dev);
424 dp_sysfs_add_if(dp->ports[port_no]);
426 err = __put_user(port_no, &port.port);
431 mutex_unlock(&dp->mutex);
438 int dp_del_port(struct net_bridge_port *p)
442 if (p->port_no != ODPP_LOCAL)
444 dp_ifinfo_notify(RTM_DELLINK, p);
448 if (is_dp_dev(p->dev)) {
449 /* Make sure that no packets arrive from now on, since
450 * dp_dev_xmit() will try to find itself through
451 * p->dp->ports[], and we're about to set that to null. */
452 netif_tx_disable(p->dev);
455 /* First drop references to device. */
456 dev_set_promiscuity(p->dev, -1);
457 list_del_rcu(&p->node);
458 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
459 rcu_assign_pointer(p->dev->br_port, NULL);
461 /* Then wait until no one is still using it, and destroy it. */
464 if (is_dp_dev(p->dev))
465 dp_dev_destroy(p->dev);
467 kobject_put(&p->kobj);
472 static int del_port(int dp_idx, int port_no)
474 struct net_bridge_port *p;
480 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
484 dp = get_dp_locked(dp_idx);
487 goto out_unlock_rtnl;
489 p = dp->ports[port_no];
494 err = dp_del_port(p);
497 mutex_unlock(&dp->mutex);
504 /* Must be called with rcu_read_lock. */
506 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
508 /* Make our own copy of the packet. Otherwise we will mangle the
509 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
510 * (No one comes after us, since we tell handle_bridge() that we took
512 skb = skb_share_check(skb, GFP_ATOMIC);
516 /* Push the Ethernet header back on. */
517 skb_push(skb, ETH_HLEN);
518 skb_reset_mac_header(skb);
519 dp_process_received_packet(skb, p);
522 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
523 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
525 struct datapath *dp = p->dp;
526 struct dp_stats_percpu *stats;
527 struct odp_flow_key key;
528 struct sw_flow *flow;
530 WARN_ON_ONCE(skb_shared(skb));
532 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
533 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
535 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
536 if (dp->drop_frags) {
543 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
545 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
546 flow_used(flow, skb);
547 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
552 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
557 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
558 * different set of devices!)
560 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
561 /* Called with rcu_read_lock and bottom-halves disabled. */
562 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
565 do_port_input(p, skb);
568 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
569 /* Called with rcu_read_lock and bottom-halves disabled. */
570 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
572 do_port_input(p, *pskb);
579 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
580 /* This code is based on a skb_checksum_setup from net/dev/core.c from a
581 * combination of Lenny's 2.6.26 Xen kernel and Xen's
582 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call this function
583 * directly because it isn't exported in all versions. */
584 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
586 if (ptr < (void *)skb->tail)
588 if (__pskb_pull_tail(skb,
589 ptr - (void *)skb->data - skb_headlen(skb))) {
596 int vswitch_skb_checksum_setup(struct sk_buff *skb)
601 __u16 csum_start, csum_offset;
603 if (!skb->proto_csum_blank)
606 if (skb->protocol != htons(ETH_P_IP))
609 if (!skb_pull_up_to(skb, skb_network_header(skb) + 1))
613 th = skb_network_header(skb) + 4 * iph->ihl;
615 csum_start = th - skb->head;
616 switch (iph->protocol) {
618 csum_offset = offsetof(struct tcphdr, check);
621 csum_offset = offsetof(struct udphdr, check);
625 printk(KERN_ERR "Attempting to checksum a non-"
626 "TCP/UDP packet, dropping a protocol"
627 " %d packet", iph->protocol);
631 if (!skb_pull_up_to(skb, th + csum_offset + 2))
634 skb->ip_summed = CHECKSUM_PARTIAL;
635 skb->proto_csum_blank = 0;
637 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
638 skb->csum_start = csum_start;
639 skb->csum_offset = csum_offset;
641 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
642 skb->csum = csum_offset;
650 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
652 /* Types of checksums that we can receive (these all refer to L4 checksums):
653 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
654 * (though not verified) checksum in packet but not in skb->csum. Packets
655 * from the bridge local port will also have this type.
656 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
657 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
658 * a valid skb->csum. Importantly, both contain a full checksum (not
659 * verified) in the packet itself. The only difference is that if the
660 * packet gets to L4 processing on this machine (not in DomU) we won't
661 * have to recompute the checksum to verify. Most hardware devices do not
662 * produce packets with this type, even if they support receive checksum
663 * offloading (they produce type #5).
664 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
665 * be computed if it is sent off box. Unfortunately on earlier kernels,
666 * this case is impossible to distinguish from #2, despite having opposite
667 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
668 * to distinguish the different states. The only real user of this type
669 * with bridging is Xen (on later kernels).
670 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
671 * generated locally by a Xen DomU and has a partial checksum. If it is
672 * handled on this machine (Dom0 or DomU), then the checksum will not be
673 * computed. If it goes off box, the checksum in the packet needs to be
674 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
675 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
676 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
677 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
678 * full checksum or using a protocol without a checksum. skb->csum is
679 * undefined. This is common from devices with receive checksum
680 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
681 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
683 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
684 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
685 * based on whether it is on the transmit or receive path. After the datapath
686 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
687 * checksum, we will panic. Since we can receive packets with checksums, we
688 * assume that all CHECKSUM_HW packets have checksums and map them to
689 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
690 * packet is processed by the local IP stack, in which case it will need to
691 * be reverified). If we receive a packet with CHECKSUM_HW that really means
692 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
693 * shouldn't be any devices that do this with bridging.
695 * The bridge has similar behavior and this function closely resembles
696 * skb_forward_csum(). It is slightly different because we are only concerned
697 * with bridging and not other types of forwarding and can get away with
698 * slightly more optimal behavior.*/
700 forward_ip_summed(struct sk_buff *skb)
703 if (skb->ip_summed == CHECKSUM_HW)
704 skb->ip_summed = CHECKSUM_NONE;
708 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
709 * unless we broke up a GSO packet. */
711 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
712 int queue_no, u32 arg)
714 struct sk_buff *nskb;
718 port_no = ODPP_LOCAL;
720 if (skb->dev->br_port)
721 port_no = skb->dev->br_port->port_no;
722 else if (is_dp_dev(skb->dev))
723 port_no = dp_dev_priv(skb->dev)->port_no;
727 struct odp_msg *header;
732 /* If a checksum-deferred packet is forwarded to the
733 * controller, correct the pointers and checksum. This happens
734 * on a regular basis only on Xen, on which VMs can pass up
735 * packets that do not have their checksum computed.
737 err = vswitch_skb_checksum_setup(skb);
741 if (skb->ip_summed == CHECKSUM_PARTIAL) {
742 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
743 /* Until 2.6.22, the start of the transport header was
744 * also the start of data to be checksummed. Linux
745 * 2.6.22 introduced the csum_start field for this
746 * purpose, but we should point the transport header to
747 * it anyway for backward compatibility, as
748 * dev_queue_xmit() does even in 2.6.28. */
749 skb_set_transport_header(skb, skb->csum_start -
752 err = skb_checksum_help(skb);
757 if (skb->ip_summed == CHECKSUM_HW) {
758 err = skb_checksum_help(skb, 0);
764 err = skb_cow(skb, sizeof *header);
768 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
769 header->type = queue_no;
770 header->length = skb->len;
771 header->port = port_no;
772 header->reserved = 0;
774 skb_queue_tail(queue, skb);
782 while ((skb = nskb) != NULL) {
790 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
793 struct dp_stats_percpu *stats;
794 struct sk_buff_head *queue;
797 WARN_ON_ONCE(skb_shared(skb));
798 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
799 queue = &dp->queues[queue_no];
801 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
804 forward_ip_summed(skb);
806 /* Break apart GSO packets into their component pieces. Otherwise
807 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
808 if (skb_is_gso(skb)) {
809 struct sk_buff *nskb = skb_gso_segment(skb, 0);
813 if (unlikely(IS_ERR(skb))) {
818 /* XXX This case might not be possible. It's hard to
819 * tell from the skb_gso_segment() code and comment. */
823 err = queue_control_packets(skb, queue, queue_no, arg);
824 wake_up_interruptible(&dp->waitqueue);
830 stats = percpu_ptr(dp->stats_percpu, get_cpu());
837 static int flush_flows(struct datapath *dp)
840 return dp_table_flush(dp);
843 static int validate_actions(const struct sw_flow_actions *actions)
847 for (i = 0; i < actions->n_actions; i++) {
848 const union odp_action *a = &actions->actions[i];
851 if (a->output.port >= DP_MAX_PORTS)
855 case ODPAT_OUTPUT_GROUP:
856 if (a->output_group.group >= DP_MAX_GROUPS)
860 case ODPAT_SET_VLAN_VID:
861 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
865 case ODPAT_SET_VLAN_PCP:
866 if (a->vlan_pcp.vlan_pcp
867 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
872 if (a->type >= ODPAT_N_ACTIONS)
881 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
883 struct sw_flow_actions *actions;
886 actions = flow_actions_alloc(flow->n_actions);
887 error = PTR_ERR(actions);
892 if (copy_from_user(actions->actions, flow->actions,
893 flow->n_actions * sizeof(union odp_action)))
894 goto error_free_actions;
895 error = validate_actions(actions);
897 goto error_free_actions;
904 return ERR_PTR(error);
907 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
909 if (flow->used.tv_sec) {
910 stats->used_sec = flow->used.tv_sec;
911 stats->used_nsec = flow->used.tv_nsec;
914 stats->used_nsec = 0;
916 stats->n_packets = flow->packet_count;
917 stats->n_bytes = flow->byte_count;
918 stats->ip_tos = flow->ip_tos;
919 stats->tcp_flags = flow->tcp_flags;
923 static void clear_stats(struct sw_flow *flow)
925 flow->used.tv_sec = flow->used.tv_nsec = 0;
928 flow->packet_count = 0;
929 flow->byte_count = 0;
932 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
934 struct odp_flow_put uf;
935 struct sw_flow *flow;
936 struct dp_table *table;
937 struct odp_flow_stats stats;
941 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
943 uf.flow.key.reserved = 0;
945 table = rcu_dereference(dp->table);
946 flow = dp_table_lookup(table, &uf.flow.key);
949 struct sw_flow_actions *acts;
952 if (!(uf.flags & ODPPF_CREATE))
955 /* Expand table, if necessary, to make room. */
956 if (dp->n_flows >= table->n_buckets) {
958 if (table->n_buckets >= DP_MAX_BUCKETS)
961 error = dp_table_expand(dp);
964 table = rcu_dereference(dp->table);
969 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
972 flow->key = uf.flow.key;
973 spin_lock_init(&flow->lock);
976 /* Obtain actions. */
977 acts = get_actions(&uf.flow);
978 error = PTR_ERR(acts);
980 goto error_free_flow;
981 rcu_assign_pointer(flow->sf_acts, acts);
983 /* Put flow in bucket. */
984 error = dp_table_insert(table, flow);
986 goto error_free_flow_acts;
988 memset(&stats, 0, sizeof(struct odp_flow_stats));
990 /* We found a matching flow. */
991 struct sw_flow_actions *old_acts, *new_acts;
992 unsigned long int flags;
994 /* Bail out if we're not allowed to modify an existing flow. */
996 if (!(uf.flags & ODPPF_MODIFY))
1000 new_acts = get_actions(&uf.flow);
1001 error = PTR_ERR(new_acts);
1002 if (IS_ERR(new_acts))
1004 old_acts = rcu_dereference(flow->sf_acts);
1005 if (old_acts->n_actions != new_acts->n_actions ||
1006 memcmp(old_acts->actions, new_acts->actions,
1007 sizeof(union odp_action) * old_acts->n_actions)) {
1008 rcu_assign_pointer(flow->sf_acts, new_acts);
1009 flow_deferred_free_acts(old_acts);
1014 /* Fetch stats, then clear them if necessary. */
1015 spin_lock_irqsave(&flow->lock, flags);
1016 get_stats(flow, &stats);
1017 if (uf.flags & ODPPF_ZERO_STATS)
1019 spin_unlock_irqrestore(&flow->lock, flags);
1022 /* Copy stats to userspace. */
1023 if (__copy_to_user(&ufp->flow.stats, &stats,
1024 sizeof(struct odp_flow_stats)))
1028 error_free_flow_acts:
1029 kfree(flow->sf_acts);
1031 kmem_cache_free(flow_cache, flow);
1036 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1038 union odp_action __user *actions;
1039 struct sw_flow_actions *sf_acts;
1042 if (__get_user(actions, &ufp->actions) ||
1043 __get_user(n_actions, &ufp->n_actions))
1049 sf_acts = rcu_dereference(flow->sf_acts);
1050 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1051 (actions && copy_to_user(actions, sf_acts->actions,
1052 sizeof(union odp_action) *
1053 min(sf_acts->n_actions, n_actions))))
1059 static int answer_query(struct sw_flow *flow, u32 query_flags,
1060 struct odp_flow __user *ufp)
1062 struct odp_flow_stats stats;
1063 unsigned long int flags;
1065 spin_lock_irqsave(&flow->lock, flags);
1066 get_stats(flow, &stats);
1068 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1069 flow->tcp_flags = 0;
1071 spin_unlock_irqrestore(&flow->lock, flags);
1073 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1075 return put_actions(flow, ufp);
1078 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1080 struct dp_table *table = rcu_dereference(dp->table);
1082 struct sw_flow *flow;
1086 if (copy_from_user(&uf, ufp, sizeof uf))
1088 uf.key.reserved = 0;
1090 flow = dp_table_lookup(table, &uf.key);
1095 /* XXX redundant lookup */
1096 error = dp_table_delete(table, flow);
1100 /* XXX These statistics might lose a few packets, since other CPUs can
1101 * be using this flow. We used to synchronize_rcu() to make sure that
1102 * we get completely accurate stats, but that blows our performance,
1105 error = answer_query(flow, 0, ufp);
1106 flow_deferred_free(flow);
1112 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1114 struct dp_table *table = rcu_dereference(dp->table);
1116 for (i = 0; i < flowvec->n_flows; i++) {
1117 struct __user odp_flow *ufp = &flowvec->flows[i];
1119 struct sw_flow *flow;
1122 if (__copy_from_user(&uf, ufp, sizeof uf))
1124 uf.key.reserved = 0;
1126 flow = dp_table_lookup(table, &uf.key);
1128 error = __put_user(ENOENT, &ufp->stats.error);
1130 error = answer_query(flow, uf.flags, ufp);
1134 return flowvec->n_flows;
1137 struct list_flows_cbdata {
1138 struct odp_flow __user *uflows;
1143 static int list_flow(struct sw_flow *flow, void *cbdata_)
1145 struct list_flows_cbdata *cbdata = cbdata_;
1146 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1149 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1151 error = answer_query(flow, 0, ufp);
1155 if (cbdata->listed_flows >= cbdata->n_flows)
1156 return cbdata->listed_flows;
1160 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1162 struct list_flows_cbdata cbdata;
1165 if (!flowvec->n_flows)
1168 cbdata.uflows = flowvec->flows;
1169 cbdata.n_flows = flowvec->n_flows;
1170 cbdata.listed_flows = 0;
1171 error = dp_table_foreach(rcu_dereference(dp->table),
1172 list_flow, &cbdata);
1173 return error ? error : cbdata.listed_flows;
1176 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1177 int (*function)(struct datapath *,
1178 const struct odp_flowvec *))
1180 struct odp_flowvec __user *uflowvec;
1181 struct odp_flowvec flowvec;
1184 uflowvec = (struct odp_flowvec __user *)argp;
1185 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1186 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1189 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1192 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1193 flowvec.n_flows * sizeof(struct odp_flow)))
1196 retval = function(dp, &flowvec);
1197 return (retval < 0 ? retval
1198 : retval == flowvec.n_flows ? 0
1199 : __put_user(retval, &uflowvec->n_flows));
1202 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1204 struct odp_execute execute;
1205 struct odp_flow_key key;
1206 struct sk_buff *skb;
1207 struct sw_flow_actions *actions;
1212 if (copy_from_user(&execute, executep, sizeof execute))
1216 if (execute.length < ETH_HLEN || execute.length > 65535)
1220 actions = flow_actions_alloc(execute.n_actions);
1225 if (copy_from_user(actions->actions, execute.actions,
1226 execute.n_actions * sizeof *execute.actions))
1227 goto error_free_actions;
1229 err = validate_actions(actions);
1231 goto error_free_actions;
1234 skb = alloc_skb(execute.length, GFP_KERNEL);
1236 goto error_free_actions;
1237 if (execute.in_port < DP_MAX_PORTS) {
1238 struct net_bridge_port *p = dp->ports[execute.in_port];
1244 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1246 goto error_free_skb;
1248 skb_reset_mac_header(skb);
1251 /* Normally, setting the skb 'protocol' field would be handled by a
1252 * call to eth_type_trans(), but it assumes there's a sending
1253 * device, which we may not have. */
1254 if (ntohs(eth->h_proto) >= 1536)
1255 skb->protocol = eth->h_proto;
1257 skb->protocol = htons(ETH_P_802_2);
1259 flow_extract(skb, execute.in_port, &key);
1260 err = execute_actions(dp, skb, &key, actions->actions,
1261 actions->n_actions, GFP_KERNEL);
1273 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1275 struct odp_stats stats;
1278 stats.n_flows = dp->n_flows;
1279 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets;
1280 stats.max_capacity = DP_MAX_BUCKETS;
1281 stats.n_ports = dp->n_ports;
1282 stats.max_ports = DP_MAX_PORTS;
1283 stats.max_groups = DP_MAX_GROUPS;
1284 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1285 for_each_possible_cpu(i) {
1286 const struct dp_stats_percpu *s;
1287 s = percpu_ptr(dp->stats_percpu, i);
1288 stats.n_frags += s->n_frags;
1289 stats.n_hit += s->n_hit;
1290 stats.n_missed += s->n_missed;
1291 stats.n_lost += s->n_lost;
1293 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1294 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1295 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1298 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1299 int dp_min_mtu(const struct datapath *dp)
1301 struct net_bridge_port *p;
1306 list_for_each_entry_rcu (p, &dp->port_list, node) {
1307 struct net_device *dev = p->dev;
1309 /* Skip any internal ports, since that's what we're trying to
1314 if (!mtu || dev->mtu < mtu)
1318 return mtu ? mtu : ETH_DATA_LEN;
1321 /* Sets the MTU of all datapath devices to the minimum of the ports. 'dev'
1322 * is the device whose MTU may have changed. Must be called with RTNL lock
1324 void set_dp_devs_mtu(const struct datapath *dp, struct net_device *dev)
1326 struct net_bridge_port *p;
1334 mtu = dp_min_mtu(dp);
1336 list_for_each_entry_rcu (p, &dp->port_list, node) {
1337 struct net_device *br_dev = p->dev;
1339 if (is_dp_dev(br_dev))
1340 dev_set_mtu(br_dev, mtu);
1345 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1348 memset(&op, 0, sizeof op);
1349 strncpy(op.devname, p->dev->name, sizeof op.devname);
1350 op.port = p->port_no;
1351 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1352 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1356 query_port(struct datapath *dp, struct odp_port __user *uport)
1358 struct odp_port port;
1360 if (copy_from_user(&port, uport, sizeof port))
1362 if (port.devname[0]) {
1363 struct net_bridge_port *p;
1364 struct net_device *dev;
1367 port.devname[IFNAMSIZ - 1] = '\0';
1369 dev = dev_get_by_name(&init_net, port.devname);
1374 if (!p && is_dp_dev(dev)) {
1375 struct dp_dev *dp_dev = dp_dev_priv(dev);
1376 if (dp_dev->dp == dp)
1377 p = dp->ports[dp_dev->port_no];
1379 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1384 if (port.port >= DP_MAX_PORTS)
1386 if (!dp->ports[port.port])
1388 return put_port(dp->ports[port.port], uport);
1393 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1395 struct odp_portvec pv;
1396 struct net_bridge_port *p;
1399 if (copy_from_user(&pv, pvp, sizeof pv))
1404 list_for_each_entry_rcu (p, &dp->port_list, node) {
1405 if (put_port(p, &pv.ports[idx]))
1407 if (idx++ >= pv.n_ports)
1411 return put_user(dp->n_ports, &pvp->n_ports);
1414 /* RCU callback for freeing a dp_port_group */
1415 static void free_port_group(struct rcu_head *rcu)
1417 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1422 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1424 struct odp_port_group pg;
1425 struct dp_port_group *new_group, *old_group;
1429 if (copy_from_user(&pg, upg, sizeof pg))
1433 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1437 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1442 new_group->n_ports = pg.n_ports;
1444 if (copy_from_user(new_group->ports, pg.ports,
1445 sizeof(u16) * pg.n_ports))
1448 old_group = rcu_dereference(dp->groups[pg.group]);
1449 rcu_assign_pointer(dp->groups[pg.group], new_group);
1451 call_rcu(&old_group->rcu, free_port_group);
1461 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1463 struct odp_port_group pg;
1464 struct dp_port_group *g;
1467 if (copy_from_user(&pg, upg, sizeof pg))
1470 if (pg.group >= DP_MAX_GROUPS)
1473 g = dp->groups[pg.group];
1474 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1475 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1478 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1484 static int get_listen_mask(const struct file *f)
1486 return (long)f->private_data;
1489 static void set_listen_mask(struct file *f, int listen_mask)
1491 f->private_data = (void*)(long)listen_mask;
1494 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1497 int dp_idx = iminor(f->f_dentry->d_inode);
1498 struct datapath *dp;
1499 int drop_frags, listeners, port_no;
1500 unsigned int sflow_probability;
1503 /* Handle commands with special locking requirements up front. */
1506 err = create_dp(dp_idx, (char __user *)argp);
1509 case ODP_DP_DESTROY:
1510 err = destroy_dp(dp_idx);
1514 err = add_port(dp_idx, (struct odp_port __user *)argp);
1518 err = get_user(port_no, (int __user *)argp);
1520 err = del_port(dp_idx, port_no);
1524 dp = get_dp_locked(dp_idx);
1531 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1534 case ODP_GET_DROP_FRAGS:
1535 err = put_user(dp->drop_frags, (int __user *)argp);
1538 case ODP_SET_DROP_FRAGS:
1539 err = get_user(drop_frags, (int __user *)argp);
1543 if (drop_frags != 0 && drop_frags != 1)
1545 dp->drop_frags = drop_frags;
1549 case ODP_GET_LISTEN_MASK:
1550 err = put_user(get_listen_mask(f), (int __user *)argp);
1553 case ODP_SET_LISTEN_MASK:
1554 err = get_user(listeners, (int __user *)argp);
1558 if (listeners & ~ODPL_ALL)
1561 set_listen_mask(f, listeners);
1564 case ODP_GET_SFLOW_PROBABILITY:
1565 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1568 case ODP_SET_SFLOW_PROBABILITY:
1569 err = get_user(sflow_probability, (unsigned int __user *)argp);
1571 dp->sflow_probability = sflow_probability;
1574 case ODP_PORT_QUERY:
1575 err = query_port(dp, (struct odp_port __user *)argp);
1579 err = list_ports(dp, (struct odp_portvec __user *)argp);
1582 case ODP_PORT_GROUP_SET:
1583 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1586 case ODP_PORT_GROUP_GET:
1587 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1590 case ODP_FLOW_FLUSH:
1591 err = flush_flows(dp);
1595 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1599 err = del_flow(dp, (struct odp_flow __user *)argp);
1603 err = do_flowvec_ioctl(dp, argp, query_flows);
1607 err = do_flowvec_ioctl(dp, argp, list_flows);
1611 err = do_execute(dp, (struct odp_execute __user *)argp);
1618 mutex_unlock(&dp->mutex);
1623 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1626 for (i = 0; i < DP_N_QUEUES; i++) {
1627 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1633 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1636 /* XXX is there sufficient synchronization here? */
1637 int listeners = get_listen_mask(f);
1638 int dp_idx = iminor(f->f_dentry->d_inode);
1639 struct datapath *dp = get_dp(dp_idx);
1640 struct sk_buff *skb;
1641 struct iovec __user iov;
1648 if (nbytes == 0 || !listeners)
1654 for (i = 0; i < DP_N_QUEUES; i++) {
1655 if (listeners & (1 << i)) {
1656 skb = skb_dequeue(&dp->queues[i]);
1662 if (f->f_flags & O_NONBLOCK) {
1667 wait_event_interruptible(dp->waitqueue,
1668 dp_has_packet_of_interest(dp,
1671 if (signal_pending(current)) {
1672 retval = -ERESTARTSYS;
1677 copy_bytes = min_t(size_t, skb->len, nbytes);
1679 iov.iov_len = copy_bytes;
1680 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1682 retval = copy_bytes;
1689 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1691 /* XXX is there sufficient synchronization here? */
1692 int dp_idx = iminor(file->f_dentry->d_inode);
1693 struct datapath *dp = get_dp(dp_idx);
1698 poll_wait(file, &dp->waitqueue, wait);
1699 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1700 mask |= POLLIN | POLLRDNORM;
1702 mask = POLLIN | POLLRDNORM | POLLHUP;
1707 struct file_operations openvswitch_fops = {
1708 /* XXX .aio_read = openvswitch_aio_read, */
1709 .read = openvswitch_read,
1710 .poll = openvswitch_poll,
1711 .unlocked_ioctl = openvswitch_ioctl,
1712 /* XXX .fasync = openvswitch_fasync, */
1717 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)
1718 static struct llc_sap *dp_stp_sap;
1720 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1721 struct packet_type *pt, struct net_device *orig_dev)
1723 /* We don't really care about STP packets, we just listen for them for
1724 * mutual exclusion with the bridge module, so this just discards
1730 static int dp_avoid_bridge_init(void)
1732 /* Register to receive STP packets because the bridge module also
1733 * attempts to do so. Since there can only be a single listener for a
1734 * given protocol, this provides mutual exclusion against the bridge
1735 * module, preventing both of them from being loaded at the same
1737 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1739 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1745 static void dp_avoid_bridge_exit(void)
1747 llc_sap_put(dp_stp_sap);
1749 #else /* Linux 2.6.27 or later. */
1750 static int dp_avoid_bridge_init(void)
1752 /* Linux 2.6.27 introduces a way for multiple clients to register for
1753 * STP packets, which interferes with what we try to do above.
1754 * Instead, just check whether there's a bridge hook defined. This is
1755 * not as safe--the bridge module is willing to load over the top of
1756 * us--but it provides a little bit of protection. */
1757 if (br_handle_frame_hook) {
1758 printk(KERN_ERR "openvswitch: bridge module is loaded, cannot load over it\n");
1764 static void dp_avoid_bridge_exit(void)
1766 /* Nothing to do. */
1768 #endif /* Linux 2.6.27 or later */
1770 static int __init dp_init(void)
1774 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1776 err = dp_avoid_bridge_init();
1784 err = register_netdevice_notifier(&dp_device_notifier);
1786 goto error_flow_exit;
1788 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1790 goto error_unreg_notifier;
1792 /* Hook into callback used by the bridge to intercept packets.
1793 * Parasites we are. */
1794 br_handle_frame_hook = dp_frame_hook;
1798 error_unreg_notifier:
1799 unregister_netdevice_notifier(&dp_device_notifier);
1806 static void dp_cleanup(void)
1809 unregister_chrdev(major, "openvswitch");
1810 unregister_netdevice_notifier(&dp_device_notifier);
1812 br_handle_frame_hook = NULL;
1813 dp_avoid_bridge_exit();
1816 module_init(dp_init);
1817 module_exit(dp_cleanup);
1819 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1820 MODULE_LICENSE("GPL");