2 * Copyright (c) 2007, 2008, 2009 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
59 * by dp_mutex. dp_mutex is almost completely redundant with genl_mutex
60 * maintained by the Generic Netlink code, but the timeout path needs mutual
63 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
66 * It is safe to access the datapath and net_bridge_port structures with just
69 static struct datapath *dps[ODP_MAX];
70 static DEFINE_MUTEX(dp_mutex);
72 /* Number of milliseconds between runs of the maintenance thread. */
73 #define MAINT_SLEEP_MSECS 1000
75 static int new_nbp(struct datapath *, struct net_device *, int port_no);
77 /* Must be called with rcu_read_lock or dp_mutex. */
78 struct datapath *get_dp(int dp_idx)
80 if (dp_idx < 0 || dp_idx >= ODP_MAX)
82 return rcu_dereference(dps[dp_idx]);
84 EXPORT_SYMBOL_GPL(get_dp);
86 struct datapath *get_dp_locked(int dp_idx)
90 mutex_lock(&dp_mutex);
93 mutex_lock(&dp->mutex);
94 mutex_unlock(&dp_mutex);
98 static inline size_t br_nlmsg_size(void)
100 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
101 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
102 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
103 + nla_total_size(4) /* IFLA_MASTER */
104 + nla_total_size(4) /* IFLA_MTU */
105 + nla_total_size(4) /* IFLA_LINK */
106 + nla_total_size(1); /* IFLA_OPERSTATE */
109 static int dp_fill_ifinfo(struct sk_buff *skb,
110 const struct net_bridge_port *port,
111 int event, unsigned int flags)
113 const struct datapath *dp = port->dp;
114 const struct net_device *dev = port->dev;
115 struct ifinfomsg *hdr;
116 struct nlmsghdr *nlh;
118 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
122 hdr = nlmsg_data(nlh);
123 hdr->ifi_family = AF_BRIDGE;
125 hdr->ifi_type = dev->type;
126 hdr->ifi_index = dev->ifindex;
127 hdr->ifi_flags = dev_get_flags(dev);
130 NLA_PUT_STRING(skb, IFLA_IFNAME, dev->name);
131 NLA_PUT_U32(skb, IFLA_MASTER, dp->ports[ODPP_LOCAL]->dev->ifindex);
132 NLA_PUT_U32(skb, IFLA_MTU, dev->mtu);
133 #ifdef IFLA_OPERSTATE
134 NLA_PUT_U8(skb, IFLA_OPERSTATE,
135 netif_running(dev) ? dev->operstate : IF_OPER_DOWN);
139 NLA_PUT(skb, IFLA_ADDRESS, dev->addr_len, dev->dev_addr);
141 if (dev->ifindex != dev->iflink)
142 NLA_PUT_U32(skb, IFLA_LINK, dev->iflink);
144 return nlmsg_end(skb, nlh);
147 nlmsg_cancel(skb, nlh);
151 static void dp_ifinfo_notify(int event, struct net_bridge_port *port)
153 struct net *net = dev_net(port->dev);
157 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
161 err = dp_fill_ifinfo(skb, port, event, 0);
163 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
164 WARN_ON(err == -EMSGSIZE);
168 rtnl_notify(skb, net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
172 rtnl_set_sk_err(net, RTNLGRP_LINK, err);
175 static void release_dp(struct kobject *kobj)
177 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
181 struct kobj_type dp_ktype = {
182 .release = release_dp
185 static int create_dp(int dp_idx, const char __user *devnamep)
187 struct net_device *dp_dev;
188 char devname[IFNAMSIZ];
195 if (strncpy_from_user(devname, devnamep, IFNAMSIZ - 1) < 0)
197 devname[IFNAMSIZ - 1] = '\0';
199 snprintf(devname, sizeof devname, "of%d", dp_idx);
203 mutex_lock(&dp_mutex);
205 if (!try_module_get(THIS_MODULE))
208 /* Exit early if a datapath with that number already exists.
209 * (We don't use -EEXIST because that's ambiguous with 'devname'
210 * conflicting with an existing network device name.) */
216 dp = kzalloc(sizeof *dp, GFP_KERNEL);
219 INIT_LIST_HEAD(&dp->port_list);
220 mutex_init(&dp->mutex);
222 for (i = 0; i < DP_N_QUEUES; i++)
223 skb_queue_head_init(&dp->queues[i]);
224 init_waitqueue_head(&dp->waitqueue);
226 /* Initialize kobject for bridge. This will be added as
227 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
228 kobject_set_name(&dp->ifobj, SYSFS_BRIDGE_PORT_SUBDIR); /* "brif" */
229 dp->ifobj.kset = NULL;
230 dp->ifobj.parent = NULL;
231 kobject_init(&dp->ifobj, &dp_ktype);
233 /* Allocate table. */
235 rcu_assign_pointer(dp->table, dp_table_create(DP_L1_SIZE));
239 /* Set up our datapath device. */
240 dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL);
241 err = PTR_ERR(dp_dev);
243 goto err_destroy_table;
245 err = new_nbp(dp, dp_dev, ODPP_LOCAL);
247 dp_dev_destroy(dp_dev);
248 goto err_destroy_table;
252 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
253 if (!dp->stats_percpu)
254 goto err_destroy_local_port;
256 rcu_assign_pointer(dps[dp_idx], dp);
257 mutex_unlock(&dp_mutex);
266 err_destroy_local_port:
267 dp_del_port(dp->ports[ODPP_LOCAL]);
269 dp_table_destroy(dp->table, 0);
273 module_put(THIS_MODULE);
275 mutex_unlock(&dp_mutex);
281 static void do_destroy_dp(struct datapath *dp)
283 struct net_bridge_port *p, *n;
286 list_for_each_entry_safe (p, n, &dp->port_list, node)
287 if (p->port_no != ODPP_LOCAL)
294 rcu_assign_pointer(dps[dp->dp_idx], NULL);
296 dp_del_port(dp->ports[ODPP_LOCAL]);
298 dp_table_destroy(dp->table, 1);
300 for (i = 0; i < DP_N_QUEUES; i++)
301 skb_queue_purge(&dp->queues[i]);
302 for (i = 0; i < DP_MAX_GROUPS; i++)
303 kfree(dp->groups[i]);
304 free_percpu(dp->stats_percpu);
305 kobject_put(&dp->ifobj);
306 module_put(THIS_MODULE);
309 static int destroy_dp(int dp_idx)
315 mutex_lock(&dp_mutex);
325 mutex_unlock(&dp_mutex);
330 static void release_nbp(struct kobject *kobj)
332 struct net_bridge_port *p = container_of(kobj, struct net_bridge_port, kobj);
336 struct kobj_type brport_ktype = {
338 .sysfs_ops = &brport_sysfs_ops,
340 .release = release_nbp
343 /* Called with RTNL lock and dp_mutex. */
344 static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no)
346 struct net_bridge_port *p;
348 if (dev->br_port != NULL)
351 p = kzalloc(sizeof(*p), GFP_KERNEL);
355 dev_set_promiscuity(dev, 1);
357 p->port_no = port_no;
361 rcu_assign_pointer(dev->br_port, p);
363 /* It would make sense to assign dev->br_port here too, but
364 * that causes packets received on internal ports to get caught
365 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
366 * back to network stack, but that's a waste of time. */
368 rcu_assign_pointer(dp->ports[port_no], p);
369 list_add_rcu(&p->node, &dp->port_list);
372 /* Initialize kobject for bridge. This will be added as
373 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
374 kobject_set_name(&p->kobj, SYSFS_BRIDGE_PORT_ATTR); /* "brport" */
376 p->kobj.parent = &p->dev->NETDEV_DEV_MEMBER.kobj;
377 kobject_init(&p->kobj, &brport_ktype);
379 dp_ifinfo_notify(RTM_NEWLINK, p);
384 static int add_port(int dp_idx, struct odp_port __user *portp)
386 struct net_device *dev;
388 struct odp_port port;
393 if (copy_from_user(&port, portp, sizeof port))
395 port.devname[IFNAMSIZ - 1] = '\0';
399 if (port_no < 0 || port_no >= DP_MAX_PORTS)
403 dp = get_dp_locked(dp_idx);
406 goto out_unlock_rtnl;
409 if (dp->ports[port_no])
412 if (!(port.flags & ODP_PORT_INTERNAL)) {
414 dev = dev_get_by_name(&init_net, port.devname);
419 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
423 dev = dp_dev_create(dp, port.devname, port_no);
430 err = new_nbp(dp, dev, port_no);
435 dp_sysfs_add_if(dp->ports[port_no]);
441 mutex_unlock(&dp->mutex);
448 int dp_del_port(struct net_bridge_port *p)
453 if (p->port_no != ODPP_LOCAL)
456 dp_ifinfo_notify(RTM_DELLINK, p);
460 if (is_dp_dev(p->dev)) {
461 /* Make sure that no packets arrive from now on, since
462 * dp_dev_xmit() will try to find itself through
463 * p->dp->ports[], and we're about to set that to null. */
464 netif_tx_disable(p->dev);
467 /* First drop references to device. */
468 dev_set_promiscuity(p->dev, -1);
469 list_del_rcu(&p->node);
470 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
471 rcu_assign_pointer(p->dev->br_port, NULL);
473 /* Then wait until no one is still using it, and destroy it. */
476 if (is_dp_dev(p->dev))
477 dp_dev_destroy(p->dev);
479 kobject_put(&p->kobj);
484 static int del_port(int dp_idx, int port_no)
486 struct net_bridge_port *p;
492 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
496 dp = get_dp_locked(dp_idx);
499 goto out_unlock_rtnl;
501 p = dp->ports[port_no];
506 err = dp_del_port(p);
509 mutex_unlock(&dp->mutex);
516 /* Must be called with rcu_read_lock. */
518 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
520 /* Make our own copy of the packet. Otherwise we will mangle the
521 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
522 * (No one comes after us, since we tell handle_bridge() that we took
524 skb = skb_share_check(skb, GFP_ATOMIC);
528 /* Push the Ethernet header back on. */
529 skb_push(skb, ETH_HLEN);
530 skb_reset_mac_header(skb);
531 dp_process_received_packet(skb, p);
534 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
535 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
537 struct datapath *dp = p->dp;
538 struct dp_stats_percpu *stats;
539 struct odp_flow_key key;
540 struct sw_flow *flow;
542 WARN_ON_ONCE(skb_shared(skb));
544 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
545 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
547 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
548 if (dp->drop_frags) {
555 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
557 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
558 flow_used(flow, skb);
559 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
564 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
569 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
570 * different set of devices!)
572 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
573 /* Called with rcu_read_lock and bottom-halves disabled. */
574 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
577 do_port_input(p, skb);
580 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
581 /* Called with rcu_read_lock and bottom-halves disabled. */
582 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
584 do_port_input(p, *pskb);
592 /* This code is copied verbatim from net/dev/core.c in Xen's
593 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
594 * directly because they aren't exported. */
595 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
597 if (ptr < (void *)skb->tail)
599 if (__pskb_pull_tail(skb,
600 ptr - (void *)skb->data - skb_headlen(skb))) {
607 int skb_checksum_setup(struct sk_buff *skb)
609 if (skb->proto_csum_blank) {
610 if (skb->protocol != htons(ETH_P_IP))
612 if (!skb_pull_up_to(skb, skb->nh.iph + 1))
614 skb->h.raw = (unsigned char *)skb->nh.iph + 4*skb->nh.iph->ihl;
615 switch (skb->nh.iph->protocol) {
617 skb->csum = offsetof(struct tcphdr, check);
620 skb->csum = offsetof(struct udphdr, check);
624 printk(KERN_ERR "Attempting to checksum a non-"
625 "TCP/UDP packet, dropping a protocol"
626 " %d packet", skb->nh.iph->protocol);
629 if (!skb_pull_up_to(skb, skb->h.raw + skb->csum + 2))
631 skb->ip_summed = CHECKSUM_HW;
632 skb->proto_csum_blank = 0;
641 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
644 struct dp_stats_percpu *stats;
645 struct sk_buff_head *queue;
649 WARN_ON_ONCE(skb_shared(skb));
650 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR);
652 queue = &dp->queues[queue_no];
654 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
657 /* If a checksum-deferred packet is forwarded to the controller,
658 * correct the pointers and checksum. This happens on a regular basis
659 * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets
660 * that do not have their checksum computed. We also implement it for
661 * the non-Xen case, but it is difficult to trigger or test this case
662 * there, hence the WARN_ON_ONCE().
664 err = skb_checksum_setup(skb);
668 if (skb->ip_summed == CHECKSUM_PARTIAL) {
670 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
671 /* Until 2.6.22, the start of the transport header was also the
672 * start of data to be checksummed. Linux 2.6.22 introduced
673 * the csum_start field for this purpose, but we should point
674 * the transport header to it anyway for backward
675 * compatibility, as dev_queue_xmit() does even in 2.6.28. */
676 skb_set_transport_header(skb, skb->csum_start -
679 err = skb_checksum_help(skb);
684 if (skb->ip_summed == CHECKSUM_HW) {
685 err = skb_checksum_help(skb, 0);
691 /* Break apart GSO packets into their component pieces. Otherwise
692 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
693 if (skb_is_gso(skb)) {
694 struct sk_buff *nskb = skb_gso_segment(skb, 0);
698 if (unlikely(IS_ERR(skb))) {
703 /* XXX This case might not be possible. It's hard to
704 * tell from the skb_gso_segment() code and comment. */
708 /* Figure out port number. */
709 port_no = ODPP_LOCAL;
711 if (skb->dev->br_port)
712 port_no = skb->dev->br_port->port_no;
713 else if (is_dp_dev(skb->dev))
714 port_no = dp_dev_priv(skb->dev)->port_no;
717 /* Append each packet to queue. There will be only one packet unless
718 * we broke up a GSO packet above. */
720 struct odp_msg *header;
721 struct sk_buff *nskb = skb->next;
724 err = skb_cow(skb, sizeof *header);
734 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
735 header->type = queue_no;
736 header->length = skb->len;
737 header->port = port_no;
738 header->reserved = 0;
740 skb_queue_tail(queue, skb);
745 wake_up_interruptible(&dp->waitqueue);
751 stats = percpu_ptr(dp->stats_percpu, get_cpu());
758 static int flush_flows(struct datapath *dp)
761 return dp_table_flush(dp);
764 static int validate_actions(const struct sw_flow_actions *actions)
768 for (i = 0; i < actions->n_actions; i++) {
769 const union odp_action *a = &actions->actions[i];
772 if (a->output.port >= DP_MAX_PORTS)
776 case ODPAT_OUTPUT_GROUP:
777 if (a->output_group.group >= DP_MAX_GROUPS)
781 case ODPAT_SET_VLAN_VID:
782 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
786 case ODPAT_SET_VLAN_PCP:
787 if (a->vlan_pcp.vlan_pcp
788 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
793 if (a->type >= ODPAT_N_ACTIONS)
802 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
804 struct sw_flow_actions *actions;
807 actions = flow_actions_alloc(flow->n_actions);
808 error = PTR_ERR(actions);
813 if (copy_from_user(actions->actions, flow->actions,
814 flow->n_actions * sizeof(union odp_action)))
815 goto error_free_actions;
816 error = validate_actions(actions);
818 goto error_free_actions;
825 return ERR_PTR(error);
828 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
830 if (flow->used.tv_sec) {
831 stats->used_sec = flow->used.tv_sec;
832 stats->used_nsec = flow->used.tv_nsec;
835 stats->used_nsec = 0;
837 stats->n_packets = flow->packet_count;
838 stats->n_bytes = flow->byte_count;
839 stats->ip_tos = flow->ip_tos;
840 stats->tcp_flags = flow->tcp_flags;
843 static void clear_stats(struct sw_flow *flow)
845 flow->used.tv_sec = flow->used.tv_nsec = 0;
848 flow->packet_count = 0;
849 flow->byte_count = 0;
852 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
854 struct odp_flow_put uf;
855 struct sw_flow *flow;
856 struct dp_table *table;
857 struct odp_flow_stats stats;
861 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
863 uf.flow.key.reserved = 0;
865 table = rcu_dereference(dp->table);
866 flow = dp_table_lookup(table, &uf.flow.key);
869 struct sw_flow_actions *acts;
872 if (!(uf.flags & ODPPF_CREATE))
875 /* Expand table, if necessary, to make room. */
876 if (dp->n_flows >= table->n_buckets) {
878 if (table->n_buckets >= DP_MAX_BUCKETS)
881 error = dp_table_expand(dp);
884 table = rcu_dereference(dp->table);
889 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
892 flow->key = uf.flow.key;
893 spin_lock_init(&flow->lock);
896 /* Obtain actions. */
897 acts = get_actions(&uf.flow);
898 error = PTR_ERR(acts);
900 goto error_free_flow;
901 rcu_assign_pointer(flow->sf_acts, acts);
903 /* Put flow in bucket. */
904 error = dp_table_insert(table, flow);
906 goto error_free_flow_acts;
908 memset(&stats, 0, sizeof(struct odp_flow_stats));
910 /* We found a matching flow. */
911 struct sw_flow_actions *old_acts, *new_acts;
912 unsigned long int flags;
914 /* Bail out if we're not allowed to modify an existing flow. */
916 if (!(uf.flags & ODPPF_MODIFY))
920 new_acts = get_actions(&uf.flow);
921 error = PTR_ERR(new_acts);
922 if (IS_ERR(new_acts))
924 old_acts = rcu_dereference(flow->sf_acts);
925 if (old_acts->n_actions != new_acts->n_actions ||
926 memcmp(old_acts->actions, new_acts->actions,
927 sizeof(union odp_action) * old_acts->n_actions)) {
928 rcu_assign_pointer(flow->sf_acts, new_acts);
929 flow_deferred_free_acts(old_acts);
934 /* Fetch stats, then clear them if necessary. */
935 spin_lock_irqsave(&flow->lock, flags);
936 get_stats(flow, &stats);
937 if (uf.flags & ODPPF_ZERO_STATS)
939 spin_unlock_irqrestore(&flow->lock, flags);
942 /* Copy stats to userspace. */
943 if (__copy_to_user(&ufp->flow.stats, &stats,
944 sizeof(struct odp_flow_stats)))
948 error_free_flow_acts:
949 kfree(flow->sf_acts);
951 kmem_cache_free(flow_cache, flow);
956 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
958 union odp_action __user *actions;
959 struct sw_flow_actions *sf_acts;
962 if (__get_user(actions, &ufp->actions) ||
963 __get_user(n_actions, &ufp->n_actions))
968 if (ufp->n_actions > INT_MAX / sizeof(union odp_action))
971 sf_acts = rcu_dereference(flow->sf_acts);
972 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
973 (actions && copy_to_user(actions, sf_acts->actions,
974 sizeof(union odp_action) *
975 min(sf_acts->n_actions, n_actions))))
981 static int answer_query(struct sw_flow *flow, struct odp_flow __user *ufp)
983 struct odp_flow_stats stats;
984 unsigned long int flags;
986 spin_lock_irqsave(&flow->lock, flags);
987 get_stats(flow, &stats);
988 spin_unlock_irqrestore(&flow->lock, flags);
990 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
992 return put_actions(flow, ufp);
995 static int del_or_query_flow(struct datapath *dp,
996 struct odp_flow __user *ufp,
999 struct dp_table *table = rcu_dereference(dp->table);
1001 struct sw_flow *flow;
1005 if (copy_from_user(&uf, ufp, sizeof uf))
1007 uf.key.reserved = 0;
1009 flow = dp_table_lookup(table, &uf.key);
1014 if (cmd == ODP_FLOW_DEL) {
1015 /* XXX redundant lookup */
1016 error = dp_table_delete(table, flow);
1020 /* XXX These statistics might lose a few packets, since other
1021 * CPUs can be using this flow. We used to synchronize_rcu()
1022 * to make sure that we get completely accurate stats, but that
1023 * blows our performance, badly. */
1025 error = answer_query(flow, ufp);
1026 flow_deferred_free(flow);
1028 error = answer_query(flow, ufp);
1035 static int query_multiple_flows(struct datapath *dp,
1036 const struct odp_flowvec *flowvec)
1038 struct dp_table *table = rcu_dereference(dp->table);
1040 for (i = 0; i < flowvec->n_flows; i++) {
1041 struct __user odp_flow *ufp = &flowvec->flows[i];
1043 struct sw_flow *flow;
1046 if (__copy_from_user(&uf, ufp, sizeof uf))
1048 uf.key.reserved = 0;
1050 flow = dp_table_lookup(table, &uf.key);
1052 error = __clear_user(&ufp->stats, sizeof ufp->stats);
1054 error = answer_query(flow, ufp);
1058 return flowvec->n_flows;
1061 struct list_flows_cbdata {
1062 struct odp_flow __user *uflows;
1067 static int list_flow(struct sw_flow *flow, void *cbdata_)
1069 struct list_flows_cbdata *cbdata = cbdata_;
1070 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1073 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1075 error = answer_query(flow, ufp);
1079 if (cbdata->listed_flows >= cbdata->n_flows)
1080 return cbdata->listed_flows;
1084 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1086 struct list_flows_cbdata cbdata;
1089 if (!flowvec->n_flows)
1092 cbdata.uflows = flowvec->flows;
1093 cbdata.n_flows = flowvec->n_flows;
1094 cbdata.listed_flows = 0;
1095 error = dp_table_foreach(rcu_dereference(dp->table),
1096 list_flow, &cbdata);
1097 return error ? error : cbdata.listed_flows;
1100 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1101 int (*function)(struct datapath *,
1102 const struct odp_flowvec *))
1104 struct odp_flowvec __user *uflowvec;
1105 struct odp_flowvec flowvec;
1108 uflowvec = (struct odp_flowvec __user *)argp;
1109 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1110 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1113 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1116 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1117 flowvec.n_flows * sizeof(struct odp_flow)))
1120 retval = function(dp, &flowvec);
1121 return (retval < 0 ? retval
1122 : retval == flowvec.n_flows ? 0
1123 : __put_user(retval, &uflowvec->n_flows));
1126 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1128 struct odp_execute execute;
1129 struct odp_flow_key key;
1130 struct sk_buff *skb;
1131 struct sw_flow_actions *actions;
1136 if (copy_from_user(&execute, executep, sizeof execute))
1140 if (execute.length < ETH_HLEN || execute.length > 65535)
1144 actions = flow_actions_alloc(execute.n_actions);
1149 if (copy_from_user(actions->actions, execute.actions,
1150 execute.n_actions * sizeof *execute.actions))
1151 goto error_free_actions;
1153 err = validate_actions(actions);
1155 goto error_free_actions;
1158 skb = alloc_skb(execute.length, GFP_KERNEL);
1160 goto error_free_actions;
1161 if (execute.in_port < DP_MAX_PORTS) {
1162 struct net_bridge_port *p = dp->ports[execute.in_port];
1168 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1170 goto error_free_skb;
1172 skb_reset_mac_header(skb);
1175 /* Normally, setting the skb 'protocol' field would be handled by a
1176 * call to eth_type_trans(), but it assumes there's a sending
1177 * device, which we may not have. */
1178 if (ntohs(eth->h_proto) >= 1536)
1179 skb->protocol = eth->h_proto;
1181 skb->protocol = htons(ETH_P_802_2);
1183 flow_extract(skb, execute.in_port, &key);
1184 err = execute_actions(dp, skb, &key, actions->actions,
1185 actions->n_actions, GFP_KERNEL);
1198 get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1200 struct odp_stats stats;
1203 stats.n_flows = dp->n_flows;
1204 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets;
1205 stats.max_capacity = DP_MAX_BUCKETS;
1206 stats.n_ports = dp->n_ports;
1207 stats.max_ports = DP_MAX_PORTS;
1208 stats.max_groups = DP_MAX_GROUPS;
1209 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1210 for_each_possible_cpu(i) {
1211 const struct dp_stats_percpu *s;
1212 s = percpu_ptr(dp->stats_percpu, i);
1213 stats.n_frags += s->n_frags;
1214 stats.n_hit += s->n_hit;
1215 stats.n_missed += s->n_missed;
1216 stats.n_lost += s->n_lost;
1218 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1219 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1220 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1223 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1224 int dp_min_mtu(const struct datapath *dp)
1226 struct net_bridge_port *p;
1231 list_for_each_entry_rcu (p, &dp->port_list, node) {
1232 struct net_device *dev = p->dev;
1234 /* Skip any internal ports, since that's what we're trying to
1239 if (!mtu || dev->mtu < mtu)
1243 return mtu ? mtu : ETH_DATA_LEN;
1247 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1250 memset(&op, 0, sizeof op);
1251 strncpy(op.devname, p->dev->name, sizeof op.devname);
1252 op.port = p->port_no;
1253 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1254 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1258 query_port(struct datapath *dp, struct odp_port __user *uport)
1260 struct odp_port port;
1262 if (copy_from_user(&port, uport, sizeof port))
1264 if (port.devname[0]) {
1265 struct net_bridge_port *p;
1266 struct net_device *dev;
1269 port.devname[IFNAMSIZ - 1] = '\0';
1271 dev = dev_get_by_name(&init_net, port.devname);
1276 if (!p && is_dp_dev(dev)) {
1277 struct dp_dev *dp_dev = dp_dev_priv(dev);
1278 if (dp_dev->dp == dp)
1279 p = dp->ports[dp_dev->port_no];
1281 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1286 if (port.port >= DP_MAX_PORTS)
1288 if (!dp->ports[port.port])
1290 return put_port(dp->ports[port.port], uport);
1295 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1297 struct odp_portvec pv;
1298 struct net_bridge_port *p;
1301 if (copy_from_user(&pv, pvp, sizeof pv))
1306 list_for_each_entry_rcu (p, &dp->port_list, node) {
1307 if (put_port(p, &pv.ports[idx]))
1309 if (idx++ >= pv.n_ports)
1313 return put_user(idx, &pvp->n_ports);
1316 /* RCU callback for freeing a dp_port_group */
1317 static void free_port_group(struct rcu_head *rcu)
1319 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1324 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1326 struct odp_port_group pg;
1327 struct dp_port_group *new_group, *old_group;
1331 if (copy_from_user(&pg, upg, sizeof pg))
1335 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1339 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1344 new_group->n_ports = pg.n_ports;
1346 if (copy_from_user(new_group->ports, pg.ports,
1347 sizeof(u16) * pg.n_ports))
1350 old_group = rcu_dereference(dp->groups[pg.group]);
1351 rcu_assign_pointer(dp->groups[pg.group], new_group);
1353 call_rcu(&old_group->rcu, free_port_group);
1363 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1365 struct odp_port_group pg;
1366 struct dp_port_group *g;
1369 if (copy_from_user(&pg, upg, sizeof pg))
1372 if (pg.group >= DP_MAX_GROUPS)
1375 g = dp->groups[pg.group];
1376 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1377 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1380 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1386 static int get_listen_mask(const struct file *f)
1388 return (long)f->private_data;
1391 static void set_listen_mask(struct file *f, int listen_mask)
1393 f->private_data = (void*)(long)listen_mask;
1396 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1399 int dp_idx = iminor(f->f_dentry->d_inode);
1400 struct datapath *dp;
1401 int drop_frags, listeners, port_no;
1404 /* Handle commands with special locking requirements up front. */
1407 return create_dp(dp_idx, (char __user *)argp);
1409 case ODP_DP_DESTROY:
1410 return destroy_dp(dp_idx);
1413 return add_port(dp_idx, (struct odp_port __user *)argp);
1416 err = get_user(port_no, (int __user *)argp);
1419 return del_port(dp_idx, port_no);
1422 dp = get_dp_locked(dp_idx);
1428 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1431 case ODP_GET_DROP_FRAGS:
1432 err = put_user(dp->drop_frags, (int __user *)argp);
1435 case ODP_SET_DROP_FRAGS:
1436 err = get_user(drop_frags, (int __user *)argp);
1440 if (drop_frags != 0 && drop_frags != 1)
1442 dp->drop_frags = drop_frags;
1446 case ODP_GET_LISTEN_MASK:
1447 err = put_user(get_listen_mask(f), (int __user *)argp);
1450 case ODP_SET_LISTEN_MASK:
1451 err = get_user(listeners, (int __user *)argp);
1455 if (listeners & ~ODPL_ALL)
1458 set_listen_mask(f, listeners);
1461 case ODP_PORT_QUERY:
1462 err = query_port(dp, (struct odp_port __user *)argp);
1466 err = list_ports(dp, (struct odp_portvec __user *)argp);
1469 case ODP_PORT_GROUP_SET:
1470 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1473 case ODP_PORT_GROUP_GET:
1474 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1477 case ODP_FLOW_FLUSH:
1478 err = flush_flows(dp);
1482 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1487 err = del_or_query_flow(dp, (struct odp_flow __user *)argp,
1491 case ODP_FLOW_GET_MULTIPLE:
1492 err = do_flowvec_ioctl(dp, argp, query_multiple_flows);
1496 err = do_flowvec_ioctl(dp, argp, list_flows);
1500 err = do_execute(dp, (struct odp_execute __user *)argp);
1507 mutex_unlock(&dp->mutex);
1511 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1514 for (i = 0; i < DP_N_QUEUES; i++) {
1515 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1521 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1524 /* XXX is there sufficient synchronization here? */
1525 int listeners = get_listen_mask(f);
1526 int dp_idx = iminor(f->f_dentry->d_inode);
1527 struct datapath *dp = get_dp(dp_idx);
1528 struct sk_buff *skb;
1529 struct iovec __user iov;
1536 if (nbytes == 0 || !listeners)
1542 for (i = 0; i < DP_N_QUEUES; i++) {
1543 if (listeners & (1 << i)) {
1544 skb = skb_dequeue(&dp->queues[i]);
1550 if (f->f_flags & O_NONBLOCK) {
1555 wait_event_interruptible(dp->waitqueue,
1556 dp_has_packet_of_interest(dp,
1559 if (signal_pending(current)) {
1560 retval = -ERESTARTSYS;
1565 copy_bytes = min_t(size_t, skb->len, nbytes);
1567 iov.iov_len = copy_bytes;
1568 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1570 retval = copy_bytes;
1577 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1579 /* XXX is there sufficient synchronization here? */
1580 int dp_idx = iminor(file->f_dentry->d_inode);
1581 struct datapath *dp = get_dp(dp_idx);
1586 poll_wait(file, &dp->waitqueue, wait);
1587 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1588 mask |= POLLIN | POLLRDNORM;
1590 mask = POLLIN | POLLRDNORM | POLLHUP;
1595 struct file_operations openvswitch_fops = {
1596 /* XXX .aio_read = openvswitch_aio_read, */
1597 .read = openvswitch_read,
1598 .poll = openvswitch_poll,
1599 .unlocked_ioctl = openvswitch_ioctl,
1600 /* XXX .fasync = openvswitch_fasync, */
1604 static struct llc_sap *dp_stp_sap;
1606 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1607 struct packet_type *pt, struct net_device *orig_dev)
1609 /* We don't really care about STP packets, we just listen for them for
1610 * mutual exclusion with the bridge module, so this just discards
1616 static int __init dp_init(void)
1620 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1622 /* Register to receive STP packets because the bridge module also
1623 * attempts to do so. Since there can only be a single listener for a
1624 * given protocol, this provides mutual exclusion against the bridge
1625 * module, preventing both of them from being loaded at the same
1627 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1629 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1637 err = register_netdevice_notifier(&dp_device_notifier);
1639 goto error_flow_exit;
1641 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1643 goto error_unreg_notifier;
1645 /* Hook into callback used by the bridge to intercept packets.
1646 * Parasites we are. */
1647 br_handle_frame_hook = dp_frame_hook;
1651 error_unreg_notifier:
1652 unregister_netdevice_notifier(&dp_device_notifier);
1659 static void dp_cleanup(void)
1662 unregister_chrdev(major, "openvswitch");
1663 unregister_netdevice_notifier(&dp_device_notifier);
1665 br_handle_frame_hook = NULL;
1666 llc_sap_put(dp_stp_sap);
1669 module_init(dp_init);
1670 module_exit(dp_cleanup);
1672 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1673 MODULE_LICENSE("GPL");