2 * Distributed under the terms of the GNU GPL version 2.
3 * Copyright (c) 2007, 2008, 2009 Nicira Networks.
6 /* Functions for managing the dp interface/device. */
8 #include <linux/init.h>
9 #include <linux/module.h>
11 #include <linux/if_arp.h>
12 #include <linux/if_bridge.h>
13 #include <linux/if_vlan.h>
16 #include <linux/delay.h>
17 #include <linux/time.h>
18 #include <linux/etherdevice.h>
19 #include <linux/kernel.h>
20 #include <linux/kthread.h>
21 #include <linux/llc.h>
22 #include <linux/mutex.h>
23 #include <linux/percpu.h>
24 #include <linux/rcupdate.h>
25 #include <linux/tcp.h>
26 #include <linux/udp.h>
27 #include <linux/version.h>
28 #include <linux/ethtool.h>
29 #include <linux/random.h>
30 #include <linux/wait.h>
31 #include <asm/system.h>
32 #include <asm/div64.h>
34 #include <linux/netfilter_bridge.h>
35 #include <linux/netfilter_ipv4.h>
36 #include <linux/inetdevice.h>
37 #include <linux/list.h>
38 #include <linux/rculist.h>
39 #include <linux/workqueue.h>
40 #include <linux/dmi.h>
43 #include "openvswitch/datapath-protocol.h"
52 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
53 EXPORT_SYMBOL(dp_ioctl_hook);
55 int (*dp_add_dp_hook)(struct datapath *dp);
56 EXPORT_SYMBOL(dp_add_dp_hook);
58 int (*dp_del_dp_hook)(struct datapath *dp);
59 EXPORT_SYMBOL(dp_del_dp_hook);
61 int (*dp_add_if_hook)(struct net_bridge_port *p);
62 EXPORT_SYMBOL(dp_add_if_hook);
64 int (*dp_del_if_hook)(struct net_bridge_port *p);
65 EXPORT_SYMBOL(dp_del_if_hook);
67 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
68 * by dp_mutex. dp_mutex is almost completely redundant with genl_mutex
69 * maintained by the Generic Netlink code, but the timeout path needs mutual
72 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
75 * It is safe to access the datapath and net_bridge_port structures with just
78 static struct datapath *dps[ODP_MAX];
79 static DEFINE_MUTEX(dp_mutex);
81 /* Number of milliseconds between runs of the maintenance thread. */
82 #define MAINT_SLEEP_MSECS 1000
84 static int new_nbp(struct datapath *, struct net_device *, int port_no);
86 /* Must be called with rcu_read_lock or dp_mutex. */
87 struct datapath *get_dp(int dp_idx)
89 if (dp_idx < 0 || dp_idx >= ODP_MAX)
91 return rcu_dereference(dps[dp_idx]);
93 EXPORT_SYMBOL_GPL(get_dp);
95 struct datapath *get_dp_locked(int dp_idx)
99 mutex_lock(&dp_mutex);
102 mutex_lock(&dp->mutex);
103 mutex_unlock(&dp_mutex);
107 static inline size_t br_nlmsg_size(void)
109 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
110 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
111 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
112 + nla_total_size(4) /* IFLA_MASTER */
113 + nla_total_size(4) /* IFLA_MTU */
114 + nla_total_size(4) /* IFLA_LINK */
115 + nla_total_size(1); /* IFLA_OPERSTATE */
118 static int dp_fill_ifinfo(struct sk_buff *skb,
119 const struct net_bridge_port *port,
120 int event, unsigned int flags)
122 const struct datapath *dp = port->dp;
123 const struct net_device *dev = port->dev;
124 struct ifinfomsg *hdr;
125 struct nlmsghdr *nlh;
127 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
131 hdr = nlmsg_data(nlh);
132 hdr->ifi_family = AF_BRIDGE;
134 hdr->ifi_type = dev->type;
135 hdr->ifi_index = dev->ifindex;
136 hdr->ifi_flags = dev_get_flags(dev);
139 NLA_PUT_STRING(skb, IFLA_IFNAME, dev->name);
140 NLA_PUT_U32(skb, IFLA_MASTER, dp->ports[ODPP_LOCAL]->dev->ifindex);
141 NLA_PUT_U32(skb, IFLA_MTU, dev->mtu);
142 #ifdef IFLA_OPERSTATE
143 NLA_PUT_U8(skb, IFLA_OPERSTATE,
144 netif_running(dev) ? dev->operstate : IF_OPER_DOWN);
148 NLA_PUT(skb, IFLA_ADDRESS, dev->addr_len, dev->dev_addr);
150 if (dev->ifindex != dev->iflink)
151 NLA_PUT_U32(skb, IFLA_LINK, dev->iflink);
153 return nlmsg_end(skb, nlh);
156 nlmsg_cancel(skb, nlh);
160 static void dp_ifinfo_notify(int event, struct net_bridge_port *port)
162 struct net *net = dev_net(port->dev);
166 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
170 err = dp_fill_ifinfo(skb, port, event, 0);
172 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
173 WARN_ON(err == -EMSGSIZE);
177 err = rtnl_notify(skb, net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
180 rtnl_set_sk_err(net, RTNLGRP_LINK, err);
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);
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 /* Setup our datapath device */
225 dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL);
226 err = PTR_ERR(dp_dev);
231 rcu_assign_pointer(dp->table, dp_table_create(DP_L1_SIZE));
233 goto err_destroy_dp_dev;
234 INIT_LIST_HEAD(&dp->port_list);
236 err = new_nbp(dp, dp_dev, ODPP_LOCAL);
238 goto err_destroy_table;
241 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
242 if (!dp->stats_percpu)
243 goto err_destroy_local_port;
245 rcu_assign_pointer(dps[dp_idx], dp);
246 mutex_unlock(&dp_mutex);
254 err_destroy_local_port:
255 dp_del_port(dp->ports[ODPP_LOCAL], NULL);
257 dp_table_destroy(dp->table, 0);
259 dp_dev_destroy(dp_dev);
263 module_put(THIS_MODULE);
265 mutex_unlock(&dp_mutex);
271 static void do_destroy_dp(struct datapath *dp, struct list_head *dp_devs)
273 struct net_bridge_port *p, *n;
279 /* Drop references to DP. */
280 list_for_each_entry_safe (p, n, &dp->port_list, node)
281 dp_del_port(p, dp_devs);
283 rcu_assign_pointer(dps[dp->dp_idx], NULL);
286 /* Wait until no longer in use, then destroy it. */
288 dp_table_destroy(dp->table, 1);
289 for (i = 0; i < DP_N_QUEUES; i++)
290 skb_queue_purge(&dp->queues[i]);
291 for (i = 0; i < DP_MAX_GROUPS; i++)
292 kfree(dp->groups[i]);
293 free_percpu(dp->stats_percpu);
295 module_put(THIS_MODULE);
298 static int destroy_dp(int dp_idx)
300 struct dp_dev *dp_dev, *next;
306 mutex_lock(&dp_mutex);
312 do_destroy_dp(dp, &dp_devs);
316 mutex_unlock(&dp_mutex);
318 list_for_each_entry_safe (dp_dev, next, &dp_devs, list)
319 free_netdev(dp_dev->dev);
323 /* Called with RTNL lock and dp_mutex. */
324 static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no)
326 struct net_bridge_port *p;
328 if (dev->br_port != NULL)
331 p = kzalloc(sizeof(*p), GFP_KERNEL);
335 dev_set_promiscuity(dev, 1);
337 p->port_no = port_no;
341 rcu_assign_pointer(dev->br_port, p);
343 /* It would make sense to assign dev->br_port here too, but
344 * that causes packets received on internal ports to get caught
345 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
346 * back to network stack, but that's a waste of time. */
348 rcu_assign_pointer(dp->ports[port_no], p);
349 list_add_rcu(&p->node, &dp->port_list);
352 dp_ifinfo_notify(RTM_NEWLINK, p);
357 static int add_port(int dp_idx, struct odp_port __user *portp)
359 struct net_device *dev;
361 struct odp_port port;
366 if (copy_from_user(&port, portp, sizeof port))
368 port.devname[IFNAMSIZ - 1] = '\0';
372 if (port_no < 0 || port_no >= DP_MAX_PORTS)
376 dp = get_dp_locked(dp_idx);
379 goto out_unlock_rtnl;
382 if (dp->ports[port_no])
385 if (!(port.flags & ODP_PORT_INTERNAL)) {
387 dev = dev_get_by_name(&init_net, port.devname);
392 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
396 dev = dp_dev_create(dp, port.devname, port_no);
403 err = new_nbp(dp, dev, port_no);
408 dp_add_if_hook(dp->ports[port_no]);
413 mutex_unlock(&dp->mutex);
420 int dp_del_port(struct net_bridge_port *p, struct list_head *dp_devs)
425 if (p->port_no != ODPP_LOCAL && dp_del_if_hook)
426 sysfs_remove_link(&p->dp->ifobj, p->dev->name);
428 dp_ifinfo_notify(RTM_DELLINK, p);
432 if (is_dp_dev(p->dev)) {
433 /* Make sure that no packets arrive from now on, since
434 * dp_dev_xmit() will try to find itself through
435 * p->dp->ports[], and we're about to set that to null. */
436 netif_tx_disable(p->dev);
439 /* First drop references to device. */
440 dev_set_promiscuity(p->dev, -1);
441 list_del_rcu(&p->node);
442 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
443 rcu_assign_pointer(p->dev->br_port, NULL);
445 /* Then wait until no one is still using it, and destroy it. */
448 if (is_dp_dev(p->dev)) {
449 dp_dev_destroy(p->dev);
451 struct dp_dev *dp_dev = dp_dev_priv(p->dev);
452 list_add(&dp_dev->list, dp_devs);
455 if (p->port_no != ODPP_LOCAL && dp_del_if_hook) {
465 static int del_port(int dp_idx, int port_no)
467 struct dp_dev *dp_dev, *next;
468 struct net_bridge_port *p;
474 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
478 dp = get_dp_locked(dp_idx);
481 goto out_unlock_rtnl;
483 p = dp->ports[port_no];
488 err = dp_del_port(p, &dp_devs);
491 mutex_unlock(&dp->mutex);
495 list_for_each_entry_safe (dp_dev, next, &dp_devs, list)
496 free_netdev(dp_dev->dev);
500 /* Must be called with rcu_read_lock. */
502 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
504 /* Make our own copy of the packet. Otherwise we will mangle the
505 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
506 * (No one comes after us, since we tell handle_bridge() that we took
508 skb = skb_share_check(skb, GFP_ATOMIC);
512 /* Push the Ethernet header back on. */
513 skb_push(skb, ETH_HLEN);
514 skb_reset_mac_header(skb);
515 dp_process_received_packet(skb, p);
518 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
519 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
521 struct datapath *dp = p->dp;
522 struct dp_stats_percpu *stats;
523 struct odp_flow_key key;
524 struct sw_flow *flow;
526 WARN_ON_ONCE(skb_shared(skb));
527 WARN_ON_ONCE(skb->destructor);
529 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
530 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
532 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
533 if (dp->drop_frags) {
540 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
542 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
543 flow_used(flow, skb);
544 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
549 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
554 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
555 * different set of devices!)
557 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
558 /* Called with rcu_read_lock and bottom-halves disabled. */
559 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
562 do_port_input(p, skb);
565 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
566 /* Called with rcu_read_lock and bottom-halves disabled. */
567 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
569 do_port_input(p, *pskb);
577 /* This code is copied verbatim from net/dev/core.c in Xen's
578 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
579 * directly because they aren't exported. */
580 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
582 if (ptr < (void *)skb->tail)
584 if (__pskb_pull_tail(skb,
585 ptr - (void *)skb->data - skb_headlen(skb))) {
592 int skb_checksum_setup(struct sk_buff *skb)
594 if (skb->proto_csum_blank) {
595 if (skb->protocol != htons(ETH_P_IP))
597 if (!skb_pull_up_to(skb, skb->nh.iph + 1))
599 skb->h.raw = (unsigned char *)skb->nh.iph + 4*skb->nh.iph->ihl;
600 switch (skb->nh.iph->protocol) {
602 skb->csum = offsetof(struct tcphdr, check);
605 skb->csum = offsetof(struct udphdr, check);
609 printk(KERN_ERR "Attempting to checksum a non-"
610 "TCP/UDP packet, dropping a protocol"
611 " %d packet", skb->nh.iph->protocol);
614 if (!skb_pull_up_to(skb, skb->h.raw + skb->csum + 2))
616 skb->ip_summed = CHECKSUM_HW;
617 skb->proto_csum_blank = 0;
626 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
629 struct dp_stats_percpu *stats;
630 struct sk_buff_head *queue;
634 WARN_ON_ONCE(skb_shared(skb));
635 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR);
637 queue = &dp->queues[queue_no];
639 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
642 /* If a checksum-deferred packet is forwarded to the controller,
643 * correct the pointers and checksum. This happens on a regular basis
644 * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets
645 * that do not have their checksum computed. We also implement it for
646 * the non-Xen case, but it is difficult to trigger or test this case
647 * there, hence the WARN_ON_ONCE().
649 err = skb_checksum_setup(skb);
653 if (skb->ip_summed == CHECKSUM_PARTIAL) {
655 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
656 /* Until 2.6.22, the start of the transport header was also the
657 * start of data to be checksummed. Linux 2.6.22 introduced
658 * the csum_start field for this purpose, but we should point
659 * the transport header to it anyway for backward
660 * compatibility, as dev_queue_xmit() does even in 2.6.28. */
661 skb_set_transport_header(skb, skb->csum_start -
664 err = skb_checksum_help(skb);
669 if (skb->ip_summed == CHECKSUM_HW) {
670 err = skb_checksum_help(skb, 0);
676 /* Break apart GSO packets into their component pieces. Otherwise
677 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
678 if (skb_is_gso(skb)) {
679 struct sk_buff *nskb = skb_gso_segment(skb, 0);
683 if (unlikely(IS_ERR(skb))) {
688 /* XXX This case might not be possible. It's hard to
689 * tell from the skb_gso_segment() code and comment. */
693 /* Figure out port number. */
694 port_no = ODPP_LOCAL;
696 if (skb->dev->br_port)
697 port_no = skb->dev->br_port->port_no;
698 else if (is_dp_dev(skb->dev))
699 port_no = dp_dev_priv(skb->dev)->port_no;
702 /* Append each packet to queue. There will be only one packet unless
703 * we broke up a GSO packet above. */
705 struct odp_msg *header;
706 struct sk_buff *nskb = skb->next;
709 err = skb_cow(skb, sizeof *header);
719 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
720 header->type = queue_no;
721 header->length = skb->len;
722 header->port = port_no;
723 header->reserved = 0;
725 skb_queue_tail(queue, skb);
730 wake_up_interruptible(&dp->waitqueue);
736 stats = percpu_ptr(dp->stats_percpu, get_cpu());
743 static int flush_flows(struct datapath *dp)
746 return dp_table_flush(dp);
749 static int validate_actions(const struct sw_flow_actions *actions)
753 for (i = 0; i < actions->n_actions; i++) {
754 const union odp_action *a = &actions->actions[i];
757 if (a->output.port >= DP_MAX_PORTS)
761 case ODPAT_OUTPUT_GROUP:
762 if (a->output_group.group >= DP_MAX_GROUPS)
766 case ODPAT_SET_VLAN_VID:
767 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
771 case ODPAT_SET_VLAN_PCP:
772 if (a->vlan_pcp.vlan_pcp & ~VLAN_PCP_MASK)
777 if (a->type >= ODPAT_N_ACTIONS)
786 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
788 struct sw_flow_actions *actions;
791 actions = flow_actions_alloc(flow->n_actions);
792 error = PTR_ERR(actions);
797 if (copy_from_user(actions->actions, flow->actions,
798 flow->n_actions * sizeof(union odp_action)))
799 goto error_free_actions;
800 error = validate_actions(actions);
802 goto error_free_actions;
809 return ERR_PTR(error);
812 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
814 if (flow->used.tv_sec) {
815 stats->used_sec = flow->used.tv_sec;
816 stats->used_nsec = flow->used.tv_nsec;
819 stats->used_nsec = 0;
821 stats->n_packets = flow->packet_count;
822 stats->n_bytes = flow->byte_count;
823 stats->ip_tos = flow->ip_tos;
824 stats->tcp_flags = flow->tcp_flags;
827 static void clear_stats(struct sw_flow *flow)
829 flow->used.tv_sec = flow->used.tv_nsec = 0;
832 flow->packet_count = 0;
833 flow->byte_count = 0;
836 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
838 struct odp_flow_put uf;
839 struct sw_flow *flow, **bucket;
840 struct dp_table *table;
841 struct odp_flow_stats stats;
845 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
847 uf.flow.key.reserved = 0;
850 table = rcu_dereference(dp->table);
851 bucket = dp_table_lookup_for_insert(table, &uf.flow.key);
853 /* No such flow, and the slots where it could go are full. */
854 error = uf.flags & ODPPF_CREATE ? -EXFULL : -ENOENT;
856 } else if (!*bucket) {
857 /* No such flow, but we found an available slot for it. */
858 struct sw_flow_actions *acts;
861 if (!(uf.flags & ODPPF_CREATE))
864 /* Expand table, if necessary, to make room. */
865 if (dp->n_flows * 4 >= table->n_buckets &&
866 table->n_buckets < DP_MAX_BUCKETS) {
867 error = dp_table_expand(dp);
871 /* The bucket's location has changed. Try again. */
877 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
880 flow->key = uf.flow.key;
881 spin_lock_init(&flow->lock);
884 /* Obtain actions. */
885 acts = get_actions(&uf.flow);
886 error = PTR_ERR(acts);
888 goto error_free_flow;
889 rcu_assign_pointer(flow->sf_acts, acts);
891 /* Put flow in bucket. */
892 rcu_assign_pointer(*bucket, flow);
894 memset(&stats, 0, sizeof(struct odp_flow_stats));
896 /* We found a matching flow. */
897 struct sw_flow *flow = *rcu_dereference(bucket);
898 struct sw_flow_actions *old_acts, *new_acts;
899 unsigned long int flags;
901 /* Bail out if we're not allowed to modify an existing flow. */
903 if (!(uf.flags & ODPPF_MODIFY))
907 new_acts = get_actions(&uf.flow);
908 error = PTR_ERR(new_acts);
909 if (IS_ERR(new_acts))
911 old_acts = rcu_dereference(flow->sf_acts);
912 if (old_acts->n_actions != new_acts->n_actions ||
913 memcmp(old_acts->actions, new_acts->actions,
914 sizeof(union odp_action) * old_acts->n_actions)) {
915 rcu_assign_pointer(flow->sf_acts, new_acts);
916 flow_deferred_free_acts(old_acts);
921 /* Fetch stats, then clear them if necessary. */
922 spin_lock_irqsave(&flow->lock, flags);
923 get_stats(flow, &stats);
924 if (uf.flags & ODPPF_ZERO_STATS)
926 spin_unlock_irqrestore(&flow->lock, flags);
929 /* Copy stats to userspace. */
930 if (__copy_to_user(&ufp->flow.stats, &stats,
931 sizeof(struct odp_flow_stats)))
936 kmem_cache_free(flow_cache, flow);
941 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
943 union odp_action __user *actions;
944 struct sw_flow_actions *sf_acts;
947 if (__get_user(actions, &ufp->actions) ||
948 __get_user(n_actions, &ufp->n_actions))
953 if (ufp->n_actions > INT_MAX / sizeof(union odp_action))
956 sf_acts = rcu_dereference(flow->sf_acts);
957 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
958 (actions && copy_to_user(actions, sf_acts->actions,
959 sizeof(union odp_action) *
960 min(sf_acts->n_actions, n_actions))))
966 static int answer_query(struct sw_flow *flow, struct odp_flow __user *ufp)
968 struct odp_flow_stats stats;
969 unsigned long int flags;
971 spin_lock_irqsave(&flow->lock, flags);
972 get_stats(flow, &stats);
973 spin_unlock_irqrestore(&flow->lock, flags);
975 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
977 return put_actions(flow, ufp);
980 static int del_or_query_flow(struct datapath *dp,
981 struct odp_flow __user *ufp,
984 struct dp_table *table = rcu_dereference(dp->table);
986 struct sw_flow *flow;
990 if (copy_from_user(&uf, ufp, sizeof uf))
994 flow = dp_table_lookup(table, &uf.key);
999 if (cmd == ODP_FLOW_DEL) {
1000 /* XXX redundant lookup */
1001 error = dp_table_delete(table, flow);
1005 /* XXX These statistics might lose a few packets, since other
1006 * CPUs can be using this flow. We used to synchronize_rcu()
1007 * to make sure that we get completely accurate stats, but that
1008 * blows our performance, badly. */
1010 error = answer_query(flow, ufp);
1011 flow_deferred_free(flow);
1013 error = answer_query(flow, ufp);
1020 static int query_multiple_flows(struct datapath *dp,
1021 const struct odp_flowvec *flowvec)
1023 struct dp_table *table = rcu_dereference(dp->table);
1025 for (i = 0; i < flowvec->n_flows; i++) {
1026 struct __user odp_flow *ufp = &flowvec->flows[i];
1028 struct sw_flow *flow;
1031 if (__copy_from_user(&uf, ufp, sizeof uf))
1033 uf.key.reserved = 0;
1035 flow = dp_table_lookup(table, &uf.key);
1037 error = __clear_user(&ufp->stats, sizeof ufp->stats);
1039 error = answer_query(flow, ufp);
1043 return flowvec->n_flows;
1046 struct list_flows_cbdata {
1047 struct odp_flow __user *uflows;
1052 static int list_flow(struct sw_flow *flow, void *cbdata_)
1054 struct list_flows_cbdata *cbdata = cbdata_;
1055 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1058 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1060 error = answer_query(flow, ufp);
1064 if (cbdata->listed_flows >= cbdata->n_flows)
1065 return cbdata->listed_flows;
1069 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1071 struct list_flows_cbdata cbdata;
1074 if (!flowvec->n_flows)
1077 cbdata.uflows = flowvec->flows;
1078 cbdata.n_flows = flowvec->n_flows;
1079 cbdata.listed_flows = 0;
1080 error = dp_table_foreach(rcu_dereference(dp->table),
1081 list_flow, &cbdata);
1082 return error ? error : cbdata.listed_flows;
1085 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1086 int (*function)(struct datapath *,
1087 const struct odp_flowvec *))
1089 struct odp_flowvec __user *uflowvec;
1090 struct odp_flowvec flowvec;
1093 uflowvec = (struct odp_flowvec __user *)argp;
1094 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1095 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1098 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1101 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1102 flowvec.n_flows * sizeof(struct odp_flow)))
1105 retval = function(dp, &flowvec);
1106 return (retval < 0 ? retval
1107 : retval == flowvec.n_flows ? 0
1108 : __put_user(retval, &uflowvec->n_flows));
1111 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1113 struct odp_execute execute;
1114 struct odp_flow_key key;
1115 struct sk_buff *skb;
1116 struct sw_flow_actions *actions;
1120 if (copy_from_user(&execute, executep, sizeof execute))
1124 if (execute.length < ETH_HLEN || execute.length > 65535)
1128 actions = flow_actions_alloc(execute.n_actions);
1133 if (copy_from_user(actions->actions, execute.actions,
1134 execute.n_actions * sizeof *execute.actions))
1135 goto error_free_actions;
1137 err = validate_actions(actions);
1139 goto error_free_actions;
1142 skb = alloc_skb(execute.length, GFP_KERNEL);
1144 goto error_free_actions;
1145 if (execute.in_port < DP_MAX_PORTS) {
1146 struct net_bridge_port *p = dp->ports[execute.in_port];
1152 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1154 goto error_free_skb;
1156 flow_extract(skb, execute.in_port, &key);
1157 err = execute_actions(dp, skb, &key, actions->actions,
1158 actions->n_actions, GFP_KERNEL);
1171 get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1173 struct odp_stats stats;
1176 stats.n_flows = dp->n_flows;
1177 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets * 2;
1178 stats.max_capacity = DP_MAX_BUCKETS * 2;
1179 stats.n_ports = dp->n_ports;
1180 stats.max_ports = DP_MAX_PORTS;
1181 stats.max_groups = DP_MAX_GROUPS;
1182 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1183 for_each_possible_cpu(i) {
1184 const struct dp_stats_percpu *s;
1185 s = percpu_ptr(dp->stats_percpu, i);
1186 stats.n_frags += s->n_frags;
1187 stats.n_hit += s->n_hit;
1188 stats.n_missed += s->n_missed;
1189 stats.n_lost += s->n_lost;
1191 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1192 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1193 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1197 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1200 memset(&op, 0, sizeof op);
1201 strncpy(op.devname, p->dev->name, sizeof op.devname);
1202 op.port = p->port_no;
1203 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1204 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1208 query_port(struct datapath *dp, struct odp_port __user *uport)
1210 struct odp_port port;
1212 if (copy_from_user(&port, uport, sizeof port))
1214 if (port.devname[0]) {
1215 struct net_bridge_port *p;
1216 struct net_device *dev;
1219 port.devname[IFNAMSIZ - 1] = '\0';
1221 dev = dev_get_by_name(&init_net, port.devname);
1226 if (!p && is_dp_dev(dev)) {
1227 struct dp_dev *dp_dev = dp_dev_priv(dev);
1228 if (dp_dev->dp == dp)
1229 p = dp->ports[dp_dev->port_no];
1231 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1236 if (port.port >= DP_MAX_PORTS)
1238 if (!dp->ports[port.port])
1240 return put_port(dp->ports[port.port], uport);
1245 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1247 struct odp_portvec pv;
1248 struct net_bridge_port *p;
1251 if (copy_from_user(&pv, pvp, sizeof pv))
1256 list_for_each_entry_rcu (p, &dp->port_list, node) {
1257 if (put_port(p, &pv.ports[idx]))
1259 if (idx++ >= pv.n_ports)
1263 return put_user(idx, &pvp->n_ports);
1266 /* RCU callback for freeing a dp_port_group */
1267 static void free_port_group(struct rcu_head *rcu)
1269 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1274 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1276 struct odp_port_group pg;
1277 struct dp_port_group *new_group, *old_group;
1281 if (copy_from_user(&pg, upg, sizeof pg))
1285 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1289 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1294 new_group->n_ports = pg.n_ports;
1296 if (copy_from_user(new_group->ports, pg.ports,
1297 sizeof(u16) * pg.n_ports))
1300 old_group = rcu_dereference(dp->groups[pg.group]);
1301 rcu_assign_pointer(dp->groups[pg.group], new_group);
1303 call_rcu(&old_group->rcu, free_port_group);
1313 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1315 struct odp_port_group pg;
1316 struct dp_port_group *g;
1319 if (copy_from_user(&pg, upg, sizeof pg))
1322 if (pg.group >= DP_MAX_GROUPS)
1325 g = dp->groups[pg.group];
1326 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1327 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1330 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1336 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1339 int dp_idx = iminor(f->f_dentry->d_inode);
1340 struct datapath *dp;
1341 int drop_frags, listeners, port_no;
1344 /* Handle commands with special locking requirements up front. */
1347 return create_dp(dp_idx, (char __user *)argp);
1349 case ODP_DP_DESTROY:
1350 return destroy_dp(dp_idx);
1353 return add_port(dp_idx, (struct odp_port __user *)argp);
1356 err = get_user(port_no, (int __user *)argp);
1359 return del_port(dp_idx, port_no);
1362 dp = get_dp_locked(dp_idx);
1368 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1371 case ODP_GET_DROP_FRAGS:
1372 err = put_user(dp->drop_frags, (int __user *)argp);
1375 case ODP_SET_DROP_FRAGS:
1376 err = get_user(drop_frags, (int __user *)argp);
1380 if (drop_frags != 0 && drop_frags != 1)
1382 dp->drop_frags = drop_frags;
1386 case ODP_GET_LISTEN_MASK:
1387 err = put_user((int)f->private_data, (int __user *)argp);
1390 case ODP_SET_LISTEN_MASK:
1391 err = get_user(listeners, (int __user *)argp);
1395 if (listeners & ~ODPL_ALL)
1398 f->private_data = (void*)listeners;
1401 case ODP_PORT_QUERY:
1402 err = query_port(dp, (struct odp_port __user *)argp);
1406 err = list_ports(dp, (struct odp_portvec __user *)argp);
1409 case ODP_PORT_GROUP_SET:
1410 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1413 case ODP_PORT_GROUP_GET:
1414 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1417 case ODP_FLOW_FLUSH:
1418 err = flush_flows(dp);
1422 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1427 err = del_or_query_flow(dp, (struct odp_flow __user *)argp,
1431 case ODP_FLOW_GET_MULTIPLE:
1432 err = do_flowvec_ioctl(dp, argp, query_multiple_flows);
1436 err = do_flowvec_ioctl(dp, argp, list_flows);
1440 err = do_execute(dp, (struct odp_execute __user *)argp);
1447 mutex_unlock(&dp->mutex);
1451 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1454 for (i = 0; i < DP_N_QUEUES; i++) {
1455 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1461 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1464 int listeners = (int) f->private_data;
1465 int dp_idx = iminor(f->f_dentry->d_inode);
1466 struct datapath *dp = get_dp(dp_idx);
1467 struct sk_buff *skb;
1468 struct iovec __user iov;
1475 if (nbytes == 0 || !listeners)
1481 for (i = 0; i < DP_N_QUEUES; i++) {
1482 if (listeners & (1 << i)) {
1483 skb = skb_dequeue(&dp->queues[i]);
1489 if (f->f_flags & O_NONBLOCK) {
1494 wait_event_interruptible(dp->waitqueue,
1495 dp_has_packet_of_interest(dp,
1498 if (signal_pending(current)) {
1499 retval = -ERESTARTSYS;
1504 copy_bytes = min(skb->len, nbytes);
1506 iov.iov_len = copy_bytes;
1507 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1509 retval = copy_bytes;
1516 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1518 int dp_idx = iminor(file->f_dentry->d_inode);
1519 struct datapath *dp = get_dp(dp_idx);
1524 poll_wait(file, &dp->waitqueue, wait);
1525 if (dp_has_packet_of_interest(dp, (int)file->private_data))
1526 mask |= POLLIN | POLLRDNORM;
1528 mask = POLLIN | POLLRDNORM | POLLHUP;
1533 struct file_operations openvswitch_fops = {
1534 /* XXX .aio_read = openvswitch_aio_read, */
1535 .read = openvswitch_read,
1536 .poll = openvswitch_poll,
1537 .unlocked_ioctl = openvswitch_ioctl,
1538 /* XXX .fasync = openvswitch_fasync, */
1542 static struct llc_sap *dp_stp_sap;
1544 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1545 struct packet_type *pt, struct net_device *orig_dev)
1547 /* We don't really care about STP packets, we just listen for them for
1548 * mutual exclusion with the bridge module, so this just discards
1554 static int __init dp_init(void)
1558 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1560 /* Register to receive STP packets because the bridge module also
1561 * attempts to do so. Since there can only be a single listener for a
1562 * given protocol, this provides mutual exclusion against the bridge
1563 * module, preventing both of them from being loaded at the same
1565 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1567 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1575 err = register_netdevice_notifier(&dp_device_notifier);
1577 goto error_flow_exit;
1579 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1581 goto error_unreg_notifier;
1583 /* Hook into callback used by the bridge to intercept packets.
1584 * Parasites we are. */
1585 br_handle_frame_hook = dp_frame_hook;
1589 error_unreg_notifier:
1590 unregister_netdevice_notifier(&dp_device_notifier);
1597 static void dp_cleanup(void)
1600 unregister_chrdev(major, "openvswitch");
1601 unregister_netdevice_notifier(&dp_device_notifier);
1603 br_handle_frame_hook = NULL;
1604 llc_sap_put(dp_stp_sap);
1607 module_init(dp_init);
1608 module_exit(dp_cleanup);
1610 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1611 MODULE_LICENSE("GPL");