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 int (*dp_add_dp_hook)(struct datapath *dp);
59 EXPORT_SYMBOL(dp_add_dp_hook);
61 int (*dp_del_dp_hook)(struct datapath *dp);
62 EXPORT_SYMBOL(dp_del_dp_hook);
64 int (*dp_add_if_hook)(struct net_bridge_port *p);
65 EXPORT_SYMBOL(dp_add_if_hook);
67 int (*dp_del_if_hook)(struct net_bridge_port *p);
68 EXPORT_SYMBOL(dp_del_if_hook);
70 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
71 * by dp_mutex. dp_mutex is almost completely redundant with genl_mutex
72 * maintained by the Generic Netlink code, but the timeout path needs mutual
75 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
78 * It is safe to access the datapath and net_bridge_port structures with just
81 static struct datapath *dps[ODP_MAX];
82 static DEFINE_MUTEX(dp_mutex);
84 /* Number of milliseconds between runs of the maintenance thread. */
85 #define MAINT_SLEEP_MSECS 1000
87 static int new_nbp(struct datapath *, struct net_device *, int port_no);
89 /* Must be called with rcu_read_lock or dp_mutex. */
90 struct datapath *get_dp(int dp_idx)
92 if (dp_idx < 0 || dp_idx >= ODP_MAX)
94 return rcu_dereference(dps[dp_idx]);
96 EXPORT_SYMBOL_GPL(get_dp);
98 struct datapath *get_dp_locked(int dp_idx)
102 mutex_lock(&dp_mutex);
105 mutex_lock(&dp->mutex);
106 mutex_unlock(&dp_mutex);
110 static inline size_t br_nlmsg_size(void)
112 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
113 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
114 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
115 + nla_total_size(4) /* IFLA_MASTER */
116 + nla_total_size(4) /* IFLA_MTU */
117 + nla_total_size(4) /* IFLA_LINK */
118 + nla_total_size(1); /* IFLA_OPERSTATE */
121 static int dp_fill_ifinfo(struct sk_buff *skb,
122 const struct net_bridge_port *port,
123 int event, unsigned int flags)
125 const struct datapath *dp = port->dp;
126 const struct net_device *dev = port->dev;
127 struct ifinfomsg *hdr;
128 struct nlmsghdr *nlh;
130 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
134 hdr = nlmsg_data(nlh);
135 hdr->ifi_family = AF_BRIDGE;
137 hdr->ifi_type = dev->type;
138 hdr->ifi_index = dev->ifindex;
139 hdr->ifi_flags = dev_get_flags(dev);
142 NLA_PUT_STRING(skb, IFLA_IFNAME, dev->name);
143 NLA_PUT_U32(skb, IFLA_MASTER, dp->ports[ODPP_LOCAL]->dev->ifindex);
144 NLA_PUT_U32(skb, IFLA_MTU, dev->mtu);
145 #ifdef IFLA_OPERSTATE
146 NLA_PUT_U8(skb, IFLA_OPERSTATE,
147 netif_running(dev) ? dev->operstate : IF_OPER_DOWN);
151 NLA_PUT(skb, IFLA_ADDRESS, dev->addr_len, dev->dev_addr);
153 if (dev->ifindex != dev->iflink)
154 NLA_PUT_U32(skb, IFLA_LINK, dev->iflink);
156 return nlmsg_end(skb, nlh);
159 nlmsg_cancel(skb, nlh);
163 static void dp_ifinfo_notify(int event, struct net_bridge_port *port)
165 struct net *net = dev_net(port->dev);
169 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
173 err = dp_fill_ifinfo(skb, port, event, 0);
175 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
176 WARN_ON(err == -EMSGSIZE);
180 err = rtnl_notify(skb, net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
183 rtnl_set_sk_err(net, RTNLGRP_LINK, err);
186 static int create_dp(int dp_idx, const char __user *devnamep)
188 struct net_device *dp_dev;
189 char devname[IFNAMSIZ];
196 if (strncpy_from_user(devname, devnamep, IFNAMSIZ - 1) < 0)
198 devname[IFNAMSIZ - 1] = '\0';
200 snprintf(devname, sizeof devname, "of%d", dp_idx);
204 mutex_lock(&dp_mutex);
206 if (!try_module_get(THIS_MODULE))
209 /* Exit early if a datapath with that number already exists.
210 * (We don't use -EEXIST because that's ambiguous with 'devname'
211 * conflicting with an existing network device name.) */
217 dp = kzalloc(sizeof *dp, GFP_KERNEL);
221 mutex_init(&dp->mutex);
223 for (i = 0; i < DP_N_QUEUES; i++)
224 skb_queue_head_init(&dp->queues[i]);
225 init_waitqueue_head(&dp->waitqueue);
227 /* Setup our datapath device */
228 dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL);
229 err = PTR_ERR(dp_dev);
234 rcu_assign_pointer(dp->table, dp_table_create(DP_L1_SIZE));
236 goto err_destroy_dp_dev;
237 INIT_LIST_HEAD(&dp->port_list);
239 err = new_nbp(dp, dp_dev, ODPP_LOCAL);
241 goto err_destroy_table;
244 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
245 if (!dp->stats_percpu)
246 goto err_destroy_local_port;
248 rcu_assign_pointer(dps[dp_idx], dp);
249 mutex_unlock(&dp_mutex);
257 err_destroy_local_port:
258 dp_del_port(dp->ports[ODPP_LOCAL], NULL);
260 dp_table_destroy(dp->table, 0);
262 dp_dev_destroy(dp_dev);
266 module_put(THIS_MODULE);
268 mutex_unlock(&dp_mutex);
274 static void do_destroy_dp(struct datapath *dp, struct list_head *dp_devs)
276 struct net_bridge_port *p, *n;
282 /* Drop references to DP. */
283 list_for_each_entry_safe (p, n, &dp->port_list, node)
284 dp_del_port(p, dp_devs);
286 rcu_assign_pointer(dps[dp->dp_idx], NULL);
288 /* Wait until no longer in use, then destroy it. */
290 dp_table_destroy(dp->table, 1);
291 for (i = 0; i < DP_N_QUEUES; i++)
292 skb_queue_purge(&dp->queues[i]);
293 for (i = 0; i < DP_MAX_GROUPS; i++)
294 kfree(dp->groups[i]);
295 free_percpu(dp->stats_percpu);
297 module_put(THIS_MODULE);
300 static int destroy_dp(int dp_idx)
302 struct dp_dev *dp_dev, *next;
308 mutex_lock(&dp_mutex);
314 do_destroy_dp(dp, &dp_devs);
318 mutex_unlock(&dp_mutex);
320 list_for_each_entry_safe (dp_dev, next, &dp_devs, list)
321 free_netdev(dp_dev->dev);
325 /* Called with RTNL lock and dp_mutex. */
326 static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no)
328 struct net_bridge_port *p;
330 if (dev->br_port != NULL)
333 p = kzalloc(sizeof(*p), GFP_KERNEL);
337 dev_set_promiscuity(dev, 1);
339 p->port_no = port_no;
343 rcu_assign_pointer(dev->br_port, p);
345 /* It would make sense to assign dev->br_port here too, but
346 * that causes packets received on internal ports to get caught
347 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
348 * back to network stack, but that's a waste of time. */
350 rcu_assign_pointer(dp->ports[port_no], p);
351 list_add_rcu(&p->node, &dp->port_list);
354 dp_ifinfo_notify(RTM_NEWLINK, p);
359 static int add_port(int dp_idx, struct odp_port __user *portp)
361 struct net_device *dev;
363 struct odp_port port;
368 if (copy_from_user(&port, portp, sizeof port))
370 port.devname[IFNAMSIZ - 1] = '\0';
374 if (port_no < 0 || port_no >= DP_MAX_PORTS)
378 dp = get_dp_locked(dp_idx);
381 goto out_unlock_rtnl;
384 if (dp->ports[port_no])
387 if (!(port.flags & ODP_PORT_INTERNAL)) {
389 dev = dev_get_by_name(&init_net, port.devname);
394 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
398 dev = dp_dev_create(dp, port.devname, port_no);
405 err = new_nbp(dp, dev, port_no);
410 dp_add_if_hook(dp->ports[port_no]);
415 mutex_unlock(&dp->mutex);
422 int dp_del_port(struct net_bridge_port *p, struct list_head *dp_devs)
427 if (p->port_no != ODPP_LOCAL && dp_del_if_hook)
428 sysfs_remove_link(&p->dp->ifobj, p->dev->name);
430 dp_ifinfo_notify(RTM_DELLINK, p);
434 if (is_dp_dev(p->dev)) {
435 /* Make sure that no packets arrive from now on, since
436 * dp_dev_xmit() will try to find itself through
437 * p->dp->ports[], and we're about to set that to null. */
438 netif_tx_disable(p->dev);
441 /* First drop references to device. */
442 dev_set_promiscuity(p->dev, -1);
443 list_del_rcu(&p->node);
444 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
445 rcu_assign_pointer(p->dev->br_port, NULL);
447 /* Then wait until no one is still using it, and destroy it. */
450 if (is_dp_dev(p->dev)) {
451 dp_dev_destroy(p->dev);
453 struct dp_dev *dp_dev = dp_dev_priv(p->dev);
454 list_add(&dp_dev->list, dp_devs);
457 if (p->port_no != ODPP_LOCAL && dp_del_if_hook) {
467 static int del_port(int dp_idx, int port_no)
469 struct dp_dev *dp_dev, *next;
470 struct net_bridge_port *p;
476 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
480 dp = get_dp_locked(dp_idx);
483 goto out_unlock_rtnl;
485 p = dp->ports[port_no];
490 err = dp_del_port(p, &dp_devs);
493 mutex_unlock(&dp->mutex);
497 list_for_each_entry_safe (dp_dev, next, &dp_devs, list)
498 free_netdev(dp_dev->dev);
502 /* Must be called with rcu_read_lock. */
504 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
506 /* Make our own copy of the packet. Otherwise we will mangle the
507 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
508 * (No one comes after us, since we tell handle_bridge() that we took
510 skb = skb_share_check(skb, GFP_ATOMIC);
514 /* Push the Ethernet header back on. */
515 skb_push(skb, ETH_HLEN);
516 skb_reset_mac_header(skb);
517 dp_process_received_packet(skb, p);
520 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
521 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
523 struct datapath *dp = p->dp;
524 struct dp_stats_percpu *stats;
525 struct odp_flow_key key;
526 struct sw_flow *flow;
528 WARN_ON_ONCE(skb_shared(skb));
529 WARN_ON_ONCE(skb->destructor);
531 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
532 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
534 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
535 if (dp->drop_frags) {
542 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
544 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
545 flow_used(flow, skb);
546 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
551 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
556 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
557 * different set of devices!)
559 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
560 /* Called with rcu_read_lock and bottom-halves disabled. */
561 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
564 do_port_input(p, skb);
567 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
568 /* Called with rcu_read_lock and bottom-halves disabled. */
569 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
571 do_port_input(p, *pskb);
579 /* This code is copied verbatim from net/dev/core.c in Xen's
580 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
581 * directly because they aren't exported. */
582 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
584 if (ptr < (void *)skb->tail)
586 if (__pskb_pull_tail(skb,
587 ptr - (void *)skb->data - skb_headlen(skb))) {
594 int skb_checksum_setup(struct sk_buff *skb)
596 if (skb->proto_csum_blank) {
597 if (skb->protocol != htons(ETH_P_IP))
599 if (!skb_pull_up_to(skb, skb->nh.iph + 1))
601 skb->h.raw = (unsigned char *)skb->nh.iph + 4*skb->nh.iph->ihl;
602 switch (skb->nh.iph->protocol) {
604 skb->csum = offsetof(struct tcphdr, check);
607 skb->csum = offsetof(struct udphdr, check);
611 printk(KERN_ERR "Attempting to checksum a non-"
612 "TCP/UDP packet, dropping a protocol"
613 " %d packet", skb->nh.iph->protocol);
616 if (!skb_pull_up_to(skb, skb->h.raw + skb->csum + 2))
618 skb->ip_summed = CHECKSUM_HW;
619 skb->proto_csum_blank = 0;
628 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
631 struct dp_stats_percpu *stats;
632 struct sk_buff_head *queue;
636 WARN_ON_ONCE(skb_shared(skb));
637 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR);
639 queue = &dp->queues[queue_no];
641 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
644 /* If a checksum-deferred packet is forwarded to the controller,
645 * correct the pointers and checksum. This happens on a regular basis
646 * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets
647 * that do not have their checksum computed. We also implement it for
648 * the non-Xen case, but it is difficult to trigger or test this case
649 * there, hence the WARN_ON_ONCE().
651 err = skb_checksum_setup(skb);
655 if (skb->ip_summed == CHECKSUM_PARTIAL) {
657 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
658 /* Until 2.6.22, the start of the transport header was also the
659 * start of data to be checksummed. Linux 2.6.22 introduced
660 * the csum_start field for this purpose, but we should point
661 * the transport header to it anyway for backward
662 * compatibility, as dev_queue_xmit() does even in 2.6.28. */
663 skb_set_transport_header(skb, skb->csum_start -
666 err = skb_checksum_help(skb);
671 if (skb->ip_summed == CHECKSUM_HW) {
672 err = skb_checksum_help(skb, 0);
678 /* Break apart GSO packets into their component pieces. Otherwise
679 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
680 if (skb_is_gso(skb)) {
681 struct sk_buff *nskb = skb_gso_segment(skb, 0);
685 if (unlikely(IS_ERR(skb))) {
690 /* XXX This case might not be possible. It's hard to
691 * tell from the skb_gso_segment() code and comment. */
695 /* Figure out port number. */
696 port_no = ODPP_LOCAL;
698 if (skb->dev->br_port)
699 port_no = skb->dev->br_port->port_no;
700 else if (is_dp_dev(skb->dev))
701 port_no = dp_dev_priv(skb->dev)->port_no;
704 /* Append each packet to queue. There will be only one packet unless
705 * we broke up a GSO packet above. */
707 struct odp_msg *header;
708 struct sk_buff *nskb = skb->next;
711 err = skb_cow(skb, sizeof *header);
721 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
722 header->type = queue_no;
723 header->length = skb->len;
724 header->port = port_no;
725 header->reserved = 0;
727 skb_queue_tail(queue, skb);
732 wake_up_interruptible(&dp->waitqueue);
738 stats = percpu_ptr(dp->stats_percpu, get_cpu());
745 static int flush_flows(struct datapath *dp)
748 return dp_table_flush(dp);
751 static int validate_actions(const struct sw_flow_actions *actions)
755 for (i = 0; i < actions->n_actions; i++) {
756 const union odp_action *a = &actions->actions[i];
759 if (a->output.port >= DP_MAX_PORTS)
763 case ODPAT_OUTPUT_GROUP:
764 if (a->output_group.group >= DP_MAX_GROUPS)
768 case ODPAT_SET_VLAN_VID:
769 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
773 case ODPAT_SET_VLAN_PCP:
774 if (a->vlan_pcp.vlan_pcp & ~VLAN_PCP_MASK)
779 if (a->type >= ODPAT_N_ACTIONS)
788 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
790 struct sw_flow_actions *actions;
793 actions = flow_actions_alloc(flow->n_actions);
794 error = PTR_ERR(actions);
799 if (copy_from_user(actions->actions, flow->actions,
800 flow->n_actions * sizeof(union odp_action)))
801 goto error_free_actions;
802 error = validate_actions(actions);
804 goto error_free_actions;
811 return ERR_PTR(error);
814 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
816 if (flow->used.tv_sec) {
817 stats->used_sec = flow->used.tv_sec;
818 stats->used_nsec = flow->used.tv_nsec;
821 stats->used_nsec = 0;
823 stats->n_packets = flow->packet_count;
824 stats->n_bytes = flow->byte_count;
825 stats->ip_tos = flow->ip_tos;
826 stats->tcp_flags = flow->tcp_flags;
829 static void clear_stats(struct sw_flow *flow)
831 flow->used.tv_sec = flow->used.tv_nsec = 0;
834 flow->packet_count = 0;
835 flow->byte_count = 0;
838 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
840 struct odp_flow_put uf;
841 struct sw_flow *flow, **bucket;
842 struct dp_table *table;
843 struct odp_flow_stats stats;
847 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
849 uf.flow.key.reserved = 0;
852 table = rcu_dereference(dp->table);
853 bucket = dp_table_lookup_for_insert(table, &uf.flow.key);
855 /* No such flow, and the slots where it could go are full. */
856 error = uf.flags & ODPPF_CREATE ? -EXFULL : -ENOENT;
858 } else if (!*bucket) {
859 /* No such flow, but we found an available slot for it. */
860 struct sw_flow_actions *acts;
863 if (!(uf.flags & ODPPF_CREATE))
866 /* Expand table, if necessary, to make room. */
867 if (dp->n_flows * 4 >= table->n_buckets &&
868 table->n_buckets < DP_MAX_BUCKETS) {
869 error = dp_table_expand(dp);
873 /* The bucket's location has changed. Try again. */
879 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
882 flow->key = uf.flow.key;
883 spin_lock_init(&flow->lock);
886 /* Obtain actions. */
887 acts = get_actions(&uf.flow);
888 error = PTR_ERR(acts);
890 goto error_free_flow;
891 rcu_assign_pointer(flow->sf_acts, acts);
893 /* Put flow in bucket. */
894 rcu_assign_pointer(*bucket, flow);
896 memset(&stats, 0, sizeof(struct odp_flow_stats));
898 /* We found a matching flow. */
899 struct sw_flow *flow = *rcu_dereference(bucket);
900 struct sw_flow_actions *old_acts, *new_acts;
901 unsigned long int flags;
903 /* Bail out if we're not allowed to modify an existing flow. */
905 if (!(uf.flags & ODPPF_MODIFY))
909 new_acts = get_actions(&uf.flow);
910 error = PTR_ERR(new_acts);
911 if (IS_ERR(new_acts))
913 old_acts = rcu_dereference(flow->sf_acts);
914 if (old_acts->n_actions != new_acts->n_actions ||
915 memcmp(old_acts->actions, new_acts->actions,
916 sizeof(union odp_action) * old_acts->n_actions)) {
917 rcu_assign_pointer(flow->sf_acts, new_acts);
918 flow_deferred_free_acts(old_acts);
923 /* Fetch stats, then clear them if necessary. */
924 spin_lock_irqsave(&flow->lock, flags);
925 get_stats(flow, &stats);
926 if (uf.flags & ODPPF_ZERO_STATS)
928 spin_unlock_irqrestore(&flow->lock, flags);
931 /* Copy stats to userspace. */
932 if (__copy_to_user(&ufp->flow.stats, &stats,
933 sizeof(struct odp_flow_stats)))
938 kmem_cache_free(flow_cache, flow);
943 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
945 union odp_action __user *actions;
946 struct sw_flow_actions *sf_acts;
949 if (__get_user(actions, &ufp->actions) ||
950 __get_user(n_actions, &ufp->n_actions))
955 if (ufp->n_actions > INT_MAX / sizeof(union odp_action))
958 sf_acts = rcu_dereference(flow->sf_acts);
959 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
960 (actions && copy_to_user(actions, sf_acts->actions,
961 sizeof(union odp_action) *
962 min(sf_acts->n_actions, n_actions))))
968 static int answer_query(struct sw_flow *flow, struct odp_flow __user *ufp)
970 struct odp_flow_stats stats;
971 unsigned long int flags;
973 spin_lock_irqsave(&flow->lock, flags);
974 get_stats(flow, &stats);
975 spin_unlock_irqrestore(&flow->lock, flags);
977 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
979 return put_actions(flow, ufp);
982 static int del_or_query_flow(struct datapath *dp,
983 struct odp_flow __user *ufp,
986 struct dp_table *table = rcu_dereference(dp->table);
988 struct sw_flow *flow;
992 if (copy_from_user(&uf, ufp, sizeof uf))
996 flow = dp_table_lookup(table, &uf.key);
1001 if (cmd == ODP_FLOW_DEL) {
1002 /* XXX redundant lookup */
1003 error = dp_table_delete(table, flow);
1007 /* XXX These statistics might lose a few packets, since other
1008 * CPUs can be using this flow. We used to synchronize_rcu()
1009 * to make sure that we get completely accurate stats, but that
1010 * blows our performance, badly. */
1012 error = answer_query(flow, ufp);
1013 flow_deferred_free(flow);
1015 error = answer_query(flow, ufp);
1022 static int query_multiple_flows(struct datapath *dp,
1023 const struct odp_flowvec *flowvec)
1025 struct dp_table *table = rcu_dereference(dp->table);
1027 for (i = 0; i < flowvec->n_flows; i++) {
1028 struct __user odp_flow *ufp = &flowvec->flows[i];
1030 struct sw_flow *flow;
1033 if (__copy_from_user(&uf, ufp, sizeof uf))
1035 uf.key.reserved = 0;
1037 flow = dp_table_lookup(table, &uf.key);
1039 error = __clear_user(&ufp->stats, sizeof ufp->stats);
1041 error = answer_query(flow, ufp);
1045 return flowvec->n_flows;
1048 struct list_flows_cbdata {
1049 struct odp_flow __user *uflows;
1054 static int list_flow(struct sw_flow *flow, void *cbdata_)
1056 struct list_flows_cbdata *cbdata = cbdata_;
1057 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1060 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1062 error = answer_query(flow, ufp);
1066 if (cbdata->listed_flows >= cbdata->n_flows)
1067 return cbdata->listed_flows;
1071 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1073 struct list_flows_cbdata cbdata;
1076 if (!flowvec->n_flows)
1079 cbdata.uflows = flowvec->flows;
1080 cbdata.n_flows = flowvec->n_flows;
1081 cbdata.listed_flows = 0;
1082 error = dp_table_foreach(rcu_dereference(dp->table),
1083 list_flow, &cbdata);
1084 return error ? error : cbdata.listed_flows;
1087 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1088 int (*function)(struct datapath *,
1089 const struct odp_flowvec *))
1091 struct odp_flowvec __user *uflowvec;
1092 struct odp_flowvec flowvec;
1095 uflowvec = (struct odp_flowvec __user *)argp;
1096 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1097 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1100 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1103 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1104 flowvec.n_flows * sizeof(struct odp_flow)))
1107 retval = function(dp, &flowvec);
1108 return (retval < 0 ? retval
1109 : retval == flowvec.n_flows ? 0
1110 : __put_user(retval, &uflowvec->n_flows));
1113 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1115 struct odp_execute execute;
1116 struct odp_flow_key key;
1117 struct sk_buff *skb;
1118 struct sw_flow_actions *actions;
1122 if (copy_from_user(&execute, executep, sizeof execute))
1126 if (execute.length < ETH_HLEN || execute.length > 65535)
1130 actions = flow_actions_alloc(execute.n_actions);
1135 if (copy_from_user(actions->actions, execute.actions,
1136 execute.n_actions * sizeof *execute.actions))
1137 goto error_free_actions;
1139 err = validate_actions(actions);
1141 goto error_free_actions;
1144 skb = alloc_skb(execute.length, GFP_KERNEL);
1146 goto error_free_actions;
1147 if (execute.in_port < DP_MAX_PORTS) {
1148 struct net_bridge_port *p = dp->ports[execute.in_port];
1154 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1156 goto error_free_skb;
1158 flow_extract(skb, execute.in_port, &key);
1159 err = execute_actions(dp, skb, &key, actions->actions,
1160 actions->n_actions, GFP_KERNEL);
1173 get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1175 struct odp_stats stats;
1178 stats.n_flows = dp->n_flows;
1179 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets * 2;
1180 stats.max_capacity = DP_MAX_BUCKETS * 2;
1181 stats.n_ports = dp->n_ports;
1182 stats.max_ports = DP_MAX_PORTS;
1183 stats.max_groups = DP_MAX_GROUPS;
1184 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1185 for_each_possible_cpu(i) {
1186 const struct dp_stats_percpu *s;
1187 s = percpu_ptr(dp->stats_percpu, i);
1188 stats.n_frags += s->n_frags;
1189 stats.n_hit += s->n_hit;
1190 stats.n_missed += s->n_missed;
1191 stats.n_lost += s->n_lost;
1193 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1194 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1195 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1199 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1202 memset(&op, 0, sizeof op);
1203 strncpy(op.devname, p->dev->name, sizeof op.devname);
1204 op.port = p->port_no;
1205 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1206 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1210 query_port(struct datapath *dp, struct odp_port __user *uport)
1212 struct odp_port port;
1214 if (copy_from_user(&port, uport, sizeof port))
1216 if (port.devname[0]) {
1217 struct net_bridge_port *p;
1218 struct net_device *dev;
1221 port.devname[IFNAMSIZ - 1] = '\0';
1223 dev = dev_get_by_name(&init_net, port.devname);
1228 if (!p && is_dp_dev(dev)) {
1229 struct dp_dev *dp_dev = dp_dev_priv(dev);
1230 if (dp_dev->dp == dp)
1231 p = dp->ports[dp_dev->port_no];
1233 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1238 if (port.port >= DP_MAX_PORTS)
1240 if (!dp->ports[port.port])
1242 return put_port(dp->ports[port.port], uport);
1247 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1249 struct odp_portvec pv;
1250 struct net_bridge_port *p;
1253 if (copy_from_user(&pv, pvp, sizeof pv))
1258 list_for_each_entry_rcu (p, &dp->port_list, node) {
1259 if (put_port(p, &pv.ports[idx]))
1261 if (idx++ >= pv.n_ports)
1265 return put_user(idx, &pvp->n_ports);
1268 /* RCU callback for freeing a dp_port_group */
1269 static void free_port_group(struct rcu_head *rcu)
1271 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1276 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1278 struct odp_port_group pg;
1279 struct dp_port_group *new_group, *old_group;
1283 if (copy_from_user(&pg, upg, sizeof pg))
1287 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1291 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1296 new_group->n_ports = pg.n_ports;
1298 if (copy_from_user(new_group->ports, pg.ports,
1299 sizeof(u16) * pg.n_ports))
1302 old_group = rcu_dereference(dp->groups[pg.group]);
1303 rcu_assign_pointer(dp->groups[pg.group], new_group);
1305 call_rcu(&old_group->rcu, free_port_group);
1315 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1317 struct odp_port_group pg;
1318 struct dp_port_group *g;
1321 if (copy_from_user(&pg, upg, sizeof pg))
1324 if (pg.group >= DP_MAX_GROUPS)
1327 g = dp->groups[pg.group];
1328 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1329 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1332 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1338 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1341 int dp_idx = iminor(f->f_dentry->d_inode);
1342 struct datapath *dp;
1343 int drop_frags, listeners, port_no;
1346 /* Handle commands with special locking requirements up front. */
1349 return create_dp(dp_idx, (char __user *)argp);
1351 case ODP_DP_DESTROY:
1352 return destroy_dp(dp_idx);
1355 return add_port(dp_idx, (struct odp_port __user *)argp);
1358 err = get_user(port_no, (int __user *)argp);
1361 return del_port(dp_idx, port_no);
1364 dp = get_dp_locked(dp_idx);
1370 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1373 case ODP_GET_DROP_FRAGS:
1374 err = put_user(dp->drop_frags, (int __user *)argp);
1377 case ODP_SET_DROP_FRAGS:
1378 err = get_user(drop_frags, (int __user *)argp);
1382 if (drop_frags != 0 && drop_frags != 1)
1384 dp->drop_frags = drop_frags;
1388 case ODP_GET_LISTEN_MASK:
1389 err = put_user((int)f->private_data, (int __user *)argp);
1392 case ODP_SET_LISTEN_MASK:
1393 err = get_user(listeners, (int __user *)argp);
1397 if (listeners & ~ODPL_ALL)
1400 f->private_data = (void*)listeners;
1403 case ODP_PORT_QUERY:
1404 err = query_port(dp, (struct odp_port __user *)argp);
1408 err = list_ports(dp, (struct odp_portvec __user *)argp);
1411 case ODP_PORT_GROUP_SET:
1412 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1415 case ODP_PORT_GROUP_GET:
1416 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1419 case ODP_FLOW_FLUSH:
1420 err = flush_flows(dp);
1424 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1429 err = del_or_query_flow(dp, (struct odp_flow __user *)argp,
1433 case ODP_FLOW_GET_MULTIPLE:
1434 err = do_flowvec_ioctl(dp, argp, query_multiple_flows);
1438 err = do_flowvec_ioctl(dp, argp, list_flows);
1442 err = do_execute(dp, (struct odp_execute __user *)argp);
1449 mutex_unlock(&dp->mutex);
1453 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1456 for (i = 0; i < DP_N_QUEUES; i++) {
1457 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1463 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1466 int listeners = (int) f->private_data;
1467 int dp_idx = iminor(f->f_dentry->d_inode);
1468 struct datapath *dp = get_dp(dp_idx);
1469 struct sk_buff *skb;
1470 struct iovec __user iov;
1477 if (nbytes == 0 || !listeners)
1483 for (i = 0; i < DP_N_QUEUES; i++) {
1484 if (listeners & (1 << i)) {
1485 skb = skb_dequeue(&dp->queues[i]);
1491 if (f->f_flags & O_NONBLOCK) {
1496 wait_event_interruptible(dp->waitqueue,
1497 dp_has_packet_of_interest(dp,
1500 if (signal_pending(current)) {
1501 retval = -ERESTARTSYS;
1506 copy_bytes = min(skb->len, nbytes);
1508 iov.iov_len = copy_bytes;
1509 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1511 retval = copy_bytes;
1518 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1520 int dp_idx = iminor(file->f_dentry->d_inode);
1521 struct datapath *dp = get_dp(dp_idx);
1526 poll_wait(file, &dp->waitqueue, wait);
1527 if (dp_has_packet_of_interest(dp, (int)file->private_data))
1528 mask |= POLLIN | POLLRDNORM;
1530 mask = POLLIN | POLLRDNORM | POLLHUP;
1535 struct file_operations openvswitch_fops = {
1536 /* XXX .aio_read = openvswitch_aio_read, */
1537 .read = openvswitch_read,
1538 .poll = openvswitch_poll,
1539 .unlocked_ioctl = openvswitch_ioctl,
1540 /* XXX .fasync = openvswitch_fasync, */
1544 static struct llc_sap *dp_stp_sap;
1546 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1547 struct packet_type *pt, struct net_device *orig_dev)
1549 /* We don't really care about STP packets, we just listen for them for
1550 * mutual exclusion with the bridge module, so this just discards
1556 static int __init dp_init(void)
1560 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1562 /* Register to receive STP packets because the bridge module also
1563 * attempts to do so. Since there can only be a single listener for a
1564 * given protocol, this provides mutual exclusion against the bridge
1565 * module, preventing both of them from being loaded at the same
1567 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1569 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1577 err = register_netdevice_notifier(&dp_device_notifier);
1579 goto error_flow_exit;
1581 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1583 goto error_unreg_notifier;
1585 /* Hook into callback used by the bridge to intercept packets.
1586 * Parasites we are. */
1587 br_handle_frame_hook = dp_frame_hook;
1591 error_unreg_notifier:
1592 unregister_netdevice_notifier(&dp_device_notifier);
1599 static void dp_cleanup(void)
1602 unregister_chrdev(major, "openvswitch");
1603 unregister_netdevice_notifier(&dp_device_notifier);
1605 br_handle_frame_hook = NULL;
1606 llc_sap_put(dp_stp_sap);
1609 module_init(dp_init);
1610 module_exit(dp_cleanup);
1612 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1613 MODULE_LICENSE("GPL");