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 dp->ifobj.kset = NULL;
229 kobject_init(&dp->ifobj, &dp_ktype);
231 /* Allocate table. */
233 rcu_assign_pointer(dp->table, dp_table_create(DP_L1_SIZE));
237 /* Setup our datapath device */
238 dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL);
239 err = PTR_ERR(dp_dev);
241 goto err_destroy_table;
243 err = new_nbp(dp, dp_dev, ODPP_LOCAL);
245 dp_dev_destroy(dp_dev);
246 goto err_destroy_table;
250 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
251 if (!dp->stats_percpu)
252 goto err_destroy_local_port;
254 rcu_assign_pointer(dps[dp_idx], dp);
255 mutex_unlock(&dp_mutex);
262 err_destroy_local_port:
263 dp_del_port(dp->ports[ODPP_LOCAL]);
265 dp_table_destroy(dp->table, 0);
269 module_put(THIS_MODULE);
271 mutex_unlock(&dp_mutex);
277 static void do_destroy_dp(struct datapath *dp)
279 struct net_bridge_port *p, *n;
282 list_for_each_entry_safe (p, n, &dp->port_list, node)
283 if (p->port_no != ODPP_LOCAL)
288 rcu_assign_pointer(dps[dp->dp_idx], NULL);
290 dp_del_port(dp->ports[ODPP_LOCAL]);
292 dp_table_destroy(dp->table, 1);
294 for (i = 0; i < DP_N_QUEUES; i++)
295 skb_queue_purge(&dp->queues[i]);
296 for (i = 0; i < DP_MAX_GROUPS; i++)
297 kfree(dp->groups[i]);
298 free_percpu(dp->stats_percpu);
299 kobject_put(&dp->ifobj);
300 module_put(THIS_MODULE);
303 static int destroy_dp(int dp_idx)
309 mutex_lock(&dp_mutex);
319 mutex_unlock(&dp_mutex);
324 static void release_nbp(struct kobject *kobj)
326 struct net_bridge_port *p = container_of(kobj, struct net_bridge_port, kobj);
330 struct kobj_type brport_ktype = {
332 .sysfs_ops = &brport_sysfs_ops,
334 .release = release_nbp
337 /* Called with RTNL lock and dp_mutex. */
338 static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no)
340 struct net_bridge_port *p;
342 if (dev->br_port != NULL)
345 p = kzalloc(sizeof(*p), GFP_KERNEL);
349 dev_set_promiscuity(dev, 1);
351 p->port_no = port_no;
355 rcu_assign_pointer(dev->br_port, p);
357 /* It would make sense to assign dev->br_port here too, but
358 * that causes packets received on internal ports to get caught
359 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
360 * back to network stack, but that's a waste of time. */
362 rcu_assign_pointer(dp->ports[port_no], p);
363 list_add_rcu(&p->node, &dp->port_list);
366 /* Initialize kobject for bridge. This will be added as
367 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
369 kobject_init(&p->kobj, &brport_ktype);
371 dp_ifinfo_notify(RTM_NEWLINK, p);
376 static int add_port(int dp_idx, struct odp_port __user *portp)
378 struct net_device *dev;
380 struct odp_port port;
385 if (copy_from_user(&port, portp, sizeof port))
387 port.devname[IFNAMSIZ - 1] = '\0';
390 dp = get_dp_locked(dp_idx);
393 goto out_unlock_rtnl;
395 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
396 if (!dp->ports[port_no])
402 if (!(port.flags & ODP_PORT_INTERNAL)) {
404 dev = dev_get_by_name(&init_net, port.devname);
409 if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER ||
413 dev = dp_dev_create(dp, port.devname, port_no);
420 err = new_nbp(dp, dev, port_no);
424 dp_sysfs_add_if(dp->ports[port_no]);
426 err = __put_user(port_no, &port.port);
431 mutex_unlock(&dp->mutex);
438 int dp_del_port(struct net_bridge_port *p)
442 if (p->port_no != ODPP_LOCAL)
444 dp_ifinfo_notify(RTM_DELLINK, p);
448 if (is_dp_dev(p->dev)) {
449 /* Make sure that no packets arrive from now on, since
450 * dp_dev_xmit() will try to find itself through
451 * p->dp->ports[], and we're about to set that to null. */
452 netif_tx_disable(p->dev);
455 /* First drop references to device. */
456 dev_set_promiscuity(p->dev, -1);
457 list_del_rcu(&p->node);
458 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
459 rcu_assign_pointer(p->dev->br_port, NULL);
461 /* Then wait until no one is still using it, and destroy it. */
464 if (is_dp_dev(p->dev))
465 dp_dev_destroy(p->dev);
467 kobject_put(&p->kobj);
472 static int del_port(int dp_idx, int port_no)
474 struct net_bridge_port *p;
480 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
484 dp = get_dp_locked(dp_idx);
487 goto out_unlock_rtnl;
489 p = dp->ports[port_no];
494 err = dp_del_port(p);
497 mutex_unlock(&dp->mutex);
504 /* Must be called with rcu_read_lock. */
506 do_port_input(struct net_bridge_port *p, struct sk_buff *skb)
508 /* Make our own copy of the packet. Otherwise we will mangle the
509 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
510 * (No one comes after us, since we tell handle_bridge() that we took
512 skb = skb_share_check(skb, GFP_ATOMIC);
516 /* Push the Ethernet header back on. */
517 skb_push(skb, ETH_HLEN);
518 skb_reset_mac_header(skb);
519 dp_process_received_packet(skb, p);
522 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
523 void dp_process_received_packet(struct sk_buff *skb, struct net_bridge_port *p)
525 struct datapath *dp = p->dp;
526 struct dp_stats_percpu *stats;
527 struct odp_flow_key key;
528 struct sw_flow *flow;
530 WARN_ON_ONCE(skb_shared(skb));
532 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
533 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
535 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
536 if (dp->drop_frags) {
543 flow = dp_table_lookup(rcu_dereference(dp->table), &key);
545 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
546 flow_used(flow, skb);
547 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
552 dp_output_control(dp, skb, _ODPL_MISS_NR, 0);
557 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
558 * different set of devices!)
560 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
561 /* Called with rcu_read_lock and bottom-halves disabled. */
562 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
565 do_port_input(p, skb);
568 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
569 /* Called with rcu_read_lock and bottom-halves disabled. */
570 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
572 do_port_input(p, *pskb);
579 #if defined(CONFIG_XEN) && LINUX_VERSION_CODE == KERNEL_VERSION(2,6,18)
580 /* This code is copied verbatim from net/dev/core.c in Xen's
581 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
582 * directly because they aren't exported. */
583 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
585 if (ptr < (void *)skb->tail)
587 if (__pskb_pull_tail(skb,
588 ptr - (void *)skb->data - skb_headlen(skb))) {
595 int vswitch_skb_checksum_setup(struct sk_buff *skb)
597 if (skb->proto_csum_blank) {
598 if (skb->protocol != htons(ETH_P_IP))
600 if (!skb_pull_up_to(skb, skb->nh.iph + 1))
602 skb->h.raw = (unsigned char *)skb->nh.iph + 4*skb->nh.iph->ihl;
603 switch (skb->nh.iph->protocol) {
605 skb->csum = offsetof(struct tcphdr, check);
608 skb->csum = offsetof(struct udphdr, check);
612 printk(KERN_ERR "Attempting to checksum a non-"
613 "TCP/UDP packet, dropping a protocol"
614 " %d packet", skb->nh.iph->protocol);
617 if (!skb_pull_up_to(skb, skb->h.raw + skb->csum + 2))
619 skb->ip_summed = CHECKSUM_HW;
620 skb->proto_csum_blank = 0;
627 int vswitch_skb_checksum_setup(struct sk_buff *skb) { return 0; }
628 #endif /* CONFIG_XEN && linux == 2.6.18 */
631 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
634 struct dp_stats_percpu *stats;
635 struct sk_buff_head *queue;
639 WARN_ON_ONCE(skb_shared(skb));
640 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR);
642 queue = &dp->queues[queue_no];
644 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
647 /* If a checksum-deferred packet is forwarded to the controller,
648 * correct the pointers and checksum. This happens on a regular basis
649 * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets
650 * that do not have their checksum computed. We also implement it for
651 * the non-Xen case, but it is difficult to trigger or test this case
652 * there, hence the WARN_ON_ONCE().
654 err = vswitch_skb_checksum_setup(skb);
658 if (skb->ip_summed == CHECKSUM_PARTIAL) {
660 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
661 /* Until 2.6.22, the start of the transport header was also the
662 * start of data to be checksummed. Linux 2.6.22 introduced
663 * the csum_start field for this purpose, but we should point
664 * the transport header to it anyway for backward
665 * compatibility, as dev_queue_xmit() does even in 2.6.28. */
666 skb_set_transport_header(skb, skb->csum_start -
669 err = skb_checksum_help(skb);
674 if (skb->ip_summed == CHECKSUM_HW) {
675 err = skb_checksum_help(skb, 0);
681 /* Break apart GSO packets into their component pieces. Otherwise
682 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
683 if (skb_is_gso(skb)) {
684 struct sk_buff *nskb = skb_gso_segment(skb, 0);
688 if (unlikely(IS_ERR(skb))) {
693 /* XXX This case might not be possible. It's hard to
694 * tell from the skb_gso_segment() code and comment. */
698 /* Figure out port number. */
699 port_no = ODPP_LOCAL;
701 if (skb->dev->br_port)
702 port_no = skb->dev->br_port->port_no;
703 else if (is_dp_dev(skb->dev))
704 port_no = dp_dev_priv(skb->dev)->port_no;
707 /* Append each packet to queue. There will be only one packet unless
708 * we broke up a GSO packet above. */
710 struct odp_msg *header;
711 struct sk_buff *nskb = skb->next;
714 err = skb_cow(skb, sizeof *header);
724 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
725 header->type = queue_no;
726 header->length = skb->len;
727 header->port = port_no;
728 header->reserved = 0;
730 skb_queue_tail(queue, skb);
735 wake_up_interruptible(&dp->waitqueue);
741 stats = percpu_ptr(dp->stats_percpu, get_cpu());
748 static int flush_flows(struct datapath *dp)
751 return dp_table_flush(dp);
754 static int validate_actions(const struct sw_flow_actions *actions)
758 for (i = 0; i < actions->n_actions; i++) {
759 const union odp_action *a = &actions->actions[i];
762 if (a->output.port >= DP_MAX_PORTS)
766 case ODPAT_OUTPUT_GROUP:
767 if (a->output_group.group >= DP_MAX_GROUPS)
771 case ODPAT_SET_VLAN_VID:
772 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
776 case ODPAT_SET_VLAN_PCP:
777 if (a->vlan_pcp.vlan_pcp & ~VLAN_PCP_MASK)
782 if (a->type >= ODPAT_N_ACTIONS)
791 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
793 struct sw_flow_actions *actions;
796 actions = flow_actions_alloc(flow->n_actions);
797 error = PTR_ERR(actions);
802 if (copy_from_user(actions->actions, flow->actions,
803 flow->n_actions * sizeof(union odp_action)))
804 goto error_free_actions;
805 error = validate_actions(actions);
807 goto error_free_actions;
814 return ERR_PTR(error);
817 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
819 if (flow->used.tv_sec) {
820 stats->used_sec = flow->used.tv_sec;
821 stats->used_nsec = flow->used.tv_nsec;
824 stats->used_nsec = 0;
826 stats->n_packets = flow->packet_count;
827 stats->n_bytes = flow->byte_count;
828 stats->ip_tos = flow->ip_tos;
829 stats->tcp_flags = flow->tcp_flags;
833 static void clear_stats(struct sw_flow *flow)
835 flow->used.tv_sec = flow->used.tv_nsec = 0;
838 flow->packet_count = 0;
839 flow->byte_count = 0;
842 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
844 struct odp_flow_put uf;
845 struct sw_flow *flow, **bucket;
846 struct dp_table *table;
847 struct odp_flow_stats stats;
851 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
853 uf.flow.key.reserved = 0;
856 table = rcu_dereference(dp->table);
857 bucket = dp_table_lookup_for_insert(table, &uf.flow.key);
859 /* No such flow, and the slots where it could go are full. */
860 error = uf.flags & ODPPF_CREATE ? -EXFULL : -ENOENT;
862 } else if (!*bucket) {
863 /* No such flow, but we found an available slot for it. */
864 struct sw_flow_actions *acts;
867 if (!(uf.flags & ODPPF_CREATE))
870 /* Expand table, if necessary, to make room. */
871 if (dp->n_flows * 4 >= table->n_buckets &&
872 table->n_buckets < DP_MAX_BUCKETS) {
873 error = dp_table_expand(dp);
877 /* The bucket's location has changed. Try again. */
883 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
886 flow->key = uf.flow.key;
887 spin_lock_init(&flow->lock);
890 /* Obtain actions. */
891 acts = get_actions(&uf.flow);
892 error = PTR_ERR(acts);
894 goto error_free_flow;
895 rcu_assign_pointer(flow->sf_acts, acts);
897 /* Put flow in bucket. */
898 rcu_assign_pointer(*bucket, flow);
900 memset(&stats, 0, sizeof(struct odp_flow_stats));
902 /* We found a matching flow. */
903 struct sw_flow *flow = *rcu_dereference(bucket);
904 struct sw_flow_actions *old_acts, *new_acts;
905 unsigned long int flags;
907 /* Bail out if we're not allowed to modify an existing flow. */
909 if (!(uf.flags & ODPPF_MODIFY))
913 new_acts = get_actions(&uf.flow);
914 error = PTR_ERR(new_acts);
915 if (IS_ERR(new_acts))
917 old_acts = rcu_dereference(flow->sf_acts);
918 if (old_acts->n_actions != new_acts->n_actions ||
919 memcmp(old_acts->actions, new_acts->actions,
920 sizeof(union odp_action) * old_acts->n_actions)) {
921 rcu_assign_pointer(flow->sf_acts, new_acts);
922 flow_deferred_free_acts(old_acts);
927 /* Fetch stats, then clear them if necessary. */
928 spin_lock_irqsave(&flow->lock, flags);
929 get_stats(flow, &stats);
930 if (uf.flags & ODPPF_ZERO_STATS)
932 spin_unlock_irqrestore(&flow->lock, flags);
935 /* Copy stats to userspace. */
936 if (__copy_to_user(&ufp->flow.stats, &stats,
937 sizeof(struct odp_flow_stats)))
942 kmem_cache_free(flow_cache, flow);
947 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
949 union odp_action __user *actions;
950 struct sw_flow_actions *sf_acts;
953 if (__get_user(actions, &ufp->actions) ||
954 __get_user(n_actions, &ufp->n_actions))
960 sf_acts = rcu_dereference(flow->sf_acts);
961 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
962 (actions && copy_to_user(actions, sf_acts->actions,
963 sizeof(union odp_action) *
964 min(sf_acts->n_actions, n_actions))))
970 static int answer_query(struct sw_flow *flow, struct odp_flow __user *ufp)
972 struct odp_flow_stats stats;
973 unsigned long int flags;
975 spin_lock_irqsave(&flow->lock, flags);
976 get_stats(flow, &stats);
977 spin_unlock_irqrestore(&flow->lock, flags);
979 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
981 return put_actions(flow, ufp);
984 static int del_flow(struct datapath *dp, 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 /* XXX redundant lookup */
1002 error = dp_table_delete(table, flow);
1006 /* XXX These statistics might lose a few packets, since other CPUs can
1007 * be using this flow. We used to synchronize_rcu() to make sure that
1008 * we get completely accurate stats, but that blows our performance,
1011 error = answer_query(flow, ufp);
1012 flow_deferred_free(flow);
1018 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1020 struct dp_table *table = rcu_dereference(dp->table);
1022 for (i = 0; i < flowvec->n_flows; i++) {
1023 struct __user odp_flow *ufp = &flowvec->flows[i];
1025 struct sw_flow *flow;
1028 if (__copy_from_user(&uf, ufp, sizeof uf))
1030 uf.key.reserved = 0;
1032 flow = dp_table_lookup(table, &uf.key);
1034 error = __put_user(ENOENT, &ufp->stats.error);
1036 error = answer_query(flow, ufp);
1040 return flowvec->n_flows;
1043 struct list_flows_cbdata {
1044 struct odp_flow __user *uflows;
1049 static int list_flow(struct sw_flow *flow, void *cbdata_)
1051 struct list_flows_cbdata *cbdata = cbdata_;
1052 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1055 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1057 error = answer_query(flow, ufp);
1061 if (cbdata->listed_flows >= cbdata->n_flows)
1062 return cbdata->listed_flows;
1066 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1068 struct list_flows_cbdata cbdata;
1071 if (!flowvec->n_flows)
1074 cbdata.uflows = flowvec->flows;
1075 cbdata.n_flows = flowvec->n_flows;
1076 cbdata.listed_flows = 0;
1077 error = dp_table_foreach(rcu_dereference(dp->table),
1078 list_flow, &cbdata);
1079 return error ? error : cbdata.listed_flows;
1082 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1083 int (*function)(struct datapath *,
1084 const struct odp_flowvec *))
1086 struct odp_flowvec __user *uflowvec;
1087 struct odp_flowvec flowvec;
1090 uflowvec = (struct odp_flowvec __user *)argp;
1091 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1092 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1095 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1098 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1099 flowvec.n_flows * sizeof(struct odp_flow)))
1102 retval = function(dp, &flowvec);
1103 return (retval < 0 ? retval
1104 : retval == flowvec.n_flows ? 0
1105 : __put_user(retval, &uflowvec->n_flows));
1108 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1110 struct odp_execute execute;
1111 struct odp_flow_key key;
1112 struct sk_buff *skb;
1113 struct sw_flow_actions *actions;
1117 if (copy_from_user(&execute, executep, sizeof execute))
1121 if (execute.length < ETH_HLEN || execute.length > 65535)
1125 actions = flow_actions_alloc(execute.n_actions);
1130 if (copy_from_user(actions->actions, execute.actions,
1131 execute.n_actions * sizeof *execute.actions))
1132 goto error_free_actions;
1134 err = validate_actions(actions);
1136 goto error_free_actions;
1139 skb = alloc_skb(execute.length, GFP_KERNEL);
1141 goto error_free_actions;
1142 if (execute.in_port < DP_MAX_PORTS) {
1143 struct net_bridge_port *p = dp->ports[execute.in_port];
1149 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1151 goto error_free_skb;
1153 flow_extract(skb, execute.in_port, &key);
1154 err = execute_actions(dp, skb, &key, actions->actions,
1155 actions->n_actions, GFP_KERNEL);
1167 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1169 struct odp_stats stats;
1172 stats.n_flows = dp->n_flows;
1173 stats.cur_capacity = rcu_dereference(dp->table)->n_buckets * 2;
1174 stats.max_capacity = DP_MAX_BUCKETS * 2;
1175 stats.n_ports = dp->n_ports;
1176 stats.max_ports = DP_MAX_PORTS;
1177 stats.max_groups = DP_MAX_GROUPS;
1178 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1179 for_each_possible_cpu(i) {
1180 const struct dp_stats_percpu *s;
1181 s = percpu_ptr(dp->stats_percpu, i);
1182 stats.n_frags += s->n_frags;
1183 stats.n_hit += s->n_hit;
1184 stats.n_missed += s->n_missed;
1185 stats.n_lost += s->n_lost;
1187 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1188 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1189 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1192 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1193 int dp_min_mtu(const struct datapath *dp)
1195 struct net_bridge_port *p;
1200 list_for_each_entry_rcu (p, &dp->port_list, node) {
1201 struct net_device *dev = p->dev;
1203 /* Skip any internal ports, since that's what we're trying to
1208 if (!mtu || dev->mtu < mtu)
1212 return mtu ? mtu : ETH_DATA_LEN;
1216 put_port(const struct net_bridge_port *p, struct odp_port __user *uop)
1219 memset(&op, 0, sizeof op);
1220 strncpy(op.devname, p->dev->name, sizeof op.devname);
1221 op.port = p->port_no;
1222 op.flags = is_dp_dev(p->dev) ? ODP_PORT_INTERNAL : 0;
1223 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1227 query_port(struct datapath *dp, struct odp_port __user *uport)
1229 struct odp_port port;
1231 if (copy_from_user(&port, uport, sizeof port))
1233 if (port.devname[0]) {
1234 struct net_bridge_port *p;
1235 struct net_device *dev;
1238 port.devname[IFNAMSIZ - 1] = '\0';
1240 dev = dev_get_by_name(&init_net, port.devname);
1245 if (!p && is_dp_dev(dev)) {
1246 struct dp_dev *dp_dev = dp_dev_priv(dev);
1247 if (dp_dev->dp == dp)
1248 p = dp->ports[dp_dev->port_no];
1250 err = p && p->dp == dp ? put_port(p, uport) : -ENOENT;
1255 if (port.port >= DP_MAX_PORTS)
1257 if (!dp->ports[port.port])
1259 return put_port(dp->ports[port.port], uport);
1264 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1266 struct odp_portvec pv;
1267 struct net_bridge_port *p;
1270 if (copy_from_user(&pv, pvp, sizeof pv))
1275 list_for_each_entry_rcu (p, &dp->port_list, node) {
1276 if (put_port(p, &pv.ports[idx]))
1278 if (idx++ >= pv.n_ports)
1282 return put_user(dp->n_ports, &pvp->n_ports);
1285 /* RCU callback for freeing a dp_port_group */
1286 static void free_port_group(struct rcu_head *rcu)
1288 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1293 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1295 struct odp_port_group pg;
1296 struct dp_port_group *new_group, *old_group;
1300 if (copy_from_user(&pg, upg, sizeof pg))
1304 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1308 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1313 new_group->n_ports = pg.n_ports;
1315 if (copy_from_user(new_group->ports, pg.ports,
1316 sizeof(u16) * pg.n_ports))
1319 old_group = rcu_dereference(dp->groups[pg.group]);
1320 rcu_assign_pointer(dp->groups[pg.group], new_group);
1322 call_rcu(&old_group->rcu, free_port_group);
1332 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1334 struct odp_port_group pg;
1335 struct dp_port_group *g;
1338 if (copy_from_user(&pg, upg, sizeof pg))
1341 if (pg.group >= DP_MAX_GROUPS)
1344 g = dp->groups[pg.group];
1345 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1346 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1349 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1355 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1358 int dp_idx = iminor(f->f_dentry->d_inode);
1359 struct datapath *dp;
1360 int drop_frags, listeners, port_no;
1363 /* Handle commands with special locking requirements up front. */
1366 err = create_dp(dp_idx, (char __user *)argp);
1369 case ODP_DP_DESTROY:
1370 err = destroy_dp(dp_idx);
1374 err = add_port(dp_idx, (struct odp_port __user *)argp);
1378 err = get_user(port_no, (int __user *)argp);
1380 err = del_port(dp_idx, port_no);
1384 dp = get_dp_locked(dp_idx);
1391 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1394 case ODP_GET_DROP_FRAGS:
1395 err = put_user(dp->drop_frags, (int __user *)argp);
1398 case ODP_SET_DROP_FRAGS:
1399 err = get_user(drop_frags, (int __user *)argp);
1403 if (drop_frags != 0 && drop_frags != 1)
1405 dp->drop_frags = drop_frags;
1409 case ODP_GET_LISTEN_MASK:
1410 err = put_user((int)f->private_data, (int __user *)argp);
1413 case ODP_SET_LISTEN_MASK:
1414 err = get_user(listeners, (int __user *)argp);
1418 if (listeners & ~ODPL_ALL)
1421 f->private_data = (void*)listeners;
1424 case ODP_PORT_QUERY:
1425 err = query_port(dp, (struct odp_port __user *)argp);
1429 err = list_ports(dp, (struct odp_portvec __user *)argp);
1432 case ODP_PORT_GROUP_SET:
1433 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1436 case ODP_PORT_GROUP_GET:
1437 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1440 case ODP_FLOW_FLUSH:
1441 err = flush_flows(dp);
1445 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1449 err = del_flow(dp, (struct odp_flow __user *)argp);
1453 err = do_flowvec_ioctl(dp, argp, query_flows);
1457 err = do_flowvec_ioctl(dp, argp, list_flows);
1461 err = do_execute(dp, (struct odp_execute __user *)argp);
1468 mutex_unlock(&dp->mutex);
1473 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1476 for (i = 0; i < DP_N_QUEUES; i++) {
1477 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1483 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1486 /* XXX is there sufficient synchronization here? */
1487 int listeners = (int) f->private_data;
1488 int dp_idx = iminor(f->f_dentry->d_inode);
1489 struct datapath *dp = get_dp(dp_idx);
1490 struct sk_buff *skb;
1491 struct iovec __user iov;
1498 if (nbytes == 0 || !listeners)
1504 for (i = 0; i < DP_N_QUEUES; i++) {
1505 if (listeners & (1 << i)) {
1506 skb = skb_dequeue(&dp->queues[i]);
1512 if (f->f_flags & O_NONBLOCK) {
1517 wait_event_interruptible(dp->waitqueue,
1518 dp_has_packet_of_interest(dp,
1521 if (signal_pending(current)) {
1522 retval = -ERESTARTSYS;
1527 copy_bytes = min(skb->len, nbytes);
1529 iov.iov_len = copy_bytes;
1530 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1532 retval = copy_bytes;
1539 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1541 /* XXX is there sufficient synchronization here? */
1542 int dp_idx = iminor(file->f_dentry->d_inode);
1543 struct datapath *dp = get_dp(dp_idx);
1548 poll_wait(file, &dp->waitqueue, wait);
1549 if (dp_has_packet_of_interest(dp, (int)file->private_data))
1550 mask |= POLLIN | POLLRDNORM;
1552 mask = POLLIN | POLLRDNORM | POLLHUP;
1557 struct file_operations openvswitch_fops = {
1558 /* XXX .aio_read = openvswitch_aio_read, */
1559 .read = openvswitch_read,
1560 .poll = openvswitch_poll,
1561 .unlocked_ioctl = openvswitch_ioctl,
1562 /* XXX .fasync = openvswitch_fasync, */
1566 static struct llc_sap *dp_stp_sap;
1568 static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev,
1569 struct packet_type *pt, struct net_device *orig_dev)
1571 /* We don't really care about STP packets, we just listen for them for
1572 * mutual exclusion with the bridge module, so this just discards
1578 static int __init dp_init(void)
1582 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1584 /* Register to receive STP packets because the bridge module also
1585 * attempts to do so. Since there can only be a single listener for a
1586 * given protocol, this provides mutual exclusion against the bridge
1587 * module, preventing both of them from being loaded at the same
1589 dp_stp_sap = llc_sap_open(LLC_SAP_BSPAN, dp_stp_rcv);
1591 printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1599 err = register_netdevice_notifier(&dp_device_notifier);
1601 goto error_flow_exit;
1603 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1605 goto error_unreg_notifier;
1607 /* Hook into callback used by the bridge to intercept packets.
1608 * Parasites we are. */
1609 br_handle_frame_hook = dp_frame_hook;
1613 error_unreg_notifier:
1614 unregister_netdevice_notifier(&dp_device_notifier);
1621 static void dp_cleanup(void)
1624 unregister_chrdev(major, "openvswitch");
1625 unregister_netdevice_notifier(&dp_device_notifier);
1627 br_handle_frame_hook = NULL;
1628 llc_sap_put(dp_stp_sap);
1631 module_init(dp_init);
1632 module_exit(dp_cleanup);
1634 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1635 MODULE_LICENSE("GPL");