2 * Copyright (c) 2007, 2008, 2009, 2010 Nicira Networks.
3 * Distributed under the terms of the GNU GPL version 2.
5 * Significant portions of this file may be copied from parts of the Linux
6 * kernel, by Linus Torvalds and others.
9 /* Functions for managing the dp interface/device. */
11 #include <linux/init.h>
12 #include <linux/module.h>
14 #include <linux/if_arp.h>
15 #include <linux/if_vlan.h>
18 #include <linux/delay.h>
19 #include <linux/time.h>
20 #include <linux/etherdevice.h>
21 #include <linux/kernel.h>
22 #include <linux/kthread.h>
23 #include <linux/mutex.h>
24 #include <linux/percpu.h>
25 #include <linux/rcupdate.h>
26 #include <linux/tcp.h>
27 #include <linux/udp.h>
28 #include <linux/version.h>
29 #include <linux/ethtool.h>
30 #include <linux/random.h>
31 #include <linux/wait.h>
32 #include <asm/system.h>
33 #include <asm/div64.h>
35 #include <linux/netfilter_bridge.h>
36 #include <linux/netfilter_ipv4.h>
37 #include <linux/inetdevice.h>
38 #include <linux/list.h>
39 #include <linux/rculist.h>
40 #include <linux/workqueue.h>
41 #include <linux/dmi.h>
42 #include <net/inet_ecn.h>
44 #include "openvswitch/datapath-protocol.h"
49 #include "vport-internal_dev.h"
54 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
55 EXPORT_SYMBOL(dp_ioctl_hook);
57 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
60 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
63 * It is safe to access the datapath and dp_port structures with just
66 static struct datapath *dps[ODP_MAX];
67 static DEFINE_MUTEX(dp_mutex);
69 /* Number of milliseconds between runs of the maintenance thread. */
70 #define MAINT_SLEEP_MSECS 1000
72 static int new_dp_port(struct datapath *, struct odp_port *, int port_no);
74 /* Must be called with rcu_read_lock or dp_mutex. */
75 struct datapath *get_dp(int dp_idx)
77 if (dp_idx < 0 || dp_idx >= ODP_MAX)
79 return rcu_dereference(dps[dp_idx]);
81 EXPORT_SYMBOL_GPL(get_dp);
83 static struct datapath *get_dp_locked(int dp_idx)
87 mutex_lock(&dp_mutex);
90 mutex_lock(&dp->mutex);
91 mutex_unlock(&dp_mutex);
95 /* Must be called with rcu_read_lock or RTNL lock. */
96 const char *dp_name(const struct datapath *dp)
98 return vport_get_name(dp->ports[ODPP_LOCAL]->vport);
101 static inline size_t br_nlmsg_size(void)
103 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
104 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
105 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
106 + nla_total_size(4) /* IFLA_MASTER */
107 + nla_total_size(4) /* IFLA_MTU */
108 + nla_total_size(4) /* IFLA_LINK */
109 + nla_total_size(1); /* IFLA_OPERSTATE */
112 static int dp_fill_ifinfo(struct sk_buff *skb,
113 const struct dp_port *port,
114 int event, unsigned int flags)
116 const struct datapath *dp = port->dp;
117 int ifindex = vport_get_ifindex(port->vport);
118 int iflink = vport_get_iflink(port->vport);
119 struct ifinfomsg *hdr;
120 struct nlmsghdr *nlh;
128 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
132 hdr = nlmsg_data(nlh);
133 hdr->ifi_family = AF_BRIDGE;
135 hdr->ifi_type = ARPHRD_ETHER;
136 hdr->ifi_index = ifindex;
137 hdr->ifi_flags = vport_get_flags(port->vport);
140 NLA_PUT_STRING(skb, IFLA_IFNAME, vport_get_name(port->vport));
141 NLA_PUT_U32(skb, IFLA_MASTER, vport_get_ifindex(dp->ports[ODPP_LOCAL]->vport));
142 NLA_PUT_U32(skb, IFLA_MTU, vport_get_mtu(port->vport));
143 #ifdef IFLA_OPERSTATE
144 NLA_PUT_U8(skb, IFLA_OPERSTATE,
145 vport_is_running(port->vport)
146 ? vport_get_operstate(port->vport)
150 NLA_PUT(skb, IFLA_ADDRESS, ETH_ALEN,
151 vport_get_addr(port->vport));
153 if (ifindex != iflink)
154 NLA_PUT_U32(skb, IFLA_LINK,iflink);
156 return nlmsg_end(skb, nlh);
159 nlmsg_cancel(skb, nlh);
163 static void dp_ifinfo_notify(int event, struct dp_port *port)
168 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
172 err = dp_fill_ifinfo(skb, port, event, 0);
174 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
175 WARN_ON(err == -EMSGSIZE);
179 rtnl_notify(skb, &init_net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
183 rtnl_set_sk_err(&init_net, RTNLGRP_LINK, err);
186 static void release_dp(struct kobject *kobj)
188 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
192 static struct kobj_type dp_ktype = {
193 .release = release_dp
196 static int create_dp(int dp_idx, const char __user *devnamep)
198 struct odp_port internal_dev_port;
199 char devname[IFNAMSIZ];
205 int retval = strncpy_from_user(devname, devnamep, IFNAMSIZ);
209 } else if (retval >= IFNAMSIZ) {
214 snprintf(devname, sizeof devname, "of%d", dp_idx);
218 mutex_lock(&dp_mutex);
220 if (!try_module_get(THIS_MODULE))
223 /* Exit early if a datapath with that number already exists.
224 * (We don't use -EEXIST because that's ambiguous with 'devname'
225 * conflicting with an existing network device name.) */
231 dp = kzalloc(sizeof *dp, GFP_KERNEL);
234 INIT_LIST_HEAD(&dp->port_list);
235 mutex_init(&dp->mutex);
237 for (i = 0; i < DP_N_QUEUES; i++)
238 skb_queue_head_init(&dp->queues[i]);
239 init_waitqueue_head(&dp->waitqueue);
241 /* Initialize kobject for bridge. This will be added as
242 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
243 dp->ifobj.kset = NULL;
244 kobject_init(&dp->ifobj, &dp_ktype);
246 /* Allocate table. */
248 rcu_assign_pointer(dp->table, tbl_create(0));
252 /* Set up our datapath device. */
253 BUILD_BUG_ON(sizeof(internal_dev_port.devname) != sizeof(devname));
254 strcpy(internal_dev_port.devname, devname);
255 internal_dev_port.flags = ODP_PORT_INTERNAL;
256 err = new_dp_port(dp, &internal_dev_port, ODPP_LOCAL);
261 goto err_destroy_table;
265 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
266 if (!dp->stats_percpu)
267 goto err_destroy_local_port;
269 rcu_assign_pointer(dps[dp_idx], dp);
270 mutex_unlock(&dp_mutex);
277 err_destroy_local_port:
278 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
280 tbl_destroy(dp->table, NULL);
284 module_put(THIS_MODULE);
286 mutex_unlock(&dp_mutex);
292 static void do_destroy_dp(struct datapath *dp)
294 struct dp_port *p, *n;
297 list_for_each_entry_safe (p, n, &dp->port_list, node)
298 if (p->port_no != ODPP_LOCAL)
299 dp_detach_port(p, 1);
303 rcu_assign_pointer(dps[dp->dp_idx], NULL);
305 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
307 tbl_destroy(dp->table, flow_free_tbl);
309 for (i = 0; i < DP_N_QUEUES; i++)
310 skb_queue_purge(&dp->queues[i]);
311 for (i = 0; i < DP_MAX_GROUPS; i++)
312 kfree(dp->groups[i]);
313 free_percpu(dp->stats_percpu);
314 kobject_put(&dp->ifobj);
315 module_put(THIS_MODULE);
318 static int destroy_dp(int dp_idx)
324 mutex_lock(&dp_mutex);
334 mutex_unlock(&dp_mutex);
339 static void release_dp_port(struct kobject *kobj)
341 struct dp_port *p = container_of(kobj, struct dp_port, kobj);
345 static struct kobj_type brport_ktype = {
347 .sysfs_ops = &brport_sysfs_ops,
349 .release = release_dp_port
352 /* Called with RTNL lock and dp_mutex. */
353 static int new_dp_port(struct datapath *dp, struct odp_port *odp_port, int port_no)
359 vport = vport_locate(odp_port->devname);
363 if (odp_port->flags & ODP_PORT_INTERNAL)
364 vport = __vport_add(odp_port->devname, "internal", NULL);
366 vport = __vport_add(odp_port->devname, "netdev", NULL);
371 return PTR_ERR(vport);
374 p = kzalloc(sizeof(*p), GFP_KERNEL);
378 p->port_no = port_no;
380 atomic_set(&p->sflow_pool, 0);
382 err = vport_attach(vport, p);
388 rcu_assign_pointer(dp->ports[port_no], p);
389 list_add_rcu(&p->node, &dp->port_list);
392 /* Initialize kobject for bridge. This will be added as
393 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
395 kobject_init(&p->kobj, &brport_ktype);
397 dp_ifinfo_notify(RTM_NEWLINK, p);
402 static int attach_port(int dp_idx, struct odp_port __user *portp)
405 struct odp_port port;
410 if (copy_from_user(&port, portp, sizeof port))
412 port.devname[IFNAMSIZ - 1] = '\0';
415 dp = get_dp_locked(dp_idx);
418 goto out_unlock_rtnl;
420 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
421 if (!dp->ports[port_no])
427 err = new_dp_port(dp, &port, port_no);
431 set_internal_devs_mtu(dp);
432 dp_sysfs_add_if(dp->ports[port_no]);
434 err = __put_user(port_no, &portp->port);
437 mutex_unlock(&dp->mutex);
444 int dp_detach_port(struct dp_port *p, int may_delete)
446 struct vport *vport = p->vport;
451 if (p->port_no != ODPP_LOCAL)
453 dp_ifinfo_notify(RTM_DELLINK, p);
455 /* First drop references to device. */
457 list_del_rcu(&p->node);
458 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
460 err = vport_detach(vport);
464 /* Then wait until no one is still using it, and destroy it. */
468 const char *port_type = vport_get_type(vport);
470 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
477 kobject_put(&p->kobj);
482 static int detach_port(int dp_idx, int port_no)
489 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
493 dp = get_dp_locked(dp_idx);
496 goto out_unlock_rtnl;
498 p = dp->ports[port_no];
503 err = dp_detach_port(p, 1);
506 mutex_unlock(&dp->mutex);
513 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
514 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
516 struct datapath *dp = p->dp;
517 struct dp_stats_percpu *stats;
518 struct odp_flow_key key;
519 struct tbl_node *flow_node;
521 WARN_ON_ONCE(skb_shared(skb));
522 skb_warn_if_lro(skb);
524 OVS_CB(skb)->dp_port = p;
525 compute_ip_summed(skb, false);
527 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
528 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
530 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
531 if (dp->drop_frags) {
538 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
540 struct sw_flow *flow = flow_cast(flow_node);
541 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
542 flow_used(flow, skb);
543 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
548 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
552 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
553 /* This code is based on skb_checksum_setup() from Xen's net/dev/core.c. We
554 * can't call this function directly because it isn't exported in all
556 int vswitch_skb_checksum_setup(struct sk_buff *skb)
561 __u16 csum_start, csum_offset;
563 if (!skb->proto_csum_blank)
566 if (skb->protocol != htons(ETH_P_IP))
569 if (!pskb_may_pull(skb, skb_network_header(skb) + sizeof(struct iphdr) - skb->data))
573 th = skb_network_header(skb) + 4 * iph->ihl;
575 csum_start = th - skb->head;
576 switch (iph->protocol) {
578 csum_offset = offsetof(struct tcphdr, check);
581 csum_offset = offsetof(struct udphdr, check);
585 printk(KERN_ERR "Attempting to checksum a non-"
586 "TCP/UDP packet, dropping a protocol"
587 " %d packet", iph->protocol);
591 if (!pskb_may_pull(skb, th + csum_offset + 2 - skb->data))
594 skb->ip_summed = CHECKSUM_PARTIAL;
595 skb->proto_csum_blank = 0;
597 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
598 skb->csum_start = csum_start;
599 skb->csum_offset = csum_offset;
601 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
602 skb->csum = csum_offset;
610 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
612 /* Types of checksums that we can receive (these all refer to L4 checksums):
613 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
614 * (though not verified) checksum in packet but not in skb->csum. Packets
615 * from the bridge local port will also have this type.
616 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
617 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
618 * a valid skb->csum. Importantly, both contain a full checksum (not
619 * verified) in the packet itself. The only difference is that if the
620 * packet gets to L4 processing on this machine (not in DomU) we won't
621 * have to recompute the checksum to verify. Most hardware devices do not
622 * produce packets with this type, even if they support receive checksum
623 * offloading (they produce type #5).
624 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
625 * be computed if it is sent off box. Unfortunately on earlier kernels,
626 * this case is impossible to distinguish from #2, despite having opposite
627 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
628 * to distinguish the different states. The only real user of this type
629 * with bridging is Xen (on later kernels).
630 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
631 * generated locally by a Xen DomU and has a partial checksum. If it is
632 * handled on this machine (Dom0 or DomU), then the checksum will not be
633 * computed. If it goes off box, the checksum in the packet needs to be
634 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
635 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
636 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
637 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
638 * full checksum or using a protocol without a checksum. skb->csum is
639 * undefined. This is common from devices with receive checksum
640 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
641 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
643 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
644 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
645 * based on whether it is on the transmit or receive path. After the datapath
646 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
647 * checksum, we will panic. Since we can receive packets with checksums, we
648 * assume that all CHECKSUM_HW packets have checksums and map them to
649 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
650 * packet is processed by the local IP stack, in which case it will need to
651 * be reverified). If we receive a packet with CHECKSUM_HW that really means
652 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
653 * shouldn't be any devices that do this with bridging.
655 * The bridge has similar behavior and this function closely resembles
656 * skb_forward_csum(). It is slightly different because we are only concerned
657 * with bridging and not other types of forwarding and can get away with
658 * slightly more optimal behavior.*/
660 compute_ip_summed(struct sk_buff *skb, bool xmit)
662 /* For our convenience these defines change repeatedly between kernel
663 * versions, so we can't just copy them over... */
664 switch (skb->ip_summed) {
666 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
668 case CHECKSUM_UNNECESSARY:
669 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
672 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
673 * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
674 * uses some special fields to represent this (see below). Since we
675 * can only make one type work, pick the one that actually happens in
678 * The one exception to this is if we are on the transmit path
679 * (basically after skb_checksum_setup() has been run) the type has
680 * already been converted, so we should stay with that. */
683 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
685 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
689 case CHECKSUM_COMPLETE:
690 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
692 case CHECKSUM_PARTIAL:
693 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
697 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
699 /* None seems the safest... */
700 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
703 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
704 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
705 * kernels. It should not be set on the transmit path though. */
706 if (skb->proto_csum_blank)
707 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
709 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
714 forward_ip_summed(struct sk_buff *skb)
717 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
718 skb->ip_summed = CHECKSUM_NONE;
722 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
723 * unless we broke up a GSO packet. */
725 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
726 int queue_no, u32 arg)
728 struct sk_buff *nskb;
732 if (OVS_CB(skb)->dp_port)
733 port_no = OVS_CB(skb)->dp_port->port_no;
735 port_no = ODPP_LOCAL;
738 struct odp_msg *header;
743 /* If a checksum-deferred packet is forwarded to the
744 * controller, correct the pointers and checksum. This happens
745 * on a regular basis only on Xen, on which VMs can pass up
746 * packets that do not have their checksum computed.
748 err = vswitch_skb_checksum_setup(skb);
752 if (skb->ip_summed == CHECKSUM_PARTIAL) {
754 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
755 /* Until 2.6.22, the start of the transport header was
756 * also the start of data to be checksummed. Linux
757 * 2.6.22 introduced the csum_start field for this
758 * purpose, but we should point the transport header to
759 * it anyway for backward compatibility, as
760 * dev_queue_xmit() does even in 2.6.28. */
761 skb_set_transport_header(skb, skb->csum_start -
765 err = skb_checksum_help(skb);
770 err = skb_cow(skb, sizeof *header);
774 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
775 header->type = queue_no;
776 header->length = skb->len;
777 header->port = port_no;
778 header->reserved = 0;
780 skb_queue_tail(queue, skb);
788 while ((skb = nskb) != NULL) {
796 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
799 struct dp_stats_percpu *stats;
800 struct sk_buff_head *queue;
803 WARN_ON_ONCE(skb_shared(skb));
804 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
805 queue = &dp->queues[queue_no];
807 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
810 forward_ip_summed(skb);
812 /* Break apart GSO packets into their component pieces. Otherwise
813 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
814 if (skb_is_gso(skb)) {
815 struct sk_buff *nskb = skb_gso_segment(skb, 0);
819 if (unlikely(IS_ERR(skb))) {
824 /* XXX This case might not be possible. It's hard to
825 * tell from the skb_gso_segment() code and comment. */
829 err = queue_control_packets(skb, queue, queue_no, arg);
830 wake_up_interruptible(&dp->waitqueue);
836 stats = percpu_ptr(dp->stats_percpu, get_cpu());
843 static int flush_flows(struct datapath *dp)
845 struct tbl *old_table = rcu_dereference(dp->table);
846 struct tbl *new_table;
848 new_table = tbl_create(0);
852 rcu_assign_pointer(dp->table, new_table);
854 tbl_deferred_destroy(old_table, flow_free_tbl);
859 static int validate_actions(const struct sw_flow_actions *actions)
863 for (i = 0; i < actions->n_actions; i++) {
864 const union odp_action *a = &actions->actions[i];
867 if (a->output.port >= DP_MAX_PORTS)
871 case ODPAT_OUTPUT_GROUP:
872 if (a->output_group.group >= DP_MAX_GROUPS)
876 case ODPAT_SET_VLAN_VID:
877 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
881 case ODPAT_SET_VLAN_PCP:
882 if (a->vlan_pcp.vlan_pcp
883 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
887 case ODPAT_SET_NW_TOS:
888 if (a->nw_tos.nw_tos & INET_ECN_MASK)
893 if (a->type >= ODPAT_N_ACTIONS)
902 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
904 struct sw_flow_actions *actions;
907 actions = flow_actions_alloc(flow->n_actions);
908 error = PTR_ERR(actions);
913 if (copy_from_user(actions->actions, flow->actions,
914 flow->n_actions * sizeof(union odp_action)))
915 goto error_free_actions;
916 error = validate_actions(actions);
918 goto error_free_actions;
925 return ERR_PTR(error);
928 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
930 if (flow->used.tv_sec) {
931 stats->used_sec = flow->used.tv_sec;
932 stats->used_nsec = flow->used.tv_nsec;
935 stats->used_nsec = 0;
937 stats->n_packets = flow->packet_count;
938 stats->n_bytes = flow->byte_count;
939 stats->ip_tos = flow->ip_tos;
940 stats->tcp_flags = flow->tcp_flags;
944 static void clear_stats(struct sw_flow *flow)
946 flow->used.tv_sec = flow->used.tv_nsec = 0;
949 flow->packet_count = 0;
950 flow->byte_count = 0;
953 static int expand_table(struct datapath *dp)
955 struct tbl *old_table = rcu_dereference(dp->table);
956 struct tbl *new_table;
958 new_table = tbl_expand(old_table);
959 if (IS_ERR(new_table))
960 return PTR_ERR(new_table);
962 rcu_assign_pointer(dp->table, new_table);
963 tbl_deferred_destroy(old_table, NULL);
968 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
970 struct odp_flow_put uf;
971 struct tbl_node *flow_node;
972 struct sw_flow *flow;
974 struct odp_flow_stats stats;
978 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
980 memset(uf.flow.key.reserved, 0, sizeof uf.flow.key.reserved);
982 table = rcu_dereference(dp->table);
983 flow_node = tbl_lookup(table, &uf.flow.key, flow_hash(&uf.flow.key), flow_cmp);
986 struct sw_flow_actions *acts;
989 if (!(uf.flags & ODPPF_CREATE))
992 /* Expand table, if necessary, to make room. */
993 if (tbl_count(table) >= tbl_n_buckets(table)) {
994 error = expand_table(dp);
997 table = rcu_dereference(dp->table);
1000 /* Allocate flow. */
1002 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
1005 flow->key = uf.flow.key;
1006 spin_lock_init(&flow->lock);
1009 /* Obtain actions. */
1010 acts = get_actions(&uf.flow);
1011 error = PTR_ERR(acts);
1013 goto error_free_flow;
1014 rcu_assign_pointer(flow->sf_acts, acts);
1016 /* Put flow in bucket. */
1017 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1019 goto error_free_flow_acts;
1021 memset(&stats, 0, sizeof(struct odp_flow_stats));
1023 /* We found a matching flow. */
1024 struct sw_flow_actions *old_acts, *new_acts;
1025 unsigned long int flags;
1027 flow = flow_cast(flow_node);
1029 /* Bail out if we're not allowed to modify an existing flow. */
1031 if (!(uf.flags & ODPPF_MODIFY))
1035 new_acts = get_actions(&uf.flow);
1036 error = PTR_ERR(new_acts);
1037 if (IS_ERR(new_acts))
1039 old_acts = rcu_dereference(flow->sf_acts);
1040 if (old_acts->n_actions != new_acts->n_actions ||
1041 memcmp(old_acts->actions, new_acts->actions,
1042 sizeof(union odp_action) * old_acts->n_actions)) {
1043 rcu_assign_pointer(flow->sf_acts, new_acts);
1044 flow_deferred_free_acts(old_acts);
1049 /* Fetch stats, then clear them if necessary. */
1050 spin_lock_irqsave(&flow->lock, flags);
1051 get_stats(flow, &stats);
1052 if (uf.flags & ODPPF_ZERO_STATS)
1054 spin_unlock_irqrestore(&flow->lock, flags);
1057 /* Copy stats to userspace. */
1058 if (__copy_to_user(&ufp->flow.stats, &stats,
1059 sizeof(struct odp_flow_stats)))
1063 error_free_flow_acts:
1064 kfree(flow->sf_acts);
1066 kmem_cache_free(flow_cache, flow);
1071 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1073 union odp_action __user *actions;
1074 struct sw_flow_actions *sf_acts;
1077 if (__get_user(actions, &ufp->actions) ||
1078 __get_user(n_actions, &ufp->n_actions))
1084 sf_acts = rcu_dereference(flow->sf_acts);
1085 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1086 (actions && copy_to_user(actions, sf_acts->actions,
1087 sizeof(union odp_action) *
1088 min(sf_acts->n_actions, n_actions))))
1094 static int answer_query(struct sw_flow *flow, u32 query_flags,
1095 struct odp_flow __user *ufp)
1097 struct odp_flow_stats stats;
1098 unsigned long int flags;
1100 spin_lock_irqsave(&flow->lock, flags);
1101 get_stats(flow, &stats);
1103 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1104 flow->tcp_flags = 0;
1106 spin_unlock_irqrestore(&flow->lock, flags);
1108 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1110 return put_actions(flow, ufp);
1113 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1115 struct tbl *table = rcu_dereference(dp->table);
1117 struct tbl_node *flow_node;
1118 struct sw_flow *flow;
1122 if (copy_from_user(&uf, ufp, sizeof uf))
1124 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1126 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1131 error = tbl_remove(table, flow_node);
1135 /* XXX These statistics might lose a few packets, since other CPUs can
1136 * be using this flow. We used to synchronize_rcu() to make sure that
1137 * we get completely accurate stats, but that blows our performance,
1140 flow = flow_cast(flow_node);
1141 error = answer_query(flow, 0, ufp);
1142 flow_deferred_free(flow);
1148 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1150 struct tbl *table = rcu_dereference(dp->table);
1152 for (i = 0; i < flowvec->n_flows; i++) {
1153 struct __user odp_flow *ufp = &flowvec->flows[i];
1155 struct tbl_node *flow_node;
1158 if (__copy_from_user(&uf, ufp, sizeof uf))
1160 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1162 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1164 error = __put_user(ENOENT, &ufp->stats.error);
1166 error = answer_query(flow_cast(flow_node), uf.flags, ufp);
1170 return flowvec->n_flows;
1173 struct list_flows_cbdata {
1174 struct odp_flow __user *uflows;
1179 static int list_flow(struct tbl_node *node, void *cbdata_)
1181 struct sw_flow *flow = flow_cast(node);
1182 struct list_flows_cbdata *cbdata = cbdata_;
1183 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1186 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1188 error = answer_query(flow, 0, ufp);
1192 if (cbdata->listed_flows >= cbdata->n_flows)
1193 return cbdata->listed_flows;
1197 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1199 struct list_flows_cbdata cbdata;
1202 if (!flowvec->n_flows)
1205 cbdata.uflows = flowvec->flows;
1206 cbdata.n_flows = flowvec->n_flows;
1207 cbdata.listed_flows = 0;
1208 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1209 return error ? error : cbdata.listed_flows;
1212 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1213 int (*function)(struct datapath *,
1214 const struct odp_flowvec *))
1216 struct odp_flowvec __user *uflowvec;
1217 struct odp_flowvec flowvec;
1220 uflowvec = (struct odp_flowvec __user *)argp;
1221 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1222 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1225 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1228 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1229 flowvec.n_flows * sizeof(struct odp_flow)))
1232 retval = function(dp, &flowvec);
1233 return (retval < 0 ? retval
1234 : retval == flowvec.n_flows ? 0
1235 : __put_user(retval, &uflowvec->n_flows));
1238 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1240 struct odp_execute execute;
1241 struct odp_flow_key key;
1242 struct sk_buff *skb;
1243 struct sw_flow_actions *actions;
1248 if (copy_from_user(&execute, executep, sizeof execute))
1252 if (execute.length < ETH_HLEN || execute.length > 65535)
1256 actions = flow_actions_alloc(execute.n_actions);
1261 if (copy_from_user(actions->actions, execute.actions,
1262 execute.n_actions * sizeof *execute.actions))
1263 goto error_free_actions;
1265 err = validate_actions(actions);
1267 goto error_free_actions;
1270 skb = alloc_skb(execute.length, GFP_KERNEL);
1272 goto error_free_actions;
1274 if (execute.in_port < DP_MAX_PORTS)
1275 OVS_CB(skb)->dp_port = dp->ports[execute.in_port];
1277 OVS_CB(skb)->dp_port = NULL;
1280 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1282 goto error_free_skb;
1284 skb_reset_mac_header(skb);
1287 /* Normally, setting the skb 'protocol' field would be handled by a
1288 * call to eth_type_trans(), but it assumes there's a sending
1289 * device, which we may not have. */
1290 if (ntohs(eth->h_proto) >= 1536)
1291 skb->protocol = eth->h_proto;
1293 skb->protocol = htons(ETH_P_802_2);
1295 flow_extract(skb, execute.in_port, &key);
1296 err = execute_actions(dp, skb, &key, actions->actions,
1297 actions->n_actions, GFP_KERNEL);
1309 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1311 struct tbl *table = rcu_dereference(dp->table);
1312 struct odp_stats stats;
1315 stats.n_flows = tbl_count(table);
1316 stats.cur_capacity = tbl_n_buckets(table);
1317 stats.max_capacity = TBL_MAX_BUCKETS;
1318 stats.n_ports = dp->n_ports;
1319 stats.max_ports = DP_MAX_PORTS;
1320 stats.max_groups = DP_MAX_GROUPS;
1321 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1322 for_each_possible_cpu(i) {
1323 const struct dp_stats_percpu *s;
1324 s = percpu_ptr(dp->stats_percpu, i);
1325 stats.n_frags += s->n_frags;
1326 stats.n_hit += s->n_hit;
1327 stats.n_missed += s->n_missed;
1328 stats.n_lost += s->n_lost;
1330 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1331 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1332 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1335 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1336 int dp_min_mtu(const struct datapath *dp)
1343 list_for_each_entry_rcu (p, &dp->port_list, node) {
1346 /* Skip any internal ports, since that's what we're trying to
1348 if (is_internal_vport(p->vport))
1351 dev_mtu = vport_get_mtu(p->vport);
1352 if (!mtu || dev_mtu < mtu)
1356 return mtu ? mtu : ETH_DATA_LEN;
1359 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1360 * be called with RTNL lock. */
1361 void set_internal_devs_mtu(const struct datapath *dp)
1368 mtu = dp_min_mtu(dp);
1370 list_for_each_entry_rcu (p, &dp->port_list, node) {
1371 if (is_internal_vport(p->vport))
1372 vport_set_mtu(p->vport, mtu);
1377 put_port(const struct dp_port *p, struct odp_port __user *uop)
1381 memset(&op, 0, sizeof op);
1384 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1387 op.port = p->port_no;
1388 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1390 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1394 query_port(struct datapath *dp, struct odp_port __user *uport)
1396 struct odp_port port;
1398 if (copy_from_user(&port, uport, sizeof port))
1401 if (port.devname[0]) {
1402 struct vport *vport;
1403 struct dp_port *dp_port;
1406 port.devname[IFNAMSIZ - 1] = '\0';
1411 vport = vport_locate(port.devname);
1417 dp_port = vport_get_dp_port(vport);
1418 if (!dp_port || dp_port->dp != dp) {
1423 port.port = dp_port->port_no;
1432 if (port.port >= DP_MAX_PORTS)
1434 if (!dp->ports[port.port])
1438 return put_port(dp->ports[port.port], uport);
1442 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1444 struct odp_portvec pv;
1448 if (copy_from_user(&pv, pvp, sizeof pv))
1453 list_for_each_entry_rcu (p, &dp->port_list, node) {
1454 if (put_port(p, &pv.ports[idx]))
1456 if (idx++ >= pv.n_ports)
1460 return put_user(dp->n_ports, &pvp->n_ports);
1463 /* RCU callback for freeing a dp_port_group */
1464 static void free_port_group(struct rcu_head *rcu)
1466 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1471 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1473 struct odp_port_group pg;
1474 struct dp_port_group *new_group, *old_group;
1478 if (copy_from_user(&pg, upg, sizeof pg))
1482 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1486 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1491 new_group->n_ports = pg.n_ports;
1493 if (copy_from_user(new_group->ports, pg.ports,
1494 sizeof(u16) * pg.n_ports))
1497 old_group = rcu_dereference(dp->groups[pg.group]);
1498 rcu_assign_pointer(dp->groups[pg.group], new_group);
1500 call_rcu(&old_group->rcu, free_port_group);
1510 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1512 struct odp_port_group pg;
1513 struct dp_port_group *g;
1516 if (copy_from_user(&pg, upg, sizeof pg))
1519 if (pg.group >= DP_MAX_GROUPS)
1522 g = dp->groups[pg.group];
1523 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1524 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1527 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1533 static int get_listen_mask(const struct file *f)
1535 return (long)f->private_data;
1538 static void set_listen_mask(struct file *f, int listen_mask)
1540 f->private_data = (void*)(long)listen_mask;
1543 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1546 int dp_idx = iminor(f->f_dentry->d_inode);
1547 struct datapath *dp;
1548 int drop_frags, listeners, port_no;
1549 unsigned int sflow_probability;
1552 /* Handle commands with special locking requirements up front. */
1555 err = create_dp(dp_idx, (char __user *)argp);
1558 case ODP_DP_DESTROY:
1559 err = destroy_dp(dp_idx);
1562 case ODP_PORT_ATTACH:
1563 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1566 case ODP_PORT_DETACH:
1567 err = get_user(port_no, (int __user *)argp);
1569 err = detach_port(dp_idx, port_no);
1573 err = vport_add((struct odp_vport_add __user *)argp);
1577 err = vport_mod((struct odp_vport_mod __user *)argp);
1581 err = vport_del((char __user *)argp);
1584 case ODP_VPORT_STATS_GET:
1585 err = vport_stats_get((struct odp_vport_stats_req __user *)argp);
1588 case ODP_VPORT_ETHER_GET:
1589 err = vport_ether_get((struct odp_vport_ether __user *)argp);
1592 case ODP_VPORT_ETHER_SET:
1593 err = vport_ether_set((struct odp_vport_ether __user *)argp);
1596 case ODP_VPORT_MTU_GET:
1597 err = vport_mtu_get((struct odp_vport_mtu __user *)argp);
1600 case ODP_VPORT_MTU_SET:
1601 err = vport_mtu_set((struct odp_vport_mtu __user *)argp);
1605 dp = get_dp_locked(dp_idx);
1612 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1615 case ODP_GET_DROP_FRAGS:
1616 err = put_user(dp->drop_frags, (int __user *)argp);
1619 case ODP_SET_DROP_FRAGS:
1620 err = get_user(drop_frags, (int __user *)argp);
1624 if (drop_frags != 0 && drop_frags != 1)
1626 dp->drop_frags = drop_frags;
1630 case ODP_GET_LISTEN_MASK:
1631 err = put_user(get_listen_mask(f), (int __user *)argp);
1634 case ODP_SET_LISTEN_MASK:
1635 err = get_user(listeners, (int __user *)argp);
1639 if (listeners & ~ODPL_ALL)
1642 set_listen_mask(f, listeners);
1645 case ODP_GET_SFLOW_PROBABILITY:
1646 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1649 case ODP_SET_SFLOW_PROBABILITY:
1650 err = get_user(sflow_probability, (unsigned int __user *)argp);
1652 dp->sflow_probability = sflow_probability;
1655 case ODP_PORT_QUERY:
1656 err = query_port(dp, (struct odp_port __user *)argp);
1660 err = list_ports(dp, (struct odp_portvec __user *)argp);
1663 case ODP_PORT_GROUP_SET:
1664 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1667 case ODP_PORT_GROUP_GET:
1668 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1671 case ODP_FLOW_FLUSH:
1672 err = flush_flows(dp);
1676 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1680 err = del_flow(dp, (struct odp_flow __user *)argp);
1684 err = do_flowvec_ioctl(dp, argp, query_flows);
1688 err = do_flowvec_ioctl(dp, argp, list_flows);
1692 err = do_execute(dp, (struct odp_execute __user *)argp);
1699 mutex_unlock(&dp->mutex);
1704 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1707 for (i = 0; i < DP_N_QUEUES; i++) {
1708 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1714 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1717 /* XXX is there sufficient synchronization here? */
1718 int listeners = get_listen_mask(f);
1719 int dp_idx = iminor(f->f_dentry->d_inode);
1720 struct datapath *dp = get_dp(dp_idx);
1721 struct sk_buff *skb;
1722 struct iovec __user iov;
1729 if (nbytes == 0 || !listeners)
1735 for (i = 0; i < DP_N_QUEUES; i++) {
1736 if (listeners & (1 << i)) {
1737 skb = skb_dequeue(&dp->queues[i]);
1743 if (f->f_flags & O_NONBLOCK) {
1748 wait_event_interruptible(dp->waitqueue,
1749 dp_has_packet_of_interest(dp,
1752 if (signal_pending(current)) {
1753 retval = -ERESTARTSYS;
1758 copy_bytes = min_t(size_t, skb->len, nbytes);
1760 iov.iov_len = copy_bytes;
1761 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1763 retval = copy_bytes;
1770 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1772 /* XXX is there sufficient synchronization here? */
1773 int dp_idx = iminor(file->f_dentry->d_inode);
1774 struct datapath *dp = get_dp(dp_idx);
1779 poll_wait(file, &dp->waitqueue, wait);
1780 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1781 mask |= POLLIN | POLLRDNORM;
1783 mask = POLLIN | POLLRDNORM | POLLHUP;
1788 struct file_operations openvswitch_fops = {
1789 /* XXX .aio_read = openvswitch_aio_read, */
1790 .read = openvswitch_read,
1791 .poll = openvswitch_poll,
1792 .unlocked_ioctl = openvswitch_ioctl,
1793 /* XXX .fasync = openvswitch_fasync, */
1798 static int __init dp_init(void)
1800 struct sk_buff *dummy_skb;
1803 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
1805 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1813 goto error_flow_exit;
1815 err = register_netdevice_notifier(&dp_device_notifier);
1817 goto error_vport_exit;
1819 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1821 goto error_unreg_notifier;
1825 error_unreg_notifier:
1826 unregister_netdevice_notifier(&dp_device_notifier);
1835 static void dp_cleanup(void)
1838 unregister_chrdev(major, "openvswitch");
1839 unregister_netdevice_notifier(&dp_device_notifier);
1844 module_init(dp_init);
1845 module_exit(dp_cleanup);
1847 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1848 MODULE_LICENSE("GPL");