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 if (!(port.flags & ODP_PORT_INTERNAL))
432 set_internal_devs_mtu(dp);
433 dp_sysfs_add_if(dp->ports[port_no]);
435 err = __put_user(port_no, &portp->port);
438 mutex_unlock(&dp->mutex);
445 int dp_detach_port(struct dp_port *p, int may_delete)
447 struct vport *vport = p->vport;
452 if (p->port_no != ODPP_LOCAL)
454 dp_ifinfo_notify(RTM_DELLINK, p);
456 /* First drop references to device. */
458 list_del_rcu(&p->node);
459 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
461 err = vport_detach(vport);
465 /* Then wait until no one is still using it, and destroy it. */
469 const char *port_type = vport_get_type(vport);
471 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
478 kobject_put(&p->kobj);
483 static int detach_port(int dp_idx, int port_no)
490 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
494 dp = get_dp_locked(dp_idx);
497 goto out_unlock_rtnl;
499 p = dp->ports[port_no];
504 err = dp_detach_port(p, 1);
507 mutex_unlock(&dp->mutex);
514 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
515 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
517 struct datapath *dp = p->dp;
518 struct dp_stats_percpu *stats;
519 struct odp_flow_key key;
520 struct tbl_node *flow_node;
522 WARN_ON_ONCE(skb_shared(skb));
523 skb_warn_if_lro(skb);
525 OVS_CB(skb)->dp_port = p;
526 compute_ip_summed(skb, false);
528 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
529 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
531 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
532 if (dp->drop_frags) {
539 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
541 struct sw_flow *flow = flow_cast(flow_node);
542 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
543 flow_used(flow, skb);
544 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
549 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
553 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
554 /* This code is based on skb_checksum_setup() from Xen's net/dev/core.c. We
555 * can't call this function directly because it isn't exported in all
557 int vswitch_skb_checksum_setup(struct sk_buff *skb)
562 __u16 csum_start, csum_offset;
564 if (!skb->proto_csum_blank)
567 if (skb->protocol != htons(ETH_P_IP))
570 if (!pskb_may_pull(skb, skb_network_header(skb) + sizeof(struct iphdr) - skb->data))
574 th = skb_network_header(skb) + 4 * iph->ihl;
576 csum_start = th - skb->head;
577 switch (iph->protocol) {
579 csum_offset = offsetof(struct tcphdr, check);
582 csum_offset = offsetof(struct udphdr, check);
586 printk(KERN_ERR "Attempting to checksum a non-"
587 "TCP/UDP packet, dropping a protocol"
588 " %d packet", iph->protocol);
592 if (!pskb_may_pull(skb, th + csum_offset + 2 - skb->data))
595 skb->ip_summed = CHECKSUM_PARTIAL;
596 skb->proto_csum_blank = 0;
598 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
599 skb->csum_start = csum_start;
600 skb->csum_offset = csum_offset;
602 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
603 skb->csum = csum_offset;
611 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
613 /* Types of checksums that we can receive (these all refer to L4 checksums):
614 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
615 * (though not verified) checksum in packet but not in skb->csum. Packets
616 * from the bridge local port will also have this type.
617 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
618 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
619 * a valid skb->csum. Importantly, both contain a full checksum (not
620 * verified) in the packet itself. The only difference is that if the
621 * packet gets to L4 processing on this machine (not in DomU) we won't
622 * have to recompute the checksum to verify. Most hardware devices do not
623 * produce packets with this type, even if they support receive checksum
624 * offloading (they produce type #5).
625 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
626 * be computed if it is sent off box. Unfortunately on earlier kernels,
627 * this case is impossible to distinguish from #2, despite having opposite
628 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
629 * to distinguish the different states. The only real user of this type
630 * with bridging is Xen (on later kernels).
631 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
632 * generated locally by a Xen DomU and has a partial checksum. If it is
633 * handled on this machine (Dom0 or DomU), then the checksum will not be
634 * computed. If it goes off box, the checksum in the packet needs to be
635 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
636 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
637 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
638 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
639 * full checksum or using a protocol without a checksum. skb->csum is
640 * undefined. This is common from devices with receive checksum
641 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
642 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
644 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
645 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
646 * based on whether it is on the transmit or receive path. After the datapath
647 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
648 * checksum, we will panic. Since we can receive packets with checksums, we
649 * assume that all CHECKSUM_HW packets have checksums and map them to
650 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
651 * packet is processed by the local IP stack, in which case it will need to
652 * be reverified). If we receive a packet with CHECKSUM_HW that really means
653 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
654 * shouldn't be any devices that do this with bridging.
656 * The bridge has similar behavior and this function closely resembles
657 * skb_forward_csum(). It is slightly different because we are only concerned
658 * with bridging and not other types of forwarding and can get away with
659 * slightly more optimal behavior.*/
661 compute_ip_summed(struct sk_buff *skb, bool xmit)
663 /* For our convenience these defines change repeatedly between kernel
664 * versions, so we can't just copy them over... */
665 switch (skb->ip_summed) {
667 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
669 case CHECKSUM_UNNECESSARY:
670 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
673 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
674 * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
675 * uses some special fields to represent this (see below). Since we
676 * can only make one type work, pick the one that actually happens in
679 * The one exception to this is if we are on the transmit path
680 * (basically after skb_checksum_setup() has been run) the type has
681 * already been converted, so we should stay with that. */
684 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
686 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
690 case CHECKSUM_COMPLETE:
691 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
693 case CHECKSUM_PARTIAL:
694 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
698 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
700 /* None seems the safest... */
701 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
704 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
705 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
706 * kernels. It should not be set on the transmit path though. */
707 if (skb->proto_csum_blank)
708 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
710 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
715 forward_ip_summed(struct sk_buff *skb)
718 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
719 skb->ip_summed = CHECKSUM_NONE;
723 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
724 * unless we broke up a GSO packet. */
726 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
727 int queue_no, u32 arg)
729 struct sk_buff *nskb;
733 if (OVS_CB(skb)->dp_port)
734 port_no = OVS_CB(skb)->dp_port->port_no;
736 port_no = ODPP_LOCAL;
739 struct odp_msg *header;
744 /* If a checksum-deferred packet is forwarded to the
745 * controller, correct the pointers and checksum. This happens
746 * on a regular basis only on Xen, on which VMs can pass up
747 * packets that do not have their checksum computed.
749 err = vswitch_skb_checksum_setup(skb);
753 if (skb->ip_summed == CHECKSUM_PARTIAL) {
755 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
756 /* Until 2.6.22, the start of the transport header was
757 * also the start of data to be checksummed. Linux
758 * 2.6.22 introduced the csum_start field for this
759 * purpose, but we should point the transport header to
760 * it anyway for backward compatibility, as
761 * dev_queue_xmit() does even in 2.6.28. */
762 skb_set_transport_header(skb, skb->csum_start -
766 err = skb_checksum_help(skb);
771 err = skb_cow(skb, sizeof *header);
775 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
776 header->type = queue_no;
777 header->length = skb->len;
778 header->port = port_no;
779 header->reserved = 0;
781 skb_queue_tail(queue, skb);
789 while ((skb = nskb) != NULL) {
797 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
800 struct dp_stats_percpu *stats;
801 struct sk_buff_head *queue;
804 WARN_ON_ONCE(skb_shared(skb));
805 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
806 queue = &dp->queues[queue_no];
808 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
811 forward_ip_summed(skb);
813 /* Break apart GSO packets into their component pieces. Otherwise
814 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
815 if (skb_is_gso(skb)) {
816 struct sk_buff *nskb = skb_gso_segment(skb, 0);
820 if (unlikely(IS_ERR(skb))) {
825 /* XXX This case might not be possible. It's hard to
826 * tell from the skb_gso_segment() code and comment. */
830 err = queue_control_packets(skb, queue, queue_no, arg);
831 wake_up_interruptible(&dp->waitqueue);
837 stats = percpu_ptr(dp->stats_percpu, get_cpu());
844 static int flush_flows(struct datapath *dp)
846 struct tbl *old_table = rcu_dereference(dp->table);
847 struct tbl *new_table;
849 new_table = tbl_create(0);
853 rcu_assign_pointer(dp->table, new_table);
855 tbl_deferred_destroy(old_table, flow_free_tbl);
860 static int validate_actions(const struct sw_flow_actions *actions)
864 for (i = 0; i < actions->n_actions; i++) {
865 const union odp_action *a = &actions->actions[i];
868 if (a->output.port >= DP_MAX_PORTS)
872 case ODPAT_OUTPUT_GROUP:
873 if (a->output_group.group >= DP_MAX_GROUPS)
877 case ODPAT_SET_VLAN_VID:
878 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
882 case ODPAT_SET_VLAN_PCP:
883 if (a->vlan_pcp.vlan_pcp
884 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
888 case ODPAT_SET_NW_TOS:
889 if (a->nw_tos.nw_tos & INET_ECN_MASK)
894 if (a->type >= ODPAT_N_ACTIONS)
903 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
905 struct sw_flow_actions *actions;
908 actions = flow_actions_alloc(flow->n_actions);
909 error = PTR_ERR(actions);
914 if (copy_from_user(actions->actions, flow->actions,
915 flow->n_actions * sizeof(union odp_action)))
916 goto error_free_actions;
917 error = validate_actions(actions);
919 goto error_free_actions;
926 return ERR_PTR(error);
929 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
931 if (flow->used.tv_sec) {
932 stats->used_sec = flow->used.tv_sec;
933 stats->used_nsec = flow->used.tv_nsec;
936 stats->used_nsec = 0;
938 stats->n_packets = flow->packet_count;
939 stats->n_bytes = flow->byte_count;
940 stats->ip_tos = flow->ip_tos;
941 stats->tcp_flags = flow->tcp_flags;
945 static void clear_stats(struct sw_flow *flow)
947 flow->used.tv_sec = flow->used.tv_nsec = 0;
950 flow->packet_count = 0;
951 flow->byte_count = 0;
954 static int expand_table(struct datapath *dp)
956 struct tbl *old_table = rcu_dereference(dp->table);
957 struct tbl *new_table;
959 new_table = tbl_expand(old_table);
960 if (IS_ERR(new_table))
961 return PTR_ERR(new_table);
963 rcu_assign_pointer(dp->table, new_table);
964 tbl_deferred_destroy(old_table, NULL);
969 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
971 struct odp_flow_put uf;
972 struct tbl_node *flow_node;
973 struct sw_flow *flow;
975 struct odp_flow_stats stats;
979 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
981 memset(uf.flow.key.reserved, 0, sizeof uf.flow.key.reserved);
983 table = rcu_dereference(dp->table);
984 flow_node = tbl_lookup(table, &uf.flow.key, flow_hash(&uf.flow.key), flow_cmp);
987 struct sw_flow_actions *acts;
990 if (!(uf.flags & ODPPF_CREATE))
993 /* Expand table, if necessary, to make room. */
994 if (tbl_count(table) >= tbl_n_buckets(table)) {
995 error = expand_table(dp);
998 table = rcu_dereference(dp->table);
1001 /* Allocate flow. */
1003 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
1006 flow->key = uf.flow.key;
1007 spin_lock_init(&flow->lock);
1010 /* Obtain actions. */
1011 acts = get_actions(&uf.flow);
1012 error = PTR_ERR(acts);
1014 goto error_free_flow;
1015 rcu_assign_pointer(flow->sf_acts, acts);
1017 /* Put flow in bucket. */
1018 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1020 goto error_free_flow_acts;
1022 memset(&stats, 0, sizeof(struct odp_flow_stats));
1024 /* We found a matching flow. */
1025 struct sw_flow_actions *old_acts, *new_acts;
1026 unsigned long int flags;
1028 flow = flow_cast(flow_node);
1030 /* Bail out if we're not allowed to modify an existing flow. */
1032 if (!(uf.flags & ODPPF_MODIFY))
1036 new_acts = get_actions(&uf.flow);
1037 error = PTR_ERR(new_acts);
1038 if (IS_ERR(new_acts))
1040 old_acts = rcu_dereference(flow->sf_acts);
1041 if (old_acts->n_actions != new_acts->n_actions ||
1042 memcmp(old_acts->actions, new_acts->actions,
1043 sizeof(union odp_action) * old_acts->n_actions)) {
1044 rcu_assign_pointer(flow->sf_acts, new_acts);
1045 flow_deferred_free_acts(old_acts);
1050 /* Fetch stats, then clear them if necessary. */
1051 spin_lock_irqsave(&flow->lock, flags);
1052 get_stats(flow, &stats);
1053 if (uf.flags & ODPPF_ZERO_STATS)
1055 spin_unlock_irqrestore(&flow->lock, flags);
1058 /* Copy stats to userspace. */
1059 if (__copy_to_user(&ufp->flow.stats, &stats,
1060 sizeof(struct odp_flow_stats)))
1064 error_free_flow_acts:
1065 kfree(flow->sf_acts);
1067 kmem_cache_free(flow_cache, flow);
1072 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1074 union odp_action __user *actions;
1075 struct sw_flow_actions *sf_acts;
1078 if (__get_user(actions, &ufp->actions) ||
1079 __get_user(n_actions, &ufp->n_actions))
1085 sf_acts = rcu_dereference(flow->sf_acts);
1086 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1087 (actions && copy_to_user(actions, sf_acts->actions,
1088 sizeof(union odp_action) *
1089 min(sf_acts->n_actions, n_actions))))
1095 static int answer_query(struct sw_flow *flow, u32 query_flags,
1096 struct odp_flow __user *ufp)
1098 struct odp_flow_stats stats;
1099 unsigned long int flags;
1101 spin_lock_irqsave(&flow->lock, flags);
1102 get_stats(flow, &stats);
1104 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1105 flow->tcp_flags = 0;
1107 spin_unlock_irqrestore(&flow->lock, flags);
1109 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1111 return put_actions(flow, ufp);
1114 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1116 struct tbl *table = rcu_dereference(dp->table);
1118 struct tbl_node *flow_node;
1119 struct sw_flow *flow;
1123 if (copy_from_user(&uf, ufp, sizeof uf))
1125 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1127 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1132 error = tbl_remove(table, flow_node);
1136 /* XXX These statistics might lose a few packets, since other CPUs can
1137 * be using this flow. We used to synchronize_rcu() to make sure that
1138 * we get completely accurate stats, but that blows our performance,
1141 flow = flow_cast(flow_node);
1142 error = answer_query(flow, 0, ufp);
1143 flow_deferred_free(flow);
1149 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1151 struct tbl *table = rcu_dereference(dp->table);
1153 for (i = 0; i < flowvec->n_flows; i++) {
1154 struct __user odp_flow *ufp = &flowvec->flows[i];
1156 struct tbl_node *flow_node;
1159 if (__copy_from_user(&uf, ufp, sizeof uf))
1161 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1163 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1165 error = __put_user(ENOENT, &ufp->stats.error);
1167 error = answer_query(flow_cast(flow_node), uf.flags, ufp);
1171 return flowvec->n_flows;
1174 struct list_flows_cbdata {
1175 struct odp_flow __user *uflows;
1180 static int list_flow(struct tbl_node *node, void *cbdata_)
1182 struct sw_flow *flow = flow_cast(node);
1183 struct list_flows_cbdata *cbdata = cbdata_;
1184 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1187 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1189 error = answer_query(flow, 0, ufp);
1193 if (cbdata->listed_flows >= cbdata->n_flows)
1194 return cbdata->listed_flows;
1198 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1200 struct list_flows_cbdata cbdata;
1203 if (!flowvec->n_flows)
1206 cbdata.uflows = flowvec->flows;
1207 cbdata.n_flows = flowvec->n_flows;
1208 cbdata.listed_flows = 0;
1209 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1210 return error ? error : cbdata.listed_flows;
1213 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1214 int (*function)(struct datapath *,
1215 const struct odp_flowvec *))
1217 struct odp_flowvec __user *uflowvec;
1218 struct odp_flowvec flowvec;
1221 uflowvec = (struct odp_flowvec __user *)argp;
1222 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1223 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1226 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1229 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1230 flowvec.n_flows * sizeof(struct odp_flow)))
1233 retval = function(dp, &flowvec);
1234 return (retval < 0 ? retval
1235 : retval == flowvec.n_flows ? 0
1236 : __put_user(retval, &uflowvec->n_flows));
1239 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1241 struct odp_execute execute;
1242 struct odp_flow_key key;
1243 struct sk_buff *skb;
1244 struct sw_flow_actions *actions;
1249 if (copy_from_user(&execute, executep, sizeof execute))
1253 if (execute.length < ETH_HLEN || execute.length > 65535)
1257 actions = flow_actions_alloc(execute.n_actions);
1262 if (copy_from_user(actions->actions, execute.actions,
1263 execute.n_actions * sizeof *execute.actions))
1264 goto error_free_actions;
1266 err = validate_actions(actions);
1268 goto error_free_actions;
1271 skb = alloc_skb(execute.length, GFP_KERNEL);
1273 goto error_free_actions;
1275 if (execute.in_port < DP_MAX_PORTS)
1276 OVS_CB(skb)->dp_port = dp->ports[execute.in_port];
1278 OVS_CB(skb)->dp_port = NULL;
1281 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1283 goto error_free_skb;
1285 skb_reset_mac_header(skb);
1288 /* Normally, setting the skb 'protocol' field would be handled by a
1289 * call to eth_type_trans(), but it assumes there's a sending
1290 * device, which we may not have. */
1291 if (ntohs(eth->h_proto) >= 1536)
1292 skb->protocol = eth->h_proto;
1294 skb->protocol = htons(ETH_P_802_2);
1296 flow_extract(skb, execute.in_port, &key);
1297 err = execute_actions(dp, skb, &key, actions->actions,
1298 actions->n_actions, GFP_KERNEL);
1310 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1312 struct tbl *table = rcu_dereference(dp->table);
1313 struct odp_stats stats;
1316 stats.n_flows = tbl_count(table);
1317 stats.cur_capacity = tbl_n_buckets(table);
1318 stats.max_capacity = TBL_MAX_BUCKETS;
1319 stats.n_ports = dp->n_ports;
1320 stats.max_ports = DP_MAX_PORTS;
1321 stats.max_groups = DP_MAX_GROUPS;
1322 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1323 for_each_possible_cpu(i) {
1324 const struct dp_stats_percpu *s;
1325 s = percpu_ptr(dp->stats_percpu, i);
1326 stats.n_frags += s->n_frags;
1327 stats.n_hit += s->n_hit;
1328 stats.n_missed += s->n_missed;
1329 stats.n_lost += s->n_lost;
1331 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1332 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1333 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1336 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1337 int dp_min_mtu(const struct datapath *dp)
1344 list_for_each_entry_rcu (p, &dp->port_list, node) {
1347 /* Skip any internal ports, since that's what we're trying to
1349 if (is_internal_vport(p->vport))
1352 dev_mtu = vport_get_mtu(p->vport);
1353 if (!mtu || dev_mtu < mtu)
1357 return mtu ? mtu : ETH_DATA_LEN;
1360 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1361 * be called with RTNL lock and dp_mutex. */
1362 void set_internal_devs_mtu(const struct datapath *dp)
1369 mtu = dp_min_mtu(dp);
1371 list_for_each_entry_rcu (p, &dp->port_list, node) {
1372 if (is_internal_vport(p->vport))
1373 vport_set_mtu(p->vport, mtu);
1378 put_port(const struct dp_port *p, struct odp_port __user *uop)
1382 memset(&op, 0, sizeof op);
1385 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1388 op.port = p->port_no;
1389 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1391 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1395 query_port(struct datapath *dp, struct odp_port __user *uport)
1397 struct odp_port port;
1399 if (copy_from_user(&port, uport, sizeof port))
1402 if (port.devname[0]) {
1403 struct vport *vport;
1404 struct dp_port *dp_port;
1407 port.devname[IFNAMSIZ - 1] = '\0';
1412 vport = vport_locate(port.devname);
1418 dp_port = vport_get_dp_port(vport);
1419 if (!dp_port || dp_port->dp != dp) {
1424 port.port = dp_port->port_no;
1433 if (port.port >= DP_MAX_PORTS)
1435 if (!dp->ports[port.port])
1439 return put_port(dp->ports[port.port], uport);
1443 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1445 struct odp_portvec pv;
1449 if (copy_from_user(&pv, pvp, sizeof pv))
1454 list_for_each_entry_rcu (p, &dp->port_list, node) {
1455 if (put_port(p, &pv.ports[idx]))
1457 if (idx++ >= pv.n_ports)
1461 return put_user(dp->n_ports, &pvp->n_ports);
1464 /* RCU callback for freeing a dp_port_group */
1465 static void free_port_group(struct rcu_head *rcu)
1467 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1472 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1474 struct odp_port_group pg;
1475 struct dp_port_group *new_group, *old_group;
1479 if (copy_from_user(&pg, upg, sizeof pg))
1483 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1487 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1492 new_group->n_ports = pg.n_ports;
1494 if (copy_from_user(new_group->ports, pg.ports,
1495 sizeof(u16) * pg.n_ports))
1498 old_group = rcu_dereference(dp->groups[pg.group]);
1499 rcu_assign_pointer(dp->groups[pg.group], new_group);
1501 call_rcu(&old_group->rcu, free_port_group);
1511 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1513 struct odp_port_group pg;
1514 struct dp_port_group *g;
1517 if (copy_from_user(&pg, upg, sizeof pg))
1520 if (pg.group >= DP_MAX_GROUPS)
1523 g = dp->groups[pg.group];
1524 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1525 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1528 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1534 static int get_listen_mask(const struct file *f)
1536 return (long)f->private_data;
1539 static void set_listen_mask(struct file *f, int listen_mask)
1541 f->private_data = (void*)(long)listen_mask;
1544 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1547 int dp_idx = iminor(f->f_dentry->d_inode);
1548 struct datapath *dp;
1549 int drop_frags, listeners, port_no;
1550 unsigned int sflow_probability;
1553 /* Handle commands with special locking requirements up front. */
1556 err = create_dp(dp_idx, (char __user *)argp);
1559 case ODP_DP_DESTROY:
1560 err = destroy_dp(dp_idx);
1563 case ODP_PORT_ATTACH:
1564 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1567 case ODP_PORT_DETACH:
1568 err = get_user(port_no, (int __user *)argp);
1570 err = detach_port(dp_idx, port_no);
1574 err = vport_add((struct odp_vport_add __user *)argp);
1578 err = vport_mod((struct odp_vport_mod __user *)argp);
1582 err = vport_del((char __user *)argp);
1585 case ODP_VPORT_STATS_GET:
1586 err = vport_stats_get((struct odp_vport_stats_req __user *)argp);
1589 case ODP_VPORT_ETHER_GET:
1590 err = vport_ether_get((struct odp_vport_ether __user *)argp);
1593 case ODP_VPORT_ETHER_SET:
1594 err = vport_ether_set((struct odp_vport_ether __user *)argp);
1597 case ODP_VPORT_MTU_GET:
1598 err = vport_mtu_get((struct odp_vport_mtu __user *)argp);
1601 case ODP_VPORT_MTU_SET:
1602 err = vport_mtu_set((struct odp_vport_mtu __user *)argp);
1606 dp = get_dp_locked(dp_idx);
1613 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1616 case ODP_GET_DROP_FRAGS:
1617 err = put_user(dp->drop_frags, (int __user *)argp);
1620 case ODP_SET_DROP_FRAGS:
1621 err = get_user(drop_frags, (int __user *)argp);
1625 if (drop_frags != 0 && drop_frags != 1)
1627 dp->drop_frags = drop_frags;
1631 case ODP_GET_LISTEN_MASK:
1632 err = put_user(get_listen_mask(f), (int __user *)argp);
1635 case ODP_SET_LISTEN_MASK:
1636 err = get_user(listeners, (int __user *)argp);
1640 if (listeners & ~ODPL_ALL)
1643 set_listen_mask(f, listeners);
1646 case ODP_GET_SFLOW_PROBABILITY:
1647 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1650 case ODP_SET_SFLOW_PROBABILITY:
1651 err = get_user(sflow_probability, (unsigned int __user *)argp);
1653 dp->sflow_probability = sflow_probability;
1656 case ODP_PORT_QUERY:
1657 err = query_port(dp, (struct odp_port __user *)argp);
1661 err = list_ports(dp, (struct odp_portvec __user *)argp);
1664 case ODP_PORT_GROUP_SET:
1665 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1668 case ODP_PORT_GROUP_GET:
1669 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1672 case ODP_FLOW_FLUSH:
1673 err = flush_flows(dp);
1677 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1681 err = del_flow(dp, (struct odp_flow __user *)argp);
1685 err = do_flowvec_ioctl(dp, argp, query_flows);
1689 err = do_flowvec_ioctl(dp, argp, list_flows);
1693 err = do_execute(dp, (struct odp_execute __user *)argp);
1700 mutex_unlock(&dp->mutex);
1705 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1708 for (i = 0; i < DP_N_QUEUES; i++) {
1709 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1715 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1718 /* XXX is there sufficient synchronization here? */
1719 int listeners = get_listen_mask(f);
1720 int dp_idx = iminor(f->f_dentry->d_inode);
1721 struct datapath *dp = get_dp(dp_idx);
1722 struct sk_buff *skb;
1723 struct iovec __user iov;
1730 if (nbytes == 0 || !listeners)
1736 for (i = 0; i < DP_N_QUEUES; i++) {
1737 if (listeners & (1 << i)) {
1738 skb = skb_dequeue(&dp->queues[i]);
1744 if (f->f_flags & O_NONBLOCK) {
1749 wait_event_interruptible(dp->waitqueue,
1750 dp_has_packet_of_interest(dp,
1753 if (signal_pending(current)) {
1754 retval = -ERESTARTSYS;
1759 copy_bytes = min_t(size_t, skb->len, nbytes);
1761 iov.iov_len = copy_bytes;
1762 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1764 retval = copy_bytes;
1771 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1773 /* XXX is there sufficient synchronization here? */
1774 int dp_idx = iminor(file->f_dentry->d_inode);
1775 struct datapath *dp = get_dp(dp_idx);
1780 poll_wait(file, &dp->waitqueue, wait);
1781 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1782 mask |= POLLIN | POLLRDNORM;
1784 mask = POLLIN | POLLRDNORM | POLLHUP;
1789 struct file_operations openvswitch_fops = {
1790 /* XXX .aio_read = openvswitch_aio_read, */
1791 .read = openvswitch_read,
1792 .poll = openvswitch_poll,
1793 .unlocked_ioctl = openvswitch_ioctl,
1794 /* XXX .fasync = openvswitch_fasync, */
1799 static int __init dp_init(void)
1801 struct sk_buff *dummy_skb;
1804 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
1806 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1814 goto error_flow_exit;
1816 err = register_netdevice_notifier(&dp_device_notifier);
1818 goto error_vport_exit;
1820 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1822 goto error_unreg_notifier;
1826 error_unreg_notifier:
1827 unregister_netdevice_notifier(&dp_device_notifier);
1836 static void dp_cleanup(void)
1839 unregister_chrdev(major, "openvswitch");
1840 unregister_netdevice_notifier(&dp_device_notifier);
1845 module_init(dp_init);
1846 module_exit(dp_cleanup);
1848 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1849 MODULE_LICENSE("GPL");