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];
206 if (strncpy_from_user(devname, devnamep, IFNAMSIZ - 1) < 0)
208 devname[IFNAMSIZ - 1] = '\0';
210 snprintf(devname, sizeof devname, "of%d", dp_idx);
214 mutex_lock(&dp_mutex);
216 if (!try_module_get(THIS_MODULE))
219 /* Exit early if a datapath with that number already exists.
220 * (We don't use -EEXIST because that's ambiguous with 'devname'
221 * conflicting with an existing network device name.) */
227 dp = kzalloc(sizeof *dp, GFP_KERNEL);
230 INIT_LIST_HEAD(&dp->port_list);
231 mutex_init(&dp->mutex);
233 for (i = 0; i < DP_N_QUEUES; i++)
234 skb_queue_head_init(&dp->queues[i]);
235 init_waitqueue_head(&dp->waitqueue);
237 /* Initialize kobject for bridge. This will be added as
238 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
239 dp->ifobj.kset = NULL;
240 kobject_init(&dp->ifobj, &dp_ktype);
242 /* Allocate table. */
244 rcu_assign_pointer(dp->table, tbl_create(0));
248 /* Set up our datapath device. */
249 strncpy(internal_dev_port.devname, devname, IFNAMSIZ - 1);
250 internal_dev_port.flags = ODP_PORT_INTERNAL;
251 err = new_dp_port(dp, &internal_dev_port, ODPP_LOCAL);
256 goto err_destroy_table;
260 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
261 if (!dp->stats_percpu)
262 goto err_destroy_local_port;
264 rcu_assign_pointer(dps[dp_idx], dp);
265 mutex_unlock(&dp_mutex);
272 err_destroy_local_port:
273 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
275 tbl_destroy(dp->table, NULL);
279 module_put(THIS_MODULE);
281 mutex_unlock(&dp_mutex);
287 static void do_destroy_dp(struct datapath *dp)
289 struct dp_port *p, *n;
292 list_for_each_entry_safe (p, n, &dp->port_list, node)
293 if (p->port_no != ODPP_LOCAL)
294 dp_detach_port(p, 1);
298 rcu_assign_pointer(dps[dp->dp_idx], NULL);
300 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
302 tbl_destroy(dp->table, flow_free_tbl);
304 for (i = 0; i < DP_N_QUEUES; i++)
305 skb_queue_purge(&dp->queues[i]);
306 for (i = 0; i < DP_MAX_GROUPS; i++)
307 kfree(dp->groups[i]);
308 free_percpu(dp->stats_percpu);
309 kobject_put(&dp->ifobj);
310 module_put(THIS_MODULE);
313 static int destroy_dp(int dp_idx)
319 mutex_lock(&dp_mutex);
329 mutex_unlock(&dp_mutex);
334 static void release_dp_port(struct kobject *kobj)
336 struct dp_port *p = container_of(kobj, struct dp_port, kobj);
340 static struct kobj_type brport_ktype = {
342 .sysfs_ops = &brport_sysfs_ops,
344 .release = release_dp_port
347 /* Called with RTNL lock and dp_mutex. */
348 static int new_dp_port(struct datapath *dp, struct odp_port *odp_port, int port_no)
354 vport = vport_locate(odp_port->devname);
358 if (odp_port->flags & ODP_PORT_INTERNAL)
359 vport = __vport_add(odp_port->devname, "internal", NULL);
361 vport = __vport_add(odp_port->devname, "netdev", NULL);
366 return PTR_ERR(vport);
369 p = kzalloc(sizeof(*p), GFP_KERNEL);
373 p->port_no = port_no;
375 atomic_set(&p->sflow_pool, 0);
377 err = vport_attach(vport, p);
383 rcu_assign_pointer(dp->ports[port_no], p);
384 list_add_rcu(&p->node, &dp->port_list);
387 /* Initialize kobject for bridge. This will be added as
388 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
390 kobject_init(&p->kobj, &brport_ktype);
392 dp_ifinfo_notify(RTM_NEWLINK, p);
397 static int attach_port(int dp_idx, struct odp_port __user *portp)
400 struct odp_port port;
405 if (copy_from_user(&port, portp, sizeof port))
407 port.devname[IFNAMSIZ - 1] = '\0';
410 dp = get_dp_locked(dp_idx);
413 goto out_unlock_rtnl;
415 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
416 if (!dp->ports[port_no])
422 err = new_dp_port(dp, &port, port_no);
426 if (!(port.flags & ODP_PORT_INTERNAL))
427 set_internal_devs_mtu(dp);
428 dp_sysfs_add_if(dp->ports[port_no]);
430 err = __put_user(port_no, &portp->port);
433 mutex_unlock(&dp->mutex);
440 int dp_detach_port(struct dp_port *p, int may_delete)
442 struct vport *vport = p->vport;
447 if (p->port_no != ODPP_LOCAL)
449 dp_ifinfo_notify(RTM_DELLINK, p);
451 /* First drop references to device. */
453 list_del_rcu(&p->node);
454 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
456 err = vport_detach(vport);
460 /* Then wait until no one is still using it, and destroy it. */
464 const char *port_type = vport_get_type(vport);
466 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
473 kobject_put(&p->kobj);
478 static int detach_port(int dp_idx, int port_no)
485 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
489 dp = get_dp_locked(dp_idx);
492 goto out_unlock_rtnl;
494 p = dp->ports[port_no];
499 err = dp_detach_port(p, 1);
502 mutex_unlock(&dp->mutex);
509 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
510 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
512 struct datapath *dp = p->dp;
513 struct dp_stats_percpu *stats;
514 struct odp_flow_key key;
515 struct tbl_node *flow_node;
517 WARN_ON_ONCE(skb_shared(skb));
518 skb_warn_if_lro(skb);
520 OVS_CB(skb)->dp_port = p;
521 compute_ip_summed(skb, false);
523 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
524 stats = percpu_ptr(dp->stats_percpu, smp_processor_id());
526 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
527 if (dp->drop_frags) {
534 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
536 struct sw_flow *flow = flow_cast(flow_node);
537 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
538 flow_used(flow, skb);
539 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
544 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
548 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
549 /* This code is based on a skb_checksum_setup from net/dev/core.c from a
550 * combination of Lenny's 2.6.26 Xen kernel and Xen's
551 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call this function
552 * directly because it isn't exported in all versions. */
553 static int skb_pull_up_to(struct sk_buff *skb, void *ptr)
555 if (ptr < (void *)skb->tail)
557 if (__pskb_pull_tail(skb,
558 ptr - (void *)skb->data - skb_headlen(skb))) {
565 int vswitch_skb_checksum_setup(struct sk_buff *skb)
570 __u16 csum_start, csum_offset;
572 if (!skb->proto_csum_blank)
575 if (skb->protocol != htons(ETH_P_IP))
578 if (!skb_pull_up_to(skb, skb_network_header(skb) + sizeof(struct iphdr)))
582 th = skb_network_header(skb) + 4 * iph->ihl;
584 csum_start = th - skb->head;
585 switch (iph->protocol) {
587 csum_offset = offsetof(struct tcphdr, check);
590 csum_offset = offsetof(struct udphdr, check);
594 printk(KERN_ERR "Attempting to checksum a non-"
595 "TCP/UDP packet, dropping a protocol"
596 " %d packet", iph->protocol);
600 if (!skb_pull_up_to(skb, th + csum_offset + 2))
603 skb->ip_summed = CHECKSUM_PARTIAL;
604 skb->proto_csum_blank = 0;
606 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
607 skb->csum_start = csum_start;
608 skb->csum_offset = csum_offset;
610 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
611 skb->csum = csum_offset;
619 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
621 /* Types of checksums that we can receive (these all refer to L4 checksums):
622 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
623 * (though not verified) checksum in packet but not in skb->csum. Packets
624 * from the bridge local port will also have this type.
625 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
626 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
627 * a valid skb->csum. Importantly, both contain a full checksum (not
628 * verified) in the packet itself. The only difference is that if the
629 * packet gets to L4 processing on this machine (not in DomU) we won't
630 * have to recompute the checksum to verify. Most hardware devices do not
631 * produce packets with this type, even if they support receive checksum
632 * offloading (they produce type #5).
633 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
634 * be computed if it is sent off box. Unfortunately on earlier kernels,
635 * this case is impossible to distinguish from #2, despite having opposite
636 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
637 * to distinguish the different states. The only real user of this type
638 * with bridging is Xen (on later kernels).
639 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
640 * generated locally by a Xen DomU and has a partial checksum. If it is
641 * handled on this machine (Dom0 or DomU), then the checksum will not be
642 * computed. If it goes off box, the checksum in the packet needs to be
643 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
644 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
645 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
646 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
647 * full checksum or using a protocol without a checksum. skb->csum is
648 * undefined. This is common from devices with receive checksum
649 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
650 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
652 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
653 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
654 * based on whether it is on the transmit or receive path. After the datapath
655 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
656 * checksum, we will panic. Since we can receive packets with checksums, we
657 * assume that all CHECKSUM_HW packets have checksums and map them to
658 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
659 * packet is processed by the local IP stack, in which case it will need to
660 * be reverified). If we receive a packet with CHECKSUM_HW that really means
661 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
662 * shouldn't be any devices that do this with bridging.
664 * The bridge has similar behavior and this function closely resembles
665 * skb_forward_csum(). It is slightly different because we are only concerned
666 * with bridging and not other types of forwarding and can get away with
667 * slightly more optimal behavior.*/
669 compute_ip_summed(struct sk_buff *skb, bool xmit)
671 /* For our convenience these defines change repeatedly between kernel
672 * versions, so we can't just copy them over... */
673 switch (skb->ip_summed) {
675 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
677 case CHECKSUM_UNNECESSARY:
678 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
681 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
682 * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
683 * uses some special fields to represent this (see below). Since we
684 * can only make one type work, pick the one that actually happens in
687 * The one exception to this is if we are on the transmit path
688 * (basically after skb_checksum_setup() has been run) the type has
689 * already been converted, so we should stay with that. */
692 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
694 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
698 case CHECKSUM_COMPLETE:
699 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
701 case CHECKSUM_PARTIAL:
702 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
706 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
708 /* None seems the safest... */
709 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
712 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
713 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
714 * kernels. It should not be set on the transmit path though. */
715 if (skb->proto_csum_blank)
716 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
718 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
723 forward_ip_summed(struct sk_buff *skb)
726 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
727 skb->ip_summed = CHECKSUM_NONE;
731 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
732 * unless we broke up a GSO packet. */
734 queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
735 int queue_no, u32 arg)
737 struct sk_buff *nskb;
741 if (OVS_CB(skb)->dp_port)
742 port_no = OVS_CB(skb)->dp_port->port_no;
744 port_no = ODPP_LOCAL;
747 struct odp_msg *header;
752 /* If a checksum-deferred packet is forwarded to the
753 * controller, correct the pointers and checksum. This happens
754 * on a regular basis only on Xen, on which VMs can pass up
755 * packets that do not have their checksum computed.
757 err = vswitch_skb_checksum_setup(skb);
761 if (skb->ip_summed == CHECKSUM_PARTIAL) {
762 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
763 /* Until 2.6.22, the start of the transport header was
764 * also the start of data to be checksummed. Linux
765 * 2.6.22 introduced the csum_start field for this
766 * purpose, but we should point the transport header to
767 * it anyway for backward compatibility, as
768 * dev_queue_xmit() does even in 2.6.28. */
769 skb_set_transport_header(skb, skb->csum_start -
772 err = skb_checksum_help(skb);
777 if (skb->ip_summed == CHECKSUM_HW) {
778 err = skb_checksum_help(skb, 0);
784 err = skb_cow(skb, sizeof *header);
788 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
789 header->type = queue_no;
790 header->length = skb->len;
791 header->port = port_no;
792 header->reserved = 0;
794 skb_queue_tail(queue, skb);
802 while ((skb = nskb) != NULL) {
810 dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
813 struct dp_stats_percpu *stats;
814 struct sk_buff_head *queue;
817 WARN_ON_ONCE(skb_shared(skb));
818 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
819 queue = &dp->queues[queue_no];
821 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
824 forward_ip_summed(skb);
826 /* Break apart GSO packets into their component pieces. Otherwise
827 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
828 if (skb_is_gso(skb)) {
829 struct sk_buff *nskb = skb_gso_segment(skb, 0);
833 if (unlikely(IS_ERR(skb))) {
838 /* XXX This case might not be possible. It's hard to
839 * tell from the skb_gso_segment() code and comment. */
843 err = queue_control_packets(skb, queue, queue_no, arg);
844 wake_up_interruptible(&dp->waitqueue);
850 stats = percpu_ptr(dp->stats_percpu, get_cpu());
857 static int flush_flows(struct datapath *dp)
859 struct tbl *old_table = rcu_dereference(dp->table);
860 struct tbl *new_table;
862 new_table = tbl_create(0);
866 rcu_assign_pointer(dp->table, new_table);
868 tbl_deferred_destroy(old_table, flow_free_tbl);
873 static int validate_actions(const struct sw_flow_actions *actions)
877 for (i = 0; i < actions->n_actions; i++) {
878 const union odp_action *a = &actions->actions[i];
881 if (a->output.port >= DP_MAX_PORTS)
885 case ODPAT_OUTPUT_GROUP:
886 if (a->output_group.group >= DP_MAX_GROUPS)
890 case ODPAT_SET_VLAN_VID:
891 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
895 case ODPAT_SET_VLAN_PCP:
896 if (a->vlan_pcp.vlan_pcp
897 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
901 case ODPAT_SET_NW_TOS:
902 if (a->nw_tos.nw_tos & INET_ECN_MASK)
907 if (a->type >= ODPAT_N_ACTIONS)
916 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
918 struct sw_flow_actions *actions;
921 actions = flow_actions_alloc(flow->n_actions);
922 error = PTR_ERR(actions);
927 if (copy_from_user(actions->actions, flow->actions,
928 flow->n_actions * sizeof(union odp_action)))
929 goto error_free_actions;
930 error = validate_actions(actions);
932 goto error_free_actions;
939 return ERR_PTR(error);
942 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats)
944 if (flow->used.tv_sec) {
945 stats->used_sec = flow->used.tv_sec;
946 stats->used_nsec = flow->used.tv_nsec;
949 stats->used_nsec = 0;
951 stats->n_packets = flow->packet_count;
952 stats->n_bytes = flow->byte_count;
953 stats->ip_tos = flow->ip_tos;
954 stats->tcp_flags = flow->tcp_flags;
958 static void clear_stats(struct sw_flow *flow)
960 flow->used.tv_sec = flow->used.tv_nsec = 0;
963 flow->packet_count = 0;
964 flow->byte_count = 0;
967 static int expand_table(struct datapath *dp)
969 struct tbl *old_table = rcu_dereference(dp->table);
970 struct tbl *new_table;
972 new_table = tbl_expand(old_table);
973 if (IS_ERR(new_table))
974 return PTR_ERR(new_table);
976 rcu_assign_pointer(dp->table, new_table);
977 tbl_deferred_destroy(old_table, NULL);
982 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
984 struct odp_flow_put uf;
985 struct tbl_node *flow_node;
986 struct sw_flow *flow;
988 struct odp_flow_stats stats;
992 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
994 memset(uf.flow.key.reserved, 0, sizeof uf.flow.key.reserved);
996 table = rcu_dereference(dp->table);
997 flow_node = tbl_lookup(table, &uf.flow.key, flow_hash(&uf.flow.key), flow_cmp);
1000 struct sw_flow_actions *acts;
1003 if (!(uf.flags & ODPPF_CREATE))
1006 /* Expand table, if necessary, to make room. */
1007 if (tbl_count(table) >= tbl_n_buckets(table)) {
1008 error = expand_table(dp);
1011 table = rcu_dereference(dp->table);
1014 /* Allocate flow. */
1016 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
1019 flow->key = uf.flow.key;
1020 spin_lock_init(&flow->lock);
1023 /* Obtain actions. */
1024 acts = get_actions(&uf.flow);
1025 error = PTR_ERR(acts);
1027 goto error_free_flow;
1028 rcu_assign_pointer(flow->sf_acts, acts);
1030 /* Put flow in bucket. */
1031 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1033 goto error_free_flow_acts;
1035 memset(&stats, 0, sizeof(struct odp_flow_stats));
1037 /* We found a matching flow. */
1038 struct sw_flow_actions *old_acts, *new_acts;
1039 unsigned long int flags;
1041 flow = flow_cast(flow_node);
1043 /* Bail out if we're not allowed to modify an existing flow. */
1045 if (!(uf.flags & ODPPF_MODIFY))
1049 new_acts = get_actions(&uf.flow);
1050 error = PTR_ERR(new_acts);
1051 if (IS_ERR(new_acts))
1053 old_acts = rcu_dereference(flow->sf_acts);
1054 if (old_acts->n_actions != new_acts->n_actions ||
1055 memcmp(old_acts->actions, new_acts->actions,
1056 sizeof(union odp_action) * old_acts->n_actions)) {
1057 rcu_assign_pointer(flow->sf_acts, new_acts);
1058 flow_deferred_free_acts(old_acts);
1063 /* Fetch stats, then clear them if necessary. */
1064 spin_lock_irqsave(&flow->lock, flags);
1065 get_stats(flow, &stats);
1066 if (uf.flags & ODPPF_ZERO_STATS)
1068 spin_unlock_irqrestore(&flow->lock, flags);
1071 /* Copy stats to userspace. */
1072 if (__copy_to_user(&ufp->flow.stats, &stats,
1073 sizeof(struct odp_flow_stats)))
1077 error_free_flow_acts:
1078 kfree(flow->sf_acts);
1080 kmem_cache_free(flow_cache, flow);
1085 static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp)
1087 union odp_action __user *actions;
1088 struct sw_flow_actions *sf_acts;
1091 if (__get_user(actions, &ufp->actions) ||
1092 __get_user(n_actions, &ufp->n_actions))
1098 sf_acts = rcu_dereference(flow->sf_acts);
1099 if (__put_user(sf_acts->n_actions, &ufp->n_actions) ||
1100 (actions && copy_to_user(actions, sf_acts->actions,
1101 sizeof(union odp_action) *
1102 min(sf_acts->n_actions, n_actions))))
1108 static int answer_query(struct sw_flow *flow, u32 query_flags,
1109 struct odp_flow __user *ufp)
1111 struct odp_flow_stats stats;
1112 unsigned long int flags;
1114 spin_lock_irqsave(&flow->lock, flags);
1115 get_stats(flow, &stats);
1117 if (query_flags & ODPFF_ZERO_TCP_FLAGS) {
1118 flow->tcp_flags = 0;
1120 spin_unlock_irqrestore(&flow->lock, flags);
1122 if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats)))
1124 return put_actions(flow, ufp);
1127 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1129 struct tbl *table = rcu_dereference(dp->table);
1131 struct tbl_node *flow_node;
1132 struct sw_flow *flow;
1136 if (copy_from_user(&uf, ufp, sizeof uf))
1138 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1140 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1145 error = tbl_remove(table, flow_node);
1149 /* XXX These statistics might lose a few packets, since other CPUs can
1150 * be using this flow. We used to synchronize_rcu() to make sure that
1151 * we get completely accurate stats, but that blows our performance,
1154 flow = flow_cast(flow_node);
1155 error = answer_query(flow, 0, ufp);
1156 flow_deferred_free(flow);
1162 static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1164 struct tbl *table = rcu_dereference(dp->table);
1166 for (i = 0; i < flowvec->n_flows; i++) {
1167 struct __user odp_flow *ufp = &flowvec->flows[i];
1169 struct tbl_node *flow_node;
1172 if (__copy_from_user(&uf, ufp, sizeof uf))
1174 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1176 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1178 error = __put_user(ENOENT, &ufp->stats.error);
1180 error = answer_query(flow_cast(flow_node), uf.flags, ufp);
1184 return flowvec->n_flows;
1187 struct list_flows_cbdata {
1188 struct odp_flow __user *uflows;
1193 static int list_flow(struct tbl_node *node, void *cbdata_)
1195 struct sw_flow *flow = flow_cast(node);
1196 struct list_flows_cbdata *cbdata = cbdata_;
1197 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1200 if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1202 error = answer_query(flow, 0, ufp);
1206 if (cbdata->listed_flows >= cbdata->n_flows)
1207 return cbdata->listed_flows;
1211 static int list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1213 struct list_flows_cbdata cbdata;
1216 if (!flowvec->n_flows)
1219 cbdata.uflows = flowvec->flows;
1220 cbdata.n_flows = flowvec->n_flows;
1221 cbdata.listed_flows = 0;
1222 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1223 return error ? error : cbdata.listed_flows;
1226 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1227 int (*function)(struct datapath *,
1228 const struct odp_flowvec *))
1230 struct odp_flowvec __user *uflowvec;
1231 struct odp_flowvec flowvec;
1234 uflowvec = (struct odp_flowvec __user *)argp;
1235 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1236 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1239 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1242 if (!access_ok(VERIFY_WRITE, flowvec.flows,
1243 flowvec.n_flows * sizeof(struct odp_flow)))
1246 retval = function(dp, &flowvec);
1247 return (retval < 0 ? retval
1248 : retval == flowvec.n_flows ? 0
1249 : __put_user(retval, &uflowvec->n_flows));
1252 static int do_execute(struct datapath *dp, const struct odp_execute *executep)
1254 struct odp_execute execute;
1255 struct odp_flow_key key;
1256 struct sk_buff *skb;
1257 struct sw_flow_actions *actions;
1262 if (copy_from_user(&execute, executep, sizeof execute))
1266 if (execute.length < ETH_HLEN || execute.length > 65535)
1270 actions = flow_actions_alloc(execute.n_actions);
1275 if (copy_from_user(actions->actions, execute.actions,
1276 execute.n_actions * sizeof *execute.actions))
1277 goto error_free_actions;
1279 err = validate_actions(actions);
1281 goto error_free_actions;
1284 skb = alloc_skb(execute.length, GFP_KERNEL);
1286 goto error_free_actions;
1288 if (execute.in_port < DP_MAX_PORTS)
1289 OVS_CB(skb)->dp_port = dp->ports[execute.in_port];
1291 OVS_CB(skb)->dp_port = NULL;
1294 if (copy_from_user(skb_put(skb, execute.length), execute.data,
1296 goto error_free_skb;
1298 skb_reset_mac_header(skb);
1301 /* Normally, setting the skb 'protocol' field would be handled by a
1302 * call to eth_type_trans(), but it assumes there's a sending
1303 * device, which we may not have. */
1304 if (ntohs(eth->h_proto) >= 1536)
1305 skb->protocol = eth->h_proto;
1307 skb->protocol = htons(ETH_P_802_2);
1309 flow_extract(skb, execute.in_port, &key);
1310 err = execute_actions(dp, skb, &key, actions->actions,
1311 actions->n_actions, GFP_KERNEL);
1323 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1325 struct tbl *table = rcu_dereference(dp->table);
1326 struct odp_stats stats;
1329 stats.n_flows = tbl_count(table);
1330 stats.cur_capacity = tbl_n_buckets(table);
1331 stats.max_capacity = TBL_MAX_BUCKETS;
1332 stats.n_ports = dp->n_ports;
1333 stats.max_ports = DP_MAX_PORTS;
1334 stats.max_groups = DP_MAX_GROUPS;
1335 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1336 for_each_possible_cpu(i) {
1337 const struct dp_stats_percpu *s;
1338 s = percpu_ptr(dp->stats_percpu, i);
1339 stats.n_frags += s->n_frags;
1340 stats.n_hit += s->n_hit;
1341 stats.n_missed += s->n_missed;
1342 stats.n_lost += s->n_lost;
1344 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1345 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1346 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1349 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1350 int dp_min_mtu(const struct datapath *dp)
1357 list_for_each_entry_rcu (p, &dp->port_list, node) {
1360 /* Skip any internal ports, since that's what we're trying to
1362 if (is_internal_vport(p->vport))
1365 dev_mtu = vport_get_mtu(p->vport);
1366 if (!mtu || dev_mtu < mtu)
1370 return mtu ? mtu : ETH_DATA_LEN;
1373 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1374 * be called with RTNL lock and dp_mutex. */
1375 void set_internal_devs_mtu(const struct datapath *dp)
1382 mtu = dp_min_mtu(dp);
1384 list_for_each_entry_rcu (p, &dp->port_list, node) {
1385 if (is_internal_vport(p->vport))
1386 vport_set_mtu(p->vport, mtu);
1391 put_port(const struct dp_port *p, struct odp_port __user *uop)
1395 memset(&op, 0, sizeof op);
1398 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1401 op.port = p->port_no;
1402 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1404 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1408 query_port(struct datapath *dp, struct odp_port __user *uport)
1410 struct odp_port port;
1412 if (copy_from_user(&port, uport, sizeof port))
1415 if (port.devname[0]) {
1416 struct vport *vport;
1417 struct dp_port *dp_port;
1420 port.devname[IFNAMSIZ - 1] = '\0';
1425 vport = vport_locate(port.devname);
1431 dp_port = vport_get_dp_port(vport);
1432 if (!dp_port || dp_port->dp != dp) {
1437 port.port = dp_port->port_no;
1446 if (port.port >= DP_MAX_PORTS)
1448 if (!dp->ports[port.port])
1452 return put_port(dp->ports[port.port], uport);
1456 list_ports(struct datapath *dp, struct odp_portvec __user *pvp)
1458 struct odp_portvec pv;
1462 if (copy_from_user(&pv, pvp, sizeof pv))
1467 list_for_each_entry_rcu (p, &dp->port_list, node) {
1468 if (put_port(p, &pv.ports[idx]))
1470 if (idx++ >= pv.n_ports)
1474 return put_user(dp->n_ports, &pvp->n_ports);
1477 /* RCU callback for freeing a dp_port_group */
1478 static void free_port_group(struct rcu_head *rcu)
1480 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1485 set_port_group(struct datapath *dp, const struct odp_port_group __user *upg)
1487 struct odp_port_group pg;
1488 struct dp_port_group *new_group, *old_group;
1492 if (copy_from_user(&pg, upg, sizeof pg))
1496 if (pg.n_ports > DP_MAX_PORTS || pg.group >= DP_MAX_GROUPS)
1500 new_group = kmalloc(sizeof *new_group + sizeof(u16) * pg.n_ports,
1505 new_group->n_ports = pg.n_ports;
1507 if (copy_from_user(new_group->ports, pg.ports,
1508 sizeof(u16) * pg.n_ports))
1511 old_group = rcu_dereference(dp->groups[pg.group]);
1512 rcu_assign_pointer(dp->groups[pg.group], new_group);
1514 call_rcu(&old_group->rcu, free_port_group);
1524 get_port_group(struct datapath *dp, struct odp_port_group *upg)
1526 struct odp_port_group pg;
1527 struct dp_port_group *g;
1530 if (copy_from_user(&pg, upg, sizeof pg))
1533 if (pg.group >= DP_MAX_GROUPS)
1536 g = dp->groups[pg.group];
1537 n_copy = g ? min_t(int, g->n_ports, pg.n_ports) : 0;
1538 if (n_copy && copy_to_user(pg.ports, g->ports, n_copy * sizeof(u16)))
1541 if (put_user(g ? g->n_ports : 0, &upg->n_ports))
1547 static int get_listen_mask(const struct file *f)
1549 return (long)f->private_data;
1552 static void set_listen_mask(struct file *f, int listen_mask)
1554 f->private_data = (void*)(long)listen_mask;
1557 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1560 int dp_idx = iminor(f->f_dentry->d_inode);
1561 struct datapath *dp;
1562 int drop_frags, listeners, port_no;
1563 unsigned int sflow_probability;
1566 /* Handle commands with special locking requirements up front. */
1569 err = create_dp(dp_idx, (char __user *)argp);
1572 case ODP_DP_DESTROY:
1573 err = destroy_dp(dp_idx);
1576 case ODP_PORT_ATTACH:
1577 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1580 case ODP_PORT_DETACH:
1581 err = get_user(port_no, (int __user *)argp);
1583 err = detach_port(dp_idx, port_no);
1587 err = vport_add((struct odp_vport_add __user *)argp);
1591 err = vport_mod((struct odp_vport_mod __user *)argp);
1595 err = vport_del((char __user *)argp);
1598 case ODP_VPORT_STATS_GET:
1599 err = vport_stats_get((struct odp_vport_stats_req __user *)argp);
1602 case ODP_VPORT_ETHER_GET:
1603 err = vport_ether_get((struct odp_vport_ether __user *)argp);
1606 case ODP_VPORT_ETHER_SET:
1607 err = vport_ether_set((struct odp_vport_ether __user *)argp);
1610 case ODP_VPORT_MTU_GET:
1611 err = vport_mtu_get((struct odp_vport_mtu __user *)argp);
1614 case ODP_VPORT_MTU_SET:
1615 err = vport_mtu_set((struct odp_vport_mtu __user *)argp);
1619 dp = get_dp_locked(dp_idx);
1626 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1629 case ODP_GET_DROP_FRAGS:
1630 err = put_user(dp->drop_frags, (int __user *)argp);
1633 case ODP_SET_DROP_FRAGS:
1634 err = get_user(drop_frags, (int __user *)argp);
1638 if (drop_frags != 0 && drop_frags != 1)
1640 dp->drop_frags = drop_frags;
1644 case ODP_GET_LISTEN_MASK:
1645 err = put_user(get_listen_mask(f), (int __user *)argp);
1648 case ODP_SET_LISTEN_MASK:
1649 err = get_user(listeners, (int __user *)argp);
1653 if (listeners & ~ODPL_ALL)
1656 set_listen_mask(f, listeners);
1659 case ODP_GET_SFLOW_PROBABILITY:
1660 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1663 case ODP_SET_SFLOW_PROBABILITY:
1664 err = get_user(sflow_probability, (unsigned int __user *)argp);
1666 dp->sflow_probability = sflow_probability;
1669 case ODP_PORT_QUERY:
1670 err = query_port(dp, (struct odp_port __user *)argp);
1674 err = list_ports(dp, (struct odp_portvec __user *)argp);
1677 case ODP_PORT_GROUP_SET:
1678 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1681 case ODP_PORT_GROUP_GET:
1682 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1685 case ODP_FLOW_FLUSH:
1686 err = flush_flows(dp);
1690 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1694 err = del_flow(dp, (struct odp_flow __user *)argp);
1698 err = do_flowvec_ioctl(dp, argp, query_flows);
1702 err = do_flowvec_ioctl(dp, argp, list_flows);
1706 err = do_execute(dp, (struct odp_execute __user *)argp);
1713 mutex_unlock(&dp->mutex);
1718 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1721 for (i = 0; i < DP_N_QUEUES; i++) {
1722 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1728 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
1731 /* XXX is there sufficient synchronization here? */
1732 int listeners = get_listen_mask(f);
1733 int dp_idx = iminor(f->f_dentry->d_inode);
1734 struct datapath *dp = get_dp(dp_idx);
1735 struct sk_buff *skb;
1736 struct iovec __user iov;
1743 if (nbytes == 0 || !listeners)
1749 for (i = 0; i < DP_N_QUEUES; i++) {
1750 if (listeners & (1 << i)) {
1751 skb = skb_dequeue(&dp->queues[i]);
1757 if (f->f_flags & O_NONBLOCK) {
1762 wait_event_interruptible(dp->waitqueue,
1763 dp_has_packet_of_interest(dp,
1766 if (signal_pending(current)) {
1767 retval = -ERESTARTSYS;
1772 copy_bytes = min_t(size_t, skb->len, nbytes);
1774 iov.iov_len = copy_bytes;
1775 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
1777 retval = copy_bytes;
1784 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
1786 /* XXX is there sufficient synchronization here? */
1787 int dp_idx = iminor(file->f_dentry->d_inode);
1788 struct datapath *dp = get_dp(dp_idx);
1793 poll_wait(file, &dp->waitqueue, wait);
1794 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
1795 mask |= POLLIN | POLLRDNORM;
1797 mask = POLLIN | POLLRDNORM | POLLHUP;
1802 struct file_operations openvswitch_fops = {
1803 /* XXX .aio_read = openvswitch_aio_read, */
1804 .read = openvswitch_read,
1805 .poll = openvswitch_poll,
1806 .unlocked_ioctl = openvswitch_ioctl,
1807 /* XXX .fasync = openvswitch_fasync, */
1812 static int __init dp_init(void)
1814 struct sk_buff *dummy_skb;
1817 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
1819 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
1827 goto error_flow_exit;
1829 err = register_netdevice_notifier(&dp_device_notifier);
1831 goto error_vport_exit;
1833 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
1835 goto error_unreg_notifier;
1839 error_unreg_notifier:
1840 unregister_netdevice_notifier(&dp_device_notifier);
1849 static void dp_cleanup(void)
1852 unregister_chrdev(major, "openvswitch");
1853 unregister_netdevice_notifier(&dp_device_notifier);
1858 module_init(dp_init);
1859 module_exit(dp_cleanup);
1861 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1862 MODULE_LICENSE("GPL");