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/highmem.h>
36 #include <linux/netfilter_bridge.h>
37 #include <linux/netfilter_ipv4.h>
38 #include <linux/inetdevice.h>
39 #include <linux/list.h>
40 #include <linux/rculist.h>
41 #include <linux/workqueue.h>
42 #include <linux/dmi.h>
43 #include <net/inet_ecn.h>
44 #include <linux/compat.h>
46 #include "openvswitch/datapath-protocol.h"
50 #include "odp-compat.h"
52 #include "vport-internal_dev.h"
57 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
58 EXPORT_SYMBOL(dp_ioctl_hook);
60 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
63 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
66 * It is safe to access the datapath and dp_port structures with just
69 static struct datapath *dps[ODP_MAX];
70 static DEFINE_MUTEX(dp_mutex);
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;
381 atomic_set(&p->sflow_pool, 0);
383 err = vport_attach(vport, p);
389 rcu_assign_pointer(dp->ports[port_no], p);
390 list_add_rcu(&p->node, &dp->port_list);
393 /* Initialize kobject for bridge. This will be added as
394 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
396 kobject_init(&p->kobj, &brport_ktype);
398 dp_ifinfo_notify(RTM_NEWLINK, p);
403 static int attach_port(int dp_idx, struct odp_port __user *portp)
406 struct odp_port port;
411 if (copy_from_user(&port, portp, sizeof port))
413 port.devname[IFNAMSIZ - 1] = '\0';
416 dp = get_dp_locked(dp_idx);
419 goto out_unlock_rtnl;
421 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
422 if (!dp->ports[port_no])
428 err = new_dp_port(dp, &port, port_no);
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. */
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 int stats_counter_off;
520 struct odp_flow_key key;
521 struct tbl_node *flow_node;
523 OVS_CB(skb)->dp_port = p;
525 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
526 if (dp->drop_frags) {
528 stats_counter_off = offsetof(struct dp_stats_percpu, n_frags);
533 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
535 struct sw_flow *flow = flow_cast(flow_node);
536 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
537 flow_used(flow, skb);
538 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
540 stats_counter_off = offsetof(struct dp_stats_percpu, n_hit);
542 stats_counter_off = offsetof(struct dp_stats_percpu, n_missed);
543 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
548 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
549 (*(u64 *)((u8 *)stats + stats_counter_off))++;
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.
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. */
654 void compute_ip_summed(struct sk_buff *skb, bool xmit)
656 /* For our convenience these defines change repeatedly between kernel
657 * versions, so we can't just copy them over... */
658 switch (skb->ip_summed) {
660 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
662 case CHECKSUM_UNNECESSARY:
663 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
666 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
667 * However, on the receive side we should only get CHECKSUM_PARTIAL
668 * packets from Xen, which uses some special fields to represent this
669 * (see below). Since we can only make one type work, pick the one
670 * that actually happens in practice.
672 * On the transmit side (basically after skb_checksum_setup()
673 * has been run or on internal dev transmit), packets with
674 * CHECKSUM_COMPLETE aren't generated, so assume CHECKSUM_PARTIAL. */
677 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
679 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
683 case CHECKSUM_COMPLETE:
684 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
686 case CHECKSUM_PARTIAL:
687 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
691 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
693 /* None seems the safest... */
694 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
697 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
698 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
699 * kernels. It should not be set on the transmit path though. */
700 if (skb->proto_csum_blank)
701 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
703 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
707 /* This function closely resembles skb_forward_csum() used by the bridge. It
708 * is slightly different because we are only concerned with bridging and not
709 * other types of forwarding and can get away with slightly more optimal
711 void forward_ip_summed(struct sk_buff *skb)
714 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
715 skb->ip_summed = CHECKSUM_NONE;
719 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
720 * unless we broke up a GSO packet. */
721 static int queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
722 int queue_no, u32 arg)
724 struct sk_buff *nskb;
728 if (OVS_CB(skb)->dp_port)
729 port_no = OVS_CB(skb)->dp_port->port_no;
731 port_no = ODPP_LOCAL;
734 struct odp_msg *header;
739 err = skb_cow(skb, sizeof *header);
743 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
744 header->type = queue_no;
745 header->length = skb->len;
746 header->port = port_no;
747 header->reserved = 0;
749 skb_queue_tail(queue, skb);
757 while ((skb = nskb) != NULL) {
764 int dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
767 struct dp_stats_percpu *stats;
768 struct sk_buff_head *queue;
771 WARN_ON_ONCE(skb_shared(skb));
772 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
773 queue = &dp->queues[queue_no];
775 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
778 forward_ip_summed(skb);
780 err = vswitch_skb_checksum_setup(skb);
784 /* Break apart GSO packets into their component pieces. Otherwise
785 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
786 if (skb_is_gso(skb)) {
787 struct sk_buff *nskb = skb_gso_segment(skb, NETIF_F_SG | NETIF_F_HW_CSUM);
791 if (unlikely(IS_ERR(skb))) {
796 /* XXX This case might not be possible. It's hard to
797 * tell from the skb_gso_segment() code and comment. */
801 err = queue_control_packets(skb, queue, queue_no, arg);
802 wake_up_interruptible(&dp->waitqueue);
809 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
816 static int flush_flows(struct datapath *dp)
818 struct tbl *old_table = rcu_dereference(dp->table);
819 struct tbl *new_table;
821 new_table = tbl_create(0);
825 rcu_assign_pointer(dp->table, new_table);
827 tbl_deferred_destroy(old_table, flow_free_tbl);
832 static int validate_actions(const struct sw_flow_actions *actions)
836 for (i = 0; i < actions->n_actions; i++) {
837 const union odp_action *a = &actions->actions[i];
840 if (a->output.port >= DP_MAX_PORTS)
844 case ODPAT_OUTPUT_GROUP:
845 if (a->output_group.group >= DP_MAX_GROUPS)
849 case ODPAT_SET_VLAN_VID:
850 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
854 case ODPAT_SET_VLAN_PCP:
855 if (a->vlan_pcp.vlan_pcp
856 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
860 case ODPAT_SET_NW_TOS:
861 if (a->nw_tos.nw_tos & INET_ECN_MASK)
866 if (a->type >= ODPAT_N_ACTIONS)
875 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
877 struct sw_flow_actions *actions;
880 actions = flow_actions_alloc(flow->n_actions);
881 error = PTR_ERR(actions);
886 if (copy_from_user(actions->actions, flow->actions,
887 flow->n_actions * sizeof(union odp_action)))
888 goto error_free_actions;
889 error = validate_actions(actions);
891 goto error_free_actions;
898 return ERR_PTR(error);
901 static struct timespec get_time_offset(void)
903 struct timespec now_mono, now_jiffies;
905 ktime_get_ts(&now_mono);
906 jiffies_to_timespec(jiffies, &now_jiffies);
907 return timespec_sub(now_mono, now_jiffies);
910 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats,
911 struct timespec time_offset)
914 struct timespec flow_ts, used;
916 jiffies_to_timespec(flow->used, &flow_ts);
917 set_normalized_timespec(&used, flow_ts.tv_sec + time_offset.tv_sec,
918 flow_ts.tv_nsec + time_offset.tv_nsec);
920 stats->used_sec = used.tv_sec;
921 stats->used_nsec = used.tv_nsec;
924 stats->used_nsec = 0;
927 stats->n_packets = flow->packet_count;
928 stats->n_bytes = flow->byte_count;
930 stats->tcp_flags = flow->tcp_flags;
934 static void clear_stats(struct sw_flow *flow)
938 flow->packet_count = 0;
939 flow->byte_count = 0;
942 static int expand_table(struct datapath *dp)
944 struct tbl *old_table = rcu_dereference(dp->table);
945 struct tbl *new_table;
947 new_table = tbl_expand(old_table);
948 if (IS_ERR(new_table))
949 return PTR_ERR(new_table);
951 rcu_assign_pointer(dp->table, new_table);
952 tbl_deferred_destroy(old_table, NULL);
957 static int do_put_flow(struct datapath *dp, struct odp_flow_put *uf,
958 struct odp_flow_stats *stats)
960 struct tbl_node *flow_node;
961 struct sw_flow *flow;
965 memset(uf->flow.key.reserved, 0, sizeof uf->flow.key.reserved);
967 table = rcu_dereference(dp->table);
968 flow_node = tbl_lookup(table, &uf->flow.key, flow_hash(&uf->flow.key), flow_cmp);
971 struct sw_flow_actions *acts;
974 if (!(uf->flags & ODPPF_CREATE))
977 /* Expand table, if necessary, to make room. */
978 if (tbl_count(table) >= tbl_n_buckets(table)) {
979 error = expand_table(dp);
982 table = rcu_dereference(dp->table);
987 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
990 flow->key = uf->flow.key;
991 spin_lock_init(&flow->lock);
994 /* Obtain actions. */
995 acts = get_actions(&uf->flow);
996 error = PTR_ERR(acts);
998 goto error_free_flow;
999 rcu_assign_pointer(flow->sf_acts, acts);
1001 /* Put flow in bucket. */
1002 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1004 goto error_free_flow_acts;
1006 memset(stats, 0, sizeof(struct odp_flow_stats));
1008 /* We found a matching flow. */
1009 struct sw_flow_actions *old_acts, *new_acts;
1011 flow = flow_cast(flow_node);
1013 /* Bail out if we're not allowed to modify an existing flow. */
1015 if (!(uf->flags & ODPPF_MODIFY))
1019 new_acts = get_actions(&uf->flow);
1020 error = PTR_ERR(new_acts);
1021 if (IS_ERR(new_acts))
1023 old_acts = rcu_dereference(flow->sf_acts);
1024 if (old_acts->n_actions != new_acts->n_actions ||
1025 memcmp(old_acts->actions, new_acts->actions,
1026 sizeof(union odp_action) * old_acts->n_actions)) {
1027 rcu_assign_pointer(flow->sf_acts, new_acts);
1028 flow_deferred_free_acts(old_acts);
1033 /* Fetch stats, then clear them if necessary. */
1034 spin_lock_bh(&flow->lock);
1035 get_stats(flow, stats, get_time_offset());
1036 if (uf->flags & ODPPF_ZERO_STATS)
1038 spin_unlock_bh(&flow->lock);
1043 error_free_flow_acts:
1044 kfree(flow->sf_acts);
1046 kmem_cache_free(flow_cache, flow);
1051 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
1053 struct odp_flow_stats stats;
1054 struct odp_flow_put uf;
1057 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
1060 error = do_put_flow(dp, &uf, &stats);
1064 if (copy_to_user(&ufp->flow.stats, &stats,
1065 sizeof(struct odp_flow_stats)))
1071 static int do_answer_query(struct sw_flow *flow, u32 query_flags,
1072 struct timespec time_offset,
1073 struct odp_flow_stats __user *ustats,
1074 union odp_action __user *actions,
1075 u32 __user *n_actionsp)
1077 struct sw_flow_actions *sf_acts;
1078 struct odp_flow_stats stats;
1081 spin_lock_bh(&flow->lock);
1082 get_stats(flow, &stats, time_offset);
1083 if (query_flags & ODPFF_ZERO_TCP_FLAGS)
1084 flow->tcp_flags = 0;
1086 spin_unlock_bh(&flow->lock);
1088 if (copy_to_user(ustats, &stats, sizeof(struct odp_flow_stats)) ||
1089 get_user(n_actions, n_actionsp))
1095 sf_acts = rcu_dereference(flow->sf_acts);
1096 if (put_user(sf_acts->n_actions, n_actionsp) ||
1097 (actions && copy_to_user(actions, sf_acts->actions,
1098 sizeof(union odp_action) *
1099 min(sf_acts->n_actions, n_actions))))
1105 static int answer_query(struct sw_flow *flow, u32 query_flags,
1106 struct timespec time_offset,
1107 struct odp_flow __user *ufp)
1109 union odp_action *actions;
1111 if (get_user(actions, &ufp->actions))
1114 return do_answer_query(flow, query_flags, time_offset,
1115 &ufp->stats, actions, &ufp->n_actions);
1118 static struct sw_flow *do_del_flow(struct datapath *dp, struct odp_flow_key *key)
1120 struct tbl *table = rcu_dereference(dp->table);
1121 struct tbl_node *flow_node;
1124 memset(key->reserved, 0, sizeof key->reserved);
1125 flow_node = tbl_lookup(table, key, flow_hash(key), flow_cmp);
1127 return ERR_PTR(-ENOENT);
1129 error = tbl_remove(table, flow_node);
1131 return ERR_PTR(error);
1133 /* XXX Returned flow_node's statistics might lose a few packets, since
1134 * other CPUs can be using this flow. We used to synchronize_rcu() to
1135 * make sure that we get completely accurate stats, but that blows our
1136 * performance, badly. */
1137 return flow_cast(flow_node);
1140 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1142 struct sw_flow *flow;
1146 if (copy_from_user(&uf, ufp, sizeof uf))
1149 flow = do_del_flow(dp, &uf.key);
1151 return PTR_ERR(flow);
1153 error = answer_query(flow, 0, get_time_offset(), ufp);
1154 flow_deferred_free(flow);
1158 static int do_query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1160 struct tbl *table = rcu_dereference(dp->table);
1161 struct timespec time_offset;
1164 time_offset = get_time_offset();
1166 for (i = 0; i < flowvec->n_flows; i++) {
1167 struct odp_flow __user *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, time_offset, ufp);
1184 return flowvec->n_flows;
1187 struct list_flows_cbdata {
1188 struct odp_flow __user *uflows;
1191 struct timespec time_offset;
1194 static int list_flow(struct tbl_node *node, void *cbdata_)
1196 struct sw_flow *flow = flow_cast(node);
1197 struct list_flows_cbdata *cbdata = cbdata_;
1198 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1201 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1203 error = answer_query(flow, 0, cbdata->time_offset, ufp);
1207 if (cbdata->listed_flows >= cbdata->n_flows)
1208 return cbdata->listed_flows;
1212 static int do_list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1214 struct list_flows_cbdata cbdata;
1217 if (!flowvec->n_flows)
1220 cbdata.uflows = flowvec->flows;
1221 cbdata.n_flows = flowvec->n_flows;
1222 cbdata.listed_flows = 0;
1223 cbdata.time_offset = get_time_offset();
1225 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1226 return error ? error : cbdata.listed_flows;
1229 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1230 int (*function)(struct datapath *,
1231 const struct odp_flowvec *))
1233 struct odp_flowvec __user *uflowvec;
1234 struct odp_flowvec flowvec;
1237 uflowvec = (struct odp_flowvec __user *)argp;
1238 if (copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1241 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1244 retval = function(dp, &flowvec);
1245 return (retval < 0 ? retval
1246 : retval == flowvec.n_flows ? 0
1247 : put_user(retval, &uflowvec->n_flows));
1250 static int do_execute(struct datapath *dp, const struct odp_execute *execute)
1252 struct odp_flow_key key;
1253 struct sk_buff *skb;
1254 struct sw_flow_actions *actions;
1259 if (execute->length < ETH_HLEN || execute->length > 65535)
1263 actions = flow_actions_alloc(execute->n_actions);
1268 if (copy_from_user(actions->actions, execute->actions,
1269 execute->n_actions * sizeof *execute->actions))
1270 goto error_free_actions;
1272 err = validate_actions(actions);
1274 goto error_free_actions;
1277 skb = alloc_skb(execute->length, GFP_KERNEL);
1279 goto error_free_actions;
1281 if (execute->in_port < DP_MAX_PORTS)
1282 OVS_CB(skb)->dp_port = dp->ports[execute->in_port];
1284 OVS_CB(skb)->dp_port = NULL;
1287 if (copy_from_user(skb_put(skb, execute->length), execute->data,
1289 goto error_free_skb;
1291 skb_reset_mac_header(skb);
1294 /* Normally, setting the skb 'protocol' field would be handled by a
1295 * call to eth_type_trans(), but it assumes there's a sending
1296 * device, which we may not have. */
1297 if (ntohs(eth->h_proto) >= 1536)
1298 skb->protocol = eth->h_proto;
1300 skb->protocol = htons(ETH_P_802_2);
1302 flow_extract(skb, execute->in_port, &key);
1305 err = execute_actions(dp, skb, &key, actions->actions,
1306 actions->n_actions, GFP_KERNEL);
1320 static int execute_packet(struct datapath *dp, const struct odp_execute __user *executep)
1322 struct odp_execute execute;
1324 if (copy_from_user(&execute, executep, sizeof execute))
1327 return do_execute(dp, &execute);
1330 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1332 struct tbl *table = rcu_dereference(dp->table);
1333 struct odp_stats stats;
1336 stats.n_flows = tbl_count(table);
1337 stats.cur_capacity = tbl_n_buckets(table);
1338 stats.max_capacity = TBL_MAX_BUCKETS;
1339 stats.n_ports = dp->n_ports;
1340 stats.max_ports = DP_MAX_PORTS;
1341 stats.max_groups = DP_MAX_GROUPS;
1342 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1343 for_each_possible_cpu(i) {
1344 const struct dp_stats_percpu *s;
1345 s = per_cpu_ptr(dp->stats_percpu, i);
1346 stats.n_frags += s->n_frags;
1347 stats.n_hit += s->n_hit;
1348 stats.n_missed += s->n_missed;
1349 stats.n_lost += s->n_lost;
1351 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1352 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1353 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1356 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1357 int dp_min_mtu(const struct datapath *dp)
1364 list_for_each_entry_rcu (p, &dp->port_list, node) {
1367 /* Skip any internal ports, since that's what we're trying to
1369 if (is_internal_vport(p->vport))
1372 dev_mtu = vport_get_mtu(p->vport);
1373 if (!mtu || dev_mtu < mtu)
1377 return mtu ? mtu : ETH_DATA_LEN;
1380 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1381 * be called with RTNL lock. */
1382 void set_internal_devs_mtu(const struct datapath *dp)
1389 mtu = dp_min_mtu(dp);
1391 list_for_each_entry_rcu (p, &dp->port_list, node) {
1392 if (is_internal_vport(p->vport))
1393 vport_set_mtu(p->vport, mtu);
1397 static int put_port(const struct dp_port *p, struct odp_port __user *uop)
1401 memset(&op, 0, sizeof op);
1404 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1407 op.port = p->port_no;
1408 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1410 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1413 static int query_port(struct datapath *dp, struct odp_port __user *uport)
1415 struct odp_port port;
1417 if (copy_from_user(&port, uport, sizeof port))
1420 if (port.devname[0]) {
1421 struct vport *vport;
1422 struct dp_port *dp_port;
1425 port.devname[IFNAMSIZ - 1] = '\0';
1430 vport = vport_locate(port.devname);
1436 dp_port = vport_get_dp_port(vport);
1437 if (!dp_port || dp_port->dp != dp) {
1442 port.port = dp_port->port_no;
1451 if (port.port >= DP_MAX_PORTS)
1453 if (!dp->ports[port.port])
1457 return put_port(dp->ports[port.port], uport);
1460 static int do_list_ports(struct datapath *dp, struct odp_port __user *uports,
1467 list_for_each_entry_rcu (p, &dp->port_list, node) {
1468 if (put_port(p, &uports[idx]))
1470 if (idx++ >= n_ports)
1477 static int list_ports(struct datapath *dp, struct odp_portvec __user *upv)
1479 struct odp_portvec pv;
1482 if (copy_from_user(&pv, upv, sizeof pv))
1485 retval = do_list_ports(dp, pv.ports, pv.n_ports);
1489 return put_user(retval, &upv->n_ports);
1492 /* RCU callback for freeing a dp_port_group */
1493 static void free_port_group(struct rcu_head *rcu)
1495 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1499 static int do_set_port_group(struct datapath *dp, u16 __user *ports,
1500 int n_ports, int group)
1502 struct dp_port_group *new_group, *old_group;
1506 if (n_ports > DP_MAX_PORTS || group >= DP_MAX_GROUPS)
1510 new_group = kmalloc(sizeof *new_group + sizeof(u16) * n_ports, GFP_KERNEL);
1514 new_group->n_ports = n_ports;
1516 if (copy_from_user(new_group->ports, ports, sizeof(u16) * n_ports))
1519 old_group = rcu_dereference(dp->groups[group]);
1520 rcu_assign_pointer(dp->groups[group], new_group);
1522 call_rcu(&old_group->rcu, free_port_group);
1531 static int set_port_group(struct datapath *dp,
1532 const struct odp_port_group __user *upg)
1534 struct odp_port_group pg;
1536 if (copy_from_user(&pg, upg, sizeof pg))
1539 return do_set_port_group(dp, pg.ports, pg.n_ports, pg.group);
1542 static int do_get_port_group(struct datapath *dp,
1543 u16 __user *ports, int n_ports, int group,
1544 u16 __user *n_portsp)
1546 struct dp_port_group *g;
1549 if (group >= DP_MAX_GROUPS)
1552 g = dp->groups[group];
1553 n_copy = g ? min_t(int, g->n_ports, n_ports) : 0;
1554 if (n_copy && copy_to_user(ports, g->ports, n_copy * sizeof(u16)))
1557 if (put_user(g ? g->n_ports : 0, n_portsp))
1563 static int get_port_group(struct datapath *dp, struct odp_port_group __user *upg)
1565 struct odp_port_group pg;
1567 if (copy_from_user(&pg, upg, sizeof pg))
1570 return do_get_port_group(dp, pg.ports, pg.n_ports, pg.group, &upg->n_ports);
1573 static int get_listen_mask(const struct file *f)
1575 return (long)f->private_data;
1578 static void set_listen_mask(struct file *f, int listen_mask)
1580 f->private_data = (void*)(long)listen_mask;
1583 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1586 int dp_idx = iminor(f->f_dentry->d_inode);
1587 struct datapath *dp;
1588 int drop_frags, listeners, port_no;
1589 unsigned int sflow_probability;
1592 /* Handle commands with special locking requirements up front. */
1595 err = create_dp(dp_idx, (char __user *)argp);
1598 case ODP_DP_DESTROY:
1599 err = destroy_dp(dp_idx);
1602 case ODP_PORT_ATTACH:
1603 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1606 case ODP_PORT_DETACH:
1607 err = get_user(port_no, (int __user *)argp);
1609 err = detach_port(dp_idx, port_no);
1613 err = vport_user_add((struct odp_vport_add __user *)argp);
1617 err = vport_user_mod((struct odp_vport_mod __user *)argp);
1621 err = vport_user_del((char __user *)argp);
1624 case ODP_VPORT_STATS_GET:
1625 err = vport_user_stats_get((struct odp_vport_stats_req __user *)argp);
1628 case ODP_VPORT_STATS_SET:
1629 err = vport_user_stats_set((struct odp_vport_stats_req __user *)argp);
1632 case ODP_VPORT_ETHER_GET:
1633 err = vport_user_ether_get((struct odp_vport_ether __user *)argp);
1636 case ODP_VPORT_ETHER_SET:
1637 err = vport_user_ether_set((struct odp_vport_ether __user *)argp);
1640 case ODP_VPORT_MTU_GET:
1641 err = vport_user_mtu_get((struct odp_vport_mtu __user *)argp);
1644 case ODP_VPORT_MTU_SET:
1645 err = vport_user_mtu_set((struct odp_vport_mtu __user *)argp);
1649 dp = get_dp_locked(dp_idx);
1656 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1659 case ODP_GET_DROP_FRAGS:
1660 err = put_user(dp->drop_frags, (int __user *)argp);
1663 case ODP_SET_DROP_FRAGS:
1664 err = get_user(drop_frags, (int __user *)argp);
1668 if (drop_frags != 0 && drop_frags != 1)
1670 dp->drop_frags = drop_frags;
1674 case ODP_GET_LISTEN_MASK:
1675 err = put_user(get_listen_mask(f), (int __user *)argp);
1678 case ODP_SET_LISTEN_MASK:
1679 err = get_user(listeners, (int __user *)argp);
1683 if (listeners & ~ODPL_ALL)
1686 set_listen_mask(f, listeners);
1689 case ODP_GET_SFLOW_PROBABILITY:
1690 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1693 case ODP_SET_SFLOW_PROBABILITY:
1694 err = get_user(sflow_probability, (unsigned int __user *)argp);
1696 dp->sflow_probability = sflow_probability;
1699 case ODP_PORT_QUERY:
1700 err = query_port(dp, (struct odp_port __user *)argp);
1704 err = list_ports(dp, (struct odp_portvec __user *)argp);
1707 case ODP_PORT_GROUP_SET:
1708 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1711 case ODP_PORT_GROUP_GET:
1712 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1715 case ODP_FLOW_FLUSH:
1716 err = flush_flows(dp);
1720 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1724 err = del_flow(dp, (struct odp_flow __user *)argp);
1728 err = do_flowvec_ioctl(dp, argp, do_query_flows);
1732 err = do_flowvec_ioctl(dp, argp, do_list_flows);
1736 err = execute_packet(dp, (struct odp_execute __user *)argp);
1743 mutex_unlock(&dp->mutex);
1748 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1751 for (i = 0; i < DP_N_QUEUES; i++) {
1752 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1758 #ifdef CONFIG_COMPAT
1759 static int compat_list_ports(struct datapath *dp, struct compat_odp_portvec __user *upv)
1761 struct compat_odp_portvec pv;
1764 if (copy_from_user(&pv, upv, sizeof pv))
1767 retval = do_list_ports(dp, compat_ptr(pv.ports), pv.n_ports);
1771 return put_user(retval, &upv->n_ports);
1774 static int compat_set_port_group(struct datapath *dp, const struct compat_odp_port_group __user *upg)
1776 struct compat_odp_port_group pg;
1778 if (copy_from_user(&pg, upg, sizeof pg))
1781 return do_set_port_group(dp, compat_ptr(pg.ports), pg.n_ports, pg.group);
1784 static int compat_get_port_group(struct datapath *dp, struct compat_odp_port_group __user *upg)
1786 struct compat_odp_port_group pg;
1788 if (copy_from_user(&pg, upg, sizeof pg))
1791 return do_get_port_group(dp, compat_ptr(pg.ports), pg.n_ports,
1792 pg.group, &upg->n_ports);
1795 static int compat_get_flow(struct odp_flow *flow, const struct compat_odp_flow __user *compat)
1797 compat_uptr_t actions;
1799 if (!access_ok(VERIFY_READ, compat, sizeof(struct compat_odp_flow)) ||
1800 __copy_from_user(&flow->stats, &compat->stats, sizeof(struct odp_flow_stats)) ||
1801 __copy_from_user(&flow->key, &compat->key, sizeof(struct odp_flow_key)) ||
1802 __get_user(actions, &compat->actions) ||
1803 __get_user(flow->n_actions, &compat->n_actions) ||
1804 __get_user(flow->flags, &compat->flags))
1807 flow->actions = compat_ptr(actions);
1811 static int compat_put_flow(struct datapath *dp, struct compat_odp_flow_put __user *ufp)
1813 struct odp_flow_stats stats;
1814 struct odp_flow_put fp;
1817 if (compat_get_flow(&fp.flow, &ufp->flow) ||
1818 get_user(fp.flags, &ufp->flags))
1821 error = do_put_flow(dp, &fp, &stats);
1825 if (copy_to_user(&ufp->flow.stats, &stats,
1826 sizeof(struct odp_flow_stats)))
1832 static int compat_answer_query(struct sw_flow *flow, u32 query_flags,
1833 struct timespec time_offset,
1834 struct compat_odp_flow __user *ufp)
1836 compat_uptr_t actions;
1838 if (get_user(actions, &ufp->actions))
1841 return do_answer_query(flow, query_flags, time_offset, &ufp->stats,
1842 compat_ptr(actions), &ufp->n_actions);
1845 static int compat_del_flow(struct datapath *dp, struct compat_odp_flow __user *ufp)
1847 struct sw_flow *flow;
1851 if (compat_get_flow(&uf, ufp))
1854 flow = do_del_flow(dp, &uf.key);
1856 return PTR_ERR(flow);
1858 error = compat_answer_query(flow, 0, get_time_offset(), ufp);
1859 flow_deferred_free(flow);
1863 static int compat_query_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1865 struct tbl *table = rcu_dereference(dp->table);
1866 struct timespec time_offset;
1869 time_offset = get_time_offset();
1871 for (i = 0; i < n_flows; i++) {
1872 struct compat_odp_flow __user *ufp = &flows[i];
1874 struct tbl_node *flow_node;
1877 if (compat_get_flow(&uf, ufp))
1879 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1881 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1883 error = put_user(ENOENT, &ufp->stats.error);
1885 error = compat_answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1892 struct compat_list_flows_cbdata {
1893 struct compat_odp_flow __user *uflows;
1896 struct timespec time_offset;
1899 static int compat_list_flow(struct tbl_node *node, void *cbdata_)
1901 struct sw_flow *flow = flow_cast(node);
1902 struct compat_list_flows_cbdata *cbdata = cbdata_;
1903 struct compat_odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1906 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1908 error = compat_answer_query(flow, 0, cbdata->time_offset, ufp);
1912 if (cbdata->listed_flows >= cbdata->n_flows)
1913 return cbdata->listed_flows;
1917 static int compat_list_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1919 struct compat_list_flows_cbdata cbdata;
1925 cbdata.uflows = flows;
1926 cbdata.n_flows = n_flows;
1927 cbdata.listed_flows = 0;
1928 cbdata.time_offset = get_time_offset();
1930 error = tbl_foreach(rcu_dereference(dp->table), compat_list_flow, &cbdata);
1931 return error ? error : cbdata.listed_flows;
1934 static int compat_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1935 int (*function)(struct datapath *,
1936 struct compat_odp_flow *,
1939 struct compat_odp_flowvec __user *uflowvec;
1940 struct compat_odp_flow __user *flows;
1941 struct compat_odp_flowvec flowvec;
1944 uflowvec = compat_ptr(argp);
1945 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1946 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1949 if (flowvec.n_flows > INT_MAX / sizeof(struct compat_odp_flow))
1952 flows = compat_ptr(flowvec.flows);
1953 if (!access_ok(VERIFY_WRITE, flows,
1954 flowvec.n_flows * sizeof(struct compat_odp_flow)))
1957 retval = function(dp, flows, flowvec.n_flows);
1958 return (retval < 0 ? retval
1959 : retval == flowvec.n_flows ? 0
1960 : put_user(retval, &uflowvec->n_flows));
1963 static int compat_execute(struct datapath *dp, const struct compat_odp_execute __user *uexecute)
1965 struct odp_execute execute;
1966 compat_uptr_t actions;
1969 if (!access_ok(VERIFY_READ, uexecute, sizeof(struct compat_odp_execute)) ||
1970 __get_user(execute.in_port, &uexecute->in_port) ||
1971 __get_user(actions, &uexecute->actions) ||
1972 __get_user(execute.n_actions, &uexecute->n_actions) ||
1973 __get_user(data, &uexecute->data) ||
1974 __get_user(execute.length, &uexecute->length))
1977 execute.actions = compat_ptr(actions);
1978 execute.data = compat_ptr(data);
1980 return do_execute(dp, &execute);
1983 static long openvswitch_compat_ioctl(struct file *f, unsigned int cmd, unsigned long argp)
1985 int dp_idx = iminor(f->f_dentry->d_inode);
1986 struct datapath *dp;
1990 case ODP_DP_DESTROY:
1991 case ODP_FLOW_FLUSH:
1992 /* Ioctls that don't need any translation at all. */
1993 return openvswitch_ioctl(f, cmd, argp);
1996 case ODP_PORT_ATTACH:
1997 case ODP_PORT_DETACH:
1999 case ODP_VPORT_MTU_SET:
2000 case ODP_VPORT_MTU_GET:
2001 case ODP_VPORT_ETHER_SET:
2002 case ODP_VPORT_ETHER_GET:
2003 case ODP_VPORT_STATS_SET:
2004 case ODP_VPORT_STATS_GET:
2006 case ODP_GET_DROP_FRAGS:
2007 case ODP_SET_DROP_FRAGS:
2008 case ODP_SET_LISTEN_MASK:
2009 case ODP_GET_LISTEN_MASK:
2010 case ODP_SET_SFLOW_PROBABILITY:
2011 case ODP_GET_SFLOW_PROBABILITY:
2012 case ODP_PORT_QUERY:
2013 /* Ioctls that just need their pointer argument extended. */
2014 return openvswitch_ioctl(f, cmd, (unsigned long)compat_ptr(argp));
2016 case ODP_VPORT_ADD32:
2017 return compat_vport_user_add(compat_ptr(argp));
2019 case ODP_VPORT_MOD32:
2020 return compat_vport_user_mod(compat_ptr(argp));
2023 dp = get_dp_locked(dp_idx);
2029 case ODP_PORT_LIST32:
2030 err = compat_list_ports(dp, compat_ptr(argp));
2033 case ODP_PORT_GROUP_SET32:
2034 err = compat_set_port_group(dp, compat_ptr(argp));
2037 case ODP_PORT_GROUP_GET32:
2038 err = compat_get_port_group(dp, compat_ptr(argp));
2041 case ODP_FLOW_PUT32:
2042 err = compat_put_flow(dp, compat_ptr(argp));
2045 case ODP_FLOW_DEL32:
2046 err = compat_del_flow(dp, compat_ptr(argp));
2049 case ODP_FLOW_GET32:
2050 err = compat_flowvec_ioctl(dp, argp, compat_query_flows);
2053 case ODP_FLOW_LIST32:
2054 err = compat_flowvec_ioctl(dp, argp, compat_list_flows);
2058 err = compat_execute(dp, compat_ptr(argp));
2065 mutex_unlock(&dp->mutex);
2071 /* Unfortunately this function is not exported so this is a verbatim copy
2072 * from net/core/datagram.c in 2.6.30. */
2073 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
2074 u8 __user *to, int len,
2077 int start = skb_headlen(skb);
2079 int i, copy = start - offset;
2086 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
2090 if ((len -= copy) == 0)
2097 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2100 WARN_ON(start > offset + len);
2102 end = start + skb_shinfo(skb)->frags[i].size;
2103 if ((copy = end - offset) > 0) {
2107 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2108 struct page *page = frag->page;
2113 csum2 = csum_and_copy_to_user(vaddr +
2120 *csump = csum_block_add(*csump, csum2, pos);
2130 if (skb_shinfo(skb)->frag_list) {
2131 struct sk_buff *list = skb_shinfo(skb)->frag_list;
2133 for (; list; list=list->next) {
2136 WARN_ON(start > offset + len);
2138 end = start + list->len;
2139 if ((copy = end - offset) > 0) {
2143 if (skb_copy_and_csum_datagram(list,
2148 *csump = csum_block_add(*csump, csum2, pos);
2149 if ((len -= copy) == 0)
2165 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
2168 /* XXX is there sufficient synchronization here? */
2169 int listeners = get_listen_mask(f);
2170 int dp_idx = iminor(f->f_dentry->d_inode);
2171 struct datapath *dp = get_dp(dp_idx);
2172 struct sk_buff *skb;
2173 size_t copy_bytes, tot_copy_bytes;
2179 if (nbytes == 0 || !listeners)
2185 for (i = 0; i < DP_N_QUEUES; i++) {
2186 if (listeners & (1 << i)) {
2187 skb = skb_dequeue(&dp->queues[i]);
2193 if (f->f_flags & O_NONBLOCK) {
2198 wait_event_interruptible(dp->waitqueue,
2199 dp_has_packet_of_interest(dp,
2202 if (signal_pending(current)) {
2203 retval = -ERESTARTSYS;
2208 copy_bytes = tot_copy_bytes = min_t(size_t, skb->len, nbytes);
2211 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2212 if (copy_bytes == skb->len) {
2214 unsigned int csum_start, csum_offset;
2216 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
2217 csum_start = skb->csum_start - skb_headroom(skb);
2218 csum_offset = skb->csum_offset;
2220 csum_start = skb_transport_header(skb) - skb->data;
2221 csum_offset = skb->csum;
2223 BUG_ON(csum_start >= skb_headlen(skb));
2224 retval = skb_copy_and_csum_datagram(skb, csum_start, buf + csum_start,
2225 copy_bytes - csum_start, &csum);
2227 __sum16 __user *csump;
2229 copy_bytes = csum_start;
2230 csump = (__sum16 __user *)(buf + csum_start + csum_offset);
2232 BUG_ON((char *)csump + sizeof(__sum16) > buf + nbytes);
2233 put_user(csum_fold(csum), csump);
2236 retval = skb_checksum_help(skb);
2240 struct iovec __user iov;
2243 iov.iov_len = copy_bytes;
2244 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
2248 retval = tot_copy_bytes;
2256 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
2258 /* XXX is there sufficient synchronization here? */
2259 int dp_idx = iminor(file->f_dentry->d_inode);
2260 struct datapath *dp = get_dp(dp_idx);
2265 poll_wait(file, &dp->waitqueue, wait);
2266 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
2267 mask |= POLLIN | POLLRDNORM;
2269 mask = POLLIN | POLLRDNORM | POLLHUP;
2274 struct file_operations openvswitch_fops = {
2275 /* XXX .aio_read = openvswitch_aio_read, */
2276 .read = openvswitch_read,
2277 .poll = openvswitch_poll,
2278 .unlocked_ioctl = openvswitch_ioctl,
2279 #ifdef CONFIG_COMPAT
2280 .compat_ioctl = openvswitch_compat_ioctl,
2282 /* XXX .fasync = openvswitch_fasync, */
2287 static int __init dp_init(void)
2289 struct sk_buff *dummy_skb;
2292 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
2294 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
2302 goto error_flow_exit;
2304 err = register_netdevice_notifier(&dp_device_notifier);
2306 goto error_vport_exit;
2308 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
2310 goto error_unreg_notifier;
2314 error_unreg_notifier:
2315 unregister_netdevice_notifier(&dp_device_notifier);
2324 static void dp_cleanup(void)
2327 unregister_chrdev(major, "openvswitch");
2328 unregister_netdevice_notifier(&dp_device_notifier);
2333 module_init(dp_init);
2334 module_exit(dp_cleanup);
2336 MODULE_DESCRIPTION("Open vSwitch switching datapath");
2337 MODULE_LICENSE("GPL");