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 /* Number of milliseconds between runs of the maintenance thread. */
73 #define MAINT_SLEEP_MSECS 1000
75 static int new_dp_port(struct datapath *, struct odp_port *, int port_no);
77 /* Must be called with rcu_read_lock or dp_mutex. */
78 struct datapath *get_dp(int dp_idx)
80 if (dp_idx < 0 || dp_idx >= ODP_MAX)
82 return rcu_dereference(dps[dp_idx]);
84 EXPORT_SYMBOL_GPL(get_dp);
86 static struct datapath *get_dp_locked(int dp_idx)
90 mutex_lock(&dp_mutex);
93 mutex_lock(&dp->mutex);
94 mutex_unlock(&dp_mutex);
98 /* Must be called with rcu_read_lock or RTNL lock. */
99 const char *dp_name(const struct datapath *dp)
101 return vport_get_name(dp->ports[ODPP_LOCAL]->vport);
104 static inline size_t br_nlmsg_size(void)
106 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
107 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
108 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
109 + nla_total_size(4) /* IFLA_MASTER */
110 + nla_total_size(4) /* IFLA_MTU */
111 + nla_total_size(4) /* IFLA_LINK */
112 + nla_total_size(1); /* IFLA_OPERSTATE */
115 static int dp_fill_ifinfo(struct sk_buff *skb,
116 const struct dp_port *port,
117 int event, unsigned int flags)
119 const struct datapath *dp = port->dp;
120 int ifindex = vport_get_ifindex(port->vport);
121 int iflink = vport_get_iflink(port->vport);
122 struct ifinfomsg *hdr;
123 struct nlmsghdr *nlh;
131 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
135 hdr = nlmsg_data(nlh);
136 hdr->ifi_family = AF_BRIDGE;
138 hdr->ifi_type = ARPHRD_ETHER;
139 hdr->ifi_index = ifindex;
140 hdr->ifi_flags = vport_get_flags(port->vport);
143 NLA_PUT_STRING(skb, IFLA_IFNAME, vport_get_name(port->vport));
144 NLA_PUT_U32(skb, IFLA_MASTER, vport_get_ifindex(dp->ports[ODPP_LOCAL]->vport));
145 NLA_PUT_U32(skb, IFLA_MTU, vport_get_mtu(port->vport));
146 #ifdef IFLA_OPERSTATE
147 NLA_PUT_U8(skb, IFLA_OPERSTATE,
148 vport_is_running(port->vport)
149 ? vport_get_operstate(port->vport)
153 NLA_PUT(skb, IFLA_ADDRESS, ETH_ALEN,
154 vport_get_addr(port->vport));
156 if (ifindex != iflink)
157 NLA_PUT_U32(skb, IFLA_LINK,iflink);
159 return nlmsg_end(skb, nlh);
162 nlmsg_cancel(skb, nlh);
166 static void dp_ifinfo_notify(int event, struct dp_port *port)
171 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
175 err = dp_fill_ifinfo(skb, port, event, 0);
177 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
178 WARN_ON(err == -EMSGSIZE);
182 rtnl_notify(skb, &init_net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
186 rtnl_set_sk_err(&init_net, RTNLGRP_LINK, err);
189 static void release_dp(struct kobject *kobj)
191 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
195 static struct kobj_type dp_ktype = {
196 .release = release_dp
199 static int create_dp(int dp_idx, const char __user *devnamep)
201 struct odp_port internal_dev_port;
202 char devname[IFNAMSIZ];
208 int retval = strncpy_from_user(devname, devnamep, IFNAMSIZ);
212 } else if (retval >= IFNAMSIZ) {
217 snprintf(devname, sizeof devname, "of%d", dp_idx);
221 mutex_lock(&dp_mutex);
223 if (!try_module_get(THIS_MODULE))
226 /* Exit early if a datapath with that number already exists.
227 * (We don't use -EEXIST because that's ambiguous with 'devname'
228 * conflicting with an existing network device name.) */
234 dp = kzalloc(sizeof *dp, GFP_KERNEL);
237 INIT_LIST_HEAD(&dp->port_list);
238 mutex_init(&dp->mutex);
240 for (i = 0; i < DP_N_QUEUES; i++)
241 skb_queue_head_init(&dp->queues[i]);
242 init_waitqueue_head(&dp->waitqueue);
244 /* Initialize kobject for bridge. This will be added as
245 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
246 dp->ifobj.kset = NULL;
247 kobject_init(&dp->ifobj, &dp_ktype);
249 /* Allocate table. */
251 rcu_assign_pointer(dp->table, tbl_create(0));
255 /* Set up our datapath device. */
256 BUILD_BUG_ON(sizeof(internal_dev_port.devname) != sizeof(devname));
257 strcpy(internal_dev_port.devname, devname);
258 internal_dev_port.flags = ODP_PORT_INTERNAL;
259 err = new_dp_port(dp, &internal_dev_port, ODPP_LOCAL);
264 goto err_destroy_table;
268 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
269 if (!dp->stats_percpu)
270 goto err_destroy_local_port;
272 rcu_assign_pointer(dps[dp_idx], dp);
273 mutex_unlock(&dp_mutex);
280 err_destroy_local_port:
281 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
283 tbl_destroy(dp->table, NULL);
287 module_put(THIS_MODULE);
289 mutex_unlock(&dp_mutex);
295 static void do_destroy_dp(struct datapath *dp)
297 struct dp_port *p, *n;
300 list_for_each_entry_safe (p, n, &dp->port_list, node)
301 if (p->port_no != ODPP_LOCAL)
302 dp_detach_port(p, 1);
306 rcu_assign_pointer(dps[dp->dp_idx], NULL);
308 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
310 tbl_destroy(dp->table, flow_free_tbl);
312 for (i = 0; i < DP_N_QUEUES; i++)
313 skb_queue_purge(&dp->queues[i]);
314 for (i = 0; i < DP_MAX_GROUPS; i++)
315 kfree(dp->groups[i]);
316 free_percpu(dp->stats_percpu);
317 kobject_put(&dp->ifobj);
318 module_put(THIS_MODULE);
321 static int destroy_dp(int dp_idx)
327 mutex_lock(&dp_mutex);
337 mutex_unlock(&dp_mutex);
342 static void release_dp_port(struct kobject *kobj)
344 struct dp_port *p = container_of(kobj, struct dp_port, kobj);
348 static struct kobj_type brport_ktype = {
350 .sysfs_ops = &brport_sysfs_ops,
352 .release = release_dp_port
355 /* Called with RTNL lock and dp_mutex. */
356 static int new_dp_port(struct datapath *dp, struct odp_port *odp_port, int port_no)
362 vport = vport_locate(odp_port->devname);
366 if (odp_port->flags & ODP_PORT_INTERNAL)
367 vport = vport_add(odp_port->devname, "internal", NULL);
369 vport = vport_add(odp_port->devname, "netdev", NULL);
374 return PTR_ERR(vport);
377 p = kzalloc(sizeof(*p), GFP_KERNEL);
381 p->port_no = port_no;
383 atomic_set(&p->sflow_pool, 0);
385 err = vport_attach(vport, p);
391 rcu_assign_pointer(dp->ports[port_no], p);
392 list_add_rcu(&p->node, &dp->port_list);
395 /* Initialize kobject for bridge. This will be added as
396 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
398 kobject_init(&p->kobj, &brport_ktype);
400 dp_ifinfo_notify(RTM_NEWLINK, p);
405 static int attach_port(int dp_idx, struct odp_port __user *portp)
408 struct odp_port port;
413 if (copy_from_user(&port, portp, sizeof port))
415 port.devname[IFNAMSIZ - 1] = '\0';
418 dp = get_dp_locked(dp_idx);
421 goto out_unlock_rtnl;
423 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
424 if (!dp->ports[port_no])
430 err = new_dp_port(dp, &port, port_no);
434 set_internal_devs_mtu(dp);
435 dp_sysfs_add_if(dp->ports[port_no]);
437 err = put_user(port_no, &portp->port);
440 mutex_unlock(&dp->mutex);
447 int dp_detach_port(struct dp_port *p, int may_delete)
449 struct vport *vport = p->vport;
454 if (p->port_no != ODPP_LOCAL)
456 dp_ifinfo_notify(RTM_DELLINK, p);
458 /* First drop references to device. */
460 list_del_rcu(&p->node);
461 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
463 err = vport_detach(vport);
467 /* Then wait until no one is still using it, and destroy it. */
471 const char *port_type = vport_get_type(vport);
473 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
480 kobject_put(&p->kobj);
485 static int detach_port(int dp_idx, int port_no)
492 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
496 dp = get_dp_locked(dp_idx);
499 goto out_unlock_rtnl;
501 p = dp->ports[port_no];
506 err = dp_detach_port(p, 1);
509 mutex_unlock(&dp->mutex);
516 /* Must be called with rcu_read_lock. */
517 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
519 struct datapath *dp = p->dp;
520 struct dp_stats_percpu *stats;
521 int stats_counter_off;
522 struct odp_flow_key key;
523 struct tbl_node *flow_node;
525 WARN_ON_ONCE(skb_shared(skb));
526 skb_warn_if_lro(skb);
528 OVS_CB(skb)->dp_port = p;
530 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
531 if (dp->drop_frags) {
533 stats_counter_off = offsetof(struct dp_stats_percpu, n_frags);
538 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
540 struct sw_flow *flow = flow_cast(flow_node);
541 struct sw_flow_actions *acts = rcu_dereference(flow->sf_acts);
542 flow_used(flow, skb);
543 execute_actions(dp, skb, &key, acts->actions, acts->n_actions,
545 stats_counter_off = offsetof(struct dp_stats_percpu, n_hit);
547 stats_counter_off = offsetof(struct dp_stats_percpu, n_missed);
548 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
553 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
554 (*(u64 *)((u8 *)stats + stats_counter_off))++;
558 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
559 /* This code is based on skb_checksum_setup() from Xen's net/dev/core.c. We
560 * can't call this function directly because it isn't exported in all
562 int vswitch_skb_checksum_setup(struct sk_buff *skb)
567 __u16 csum_start, csum_offset;
569 if (!skb->proto_csum_blank)
572 if (skb->protocol != htons(ETH_P_IP))
575 if (!pskb_may_pull(skb, skb_network_header(skb) + sizeof(struct iphdr) - skb->data))
579 th = skb_network_header(skb) + 4 * iph->ihl;
581 csum_start = th - skb->head;
582 switch (iph->protocol) {
584 csum_offset = offsetof(struct tcphdr, check);
587 csum_offset = offsetof(struct udphdr, check);
591 printk(KERN_ERR "Attempting to checksum a non-"
592 "TCP/UDP packet, dropping a protocol"
593 " %d packet", iph->protocol);
597 if (!pskb_may_pull(skb, th + csum_offset + 2 - skb->data))
600 skb->ip_summed = CHECKSUM_PARTIAL;
601 skb->proto_csum_blank = 0;
603 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
604 skb->csum_start = csum_start;
605 skb->csum_offset = csum_offset;
607 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
608 skb->csum = csum_offset;
616 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
618 /* Types of checksums that we can receive (these all refer to L4 checksums):
619 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
620 * (though not verified) checksum in packet but not in skb->csum. Packets
621 * from the bridge local port will also have this type.
622 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
623 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
624 * a valid skb->csum. Importantly, both contain a full checksum (not
625 * verified) in the packet itself. The only difference is that if the
626 * packet gets to L4 processing on this machine (not in DomU) we won't
627 * have to recompute the checksum to verify. Most hardware devices do not
628 * produce packets with this type, even if they support receive checksum
629 * offloading (they produce type #5).
630 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
631 * be computed if it is sent off box. Unfortunately on earlier kernels,
632 * this case is impossible to distinguish from #2, despite having opposite
633 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
634 * to distinguish the different states.
635 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
636 * generated locally by a Xen DomU and has a partial checksum. If it is
637 * handled on this machine (Dom0 or DomU), then the checksum will not be
638 * computed. If it goes off box, the checksum in the packet needs to be
639 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
640 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
641 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
642 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
643 * full checksum or using a protocol without a checksum. skb->csum is
644 * undefined. This is common from devices with receive checksum
645 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
646 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
648 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
649 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
650 * based on whether it is on the transmit or receive path. After the datapath
651 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
652 * checksum, we will panic. Since we can receive packets with checksums, we
653 * assume that all CHECKSUM_HW packets have checksums and map them to
654 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
655 * packet is processed by the local IP stack, in which case it will need to
656 * be reverified). If we receive a packet with CHECKSUM_HW that really means
657 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
658 * shouldn't be any devices that do this with bridging. */
659 void compute_ip_summed(struct sk_buff *skb, bool xmit)
661 /* For our convenience these defines change repeatedly between kernel
662 * versions, so we can't just copy them over... */
663 switch (skb->ip_summed) {
665 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
667 case CHECKSUM_UNNECESSARY:
668 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
671 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
672 * However, on the receive side we should only get CHECKSUM_PARTIAL
673 * packets from Xen, which uses some special fields to represent this
674 * (see below). Since we can only make one type work, pick the one
675 * that actually happens in practice.
677 * On the transmit side (basically after skb_checksum_setup()
678 * has been run or on internal dev transmit), packets with
679 * CHECKSUM_COMPLETE aren't generated, so assume CHECKSUM_PARTIAL. */
682 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
684 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
688 case CHECKSUM_COMPLETE:
689 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
691 case CHECKSUM_PARTIAL:
692 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
696 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
698 /* None seems the safest... */
699 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
702 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
703 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
704 * kernels. It should not be set on the transmit path though. */
705 if (skb->proto_csum_blank)
706 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
708 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
712 /* This function closely resembles skb_forward_csum() used by the bridge. It
713 * is slightly different because we are only concerned with bridging and not
714 * other types of forwarding and can get away with slightly more optimal
716 void forward_ip_summed(struct sk_buff *skb)
719 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
720 skb->ip_summed = CHECKSUM_NONE;
724 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
725 * unless we broke up a GSO packet. */
726 static int 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 err = skb_cow(skb, sizeof *header);
748 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
749 header->type = queue_no;
750 header->length = skb->len;
751 header->port = port_no;
752 header->reserved = 0;
754 skb_queue_tail(queue, skb);
762 while ((skb = nskb) != NULL) {
769 int dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
772 struct dp_stats_percpu *stats;
773 struct sk_buff_head *queue;
776 WARN_ON_ONCE(skb_shared(skb));
777 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
778 queue = &dp->queues[queue_no];
780 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
783 forward_ip_summed(skb);
785 err = vswitch_skb_checksum_setup(skb);
789 /* Break apart GSO packets into their component pieces. Otherwise
790 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
791 if (skb_is_gso(skb)) {
792 struct sk_buff *nskb = skb_gso_segment(skb, NETIF_F_SG | NETIF_F_HW_CSUM);
796 if (unlikely(IS_ERR(skb))) {
801 /* XXX This case might not be possible. It's hard to
802 * tell from the skb_gso_segment() code and comment. */
806 err = queue_control_packets(skb, queue, queue_no, arg);
807 wake_up_interruptible(&dp->waitqueue);
814 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
821 static int flush_flows(struct datapath *dp)
823 struct tbl *old_table = rcu_dereference(dp->table);
824 struct tbl *new_table;
826 new_table = tbl_create(0);
830 rcu_assign_pointer(dp->table, new_table);
832 tbl_deferred_destroy(old_table, flow_free_tbl);
837 static int validate_actions(const struct sw_flow_actions *actions)
841 for (i = 0; i < actions->n_actions; i++) {
842 const union odp_action *a = &actions->actions[i];
845 if (a->output.port >= DP_MAX_PORTS)
849 case ODPAT_OUTPUT_GROUP:
850 if (a->output_group.group >= DP_MAX_GROUPS)
854 case ODPAT_SET_VLAN_VID:
855 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
859 case ODPAT_SET_VLAN_PCP:
860 if (a->vlan_pcp.vlan_pcp
861 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
865 case ODPAT_SET_NW_TOS:
866 if (a->nw_tos.nw_tos & INET_ECN_MASK)
871 if (a->type >= ODPAT_N_ACTIONS)
880 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
882 struct sw_flow_actions *actions;
885 actions = flow_actions_alloc(flow->n_actions);
886 error = PTR_ERR(actions);
891 if (copy_from_user(actions->actions, flow->actions,
892 flow->n_actions * sizeof(union odp_action)))
893 goto error_free_actions;
894 error = validate_actions(actions);
896 goto error_free_actions;
903 return ERR_PTR(error);
906 static struct timespec get_time_offset(void)
908 struct timespec now_mono, now_jiffies;
910 ktime_get_ts(&now_mono);
911 jiffies_to_timespec(jiffies, &now_jiffies);
912 return timespec_sub(now_mono, now_jiffies);
915 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats,
916 struct timespec time_offset)
919 struct timespec flow_ts, used;
921 jiffies_to_timespec(flow->used, &flow_ts);
922 set_normalized_timespec(&used, flow_ts.tv_sec + time_offset.tv_sec,
923 flow_ts.tv_nsec + time_offset.tv_nsec);
925 stats->used_sec = used.tv_sec;
926 stats->used_nsec = used.tv_nsec;
929 stats->used_nsec = 0;
932 stats->n_packets = flow->packet_count;
933 stats->n_bytes = flow->byte_count;
934 stats->ip_tos = flow->ip_tos;
935 stats->tcp_flags = flow->tcp_flags;
939 static void clear_stats(struct sw_flow *flow)
944 flow->packet_count = 0;
945 flow->byte_count = 0;
948 static int expand_table(struct datapath *dp)
950 struct tbl *old_table = rcu_dereference(dp->table);
951 struct tbl *new_table;
953 new_table = tbl_expand(old_table);
954 if (IS_ERR(new_table))
955 return PTR_ERR(new_table);
957 rcu_assign_pointer(dp->table, new_table);
958 tbl_deferred_destroy(old_table, NULL);
963 static int do_put_flow(struct datapath *dp, struct odp_flow_put *uf,
964 struct odp_flow_stats *stats)
966 struct tbl_node *flow_node;
967 struct sw_flow *flow;
971 memset(uf->flow.key.reserved, 0, sizeof uf->flow.key.reserved);
973 table = rcu_dereference(dp->table);
974 flow_node = tbl_lookup(table, &uf->flow.key, flow_hash(&uf->flow.key), flow_cmp);
977 struct sw_flow_actions *acts;
980 if (!(uf->flags & ODPPF_CREATE))
983 /* Expand table, if necessary, to make room. */
984 if (tbl_count(table) >= tbl_n_buckets(table)) {
985 error = expand_table(dp);
988 table = rcu_dereference(dp->table);
993 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
996 flow->key = uf->flow.key;
997 spin_lock_init(&flow->lock);
1000 /* Obtain actions. */
1001 acts = get_actions(&uf->flow);
1002 error = PTR_ERR(acts);
1004 goto error_free_flow;
1005 rcu_assign_pointer(flow->sf_acts, acts);
1007 /* Put flow in bucket. */
1008 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1010 goto error_free_flow_acts;
1012 memset(stats, 0, sizeof(struct odp_flow_stats));
1014 /* We found a matching flow. */
1015 struct sw_flow_actions *old_acts, *new_acts;
1017 flow = flow_cast(flow_node);
1019 /* Bail out if we're not allowed to modify an existing flow. */
1021 if (!(uf->flags & ODPPF_MODIFY))
1025 new_acts = get_actions(&uf->flow);
1026 error = PTR_ERR(new_acts);
1027 if (IS_ERR(new_acts))
1029 old_acts = rcu_dereference(flow->sf_acts);
1030 if (old_acts->n_actions != new_acts->n_actions ||
1031 memcmp(old_acts->actions, new_acts->actions,
1032 sizeof(union odp_action) * old_acts->n_actions)) {
1033 rcu_assign_pointer(flow->sf_acts, new_acts);
1034 flow_deferred_free_acts(old_acts);
1039 /* Fetch stats, then clear them if necessary. */
1040 spin_lock_bh(&flow->lock);
1041 get_stats(flow, stats, get_time_offset());
1042 if (uf->flags & ODPPF_ZERO_STATS)
1044 spin_unlock_bh(&flow->lock);
1049 error_free_flow_acts:
1050 kfree(flow->sf_acts);
1052 kmem_cache_free(flow_cache, flow);
1057 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
1059 struct odp_flow_stats stats;
1060 struct odp_flow_put uf;
1063 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
1066 error = do_put_flow(dp, &uf, &stats);
1070 if (copy_to_user(&ufp->flow.stats, &stats,
1071 sizeof(struct odp_flow_stats)))
1077 static int do_answer_query(struct sw_flow *flow, u32 query_flags,
1078 struct timespec time_offset,
1079 struct odp_flow_stats __user *ustats,
1080 union odp_action __user *actions,
1081 u32 __user *n_actionsp)
1083 struct sw_flow_actions *sf_acts;
1084 struct odp_flow_stats stats;
1087 spin_lock_bh(&flow->lock);
1088 get_stats(flow, &stats, time_offset);
1089 if (query_flags & ODPFF_ZERO_TCP_FLAGS)
1090 flow->tcp_flags = 0;
1092 spin_unlock_bh(&flow->lock);
1094 if (copy_to_user(ustats, &stats, sizeof(struct odp_flow_stats)) ||
1095 get_user(n_actions, n_actionsp))
1101 sf_acts = rcu_dereference(flow->sf_acts);
1102 if (put_user(sf_acts->n_actions, n_actionsp) ||
1103 (actions && copy_to_user(actions, sf_acts->actions,
1104 sizeof(union odp_action) *
1105 min(sf_acts->n_actions, n_actions))))
1111 static int answer_query(struct sw_flow *flow, u32 query_flags,
1112 struct timespec time_offset,
1113 struct odp_flow __user *ufp)
1115 union odp_action *actions;
1117 if (get_user(actions, &ufp->actions))
1120 return do_answer_query(flow, query_flags, time_offset,
1121 &ufp->stats, actions, &ufp->n_actions);
1124 static struct sw_flow *do_del_flow(struct datapath *dp, struct odp_flow_key *key)
1126 struct tbl *table = rcu_dereference(dp->table);
1127 struct tbl_node *flow_node;
1130 memset(key->reserved, 0, sizeof key->reserved);
1131 flow_node = tbl_lookup(table, key, flow_hash(key), flow_cmp);
1133 return ERR_PTR(-ENOENT);
1135 error = tbl_remove(table, flow_node);
1137 return ERR_PTR(error);
1139 /* XXX Returned flow_node's statistics might lose a few packets, since
1140 * other CPUs can be using this flow. We used to synchronize_rcu() to
1141 * make sure that we get completely accurate stats, but that blows our
1142 * performance, badly. */
1143 return flow_cast(flow_node);
1146 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1148 struct sw_flow *flow;
1152 if (copy_from_user(&uf, ufp, sizeof uf))
1155 flow = do_del_flow(dp, &uf.key);
1157 return PTR_ERR(flow);
1159 error = answer_query(flow, 0, get_time_offset(), ufp);
1160 flow_deferred_free(flow);
1164 static int do_query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1166 struct tbl *table = rcu_dereference(dp->table);
1167 struct timespec time_offset;
1170 time_offset = get_time_offset();
1172 for (i = 0; i < flowvec->n_flows; i++) {
1173 struct odp_flow __user *ufp = &flowvec->flows[i];
1175 struct tbl_node *flow_node;
1178 if (copy_from_user(&uf, ufp, sizeof uf))
1180 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1182 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1184 error = put_user(ENOENT, &ufp->stats.error);
1186 error = answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1190 return flowvec->n_flows;
1193 struct list_flows_cbdata {
1194 struct odp_flow __user *uflows;
1197 struct timespec time_offset;
1200 static int list_flow(struct tbl_node *node, void *cbdata_)
1202 struct sw_flow *flow = flow_cast(node);
1203 struct list_flows_cbdata *cbdata = cbdata_;
1204 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1207 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1209 error = answer_query(flow, 0, cbdata->time_offset, ufp);
1213 if (cbdata->listed_flows >= cbdata->n_flows)
1214 return cbdata->listed_flows;
1218 static int do_list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1220 struct list_flows_cbdata cbdata;
1223 if (!flowvec->n_flows)
1226 cbdata.uflows = flowvec->flows;
1227 cbdata.n_flows = flowvec->n_flows;
1228 cbdata.listed_flows = 0;
1229 cbdata.time_offset = get_time_offset();
1231 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1232 return error ? error : cbdata.listed_flows;
1235 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1236 int (*function)(struct datapath *,
1237 const struct odp_flowvec *))
1239 struct odp_flowvec __user *uflowvec;
1240 struct odp_flowvec flowvec;
1243 uflowvec = (struct odp_flowvec __user *)argp;
1244 if (copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1247 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1250 retval = function(dp, &flowvec);
1251 return (retval < 0 ? retval
1252 : retval == flowvec.n_flows ? 0
1253 : put_user(retval, &uflowvec->n_flows));
1256 static int do_execute(struct datapath *dp, const struct odp_execute *execute)
1258 struct odp_flow_key key;
1259 struct sk_buff *skb;
1260 struct sw_flow_actions *actions;
1265 if (execute->length < ETH_HLEN || execute->length > 65535)
1269 actions = flow_actions_alloc(execute->n_actions);
1274 if (copy_from_user(actions->actions, execute->actions,
1275 execute->n_actions * sizeof *execute->actions))
1276 goto error_free_actions;
1278 err = validate_actions(actions);
1280 goto error_free_actions;
1283 skb = alloc_skb(execute->length, GFP_KERNEL);
1285 goto error_free_actions;
1287 if (execute->in_port < DP_MAX_PORTS)
1288 OVS_CB(skb)->dp_port = dp->ports[execute->in_port];
1290 OVS_CB(skb)->dp_port = NULL;
1293 if (copy_from_user(skb_put(skb, execute->length), execute->data,
1295 goto error_free_skb;
1297 skb_reset_mac_header(skb);
1300 /* Normally, setting the skb 'protocol' field would be handled by a
1301 * call to eth_type_trans(), but it assumes there's a sending
1302 * device, which we may not have. */
1303 if (ntohs(eth->h_proto) >= 1536)
1304 skb->protocol = eth->h_proto;
1306 skb->protocol = htons(ETH_P_802_2);
1308 flow_extract(skb, execute->in_port, &key);
1311 err = execute_actions(dp, skb, &key, actions->actions,
1312 actions->n_actions, GFP_KERNEL);
1326 static int execute_packet(struct datapath *dp, const struct odp_execute __user *executep)
1328 struct odp_execute execute;
1330 if (copy_from_user(&execute, executep, sizeof execute))
1333 return do_execute(dp, &execute);
1336 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1338 struct tbl *table = rcu_dereference(dp->table);
1339 struct odp_stats stats;
1342 stats.n_flows = tbl_count(table);
1343 stats.cur_capacity = tbl_n_buckets(table);
1344 stats.max_capacity = TBL_MAX_BUCKETS;
1345 stats.n_ports = dp->n_ports;
1346 stats.max_ports = DP_MAX_PORTS;
1347 stats.max_groups = DP_MAX_GROUPS;
1348 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1349 for_each_possible_cpu(i) {
1350 const struct dp_stats_percpu *s;
1351 s = per_cpu_ptr(dp->stats_percpu, i);
1352 stats.n_frags += s->n_frags;
1353 stats.n_hit += s->n_hit;
1354 stats.n_missed += s->n_missed;
1355 stats.n_lost += s->n_lost;
1357 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1358 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1359 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1362 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1363 int dp_min_mtu(const struct datapath *dp)
1370 list_for_each_entry_rcu (p, &dp->port_list, node) {
1373 /* Skip any internal ports, since that's what we're trying to
1375 if (is_internal_vport(p->vport))
1378 dev_mtu = vport_get_mtu(p->vport);
1379 if (!mtu || dev_mtu < mtu)
1383 return mtu ? mtu : ETH_DATA_LEN;
1386 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1387 * be called with RTNL lock. */
1388 void set_internal_devs_mtu(const struct datapath *dp)
1395 mtu = dp_min_mtu(dp);
1397 list_for_each_entry_rcu (p, &dp->port_list, node) {
1398 if (is_internal_vport(p->vport))
1399 vport_set_mtu(p->vport, mtu);
1403 static int put_port(const struct dp_port *p, struct odp_port __user *uop)
1407 memset(&op, 0, sizeof op);
1410 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1413 op.port = p->port_no;
1414 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1416 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1419 static int query_port(struct datapath *dp, struct odp_port __user *uport)
1421 struct odp_port port;
1423 if (copy_from_user(&port, uport, sizeof port))
1426 if (port.devname[0]) {
1427 struct vport *vport;
1428 struct dp_port *dp_port;
1431 port.devname[IFNAMSIZ - 1] = '\0';
1436 vport = vport_locate(port.devname);
1442 dp_port = vport_get_dp_port(vport);
1443 if (!dp_port || dp_port->dp != dp) {
1448 port.port = dp_port->port_no;
1457 if (port.port >= DP_MAX_PORTS)
1459 if (!dp->ports[port.port])
1463 return put_port(dp->ports[port.port], uport);
1466 static int do_list_ports(struct datapath *dp, struct odp_port __user *uports,
1473 list_for_each_entry_rcu (p, &dp->port_list, node) {
1474 if (put_port(p, &uports[idx]))
1476 if (idx++ >= n_ports)
1483 static int list_ports(struct datapath *dp, struct odp_portvec __user *upv)
1485 struct odp_portvec pv;
1488 if (copy_from_user(&pv, upv, sizeof pv))
1491 retval = do_list_ports(dp, pv.ports, pv.n_ports);
1495 return put_user(retval, &upv->n_ports);
1498 /* RCU callback for freeing a dp_port_group */
1499 static void free_port_group(struct rcu_head *rcu)
1501 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1505 static int do_set_port_group(struct datapath *dp, u16 __user *ports,
1506 int n_ports, int group)
1508 struct dp_port_group *new_group, *old_group;
1512 if (n_ports > DP_MAX_PORTS || group >= DP_MAX_GROUPS)
1516 new_group = kmalloc(sizeof *new_group + sizeof(u16) * n_ports, GFP_KERNEL);
1520 new_group->n_ports = n_ports;
1522 if (copy_from_user(new_group->ports, ports, sizeof(u16) * n_ports))
1525 old_group = rcu_dereference(dp->groups[group]);
1526 rcu_assign_pointer(dp->groups[group], new_group);
1528 call_rcu(&old_group->rcu, free_port_group);
1537 static int set_port_group(struct datapath *dp,
1538 const struct odp_port_group __user *upg)
1540 struct odp_port_group pg;
1542 if (copy_from_user(&pg, upg, sizeof pg))
1545 return do_set_port_group(dp, pg.ports, pg.n_ports, pg.group);
1548 static int do_get_port_group(struct datapath *dp,
1549 u16 __user *ports, int n_ports, int group,
1550 u16 __user *n_portsp)
1552 struct dp_port_group *g;
1555 if (group >= DP_MAX_GROUPS)
1558 g = dp->groups[group];
1559 n_copy = g ? min_t(int, g->n_ports, n_ports) : 0;
1560 if (n_copy && copy_to_user(ports, g->ports, n_copy * sizeof(u16)))
1563 if (put_user(g ? g->n_ports : 0, n_portsp))
1569 static int get_port_group(struct datapath *dp, struct odp_port_group __user *upg)
1571 struct odp_port_group pg;
1573 if (copy_from_user(&pg, upg, sizeof pg))
1576 return do_get_port_group(dp, pg.ports, pg.n_ports, pg.group, &upg->n_ports);
1579 static int get_listen_mask(const struct file *f)
1581 return (long)f->private_data;
1584 static void set_listen_mask(struct file *f, int listen_mask)
1586 f->private_data = (void*)(long)listen_mask;
1589 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1592 int dp_idx = iminor(f->f_dentry->d_inode);
1593 struct datapath *dp;
1594 int drop_frags, listeners, port_no;
1595 unsigned int sflow_probability;
1598 /* Handle commands with special locking requirements up front. */
1601 err = create_dp(dp_idx, (char __user *)argp);
1604 case ODP_DP_DESTROY:
1605 err = destroy_dp(dp_idx);
1608 case ODP_PORT_ATTACH:
1609 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1612 case ODP_PORT_DETACH:
1613 err = get_user(port_no, (int __user *)argp);
1615 err = detach_port(dp_idx, port_no);
1619 err = vport_user_add((struct odp_vport_add __user *)argp);
1623 err = vport_user_mod((struct odp_vport_mod __user *)argp);
1627 err = vport_user_del((char __user *)argp);
1630 case ODP_VPORT_STATS_GET:
1631 err = vport_user_stats_get((struct odp_vport_stats_req __user *)argp);
1634 case ODP_VPORT_STATS_SET:
1635 err = vport_user_stats_set((struct odp_vport_stats_req __user *)argp);
1638 case ODP_VPORT_ETHER_GET:
1639 err = vport_user_ether_get((struct odp_vport_ether __user *)argp);
1642 case ODP_VPORT_ETHER_SET:
1643 err = vport_user_ether_set((struct odp_vport_ether __user *)argp);
1646 case ODP_VPORT_MTU_GET:
1647 err = vport_user_mtu_get((struct odp_vport_mtu __user *)argp);
1650 case ODP_VPORT_MTU_SET:
1651 err = vport_user_mtu_set((struct odp_vport_mtu __user *)argp);
1655 dp = get_dp_locked(dp_idx);
1662 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1665 case ODP_GET_DROP_FRAGS:
1666 err = put_user(dp->drop_frags, (int __user *)argp);
1669 case ODP_SET_DROP_FRAGS:
1670 err = get_user(drop_frags, (int __user *)argp);
1674 if (drop_frags != 0 && drop_frags != 1)
1676 dp->drop_frags = drop_frags;
1680 case ODP_GET_LISTEN_MASK:
1681 err = put_user(get_listen_mask(f), (int __user *)argp);
1684 case ODP_SET_LISTEN_MASK:
1685 err = get_user(listeners, (int __user *)argp);
1689 if (listeners & ~ODPL_ALL)
1692 set_listen_mask(f, listeners);
1695 case ODP_GET_SFLOW_PROBABILITY:
1696 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1699 case ODP_SET_SFLOW_PROBABILITY:
1700 err = get_user(sflow_probability, (unsigned int __user *)argp);
1702 dp->sflow_probability = sflow_probability;
1705 case ODP_PORT_QUERY:
1706 err = query_port(dp, (struct odp_port __user *)argp);
1710 err = list_ports(dp, (struct odp_portvec __user *)argp);
1713 case ODP_PORT_GROUP_SET:
1714 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1717 case ODP_PORT_GROUP_GET:
1718 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1721 case ODP_FLOW_FLUSH:
1722 err = flush_flows(dp);
1726 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1730 err = del_flow(dp, (struct odp_flow __user *)argp);
1734 err = do_flowvec_ioctl(dp, argp, do_query_flows);
1738 err = do_flowvec_ioctl(dp, argp, do_list_flows);
1742 err = execute_packet(dp, (struct odp_execute __user *)argp);
1749 mutex_unlock(&dp->mutex);
1754 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1757 for (i = 0; i < DP_N_QUEUES; i++) {
1758 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1764 #ifdef CONFIG_COMPAT
1765 static int compat_list_ports(struct datapath *dp, struct compat_odp_portvec __user *upv)
1767 struct compat_odp_portvec pv;
1770 if (copy_from_user(&pv, upv, sizeof pv))
1773 retval = do_list_ports(dp, compat_ptr(pv.ports), pv.n_ports);
1777 return put_user(retval, &upv->n_ports);
1780 static int compat_set_port_group(struct datapath *dp, const struct compat_odp_port_group __user *upg)
1782 struct compat_odp_port_group pg;
1784 if (copy_from_user(&pg, upg, sizeof pg))
1787 return do_set_port_group(dp, compat_ptr(pg.ports), pg.n_ports, pg.group);
1790 static int compat_get_port_group(struct datapath *dp, struct compat_odp_port_group __user *upg)
1792 struct compat_odp_port_group pg;
1794 if (copy_from_user(&pg, upg, sizeof pg))
1797 return do_get_port_group(dp, compat_ptr(pg.ports), pg.n_ports,
1798 pg.group, &upg->n_ports);
1801 static int compat_get_flow(struct odp_flow *flow, const struct compat_odp_flow __user *compat)
1803 compat_uptr_t actions;
1805 if (!access_ok(VERIFY_READ, compat, sizeof(struct compat_odp_flow)) ||
1806 __copy_from_user(&flow->stats, &compat->stats, sizeof(struct odp_flow_stats)) ||
1807 __copy_from_user(&flow->key, &compat->key, sizeof(struct odp_flow_key)) ||
1808 __get_user(actions, &compat->actions) ||
1809 __get_user(flow->n_actions, &compat->n_actions) ||
1810 __get_user(flow->flags, &compat->flags))
1813 flow->actions = compat_ptr(actions);
1817 static int compat_put_flow(struct datapath *dp, struct compat_odp_flow_put __user *ufp)
1819 struct odp_flow_stats stats;
1820 struct odp_flow_put fp;
1823 if (compat_get_flow(&fp.flow, &ufp->flow) ||
1824 get_user(fp.flags, &ufp->flags))
1827 error = do_put_flow(dp, &fp, &stats);
1831 if (copy_to_user(&ufp->flow.stats, &stats,
1832 sizeof(struct odp_flow_stats)))
1838 static int compat_answer_query(struct sw_flow *flow, u32 query_flags,
1839 struct timespec time_offset,
1840 struct compat_odp_flow __user *ufp)
1842 compat_uptr_t actions;
1844 if (get_user(actions, &ufp->actions))
1847 return do_answer_query(flow, query_flags, time_offset, &ufp->stats,
1848 compat_ptr(actions), &ufp->n_actions);
1851 static int compat_del_flow(struct datapath *dp, struct compat_odp_flow __user *ufp)
1853 struct sw_flow *flow;
1857 if (compat_get_flow(&uf, ufp))
1860 flow = do_del_flow(dp, &uf.key);
1862 return PTR_ERR(flow);
1864 error = compat_answer_query(flow, 0, get_time_offset(), ufp);
1865 flow_deferred_free(flow);
1869 static int compat_query_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1871 struct tbl *table = rcu_dereference(dp->table);
1872 struct timespec time_offset;
1875 time_offset = get_time_offset();
1877 for (i = 0; i < n_flows; i++) {
1878 struct compat_odp_flow __user *ufp = &flows[i];
1880 struct tbl_node *flow_node;
1883 if (compat_get_flow(&uf, ufp))
1885 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1887 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1889 error = put_user(ENOENT, &ufp->stats.error);
1891 error = compat_answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1898 struct compat_list_flows_cbdata {
1899 struct compat_odp_flow __user *uflows;
1902 struct timespec time_offset;
1905 static int compat_list_flow(struct tbl_node *node, void *cbdata_)
1907 struct sw_flow *flow = flow_cast(node);
1908 struct compat_list_flows_cbdata *cbdata = cbdata_;
1909 struct compat_odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1912 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1914 error = compat_answer_query(flow, 0, cbdata->time_offset, ufp);
1918 if (cbdata->listed_flows >= cbdata->n_flows)
1919 return cbdata->listed_flows;
1923 static int compat_list_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1925 struct compat_list_flows_cbdata cbdata;
1931 cbdata.uflows = flows;
1932 cbdata.n_flows = n_flows;
1933 cbdata.listed_flows = 0;
1934 cbdata.time_offset = get_time_offset();
1936 error = tbl_foreach(rcu_dereference(dp->table), compat_list_flow, &cbdata);
1937 return error ? error : cbdata.listed_flows;
1940 static int compat_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1941 int (*function)(struct datapath *,
1942 struct compat_odp_flow *,
1945 struct compat_odp_flowvec __user *uflowvec;
1946 struct compat_odp_flow __user *flows;
1947 struct compat_odp_flowvec flowvec;
1950 uflowvec = compat_ptr(argp);
1951 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
1952 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1955 if (flowvec.n_flows > INT_MAX / sizeof(struct compat_odp_flow))
1958 flows = compat_ptr(flowvec.flows);
1959 if (!access_ok(VERIFY_WRITE, flows,
1960 flowvec.n_flows * sizeof(struct compat_odp_flow)))
1963 retval = function(dp, flows, flowvec.n_flows);
1964 return (retval < 0 ? retval
1965 : retval == flowvec.n_flows ? 0
1966 : put_user(retval, &uflowvec->n_flows));
1969 static int compat_execute(struct datapath *dp, const struct compat_odp_execute __user *uexecute)
1971 struct odp_execute execute;
1972 compat_uptr_t actions;
1975 if (!access_ok(VERIFY_READ, uexecute, sizeof(struct compat_odp_execute)) ||
1976 __get_user(execute.in_port, &uexecute->in_port) ||
1977 __get_user(actions, &uexecute->actions) ||
1978 __get_user(execute.n_actions, &uexecute->n_actions) ||
1979 __get_user(data, &uexecute->data) ||
1980 __get_user(execute.length, &uexecute->length))
1983 execute.actions = compat_ptr(actions);
1984 execute.data = compat_ptr(data);
1986 return do_execute(dp, &execute);
1989 static long openvswitch_compat_ioctl(struct file *f, unsigned int cmd, unsigned long argp)
1991 int dp_idx = iminor(f->f_dentry->d_inode);
1992 struct datapath *dp;
1996 case ODP_DP_DESTROY:
1997 case ODP_FLOW_FLUSH:
1998 /* Ioctls that don't need any translation at all. */
1999 return openvswitch_ioctl(f, cmd, argp);
2002 case ODP_PORT_ATTACH:
2003 case ODP_PORT_DETACH:
2005 case ODP_VPORT_MTU_SET:
2006 case ODP_VPORT_MTU_GET:
2007 case ODP_VPORT_ETHER_SET:
2008 case ODP_VPORT_ETHER_GET:
2009 case ODP_VPORT_STATS_SET:
2010 case ODP_VPORT_STATS_GET:
2012 case ODP_GET_DROP_FRAGS:
2013 case ODP_SET_DROP_FRAGS:
2014 case ODP_SET_LISTEN_MASK:
2015 case ODP_GET_LISTEN_MASK:
2016 case ODP_SET_SFLOW_PROBABILITY:
2017 case ODP_GET_SFLOW_PROBABILITY:
2018 case ODP_PORT_QUERY:
2019 /* Ioctls that just need their pointer argument extended. */
2020 return openvswitch_ioctl(f, cmd, (unsigned long)compat_ptr(argp));
2022 case ODP_VPORT_ADD32:
2023 return compat_vport_user_add(compat_ptr(argp));
2025 case ODP_VPORT_MOD32:
2026 return compat_vport_user_mod(compat_ptr(argp));
2029 dp = get_dp_locked(dp_idx);
2035 case ODP_PORT_LIST32:
2036 err = compat_list_ports(dp, compat_ptr(argp));
2039 case ODP_PORT_GROUP_SET32:
2040 err = compat_set_port_group(dp, compat_ptr(argp));
2043 case ODP_PORT_GROUP_GET32:
2044 err = compat_get_port_group(dp, compat_ptr(argp));
2047 case ODP_FLOW_PUT32:
2048 err = compat_put_flow(dp, compat_ptr(argp));
2051 case ODP_FLOW_DEL32:
2052 err = compat_del_flow(dp, compat_ptr(argp));
2055 case ODP_FLOW_GET32:
2056 err = compat_flowvec_ioctl(dp, argp, compat_query_flows);
2059 case ODP_FLOW_LIST32:
2060 err = compat_flowvec_ioctl(dp, argp, compat_list_flows);
2064 err = compat_execute(dp, compat_ptr(argp));
2071 mutex_unlock(&dp->mutex);
2077 /* Unfortunately this function is not exported so this is a verbatim copy
2078 * from net/core/datagram.c in 2.6.30. */
2079 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
2080 u8 __user *to, int len,
2083 int start = skb_headlen(skb);
2085 int i, copy = start - offset;
2092 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
2096 if ((len -= copy) == 0)
2103 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2106 WARN_ON(start > offset + len);
2108 end = start + skb_shinfo(skb)->frags[i].size;
2109 if ((copy = end - offset) > 0) {
2113 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2114 struct page *page = frag->page;
2119 csum2 = csum_and_copy_to_user(vaddr +
2126 *csump = csum_block_add(*csump, csum2, pos);
2136 if (skb_shinfo(skb)->frag_list) {
2137 struct sk_buff *list = skb_shinfo(skb)->frag_list;
2139 for (; list; list=list->next) {
2142 WARN_ON(start > offset + len);
2144 end = start + list->len;
2145 if ((copy = end - offset) > 0) {
2149 if (skb_copy_and_csum_datagram(list,
2154 *csump = csum_block_add(*csump, csum2, pos);
2155 if ((len -= copy) == 0)
2171 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
2174 /* XXX is there sufficient synchronization here? */
2175 int listeners = get_listen_mask(f);
2176 int dp_idx = iminor(f->f_dentry->d_inode);
2177 struct datapath *dp = get_dp(dp_idx);
2178 struct sk_buff *skb;
2179 size_t copy_bytes, tot_copy_bytes;
2185 if (nbytes == 0 || !listeners)
2191 for (i = 0; i < DP_N_QUEUES; i++) {
2192 if (listeners & (1 << i)) {
2193 skb = skb_dequeue(&dp->queues[i]);
2199 if (f->f_flags & O_NONBLOCK) {
2204 wait_event_interruptible(dp->waitqueue,
2205 dp_has_packet_of_interest(dp,
2208 if (signal_pending(current)) {
2209 retval = -ERESTARTSYS;
2214 copy_bytes = tot_copy_bytes = min_t(size_t, skb->len, nbytes);
2217 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2218 if (copy_bytes == skb->len) {
2220 unsigned int csum_start, csum_offset;
2222 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
2223 csum_start = skb->csum_start - skb_headroom(skb);
2224 csum_offset = skb->csum_offset;
2226 csum_start = skb_transport_header(skb) - skb->data;
2227 csum_offset = skb->csum;
2229 BUG_ON(csum_start >= skb_headlen(skb));
2230 retval = skb_copy_and_csum_datagram(skb, csum_start, buf + csum_start,
2231 copy_bytes - csum_start, &csum);
2233 __sum16 __user *csump;
2235 copy_bytes = csum_start;
2236 csump = (__sum16 __user *)(buf + csum_start + csum_offset);
2238 BUG_ON((char *)csump + sizeof(__sum16) > buf + nbytes);
2239 put_user(csum_fold(csum), csump);
2242 retval = skb_checksum_help(skb);
2246 struct iovec __user iov;
2249 iov.iov_len = copy_bytes;
2250 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
2254 retval = tot_copy_bytes;
2262 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
2264 /* XXX is there sufficient synchronization here? */
2265 int dp_idx = iminor(file->f_dentry->d_inode);
2266 struct datapath *dp = get_dp(dp_idx);
2271 poll_wait(file, &dp->waitqueue, wait);
2272 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
2273 mask |= POLLIN | POLLRDNORM;
2275 mask = POLLIN | POLLRDNORM | POLLHUP;
2280 struct file_operations openvswitch_fops = {
2281 /* XXX .aio_read = openvswitch_aio_read, */
2282 .read = openvswitch_read,
2283 .poll = openvswitch_poll,
2284 .unlocked_ioctl = openvswitch_ioctl,
2285 #ifdef CONFIG_COMPAT
2286 .compat_ioctl = openvswitch_compat_ioctl,
2288 /* XXX .fasync = openvswitch_fasync, */
2293 static int __init dp_init(void)
2295 struct sk_buff *dummy_skb;
2298 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
2300 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
2308 goto error_flow_exit;
2310 err = register_netdevice_notifier(&dp_device_notifier);
2312 goto error_vport_exit;
2314 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
2316 goto error_unreg_notifier;
2320 error_unreg_notifier:
2321 unregister_netdevice_notifier(&dp_device_notifier);
2330 static void dp_cleanup(void)
2333 unregister_chrdev(major, "openvswitch");
2334 unregister_netdevice_notifier(&dp_device_notifier);
2339 module_init(dp_init);
2340 module_exit(dp_cleanup);
2342 MODULE_DESCRIPTION("Open vSwitch switching datapath");
2343 MODULE_LICENSE("GPL");