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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/init.h>
14 #include <linux/module.h>
16 #include <linux/if_arp.h>
17 #include <linux/if_vlan.h>
20 #include <linux/delay.h>
21 #include <linux/time.h>
22 #include <linux/etherdevice.h>
23 #include <linux/kernel.h>
24 #include <linux/kthread.h>
25 #include <linux/mutex.h>
26 #include <linux/percpu.h>
27 #include <linux/rcupdate.h>
28 #include <linux/tcp.h>
29 #include <linux/udp.h>
30 #include <linux/version.h>
31 #include <linux/ethtool.h>
32 #include <linux/random.h>
33 #include <linux/wait.h>
34 #include <asm/system.h>
35 #include <asm/div64.h>
37 #include <linux/highmem.h>
38 #include <linux/netfilter_bridge.h>
39 #include <linux/netfilter_ipv4.h>
40 #include <linux/inetdevice.h>
41 #include <linux/list.h>
42 #include <linux/rculist.h>
43 #include <linux/workqueue.h>
44 #include <linux/dmi.h>
45 #include <net/inet_ecn.h>
46 #include <linux/compat.h>
48 #include "openvswitch/datapath-protocol.h"
52 #include "odp-compat.h"
54 #include "vport-internal_dev.h"
59 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
60 EXPORT_SYMBOL(dp_ioctl_hook);
62 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
65 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
68 * It is safe to access the datapath and dp_port structures with just
71 static struct datapath *dps[ODP_MAX];
72 static DEFINE_MUTEX(dp_mutex);
74 /* We limit the number of times that we pass into dp_process_received_packet()
75 * to avoid blowing out the stack in the event that we have a loop. */
77 int count; /* Count. */
78 bool looping; /* Loop detected? */
81 #define DP_MAX_LOOPS 5
83 /* We use a separate counter for each CPU for both interrupt and non-interrupt
84 * context in order to keep the limit deterministic for a given packet. */
85 struct percpu_loop_counters {
86 struct loop_counter counters[2];
89 static DEFINE_PER_CPU(struct percpu_loop_counters, dp_loop_counters);
91 static int new_dp_port(struct datapath *, struct odp_port *, int port_no);
93 /* Must be called with rcu_read_lock or dp_mutex. */
94 struct datapath *get_dp(int dp_idx)
96 if (dp_idx < 0 || dp_idx >= ODP_MAX)
98 return rcu_dereference(dps[dp_idx]);
100 EXPORT_SYMBOL_GPL(get_dp);
102 static struct datapath *get_dp_locked(int dp_idx)
106 mutex_lock(&dp_mutex);
109 mutex_lock(&dp->mutex);
110 mutex_unlock(&dp_mutex);
114 /* Must be called with rcu_read_lock or RTNL lock. */
115 const char *dp_name(const struct datapath *dp)
117 return vport_get_name(dp->ports[ODPP_LOCAL]->vport);
120 static inline size_t br_nlmsg_size(void)
122 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
123 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
124 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
125 + nla_total_size(4) /* IFLA_MASTER */
126 + nla_total_size(4) /* IFLA_MTU */
127 + nla_total_size(4) /* IFLA_LINK */
128 + nla_total_size(1); /* IFLA_OPERSTATE */
131 static int dp_fill_ifinfo(struct sk_buff *skb,
132 const struct dp_port *port,
133 int event, unsigned int flags)
135 const struct datapath *dp = port->dp;
136 int ifindex = vport_get_ifindex(port->vport);
137 int iflink = vport_get_iflink(port->vport);
138 struct ifinfomsg *hdr;
139 struct nlmsghdr *nlh;
147 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
151 hdr = nlmsg_data(nlh);
152 hdr->ifi_family = AF_BRIDGE;
154 hdr->ifi_type = ARPHRD_ETHER;
155 hdr->ifi_index = ifindex;
156 hdr->ifi_flags = vport_get_flags(port->vport);
159 NLA_PUT_STRING(skb, IFLA_IFNAME, vport_get_name(port->vport));
160 NLA_PUT_U32(skb, IFLA_MASTER, vport_get_ifindex(dp->ports[ODPP_LOCAL]->vport));
161 NLA_PUT_U32(skb, IFLA_MTU, vport_get_mtu(port->vport));
162 #ifdef IFLA_OPERSTATE
163 NLA_PUT_U8(skb, IFLA_OPERSTATE,
164 vport_is_running(port->vport)
165 ? vport_get_operstate(port->vport)
169 NLA_PUT(skb, IFLA_ADDRESS, ETH_ALEN,
170 vport_get_addr(port->vport));
172 if (ifindex != iflink)
173 NLA_PUT_U32(skb, IFLA_LINK,iflink);
175 return nlmsg_end(skb, nlh);
178 nlmsg_cancel(skb, nlh);
182 static void dp_ifinfo_notify(int event, struct dp_port *port)
187 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
191 err = dp_fill_ifinfo(skb, port, event, 0);
193 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
194 WARN_ON(err == -EMSGSIZE);
198 rtnl_notify(skb, &init_net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
202 rtnl_set_sk_err(&init_net, RTNLGRP_LINK, err);
205 static void release_dp(struct kobject *kobj)
207 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
211 static struct kobj_type dp_ktype = {
212 .release = release_dp
215 static int create_dp(int dp_idx, const char __user *devnamep)
217 struct odp_port internal_dev_port;
218 char devname[IFNAMSIZ];
224 int retval = strncpy_from_user(devname, devnamep, IFNAMSIZ);
228 } else if (retval >= IFNAMSIZ) {
233 snprintf(devname, sizeof devname, "of%d", dp_idx);
237 mutex_lock(&dp_mutex);
239 if (!try_module_get(THIS_MODULE))
242 /* Exit early if a datapath with that number already exists.
243 * (We don't use -EEXIST because that's ambiguous with 'devname'
244 * conflicting with an existing network device name.) */
250 dp = kzalloc(sizeof *dp, GFP_KERNEL);
253 INIT_LIST_HEAD(&dp->port_list);
254 mutex_init(&dp->mutex);
256 for (i = 0; i < DP_N_QUEUES; i++)
257 skb_queue_head_init(&dp->queues[i]);
258 init_waitqueue_head(&dp->waitqueue);
260 /* Initialize kobject for bridge. This will be added as
261 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
262 dp->ifobj.kset = NULL;
263 kobject_init(&dp->ifobj, &dp_ktype);
265 /* Allocate table. */
267 rcu_assign_pointer(dp->table, tbl_create(0));
271 /* Set up our datapath device. */
272 BUILD_BUG_ON(sizeof(internal_dev_port.devname) != sizeof(devname));
273 strcpy(internal_dev_port.devname, devname);
274 internal_dev_port.flags = ODP_PORT_INTERNAL;
275 err = new_dp_port(dp, &internal_dev_port, ODPP_LOCAL);
280 goto err_destroy_table;
284 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
285 if (!dp->stats_percpu)
286 goto err_destroy_local_port;
288 rcu_assign_pointer(dps[dp_idx], dp);
289 mutex_unlock(&dp_mutex);
296 err_destroy_local_port:
297 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
299 tbl_destroy(dp->table, NULL);
303 module_put(THIS_MODULE);
305 mutex_unlock(&dp_mutex);
311 static void do_destroy_dp(struct datapath *dp)
313 struct dp_port *p, *n;
316 list_for_each_entry_safe (p, n, &dp->port_list, node)
317 if (p->port_no != ODPP_LOCAL)
318 dp_detach_port(p, 1);
322 rcu_assign_pointer(dps[dp->dp_idx], NULL);
324 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
326 tbl_destroy(dp->table, flow_free_tbl);
328 for (i = 0; i < DP_N_QUEUES; i++)
329 skb_queue_purge(&dp->queues[i]);
330 for (i = 0; i < DP_MAX_GROUPS; i++)
331 kfree(dp->groups[i]);
332 free_percpu(dp->stats_percpu);
333 kobject_put(&dp->ifobj);
334 module_put(THIS_MODULE);
337 static int destroy_dp(int dp_idx)
343 mutex_lock(&dp_mutex);
353 mutex_unlock(&dp_mutex);
358 static void release_dp_port(struct kobject *kobj)
360 struct dp_port *p = container_of(kobj, struct dp_port, kobj);
364 static struct kobj_type brport_ktype = {
366 .sysfs_ops = &brport_sysfs_ops,
368 .release = release_dp_port
371 /* Called with RTNL lock and dp_mutex. */
372 static int new_dp_port(struct datapath *dp, struct odp_port *odp_port, int port_no)
378 vport = vport_locate(odp_port->devname);
382 if (odp_port->flags & ODP_PORT_INTERNAL)
383 vport = vport_add(odp_port->devname, "internal", NULL);
385 vport = vport_add(odp_port->devname, "netdev", NULL);
390 return PTR_ERR(vport);
393 p = kzalloc(sizeof(*p), GFP_KERNEL);
397 p->port_no = port_no;
400 atomic_set(&p->sflow_pool, 0);
402 err = vport_attach(vport, p);
408 rcu_assign_pointer(dp->ports[port_no], p);
409 list_add_rcu(&p->node, &dp->port_list);
412 /* Initialize kobject for bridge. This will be added as
413 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
415 kobject_init(&p->kobj, &brport_ktype);
417 dp_ifinfo_notify(RTM_NEWLINK, p);
422 static int attach_port(int dp_idx, struct odp_port __user *portp)
425 struct odp_port port;
430 if (copy_from_user(&port, portp, sizeof port))
432 port.devname[IFNAMSIZ - 1] = '\0';
435 dp = get_dp_locked(dp_idx);
438 goto out_unlock_rtnl;
440 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
441 if (!dp->ports[port_no])
447 err = new_dp_port(dp, &port, port_no);
451 set_internal_devs_mtu(dp);
452 dp_sysfs_add_if(dp->ports[port_no]);
454 err = put_user(port_no, &portp->port);
457 mutex_unlock(&dp->mutex);
464 int dp_detach_port(struct dp_port *p, int may_delete)
466 struct vport *vport = p->vport;
471 if (p->port_no != ODPP_LOCAL)
473 dp_ifinfo_notify(RTM_DELLINK, p);
475 /* First drop references to device. */
477 list_del_rcu(&p->node);
478 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
480 err = vport_detach(vport);
484 /* Then wait until no one is still using it, and destroy it. */
488 const char *port_type = vport_get_type(vport);
490 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
497 kobject_put(&p->kobj);
502 static int detach_port(int dp_idx, int port_no)
509 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
513 dp = get_dp_locked(dp_idx);
516 goto out_unlock_rtnl;
518 p = dp->ports[port_no];
523 err = dp_detach_port(p, 1);
526 mutex_unlock(&dp->mutex);
533 static void suppress_loop(struct datapath *dp, struct sw_flow_actions *actions)
536 pr_warn("%s: flow looped %d times, dropping\n",
537 dp_name(dp), DP_MAX_LOOPS);
538 actions->n_actions = 0;
541 /* Must be called with rcu_read_lock. */
542 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
544 struct datapath *dp = p->dp;
545 struct dp_stats_percpu *stats;
546 int stats_counter_off;
547 struct odp_flow_key key;
548 struct tbl_node *flow_node;
549 struct sw_flow *flow;
550 struct sw_flow_actions *acts;
551 struct loop_counter *loop;
554 OVS_CB(skb)->dp_port = p;
556 /* Extract flow from 'skb' into 'key'. */
557 error = flow_extract(skb, p ? p->port_no : ODPP_NONE, &key);
558 if (unlikely(error)) {
563 if (OVS_CB(skb)->is_frag && dp->drop_frags) {
565 stats_counter_off = offsetof(struct dp_stats_percpu, n_frags);
570 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
571 if (unlikely(!flow_node)) {
572 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
573 stats_counter_off = offsetof(struct dp_stats_percpu, n_missed);
577 flow = flow_cast(flow_node);
578 flow_used(flow, skb);
580 acts = rcu_dereference(flow->sf_acts);
582 /* Check whether we've looped too much. */
583 loop = &get_cpu_var(dp_loop_counters).counters[!!in_interrupt()];
584 if (unlikely(++loop->count > DP_MAX_LOOPS))
585 loop->looping = true;
586 if (unlikely(loop->looping)) {
587 suppress_loop(dp, acts);
591 /* Execute actions. */
592 execute_actions(dp, skb, &key, acts->actions, acts->n_actions, GFP_ATOMIC);
593 stats_counter_off = offsetof(struct dp_stats_percpu, n_hit);
595 /* Check whether sub-actions looped too much. */
596 if (unlikely(loop->looping))
597 suppress_loop(dp, acts);
600 /* Decrement loop counter. */
602 loop->looping = false;
603 put_cpu_var(dp_loop_counters);
606 /* Update datapath statistics. */
608 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
610 write_seqcount_begin(&stats->seqlock);
611 (*(u64 *)((u8 *)stats + stats_counter_off))++;
612 write_seqcount_end(&stats->seqlock);
617 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
618 /* This code is based on skb_checksum_setup() from Xen's net/dev/core.c. We
619 * can't call this function directly because it isn't exported in all
621 int vswitch_skb_checksum_setup(struct sk_buff *skb)
626 __u16 csum_start, csum_offset;
628 if (!skb->proto_csum_blank)
631 if (skb->protocol != htons(ETH_P_IP))
634 if (!pskb_may_pull(skb, skb_network_header(skb) + sizeof(struct iphdr) - skb->data))
638 th = skb_network_header(skb) + 4 * iph->ihl;
640 csum_start = th - skb->head;
641 switch (iph->protocol) {
643 csum_offset = offsetof(struct tcphdr, check);
646 csum_offset = offsetof(struct udphdr, check);
650 pr_err("Attempting to checksum a non-TCP/UDP packet, "
651 "dropping a protocol %d packet",
656 if (!pskb_may_pull(skb, th + csum_offset + 2 - skb->data))
659 skb->ip_summed = CHECKSUM_PARTIAL;
660 skb->proto_csum_blank = 0;
662 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
663 skb->csum_start = csum_start;
664 skb->csum_offset = csum_offset;
666 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
667 skb->csum = csum_offset;
675 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
677 /* Types of checksums that we can receive (these all refer to L4 checksums):
678 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
679 * (though not verified) checksum in packet but not in skb->csum. Packets
680 * from the bridge local port will also have this type.
681 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
682 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
683 * a valid skb->csum. Importantly, both contain a full checksum (not
684 * verified) in the packet itself. The only difference is that if the
685 * packet gets to L4 processing on this machine (not in DomU) we won't
686 * have to recompute the checksum to verify. Most hardware devices do not
687 * produce packets with this type, even if they support receive checksum
688 * offloading (they produce type #5).
689 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
690 * be computed if it is sent off box. Unfortunately on earlier kernels,
691 * this case is impossible to distinguish from #2, despite having opposite
692 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
693 * to distinguish the different states.
694 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
695 * generated locally by a Xen DomU and has a partial checksum. If it is
696 * handled on this machine (Dom0 or DomU), then the checksum will not be
697 * computed. If it goes off box, the checksum in the packet needs to be
698 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
699 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
700 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
701 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
702 * full checksum or using a protocol without a checksum. skb->csum is
703 * undefined. This is common from devices with receive checksum
704 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
705 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
707 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
708 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
709 * based on whether it is on the transmit or receive path. After the datapath
710 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
711 * checksum, we will panic. Since we can receive packets with checksums, we
712 * assume that all CHECKSUM_HW packets have checksums and map them to
713 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
714 * packet is processed by the local IP stack, in which case it will need to
715 * be reverified). If we receive a packet with CHECKSUM_HW that really means
716 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
717 * shouldn't be any devices that do this with bridging. */
718 void compute_ip_summed(struct sk_buff *skb, bool xmit)
720 /* For our convenience these defines change repeatedly between kernel
721 * versions, so we can't just copy them over... */
722 switch (skb->ip_summed) {
724 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
726 case CHECKSUM_UNNECESSARY:
727 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
730 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
731 * However, on the receive side we should only get CHECKSUM_PARTIAL
732 * packets from Xen, which uses some special fields to represent this
733 * (see below). Since we can only make one type work, pick the one
734 * that actually happens in practice.
736 * On the transmit side (basically after skb_checksum_setup()
737 * has been run or on internal dev transmit), packets with
738 * CHECKSUM_COMPLETE aren't generated, so assume CHECKSUM_PARTIAL. */
741 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
743 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
747 case CHECKSUM_COMPLETE:
748 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
750 case CHECKSUM_PARTIAL:
751 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
755 pr_err("unknown checksum type %d\n", skb->ip_summed);
756 /* None seems the safest... */
757 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
760 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
761 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
762 * kernels. It should not be set on the transmit path though. */
763 if (skb->proto_csum_blank)
764 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
766 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
770 /* This function closely resembles skb_forward_csum() used by the bridge. It
771 * is slightly different because we are only concerned with bridging and not
772 * other types of forwarding and can get away with slightly more optimal
774 void forward_ip_summed(struct sk_buff *skb)
777 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
778 skb->ip_summed = CHECKSUM_NONE;
782 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
783 * unless we broke up a GSO packet. */
784 static int queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
785 int queue_no, u32 arg)
787 struct sk_buff *nskb;
791 if (OVS_CB(skb)->dp_port)
792 port_no = OVS_CB(skb)->dp_port->port_no;
794 port_no = ODPP_LOCAL;
797 struct odp_msg *header;
802 err = skb_cow(skb, sizeof *header);
806 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
807 header->type = queue_no;
808 header->length = skb->len;
809 header->port = port_no;
810 header->reserved = 0;
812 skb_queue_tail(queue, skb);
820 while ((skb = nskb) != NULL) {
827 int dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
830 struct dp_stats_percpu *stats;
831 struct sk_buff_head *queue;
834 WARN_ON_ONCE(skb_shared(skb));
835 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
836 queue = &dp->queues[queue_no];
838 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
841 forward_ip_summed(skb);
843 err = vswitch_skb_checksum_setup(skb);
847 /* Break apart GSO packets into their component pieces. Otherwise
848 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
849 if (skb_is_gso(skb)) {
850 struct sk_buff *nskb = skb_gso_segment(skb, NETIF_F_SG | NETIF_F_HW_CSUM);
854 if (unlikely(IS_ERR(skb))) {
859 /* XXX This case might not be possible. It's hard to
860 * tell from the skb_gso_segment() code and comment. */
864 err = queue_control_packets(skb, queue, queue_no, arg);
865 wake_up_interruptible(&dp->waitqueue);
872 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
874 write_seqcount_begin(&stats->seqlock);
876 write_seqcount_end(&stats->seqlock);
883 static int flush_flows(struct datapath *dp)
885 struct tbl *old_table = rcu_dereference(dp->table);
886 struct tbl *new_table;
888 new_table = tbl_create(0);
892 rcu_assign_pointer(dp->table, new_table);
894 tbl_deferred_destroy(old_table, flow_free_tbl);
899 static int validate_actions(const struct sw_flow_actions *actions)
903 for (i = 0; i < actions->n_actions; i++) {
904 const union odp_action *a = &actions->actions[i];
907 if (a->output.port >= DP_MAX_PORTS)
911 case ODPAT_OUTPUT_GROUP:
912 if (a->output_group.group >= DP_MAX_GROUPS)
916 case ODPAT_SET_VLAN_VID:
917 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
921 case ODPAT_SET_VLAN_PCP:
922 if (a->vlan_pcp.vlan_pcp
923 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
927 case ODPAT_SET_NW_TOS:
928 if (a->nw_tos.nw_tos & INET_ECN_MASK)
933 if (a->type >= ODPAT_N_ACTIONS)
942 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
944 struct sw_flow_actions *actions;
947 actions = flow_actions_alloc(flow->n_actions);
948 error = PTR_ERR(actions);
953 if (copy_from_user(actions->actions, flow->actions,
954 flow->n_actions * sizeof(union odp_action)))
955 goto error_free_actions;
956 error = validate_actions(actions);
958 goto error_free_actions;
965 return ERR_PTR(error);
968 static struct timespec get_time_offset(void)
970 struct timespec now_mono, now_jiffies;
972 ktime_get_ts(&now_mono);
973 jiffies_to_timespec(jiffies, &now_jiffies);
974 return timespec_sub(now_mono, now_jiffies);
977 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats,
978 struct timespec time_offset)
981 struct timespec flow_ts, used;
983 jiffies_to_timespec(flow->used, &flow_ts);
984 set_normalized_timespec(&used, flow_ts.tv_sec + time_offset.tv_sec,
985 flow_ts.tv_nsec + time_offset.tv_nsec);
987 stats->used_sec = used.tv_sec;
988 stats->used_nsec = used.tv_nsec;
991 stats->used_nsec = 0;
994 stats->n_packets = flow->packet_count;
995 stats->n_bytes = flow->byte_count;
997 stats->tcp_flags = flow->tcp_flags;
1001 static void clear_stats(struct sw_flow *flow)
1004 flow->tcp_flags = 0;
1005 flow->packet_count = 0;
1006 flow->byte_count = 0;
1009 static int expand_table(struct datapath *dp)
1011 struct tbl *old_table = rcu_dereference(dp->table);
1012 struct tbl *new_table;
1014 new_table = tbl_expand(old_table);
1015 if (IS_ERR(new_table))
1016 return PTR_ERR(new_table);
1018 rcu_assign_pointer(dp->table, new_table);
1019 tbl_deferred_destroy(old_table, NULL);
1024 static int do_put_flow(struct datapath *dp, struct odp_flow_put *uf,
1025 struct odp_flow_stats *stats)
1027 struct tbl_node *flow_node;
1028 struct sw_flow *flow;
1032 memset(uf->flow.key.reserved, 0, sizeof uf->flow.key.reserved);
1034 table = rcu_dereference(dp->table);
1035 flow_node = tbl_lookup(table, &uf->flow.key, flow_hash(&uf->flow.key), flow_cmp);
1038 struct sw_flow_actions *acts;
1041 if (!(uf->flags & ODPPF_CREATE))
1044 /* Expand table, if necessary, to make room. */
1045 if (tbl_count(table) >= tbl_n_buckets(table)) {
1046 error = expand_table(dp);
1049 table = rcu_dereference(dp->table);
1052 /* Allocate flow. */
1054 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
1057 flow->key = uf->flow.key;
1058 spin_lock_init(&flow->lock);
1061 /* Obtain actions. */
1062 acts = get_actions(&uf->flow);
1063 error = PTR_ERR(acts);
1065 goto error_free_flow;
1066 rcu_assign_pointer(flow->sf_acts, acts);
1068 /* Put flow in bucket. */
1069 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1071 goto error_free_flow_acts;
1073 memset(stats, 0, sizeof(struct odp_flow_stats));
1075 /* We found a matching flow. */
1076 struct sw_flow_actions *old_acts, *new_acts;
1078 flow = flow_cast(flow_node);
1080 /* Bail out if we're not allowed to modify an existing flow. */
1082 if (!(uf->flags & ODPPF_MODIFY))
1086 new_acts = get_actions(&uf->flow);
1087 error = PTR_ERR(new_acts);
1088 if (IS_ERR(new_acts))
1090 old_acts = rcu_dereference(flow->sf_acts);
1091 if (old_acts->n_actions != new_acts->n_actions ||
1092 memcmp(old_acts->actions, new_acts->actions,
1093 sizeof(union odp_action) * old_acts->n_actions)) {
1094 rcu_assign_pointer(flow->sf_acts, new_acts);
1095 flow_deferred_free_acts(old_acts);
1100 /* Fetch stats, then clear them if necessary. */
1101 spin_lock_bh(&flow->lock);
1102 get_stats(flow, stats, get_time_offset());
1103 if (uf->flags & ODPPF_ZERO_STATS)
1105 spin_unlock_bh(&flow->lock);
1110 error_free_flow_acts:
1111 kfree(flow->sf_acts);
1113 kmem_cache_free(flow_cache, flow);
1118 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
1120 struct odp_flow_stats stats;
1121 struct odp_flow_put uf;
1124 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
1127 error = do_put_flow(dp, &uf, &stats);
1131 if (copy_to_user(&ufp->flow.stats, &stats,
1132 sizeof(struct odp_flow_stats)))
1138 static int do_answer_query(struct sw_flow *flow, u32 query_flags,
1139 struct timespec time_offset,
1140 struct odp_flow_stats __user *ustats,
1141 union odp_action __user *actions,
1142 u32 __user *n_actionsp)
1144 struct sw_flow_actions *sf_acts;
1145 struct odp_flow_stats stats;
1148 spin_lock_bh(&flow->lock);
1149 get_stats(flow, &stats, time_offset);
1150 if (query_flags & ODPFF_ZERO_TCP_FLAGS)
1151 flow->tcp_flags = 0;
1153 spin_unlock_bh(&flow->lock);
1155 if (copy_to_user(ustats, &stats, sizeof(struct odp_flow_stats)) ||
1156 get_user(n_actions, n_actionsp))
1162 sf_acts = rcu_dereference(flow->sf_acts);
1163 if (put_user(sf_acts->n_actions, n_actionsp) ||
1164 (actions && copy_to_user(actions, sf_acts->actions,
1165 sizeof(union odp_action) *
1166 min(sf_acts->n_actions, n_actions))))
1172 static int answer_query(struct sw_flow *flow, u32 query_flags,
1173 struct timespec time_offset,
1174 struct odp_flow __user *ufp)
1176 union odp_action *actions;
1178 if (get_user(actions, &ufp->actions))
1181 return do_answer_query(flow, query_flags, time_offset,
1182 &ufp->stats, actions, &ufp->n_actions);
1185 static struct sw_flow *do_del_flow(struct datapath *dp, struct odp_flow_key *key)
1187 struct tbl *table = rcu_dereference(dp->table);
1188 struct tbl_node *flow_node;
1191 memset(key->reserved, 0, sizeof key->reserved);
1192 flow_node = tbl_lookup(table, key, flow_hash(key), flow_cmp);
1194 return ERR_PTR(-ENOENT);
1196 error = tbl_remove(table, flow_node);
1198 return ERR_PTR(error);
1200 /* XXX Returned flow_node's statistics might lose a few packets, since
1201 * other CPUs can be using this flow. We used to synchronize_rcu() to
1202 * make sure that we get completely accurate stats, but that blows our
1203 * performance, badly. */
1204 return flow_cast(flow_node);
1207 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1209 struct sw_flow *flow;
1213 if (copy_from_user(&uf, ufp, sizeof uf))
1216 flow = do_del_flow(dp, &uf.key);
1218 return PTR_ERR(flow);
1220 error = answer_query(flow, 0, get_time_offset(), ufp);
1221 flow_deferred_free(flow);
1225 static int do_query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1227 struct tbl *table = rcu_dereference(dp->table);
1228 struct timespec time_offset;
1231 time_offset = get_time_offset();
1233 for (i = 0; i < flowvec->n_flows; i++) {
1234 struct odp_flow __user *ufp = &flowvec->flows[i];
1236 struct tbl_node *flow_node;
1239 if (copy_from_user(&uf, ufp, sizeof uf))
1241 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1243 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1245 error = put_user(ENOENT, &ufp->stats.error);
1247 error = answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1251 return flowvec->n_flows;
1254 struct list_flows_cbdata {
1255 struct odp_flow __user *uflows;
1258 struct timespec time_offset;
1261 static int list_flow(struct tbl_node *node, void *cbdata_)
1263 struct sw_flow *flow = flow_cast(node);
1264 struct list_flows_cbdata *cbdata = cbdata_;
1265 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1268 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1270 error = answer_query(flow, 0, cbdata->time_offset, ufp);
1274 if (cbdata->listed_flows >= cbdata->n_flows)
1275 return cbdata->listed_flows;
1279 static int do_list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1281 struct list_flows_cbdata cbdata;
1284 if (!flowvec->n_flows)
1287 cbdata.uflows = flowvec->flows;
1288 cbdata.n_flows = flowvec->n_flows;
1289 cbdata.listed_flows = 0;
1290 cbdata.time_offset = get_time_offset();
1292 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1293 return error ? error : cbdata.listed_flows;
1296 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1297 int (*function)(struct datapath *,
1298 const struct odp_flowvec *))
1300 struct odp_flowvec __user *uflowvec;
1301 struct odp_flowvec flowvec;
1304 uflowvec = (struct odp_flowvec __user *)argp;
1305 if (copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1308 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1311 retval = function(dp, &flowvec);
1312 return (retval < 0 ? retval
1313 : retval == flowvec.n_flows ? 0
1314 : put_user(retval, &uflowvec->n_flows));
1317 static int do_execute(struct datapath *dp, const struct odp_execute *execute)
1319 struct odp_flow_key key;
1320 struct sk_buff *skb;
1321 struct sw_flow_actions *actions;
1326 if (execute->length < ETH_HLEN || execute->length > 65535)
1329 actions = flow_actions_alloc(execute->n_actions);
1330 if (IS_ERR(actions)) {
1331 err = PTR_ERR(actions);
1336 if (copy_from_user(actions->actions, execute->actions,
1337 execute->n_actions * sizeof *execute->actions))
1338 goto error_free_actions;
1340 err = validate_actions(actions);
1342 goto error_free_actions;
1345 skb = alloc_skb(execute->length, GFP_KERNEL);
1347 goto error_free_actions;
1349 if (execute->in_port < DP_MAX_PORTS)
1350 OVS_CB(skb)->dp_port = dp->ports[execute->in_port];
1352 OVS_CB(skb)->dp_port = NULL;
1355 if (copy_from_user(skb_put(skb, execute->length), execute->data,
1357 goto error_free_skb;
1359 skb_reset_mac_header(skb);
1362 /* Normally, setting the skb 'protocol' field would be handled by a
1363 * call to eth_type_trans(), but it assumes there's a sending
1364 * device, which we may not have. */
1365 if (ntohs(eth->h_proto) >= 1536)
1366 skb->protocol = eth->h_proto;
1368 skb->protocol = htons(ETH_P_802_2);
1370 err = flow_extract(skb, execute->in_port, &key);
1372 goto error_free_skb;
1375 err = execute_actions(dp, skb, &key, actions->actions,
1376 actions->n_actions, GFP_KERNEL);
1390 static int execute_packet(struct datapath *dp, const struct odp_execute __user *executep)
1392 struct odp_execute execute;
1394 if (copy_from_user(&execute, executep, sizeof execute))
1397 return do_execute(dp, &execute);
1400 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1402 struct tbl *table = rcu_dereference(dp->table);
1403 struct odp_stats stats;
1406 stats.n_flows = tbl_count(table);
1407 stats.cur_capacity = tbl_n_buckets(table);
1408 stats.max_capacity = TBL_MAX_BUCKETS;
1409 stats.n_ports = dp->n_ports;
1410 stats.max_ports = DP_MAX_PORTS;
1411 stats.max_groups = DP_MAX_GROUPS;
1412 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1413 for_each_possible_cpu(i) {
1414 const struct dp_stats_percpu *percpu_stats;
1415 struct dp_stats_percpu local_stats;
1418 percpu_stats = per_cpu_ptr(dp->stats_percpu, i);
1421 seqcount = read_seqcount_begin(&percpu_stats->seqlock);
1422 local_stats = *percpu_stats;
1423 } while (read_seqcount_retry(&percpu_stats->seqlock, seqcount));
1425 stats.n_frags += local_stats.n_frags;
1426 stats.n_hit += local_stats.n_hit;
1427 stats.n_missed += local_stats.n_missed;
1428 stats.n_lost += local_stats.n_lost;
1430 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1431 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1432 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1435 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1436 int dp_min_mtu(const struct datapath *dp)
1443 list_for_each_entry_rcu (p, &dp->port_list, node) {
1446 /* Skip any internal ports, since that's what we're trying to
1448 if (is_internal_vport(p->vport))
1451 dev_mtu = vport_get_mtu(p->vport);
1452 if (!mtu || dev_mtu < mtu)
1456 return mtu ? mtu : ETH_DATA_LEN;
1459 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1460 * be called with RTNL lock. */
1461 void set_internal_devs_mtu(const struct datapath *dp)
1468 mtu = dp_min_mtu(dp);
1470 list_for_each_entry_rcu (p, &dp->port_list, node) {
1471 if (is_internal_vport(p->vport))
1472 vport_set_mtu(p->vport, mtu);
1476 static int put_port(const struct dp_port *p, struct odp_port __user *uop)
1480 memset(&op, 0, sizeof op);
1483 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1486 op.port = p->port_no;
1487 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1489 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1492 static int query_port(struct datapath *dp, struct odp_port __user *uport)
1494 struct odp_port port;
1496 if (copy_from_user(&port, uport, sizeof port))
1499 if (port.devname[0]) {
1500 struct vport *vport;
1501 struct dp_port *dp_port;
1504 port.devname[IFNAMSIZ - 1] = '\0';
1509 vport = vport_locate(port.devname);
1515 dp_port = vport_get_dp_port(vport);
1516 if (!dp_port || dp_port->dp != dp) {
1521 port.port = dp_port->port_no;
1530 if (port.port >= DP_MAX_PORTS)
1532 if (!dp->ports[port.port])
1536 return put_port(dp->ports[port.port], uport);
1539 static int do_list_ports(struct datapath *dp, struct odp_port __user *uports,
1546 list_for_each_entry_rcu (p, &dp->port_list, node) {
1547 if (put_port(p, &uports[idx]))
1549 if (idx++ >= n_ports)
1556 static int list_ports(struct datapath *dp, struct odp_portvec __user *upv)
1558 struct odp_portvec pv;
1561 if (copy_from_user(&pv, upv, sizeof pv))
1564 retval = do_list_ports(dp, pv.ports, pv.n_ports);
1568 return put_user(retval, &upv->n_ports);
1571 /* RCU callback for freeing a dp_port_group */
1572 static void free_port_group(struct rcu_head *rcu)
1574 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1578 static int do_set_port_group(struct datapath *dp, u16 __user *ports,
1579 int n_ports, int group)
1581 struct dp_port_group *new_group, *old_group;
1585 if (n_ports > DP_MAX_PORTS || group >= DP_MAX_GROUPS)
1589 new_group = kmalloc(sizeof *new_group + sizeof(u16) * n_ports, GFP_KERNEL);
1593 new_group->n_ports = n_ports;
1595 if (copy_from_user(new_group->ports, ports, sizeof(u16) * n_ports))
1598 old_group = rcu_dereference(dp->groups[group]);
1599 rcu_assign_pointer(dp->groups[group], new_group);
1601 call_rcu(&old_group->rcu, free_port_group);
1610 static int set_port_group(struct datapath *dp,
1611 const struct odp_port_group __user *upg)
1613 struct odp_port_group pg;
1615 if (copy_from_user(&pg, upg, sizeof pg))
1618 return do_set_port_group(dp, pg.ports, pg.n_ports, pg.group);
1621 static int do_get_port_group(struct datapath *dp,
1622 u16 __user *ports, int n_ports, int group,
1623 u16 __user *n_portsp)
1625 struct dp_port_group *g;
1628 if (group >= DP_MAX_GROUPS)
1631 g = dp->groups[group];
1632 n_copy = g ? min_t(int, g->n_ports, n_ports) : 0;
1633 if (n_copy && copy_to_user(ports, g->ports, n_copy * sizeof(u16)))
1636 if (put_user(g ? g->n_ports : 0, n_portsp))
1642 static int get_port_group(struct datapath *dp, struct odp_port_group __user *upg)
1644 struct odp_port_group pg;
1646 if (copy_from_user(&pg, upg, sizeof pg))
1649 return do_get_port_group(dp, pg.ports, pg.n_ports, pg.group, &upg->n_ports);
1652 static int get_listen_mask(const struct file *f)
1654 return (long)f->private_data;
1657 static void set_listen_mask(struct file *f, int listen_mask)
1659 f->private_data = (void*)(long)listen_mask;
1662 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1665 int dp_idx = iminor(f->f_dentry->d_inode);
1666 struct datapath *dp;
1667 int drop_frags, listeners, port_no;
1668 unsigned int sflow_probability;
1671 /* Handle commands with special locking requirements up front. */
1674 err = create_dp(dp_idx, (char __user *)argp);
1677 case ODP_DP_DESTROY:
1678 err = destroy_dp(dp_idx);
1681 case ODP_PORT_ATTACH:
1682 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1685 case ODP_PORT_DETACH:
1686 err = get_user(port_no, (int __user *)argp);
1688 err = detach_port(dp_idx, port_no);
1692 err = vport_user_add((struct odp_vport_add __user *)argp);
1696 err = vport_user_mod((struct odp_vport_mod __user *)argp);
1700 err = vport_user_del((char __user *)argp);
1703 case ODP_VPORT_STATS_GET:
1704 err = vport_user_stats_get((struct odp_vport_stats_req __user *)argp);
1707 case ODP_VPORT_STATS_SET:
1708 err = vport_user_stats_set((struct odp_vport_stats_req __user *)argp);
1711 case ODP_VPORT_ETHER_GET:
1712 err = vport_user_ether_get((struct odp_vport_ether __user *)argp);
1715 case ODP_VPORT_ETHER_SET:
1716 err = vport_user_ether_set((struct odp_vport_ether __user *)argp);
1719 case ODP_VPORT_MTU_GET:
1720 err = vport_user_mtu_get((struct odp_vport_mtu __user *)argp);
1723 case ODP_VPORT_MTU_SET:
1724 err = vport_user_mtu_set((struct odp_vport_mtu __user *)argp);
1728 dp = get_dp_locked(dp_idx);
1735 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1738 case ODP_GET_DROP_FRAGS:
1739 err = put_user(dp->drop_frags, (int __user *)argp);
1742 case ODP_SET_DROP_FRAGS:
1743 err = get_user(drop_frags, (int __user *)argp);
1747 if (drop_frags != 0 && drop_frags != 1)
1749 dp->drop_frags = drop_frags;
1753 case ODP_GET_LISTEN_MASK:
1754 err = put_user(get_listen_mask(f), (int __user *)argp);
1757 case ODP_SET_LISTEN_MASK:
1758 err = get_user(listeners, (int __user *)argp);
1762 if (listeners & ~ODPL_ALL)
1765 set_listen_mask(f, listeners);
1768 case ODP_GET_SFLOW_PROBABILITY:
1769 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1772 case ODP_SET_SFLOW_PROBABILITY:
1773 err = get_user(sflow_probability, (unsigned int __user *)argp);
1775 dp->sflow_probability = sflow_probability;
1778 case ODP_PORT_QUERY:
1779 err = query_port(dp, (struct odp_port __user *)argp);
1783 err = list_ports(dp, (struct odp_portvec __user *)argp);
1786 case ODP_PORT_GROUP_SET:
1787 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1790 case ODP_PORT_GROUP_GET:
1791 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1794 case ODP_FLOW_FLUSH:
1795 err = flush_flows(dp);
1799 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1803 err = del_flow(dp, (struct odp_flow __user *)argp);
1807 err = do_flowvec_ioctl(dp, argp, do_query_flows);
1811 err = do_flowvec_ioctl(dp, argp, do_list_flows);
1815 err = execute_packet(dp, (struct odp_execute __user *)argp);
1822 mutex_unlock(&dp->mutex);
1827 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1830 for (i = 0; i < DP_N_QUEUES; i++) {
1831 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1837 #ifdef CONFIG_COMPAT
1838 static int compat_list_ports(struct datapath *dp, struct compat_odp_portvec __user *upv)
1840 struct compat_odp_portvec pv;
1843 if (copy_from_user(&pv, upv, sizeof pv))
1846 retval = do_list_ports(dp, compat_ptr(pv.ports), pv.n_ports);
1850 return put_user(retval, &upv->n_ports);
1853 static int compat_set_port_group(struct datapath *dp, const struct compat_odp_port_group __user *upg)
1855 struct compat_odp_port_group pg;
1857 if (copy_from_user(&pg, upg, sizeof pg))
1860 return do_set_port_group(dp, compat_ptr(pg.ports), pg.n_ports, pg.group);
1863 static int compat_get_port_group(struct datapath *dp, struct compat_odp_port_group __user *upg)
1865 struct compat_odp_port_group pg;
1867 if (copy_from_user(&pg, upg, sizeof pg))
1870 return do_get_port_group(dp, compat_ptr(pg.ports), pg.n_ports,
1871 pg.group, &upg->n_ports);
1874 static int compat_get_flow(struct odp_flow *flow, const struct compat_odp_flow __user *compat)
1876 compat_uptr_t actions;
1878 if (!access_ok(VERIFY_READ, compat, sizeof(struct compat_odp_flow)) ||
1879 __copy_from_user(&flow->stats, &compat->stats, sizeof(struct odp_flow_stats)) ||
1880 __copy_from_user(&flow->key, &compat->key, sizeof(struct odp_flow_key)) ||
1881 __get_user(actions, &compat->actions) ||
1882 __get_user(flow->n_actions, &compat->n_actions) ||
1883 __get_user(flow->flags, &compat->flags))
1886 flow->actions = compat_ptr(actions);
1890 static int compat_put_flow(struct datapath *dp, struct compat_odp_flow_put __user *ufp)
1892 struct odp_flow_stats stats;
1893 struct odp_flow_put fp;
1896 if (compat_get_flow(&fp.flow, &ufp->flow) ||
1897 get_user(fp.flags, &ufp->flags))
1900 error = do_put_flow(dp, &fp, &stats);
1904 if (copy_to_user(&ufp->flow.stats, &stats,
1905 sizeof(struct odp_flow_stats)))
1911 static int compat_answer_query(struct sw_flow *flow, u32 query_flags,
1912 struct timespec time_offset,
1913 struct compat_odp_flow __user *ufp)
1915 compat_uptr_t actions;
1917 if (get_user(actions, &ufp->actions))
1920 return do_answer_query(flow, query_flags, time_offset, &ufp->stats,
1921 compat_ptr(actions), &ufp->n_actions);
1924 static int compat_del_flow(struct datapath *dp, struct compat_odp_flow __user *ufp)
1926 struct sw_flow *flow;
1930 if (compat_get_flow(&uf, ufp))
1933 flow = do_del_flow(dp, &uf.key);
1935 return PTR_ERR(flow);
1937 error = compat_answer_query(flow, 0, get_time_offset(), ufp);
1938 flow_deferred_free(flow);
1942 static int compat_query_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1944 struct tbl *table = rcu_dereference(dp->table);
1945 struct timespec time_offset;
1948 time_offset = get_time_offset();
1950 for (i = 0; i < n_flows; i++) {
1951 struct compat_odp_flow __user *ufp = &flows[i];
1953 struct tbl_node *flow_node;
1956 if (compat_get_flow(&uf, ufp))
1958 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1960 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1962 error = put_user(ENOENT, &ufp->stats.error);
1964 error = compat_answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1971 struct compat_list_flows_cbdata {
1972 struct compat_odp_flow __user *uflows;
1975 struct timespec time_offset;
1978 static int compat_list_flow(struct tbl_node *node, void *cbdata_)
1980 struct sw_flow *flow = flow_cast(node);
1981 struct compat_list_flows_cbdata *cbdata = cbdata_;
1982 struct compat_odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1985 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1987 error = compat_answer_query(flow, 0, cbdata->time_offset, ufp);
1991 if (cbdata->listed_flows >= cbdata->n_flows)
1992 return cbdata->listed_flows;
1996 static int compat_list_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1998 struct compat_list_flows_cbdata cbdata;
2004 cbdata.uflows = flows;
2005 cbdata.n_flows = n_flows;
2006 cbdata.listed_flows = 0;
2007 cbdata.time_offset = get_time_offset();
2009 error = tbl_foreach(rcu_dereference(dp->table), compat_list_flow, &cbdata);
2010 return error ? error : cbdata.listed_flows;
2013 static int compat_flowvec_ioctl(struct datapath *dp, unsigned long argp,
2014 int (*function)(struct datapath *,
2015 struct compat_odp_flow *,
2018 struct compat_odp_flowvec __user *uflowvec;
2019 struct compat_odp_flow __user *flows;
2020 struct compat_odp_flowvec flowvec;
2023 uflowvec = compat_ptr(argp);
2024 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
2025 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
2028 if (flowvec.n_flows > INT_MAX / sizeof(struct compat_odp_flow))
2031 flows = compat_ptr(flowvec.flows);
2032 if (!access_ok(VERIFY_WRITE, flows,
2033 flowvec.n_flows * sizeof(struct compat_odp_flow)))
2036 retval = function(dp, flows, flowvec.n_flows);
2037 return (retval < 0 ? retval
2038 : retval == flowvec.n_flows ? 0
2039 : put_user(retval, &uflowvec->n_flows));
2042 static int compat_execute(struct datapath *dp, const struct compat_odp_execute __user *uexecute)
2044 struct odp_execute execute;
2045 compat_uptr_t actions;
2048 if (!access_ok(VERIFY_READ, uexecute, sizeof(struct compat_odp_execute)) ||
2049 __get_user(execute.in_port, &uexecute->in_port) ||
2050 __get_user(actions, &uexecute->actions) ||
2051 __get_user(execute.n_actions, &uexecute->n_actions) ||
2052 __get_user(data, &uexecute->data) ||
2053 __get_user(execute.length, &uexecute->length))
2056 execute.actions = compat_ptr(actions);
2057 execute.data = compat_ptr(data);
2059 return do_execute(dp, &execute);
2062 static long openvswitch_compat_ioctl(struct file *f, unsigned int cmd, unsigned long argp)
2064 int dp_idx = iminor(f->f_dentry->d_inode);
2065 struct datapath *dp;
2069 case ODP_DP_DESTROY:
2070 case ODP_FLOW_FLUSH:
2071 /* Ioctls that don't need any translation at all. */
2072 return openvswitch_ioctl(f, cmd, argp);
2075 case ODP_PORT_ATTACH:
2076 case ODP_PORT_DETACH:
2078 case ODP_VPORT_MTU_SET:
2079 case ODP_VPORT_MTU_GET:
2080 case ODP_VPORT_ETHER_SET:
2081 case ODP_VPORT_ETHER_GET:
2082 case ODP_VPORT_STATS_SET:
2083 case ODP_VPORT_STATS_GET:
2085 case ODP_GET_DROP_FRAGS:
2086 case ODP_SET_DROP_FRAGS:
2087 case ODP_SET_LISTEN_MASK:
2088 case ODP_GET_LISTEN_MASK:
2089 case ODP_SET_SFLOW_PROBABILITY:
2090 case ODP_GET_SFLOW_PROBABILITY:
2091 case ODP_PORT_QUERY:
2092 /* Ioctls that just need their pointer argument extended. */
2093 return openvswitch_ioctl(f, cmd, (unsigned long)compat_ptr(argp));
2095 case ODP_VPORT_ADD32:
2096 return compat_vport_user_add(compat_ptr(argp));
2098 case ODP_VPORT_MOD32:
2099 return compat_vport_user_mod(compat_ptr(argp));
2102 dp = get_dp_locked(dp_idx);
2108 case ODP_PORT_LIST32:
2109 err = compat_list_ports(dp, compat_ptr(argp));
2112 case ODP_PORT_GROUP_SET32:
2113 err = compat_set_port_group(dp, compat_ptr(argp));
2116 case ODP_PORT_GROUP_GET32:
2117 err = compat_get_port_group(dp, compat_ptr(argp));
2120 case ODP_FLOW_PUT32:
2121 err = compat_put_flow(dp, compat_ptr(argp));
2124 case ODP_FLOW_DEL32:
2125 err = compat_del_flow(dp, compat_ptr(argp));
2128 case ODP_FLOW_GET32:
2129 err = compat_flowvec_ioctl(dp, argp, compat_query_flows);
2132 case ODP_FLOW_LIST32:
2133 err = compat_flowvec_ioctl(dp, argp, compat_list_flows);
2137 err = compat_execute(dp, compat_ptr(argp));
2144 mutex_unlock(&dp->mutex);
2150 /* Unfortunately this function is not exported so this is a verbatim copy
2151 * from net/core/datagram.c in 2.6.30. */
2152 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
2153 u8 __user *to, int len,
2156 int start = skb_headlen(skb);
2158 int i, copy = start - offset;
2165 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
2169 if ((len -= copy) == 0)
2176 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2179 WARN_ON(start > offset + len);
2181 end = start + skb_shinfo(skb)->frags[i].size;
2182 if ((copy = end - offset) > 0) {
2186 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2187 struct page *page = frag->page;
2192 csum2 = csum_and_copy_to_user(vaddr +
2199 *csump = csum_block_add(*csump, csum2, pos);
2209 if (skb_shinfo(skb)->frag_list) {
2210 struct sk_buff *list = skb_shinfo(skb)->frag_list;
2212 for (; list; list=list->next) {
2215 WARN_ON(start > offset + len);
2217 end = start + list->len;
2218 if ((copy = end - offset) > 0) {
2222 if (skb_copy_and_csum_datagram(list,
2227 *csump = csum_block_add(*csump, csum2, pos);
2228 if ((len -= copy) == 0)
2244 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
2247 /* XXX is there sufficient synchronization here? */
2248 int listeners = get_listen_mask(f);
2249 int dp_idx = iminor(f->f_dentry->d_inode);
2250 struct datapath *dp = get_dp(dp_idx);
2251 struct sk_buff *skb;
2252 size_t copy_bytes, tot_copy_bytes;
2258 if (nbytes == 0 || !listeners)
2264 for (i = 0; i < DP_N_QUEUES; i++) {
2265 if (listeners & (1 << i)) {
2266 skb = skb_dequeue(&dp->queues[i]);
2272 if (f->f_flags & O_NONBLOCK) {
2277 wait_event_interruptible(dp->waitqueue,
2278 dp_has_packet_of_interest(dp,
2281 if (signal_pending(current)) {
2282 retval = -ERESTARTSYS;
2287 copy_bytes = tot_copy_bytes = min_t(size_t, skb->len, nbytes);
2290 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2291 if (copy_bytes == skb->len) {
2293 unsigned int csum_start, csum_offset;
2295 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
2296 csum_start = skb->csum_start - skb_headroom(skb);
2297 csum_offset = skb->csum_offset;
2299 csum_start = skb_transport_header(skb) - skb->data;
2300 csum_offset = skb->csum;
2302 BUG_ON(csum_start >= skb_headlen(skb));
2303 retval = skb_copy_and_csum_datagram(skb, csum_start, buf + csum_start,
2304 copy_bytes - csum_start, &csum);
2306 __sum16 __user *csump;
2308 copy_bytes = csum_start;
2309 csump = (__sum16 __user *)(buf + csum_start + csum_offset);
2311 BUG_ON((char *)csump + sizeof(__sum16) > buf + nbytes);
2312 put_user(csum_fold(csum), csump);
2315 retval = skb_checksum_help(skb);
2319 struct iovec __user iov;
2322 iov.iov_len = copy_bytes;
2323 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
2327 retval = tot_copy_bytes;
2335 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
2337 /* XXX is there sufficient synchronization here? */
2338 int dp_idx = iminor(file->f_dentry->d_inode);
2339 struct datapath *dp = get_dp(dp_idx);
2344 poll_wait(file, &dp->waitqueue, wait);
2345 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
2346 mask |= POLLIN | POLLRDNORM;
2348 mask = POLLIN | POLLRDNORM | POLLHUP;
2353 struct file_operations openvswitch_fops = {
2354 /* XXX .aio_read = openvswitch_aio_read, */
2355 .read = openvswitch_read,
2356 .poll = openvswitch_poll,
2357 .unlocked_ioctl = openvswitch_ioctl,
2358 #ifdef CONFIG_COMPAT
2359 .compat_ioctl = openvswitch_compat_ioctl,
2361 /* XXX .fasync = openvswitch_fasync, */
2366 static int __init dp_init(void)
2368 struct sk_buff *dummy_skb;
2371 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
2373 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
2381 goto error_flow_exit;
2383 err = register_netdevice_notifier(&dp_device_notifier);
2385 goto error_vport_exit;
2387 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
2389 goto error_unreg_notifier;
2393 error_unreg_notifier:
2394 unregister_netdevice_notifier(&dp_device_notifier);
2403 static void dp_cleanup(void)
2406 unregister_chrdev(major, "openvswitch");
2407 unregister_netdevice_notifier(&dp_device_notifier);
2412 module_init(dp_init);
2413 module_exit(dp_cleanup);
2415 MODULE_DESCRIPTION("Open vSwitch switching datapath");
2416 MODULE_LICENSE("GPL");