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 /* We limit the number of times that we pass into dp_process_received_packet()
73 * to avoid blowing out the stack in the event that we have a loop. */
75 int count; /* Count. */
76 bool looping; /* Loop detected? */
79 #define DP_MAX_LOOPS 5
81 /* We use a separate counter for each CPU for both interrupt and non-interrupt
82 * context in order to keep the limit deterministic for a given packet. */
83 struct percpu_loop_counters {
84 struct loop_counter counters[2];
87 static DEFINE_PER_CPU(struct percpu_loop_counters, dp_loop_counters);
89 static int new_dp_port(struct datapath *, struct odp_port *, int port_no);
91 /* Must be called with rcu_read_lock or dp_mutex. */
92 struct datapath *get_dp(int dp_idx)
94 if (dp_idx < 0 || dp_idx >= ODP_MAX)
96 return rcu_dereference(dps[dp_idx]);
98 EXPORT_SYMBOL_GPL(get_dp);
100 static struct datapath *get_dp_locked(int dp_idx)
104 mutex_lock(&dp_mutex);
107 mutex_lock(&dp->mutex);
108 mutex_unlock(&dp_mutex);
112 /* Must be called with rcu_read_lock or RTNL lock. */
113 const char *dp_name(const struct datapath *dp)
115 return vport_get_name(dp->ports[ODPP_LOCAL]->vport);
118 static inline size_t br_nlmsg_size(void)
120 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
121 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
122 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
123 + nla_total_size(4) /* IFLA_MASTER */
124 + nla_total_size(4) /* IFLA_MTU */
125 + nla_total_size(4) /* IFLA_LINK */
126 + nla_total_size(1); /* IFLA_OPERSTATE */
129 static int dp_fill_ifinfo(struct sk_buff *skb,
130 const struct dp_port *port,
131 int event, unsigned int flags)
133 const struct datapath *dp = port->dp;
134 int ifindex = vport_get_ifindex(port->vport);
135 int iflink = vport_get_iflink(port->vport);
136 struct ifinfomsg *hdr;
137 struct nlmsghdr *nlh;
145 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
149 hdr = nlmsg_data(nlh);
150 hdr->ifi_family = AF_BRIDGE;
152 hdr->ifi_type = ARPHRD_ETHER;
153 hdr->ifi_index = ifindex;
154 hdr->ifi_flags = vport_get_flags(port->vport);
157 NLA_PUT_STRING(skb, IFLA_IFNAME, vport_get_name(port->vport));
158 NLA_PUT_U32(skb, IFLA_MASTER, vport_get_ifindex(dp->ports[ODPP_LOCAL]->vport));
159 NLA_PUT_U32(skb, IFLA_MTU, vport_get_mtu(port->vport));
160 #ifdef IFLA_OPERSTATE
161 NLA_PUT_U8(skb, IFLA_OPERSTATE,
162 vport_is_running(port->vport)
163 ? vport_get_operstate(port->vport)
167 NLA_PUT(skb, IFLA_ADDRESS, ETH_ALEN,
168 vport_get_addr(port->vport));
170 if (ifindex != iflink)
171 NLA_PUT_U32(skb, IFLA_LINK,iflink);
173 return nlmsg_end(skb, nlh);
176 nlmsg_cancel(skb, nlh);
180 static void dp_ifinfo_notify(int event, struct dp_port *port)
185 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
189 err = dp_fill_ifinfo(skb, port, event, 0);
191 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
192 WARN_ON(err == -EMSGSIZE);
196 rtnl_notify(skb, &init_net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
200 rtnl_set_sk_err(&init_net, RTNLGRP_LINK, err);
203 static void release_dp(struct kobject *kobj)
205 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
209 static struct kobj_type dp_ktype = {
210 .release = release_dp
213 static int create_dp(int dp_idx, const char __user *devnamep)
215 struct odp_port internal_dev_port;
216 char devname[IFNAMSIZ];
222 int retval = strncpy_from_user(devname, devnamep, IFNAMSIZ);
226 } else if (retval >= IFNAMSIZ) {
231 snprintf(devname, sizeof devname, "of%d", dp_idx);
235 mutex_lock(&dp_mutex);
237 if (!try_module_get(THIS_MODULE))
240 /* Exit early if a datapath with that number already exists.
241 * (We don't use -EEXIST because that's ambiguous with 'devname'
242 * conflicting with an existing network device name.) */
248 dp = kzalloc(sizeof *dp, GFP_KERNEL);
251 INIT_LIST_HEAD(&dp->port_list);
252 mutex_init(&dp->mutex);
254 for (i = 0; i < DP_N_QUEUES; i++)
255 skb_queue_head_init(&dp->queues[i]);
256 init_waitqueue_head(&dp->waitqueue);
258 /* Initialize kobject for bridge. This will be added as
259 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
260 dp->ifobj.kset = NULL;
261 kobject_init(&dp->ifobj, &dp_ktype);
263 /* Allocate table. */
265 rcu_assign_pointer(dp->table, tbl_create(0));
269 /* Set up our datapath device. */
270 BUILD_BUG_ON(sizeof(internal_dev_port.devname) != sizeof(devname));
271 strcpy(internal_dev_port.devname, devname);
272 internal_dev_port.flags = ODP_PORT_INTERNAL;
273 err = new_dp_port(dp, &internal_dev_port, ODPP_LOCAL);
278 goto err_destroy_table;
282 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
283 if (!dp->stats_percpu)
284 goto err_destroy_local_port;
286 rcu_assign_pointer(dps[dp_idx], dp);
287 mutex_unlock(&dp_mutex);
294 err_destroy_local_port:
295 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
297 tbl_destroy(dp->table, NULL);
301 module_put(THIS_MODULE);
303 mutex_unlock(&dp_mutex);
309 static void do_destroy_dp(struct datapath *dp)
311 struct dp_port *p, *n;
314 list_for_each_entry_safe (p, n, &dp->port_list, node)
315 if (p->port_no != ODPP_LOCAL)
316 dp_detach_port(p, 1);
320 rcu_assign_pointer(dps[dp->dp_idx], NULL);
322 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
324 tbl_destroy(dp->table, flow_free_tbl);
326 for (i = 0; i < DP_N_QUEUES; i++)
327 skb_queue_purge(&dp->queues[i]);
328 for (i = 0; i < DP_MAX_GROUPS; i++)
329 kfree(dp->groups[i]);
330 free_percpu(dp->stats_percpu);
331 kobject_put(&dp->ifobj);
332 module_put(THIS_MODULE);
335 static int destroy_dp(int dp_idx)
341 mutex_lock(&dp_mutex);
351 mutex_unlock(&dp_mutex);
356 static void release_dp_port(struct kobject *kobj)
358 struct dp_port *p = container_of(kobj, struct dp_port, kobj);
362 static struct kobj_type brport_ktype = {
364 .sysfs_ops = &brport_sysfs_ops,
366 .release = release_dp_port
369 /* Called with RTNL lock and dp_mutex. */
370 static int new_dp_port(struct datapath *dp, struct odp_port *odp_port, int port_no)
376 vport = vport_locate(odp_port->devname);
380 if (odp_port->flags & ODP_PORT_INTERNAL)
381 vport = vport_add(odp_port->devname, "internal", NULL);
383 vport = vport_add(odp_port->devname, "netdev", NULL);
388 return PTR_ERR(vport);
391 p = kzalloc(sizeof(*p), GFP_KERNEL);
395 p->port_no = port_no;
398 atomic_set(&p->sflow_pool, 0);
400 err = vport_attach(vport, p);
406 rcu_assign_pointer(dp->ports[port_no], p);
407 list_add_rcu(&p->node, &dp->port_list);
410 /* Initialize kobject for bridge. This will be added as
411 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
413 kobject_init(&p->kobj, &brport_ktype);
415 dp_ifinfo_notify(RTM_NEWLINK, p);
420 static int attach_port(int dp_idx, struct odp_port __user *portp)
423 struct odp_port port;
428 if (copy_from_user(&port, portp, sizeof port))
430 port.devname[IFNAMSIZ - 1] = '\0';
433 dp = get_dp_locked(dp_idx);
436 goto out_unlock_rtnl;
438 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
439 if (!dp->ports[port_no])
445 err = new_dp_port(dp, &port, port_no);
449 set_internal_devs_mtu(dp);
450 dp_sysfs_add_if(dp->ports[port_no]);
452 err = put_user(port_no, &portp->port);
455 mutex_unlock(&dp->mutex);
462 int dp_detach_port(struct dp_port *p, int may_delete)
464 struct vport *vport = p->vport;
469 if (p->port_no != ODPP_LOCAL)
471 dp_ifinfo_notify(RTM_DELLINK, p);
473 /* First drop references to device. */
475 list_del_rcu(&p->node);
476 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
478 err = vport_detach(vport);
482 /* Then wait until no one is still using it, and destroy it. */
486 const char *port_type = vport_get_type(vport);
488 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
495 kobject_put(&p->kobj);
500 static int detach_port(int dp_idx, int port_no)
507 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
511 dp = get_dp_locked(dp_idx);
514 goto out_unlock_rtnl;
516 p = dp->ports[port_no];
521 err = dp_detach_port(p, 1);
524 mutex_unlock(&dp->mutex);
531 static void suppress_loop(struct datapath *dp, struct sw_flow_actions *actions)
534 printk(KERN_WARNING "%s: flow looped %d times, dropping\n",
535 dp_name(dp), DP_MAX_LOOPS);
536 actions->n_actions = 0;
539 /* Must be called with rcu_read_lock. */
540 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
542 struct datapath *dp = p->dp;
543 struct dp_stats_percpu *stats;
544 int stats_counter_off;
545 struct odp_flow_key key;
546 struct tbl_node *flow_node;
547 struct sw_flow *flow;
548 struct sw_flow_actions *acts;
549 struct loop_counter *loop;
551 OVS_CB(skb)->dp_port = p;
553 /* Extract flow from 'skb' into 'key'. */
554 if (flow_extract(skb, p ? p->port_no : ODPP_NONE, &key)) {
555 if (dp->drop_frags) {
557 stats_counter_off = offsetof(struct dp_stats_percpu, n_frags);
563 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
564 if (unlikely(!flow_node)) {
565 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
566 stats_counter_off = offsetof(struct dp_stats_percpu, n_missed);
570 flow = flow_cast(flow_node);
571 flow_used(flow, skb);
573 acts = rcu_dereference(flow->sf_acts);
575 /* Check whether we've looped too much. */
576 loop = &get_cpu_var(dp_loop_counters).counters[!!in_interrupt()];
577 if (unlikely(++loop->count > DP_MAX_LOOPS))
578 loop->looping = true;
579 if (unlikely(loop->looping)) {
580 suppress_loop(dp, acts);
584 /* Execute actions. */
585 execute_actions(dp, skb, &key, acts->actions, acts->n_actions, GFP_ATOMIC);
586 stats_counter_off = offsetof(struct dp_stats_percpu, n_hit);
588 /* Check whether sub-actions looped too much. */
589 if (unlikely(loop->looping))
590 suppress_loop(dp, acts);
593 /* Decrement loop counter. */
595 loop->looping = false;
596 put_cpu_var(dp_loop_counters);
599 /* Update datapath statistics. */
601 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
603 write_seqcount_begin(&stats->seqlock);
604 (*(u64 *)((u8 *)stats + stats_counter_off))++;
605 write_seqcount_end(&stats->seqlock);
610 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
611 /* This code is based on skb_checksum_setup() from Xen's net/dev/core.c. We
612 * can't call this function directly because it isn't exported in all
614 int vswitch_skb_checksum_setup(struct sk_buff *skb)
619 __u16 csum_start, csum_offset;
621 if (!skb->proto_csum_blank)
624 if (skb->protocol != htons(ETH_P_IP))
627 if (!pskb_may_pull(skb, skb_network_header(skb) + sizeof(struct iphdr) - skb->data))
631 th = skb_network_header(skb) + 4 * iph->ihl;
633 csum_start = th - skb->head;
634 switch (iph->protocol) {
636 csum_offset = offsetof(struct tcphdr, check);
639 csum_offset = offsetof(struct udphdr, check);
643 printk(KERN_ERR "Attempting to checksum a non-"
644 "TCP/UDP packet, dropping a protocol"
645 " %d packet", iph->protocol);
649 if (!pskb_may_pull(skb, th + csum_offset + 2 - skb->data))
652 skb->ip_summed = CHECKSUM_PARTIAL;
653 skb->proto_csum_blank = 0;
655 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
656 skb->csum_start = csum_start;
657 skb->csum_offset = csum_offset;
659 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
660 skb->csum = csum_offset;
668 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
670 /* Types of checksums that we can receive (these all refer to L4 checksums):
671 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
672 * (though not verified) checksum in packet but not in skb->csum. Packets
673 * from the bridge local port will also have this type.
674 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
675 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
676 * a valid skb->csum. Importantly, both contain a full checksum (not
677 * verified) in the packet itself. The only difference is that if the
678 * packet gets to L4 processing on this machine (not in DomU) we won't
679 * have to recompute the checksum to verify. Most hardware devices do not
680 * produce packets with this type, even if they support receive checksum
681 * offloading (they produce type #5).
682 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
683 * be computed if it is sent off box. Unfortunately on earlier kernels,
684 * this case is impossible to distinguish from #2, despite having opposite
685 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
686 * to distinguish the different states.
687 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
688 * generated locally by a Xen DomU and has a partial checksum. If it is
689 * handled on this machine (Dom0 or DomU), then the checksum will not be
690 * computed. If it goes off box, the checksum in the packet needs to be
691 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
692 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
693 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
694 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
695 * full checksum or using a protocol without a checksum. skb->csum is
696 * undefined. This is common from devices with receive checksum
697 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
698 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
700 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
701 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
702 * based on whether it is on the transmit or receive path. After the datapath
703 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
704 * checksum, we will panic. Since we can receive packets with checksums, we
705 * assume that all CHECKSUM_HW packets have checksums and map them to
706 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
707 * packet is processed by the local IP stack, in which case it will need to
708 * be reverified). If we receive a packet with CHECKSUM_HW that really means
709 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
710 * shouldn't be any devices that do this with bridging. */
711 void compute_ip_summed(struct sk_buff *skb, bool xmit)
713 /* For our convenience these defines change repeatedly between kernel
714 * versions, so we can't just copy them over... */
715 switch (skb->ip_summed) {
717 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
719 case CHECKSUM_UNNECESSARY:
720 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
723 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
724 * However, on the receive side we should only get CHECKSUM_PARTIAL
725 * packets from Xen, which uses some special fields to represent this
726 * (see below). Since we can only make one type work, pick the one
727 * that actually happens in practice.
729 * On the transmit side (basically after skb_checksum_setup()
730 * has been run or on internal dev transmit), packets with
731 * CHECKSUM_COMPLETE aren't generated, so assume CHECKSUM_PARTIAL. */
734 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
736 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
740 case CHECKSUM_COMPLETE:
741 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
743 case CHECKSUM_PARTIAL:
744 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
748 printk(KERN_ERR "openvswitch: unknown checksum type %d\n",
750 /* None seems the safest... */
751 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
754 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
755 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
756 * kernels. It should not be set on the transmit path though. */
757 if (skb->proto_csum_blank)
758 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
760 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
764 /* This function closely resembles skb_forward_csum() used by the bridge. It
765 * is slightly different because we are only concerned with bridging and not
766 * other types of forwarding and can get away with slightly more optimal
768 void forward_ip_summed(struct sk_buff *skb)
771 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
772 skb->ip_summed = CHECKSUM_NONE;
776 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
777 * unless we broke up a GSO packet. */
778 static int queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
779 int queue_no, u32 arg)
781 struct sk_buff *nskb;
785 if (OVS_CB(skb)->dp_port)
786 port_no = OVS_CB(skb)->dp_port->port_no;
788 port_no = ODPP_LOCAL;
791 struct odp_msg *header;
796 err = skb_cow(skb, sizeof *header);
800 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
801 header->type = queue_no;
802 header->length = skb->len;
803 header->port = port_no;
804 header->reserved = 0;
806 skb_queue_tail(queue, skb);
814 while ((skb = nskb) != NULL) {
821 int dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
824 struct dp_stats_percpu *stats;
825 struct sk_buff_head *queue;
828 WARN_ON_ONCE(skb_shared(skb));
829 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
830 queue = &dp->queues[queue_no];
832 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
835 forward_ip_summed(skb);
837 err = vswitch_skb_checksum_setup(skb);
841 /* Break apart GSO packets into their component pieces. Otherwise
842 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
843 if (skb_is_gso(skb)) {
844 struct sk_buff *nskb = skb_gso_segment(skb, NETIF_F_SG | NETIF_F_HW_CSUM);
848 if (unlikely(IS_ERR(skb))) {
853 /* XXX This case might not be possible. It's hard to
854 * tell from the skb_gso_segment() code and comment. */
858 err = queue_control_packets(skb, queue, queue_no, arg);
859 wake_up_interruptible(&dp->waitqueue);
866 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
868 write_seqcount_begin(&stats->seqlock);
870 write_seqcount_end(&stats->seqlock);
877 static int flush_flows(struct datapath *dp)
879 struct tbl *old_table = rcu_dereference(dp->table);
880 struct tbl *new_table;
882 new_table = tbl_create(0);
886 rcu_assign_pointer(dp->table, new_table);
888 tbl_deferred_destroy(old_table, flow_free_tbl);
893 static int validate_actions(const struct sw_flow_actions *actions)
897 for (i = 0; i < actions->n_actions; i++) {
898 const union odp_action *a = &actions->actions[i];
901 if (a->output.port >= DP_MAX_PORTS)
905 case ODPAT_OUTPUT_GROUP:
906 if (a->output_group.group >= DP_MAX_GROUPS)
910 case ODPAT_SET_VLAN_VID:
911 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
915 case ODPAT_SET_VLAN_PCP:
916 if (a->vlan_pcp.vlan_pcp
917 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
921 case ODPAT_SET_NW_TOS:
922 if (a->nw_tos.nw_tos & INET_ECN_MASK)
927 if (a->type >= ODPAT_N_ACTIONS)
936 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
938 struct sw_flow_actions *actions;
941 actions = flow_actions_alloc(flow->n_actions);
942 error = PTR_ERR(actions);
947 if (copy_from_user(actions->actions, flow->actions,
948 flow->n_actions * sizeof(union odp_action)))
949 goto error_free_actions;
950 error = validate_actions(actions);
952 goto error_free_actions;
959 return ERR_PTR(error);
962 static struct timespec get_time_offset(void)
964 struct timespec now_mono, now_jiffies;
966 ktime_get_ts(&now_mono);
967 jiffies_to_timespec(jiffies, &now_jiffies);
968 return timespec_sub(now_mono, now_jiffies);
971 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats,
972 struct timespec time_offset)
975 struct timespec flow_ts, used;
977 jiffies_to_timespec(flow->used, &flow_ts);
978 set_normalized_timespec(&used, flow_ts.tv_sec + time_offset.tv_sec,
979 flow_ts.tv_nsec + time_offset.tv_nsec);
981 stats->used_sec = used.tv_sec;
982 stats->used_nsec = used.tv_nsec;
985 stats->used_nsec = 0;
988 stats->n_packets = flow->packet_count;
989 stats->n_bytes = flow->byte_count;
991 stats->tcp_flags = flow->tcp_flags;
995 static void clear_stats(struct sw_flow *flow)
999 flow->packet_count = 0;
1000 flow->byte_count = 0;
1003 static int expand_table(struct datapath *dp)
1005 struct tbl *old_table = rcu_dereference(dp->table);
1006 struct tbl *new_table;
1008 new_table = tbl_expand(old_table);
1009 if (IS_ERR(new_table))
1010 return PTR_ERR(new_table);
1012 rcu_assign_pointer(dp->table, new_table);
1013 tbl_deferred_destroy(old_table, NULL);
1018 static int do_put_flow(struct datapath *dp, struct odp_flow_put *uf,
1019 struct odp_flow_stats *stats)
1021 struct tbl_node *flow_node;
1022 struct sw_flow *flow;
1026 memset(uf->flow.key.reserved, 0, sizeof uf->flow.key.reserved);
1028 table = rcu_dereference(dp->table);
1029 flow_node = tbl_lookup(table, &uf->flow.key, flow_hash(&uf->flow.key), flow_cmp);
1032 struct sw_flow_actions *acts;
1035 if (!(uf->flags & ODPPF_CREATE))
1038 /* Expand table, if necessary, to make room. */
1039 if (tbl_count(table) >= tbl_n_buckets(table)) {
1040 error = expand_table(dp);
1043 table = rcu_dereference(dp->table);
1046 /* Allocate flow. */
1048 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
1051 flow->key = uf->flow.key;
1052 spin_lock_init(&flow->lock);
1055 /* Obtain actions. */
1056 acts = get_actions(&uf->flow);
1057 error = PTR_ERR(acts);
1059 goto error_free_flow;
1060 rcu_assign_pointer(flow->sf_acts, acts);
1062 /* Put flow in bucket. */
1063 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1065 goto error_free_flow_acts;
1067 memset(stats, 0, sizeof(struct odp_flow_stats));
1069 /* We found a matching flow. */
1070 struct sw_flow_actions *old_acts, *new_acts;
1072 flow = flow_cast(flow_node);
1074 /* Bail out if we're not allowed to modify an existing flow. */
1076 if (!(uf->flags & ODPPF_MODIFY))
1080 new_acts = get_actions(&uf->flow);
1081 error = PTR_ERR(new_acts);
1082 if (IS_ERR(new_acts))
1084 old_acts = rcu_dereference(flow->sf_acts);
1085 if (old_acts->n_actions != new_acts->n_actions ||
1086 memcmp(old_acts->actions, new_acts->actions,
1087 sizeof(union odp_action) * old_acts->n_actions)) {
1088 rcu_assign_pointer(flow->sf_acts, new_acts);
1089 flow_deferred_free_acts(old_acts);
1094 /* Fetch stats, then clear them if necessary. */
1095 spin_lock_bh(&flow->lock);
1096 get_stats(flow, stats, get_time_offset());
1097 if (uf->flags & ODPPF_ZERO_STATS)
1099 spin_unlock_bh(&flow->lock);
1104 error_free_flow_acts:
1105 kfree(flow->sf_acts);
1107 kmem_cache_free(flow_cache, flow);
1112 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
1114 struct odp_flow_stats stats;
1115 struct odp_flow_put uf;
1118 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
1121 error = do_put_flow(dp, &uf, &stats);
1125 if (copy_to_user(&ufp->flow.stats, &stats,
1126 sizeof(struct odp_flow_stats)))
1132 static int do_answer_query(struct sw_flow *flow, u32 query_flags,
1133 struct timespec time_offset,
1134 struct odp_flow_stats __user *ustats,
1135 union odp_action __user *actions,
1136 u32 __user *n_actionsp)
1138 struct sw_flow_actions *sf_acts;
1139 struct odp_flow_stats stats;
1142 spin_lock_bh(&flow->lock);
1143 get_stats(flow, &stats, time_offset);
1144 if (query_flags & ODPFF_ZERO_TCP_FLAGS)
1145 flow->tcp_flags = 0;
1147 spin_unlock_bh(&flow->lock);
1149 if (copy_to_user(ustats, &stats, sizeof(struct odp_flow_stats)) ||
1150 get_user(n_actions, n_actionsp))
1156 sf_acts = rcu_dereference(flow->sf_acts);
1157 if (put_user(sf_acts->n_actions, n_actionsp) ||
1158 (actions && copy_to_user(actions, sf_acts->actions,
1159 sizeof(union odp_action) *
1160 min(sf_acts->n_actions, n_actions))))
1166 static int answer_query(struct sw_flow *flow, u32 query_flags,
1167 struct timespec time_offset,
1168 struct odp_flow __user *ufp)
1170 union odp_action *actions;
1172 if (get_user(actions, &ufp->actions))
1175 return do_answer_query(flow, query_flags, time_offset,
1176 &ufp->stats, actions, &ufp->n_actions);
1179 static struct sw_flow *do_del_flow(struct datapath *dp, struct odp_flow_key *key)
1181 struct tbl *table = rcu_dereference(dp->table);
1182 struct tbl_node *flow_node;
1185 memset(key->reserved, 0, sizeof key->reserved);
1186 flow_node = tbl_lookup(table, key, flow_hash(key), flow_cmp);
1188 return ERR_PTR(-ENOENT);
1190 error = tbl_remove(table, flow_node);
1192 return ERR_PTR(error);
1194 /* XXX Returned flow_node's statistics might lose a few packets, since
1195 * other CPUs can be using this flow. We used to synchronize_rcu() to
1196 * make sure that we get completely accurate stats, but that blows our
1197 * performance, badly. */
1198 return flow_cast(flow_node);
1201 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1203 struct sw_flow *flow;
1207 if (copy_from_user(&uf, ufp, sizeof uf))
1210 flow = do_del_flow(dp, &uf.key);
1212 return PTR_ERR(flow);
1214 error = answer_query(flow, 0, get_time_offset(), ufp);
1215 flow_deferred_free(flow);
1219 static int do_query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1221 struct tbl *table = rcu_dereference(dp->table);
1222 struct timespec time_offset;
1225 time_offset = get_time_offset();
1227 for (i = 0; i < flowvec->n_flows; i++) {
1228 struct odp_flow __user *ufp = &flowvec->flows[i];
1230 struct tbl_node *flow_node;
1233 if (copy_from_user(&uf, ufp, sizeof uf))
1235 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1237 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1239 error = put_user(ENOENT, &ufp->stats.error);
1241 error = answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1245 return flowvec->n_flows;
1248 struct list_flows_cbdata {
1249 struct odp_flow __user *uflows;
1252 struct timespec time_offset;
1255 static int list_flow(struct tbl_node *node, void *cbdata_)
1257 struct sw_flow *flow = flow_cast(node);
1258 struct list_flows_cbdata *cbdata = cbdata_;
1259 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1262 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1264 error = answer_query(flow, 0, cbdata->time_offset, ufp);
1268 if (cbdata->listed_flows >= cbdata->n_flows)
1269 return cbdata->listed_flows;
1273 static int do_list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1275 struct list_flows_cbdata cbdata;
1278 if (!flowvec->n_flows)
1281 cbdata.uflows = flowvec->flows;
1282 cbdata.n_flows = flowvec->n_flows;
1283 cbdata.listed_flows = 0;
1284 cbdata.time_offset = get_time_offset();
1286 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1287 return error ? error : cbdata.listed_flows;
1290 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1291 int (*function)(struct datapath *,
1292 const struct odp_flowvec *))
1294 struct odp_flowvec __user *uflowvec;
1295 struct odp_flowvec flowvec;
1298 uflowvec = (struct odp_flowvec __user *)argp;
1299 if (copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1302 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1305 retval = function(dp, &flowvec);
1306 return (retval < 0 ? retval
1307 : retval == flowvec.n_flows ? 0
1308 : put_user(retval, &uflowvec->n_flows));
1311 static int do_execute(struct datapath *dp, const struct odp_execute *execute)
1313 struct odp_flow_key key;
1314 struct sk_buff *skb;
1315 struct sw_flow_actions *actions;
1320 if (execute->length < ETH_HLEN || execute->length > 65535)
1324 actions = flow_actions_alloc(execute->n_actions);
1329 if (copy_from_user(actions->actions, execute->actions,
1330 execute->n_actions * sizeof *execute->actions))
1331 goto error_free_actions;
1333 err = validate_actions(actions);
1335 goto error_free_actions;
1338 skb = alloc_skb(execute->length, GFP_KERNEL);
1340 goto error_free_actions;
1342 if (execute->in_port < DP_MAX_PORTS)
1343 OVS_CB(skb)->dp_port = dp->ports[execute->in_port];
1345 OVS_CB(skb)->dp_port = NULL;
1348 if (copy_from_user(skb_put(skb, execute->length), execute->data,
1350 goto error_free_skb;
1352 skb_reset_mac_header(skb);
1355 /* Normally, setting the skb 'protocol' field would be handled by a
1356 * call to eth_type_trans(), but it assumes there's a sending
1357 * device, which we may not have. */
1358 if (ntohs(eth->h_proto) >= 1536)
1359 skb->protocol = eth->h_proto;
1361 skb->protocol = htons(ETH_P_802_2);
1363 flow_extract(skb, execute->in_port, &key);
1366 err = execute_actions(dp, skb, &key, actions->actions,
1367 actions->n_actions, GFP_KERNEL);
1381 static int execute_packet(struct datapath *dp, const struct odp_execute __user *executep)
1383 struct odp_execute execute;
1385 if (copy_from_user(&execute, executep, sizeof execute))
1388 return do_execute(dp, &execute);
1391 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1393 struct tbl *table = rcu_dereference(dp->table);
1394 struct odp_stats stats;
1397 stats.n_flows = tbl_count(table);
1398 stats.cur_capacity = tbl_n_buckets(table);
1399 stats.max_capacity = TBL_MAX_BUCKETS;
1400 stats.n_ports = dp->n_ports;
1401 stats.max_ports = DP_MAX_PORTS;
1402 stats.max_groups = DP_MAX_GROUPS;
1403 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1404 for_each_possible_cpu(i) {
1405 const struct dp_stats_percpu *percpu_stats;
1406 struct dp_stats_percpu local_stats;
1409 percpu_stats = per_cpu_ptr(dp->stats_percpu, i);
1412 seqcount = read_seqcount_begin(&percpu_stats->seqlock);
1413 local_stats = *percpu_stats;
1414 } while (read_seqcount_retry(&percpu_stats->seqlock, seqcount));
1416 stats.n_frags += local_stats.n_frags;
1417 stats.n_hit += local_stats.n_hit;
1418 stats.n_missed += local_stats.n_missed;
1419 stats.n_lost += local_stats.n_lost;
1421 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1422 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1423 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1426 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1427 int dp_min_mtu(const struct datapath *dp)
1434 list_for_each_entry_rcu (p, &dp->port_list, node) {
1437 /* Skip any internal ports, since that's what we're trying to
1439 if (is_internal_vport(p->vport))
1442 dev_mtu = vport_get_mtu(p->vport);
1443 if (!mtu || dev_mtu < mtu)
1447 return mtu ? mtu : ETH_DATA_LEN;
1450 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1451 * be called with RTNL lock. */
1452 void set_internal_devs_mtu(const struct datapath *dp)
1459 mtu = dp_min_mtu(dp);
1461 list_for_each_entry_rcu (p, &dp->port_list, node) {
1462 if (is_internal_vport(p->vport))
1463 vport_set_mtu(p->vport, mtu);
1467 static int put_port(const struct dp_port *p, struct odp_port __user *uop)
1471 memset(&op, 0, sizeof op);
1474 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1477 op.port = p->port_no;
1478 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1480 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1483 static int query_port(struct datapath *dp, struct odp_port __user *uport)
1485 struct odp_port port;
1487 if (copy_from_user(&port, uport, sizeof port))
1490 if (port.devname[0]) {
1491 struct vport *vport;
1492 struct dp_port *dp_port;
1495 port.devname[IFNAMSIZ - 1] = '\0';
1500 vport = vport_locate(port.devname);
1506 dp_port = vport_get_dp_port(vport);
1507 if (!dp_port || dp_port->dp != dp) {
1512 port.port = dp_port->port_no;
1521 if (port.port >= DP_MAX_PORTS)
1523 if (!dp->ports[port.port])
1527 return put_port(dp->ports[port.port], uport);
1530 static int do_list_ports(struct datapath *dp, struct odp_port __user *uports,
1537 list_for_each_entry_rcu (p, &dp->port_list, node) {
1538 if (put_port(p, &uports[idx]))
1540 if (idx++ >= n_ports)
1547 static int list_ports(struct datapath *dp, struct odp_portvec __user *upv)
1549 struct odp_portvec pv;
1552 if (copy_from_user(&pv, upv, sizeof pv))
1555 retval = do_list_ports(dp, pv.ports, pv.n_ports);
1559 return put_user(retval, &upv->n_ports);
1562 /* RCU callback for freeing a dp_port_group */
1563 static void free_port_group(struct rcu_head *rcu)
1565 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1569 static int do_set_port_group(struct datapath *dp, u16 __user *ports,
1570 int n_ports, int group)
1572 struct dp_port_group *new_group, *old_group;
1576 if (n_ports > DP_MAX_PORTS || group >= DP_MAX_GROUPS)
1580 new_group = kmalloc(sizeof *new_group + sizeof(u16) * n_ports, GFP_KERNEL);
1584 new_group->n_ports = n_ports;
1586 if (copy_from_user(new_group->ports, ports, sizeof(u16) * n_ports))
1589 old_group = rcu_dereference(dp->groups[group]);
1590 rcu_assign_pointer(dp->groups[group], new_group);
1592 call_rcu(&old_group->rcu, free_port_group);
1601 static int set_port_group(struct datapath *dp,
1602 const struct odp_port_group __user *upg)
1604 struct odp_port_group pg;
1606 if (copy_from_user(&pg, upg, sizeof pg))
1609 return do_set_port_group(dp, pg.ports, pg.n_ports, pg.group);
1612 static int do_get_port_group(struct datapath *dp,
1613 u16 __user *ports, int n_ports, int group,
1614 u16 __user *n_portsp)
1616 struct dp_port_group *g;
1619 if (group >= DP_MAX_GROUPS)
1622 g = dp->groups[group];
1623 n_copy = g ? min_t(int, g->n_ports, n_ports) : 0;
1624 if (n_copy && copy_to_user(ports, g->ports, n_copy * sizeof(u16)))
1627 if (put_user(g ? g->n_ports : 0, n_portsp))
1633 static int get_port_group(struct datapath *dp, struct odp_port_group __user *upg)
1635 struct odp_port_group pg;
1637 if (copy_from_user(&pg, upg, sizeof pg))
1640 return do_get_port_group(dp, pg.ports, pg.n_ports, pg.group, &upg->n_ports);
1643 static int get_listen_mask(const struct file *f)
1645 return (long)f->private_data;
1648 static void set_listen_mask(struct file *f, int listen_mask)
1650 f->private_data = (void*)(long)listen_mask;
1653 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1656 int dp_idx = iminor(f->f_dentry->d_inode);
1657 struct datapath *dp;
1658 int drop_frags, listeners, port_no;
1659 unsigned int sflow_probability;
1662 /* Handle commands with special locking requirements up front. */
1665 err = create_dp(dp_idx, (char __user *)argp);
1668 case ODP_DP_DESTROY:
1669 err = destroy_dp(dp_idx);
1672 case ODP_PORT_ATTACH:
1673 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1676 case ODP_PORT_DETACH:
1677 err = get_user(port_no, (int __user *)argp);
1679 err = detach_port(dp_idx, port_no);
1683 err = vport_user_add((struct odp_vport_add __user *)argp);
1687 err = vport_user_mod((struct odp_vport_mod __user *)argp);
1691 err = vport_user_del((char __user *)argp);
1694 case ODP_VPORT_STATS_GET:
1695 err = vport_user_stats_get((struct odp_vport_stats_req __user *)argp);
1698 case ODP_VPORT_STATS_SET:
1699 err = vport_user_stats_set((struct odp_vport_stats_req __user *)argp);
1702 case ODP_VPORT_ETHER_GET:
1703 err = vport_user_ether_get((struct odp_vport_ether __user *)argp);
1706 case ODP_VPORT_ETHER_SET:
1707 err = vport_user_ether_set((struct odp_vport_ether __user *)argp);
1710 case ODP_VPORT_MTU_GET:
1711 err = vport_user_mtu_get((struct odp_vport_mtu __user *)argp);
1714 case ODP_VPORT_MTU_SET:
1715 err = vport_user_mtu_set((struct odp_vport_mtu __user *)argp);
1719 dp = get_dp_locked(dp_idx);
1726 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1729 case ODP_GET_DROP_FRAGS:
1730 err = put_user(dp->drop_frags, (int __user *)argp);
1733 case ODP_SET_DROP_FRAGS:
1734 err = get_user(drop_frags, (int __user *)argp);
1738 if (drop_frags != 0 && drop_frags != 1)
1740 dp->drop_frags = drop_frags;
1744 case ODP_GET_LISTEN_MASK:
1745 err = put_user(get_listen_mask(f), (int __user *)argp);
1748 case ODP_SET_LISTEN_MASK:
1749 err = get_user(listeners, (int __user *)argp);
1753 if (listeners & ~ODPL_ALL)
1756 set_listen_mask(f, listeners);
1759 case ODP_GET_SFLOW_PROBABILITY:
1760 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1763 case ODP_SET_SFLOW_PROBABILITY:
1764 err = get_user(sflow_probability, (unsigned int __user *)argp);
1766 dp->sflow_probability = sflow_probability;
1769 case ODP_PORT_QUERY:
1770 err = query_port(dp, (struct odp_port __user *)argp);
1774 err = list_ports(dp, (struct odp_portvec __user *)argp);
1777 case ODP_PORT_GROUP_SET:
1778 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1781 case ODP_PORT_GROUP_GET:
1782 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1785 case ODP_FLOW_FLUSH:
1786 err = flush_flows(dp);
1790 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1794 err = del_flow(dp, (struct odp_flow __user *)argp);
1798 err = do_flowvec_ioctl(dp, argp, do_query_flows);
1802 err = do_flowvec_ioctl(dp, argp, do_list_flows);
1806 err = execute_packet(dp, (struct odp_execute __user *)argp);
1813 mutex_unlock(&dp->mutex);
1818 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1821 for (i = 0; i < DP_N_QUEUES; i++) {
1822 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1828 #ifdef CONFIG_COMPAT
1829 static int compat_list_ports(struct datapath *dp, struct compat_odp_portvec __user *upv)
1831 struct compat_odp_portvec pv;
1834 if (copy_from_user(&pv, upv, sizeof pv))
1837 retval = do_list_ports(dp, compat_ptr(pv.ports), pv.n_ports);
1841 return put_user(retval, &upv->n_ports);
1844 static int compat_set_port_group(struct datapath *dp, const struct compat_odp_port_group __user *upg)
1846 struct compat_odp_port_group pg;
1848 if (copy_from_user(&pg, upg, sizeof pg))
1851 return do_set_port_group(dp, compat_ptr(pg.ports), pg.n_ports, pg.group);
1854 static int compat_get_port_group(struct datapath *dp, struct compat_odp_port_group __user *upg)
1856 struct compat_odp_port_group pg;
1858 if (copy_from_user(&pg, upg, sizeof pg))
1861 return do_get_port_group(dp, compat_ptr(pg.ports), pg.n_ports,
1862 pg.group, &upg->n_ports);
1865 static int compat_get_flow(struct odp_flow *flow, const struct compat_odp_flow __user *compat)
1867 compat_uptr_t actions;
1869 if (!access_ok(VERIFY_READ, compat, sizeof(struct compat_odp_flow)) ||
1870 __copy_from_user(&flow->stats, &compat->stats, sizeof(struct odp_flow_stats)) ||
1871 __copy_from_user(&flow->key, &compat->key, sizeof(struct odp_flow_key)) ||
1872 __get_user(actions, &compat->actions) ||
1873 __get_user(flow->n_actions, &compat->n_actions) ||
1874 __get_user(flow->flags, &compat->flags))
1877 flow->actions = compat_ptr(actions);
1881 static int compat_put_flow(struct datapath *dp, struct compat_odp_flow_put __user *ufp)
1883 struct odp_flow_stats stats;
1884 struct odp_flow_put fp;
1887 if (compat_get_flow(&fp.flow, &ufp->flow) ||
1888 get_user(fp.flags, &ufp->flags))
1891 error = do_put_flow(dp, &fp, &stats);
1895 if (copy_to_user(&ufp->flow.stats, &stats,
1896 sizeof(struct odp_flow_stats)))
1902 static int compat_answer_query(struct sw_flow *flow, u32 query_flags,
1903 struct timespec time_offset,
1904 struct compat_odp_flow __user *ufp)
1906 compat_uptr_t actions;
1908 if (get_user(actions, &ufp->actions))
1911 return do_answer_query(flow, query_flags, time_offset, &ufp->stats,
1912 compat_ptr(actions), &ufp->n_actions);
1915 static int compat_del_flow(struct datapath *dp, struct compat_odp_flow __user *ufp)
1917 struct sw_flow *flow;
1921 if (compat_get_flow(&uf, ufp))
1924 flow = do_del_flow(dp, &uf.key);
1926 return PTR_ERR(flow);
1928 error = compat_answer_query(flow, 0, get_time_offset(), ufp);
1929 flow_deferred_free(flow);
1933 static int compat_query_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1935 struct tbl *table = rcu_dereference(dp->table);
1936 struct timespec time_offset;
1939 time_offset = get_time_offset();
1941 for (i = 0; i < n_flows; i++) {
1942 struct compat_odp_flow __user *ufp = &flows[i];
1944 struct tbl_node *flow_node;
1947 if (compat_get_flow(&uf, ufp))
1949 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1951 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1953 error = put_user(ENOENT, &ufp->stats.error);
1955 error = compat_answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1962 struct compat_list_flows_cbdata {
1963 struct compat_odp_flow __user *uflows;
1966 struct timespec time_offset;
1969 static int compat_list_flow(struct tbl_node *node, void *cbdata_)
1971 struct sw_flow *flow = flow_cast(node);
1972 struct compat_list_flows_cbdata *cbdata = cbdata_;
1973 struct compat_odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1976 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1978 error = compat_answer_query(flow, 0, cbdata->time_offset, ufp);
1982 if (cbdata->listed_flows >= cbdata->n_flows)
1983 return cbdata->listed_flows;
1987 static int compat_list_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1989 struct compat_list_flows_cbdata cbdata;
1995 cbdata.uflows = flows;
1996 cbdata.n_flows = n_flows;
1997 cbdata.listed_flows = 0;
1998 cbdata.time_offset = get_time_offset();
2000 error = tbl_foreach(rcu_dereference(dp->table), compat_list_flow, &cbdata);
2001 return error ? error : cbdata.listed_flows;
2004 static int compat_flowvec_ioctl(struct datapath *dp, unsigned long argp,
2005 int (*function)(struct datapath *,
2006 struct compat_odp_flow *,
2009 struct compat_odp_flowvec __user *uflowvec;
2010 struct compat_odp_flow __user *flows;
2011 struct compat_odp_flowvec flowvec;
2014 uflowvec = compat_ptr(argp);
2015 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
2016 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
2019 if (flowvec.n_flows > INT_MAX / sizeof(struct compat_odp_flow))
2022 flows = compat_ptr(flowvec.flows);
2023 if (!access_ok(VERIFY_WRITE, flows,
2024 flowvec.n_flows * sizeof(struct compat_odp_flow)))
2027 retval = function(dp, flows, flowvec.n_flows);
2028 return (retval < 0 ? retval
2029 : retval == flowvec.n_flows ? 0
2030 : put_user(retval, &uflowvec->n_flows));
2033 static int compat_execute(struct datapath *dp, const struct compat_odp_execute __user *uexecute)
2035 struct odp_execute execute;
2036 compat_uptr_t actions;
2039 if (!access_ok(VERIFY_READ, uexecute, sizeof(struct compat_odp_execute)) ||
2040 __get_user(execute.in_port, &uexecute->in_port) ||
2041 __get_user(actions, &uexecute->actions) ||
2042 __get_user(execute.n_actions, &uexecute->n_actions) ||
2043 __get_user(data, &uexecute->data) ||
2044 __get_user(execute.length, &uexecute->length))
2047 execute.actions = compat_ptr(actions);
2048 execute.data = compat_ptr(data);
2050 return do_execute(dp, &execute);
2053 static long openvswitch_compat_ioctl(struct file *f, unsigned int cmd, unsigned long argp)
2055 int dp_idx = iminor(f->f_dentry->d_inode);
2056 struct datapath *dp;
2060 case ODP_DP_DESTROY:
2061 case ODP_FLOW_FLUSH:
2062 /* Ioctls that don't need any translation at all. */
2063 return openvswitch_ioctl(f, cmd, argp);
2066 case ODP_PORT_ATTACH:
2067 case ODP_PORT_DETACH:
2069 case ODP_VPORT_MTU_SET:
2070 case ODP_VPORT_MTU_GET:
2071 case ODP_VPORT_ETHER_SET:
2072 case ODP_VPORT_ETHER_GET:
2073 case ODP_VPORT_STATS_SET:
2074 case ODP_VPORT_STATS_GET:
2076 case ODP_GET_DROP_FRAGS:
2077 case ODP_SET_DROP_FRAGS:
2078 case ODP_SET_LISTEN_MASK:
2079 case ODP_GET_LISTEN_MASK:
2080 case ODP_SET_SFLOW_PROBABILITY:
2081 case ODP_GET_SFLOW_PROBABILITY:
2082 case ODP_PORT_QUERY:
2083 /* Ioctls that just need their pointer argument extended. */
2084 return openvswitch_ioctl(f, cmd, (unsigned long)compat_ptr(argp));
2086 case ODP_VPORT_ADD32:
2087 return compat_vport_user_add(compat_ptr(argp));
2089 case ODP_VPORT_MOD32:
2090 return compat_vport_user_mod(compat_ptr(argp));
2093 dp = get_dp_locked(dp_idx);
2099 case ODP_PORT_LIST32:
2100 err = compat_list_ports(dp, compat_ptr(argp));
2103 case ODP_PORT_GROUP_SET32:
2104 err = compat_set_port_group(dp, compat_ptr(argp));
2107 case ODP_PORT_GROUP_GET32:
2108 err = compat_get_port_group(dp, compat_ptr(argp));
2111 case ODP_FLOW_PUT32:
2112 err = compat_put_flow(dp, compat_ptr(argp));
2115 case ODP_FLOW_DEL32:
2116 err = compat_del_flow(dp, compat_ptr(argp));
2119 case ODP_FLOW_GET32:
2120 err = compat_flowvec_ioctl(dp, argp, compat_query_flows);
2123 case ODP_FLOW_LIST32:
2124 err = compat_flowvec_ioctl(dp, argp, compat_list_flows);
2128 err = compat_execute(dp, compat_ptr(argp));
2135 mutex_unlock(&dp->mutex);
2141 /* Unfortunately this function is not exported so this is a verbatim copy
2142 * from net/core/datagram.c in 2.6.30. */
2143 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
2144 u8 __user *to, int len,
2147 int start = skb_headlen(skb);
2149 int i, copy = start - offset;
2156 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
2160 if ((len -= copy) == 0)
2167 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2170 WARN_ON(start > offset + len);
2172 end = start + skb_shinfo(skb)->frags[i].size;
2173 if ((copy = end - offset) > 0) {
2177 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2178 struct page *page = frag->page;
2183 csum2 = csum_and_copy_to_user(vaddr +
2190 *csump = csum_block_add(*csump, csum2, pos);
2200 if (skb_shinfo(skb)->frag_list) {
2201 struct sk_buff *list = skb_shinfo(skb)->frag_list;
2203 for (; list; list=list->next) {
2206 WARN_ON(start > offset + len);
2208 end = start + list->len;
2209 if ((copy = end - offset) > 0) {
2213 if (skb_copy_and_csum_datagram(list,
2218 *csump = csum_block_add(*csump, csum2, pos);
2219 if ((len -= copy) == 0)
2235 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
2238 /* XXX is there sufficient synchronization here? */
2239 int listeners = get_listen_mask(f);
2240 int dp_idx = iminor(f->f_dentry->d_inode);
2241 struct datapath *dp = get_dp(dp_idx);
2242 struct sk_buff *skb;
2243 size_t copy_bytes, tot_copy_bytes;
2249 if (nbytes == 0 || !listeners)
2255 for (i = 0; i < DP_N_QUEUES; i++) {
2256 if (listeners & (1 << i)) {
2257 skb = skb_dequeue(&dp->queues[i]);
2263 if (f->f_flags & O_NONBLOCK) {
2268 wait_event_interruptible(dp->waitqueue,
2269 dp_has_packet_of_interest(dp,
2272 if (signal_pending(current)) {
2273 retval = -ERESTARTSYS;
2278 copy_bytes = tot_copy_bytes = min_t(size_t, skb->len, nbytes);
2281 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2282 if (copy_bytes == skb->len) {
2284 unsigned int csum_start, csum_offset;
2286 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
2287 csum_start = skb->csum_start - skb_headroom(skb);
2288 csum_offset = skb->csum_offset;
2290 csum_start = skb_transport_header(skb) - skb->data;
2291 csum_offset = skb->csum;
2293 BUG_ON(csum_start >= skb_headlen(skb));
2294 retval = skb_copy_and_csum_datagram(skb, csum_start, buf + csum_start,
2295 copy_bytes - csum_start, &csum);
2297 __sum16 __user *csump;
2299 copy_bytes = csum_start;
2300 csump = (__sum16 __user *)(buf + csum_start + csum_offset);
2302 BUG_ON((char *)csump + sizeof(__sum16) > buf + nbytes);
2303 put_user(csum_fold(csum), csump);
2306 retval = skb_checksum_help(skb);
2310 struct iovec __user iov;
2313 iov.iov_len = copy_bytes;
2314 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
2318 retval = tot_copy_bytes;
2326 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
2328 /* XXX is there sufficient synchronization here? */
2329 int dp_idx = iminor(file->f_dentry->d_inode);
2330 struct datapath *dp = get_dp(dp_idx);
2335 poll_wait(file, &dp->waitqueue, wait);
2336 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
2337 mask |= POLLIN | POLLRDNORM;
2339 mask = POLLIN | POLLRDNORM | POLLHUP;
2344 struct file_operations openvswitch_fops = {
2345 /* XXX .aio_read = openvswitch_aio_read, */
2346 .read = openvswitch_read,
2347 .poll = openvswitch_poll,
2348 .unlocked_ioctl = openvswitch_ioctl,
2349 #ifdef CONFIG_COMPAT
2350 .compat_ioctl = openvswitch_compat_ioctl,
2352 /* XXX .fasync = openvswitch_fasync, */
2357 static int __init dp_init(void)
2359 struct sk_buff *dummy_skb;
2362 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
2364 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
2372 goto error_flow_exit;
2374 err = register_netdevice_notifier(&dp_device_notifier);
2376 goto error_vport_exit;
2378 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
2380 goto error_unreg_notifier;
2384 error_unreg_notifier:
2385 unregister_netdevice_notifier(&dp_device_notifier);
2394 static void dp_cleanup(void)
2397 unregister_chrdev(major, "openvswitch");
2398 unregister_netdevice_notifier(&dp_device_notifier);
2403 module_init(dp_init);
2404 module_exit(dp_cleanup);
2406 MODULE_DESCRIPTION("Open vSwitch switching datapath");
2407 MODULE_LICENSE("GPL");