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/wait.h>
33 #include <asm/system.h>
34 #include <asm/div64.h>
36 #include <linux/highmem.h>
37 #include <linux/netfilter_bridge.h>
38 #include <linux/netfilter_ipv4.h>
39 #include <linux/inetdevice.h>
40 #include <linux/list.h>
41 #include <linux/rculist.h>
42 #include <linux/workqueue.h>
43 #include <linux/dmi.h>
44 #include <net/inet_ecn.h>
45 #include <linux/compat.h>
47 #include "openvswitch/datapath-protocol.h"
51 #include "odp-compat.h"
53 #include "vport-internal_dev.h"
58 int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd);
59 EXPORT_SYMBOL(dp_ioctl_hook);
61 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
64 * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL
67 * It is safe to access the datapath and dp_port structures with just
70 static struct datapath *dps[ODP_MAX];
71 static DEFINE_MUTEX(dp_mutex);
73 /* We limit the number of times that we pass into dp_process_received_packet()
74 * to avoid blowing out the stack in the event that we have a loop. */
76 int count; /* Count. */
77 bool looping; /* Loop detected? */
80 #define DP_MAX_LOOPS 5
82 /* We use a separate counter for each CPU for both interrupt and non-interrupt
83 * context in order to keep the limit deterministic for a given packet. */
84 struct percpu_loop_counters {
85 struct loop_counter counters[2];
88 static DEFINE_PER_CPU(struct percpu_loop_counters, dp_loop_counters);
90 static int new_dp_port(struct datapath *, struct odp_port *, int port_no);
92 /* Must be called with rcu_read_lock or dp_mutex. */
93 struct datapath *get_dp(int dp_idx)
95 if (dp_idx < 0 || dp_idx >= ODP_MAX)
97 return rcu_dereference(dps[dp_idx]);
99 EXPORT_SYMBOL_GPL(get_dp);
101 static struct datapath *get_dp_locked(int dp_idx)
105 mutex_lock(&dp_mutex);
108 mutex_lock(&dp->mutex);
109 mutex_unlock(&dp_mutex);
113 /* Must be called with rcu_read_lock or RTNL lock. */
114 const char *dp_name(const struct datapath *dp)
116 return vport_get_name(dp->ports[ODPP_LOCAL]->vport);
119 static inline size_t br_nlmsg_size(void)
121 return NLMSG_ALIGN(sizeof(struct ifinfomsg))
122 + nla_total_size(IFNAMSIZ) /* IFLA_IFNAME */
123 + nla_total_size(MAX_ADDR_LEN) /* IFLA_ADDRESS */
124 + nla_total_size(4) /* IFLA_MASTER */
125 + nla_total_size(4) /* IFLA_MTU */
126 + nla_total_size(4) /* IFLA_LINK */
127 + nla_total_size(1); /* IFLA_OPERSTATE */
130 static int dp_fill_ifinfo(struct sk_buff *skb,
131 const struct dp_port *port,
132 int event, unsigned int flags)
134 const struct datapath *dp = port->dp;
135 int ifindex = vport_get_ifindex(port->vport);
136 int iflink = vport_get_iflink(port->vport);
137 struct ifinfomsg *hdr;
138 struct nlmsghdr *nlh;
146 nlh = nlmsg_put(skb, 0, 0, event, sizeof(*hdr), flags);
150 hdr = nlmsg_data(nlh);
151 hdr->ifi_family = AF_BRIDGE;
153 hdr->ifi_type = ARPHRD_ETHER;
154 hdr->ifi_index = ifindex;
155 hdr->ifi_flags = vport_get_flags(port->vport);
158 NLA_PUT_STRING(skb, IFLA_IFNAME, vport_get_name(port->vport));
159 NLA_PUT_U32(skb, IFLA_MASTER, vport_get_ifindex(dp->ports[ODPP_LOCAL]->vport));
160 NLA_PUT_U32(skb, IFLA_MTU, vport_get_mtu(port->vport));
161 #ifdef IFLA_OPERSTATE
162 NLA_PUT_U8(skb, IFLA_OPERSTATE,
163 vport_is_running(port->vport)
164 ? vport_get_operstate(port->vport)
168 NLA_PUT(skb, IFLA_ADDRESS, ETH_ALEN,
169 vport_get_addr(port->vport));
171 if (ifindex != iflink)
172 NLA_PUT_U32(skb, IFLA_LINK,iflink);
174 return nlmsg_end(skb, nlh);
177 nlmsg_cancel(skb, nlh);
181 static void dp_ifinfo_notify(int event, struct dp_port *port)
186 skb = nlmsg_new(br_nlmsg_size(), GFP_KERNEL);
190 err = dp_fill_ifinfo(skb, port, event, 0);
192 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
193 WARN_ON(err == -EMSGSIZE);
197 rtnl_notify(skb, &init_net, 0, RTNLGRP_LINK, NULL, GFP_KERNEL);
201 rtnl_set_sk_err(&init_net, RTNLGRP_LINK, err);
204 static void release_dp(struct kobject *kobj)
206 struct datapath *dp = container_of(kobj, struct datapath, ifobj);
210 static struct kobj_type dp_ktype = {
211 .release = release_dp
214 static int create_dp(int dp_idx, const char __user *devnamep)
216 struct odp_port internal_dev_port;
217 char devname[IFNAMSIZ];
223 int retval = strncpy_from_user(devname, devnamep, IFNAMSIZ);
227 } else if (retval >= IFNAMSIZ) {
232 snprintf(devname, sizeof devname, "of%d", dp_idx);
236 mutex_lock(&dp_mutex);
238 if (!try_module_get(THIS_MODULE))
241 /* Exit early if a datapath with that number already exists.
242 * (We don't use -EEXIST because that's ambiguous with 'devname'
243 * conflicting with an existing network device name.) */
249 dp = kzalloc(sizeof *dp, GFP_KERNEL);
252 INIT_LIST_HEAD(&dp->port_list);
253 mutex_init(&dp->mutex);
255 for (i = 0; i < DP_N_QUEUES; i++)
256 skb_queue_head_init(&dp->queues[i]);
257 init_waitqueue_head(&dp->waitqueue);
259 /* Initialize kobject for bridge. This will be added as
260 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
261 dp->ifobj.kset = NULL;
262 kobject_init(&dp->ifobj, &dp_ktype);
264 /* Allocate table. */
266 rcu_assign_pointer(dp->table, tbl_create(0));
270 /* Set up our datapath device. */
271 BUILD_BUG_ON(sizeof(internal_dev_port.devname) != sizeof(devname));
272 strcpy(internal_dev_port.devname, devname);
273 internal_dev_port.flags = ODP_PORT_INTERNAL;
274 err = new_dp_port(dp, &internal_dev_port, ODPP_LOCAL);
279 goto err_destroy_table;
283 dp->stats_percpu = alloc_percpu(struct dp_stats_percpu);
284 if (!dp->stats_percpu)
285 goto err_destroy_local_port;
287 rcu_assign_pointer(dps[dp_idx], dp);
288 mutex_unlock(&dp_mutex);
295 err_destroy_local_port:
296 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
298 tbl_destroy(dp->table, NULL);
302 module_put(THIS_MODULE);
304 mutex_unlock(&dp_mutex);
310 static void do_destroy_dp(struct datapath *dp)
312 struct dp_port *p, *n;
315 list_for_each_entry_safe (p, n, &dp->port_list, node)
316 if (p->port_no != ODPP_LOCAL)
317 dp_detach_port(p, 1);
321 rcu_assign_pointer(dps[dp->dp_idx], NULL);
323 dp_detach_port(dp->ports[ODPP_LOCAL], 1);
325 tbl_destroy(dp->table, flow_free_tbl);
327 for (i = 0; i < DP_N_QUEUES; i++)
328 skb_queue_purge(&dp->queues[i]);
329 for (i = 0; i < DP_MAX_GROUPS; i++)
330 kfree(dp->groups[i]);
331 free_percpu(dp->stats_percpu);
332 kobject_put(&dp->ifobj);
333 module_put(THIS_MODULE);
336 static int destroy_dp(int dp_idx)
342 mutex_lock(&dp_mutex);
352 mutex_unlock(&dp_mutex);
357 static void release_dp_port(struct kobject *kobj)
359 struct dp_port *p = container_of(kobj, struct dp_port, kobj);
363 static struct kobj_type brport_ktype = {
365 .sysfs_ops = &brport_sysfs_ops,
367 .release = release_dp_port
370 /* Called with RTNL lock and dp_mutex. */
371 static int new_dp_port(struct datapath *dp, struct odp_port *odp_port, int port_no)
377 vport = vport_locate(odp_port->devname);
381 if (odp_port->flags & ODP_PORT_INTERNAL)
382 vport = vport_add(odp_port->devname, "internal", NULL);
384 vport = vport_add(odp_port->devname, "netdev", NULL);
389 return PTR_ERR(vport);
392 p = kzalloc(sizeof(*p), GFP_KERNEL);
396 p->port_no = port_no;
399 atomic_set(&p->sflow_pool, 0);
401 err = vport_attach(vport, p);
407 rcu_assign_pointer(dp->ports[port_no], p);
408 list_add_rcu(&p->node, &dp->port_list);
411 /* Initialize kobject for bridge. This will be added as
412 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
414 kobject_init(&p->kobj, &brport_ktype);
416 dp_ifinfo_notify(RTM_NEWLINK, p);
421 static int attach_port(int dp_idx, struct odp_port __user *portp)
424 struct odp_port port;
429 if (copy_from_user(&port, portp, sizeof port))
431 port.devname[IFNAMSIZ - 1] = '\0';
434 dp = get_dp_locked(dp_idx);
437 goto out_unlock_rtnl;
439 for (port_no = 1; port_no < DP_MAX_PORTS; port_no++)
440 if (!dp->ports[port_no])
446 err = new_dp_port(dp, &port, port_no);
450 set_internal_devs_mtu(dp);
451 dp_sysfs_add_if(dp->ports[port_no]);
453 err = put_user(port_no, &portp->port);
456 mutex_unlock(&dp->mutex);
463 int dp_detach_port(struct dp_port *p, int may_delete)
465 struct vport *vport = p->vport;
470 if (p->port_no != ODPP_LOCAL)
472 dp_ifinfo_notify(RTM_DELLINK, p);
474 /* First drop references to device. */
476 list_del_rcu(&p->node);
477 rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
479 err = vport_detach(vport);
483 /* Then wait until no one is still using it, and destroy it. */
487 const char *port_type = vport_get_type(vport);
489 if (!strcmp(port_type, "netdev") || !strcmp(port_type, "internal")) {
496 kobject_put(&p->kobj);
501 static int detach_port(int dp_idx, int port_no)
508 if (port_no < 0 || port_no >= DP_MAX_PORTS || port_no == ODPP_LOCAL)
512 dp = get_dp_locked(dp_idx);
515 goto out_unlock_rtnl;
517 p = dp->ports[port_no];
522 err = dp_detach_port(p, 1);
525 mutex_unlock(&dp->mutex);
532 static void suppress_loop(struct datapath *dp, struct sw_flow_actions *actions)
535 pr_warn("%s: flow looped %d times, dropping\n",
536 dp_name(dp), DP_MAX_LOOPS);
537 actions->n_actions = 0;
540 /* Must be called with rcu_read_lock. */
541 void dp_process_received_packet(struct dp_port *p, struct sk_buff *skb)
543 struct datapath *dp = p->dp;
544 struct dp_stats_percpu *stats;
545 int stats_counter_off;
546 struct odp_flow_key key;
547 struct tbl_node *flow_node;
548 struct sw_flow *flow;
549 struct sw_flow_actions *acts;
550 struct loop_counter *loop;
553 OVS_CB(skb)->dp_port = p;
555 /* Extract flow from 'skb' into 'key'. */
556 error = flow_extract(skb, p ? p->port_no : ODPP_NONE, &key);
557 if (unlikely(error)) {
562 if (OVS_CB(skb)->is_frag && dp->drop_frags) {
564 stats_counter_off = offsetof(struct dp_stats_percpu, n_frags);
569 flow_node = tbl_lookup(rcu_dereference(dp->table), &key, flow_hash(&key), flow_cmp);
570 if (unlikely(!flow_node)) {
571 dp_output_control(dp, skb, _ODPL_MISS_NR, OVS_CB(skb)->tun_id);
572 stats_counter_off = offsetof(struct dp_stats_percpu, n_missed);
576 flow = flow_cast(flow_node);
577 flow_used(flow, skb);
579 acts = rcu_dereference(flow->sf_acts);
581 /* Check whether we've looped too much. */
582 loop = &get_cpu_var(dp_loop_counters).counters[!!in_interrupt()];
583 if (unlikely(++loop->count > DP_MAX_LOOPS))
584 loop->looping = true;
585 if (unlikely(loop->looping)) {
586 suppress_loop(dp, acts);
590 /* Execute actions. */
591 execute_actions(dp, skb, &key, acts->actions, acts->n_actions, GFP_ATOMIC);
592 stats_counter_off = offsetof(struct dp_stats_percpu, n_hit);
594 /* Check whether sub-actions looped too much. */
595 if (unlikely(loop->looping))
596 suppress_loop(dp, acts);
599 /* Decrement loop counter. */
601 loop->looping = false;
602 put_cpu_var(dp_loop_counters);
605 /* Update datapath statistics. */
607 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
609 write_seqcount_begin(&stats->seqlock);
610 (*(u64 *)((u8 *)stats + stats_counter_off))++;
611 write_seqcount_end(&stats->seqlock);
616 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
617 /* This code is based on skb_checksum_setup() from Xen's net/dev/core.c. We
618 * can't call this function directly because it isn't exported in all
620 int vswitch_skb_checksum_setup(struct sk_buff *skb)
625 __u16 csum_start, csum_offset;
627 if (!skb->proto_csum_blank)
630 if (skb->protocol != htons(ETH_P_IP))
633 if (!pskb_may_pull(skb, skb_network_header(skb) + sizeof(struct iphdr) - skb->data))
637 th = skb_network_header(skb) + 4 * iph->ihl;
639 csum_start = th - skb->head;
640 switch (iph->protocol) {
642 csum_offset = offsetof(struct tcphdr, check);
645 csum_offset = offsetof(struct udphdr, check);
649 pr_err("Attempting to checksum a non-TCP/UDP packet, "
650 "dropping a protocol %d packet",
655 if (!pskb_may_pull(skb, th + csum_offset + 2 - skb->data))
658 skb->ip_summed = CHECKSUM_PARTIAL;
659 skb->proto_csum_blank = 0;
661 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
662 skb->csum_start = csum_start;
663 skb->csum_offset = csum_offset;
665 skb_set_transport_header(skb, csum_start - skb_headroom(skb));
666 skb->csum = csum_offset;
674 #endif /* CONFIG_XEN && HAVE_PROTO_DATA_VALID */
676 /* Types of checksums that we can receive (these all refer to L4 checksums):
677 * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full
678 * (though not verified) checksum in packet but not in skb->csum. Packets
679 * from the bridge local port will also have this type.
680 * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums,
681 * also the GRE module. This is the same as CHECKSUM_NONE, except it has
682 * a valid skb->csum. Importantly, both contain a full checksum (not
683 * verified) in the packet itself. The only difference is that if the
684 * packet gets to L4 processing on this machine (not in DomU) we won't
685 * have to recompute the checksum to verify. Most hardware devices do not
686 * produce packets with this type, even if they support receive checksum
687 * offloading (they produce type #5).
688 * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to
689 * be computed if it is sent off box. Unfortunately on earlier kernels,
690 * this case is impossible to distinguish from #2, despite having opposite
691 * meanings. Xen adds an extra field on earlier kernels (see #4) in order
692 * to distinguish the different states.
693 * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was
694 * generated locally by a Xen DomU and has a partial checksum. If it is
695 * handled on this machine (Dom0 or DomU), then the checksum will not be
696 * computed. If it goes off box, the checksum in the packet needs to be
697 * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW
698 * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later
699 * kernels, this combination is replaced with CHECKSUM_PARTIAL.
700 * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct
701 * full checksum or using a protocol without a checksum. skb->csum is
702 * undefined. This is common from devices with receive checksum
703 * offloading. This is somewhat similar to CHECKSUM_NONE, except that
704 * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY.
706 * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are
707 * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear
708 * based on whether it is on the transmit or receive path. After the datapath
709 * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a
710 * checksum, we will panic. Since we can receive packets with checksums, we
711 * assume that all CHECKSUM_HW packets have checksums and map them to
712 * CHECKSUM_NONE, which has a similar meaning (the it is only different if the
713 * packet is processed by the local IP stack, in which case it will need to
714 * be reverified). If we receive a packet with CHECKSUM_HW that really means
715 * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there
716 * shouldn't be any devices that do this with bridging. */
717 void compute_ip_summed(struct sk_buff *skb, bool xmit)
719 /* For our convenience these defines change repeatedly between kernel
720 * versions, so we can't just copy them over... */
721 switch (skb->ip_summed) {
723 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
725 case CHECKSUM_UNNECESSARY:
726 OVS_CB(skb)->ip_summed = OVS_CSUM_UNNECESSARY;
729 /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
730 * However, on the receive side we should only get CHECKSUM_PARTIAL
731 * packets from Xen, which uses some special fields to represent this
732 * (see below). Since we can only make one type work, pick the one
733 * that actually happens in practice.
735 * On the transmit side (basically after skb_checksum_setup()
736 * has been run or on internal dev transmit), packets with
737 * CHECKSUM_COMPLETE aren't generated, so assume CHECKSUM_PARTIAL. */
740 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
742 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
746 case CHECKSUM_COMPLETE:
747 OVS_CB(skb)->ip_summed = OVS_CSUM_COMPLETE;
749 case CHECKSUM_PARTIAL:
750 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
754 pr_err("unknown checksum type %d\n", skb->ip_summed);
755 /* None seems the safest... */
756 OVS_CB(skb)->ip_summed = OVS_CSUM_NONE;
759 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
760 /* Xen has a special way of representing CHECKSUM_PARTIAL on older
761 * kernels. It should not be set on the transmit path though. */
762 if (skb->proto_csum_blank)
763 OVS_CB(skb)->ip_summed = OVS_CSUM_PARTIAL;
765 WARN_ON_ONCE(skb->proto_csum_blank && xmit);
769 /* This function closely resembles skb_forward_csum() used by the bridge. It
770 * is slightly different because we are only concerned with bridging and not
771 * other types of forwarding and can get away with slightly more optimal
773 void forward_ip_summed(struct sk_buff *skb)
776 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
777 skb->ip_summed = CHECKSUM_NONE;
781 /* Append each packet in 'skb' list to 'queue'. There will be only one packet
782 * unless we broke up a GSO packet. */
783 static int queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue,
784 int queue_no, u32 arg)
786 struct sk_buff *nskb;
790 if (OVS_CB(skb)->dp_port)
791 port_no = OVS_CB(skb)->dp_port->port_no;
793 port_no = ODPP_LOCAL;
796 struct odp_msg *header;
801 err = skb_cow(skb, sizeof *header);
805 header = (struct odp_msg*)__skb_push(skb, sizeof *header);
806 header->type = queue_no;
807 header->length = skb->len;
808 header->port = port_no;
809 header->reserved = 0;
811 skb_queue_tail(queue, skb);
819 while ((skb = nskb) != NULL) {
826 int dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no,
829 struct dp_stats_percpu *stats;
830 struct sk_buff_head *queue;
833 WARN_ON_ONCE(skb_shared(skb));
834 BUG_ON(queue_no != _ODPL_MISS_NR && queue_no != _ODPL_ACTION_NR && queue_no != _ODPL_SFLOW_NR);
835 queue = &dp->queues[queue_no];
837 if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN)
840 forward_ip_summed(skb);
842 err = vswitch_skb_checksum_setup(skb);
846 /* Break apart GSO packets into their component pieces. Otherwise
847 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
848 if (skb_is_gso(skb)) {
849 struct sk_buff *nskb = skb_gso_segment(skb, NETIF_F_SG | NETIF_F_HW_CSUM);
853 if (unlikely(IS_ERR(skb))) {
858 /* XXX This case might not be possible. It's hard to
859 * tell from the skb_gso_segment() code and comment. */
863 err = queue_control_packets(skb, queue, queue_no, arg);
864 wake_up_interruptible(&dp->waitqueue);
871 stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
873 write_seqcount_begin(&stats->seqlock);
875 write_seqcount_end(&stats->seqlock);
882 static int flush_flows(struct datapath *dp)
884 struct tbl *old_table = rcu_dereference(dp->table);
885 struct tbl *new_table;
887 new_table = tbl_create(0);
891 rcu_assign_pointer(dp->table, new_table);
893 tbl_deferred_destroy(old_table, flow_free_tbl);
898 static int validate_actions(const struct sw_flow_actions *actions)
902 for (i = 0; i < actions->n_actions; i++) {
903 const union odp_action *a = &actions->actions[i];
906 if (a->output.port >= DP_MAX_PORTS)
910 case ODPAT_OUTPUT_GROUP:
911 if (a->output_group.group >= DP_MAX_GROUPS)
915 case ODPAT_SET_VLAN_VID:
916 if (a->vlan_vid.vlan_vid & htons(~VLAN_VID_MASK))
920 case ODPAT_SET_VLAN_PCP:
921 if (a->vlan_pcp.vlan_pcp
922 & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT))
926 case ODPAT_SET_NW_TOS:
927 if (a->nw_tos.nw_tos & INET_ECN_MASK)
932 if (a->type >= ODPAT_N_ACTIONS)
941 static struct sw_flow_actions *get_actions(const struct odp_flow *flow)
943 struct sw_flow_actions *actions;
946 actions = flow_actions_alloc(flow->n_actions);
947 error = PTR_ERR(actions);
952 if (copy_from_user(actions->actions, flow->actions,
953 flow->n_actions * sizeof(union odp_action)))
954 goto error_free_actions;
955 error = validate_actions(actions);
957 goto error_free_actions;
964 return ERR_PTR(error);
967 static struct timespec get_time_offset(void)
969 struct timespec now_mono, now_jiffies;
971 ktime_get_ts(&now_mono);
972 jiffies_to_timespec(jiffies, &now_jiffies);
973 return timespec_sub(now_mono, now_jiffies);
976 static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats,
977 struct timespec time_offset)
980 struct timespec flow_ts, used;
982 jiffies_to_timespec(flow->used, &flow_ts);
983 set_normalized_timespec(&used, flow_ts.tv_sec + time_offset.tv_sec,
984 flow_ts.tv_nsec + time_offset.tv_nsec);
986 stats->used_sec = used.tv_sec;
987 stats->used_nsec = used.tv_nsec;
990 stats->used_nsec = 0;
993 stats->n_packets = flow->packet_count;
994 stats->n_bytes = flow->byte_count;
996 stats->tcp_flags = flow->tcp_flags;
1000 static void clear_stats(struct sw_flow *flow)
1003 flow->tcp_flags = 0;
1004 flow->packet_count = 0;
1005 flow->byte_count = 0;
1008 static int expand_table(struct datapath *dp)
1010 struct tbl *old_table = rcu_dereference(dp->table);
1011 struct tbl *new_table;
1013 new_table = tbl_expand(old_table);
1014 if (IS_ERR(new_table))
1015 return PTR_ERR(new_table);
1017 rcu_assign_pointer(dp->table, new_table);
1018 tbl_deferred_destroy(old_table, NULL);
1023 static int do_put_flow(struct datapath *dp, struct odp_flow_put *uf,
1024 struct odp_flow_stats *stats)
1026 struct tbl_node *flow_node;
1027 struct sw_flow *flow;
1031 memset(uf->flow.key.reserved, 0, sizeof uf->flow.key.reserved);
1033 table = rcu_dereference(dp->table);
1034 flow_node = tbl_lookup(table, &uf->flow.key, flow_hash(&uf->flow.key), flow_cmp);
1037 struct sw_flow_actions *acts;
1040 if (!(uf->flags & ODPPF_CREATE))
1043 /* Expand table, if necessary, to make room. */
1044 if (tbl_count(table) >= tbl_n_buckets(table)) {
1045 error = expand_table(dp);
1048 table = rcu_dereference(dp->table);
1051 /* Allocate flow. */
1052 flow = flow_alloc();
1054 error = PTR_ERR(flow);
1057 flow->key = uf->flow.key;
1060 /* Obtain actions. */
1061 acts = get_actions(&uf->flow);
1062 error = PTR_ERR(acts);
1064 goto error_free_flow;
1065 rcu_assign_pointer(flow->sf_acts, acts);
1067 /* Put flow in bucket. */
1068 error = tbl_insert(table, &flow->tbl_node, flow_hash(&flow->key));
1070 goto error_free_flow_acts;
1072 memset(stats, 0, sizeof(struct odp_flow_stats));
1074 /* We found a matching flow. */
1075 struct sw_flow_actions *old_acts, *new_acts;
1077 flow = flow_cast(flow_node);
1079 /* Bail out if we're not allowed to modify an existing flow. */
1081 if (!(uf->flags & ODPPF_MODIFY))
1085 new_acts = get_actions(&uf->flow);
1086 error = PTR_ERR(new_acts);
1087 if (IS_ERR(new_acts))
1089 old_acts = rcu_dereference(flow->sf_acts);
1090 if (old_acts->n_actions != new_acts->n_actions ||
1091 memcmp(old_acts->actions, new_acts->actions,
1092 sizeof(union odp_action) * old_acts->n_actions)) {
1093 rcu_assign_pointer(flow->sf_acts, new_acts);
1094 flow_deferred_free_acts(old_acts);
1099 /* Fetch stats, then clear them if necessary. */
1100 spin_lock_bh(&flow->lock);
1101 get_stats(flow, stats, get_time_offset());
1102 if (uf->flags & ODPPF_ZERO_STATS)
1104 spin_unlock_bh(&flow->lock);
1109 error_free_flow_acts:
1110 kfree(flow->sf_acts);
1117 static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp)
1119 struct odp_flow_stats stats;
1120 struct odp_flow_put uf;
1123 if (copy_from_user(&uf, ufp, sizeof(struct odp_flow_put)))
1126 error = do_put_flow(dp, &uf, &stats);
1130 if (copy_to_user(&ufp->flow.stats, &stats,
1131 sizeof(struct odp_flow_stats)))
1137 static int do_answer_query(struct sw_flow *flow, u32 query_flags,
1138 struct timespec time_offset,
1139 struct odp_flow_stats __user *ustats,
1140 union odp_action __user *actions,
1141 u32 __user *n_actionsp)
1143 struct sw_flow_actions *sf_acts;
1144 struct odp_flow_stats stats;
1147 spin_lock_bh(&flow->lock);
1148 get_stats(flow, &stats, time_offset);
1149 if (query_flags & ODPFF_ZERO_TCP_FLAGS)
1150 flow->tcp_flags = 0;
1152 spin_unlock_bh(&flow->lock);
1154 if (copy_to_user(ustats, &stats, sizeof(struct odp_flow_stats)) ||
1155 get_user(n_actions, n_actionsp))
1161 sf_acts = rcu_dereference(flow->sf_acts);
1162 if (put_user(sf_acts->n_actions, n_actionsp) ||
1163 (actions && copy_to_user(actions, sf_acts->actions,
1164 sizeof(union odp_action) *
1165 min(sf_acts->n_actions, n_actions))))
1171 static int answer_query(struct sw_flow *flow, u32 query_flags,
1172 struct timespec time_offset,
1173 struct odp_flow __user *ufp)
1175 union odp_action *actions;
1177 if (get_user(actions, &ufp->actions))
1180 return do_answer_query(flow, query_flags, time_offset,
1181 &ufp->stats, actions, &ufp->n_actions);
1184 static struct sw_flow *do_del_flow(struct datapath *dp, struct odp_flow_key *key)
1186 struct tbl *table = rcu_dereference(dp->table);
1187 struct tbl_node *flow_node;
1190 memset(key->reserved, 0, sizeof key->reserved);
1191 flow_node = tbl_lookup(table, key, flow_hash(key), flow_cmp);
1193 return ERR_PTR(-ENOENT);
1195 error = tbl_remove(table, flow_node);
1197 return ERR_PTR(error);
1199 /* XXX Returned flow_node's statistics might lose a few packets, since
1200 * other CPUs can be using this flow. We used to synchronize_rcu() to
1201 * make sure that we get completely accurate stats, but that blows our
1202 * performance, badly. */
1203 return flow_cast(flow_node);
1206 static int del_flow(struct datapath *dp, struct odp_flow __user *ufp)
1208 struct sw_flow *flow;
1212 if (copy_from_user(&uf, ufp, sizeof uf))
1215 flow = do_del_flow(dp, &uf.key);
1217 return PTR_ERR(flow);
1219 error = answer_query(flow, 0, get_time_offset(), ufp);
1220 flow_deferred_free(flow);
1224 static int do_query_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1226 struct tbl *table = rcu_dereference(dp->table);
1227 struct timespec time_offset;
1230 time_offset = get_time_offset();
1232 for (i = 0; i < flowvec->n_flows; i++) {
1233 struct odp_flow __user *ufp = &flowvec->flows[i];
1235 struct tbl_node *flow_node;
1238 if (copy_from_user(&uf, ufp, sizeof uf))
1240 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1242 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1244 error = put_user(ENOENT, &ufp->stats.error);
1246 error = answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1250 return flowvec->n_flows;
1253 struct list_flows_cbdata {
1254 struct odp_flow __user *uflows;
1257 struct timespec time_offset;
1260 static int list_flow(struct tbl_node *node, void *cbdata_)
1262 struct sw_flow *flow = flow_cast(node);
1263 struct list_flows_cbdata *cbdata = cbdata_;
1264 struct odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1267 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1269 error = answer_query(flow, 0, cbdata->time_offset, ufp);
1273 if (cbdata->listed_flows >= cbdata->n_flows)
1274 return cbdata->listed_flows;
1278 static int do_list_flows(struct datapath *dp, const struct odp_flowvec *flowvec)
1280 struct list_flows_cbdata cbdata;
1283 if (!flowvec->n_flows)
1286 cbdata.uflows = flowvec->flows;
1287 cbdata.n_flows = flowvec->n_flows;
1288 cbdata.listed_flows = 0;
1289 cbdata.time_offset = get_time_offset();
1291 error = tbl_foreach(rcu_dereference(dp->table), list_flow, &cbdata);
1292 return error ? error : cbdata.listed_flows;
1295 static int do_flowvec_ioctl(struct datapath *dp, unsigned long argp,
1296 int (*function)(struct datapath *,
1297 const struct odp_flowvec *))
1299 struct odp_flowvec __user *uflowvec;
1300 struct odp_flowvec flowvec;
1303 uflowvec = (struct odp_flowvec __user *)argp;
1304 if (copy_from_user(&flowvec, uflowvec, sizeof flowvec))
1307 if (flowvec.n_flows > INT_MAX / sizeof(struct odp_flow))
1310 retval = function(dp, &flowvec);
1311 return (retval < 0 ? retval
1312 : retval == flowvec.n_flows ? 0
1313 : put_user(retval, &uflowvec->n_flows));
1316 static int do_execute(struct datapath *dp, const struct odp_execute *execute)
1318 struct odp_flow_key key;
1319 struct sk_buff *skb;
1320 struct sw_flow_actions *actions;
1325 if (execute->length < ETH_HLEN || execute->length > 65535)
1328 actions = flow_actions_alloc(execute->n_actions);
1329 if (IS_ERR(actions)) {
1330 err = PTR_ERR(actions);
1335 if (copy_from_user(actions->actions, execute->actions,
1336 execute->n_actions * sizeof *execute->actions))
1337 goto error_free_actions;
1339 err = validate_actions(actions);
1341 goto error_free_actions;
1344 skb = alloc_skb(execute->length, GFP_KERNEL);
1346 goto error_free_actions;
1348 if (execute->in_port < DP_MAX_PORTS)
1349 OVS_CB(skb)->dp_port = dp->ports[execute->in_port];
1351 OVS_CB(skb)->dp_port = NULL;
1354 if (copy_from_user(skb_put(skb, execute->length), execute->data,
1356 goto error_free_skb;
1358 skb_reset_mac_header(skb);
1361 /* Normally, setting the skb 'protocol' field would be handled by a
1362 * call to eth_type_trans(), but it assumes there's a sending
1363 * device, which we may not have. */
1364 if (ntohs(eth->h_proto) >= 1536)
1365 skb->protocol = eth->h_proto;
1367 skb->protocol = htons(ETH_P_802_2);
1369 err = flow_extract(skb, execute->in_port, &key);
1371 goto error_free_skb;
1374 err = execute_actions(dp, skb, &key, actions->actions,
1375 actions->n_actions, GFP_KERNEL);
1389 static int execute_packet(struct datapath *dp, const struct odp_execute __user *executep)
1391 struct odp_execute execute;
1393 if (copy_from_user(&execute, executep, sizeof execute))
1396 return do_execute(dp, &execute);
1399 static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp)
1401 struct tbl *table = rcu_dereference(dp->table);
1402 struct odp_stats stats;
1405 stats.n_flows = tbl_count(table);
1406 stats.cur_capacity = tbl_n_buckets(table);
1407 stats.max_capacity = TBL_MAX_BUCKETS;
1408 stats.n_ports = dp->n_ports;
1409 stats.max_ports = DP_MAX_PORTS;
1410 stats.max_groups = DP_MAX_GROUPS;
1411 stats.n_frags = stats.n_hit = stats.n_missed = stats.n_lost = 0;
1412 for_each_possible_cpu(i) {
1413 const struct dp_stats_percpu *percpu_stats;
1414 struct dp_stats_percpu local_stats;
1417 percpu_stats = per_cpu_ptr(dp->stats_percpu, i);
1420 seqcount = read_seqcount_begin(&percpu_stats->seqlock);
1421 local_stats = *percpu_stats;
1422 } while (read_seqcount_retry(&percpu_stats->seqlock, seqcount));
1424 stats.n_frags += local_stats.n_frags;
1425 stats.n_hit += local_stats.n_hit;
1426 stats.n_missed += local_stats.n_missed;
1427 stats.n_lost += local_stats.n_lost;
1429 stats.max_miss_queue = DP_MAX_QUEUE_LEN;
1430 stats.max_action_queue = DP_MAX_QUEUE_LEN;
1431 return copy_to_user(statsp, &stats, sizeof stats) ? -EFAULT : 0;
1434 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1435 int dp_min_mtu(const struct datapath *dp)
1442 list_for_each_entry_rcu (p, &dp->port_list, node) {
1445 /* Skip any internal ports, since that's what we're trying to
1447 if (is_internal_vport(p->vport))
1450 dev_mtu = vport_get_mtu(p->vport);
1451 if (!mtu || dev_mtu < mtu)
1455 return mtu ? mtu : ETH_DATA_LEN;
1458 /* Sets the MTU of all datapath devices to the minimum of the ports. Must
1459 * be called with RTNL lock. */
1460 void set_internal_devs_mtu(const struct datapath *dp)
1467 mtu = dp_min_mtu(dp);
1469 list_for_each_entry_rcu (p, &dp->port_list, node) {
1470 if (is_internal_vport(p->vport))
1471 vport_set_mtu(p->vport, mtu);
1475 static int put_port(const struct dp_port *p, struct odp_port __user *uop)
1479 memset(&op, 0, sizeof op);
1482 strncpy(op.devname, vport_get_name(p->vport), sizeof op.devname);
1485 op.port = p->port_no;
1486 op.flags = is_internal_vport(p->vport) ? ODP_PORT_INTERNAL : 0;
1488 return copy_to_user(uop, &op, sizeof op) ? -EFAULT : 0;
1491 static int query_port(struct datapath *dp, struct odp_port __user *uport)
1493 struct odp_port port;
1495 if (copy_from_user(&port, uport, sizeof port))
1498 if (port.devname[0]) {
1499 struct vport *vport;
1500 struct dp_port *dp_port;
1503 port.devname[IFNAMSIZ - 1] = '\0';
1508 vport = vport_locate(port.devname);
1514 dp_port = vport_get_dp_port(vport);
1515 if (!dp_port || dp_port->dp != dp) {
1520 port.port = dp_port->port_no;
1529 if (port.port >= DP_MAX_PORTS)
1531 if (!dp->ports[port.port])
1535 return put_port(dp->ports[port.port], uport);
1538 static int do_list_ports(struct datapath *dp, struct odp_port __user *uports,
1545 list_for_each_entry_rcu (p, &dp->port_list, node) {
1546 if (put_port(p, &uports[idx]))
1548 if (idx++ >= n_ports)
1555 static int list_ports(struct datapath *dp, struct odp_portvec __user *upv)
1557 struct odp_portvec pv;
1560 if (copy_from_user(&pv, upv, sizeof pv))
1563 retval = do_list_ports(dp, pv.ports, pv.n_ports);
1567 return put_user(retval, &upv->n_ports);
1570 /* RCU callback for freeing a dp_port_group */
1571 static void free_port_group(struct rcu_head *rcu)
1573 struct dp_port_group *g = container_of(rcu, struct dp_port_group, rcu);
1577 static int do_set_port_group(struct datapath *dp, u16 __user *ports,
1578 int n_ports, int group)
1580 struct dp_port_group *new_group, *old_group;
1584 if (n_ports > DP_MAX_PORTS || group >= DP_MAX_GROUPS)
1588 new_group = kmalloc(sizeof *new_group + sizeof(u16) * n_ports, GFP_KERNEL);
1592 new_group->n_ports = n_ports;
1594 if (copy_from_user(new_group->ports, ports, sizeof(u16) * n_ports))
1597 old_group = rcu_dereference(dp->groups[group]);
1598 rcu_assign_pointer(dp->groups[group], new_group);
1600 call_rcu(&old_group->rcu, free_port_group);
1609 static int set_port_group(struct datapath *dp,
1610 const struct odp_port_group __user *upg)
1612 struct odp_port_group pg;
1614 if (copy_from_user(&pg, upg, sizeof pg))
1617 return do_set_port_group(dp, pg.ports, pg.n_ports, pg.group);
1620 static int do_get_port_group(struct datapath *dp,
1621 u16 __user *ports, int n_ports, int group,
1622 u16 __user *n_portsp)
1624 struct dp_port_group *g;
1627 if (group >= DP_MAX_GROUPS)
1630 g = dp->groups[group];
1631 n_copy = g ? min_t(int, g->n_ports, n_ports) : 0;
1632 if (n_copy && copy_to_user(ports, g->ports, n_copy * sizeof(u16)))
1635 if (put_user(g ? g->n_ports : 0, n_portsp))
1641 static int get_port_group(struct datapath *dp, struct odp_port_group __user *upg)
1643 struct odp_port_group pg;
1645 if (copy_from_user(&pg, upg, sizeof pg))
1648 return do_get_port_group(dp, pg.ports, pg.n_ports, pg.group, &upg->n_ports);
1651 static int get_listen_mask(const struct file *f)
1653 return (long)f->private_data;
1656 static void set_listen_mask(struct file *f, int listen_mask)
1658 f->private_data = (void*)(long)listen_mask;
1661 static long openvswitch_ioctl(struct file *f, unsigned int cmd,
1664 int dp_idx = iminor(f->f_dentry->d_inode);
1665 struct datapath *dp;
1666 int drop_frags, listeners, port_no;
1667 unsigned int sflow_probability;
1670 /* Handle commands with special locking requirements up front. */
1673 err = create_dp(dp_idx, (char __user *)argp);
1676 case ODP_DP_DESTROY:
1677 err = destroy_dp(dp_idx);
1680 case ODP_PORT_ATTACH:
1681 err = attach_port(dp_idx, (struct odp_port __user *)argp);
1684 case ODP_PORT_DETACH:
1685 err = get_user(port_no, (int __user *)argp);
1687 err = detach_port(dp_idx, port_no);
1691 err = vport_user_add((struct odp_vport_add __user *)argp);
1695 err = vport_user_mod((struct odp_vport_mod __user *)argp);
1699 err = vport_user_del((char __user *)argp);
1702 case ODP_VPORT_STATS_GET:
1703 err = vport_user_stats_get((struct odp_vport_stats_req __user *)argp);
1706 case ODP_VPORT_STATS_SET:
1707 err = vport_user_stats_set((struct odp_vport_stats_req __user *)argp);
1710 case ODP_VPORT_ETHER_GET:
1711 err = vport_user_ether_get((struct odp_vport_ether __user *)argp);
1714 case ODP_VPORT_ETHER_SET:
1715 err = vport_user_ether_set((struct odp_vport_ether __user *)argp);
1718 case ODP_VPORT_MTU_GET:
1719 err = vport_user_mtu_get((struct odp_vport_mtu __user *)argp);
1722 case ODP_VPORT_MTU_SET:
1723 err = vport_user_mtu_set((struct odp_vport_mtu __user *)argp);
1727 dp = get_dp_locked(dp_idx);
1734 err = get_dp_stats(dp, (struct odp_stats __user *)argp);
1737 case ODP_GET_DROP_FRAGS:
1738 err = put_user(dp->drop_frags, (int __user *)argp);
1741 case ODP_SET_DROP_FRAGS:
1742 err = get_user(drop_frags, (int __user *)argp);
1746 if (drop_frags != 0 && drop_frags != 1)
1748 dp->drop_frags = drop_frags;
1752 case ODP_GET_LISTEN_MASK:
1753 err = put_user(get_listen_mask(f), (int __user *)argp);
1756 case ODP_SET_LISTEN_MASK:
1757 err = get_user(listeners, (int __user *)argp);
1761 if (listeners & ~ODPL_ALL)
1764 set_listen_mask(f, listeners);
1767 case ODP_GET_SFLOW_PROBABILITY:
1768 err = put_user(dp->sflow_probability, (unsigned int __user *)argp);
1771 case ODP_SET_SFLOW_PROBABILITY:
1772 err = get_user(sflow_probability, (unsigned int __user *)argp);
1774 dp->sflow_probability = sflow_probability;
1777 case ODP_PORT_QUERY:
1778 err = query_port(dp, (struct odp_port __user *)argp);
1782 err = list_ports(dp, (struct odp_portvec __user *)argp);
1785 case ODP_PORT_GROUP_SET:
1786 err = set_port_group(dp, (struct odp_port_group __user *)argp);
1789 case ODP_PORT_GROUP_GET:
1790 err = get_port_group(dp, (struct odp_port_group __user *)argp);
1793 case ODP_FLOW_FLUSH:
1794 err = flush_flows(dp);
1798 err = put_flow(dp, (struct odp_flow_put __user *)argp);
1802 err = del_flow(dp, (struct odp_flow __user *)argp);
1806 err = do_flowvec_ioctl(dp, argp, do_query_flows);
1810 err = do_flowvec_ioctl(dp, argp, do_list_flows);
1814 err = execute_packet(dp, (struct odp_execute __user *)argp);
1821 mutex_unlock(&dp->mutex);
1826 static int dp_has_packet_of_interest(struct datapath *dp, int listeners)
1829 for (i = 0; i < DP_N_QUEUES; i++) {
1830 if (listeners & (1 << i) && !skb_queue_empty(&dp->queues[i]))
1836 #ifdef CONFIG_COMPAT
1837 static int compat_list_ports(struct datapath *dp, struct compat_odp_portvec __user *upv)
1839 struct compat_odp_portvec pv;
1842 if (copy_from_user(&pv, upv, sizeof pv))
1845 retval = do_list_ports(dp, compat_ptr(pv.ports), pv.n_ports);
1849 return put_user(retval, &upv->n_ports);
1852 static int compat_set_port_group(struct datapath *dp, const struct compat_odp_port_group __user *upg)
1854 struct compat_odp_port_group pg;
1856 if (copy_from_user(&pg, upg, sizeof pg))
1859 return do_set_port_group(dp, compat_ptr(pg.ports), pg.n_ports, pg.group);
1862 static int compat_get_port_group(struct datapath *dp, struct compat_odp_port_group __user *upg)
1864 struct compat_odp_port_group pg;
1866 if (copy_from_user(&pg, upg, sizeof pg))
1869 return do_get_port_group(dp, compat_ptr(pg.ports), pg.n_ports,
1870 pg.group, &upg->n_ports);
1873 static int compat_get_flow(struct odp_flow *flow, const struct compat_odp_flow __user *compat)
1875 compat_uptr_t actions;
1877 if (!access_ok(VERIFY_READ, compat, sizeof(struct compat_odp_flow)) ||
1878 __copy_from_user(&flow->stats, &compat->stats, sizeof(struct odp_flow_stats)) ||
1879 __copy_from_user(&flow->key, &compat->key, sizeof(struct odp_flow_key)) ||
1880 __get_user(actions, &compat->actions) ||
1881 __get_user(flow->n_actions, &compat->n_actions) ||
1882 __get_user(flow->flags, &compat->flags))
1885 flow->actions = compat_ptr(actions);
1889 static int compat_put_flow(struct datapath *dp, struct compat_odp_flow_put __user *ufp)
1891 struct odp_flow_stats stats;
1892 struct odp_flow_put fp;
1895 if (compat_get_flow(&fp.flow, &ufp->flow) ||
1896 get_user(fp.flags, &ufp->flags))
1899 error = do_put_flow(dp, &fp, &stats);
1903 if (copy_to_user(&ufp->flow.stats, &stats,
1904 sizeof(struct odp_flow_stats)))
1910 static int compat_answer_query(struct sw_flow *flow, u32 query_flags,
1911 struct timespec time_offset,
1912 struct compat_odp_flow __user *ufp)
1914 compat_uptr_t actions;
1916 if (get_user(actions, &ufp->actions))
1919 return do_answer_query(flow, query_flags, time_offset, &ufp->stats,
1920 compat_ptr(actions), &ufp->n_actions);
1923 static int compat_del_flow(struct datapath *dp, struct compat_odp_flow __user *ufp)
1925 struct sw_flow *flow;
1929 if (compat_get_flow(&uf, ufp))
1932 flow = do_del_flow(dp, &uf.key);
1934 return PTR_ERR(flow);
1936 error = compat_answer_query(flow, 0, get_time_offset(), ufp);
1937 flow_deferred_free(flow);
1941 static int compat_query_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1943 struct tbl *table = rcu_dereference(dp->table);
1944 struct timespec time_offset;
1947 time_offset = get_time_offset();
1949 for (i = 0; i < n_flows; i++) {
1950 struct compat_odp_flow __user *ufp = &flows[i];
1952 struct tbl_node *flow_node;
1955 if (compat_get_flow(&uf, ufp))
1957 memset(uf.key.reserved, 0, sizeof uf.key.reserved);
1959 flow_node = tbl_lookup(table, &uf.key, flow_hash(&uf.key), flow_cmp);
1961 error = put_user(ENOENT, &ufp->stats.error);
1963 error = compat_answer_query(flow_cast(flow_node), uf.flags, time_offset, ufp);
1970 struct compat_list_flows_cbdata {
1971 struct compat_odp_flow __user *uflows;
1974 struct timespec time_offset;
1977 static int compat_list_flow(struct tbl_node *node, void *cbdata_)
1979 struct sw_flow *flow = flow_cast(node);
1980 struct compat_list_flows_cbdata *cbdata = cbdata_;
1981 struct compat_odp_flow __user *ufp = &cbdata->uflows[cbdata->listed_flows++];
1984 if (copy_to_user(&ufp->key, &flow->key, sizeof flow->key))
1986 error = compat_answer_query(flow, 0, cbdata->time_offset, ufp);
1990 if (cbdata->listed_flows >= cbdata->n_flows)
1991 return cbdata->listed_flows;
1995 static int compat_list_flows(struct datapath *dp, struct compat_odp_flow *flows, u32 n_flows)
1997 struct compat_list_flows_cbdata cbdata;
2003 cbdata.uflows = flows;
2004 cbdata.n_flows = n_flows;
2005 cbdata.listed_flows = 0;
2006 cbdata.time_offset = get_time_offset();
2008 error = tbl_foreach(rcu_dereference(dp->table), compat_list_flow, &cbdata);
2009 return error ? error : cbdata.listed_flows;
2012 static int compat_flowvec_ioctl(struct datapath *dp, unsigned long argp,
2013 int (*function)(struct datapath *,
2014 struct compat_odp_flow *,
2017 struct compat_odp_flowvec __user *uflowvec;
2018 struct compat_odp_flow __user *flows;
2019 struct compat_odp_flowvec flowvec;
2022 uflowvec = compat_ptr(argp);
2023 if (!access_ok(VERIFY_WRITE, uflowvec, sizeof *uflowvec) ||
2024 copy_from_user(&flowvec, uflowvec, sizeof flowvec))
2027 if (flowvec.n_flows > INT_MAX / sizeof(struct compat_odp_flow))
2030 flows = compat_ptr(flowvec.flows);
2031 if (!access_ok(VERIFY_WRITE, flows,
2032 flowvec.n_flows * sizeof(struct compat_odp_flow)))
2035 retval = function(dp, flows, flowvec.n_flows);
2036 return (retval < 0 ? retval
2037 : retval == flowvec.n_flows ? 0
2038 : put_user(retval, &uflowvec->n_flows));
2041 static int compat_execute(struct datapath *dp, const struct compat_odp_execute __user *uexecute)
2043 struct odp_execute execute;
2044 compat_uptr_t actions;
2047 if (!access_ok(VERIFY_READ, uexecute, sizeof(struct compat_odp_execute)) ||
2048 __get_user(execute.in_port, &uexecute->in_port) ||
2049 __get_user(actions, &uexecute->actions) ||
2050 __get_user(execute.n_actions, &uexecute->n_actions) ||
2051 __get_user(data, &uexecute->data) ||
2052 __get_user(execute.length, &uexecute->length))
2055 execute.actions = compat_ptr(actions);
2056 execute.data = compat_ptr(data);
2058 return do_execute(dp, &execute);
2061 static long openvswitch_compat_ioctl(struct file *f, unsigned int cmd, unsigned long argp)
2063 int dp_idx = iminor(f->f_dentry->d_inode);
2064 struct datapath *dp;
2068 case ODP_DP_DESTROY:
2069 case ODP_FLOW_FLUSH:
2070 /* Ioctls that don't need any translation at all. */
2071 return openvswitch_ioctl(f, cmd, argp);
2074 case ODP_PORT_ATTACH:
2075 case ODP_PORT_DETACH:
2077 case ODP_VPORT_MTU_SET:
2078 case ODP_VPORT_MTU_GET:
2079 case ODP_VPORT_ETHER_SET:
2080 case ODP_VPORT_ETHER_GET:
2081 case ODP_VPORT_STATS_SET:
2082 case ODP_VPORT_STATS_GET:
2084 case ODP_GET_DROP_FRAGS:
2085 case ODP_SET_DROP_FRAGS:
2086 case ODP_SET_LISTEN_MASK:
2087 case ODP_GET_LISTEN_MASK:
2088 case ODP_SET_SFLOW_PROBABILITY:
2089 case ODP_GET_SFLOW_PROBABILITY:
2090 case ODP_PORT_QUERY:
2091 /* Ioctls that just need their pointer argument extended. */
2092 return openvswitch_ioctl(f, cmd, (unsigned long)compat_ptr(argp));
2094 case ODP_VPORT_ADD32:
2095 return compat_vport_user_add(compat_ptr(argp));
2097 case ODP_VPORT_MOD32:
2098 return compat_vport_user_mod(compat_ptr(argp));
2101 dp = get_dp_locked(dp_idx);
2107 case ODP_PORT_LIST32:
2108 err = compat_list_ports(dp, compat_ptr(argp));
2111 case ODP_PORT_GROUP_SET32:
2112 err = compat_set_port_group(dp, compat_ptr(argp));
2115 case ODP_PORT_GROUP_GET32:
2116 err = compat_get_port_group(dp, compat_ptr(argp));
2119 case ODP_FLOW_PUT32:
2120 err = compat_put_flow(dp, compat_ptr(argp));
2123 case ODP_FLOW_DEL32:
2124 err = compat_del_flow(dp, compat_ptr(argp));
2127 case ODP_FLOW_GET32:
2128 err = compat_flowvec_ioctl(dp, argp, compat_query_flows);
2131 case ODP_FLOW_LIST32:
2132 err = compat_flowvec_ioctl(dp, argp, compat_list_flows);
2136 err = compat_execute(dp, compat_ptr(argp));
2143 mutex_unlock(&dp->mutex);
2149 /* Unfortunately this function is not exported so this is a verbatim copy
2150 * from net/core/datagram.c in 2.6.30. */
2151 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
2152 u8 __user *to, int len,
2155 int start = skb_headlen(skb);
2157 int i, copy = start - offset;
2164 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
2168 if ((len -= copy) == 0)
2175 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2178 WARN_ON(start > offset + len);
2180 end = start + skb_shinfo(skb)->frags[i].size;
2181 if ((copy = end - offset) > 0) {
2185 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2186 struct page *page = frag->page;
2191 csum2 = csum_and_copy_to_user(vaddr +
2198 *csump = csum_block_add(*csump, csum2, pos);
2208 if (skb_shinfo(skb)->frag_list) {
2209 struct sk_buff *list = skb_shinfo(skb)->frag_list;
2211 for (; list; list=list->next) {
2214 WARN_ON(start > offset + len);
2216 end = start + list->len;
2217 if ((copy = end - offset) > 0) {
2221 if (skb_copy_and_csum_datagram(list,
2226 *csump = csum_block_add(*csump, csum2, pos);
2227 if ((len -= copy) == 0)
2243 ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes,
2246 /* XXX is there sufficient synchronization here? */
2247 int listeners = get_listen_mask(f);
2248 int dp_idx = iminor(f->f_dentry->d_inode);
2249 struct datapath *dp = get_dp(dp_idx);
2250 struct sk_buff *skb;
2251 size_t copy_bytes, tot_copy_bytes;
2257 if (nbytes == 0 || !listeners)
2263 for (i = 0; i < DP_N_QUEUES; i++) {
2264 if (listeners & (1 << i)) {
2265 skb = skb_dequeue(&dp->queues[i]);
2271 if (f->f_flags & O_NONBLOCK) {
2276 wait_event_interruptible(dp->waitqueue,
2277 dp_has_packet_of_interest(dp,
2280 if (signal_pending(current)) {
2281 retval = -ERESTARTSYS;
2286 copy_bytes = tot_copy_bytes = min_t(size_t, skb->len, nbytes);
2289 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2290 if (copy_bytes == skb->len) {
2292 unsigned int csum_start, csum_offset;
2294 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
2295 csum_start = skb->csum_start - skb_headroom(skb);
2296 csum_offset = skb->csum_offset;
2298 csum_start = skb_transport_header(skb) - skb->data;
2299 csum_offset = skb->csum;
2301 BUG_ON(csum_start >= skb_headlen(skb));
2302 retval = skb_copy_and_csum_datagram(skb, csum_start, buf + csum_start,
2303 copy_bytes - csum_start, &csum);
2305 __sum16 __user *csump;
2307 copy_bytes = csum_start;
2308 csump = (__sum16 __user *)(buf + csum_start + csum_offset);
2310 BUG_ON((char *)csump + sizeof(__sum16) > buf + nbytes);
2311 put_user(csum_fold(csum), csump);
2314 retval = skb_checksum_help(skb);
2318 struct iovec __user iov;
2321 iov.iov_len = copy_bytes;
2322 retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len);
2326 retval = tot_copy_bytes;
2334 static unsigned int openvswitch_poll(struct file *file, poll_table *wait)
2336 /* XXX is there sufficient synchronization here? */
2337 int dp_idx = iminor(file->f_dentry->d_inode);
2338 struct datapath *dp = get_dp(dp_idx);
2343 poll_wait(file, &dp->waitqueue, wait);
2344 if (dp_has_packet_of_interest(dp, get_listen_mask(file)))
2345 mask |= POLLIN | POLLRDNORM;
2347 mask = POLLIN | POLLRDNORM | POLLHUP;
2352 struct file_operations openvswitch_fops = {
2353 /* XXX .aio_read = openvswitch_aio_read, */
2354 .read = openvswitch_read,
2355 .poll = openvswitch_poll,
2356 .unlocked_ioctl = openvswitch_ioctl,
2357 #ifdef CONFIG_COMPAT
2358 .compat_ioctl = openvswitch_compat_ioctl,
2360 /* XXX .fasync = openvswitch_fasync, */
2365 static int __init dp_init(void)
2367 struct sk_buff *dummy_skb;
2370 BUILD_BUG_ON(sizeof(struct ovs_skb_cb) > sizeof(dummy_skb->cb));
2372 printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR);
2380 goto error_flow_exit;
2382 err = register_netdevice_notifier(&dp_device_notifier);
2384 goto error_vport_exit;
2386 major = register_chrdev(0, "openvswitch", &openvswitch_fops);
2388 goto error_unreg_notifier;
2392 error_unreg_notifier:
2393 unregister_netdevice_notifier(&dp_device_notifier);
2402 static void dp_cleanup(void)
2405 unregister_chrdev(major, "openvswitch");
2406 unregister_netdevice_notifier(&dp_device_notifier);
2411 module_init(dp_init);
2412 module_exit(dp_cleanup);
2414 MODULE_DESCRIPTION("Open vSwitch switching datapath");
2415 MODULE_LICENSE("GPL");