2 * Copyright (c) 2007, 2008, 2009 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_bridge.h>
16 #include <linux/if_vlan.h>
19 #include <linux/delay.h>
20 #include <linux/time.h>
21 #include <linux/etherdevice.h>
22 #include <linux/kernel.h>
23 #include <linux/kthread.h>
24 #include <linux/llc.h>
25 #include <linux/mutex.h>
26 #include <linux/percpu.h>
27 #include <linux/rcupdate.h>
28 #include <linux/tcp.h>
29 #include <linux/udp.h>
30 #include <linux/version.h>
31 #include <linux/ethtool.h>
32 #include <linux/random.h>
33 #include <linux/wait.h>
34 #include <asm/system.h>
35 #include <asm/div64.h>
37 #include <linux/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>
46 #include "openvswitch/datapath-protocol.h"
55 int (*dp_ioctl_hook
)(struct net_device
*dev
, struct ifreq
*rq
, int cmd
);
56 EXPORT_SYMBOL(dp_ioctl_hook
);
58 /* Datapaths. Protected on the read side by rcu_read_lock, on the write side
59 * by dp_mutex. dp_mutex is almost completely redundant with genl_mutex
60 * maintained by the Generic Netlink code, but the timeout path needs mutual
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 net_bridge_port structures with just
69 static struct datapath
*dps
[ODP_MAX
];
70 static DEFINE_MUTEX(dp_mutex
);
72 /* Number of milliseconds between runs of the maintenance thread. */
73 #define MAINT_SLEEP_MSECS 1000
75 static int new_nbp(struct datapath
*, struct net_device
*, int port_no
);
77 /* Must be called with rcu_read_lock or dp_mutex. */
78 struct datapath
*get_dp(int dp_idx
)
80 if (dp_idx
< 0 || dp_idx
>= ODP_MAX
)
82 return rcu_dereference(dps
[dp_idx
]);
84 EXPORT_SYMBOL_GPL(get_dp
);
86 struct datapath
*get_dp_locked(int dp_idx
)
90 mutex_lock(&dp_mutex
);
93 mutex_lock(&dp
->mutex
);
94 mutex_unlock(&dp_mutex
);
98 static inline size_t br_nlmsg_size(void)
100 return NLMSG_ALIGN(sizeof(struct ifinfomsg
))
101 + nla_total_size(IFNAMSIZ
) /* IFLA_IFNAME */
102 + nla_total_size(MAX_ADDR_LEN
) /* IFLA_ADDRESS */
103 + nla_total_size(4) /* IFLA_MASTER */
104 + nla_total_size(4) /* IFLA_MTU */
105 + nla_total_size(4) /* IFLA_LINK */
106 + nla_total_size(1); /* IFLA_OPERSTATE */
109 static int dp_fill_ifinfo(struct sk_buff
*skb
,
110 const struct net_bridge_port
*port
,
111 int event
, unsigned int flags
)
113 const struct datapath
*dp
= port
->dp
;
114 const struct net_device
*dev
= port
->dev
;
115 struct ifinfomsg
*hdr
;
116 struct nlmsghdr
*nlh
;
118 nlh
= nlmsg_put(skb
, 0, 0, event
, sizeof(*hdr
), flags
);
122 hdr
= nlmsg_data(nlh
);
123 hdr
->ifi_family
= AF_BRIDGE
;
125 hdr
->ifi_type
= dev
->type
;
126 hdr
->ifi_index
= dev
->ifindex
;
127 hdr
->ifi_flags
= dev_get_flags(dev
);
130 NLA_PUT_STRING(skb
, IFLA_IFNAME
, dev
->name
);
131 NLA_PUT_U32(skb
, IFLA_MASTER
, dp
->ports
[ODPP_LOCAL
]->dev
->ifindex
);
132 NLA_PUT_U32(skb
, IFLA_MTU
, dev
->mtu
);
133 #ifdef IFLA_OPERSTATE
134 NLA_PUT_U8(skb
, IFLA_OPERSTATE
,
135 netif_running(dev
) ? dev
->operstate
: IF_OPER_DOWN
);
139 NLA_PUT(skb
, IFLA_ADDRESS
, dev
->addr_len
, dev
->dev_addr
);
141 if (dev
->ifindex
!= dev
->iflink
)
142 NLA_PUT_U32(skb
, IFLA_LINK
, dev
->iflink
);
144 return nlmsg_end(skb
, nlh
);
147 nlmsg_cancel(skb
, nlh
);
151 static void dp_ifinfo_notify(int event
, struct net_bridge_port
*port
)
153 struct net
*net
= dev_net(port
->dev
);
157 skb
= nlmsg_new(br_nlmsg_size(), GFP_KERNEL
);
161 err
= dp_fill_ifinfo(skb
, port
, event
, 0);
163 /* -EMSGSIZE implies BUG in br_nlmsg_size() */
164 WARN_ON(err
== -EMSGSIZE
);
168 err
= rtnl_notify(skb
, net
, 0, RTNLGRP_LINK
, NULL
, GFP_KERNEL
);
171 rtnl_set_sk_err(net
, RTNLGRP_LINK
, err
);
174 static void release_dp(struct kobject
*kobj
)
176 struct datapath
*dp
= container_of(kobj
, struct datapath
, ifobj
);
180 struct kobj_type dp_ktype
= {
181 .release
= release_dp
184 static int create_dp(int dp_idx
, const char __user
*devnamep
)
186 struct net_device
*dp_dev
;
187 char devname
[IFNAMSIZ
];
194 if (strncpy_from_user(devname
, devnamep
, IFNAMSIZ
- 1) < 0)
196 devname
[IFNAMSIZ
- 1] = '\0';
198 snprintf(devname
, sizeof devname
, "of%d", dp_idx
);
202 mutex_lock(&dp_mutex
);
204 if (!try_module_get(THIS_MODULE
))
207 /* Exit early if a datapath with that number already exists.
208 * (We don't use -EEXIST because that's ambiguous with 'devname'
209 * conflicting with an existing network device name.) */
215 dp
= kzalloc(sizeof *dp
, GFP_KERNEL
);
218 INIT_LIST_HEAD(&dp
->port_list
);
219 mutex_init(&dp
->mutex
);
221 for (i
= 0; i
< DP_N_QUEUES
; i
++)
222 skb_queue_head_init(&dp
->queues
[i
]);
223 init_waitqueue_head(&dp
->waitqueue
);
225 /* Initialize kobject for bridge. This will be added as
226 * /sys/class/net/<devname>/brif later, if sysfs is enabled. */
227 kobject_set_name(&dp
->ifobj
, SYSFS_BRIDGE_PORT_SUBDIR
); /* "brif" */
228 dp
->ifobj
.kset
= NULL
;
229 dp
->ifobj
.parent
= NULL
;
230 kobject_init(&dp
->ifobj
, &dp_ktype
);
232 /* Allocate table. */
234 rcu_assign_pointer(dp
->table
, dp_table_create(DP_L1_SIZE
));
238 /* Setup our datapath device */
239 dp_dev
= dp_dev_create(dp
, devname
, ODPP_LOCAL
);
240 err
= PTR_ERR(dp_dev
);
242 goto err_destroy_table
;
244 err
= new_nbp(dp
, dp_dev
, ODPP_LOCAL
);
246 dp_dev_destroy(dp_dev
);
247 goto err_destroy_table
;
251 dp
->stats_percpu
= alloc_percpu(struct dp_stats_percpu
);
252 if (!dp
->stats_percpu
)
253 goto err_destroy_local_port
;
255 rcu_assign_pointer(dps
[dp_idx
], dp
);
256 mutex_unlock(&dp_mutex
);
265 err_destroy_local_port
:
266 dp_del_port(dp
->ports
[ODPP_LOCAL
]);
268 dp_table_destroy(dp
->table
, 0);
272 module_put(THIS_MODULE
);
274 mutex_unlock(&dp_mutex
);
280 static void do_destroy_dp(struct datapath
*dp
)
282 struct net_bridge_port
*p
, *n
;
285 list_for_each_entry_safe (p
, n
, &dp
->port_list
, node
)
286 if (p
->port_no
!= ODPP_LOCAL
)
293 rcu_assign_pointer(dps
[dp
->dp_idx
], NULL
);
295 dp_del_port(dp
->ports
[ODPP_LOCAL
]);
297 dp_table_destroy(dp
->table
, 1);
299 for (i
= 0; i
< DP_N_QUEUES
; i
++)
300 skb_queue_purge(&dp
->queues
[i
]);
301 for (i
= 0; i
< DP_MAX_GROUPS
; i
++)
302 kfree(dp
->groups
[i
]);
303 free_percpu(dp
->stats_percpu
);
304 kobject_put(&dp
->ifobj
);
305 module_put(THIS_MODULE
);
308 static int destroy_dp(int dp_idx
)
314 mutex_lock(&dp_mutex
);
324 mutex_unlock(&dp_mutex
);
329 static void release_nbp(struct kobject
*kobj
)
331 struct net_bridge_port
*p
= container_of(kobj
, struct net_bridge_port
, kobj
);
335 struct kobj_type brport_ktype
= {
337 .sysfs_ops
= &brport_sysfs_ops
,
339 .release
= release_nbp
342 /* Called with RTNL lock and dp_mutex. */
343 static int new_nbp(struct datapath
*dp
, struct net_device
*dev
, int port_no
)
345 struct net_bridge_port
*p
;
347 if (dev
->br_port
!= NULL
)
350 p
= kzalloc(sizeof(*p
), GFP_KERNEL
);
354 dev_set_promiscuity(dev
, 1);
356 p
->port_no
= port_no
;
360 rcu_assign_pointer(dev
->br_port
, p
);
362 /* It would make sense to assign dev->br_port here too, but
363 * that causes packets received on internal ports to get caught
364 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
365 * back to network stack, but that's a waste of time. */
367 rcu_assign_pointer(dp
->ports
[port_no
], p
);
368 list_add_rcu(&p
->node
, &dp
->port_list
);
371 /* Initialize kobject for bridge. This will be added as
372 * /sys/class/net/<devname>/brport later, if sysfs is enabled. */
373 kobject_set_name(&p
->kobj
, SYSFS_BRIDGE_PORT_ATTR
); /* "brport" */
375 p
->kobj
.parent
= &p
->dev
->NETDEV_DEV_MEMBER
.kobj
;
376 kobject_init(&p
->kobj
, &brport_ktype
);
378 dp_ifinfo_notify(RTM_NEWLINK
, p
);
383 static int add_port(int dp_idx
, struct odp_port __user
*portp
)
385 struct net_device
*dev
;
387 struct odp_port port
;
392 if (copy_from_user(&port
, portp
, sizeof port
))
394 port
.devname
[IFNAMSIZ
- 1] = '\0';
398 if (port_no
< 0 || port_no
>= DP_MAX_PORTS
)
402 dp
= get_dp_locked(dp_idx
);
405 goto out_unlock_rtnl
;
408 if (dp
->ports
[port_no
])
411 if (!(port
.flags
& ODP_PORT_INTERNAL
)) {
413 dev
= dev_get_by_name(&init_net
, port
.devname
);
418 if (dev
->flags
& IFF_LOOPBACK
|| dev
->type
!= ARPHRD_ETHER
||
422 dev
= dp_dev_create(dp
, port
.devname
, port_no
);
429 err
= new_nbp(dp
, dev
, port_no
);
434 dp_sysfs_add_if(dp
->ports
[port_no
]);
440 mutex_unlock(&dp
->mutex
);
447 int dp_del_port(struct net_bridge_port
*p
)
452 if (p
->port_no
!= ODPP_LOCAL
)
455 dp_ifinfo_notify(RTM_DELLINK
, p
);
459 if (is_dp_dev(p
->dev
)) {
460 /* Make sure that no packets arrive from now on, since
461 * dp_dev_xmit() will try to find itself through
462 * p->dp->ports[], and we're about to set that to null. */
463 netif_tx_disable(p
->dev
);
466 /* First drop references to device. */
467 dev_set_promiscuity(p
->dev
, -1);
468 list_del_rcu(&p
->node
);
469 rcu_assign_pointer(p
->dp
->ports
[p
->port_no
], NULL
);
470 rcu_assign_pointer(p
->dev
->br_port
, NULL
);
472 /* Then wait until no one is still using it, and destroy it. */
475 if (is_dp_dev(p
->dev
))
476 dp_dev_destroy(p
->dev
);
478 kobject_put(&p
->kobj
);
483 static int del_port(int dp_idx
, int port_no
)
485 struct net_bridge_port
*p
;
491 if (port_no
< 0 || port_no
>= DP_MAX_PORTS
|| port_no
== ODPP_LOCAL
)
495 dp
= get_dp_locked(dp_idx
);
498 goto out_unlock_rtnl
;
500 p
= dp
->ports
[port_no
];
505 err
= dp_del_port(p
);
508 mutex_unlock(&dp
->mutex
);
515 /* Must be called with rcu_read_lock. */
517 do_port_input(struct net_bridge_port
*p
, struct sk_buff
*skb
)
519 /* Make our own copy of the packet. Otherwise we will mangle the
520 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
521 * (No one comes after us, since we tell handle_bridge() that we took
523 skb
= skb_share_check(skb
, GFP_ATOMIC
);
527 /* Push the Ethernet header back on. */
528 skb_push(skb
, ETH_HLEN
);
529 skb_reset_mac_header(skb
);
530 dp_process_received_packet(skb
, p
);
533 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
534 void dp_process_received_packet(struct sk_buff
*skb
, struct net_bridge_port
*p
)
536 struct datapath
*dp
= p
->dp
;
537 struct dp_stats_percpu
*stats
;
538 struct odp_flow_key key
;
539 struct sw_flow
*flow
;
541 WARN_ON_ONCE(skb_shared(skb
));
543 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
544 stats
= percpu_ptr(dp
->stats_percpu
, smp_processor_id());
546 if (flow_extract(skb
, p
? p
->port_no
: ODPP_NONE
, &key
)) {
547 if (dp
->drop_frags
) {
554 flow
= dp_table_lookup(rcu_dereference(dp
->table
), &key
);
556 struct sw_flow_actions
*acts
= rcu_dereference(flow
->sf_acts
);
557 flow_used(flow
, skb
);
558 execute_actions(dp
, skb
, &key
, acts
->actions
, acts
->n_actions
,
563 dp_output_control(dp
, skb
, _ODPL_MISS_NR
, 0);
568 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
569 * different set of devices!)
571 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
572 /* Called with rcu_read_lock and bottom-halves disabled. */
573 static struct sk_buff
*dp_frame_hook(struct net_bridge_port
*p
,
576 do_port_input(p
, skb
);
579 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
580 /* Called with rcu_read_lock and bottom-halves disabled. */
581 static int dp_frame_hook(struct net_bridge_port
*p
, struct sk_buff
**pskb
)
583 do_port_input(p
, *pskb
);
591 /* This code is copied verbatim from net/dev/core.c in Xen's
592 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
593 * directly because they aren't exported. */
594 static int skb_pull_up_to(struct sk_buff
*skb
, void *ptr
)
596 if (ptr
< (void *)skb
->tail
)
598 if (__pskb_pull_tail(skb
,
599 ptr
- (void *)skb
->data
- skb_headlen(skb
))) {
606 int skb_checksum_setup(struct sk_buff
*skb
)
608 if (skb
->proto_csum_blank
) {
609 if (skb
->protocol
!= htons(ETH_P_IP
))
611 if (!skb_pull_up_to(skb
, skb
->nh
.iph
+ 1))
613 skb
->h
.raw
= (unsigned char *)skb
->nh
.iph
+ 4*skb
->nh
.iph
->ihl
;
614 switch (skb
->nh
.iph
->protocol
) {
616 skb
->csum
= offsetof(struct tcphdr
, check
);
619 skb
->csum
= offsetof(struct udphdr
, check
);
623 printk(KERN_ERR
"Attempting to checksum a non-"
624 "TCP/UDP packet, dropping a protocol"
625 " %d packet", skb
->nh
.iph
->protocol
);
628 if (!skb_pull_up_to(skb
, skb
->h
.raw
+ skb
->csum
+ 2))
630 skb
->ip_summed
= CHECKSUM_HW
;
631 skb
->proto_csum_blank
= 0;
640 dp_output_control(struct datapath
*dp
, struct sk_buff
*skb
, int queue_no
,
643 struct dp_stats_percpu
*stats
;
644 struct sk_buff_head
*queue
;
648 WARN_ON_ONCE(skb_shared(skb
));
649 BUG_ON(queue_no
!= _ODPL_MISS_NR
&& queue_no
!= _ODPL_ACTION_NR
);
651 queue
= &dp
->queues
[queue_no
];
653 if (skb_queue_len(queue
) >= DP_MAX_QUEUE_LEN
)
656 /* If a checksum-deferred packet is forwarded to the controller,
657 * correct the pointers and checksum. This happens on a regular basis
658 * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets
659 * that do not have their checksum computed. We also implement it for
660 * the non-Xen case, but it is difficult to trigger or test this case
661 * there, hence the WARN_ON_ONCE().
663 err
= skb_checksum_setup(skb
);
667 if (skb
->ip_summed
== CHECKSUM_PARTIAL
) {
669 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
670 /* Until 2.6.22, the start of the transport header was also the
671 * start of data to be checksummed. Linux 2.6.22 introduced
672 * the csum_start field for this purpose, but we should point
673 * the transport header to it anyway for backward
674 * compatibility, as dev_queue_xmit() does even in 2.6.28. */
675 skb_set_transport_header(skb
, skb
->csum_start
-
678 err
= skb_checksum_help(skb
);
683 if (skb
->ip_summed
== CHECKSUM_HW
) {
684 err
= skb_checksum_help(skb
, 0);
690 /* Break apart GSO packets into their component pieces. Otherwise
691 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
692 if (skb_is_gso(skb
)) {
693 struct sk_buff
*nskb
= skb_gso_segment(skb
, 0);
697 if (unlikely(IS_ERR(skb
))) {
702 /* XXX This case might not be possible. It's hard to
703 * tell from the skb_gso_segment() code and comment. */
707 /* Figure out port number. */
708 port_no
= ODPP_LOCAL
;
710 if (skb
->dev
->br_port
)
711 port_no
= skb
->dev
->br_port
->port_no
;
712 else if (is_dp_dev(skb
->dev
))
713 port_no
= dp_dev_priv(skb
->dev
)->port_no
;
716 /* Append each packet to queue. There will be only one packet unless
717 * we broke up a GSO packet above. */
719 struct odp_msg
*header
;
720 struct sk_buff
*nskb
= skb
->next
;
723 err
= skb_cow(skb
, sizeof *header
);
733 header
= (struct odp_msg
*)__skb_push(skb
, sizeof *header
);
734 header
->type
= queue_no
;
735 header
->length
= skb
->len
;
736 header
->port
= port_no
;
737 header
->reserved
= 0;
739 skb_queue_tail(queue
, skb
);
744 wake_up_interruptible(&dp
->waitqueue
);
750 stats
= percpu_ptr(dp
->stats_percpu
, get_cpu());
757 static int flush_flows(struct datapath
*dp
)
760 return dp_table_flush(dp
);
763 static int validate_actions(const struct sw_flow_actions
*actions
)
767 for (i
= 0; i
< actions
->n_actions
; i
++) {
768 const union odp_action
*a
= &actions
->actions
[i
];
771 if (a
->output
.port
>= DP_MAX_PORTS
)
775 case ODPAT_OUTPUT_GROUP
:
776 if (a
->output_group
.group
>= DP_MAX_GROUPS
)
780 case ODPAT_SET_VLAN_VID
:
781 if (a
->vlan_vid
.vlan_vid
& htons(~VLAN_VID_MASK
))
785 case ODPAT_SET_VLAN_PCP
:
786 if (a
->vlan_pcp
.vlan_pcp
787 & ~(VLAN_PCP_MASK
>> VLAN_PCP_SHIFT
))
792 if (a
->type
>= ODPAT_N_ACTIONS
)
801 static struct sw_flow_actions
*get_actions(const struct odp_flow
*flow
)
803 struct sw_flow_actions
*actions
;
806 actions
= flow_actions_alloc(flow
->n_actions
);
807 error
= PTR_ERR(actions
);
812 if (copy_from_user(actions
->actions
, flow
->actions
,
813 flow
->n_actions
* sizeof(union odp_action
)))
814 goto error_free_actions
;
815 error
= validate_actions(actions
);
817 goto error_free_actions
;
824 return ERR_PTR(error
);
827 static void get_stats(struct sw_flow
*flow
, struct odp_flow_stats
*stats
)
829 if (flow
->used
.tv_sec
) {
830 stats
->used_sec
= flow
->used
.tv_sec
;
831 stats
->used_nsec
= flow
->used
.tv_nsec
;
834 stats
->used_nsec
= 0;
836 stats
->n_packets
= flow
->packet_count
;
837 stats
->n_bytes
= flow
->byte_count
;
838 stats
->ip_tos
= flow
->ip_tos
;
839 stats
->tcp_flags
= flow
->tcp_flags
;
842 static void clear_stats(struct sw_flow
*flow
)
844 flow
->used
.tv_sec
= flow
->used
.tv_nsec
= 0;
847 flow
->packet_count
= 0;
848 flow
->byte_count
= 0;
851 static int put_flow(struct datapath
*dp
, struct odp_flow_put __user
*ufp
)
853 struct odp_flow_put uf
;
854 struct sw_flow
*flow
;
855 struct dp_table
*table
;
856 struct odp_flow_stats stats
;
860 if (copy_from_user(&uf
, ufp
, sizeof(struct odp_flow_put
)))
862 uf
.flow
.key
.reserved
= 0;
864 table
= rcu_dereference(dp
->table
);
865 flow
= dp_table_lookup(table
, &uf
.flow
.key
);
868 struct sw_flow_actions
*acts
;
871 if (!(uf
.flags
& ODPPF_CREATE
))
874 /* Expand table, if necessary, to make room. */
875 if (dp
->n_flows
>= table
->n_buckets
) {
877 if (table
->n_buckets
>= DP_MAX_BUCKETS
)
880 error
= dp_table_expand(dp
);
883 table
= rcu_dereference(dp
->table
);
888 flow
= kmem_cache_alloc(flow_cache
, GFP_KERNEL
);
891 flow
->key
= uf
.flow
.key
;
892 spin_lock_init(&flow
->lock
);
895 /* Obtain actions. */
896 acts
= get_actions(&uf
.flow
);
897 error
= PTR_ERR(acts
);
899 goto error_free_flow
;
900 rcu_assign_pointer(flow
->sf_acts
, acts
);
902 /* Put flow in bucket. */
903 error
= dp_table_insert(table
, flow
);
905 goto error_free_flow_acts
;
907 memset(&stats
, 0, sizeof(struct odp_flow_stats
));
909 /* We found a matching flow. */
910 struct sw_flow_actions
*old_acts
, *new_acts
;
911 unsigned long int flags
;
913 /* Bail out if we're not allowed to modify an existing flow. */
915 if (!(uf
.flags
& ODPPF_MODIFY
))
919 new_acts
= get_actions(&uf
.flow
);
920 error
= PTR_ERR(new_acts
);
921 if (IS_ERR(new_acts
))
923 old_acts
= rcu_dereference(flow
->sf_acts
);
924 if (old_acts
->n_actions
!= new_acts
->n_actions
||
925 memcmp(old_acts
->actions
, new_acts
->actions
,
926 sizeof(union odp_action
) * old_acts
->n_actions
)) {
927 rcu_assign_pointer(flow
->sf_acts
, new_acts
);
928 flow_deferred_free_acts(old_acts
);
933 /* Fetch stats, then clear them if necessary. */
934 spin_lock_irqsave(&flow
->lock
, flags
);
935 get_stats(flow
, &stats
);
936 if (uf
.flags
& ODPPF_ZERO_STATS
)
938 spin_unlock_irqrestore(&flow
->lock
, flags
);
941 /* Copy stats to userspace. */
942 if (__copy_to_user(&ufp
->flow
.stats
, &stats
,
943 sizeof(struct odp_flow_stats
)))
947 error_free_flow_acts
:
948 kfree(flow
->sf_acts
);
950 kmem_cache_free(flow_cache
, flow
);
955 static int put_actions(const struct sw_flow
*flow
, struct odp_flow __user
*ufp
)
957 union odp_action __user
*actions
;
958 struct sw_flow_actions
*sf_acts
;
961 if (__get_user(actions
, &ufp
->actions
) ||
962 __get_user(n_actions
, &ufp
->n_actions
))
967 if (ufp
->n_actions
> INT_MAX
/ sizeof(union odp_action
))
970 sf_acts
= rcu_dereference(flow
->sf_acts
);
971 if (__put_user(sf_acts
->n_actions
, &ufp
->n_actions
) ||
972 (actions
&& copy_to_user(actions
, sf_acts
->actions
,
973 sizeof(union odp_action
) *
974 min(sf_acts
->n_actions
, n_actions
))))
980 static int answer_query(struct sw_flow
*flow
, struct odp_flow __user
*ufp
)
982 struct odp_flow_stats stats
;
983 unsigned long int flags
;
985 spin_lock_irqsave(&flow
->lock
, flags
);
986 get_stats(flow
, &stats
);
987 spin_unlock_irqrestore(&flow
->lock
, flags
);
989 if (__copy_to_user(&ufp
->stats
, &stats
, sizeof(struct odp_flow_stats
)))
991 return put_actions(flow
, ufp
);
994 static int del_or_query_flow(struct datapath
*dp
,
995 struct odp_flow __user
*ufp
,
998 struct dp_table
*table
= rcu_dereference(dp
->table
);
1000 struct sw_flow
*flow
;
1004 if (copy_from_user(&uf
, ufp
, sizeof uf
))
1006 uf
.key
.reserved
= 0;
1008 flow
= dp_table_lookup(table
, &uf
.key
);
1013 if (cmd
== ODP_FLOW_DEL
) {
1014 /* XXX redundant lookup */
1015 error
= dp_table_delete(table
, flow
);
1019 /* XXX These statistics might lose a few packets, since other
1020 * CPUs can be using this flow. We used to synchronize_rcu()
1021 * to make sure that we get completely accurate stats, but that
1022 * blows our performance, badly. */
1024 error
= answer_query(flow
, ufp
);
1025 flow_deferred_free(flow
);
1027 error
= answer_query(flow
, ufp
);
1034 static int query_multiple_flows(struct datapath
*dp
,
1035 const struct odp_flowvec
*flowvec
)
1037 struct dp_table
*table
= rcu_dereference(dp
->table
);
1039 for (i
= 0; i
< flowvec
->n_flows
; i
++) {
1040 struct __user odp_flow
*ufp
= &flowvec
->flows
[i
];
1042 struct sw_flow
*flow
;
1045 if (__copy_from_user(&uf
, ufp
, sizeof uf
))
1047 uf
.key
.reserved
= 0;
1049 flow
= dp_table_lookup(table
, &uf
.key
);
1051 error
= __clear_user(&ufp
->stats
, sizeof ufp
->stats
);
1053 error
= answer_query(flow
, ufp
);
1057 return flowvec
->n_flows
;
1060 struct list_flows_cbdata
{
1061 struct odp_flow __user
*uflows
;
1066 static int list_flow(struct sw_flow
*flow
, void *cbdata_
)
1068 struct list_flows_cbdata
*cbdata
= cbdata_
;
1069 struct odp_flow __user
*ufp
= &cbdata
->uflows
[cbdata
->listed_flows
++];
1072 if (__copy_to_user(&ufp
->key
, &flow
->key
, sizeof flow
->key
))
1074 error
= answer_query(flow
, ufp
);
1078 if (cbdata
->listed_flows
>= cbdata
->n_flows
)
1079 return cbdata
->listed_flows
;
1083 static int list_flows(struct datapath
*dp
, const struct odp_flowvec
*flowvec
)
1085 struct list_flows_cbdata cbdata
;
1088 if (!flowvec
->n_flows
)
1091 cbdata
.uflows
= flowvec
->flows
;
1092 cbdata
.n_flows
= flowvec
->n_flows
;
1093 cbdata
.listed_flows
= 0;
1094 error
= dp_table_foreach(rcu_dereference(dp
->table
),
1095 list_flow
, &cbdata
);
1096 return error
? error
: cbdata
.listed_flows
;
1099 static int do_flowvec_ioctl(struct datapath
*dp
, unsigned long argp
,
1100 int (*function
)(struct datapath
*,
1101 const struct odp_flowvec
*))
1103 struct odp_flowvec __user
*uflowvec
;
1104 struct odp_flowvec flowvec
;
1107 uflowvec
= (struct odp_flowvec __user
*)argp
;
1108 if (!access_ok(VERIFY_WRITE
, uflowvec
, sizeof *uflowvec
) ||
1109 copy_from_user(&flowvec
, uflowvec
, sizeof flowvec
))
1112 if (flowvec
.n_flows
> INT_MAX
/ sizeof(struct odp_flow
))
1115 if (!access_ok(VERIFY_WRITE
, flowvec
.flows
,
1116 flowvec
.n_flows
* sizeof(struct odp_flow
)))
1119 retval
= function(dp
, &flowvec
);
1120 return (retval
< 0 ? retval
1121 : retval
== flowvec
.n_flows
? 0
1122 : __put_user(retval
, &uflowvec
->n_flows
));
1125 static int do_execute(struct datapath
*dp
, const struct odp_execute
*executep
)
1127 struct odp_execute execute
;
1128 struct odp_flow_key key
;
1129 struct sk_buff
*skb
;
1130 struct sw_flow_actions
*actions
;
1135 if (copy_from_user(&execute
, executep
, sizeof execute
))
1139 if (execute
.length
< ETH_HLEN
|| execute
.length
> 65535)
1143 actions
= flow_actions_alloc(execute
.n_actions
);
1148 if (copy_from_user(actions
->actions
, execute
.actions
,
1149 execute
.n_actions
* sizeof *execute
.actions
))
1150 goto error_free_actions
;
1152 err
= validate_actions(actions
);
1154 goto error_free_actions
;
1157 skb
= alloc_skb(execute
.length
, GFP_KERNEL
);
1159 goto error_free_actions
;
1160 if (execute
.in_port
< DP_MAX_PORTS
) {
1161 struct net_bridge_port
*p
= dp
->ports
[execute
.in_port
];
1167 if (copy_from_user(skb_put(skb
, execute
.length
), execute
.data
,
1169 goto error_free_skb
;
1171 skb_reset_mac_header(skb
);
1174 /* Normally, setting the skb 'protocol' field would be handled by a
1175 * call to eth_type_trans(), but it assumes there's a sending
1176 * device, which we may not have. */
1177 if (ntohs(eth
->h_proto
) >= 1536)
1178 skb
->protocol
= eth
->h_proto
;
1180 skb
->protocol
= htons(ETH_P_802_2
);
1182 flow_extract(skb
, execute
.in_port
, &key
);
1183 err
= execute_actions(dp
, skb
, &key
, actions
->actions
,
1184 actions
->n_actions
, GFP_KERNEL
);
1197 get_dp_stats(struct datapath
*dp
, struct odp_stats __user
*statsp
)
1199 struct odp_stats stats
;
1202 stats
.n_flows
= dp
->n_flows
;
1203 stats
.cur_capacity
= rcu_dereference(dp
->table
)->n_buckets
;
1204 stats
.max_capacity
= DP_MAX_BUCKETS
;
1205 stats
.n_ports
= dp
->n_ports
;
1206 stats
.max_ports
= DP_MAX_PORTS
;
1207 stats
.max_groups
= DP_MAX_GROUPS
;
1208 stats
.n_frags
= stats
.n_hit
= stats
.n_missed
= stats
.n_lost
= 0;
1209 for_each_possible_cpu(i
) {
1210 const struct dp_stats_percpu
*s
;
1211 s
= percpu_ptr(dp
->stats_percpu
, i
);
1212 stats
.n_frags
+= s
->n_frags
;
1213 stats
.n_hit
+= s
->n_hit
;
1214 stats
.n_missed
+= s
->n_missed
;
1215 stats
.n_lost
+= s
->n_lost
;
1217 stats
.max_miss_queue
= DP_MAX_QUEUE_LEN
;
1218 stats
.max_action_queue
= DP_MAX_QUEUE_LEN
;
1219 return copy_to_user(statsp
, &stats
, sizeof stats
) ? -EFAULT
: 0;
1222 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1223 int dp_min_mtu(const struct datapath
*dp
)
1225 struct net_bridge_port
*p
;
1230 list_for_each_entry_rcu (p
, &dp
->port_list
, node
) {
1231 struct net_device
*dev
= p
->dev
;
1233 /* Skip any internal ports, since that's what we're trying to
1238 if (!mtu
|| dev
->mtu
< mtu
)
1242 return mtu
? mtu
: ETH_DATA_LEN
;
1246 put_port(const struct net_bridge_port
*p
, struct odp_port __user
*uop
)
1249 memset(&op
, 0, sizeof op
);
1250 strncpy(op
.devname
, p
->dev
->name
, sizeof op
.devname
);
1251 op
.port
= p
->port_no
;
1252 op
.flags
= is_dp_dev(p
->dev
) ? ODP_PORT_INTERNAL
: 0;
1253 return copy_to_user(uop
, &op
, sizeof op
) ? -EFAULT
: 0;
1257 query_port(struct datapath
*dp
, struct odp_port __user
*uport
)
1259 struct odp_port port
;
1261 if (copy_from_user(&port
, uport
, sizeof port
))
1263 if (port
.devname
[0]) {
1264 struct net_bridge_port
*p
;
1265 struct net_device
*dev
;
1268 port
.devname
[IFNAMSIZ
- 1] = '\0';
1270 dev
= dev_get_by_name(&init_net
, port
.devname
);
1275 if (!p
&& is_dp_dev(dev
)) {
1276 struct dp_dev
*dp_dev
= dp_dev_priv(dev
);
1277 if (dp_dev
->dp
== dp
)
1278 p
= dp
->ports
[dp_dev
->port_no
];
1280 err
= p
&& p
->dp
== dp
? put_port(p
, uport
) : -ENOENT
;
1285 if (port
.port
>= DP_MAX_PORTS
)
1287 if (!dp
->ports
[port
.port
])
1289 return put_port(dp
->ports
[port
.port
], uport
);
1294 list_ports(struct datapath
*dp
, struct odp_portvec __user
*pvp
)
1296 struct odp_portvec pv
;
1297 struct net_bridge_port
*p
;
1300 if (copy_from_user(&pv
, pvp
, sizeof pv
))
1305 list_for_each_entry_rcu (p
, &dp
->port_list
, node
) {
1306 if (put_port(p
, &pv
.ports
[idx
]))
1308 if (idx
++ >= pv
.n_ports
)
1312 return put_user(idx
, &pvp
->n_ports
);
1315 /* RCU callback for freeing a dp_port_group */
1316 static void free_port_group(struct rcu_head
*rcu
)
1318 struct dp_port_group
*g
= container_of(rcu
, struct dp_port_group
, rcu
);
1323 set_port_group(struct datapath
*dp
, const struct odp_port_group __user
*upg
)
1325 struct odp_port_group pg
;
1326 struct dp_port_group
*new_group
, *old_group
;
1330 if (copy_from_user(&pg
, upg
, sizeof pg
))
1334 if (pg
.n_ports
> DP_MAX_PORTS
|| pg
.group
>= DP_MAX_GROUPS
)
1338 new_group
= kmalloc(sizeof *new_group
+ sizeof(u16
) * pg
.n_ports
,
1343 new_group
->n_ports
= pg
.n_ports
;
1345 if (copy_from_user(new_group
->ports
, pg
.ports
,
1346 sizeof(u16
) * pg
.n_ports
))
1349 old_group
= rcu_dereference(dp
->groups
[pg
.group
]);
1350 rcu_assign_pointer(dp
->groups
[pg
.group
], new_group
);
1352 call_rcu(&old_group
->rcu
, free_port_group
);
1362 get_port_group(struct datapath
*dp
, struct odp_port_group
*upg
)
1364 struct odp_port_group pg
;
1365 struct dp_port_group
*g
;
1368 if (copy_from_user(&pg
, upg
, sizeof pg
))
1371 if (pg
.group
>= DP_MAX_GROUPS
)
1374 g
= dp
->groups
[pg
.group
];
1375 n_copy
= g
? min_t(int, g
->n_ports
, pg
.n_ports
) : 0;
1376 if (n_copy
&& copy_to_user(pg
.ports
, g
->ports
, n_copy
* sizeof(u16
)))
1379 if (put_user(g
? g
->n_ports
: 0, &upg
->n_ports
))
1385 static int get_listen_mask(const struct file
*f
)
1387 return (long)f
->private_data
;
1390 static void set_listen_mask(struct file
*f
, int listen_mask
)
1392 f
->private_data
= (void*)(long)listen_mask
;
1395 static long openvswitch_ioctl(struct file
*f
, unsigned int cmd
,
1398 int dp_idx
= iminor(f
->f_dentry
->d_inode
);
1399 struct datapath
*dp
;
1400 int drop_frags
, listeners
, port_no
;
1403 /* Handle commands with special locking requirements up front. */
1406 return create_dp(dp_idx
, (char __user
*)argp
);
1408 case ODP_DP_DESTROY
:
1409 return destroy_dp(dp_idx
);
1412 return add_port(dp_idx
, (struct odp_port __user
*)argp
);
1415 err
= get_user(port_no
, (int __user
*)argp
);
1418 return del_port(dp_idx
, port_no
);
1421 dp
= get_dp_locked(dp_idx
);
1427 err
= get_dp_stats(dp
, (struct odp_stats __user
*)argp
);
1430 case ODP_GET_DROP_FRAGS
:
1431 err
= put_user(dp
->drop_frags
, (int __user
*)argp
);
1434 case ODP_SET_DROP_FRAGS
:
1435 err
= get_user(drop_frags
, (int __user
*)argp
);
1439 if (drop_frags
!= 0 && drop_frags
!= 1)
1441 dp
->drop_frags
= drop_frags
;
1445 case ODP_GET_LISTEN_MASK
:
1446 err
= put_user(get_listen_mask(f
), (int __user
*)argp
);
1449 case ODP_SET_LISTEN_MASK
:
1450 err
= get_user(listeners
, (int __user
*)argp
);
1454 if (listeners
& ~ODPL_ALL
)
1457 set_listen_mask(f
, listeners
);
1460 case ODP_PORT_QUERY
:
1461 err
= query_port(dp
, (struct odp_port __user
*)argp
);
1465 err
= list_ports(dp
, (struct odp_portvec __user
*)argp
);
1468 case ODP_PORT_GROUP_SET
:
1469 err
= set_port_group(dp
, (struct odp_port_group __user
*)argp
);
1472 case ODP_PORT_GROUP_GET
:
1473 err
= get_port_group(dp
, (struct odp_port_group __user
*)argp
);
1476 case ODP_FLOW_FLUSH
:
1477 err
= flush_flows(dp
);
1481 err
= put_flow(dp
, (struct odp_flow_put __user
*)argp
);
1486 err
= del_or_query_flow(dp
, (struct odp_flow __user
*)argp
,
1490 case ODP_FLOW_GET_MULTIPLE
:
1491 err
= do_flowvec_ioctl(dp
, argp
, query_multiple_flows
);
1495 err
= do_flowvec_ioctl(dp
, argp
, list_flows
);
1499 err
= do_execute(dp
, (struct odp_execute __user
*)argp
);
1506 mutex_unlock(&dp
->mutex
);
1510 static int dp_has_packet_of_interest(struct datapath
*dp
, int listeners
)
1513 for (i
= 0; i
< DP_N_QUEUES
; i
++) {
1514 if (listeners
& (1 << i
) && !skb_queue_empty(&dp
->queues
[i
]))
1520 ssize_t
openvswitch_read(struct file
*f
, char __user
*buf
, size_t nbytes
,
1523 /* XXX is there sufficient synchronization here? */
1524 int listeners
= get_listen_mask(f
);
1525 int dp_idx
= iminor(f
->f_dentry
->d_inode
);
1526 struct datapath
*dp
= get_dp(dp_idx
);
1527 struct sk_buff
*skb
;
1528 struct iovec __user iov
;
1535 if (nbytes
== 0 || !listeners
)
1541 for (i
= 0; i
< DP_N_QUEUES
; i
++) {
1542 if (listeners
& (1 << i
)) {
1543 skb
= skb_dequeue(&dp
->queues
[i
]);
1549 if (f
->f_flags
& O_NONBLOCK
) {
1554 wait_event_interruptible(dp
->waitqueue
,
1555 dp_has_packet_of_interest(dp
,
1558 if (signal_pending(current
)) {
1559 retval
= -ERESTARTSYS
;
1564 copy_bytes
= min_t(size_t, skb
->len
, nbytes
);
1566 iov
.iov_len
= copy_bytes
;
1567 retval
= skb_copy_datagram_iovec(skb
, 0, &iov
, iov
.iov_len
);
1569 retval
= copy_bytes
;
1576 static unsigned int openvswitch_poll(struct file
*file
, poll_table
*wait
)
1578 /* XXX is there sufficient synchronization here? */
1579 int dp_idx
= iminor(file
->f_dentry
->d_inode
);
1580 struct datapath
*dp
= get_dp(dp_idx
);
1585 poll_wait(file
, &dp
->waitqueue
, wait
);
1586 if (dp_has_packet_of_interest(dp
, get_listen_mask(file
)))
1587 mask
|= POLLIN
| POLLRDNORM
;
1589 mask
= POLLIN
| POLLRDNORM
| POLLHUP
;
1594 struct file_operations openvswitch_fops
= {
1595 /* XXX .aio_read = openvswitch_aio_read, */
1596 .read
= openvswitch_read
,
1597 .poll
= openvswitch_poll
,
1598 .unlocked_ioctl
= openvswitch_ioctl
,
1599 /* XXX .fasync = openvswitch_fasync, */
1603 static struct llc_sap
*dp_stp_sap
;
1605 static int dp_stp_rcv(struct sk_buff
*skb
, struct net_device
*dev
,
1606 struct packet_type
*pt
, struct net_device
*orig_dev
)
1608 /* We don't really care about STP packets, we just listen for them for
1609 * mutual exclusion with the bridge module, so this just discards
1615 static int __init
dp_init(void)
1619 printk("Open vSwitch %s, built "__DATE__
" "__TIME__
"\n", VERSION BUILDNR
);
1621 /* Register to receive STP packets because the bridge module also
1622 * attempts to do so. Since there can only be a single listener for a
1623 * given protocol, this provides mutual exclusion against the bridge
1624 * module, preventing both of them from being loaded at the same
1626 dp_stp_sap
= llc_sap_open(LLC_SAP_BSPAN
, dp_stp_rcv
);
1628 printk(KERN_ERR
"openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1636 err
= register_netdevice_notifier(&dp_device_notifier
);
1638 goto error_flow_exit
;
1640 major
= register_chrdev(0, "openvswitch", &openvswitch_fops
);
1642 goto error_unreg_notifier
;
1644 /* Hook into callback used by the bridge to intercept packets.
1645 * Parasites we are. */
1646 br_handle_frame_hook
= dp_frame_hook
;
1650 error_unreg_notifier
:
1651 unregister_netdevice_notifier(&dp_device_notifier
);
1658 static void dp_cleanup(void)
1661 unregister_chrdev(major
, "openvswitch");
1662 unregister_netdevice_notifier(&dp_device_notifier
);
1664 br_handle_frame_hook
= NULL
;
1665 llc_sap_put(dp_stp_sap
);
1668 module_init(dp_init
);
1669 module_exit(dp_cleanup
);
1671 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1672 MODULE_LICENSE("GPL");