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 int create_dp(int dp_idx
, const char __user
*devnamep
)
176 struct net_device
*dp_dev
;
177 char devname
[IFNAMSIZ
];
184 if (strncpy_from_user(devname
, devnamep
, IFNAMSIZ
- 1) < 0)
186 devname
[IFNAMSIZ
- 1] = '\0';
188 snprintf(devname
, sizeof devname
, "of%d", dp_idx
);
192 mutex_lock(&dp_mutex
);
194 if (!try_module_get(THIS_MODULE
))
197 /* Exit early if a datapath with that number already exists.
198 * (We don't use -EEXIST because that's ambiguous with 'devname'
199 * conflicting with an existing network device name.) */
205 dp
= kzalloc(sizeof *dp
, GFP_KERNEL
);
208 INIT_LIST_HEAD(&dp
->port_list
);
209 mutex_init(&dp
->mutex
);
211 for (i
= 0; i
< DP_N_QUEUES
; i
++)
212 skb_queue_head_init(&dp
->queues
[i
]);
213 init_waitqueue_head(&dp
->waitqueue
);
215 /* Allocate table. */
217 rcu_assign_pointer(dp
->table
, dp_table_create(DP_L1_SIZE
));
221 /* Setup our datapath device */
222 dp_dev
= dp_dev_create(dp
, devname
, ODPP_LOCAL
);
223 err
= PTR_ERR(dp_dev
);
225 goto err_destroy_table
;
227 err
= new_nbp(dp
, dp_dev
, ODPP_LOCAL
);
229 dp_dev_destroy(dp_dev
);
230 goto err_destroy_table
;
234 dp
->stats_percpu
= alloc_percpu(struct dp_stats_percpu
);
235 if (!dp
->stats_percpu
)
236 goto err_destroy_local_port
;
238 rcu_assign_pointer(dps
[dp_idx
], dp
);
239 mutex_unlock(&dp_mutex
);
248 err_destroy_local_port
:
249 dp_del_port(dp
->ports
[ODPP_LOCAL
]);
251 dp_table_destroy(dp
->table
, 0);
255 module_put(THIS_MODULE
);
257 mutex_unlock(&dp_mutex
);
263 static void do_destroy_dp(struct datapath
*dp
)
265 struct net_bridge_port
*p
, *n
;
268 list_for_each_entry_safe (p
, n
, &dp
->port_list
, node
)
269 if (p
->port_no
!= ODPP_LOCAL
)
276 rcu_assign_pointer(dps
[dp
->dp_idx
], NULL
);
278 dp_del_port(dp
->ports
[ODPP_LOCAL
]);
280 dp_table_destroy(dp
->table
, 1);
282 for (i
= 0; i
< DP_N_QUEUES
; i
++)
283 skb_queue_purge(&dp
->queues
[i
]);
284 for (i
= 0; i
< DP_MAX_GROUPS
; i
++)
285 kfree(dp
->groups
[i
]);
286 free_percpu(dp
->stats_percpu
);
288 module_put(THIS_MODULE
);
291 static int destroy_dp(int dp_idx
)
297 mutex_lock(&dp_mutex
);
307 mutex_unlock(&dp_mutex
);
312 /* Called with RTNL lock and dp_mutex. */
313 static int new_nbp(struct datapath
*dp
, struct net_device
*dev
, int port_no
)
315 struct net_bridge_port
*p
;
317 if (dev
->br_port
!= NULL
)
320 p
= kzalloc(sizeof(*p
), GFP_KERNEL
);
324 dev_set_promiscuity(dev
, 1);
326 p
->port_no
= port_no
;
330 rcu_assign_pointer(dev
->br_port
, p
);
332 /* It would make sense to assign dev->br_port here too, but
333 * that causes packets received on internal ports to get caught
334 * in dp_frame_hook(). In turn dp_frame_hook() can reject them
335 * back to network stack, but that's a waste of time. */
337 rcu_assign_pointer(dp
->ports
[port_no
], p
);
338 list_add_rcu(&p
->node
, &dp
->port_list
);
341 dp_ifinfo_notify(RTM_NEWLINK
, p
);
346 static int add_port(int dp_idx
, struct odp_port __user
*portp
)
348 struct net_device
*dev
;
350 struct odp_port port
;
355 if (copy_from_user(&port
, portp
, sizeof port
))
357 port
.devname
[IFNAMSIZ
- 1] = '\0';
361 if (port_no
< 0 || port_no
>= DP_MAX_PORTS
)
365 dp
= get_dp_locked(dp_idx
);
368 goto out_unlock_rtnl
;
371 if (dp
->ports
[port_no
])
374 if (!(port
.flags
& ODP_PORT_INTERNAL
)) {
376 dev
= dev_get_by_name(&init_net
, port
.devname
);
381 if (dev
->flags
& IFF_LOOPBACK
|| dev
->type
!= ARPHRD_ETHER
||
385 dev
= dp_dev_create(dp
, port
.devname
, port_no
);
392 err
= new_nbp(dp
, dev
, port_no
);
397 dp_sysfs_add_if(dp
->ports
[port_no
]);
403 mutex_unlock(&dp
->mutex
);
410 int dp_del_port(struct net_bridge_port
*p
)
415 if (p
->port_no
!= ODPP_LOCAL
)
416 sysfs_remove_link(&p
->dp
->ifobj
, p
->dev
->name
);
418 dp_ifinfo_notify(RTM_DELLINK
, p
);
422 if (is_dp_dev(p
->dev
)) {
423 /* Make sure that no packets arrive from now on, since
424 * dp_dev_xmit() will try to find itself through
425 * p->dp->ports[], and we're about to set that to null. */
426 netif_tx_disable(p
->dev
);
429 /* First drop references to device. */
430 dev_set_promiscuity(p
->dev
, -1);
431 list_del_rcu(&p
->node
);
432 rcu_assign_pointer(p
->dp
->ports
[p
->port_no
], NULL
);
433 rcu_assign_pointer(p
->dev
->br_port
, NULL
);
435 /* Then wait until no one is still using it, and destroy it. */
438 if (is_dp_dev(p
->dev
)) {
439 dp_dev_destroy(p
->dev
);
441 if (p
->port_no
!= ODPP_LOCAL
) {
451 static int del_port(int dp_idx
, int port_no
)
453 struct net_bridge_port
*p
;
459 if (port_no
< 0 || port_no
>= DP_MAX_PORTS
|| port_no
== ODPP_LOCAL
)
463 dp
= get_dp_locked(dp_idx
);
466 goto out_unlock_rtnl
;
468 p
= dp
->ports
[port_no
];
473 err
= dp_del_port(p
);
476 mutex_unlock(&dp
->mutex
);
483 /* Must be called with rcu_read_lock. */
485 do_port_input(struct net_bridge_port
*p
, struct sk_buff
*skb
)
487 /* Make our own copy of the packet. Otherwise we will mangle the
488 * packet for anyone who came before us (e.g. tcpdump via AF_PACKET).
489 * (No one comes after us, since we tell handle_bridge() that we took
491 skb
= skb_share_check(skb
, GFP_ATOMIC
);
495 /* Push the Ethernet header back on. */
496 skb_push(skb
, ETH_HLEN
);
497 skb_reset_mac_header(skb
);
498 dp_process_received_packet(skb
, p
);
501 /* Must be called with rcu_read_lock and with bottom-halves disabled. */
502 void dp_process_received_packet(struct sk_buff
*skb
, struct net_bridge_port
*p
)
504 struct datapath
*dp
= p
->dp
;
505 struct dp_stats_percpu
*stats
;
506 struct odp_flow_key key
;
507 struct sw_flow
*flow
;
509 WARN_ON_ONCE(skb_shared(skb
));
511 /* BHs are off so we don't have to use get_cpu()/put_cpu() here. */
512 stats
= percpu_ptr(dp
->stats_percpu
, smp_processor_id());
514 if (flow_extract(skb
, p
? p
->port_no
: ODPP_NONE
, &key
)) {
515 if (dp
->drop_frags
) {
522 flow
= dp_table_lookup(rcu_dereference(dp
->table
), &key
);
524 struct sw_flow_actions
*acts
= rcu_dereference(flow
->sf_acts
);
525 flow_used(flow
, skb
);
526 execute_actions(dp
, skb
, &key
, acts
->actions
, acts
->n_actions
,
531 dp_output_control(dp
, skb
, _ODPL_MISS_NR
, 0);
536 * Used as br_handle_frame_hook. (Cannot run bridge at the same time, even on
537 * different set of devices!)
539 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
540 /* Called with rcu_read_lock and bottom-halves disabled. */
541 static struct sk_buff
*dp_frame_hook(struct net_bridge_port
*p
,
544 do_port_input(p
, skb
);
547 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
548 /* Called with rcu_read_lock and bottom-halves disabled. */
549 static int dp_frame_hook(struct net_bridge_port
*p
, struct sk_buff
**pskb
)
551 do_port_input(p
, *pskb
);
559 /* This code is copied verbatim from net/dev/core.c in Xen's
560 * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions
561 * directly because they aren't exported. */
562 static int skb_pull_up_to(struct sk_buff
*skb
, void *ptr
)
564 if (ptr
< (void *)skb
->tail
)
566 if (__pskb_pull_tail(skb
,
567 ptr
- (void *)skb
->data
- skb_headlen(skb
))) {
574 int skb_checksum_setup(struct sk_buff
*skb
)
576 if (skb
->proto_csum_blank
) {
577 if (skb
->protocol
!= htons(ETH_P_IP
))
579 if (!skb_pull_up_to(skb
, skb
->nh
.iph
+ 1))
581 skb
->h
.raw
= (unsigned char *)skb
->nh
.iph
+ 4*skb
->nh
.iph
->ihl
;
582 switch (skb
->nh
.iph
->protocol
) {
584 skb
->csum
= offsetof(struct tcphdr
, check
);
587 skb
->csum
= offsetof(struct udphdr
, check
);
591 printk(KERN_ERR
"Attempting to checksum a non-"
592 "TCP/UDP packet, dropping a protocol"
593 " %d packet", skb
->nh
.iph
->protocol
);
596 if (!skb_pull_up_to(skb
, skb
->h
.raw
+ skb
->csum
+ 2))
598 skb
->ip_summed
= CHECKSUM_HW
;
599 skb
->proto_csum_blank
= 0;
608 dp_output_control(struct datapath
*dp
, struct sk_buff
*skb
, int queue_no
,
611 struct dp_stats_percpu
*stats
;
612 struct sk_buff_head
*queue
;
616 WARN_ON_ONCE(skb_shared(skb
));
617 BUG_ON(queue_no
!= _ODPL_MISS_NR
&& queue_no
!= _ODPL_ACTION_NR
);
619 queue
= &dp
->queues
[queue_no
];
621 if (skb_queue_len(queue
) >= DP_MAX_QUEUE_LEN
)
624 /* If a checksum-deferred packet is forwarded to the controller,
625 * correct the pointers and checksum. This happens on a regular basis
626 * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets
627 * that do not have their checksum computed. We also implement it for
628 * the non-Xen case, but it is difficult to trigger or test this case
629 * there, hence the WARN_ON_ONCE().
631 err
= skb_checksum_setup(skb
);
635 if (skb
->ip_summed
== CHECKSUM_PARTIAL
) {
637 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
638 /* Until 2.6.22, the start of the transport header was also the
639 * start of data to be checksummed. Linux 2.6.22 introduced
640 * the csum_start field for this purpose, but we should point
641 * the transport header to it anyway for backward
642 * compatibility, as dev_queue_xmit() does even in 2.6.28. */
643 skb_set_transport_header(skb
, skb
->csum_start
-
646 err
= skb_checksum_help(skb
);
651 if (skb
->ip_summed
== CHECKSUM_HW
) {
652 err
= skb_checksum_help(skb
, 0);
658 /* Break apart GSO packets into their component pieces. Otherwise
659 * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */
660 if (skb_is_gso(skb
)) {
661 struct sk_buff
*nskb
= skb_gso_segment(skb
, 0);
665 if (unlikely(IS_ERR(skb
))) {
670 /* XXX This case might not be possible. It's hard to
671 * tell from the skb_gso_segment() code and comment. */
675 /* Figure out port number. */
676 port_no
= ODPP_LOCAL
;
678 if (skb
->dev
->br_port
)
679 port_no
= skb
->dev
->br_port
->port_no
;
680 else if (is_dp_dev(skb
->dev
))
681 port_no
= dp_dev_priv(skb
->dev
)->port_no
;
684 /* Append each packet to queue. There will be only one packet unless
685 * we broke up a GSO packet above. */
687 struct odp_msg
*header
;
688 struct sk_buff
*nskb
= skb
->next
;
691 err
= skb_cow(skb
, sizeof *header
);
701 header
= (struct odp_msg
*)__skb_push(skb
, sizeof *header
);
702 header
->type
= queue_no
;
703 header
->length
= skb
->len
;
704 header
->port
= port_no
;
705 header
->reserved
= 0;
707 skb_queue_tail(queue
, skb
);
712 wake_up_interruptible(&dp
->waitqueue
);
718 stats
= percpu_ptr(dp
->stats_percpu
, get_cpu());
725 static int flush_flows(struct datapath
*dp
)
728 return dp_table_flush(dp
);
731 static int validate_actions(const struct sw_flow_actions
*actions
)
735 for (i
= 0; i
< actions
->n_actions
; i
++) {
736 const union odp_action
*a
= &actions
->actions
[i
];
739 if (a
->output
.port
>= DP_MAX_PORTS
)
743 case ODPAT_OUTPUT_GROUP
:
744 if (a
->output_group
.group
>= DP_MAX_GROUPS
)
748 case ODPAT_SET_VLAN_VID
:
749 if (a
->vlan_vid
.vlan_vid
& htons(~VLAN_VID_MASK
))
753 case ODPAT_SET_VLAN_PCP
:
754 if (a
->vlan_pcp
.vlan_pcp
& ~VLAN_PCP_MASK
)
759 if (a
->type
>= ODPAT_N_ACTIONS
)
768 static struct sw_flow_actions
*get_actions(const struct odp_flow
*flow
)
770 struct sw_flow_actions
*actions
;
773 actions
= flow_actions_alloc(flow
->n_actions
);
774 error
= PTR_ERR(actions
);
779 if (copy_from_user(actions
->actions
, flow
->actions
,
780 flow
->n_actions
* sizeof(union odp_action
)))
781 goto error_free_actions
;
782 error
= validate_actions(actions
);
784 goto error_free_actions
;
791 return ERR_PTR(error
);
794 static void get_stats(struct sw_flow
*flow
, struct odp_flow_stats
*stats
)
796 if (flow
->used
.tv_sec
) {
797 stats
->used_sec
= flow
->used
.tv_sec
;
798 stats
->used_nsec
= flow
->used
.tv_nsec
;
801 stats
->used_nsec
= 0;
803 stats
->n_packets
= flow
->packet_count
;
804 stats
->n_bytes
= flow
->byte_count
;
805 stats
->ip_tos
= flow
->ip_tos
;
806 stats
->tcp_flags
= flow
->tcp_flags
;
809 static void clear_stats(struct sw_flow
*flow
)
811 flow
->used
.tv_sec
= flow
->used
.tv_nsec
= 0;
814 flow
->packet_count
= 0;
815 flow
->byte_count
= 0;
818 static int put_flow(struct datapath
*dp
, struct odp_flow_put __user
*ufp
)
820 struct odp_flow_put uf
;
821 struct sw_flow
*flow
, **bucket
;
822 struct dp_table
*table
;
823 struct odp_flow_stats stats
;
827 if (copy_from_user(&uf
, ufp
, sizeof(struct odp_flow_put
)))
829 uf
.flow
.key
.reserved
= 0;
832 table
= rcu_dereference(dp
->table
);
833 bucket
= dp_table_lookup_for_insert(table
, &uf
.flow
.key
);
835 /* No such flow, and the slots where it could go are full. */
836 error
= uf
.flags
& ODPPF_CREATE
? -EXFULL
: -ENOENT
;
838 } else if (!*bucket
) {
839 /* No such flow, but we found an available slot for it. */
840 struct sw_flow_actions
*acts
;
843 if (!(uf
.flags
& ODPPF_CREATE
))
846 /* Expand table, if necessary, to make room. */
847 if (dp
->n_flows
* 4 >= table
->n_buckets
&&
848 table
->n_buckets
< DP_MAX_BUCKETS
) {
849 error
= dp_table_expand(dp
);
853 /* The bucket's location has changed. Try again. */
859 flow
= kmem_cache_alloc(flow_cache
, GFP_KERNEL
);
862 flow
->key
= uf
.flow
.key
;
863 spin_lock_init(&flow
->lock
);
866 /* Obtain actions. */
867 acts
= get_actions(&uf
.flow
);
868 error
= PTR_ERR(acts
);
870 goto error_free_flow
;
871 rcu_assign_pointer(flow
->sf_acts
, acts
);
873 /* Put flow in bucket. */
874 rcu_assign_pointer(*bucket
, flow
);
876 memset(&stats
, 0, sizeof(struct odp_flow_stats
));
878 /* We found a matching flow. */
879 struct sw_flow
*flow
= *rcu_dereference(bucket
);
880 struct sw_flow_actions
*old_acts
, *new_acts
;
881 unsigned long int flags
;
883 /* Bail out if we're not allowed to modify an existing flow. */
885 if (!(uf
.flags
& ODPPF_MODIFY
))
889 new_acts
= get_actions(&uf
.flow
);
890 error
= PTR_ERR(new_acts
);
891 if (IS_ERR(new_acts
))
893 old_acts
= rcu_dereference(flow
->sf_acts
);
894 if (old_acts
->n_actions
!= new_acts
->n_actions
||
895 memcmp(old_acts
->actions
, new_acts
->actions
,
896 sizeof(union odp_action
) * old_acts
->n_actions
)) {
897 rcu_assign_pointer(flow
->sf_acts
, new_acts
);
898 flow_deferred_free_acts(old_acts
);
903 /* Fetch stats, then clear them if necessary. */
904 spin_lock_irqsave(&flow
->lock
, flags
);
905 get_stats(flow
, &stats
);
906 if (uf
.flags
& ODPPF_ZERO_STATS
)
908 spin_unlock_irqrestore(&flow
->lock
, flags
);
911 /* Copy stats to userspace. */
912 if (__copy_to_user(&ufp
->flow
.stats
, &stats
,
913 sizeof(struct odp_flow_stats
)))
918 kmem_cache_free(flow_cache
, flow
);
923 static int put_actions(const struct sw_flow
*flow
, struct odp_flow __user
*ufp
)
925 union odp_action __user
*actions
;
926 struct sw_flow_actions
*sf_acts
;
929 if (__get_user(actions
, &ufp
->actions
) ||
930 __get_user(n_actions
, &ufp
->n_actions
))
935 if (ufp
->n_actions
> INT_MAX
/ sizeof(union odp_action
))
938 sf_acts
= rcu_dereference(flow
->sf_acts
);
939 if (__put_user(sf_acts
->n_actions
, &ufp
->n_actions
) ||
940 (actions
&& copy_to_user(actions
, sf_acts
->actions
,
941 sizeof(union odp_action
) *
942 min(sf_acts
->n_actions
, n_actions
))))
948 static int answer_query(struct sw_flow
*flow
, struct odp_flow __user
*ufp
)
950 struct odp_flow_stats stats
;
951 unsigned long int flags
;
953 spin_lock_irqsave(&flow
->lock
, flags
);
954 get_stats(flow
, &stats
);
955 spin_unlock_irqrestore(&flow
->lock
, flags
);
957 if (__copy_to_user(&ufp
->stats
, &stats
, sizeof(struct odp_flow_stats
)))
959 return put_actions(flow
, ufp
);
962 static int del_or_query_flow(struct datapath
*dp
,
963 struct odp_flow __user
*ufp
,
966 struct dp_table
*table
= rcu_dereference(dp
->table
);
968 struct sw_flow
*flow
;
972 if (copy_from_user(&uf
, ufp
, sizeof uf
))
976 flow
= dp_table_lookup(table
, &uf
.key
);
981 if (cmd
== ODP_FLOW_DEL
) {
982 /* XXX redundant lookup */
983 error
= dp_table_delete(table
, flow
);
987 /* XXX These statistics might lose a few packets, since other
988 * CPUs can be using this flow. We used to synchronize_rcu()
989 * to make sure that we get completely accurate stats, but that
990 * blows our performance, badly. */
992 error
= answer_query(flow
, ufp
);
993 flow_deferred_free(flow
);
995 error
= answer_query(flow
, ufp
);
1002 static int query_multiple_flows(struct datapath
*dp
,
1003 const struct odp_flowvec
*flowvec
)
1005 struct dp_table
*table
= rcu_dereference(dp
->table
);
1007 for (i
= 0; i
< flowvec
->n_flows
; i
++) {
1008 struct __user odp_flow
*ufp
= &flowvec
->flows
[i
];
1010 struct sw_flow
*flow
;
1013 if (__copy_from_user(&uf
, ufp
, sizeof uf
))
1015 uf
.key
.reserved
= 0;
1017 flow
= dp_table_lookup(table
, &uf
.key
);
1019 error
= __clear_user(&ufp
->stats
, sizeof ufp
->stats
);
1021 error
= answer_query(flow
, ufp
);
1025 return flowvec
->n_flows
;
1028 struct list_flows_cbdata
{
1029 struct odp_flow __user
*uflows
;
1034 static int list_flow(struct sw_flow
*flow
, void *cbdata_
)
1036 struct list_flows_cbdata
*cbdata
= cbdata_
;
1037 struct odp_flow __user
*ufp
= &cbdata
->uflows
[cbdata
->listed_flows
++];
1040 if (__copy_to_user(&ufp
->key
, &flow
->key
, sizeof flow
->key
))
1042 error
= answer_query(flow
, ufp
);
1046 if (cbdata
->listed_flows
>= cbdata
->n_flows
)
1047 return cbdata
->listed_flows
;
1051 static int list_flows(struct datapath
*dp
, const struct odp_flowvec
*flowvec
)
1053 struct list_flows_cbdata cbdata
;
1056 if (!flowvec
->n_flows
)
1059 cbdata
.uflows
= flowvec
->flows
;
1060 cbdata
.n_flows
= flowvec
->n_flows
;
1061 cbdata
.listed_flows
= 0;
1062 error
= dp_table_foreach(rcu_dereference(dp
->table
),
1063 list_flow
, &cbdata
);
1064 return error
? error
: cbdata
.listed_flows
;
1067 static int do_flowvec_ioctl(struct datapath
*dp
, unsigned long argp
,
1068 int (*function
)(struct datapath
*,
1069 const struct odp_flowvec
*))
1071 struct odp_flowvec __user
*uflowvec
;
1072 struct odp_flowvec flowvec
;
1075 uflowvec
= (struct odp_flowvec __user
*)argp
;
1076 if (!access_ok(VERIFY_WRITE
, uflowvec
, sizeof *uflowvec
) ||
1077 copy_from_user(&flowvec
, uflowvec
, sizeof flowvec
))
1080 if (flowvec
.n_flows
> INT_MAX
/ sizeof(struct odp_flow
))
1083 if (!access_ok(VERIFY_WRITE
, flowvec
.flows
,
1084 flowvec
.n_flows
* sizeof(struct odp_flow
)))
1087 retval
= function(dp
, &flowvec
);
1088 return (retval
< 0 ? retval
1089 : retval
== flowvec
.n_flows
? 0
1090 : __put_user(retval
, &uflowvec
->n_flows
));
1093 static int do_execute(struct datapath
*dp
, const struct odp_execute
*executep
)
1095 struct odp_execute execute
;
1096 struct odp_flow_key key
;
1097 struct sk_buff
*skb
;
1098 struct sw_flow_actions
*actions
;
1102 if (copy_from_user(&execute
, executep
, sizeof execute
))
1106 if (execute
.length
< ETH_HLEN
|| execute
.length
> 65535)
1110 actions
= flow_actions_alloc(execute
.n_actions
);
1115 if (copy_from_user(actions
->actions
, execute
.actions
,
1116 execute
.n_actions
* sizeof *execute
.actions
))
1117 goto error_free_actions
;
1119 err
= validate_actions(actions
);
1121 goto error_free_actions
;
1124 skb
= alloc_skb(execute
.length
, GFP_KERNEL
);
1126 goto error_free_actions
;
1127 if (execute
.in_port
< DP_MAX_PORTS
) {
1128 struct net_bridge_port
*p
= dp
->ports
[execute
.in_port
];
1134 if (copy_from_user(skb_put(skb
, execute
.length
), execute
.data
,
1136 goto error_free_skb
;
1138 flow_extract(skb
, execute
.in_port
, &key
);
1139 err
= execute_actions(dp
, skb
, &key
, actions
->actions
,
1140 actions
->n_actions
, GFP_KERNEL
);
1153 get_dp_stats(struct datapath
*dp
, struct odp_stats __user
*statsp
)
1155 struct odp_stats stats
;
1158 stats
.n_flows
= dp
->n_flows
;
1159 stats
.cur_capacity
= rcu_dereference(dp
->table
)->n_buckets
* 2;
1160 stats
.max_capacity
= DP_MAX_BUCKETS
* 2;
1161 stats
.n_ports
= dp
->n_ports
;
1162 stats
.max_ports
= DP_MAX_PORTS
;
1163 stats
.max_groups
= DP_MAX_GROUPS
;
1164 stats
.n_frags
= stats
.n_hit
= stats
.n_missed
= stats
.n_lost
= 0;
1165 for_each_possible_cpu(i
) {
1166 const struct dp_stats_percpu
*s
;
1167 s
= percpu_ptr(dp
->stats_percpu
, i
);
1168 stats
.n_frags
+= s
->n_frags
;
1169 stats
.n_hit
+= s
->n_hit
;
1170 stats
.n_missed
+= s
->n_missed
;
1171 stats
.n_lost
+= s
->n_lost
;
1173 stats
.max_miss_queue
= DP_MAX_QUEUE_LEN
;
1174 stats
.max_action_queue
= DP_MAX_QUEUE_LEN
;
1175 return copy_to_user(statsp
, &stats
, sizeof stats
) ? -EFAULT
: 0;
1178 /* MTU of the dp pseudo-device: ETH_DATA_LEN or the minimum of the ports */
1179 int dp_min_mtu(const struct datapath
*dp
)
1181 struct net_bridge_port
*p
;
1186 list_for_each_entry_rcu (p
, &dp
->port_list
, node
) {
1187 struct net_device
*dev
= p
->dev
;
1189 /* Skip any internal ports, since that's what we're trying to
1194 if (!mtu
|| dev
->mtu
< mtu
)
1198 return mtu
? mtu
: ETH_DATA_LEN
;
1202 put_port(const struct net_bridge_port
*p
, struct odp_port __user
*uop
)
1205 memset(&op
, 0, sizeof op
);
1206 strncpy(op
.devname
, p
->dev
->name
, sizeof op
.devname
);
1207 op
.port
= p
->port_no
;
1208 op
.flags
= is_dp_dev(p
->dev
) ? ODP_PORT_INTERNAL
: 0;
1209 return copy_to_user(uop
, &op
, sizeof op
) ? -EFAULT
: 0;
1213 query_port(struct datapath
*dp
, struct odp_port __user
*uport
)
1215 struct odp_port port
;
1217 if (copy_from_user(&port
, uport
, sizeof port
))
1219 if (port
.devname
[0]) {
1220 struct net_bridge_port
*p
;
1221 struct net_device
*dev
;
1224 port
.devname
[IFNAMSIZ
- 1] = '\0';
1226 dev
= dev_get_by_name(&init_net
, port
.devname
);
1231 if (!p
&& is_dp_dev(dev
)) {
1232 struct dp_dev
*dp_dev
= dp_dev_priv(dev
);
1233 if (dp_dev
->dp
== dp
)
1234 p
= dp
->ports
[dp_dev
->port_no
];
1236 err
= p
&& p
->dp
== dp
? put_port(p
, uport
) : -ENOENT
;
1241 if (port
.port
>= DP_MAX_PORTS
)
1243 if (!dp
->ports
[port
.port
])
1245 return put_port(dp
->ports
[port
.port
], uport
);
1250 list_ports(struct datapath
*dp
, struct odp_portvec __user
*pvp
)
1252 struct odp_portvec pv
;
1253 struct net_bridge_port
*p
;
1256 if (copy_from_user(&pv
, pvp
, sizeof pv
))
1261 list_for_each_entry_rcu (p
, &dp
->port_list
, node
) {
1262 if (put_port(p
, &pv
.ports
[idx
]))
1264 if (idx
++ >= pv
.n_ports
)
1268 return put_user(idx
, &pvp
->n_ports
);
1271 /* RCU callback for freeing a dp_port_group */
1272 static void free_port_group(struct rcu_head
*rcu
)
1274 struct dp_port_group
*g
= container_of(rcu
, struct dp_port_group
, rcu
);
1279 set_port_group(struct datapath
*dp
, const struct odp_port_group __user
*upg
)
1281 struct odp_port_group pg
;
1282 struct dp_port_group
*new_group
, *old_group
;
1286 if (copy_from_user(&pg
, upg
, sizeof pg
))
1290 if (pg
.n_ports
> DP_MAX_PORTS
|| pg
.group
>= DP_MAX_GROUPS
)
1294 new_group
= kmalloc(sizeof *new_group
+ sizeof(u16
) * pg
.n_ports
,
1299 new_group
->n_ports
= pg
.n_ports
;
1301 if (copy_from_user(new_group
->ports
, pg
.ports
,
1302 sizeof(u16
) * pg
.n_ports
))
1305 old_group
= rcu_dereference(dp
->groups
[pg
.group
]);
1306 rcu_assign_pointer(dp
->groups
[pg
.group
], new_group
);
1308 call_rcu(&old_group
->rcu
, free_port_group
);
1318 get_port_group(struct datapath
*dp
, struct odp_port_group
*upg
)
1320 struct odp_port_group pg
;
1321 struct dp_port_group
*g
;
1324 if (copy_from_user(&pg
, upg
, sizeof pg
))
1327 if (pg
.group
>= DP_MAX_GROUPS
)
1330 g
= dp
->groups
[pg
.group
];
1331 n_copy
= g
? min_t(int, g
->n_ports
, pg
.n_ports
) : 0;
1332 if (n_copy
&& copy_to_user(pg
.ports
, g
->ports
, n_copy
* sizeof(u16
)))
1335 if (put_user(g
? g
->n_ports
: 0, &upg
->n_ports
))
1341 static long openvswitch_ioctl(struct file
*f
, unsigned int cmd
,
1344 int dp_idx
= iminor(f
->f_dentry
->d_inode
);
1345 struct datapath
*dp
;
1346 int drop_frags
, listeners
, port_no
;
1349 /* Handle commands with special locking requirements up front. */
1352 return create_dp(dp_idx
, (char __user
*)argp
);
1354 case ODP_DP_DESTROY
:
1355 return destroy_dp(dp_idx
);
1358 return add_port(dp_idx
, (struct odp_port __user
*)argp
);
1361 err
= get_user(port_no
, (int __user
*)argp
);
1364 return del_port(dp_idx
, port_no
);
1367 dp
= get_dp_locked(dp_idx
);
1373 err
= get_dp_stats(dp
, (struct odp_stats __user
*)argp
);
1376 case ODP_GET_DROP_FRAGS
:
1377 err
= put_user(dp
->drop_frags
, (int __user
*)argp
);
1380 case ODP_SET_DROP_FRAGS
:
1381 err
= get_user(drop_frags
, (int __user
*)argp
);
1385 if (drop_frags
!= 0 && drop_frags
!= 1)
1387 dp
->drop_frags
= drop_frags
;
1391 case ODP_GET_LISTEN_MASK
:
1392 err
= put_user((int)f
->private_data
, (int __user
*)argp
);
1395 case ODP_SET_LISTEN_MASK
:
1396 err
= get_user(listeners
, (int __user
*)argp
);
1400 if (listeners
& ~ODPL_ALL
)
1403 f
->private_data
= (void*)listeners
;
1406 case ODP_PORT_QUERY
:
1407 err
= query_port(dp
, (struct odp_port __user
*)argp
);
1411 err
= list_ports(dp
, (struct odp_portvec __user
*)argp
);
1414 case ODP_PORT_GROUP_SET
:
1415 err
= set_port_group(dp
, (struct odp_port_group __user
*)argp
);
1418 case ODP_PORT_GROUP_GET
:
1419 err
= get_port_group(dp
, (struct odp_port_group __user
*)argp
);
1422 case ODP_FLOW_FLUSH
:
1423 err
= flush_flows(dp
);
1427 err
= put_flow(dp
, (struct odp_flow_put __user
*)argp
);
1432 err
= del_or_query_flow(dp
, (struct odp_flow __user
*)argp
,
1436 case ODP_FLOW_GET_MULTIPLE
:
1437 err
= do_flowvec_ioctl(dp
, argp
, query_multiple_flows
);
1441 err
= do_flowvec_ioctl(dp
, argp
, list_flows
);
1445 err
= do_execute(dp
, (struct odp_execute __user
*)argp
);
1452 mutex_unlock(&dp
->mutex
);
1456 static int dp_has_packet_of_interest(struct datapath
*dp
, int listeners
)
1459 for (i
= 0; i
< DP_N_QUEUES
; i
++) {
1460 if (listeners
& (1 << i
) && !skb_queue_empty(&dp
->queues
[i
]))
1466 ssize_t
openvswitch_read(struct file
*f
, char __user
*buf
, size_t nbytes
,
1469 /* XXX is there sufficient synchronization here? */
1470 int listeners
= (int) f
->private_data
;
1471 int dp_idx
= iminor(f
->f_dentry
->d_inode
);
1472 struct datapath
*dp
= get_dp(dp_idx
);
1473 struct sk_buff
*skb
;
1474 struct iovec __user iov
;
1481 if (nbytes
== 0 || !listeners
)
1487 for (i
= 0; i
< DP_N_QUEUES
; i
++) {
1488 if (listeners
& (1 << i
)) {
1489 skb
= skb_dequeue(&dp
->queues
[i
]);
1495 if (f
->f_flags
& O_NONBLOCK
) {
1500 wait_event_interruptible(dp
->waitqueue
,
1501 dp_has_packet_of_interest(dp
,
1504 if (signal_pending(current
)) {
1505 retval
= -ERESTARTSYS
;
1510 copy_bytes
= min(skb
->len
, nbytes
);
1512 iov
.iov_len
= copy_bytes
;
1513 retval
= skb_copy_datagram_iovec(skb
, 0, &iov
, iov
.iov_len
);
1515 retval
= copy_bytes
;
1522 static unsigned int openvswitch_poll(struct file
*file
, poll_table
*wait
)
1524 /* XXX is there sufficient synchronization here? */
1525 int dp_idx
= iminor(file
->f_dentry
->d_inode
);
1526 struct datapath
*dp
= get_dp(dp_idx
);
1531 poll_wait(file
, &dp
->waitqueue
, wait
);
1532 if (dp_has_packet_of_interest(dp
, (int)file
->private_data
))
1533 mask
|= POLLIN
| POLLRDNORM
;
1535 mask
= POLLIN
| POLLRDNORM
| POLLHUP
;
1540 struct file_operations openvswitch_fops
= {
1541 /* XXX .aio_read = openvswitch_aio_read, */
1542 .read
= openvswitch_read
,
1543 .poll
= openvswitch_poll
,
1544 .unlocked_ioctl
= openvswitch_ioctl
,
1545 /* XXX .fasync = openvswitch_fasync, */
1549 static struct llc_sap
*dp_stp_sap
;
1551 static int dp_stp_rcv(struct sk_buff
*skb
, struct net_device
*dev
,
1552 struct packet_type
*pt
, struct net_device
*orig_dev
)
1554 /* We don't really care about STP packets, we just listen for them for
1555 * mutual exclusion with the bridge module, so this just discards
1561 static int __init
dp_init(void)
1565 printk("Open vSwitch %s, built "__DATE__
" "__TIME__
"\n", VERSION BUILDNR
);
1567 /* Register to receive STP packets because the bridge module also
1568 * attempts to do so. Since there can only be a single listener for a
1569 * given protocol, this provides mutual exclusion against the bridge
1570 * module, preventing both of them from being loaded at the same
1572 dp_stp_sap
= llc_sap_open(LLC_SAP_BSPAN
, dp_stp_rcv
);
1574 printk(KERN_ERR
"openvswitch: can't register sap for STP (probably the bridge module is loaded)\n");
1582 err
= register_netdevice_notifier(&dp_device_notifier
);
1584 goto error_flow_exit
;
1586 major
= register_chrdev(0, "openvswitch", &openvswitch_fops
);
1588 goto error_unreg_notifier
;
1590 /* Hook into callback used by the bridge to intercept packets.
1591 * Parasites we are. */
1592 br_handle_frame_hook
= dp_frame_hook
;
1596 error_unreg_notifier
:
1597 unregister_netdevice_notifier(&dp_device_notifier
);
1604 static void dp_cleanup(void)
1607 unregister_chrdev(major
, "openvswitch");
1608 unregister_netdevice_notifier(&dp_device_notifier
);
1610 br_handle_frame_hook
= NULL
;
1611 llc_sap_put(dp_stp_sap
);
1614 module_init(dp_init
);
1615 module_exit(dp_cleanup
);
1617 MODULE_DESCRIPTION("Open vSwitch switching datapath");
1618 MODULE_LICENSE("GPL");