1 /*-
2 * Copyright (c) 2014, Bryan Venteicher <[email protected]>
3 * All rights reserved.
4 * Copyright (c) 2020, Chelsio Communications.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice unmodified, this list of conditions, and the following
11 * disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 */
27
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33
34 #include <sys/param.h>
35 #include <sys/eventhandler.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/hash.h>
39 #include <sys/malloc.h>
40 #include <sys/mbuf.h>
41 #include <sys/module.h>
42 #include <sys/refcount.h>
43 #include <sys/rmlock.h>
44 #include <sys/priv.h>
45 #include <sys/proc.h>
46 #include <sys/queue.h>
47 #include <sys/sbuf.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sockio.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53
54 #include <net/bpf.h>
55 #include <net/ethernet.h>
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #include <net/if_clone.h>
59 #include <net/if_dl.h>
60 #include <net/if_media.h>
61 #include <net/if_types.h>
62 #include <net/if_vxlan.h>
63 #include <net/netisr.h>
64 #include <net/route.h>
65 #include <net/route/nhop.h>
66
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/in_var.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/udp.h>
75 #include <netinet/udp_var.h>
76 #include <netinet/in_fib.h>
77 #include <netinet6/in6_fib.h>
78
79 #include <netinet6/ip6_var.h>
80 #include <netinet6/scope6_var.h>
81
82 struct vxlan_softc;
83 LIST_HEAD(vxlan_softc_head, vxlan_softc);
84
85 struct sx vxlan_sx;
86 SX_SYSINIT(vxlan, &vxlan_sx, "VXLAN global start/stop lock");
87
88 struct vxlan_socket_mc_info {
89 union vxlan_sockaddr vxlsomc_saddr;
90 union vxlan_sockaddr vxlsomc_gaddr;
91 int vxlsomc_ifidx;
92 int vxlsomc_users;
93 };
94
95 /*
96 * The maximum MTU of encapsulated ethernet frame within IPv4/UDP packet.
97 */
98 #define VXLAN_MAX_MTU (IP_MAXPACKET - \
99 60 /* Maximum IPv4 header len */ - \
100 sizeof(struct udphdr) - \
101 sizeof(struct vxlan_header) - \
102 ETHER_HDR_LEN - ETHER_CRC_LEN - ETHER_VLAN_ENCAP_LEN)
103 #define VXLAN_BASIC_IFCAPS (IFCAP_LINKSTATE | IFCAP_JUMBO_MTU)
104
105 #define VXLAN_SO_MC_MAX_GROUPS 32
106
107 #define VXLAN_SO_VNI_HASH_SHIFT 6
108 #define VXLAN_SO_VNI_HASH_SIZE (1 << VXLAN_SO_VNI_HASH_SHIFT)
109 #define VXLAN_SO_VNI_HASH(_vni) ((_vni) % VXLAN_SO_VNI_HASH_SIZE)
110
111 struct vxlan_socket {
112 struct socket *vxlso_sock;
113 struct rmlock vxlso_lock;
114 u_int vxlso_refcnt;
115 union vxlan_sockaddr vxlso_laddr;
116 LIST_ENTRY(vxlan_socket) vxlso_entry;
117 struct vxlan_softc_head vxlso_vni_hash[VXLAN_SO_VNI_HASH_SIZE];
118 struct vxlan_socket_mc_info vxlso_mc[VXLAN_SO_MC_MAX_GROUPS];
119 };
120
121 #define VXLAN_SO_RLOCK(_vso, _p) rm_rlock(&(_vso)->vxlso_lock, (_p))
122 #define VXLAN_SO_RUNLOCK(_vso, _p) rm_runlock(&(_vso)->vxlso_lock, (_p))
123 #define VXLAN_SO_WLOCK(_vso) rm_wlock(&(_vso)->vxlso_lock)
124 #define VXLAN_SO_WUNLOCK(_vso) rm_wunlock(&(_vso)->vxlso_lock)
125 #define VXLAN_SO_LOCK_ASSERT(_vso) \
126 rm_assert(&(_vso)->vxlso_lock, RA_LOCKED)
127 #define VXLAN_SO_LOCK_WASSERT(_vso) \
128 rm_assert(&(_vso)->vxlso_lock, RA_WLOCKED)
129
130 #define VXLAN_SO_ACQUIRE(_vso) refcount_acquire(&(_vso)->vxlso_refcnt)
131 #define VXLAN_SO_RELEASE(_vso) refcount_release(&(_vso)->vxlso_refcnt)
132
133 struct vxlan_ftable_entry {
134 LIST_ENTRY(vxlan_ftable_entry) vxlfe_hash;
135 uint16_t vxlfe_flags;
136 uint8_t vxlfe_mac[ETHER_ADDR_LEN];
137 union vxlan_sockaddr vxlfe_raddr;
138 time_t vxlfe_expire;
139 };
140
141 #define VXLAN_FE_FLAG_DYNAMIC 0x01
142 #define VXLAN_FE_FLAG_STATIC 0x02
143
144 #define VXLAN_FE_IS_DYNAMIC(_fe) \
145 ((_fe)->vxlfe_flags & VXLAN_FE_FLAG_DYNAMIC)
146
147 #define VXLAN_SC_FTABLE_SHIFT 9
148 #define VXLAN_SC_FTABLE_SIZE (1 << VXLAN_SC_FTABLE_SHIFT)
149 #define VXLAN_SC_FTABLE_MASK (VXLAN_SC_FTABLE_SIZE - 1)
150 #define VXLAN_SC_FTABLE_HASH(_sc, _mac) \
151 (vxlan_mac_hash(_sc, _mac) % VXLAN_SC_FTABLE_SIZE)
152
153 LIST_HEAD(vxlan_ftable_head, vxlan_ftable_entry);
154
155 struct vxlan_statistics {
156 uint32_t ftable_nospace;
157 uint32_t ftable_lock_upgrade_failed;
158 counter_u64_t txcsum;
159 counter_u64_t tso;
160 counter_u64_t rxcsum;
161 };
162
163 struct vxlan_softc {
164 struct ifnet *vxl_ifp;
165 int vxl_reqcap;
166 struct vxlan_socket *vxl_sock;
167 uint32_t vxl_vni;
168 union vxlan_sockaddr vxl_src_addr;
169 union vxlan_sockaddr vxl_dst_addr;
170 uint32_t vxl_flags;
171 #define VXLAN_FLAG_INIT 0x0001
172 #define VXLAN_FLAG_TEARDOWN 0x0002
173 #define VXLAN_FLAG_LEARN 0x0004
174 #define VXLAN_FLAG_USER_MTU 0x0008
175
176 uint32_t vxl_port_hash_key;
177 uint16_t vxl_min_port;
178 uint16_t vxl_max_port;
179 uint8_t vxl_ttl;
180
181 /* Lookup table from MAC address to forwarding entry. */
182 uint32_t vxl_ftable_cnt;
183 uint32_t vxl_ftable_max;
184 uint32_t vxl_ftable_timeout;
185 uint32_t vxl_ftable_hash_key;
186 struct vxlan_ftable_head *vxl_ftable;
187
188 /* Derived from vxl_dst_addr. */
189 struct vxlan_ftable_entry vxl_default_fe;
190
191 struct ip_moptions *vxl_im4o;
192 struct ip6_moptions *vxl_im6o;
193
194 struct rmlock vxl_lock;
195 volatile u_int vxl_refcnt;
196
197 int vxl_unit;
198 int vxl_vso_mc_index;
199 struct vxlan_statistics vxl_stats;
200 struct sysctl_oid *vxl_sysctl_node;
201 struct sysctl_ctx_list vxl_sysctl_ctx;
202 struct callout vxl_callout;
203 struct ether_addr vxl_hwaddr;
204 int vxl_mc_ifindex;
205 struct ifnet *vxl_mc_ifp;
206 struct ifmedia vxl_media;
207 char vxl_mc_ifname[IFNAMSIZ];
208 LIST_ENTRY(vxlan_softc) vxl_entry;
209 LIST_ENTRY(vxlan_softc) vxl_ifdetach_list;
210
211 /* For rate limiting errors on the tx fast path. */
212 struct timeval err_time;
213 int err_pps;
214 };
215
216 #define VXLAN_RLOCK(_sc, _p) rm_rlock(&(_sc)->vxl_lock, (_p))
217 #define VXLAN_RUNLOCK(_sc, _p) rm_runlock(&(_sc)->vxl_lock, (_p))
218 #define VXLAN_WLOCK(_sc) rm_wlock(&(_sc)->vxl_lock)
219 #define VXLAN_WUNLOCK(_sc) rm_wunlock(&(_sc)->vxl_lock)
220 #define VXLAN_LOCK_WOWNED(_sc) rm_wowned(&(_sc)->vxl_lock)
221 #define VXLAN_LOCK_ASSERT(_sc) rm_assert(&(_sc)->vxl_lock, RA_LOCKED)
222 #define VXLAN_LOCK_WASSERT(_sc) rm_assert(&(_sc)->vxl_lock, RA_WLOCKED)
223 #define VXLAN_UNLOCK(_sc, _p) do { \
224 if (VXLAN_LOCK_WOWNED(_sc)) \
225 VXLAN_WUNLOCK(_sc); \
226 else \
227 VXLAN_RUNLOCK(_sc, _p); \
228 } while (0)
229
230 #define VXLAN_ACQUIRE(_sc) refcount_acquire(&(_sc)->vxl_refcnt)
231 #define VXLAN_RELEASE(_sc) refcount_release(&(_sc)->vxl_refcnt)
232
233 #define satoconstsin(sa) ((const struct sockaddr_in *)(sa))
234 #define satoconstsin6(sa) ((const struct sockaddr_in6 *)(sa))
235
236 struct vxlanudphdr {
237 struct udphdr vxlh_udp;
238 struct vxlan_header vxlh_hdr;
239 } __packed;
240
241 static int vxlan_ftable_addr_cmp(const uint8_t *, const uint8_t *);
242 static void vxlan_ftable_init(struct vxlan_softc *);
243 static void vxlan_ftable_fini(struct vxlan_softc *);
244 static void vxlan_ftable_flush(struct vxlan_softc *, int);
245 static void vxlan_ftable_expire(struct vxlan_softc *);
246 static int vxlan_ftable_update_locked(struct vxlan_softc *,
247 const union vxlan_sockaddr *, const uint8_t *,
248 struct rm_priotracker *);
249 static int vxlan_ftable_learn(struct vxlan_softc *,
250 const struct sockaddr *, const uint8_t *);
251 static int vxlan_ftable_sysctl_dump(SYSCTL_HANDLER_ARGS);
252
253 static struct vxlan_ftable_entry *
254 vxlan_ftable_entry_alloc(void);
255 static void vxlan_ftable_entry_free(struct vxlan_ftable_entry *);
256 static void vxlan_ftable_entry_init(struct vxlan_softc *,
257 struct vxlan_ftable_entry *, const uint8_t *,
258 const struct sockaddr *, uint32_t);
259 static void vxlan_ftable_entry_destroy(struct vxlan_softc *,
260 struct vxlan_ftable_entry *);
261 static int vxlan_ftable_entry_insert(struct vxlan_softc *,
262 struct vxlan_ftable_entry *);
263 static struct vxlan_ftable_entry *
264 vxlan_ftable_entry_lookup(struct vxlan_softc *,
265 const uint8_t *);
266 static void vxlan_ftable_entry_dump(struct vxlan_ftable_entry *,
267 struct sbuf *);
268
269 static struct vxlan_socket *
270 vxlan_socket_alloc(const union vxlan_sockaddr *);
271 static void vxlan_socket_destroy(struct vxlan_socket *);
272 static void vxlan_socket_release(struct vxlan_socket *);
273 static struct vxlan_socket *
274 vxlan_socket_lookup(union vxlan_sockaddr *vxlsa);
275 static void vxlan_socket_insert(struct vxlan_socket *);
276 static int vxlan_socket_init(struct vxlan_socket *, struct ifnet *);
277 static int vxlan_socket_bind(struct vxlan_socket *, struct ifnet *);
278 static int vxlan_socket_create(struct ifnet *, int,
279 const union vxlan_sockaddr *, struct vxlan_socket **);
280 static void vxlan_socket_ifdetach(struct vxlan_socket *,
281 struct ifnet *, struct vxlan_softc_head *);
282
283 static struct vxlan_socket *
284 vxlan_socket_mc_lookup(const union vxlan_sockaddr *);
285 static int vxlan_sockaddr_mc_info_match(
286 const struct vxlan_socket_mc_info *,
287 const union vxlan_sockaddr *,
288 const union vxlan_sockaddr *, int);
289 static int vxlan_socket_mc_join_group(struct vxlan_socket *,
290 const union vxlan_sockaddr *, const union vxlan_sockaddr *,
291 int *, union vxlan_sockaddr *);
292 static int vxlan_socket_mc_leave_group(struct vxlan_socket *,
293 const union vxlan_sockaddr *,
294 const union vxlan_sockaddr *, int);
295 static int vxlan_socket_mc_add_group(struct vxlan_socket *,
296 const union vxlan_sockaddr *, const union vxlan_sockaddr *,
297 int, int *);
298 static void vxlan_socket_mc_release_group_by_idx(struct vxlan_socket *,
299 int);
300
301 static struct vxlan_softc *
302 vxlan_socket_lookup_softc_locked(struct vxlan_socket *,
303 uint32_t);
304 static struct vxlan_softc *
305 vxlan_socket_lookup_softc(struct vxlan_socket *, uint32_t);
306 static int vxlan_socket_insert_softc(struct vxlan_socket *,
307 struct vxlan_softc *);
308 static void vxlan_socket_remove_softc(struct vxlan_socket *,
309 struct vxlan_softc *);
310
311 static struct ifnet *
312 vxlan_multicast_if_ref(struct vxlan_softc *, int);
313 static void vxlan_free_multicast(struct vxlan_softc *);
314 static int vxlan_setup_multicast_interface(struct vxlan_softc *);
315
316 static int vxlan_setup_multicast(struct vxlan_softc *);
317 static int vxlan_setup_socket(struct vxlan_softc *);
318 #ifdef INET6
319 static void vxlan_setup_zero_checksum_port(struct vxlan_softc *);
320 #endif
321 static void vxlan_setup_interface_hdrlen(struct vxlan_softc *);
322 static int vxlan_valid_init_config(struct vxlan_softc *);
323 static void vxlan_init_wait(struct vxlan_softc *);
324 static void vxlan_init_complete(struct vxlan_softc *);
325 static void vxlan_init(void *);
326 static void vxlan_release(struct vxlan_softc *);
327 static void vxlan_teardown_wait(struct vxlan_softc *);
328 static void vxlan_teardown_complete(struct vxlan_softc *);
329 static void vxlan_teardown_locked(struct vxlan_softc *);
330 static void vxlan_teardown(struct vxlan_softc *);
331 static void vxlan_ifdetach(struct vxlan_softc *, struct ifnet *,
332 struct vxlan_softc_head *);
333 static void vxlan_timer(void *);
334
335 static int vxlan_ctrl_get_config(struct vxlan_softc *, void *);
336 static int vxlan_ctrl_set_vni(struct vxlan_softc *, void *);
337 static int vxlan_ctrl_set_local_addr(struct vxlan_softc *, void *);
338 static int vxlan_ctrl_set_remote_addr(struct vxlan_softc *, void *);
339 static int vxlan_ctrl_set_local_port(struct vxlan_softc *, void *);
340 static int vxlan_ctrl_set_remote_port(struct vxlan_softc *, void *);
341 static int vxlan_ctrl_set_port_range(struct vxlan_softc *, void *);
342 static int vxlan_ctrl_set_ftable_timeout(struct vxlan_softc *, void *);
343 static int vxlan_ctrl_set_ftable_max(struct vxlan_softc *, void *);
344 static int vxlan_ctrl_set_multicast_if(struct vxlan_softc * , void *);
345 static int vxlan_ctrl_set_ttl(struct vxlan_softc *, void *);
346 static int vxlan_ctrl_set_learn(struct vxlan_softc *, void *);
347 static int vxlan_ctrl_ftable_entry_add(struct vxlan_softc *, void *);
348 static int vxlan_ctrl_ftable_entry_rem(struct vxlan_softc *, void *);
349 static int vxlan_ctrl_flush(struct vxlan_softc *, void *);
350 static int vxlan_ioctl_drvspec(struct vxlan_softc *,
351 struct ifdrv *, int);
352 static int vxlan_ioctl_ifflags(struct vxlan_softc *);
353 static int vxlan_ioctl(struct ifnet *, u_long, caddr_t);
354
355 #if defined(INET) || defined(INET6)
356 static uint16_t vxlan_pick_source_port(struct vxlan_softc *, struct mbuf *);
357 static void vxlan_encap_header(struct vxlan_softc *, struct mbuf *,
358 int, uint16_t, uint16_t);
359 #endif
360 static int vxlan_encap4(struct vxlan_softc *,
361 const union vxlan_sockaddr *, struct mbuf *);
362 static int vxlan_encap6(struct vxlan_softc *,
363 const union vxlan_sockaddr *, struct mbuf *);
364 static int vxlan_transmit(struct ifnet *, struct mbuf *);
365 static void vxlan_qflush(struct ifnet *);
366 static void vxlan_rcv_udp_packet(struct mbuf *, int, struct inpcb *,
367 const struct sockaddr *, void *);
368 static int vxlan_input(struct vxlan_socket *, uint32_t, struct mbuf **,
369 const struct sockaddr *);
370
371 static int vxlan_stats_alloc(struct vxlan_softc *);
372 static void vxlan_stats_free(struct vxlan_softc *);
373 static void vxlan_set_default_config(struct vxlan_softc *);
374 static int vxlan_set_user_config(struct vxlan_softc *,
375 struct ifvxlanparam *);
376 static int vxlan_set_reqcap(struct vxlan_softc *, struct ifnet *, int);
377 static void vxlan_set_hwcaps(struct vxlan_softc *);
378 static int vxlan_clone_create(struct if_clone *, int, caddr_t);
379 static void vxlan_clone_destroy(struct ifnet *);
380
381 static uint32_t vxlan_mac_hash(struct vxlan_softc *, const uint8_t *);
382 static int vxlan_media_change(struct ifnet *);
383 static void vxlan_media_status(struct ifnet *, struct ifmediareq *);
384
385 static int vxlan_sockaddr_cmp(const union vxlan_sockaddr *,
386 const struct sockaddr *);
387 static void vxlan_sockaddr_copy(union vxlan_sockaddr *,
388 const struct sockaddr *);
389 static int vxlan_sockaddr_in_equal(const union vxlan_sockaddr *,
390 const struct sockaddr *);
391 static void vxlan_sockaddr_in_copy(union vxlan_sockaddr *,
392 const struct sockaddr *);
393 static int vxlan_sockaddr_supported(const union vxlan_sockaddr *, int);
394 static int vxlan_sockaddr_in_any(const union vxlan_sockaddr *);
395 static int vxlan_sockaddr_in_multicast(const union vxlan_sockaddr *);
396 static int vxlan_sockaddr_in6_embedscope(union vxlan_sockaddr *);
397
398 static int vxlan_can_change_config(struct vxlan_softc *);
399 static int vxlan_check_vni(uint32_t);
400 static int vxlan_check_ttl(int);
401 static int vxlan_check_ftable_timeout(uint32_t);
402 static int vxlan_check_ftable_max(uint32_t);
403
404 static void vxlan_sysctl_setup(struct vxlan_softc *);
405 static void vxlan_sysctl_destroy(struct vxlan_softc *);
406 static int vxlan_tunable_int(struct vxlan_softc *, const char *, int);
407
408 static void vxlan_ifdetach_event(void *, struct ifnet *);
409 static void vxlan_load(void);
410 static void vxlan_unload(void);
411 static int vxlan_modevent(module_t, int, void *);
412
413 static const char vxlan_name[] = "vxlan";
414 static MALLOC_DEFINE(M_VXLAN, vxlan_name,
415 "Virtual eXtensible LAN Interface");
416 static struct if_clone *vxlan_cloner;
417
418 static struct mtx vxlan_list_mtx;
419 #define VXLAN_LIST_LOCK() mtx_lock(&vxlan_list_mtx)
420 #define VXLAN_LIST_UNLOCK() mtx_unlock(&vxlan_list_mtx)
421
422 static LIST_HEAD(, vxlan_socket) vxlan_socket_list;
423
424 static eventhandler_tag vxlan_ifdetach_event_tag;
425
426 SYSCTL_DECL(_net_link);
427 SYSCTL_NODE(_net_link, OID_AUTO, vxlan, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
428 "Virtual eXtensible Local Area Network");
429
430 static int vxlan_legacy_port = 0;
431 TUNABLE_INT("net.link.vxlan.legacy_port", &vxlan_legacy_port);
432 static int vxlan_reuse_port = 0;
433 TUNABLE_INT("net.link.vxlan.reuse_port", &vxlan_reuse_port);
434
435 /* Default maximum number of addresses in the forwarding table. */
436 #ifndef VXLAN_FTABLE_MAX
437 #define VXLAN_FTABLE_MAX 2000
438 #endif
439
440 /* Timeout (in seconds) of addresses learned in the forwarding table. */
441 #ifndef VXLAN_FTABLE_TIMEOUT
442 #define VXLAN_FTABLE_TIMEOUT (20 * 60)
443 #endif
444
445 /*
446 * Maximum timeout (in seconds) of addresses learned in the forwarding
447 * table.
448 */
449 #ifndef VXLAN_FTABLE_MAX_TIMEOUT
450 #define VXLAN_FTABLE_MAX_TIMEOUT (60 * 60 * 24)
451 #endif
452
453 /* Number of seconds between pruning attempts of the forwarding table. */
454 #ifndef VXLAN_FTABLE_PRUNE
455 #define VXLAN_FTABLE_PRUNE (5 * 60)
456 #endif
457
458 static int vxlan_ftable_prune_period = VXLAN_FTABLE_PRUNE;
459
460 struct vxlan_control {
461 int (*vxlc_func)(struct vxlan_softc *, void *);
462 int vxlc_argsize;
463 int vxlc_flags;
464 #define VXLAN_CTRL_FLAG_COPYIN 0x01
465 #define VXLAN_CTRL_FLAG_COPYOUT 0x02
466 #define VXLAN_CTRL_FLAG_SUSER 0x04
467 };
468
469 static const struct vxlan_control vxlan_control_table[] = {
470 [VXLAN_CMD_GET_CONFIG] =
471 { vxlan_ctrl_get_config, sizeof(struct ifvxlancfg),
472 VXLAN_CTRL_FLAG_COPYOUT
473 },
474
475 [VXLAN_CMD_SET_VNI] =
476 { vxlan_ctrl_set_vni, sizeof(struct ifvxlancmd),
477 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
478 },
479
480 [VXLAN_CMD_SET_LOCAL_ADDR] =
481 { vxlan_ctrl_set_local_addr, sizeof(struct ifvxlancmd),
482 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
483 },
484
485 [VXLAN_CMD_SET_REMOTE_ADDR] =
486 { vxlan_ctrl_set_remote_addr, sizeof(struct ifvxlancmd),
487 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
488 },
489
490 [VXLAN_CMD_SET_LOCAL_PORT] =
491 { vxlan_ctrl_set_local_port, sizeof(struct ifvxlancmd),
492 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
493 },
494
495 [VXLAN_CMD_SET_REMOTE_PORT] =
496 { vxlan_ctrl_set_remote_port, sizeof(struct ifvxlancmd),
497 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
498 },
499
500 [VXLAN_CMD_SET_PORT_RANGE] =
501 { vxlan_ctrl_set_port_range, sizeof(struct ifvxlancmd),
502 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
503 },
504
505 [VXLAN_CMD_SET_FTABLE_TIMEOUT] =
506 { vxlan_ctrl_set_ftable_timeout, sizeof(struct ifvxlancmd),
507 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
508 },
509
510 [VXLAN_CMD_SET_FTABLE_MAX] =
511 { vxlan_ctrl_set_ftable_max, sizeof(struct ifvxlancmd),
512 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
513 },
514
515 [VXLAN_CMD_SET_MULTICAST_IF] =
516 { vxlan_ctrl_set_multicast_if, sizeof(struct ifvxlancmd),
517 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
518 },
519
520 [VXLAN_CMD_SET_TTL] =
521 { vxlan_ctrl_set_ttl, sizeof(struct ifvxlancmd),
522 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
523 },
524
525 [VXLAN_CMD_SET_LEARN] =
526 { vxlan_ctrl_set_learn, sizeof(struct ifvxlancmd),
527 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
528 },
529
530 [VXLAN_CMD_FTABLE_ENTRY_ADD] =
531 { vxlan_ctrl_ftable_entry_add, sizeof(struct ifvxlancmd),
532 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
533 },
534
535 [VXLAN_CMD_FTABLE_ENTRY_REM] =
536 { vxlan_ctrl_ftable_entry_rem, sizeof(struct ifvxlancmd),
537 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
538 },
539
540 [VXLAN_CMD_FLUSH] =
541 { vxlan_ctrl_flush, sizeof(struct ifvxlancmd),
542 VXLAN_CTRL_FLAG_COPYIN | VXLAN_CTRL_FLAG_SUSER,
543 },
544 };
545
546 static const int vxlan_control_table_size = nitems(vxlan_control_table);
547
548 static int
vxlan_ftable_addr_cmp(const uint8_t * a,const uint8_t * b)549 vxlan_ftable_addr_cmp(const uint8_t *a, const uint8_t *b)
550 {
551 int i, d;
552
553 for (i = 0, d = 0; i < ETHER_ADDR_LEN && d == 0; i++)
554 d = ((int)a[i]) - ((int)b[i]);
555
556 return (d);
557 }
558
559 static void
vxlan_ftable_init(struct vxlan_softc * sc)560 vxlan_ftable_init(struct vxlan_softc *sc)
561 {
562 int i;
563
564 sc->vxl_ftable = malloc(sizeof(struct vxlan_ftable_head) *
565 VXLAN_SC_FTABLE_SIZE, M_VXLAN, M_ZERO | M_WAITOK);
566
567 for (i = 0; i < VXLAN_SC_FTABLE_SIZE; i++)
568 LIST_INIT(&sc->vxl_ftable[i]);
569 sc->vxl_ftable_hash_key = arc4random();
570 }
571
572 static void
vxlan_ftable_fini(struct vxlan_softc * sc)573 vxlan_ftable_fini(struct vxlan_softc *sc)
574 {
575 int i;
576
577 for (i = 0; i < VXLAN_SC_FTABLE_SIZE; i++) {
578 KASSERT(LIST_EMPTY(&sc->vxl_ftable[i]),
579 ("%s: vxlan %p ftable[%d] not empty", __func__, sc, i));
580 }
581 MPASS(sc->vxl_ftable_cnt == 0);
582
583 free(sc->vxl_ftable, M_VXLAN);
584 sc->vxl_ftable = NULL;
585 }
586
587 static void
vxlan_ftable_flush(struct vxlan_softc * sc,int all)588 vxlan_ftable_flush(struct vxlan_softc *sc, int all)
589 {
590 struct vxlan_ftable_entry *fe, *tfe;
591 int i;
592
593 for (i = 0; i < VXLAN_SC_FTABLE_SIZE; i++) {
594 LIST_FOREACH_SAFE(fe, &sc->vxl_ftable[i], vxlfe_hash, tfe) {
595 if (all || VXLAN_FE_IS_DYNAMIC(fe))
596 vxlan_ftable_entry_destroy(sc, fe);
597 }
598 }
599 }
600
601 static void
vxlan_ftable_expire(struct vxlan_softc * sc)602 vxlan_ftable_expire(struct vxlan_softc *sc)
603 {
604 struct vxlan_ftable_entry *fe, *tfe;
605 int i;
606
607 VXLAN_LOCK_WASSERT(sc);
608
609 for (i = 0; i < VXLAN_SC_FTABLE_SIZE; i++) {
610 LIST_FOREACH_SAFE(fe, &sc->vxl_ftable[i], vxlfe_hash, tfe) {
611 if (VXLAN_FE_IS_DYNAMIC(fe) &&
612 time_uptime >= fe->vxlfe_expire)
613 vxlan_ftable_entry_destroy(sc, fe);
614 }
615 }
616 }
617
618 static int
vxlan_ftable_update_locked(struct vxlan_softc * sc,const union vxlan_sockaddr * vxlsa,const uint8_t * mac,struct rm_priotracker * tracker)619 vxlan_ftable_update_locked(struct vxlan_softc *sc,
620 const union vxlan_sockaddr *vxlsa, const uint8_t *mac,
621 struct rm_priotracker *tracker)
622 {
623 struct vxlan_ftable_entry *fe;
624 int error __unused;
625
626 VXLAN_LOCK_ASSERT(sc);
627
628 again:
629 /*
630 * A forwarding entry for this MAC address might already exist. If
631 * so, update it, otherwise create a new one. We may have to upgrade
632 * the lock if we have to change or create an entry.
633 */
634 fe = vxlan_ftable_entry_lookup(sc, mac);
635 if (fe != NULL) {
636 fe->vxlfe_expire = time_uptime + sc->vxl_ftable_timeout;
637
638 if (!VXLAN_FE_IS_DYNAMIC(fe) ||
639 vxlan_sockaddr_in_equal(&fe->vxlfe_raddr, &vxlsa->sa))
640 return (0);
641 if (!VXLAN_LOCK_WOWNED(sc)) {
642 VXLAN_RUNLOCK(sc, tracker);
643 VXLAN_WLOCK(sc);
644 sc->vxl_stats.ftable_lock_upgrade_failed++;
645 goto again;
646 }
647 vxlan_sockaddr_in_copy(&fe->vxlfe_raddr, &vxlsa->sa);
648 return (0);
649 }
650
651 if (!VXLAN_LOCK_WOWNED(sc)) {
652 VXLAN_RUNLOCK(sc, tracker);
653 VXLAN_WLOCK(sc);
654 sc->vxl_stats.ftable_lock_upgrade_failed++;
655 goto again;
656 }
657
658 if (sc->vxl_ftable_cnt >= sc->vxl_ftable_max) {
659 sc->vxl_stats.ftable_nospace++;
660 return (ENOSPC);
661 }
662
663 fe = vxlan_ftable_entry_alloc();
664 if (fe == NULL)
665 return (ENOMEM);
666
667 vxlan_ftable_entry_init(sc, fe, mac, &vxlsa->sa, VXLAN_FE_FLAG_DYNAMIC);
668
669 /* The prior lookup failed, so the insert should not. */
670 error = vxlan_ftable_entry_insert(sc, fe);
671 MPASS(error == 0);
672
673 return (0);
674 }
675
676 static int
vxlan_ftable_learn(struct vxlan_softc * sc,const struct sockaddr * sa,const uint8_t * mac)677 vxlan_ftable_learn(struct vxlan_softc *sc, const struct sockaddr *sa,
678 const uint8_t *mac)
679 {
680 struct rm_priotracker tracker;
681 union vxlan_sockaddr vxlsa;
682 int error;
683
684 /*
685 * The source port may be randomly selected by the remote host, so
686 * use the port of the default destination address.
687 */
688 vxlan_sockaddr_copy(&vxlsa, sa);
689 vxlsa.in4.sin_port = sc->vxl_dst_addr.in4.sin_port;
690
691 if (VXLAN_SOCKADDR_IS_IPV6(&vxlsa)) {
692 error = vxlan_sockaddr_in6_embedscope(&vxlsa);
693 if (error)
694 return (error);
695 }
696
697 VXLAN_RLOCK(sc, &tracker);
698 error = vxlan_ftable_update_locked(sc, &vxlsa, mac, &tracker);
699 VXLAN_UNLOCK(sc, &tracker);
700
701 return (error);
702 }
703
704 static int
vxlan_ftable_sysctl_dump(SYSCTL_HANDLER_ARGS)705 vxlan_ftable_sysctl_dump(SYSCTL_HANDLER_ARGS)
706 {
707 struct rm_priotracker tracker;
708 struct sbuf sb;
709 struct vxlan_softc *sc;
710 struct vxlan_ftable_entry *fe;
711 size_t size;
712 int i, error;
713
714 /*
715 * This is mostly intended for debugging during development. It is
716 * not practical to dump an entire large table this way.
717 */
718
719 sc = arg1;
720 size = PAGE_SIZE; /* Calculate later. */
721
722 sbuf_new(&sb, NULL, size, SBUF_FIXEDLEN);
723 sbuf_putc(&sb, '\n');
724
725 VXLAN_RLOCK(sc, &tracker);
726 for (i = 0; i < VXLAN_SC_FTABLE_SIZE; i++) {
727 LIST_FOREACH(fe, &sc->vxl_ftable[i], vxlfe_hash) {
728 if (sbuf_error(&sb) != 0)
729 break;
730 vxlan_ftable_entry_dump(fe, &sb);
731 }
732 }
733 VXLAN_RUNLOCK(sc, &tracker);
734
735 if (sbuf_len(&sb) == 1)
736 sbuf_setpos(&sb, 0);
737
738 sbuf_finish(&sb);
739 error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
740 sbuf_delete(&sb);
741
742 return (error);
743 }
744
745 static struct vxlan_ftable_entry *
vxlan_ftable_entry_alloc(void)746 vxlan_ftable_entry_alloc(void)
747 {
748 struct vxlan_ftable_entry *fe;
749
750 fe = malloc(sizeof(*fe), M_VXLAN, M_ZERO | M_NOWAIT);
751
752 return (fe);
753 }
754
755 static void
vxlan_ftable_entry_free(struct vxlan_ftable_entry * fe)756 vxlan_ftable_entry_free(struct vxlan_ftable_entry *fe)
757 {
758
759 free(fe, M_VXLAN);
760 }
761
762 static void
vxlan_ftable_entry_init(struct vxlan_softc * sc,struct vxlan_ftable_entry * fe,const uint8_t * mac,const struct sockaddr * sa,uint32_t flags)763 vxlan_ftable_entry_init(struct vxlan_softc *sc, struct vxlan_ftable_entry *fe,
764 const uint8_t *mac, const struct sockaddr *sa, uint32_t flags)
765 {
766
767 fe->vxlfe_flags = flags;
768 fe->vxlfe_expire = time_uptime + sc->vxl_ftable_timeout;
769 memcpy(fe->vxlfe_mac, mac, ETHER_ADDR_LEN);
770 vxlan_sockaddr_copy(&fe->vxlfe_raddr, sa);
771 }
772
773 static void
vxlan_ftable_entry_destroy(struct vxlan_softc * sc,struct vxlan_ftable_entry * fe)774 vxlan_ftable_entry_destroy(struct vxlan_softc *sc,
775 struct vxlan_ftable_entry *fe)
776 {
777
778 sc->vxl_ftable_cnt--;
779 LIST_REMOVE(fe, vxlfe_hash);
780 vxlan_ftable_entry_free(fe);
781 }
782
783 static int
vxlan_ftable_entry_insert(struct vxlan_softc * sc,struct vxlan_ftable_entry * fe)784 vxlan_ftable_entry_insert(struct vxlan_softc *sc,
785 struct vxlan_ftable_entry *fe)
786 {
787 struct vxlan_ftable_entry *lfe;
788 uint32_t hash;
789 int dir;
790
791 VXLAN_LOCK_WASSERT(sc);
792 hash = VXLAN_SC_FTABLE_HASH(sc, fe->vxlfe_mac);
793
794 lfe = LIST_FIRST(&sc->vxl_ftable[hash]);
795 if (lfe == NULL) {
796 LIST_INSERT_HEAD(&sc->vxl_ftable[hash], fe, vxlfe_hash);
797 goto out;
798 }
799
800 do {
801 dir = vxlan_ftable_addr_cmp(fe->vxlfe_mac, lfe->vxlfe_mac);
802 if (dir == 0)
803 return (EEXIST);
804 if (dir > 0) {
805 LIST_INSERT_BEFORE(lfe, fe, vxlfe_hash);
806 goto out;
807 } else if (LIST_NEXT(lfe, vxlfe_hash) == NULL) {
808 LIST_INSERT_AFTER(lfe, fe, vxlfe_hash);
809 goto out;
810 } else
811 lfe = LIST_NEXT(lfe, vxlfe_hash);
812 } while (lfe != NULL);
813
814 out:
815 sc->vxl_ftable_cnt++;
816
817 return (0);
818 }
819
820 static struct vxlan_ftable_entry *
vxlan_ftable_entry_lookup(struct vxlan_softc * sc,const uint8_t * mac)821 vxlan_ftable_entry_lookup(struct vxlan_softc *sc, const uint8_t *mac)
822 {
823 struct vxlan_ftable_entry *fe;
824 uint32_t hash;
825 int dir;
826
827 VXLAN_LOCK_ASSERT(sc);
828 hash = VXLAN_SC_FTABLE_HASH(sc, mac);
829
830 LIST_FOREACH(fe, &sc->vxl_ftable[hash], vxlfe_hash) {
831 dir = vxlan_ftable_addr_cmp(mac, fe->vxlfe_mac);
832 if (dir == 0)
833 return (fe);
834 if (dir > 0)
835 break;
836 }
837
838 return (NULL);
839 }
840
841 static void
vxlan_ftable_entry_dump(struct vxlan_ftable_entry * fe,struct sbuf * sb)842 vxlan_ftable_entry_dump(struct vxlan_ftable_entry *fe, struct sbuf *sb)
843 {
844 char buf[64];
845 const union vxlan_sockaddr *sa;
846 const void *addr;
847 int i, len, af, width;
848
849 sa = &fe->vxlfe_raddr;
850 af = sa->sa.sa_family;
851 len = sbuf_len(sb);
852
853 sbuf_printf(sb, "%c 0x%02X ", VXLAN_FE_IS_DYNAMIC(fe) ? 'D' : 'S',
854 fe->vxlfe_flags);
855
856 for (i = 0; i < ETHER_ADDR_LEN - 1; i++)
857 sbuf_printf(sb, "%02X:", fe->vxlfe_mac[i]);
858 sbuf_printf(sb, "%02X ", fe->vxlfe_mac[i]);
859
860 if (af == AF_INET) {
861 addr = &sa->in4.sin_addr;
862 width = INET_ADDRSTRLEN - 1;
863 } else {
864 addr = &sa->in6.sin6_addr;
865 width = INET6_ADDRSTRLEN - 1;
866 }
867 inet_ntop(af, addr, buf, sizeof(buf));
868 sbuf_printf(sb, "%*s ", width, buf);
869
870 sbuf_printf(sb, "%08jd", (intmax_t)fe->vxlfe_expire);
871
872 sbuf_putc(sb, '\n');
873
874 /* Truncate a partial line. */
875 if (sbuf_error(sb) != 0)
876 sbuf_setpos(sb, len);
877 }
878
879 static struct vxlan_socket *
vxlan_socket_alloc(const union vxlan_sockaddr * sa)880 vxlan_socket_alloc(const union vxlan_sockaddr *sa)
881 {
882 struct vxlan_socket *vso;
883 int i;
884
885 vso = malloc(sizeof(*vso), M_VXLAN, M_WAITOK | M_ZERO);
886 rm_init(&vso->vxlso_lock, "vxlansorm");
887 refcount_init(&vso->vxlso_refcnt, 0);
888 for (i = 0; i < VXLAN_SO_VNI_HASH_SIZE; i++)
889 LIST_INIT(&vso->vxlso_vni_hash[i]);
890 vso->vxlso_laddr = *sa;
891
892 return (vso);
893 }
894
895 static void
vxlan_socket_destroy(struct vxlan_socket * vso)896 vxlan_socket_destroy(struct vxlan_socket *vso)
897 {
898 struct socket *so;
899 #ifdef INVARIANTS
900 int i;
901 struct vxlan_socket_mc_info *mc;
902
903 for (i = 0; i < VXLAN_SO_MC_MAX_GROUPS; i++) {
904 mc = &vso->vxlso_mc[i];
905 KASSERT(mc->vxlsomc_gaddr.sa.sa_family == AF_UNSPEC,
906 ("%s: socket %p mc[%d] still has address",
907 __func__, vso, i));
908 }
909
910 for (i = 0; i < VXLAN_SO_VNI_HASH_SIZE; i++) {
911 KASSERT(LIST_EMPTY(&vso->vxlso_vni_hash[i]),
912 ("%s: socket %p vni_hash[%d] not empty",
913 __func__, vso, i));
914 }
915 #endif
916 so = vso->vxlso_sock;
917 if (so != NULL) {
918 vso->vxlso_sock = NULL;
919 soclose(so);
920 }
921
922 rm_destroy(&vso->vxlso_lock);
923 free(vso, M_VXLAN);
924 }
925
926 static void
vxlan_socket_release(struct vxlan_socket * vso)927 vxlan_socket_release(struct vxlan_socket *vso)
928 {
929 int destroy;
930
931 VXLAN_LIST_LOCK();
932 destroy = VXLAN_SO_RELEASE(vso);
933 if (destroy != 0)
934 LIST_REMOVE(vso, vxlso_entry);
935 VXLAN_LIST_UNLOCK();
936
937 if (destroy != 0)
938 vxlan_socket_destroy(vso);
939 }
940
941 static struct vxlan_socket *
vxlan_socket_lookup(union vxlan_sockaddr * vxlsa)942 vxlan_socket_lookup(union vxlan_sockaddr *vxlsa)
943 {
944 struct vxlan_socket *vso;
945
946 VXLAN_LIST_LOCK();
947 LIST_FOREACH(vso, &vxlan_socket_list, vxlso_entry) {
948 if (vxlan_sockaddr_cmp(&vso->vxlso_laddr, &vxlsa->sa) == 0) {
949 VXLAN_SO_ACQUIRE(vso);
950 break;
951 }
952 }
953 VXLAN_LIST_UNLOCK();
954
955 return (vso);
956 }
957
958 static void
vxlan_socket_insert(struct vxlan_socket * vso)959 vxlan_socket_insert(struct vxlan_socket *vso)
960 {
961
962 VXLAN_LIST_LOCK();
963 VXLAN_SO_ACQUIRE(vso);
964 LIST_INSERT_HEAD(&vxlan_socket_list, vso, vxlso_entry);
965 VXLAN_LIST_UNLOCK();
966 }
967
968 static int
vxlan_socket_init(struct vxlan_socket * vso,struct ifnet * ifp)969 vxlan_socket_init(struct vxlan_socket *vso, struct ifnet *ifp)
970 {
971 struct thread *td;
972 int error;
973
974 td = curthread;
975
976 error = socreate(vso->vxlso_laddr.sa.sa_family, &vso->vxlso_sock,
977 SOCK_DGRAM, IPPROTO_UDP, td->td_ucred, td);
978 if (error) {
979 if_printf(ifp, "cannot create socket: %d\n", error);
980 return (error);
981 }
982
983 error = udp_set_kernel_tunneling(vso->vxlso_sock,
984 vxlan_rcv_udp_packet, NULL, vso);
985 if (error) {
986 if_printf(ifp, "cannot set tunneling function: %d\n", error);
987 return (error);
988 }
989
990 if (vxlan_reuse_port != 0) {
991 struct sockopt sopt;
992 int val = 1;
993
994 bzero(&sopt, sizeof(sopt));
995 sopt.sopt_dir = SOPT_SET;
996 sopt.sopt_level = IPPROTO_IP;
997 sopt.sopt_name = SO_REUSEPORT;
998 sopt.sopt_val = &val;
999 sopt.sopt_valsize = sizeof(val);
1000 error = sosetopt(vso->vxlso_sock, &sopt);
1001 if (error) {
1002 if_printf(ifp,
1003 "cannot set REUSEADDR socket opt: %d\n", error);
1004 return (error);
1005 }
1006 }
1007
1008 return (0);
1009 }
1010
1011 static int
vxlan_socket_bind(struct vxlan_socket * vso,struct ifnet * ifp)1012 vxlan_socket_bind(struct vxlan_socket *vso, struct ifnet *ifp)
1013 {
1014 union vxlan_sockaddr laddr;
1015 struct thread *td;
1016 int error;
1017
1018 td = curthread;
1019 laddr = vso->vxlso_laddr;
1020
1021 error = sobind(vso->vxlso_sock, &laddr.sa, td);
1022 if (error) {
1023 if (error != EADDRINUSE)
1024 if_printf(ifp, "cannot bind socket: %d\n", error);
1025 return (error);
1026 }
1027
1028 return (0);
1029 }
1030
1031 static int
vxlan_socket_create(struct ifnet * ifp,int multicast,const union vxlan_sockaddr * saddr,struct vxlan_socket ** vsop)1032 vxlan_socket_create(struct ifnet *ifp, int multicast,
1033 const union vxlan_sockaddr *saddr, struct vxlan_socket **vsop)
1034 {
1035 union vxlan_sockaddr laddr;
1036 struct vxlan_socket *vso;
1037 int error;
1038
1039 laddr = *saddr;
1040
1041 /*
1042 * If this socket will be multicast, then only the local port
1043 * must be specified when binding.
1044 */
1045 if (multicast != 0) {
1046 if (VXLAN_SOCKADDR_IS_IPV4(&laddr))
1047 laddr.in4.sin_addr.s_addr = INADDR_ANY;
1048 #ifdef INET6
1049 else
1050 laddr.in6.sin6_addr = in6addr_any;
1051 #endif
1052 }
1053
1054 vso = vxlan_socket_alloc(&laddr);
1055 if (vso == NULL)
1056 return (ENOMEM);
1057
1058 error = vxlan_socket_init(vso, ifp);
1059 if (error)
1060 goto fail;
1061
1062 error = vxlan_socket_bind(vso, ifp);
1063 if (error)
1064 goto fail;
1065
1066 /*
1067 * There is a small window between the bind completing and
1068 * inserting the socket, so that a concurrent create may fail.
1069 * Let's not worry about that for now.
1070 */
1071 vxlan_socket_insert(vso);
1072 *vsop = vso;
1073
1074 return (0);
1075
1076 fail:
1077 vxlan_socket_destroy(vso);
1078
1079 return (error);
1080 }
1081
1082 static void
vxlan_socket_ifdetach(struct vxlan_socket * vso,struct ifnet * ifp,struct vxlan_softc_head * list)1083 vxlan_socket_ifdetach(struct vxlan_socket *vso, struct ifnet *ifp,
1084 struct vxlan_softc_head *list)
1085 {
1086 struct rm_priotracker tracker;
1087 struct vxlan_softc *sc;
1088 int i;
1089
1090 VXLAN_SO_RLOCK(vso, &tracker);
1091 for (i = 0; i < VXLAN_SO_VNI_HASH_SIZE; i++) {
1092 LIST_FOREACH(sc, &vso->vxlso_vni_hash[i], vxl_entry)
1093 vxlan_ifdetach(sc, ifp, list);
1094 }
1095 VXLAN_SO_RUNLOCK(vso, &tracker);
1096 }
1097
1098 static struct vxlan_socket *
vxlan_socket_mc_lookup(const union vxlan_sockaddr * vxlsa)1099 vxlan_socket_mc_lookup(const union vxlan_sockaddr *vxlsa)
1100 {
1101 union vxlan_sockaddr laddr;
1102 struct vxlan_socket *vso;
1103
1104 laddr = *vxlsa;
1105
1106 if (VXLAN_SOCKADDR_IS_IPV4(&laddr))
1107 laddr.in4.sin_addr.s_addr = INADDR_ANY;
1108 #ifdef INET6
1109 else
1110 laddr.in6.sin6_addr = in6addr_any;
1111 #endif
1112
1113 vso = vxlan_socket_lookup(&laddr);
1114
1115 return (vso);
1116 }
1117
1118 static int
vxlan_sockaddr_mc_info_match(const struct vxlan_socket_mc_info * mc,const union vxlan_sockaddr * group,const union vxlan_sockaddr * local,int ifidx)1119 vxlan_sockaddr_mc_info_match(const struct vxlan_socket_mc_info *mc,
1120 const union vxlan_sockaddr *group, const union vxlan_sockaddr *local,
1121 int ifidx)
1122 {
1123
1124 if (!vxlan_sockaddr_in_any(local) &&
1125 !vxlan_sockaddr_in_equal(&mc->vxlsomc_saddr, &local->sa))
1126 return (0);
1127 if (!vxlan_sockaddr_in_equal(&mc->vxlsomc_gaddr, &group->sa))
1128 return (0);
1129 if (ifidx != 0 && ifidx != mc->vxlsomc_ifidx)
1130 return (0);
1131
1132 return (1);
1133 }
1134
1135 static int
vxlan_socket_mc_join_group(struct vxlan_socket * vso,const union vxlan_sockaddr * group,const union vxlan_sockaddr * local,int * ifidx,union vxlan_sockaddr * source)1136 vxlan_socket_mc_join_group(struct vxlan_socket *vso,
1137 const union vxlan_sockaddr *group, const union vxlan_sockaddr *local,
1138 int *ifidx, union vxlan_sockaddr *source)
1139 {
1140 struct sockopt sopt;
1141 int error;
1142
1143 *source = *local;
1144
1145 if (VXLAN_SOCKADDR_IS_IPV4(group)) {
1146 struct ip_mreq mreq;
1147
1148 mreq.imr_multiaddr = group->in4.sin_addr;
1149 mreq.imr_interface = local->in4.sin_addr;
1150
1151 bzero(&sopt, sizeof(sopt));
1152 sopt.sopt_dir = SOPT_SET;
1153 sopt.sopt_level = IPPROTO_IP;
1154 sopt.sopt_name = IP_ADD_MEMBERSHIP;
1155 sopt.sopt_val = &mreq;
1156 sopt.sopt_valsize = sizeof(mreq);
1157 error = sosetopt(vso->vxlso_sock, &sopt);
1158 if (error)
1159 return (error);
1160
1161 /*
1162 * BMV: Ideally, there would be a formal way for us to get
1163 * the local interface that was selected based on the
1164 * imr_interface address. We could then update *ifidx so
1165 * vxlan_sockaddr_mc_info_match() would return a match for
1166 * later creates that explicitly set the multicast interface.
1167 *
1168 * If we really need to, we can of course look in the INP's
1169 * membership list:
1170 * sotoinpcb(vso->vxlso_sock)->inp_moptions->
1171 * imo_head[]->imf_inm->inm_ifp
1172 * similarly to imo_match_group().
1173 */
1174 source->in4.sin_addr = local->in4.sin_addr;
1175
1176 } else if (VXLAN_SOCKADDR_IS_IPV6(group)) {
1177 struct ipv6_mreq mreq;
1178
1179 mreq.ipv6mr_multiaddr = group->in6.sin6_addr;
1180 mreq.ipv6mr_interface = *ifidx;
1181
1182 bzero(&sopt, sizeof(sopt));
1183 sopt.sopt_dir = SOPT_SET;
1184 sopt.sopt_level = IPPROTO_IPV6;
1185 sopt.sopt_name = IPV6_JOIN_GROUP;
1186 sopt.sopt_val = &mreq;
1187 sopt.sopt_valsize = sizeof(mreq);
1188 error = sosetopt(vso->vxlso_sock, &sopt);
1189 if (error)
1190 return (error);
1191
1192 /*
1193 * BMV: As with IPv4, we would really like to know what
1194 * interface in6p_lookup_mcast_ifp() selected.
1195 */
1196 } else
1197 error = EAFNOSUPPORT;
1198
1199 return (error);
1200 }
1201
1202 static int
vxlan_socket_mc_leave_group(struct vxlan_socket * vso,const union vxlan_sockaddr * group,const union vxlan_sockaddr * source,int ifidx)1203 vxlan_socket_mc_leave_group(struct vxlan_socket *vso,
1204 const union vxlan_sockaddr *group, const union vxlan_sockaddr *source,
1205 int ifidx)
1206 {
1207 struct sockopt sopt;
1208 int error;
1209
1210 bzero(&sopt, sizeof(sopt));
1211 sopt.sopt_dir = SOPT_SET;
1212
1213 if (VXLAN_SOCKADDR_IS_IPV4(group)) {
1214 struct ip_mreq mreq;
1215
1216 mreq.imr_multiaddr = group->in4.sin_addr;
1217 mreq.imr_interface = source->in4.sin_addr;
1218
1219 sopt.sopt_level = IPPROTO_IP;
1220 sopt.sopt_name = IP_DROP_MEMBERSHIP;
1221 sopt.sopt_val = &mreq;
1222 sopt.sopt_valsize = sizeof(mreq);
1223 error = sosetopt(vso->vxlso_sock, &sopt);
1224
1225 } else if (VXLAN_SOCKADDR_IS_IPV6(group)) {
1226 struct ipv6_mreq mreq;
1227
1228 mreq.ipv6mr_multiaddr = group->in6.sin6_addr;
1229 mreq.ipv6mr_interface = ifidx;
1230
1231 sopt.sopt_level = IPPROTO_IPV6;
1232 sopt.sopt_name = IPV6_LEAVE_GROUP;
1233 sopt.sopt_val = &mreq;
1234 sopt.sopt_valsize = sizeof(mreq);
1235 error = sosetopt(vso->vxlso_sock, &sopt);
1236
1237 } else
1238 error = EAFNOSUPPORT;
1239
1240 return (error);
1241 }
1242
1243 static int
vxlan_socket_mc_add_group(struct vxlan_socket * vso,const union vxlan_sockaddr * group,const union vxlan_sockaddr * local,int ifidx,int * idx)1244 vxlan_socket_mc_add_group(struct vxlan_socket *vso,
1245 const union vxlan_sockaddr *group, const union vxlan_sockaddr *local,
1246 int ifidx, int *idx)
1247 {
1248 union vxlan_sockaddr source;
1249 struct vxlan_socket_mc_info *mc;
1250 int i, empty, error;
1251
1252 /*
1253 * Within a socket, the same multicast group may be used by multiple
1254 * interfaces, each with a different network identifier. But a socket
1255 * may only join a multicast group once, so keep track of the users
1256 * here.
1257 */
1258
1259 VXLAN_SO_WLOCK(vso);
1260 for (empty = 0, i = 0; i < VXLAN_SO_MC_MAX_GROUPS; i++) {
1261 mc = &vso->vxlso_mc[i];
1262
1263 if (mc->vxlsomc_gaddr.sa.sa_family == AF_UNSPEC) {
1264 empty++;
1265 continue;
1266 }
1267
1268 if (vxlan_sockaddr_mc_info_match(mc, group, local, ifidx))
1269 goto out;
1270 }
1271 VXLAN_SO_WUNLOCK(vso);
1272
1273 if (empty == 0)
1274 return (ENOSPC);
1275
1276 error = vxlan_socket_mc_join_group(vso, group, local, &ifidx, &source);
1277 if (error)
1278 return (error);
1279
1280 VXLAN_SO_WLOCK(vso);
1281 for (i = 0; i < VXLAN_SO_MC_MAX_GROUPS; i++) {
1282 mc = &vso->vxlso_mc[i];
1283
1284 if (mc->vxlsomc_gaddr.sa.sa_family == AF_UNSPEC) {
1285 vxlan_sockaddr_copy(&mc->vxlsomc_gaddr, &group->sa);
1286 vxlan_sockaddr_copy(&mc->vxlsomc_saddr, &source.sa);
1287 mc->vxlsomc_ifidx = ifidx;
1288 goto out;
1289 }
1290 }
1291 VXLAN_SO_WUNLOCK(vso);
1292
1293 error = vxlan_socket_mc_leave_group(vso, group, &source, ifidx);
1294 MPASS(error == 0);
1295
1296 return (ENOSPC);
1297
1298 out:
1299 mc->vxlsomc_users++;
1300 VXLAN_SO_WUNLOCK(vso);
1301
1302 *idx = i;
1303
1304 return (0);
1305 }
1306
1307 static void
vxlan_socket_mc_release_group_by_idx(struct vxlan_socket * vso,int idx)1308 vxlan_socket_mc_release_group_by_idx(struct vxlan_socket *vso, int idx)
1309 {
1310 union vxlan_sockaddr group, source;
1311 struct vxlan_socket_mc_info *mc;
1312 int ifidx, leave;
1313
1314 KASSERT(idx >= 0 && idx < VXLAN_SO_MC_MAX_GROUPS,
1315 ("%s: vso %p idx %d out of bounds", __func__, vso, idx));
1316
1317 leave = 0;
1318 mc = &vso->vxlso_mc[idx];
1319
1320 VXLAN_SO_WLOCK(vso);
1321 mc->vxlsomc_users--;
1322 if (mc->vxlsomc_users == 0) {
1323 group = mc->vxlsomc_gaddr;
1324 source = mc->vxlsomc_saddr;
1325 ifidx = mc->vxlsomc_ifidx;
1326 bzero(mc, sizeof(*mc));
1327 leave = 1;
1328 }
1329 VXLAN_SO_WUNLOCK(vso);
1330
1331 if (leave != 0) {
1332 /*
1333 * Our socket's membership in this group may have already
1334 * been removed if we joined through an interface that's
1335 * been detached.
1336 */
1337 vxlan_socket_mc_leave_group(vso, &group, &source, ifidx);
1338 }
1339 }
1340
1341 static struct vxlan_softc *
vxlan_socket_lookup_softc_locked(struct vxlan_socket * vso,uint32_t vni)1342 vxlan_socket_lookup_softc_locked(struct vxlan_socket *vso, uint32_t vni)
1343 {
1344 struct vxlan_softc *sc;
1345 uint32_t hash;
1346
1347 VXLAN_SO_LOCK_ASSERT(vso);
1348 hash = VXLAN_SO_VNI_HASH(vni);
1349
1350 LIST_FOREACH(sc, &vso->vxlso_vni_hash[hash], vxl_entry) {
1351 if (sc->vxl_vni == vni) {
1352 VXLAN_ACQUIRE(sc);
1353 break;
1354 }
1355 }
1356
1357 return (sc);
1358 }
1359
1360 static struct vxlan_softc *
vxlan_socket_lookup_softc(struct vxlan_socket * vso,uint32_t vni)1361 vxlan_socket_lookup_softc(struct vxlan_socket *vso, uint32_t vni)
1362 {
1363 struct rm_priotracker tracker;
1364 struct vxlan_softc *sc;
1365
1366 VXLAN_SO_RLOCK(vso, &tracker);
1367 sc = vxlan_socket_lookup_softc_locked(vso, vni);
1368 VXLAN_SO_RUNLOCK(vso, &tracker);
1369
1370 return (sc);
1371 }
1372
1373 static int
vxlan_socket_insert_softc(struct vxlan_socket * vso,struct vxlan_softc * sc)1374 vxlan_socket_insert_softc(struct vxlan_socket *vso, struct vxlan_softc *sc)
1375 {
1376 struct vxlan_softc *tsc;
1377 uint32_t vni, hash;
1378
1379 vni = sc->vxl_vni;
1380 hash = VXLAN_SO_VNI_HASH(vni);
1381
1382 VXLAN_SO_WLOCK(vso);
1383 tsc = vxlan_socket_lookup_softc_locked(vso, vni);
1384 if (tsc != NULL) {
1385 VXLAN_SO_WUNLOCK(vso);
1386 vxlan_release(tsc);
1387 return (EEXIST);
1388 }
1389
1390 VXLAN_ACQUIRE(sc);
1391 LIST_INSERT_HEAD(&vso->vxlso_vni_hash[hash], sc, vxl_entry);
1392 VXLAN_SO_WUNLOCK(vso);
1393
1394 return (0);
1395 }
1396
1397 static void
vxlan_socket_remove_softc(struct vxlan_socket * vso,struct vxlan_softc * sc)1398 vxlan_socket_remove_softc(struct vxlan_socket *vso, struct vxlan_softc *sc)
1399 {
1400
1401 VXLAN_SO_WLOCK(vso);
1402 LIST_REMOVE(sc, vxl_entry);
1403 VXLAN_SO_WUNLOCK(vso);
1404
1405 vxlan_release(sc);
1406 }
1407
1408 static struct ifnet *
vxlan_multicast_if_ref(struct vxlan_softc * sc,int ipv4)1409 vxlan_multicast_if_ref(struct vxlan_softc *sc, int ipv4)
1410 {
1411 struct ifnet *ifp;
1412
1413 VXLAN_LOCK_ASSERT(sc);
1414
1415 if (ipv4 && sc->vxl_im4o != NULL)
1416 ifp = sc->vxl_im4o->imo_multicast_ifp;
1417 else if (!ipv4 && sc->vxl_im6o != NULL)
1418 ifp = sc->vxl_im6o->im6o_multicast_ifp;
1419 else
1420 ifp = NULL;
1421
1422 if (ifp != NULL)
1423 if_ref(ifp);
1424
1425 return (ifp);
1426 }
1427
1428 static void
vxlan_free_multicast(struct vxlan_softc * sc)1429 vxlan_free_multicast(struct vxlan_softc *sc)
1430 {
1431
1432 if (sc->vxl_mc_ifp != NULL) {
1433 if_rele(sc->vxl_mc_ifp);
1434 sc->vxl_mc_ifp = NULL;
1435 sc->vxl_mc_ifindex = 0;
1436 }
1437
1438 if (sc->vxl_im4o != NULL) {
1439 free(sc->vxl_im4o, M_VXLAN);
1440 sc->vxl_im4o = NULL;
1441 }
1442
1443 if (sc->vxl_im6o != NULL) {
1444 free(sc->vxl_im6o, M_VXLAN);
1445 sc->vxl_im6o = NULL;
1446 }
1447 }
1448
1449 static int
vxlan_setup_multicast_interface(struct vxlan_softc * sc)1450 vxlan_setup_multicast_interface(struct vxlan_softc *sc)
1451 {
1452 struct ifnet *ifp;
1453
1454 ifp = ifunit_ref(sc->vxl_mc_ifname);
1455 if (ifp == NULL) {
1456 if_printf(sc->vxl_ifp, "multicast interface %s does "
1457 "not exist\n", sc->vxl_mc_ifname);
1458 return (ENOENT);
1459 }
1460
1461 if ((ifp->if_flags & IFF_MULTICAST) == 0) {
1462 if_printf(sc->vxl_ifp, "interface %s does not support "
1463 "multicast\n", sc->vxl_mc_ifname);
1464 if_rele(ifp);
1465 return (ENOTSUP);
1466 }
1467
1468 sc->vxl_mc_ifp = ifp;
1469 sc->vxl_mc_ifindex = ifp->if_index;
1470
1471 return (0);
1472 }
1473
1474 static int
vxlan_setup_multicast(struct vxlan_softc * sc)1475 vxlan_setup_multicast(struct vxlan_softc *sc)
1476 {
1477 const union vxlan_sockaddr *group;
1478 int error;
1479
1480 group = &sc->vxl_dst_addr;
1481 error = 0;
1482
1483 if (sc->vxl_mc_ifname[0] != '\0') {
1484 error = vxlan_setup_multicast_interface(sc);
1485 if (error)
1486 return (error);
1487 }
1488
1489 /*
1490 * Initialize an multicast options structure that is sufficiently
1491 * populated for use in the respective IP output routine. This
1492 * structure is typically stored in the socket, but our sockets
1493 * may be shared among multiple interfaces.
1494 */
1495 if (VXLAN_SOCKADDR_IS_IPV4(group)) {
1496 sc->vxl_im4o = malloc(sizeof(struct ip_moptions), M_VXLAN,
1497 M_ZERO | M_WAITOK);
1498 sc->vxl_im4o->imo_multicast_ifp = sc->vxl_mc_ifp;
1499 sc->vxl_im4o->imo_multicast_ttl = sc->vxl_ttl;
1500 sc->vxl_im4o->imo_multicast_vif = -1;
1501 } else if (VXLAN_SOCKADDR_IS_IPV6(group)) {
1502 sc->vxl_im6o = malloc(sizeof(struct ip6_moptions), M_VXLAN,
1503 M_ZERO | M_WAITOK);
1504 sc->vxl_im6o->im6o_multicast_ifp = sc->vxl_mc_ifp;
1505 sc->vxl_im6o->im6o_multicast_hlim = sc->vxl_ttl;
1506 }
1507
1508 return (error);
1509 }
1510
1511 static int
vxlan_setup_socket(struct vxlan_softc * sc)1512 vxlan_setup_socket(struct vxlan_softc *sc)
1513 {
1514 struct vxlan_socket *vso;
1515 struct ifnet *ifp;
1516 union vxlan_sockaddr *saddr, *daddr;
1517 int multicast, error;
1518
1519 vso = NULL;
1520 ifp = sc->vxl_ifp;
1521 saddr = &sc->vxl_src_addr;
1522 daddr = &sc->vxl_dst_addr;
1523
1524 multicast = vxlan_sockaddr_in_multicast(daddr);
1525 MPASS(multicast != -1);
1526 sc->vxl_vso_mc_index = -1;
1527
1528 /*
1529 * Try to create the socket. If that fails, attempt to use an
1530 * existing socket.
1531 */
1532 error = vxlan_socket_create(ifp, multicast, saddr, &vso);
1533 if (error) {
1534 if (multicast != 0)
1535 vso = vxlan_socket_mc_lookup(saddr);
1536 else
1537 vso = vxlan_socket_lookup(saddr);
1538
1539 if (vso == NULL) {
1540 if_printf(ifp, "cannot create socket (error: %d), "
1541 "and no existing socket found\n", error);
1542 goto out;
1543 }
1544 }
1545
1546 if (multicast != 0) {
1547 error = vxlan_setup_multicast(sc);
1548 if (error)
1549 goto out;
1550
1551 error = vxlan_socket_mc_add_group(vso, daddr, saddr,
1552 sc->vxl_mc_ifindex, &sc->vxl_vso_mc_index);
1553 if (error)
1554 goto out;
1555 }
1556
1557 sc->vxl_sock = vso;
1558 error = vxlan_socket_insert_softc(vso, sc);
1559 if (error) {
1560 sc->vxl_sock = NULL;
1561 if_printf(ifp, "network identifier %d already exists in "
1562 "this socket\n", sc->vxl_vni);
1563 goto out;
1564 }
1565
1566 return (0);
1567
1568 out:
1569 if (vso != NULL) {
1570 if (sc->vxl_vso_mc_index != -1) {
1571 vxlan_socket_mc_release_group_by_idx(vso,
1572 sc->vxl_vso_mc_index);
1573 sc->vxl_vso_mc_index = -1;
1574 }
1575 if (multicast != 0)
1576 vxlan_free_multicast(sc);
1577 vxlan_socket_release(vso);
1578 }
1579
1580 return (error);
1581 }
1582
1583 #ifdef INET6
1584 static void
vxlan_setup_zero_checksum_port(struct vxlan_softc * sc)1585 vxlan_setup_zero_checksum_port(struct vxlan_softc *sc)
1586 {
1587
1588 if (!VXLAN_SOCKADDR_IS_IPV6(&sc->vxl_src_addr))
1589 return;
1590
1591 MPASS(sc->vxl_src_addr.in6.sin6_port != 0);
1592 MPASS(sc->vxl_dst_addr.in6.sin6_port != 0);
1593
1594 if (sc->vxl_src_addr.in6.sin6_port != sc->vxl_dst_addr.in6.sin6_port) {
1595 if_printf(sc->vxl_ifp, "port %d in src address does not match "
1596 "port %d in dst address, rfc6935_port (%d) not updated.\n",
1597 ntohs(sc->vxl_src_addr.in6.sin6_port),
1598 ntohs(sc->vxl_dst_addr.in6.sin6_port),
1599 V_zero_checksum_port);
1600 return;
1601 }
1602
1603 if (V_zero_checksum_port != 0) {
1604 if (V_zero_checksum_port !=
1605 ntohs(sc->vxl_src_addr.in6.sin6_port)) {
1606 if_printf(sc->vxl_ifp, "rfc6935_port is already set to "
1607 "%d, cannot set it to %d.\n", V_zero_checksum_port,
1608 ntohs(sc->vxl_src_addr.in6.sin6_port));
1609 }
1610 return;
1611 }
1612
1613 V_zero_checksum_port = ntohs(sc->vxl_src_addr.in6.sin6_port);
1614 if_printf(sc->vxl_ifp, "rfc6935_port set to %d\n",
1615 V_zero_checksum_port);
1616 }
1617 #endif
1618
1619 static void
vxlan_setup_interface_hdrlen(struct vxlan_softc * sc)1620 vxlan_setup_interface_hdrlen(struct vxlan_softc *sc)
1621 {
1622 struct ifnet *ifp;
1623
1624 VXLAN_LOCK_WASSERT(sc);
1625
1626 ifp = sc->vxl_ifp;
1627 ifp->if_hdrlen = ETHER_HDR_LEN + sizeof(struct vxlanudphdr);
1628
1629 if (VXLAN_SOCKADDR_IS_IPV4(&sc->vxl_dst_addr) != 0)
1630 ifp->if_hdrlen += sizeof(struct ip);
1631 else if (VXLAN_SOCKADDR_IS_IPV6(&sc->vxl_dst_addr) != 0)
1632 ifp->if_hdrlen += sizeof(struct ip6_hdr);
1633
1634 if ((sc->vxl_flags & VXLAN_FLAG_USER_MTU) == 0)
1635 ifp->if_mtu = ETHERMTU - ifp->if_hdrlen;
1636 }
1637
1638 static int
vxlan_valid_init_config(struct vxlan_softc * sc)1639 vxlan_valid_init_config(struct vxlan_softc *sc)
1640 {
1641 const char *reason;
1642
1643 if (vxlan_check_vni(sc->vxl_vni) != 0) {
1644 reason = "invalid virtual network identifier specified";
1645 goto fail;
1646 }
1647
1648 if (vxlan_sockaddr_supported(&sc->vxl_src_addr, 1) == 0) {
1649 reason = "source address type is not supported";
1650 goto fail;
1651 }
1652
1653 if (vxlan_sockaddr_supported(&sc->vxl_dst_addr, 0) == 0) {
1654 reason = "destination address type is not supported";
1655 goto fail;
1656 }
1657
1658 if (vxlan_sockaddr_in_any(&sc->vxl_dst_addr) != 0) {
1659 reason = "no valid destination address specified";
1660 goto fail;
1661 }
1662
1663 if (vxlan_sockaddr_in_multicast(&sc->vxl_dst_addr) == 0 &&
1664 sc->vxl_mc_ifname[0] != '\0') {
1665 reason = "can only specify interface with a group address";
1666 goto fail;
1667 }
1668
1669 if (vxlan_sockaddr_in_any(&sc->vxl_src_addr) == 0) {
1670 if (VXLAN_SOCKADDR_IS_IPV4(&sc->vxl_src_addr) ^
1671 VXLAN_SOCKADDR_IS_IPV4(&sc->vxl_dst_addr)) {
1672 reason = "source and destination address must both "
1673 "be either IPv4 or IPv6";
1674 goto fail;
1675 }
1676 }
1677
1678 if (sc->vxl_src_addr.in4.sin_port == 0) {
1679 reason = "local port not specified";
1680 goto fail;
1681 }
1682
1683 if (sc->vxl_dst_addr.in4.sin_port == 0) {
1684 reason = "remote port not specified";
1685 goto fail;
1686 }
1687
1688 return (0);
1689
1690 fail:
1691 if_printf(sc->vxl_ifp, "cannot initialize interface: %s\n", reason);
1692 return (EINVAL);
1693 }
1694
1695 static void
vxlan_init_wait(struct vxlan_softc * sc)1696 vxlan_init_wait(struct vxlan_softc *sc)
1697 {
1698
1699 VXLAN_LOCK_WASSERT(sc);
1700 while (sc->vxl_flags & VXLAN_FLAG_INIT)
1701 rm_sleep(sc, &sc->vxl_lock, 0, "vxlint", hz);
1702 }
1703
1704 static void
vxlan_init_complete(struct vxlan_softc * sc)1705 vxlan_init_complete(struct vxlan_softc *sc)
1706 {
1707
1708 VXLAN_WLOCK(sc);
1709 sc->vxl_flags &= ~VXLAN_FLAG_INIT;
1710 wakeup(sc);
1711 VXLAN_WUNLOCK(sc);
1712 }
1713
1714 static void
vxlan_init(void * xsc)1715 vxlan_init(void *xsc)
1716 {
1717 static const uint8_t empty_mac[ETHER_ADDR_LEN];
1718 struct vxlan_softc *sc;
1719 struct ifnet *ifp;
1720
1721 sc = xsc;
1722 ifp = sc->vxl_ifp;
1723
1724 sx_xlock(&vxlan_sx);
1725 VXLAN_WLOCK(sc);
1726 if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
1727 VXLAN_WUNLOCK(sc);
1728 sx_xunlock(&vxlan_sx);
1729 return;
1730 }
1731 sc->vxl_flags |= VXLAN_FLAG_INIT;
1732 VXLAN_WUNLOCK(sc);
1733
1734 if (vxlan_valid_init_config(sc) != 0)
1735 goto out;
1736
1737 if (vxlan_setup_socket(sc) != 0)
1738 goto out;
1739
1740 #ifdef INET6
1741 vxlan_setup_zero_checksum_port(sc);
1742 #endif
1743
1744 /* Initialize the default forwarding entry. */
1745 vxlan_ftable_entry_init(sc, &sc->vxl_default_fe, empty_mac,
1746 &sc->vxl_dst_addr.sa, VXLAN_FE_FLAG_STATIC);
1747
1748 VXLAN_WLOCK(sc);
1749 ifp->if_drv_flags |= IFF_DRV_RUNNING;
1750 callout_reset(&sc->vxl_callout, vxlan_ftable_prune_period * hz,
1751 vxlan_timer, sc);
1752 VXLAN_WUNLOCK(sc);
1753
1754 if_link_state_change(ifp, LINK_STATE_UP);
1755
1756 EVENTHANDLER_INVOKE(vxlan_start, ifp, sc->vxl_src_addr.in4.sin_family,
1757 ntohs(sc->vxl_src_addr.in4.sin_port));
1758 out:
1759 vxlan_init_complete(sc);
1760 sx_xunlock(&vxlan_sx);
1761 }
1762
1763 static void
vxlan_release(struct vxlan_softc * sc)1764 vxlan_release(struct vxlan_softc *sc)
1765 {
1766
1767 /*
1768 * The softc may be destroyed as soon as we release our reference,
1769 * so we cannot serialize the wakeup with the softc lock. We use a
1770 * timeout in our sleeps so a missed wakeup is unfortunate but not
1771 * fatal.
1772 */
1773 if (VXLAN_RELEASE(sc) != 0)
1774 wakeup(sc);
1775 }
1776
1777 static void
vxlan_teardown_wait(struct vxlan_softc * sc)1778 vxlan_teardown_wait(struct vxlan_softc *sc)
1779 {
1780
1781 VXLAN_LOCK_WASSERT(sc);
1782 while (sc->vxl_flags & VXLAN_FLAG_TEARDOWN)
1783 rm_sleep(sc, &sc->vxl_lock, 0, "vxltrn", hz);
1784 }
1785
1786 static void
vxlan_teardown_complete(struct vxlan_softc * sc)1787 vxlan_teardown_complete(struct vxlan_softc *sc)
1788 {
1789
1790 VXLAN_WLOCK(sc);
1791 sc->vxl_flags &= ~VXLAN_FLAG_TEARDOWN;
1792 wakeup(sc);
1793 VXLAN_WUNLOCK(sc);
1794 }
1795
1796 static void
vxlan_teardown_locked(struct vxlan_softc * sc)1797 vxlan_teardown_locked(struct vxlan_softc *sc)
1798 {
1799 struct ifnet *ifp;
1800 struct vxlan_socket *vso;
1801
1802 sx_assert(&vxlan_sx, SA_XLOCKED);
1803 VXLAN_LOCK_WASSERT(sc);
1804 MPASS(sc->vxl_flags & VXLAN_FLAG_TEARDOWN);
1805
1806 ifp = sc->vxl_ifp;
1807 ifp->if_flags &= ~IFF_UP;
1808 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
1809 callout_stop(&sc->vxl_callout);
1810 vso = sc->vxl_sock;
1811 sc->vxl_sock = NULL;
1812
1813 VXLAN_WUNLOCK(sc);
1814 if_link_state_change(ifp, LINK_STATE_DOWN);
1815 EVENTHANDLER_INVOKE(vxlan_stop, ifp, sc->vxl_src_addr.in4.sin_family,
1816 ntohs(sc->vxl_src_addr.in4.sin_port));
1817
1818 if (vso != NULL) {
1819 vxlan_socket_remove_softc(vso, sc);
1820
1821 if (sc->vxl_vso_mc_index != -1) {
1822 vxlan_socket_mc_release_group_by_idx(vso,
1823 sc->vxl_vso_mc_index);
1824 sc->vxl_vso_mc_index = -1;
1825 }
1826 }
1827
1828 VXLAN_WLOCK(sc);
1829 while (sc->vxl_refcnt != 0)
1830 rm_sleep(sc, &sc->vxl_lock, 0, "vxldrn", hz);
1831 VXLAN_WUNLOCK(sc);
1832
1833 callout_drain(&sc->vxl_callout);
1834
1835 vxlan_free_multicast(sc);
1836 if (vso != NULL)
1837 vxlan_socket_release(vso);
1838
1839 vxlan_teardown_complete(sc);
1840 }
1841
1842 static void
vxlan_teardown(struct vxlan_softc * sc)1843 vxlan_teardown(struct vxlan_softc *sc)
1844 {
1845
1846 sx_xlock(&vxlan_sx);
1847 VXLAN_WLOCK(sc);
1848 if (sc->vxl_flags & VXLAN_FLAG_TEARDOWN) {
1849 vxlan_teardown_wait(sc);
1850 VXLAN_WUNLOCK(sc);
1851 sx_xunlock(&vxlan_sx);
1852 return;
1853 }
1854
1855 sc->vxl_flags |= VXLAN_FLAG_TEARDOWN;
1856 vxlan_teardown_locked(sc);
1857 sx_xunlock(&vxlan_sx);
1858 }
1859
1860 static void
vxlan_ifdetach(struct vxlan_softc * sc,struct ifnet * ifp,struct vxlan_softc_head * list)1861 vxlan_ifdetach(struct vxlan_softc *sc, struct ifnet *ifp,
1862 struct vxlan_softc_head *list)
1863 {
1864
1865 VXLAN_WLOCK(sc);
1866
1867 if (sc->vxl_mc_ifp != ifp)
1868 goto out;
1869 if (sc->vxl_flags & VXLAN_FLAG_TEARDOWN)
1870 goto out;
1871
1872 sc->vxl_flags |= VXLAN_FLAG_TEARDOWN;
1873 LIST_INSERT_HEAD(list, sc, vxl_ifdetach_list);
1874
1875 out:
1876 VXLAN_WUNLOCK(sc);
1877 }
1878
1879 static void
vxlan_timer(void * xsc)1880 vxlan_timer(void *xsc)
1881 {
1882 struct vxlan_softc *sc;
1883
1884 sc = xsc;
1885 VXLAN_LOCK_WASSERT(sc);
1886
1887 vxlan_ftable_expire(sc);
1888 callout_schedule(&sc->vxl_callout, vxlan_ftable_prune_period * hz);
1889 }
1890
1891 static int
vxlan_ioctl_ifflags(struct vxlan_softc * sc)1892 vxlan_ioctl_ifflags(struct vxlan_softc *sc)
1893 {
1894 struct ifnet *ifp;
1895
1896 ifp = sc->vxl_ifp;
1897
1898 if (ifp->if_flags & IFF_UP) {
1899 if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
1900 vxlan_init(sc);
1901 } else {
1902 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1903 vxlan_teardown(sc);
1904 }
1905
1906 return (0);
1907 }
1908
1909 static int
vxlan_ctrl_get_config(struct vxlan_softc * sc,void * arg)1910 vxlan_ctrl_get_config(struct vxlan_softc *sc, void *arg)
1911 {
1912 struct rm_priotracker tracker;
1913 struct ifvxlancfg *cfg;
1914
1915 cfg = arg;
1916 bzero(cfg, sizeof(*cfg));
1917
1918 VXLAN_RLOCK(sc, &tracker);
1919 cfg->vxlc_vni = sc->vxl_vni;
1920 memcpy(&cfg->vxlc_local_sa, &sc->vxl_src_addr,
1921 sizeof(union vxlan_sockaddr));
1922 memcpy(&cfg->vxlc_remote_sa, &sc->vxl_dst_addr,
1923 sizeof(union vxlan_sockaddr));
1924 cfg->vxlc_mc_ifindex = sc->vxl_mc_ifindex;
1925 cfg->vxlc_ftable_cnt = sc->vxl_ftable_cnt;
1926 cfg->vxlc_ftable_max = sc->vxl_ftable_max;
1927 cfg->vxlc_ftable_timeout = sc->vxl_ftable_timeout;
1928 cfg->vxlc_port_min = sc->vxl_min_port;
1929 cfg->vxlc_port_max = sc->vxl_max_port;
1930 cfg->vxlc_learn = (sc->vxl_flags & VXLAN_FLAG_LEARN) != 0;
1931 cfg->vxlc_ttl = sc->vxl_ttl;
1932 VXLAN_RUNLOCK(sc, &tracker);
1933
1934 #ifdef INET6
1935 if (VXLAN_SOCKADDR_IS_IPV6(&cfg->vxlc_local_sa))
1936 sa6_recoverscope(&cfg->vxlc_local_sa.in6);
1937 if (VXLAN_SOCKADDR_IS_IPV6(&cfg->vxlc_remote_sa))
1938 sa6_recoverscope(&cfg->vxlc_remote_sa.in6);
1939 #endif
1940
1941 return (0);
1942 }
1943
1944 static int
vxlan_ctrl_set_vni(struct vxlan_softc * sc,void * arg)1945 vxlan_ctrl_set_vni(struct vxlan_softc *sc, void *arg)
1946 {
1947 struct ifvxlancmd *cmd;
1948 int error;
1949
1950 cmd = arg;
1951
1952 if (vxlan_check_vni(cmd->vxlcmd_vni) != 0)
1953 return (EINVAL);
1954
1955 VXLAN_WLOCK(sc);
1956 if (vxlan_can_change_config(sc)) {
1957 sc->vxl_vni = cmd->vxlcmd_vni;
1958 error = 0;
1959 } else
1960 error = EBUSY;
1961 VXLAN_WUNLOCK(sc);
1962
1963 return (error);
1964 }
1965
1966 static int
vxlan_ctrl_set_local_addr(struct vxlan_softc * sc,void * arg)1967 vxlan_ctrl_set_local_addr(struct vxlan_softc *sc, void *arg)
1968 {
1969 struct ifvxlancmd *cmd;
1970 union vxlan_sockaddr *vxlsa;
1971 int error;
1972
1973 cmd = arg;
1974 vxlsa = &cmd->vxlcmd_sa;
1975
1976 if (!VXLAN_SOCKADDR_IS_IPV46(vxlsa))
1977 return (EINVAL);
1978 if (vxlan_sockaddr_in_multicast(vxlsa) != 0)
1979 return (EINVAL);
1980 if (VXLAN_SOCKADDR_IS_IPV6(vxlsa)) {
1981 error = vxlan_sockaddr_in6_embedscope(vxlsa);
1982 if (error)
1983 return (error);
1984 }
1985
1986 VXLAN_WLOCK(sc);
1987 if (vxlan_can_change_config(sc)) {
1988 vxlan_sockaddr_in_copy(&sc->vxl_src_addr, &vxlsa->sa);
1989 vxlan_set_hwcaps(sc);
1990 error = 0;
1991 } else
1992 error = EBUSY;
1993 VXLAN_WUNLOCK(sc);
1994
1995 return (error);
1996 }
1997
1998 static int
vxlan_ctrl_set_remote_addr(struct vxlan_softc * sc,void * arg)1999 vxlan_ctrl_set_remote_addr(struct vxlan_softc *sc, void *arg)
2000 {
2001 struct ifvxlancmd *cmd;
2002 union vxlan_sockaddr *vxlsa;
2003 int error;
2004
2005 cmd = arg;
2006 vxlsa = &cmd->vxlcmd_sa;
2007
2008 if (!VXLAN_SOCKADDR_IS_IPV46(vxlsa))
2009 return (EINVAL);
2010 if (VXLAN_SOCKADDR_IS_IPV6(vxlsa)) {
2011 error = vxlan_sockaddr_in6_embedscope(vxlsa);
2012 if (error)
2013 return (error);
2014 }
2015
2016 VXLAN_WLOCK(sc);
2017 if (vxlan_can_change_config(sc)) {
2018 vxlan_sockaddr_in_copy(&sc->vxl_dst_addr, &vxlsa->sa);
2019 vxlan_setup_interface_hdrlen(sc);
2020 error = 0;
2021 } else
2022 error = EBUSY;
2023 VXLAN_WUNLOCK(sc);
2024
2025 return (error);
2026 }
2027
2028 static int
vxlan_ctrl_set_local_port(struct vxlan_softc * sc,void * arg)2029 vxlan_ctrl_set_local_port(struct vxlan_softc *sc, void *arg)
2030 {
2031 struct ifvxlancmd *cmd;
2032 int error;
2033
2034 cmd = arg;
2035
2036 if (cmd->vxlcmd_port == 0)
2037 return (EINVAL);
2038
2039 VXLAN_WLOCK(sc);
2040 if (vxlan_can_change_config(sc)) {
2041 sc->vxl_src_addr.in4.sin_port = htons(cmd->vxlcmd_port);
2042 error = 0;
2043 } else
2044 error = EBUSY;
2045 VXLAN_WUNLOCK(sc);
2046
2047 return (error);
2048 }
2049
2050 static int
vxlan_ctrl_set_remote_port(struct vxlan_softc * sc,void * arg)2051 vxlan_ctrl_set_remote_port(struct vxlan_softc *sc, void *arg)
2052 {
2053 struct ifvxlancmd *cmd;
2054 int error;
2055
2056 cmd = arg;
2057
2058 if (cmd->vxlcmd_port == 0)
2059 return (EINVAL);
2060
2061 VXLAN_WLOCK(sc);
2062 if (vxlan_can_change_config(sc)) {
2063 sc->vxl_dst_addr.in4.sin_port = htons(cmd->vxlcmd_port);
2064 error = 0;
2065 } else
2066 error = EBUSY;
2067 VXLAN_WUNLOCK(sc);
2068
2069 return (error);
2070 }
2071
2072 static int
vxlan_ctrl_set_port_range(struct vxlan_softc * sc,void * arg)2073 vxlan_ctrl_set_port_range(struct vxlan_softc *sc, void *arg)
2074 {
2075 struct ifvxlancmd *cmd;
2076 uint16_t min, max;
2077 int error;
2078
2079 cmd = arg;
2080 min = cmd->vxlcmd_port_min;
2081 max = cmd->vxlcmd_port_max;
2082
2083 if (max < min)
2084 return (EINVAL);
2085
2086 VXLAN_WLOCK(sc);
2087 if (vxlan_can_change_config(sc)) {
2088 sc->vxl_min_port = min;
2089 sc->vxl_max_port = max;
2090 error = 0;
2091 } else
2092 error = EBUSY;
2093 VXLAN_WUNLOCK(sc);
2094
2095 return (error);
2096 }
2097
2098 static int
vxlan_ctrl_set_ftable_timeout(struct vxlan_softc * sc,void * arg)2099 vxlan_ctrl_set_ftable_timeout(struct vxlan_softc *sc, void *arg)
2100 {
2101 struct ifvxlancmd *cmd;
2102 int error;
2103
2104 cmd = arg;
2105
2106 VXLAN_WLOCK(sc);
2107 if (vxlan_check_ftable_timeout(cmd->vxlcmd_ftable_timeout) == 0) {
2108 sc->vxl_ftable_timeout = cmd->vxlcmd_ftable_timeout;
2109 error = 0;
2110 } else
2111 error = EINVAL;
2112 VXLAN_WUNLOCK(sc);
2113
2114 return (error);
2115 }
2116
2117 static int
vxlan_ctrl_set_ftable_max(struct vxlan_softc * sc,void * arg)2118 vxlan_ctrl_set_ftable_max(struct vxlan_softc *sc, void *arg)
2119 {
2120 struct ifvxlancmd *cmd;
2121 int error;
2122
2123 cmd = arg;
2124
2125 VXLAN_WLOCK(sc);
2126 if (vxlan_check_ftable_max(cmd->vxlcmd_ftable_max) == 0) {
2127 sc->vxl_ftable_max = cmd->vxlcmd_ftable_max;
2128 error = 0;
2129 } else
2130 error = EINVAL;
2131 VXLAN_WUNLOCK(sc);
2132
2133 return (error);
2134 }
2135
2136 static int
vxlan_ctrl_set_multicast_if(struct vxlan_softc * sc,void * arg)2137 vxlan_ctrl_set_multicast_if(struct vxlan_softc * sc, void *arg)
2138 {
2139 struct ifvxlancmd *cmd;
2140 int error;
2141
2142 cmd = arg;
2143
2144 VXLAN_WLOCK(sc);
2145 if (vxlan_can_change_config(sc)) {
2146 strlcpy(sc->vxl_mc_ifname, cmd->vxlcmd_ifname, IFNAMSIZ);
2147 vxlan_set_hwcaps(sc);
2148 error = 0;
2149 } else
2150 error = EBUSY;
2151 VXLAN_WUNLOCK(sc);
2152
2153 return (error);
2154 }
2155
2156 static int
vxlan_ctrl_set_ttl(struct vxlan_softc * sc,void * arg)2157 vxlan_ctrl_set_ttl(struct vxlan_softc *sc, void *arg)
2158 {
2159 struct ifvxlancmd *cmd;
2160 int error;
2161
2162 cmd = arg;
2163
2164 VXLAN_WLOCK(sc);
2165 if (vxlan_check_ttl(cmd->vxlcmd_ttl) == 0) {
2166 sc->vxl_ttl = cmd->vxlcmd_ttl;
2167 if (sc->vxl_im4o != NULL)
2168 sc->vxl_im4o->imo_multicast_ttl = sc->vxl_ttl;
2169 if (sc->vxl_im6o != NULL)
2170 sc->vxl_im6o->im6o_multicast_hlim = sc->vxl_ttl;
2171 error = 0;
2172 } else
2173 error = EINVAL;
2174 VXLAN_WUNLOCK(sc);
2175
2176 return (error);
2177 }
2178
2179 static int
vxlan_ctrl_set_learn(struct vxlan_softc * sc,void * arg)2180 vxlan_ctrl_set_learn(struct vxlan_softc *sc, void *arg)
2181 {
2182 struct ifvxlancmd *cmd;
2183
2184 cmd = arg;
2185
2186 VXLAN_WLOCK(sc);
2187 if (cmd->vxlcmd_flags & VXLAN_CMD_FLAG_LEARN)
2188 sc->vxl_flags |= VXLAN_FLAG_LEARN;
2189 else
2190 sc->vxl_flags &= ~VXLAN_FLAG_LEARN;
2191 VXLAN_WUNLOCK(sc);
2192
2193 return (0);
2194 }
2195
2196 static int
vxlan_ctrl_ftable_entry_add(struct vxlan_softc * sc,void * arg)2197 vxlan_ctrl_ftable_entry_add(struct vxlan_softc *sc, void *arg)
2198 {
2199 union vxlan_sockaddr vxlsa;
2200 struct ifvxlancmd *cmd;
2201 struct vxlan_ftable_entry *fe;
2202 int error;
2203
2204 cmd = arg;
2205 vxlsa = cmd->vxlcmd_sa;
2206
2207 if (!VXLAN_SOCKADDR_IS_IPV46(&vxlsa))
2208 return (EINVAL);
2209 if (vxlan_sockaddr_in_any(&vxlsa) != 0)
2210 return (EINVAL);
2211 if (vxlan_sockaddr_in_multicast(&vxlsa) != 0)
2212 return (EINVAL);
2213 /* BMV: We could support both IPv4 and IPv6 later. */
2214 if (vxlsa.sa.sa_family != sc->vxl_dst_addr.sa.sa_family)
2215 return (EAFNOSUPPORT);
2216
2217 if (VXLAN_SOCKADDR_IS_IPV6(&vxlsa)) {
2218 error = vxlan_sockaddr_in6_embedscope(&vxlsa);
2219 if (error)
2220 return (error);
2221 }
2222
2223 fe = vxlan_ftable_entry_alloc();
2224 if (fe == NULL)
2225 return (ENOMEM);
2226
2227 if (vxlsa.in4.sin_port == 0)
2228 vxlsa.in4.sin_port = sc->vxl_dst_addr.in4.sin_port;
2229
2230 vxlan_ftable_entry_init(sc, fe, cmd->vxlcmd_mac, &vxlsa.sa,
2231 VXLAN_FE_FLAG_STATIC);
2232
2233 VXLAN_WLOCK(sc);
2234 error = vxlan_ftable_entry_insert(sc, fe);
2235 VXLAN_WUNLOCK(sc);
2236
2237 if (error)
2238 vxlan_ftable_entry_free(fe);
2239
2240 return (error);
2241 }
2242
2243 static int
vxlan_ctrl_ftable_entry_rem(struct vxlan_softc * sc,void * arg)2244 vxlan_ctrl_ftable_entry_rem(struct vxlan_softc *sc, void *arg)
2245 {
2246 struct ifvxlancmd *cmd;
2247 struct vxlan_ftable_entry *fe;
2248 int error;
2249
2250 cmd = arg;
2251
2252 VXLAN_WLOCK(sc);
2253 fe = vxlan_ftable_entry_lookup(sc, cmd->vxlcmd_mac);
2254 if (fe != NULL) {
2255 vxlan_ftable_entry_destroy(sc, fe);
2256 error = 0;
2257 } else
2258 error = ENOENT;
2259 VXLAN_WUNLOCK(sc);
2260
2261 return (error);
2262 }
2263
2264 static int
vxlan_ctrl_flush(struct vxlan_softc * sc,void * arg)2265 vxlan_ctrl_flush(struct vxlan_softc *sc, void *arg)
2266 {
2267 struct ifvxlancmd *cmd;
2268 int all;
2269
2270 cmd = arg;
2271 all = cmd->vxlcmd_flags & VXLAN_CMD_FLAG_FLUSH_ALL;
2272
2273 VXLAN_WLOCK(sc);
2274 vxlan_ftable_flush(sc, all);
2275 VXLAN_WUNLOCK(sc);
2276
2277 return (0);
2278 }
2279
2280 static int
vxlan_ioctl_drvspec(struct vxlan_softc * sc,struct ifdrv * ifd,int get)2281 vxlan_ioctl_drvspec(struct vxlan_softc *sc, struct ifdrv *ifd, int get)
2282 {
2283 const struct vxlan_control *vc;
2284 union {
2285 struct ifvxlancfg cfg;
2286 struct ifvxlancmd cmd;
2287 } args;
2288 int out, error;
2289
2290 if (ifd->ifd_cmd >= vxlan_control_table_size)
2291 return (EINVAL);
2292
2293 bzero(&args, sizeof(args));
2294 vc = &vxlan_control_table[ifd->ifd_cmd];
2295 out = (vc->vxlc_flags & VXLAN_CTRL_FLAG_COPYOUT) != 0;
2296
2297 if ((get != 0 && out == 0) || (get == 0 && out != 0))
2298 return (EINVAL);
2299
2300 if (vc->vxlc_flags & VXLAN_CTRL_FLAG_SUSER) {
2301 error = priv_check(curthread, PRIV_NET_VXLAN);
2302 if (error)
2303 return (error);
2304 }
2305
2306 if (ifd->ifd_len != vc->vxlc_argsize ||
2307 ifd->ifd_len > sizeof(args))
2308 return (EINVAL);
2309
2310 if (vc->vxlc_flags & VXLAN_CTRL_FLAG_COPYIN) {
2311 error = copyin(ifd->ifd_data, &args, ifd->ifd_len);
2312 if (error)
2313 return (error);
2314 }
2315
2316 error = vc->vxlc_func(sc, &args);
2317 if (error)
2318 return (error);
2319
2320 if (vc->vxlc_flags & VXLAN_CTRL_FLAG_COPYOUT) {
2321 error = copyout(&args, ifd->ifd_data, ifd->ifd_len);
2322 if (error)
2323 return (error);
2324 }
2325
2326 return (0);
2327 }
2328
2329 static int
vxlan_ioctl(struct ifnet * ifp,u_long cmd,caddr_t data)2330 vxlan_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
2331 {
2332 struct vxlan_softc *sc;
2333 struct ifreq *ifr;
2334 struct ifdrv *ifd;
2335 int error;
2336
2337 sc = ifp->if_softc;
2338 ifr = (struct ifreq *) data;
2339 ifd = (struct ifdrv *) data;
2340
2341 error = 0;
2342
2343 switch (cmd) {
2344 case SIOCADDMULTI:
2345 case SIOCDELMULTI:
2346 break;
2347
2348 case SIOCGDRVSPEC:
2349 case SIOCSDRVSPEC:
2350 error = vxlan_ioctl_drvspec(sc, ifd, cmd == SIOCGDRVSPEC);
2351 break;
2352
2353 case SIOCSIFFLAGS:
2354 error = vxlan_ioctl_ifflags(sc);
2355 break;
2356
2357 case SIOCSIFMEDIA:
2358 case SIOCGIFMEDIA:
2359 error = ifmedia_ioctl(ifp, ifr, &sc->vxl_media, cmd);
2360 break;
2361
2362 case SIOCSIFMTU:
2363 if (ifr->ifr_mtu < ETHERMIN || ifr->ifr_mtu > VXLAN_MAX_MTU) {
2364 error = EINVAL;
2365 } else {
2366 VXLAN_WLOCK(sc);
2367 ifp->if_mtu = ifr->ifr_mtu;
2368 sc->vxl_flags |= VXLAN_FLAG_USER_MTU;
2369 VXLAN_WUNLOCK(sc);
2370 }
2371 break;
2372
2373 case SIOCSIFCAP:
2374 VXLAN_WLOCK(sc);
2375 error = vxlan_set_reqcap(sc, ifp, ifr->ifr_reqcap);
2376 if (error == 0)
2377 vxlan_set_hwcaps(sc);
2378 VXLAN_WUNLOCK(sc);
2379 break;
2380
2381 default:
2382 error = ether_ioctl(ifp, cmd, data);
2383 break;
2384 }
2385
2386 return (error);
2387 }
2388
2389 #if defined(INET) || defined(INET6)
2390 static uint16_t
vxlan_pick_source_port(struct vxlan_softc * sc,struct mbuf * m)2391 vxlan_pick_source_port(struct vxlan_softc *sc, struct mbuf *m)
2392 {
2393 int range;
2394 uint32_t hash;
2395
2396 range = sc->vxl_max_port - sc->vxl_min_port + 1;
2397
2398 if (M_HASHTYPE_ISHASH(m))
2399 hash = m->m_pkthdr.flowid;
2400 else
2401 hash = jenkins_hash(m->m_data, ETHER_HDR_LEN,
2402 sc->vxl_port_hash_key);
2403
2404 return (sc->vxl_min_port + (hash % range));
2405 }
2406
2407 static void
vxlan_encap_header(struct vxlan_softc * sc,struct mbuf * m,int ipoff,uint16_t srcport,uint16_t dstport)2408 vxlan_encap_header(struct vxlan_softc *sc, struct mbuf *m, int ipoff,
2409 uint16_t srcport, uint16_t dstport)
2410 {
2411 struct vxlanudphdr *hdr;
2412 struct udphdr *udph;
2413 struct vxlan_header *vxh;
2414 int len;
2415
2416 len = m->m_pkthdr.len - ipoff;
2417 MPASS(len >= sizeof(struct vxlanudphdr));
2418 hdr = mtodo(m, ipoff);
2419
2420 udph = &hdr->vxlh_udp;
2421 udph->uh_sport = srcport;
2422 udph->uh_dport = dstport;
2423 udph->uh_ulen = htons(len);
2424 udph->uh_sum = 0;
2425
2426 vxh = &hdr->vxlh_hdr;
2427 vxh->vxlh_flags = htonl(VXLAN_HDR_FLAGS_VALID_VNI);
2428 vxh->vxlh_vni = htonl(sc->vxl_vni << VXLAN_HDR_VNI_SHIFT);
2429 }
2430 #endif
2431
2432 /*
2433 * Return the CSUM_INNER_* equivalent of CSUM_* caps.
2434 */
2435 static uint32_t
csum_flags_to_inner_flags(uint32_t csum_flags_in,const uint32_t encap)2436 csum_flags_to_inner_flags(uint32_t csum_flags_in, const uint32_t encap)
2437 {
2438 uint32_t csum_flags = encap;
2439 const uint32_t v4 = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP;
2440
2441 /*
2442 * csum_flags can request either v4 or v6 offload but not both.
2443 * tcp_output always sets CSUM_TSO (both CSUM_IP_TSO and CSUM_IP6_TSO)
2444 * so those bits are no good to detect the IP version. Other bits are
2445 * always set with CSUM_TSO and we use those to figure out the IP
2446 * version.
2447 */
2448 if (csum_flags_in & v4) {
2449 if (csum_flags_in & CSUM_IP)
2450 csum_flags |= CSUM_INNER_IP;
2451 if (csum_flags_in & CSUM_IP_UDP)
2452 csum_flags |= CSUM_INNER_IP_UDP;
2453 if (csum_flags_in & CSUM_IP_TCP)
2454 csum_flags |= CSUM_INNER_IP_TCP;
2455 if (csum_flags_in & CSUM_IP_TSO)
2456 csum_flags |= CSUM_INNER_IP_TSO;
2457 } else {
2458 #ifdef INVARIANTS
2459 const uint32_t v6 = CSUM_IP6_UDP | CSUM_IP6_TCP;
2460
2461 MPASS((csum_flags_in & v6) != 0);
2462 #endif
2463 if (csum_flags_in & CSUM_IP6_UDP)
2464 csum_flags |= CSUM_INNER_IP6_UDP;
2465 if (csum_flags_in & CSUM_IP6_TCP)
2466 csum_flags |= CSUM_INNER_IP6_TCP;
2467 if (csum_flags_in & CSUM_IP6_TSO)
2468 csum_flags |= CSUM_INNER_IP6_TSO;
2469 }
2470
2471 return (csum_flags);
2472 }
2473
2474 static int
vxlan_encap4(struct vxlan_softc * sc,const union vxlan_sockaddr * fvxlsa,struct mbuf * m)2475 vxlan_encap4(struct vxlan_softc *sc, const union vxlan_sockaddr *fvxlsa,
2476 struct mbuf *m)
2477 {
2478 #ifdef INET
2479 struct ifnet *ifp;
2480 struct ip *ip;
2481 struct in_addr srcaddr, dstaddr;
2482 uint16_t srcport, dstport;
2483 int len, mcast, error;
2484 struct route route, *ro;
2485 struct sockaddr_in *sin;
2486 uint32_t csum_flags;
2487
2488 NET_EPOCH_ASSERT();
2489
2490 ifp = sc->vxl_ifp;
2491 srcaddr = sc->vxl_src_addr.in4.sin_addr;
2492 srcport = vxlan_pick_source_port(sc, m);
2493 dstaddr = fvxlsa->in4.sin_addr;
2494 dstport = fvxlsa->in4.sin_port;
2495
2496 M_PREPEND(m, sizeof(struct ip) + sizeof(struct vxlanudphdr),
2497 M_NOWAIT);
2498 if (m == NULL) {
2499 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2500 return (ENOBUFS);
2501 }
2502
2503 len = m->m_pkthdr.len;
2504
2505 ip = mtod(m, struct ip *);
2506 ip->ip_tos = 0;
2507 ip->ip_len = htons(len);
2508 ip->ip_off = 0;
2509 ip->ip_ttl = sc->vxl_ttl;
2510 ip->ip_p = IPPROTO_UDP;
2511 ip->ip_sum = 0;
2512 ip->ip_src = srcaddr;
2513 ip->ip_dst = dstaddr;
2514
2515 vxlan_encap_header(sc, m, sizeof(struct ip), srcport, dstport);
2516
2517 mcast = (m->m_flags & (M_MCAST | M_BCAST)) ? 1 : 0;
2518 m->m_flags &= ~(M_MCAST | M_BCAST);
2519
2520 m->m_pkthdr.csum_flags &= CSUM_FLAGS_TX;
2521 if (m->m_pkthdr.csum_flags != 0) {
2522 /*
2523 * HW checksum (L3 and/or L4) or TSO has been requested. Look
2524 * up the ifnet for the outbound route and verify that the
2525 * outbound ifnet can perform the requested operation on the
2526 * inner frame.
2527 */
2528 bzero(&route, sizeof(route));
2529 ro = &route;
2530 sin = (struct sockaddr_in *)&ro->ro_dst;
2531 sin->sin_family = AF_INET;
2532 sin->sin_len = sizeof(*sin);
2533 sin->sin_addr = ip->ip_dst;
2534 ro->ro_nh = fib4_lookup(RT_DEFAULT_FIB, ip->ip_dst, 0, NHR_NONE,
2535 0);
2536 if (ro->ro_nh == NULL) {
2537 m_freem(m);
2538 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2539 return (EHOSTUNREACH);
2540 }
2541
2542 csum_flags = csum_flags_to_inner_flags(m->m_pkthdr.csum_flags,
2543 CSUM_ENCAP_VXLAN);
2544 if ((csum_flags & ro->ro_nh->nh_ifp->if_hwassist) !=
2545 csum_flags) {
2546 if (ppsratecheck(&sc->err_time, &sc->err_pps, 1)) {
2547 const struct ifnet *nh_ifp = ro->ro_nh->nh_ifp;
2548
2549 if_printf(ifp, "interface %s is missing hwcaps "
2550 "0x%08x, csum_flags 0x%08x -> 0x%08x, "
2551 "hwassist 0x%08x\n", nh_ifp->if_xname,
2552 csum_flags & ~(uint32_t)nh_ifp->if_hwassist,
2553 m->m_pkthdr.csum_flags, csum_flags,
2554 (uint32_t)nh_ifp->if_hwassist);
2555 }
2556 m_freem(m);
2557 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2558 return (ENXIO);
2559 }
2560 m->m_pkthdr.csum_flags = csum_flags;
2561 if (csum_flags &
2562 (CSUM_INNER_IP | CSUM_INNER_IP_UDP | CSUM_INNER_IP6_UDP |
2563 CSUM_INNER_IP_TCP | CSUM_INNER_IP6_TCP)) {
2564 counter_u64_add(sc->vxl_stats.txcsum, 1);
2565 if (csum_flags & CSUM_INNER_TSO)
2566 counter_u64_add(sc->vxl_stats.tso, 1);
2567 }
2568 } else
2569 ro = NULL;
2570 error = ip_output(m, NULL, ro, 0, sc->vxl_im4o, NULL);
2571 if (error == 0) {
2572 if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
2573 if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
2574 if (mcast != 0)
2575 if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
2576 } else
2577 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2578
2579 return (error);
2580 #else
2581 m_freem(m);
2582 return (ENOTSUP);
2583 #endif
2584 }
2585
2586 static int
vxlan_encap6(struct vxlan_softc * sc,const union vxlan_sockaddr * fvxlsa,struct mbuf * m)2587 vxlan_encap6(struct vxlan_softc *sc, const union vxlan_sockaddr *fvxlsa,
2588 struct mbuf *m)
2589 {
2590 #ifdef INET6
2591 struct ifnet *ifp;
2592 struct ip6_hdr *ip6;
2593 const struct in6_addr *srcaddr, *dstaddr;
2594 uint16_t srcport, dstport;
2595 int len, mcast, error;
2596 struct route_in6 route, *ro;
2597 struct sockaddr_in6 *sin6;
2598 uint32_t csum_flags;
2599
2600 NET_EPOCH_ASSERT();
2601
2602 ifp = sc->vxl_ifp;
2603 srcaddr = &sc->vxl_src_addr.in6.sin6_addr;
2604 srcport = vxlan_pick_source_port(sc, m);
2605 dstaddr = &fvxlsa->in6.sin6_addr;
2606 dstport = fvxlsa->in6.sin6_port;
2607
2608 M_PREPEND(m, sizeof(struct ip6_hdr) + sizeof(struct vxlanudphdr),
2609 M_NOWAIT);
2610 if (m == NULL) {
2611 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2612 return (ENOBUFS);
2613 }
2614
2615 len = m->m_pkthdr.len;
2616
2617 ip6 = mtod(m, struct ip6_hdr *);
2618 ip6->ip6_flow = 0; /* BMV: Keep in forwarding entry? */
2619 ip6->ip6_vfc = IPV6_VERSION;
2620 ip6->ip6_plen = 0;
2621 ip6->ip6_nxt = IPPROTO_UDP;
2622 ip6->ip6_hlim = sc->vxl_ttl;
2623 ip6->ip6_src = *srcaddr;
2624 ip6->ip6_dst = *dstaddr;
2625
2626 vxlan_encap_header(sc, m, sizeof(struct ip6_hdr), srcport, dstport);
2627
2628 mcast = (m->m_flags & (M_MCAST | M_BCAST)) ? 1 : 0;
2629 m->m_flags &= ~(M_MCAST | M_BCAST);
2630
2631 ro = NULL;
2632 m->m_pkthdr.csum_flags &= CSUM_FLAGS_TX;
2633 if (m->m_pkthdr.csum_flags != 0) {
2634 /*
2635 * HW checksum (L3 and/or L4) or TSO has been requested. Look
2636 * up the ifnet for the outbound route and verify that the
2637 * outbound ifnet can perform the requested operation on the
2638 * inner frame.
2639 */
2640 bzero(&route, sizeof(route));
2641 ro = &route;
2642 sin6 = (struct sockaddr_in6 *)&ro->ro_dst;
2643 sin6->sin6_family = AF_INET6;
2644 sin6->sin6_len = sizeof(*sin6);
2645 sin6->sin6_addr = ip6->ip6_dst;
2646 ro->ro_nh = fib6_lookup(RT_DEFAULT_FIB, &ip6->ip6_dst, 0,
2647 NHR_NONE, 0);
2648 if (ro->ro_nh == NULL) {
2649 m_freem(m);
2650 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2651 return (EHOSTUNREACH);
2652 }
2653
2654 csum_flags = csum_flags_to_inner_flags(m->m_pkthdr.csum_flags,
2655 CSUM_ENCAP_VXLAN);
2656 if ((csum_flags & ro->ro_nh->nh_ifp->if_hwassist) !=
2657 csum_flags) {
2658 if (ppsratecheck(&sc->err_time, &sc->err_pps, 1)) {
2659 const struct ifnet *nh_ifp = ro->ro_nh->nh_ifp;
2660
2661 if_printf(ifp, "interface %s is missing hwcaps "
2662 "0x%08x, csum_flags 0x%08x -> 0x%08x, "
2663 "hwassist 0x%08x\n", nh_ifp->if_xname,
2664 csum_flags & ~(uint32_t)nh_ifp->if_hwassist,
2665 m->m_pkthdr.csum_flags, csum_flags,
2666 (uint32_t)nh_ifp->if_hwassist);
2667 }
2668 m_freem(m);
2669 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2670 return (ENXIO);
2671 }
2672 m->m_pkthdr.csum_flags = csum_flags;
2673 if (csum_flags &
2674 (CSUM_INNER_IP | CSUM_INNER_IP_UDP | CSUM_INNER_IP6_UDP |
2675 CSUM_INNER_IP_TCP | CSUM_INNER_IP6_TCP)) {
2676 counter_u64_add(sc->vxl_stats.txcsum, 1);
2677 if (csum_flags & CSUM_INNER_TSO)
2678 counter_u64_add(sc->vxl_stats.tso, 1);
2679 }
2680 } else if (ntohs(dstport) != V_zero_checksum_port) {
2681 struct udphdr *hdr = mtodo(m, sizeof(struct ip6_hdr));
2682
2683 hdr->uh_sum = in6_cksum_pseudo(ip6,
2684 m->m_pkthdr.len - sizeof(struct ip6_hdr), IPPROTO_UDP, 0);
2685 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
2686 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
2687 }
2688 error = ip6_output(m, NULL, ro, 0, sc->vxl_im6o, NULL, NULL);
2689 if (error == 0) {
2690 if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
2691 if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
2692 if (mcast != 0)
2693 if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
2694 } else
2695 if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
2696
2697 return (error);
2698 #else
2699 m_freem(m);
2700 return (ENOTSUP);
2701 #endif
2702 }
2703
2704 static int
vxlan_transmit(struct ifnet * ifp,struct mbuf * m)2705 vxlan_transmit(struct ifnet *ifp, struct mbuf *m)
2706 {
2707 struct rm_priotracker tracker;
2708 union vxlan_sockaddr vxlsa;
2709 struct vxlan_softc *sc;
2710 struct vxlan_ftable_entry *fe;
2711 struct ifnet *mcifp;
2712 struct ether_header *eh;
2713 int ipv4, error;
2714
2715 sc = ifp->if_softc;
2716 eh = mtod(m, struct ether_header *);
2717 fe = NULL;
2718 mcifp = NULL;
2719
2720 ETHER_BPF_MTAP(ifp, m);
2721
2722 VXLAN_RLOCK(sc, &tracker);
2723 if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
2724 VXLAN_RUNLOCK(sc, &tracker);
2725 m_freem(m);
2726 return (ENETDOWN);
2727 }
2728
2729 if ((m->m_flags & (M_BCAST | M_MCAST)) == 0)
2730 fe = vxlan_ftable_entry_lookup(sc, eh->ether_dhost);
2731 if (fe == NULL)
2732 fe = &sc->vxl_default_fe;
2733 vxlan_sockaddr_copy(&vxlsa, &fe->vxlfe_raddr.sa);
2734
2735 ipv4 = VXLAN_SOCKADDR_IS_IPV4(&vxlsa) != 0;
2736 if (vxlan_sockaddr_in_multicast(&vxlsa) != 0)
2737 mcifp = vxlan_multicast_if_ref(sc, ipv4);
2738
2739 VXLAN_ACQUIRE(sc);
2740 VXLAN_RUNLOCK(sc, &tracker);
2741
2742 if (ipv4 != 0)
2743 error = vxlan_encap4(sc, &vxlsa, m);
2744 else
2745 error = vxlan_encap6(sc, &vxlsa, m);
2746
2747 vxlan_release(sc);
2748 if (mcifp != NULL)
2749 if_rele(mcifp);
2750
2751 return (error);
2752 }
2753
2754 static void
vxlan_qflush(struct ifnet * ifp __unused)2755 vxlan_qflush(struct ifnet *ifp __unused)
2756 {
2757 }
2758
2759 static void
vxlan_rcv_udp_packet(struct mbuf * m,int offset,struct inpcb * inpcb,const struct sockaddr * srcsa,void * xvso)2760 vxlan_rcv_udp_packet(struct mbuf *m, int offset, struct inpcb *inpcb,
2761 const struct sockaddr *srcsa, void *xvso)
2762 {
2763 struct vxlan_socket *vso;
2764 struct vxlan_header *vxh, vxlanhdr;
2765 uint32_t vni;
2766 int error __unused;
2767
2768 M_ASSERTPKTHDR(m);
2769 vso = xvso;
2770 offset += sizeof(struct udphdr);
2771
2772 if (m->m_pkthdr.len < offset + sizeof(struct vxlan_header))
2773 goto out;
2774
2775 if (__predict_false(m->m_len < offset + sizeof(struct vxlan_header))) {
2776 m_copydata(m, offset, sizeof(struct vxlan_header),
2777 (caddr_t) &vxlanhdr);
2778 vxh = &vxlanhdr;
2779 } else
2780 vxh = mtodo(m, offset);
2781
2782 /*
2783 * Drop if there is a reserved bit set in either the flags or VNI
2784 * fields of the header. This goes against the specification, but
2785 * a bit set may indicate an unsupported new feature. This matches
2786 * the behavior of the Linux implementation.
2787 */
2788 if (vxh->vxlh_flags != htonl(VXLAN_HDR_FLAGS_VALID_VNI) ||
2789 vxh->vxlh_vni & ~VXLAN_VNI_MASK)
2790 goto out;
2791
2792 vni = ntohl(vxh->vxlh_vni) >> VXLAN_HDR_VNI_SHIFT;
2793
2794 /* Adjust to the start of the inner Ethernet frame. */
2795 m_adj_decap(m, offset + sizeof(struct vxlan_header));
2796
2797 error = vxlan_input(vso, vni, &m, srcsa);
2798 MPASS(error != 0 || m == NULL);
2799
2800 out:
2801 if (m != NULL)
2802 m_freem(m);
2803 }
2804
2805 static int
vxlan_input(struct vxlan_socket * vso,uint32_t vni,struct mbuf ** m0,const struct sockaddr * sa)2806 vxlan_input(struct vxlan_socket *vso, uint32_t vni, struct mbuf **m0,
2807 const struct sockaddr *sa)
2808 {
2809 struct vxlan_softc *sc;
2810 struct ifnet *ifp;
2811 struct mbuf *m;
2812 struct ether_header *eh;
2813 int error;
2814
2815 m = *m0;
2816
2817 if (m->m_pkthdr.len < ETHER_HDR_LEN)
2818 return (EINVAL);
2819
2820 sc = vxlan_socket_lookup_softc(vso, vni);
2821 if (sc == NULL)
2822 return (ENOENT);
2823
2824 ifp = sc->vxl_ifp;
2825 eh = mtod(m, struct ether_header *);
2826
2827 if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
2828 error = ENETDOWN;
2829 goto out;
2830 } else if (ifp == m->m_pkthdr.rcvif) {
2831 /* XXX Does not catch more complex loops. */
2832 error = EDEADLK;
2833 goto out;
2834 }
2835
2836 if (sc->vxl_flags & VXLAN_FLAG_LEARN)
2837 vxlan_ftable_learn(sc, sa, eh->ether_shost);
2838
2839 m_clrprotoflags(m);
2840 m->m_pkthdr.rcvif = ifp;
2841 M_SETFIB(m, ifp->if_fib);
2842 if (((ifp->if_capenable & IFCAP_RXCSUM &&
2843 m->m_pkthdr.csum_flags & CSUM_INNER_L3_CALC) ||
2844 (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2845 !(m->m_pkthdr.csum_flags & CSUM_INNER_L3_CALC)))) {
2846 uint32_t csum_flags = 0;
2847
2848 if (m->m_pkthdr.csum_flags & CSUM_INNER_L3_CALC)
2849 csum_flags |= CSUM_L3_CALC;
2850 if (m->m_pkthdr.csum_flags & CSUM_INNER_L3_VALID)
2851 csum_flags |= CSUM_L3_VALID;
2852 if (m->m_pkthdr.csum_flags & CSUM_INNER_L4_CALC)
2853 csum_flags |= CSUM_L4_CALC;
2854 if (m->m_pkthdr.csum_flags & CSUM_INNER_L4_VALID)
2855 csum_flags |= CSUM_L4_VALID;
2856 m->m_pkthdr.csum_flags = csum_flags;
2857 counter_u64_add(sc->vxl_stats.rxcsum, 1);
2858 } else {
2859 /* clear everything */
2860 m->m_pkthdr.csum_flags = 0;
2861 m->m_pkthdr.csum_data = 0;
2862 }
2863
2864 (*ifp->if_input)(ifp, m);
2865 *m0 = NULL;
2866 error = 0;
2867
2868 out:
2869 vxlan_release(sc);
2870 return (error);
2871 }
2872
2873 static int
vxlan_stats_alloc(struct vxlan_softc * sc)2874 vxlan_stats_alloc(struct vxlan_softc *sc)
2875 {
2876 struct vxlan_statistics *stats = &sc->vxl_stats;
2877
2878 stats->txcsum = counter_u64_alloc(M_WAITOK);
2879 if (stats->txcsum == NULL)
2880 goto failed;
2881
2882 stats->tso = counter_u64_alloc(M_WAITOK);
2883 if (stats->tso == NULL)
2884 goto failed;
2885
2886 stats->rxcsum = counter_u64_alloc(M_WAITOK);
2887 if (stats->rxcsum == NULL)
2888 goto failed;
2889
2890 return (0);
2891 failed:
2892 vxlan_stats_free(sc);
2893 return (ENOMEM);
2894 }
2895
2896 static void
vxlan_stats_free(struct vxlan_softc * sc)2897 vxlan_stats_free(struct vxlan_softc *sc)
2898 {
2899 struct vxlan_statistics *stats = &sc->vxl_stats;
2900
2901 if (stats->txcsum != NULL) {
2902 counter_u64_free(stats->txcsum);
2903 stats->txcsum = NULL;
2904 }
2905 if (stats->tso != NULL) {
2906 counter_u64_free(stats->tso);
2907 stats->tso = NULL;
2908 }
2909 if (stats->rxcsum != NULL) {
2910 counter_u64_free(stats->rxcsum);
2911 stats->rxcsum = NULL;
2912 }
2913 }
2914
2915 static void
vxlan_set_default_config(struct vxlan_softc * sc)2916 vxlan_set_default_config(struct vxlan_softc *sc)
2917 {
2918
2919 sc->vxl_flags |= VXLAN_FLAG_LEARN;
2920
2921 sc->vxl_vni = VXLAN_VNI_MAX;
2922 sc->vxl_ttl = IPDEFTTL;
2923
2924 if (!vxlan_tunable_int(sc, "legacy_port", vxlan_legacy_port)) {
2925 sc->vxl_src_addr.in4.sin_port = htons(VXLAN_PORT);
2926 sc->vxl_dst_addr.in4.sin_port = htons(VXLAN_PORT);
2927 } else {
2928 sc->vxl_src_addr.in4.sin_port = htons(VXLAN_LEGACY_PORT);
2929 sc->vxl_dst_addr.in4.sin_port = htons(VXLAN_LEGACY_PORT);
2930 }
2931
2932 sc->vxl_min_port = V_ipport_firstauto;
2933 sc->vxl_max_port = V_ipport_lastauto;
2934
2935 sc->vxl_ftable_max = VXLAN_FTABLE_MAX;
2936 sc->vxl_ftable_timeout = VXLAN_FTABLE_TIMEOUT;
2937 }
2938
2939 static int
vxlan_set_user_config(struct vxlan_softc * sc,struct ifvxlanparam * vxlp)2940 vxlan_set_user_config(struct vxlan_softc *sc, struct ifvxlanparam *vxlp)
2941 {
2942
2943 #ifndef INET
2944 if (vxlp->vxlp_with & (VXLAN_PARAM_WITH_LOCAL_ADDR4 |
2945 VXLAN_PARAM_WITH_REMOTE_ADDR4))
2946 return (EAFNOSUPPORT);
2947 #endif
2948
2949 #ifndef INET6
2950 if (vxlp->vxlp_with & (VXLAN_PARAM_WITH_LOCAL_ADDR6 |
2951 VXLAN_PARAM_WITH_REMOTE_ADDR6))
2952 return (EAFNOSUPPORT);
2953 #else
2954 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_LOCAL_ADDR6) {
2955 int error = vxlan_sockaddr_in6_embedscope(&vxlp->vxlp_local_sa);
2956 if (error)
2957 return (error);
2958 }
2959 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_REMOTE_ADDR6) {
2960 int error = vxlan_sockaddr_in6_embedscope(
2961 &vxlp->vxlp_remote_sa);
2962 if (error)
2963 return (error);
2964 }
2965 #endif
2966
2967 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_VNI) {
2968 if (vxlan_check_vni(vxlp->vxlp_vni) == 0)
2969 sc->vxl_vni = vxlp->vxlp_vni;
2970 }
2971
2972 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_LOCAL_ADDR4) {
2973 sc->vxl_src_addr.in4.sin_len = sizeof(struct sockaddr_in);
2974 sc->vxl_src_addr.in4.sin_family = AF_INET;
2975 sc->vxl_src_addr.in4.sin_addr =
2976 vxlp->vxlp_local_sa.in4.sin_addr;
2977 } else if (vxlp->vxlp_with & VXLAN_PARAM_WITH_LOCAL_ADDR6) {
2978 sc->vxl_src_addr.in6.sin6_len = sizeof(struct sockaddr_in6);
2979 sc->vxl_src_addr.in6.sin6_family = AF_INET6;
2980 sc->vxl_src_addr.in6.sin6_addr =
2981 vxlp->vxlp_local_sa.in6.sin6_addr;
2982 }
2983
2984 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_REMOTE_ADDR4) {
2985 sc->vxl_dst_addr.in4.sin_len = sizeof(struct sockaddr_in);
2986 sc->vxl_dst_addr.in4.sin_family = AF_INET;
2987 sc->vxl_dst_addr.in4.sin_addr =
2988 vxlp->vxlp_remote_sa.in4.sin_addr;
2989 } else if (vxlp->vxlp_with & VXLAN_PARAM_WITH_REMOTE_ADDR6) {
2990 sc->vxl_dst_addr.in6.sin6_len = sizeof(struct sockaddr_in6);
2991 sc->vxl_dst_addr.in6.sin6_family = AF_INET6;
2992 sc->vxl_dst_addr.in6.sin6_addr =
2993 vxlp->vxlp_remote_sa.in6.sin6_addr;
2994 }
2995
2996 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_LOCAL_PORT)
2997 sc->vxl_src_addr.in4.sin_port = htons(vxlp->vxlp_local_port);
2998 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_REMOTE_PORT)
2999 sc->vxl_dst_addr.in4.sin_port = htons(vxlp->vxlp_remote_port);
3000
3001 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_PORT_RANGE) {
3002 if (vxlp->vxlp_min_port <= vxlp->vxlp_max_port) {
3003 sc->vxl_min_port = vxlp->vxlp_min_port;
3004 sc->vxl_max_port = vxlp->vxlp_max_port;
3005 }
3006 }
3007
3008 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_MULTICAST_IF)
3009 strlcpy(sc->vxl_mc_ifname, vxlp->vxlp_mc_ifname, IFNAMSIZ);
3010
3011 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_FTABLE_TIMEOUT) {
3012 if (vxlan_check_ftable_timeout(vxlp->vxlp_ftable_timeout) == 0)
3013 sc->vxl_ftable_timeout = vxlp->vxlp_ftable_timeout;
3014 }
3015
3016 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_FTABLE_MAX) {
3017 if (vxlan_check_ftable_max(vxlp->vxlp_ftable_max) == 0)
3018 sc->vxl_ftable_max = vxlp->vxlp_ftable_max;
3019 }
3020
3021 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_TTL) {
3022 if (vxlan_check_ttl(vxlp->vxlp_ttl) == 0)
3023 sc->vxl_ttl = vxlp->vxlp_ttl;
3024 }
3025
3026 if (vxlp->vxlp_with & VXLAN_PARAM_WITH_LEARN) {
3027 if (vxlp->vxlp_learn == 0)
3028 sc->vxl_flags &= ~VXLAN_FLAG_LEARN;
3029 }
3030
3031 return (0);
3032 }
3033
3034 static int
vxlan_set_reqcap(struct vxlan_softc * sc,struct ifnet * ifp,int reqcap)3035 vxlan_set_reqcap(struct vxlan_softc *sc, struct ifnet *ifp, int reqcap)
3036 {
3037 int mask = reqcap ^ ifp->if_capenable;
3038
3039 /* Disable TSO if tx checksums are disabled. */
3040 if (mask & IFCAP_TXCSUM && !(reqcap & IFCAP_TXCSUM) &&
3041 reqcap & IFCAP_TSO4) {
3042 reqcap &= ~IFCAP_TSO4;
3043 if_printf(ifp, "tso4 disabled due to -txcsum.\n");
3044 }
3045 if (mask & IFCAP_TXCSUM_IPV6 && !(reqcap & IFCAP_TXCSUM_IPV6) &&
3046 reqcap & IFCAP_TSO6) {
3047 reqcap &= ~IFCAP_TSO6;
3048 if_printf(ifp, "tso6 disabled due to -txcsum6.\n");
3049 }
3050
3051 /* Do not enable TSO if tx checksums are disabled. */
3052 if (mask & IFCAP_TSO4 && reqcap & IFCAP_TSO4 &&
3053 !(reqcap & IFCAP_TXCSUM)) {
3054 if_printf(ifp, "enable txcsum first.\n");
3055 return (EAGAIN);
3056 }
3057 if (mask & IFCAP_TSO6 && reqcap & IFCAP_TSO6 &&
3058 !(reqcap & IFCAP_TXCSUM_IPV6)) {
3059 if_printf(ifp, "enable txcsum6 first.\n");
3060 return (EAGAIN);
3061 }
3062
3063 sc->vxl_reqcap = reqcap;
3064 return (0);
3065 }
3066
3067 /*
3068 * A VXLAN interface inherits the capabilities of the vxlandev or the interface
3069 * hosting the vxlanlocal address.
3070 */
3071 static void
vxlan_set_hwcaps(struct vxlan_softc * sc)3072 vxlan_set_hwcaps(struct vxlan_softc *sc)
3073 {
3074 struct epoch_tracker et;
3075 struct ifnet *p;
3076 struct ifaddr *ifa;
3077 u_long hwa;
3078 int cap, ena;
3079 bool rel;
3080 struct ifnet *ifp = sc->vxl_ifp;
3081
3082 /* reset caps */
3083 ifp->if_capabilities &= VXLAN_BASIC_IFCAPS;
3084 ifp->if_capenable &= VXLAN_BASIC_IFCAPS;
3085 ifp->if_hwassist = 0;
3086
3087 NET_EPOCH_ENTER(et);
3088 CURVNET_SET(ifp->if_vnet);
3089
3090 rel = false;
3091 p = NULL;
3092 if (sc->vxl_mc_ifname[0] != '\0') {
3093 rel = true;
3094 p = ifunit_ref(sc->vxl_mc_ifname);
3095 } else if (vxlan_sockaddr_in_any(&sc->vxl_src_addr) == 0) {
3096 if (sc->vxl_src_addr.sa.sa_family == AF_INET) {
3097 struct sockaddr_in in4 = sc->vxl_src_addr.in4;
3098
3099 in4.sin_port = 0;
3100 ifa = ifa_ifwithaddr((struct sockaddr *)&in4);
3101 if (ifa != NULL)
3102 p = ifa->ifa_ifp;
3103 } else if (sc->vxl_src_addr.sa.sa_family == AF_INET6) {
3104 struct sockaddr_in6 in6 = sc->vxl_src_addr.in6;
3105
3106 in6.sin6_port = 0;
3107 ifa = ifa_ifwithaddr((struct sockaddr *)&in6);
3108 if (ifa != NULL)
3109 p = ifa->ifa_ifp;
3110 }
3111 }
3112 if (p == NULL)
3113 goto done;
3114
3115 cap = ena = hwa = 0;
3116
3117 /* checksum offload */
3118 if (p->if_capabilities & IFCAP_VXLAN_HWCSUM)
3119 cap |= p->if_capabilities & (IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6);
3120 if (p->if_capenable & IFCAP_VXLAN_HWCSUM) {
3121 ena |= sc->vxl_reqcap & p->if_capenable &
3122 (IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6);
3123 if (ena & IFCAP_TXCSUM) {
3124 if (p->if_hwassist & CSUM_INNER_IP)
3125 hwa |= CSUM_IP;
3126 if (p->if_hwassist & CSUM_INNER_IP_UDP)
3127 hwa |= CSUM_IP_UDP;
3128 if (p->if_hwassist & CSUM_INNER_IP_TCP)
3129 hwa |= CSUM_IP_TCP;
3130 }
3131 if (ena & IFCAP_TXCSUM_IPV6) {
3132 if (p->if_hwassist & CSUM_INNER_IP6_UDP)
3133 hwa |= CSUM_IP6_UDP;
3134 if (p->if_hwassist & CSUM_INNER_IP6_TCP)
3135 hwa |= CSUM_IP6_TCP;
3136 }
3137 }
3138
3139 /* hardware TSO */
3140 if (p->if_capabilities & IFCAP_VXLAN_HWTSO) {
3141 cap |= p->if_capabilities & IFCAP_TSO;
3142 if (p->if_hw_tsomax > IP_MAXPACKET - ifp->if_hdrlen)
3143 ifp->if_hw_tsomax = IP_MAXPACKET - ifp->if_hdrlen;
3144 else
3145 ifp->if_hw_tsomax = p->if_hw_tsomax;
3146 /* XXX: tsomaxsegcount decrement is cxgbe specific */
3147 ifp->if_hw_tsomaxsegcount = p->if_hw_tsomaxsegcount - 1;
3148 ifp->if_hw_tsomaxsegsize = p->if_hw_tsomaxsegsize;
3149 }
3150 if (p->if_capenable & IFCAP_VXLAN_HWTSO) {
3151 ena |= sc->vxl_reqcap & p->if_capenable & IFCAP_TSO;
3152 if (ena & IFCAP_TSO) {
3153 if (p->if_hwassist & CSUM_INNER_IP_TSO)
3154 hwa |= CSUM_IP_TSO;
3155 if (p->if_hwassist & CSUM_INNER_IP6_TSO)
3156 hwa |= CSUM_IP6_TSO;
3157 }
3158 }
3159
3160 ifp->if_capabilities |= cap;
3161 ifp->if_capenable |= ena;
3162 ifp->if_hwassist |= hwa;
3163 if (rel)
3164 if_rele(p);
3165 done:
3166 CURVNET_RESTORE();
3167 NET_EPOCH_EXIT(et);
3168 }
3169
3170 static int
vxlan_clone_create(struct if_clone * ifc,int unit,caddr_t params)3171 vxlan_clone_create(struct if_clone *ifc, int unit, caddr_t params)
3172 {
3173 struct vxlan_softc *sc;
3174 struct ifnet *ifp;
3175 struct ifvxlanparam vxlp;
3176 int error;
3177
3178 sc = malloc(sizeof(struct vxlan_softc), M_VXLAN, M_WAITOK | M_ZERO);
3179 sc->vxl_unit = unit;
3180 vxlan_set_default_config(sc);
3181 error = vxlan_stats_alloc(sc);
3182 if (error != 0)
3183 goto fail;
3184
3185 if (params != 0) {
3186 error = copyin(params, &vxlp, sizeof(vxlp));
3187 if (error)
3188 goto fail;
3189
3190 error = vxlan_set_user_config(sc, &vxlp);
3191 if (error)
3192 goto fail;
3193 }
3194
3195 ifp = if_alloc(IFT_ETHER);
3196 if (ifp == NULL) {
3197 error = ENOSPC;
3198 goto fail;
3199 }
3200
3201 sc->vxl_ifp = ifp;
3202 rm_init(&sc->vxl_lock, "vxlanrm");
3203 callout_init_rw(&sc->vxl_callout, &sc->vxl_lock, 0);
3204 sc->vxl_port_hash_key = arc4random();
3205 vxlan_ftable_init(sc);
3206
3207 vxlan_sysctl_setup(sc);
3208
3209 ifp->if_softc = sc;
3210 if_initname(ifp, vxlan_name, unit);
3211 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
3212 ifp->if_init = vxlan_init;
3213 ifp->if_ioctl = vxlan_ioctl;
3214 ifp->if_transmit = vxlan_transmit;
3215 ifp->if_qflush = vxlan_qflush;
3216 ifp->if_capabilities = VXLAN_BASIC_IFCAPS;
3217 ifp->if_capenable = VXLAN_BASIC_IFCAPS;
3218 sc->vxl_reqcap = -1;
3219 vxlan_set_hwcaps(sc);
3220
3221 ifmedia_init(&sc->vxl_media, 0, vxlan_media_change, vxlan_media_status);
3222 ifmedia_add(&sc->vxl_media, IFM_ETHER | IFM_AUTO, 0, NULL);
3223 ifmedia_set(&sc->vxl_media, IFM_ETHER | IFM_AUTO);
3224
3225 ether_gen_addr(ifp, &sc->vxl_hwaddr);
3226 ether_ifattach(ifp, sc->vxl_hwaddr.octet);
3227
3228 ifp->if_baudrate = 0;
3229
3230 VXLAN_WLOCK(sc);
3231 vxlan_setup_interface_hdrlen(sc);
3232 VXLAN_WUNLOCK(sc);
3233
3234 return (0);
3235
3236 fail:
3237 free(sc, M_VXLAN);
3238 return (error);
3239 }
3240
3241 static void
vxlan_clone_destroy(struct ifnet * ifp)3242 vxlan_clone_destroy(struct ifnet *ifp)
3243 {
3244 struct vxlan_softc *sc;
3245
3246 sc = ifp->if_softc;
3247
3248 vxlan_teardown(sc);
3249
3250 vxlan_ftable_flush(sc, 1);
3251
3252 ether_ifdetach(ifp);
3253 if_free(ifp);
3254 ifmedia_removeall(&sc->vxl_media);
3255
3256 vxlan_ftable_fini(sc);
3257
3258 vxlan_sysctl_destroy(sc);
3259 rm_destroy(&sc->vxl_lock);
3260 vxlan_stats_free(sc);
3261 free(sc, M_VXLAN);
3262 }
3263
3264 /* BMV: Taken from if_bridge. */
3265 static uint32_t
vxlan_mac_hash(struct vxlan_softc * sc,const uint8_t * addr)3266 vxlan_mac_hash(struct vxlan_softc *sc, const uint8_t *addr)
3267 {
3268 uint32_t a = 0x9e3779b9, b = 0x9e3779b9, c = sc->vxl_ftable_hash_key;
3269
3270 b += addr[5] << 8;
3271 b += addr[4];
3272 a += addr[3] << 24;
3273 a += addr[2] << 16;
3274 a += addr[1] << 8;
3275 a += addr[0];
3276
3277 /*
3278 * The following hash function is adapted from "Hash Functions" by Bob Jenkins
3279 * ("Algorithm Alley", Dr. Dobbs Journal, September 1997).
3280 */
3281 #define mix(a, b, c) \
3282 do { \
3283 a -= b; a -= c; a ^= (c >> 13); \
3284 b -= c; b -= a; b ^= (a << 8); \
3285 c -= a; c -= b; c ^= (b >> 13); \
3286 a -= b; a -= c; a ^= (c >> 12); \
3287 b -= c; b -= a; b ^= (a << 16); \
3288 c -= a; c -= b; c ^= (b >> 5); \
3289 a -= b; a -= c; a ^= (c >> 3); \
3290 b -= c; b -= a; b ^= (a << 10); \
3291 c -= a; c -= b; c ^= (b >> 15); \
3292 } while (0)
3293
3294 mix(a, b, c);
3295
3296 #undef mix
3297
3298 return (c);
3299 }
3300
3301 static int
vxlan_media_change(struct ifnet * ifp)3302 vxlan_media_change(struct ifnet *ifp)
3303 {
3304
3305 /* Ignore. */
3306 return (0);
3307 }
3308
3309 static void
vxlan_media_status(struct ifnet * ifp,struct ifmediareq * ifmr)3310 vxlan_media_status(struct ifnet *ifp, struct ifmediareq *ifmr)
3311 {
3312
3313 ifmr->ifm_status = IFM_ACTIVE | IFM_AVALID;
3314 ifmr->ifm_active = IFM_ETHER | IFM_FDX;
3315 }
3316
3317 static int
vxlan_sockaddr_cmp(const union vxlan_sockaddr * vxladdr,const struct sockaddr * sa)3318 vxlan_sockaddr_cmp(const union vxlan_sockaddr *vxladdr,
3319 const struct sockaddr *sa)
3320 {
3321
3322 return (bcmp(&vxladdr->sa, sa, vxladdr->sa.sa_len));
3323 }
3324
3325 static void
vxlan_sockaddr_copy(union vxlan_sockaddr * vxladdr,const struct sockaddr * sa)3326 vxlan_sockaddr_copy(union vxlan_sockaddr *vxladdr,
3327 const struct sockaddr *sa)
3328 {
3329
3330 MPASS(sa->sa_family == AF_INET || sa->sa_family == AF_INET6);
3331 bzero(vxladdr, sizeof(*vxladdr));
3332
3333 if (sa->sa_family == AF_INET) {
3334 vxladdr->in4 = *satoconstsin(sa);
3335 vxladdr->in4.sin_len = sizeof(struct sockaddr_in);
3336 } else if (sa->sa_family == AF_INET6) {
3337 vxladdr->in6 = *satoconstsin6(sa);
3338 vxladdr->in6.sin6_len = sizeof(struct sockaddr_in6);
3339 }
3340 }
3341
3342 static int
vxlan_sockaddr_in_equal(const union vxlan_sockaddr * vxladdr,const struct sockaddr * sa)3343 vxlan_sockaddr_in_equal(const union vxlan_sockaddr *vxladdr,
3344 const struct sockaddr *sa)
3345 {
3346 int equal;
3347
3348 if (sa->sa_family == AF_INET) {
3349 const struct in_addr *in4 = &satoconstsin(sa)->sin_addr;
3350 equal = in4->s_addr == vxladdr->in4.sin_addr.s_addr;
3351 } else if (sa->sa_family == AF_INET6) {
3352 const struct in6_addr *in6 = &satoconstsin6(sa)->sin6_addr;
3353 equal = IN6_ARE_ADDR_EQUAL(in6, &vxladdr->in6.sin6_addr);
3354 } else
3355 equal = 0;
3356
3357 return (equal);
3358 }
3359
3360 static void
vxlan_sockaddr_in_copy(union vxlan_sockaddr * vxladdr,const struct sockaddr * sa)3361 vxlan_sockaddr_in_copy(union vxlan_sockaddr *vxladdr,
3362 const struct sockaddr *sa)
3363 {
3364
3365 MPASS(sa->sa_family == AF_INET || sa->sa_family == AF_INET6);
3366
3367 if (sa->sa_family == AF_INET) {
3368 const struct in_addr *in4 = &satoconstsin(sa)->sin_addr;
3369 vxladdr->in4.sin_family = AF_INET;
3370 vxladdr->in4.sin_len = sizeof(struct sockaddr_in);
3371 vxladdr->in4.sin_addr = *in4;
3372 } else if (sa->sa_family == AF_INET6) {
3373 const struct in6_addr *in6 = &satoconstsin6(sa)->sin6_addr;
3374 vxladdr->in6.sin6_family = AF_INET6;
3375 vxladdr->in6.sin6_len = sizeof(struct sockaddr_in6);
3376 vxladdr->in6.sin6_addr = *in6;
3377 }
3378 }
3379
3380 static int
vxlan_sockaddr_supported(const union vxlan_sockaddr * vxladdr,int unspec)3381 vxlan_sockaddr_supported(const union vxlan_sockaddr *vxladdr, int unspec)
3382 {
3383 const struct sockaddr *sa;
3384 int supported;
3385
3386 sa = &vxladdr->sa;
3387 supported = 0;
3388
3389 if (sa->sa_family == AF_UNSPEC && unspec != 0) {
3390 supported = 1;
3391 } else if (sa->sa_family == AF_INET) {
3392 #ifdef INET
3393 supported = 1;
3394 #endif
3395 } else if (sa->sa_family == AF_INET6) {
3396 #ifdef INET6
3397 supported = 1;
3398 #endif
3399 }
3400
3401 return (supported);
3402 }
3403
3404 static int
vxlan_sockaddr_in_any(const union vxlan_sockaddr * vxladdr)3405 vxlan_sockaddr_in_any(const union vxlan_sockaddr *vxladdr)
3406 {
3407 const struct sockaddr *sa;
3408 int any;
3409
3410 sa = &vxladdr->sa;
3411
3412 if (sa->sa_family == AF_INET) {
3413 const struct in_addr *in4 = &satoconstsin(sa)->sin_addr;
3414 any = in4->s_addr == INADDR_ANY;
3415 } else if (sa->sa_family == AF_INET6) {
3416 const struct in6_addr *in6 = &satoconstsin6(sa)->sin6_addr;
3417 any = IN6_IS_ADDR_UNSPECIFIED(in6);
3418 } else
3419 any = -1;
3420
3421 return (any);
3422 }
3423
3424 static int
vxlan_sockaddr_in_multicast(const union vxlan_sockaddr * vxladdr)3425 vxlan_sockaddr_in_multicast(const union vxlan_sockaddr *vxladdr)
3426 {
3427 const struct sockaddr *sa;
3428 int mc;
3429
3430 sa = &vxladdr->sa;
3431
3432 if (sa->sa_family == AF_INET) {
3433 const struct in_addr *in4 = &satoconstsin(sa)->sin_addr;
3434 mc = IN_MULTICAST(ntohl(in4->s_addr));
3435 } else if (sa->sa_family == AF_INET6) {
3436 const struct in6_addr *in6 = &satoconstsin6(sa)->sin6_addr;
3437 mc = IN6_IS_ADDR_MULTICAST(in6);
3438 } else
3439 mc = -1;
3440
3441 return (mc);
3442 }
3443
3444 static int
vxlan_sockaddr_in6_embedscope(union vxlan_sockaddr * vxladdr)3445 vxlan_sockaddr_in6_embedscope(union vxlan_sockaddr *vxladdr)
3446 {
3447 int error;
3448
3449 MPASS(VXLAN_SOCKADDR_IS_IPV6(vxladdr));
3450 #ifdef INET6
3451 error = sa6_embedscope(&vxladdr->in6, V_ip6_use_defzone);
3452 #else
3453 error = EAFNOSUPPORT;
3454 #endif
3455
3456 return (error);
3457 }
3458
3459 static int
vxlan_can_change_config(struct vxlan_softc * sc)3460 vxlan_can_change_config(struct vxlan_softc *sc)
3461 {
3462 struct ifnet *ifp;
3463
3464 ifp = sc->vxl_ifp;
3465 VXLAN_LOCK_ASSERT(sc);
3466
3467 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
3468 return (0);
3469 if (sc->vxl_flags & (VXLAN_FLAG_INIT | VXLAN_FLAG_TEARDOWN))
3470 return (0);
3471
3472 return (1);
3473 }
3474
3475 static int
vxlan_check_vni(uint32_t vni)3476 vxlan_check_vni(uint32_t vni)
3477 {
3478
3479 return (vni >= VXLAN_VNI_MAX);
3480 }
3481
3482 static int
vxlan_check_ttl(int ttl)3483 vxlan_check_ttl(int ttl)
3484 {
3485
3486 return (ttl > MAXTTL);
3487 }
3488
3489 static int
vxlan_check_ftable_timeout(uint32_t timeout)3490 vxlan_check_ftable_timeout(uint32_t timeout)
3491 {
3492
3493 return (timeout > VXLAN_FTABLE_MAX_TIMEOUT);
3494 }
3495
3496 static int
vxlan_check_ftable_max(uint32_t max)3497 vxlan_check_ftable_max(uint32_t max)
3498 {
3499
3500 return (max > VXLAN_FTABLE_MAX);
3501 }
3502
3503 static void
vxlan_sysctl_setup(struct vxlan_softc * sc)3504 vxlan_sysctl_setup(struct vxlan_softc *sc)
3505 {
3506 struct sysctl_ctx_list *ctx;
3507 struct sysctl_oid *node;
3508 struct vxlan_statistics *stats;
3509 char namebuf[8];
3510
3511 ctx = &sc->vxl_sysctl_ctx;
3512 stats = &sc->vxl_stats;
3513 snprintf(namebuf, sizeof(namebuf), "%d", sc->vxl_unit);
3514
3515 sysctl_ctx_init(ctx);
3516 sc->vxl_sysctl_node = SYSCTL_ADD_NODE(ctx,
3517 SYSCTL_STATIC_CHILDREN(_net_link_vxlan), OID_AUTO, namebuf,
3518 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3519
3520 node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->vxl_sysctl_node),
3521 OID_AUTO, "ftable", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3522 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "count",
3523 CTLFLAG_RD, &sc->vxl_ftable_cnt, 0,
3524 "Number of entries in fowarding table");
3525 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "max",
3526 CTLFLAG_RD, &sc->vxl_ftable_max, 0,
3527 "Maximum number of entries allowed in fowarding table");
3528 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "timeout",
3529 CTLFLAG_RD, &sc->vxl_ftable_timeout, 0,
3530 "Number of seconds between prunes of the forwarding table");
3531 SYSCTL_ADD_PROC(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "dump",
3532 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE | CTLFLAG_SKIP,
3533 sc, 0, vxlan_ftable_sysctl_dump, "A",
3534 "Dump the forwarding table entries");
3535
3536 node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->vxl_sysctl_node),
3537 OID_AUTO, "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3538 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(node), OID_AUTO,
3539 "ftable_nospace", CTLFLAG_RD, &stats->ftable_nospace, 0,
3540 "Fowarding table reached maximum entries");
3541 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(node), OID_AUTO,
3542 "ftable_lock_upgrade_failed", CTLFLAG_RD,
3543 &stats->ftable_lock_upgrade_failed, 0,
3544 "Forwarding table update required lock upgrade");
3545
3546 SYSCTL_ADD_COUNTER_U64(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "txcsum",
3547 CTLFLAG_RD, &stats->txcsum,
3548 "# of times hardware assisted with tx checksum");
3549 SYSCTL_ADD_COUNTER_U64(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "tso",
3550 CTLFLAG_RD, &stats->tso, "# of times hardware assisted with TSO");
3551 SYSCTL_ADD_COUNTER_U64(ctx, SYSCTL_CHILDREN(node), OID_AUTO, "rxcsum",
3552 CTLFLAG_RD, &stats->rxcsum,
3553 "# of times hardware assisted with rx checksum");
3554 }
3555
3556 static void
vxlan_sysctl_destroy(struct vxlan_softc * sc)3557 vxlan_sysctl_destroy(struct vxlan_softc *sc)
3558 {
3559
3560 sysctl_ctx_free(&sc->vxl_sysctl_ctx);
3561 sc->vxl_sysctl_node = NULL;
3562 }
3563
3564 static int
vxlan_tunable_int(struct vxlan_softc * sc,const char * knob,int def)3565 vxlan_tunable_int(struct vxlan_softc *sc, const char *knob, int def)
3566 {
3567 char path[64];
3568
3569 snprintf(path, sizeof(path), "net.link.vxlan.%d.%s",
3570 sc->vxl_unit, knob);
3571 TUNABLE_INT_FETCH(path, &def);
3572
3573 return (def);
3574 }
3575
3576 static void
vxlan_ifdetach_event(void * arg __unused,struct ifnet * ifp)3577 vxlan_ifdetach_event(void *arg __unused, struct ifnet *ifp)
3578 {
3579 struct vxlan_softc_head list;
3580 struct vxlan_socket *vso;
3581 struct vxlan_softc *sc, *tsc;
3582
3583 LIST_INIT(&list);
3584
3585 if (ifp->if_flags & IFF_RENAMING)
3586 return;
3587 if ((ifp->if_flags & IFF_MULTICAST) == 0)
3588 return;
3589
3590 VXLAN_LIST_LOCK();
3591 LIST_FOREACH(vso, &vxlan_socket_list, vxlso_entry)
3592 vxlan_socket_ifdetach(vso, ifp, &list);
3593 VXLAN_LIST_UNLOCK();
3594
3595 LIST_FOREACH_SAFE(sc, &list, vxl_ifdetach_list, tsc) {
3596 LIST_REMOVE(sc, vxl_ifdetach_list);
3597
3598 sx_xlock(&vxlan_sx);
3599 VXLAN_WLOCK(sc);
3600 if (sc->vxl_flags & VXLAN_FLAG_INIT)
3601 vxlan_init_wait(sc);
3602 vxlan_teardown_locked(sc);
3603 sx_xunlock(&vxlan_sx);
3604 }
3605 }
3606
3607 static void
vxlan_load(void)3608 vxlan_load(void)
3609 {
3610
3611 mtx_init(&vxlan_list_mtx, "vxlan list", NULL, MTX_DEF);
3612 LIST_INIT(&vxlan_socket_list);
3613 vxlan_ifdetach_event_tag = EVENTHANDLER_REGISTER(ifnet_departure_event,
3614 vxlan_ifdetach_event, NULL, EVENTHANDLER_PRI_ANY);
3615 vxlan_cloner = if_clone_simple(vxlan_name, vxlan_clone_create,
3616 vxlan_clone_destroy, 0);
3617 }
3618
3619 static void
vxlan_unload(void)3620 vxlan_unload(void)
3621 {
3622
3623 EVENTHANDLER_DEREGISTER(ifnet_departure_event,
3624 vxlan_ifdetach_event_tag);
3625 if_clone_detach(vxlan_cloner);
3626 mtx_destroy(&vxlan_list_mtx);
3627 MPASS(LIST_EMPTY(&vxlan_socket_list));
3628 }
3629
3630 static int
vxlan_modevent(module_t mod,int type,void * unused)3631 vxlan_modevent(module_t mod, int type, void *unused)
3632 {
3633 int error;
3634
3635 error = 0;
3636
3637 switch (type) {
3638 case MOD_LOAD:
3639 vxlan_load();
3640 break;
3641 case MOD_UNLOAD:
3642 vxlan_unload();
3643 break;
3644 default:
3645 error = ENOTSUP;
3646 break;
3647 }
3648
3649 return (error);
3650 }
3651
3652 static moduledata_t vxlan_mod = {
3653 "if_vxlan",
3654 vxlan_modevent,
3655 0
3656 };
3657
3658 DECLARE_MODULE(if_vxlan, vxlan_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
3659 MODULE_VERSION(if_vxlan, 1);
3660