]> git.saurik.com Git - apple/xnu.git/blame - bsd/net/if_utun.c
projects / apple / xnu.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
xnu-4903.241.1.tar.gz
[apple/xnu.git] / bsd / net / if_utun.c
CommitLineData
b0d623f7 1/*
a39ff7e2 2 * Copyright (c) 2008-2018 Apple Inc. All rights reserved.
b0d623f7
A		 3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28
29
30
31/* ----------------------------------------------------------------------------------
32Application of kernel control for interface creation
33
34Theory of operation:
35utun (user tunnel) acts as glue between kernel control sockets and network interfaces.
36This kernel control will register an interface for every client that connects.
37---------------------------------------------------------------------------------- */
38
39#include <sys/systm.h>
40#include <sys/kern_control.h>
41#include <net/kpi_protocol.h>
42#include <net/kpi_interface.h>
43#include <sys/socket.h>
44#include <net/if.h>
45#include <net/if_types.h>
46#include <net/bpf.h>
47#include <net/if_utun.h>
b0d623f7
A		 48#include <sys/mbuf.h>
49#include <sys/sockio.h>
50#include <netinet/in.h>
5ba3f43e 51#include <netinet/ip.h>
b0d623f7
A		 52#include <netinet6/in6_var.h>
53#include <netinet6/in6_var.h>
54#include <sys/kauth.h>
5ba3f43e
A		 55#include <net/necp.h>
56#include <kern/zalloc.h>
57
58#define UTUN_NEXUS 0
59
5ba3f43e
A		 60#if UTUN_NEXUS
61static nexus_controller_t utun_ncd;
62static int utun_ncd_refcount;
63static uuid_t utun_kpipe_uuid;
64static uuid_t utun_nx_dom_prov;
65
66typedef struct utun_nx {
67 uuid_t if_provider;
68 uuid_t if_instance;
69 uuid_t ms_provider;
70 uuid_t ms_instance;
71 uuid_t ms_device;
72 uuid_t ms_host;
73 uuid_t ms_agent;
74} *utun_nx_t;
75
76#endif // UTUN_NEXUS
77
78/* Control block allocated for each kernel control connection */
79struct utun_pcb {
80 TAILQ_ENTRY(utun_pcb) utun_chain;
81 kern_ctl_ref utun_ctlref;
82 ifnet_t utun_ifp;
83 u_int32_t utun_unit;
84 u_int32_t utun_unique_id;
85 u_int32_t utun_flags;
86 int utun_ext_ifdata_stats;
87 u_int32_t utun_max_pending_packets;
88 char utun_if_xname[IFXNAMSIZ];
89 char utun_unique_name[IFXNAMSIZ];
90 // PCB lock protects state fields and rings
91 decl_lck_rw_data(, utun_pcb_lock);
92 struct mbuf * utun_input_chain;
93 struct mbuf * utun_input_chain_last;
94 // Input chain lock protects the list of input mbufs
95 // The input chain lock must be taken AFTER the PCB lock if both are held
96 lck_mtx_t utun_input_chain_lock;
5ba3f43e
A		 97
98#if UTUN_NEXUS
99 struct utun_nx utun_nx;
100 int utun_kpipe_enabled;
101 uuid_t utun_kpipe_uuid;
102 void * utun_kpipe_rxring;
103 void * utun_kpipe_txring;
a39ff7e2 104 kern_pbufpool_t utun_kpipe_pp;
5ba3f43e
A		 105
106 kern_nexus_t utun_netif_nexus;
a39ff7e2 107 kern_pbufpool_t utun_netif_pp;
5ba3f43e
A		 108 void * utun_netif_rxring;
109 void * utun_netif_txring;
110 uint64_t utun_netif_txring_size;
5c9f4661
A		 111
112 u_int32_t utun_slot_size;
113 u_int32_t utun_netif_ring_size;
114 u_int32_t utun_tx_fsw_ring_size;
115 u_int32_t utun_rx_fsw_ring_size;
116 bool utun_use_netif;
9d749ea3 117 bool utun_needs_netagent;
5ba3f43e
A		 118#endif // UTUN_NEXUS
119};
39037602 120
b0d623f7 121/* Kernel Control functions */
5c9f4661
A		 122static errno_t utun_ctl_bind(kern_ctl_ref kctlref, struct sockaddr_ctl *sac,
123 void **unitinfo);
b0d623f7
A		 124static errno_t utun_ctl_connect(kern_ctl_ref kctlref, struct sockaddr_ctl *sac,
125 void **unitinfo);
126static errno_t utun_ctl_disconnect(kern_ctl_ref kctlref, u_int32_t unit,
127 void *unitinfo);
128static errno_t utun_ctl_send(kern_ctl_ref kctlref, u_int32_t unit,
129 void *unitinfo, mbuf_t m, int flags);
130static errno_t utun_ctl_getopt(kern_ctl_ref kctlref, u_int32_t unit, void *unitinfo,
131 int opt, void *data, size_t *len);
132static errno_t utun_ctl_setopt(kern_ctl_ref kctlref, u_int32_t unit, void *unitinfo,
133 int opt, void *data, size_t len);
fe8ab488
A		 134static void utun_ctl_rcvd(kern_ctl_ref kctlref, u_int32_t unit, void *unitinfo,
135 int flags);
b0d623f7
A		 136
137/* Network Interface functions */
fe8ab488 138static void utun_start(ifnet_t interface);
5ba3f43e
A		 139static errno_t utun_framer(ifnet_t interface, mbuf_t *packet,
140 const struct sockaddr *dest, const char *desk_linkaddr,
141 const char *frame_type, u_int32_t *prepend_len, u_int32_t *postpend_len);
b0d623f7
A		 142static errno_t utun_output(ifnet_t interface, mbuf_t data);
143static errno_t utun_demux(ifnet_t interface, mbuf_t data, char *frame_header,
144 protocol_family_t *protocol);
b0d623f7
A		 145static errno_t utun_add_proto(ifnet_t interface, protocol_family_t protocol,
146 const struct ifnet_demux_desc *demux_array,
147 u_int32_t demux_count);
148static errno_t utun_del_proto(ifnet_t interface, protocol_family_t protocol);
149static errno_t utun_ioctl(ifnet_t interface, u_long cmd, void *data);
150static void utun_detached(ifnet_t interface);
151
152/* Protocol handlers */
153static errno_t utun_attach_proto(ifnet_t interface, protocol_family_t proto);
154static errno_t utun_proto_input(ifnet_t interface, protocol_family_t protocol,
155 mbuf_t m, char *frame_header);
156static errno_t utun_proto_pre_output(ifnet_t interface, protocol_family_t protocol,
157 mbuf_t *packet, const struct sockaddr *dest, void *route,
158 char *frame_type, char *link_layer_dest);
5c9f4661 159static errno_t utun_pkt_input(struct utun_pcb *pcb, mbuf_t m);
39037602 160
5ba3f43e
A		 161#if UTUN_NEXUS
162
5c9f4661 163#define UTUN_IF_DEFAULT_SLOT_SIZE 2048
5ba3f43e
A		 164#define UTUN_IF_DEFAULT_RING_SIZE 64
165#define UTUN_IF_DEFAULT_TX_FSW_RING_SIZE 64
166#define UTUN_IF_DEFAULT_RX_FSW_RING_SIZE 128
a39ff7e2 167#define UTUN_IF_DEFAULT_BUF_SEG_SIZE skmem_usr_buf_seg_size
5ba3f43e
A		 168#define UTUN_IF_HEADROOM_SIZE 32
169
170#define UTUN_IF_MIN_RING_SIZE 16
171#define UTUN_IF_MAX_RING_SIZE 1024
172
5c9f4661
A		 173#define UTUN_IF_MIN_SLOT_SIZE 1024
174#define UTUN_IF_MAX_SLOT_SIZE 4096
175
5ba3f43e
A		 176static int sysctl_if_utun_ring_size SYSCTL_HANDLER_ARGS;
177static int sysctl_if_utun_tx_fsw_ring_size SYSCTL_HANDLER_ARGS;
178static int sysctl_if_utun_rx_fsw_ring_size SYSCTL_HANDLER_ARGS;
179
180static int if_utun_ring_size = UTUN_IF_DEFAULT_RING_SIZE;
181static int if_utun_tx_fsw_ring_size = UTUN_IF_DEFAULT_TX_FSW_RING_SIZE;
182static int if_utun_rx_fsw_ring_size = UTUN_IF_DEFAULT_RX_FSW_RING_SIZE;
183
184SYSCTL_DECL(_net_utun);
185SYSCTL_NODE(_net, OID_AUTO, utun, CTLFLAG_RW | CTLFLAG_LOCKED, 0, "UTun");
186
187SYSCTL_PROC(_net_utun, OID_AUTO, ring_size, CTLTYPE_INT | CTLFLAG_LOCKED | CTLFLAG_RW,
188 &if_utun_ring_size, UTUN_IF_DEFAULT_RING_SIZE, &sysctl_if_utun_ring_size, "I", "");
189SYSCTL_PROC(_net_utun, OID_AUTO, tx_fsw_ring_size, CTLTYPE_INT | CTLFLAG_LOCKED | CTLFLAG_RW,
190 &if_utun_tx_fsw_ring_size, UTUN_IF_DEFAULT_TX_FSW_RING_SIZE, &sysctl_if_utun_tx_fsw_ring_size, "I", "");
191SYSCTL_PROC(_net_utun, OID_AUTO, rx_fsw_ring_size, CTLTYPE_INT | CTLFLAG_LOCKED | CTLFLAG_RW,
192 &if_utun_rx_fsw_ring_size, UTUN_IF_DEFAULT_RX_FSW_RING_SIZE, &sysctl_if_utun_rx_fsw_ring_size, "I", "");
193
194static errno_t
195utun_register_nexus(void);
196
197static errno_t
198utun_netif_prepare(__unused kern_nexus_t nexus, ifnet_t ifp);
199static errno_t
200utun_nexus_pre_connect(kern_nexus_provider_t nxprov,
201 proc_t p, kern_nexus_t nexus,
202 nexus_port_t nexus_port, kern_channel_t channel, void **ch_ctx);
203static errno_t
204utun_nexus_connected(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
205 kern_channel_t channel);
206static void
207utun_netif_pre_disconnect(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
208 kern_channel_t channel);
209static void
210utun_nexus_pre_disconnect(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
211 kern_channel_t channel);
212static void
213utun_nexus_disconnected(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
214 kern_channel_t channel);
215static errno_t
216utun_kpipe_ring_init(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
217 kern_channel_t channel, kern_channel_ring_t ring, boolean_t is_tx_ring,
218 void **ring_ctx);
219static void
220utun_kpipe_ring_fini(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
221 kern_channel_ring_t ring);
222static errno_t
223utun_kpipe_sync_tx(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
224 kern_channel_ring_t ring, uint32_t flags);
225static errno_t
226utun_kpipe_sync_rx(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
227 kern_channel_ring_t ring, uint32_t flags);
228#endif // UTUN_NEXUS
39037602
A		 229
230#define UTUN_DEFAULT_MTU 1500
231#define UTUN_HEADER_SIZE(_pcb) (sizeof(u_int32_t) + (((_pcb)->utun_flags & UTUN_FLAGS_ENABLE_PROC_UUID) ? sizeof(uuid_t) : 0))
b0d623f7
A		 232
233static kern_ctl_ref utun_kctlref;
234static u_int32_t utun_family;
5ba3f43e
A		 235static lck_attr_t *utun_lck_attr;
236static lck_grp_attr_t *utun_lck_grp_attr;
237static lck_grp_t *utun_lck_grp;
238static lck_mtx_t utun_lock;
239
240TAILQ_HEAD(utun_list, utun_pcb) utun_head;
241
242#define UTUN_PCB_ZONE_MAX 32
243#define UTUN_PCB_ZONE_NAME "net.if_utun"
244
245static unsigned int utun_pcb_size; /* size of zone element */
246static struct zone *utun_pcb_zone; /* zone for utun_pcb */
247
248#if UTUN_NEXUS
249
250static int
251sysctl_if_utun_ring_size SYSCTL_HANDLER_ARGS
252{
253#pragma unused(arg1, arg2)
254 int value = if_utun_ring_size;
255
256 int error = sysctl_handle_int(oidp, &value, 0, req);
257 if (error || !req->newptr) {
258 return (error);
259 }
260
261 if (value < UTUN_IF_MIN_RING_SIZE ||
262 value > UTUN_IF_MAX_RING_SIZE) {
263 return (EINVAL);
264 }
265
266 if_utun_ring_size = value;
267
268 return (0);
269}
270
271static int
272sysctl_if_utun_tx_fsw_ring_size SYSCTL_HANDLER_ARGS
273{
274#pragma unused(arg1, arg2)
275 int value = if_utun_tx_fsw_ring_size;
276
277 int error = sysctl_handle_int(oidp, &value, 0, req);
278 if (error || !req->newptr) {
279 return (error);
280 }
281
282 if (value < UTUN_IF_MIN_RING_SIZE ||
283 value > UTUN_IF_MAX_RING_SIZE) {
284 return (EINVAL);
285 }
286
287 if_utun_tx_fsw_ring_size = value;
288
289 return (0);
290}
291
292static int
293sysctl_if_utun_rx_fsw_ring_size SYSCTL_HANDLER_ARGS
294{
295#pragma unused(arg1, arg2)
296 int value = if_utun_rx_fsw_ring_size;
297
298 int error = sysctl_handle_int(oidp, &value, 0, req);
299 if (error || !req->newptr) {
300 return (error);
301 }
302
303 if (value < UTUN_IF_MIN_RING_SIZE ||
304 value > UTUN_IF_MAX_RING_SIZE) {
305 return (EINVAL);
306 }
307
308 if_utun_rx_fsw_ring_size = value;
309
310 return (0);
311}
312
313static errno_t
314utun_netif_ring_init(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
315 kern_channel_t channel, kern_channel_ring_t ring, boolean_t is_tx_ring,
316 void **ring_ctx)
317{
318#pragma unused(nxprov)
319#pragma unused(channel)
320#pragma unused(ring_ctx)
321 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
322 if (!is_tx_ring) {
323 VERIFY(pcb->utun_netif_rxring == NULL);
324 pcb->utun_netif_rxring = ring;
325 } else {
326 VERIFY(pcb->utun_netif_txring == NULL);
327 pcb->utun_netif_txring = ring;
328 }
329 return 0;
330}
331
332static void
333utun_netif_ring_fini(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
334 kern_channel_ring_t ring)
335{
336#pragma unused(nxprov)
337 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
338 if (pcb->utun_netif_rxring == ring) {
339 pcb->utun_netif_rxring = NULL;
340 } else if (pcb->utun_netif_txring == ring) {
341 pcb->utun_netif_txring = NULL;
342 }
343}
344
345static errno_t
346utun_netif_sync_tx(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
347 kern_channel_ring_t tx_ring, uint32_t flags)
348{
349#pragma unused(nxprov)
350#pragma unused(flags)
351 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
352
353 struct netif_stats *nifs = &NX_NETIF_PRIVATE(nexus)->nif_stats;
354
355 lck_rw_lock_shared(&pcb->utun_pcb_lock);
356
357 struct kern_channel_ring_stat_increment tx_ring_stats;
358 bzero(&tx_ring_stats, sizeof(tx_ring_stats));
359 kern_channel_slot_t tx_pslot = NULL;
360 kern_channel_slot_t tx_slot = kern_channel_get_next_slot(tx_ring, NULL, NULL);
361
362 STATS_INC(nifs, NETIF_STATS_TXSYNC);
363
364 if (tx_slot == NULL) {
365 // Nothing to write, don't bother signalling
366 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
367 return 0;
368 }
369
370 if (pcb->utun_kpipe_enabled) {
371 kern_channel_ring_t rx_ring = pcb->utun_kpipe_rxring;
372 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
373
374 // Signal the kernel pipe ring to read
375 if (rx_ring != NULL) {
376 kern_channel_notify(rx_ring, 0);
377 }
378 return 0;
379 }
380
381 // If we're here, we're injecting into the utun kernel control socket
382 while (tx_slot != NULL) {
383 size_t length = 0;
384 mbuf_t data = NULL;
385
386 kern_packet_t tx_ph = kern_channel_slot_get_packet(tx_ring, tx_slot);
387
388 if (tx_ph == 0) {
389 // Advance TX ring
390 tx_pslot = tx_slot;
391 tx_slot = kern_channel_get_next_slot(tx_ring, tx_slot, NULL);
392 continue;
393 }
394 (void) kern_channel_slot_detach_packet(tx_ring, tx_slot, tx_ph);
395
396 // Advance TX ring
397 tx_pslot = tx_slot;
398 tx_slot = kern_channel_get_next_slot(tx_ring, tx_slot, NULL);
399
400 kern_buflet_t tx_buf = kern_packet_get_next_buflet(tx_ph, NULL);
401 VERIFY(tx_buf != NULL);
402
403 /* tx_baddr is the absolute buffer address */
404 uint8_t *tx_baddr = kern_buflet_get_object_address(tx_buf);
405 VERIFY(tx_baddr != 0);
406
407 bpf_tap_packet_out(pcb->utun_ifp, DLT_RAW, tx_ph, NULL, 0);
408
409 uint16_t tx_offset = kern_buflet_get_data_offset(tx_buf);
410 uint32_t tx_length = kern_buflet_get_data_length(tx_buf);
411
412 // The offset must be large enough for the headers
413 VERIFY(tx_offset >= UTUN_HEADER_SIZE(pcb));
414
415 // Find family
416 uint32_t af = 0;
417 uint8_t vhl = *(uint8_t *)(tx_baddr + tx_offset);
418 u_int ip_version = (vhl >> 4);
419 switch (ip_version) {
420 case 4: {
421 af = AF_INET;
422 break;
423 }
424 case 6: {
425 af = AF_INET6;
426 break;
427 }
428 default: {
429 printf("utun_netif_sync_tx %s: unknown ip version %u vhl %u tx_offset %u len %u header_size %zu\n",
430 pcb->utun_ifp->if_xname, ip_version, vhl, tx_offset, tx_length,
431 UTUN_HEADER_SIZE(pcb));
432 break;
433 }
434 }
435
436 tx_offset -= UTUN_HEADER_SIZE(pcb);
437 tx_length += UTUN_HEADER_SIZE(pcb);
438 tx_baddr += tx_offset;
439
5c9f4661 440 length = MIN(tx_length, pcb->utun_slot_size);
5ba3f43e
A		 441
442 // Copy in family
443 memcpy(tx_baddr, &af, sizeof(af));
444 if (pcb->utun_flags & UTUN_FLAGS_ENABLE_PROC_UUID) {
445 kern_packet_get_euuid(tx_ph, (void *)(tx_baddr + sizeof(af)));
446 }
447
448 if (length > 0) {
449 errno_t error = mbuf_gethdr(MBUF_DONTWAIT, MBUF_TYPE_HEADER, &data);
450 if (error == 0) {
451 error = mbuf_copyback(data, 0, length, tx_baddr, MBUF_DONTWAIT);
452 if (error == 0) {
453 error = utun_output(pcb->utun_ifp, data);
454 if (error != 0) {
455 printf("utun_netif_sync_tx %s - utun_output error %d\n", pcb->utun_ifp->if_xname, error);
456 }
457 } else {
458 printf("utun_netif_sync_tx %s - mbuf_copyback(%zu) error %d\n", pcb->utun_ifp->if_xname, length, error);
459 STATS_INC(nifs, NETIF_STATS_NOMEM_MBUF);
460 STATS_INC(nifs, NETIF_STATS_DROPPED);
461 mbuf_freem(data);
462 data = NULL;
463 }
464 } else {
465 printf("utun_netif_sync_tx %s - mbuf_gethdr error %d\n", pcb->utun_ifp->if_xname, error);
466 STATS_INC(nifs, NETIF_STATS_NOMEM_MBUF);
467 STATS_INC(nifs, NETIF_STATS_DROPPED);
468 }
469 } else {
470 printf("utun_netif_sync_tx %s - 0 length packet\n", pcb->utun_ifp->if_xname);
471 STATS_INC(nifs, NETIF_STATS_NOMEM_MBUF);
472 STATS_INC(nifs, NETIF_STATS_DROPPED);
473 }
474
475 kern_pbufpool_free(tx_ring->ckr_pp, tx_ph);
476
477 if (data == NULL) {
478 continue;
479 }
480
481 STATS_INC(nifs, NETIF_STATS_TXPKTS);
482 STATS_INC(nifs, NETIF_STATS_TXCOPY_MBUF);
483
484 tx_ring_stats.kcrsi_slots_transferred++;
485 tx_ring_stats.kcrsi_bytes_transferred += length;
486 }
487
488 if (tx_pslot) {
489 kern_channel_advance_slot(tx_ring, tx_pslot);
490 kern_channel_increment_ring_net_stats(tx_ring, pcb->utun_ifp, &tx_ring_stats);
491 (void)kern_channel_reclaim(tx_ring);
492 }
493
494 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
495
496 return 0;
497}
498
499static errno_t
500utun_netif_tx_doorbell(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
501 kern_channel_ring_t ring, __unused uint32_t flags)
502{
503#pragma unused(nxprov)
504 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
5ba3f43e
A		 505 boolean_t more = false;
506 errno_t rc = 0;
5ba3f43e 507
5c9f4661
A		 508 /*
509 * Refill and sync the ring; we may be racing against another thread doing
510 * an RX sync that also wants to do kr_enter(), and so use the blocking
511 * variant here.
512 */
513 rc = kern_channel_tx_refill_canblock(ring, UINT32_MAX, UINT32_MAX, true, &more);
514 if (rc != 0 && rc != EAGAIN && rc != EBUSY) {
515 printf("%s, tx refill failed %d\n", __func__, rc);
516 }
517
518 (void) kr_enter(ring, TRUE);
519 lck_rw_lock_shared(&pcb->utun_pcb_lock);
520
521 if (pcb->utun_kpipe_enabled) {
5ba3f43e
A		 522 uint32_t tx_available = kern_channel_available_slot_count(ring);
523 if (pcb->utun_netif_txring_size > 0 &&
524 tx_available >= pcb->utun_netif_txring_size - 1) {
525 // No room left in tx ring, disable output for now
526 errno_t error = ifnet_disable_output(pcb->utun_ifp);
527 if (error != 0) {
528 printf("utun_netif_tx_doorbell: ifnet_disable_output returned error %d\n", error);
5ba3f43e
A		 529 }
530 }
531 }
b0d623f7 532
5c9f4661 533 if (pcb->utun_kpipe_enabled) {
5ba3f43e
A		 534 kern_channel_ring_t rx_ring = pcb->utun_kpipe_rxring;
535
536 // Unlock while calling notify
537 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
538 // Signal the kernel pipe ring to read
539 if (rx_ring != NULL) {
540 kern_channel_notify(rx_ring, 0);
541 }
542 } else {
543 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
544 }
545
5c9f4661
A		 546 kr_exit(ring);
547
5ba3f43e
A		 548 return (0);
549}
550
551static errno_t
552utun_netif_sync_rx(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
553 kern_channel_ring_t rx_ring, uint32_t flags)
554{
555#pragma unused(nxprov)
556#pragma unused(flags)
557 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
558 struct kern_channel_ring_stat_increment rx_ring_stats;
559
560 struct netif_stats *nifs = &NX_NETIF_PRIVATE(nexus)->nif_stats;
561
562 lck_rw_lock_shared(&pcb->utun_pcb_lock);
563
564 // Reclaim user-released slots
565 (void) kern_channel_reclaim(rx_ring);
566
567 STATS_INC(nifs, NETIF_STATS_RXSYNC);
568
569 uint32_t avail = kern_channel_available_slot_count(rx_ring);
570 if (avail == 0) {
571 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
572 return 0;
573 }
574
575 struct kern_pbufpool *rx_pp = rx_ring->ckr_pp;
576 VERIFY(rx_pp != NULL);
577 bzero(&rx_ring_stats, sizeof(rx_ring_stats));
578 kern_channel_slot_t rx_pslot = NULL;
579 kern_channel_slot_t rx_slot = kern_channel_get_next_slot(rx_ring, NULL, NULL);
580
581 while (rx_slot != NULL) {
582 // Check for a waiting packet
583 lck_mtx_lock(&pcb->utun_input_chain_lock);
584 mbuf_t data = pcb->utun_input_chain;
585 if (data == NULL) {
586 lck_mtx_unlock(&pcb->utun_input_chain_lock);
587 break;
588 }
589
590 // Allocate rx packet
591 kern_packet_t rx_ph = 0;
592 errno_t error = kern_pbufpool_alloc_nosleep(rx_pp, 1, &rx_ph);
a39ff7e2 593 if (__improbable(error != 0)) {
5ba3f43e
A		 594 STATS_INC(nifs, NETIF_STATS_NOMEM_PKT);
595 STATS_INC(nifs, NETIF_STATS_DROPPED);
5ba3f43e
A		 596 lck_mtx_unlock(&pcb->utun_input_chain_lock);
597 break;
598 }
599
600 // Advance waiting packets
601 pcb->utun_input_chain = data->m_nextpkt;
602 data->m_nextpkt = NULL;
603 if (pcb->utun_input_chain == NULL) {
604 pcb->utun_input_chain_last = NULL;
605 }
606 lck_mtx_unlock(&pcb->utun_input_chain_lock);
607
608 size_t header_offset = UTUN_HEADER_SIZE(pcb);
609 size_t length = mbuf_pkthdr_len(data);
610
611 if (length < header_offset) {
612 // mbuf is too small
613 mbuf_freem(data);
614 kern_pbufpool_free(rx_pp, rx_ph);
615 STATS_INC(nifs, NETIF_STATS_BADLEN);
616 STATS_INC(nifs, NETIF_STATS_DROPPED);
617 printf("utun_netif_sync_rx %s: legacy packet length too short for header %zu < %zu\n",
618 pcb->utun_ifp->if_xname, length, header_offset);
619 continue;
620 }
621
622 length -= header_offset;
623 if (length > rx_pp->pp_buflet_size) {
624 // Flush data
625 mbuf_freem(data);
626 kern_pbufpool_free(rx_pp, rx_ph);
627 STATS_INC(nifs, NETIF_STATS_BADLEN);
628 STATS_INC(nifs, NETIF_STATS_DROPPED);
629 printf("utun_netif_sync_rx %s: legacy packet length %zu > %u\n",
630 pcb->utun_ifp->if_xname, length, rx_pp->pp_buflet_size);
631 continue;
632 }
633
634 mbuf_pkthdr_setrcvif(data, pcb->utun_ifp);
635
636 // Fillout rx packet
637 kern_buflet_t rx_buf = kern_packet_get_next_buflet(rx_ph, NULL);
638 VERIFY(rx_buf != NULL);
639 void *rx_baddr = kern_buflet_get_object_address(rx_buf);
640 VERIFY(rx_baddr != NULL);
641
642 // Copy-in data from mbuf to buflet
643 mbuf_copydata(data, header_offset, length, (void *)rx_baddr);
644 kern_packet_clear_flow_uuid(rx_ph); // Zero flow id
645
646 // Finalize and attach the packet
647 error = kern_buflet_set_data_offset(rx_buf, 0);
648 VERIFY(error == 0);
649 error = kern_buflet_set_data_length(rx_buf, length);
650 VERIFY(error == 0);
651 error = kern_packet_set_link_header_offset(rx_ph, 0);
652 VERIFY(error == 0);
653 error = kern_packet_set_network_header_offset(rx_ph, 0);
654 VERIFY(error == 0);
655 error = kern_packet_finalize(rx_ph);
656 VERIFY(error == 0);
657 error = kern_channel_slot_attach_packet(rx_ring, rx_slot, rx_ph);
658 VERIFY(error == 0);
659
660 STATS_INC(nifs, NETIF_STATS_RXPKTS);
661 STATS_INC(nifs, NETIF_STATS_RXCOPY_MBUF);
662 bpf_tap_packet_in(pcb->utun_ifp, DLT_RAW, rx_ph, NULL, 0);
663
664 rx_ring_stats.kcrsi_slots_transferred++;
665 rx_ring_stats.kcrsi_bytes_transferred += length;
666
667 mbuf_freem(data);
668
669 // Advance ring
670 rx_pslot = rx_slot;
671 rx_slot = kern_channel_get_next_slot(rx_ring, rx_slot, NULL);
672 }
673
674 struct kern_channel_ring_stat_increment tx_ring_stats;
675 bzero(&tx_ring_stats, sizeof(tx_ring_stats));
676 kern_channel_ring_t tx_ring = pcb->utun_kpipe_txring;
677 kern_channel_slot_t tx_pslot = NULL;
678 kern_channel_slot_t tx_slot = NULL;
679 if (tx_ring == NULL) {
680 // Net-If TX ring not set up yet, nothing to read
681 goto done;
682 }
683
684 // Unlock utun before entering ring
685 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
686
687 (void)kr_enter(tx_ring, TRUE);
688
689 // Lock again after entering and validate
690 lck_rw_lock_shared(&pcb->utun_pcb_lock);
691 if (tx_ring != pcb->utun_kpipe_txring) {
692 goto done;
693 }
694
695 tx_slot = kern_channel_get_next_slot(tx_ring, NULL, NULL);
696 if (tx_slot == NULL) {
697 // Nothing to read, don't bother signalling
698 goto done;
699 }
700
701 while (rx_slot != NULL && tx_slot != NULL) {
702 // Allocate rx packet
703 kern_packet_t rx_ph = 0;
704 kern_packet_t tx_ph = kern_channel_slot_get_packet(tx_ring, tx_slot);
705
706 // Advance TX ring
707 tx_pslot = tx_slot;
708 tx_slot = kern_channel_get_next_slot(tx_ring, tx_slot, NULL);
709
710 /* Skip slot if packet is zero-length or marked as dropped (QUMF_DROPPED) */
711 if (tx_ph == 0) {
712 continue;
713 }
714
a39ff7e2
A		 715 /* XXX We could try this alloc before advancing the slot to avoid
716 * dropping the packet on failure to allocate.
717 */
5ba3f43e 718 errno_t error = kern_pbufpool_alloc_nosleep(rx_pp, 1, &rx_ph);
a39ff7e2 719 if (__improbable(error != 0)) {
5ba3f43e
A		 720 STATS_INC(nifs, NETIF_STATS_NOMEM_PKT);
721 STATS_INC(nifs, NETIF_STATS_DROPPED);
5ba3f43e
A		 722 break;
723 }
724
725 kern_buflet_t tx_buf = kern_packet_get_next_buflet(tx_ph, NULL);
726 VERIFY(tx_buf != NULL);
727 uint8_t *tx_baddr = kern_buflet_get_object_address(tx_buf);
728 VERIFY(tx_baddr != 0);
729 tx_baddr += kern_buflet_get_data_offset(tx_buf);
730
731 // Check packet length
732 size_t header_offset = UTUN_HEADER_SIZE(pcb);
733 uint32_t tx_length = kern_packet_get_data_length(tx_ph);
734 if (tx_length < header_offset) {
735 // Packet is too small
736 kern_pbufpool_free(rx_pp, rx_ph);
737 STATS_INC(nifs, NETIF_STATS_BADLEN);
738 STATS_INC(nifs, NETIF_STATS_DROPPED);
739 printf("utun_netif_sync_rx %s: packet length too short for header %u < %zu\n",
740 pcb->utun_ifp->if_xname, tx_length, header_offset);
741 continue;
742 }
743
744 size_t length = MIN(tx_length - header_offset,
5c9f4661 745 pcb->utun_slot_size);
5ba3f43e
A		 746
747 tx_ring_stats.kcrsi_slots_transferred++;
748 tx_ring_stats.kcrsi_bytes_transferred += length;
749
750 // Fillout rx packet
751 kern_buflet_t rx_buf = kern_packet_get_next_buflet(rx_ph, NULL);
752 VERIFY(rx_buf != NULL);
753 void *rx_baddr = kern_buflet_get_object_address(rx_buf);
754 VERIFY(rx_baddr != NULL);
755
756 // Copy-in data from tx to rx
757 memcpy((void *)rx_baddr, (void *)(tx_baddr + header_offset), length);
758 kern_packet_clear_flow_uuid(rx_ph); // Zero flow id
759
760 // Finalize and attach the packet
761 error = kern_buflet_set_data_offset(rx_buf, 0);
762 VERIFY(error == 0);
763 error = kern_buflet_set_data_length(rx_buf, length);
764 VERIFY(error == 0);
765 error = kern_packet_set_link_header_offset(rx_ph, 0);
766 VERIFY(error == 0);
767 error = kern_packet_set_network_header_offset(rx_ph, 0);
768 VERIFY(error == 0);
769 error = kern_packet_finalize(rx_ph);
770 VERIFY(error == 0);
771 error = kern_channel_slot_attach_packet(rx_ring, rx_slot, rx_ph);
772 VERIFY(error == 0);
773
774 STATS_INC(nifs, NETIF_STATS_RXPKTS);
775 STATS_INC(nifs, NETIF_STATS_RXCOPY_DIRECT);
776 bpf_tap_packet_in(pcb->utun_ifp, DLT_RAW, rx_ph, NULL, 0);
777
778 rx_ring_stats.kcrsi_slots_transferred++;
779 rx_ring_stats.kcrsi_bytes_transferred += length;
780
781 rx_pslot = rx_slot;
782 rx_slot = kern_channel_get_next_slot(rx_ring, rx_slot, NULL);
783 }
784
785done:
786 if (rx_pslot) {
787 kern_channel_advance_slot(rx_ring, rx_pslot);
788 kern_channel_increment_ring_net_stats(rx_ring, pcb->utun_ifp, &rx_ring_stats);
789 }
790
791 if (tx_pslot) {
792 kern_channel_advance_slot(tx_ring, tx_pslot);
793 kern_channel_increment_ring_net_stats(tx_ring, pcb->utun_ifp, &tx_ring_stats);
794 (void)kern_channel_reclaim(tx_ring);
795 }
796
797 // Unlock first, then exit ring
798 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
799 if (tx_ring != NULL) {
800 if (tx_pslot != NULL) {
801 kern_channel_notify(tx_ring, 0);
802 }
803 kr_exit(tx_ring);
804 }
805
806 return 0;
807}
808
809static errno_t
810utun_nexus_ifattach(struct utun_pcb *pcb,
811 struct ifnet_init_eparams *init_params,
812 struct ifnet **ifp)
813{
814 errno_t err;
815 nexus_controller_t controller = kern_nexus_shared_controller();
816 struct kern_nexus_net_init net_init;
a39ff7e2 817 struct kern_pbufpool_init pp_init;
5ba3f43e
A		 818
819 nexus_name_t provider_name;
820 snprintf((char *)provider_name, sizeof(provider_name),
a39ff7e2 821 "com.apple.netif.%s", pcb->utun_if_xname);
5ba3f43e
A		 822
823 struct kern_nexus_provider_init prov_init = {
824 .nxpi_version = KERN_NEXUS_DOMAIN_PROVIDER_CURRENT_VERSION,
825 .nxpi_flags = NXPIF_VIRTUAL_DEVICE,
826 .nxpi_pre_connect = utun_nexus_pre_connect,
827 .nxpi_connected = utun_nexus_connected,
828 .nxpi_pre_disconnect = utun_netif_pre_disconnect,
829 .nxpi_disconnected = utun_nexus_disconnected,
830 .nxpi_ring_init = utun_netif_ring_init,
831 .nxpi_ring_fini = utun_netif_ring_fini,
832 .nxpi_slot_init = NULL,
833 .nxpi_slot_fini = NULL,
834 .nxpi_sync_tx = utun_netif_sync_tx,
835 .nxpi_sync_rx = utun_netif_sync_rx,
836 .nxpi_tx_doorbell = utun_netif_tx_doorbell,
837 };
838
839 nexus_attr_t nxa = NULL;
840 err = kern_nexus_attr_create(&nxa);
841 if (err != 0) {
842 printf("%s: kern_nexus_attr_create failed: %d\n",
843 __func__, err);
844 goto failed;
845 }
846
5c9f4661 847 uint64_t slot_buffer_size = pcb->utun_slot_size;
5ba3f43e
A		 848 err = kern_nexus_attr_set(nxa, NEXUS_ATTR_SLOT_BUF_SIZE, slot_buffer_size);
849 VERIFY(err == 0);
850
851 // Reset ring size for netif nexus to limit memory usage
5c9f4661 852 uint64_t ring_size = pcb->utun_netif_ring_size;
5ba3f43e
A		 853 err = kern_nexus_attr_set(nxa, NEXUS_ATTR_TX_SLOTS, ring_size);
854 VERIFY(err == 0);
855 err = kern_nexus_attr_set(nxa, NEXUS_ATTR_RX_SLOTS, ring_size);
856 VERIFY(err == 0);
857
858 pcb->utun_netif_txring_size = ring_size;
859
a39ff7e2
A		 860 bzero(&pp_init, sizeof (pp_init));
861 pp_init.kbi_version = KERN_PBUFPOOL_CURRENT_VERSION;
862 pp_init.kbi_packets = pcb->utun_netif_ring_size * 2;
863 pp_init.kbi_bufsize = pcb->utun_slot_size;
864 pp_init.kbi_buf_seg_size = UTUN_IF_DEFAULT_BUF_SEG_SIZE;
865 pp_init.kbi_max_frags = 1;
866 (void) snprintf((char *)pp_init.kbi_name, sizeof (pp_init.kbi_name),
867 "%s", provider_name);
868
869 err = kern_pbufpool_create(&pp_init, &pp_init, &pcb->utun_netif_pp, NULL);
870 if (err != 0) {
871 printf("%s pbufbool create failed, error %d\n", __func__, err);
872 goto failed;
873 }
874
5ba3f43e
A		 875 err = kern_nexus_controller_register_provider(controller,
876 utun_nx_dom_prov,
877 provider_name,
878 &prov_init,
879 sizeof(prov_init),
880 nxa,
881 &pcb->utun_nx.if_provider);
882 if (err != 0) {
883 printf("%s register provider failed, error %d\n",
884 __func__, err);
885 goto failed;
886 }
887
888 bzero(&net_init, sizeof(net_init));
889 net_init.nxneti_version = KERN_NEXUS_NET_CURRENT_VERSION;
890 net_init.nxneti_flags = 0;
891 net_init.nxneti_eparams = init_params;
892 net_init.nxneti_lladdr = NULL;
893 net_init.nxneti_prepare = utun_netif_prepare;
a39ff7e2 894 net_init.nxneti_tx_pbufpool = pcb->utun_netif_pp;
5ba3f43e
A		 895 err = kern_nexus_controller_alloc_net_provider_instance(controller,
896 pcb->utun_nx.if_provider,
897 pcb,
898 &pcb->utun_nx.if_instance,
899 &net_init,
900 ifp);
901 if (err != 0) {
902 printf("%s alloc_net_provider_instance failed, %d\n",
903 __func__, err);
904 kern_nexus_controller_deregister_provider(controller,
905 pcb->utun_nx.if_provider);
906 uuid_clear(pcb->utun_nx.if_provider);
907 goto failed;
908 }
909
910failed:
911 if (nxa) {
912 kern_nexus_attr_destroy(nxa);
913 }
a39ff7e2
A		 914 if (err && pcb->utun_netif_pp != NULL) {
915 kern_pbufpool_destroy(pcb->utun_netif_pp);
916 pcb->utun_netif_pp = NULL;
917 }
5ba3f43e
A		 918 return (err);
919}
920
921static void
922utun_detach_provider_and_instance(uuid_t provider, uuid_t instance)
923{
924 nexus_controller_t controller = kern_nexus_shared_controller();
925 errno_t err;
926
927 if (!uuid_is_null(instance)) {
928 err = kern_nexus_controller_free_provider_instance(controller,
929 instance);
930 if (err != 0) {
931 printf("%s free_provider_instance failed %d\n",
932 __func__, err);
933 }
934 uuid_clear(instance);
935 }
936 if (!uuid_is_null(provider)) {
937 err = kern_nexus_controller_deregister_provider(controller,
938 provider);
939 if (err != 0) {
940 printf("%s deregister_provider %d\n", __func__, err);
941 }
942 uuid_clear(provider);
943 }
944 return;
945}
946
947static void
a39ff7e2 948utun_nexus_detach(struct utun_pcb *pcb)
5ba3f43e 949{
a39ff7e2 950 utun_nx_t nx = &pcb->utun_nx;
5ba3f43e
A		 951 nexus_controller_t controller = kern_nexus_shared_controller();
952 errno_t err;
953
954 if (!uuid_is_null(nx->ms_host)) {
955 err = kern_nexus_ifdetach(controller,
956 nx->ms_instance,
957 nx->ms_host);
958 if (err != 0) {
959 printf("%s: kern_nexus_ifdetach ms host failed %d\n",
960 __func__, err);
961 }
962 }
963
964 if (!uuid_is_null(nx->ms_device)) {
965 err = kern_nexus_ifdetach(controller,
966 nx->ms_instance,
967 nx->ms_device);
968 if (err != 0) {
969 printf("%s: kern_nexus_ifdetach ms device failed %d\n",
970 __func__, err);
971 }
972 }
973
974 utun_detach_provider_and_instance(nx->if_provider,
975 nx->if_instance);
976 utun_detach_provider_and_instance(nx->ms_provider,
977 nx->ms_instance);
978
a39ff7e2
A		 979 if (pcb->utun_netif_pp != NULL) {
980 kern_pbufpool_destroy(pcb->utun_netif_pp);
981 pcb->utun_netif_pp = NULL;
982
983 }
5ba3f43e
A		 984 memset(nx, 0, sizeof(*nx));
985}
986
987static errno_t
5c9f4661
A		 988utun_create_fs_provider_and_instance(struct utun_pcb *pcb,
989 uint32_t subtype, const char *type_name,
5ba3f43e
A		 990 const char *ifname,
991 uuid_t *provider, uuid_t *instance)
992{
993 nexus_attr_t attr = NULL;
994 nexus_controller_t controller = kern_nexus_shared_controller();
995 uuid_t dom_prov;
996 errno_t err;
997 struct kern_nexus_init init;
998 nexus_name_t provider_name;
999
1000 err = kern_nexus_get_builtin_domain_provider(NEXUS_TYPE_FLOW_SWITCH,
1001 &dom_prov);
1002 if (err != 0) {
1003 printf("%s can't get %s provider, error %d\n",
1004 __func__, type_name, err);
1005 goto failed;
1006 }
1007
1008 err = kern_nexus_attr_create(&attr);
1009 if (err != 0) {
1010 printf("%s: kern_nexus_attr_create failed: %d\n",
1011 __func__, err);
1012 goto failed;
1013 }
1014
1015 err = kern_nexus_attr_set(attr, NEXUS_ATTR_EXTENSIONS, subtype);
1016 VERIFY(err == 0);
1017
5c9f4661 1018 uint64_t slot_buffer_size = pcb->utun_slot_size;
5ba3f43e
A		 1019 err = kern_nexus_attr_set(attr, NEXUS_ATTR_SLOT_BUF_SIZE, slot_buffer_size);
1020 VERIFY(err == 0);
1021
1022 // Reset ring size for flowswitch nexus to limit memory usage. Larger RX than netif.
5c9f4661 1023 uint64_t tx_ring_size = pcb->utun_tx_fsw_ring_size;
5ba3f43e
A		 1024 err = kern_nexus_attr_set(attr, NEXUS_ATTR_TX_SLOTS, tx_ring_size);
1025 VERIFY(err == 0);
5c9f4661 1026 uint64_t rx_ring_size = pcb->utun_rx_fsw_ring_size;
5ba3f43e
A		 1027 err = kern_nexus_attr_set(attr, NEXUS_ATTR_RX_SLOTS, rx_ring_size);
1028 VERIFY(err == 0);
1029
1030 snprintf((char *)provider_name, sizeof(provider_name),
1031 "com.apple.%s.%s", type_name, ifname);
1032 err = kern_nexus_controller_register_provider(controller,
1033 dom_prov,
1034 provider_name,
1035 NULL,
1036 0,
1037 attr,
1038 provider);
1039 kern_nexus_attr_destroy(attr);
1040 attr = NULL;
1041 if (err != 0) {
1042 printf("%s register %s provider failed, error %d\n",
1043 __func__, type_name, err);
1044 goto failed;
1045 }
1046 bzero(&init, sizeof (init));
1047 init.nxi_version = KERN_NEXUS_CURRENT_VERSION;
1048 err = kern_nexus_controller_alloc_provider_instance(controller,
1049 *provider,
1050 NULL,
1051 instance, &init);
1052 if (err != 0) {
1053 printf("%s alloc_provider_instance %s failed, %d\n",
1054 __func__, type_name, err);
1055 kern_nexus_controller_deregister_provider(controller,
1056 *provider);
1057 uuid_clear(*provider);
1058 }
1059failed:
1060 return (err);
1061}
1062
1063static errno_t
1064utun_multistack_attach(struct utun_pcb *pcb)
1065{
1066 nexus_controller_t controller = kern_nexus_shared_controller();
1067 errno_t err = 0;
1068 utun_nx_t nx = &pcb->utun_nx;
1069
1070 // Allocate multistack flowswitch
5c9f4661
A		 1071 err = utun_create_fs_provider_and_instance(pcb,
1072 NEXUS_EXTENSION_FSW_TYPE_MULTISTACK,
5ba3f43e
A		 1073 "multistack",
1074 pcb->utun_ifp->if_xname,
1075 &nx->ms_provider,
1076 &nx->ms_instance);
1077 if (err != 0) {
1078 printf("%s: failed to create bridge provider and instance\n",
1079 __func__);
1080 goto failed;
1081 }
1082
1083 // Attach multistack to device port
1084 err = kern_nexus_ifattach(controller, nx->ms_instance,
1085 NULL, nx->if_instance,
1086 FALSE, &nx->ms_device);
1087 if (err != 0) {
1088 printf("%s kern_nexus_ifattach ms device %d\n", __func__, err);
1089 goto failed;
1090 }
1091
1092 // Attach multistack to host port
1093 err = kern_nexus_ifattach(controller, nx->ms_instance,
1094 NULL, nx->if_instance,
1095 TRUE, &nx->ms_host);
1096 if (err != 0) {
1097 printf("%s kern_nexus_ifattach ms host %d\n", __func__, err);
1098 goto failed;
1099 }
1100
1101 // Extract the agent UUID and save for later
1102 struct kern_nexus *multistack_nx = nx_find(nx->ms_instance, false);
1103 if (multistack_nx != NULL) {
1104 struct nx_flowswitch *flowswitch = NX_FSW_PRIVATE(multistack_nx);
1105 if (flowswitch != NULL) {
1106 FSW_RLOCK(flowswitch);
1107 struct fsw_ms_context *ms_context = (struct fsw_ms_context *)flowswitch->fsw_ops_private;
1108 if (ms_context != NULL) {
1109 uuid_copy(nx->ms_agent, ms_context->mc_agent_uuid);
1110 } else {
1111 printf("utun_multistack_attach - fsw_ms_context is NULL\n");
1112 }
1113 FSW_UNLOCK(flowswitch);
1114 } else {
1115 printf("utun_multistack_attach - flowswitch is NULL\n");
1116 }
1117 nx_release(multistack_nx);
1118 } else {
1119 printf("utun_multistack_attach - unable to find multistack nexus\n");
1120 }
1121
1122 return (0);
1123
1124failed:
a39ff7e2 1125 utun_nexus_detach(pcb);
5ba3f43e
A		 1126
1127 errno_t detach_error = 0;
1128 if ((detach_error = ifnet_detach(pcb->utun_ifp)) != 0) {
1129 panic("utun_multistack_attach - ifnet_detach failed: %d\n", detach_error);
1130 /* NOT REACHED */
1131 }
1132
1133 return (err);
1134}
1135
1136static errno_t
1137utun_register_kernel_pipe_nexus(void)
1138{
1139 nexus_attr_t nxa = NULL;
1140 errno_t result;
1141
1142 lck_mtx_lock(&utun_lock);
1143 if (utun_ncd_refcount++) {
1144 lck_mtx_unlock(&utun_lock);
1145 return 0;
1146 }
1147
1148 result = kern_nexus_controller_create(&utun_ncd);
1149 if (result) {
1150 printf("%s: kern_nexus_controller_create failed: %d\n",
1151 __FUNCTION__, result);
1152 goto done;
1153 }
1154
1155 uuid_t dom_prov;
1156 result = kern_nexus_get_builtin_domain_provider(
1157 NEXUS_TYPE_KERNEL_PIPE, &dom_prov);
1158 if (result) {
1159 printf("%s: kern_nexus_get_builtin_domain_provider failed: %d\n",
1160 __FUNCTION__, result);
1161 goto done;
1162 }
1163
1164 struct kern_nexus_provider_init prov_init = {
1165 .nxpi_version = KERN_NEXUS_DOMAIN_PROVIDER_CURRENT_VERSION,
1166 .nxpi_flags = NXPIF_VIRTUAL_DEVICE,
1167 .nxpi_pre_connect = utun_nexus_pre_connect,
1168 .nxpi_connected = utun_nexus_connected,
1169 .nxpi_pre_disconnect = utun_nexus_pre_disconnect,
1170 .nxpi_disconnected = utun_nexus_disconnected,
1171 .nxpi_ring_init = utun_kpipe_ring_init,
1172 .nxpi_ring_fini = utun_kpipe_ring_fini,
1173 .nxpi_slot_init = NULL,
1174 .nxpi_slot_fini = NULL,
1175 .nxpi_sync_tx = utun_kpipe_sync_tx,
1176 .nxpi_sync_rx = utun_kpipe_sync_rx,
1177 .nxpi_tx_doorbell = NULL,
1178 };
1179
1180 result = kern_nexus_attr_create(&nxa);
1181 if (result) {
1182 printf("%s: kern_nexus_attr_create failed: %d\n",
1183 __FUNCTION__, result);
1184 goto done;
1185 }
1186
1187 uint64_t slot_buffer_size = UTUN_IF_DEFAULT_SLOT_SIZE;
1188 result = kern_nexus_attr_set(nxa, NEXUS_ATTR_SLOT_BUF_SIZE, slot_buffer_size);
1189 VERIFY(result == 0);
1190
1191 // Reset ring size for kernel pipe nexus to limit memory usage
1192 uint64_t ring_size = if_utun_ring_size;
1193 result = kern_nexus_attr_set(nxa, NEXUS_ATTR_TX_SLOTS, ring_size);
1194 VERIFY(result == 0);
1195 result = kern_nexus_attr_set(nxa, NEXUS_ATTR_RX_SLOTS, ring_size);
1196 VERIFY(result == 0);
1197
1198 result = kern_nexus_controller_register_provider(utun_ncd,
1199 dom_prov,
1200 (const uint8_t *)"com.apple.nexus.utun.kpipe",
1201 &prov_init,
1202 sizeof(prov_init),
1203 nxa,
1204 &utun_kpipe_uuid);
1205 if (result) {
1206 printf("%s: kern_nexus_controller_register_provider failed: %d\n",
1207 __FUNCTION__, result);
1208 goto done;
1209 }
1210
1211done:
1212 if (nxa) {
1213 kern_nexus_attr_destroy(nxa);
1214 }
1215
1216 if (result) {
1217 if (utun_ncd) {
1218 kern_nexus_controller_destroy(utun_ncd);
1219 utun_ncd = NULL;
1220 }
1221 utun_ncd_refcount = 0;
1222 }
1223
1224 lck_mtx_unlock(&utun_lock);
1225
1226 return result;
1227}
1228
1229static void
1230utun_unregister_kernel_pipe_nexus(void)
1231{
1232 lck_mtx_lock(&utun_lock);
1233
1234 VERIFY(utun_ncd_refcount > 0);
1235
1236 if (--utun_ncd_refcount == 0) {
1237 kern_nexus_controller_destroy(utun_ncd);
1238 utun_ncd = NULL;
1239 }
1240
1241 lck_mtx_unlock(&utun_lock);
1242}
1243
1244// For use by socket option, not internally
1245static errno_t
1246utun_disable_channel(struct utun_pcb *pcb)
1247{
1248 errno_t result;
1249 int enabled;
1250 uuid_t uuid;
1251
1252 lck_rw_lock_exclusive(&pcb->utun_pcb_lock);
1253
1254 enabled = pcb->utun_kpipe_enabled;
1255 uuid_copy(uuid, pcb->utun_kpipe_uuid);
1256
1257 VERIFY(uuid_is_null(pcb->utun_kpipe_uuid) == !enabled);
1258
1259 pcb->utun_kpipe_enabled = 0;
1260 uuid_clear(pcb->utun_kpipe_uuid);
1261
1262 lck_rw_unlock_exclusive(&pcb->utun_pcb_lock);
1263
1264 if (enabled) {
1265 result = kern_nexus_controller_free_provider_instance(utun_ncd, uuid);
1266 } else {
1267 result = ENXIO;
1268 }
1269
1270 if (!result) {
a39ff7e2
A		 1271 if (pcb->utun_kpipe_pp != NULL) {
1272 kern_pbufpool_destroy(pcb->utun_kpipe_pp);
1273 pcb->utun_kpipe_pp = NULL;
1274 }
5ba3f43e
A		 1275 utun_unregister_kernel_pipe_nexus();
1276 }
1277
1278 return result;
1279}
1280
1281static errno_t
1282utun_enable_channel(struct utun_pcb *pcb, struct proc *proc)
1283{
1284 struct kern_nexus_init init;
a39ff7e2 1285 struct kern_pbufpool_init pp_init;
5ba3f43e
A		 1286 errno_t result;
1287
d9a64523
A		 1288 kauth_cred_t cred = kauth_cred_get();
1289 result = priv_check_cred(cred, PRIV_SKYWALK_REGISTER_KERNEL_PIPE, 0);
1290 if (result) {
1291 return result;
1292 }
1293
5ba3f43e
A		 1294 result = utun_register_kernel_pipe_nexus();
1295 if (result) {
1296 return result;
1297 }
1298
1299 VERIFY(utun_ncd);
1300
1301 lck_rw_lock_exclusive(&pcb->utun_pcb_lock);
1302
1303 if (pcb->utun_kpipe_enabled) {
1304 result = EEXIST; // return success instead?
1305 goto done;
1306 }
1307
1308 /*
1309 * Make sure we can fit packets in the channel buffers and
1310 * Allow an extra 4 bytes for the protocol number header in the channel
1311 */
5c9f4661 1312 if (pcb->utun_ifp->if_mtu + UTUN_HEADER_SIZE(pcb) > pcb->utun_slot_size) {
5ba3f43e
A		 1313 result = EOPNOTSUPP;
1314 goto done;
1315 }
1316
a39ff7e2
A		 1317 bzero(&pp_init, sizeof (pp_init));
1318 pp_init.kbi_version = KERN_PBUFPOOL_CURRENT_VERSION;
1319 pp_init.kbi_packets = pcb->utun_netif_ring_size * 2;
1320 pp_init.kbi_bufsize = pcb->utun_slot_size;
1321 pp_init.kbi_buf_seg_size = UTUN_IF_DEFAULT_BUF_SEG_SIZE;
1322 pp_init.kbi_max_frags = 1;
1323 pp_init.kbi_flags |= KBIF_QUANTUM;
1324 (void) snprintf((char *)pp_init.kbi_name, sizeof (pp_init.kbi_name),
1325 "com.apple.kpipe.%s", pcb->utun_if_xname);
1326
1327 result = kern_pbufpool_create(&pp_init, &pp_init, &pcb->utun_kpipe_pp,
1328 NULL);
1329 if (result != 0) {
1330 printf("%s pbufbool create failed, error %d\n", __func__, result);
1331 goto done;
1332 }
1333
5ba3f43e
A		 1334 VERIFY(uuid_is_null(pcb->utun_kpipe_uuid));
1335 bzero(&init, sizeof (init));
1336 init.nxi_version = KERN_NEXUS_CURRENT_VERSION;
a39ff7e2 1337 init.nxi_tx_pbufpool = pcb->utun_kpipe_pp;
5ba3f43e
A		 1338 result = kern_nexus_controller_alloc_provider_instance(utun_ncd,
1339 utun_kpipe_uuid, pcb, &pcb->utun_kpipe_uuid, &init);
1340 if (result) {
1341 goto done;
1342 }
1343
1344 nexus_port_t port = NEXUS_PORT_KERNEL_PIPE_CLIENT;
1345 result = kern_nexus_controller_bind_provider_instance(utun_ncd,
1346 pcb->utun_kpipe_uuid, &port,
1347 proc_pid(proc), NULL, NULL, 0, NEXUS_BIND_PID);
1348 if (result) {
1349 kern_nexus_controller_free_provider_instance(utun_ncd,
1350 pcb->utun_kpipe_uuid);
1351 uuid_clear(pcb->utun_kpipe_uuid);
1352 goto done;
1353 }
1354
1355 pcb->utun_kpipe_enabled = 1;
1356
1357done:
1358 lck_rw_unlock_exclusive(&pcb->utun_pcb_lock);
1359
1360 if (result) {
a39ff7e2
A		 1361 if (pcb->utun_kpipe_pp != NULL) {
1362 kern_pbufpool_destroy(pcb->utun_kpipe_pp);
1363 pcb->utun_kpipe_pp = NULL;
1364 }
5ba3f43e
A		 1365 utun_unregister_kernel_pipe_nexus();
1366 }
1367
1368 return result;
1369}
1370
1371#endif // UTUN_NEXUS
b0d623f7
A		 1372
1373errno_t
1374utun_register_control(void)
1375{
5ba3f43e
A		 1376 struct kern_ctl_reg kern_ctl;
1377 errno_t result = 0;
b0d623f7 1378
b0d623f7
A		 1379 /* Find a unique value for our interface family */
1380 result = mbuf_tag_id_find(UTUN_CONTROL_NAME, &utun_family);
1381 if (result != 0) {
1382 printf("utun_register_control - mbuf_tag_id_find_internal failed: %d\n", result);
1383 return result;
1384 }
5ba3f43e
A		 1385
1386 utun_pcb_size = sizeof(struct utun_pcb);
1387 utun_pcb_zone = zinit(utun_pcb_size,
1388 UTUN_PCB_ZONE_MAX * utun_pcb_size,
1389 0, UTUN_PCB_ZONE_NAME);
1390 if (utun_pcb_zone == NULL) {
1391 printf("utun_register_control - zinit(utun_pcb) failed");
1392 return ENOMEM;
1393 }
1394
1395#if UTUN_NEXUS
1396 utun_register_nexus();
1397#endif // UTUN_NEXUS
1398
1399 TAILQ_INIT(&utun_head);
b0d623f7
A		 1400
1401 bzero(&kern_ctl, sizeof(kern_ctl));
fe8ab488 1402 strlcpy(kern_ctl.ctl_name, UTUN_CONTROL_NAME, sizeof(kern_ctl.ctl_name));
b0d623f7 1403 kern_ctl.ctl_name[sizeof(kern_ctl.ctl_name) - 1] = 0;
fe8ab488 1404 kern_ctl.ctl_flags = CTL_FLAG_PRIVILEGED | CTL_FLAG_REG_EXTENDED; /* Require root */
39236c6e
A		 1405 kern_ctl.ctl_sendsize = 512 * 1024;
1406 kern_ctl.ctl_recvsize = 512 * 1024;
5c9f4661 1407 kern_ctl.ctl_bind = utun_ctl_bind;
b0d623f7
A		 1408 kern_ctl.ctl_connect = utun_ctl_connect;
1409 kern_ctl.ctl_disconnect = utun_ctl_disconnect;
1410 kern_ctl.ctl_send = utun_ctl_send;
1411 kern_ctl.ctl_setopt = utun_ctl_setopt;
1412 kern_ctl.ctl_getopt = utun_ctl_getopt;
fe8ab488 1413 kern_ctl.ctl_rcvd = utun_ctl_rcvd;
39236c6e 1414
b0d623f7
A		 1415 result = ctl_register(&kern_ctl, &utun_kctlref);
1416 if (result != 0) {
1417 printf("utun_register_control - ctl_register failed: %d\n", result);
1418 return result;
1419 }
1420
1421 /* Register the protocol plumbers */
1422 if ((result = proto_register_plumber(PF_INET, utun_family,
1423 utun_attach_proto, NULL)) != 0) {
1424 printf("utun_register_control - proto_register_plumber(PF_INET, %d) failed: %d\n",
1425 utun_family, result);
1426 ctl_deregister(utun_kctlref);
1427 return result;
1428 }
1429
1430 /* Register the protocol plumbers */
1431 if ((result = proto_register_plumber(PF_INET6, utun_family,
1432 utun_attach_proto, NULL)) != 0) {
1433 proto_unregister_plumber(PF_INET, utun_family);
1434 ctl_deregister(utun_kctlref);
1435 printf("utun_register_control - proto_register_plumber(PF_INET6, %d) failed: %d\n",
1436 utun_family, result);
1437 return result;
1438 }
39037602 1439
5ba3f43e
A		 1440 utun_lck_attr = lck_attr_alloc_init();
1441 utun_lck_grp_attr = lck_grp_attr_alloc_init();
1442 utun_lck_grp = lck_grp_alloc_init("utun", utun_lck_grp_attr);
1443
5ba3f43e 1444 lck_mtx_init(&utun_lock, utun_lck_grp, utun_lck_attr);
5ba3f43e 1445
b0d623f7
A		 1446 return 0;
1447}
1448
1449/* Kernel control functions */
1450
5ba3f43e 1451static inline void
5c9f4661 1452utun_free_pcb(struct utun_pcb *pcb, bool in_list)
5ba3f43e
A		 1453{
1454#ifdef UTUN_NEXUS
1455 mbuf_freem_list(pcb->utun_input_chain);
1456 lck_mtx_destroy(&pcb->utun_input_chain_lock, utun_lck_grp);
1457#endif // UTUN_NEXUS
1458 lck_rw_destroy(&pcb->utun_pcb_lock, utun_lck_grp);
5c9f4661
A		 1459 if (in_list) {
1460 lck_mtx_lock(&utun_lock);
1461 TAILQ_REMOVE(&utun_head, pcb, utun_chain);
1462 lck_mtx_unlock(&utun_lock);
1463 }
5ba3f43e
A		 1464 zfree(utun_pcb_zone, pcb);
1465}
1466
b0d623f7 1467static errno_t
5c9f4661
A		 1468utun_ctl_bind(kern_ctl_ref kctlref,
1469 struct sockaddr_ctl *sac,
1470 void **unitinfo)
b0d623f7 1471{
5ba3f43e
A		 1472 struct utun_pcb *pcb = zalloc(utun_pcb_zone);
1473 memset(pcb, 0, sizeof(*pcb));
39037602 1474
b0d623f7 1475 *unitinfo = pcb;
d1ecb069
A		 1476 pcb->utun_ctlref = kctlref;
1477 pcb->utun_unit = sac->sc_unit;
fe8ab488 1478 pcb->utun_max_pending_packets = 1;
5ba3f43e 1479
5c9f4661
A		 1480#if UTUN_NEXUS
1481 pcb->utun_use_netif = false;
1482 pcb->utun_slot_size = UTUN_IF_DEFAULT_SLOT_SIZE;
1483 pcb->utun_netif_ring_size = UTUN_IF_DEFAULT_RING_SIZE;
1484 pcb->utun_tx_fsw_ring_size = UTUN_IF_DEFAULT_TX_FSW_RING_SIZE;
1485 pcb->utun_rx_fsw_ring_size = UTUN_IF_DEFAULT_RX_FSW_RING_SIZE;
1486#endif // UTUN_NEXUS
1487
5ba3f43e
A		 1488 lck_mtx_init(&pcb->utun_input_chain_lock, utun_lck_grp, utun_lck_attr);
1489 lck_rw_init(&pcb->utun_pcb_lock, utun_lck_grp, utun_lck_attr);
1490
5c9f4661
A		 1491 return (0);
1492}
1493
1494static errno_t
1495utun_ctl_connect(kern_ctl_ref kctlref,
1496 struct sockaddr_ctl *sac,
1497 void **unitinfo)
1498{
1499 struct ifnet_init_eparams utun_init = {};
1500 errno_t result = 0;
1501
1502 if (*unitinfo == NULL) {
1503 (void)utun_ctl_bind(kctlref, sac, unitinfo);
1504 }
1505
1506 struct utun_pcb *pcb = *unitinfo;
1507
5ba3f43e
A		 1508 lck_mtx_lock(&utun_lock);
1509
1510 /* Find some open interface id */
1511 u_int32_t chosen_unique_id = 1;
1512 struct utun_pcb *next_pcb = TAILQ_LAST(&utun_head, utun_list);
1513 if (next_pcb != NULL) {
1514 /* List was not empty, add one to the last item */
1515 chosen_unique_id = next_pcb->utun_unique_id + 1;
1516 next_pcb = NULL;
1517
1518 /*
1519 * If this wrapped the id number, start looking at
1520 * the front of the list for an unused id.
1521 */
1522 if (chosen_unique_id == 0) {
1523 /* Find the next unused ID */
1524 chosen_unique_id = 1;
1525 TAILQ_FOREACH(next_pcb, &utun_head, utun_chain) {
1526 if (next_pcb->utun_unique_id > chosen_unique_id) {
1527 /* We found a gap */
1528 break;
1529 }
1530
1531 chosen_unique_id = next_pcb->utun_unique_id + 1;
1532 }
1533 }
1534 }
1535
1536 pcb->utun_unique_id = chosen_unique_id;
1537
1538 if (next_pcb != NULL) {
1539 TAILQ_INSERT_BEFORE(next_pcb, pcb, utun_chain);
1540 } else {
1541 TAILQ_INSERT_TAIL(&utun_head, pcb, utun_chain);
1542 }
1543 lck_mtx_unlock(&utun_lock);
1544
1545 snprintf(pcb->utun_if_xname, sizeof(pcb->utun_if_xname), "utun%d", pcb->utun_unit - 1);
1546 snprintf(pcb->utun_unique_name, sizeof(pcb->utun_unique_name), "utunid%d", pcb->utun_unique_id - 1);
1547 printf("utun_ctl_connect: creating interface %s (id %s)\n", pcb->utun_if_xname, pcb->utun_unique_name);
b0d623f7
A		 1548
1549 /* Create the interface */
1550 bzero(&utun_init, sizeof(utun_init));
39236c6e
A		 1551 utun_init.ver = IFNET_INIT_CURRENT_VERSION;
1552 utun_init.len = sizeof (utun_init);
5ba3f43e
A		 1553
1554#if UTUN_NEXUS
5c9f4661
A		 1555 if (pcb->utun_use_netif) {
1556 utun_init.flags = (IFNET_INIT_SKYWALK_NATIVE | IFNET_INIT_NX_NOAUTO);
1557 utun_init.tx_headroom = UTUN_IF_HEADROOM_SIZE;
1558 } else
5ba3f43e 1559#endif // UTUN_NEXUS
5c9f4661
A		 1560 {
1561 utun_init.flags = IFNET_INIT_NX_NOAUTO;
1562 utun_init.start = utun_start;
1563 utun_init.framer_extended = utun_framer;
1564 }
5ba3f43e 1565 utun_init.name = "utun";
d1ecb069 1566 utun_init.unit = pcb->utun_unit - 1;
5ba3f43e
A		 1567 utun_init.uniqueid = pcb->utun_unique_name;
1568 utun_init.uniqueid_len = strlen(pcb->utun_unique_name);
b0d623f7 1569 utun_init.family = utun_family;
39037602 1570 utun_init.subfamily = IFNET_SUBFAMILY_UTUN;
b0d623f7 1571 utun_init.type = IFT_OTHER;
b0d623f7 1572 utun_init.demux = utun_demux;
b0d623f7
A		 1573 utun_init.add_proto = utun_add_proto;
1574 utun_init.del_proto = utun_del_proto;
1575 utun_init.softc = pcb;
1576 utun_init.ioctl = utun_ioctl;
1577 utun_init.detach = utun_detached;
39037602 1578
5ba3f43e 1579#if UTUN_NEXUS
5c9f4661
A		 1580 if (pcb->utun_use_netif) {
1581 result = utun_nexus_ifattach(pcb, &utun_init, &pcb->utun_ifp);
1582 if (result != 0) {
1583 printf("utun_ctl_connect - utun_nexus_ifattach failed: %d\n", result);
1584 utun_free_pcb(pcb, true);
1585 *unitinfo = NULL;
1586 return result;
1587 }
5ba3f43e 1588
5c9f4661
A		 1589 result = utun_multistack_attach(pcb);
1590 if (result != 0) {
1591 printf("utun_ctl_connect - utun_multistack_attach failed: %d\n", result);
1592 *unitinfo = NULL;
1593 return result;
1594 }
5ba3f43e 1595
5c9f4661
A		 1596 /* Attach to bpf */
1597 bpfattach(pcb->utun_ifp, DLT_RAW, 0);
1598 } else
1599#endif // UTUN_NEXUS
1600 {
1601 /*
1602 * Upon success, this holds an ifnet reference which we will
1603 * release via ifnet_release() at final detach time.
1604 */
1605 result = ifnet_allocate_extended(&utun_init, &pcb->utun_ifp);
1606 if (result != 0) {
1607 printf("utun_ctl_connect - ifnet_allocate failed: %d\n", result);
1608 utun_free_pcb(pcb, true);
1609 *unitinfo = NULL;
1610 return result;
1611 }
b0d623f7 1612
5c9f4661
A		 1613 /* Set flags and additional information. */
1614 ifnet_set_mtu(pcb->utun_ifp, UTUN_DEFAULT_MTU);
1615 ifnet_set_flags(pcb->utun_ifp, IFF_UP | IFF_MULTICAST | IFF_POINTOPOINT, 0xffff);
d1ecb069 1616
5c9f4661
A		 1617 /* The interface must generate its own IPv6 LinkLocal address,
1618 * if possible following the recommendation of RFC2472 to the 64bit interface ID
1619 */
1620 ifnet_set_eflags(pcb->utun_ifp, IFEF_NOAUTOIPV6LL, IFEF_NOAUTOIPV6LL);
1621
1622 /* Reset the stats in case as the interface may have been recycled */
1623 struct ifnet_stats_param stats;
1624 bzero(&stats, sizeof(struct ifnet_stats_param));
1625 ifnet_set_stat(pcb->utun_ifp, &stats);
1626
1627 /* Attach the interface */
1628 result = ifnet_attach(pcb->utun_ifp, NULL);
1629 if (result != 0) {
1630 printf("utun_ctl_connect - ifnet_attach failed: %d\n", result);
1631 /* Release reference now since attach failed */
1632 ifnet_release(pcb->utun_ifp);
1633 utun_free_pcb(pcb, true);
1634 *unitinfo = NULL;
1635 return (result);
1636 }
1637
1638 /* Attach to bpf */
1639 bpfattach(pcb->utun_ifp, DLT_NULL, UTUN_HEADER_SIZE(pcb));
39037602 1640 }
5ba3f43e 1641
5ba3f43e
A		 1642 /* The interfaces resoures allocated, mark it as running */
1643 ifnet_set_flags(pcb->utun_ifp, IFF_RUNNING, IFF_RUNNING);
1644
b0d623f7
A		 1645 return result;
1646}
1647
1648static errno_t
5ba3f43e
A		 1649utun_detach_ip(ifnet_t interface,
1650 protocol_family_t protocol,
1651 socket_t pf_socket)
b0d623f7
A		 1652{
1653 errno_t result = EPROTONOSUPPORT;
1654
1655 /* Attempt a detach */
1656 if (protocol == PF_INET) {
1657 struct ifreq ifr;
1658
1659 bzero(&ifr, sizeof(ifr));
1660 snprintf(ifr.ifr_name, sizeof(ifr.ifr_name), "%s%d",
1661 ifnet_name(interface), ifnet_unit(interface));
1662
1663 result = sock_ioctl(pf_socket, SIOCPROTODETACH, &ifr);
5ba3f43e 1664 } else if (protocol == PF_INET6) {
b0d623f7
A		 1665 struct in6_ifreq ifr6;
1666
1667 bzero(&ifr6, sizeof(ifr6));
1668 snprintf(ifr6.ifr_name, sizeof(ifr6.ifr_name), "%s%d",
1669 ifnet_name(interface), ifnet_unit(interface));
1670
1671 result = sock_ioctl(pf_socket, SIOCPROTODETACH_IN6, &ifr6);
1672 }
1673
1674 return result;
1675}
1676
1677static void
5ba3f43e
A		 1678utun_remove_address(ifnet_t interface,
1679 protocol_family_t protocol,
1680 ifaddr_t address,
1681 socket_t pf_socket)
b0d623f7
A		 1682{
1683 errno_t result = 0;
1684
1685 /* Attempt a detach */
1686 if (protocol == PF_INET) {
5ba3f43e 1687 struct ifreq ifr;
b0d623f7
A		 1688
1689 bzero(&ifr, sizeof(ifr));
1690 snprintf(ifr.ifr_name, sizeof(ifr.ifr_name), "%s%d",
1691 ifnet_name(interface), ifnet_unit(interface));
1692 result = ifaddr_address(address, &ifr.ifr_addr, sizeof(ifr.ifr_addr));
1693 if (result != 0) {
1694 printf("utun_remove_address - ifaddr_address failed: %d", result);
5ba3f43e 1695 } else {
b0d623f7
A		 1696 result = sock_ioctl(pf_socket, SIOCDIFADDR, &ifr);
1697 if (result != 0) {
1698 printf("utun_remove_address - SIOCDIFADDR failed: %d", result);
1699 }
1700 }
5ba3f43e
A		 1701 } else if (protocol == PF_INET6) {
1702 struct in6_ifreq ifr6;
b0d623f7
A		 1703
1704 bzero(&ifr6, sizeof(ifr6));
1705 snprintf(ifr6.ifr_name, sizeof(ifr6.ifr_name), "%s%d",
1706 ifnet_name(interface), ifnet_unit(interface));
1707 result = ifaddr_address(address, (struct sockaddr*)&ifr6.ifr_addr,
1708 sizeof(ifr6.ifr_addr));
1709 if (result != 0) {
1710 printf("utun_remove_address - ifaddr_address failed (v6): %d",
1711 result);
5ba3f43e 1712 } else {
b0d623f7
A		 1713 result = sock_ioctl(pf_socket, SIOCDIFADDR_IN6, &ifr6);
1714 if (result != 0) {
1715 printf("utun_remove_address - SIOCDIFADDR_IN6 failed: %d",
1716 result);
1717 }
1718 }
1719 }
1720}
1721
1722static void
5ba3f43e
A		 1723utun_cleanup_family(ifnet_t interface,
1724 protocol_family_t protocol)
1725{
1726 errno_t result = 0;
1727 socket_t pf_socket = NULL;
1728 ifaddr_t *addresses = NULL;
1729 int i;
b0d623f7
A		 1730
1731 if (protocol != PF_INET && protocol != PF_INET6) {
1732 printf("utun_cleanup_family - invalid protocol family %d\n", protocol);
1733 return;
1734 }
1735
1736 /* Create a socket for removing addresses and detaching the protocol */
1737 result = sock_socket(protocol, SOCK_DGRAM, 0, NULL, NULL, &pf_socket);
1738 if (result != 0) {
1739 if (result != EAFNOSUPPORT)
1740 printf("utun_cleanup_family - failed to create %s socket: %d\n",
1741 protocol == PF_INET ? "IP" : "IPv6", result);
1742 goto cleanup;
1743 }
1744
6d2010ae
A		 1745 /* always set SS_PRIV, we want to close and detach regardless */
1746 sock_setpriv(pf_socket, 1);
1747
b0d623f7
A		 1748 result = utun_detach_ip(interface, protocol, pf_socket);
1749 if (result == 0 || result == ENXIO) {
1750 /* We are done! We either detached or weren't attached. */
1751 goto cleanup;
5ba3f43e 1752 } else if (result != EBUSY) {
b0d623f7
A		 1753 /* Uh, not really sure what happened here... */
1754 printf("utun_cleanup_family - utun_detach_ip failed: %d\n", result);
1755 goto cleanup;
1756 }
1757
1758 /*
1759 * At this point, we received an EBUSY error. This means there are
1760 * addresses attached. We should detach them and then try again.
1761 */
1762 result = ifnet_get_address_list_family(interface, &addresses, protocol);
1763 if (result != 0) {
1764 printf("fnet_get_address_list_family(%s%d, 0xblah, %s) - failed: %d\n",
1765 ifnet_name(interface), ifnet_unit(interface),
1766 protocol == PF_INET ? "PF_INET" : "PF_INET6", result);
1767 goto cleanup;
1768 }
1769
1770 for (i = 0; addresses[i] != 0; i++) {
1771 utun_remove_address(interface, protocol, addresses[i], pf_socket);
1772 }
1773 ifnet_free_address_list(addresses);
1774 addresses = NULL;
1775
1776 /*
1777 * The addresses should be gone, we should try the remove again.
1778 */
1779 result = utun_detach_ip(interface, protocol, pf_socket);
1780 if (result != 0 && result != ENXIO) {
1781 printf("utun_cleanup_family - utun_detach_ip failed: %d\n", result);
1782 }
1783
1784cleanup:
5ba3f43e 1785 if (pf_socket != NULL) {
b0d623f7 1786 sock_close(pf_socket);
5ba3f43e 1787 }
b0d623f7 1788
5ba3f43e 1789 if (addresses != NULL) {
b0d623f7 1790 ifnet_free_address_list(addresses);
5ba3f43e 1791 }
b0d623f7
A		 1792}
1793
1794static errno_t
5ba3f43e
A		 1795utun_ctl_disconnect(__unused kern_ctl_ref kctlref,
1796 __unused u_int32_t unit,
1797 void *unitinfo)
b0d623f7
A		 1798{
1799 struct utun_pcb *pcb = unitinfo;
5ba3f43e
A		 1800 ifnet_t ifp = NULL;
1801 errno_t result = 0;
1802
1803 if (pcb == NULL) {
1804 return EINVAL;
1805 }
1806
1807#if UTUN_NEXUS
1808 // Tell the nexus to stop all rings
1809 if (pcb->utun_netif_nexus != NULL) {
1810 kern_nexus_stop(pcb->utun_netif_nexus);
1811 }
1812#endif // UTUN_NEXUS
316670eb 1813
5ba3f43e 1814 lck_rw_lock_exclusive(&pcb->utun_pcb_lock);
316670eb 1815
5ba3f43e
A		 1816#if UTUN_NEXUS
1817 uuid_t kpipe_uuid;
1818 uuid_copy(kpipe_uuid, pcb->utun_kpipe_uuid);
1819 uuid_clear(pcb->utun_kpipe_uuid);
1820 pcb->utun_kpipe_enabled = FALSE;
1821#endif // UTUN_NEXUS
39037602 1822
d1ecb069 1823 pcb->utun_ctlref = NULL;
5ba3f43e 1824
5c9f4661
A		 1825 ifp = pcb->utun_ifp;
1826 if (ifp != NULL) {
1827#if UTUN_NEXUS
1828 // Tell the nexus to stop all rings
1829 if (pcb->utun_netif_nexus != NULL) {
1830 /*
1831 * Quiesce the interface and flush any pending outbound packets.
1832 */
1833 if_down(ifp);
1834
1835 /* Increment refcnt, but detach interface */
1836 ifnet_incr_iorefcnt(ifp);
1837 if ((result = ifnet_detach(ifp)) != 0) {
1838 panic("utun_ctl_disconnect - ifnet_detach failed: %d\n", result);
1839 }
5ba3f43e 1840
5c9f4661
A		 1841 /*
1842 * We want to do everything in our power to ensure that the interface
1843 * really goes away when the socket is closed. We must remove IP/IPv6
1844 * addresses and detach the protocols. Finally, we can remove and
1845 * release the interface.
1846 */
1847 utun_cleanup_family(ifp, AF_INET);
1848 utun_cleanup_family(ifp, AF_INET6);
39037602 1849
5c9f4661 1850 lck_rw_unlock_exclusive(&pcb->utun_pcb_lock);
39037602 1851
5c9f4661
A		 1852 if (!uuid_is_null(kpipe_uuid)) {
1853 if (kern_nexus_controller_free_provider_instance(utun_ncd, kpipe_uuid) == 0) {
a39ff7e2
A		 1854 if (pcb->utun_kpipe_pp != NULL) {
1855 kern_pbufpool_destroy(pcb->utun_kpipe_pp);
1856 pcb->utun_kpipe_pp = NULL;
1857 }
5c9f4661
A		 1858 utun_unregister_kernel_pipe_nexus();
1859 }
1860 }
a39ff7e2 1861 utun_nexus_detach(pcb);
5c9f4661
A		 1862
1863 /* Decrement refcnt to finish detaching and freeing */
1864 ifnet_decr_iorefcnt(ifp);
1865 } else
1866#endif // UTUN_NEXUS
1867 {
1868 lck_rw_unlock_exclusive(&pcb->utun_pcb_lock);
5ba3f43e
A		 1869
1870#if UTUN_NEXUS
5c9f4661
A		 1871 if (!uuid_is_null(kpipe_uuid)) {
1872 if (kern_nexus_controller_free_provider_instance(utun_ncd, kpipe_uuid) == 0) {
a39ff7e2
A		 1873 if (pcb->utun_kpipe_pp != NULL) {
1874 kern_pbufpool_destroy(pcb->utun_kpipe_pp);
1875 pcb->utun_kpipe_pp = NULL;
1876 }
5c9f4661
A		 1877 utun_unregister_kernel_pipe_nexus();
1878 }
1879 }
5ba3f43e
A		 1880#endif // UTUN_NEXUS
1881
5c9f4661
A		 1882 /*
1883 * We want to do everything in our power to ensure that the interface
1884 * really goes away when the socket is closed. We must remove IP/IPv6
1885 * addresses and detach the protocols. Finally, we can remove and
1886 * release the interface.
1887 */
1888 utun_cleanup_family(ifp, AF_INET);
1889 utun_cleanup_family(ifp, AF_INET6);
1890
1891 /*
1892 * Detach now; utun_detach() will be called asynchronously once
1893 * the I/O reference count drops to 0. There we will invoke
1894 * ifnet_release().
1895 */
1896 if ((result = ifnet_detach(ifp)) != 0) {
1897 printf("utun_ctl_disconnect - ifnet_detach failed: %d\n", result);
1898 }
1899 }
1900 } else {
1901 // Bound, but not connected
1902 lck_rw_unlock_exclusive(&pcb->utun_pcb_lock);
1903 utun_free_pcb(pcb, false);
1904 }
b0d623f7 1905
b0d623f7 1906 return 0;
39037602 1907}
b0d623f7
A		 1908
1909static errno_t
5ba3f43e
A		 1910utun_ctl_send(__unused kern_ctl_ref kctlref,
1911 __unused u_int32_t unit,
1912 void *unitinfo,
1913 mbuf_t m,
1914 __unused int flags)
b0d623f7 1915{
39236c6e
A		 1916 /*
1917 * The userland ABI requires the first four bytes have the protocol family
1918 * in network byte order: swap them
1919 */
39037602 1920 if (m_pktlen(m) >= (int32_t)UTUN_HEADER_SIZE((struct utun_pcb *)unitinfo)) {
39236c6e 1921 *(protocol_family_t *)mbuf_data(m) = ntohl(*(protocol_family_t *)mbuf_data(m));
39037602 1922 } else {
39236c6e 1923 printf("%s - unexpected short mbuf pkt len %d\n", __func__, m_pktlen(m) );
39037602 1924 }
39236c6e 1925
316670eb 1926 return utun_pkt_input((struct utun_pcb *)unitinfo, m);
b0d623f7
A		 1927}
1928
1929static errno_t
5ba3f43e
A		 1930utun_ctl_setopt(__unused kern_ctl_ref kctlref,
1931 __unused u_int32_t unit,
1932 void *unitinfo,
1933 int opt,
1934 void *data,
1935 size_t len)
b0d623f7 1936{
5ba3f43e
A		 1937 struct utun_pcb *pcb = unitinfo;
1938 errno_t result = 0;
b0d623f7
A		 1939 /* check for privileges for privileged options */
1940 switch (opt) {
1941 case UTUN_OPT_FLAGS:
d1ecb069 1942 case UTUN_OPT_EXT_IFDATA_STATS:
39236c6e 1943 case UTUN_OPT_SET_DELEGATE_INTERFACE:
b0d623f7
A		 1944 if (kauth_cred_issuser(kauth_cred_get()) == 0) {
1945 return EPERM;
1946 }
1947 break;
1948 }
1949
1950 switch (opt) {
1951 case UTUN_OPT_FLAGS:
39037602 1952 if (len != sizeof(u_int32_t)) {
b0d623f7 1953 result = EMSGSIZE;
39037602 1954 } else {
5c9f4661
A		 1955 if (pcb->utun_ifp == NULL) {
1956 // Only can set after connecting
1957 result = EINVAL;
1958 break;
1959 }
1960#if UTUN_NEXUS
1961 if (pcb->utun_use_netif) {
1962 pcb->utun_flags = *(u_int32_t *)data;
1963 } else
1964#endif // UTUN_NEXUS
1965 {
1966 u_int32_t old_flags = pcb->utun_flags;
1967 pcb->utun_flags = *(u_int32_t *)data;
1968 if (((old_flags ^ pcb->utun_flags) & UTUN_FLAGS_ENABLE_PROC_UUID)) {
1969 // If UTUN_FLAGS_ENABLE_PROC_UUID flag changed, update bpf
1970 bpfdetach(pcb->utun_ifp);
1971 bpfattach(pcb->utun_ifp, DLT_NULL, UTUN_HEADER_SIZE(pcb));
1972 }
1973 }
39037602 1974 }
39236c6e
A		 1975 break;
1976
d1ecb069
A		 1977 case UTUN_OPT_EXT_IFDATA_STATS:
1978 if (len != sizeof(int)) {
1979 result = EMSGSIZE;
1980 break;
1981 }
5c9f4661
A		 1982 if (pcb->utun_ifp == NULL) {
1983 // Only can set after connecting
1984 result = EINVAL;
1985 break;
1986 }
d1ecb069
A		 1987 pcb->utun_ext_ifdata_stats = (*(int *)data) ? 1 : 0;
1988 break;
1989
1990 case UTUN_OPT_INC_IFDATA_STATS_IN:
1991 case UTUN_OPT_INC_IFDATA_STATS_OUT: {
1992 struct utun_stats_param *utsp = (struct utun_stats_param *)data;
1993
1994 if (utsp == NULL || len < sizeof(struct utun_stats_param)) {
1995 result = EINVAL;
1996 break;
1997 }
5c9f4661
A		 1998 if (pcb->utun_ifp == NULL) {
1999 // Only can set after connecting
2000 result = EINVAL;
2001 break;
2002 }
d1ecb069
A		 2003 if (!pcb->utun_ext_ifdata_stats) {
2004 result = EINVAL;
2005 break;
2006 }
2007 if (opt == UTUN_OPT_INC_IFDATA_STATS_IN)
2008 ifnet_stat_increment_in(pcb->utun_ifp, utsp->utsp_packets,
2009 utsp->utsp_bytes, utsp->utsp_errors);
2010 else
2011 ifnet_stat_increment_out(pcb->utun_ifp, utsp->utsp_packets,
2012 utsp->utsp_bytes, utsp->utsp_errors);
b0d623f7 2013 break;
d1ecb069 2014 }
fe8ab488 2015 case UTUN_OPT_SET_DELEGATE_INTERFACE: {
39236c6e
A		 2016 ifnet_t del_ifp = NULL;
2017 char name[IFNAMSIZ];
2018
2019 if (len > IFNAMSIZ - 1) {
2020 result = EMSGSIZE;
2021 break;
2022 }
5c9f4661
A		 2023 if (pcb->utun_ifp == NULL) {
2024 // Only can set after connecting
2025 result = EINVAL;
2026 break;
2027 }
39236c6e
A		 2028 if (len != 0) { /* if len==0, del_ifp will be NULL causing the delegate to be removed */
2029 bcopy(data, name, len);
2030 name[len] = 0;
2031 result = ifnet_find_by_name(name, &del_ifp);
2032 }
2033 if (result == 0) {
2034 result = ifnet_set_delegate(pcb->utun_ifp, del_ifp);
2035 if (del_ifp)
2036 ifnet_release(del_ifp);
2037 }
2038 break;
2039 }
fe8ab488
A		 2040 case UTUN_OPT_MAX_PENDING_PACKETS: {
2041 u_int32_t max_pending_packets = 0;
2042 if (len != sizeof(u_int32_t)) {
2043 result = EMSGSIZE;
2044 break;
2045 }
2046 max_pending_packets = *(u_int32_t *)data;
2047 if (max_pending_packets == 0) {
2048 result = EINVAL;
2049 break;
2050 }
2051 pcb->utun_max_pending_packets = max_pending_packets;
2052 break;
2053 }
5ba3f43e
A		 2054#if UTUN_NEXUS
2055 case UTUN_OPT_ENABLE_CHANNEL: {
2056 if (len != sizeof(int)) {
2057 result = EMSGSIZE;
2058 break;
2059 }
5c9f4661
A		 2060 if (pcb->utun_ifp == NULL) {
2061 // Only can set after connecting
2062 result = EINVAL;
2063 break;
2064 }
5ba3f43e
A		 2065 if (*(int *)data) {
2066 result = utun_enable_channel(pcb, current_proc());
2067 } else {
2068 result = utun_disable_channel(pcb);
2069 }
2070 break;
2071 }
2072 case UTUN_OPT_ENABLE_FLOWSWITCH: {
2073 if (len != sizeof(int)) {
2074 result = EMSGSIZE;
2075 break;
2076 }
5c9f4661
A		 2077 if (pcb->utun_ifp == NULL) {
2078 // Only can set after connecting
2079 result = EINVAL;
2080 break;
2081 }
a39ff7e2 2082 if (!if_is_netagent_enabled()) {
5ba3f43e
A		 2083 result = ENOTSUP;
2084 break;
2085 }
2086 if (uuid_is_null(pcb->utun_nx.ms_agent)) {
2087 result = ENOENT;
2088 break;
2089 }
2090
2091 if (*(int *)data) {
2092 if_add_netagent(pcb->utun_ifp, pcb->utun_nx.ms_agent);
9d749ea3 2093 pcb->utun_needs_netagent = true;
5ba3f43e 2094 } else {
9d749ea3 2095 pcb->utun_needs_netagent = false;
5ba3f43e
A		 2096 if_delete_netagent(pcb->utun_ifp, pcb->utun_nx.ms_agent);
2097 }
2098 break;
2099 }
5c9f4661
A		 2100 case UTUN_OPT_ENABLE_NETIF: {
2101 if (len != sizeof(int)) {
2102 result = EMSGSIZE;
2103 break;
2104 }
2105 if (pcb->utun_ifp != NULL) {
2106 // Only can set before connecting
2107 result = EINVAL;
2108 break;
2109 }
a39ff7e2
A		 2110 lck_rw_lock_exclusive(&pcb->utun_pcb_lock);
2111 pcb->utun_use_netif = !!(*(int *)data);
2112 lck_rw_unlock_exclusive(&pcb->utun_pcb_lock);
5c9f4661
A		 2113 break;
2114 }
2115 case UTUN_OPT_SLOT_SIZE: {
2116 if (len != sizeof(u_int32_t)) {
2117 result = EMSGSIZE;
2118 break;
2119 }
2120 if (pcb->utun_ifp != NULL) {
2121 // Only can set before connecting
2122 result = EINVAL;
2123 break;
2124 }
2125 u_int32_t slot_size = *(u_int32_t *)data;
2126 if (slot_size < UTUN_IF_MIN_SLOT_SIZE ||
2127 slot_size > UTUN_IF_MAX_SLOT_SIZE) {
2128 return (EINVAL);
2129 }
2130 pcb->utun_slot_size = slot_size;
2131 break;
2132 }
2133 case UTUN_OPT_NETIF_RING_SIZE: {
2134 if (len != sizeof(u_int32_t)) {
2135 result = EMSGSIZE;
2136 break;
2137 }
2138 if (pcb->utun_ifp != NULL) {
2139 // Only can set before connecting
2140 result = EINVAL;
2141 break;
2142 }
2143 u_int32_t ring_size = *(u_int32_t *)data;
2144 if (ring_size < UTUN_IF_MIN_RING_SIZE ||
2145 ring_size > UTUN_IF_MAX_RING_SIZE) {
2146 return (EINVAL);
2147 }
2148 pcb->utun_netif_ring_size = ring_size;
2149 break;
2150 }
2151 case UTUN_OPT_TX_FSW_RING_SIZE: {
2152 if (len != sizeof(u_int32_t)) {
2153 result = EMSGSIZE;
2154 break;
2155 }
2156 if (pcb->utun_ifp != NULL) {
2157 // Only can set before connecting
2158 result = EINVAL;
2159 break;
2160 }
2161 u_int32_t ring_size = *(u_int32_t *)data;
2162 if (ring_size < UTUN_IF_MIN_RING_SIZE ||
2163 ring_size > UTUN_IF_MAX_RING_SIZE) {
2164 return (EINVAL);
2165 }
2166 pcb->utun_tx_fsw_ring_size = ring_size;
2167 break;
2168 }
2169 case UTUN_OPT_RX_FSW_RING_SIZE: {
2170 if (len != sizeof(u_int32_t)) {
2171 result = EMSGSIZE;
2172 break;
2173 }
2174 if (pcb->utun_ifp != NULL) {
2175 // Only can set before connecting
2176 result = EINVAL;
2177 break;
2178 }
2179 u_int32_t ring_size = *(u_int32_t *)data;
2180 if (ring_size < UTUN_IF_MIN_RING_SIZE ||
2181 ring_size > UTUN_IF_MAX_RING_SIZE) {
2182 return (EINVAL);
2183 }
2184 pcb->utun_rx_fsw_ring_size = ring_size;
2185 break;
2186 }
5ba3f43e 2187#endif // UTUN_NEXUS
fe8ab488 2188 default: {
b0d623f7
A		 2189 result = ENOPROTOOPT;
2190 break;
fe8ab488 2191 }
b0d623f7
A		 2192 }
2193
2194 return result;
2195}
2196
2197static errno_t
5ba3f43e
A		 2198utun_ctl_getopt(__unused kern_ctl_ref kctlref,
2199 __unused u_int32_t unit,
2200 void *unitinfo,
2201 int opt,
2202 void *data,
2203 size_t *len)
b0d623f7 2204{
5ba3f43e
A		 2205 struct utun_pcb *pcb = unitinfo;
2206 errno_t result = 0;
b0d623f7
A		 2207
2208 switch (opt) {
2209 case UTUN_OPT_FLAGS:
5ba3f43e 2210 if (*len != sizeof(u_int32_t)) {
b0d623f7 2211 result = EMSGSIZE;
5ba3f43e 2212 } else {
d1ecb069 2213 *(u_int32_t *)data = pcb->utun_flags;
5ba3f43e 2214 }
b0d623f7 2215 break;
d1ecb069
A		 2216
2217 case UTUN_OPT_EXT_IFDATA_STATS:
5ba3f43e 2218 if (*len != sizeof(int)) {
d1ecb069 2219 result = EMSGSIZE;
5ba3f43e 2220 } else {
d1ecb069 2221 *(int *)data = (pcb->utun_ext_ifdata_stats) ? 1 : 0;
5ba3f43e 2222 }
d1ecb069
A		 2223 break;
2224
b0d623f7 2225 case UTUN_OPT_IFNAME:
5ba3f43e
A		 2226 if (*len < MIN(strlen(pcb->utun_if_xname) + 1, sizeof(pcb->utun_if_xname))) {
2227 result = EMSGSIZE;
2228 } else {
5c9f4661
A		 2229 if (pcb->utun_ifp == NULL) {
2230 // Only can get after connecting
2231 result = EINVAL;
2232 break;
2233 }
5ba3f43e
A		 2234 *len = snprintf(data, *len, "%s", pcb->utun_if_xname) + 1;
2235 }
b0d623f7 2236 break;
d1ecb069 2237
fe8ab488 2238 case UTUN_OPT_MAX_PENDING_PACKETS: {
5ba3f43e
A		 2239 if (*len != sizeof(u_int32_t)) {
2240 result = EMSGSIZE;
2241 } else {
2242 *((u_int32_t *)data) = pcb->utun_max_pending_packets;
2243 }
fe8ab488
A		 2244 break;
2245 }
39037602 2246
5ba3f43e 2247#if UTUN_NEXUS
a39ff7e2
A		 2248 case UTUN_OPT_ENABLE_CHANNEL: {
2249 if (*len != sizeof(int)) {
2250 result = EMSGSIZE;
2251 } else {
2252 lck_rw_lock_shared(&pcb->utun_pcb_lock);
2253 *(int *)data = pcb->utun_kpipe_enabled;
2254 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2255 }
2256 break;
2257 }
2258
2259 case UTUN_OPT_ENABLE_FLOWSWITCH: {
2260 if (*len != sizeof(int)) {
2261 result = EMSGSIZE;
2262 } else {
2263 *(int *)data = if_check_netagent(pcb->utun_ifp, pcb->utun_nx.ms_agent);
2264 }
2265 break;
2266 }
2267
2268 case UTUN_OPT_ENABLE_NETIF: {
2269 if (*len != sizeof(int)) {
2270 result = EMSGSIZE;
2271 } else {
2272 lck_rw_lock_shared(&pcb->utun_pcb_lock);
2273 *(int *)data = !!pcb->utun_use_netif;
2274 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2275 }
2276 break;
2277 }
2278
5c9f4661 2279 case UTUN_OPT_GET_CHANNEL_UUID: {
5ba3f43e
A		 2280 lck_rw_lock_shared(&pcb->utun_pcb_lock);
2281 if (uuid_is_null(pcb->utun_kpipe_uuid)) {
2282 result = ENXIO;
2283 } else if (*len != sizeof(uuid_t)) {
2284 result = EMSGSIZE;
2285 } else {
2286 uuid_copy(data, pcb->utun_kpipe_uuid);
2287 }
2288 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2289 break;
5c9f4661
A		 2290 }
2291 case UTUN_OPT_SLOT_SIZE: {
2292 if (*len != sizeof(u_int32_t)) {
2293 result = EMSGSIZE;
2294 } else {
2295 *(u_int32_t *)data = pcb->utun_slot_size;
2296 }
2297 break;
2298 }
2299 case UTUN_OPT_NETIF_RING_SIZE: {
2300 if (*len != sizeof(u_int32_t)) {
2301 result = EMSGSIZE;
2302 } else {
2303 *(u_int32_t *)data = pcb->utun_netif_ring_size;
2304 }
2305 break;
2306 }
2307 case UTUN_OPT_TX_FSW_RING_SIZE: {
2308 if (*len != sizeof(u_int32_t)) {
2309 result = EMSGSIZE;
2310 } else {
2311 *(u_int32_t *)data = pcb->utun_tx_fsw_ring_size;
2312 }
2313 break;
2314 }
2315 case UTUN_OPT_RX_FSW_RING_SIZE: {
2316 if (*len != sizeof(u_int32_t)) {
2317 result = EMSGSIZE;
2318 } else {
2319 *(u_int32_t *)data = pcb->utun_rx_fsw_ring_size;
2320 }
2321 break;
2322 }
5ba3f43e 2323#endif // UTUN_NEXUS
39037602 2324
b0d623f7
A		 2325 default:
2326 result = ENOPROTOOPT;
2327 break;
2328 }
2329
2330 return result;
2331}
2332
fe8ab488
A		 2333static void
2334utun_ctl_rcvd(kern_ctl_ref kctlref, u_int32_t unit, void *unitinfo, int flags)
2335{
3e170ce0 2336#pragma unused(flags)
fe8ab488
A		 2337 bool reenable_output = false;
2338 struct utun_pcb *pcb = unitinfo;
2339 if (pcb == NULL) {
2340 return;
2341 }
2342 ifnet_lock_exclusive(pcb->utun_ifp);
3e170ce0
A		 2343
2344 u_int32_t utun_packet_cnt;
2345 errno_t error_pc = ctl_getenqueuepacketcount(kctlref, unit, &utun_packet_cnt);
2346 if (error_pc != 0) {
2347 printf("utun_ctl_rcvd: ctl_getenqueuepacketcount returned error %d\n", error_pc);
2348 utun_packet_cnt = 0;
fe8ab488 2349 }
3e170ce0
A		 2350
2351 if (utun_packet_cnt < pcb->utun_max_pending_packets) {
2352 reenable_output = true;
2353 }
2354
fe8ab488
A		 2355 if (reenable_output) {
2356 errno_t error = ifnet_enable_output(pcb->utun_ifp);
2357 if (error != 0) {
2358 printf("utun_ctl_rcvd: ifnet_enable_output returned error %d\n", error);
2359 }
2360 }
2361 ifnet_lock_done(pcb->utun_ifp);
2362}
2363
b0d623f7 2364/* Network Interface functions */
fe8ab488
A		 2365static void
2366utun_start(ifnet_t interface)
2367{
2368 mbuf_t data;
39037602
A		 2369 struct utun_pcb *pcb = ifnet_softc(interface);
2370
2371 VERIFY(pcb != NULL);
2372
5ba3f43e
A		 2373#if UTUN_NEXUS
2374 lck_rw_lock_shared(&pcb->utun_pcb_lock);
2375 if (pcb->utun_kpipe_enabled) {
2376 /* It's possible to have channels enabled, but not yet have the channel opened,
2377 * in which case the rxring will not be set
2378 */
2379 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2380 if (pcb->utun_kpipe_rxring != NULL) {
2381 kern_channel_notify(pcb->utun_kpipe_rxring, 0);
2382 }
2383 return;
2384 }
2385 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2386#endif // UTUN_NEXUS
39037602 2387
fe8ab488
A		 2388 for (;;) {
2389 bool can_accept_packets = true;
2390 ifnet_lock_shared(pcb->utun_ifp);
3e170ce0
A		 2391
2392 u_int32_t utun_packet_cnt;
2393 errno_t error_pc = ctl_getenqueuepacketcount(pcb->utun_ctlref, pcb->utun_unit, &utun_packet_cnt);
2394 if (error_pc != 0) {
2395 printf("utun_start: ctl_getenqueuepacketcount returned error %d\n", error_pc);
2396 utun_packet_cnt = 0;
2397 }
2398
2399 can_accept_packets = (utun_packet_cnt < pcb->utun_max_pending_packets);
fe8ab488
A		 2400 if (!can_accept_packets && pcb->utun_ctlref) {
2401 u_int32_t difference = 0;
2402 if (ctl_getenqueuereadable(pcb->utun_ctlref, pcb->utun_unit, &difference) == 0) {
2403 if (difference > 0) {
2404 // If the low-water mark has not yet been reached, we still need to enqueue data
2405 // into the buffer
2406 can_accept_packets = true;
2407 }
2408 }
2409 }
2410 if (!can_accept_packets) {
2411 errno_t error = ifnet_disable_output(interface);
2412 if (error != 0) {
2413 printf("utun_start: ifnet_disable_output returned error %d\n", error);
2414 }
2415 ifnet_lock_done(pcb->utun_ifp);
2416 break;
2417 }
2418 ifnet_lock_done(pcb->utun_ifp);
5ba3f43e 2419 if (ifnet_dequeue(interface, &data) != 0) {
fe8ab488 2420 break;
5ba3f43e
A		 2421 }
2422 if (utun_output(interface, data) != 0) {
fe8ab488 2423 break;
5ba3f43e 2424 }
fe8ab488
A		 2425 }
2426}
2427
b0d623f7 2428static errno_t
39037602
A		 2429utun_output(ifnet_t interface,
2430 mbuf_t data)
b0d623f7
A		 2431{
2432 struct utun_pcb *pcb = ifnet_softc(interface);
5ba3f43e 2433 errno_t result;
39037602
A		 2434
2435 VERIFY(interface == pcb->utun_ifp);
5ba3f43e 2436
5c9f4661
A		 2437#if UTUN_NEXUS
2438 if (!pcb->utun_use_netif)
2439#endif // UTUN_NEXUS
2440 {
2441 if (m_pktlen(data) >= (int32_t)UTUN_HEADER_SIZE(pcb)) {
2442 bpf_tap_out(pcb->utun_ifp, DLT_NULL, data, 0, 0);
2443 }
2444 }
2445
d1ecb069 2446 if (pcb->utun_flags & UTUN_FLAGS_NO_OUTPUT) {
b0d623f7
A		 2447 /* flush data */
2448 mbuf_freem(data);
2449 return 0;
2450 }
2451
316670eb 2452 // otherwise, fall thru to ctl_enqueumbuf
d1ecb069 2453 if (pcb->utun_ctlref) {
316670eb
A		 2454 int length;
2455
39236c6e
A		 2456 /*
2457 * The ABI requires the protocol in network byte order
2458 */
39037602 2459 if (m_pktlen(data) >= (int32_t)UTUN_HEADER_SIZE(pcb)) {
39236c6e 2460 *(u_int32_t *)mbuf_data(data) = htonl(*(u_int32_t *)mbuf_data(data));
39037602 2461 }
39236c6e 2462
316670eb 2463 length = mbuf_pkthdr_len(data);
d1ecb069 2464 result = ctl_enqueuembuf(pcb->utun_ctlref, pcb->utun_unit, data, CTL_DATA_EOR);
b0d623f7
A		 2465 if (result != 0) {
2466 mbuf_freem(data);
2467 printf("utun_output - ctl_enqueuembuf failed: %d\n", result);
5c9f4661
A		 2468#if UTUN_NEXUS
2469 if (!pcb->utun_use_netif)
2470#endif // UTUN_NEXUS
2471 {
2472 ifnet_stat_increment_out(interface, 0, 0, 1);
2473 }
5ba3f43e 2474 } else {
5c9f4661
A		 2475#if UTUN_NEXUS
2476 if (!pcb->utun_use_netif)
2477#endif // UTUN_NEXUS
2478 {
2479 if (!pcb->utun_ext_ifdata_stats) {
2480 ifnet_stat_increment_out(interface, 1, length, 0);
2481 }
5ba3f43e 2482 }
b0d623f7 2483 }
5ba3f43e 2484 } else {
b0d623f7 2485 mbuf_freem(data);
5ba3f43e 2486 }
b0d623f7
A		 2487
2488 return 0;
2489}
2490
b0d623f7 2491static errno_t
5ba3f43e
A		 2492utun_demux(__unused ifnet_t interface,
2493 mbuf_t data,
2494 __unused char *frame_header,
2495 protocol_family_t *protocol)
b0d623f7 2496{
d9a64523
A		 2497#if UTUN_NEXUS
2498 struct utun_pcb *pcb = ifnet_softc(interface);
2499 struct ip *ip;
2500 u_int ip_version;
2501#endif
2502
b0d623f7
A		 2503 while (data != NULL && mbuf_len(data) < 1) {
2504 data = mbuf_next(data);
2505 }
5ba3f43e 2506
5c9f4661 2507 if (data == NULL) {
b0d623f7 2508 return ENOENT;
5c9f4661 2509 }
5ba3f43e 2510
5c9f4661
A		 2511#if UTUN_NEXUS
2512 if (pcb->utun_use_netif) {
2513 ip = mtod(data, struct ip *);
2514 ip_version = ip->ip_v;
2515
2516 switch(ip_version) {
2517 case 4:
2518 *protocol = PF_INET;
2519 return 0;
2520 case 6:
2521 *protocol = PF_INET6;
2522 return 0;
2523 default:
2524 *protocol = 0;
2525 break;
2526 }
2527 } else
2528#endif // UTUN_NEXUS
2529 {
2530 *protocol = *(u_int32_t *)mbuf_data(data);
5ba3f43e
A		 2531 }
2532
b0d623f7
A		 2533 return 0;
2534}
2535
2536static errno_t
5ba3f43e
A		 2537utun_framer(ifnet_t interface,
2538 mbuf_t *packet,
2539 __unused const struct sockaddr *dest,
b0d623f7 2540 __unused const char *desk_linkaddr,
39236c6e 2541 const char *frame_type,
5ba3f43e 2542 u_int32_t *prepend_len,
39236c6e 2543 u_int32_t *postpend_len)
b0d623f7 2544{
39037602
A		 2545 struct utun_pcb *pcb = ifnet_softc(interface);
2546 VERIFY(interface == pcb->utun_ifp);
2547
2548 u_int32_t header_length = UTUN_HEADER_SIZE(pcb);
d9a64523 2549 if (mbuf_prepend(packet, header_length, MBUF_DONTWAIT) != 0) {
b0d623f7 2550 printf("utun_framer - ifnet_output prepend failed\n");
d1ecb069 2551
b0d623f7 2552 ifnet_stat_increment_out(interface, 0, 0, 1);
d1ecb069 2553
b0d623f7 2554 // just return, because the buffer was freed in mbuf_prepend
d9a64523
A		 2555 return EJUSTRETURN;
2556 }
5ba3f43e 2557 if (prepend_len != NULL) {
39037602 2558 *prepend_len = header_length;
5ba3f43e
A		 2559 }
2560 if (postpend_len != NULL) {
39236c6e 2561 *postpend_len = 0;
5ba3f43e 2562 }
b0d623f7 2563
d9a64523
A		 2564 // place protocol number at the beginning of the mbuf
2565 *(protocol_family_t *)mbuf_data(*packet) = *(protocol_family_t *)(uintptr_t)(size_t)frame_type;
39037602
A		 2566
2567
b0d623f7
A		 2568 return 0;
2569}
2570
2571static errno_t
5ba3f43e
A		 2572utun_add_proto(__unused ifnet_t interface,
2573 protocol_family_t protocol,
2574 __unused const struct ifnet_demux_desc *demux_array,
2575 __unused u_int32_t demux_count)
b0d623f7
A		 2576{
2577 switch(protocol) {
2578 case PF_INET:
2579 return 0;
2580 case PF_INET6:
2581 return 0;
2582 default:
2583 break;
2584 }
2585
2586 return ENOPROTOOPT;
2587}
2588
2589static errno_t
5ba3f43e
A		 2590utun_del_proto(__unused ifnet_t interface,
2591 __unused protocol_family_t protocol)
b0d623f7
A		 2592{
2593 return 0;
2594}
2595
2596static errno_t
5ba3f43e
A		 2597utun_ioctl(ifnet_t interface,
2598 u_long command,
2599 void *data)
b0d623f7 2600{
d9a64523
A		 2601#if UTUN_NEXUS
2602 struct utun_pcb *pcb = ifnet_softc(interface);
2603#endif
b0d623f7 2604 errno_t result = 0;
d9a64523 2605
b0d623f7 2606 switch(command) {
5c9f4661 2607 case SIOCSIFMTU: {
5ba3f43e 2608#if UTUN_NEXUS
5c9f4661
A		 2609 if (pcb->utun_use_netif) {
2610 // Make sure we can fit packets in the channel buffers
2611 // Allow for the headroom in the slot
2612 if (((uint64_t)((struct ifreq*)data)->ifr_mtu) + UTUN_IF_HEADROOM_SIZE > pcb->utun_slot_size) {
2613 result = EINVAL;
2614 } else {
2615 ifnet_set_mtu(interface, (uint32_t)((struct ifreq*)data)->ifr_mtu);
2616 }
2617 } else
5ba3f43e 2618#endif // UTUN_NEXUS
5c9f4661
A		 2619 {
2620 ifnet_set_mtu(interface, ((struct ifreq*)data)->ifr_mtu);
2621 }
b0d623f7 2622 break;
5c9f4661 2623 }
d1ecb069
A		 2624
2625 case SIOCSIFFLAGS:
2626 /* ifioctl() takes care of it */
2627 break;
2628
b0d623f7
A		 2629 default:
2630 result = EOPNOTSUPP;
2631 }
2632
2633 return result;
2634}
2635
2636static void
5ba3f43e 2637utun_detached(ifnet_t interface)
b0d623f7
A		 2638{
2639 struct utun_pcb *pcb = ifnet_softc(interface);
5ba3f43e 2640 (void)ifnet_release(interface);
5c9f4661 2641 utun_free_pcb(pcb, true);
b0d623f7
A		 2642}
2643
2644/* Protocol Handlers */
2645
2646static errno_t
5ba3f43e
A		 2647utun_proto_input(__unused ifnet_t interface,
2648 protocol_family_t protocol,
2649 mbuf_t m,
2650 __unused char *frame_header)
b0d623f7 2651{
5c9f4661
A		 2652 struct utun_pcb *pcb = ifnet_softc(interface);
2653#if UTUN_NEXUS
2654 if (!pcb->utun_use_netif)
2655#endif // UTUN_NEXUS
2656 {
2657 mbuf_adj(m, UTUN_HEADER_SIZE(pcb));
2658 }
a39ff7e2 2659 int32_t pktlen = m->m_pkthdr.len;
39037602 2660 if (proto_input(protocol, m) != 0) {
6d2010ae 2661 m_freem(m);
5c9f4661
A		 2662#if UTUN_NEXUS
2663 if (!pcb->utun_use_netif)
2664#endif // UTUN_NEXUS
2665 {
2666 ifnet_stat_increment_in(interface, 0, 0, 1);
2667 }
5ba3f43e 2668 } else {
5c9f4661
A		 2669#if UTUN_NEXUS
2670 if (!pcb->utun_use_netif)
5ba3f43e 2671#endif // UTUN_NEXUS
5c9f4661 2672 {
a39ff7e2 2673 ifnet_stat_increment_in(interface, 1, pktlen, 0);
5c9f4661 2674 }
39037602 2675 }
b0d623f7
A		 2676
2677 return 0;
2678}
2679
2680static errno_t
5ba3f43e
A		 2681utun_proto_pre_output(__unused ifnet_t interface,
2682 protocol_family_t protocol,
2683 __unused mbuf_t *packet,
2684 __unused const struct sockaddr *dest,
2685 __unused void *route,
2686 char *frame_type,
2687 __unused char *link_layer_dest)
b0d623f7 2688{
6d2010ae 2689 *(protocol_family_t *)(void *)frame_type = protocol;
39037602 2690 return 0;
b0d623f7
A		 2691}
2692
2693static errno_t
5ba3f43e
A		 2694utun_attach_proto(ifnet_t interface,
2695 protocol_family_t protocol)
b0d623f7
A		 2696{
2697 struct ifnet_attach_proto_param proto;
b0d623f7
A		 2698
2699 bzero(&proto, sizeof(proto));
2700 proto.input = utun_proto_input;
2701 proto.pre_output = utun_proto_pre_output;
2702
5ba3f43e 2703 errno_t result = ifnet_attach_protocol(interface, protocol, &proto);
b0d623f7
A		 2704 if (result != 0 && result != EEXIST) {
2705 printf("utun_attach_inet - ifnet_attach_protocol %d failed: %d\n",
2706 protocol, result);
2707 }
2708
2709 return result;
2710}
2711
5ba3f43e
A		 2712static errno_t
2713utun_pkt_input(struct utun_pcb *pcb, mbuf_t packet)
2714{
5c9f4661
A		 2715#if UTUN_NEXUS
2716 if (pcb->utun_use_netif) {
2717 lck_rw_lock_shared(&pcb->utun_pcb_lock);
5ba3f43e 2718
5c9f4661
A		 2719 lck_mtx_lock(&pcb->utun_input_chain_lock);
2720 if (pcb->utun_input_chain != NULL) {
2721 pcb->utun_input_chain_last->m_nextpkt = packet;
2722 } else {
2723 pcb->utun_input_chain = packet;
2724 }
2725 while (packet->m_nextpkt) {
2726 VERIFY(packet != packet->m_nextpkt);
2727 packet = packet->m_nextpkt;
2728 }
2729 pcb->utun_input_chain_last = packet;
2730 lck_mtx_unlock(&pcb->utun_input_chain_lock);
5ba3f43e 2731
5c9f4661
A		 2732 kern_channel_ring_t rx_ring = pcb->utun_netif_rxring;
2733 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
316670eb 2734
5c9f4661
A		 2735 if (rx_ring != NULL) {
2736 kern_channel_notify(rx_ring, 0);
2737 }
316670eb 2738
5c9f4661
A		 2739 return (0);
2740 } else
2741#endif // IPSEC_NEXUS
2742 {
2743 mbuf_pkthdr_setrcvif(packet, pcb->utun_ifp);
5ba3f43e 2744
5c9f4661
A		 2745 if (m_pktlen(packet) >= (int32_t)UTUN_HEADER_SIZE(pcb)) {
2746 bpf_tap_in(pcb->utun_ifp, DLT_NULL, packet, 0, 0);
2747 }
2748 if (pcb->utun_flags & UTUN_FLAGS_NO_INPUT) {
2749 /* flush data */
2750 mbuf_freem(packet);
2751 return 0;
2752 }
316670eb 2753
5c9f4661
A		 2754 errno_t result = 0;
2755 if (!pcb->utun_ext_ifdata_stats) {
2756 struct ifnet_stat_increment_param incs = {};
2757 incs.packets_in = 1;
2758 incs.bytes_in = mbuf_pkthdr_len(packet);
2759 result = ifnet_input(pcb->utun_ifp, packet, &incs);
2760 } else {
2761 result = ifnet_input(pcb->utun_ifp, packet, NULL);
2762 }
2763 if (result != 0) {
2764 ifnet_stat_increment_in(pcb->utun_ifp, 0, 0, 1);
5ba3f43e 2765
5c9f4661
A		 2766 printf("%s - ifnet_input failed: %d\n", __FUNCTION__, result);
2767 mbuf_freem(packet);
2768 }
5ba3f43e 2769
5c9f4661 2770 return (0);
316670eb 2771 }
316670eb 2772}
5ba3f43e
A		 2773
2774#if UTUN_NEXUS
2775
2776static errno_t
2777utun_nxdp_init(__unused kern_nexus_domain_provider_t domprov)
2778{
2779 return 0;
2780}
2781
2782static void
2783utun_nxdp_fini(__unused kern_nexus_domain_provider_t domprov)
2784{
2785 // Ignore
2786}
2787
2788static errno_t
2789utun_register_nexus(void)
2790{
2791 const struct kern_nexus_domain_provider_init dp_init = {
2792 .nxdpi_version = KERN_NEXUS_DOMAIN_PROVIDER_CURRENT_VERSION,
2793 .nxdpi_flags = 0,
2794 .nxdpi_init = utun_nxdp_init,
2795 .nxdpi_fini = utun_nxdp_fini
2796 };
2797 errno_t err = 0;
2798
2799 /* utun_nxdp_init() is called before this function returns */
2800 err = kern_nexus_register_domain_provider(NEXUS_TYPE_NET_IF,
2801 (const uint8_t *) "com.apple.utun",
2802 &dp_init, sizeof(dp_init),
2803 &utun_nx_dom_prov);
2804 if (err != 0) {
2805 printf("%s: failed to register domain provider\n", __func__);
2806 return (err);
2807 }
2808 return (0);
2809}
9d749ea3
A		 2810boolean_t
2811utun_interface_needs_netagent(ifnet_t interface)
2812{
2813 struct utun_pcb *pcb = NULL;
2814
2815 if (interface == NULL) {
2816 return (FALSE);
2817 }
2818
2819 pcb = ifnet_softc(interface);
2820
2821 if (pcb == NULL) {
2822 return (FALSE);
2823 }
2824
2825 return (pcb->utun_needs_netagent == true);
2826}
5ba3f43e
A		 2827
2828static errno_t
2829utun_ifnet_set_attrs(ifnet_t ifp)
2830{
2831 /* Set flags and additional information. */
2832 ifnet_set_mtu(ifp, 1500);
2833 ifnet_set_flags(ifp, IFF_UP | IFF_MULTICAST | IFF_POINTOPOINT, 0xffff);
2834
2835 /* The interface must generate its own IPv6 LinkLocal address,
2836 * if possible following the recommendation of RFC2472 to the 64bit interface ID
2837 */
2838 ifnet_set_eflags(ifp, IFEF_NOAUTOIPV6LL, IFEF_NOAUTOIPV6LL);
2839
2840 return (0);
2841}
2842
2843static errno_t
2844utun_netif_prepare(kern_nexus_t nexus, ifnet_t ifp)
2845{
2846 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
2847 pcb->utun_netif_nexus = nexus;
2848 return (utun_ifnet_set_attrs(ifp));
2849}
2850
2851static errno_t
2852utun_nexus_pre_connect(kern_nexus_provider_t nxprov,
2853 proc_t p, kern_nexus_t nexus,
2854 nexus_port_t nexus_port, kern_channel_t channel, void **ch_ctx)
2855{
2856#pragma unused(nxprov, p)
2857#pragma unused(nexus, nexus_port, channel, ch_ctx)
2858 return (0);
2859}
2860
2861static errno_t
2862utun_nexus_connected(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2863 kern_channel_t channel)
2864{
2865#pragma unused(nxprov, channel)
2866 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
2867 boolean_t ok = ifnet_is_attached(pcb->utun_ifp, 1);
2868 return (ok ? 0 : ENXIO);
2869}
2870
2871static void
2872utun_nexus_pre_disconnect(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2873 kern_channel_t channel)
2874{
2875#pragma unused(nxprov, nexus, channel)
2876}
2877
2878static void
2879utun_netif_pre_disconnect(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2880 kern_channel_t channel)
2881{
2882#pragma unused(nxprov, nexus, channel)
2883}
2884
2885static void
2886utun_nexus_disconnected(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2887 kern_channel_t channel)
2888{
2889#pragma unused(nxprov, channel)
2890 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
2891 if (pcb->utun_netif_nexus == nexus) {
2892 pcb->utun_netif_nexus = NULL;
2893 }
2894 ifnet_decr_iorefcnt(pcb->utun_ifp);
2895}
2896
2897static errno_t
2898utun_kpipe_ring_init(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2899 kern_channel_t channel, kern_channel_ring_t ring,
2900 boolean_t is_tx_ring, void **ring_ctx)
2901{
2902#pragma unused(nxprov)
2903#pragma unused(channel)
2904#pragma unused(ring_ctx)
2905 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
2906 if (!is_tx_ring) {
2907 VERIFY(pcb->utun_kpipe_rxring == NULL);
2908 pcb->utun_kpipe_rxring = ring;
2909 } else {
2910 VERIFY(pcb->utun_kpipe_txring == NULL);
2911 pcb->utun_kpipe_txring = ring;
2912 }
2913 return 0;
2914}
2915
2916static void
2917utun_kpipe_ring_fini(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2918 kern_channel_ring_t ring)
2919{
2920#pragma unused(nxprov)
2921 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
2922 if (pcb->utun_kpipe_rxring == ring) {
2923 pcb->utun_kpipe_rxring = NULL;
2924 } else if (pcb->utun_kpipe_txring == ring) {
2925 pcb->utun_kpipe_txring = NULL;
2926 }
2927}
2928
2929static errno_t
2930utun_kpipe_sync_tx(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
2931 kern_channel_ring_t tx_ring, uint32_t flags)
2932{
2933#pragma unused(nxprov)
2934#pragma unused(flags)
2935 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
2936
2937 lck_rw_lock_shared(&pcb->utun_pcb_lock);
2938 int channel_enabled = pcb->utun_kpipe_enabled;
2939 if (!channel_enabled) {
2940 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2941 return 0;
2942 }
2943
5c9f4661
A		 2944 if (pcb->utun_use_netif) {
2945 kern_channel_slot_t tx_slot = kern_channel_get_next_slot(tx_ring, NULL, NULL);
2946 if (tx_slot == NULL) {
2947 // Nothing to write, bail
2948 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2949 return 0;
2950 }
2951
2952 // Signal the netif ring to read
2953 kern_channel_ring_t rx_ring = pcb->utun_netif_rxring;
2954 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
2955 if (rx_ring != NULL) {
2956 kern_channel_notify(rx_ring, 0);
2957 }
2958 } else {
5ba3f43e 2959 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
5ba3f43e 2960
5c9f4661
A		 2961 struct ifnet_stat_increment_param incs = {};
2962 struct kern_channel_ring_stat_increment tx_ring_stats = {};
2963 MBUFQ_HEAD(mbufq) mbq;
2964 MBUFQ_INIT(&mbq);
2965 kern_channel_slot_t tx_pslot = NULL;
2966 kern_channel_slot_t tx_slot = kern_channel_get_next_slot(tx_ring, NULL, NULL);
2967 while (tx_slot != NULL) {
2968 kern_packet_t tx_ph = kern_channel_slot_get_packet(tx_ring, tx_slot);
2969
2970 // Advance TX ring
2971 tx_pslot = tx_slot;
2972 tx_slot = kern_channel_get_next_slot(tx_ring, tx_slot, NULL);
2973
2974 if (tx_ph == 0) {
2975 continue;
2976 }
2977
2978 kern_buflet_t tx_buf = kern_packet_get_next_buflet(tx_ph, NULL);
2979 VERIFY(tx_buf != NULL);
2980 uint8_t *tx_baddr = kern_buflet_get_object_address(tx_buf);
2981 VERIFY(tx_baddr != 0);
2982 tx_baddr += kern_buflet_get_data_offset(tx_buf);
2983
2984 size_t length = MIN(kern_packet_get_data_length(tx_ph),
2985 pcb->utun_slot_size);
2986
2987 mbuf_t data = NULL;
2988 if (length >= UTUN_HEADER_SIZE(pcb) &&
2989 !(pcb->utun_flags & UTUN_FLAGS_NO_INPUT)) {
2990 errno_t error = mbuf_gethdr(MBUF_WAITOK, MBUF_TYPE_HEADER, &data);
2991 VERIFY(0 == error);
2992 error = mbuf_copyback(data, 0, length, tx_baddr, MBUF_WAITOK);
2993 VERIFY(0 == error);
2994 /*
2995 * The userland ABI requires the first four bytes have
2996 * the protocol family in network byte order: swap them
2997 */
2998 *(uint32_t *)mbuf_data(data) = ntohl(*(uint32_t *)mbuf_data(data));
2999 mbuf_pkthdr_setrcvif(data, pcb->utun_ifp);
3000 bpf_tap_in(pcb->utun_ifp, DLT_NULL, data, 0, 0);
3001 incs.packets_in++;
3002 incs.bytes_in += length;
3003 MBUFQ_ENQUEUE(&mbq, data);
3004 }
3005 }
3006 if (tx_pslot) {
3007 kern_channel_advance_slot(tx_ring, tx_pslot);
3008 tx_ring_stats.kcrsi_slots_transferred = incs.packets_in;
3009 tx_ring_stats.kcrsi_bytes_transferred = incs.bytes_in;
3010 kern_channel_increment_ring_net_stats(tx_ring, pcb->utun_ifp, &tx_ring_stats);
3011 (void) kern_channel_reclaim(tx_ring);
3012 }
3013 if (!MBUFQ_EMPTY(&mbq)) {
3014 (void) ifnet_input_extended(pcb->utun_ifp, MBUFQ_FIRST(&mbq),
3015 MBUFQ_LAST(&mbq), &incs);
3016 MBUFQ_INIT(&mbq);
3017 }
5ba3f43e
A		 3018 }
3019
3020 return 0;
3021}
3022
3023static errno_t
3024utun_kpipe_sync_rx(kern_nexus_provider_t nxprov, kern_nexus_t nexus,
3025 kern_channel_ring_t rx_ring, uint32_t flags)
3026{
3027#pragma unused(nxprov)
3028#pragma unused(flags)
3029 struct utun_pcb *pcb = kern_nexus_get_context(nexus);
5c9f4661 3030 struct kern_channel_ring_stat_increment rx_ring_stats = {};
5ba3f43e
A		 3031
3032 lck_rw_lock_shared(&pcb->utun_pcb_lock);
3033
3034 int channel_enabled = pcb->utun_kpipe_enabled;
3035 if (!channel_enabled) {
3036 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
3037 return 0;
3038 }
3039
3040 /* reclaim user-released slots */
3041 (void) kern_channel_reclaim(rx_ring);
3042
3043 uint32_t avail = kern_channel_available_slot_count(rx_ring);
3044 if (avail == 0) {
3045 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
3046 return 0;
3047 }
3048
5c9f4661
A		 3049 if (pcb->utun_use_netif) {
3050 kern_channel_ring_t tx_ring = pcb->utun_netif_txring;
3051 if (tx_ring == NULL ||
3052 pcb->utun_netif_nexus == NULL) {
3053 // Net-If TX ring not set up yet, nothing to read
3054 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
3055 return 0;
3056 }
5ba3f43e 3057
5c9f4661 3058 struct netif_stats *nifs = &NX_NETIF_PRIVATE(pcb->utun_netif_nexus)->nif_stats;
5ba3f43e 3059
5c9f4661 3060 // Unlock utun before entering ring
5ba3f43e 3061 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
5ba3f43e 3062
5c9f4661 3063 (void)kr_enter(tx_ring, TRUE);
5ba3f43e 3064
5c9f4661
A		 3065 // Lock again after entering and validate
3066 lck_rw_lock_shared(&pcb->utun_pcb_lock);
3067 if (tx_ring != pcb->utun_netif_txring) {
3068 // Ring no longer valid
3069 // Unlock first, then exit ring
3070 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
3071 kr_exit(tx_ring);
3072 return 0;
5ba3f43e
A		 3073 }
3074
5c9f4661
A		 3075 struct kern_channel_ring_stat_increment tx_ring_stats;
3076 bzero(&tx_ring_stats, sizeof(tx_ring_stats));
3077 kern_channel_slot_t tx_pslot = NULL;
3078 kern_channel_slot_t tx_slot = kern_channel_get_next_slot(tx_ring, NULL, NULL);
3079 if (tx_slot == NULL) {
3080 // Nothing to read, don't bother signalling
3081 // Unlock first, then exit ring
3082 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
3083 kr_exit(tx_ring);
3084 return 0;
5ba3f43e
A		 3085 }
3086
5c9f4661
A		 3087 struct kern_pbufpool *rx_pp = rx_ring->ckr_pp;
3088 VERIFY(rx_pp != NULL);
3089 kern_channel_slot_t rx_pslot = NULL;
3090 kern_channel_slot_t rx_slot = kern_channel_get_next_slot(rx_ring, NULL, NULL);
5ba3f43e 3091
5c9f4661
A		 3092 while (rx_slot != NULL && tx_slot != NULL) {
3093 size_t length;
3094 kern_buflet_t rx_buf;
3095 void *rx_baddr;
5ba3f43e 3096
5c9f4661 3097 kern_packet_t tx_ph = kern_channel_slot_get_packet(tx_ring, tx_slot);
5ba3f43e 3098
5c9f4661
A		 3099 // Advance TX ring
3100 tx_pslot = tx_slot;
3101 tx_slot = kern_channel_get_next_slot(tx_ring, tx_slot, NULL);
5ba3f43e 3102
5c9f4661
A		 3103 /* Skip slot if packet is zero-length or marked as dropped (QUMF_DROPPED) */
3104 if (tx_ph == 0) {
3105 continue;
3106 }
5ba3f43e 3107
5c9f4661
A		 3108 // Allocate rx packet
3109 kern_packet_t rx_ph = 0;
3110 errno_t error = kern_pbufpool_alloc_nosleep(rx_pp, 1, &rx_ph);
a39ff7e2 3111 if (__improbable(error != 0)) {
5c9f4661
A		 3112 printf("utun_kpipe_sync_rx %s: failed to allocate packet\n",
3113 pcb->utun_ifp->if_xname);
5ba3f43e
A		 3114 break;
3115 }
5c9f4661
A		 3116
3117 kern_buflet_t tx_buf = kern_packet_get_next_buflet(tx_ph, NULL);
3118 VERIFY(tx_buf != NULL);
3119 uint8_t *tx_baddr = kern_buflet_get_object_address(tx_buf);
3120 VERIFY(tx_baddr != NULL);
3121 tx_baddr += kern_buflet_get_data_offset(tx_buf);
3122
3123 bpf_tap_packet_out(pcb->utun_ifp, DLT_RAW, tx_ph, NULL, 0);
3124
3125 length = MIN(kern_packet_get_data_length(tx_ph) + UTUN_HEADER_SIZE(pcb),
3126 pcb->utun_slot_size);
3127
3128 tx_ring_stats.kcrsi_slots_transferred++;
3129 tx_ring_stats.kcrsi_bytes_transferred += length;
3130
3131 if (length < UTUN_HEADER_SIZE(pcb) ||
3132 length > pcb->utun_slot_size ||
3133 length > rx_pp->pp_buflet_size ||
3134 (pcb->utun_flags & UTUN_FLAGS_NO_OUTPUT)) {
3135 /* flush data */
3136 kern_pbufpool_free(rx_pp, rx_ph);
3137 printf("utun_kpipe_sync_rx %s: invalid length %zu header_size %zu\n",
3138 pcb->utun_ifp->if_xname, length, UTUN_HEADER_SIZE(pcb));
3139 STATS_INC(nifs, NETIF_STATS_BADLEN);
3140 STATS_INC(nifs, NETIF_STATS_DROPPED);
3141 continue;
5ba3f43e 3142 }
5c9f4661
A		 3143
3144 /* fillout packet */
3145 rx_buf = kern_packet_get_next_buflet(rx_ph, NULL);
3146 VERIFY(rx_buf != NULL);
3147 rx_baddr = kern_buflet_get_object_address(rx_buf);
3148 VERIFY(rx_baddr != NULL);
3149
3150 // Find family
3151 uint32_t af = 0;
3152 uint8_t vhl = *(uint8_t *)(tx_baddr);
3153 u_int ip_version = (vhl >> 4);
3154 switch (ip_version) {
3155 case 4: {
3156 af = AF_INET;
3157 break;
3158 }
3159 case 6: {
3160 af = AF_INET6;
3161 break;
3162 }
3163 default: {
3164 printf("utun_kpipe_sync_rx %s: unknown ip version %u vhl %u header_size %zu\n",
3165 pcb->utun_ifp->if_xname, ip_version, vhl, UTUN_HEADER_SIZE(pcb));
3166 break;
3167 }
5ba3f43e 3168 }
5ba3f43e 3169
5c9f4661
A		 3170 // Copy header
3171 af = htonl(af);
3172 memcpy((void *)rx_baddr, &af, sizeof(af));
3173 if (pcb->utun_flags & UTUN_FLAGS_ENABLE_PROC_UUID) {
3174 kern_packet_get_euuid(tx_ph, (void *)(rx_baddr + sizeof(af)));
3175 }
5ba3f43e 3176
5c9f4661
A		 3177 // Copy data from tx to rx
3178 memcpy((void *)(rx_baddr + UTUN_HEADER_SIZE(pcb)), (void *)tx_baddr, length - UTUN_HEADER_SIZE(pcb));
3179 kern_packet_clear_flow_uuid(rx_ph); // zero flow id
5ba3f43e 3180
5c9f4661
A		 3181 /* finalize and attach the packet */
3182 error = kern_buflet_set_data_offset(rx_buf, 0);
3183 VERIFY(error == 0);
3184 error = kern_buflet_set_data_length(rx_buf, length);
3185 VERIFY(error == 0);
3186 error = kern_packet_finalize(rx_ph);
3187 VERIFY(error == 0);
3188 error = kern_channel_slot_attach_packet(rx_ring, rx_slot, rx_ph);
3189 VERIFY(error == 0);
5ba3f43e 3190
5c9f4661
A		 3191 STATS_INC(nifs, NETIF_STATS_TXPKTS);
3192 STATS_INC(nifs, NETIF_STATS_TXCOPY_DIRECT);
5ba3f43e 3193
5c9f4661
A		 3194 rx_ring_stats.kcrsi_slots_transferred++;
3195 rx_ring_stats.kcrsi_bytes_transferred += length;
5ba3f43e 3196
5c9f4661
A		 3197 rx_pslot = rx_slot;
3198 rx_slot = kern_channel_get_next_slot(rx_ring, rx_slot, NULL);
3199 }
5ba3f43e 3200
5c9f4661
A		 3201 if (rx_pslot) {
3202 kern_channel_advance_slot(rx_ring, rx_pslot);
3203 kern_channel_increment_ring_net_stats(rx_ring, pcb->utun_ifp, &rx_ring_stats);
3204 }
5ba3f43e 3205
5c9f4661
A		 3206 if (tx_pslot) {
3207 kern_channel_advance_slot(tx_ring, tx_pslot);
3208 kern_channel_increment_ring_net_stats(tx_ring, pcb->utun_ifp, &tx_ring_stats);
3209 (void)kern_channel_reclaim(tx_ring);
3210 }
5ba3f43e 3211
5c9f4661 3212 /* just like utun_ctl_rcvd(), always reenable output */
5ba3f43e
A		 3213 errno_t error = ifnet_enable_output(pcb->utun_ifp);
3214 if (error != 0) {
3215 printf("utun_kpipe_sync_rx: ifnet_enable_output returned error %d\n", error);
5ba3f43e 3216 }
5ba3f43e 3217
5c9f4661
A		 3218 // Unlock first, then exit ring
3219 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
5ba3f43e 3220
5c9f4661
A		 3221 if (tx_pslot != NULL) {
3222 kern_channel_notify(tx_ring, 0);
3223 }
3224 kr_exit(tx_ring);
3225 } else {
3226 lck_rw_unlock_shared(&pcb->utun_pcb_lock);
3227
3228 uint32_t mb_cnt = 0;
3229 uint32_t mb_len = 0;
3230 struct mbuf *mb_head = NULL;
3231 struct mbuf *mb_tail = NULL;
3232
3233 if (ifnet_dequeue_multi(pcb->utun_ifp, avail, &mb_head,
3234 &mb_tail, &mb_cnt, &mb_len) != 0) {
3235 return 0;
3236 }
3237 VERIFY(mb_cnt <= avail);
3238
3239 struct kern_pbufpool *rx_pp = rx_ring->ckr_pp;
3240 VERIFY(rx_pp != NULL);
3241 kern_channel_slot_t rx_pslot = NULL;
3242 kern_channel_slot_t rx_slot = kern_channel_get_next_slot(rx_ring, NULL, NULL);
3243 while (rx_slot) {
3244 size_t length = 0;
3245 mbuf_t data = NULL;
3246 if ((data = mb_head) == NULL) {
3247 VERIFY(mb_cnt == 0);
3248 break;
3249 }
3250 mb_head = mbuf_nextpkt(mb_head);
3251 mbuf_setnextpkt(data, NULL);
3252 VERIFY(mb_cnt != 0);
3253 --mb_cnt;
3254 length = mbuf_pkthdr_len(data);
3255 if (length < UTUN_HEADER_SIZE(pcb) ||
3256 length > pcb->utun_slot_size ||
3257 (pcb->utun_flags & UTUN_FLAGS_NO_OUTPUT)) {
3258 /* flush data */
3259 mbuf_freem(data);
3260 continue;
3261 }
3262 bpf_tap_out(pcb->utun_ifp, DLT_NULL, data, 0, 0);
3263
3264 // Allocate rx packet
3265 kern_packet_t rx_ph = 0;
3266 errno_t error = kern_pbufpool_alloc_nosleep(rx_pp, 1, &rx_ph);
a39ff7e2 3267 if (__improbable(error != 0)) {
5c9f4661
A		 3268 printf("utun_kpipe_sync_rx %s: failed to allocate packet\n",
3269 pcb->utun_ifp->if_xname);
3270 break;
3271 }
3272
3273 /*
3274 * The ABI requires the protocol in network byte order
3275 */
3276 *(u_int32_t *)mbuf_data(data) = htonl(*(u_int32_t *)mbuf_data(data));
3277
3278 // Fillout rx packet
3279 kern_buflet_t rx_buf = kern_packet_get_next_buflet(rx_ph, NULL);
3280 VERIFY(rx_buf != NULL);
3281 void *rx_baddr = kern_buflet_get_object_address(rx_buf);
3282 VERIFY(rx_baddr != NULL);
3283
3284 // Copy-in data from mbuf to buflet
3285 mbuf_copydata(data, 0, length, (void *)rx_baddr);
3286 kern_packet_clear_flow_uuid(rx_ph); // Zero flow id
3287
3288 // Finalize and attach the packet
3289 error = kern_buflet_set_data_offset(rx_buf, 0);
3290 VERIFY(error == 0);
3291 error = kern_buflet_set_data_length(rx_buf, length);
3292 VERIFY(error == 0);
3293 error = kern_packet_finalize(rx_ph);
3294 VERIFY(error == 0);
3295 error = kern_channel_slot_attach_packet(rx_ring, rx_slot, rx_ph);
3296 VERIFY(error == 0);
3297
3298 rx_ring_stats.kcrsi_slots_transferred++;
3299 rx_ring_stats.kcrsi_bytes_transferred += length;
3300
3301 if (!pcb->utun_ext_ifdata_stats) {
3302 ifnet_stat_increment_out(pcb->utun_ifp, 1, length, 0);
3303 }
3304
3305 mbuf_freem(data);
3306
3307 rx_pslot = rx_slot;
3308 rx_slot = kern_channel_get_next_slot(rx_ring, rx_slot, NULL);
3309 }
3310 if (rx_pslot) {
3311 kern_channel_advance_slot(rx_ring, rx_pslot);
3312 kern_channel_increment_ring_stats(rx_ring, &rx_ring_stats);
3313 }
3314 if (mb_head != NULL) {
3315 VERIFY(mb_cnt != 0);
3316 mbuf_freem_list(mb_head);
3317 }
5ba3f43e 3318 }
5ba3f43e
A		 3319
3320 return 0;
3321}
39037602 3322
5ba3f43e 3323#endif // UTUN_NEXUS
39037602
A		 3324
3325
3326/*
5ba3f43e 3327 * These are place holders until coreTLS kext stops calling them
39037602
A		 3328 */
3329errno_t utun_ctl_register_dtls (void *reg);
3330int utun_pkt_dtls_input(struct utun_pcb *pcb, mbuf_t *pkt, protocol_family_t family);
3331void utun_ctl_disable_crypto_dtls(struct utun_pcb *pcb);
3332
3333errno_t
3334utun_ctl_register_dtls (void *reg)
3335{
3336#pragma unused(reg)
3337 return 0;
3338}
3339
3340int
3341utun_pkt_dtls_input(struct utun_pcb *pcb, mbuf_t *pkt, protocol_family_t family)
3342{
3343#pragma unused(pcb)
3344#pragma unused(pkt)
3345#pragma unused(family)
3346 return 0;
3347}
3348
3349void
3350utun_ctl_disable_crypto_dtls(struct utun_pcb *pcb)
3351{
3352#pragma unused(pcb)
3353}
mirror of XNU