xref: /freebsd/sys/kern/uipc_ktls.c (revision 2ed3236082a4473c1da8f72c1ebc071a7b54321f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_rss.h"
34 
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/ktls.h>
38 #include <sys/lock.h>
39 #include <sys/mbuf.h>
40 #include <sys/mutex.h>
41 #include <sys/rmlock.h>
42 #include <sys/proc.h>
43 #include <sys/protosw.h>
44 #include <sys/refcount.h>
45 #include <sys/smp.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/taskqueue.h>
50 #include <sys/kthread.h>
51 #include <sys/uio.h>
52 #include <sys/vmmeter.h>
53 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
54 #include <machine/pcb.h>
55 #endif
56 #include <machine/vmparam.h>
57 #include <net/if.h>
58 #include <net/if_var.h>
59 #ifdef RSS
60 #include <net/netisr.h>
61 #include <net/rss_config.h>
62 #endif
63 #include <net/route.h>
64 #include <net/route/nhop.h>
65 #if defined(INET) || defined(INET6)
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #endif
69 #include <netinet/tcp_var.h>
70 #ifdef TCP_OFFLOAD
71 #include <netinet/tcp_offload.h>
72 #endif
73 #include <opencrypto/xform.h>
74 #include <vm/uma_dbg.h>
75 #include <vm/vm.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_page.h>
78 
79 struct ktls_wq {
80 	struct mtx	mtx;
81 	STAILQ_HEAD(, mbuf) m_head;
82 	STAILQ_HEAD(, socket) so_head;
83 	bool		running;
84 } __aligned(CACHE_LINE_SIZE);
85 
86 static struct ktls_wq *ktls_wq;
87 static struct proc *ktls_proc;
88 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
89 static struct rmlock ktls_backends_lock;
90 static uma_zone_t ktls_session_zone;
91 static uint16_t ktls_cpuid_lookup[MAXCPU];
92 
93 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
94     "Kernel TLS offload");
95 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
96     "Kernel TLS offload stats");
97 
98 static int ktls_allow_unload;
99 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
100     &ktls_allow_unload, 0, "Allow software crypto modules to unload");
101 
102 #ifdef RSS
103 static int ktls_bind_threads = 1;
104 #else
105 static int ktls_bind_threads;
106 #endif
107 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
108     &ktls_bind_threads, 0,
109     "Bind crypto threads to cores or domains at boot");
110 
111 static u_int ktls_maxlen = 16384;
112 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
113     &ktls_maxlen, 0, "Maximum TLS record size");
114 
115 static int ktls_number_threads;
116 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
117     &ktls_number_threads, 0,
118     "Number of TLS threads in thread-pool");
119 
120 static bool ktls_offload_enable;
121 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW,
122     &ktls_offload_enable, 0,
123     "Enable support for kernel TLS offload");
124 
125 static bool ktls_cbc_enable = true;
126 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW,
127     &ktls_cbc_enable, 1,
128     "Enable Support of AES-CBC crypto for kernel TLS");
129 
130 static counter_u64_t ktls_tasks_active;
131 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
132     &ktls_tasks_active, "Number of active tasks");
133 
134 static counter_u64_t ktls_cnt_tx_queued;
135 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
136     &ktls_cnt_tx_queued,
137     "Number of TLS records in queue to tasks for SW encryption");
138 
139 static counter_u64_t ktls_cnt_rx_queued;
140 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
141     &ktls_cnt_rx_queued,
142     "Number of TLS sockets in queue to tasks for SW decryption");
143 
144 static counter_u64_t ktls_offload_total;
145 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
146     CTLFLAG_RD, &ktls_offload_total,
147     "Total successful TLS setups (parameters set)");
148 
149 static counter_u64_t ktls_offload_enable_calls;
150 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
151     CTLFLAG_RD, &ktls_offload_enable_calls,
152     "Total number of TLS enable calls made");
153 
154 static counter_u64_t ktls_offload_active;
155 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
156     &ktls_offload_active, "Total Active TLS sessions");
157 
158 static counter_u64_t ktls_offload_corrupted_records;
159 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
160     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
161 
162 static counter_u64_t ktls_offload_failed_crypto;
163 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
164     &ktls_offload_failed_crypto, "Total TLS crypto failures");
165 
166 static counter_u64_t ktls_switch_to_ifnet;
167 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
168     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
169 
170 static counter_u64_t ktls_switch_to_sw;
171 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
172     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
173 
174 static counter_u64_t ktls_switch_failed;
175 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
176     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
177 
178 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
179     "Software TLS session stats");
180 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
181     "Hardware (ifnet) TLS session stats");
182 #ifdef TCP_OFFLOAD
183 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
184     "TOE TLS session stats");
185 #endif
186 
187 static counter_u64_t ktls_sw_cbc;
188 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
189     "Active number of software TLS sessions using AES-CBC");
190 
191 static counter_u64_t ktls_sw_gcm;
192 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
193     "Active number of software TLS sessions using AES-GCM");
194 
195 static counter_u64_t ktls_ifnet_cbc;
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
197     &ktls_ifnet_cbc,
198     "Active number of ifnet TLS sessions using AES-CBC");
199 
200 static counter_u64_t ktls_ifnet_gcm;
201 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
202     &ktls_ifnet_gcm,
203     "Active number of ifnet TLS sessions using AES-GCM");
204 
205 static counter_u64_t ktls_ifnet_reset;
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
207     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
208 
209 static counter_u64_t ktls_ifnet_reset_dropped;
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
211     &ktls_ifnet_reset_dropped,
212     "TLS sessions dropped after failing to update ifnet send tag");
213 
214 static counter_u64_t ktls_ifnet_reset_failed;
215 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
216     &ktls_ifnet_reset_failed,
217     "TLS sessions that failed to allocate a new ifnet send tag");
218 
219 static int ktls_ifnet_permitted;
220 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
221     &ktls_ifnet_permitted, 1,
222     "Whether to permit hardware (ifnet) TLS sessions");
223 
224 #ifdef TCP_OFFLOAD
225 static counter_u64_t ktls_toe_cbc;
226 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
227     &ktls_toe_cbc,
228     "Active number of TOE TLS sessions using AES-CBC");
229 
230 static counter_u64_t ktls_toe_gcm;
231 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
232     &ktls_toe_gcm,
233     "Active number of TOE TLS sessions using AES-GCM");
234 #endif
235 
236 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
237 
238 static void ktls_cleanup(struct ktls_session *tls);
239 #if defined(INET) || defined(INET6)
240 static void ktls_reset_send_tag(void *context, int pending);
241 #endif
242 static void ktls_work_thread(void *ctx);
243 
244 int
245 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
246 {
247 	struct ktls_crypto_backend *curr_be, *tmp;
248 
249 	if (be->api_version != KTLS_API_VERSION) {
250 		printf("KTLS: API version mismatch (%d vs %d) for %s\n",
251 		    be->api_version, KTLS_API_VERSION,
252 		    be->name);
253 		return (EINVAL);
254 	}
255 
256 	rm_wlock(&ktls_backends_lock);
257 	printf("KTLS: Registering crypto method %s with prio %d\n",
258 	       be->name, be->prio);
259 	if (LIST_EMPTY(&ktls_backends)) {
260 		LIST_INSERT_HEAD(&ktls_backends, be, next);
261 	} else {
262 		LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
263 			if (curr_be->prio < be->prio) {
264 				LIST_INSERT_BEFORE(curr_be, be, next);
265 				break;
266 			}
267 			if (LIST_NEXT(curr_be, next) == NULL) {
268 				LIST_INSERT_AFTER(curr_be, be, next);
269 				break;
270 			}
271 		}
272 	}
273 	rm_wunlock(&ktls_backends_lock);
274 	return (0);
275 }
276 
277 int
278 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
279 {
280 	struct ktls_crypto_backend *tmp;
281 
282 	/*
283 	 * Don't error if the backend isn't registered.  This permits
284 	 * MOD_UNLOAD handlers to use this function unconditionally.
285 	 */
286 	rm_wlock(&ktls_backends_lock);
287 	LIST_FOREACH(tmp, &ktls_backends, next) {
288 		if (tmp == be)
289 			break;
290 	}
291 	if (tmp == NULL) {
292 		rm_wunlock(&ktls_backends_lock);
293 		return (0);
294 	}
295 
296 	if (!ktls_allow_unload) {
297 		rm_wunlock(&ktls_backends_lock);
298 		printf(
299 		    "KTLS: Deregistering crypto method %s is not supported\n",
300 		    be->name);
301 		return (EBUSY);
302 	}
303 
304 	if (be->use_count) {
305 		rm_wunlock(&ktls_backends_lock);
306 		return (EBUSY);
307 	}
308 
309 	LIST_REMOVE(be, next);
310 	rm_wunlock(&ktls_backends_lock);
311 	return (0);
312 }
313 
314 #if defined(INET) || defined(INET6)
315 static u_int
316 ktls_get_cpu(struct socket *so)
317 {
318 	struct inpcb *inp;
319 	u_int cpuid;
320 
321 	inp = sotoinpcb(so);
322 #ifdef RSS
323 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
324 	if (cpuid != NETISR_CPUID_NONE)
325 		return (cpuid);
326 #endif
327 	/*
328 	 * Just use the flowid to shard connections in a repeatable
329 	 * fashion.  Note that some crypto backends rely on the
330 	 * serialization provided by having the same connection use
331 	 * the same queue.
332 	 */
333 	cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
334 	return (cpuid);
335 }
336 #endif
337 
338 static void
339 ktls_init(void *dummy __unused)
340 {
341 	struct thread *td;
342 	struct pcpu *pc;
343 	cpuset_t mask;
344 	int error, i;
345 
346 	ktls_tasks_active = counter_u64_alloc(M_WAITOK);
347 	ktls_cnt_tx_queued = counter_u64_alloc(M_WAITOK);
348 	ktls_cnt_rx_queued = counter_u64_alloc(M_WAITOK);
349 	ktls_offload_total = counter_u64_alloc(M_WAITOK);
350 	ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
351 	ktls_offload_active = counter_u64_alloc(M_WAITOK);
352 	ktls_offload_corrupted_records = counter_u64_alloc(M_WAITOK);
353 	ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
354 	ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
355 	ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
356 	ktls_switch_failed = counter_u64_alloc(M_WAITOK);
357 	ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
358 	ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
359 	ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
360 	ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
361 	ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
362 	ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
363 	ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
364 #ifdef TCP_OFFLOAD
365 	ktls_toe_cbc = counter_u64_alloc(M_WAITOK);
366 	ktls_toe_gcm = counter_u64_alloc(M_WAITOK);
367 #endif
368 
369 	rm_init(&ktls_backends_lock, "ktls backends");
370 	LIST_INIT(&ktls_backends);
371 
372 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
373 	    M_WAITOK | M_ZERO);
374 
375 	ktls_session_zone = uma_zcreate("ktls_session",
376 	    sizeof(struct ktls_session),
377 	    NULL, NULL, NULL, NULL,
378 	    UMA_ALIGN_CACHE, 0);
379 
380 	/*
381 	 * Initialize the workqueues to run the TLS work.  We create a
382 	 * work queue for each CPU.
383 	 */
384 	CPU_FOREACH(i) {
385 		STAILQ_INIT(&ktls_wq[i].m_head);
386 		STAILQ_INIT(&ktls_wq[i].so_head);
387 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
388 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
389 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
390 		if (error)
391 			panic("Can't add KTLS thread %d error %d", i, error);
392 
393 		/*
394 		 * Bind threads to cores.  If ktls_bind_threads is >
395 		 * 1, then we bind to the NUMA domain.
396 		 */
397 		if (ktls_bind_threads) {
398 			if (ktls_bind_threads > 1) {
399 				pc = pcpu_find(i);
400 				CPU_COPY(&cpuset_domain[pc->pc_domain], &mask);
401 			} else {
402 				CPU_SETOF(i, &mask);
403 			}
404 			error = cpuset_setthread(td->td_tid, &mask);
405 			if (error)
406 				panic(
407 			    "Unable to bind KTLS thread for CPU %d error %d",
408 				     i, error);
409 		}
410 		ktls_cpuid_lookup[ktls_number_threads] = i;
411 		ktls_number_threads++;
412 	}
413 	printf("KTLS: Initialized %d threads\n", ktls_number_threads);
414 }
415 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
416 
417 #if defined(INET) || defined(INET6)
418 static int
419 ktls_create_session(struct socket *so, struct tls_enable *en,
420     struct ktls_session **tlsp)
421 {
422 	struct ktls_session *tls;
423 	int error;
424 
425 	/* Only TLS 1.0 - 1.3 are supported. */
426 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
427 		return (EINVAL);
428 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
429 	    en->tls_vminor > TLS_MINOR_VER_THREE)
430 		return (EINVAL);
431 
432 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
433 		return (EINVAL);
434 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
435 		return (EINVAL);
436 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
437 		return (EINVAL);
438 
439 	/* All supported algorithms require a cipher key. */
440 	if (en->cipher_key_len == 0)
441 		return (EINVAL);
442 
443 	/* No flags are currently supported. */
444 	if (en->flags != 0)
445 		return (EINVAL);
446 
447 	/* Common checks for supported algorithms. */
448 	switch (en->cipher_algorithm) {
449 	case CRYPTO_AES_NIST_GCM_16:
450 		/*
451 		 * auth_algorithm isn't used, but permit GMAC values
452 		 * for compatibility.
453 		 */
454 		switch (en->auth_algorithm) {
455 		case 0:
456 #ifdef COMPAT_FREEBSD12
457 		/* XXX: Really 13.0-current COMPAT. */
458 		case CRYPTO_AES_128_NIST_GMAC:
459 		case CRYPTO_AES_192_NIST_GMAC:
460 		case CRYPTO_AES_256_NIST_GMAC:
461 #endif
462 			break;
463 		default:
464 			return (EINVAL);
465 		}
466 		if (en->auth_key_len != 0)
467 			return (EINVAL);
468 		if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
469 			en->iv_len != TLS_AEAD_GCM_LEN) ||
470 		    (en->tls_vminor == TLS_MINOR_VER_THREE &&
471 			en->iv_len != TLS_1_3_GCM_IV_LEN))
472 			return (EINVAL);
473 		break;
474 	case CRYPTO_AES_CBC:
475 		switch (en->auth_algorithm) {
476 		case CRYPTO_SHA1_HMAC:
477 			/*
478 			 * TLS 1.0 requires an implicit IV.  TLS 1.1+
479 			 * all use explicit IVs.
480 			 */
481 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
482 				if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
483 					return (EINVAL);
484 				break;
485 			}
486 
487 			/* FALLTHROUGH */
488 		case CRYPTO_SHA2_256_HMAC:
489 		case CRYPTO_SHA2_384_HMAC:
490 			/* Ignore any supplied IV. */
491 			en->iv_len = 0;
492 			break;
493 		default:
494 			return (EINVAL);
495 		}
496 		if (en->auth_key_len == 0)
497 			return (EINVAL);
498 		break;
499 	default:
500 		return (EINVAL);
501 	}
502 
503 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
504 
505 	counter_u64_add(ktls_offload_active, 1);
506 
507 	refcount_init(&tls->refcount, 1);
508 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
509 
510 	tls->wq_index = ktls_get_cpu(so);
511 
512 	tls->params.cipher_algorithm = en->cipher_algorithm;
513 	tls->params.auth_algorithm = en->auth_algorithm;
514 	tls->params.tls_vmajor = en->tls_vmajor;
515 	tls->params.tls_vminor = en->tls_vminor;
516 	tls->params.flags = en->flags;
517 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
518 
519 	/* Set the header and trailer lengths. */
520 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
521 	switch (en->cipher_algorithm) {
522 	case CRYPTO_AES_NIST_GCM_16:
523 		/*
524 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
525 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
526 		 */
527 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
528 			tls->params.tls_hlen += sizeof(uint64_t);
529 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
530 
531 		/*
532 		 * TLS 1.3 includes optional padding which we
533 		 * do not support, and also puts the "real" record
534 		 * type at the end of the encrypted data.
535 		 */
536 		if (en->tls_vminor == TLS_MINOR_VER_THREE)
537 			tls->params.tls_tlen += sizeof(uint8_t);
538 
539 		tls->params.tls_bs = 1;
540 		break;
541 	case CRYPTO_AES_CBC:
542 		switch (en->auth_algorithm) {
543 		case CRYPTO_SHA1_HMAC:
544 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
545 				/* Implicit IV, no nonce. */
546 			} else {
547 				tls->params.tls_hlen += AES_BLOCK_LEN;
548 			}
549 			tls->params.tls_tlen = AES_BLOCK_LEN +
550 			    SHA1_HASH_LEN;
551 			break;
552 		case CRYPTO_SHA2_256_HMAC:
553 			tls->params.tls_hlen += AES_BLOCK_LEN;
554 			tls->params.tls_tlen = AES_BLOCK_LEN +
555 			    SHA2_256_HASH_LEN;
556 			break;
557 		case CRYPTO_SHA2_384_HMAC:
558 			tls->params.tls_hlen += AES_BLOCK_LEN;
559 			tls->params.tls_tlen = AES_BLOCK_LEN +
560 			    SHA2_384_HASH_LEN;
561 			break;
562 		default:
563 			panic("invalid hmac");
564 		}
565 		tls->params.tls_bs = AES_BLOCK_LEN;
566 		break;
567 	default:
568 		panic("invalid cipher");
569 	}
570 
571 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
572 	    ("TLS header length too long: %d", tls->params.tls_hlen));
573 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
574 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
575 
576 	if (en->auth_key_len != 0) {
577 		tls->params.auth_key_len = en->auth_key_len;
578 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
579 		    M_WAITOK);
580 		error = copyin(en->auth_key, tls->params.auth_key,
581 		    en->auth_key_len);
582 		if (error)
583 			goto out;
584 	}
585 
586 	tls->params.cipher_key_len = en->cipher_key_len;
587 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
588 	error = copyin(en->cipher_key, tls->params.cipher_key,
589 	    en->cipher_key_len);
590 	if (error)
591 		goto out;
592 
593 	/*
594 	 * This holds the implicit portion of the nonce for GCM and
595 	 * the initial implicit IV for TLS 1.0.  The explicit portions
596 	 * of the IV are generated in ktls_frame().
597 	 */
598 	if (en->iv_len != 0) {
599 		tls->params.iv_len = en->iv_len;
600 		error = copyin(en->iv, tls->params.iv, en->iv_len);
601 		if (error)
602 			goto out;
603 
604 		/*
605 		 * For TLS 1.2, generate an 8-byte nonce as a counter
606 		 * to generate unique explicit IVs.
607 		 *
608 		 * Store this counter in the last 8 bytes of the IV
609 		 * array so that it is 8-byte aligned.
610 		 */
611 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
612 		    en->tls_vminor == TLS_MINOR_VER_TWO)
613 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
614 	}
615 
616 	*tlsp = tls;
617 	return (0);
618 
619 out:
620 	ktls_cleanup(tls);
621 	return (error);
622 }
623 
624 static struct ktls_session *
625 ktls_clone_session(struct ktls_session *tls)
626 {
627 	struct ktls_session *tls_new;
628 
629 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
630 
631 	counter_u64_add(ktls_offload_active, 1);
632 
633 	refcount_init(&tls_new->refcount, 1);
634 
635 	/* Copy fields from existing session. */
636 	tls_new->params = tls->params;
637 	tls_new->wq_index = tls->wq_index;
638 
639 	/* Deep copy keys. */
640 	if (tls_new->params.auth_key != NULL) {
641 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
642 		    M_KTLS, M_WAITOK);
643 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
644 		    tls->params.auth_key_len);
645 	}
646 
647 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
648 	    M_WAITOK);
649 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
650 	    tls->params.cipher_key_len);
651 
652 	return (tls_new);
653 }
654 #endif
655 
656 static void
657 ktls_cleanup(struct ktls_session *tls)
658 {
659 
660 	counter_u64_add(ktls_offload_active, -1);
661 	switch (tls->mode) {
662 	case TCP_TLS_MODE_SW:
663 		MPASS(tls->be != NULL);
664 		switch (tls->params.cipher_algorithm) {
665 		case CRYPTO_AES_CBC:
666 			counter_u64_add(ktls_sw_cbc, -1);
667 			break;
668 		case CRYPTO_AES_NIST_GCM_16:
669 			counter_u64_add(ktls_sw_gcm, -1);
670 			break;
671 		}
672 		tls->free(tls);
673 		break;
674 	case TCP_TLS_MODE_IFNET:
675 		switch (tls->params.cipher_algorithm) {
676 		case CRYPTO_AES_CBC:
677 			counter_u64_add(ktls_ifnet_cbc, -1);
678 			break;
679 		case CRYPTO_AES_NIST_GCM_16:
680 			counter_u64_add(ktls_ifnet_gcm, -1);
681 			break;
682 		}
683 		if (tls->snd_tag != NULL)
684 			m_snd_tag_rele(tls->snd_tag);
685 		break;
686 #ifdef TCP_OFFLOAD
687 	case TCP_TLS_MODE_TOE:
688 		switch (tls->params.cipher_algorithm) {
689 		case CRYPTO_AES_CBC:
690 			counter_u64_add(ktls_toe_cbc, -1);
691 			break;
692 		case CRYPTO_AES_NIST_GCM_16:
693 			counter_u64_add(ktls_toe_gcm, -1);
694 			break;
695 		}
696 		break;
697 #endif
698 	}
699 	if (tls->params.auth_key != NULL) {
700 		zfree(tls->params.auth_key, M_KTLS);
701 		tls->params.auth_key = NULL;
702 		tls->params.auth_key_len = 0;
703 	}
704 	if (tls->params.cipher_key != NULL) {
705 		zfree(tls->params.cipher_key, M_KTLS);
706 		tls->params.cipher_key = NULL;
707 		tls->params.cipher_key_len = 0;
708 	}
709 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
710 }
711 
712 #if defined(INET) || defined(INET6)
713 
714 #ifdef TCP_OFFLOAD
715 static int
716 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
717 {
718 	struct inpcb *inp;
719 	struct tcpcb *tp;
720 	int error;
721 
722 	inp = so->so_pcb;
723 	INP_WLOCK(inp);
724 	if (inp->inp_flags2 & INP_FREED) {
725 		INP_WUNLOCK(inp);
726 		return (ECONNRESET);
727 	}
728 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
729 		INP_WUNLOCK(inp);
730 		return (ECONNRESET);
731 	}
732 	if (inp->inp_socket == NULL) {
733 		INP_WUNLOCK(inp);
734 		return (ECONNRESET);
735 	}
736 	tp = intotcpcb(inp);
737 	if (!(tp->t_flags & TF_TOE)) {
738 		INP_WUNLOCK(inp);
739 		return (EOPNOTSUPP);
740 	}
741 
742 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
743 	INP_WUNLOCK(inp);
744 	if (error == 0) {
745 		tls->mode = TCP_TLS_MODE_TOE;
746 		switch (tls->params.cipher_algorithm) {
747 		case CRYPTO_AES_CBC:
748 			counter_u64_add(ktls_toe_cbc, 1);
749 			break;
750 		case CRYPTO_AES_NIST_GCM_16:
751 			counter_u64_add(ktls_toe_gcm, 1);
752 			break;
753 		}
754 	}
755 	return (error);
756 }
757 #endif
758 
759 /*
760  * Common code used when first enabling ifnet TLS on a connection or
761  * when allocating a new ifnet TLS session due to a routing change.
762  * This function allocates a new TLS send tag on whatever interface
763  * the connection is currently routed over.
764  */
765 static int
766 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
767     struct m_snd_tag **mstp)
768 {
769 	union if_snd_tag_alloc_params params;
770 	struct ifnet *ifp;
771 	struct nhop_object *nh;
772 	struct tcpcb *tp;
773 	int error;
774 
775 	INP_RLOCK(inp);
776 	if (inp->inp_flags2 & INP_FREED) {
777 		INP_RUNLOCK(inp);
778 		return (ECONNRESET);
779 	}
780 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
781 		INP_RUNLOCK(inp);
782 		return (ECONNRESET);
783 	}
784 	if (inp->inp_socket == NULL) {
785 		INP_RUNLOCK(inp);
786 		return (ECONNRESET);
787 	}
788 	tp = intotcpcb(inp);
789 
790 	/*
791 	 * Check administrative controls on ifnet TLS to determine if
792 	 * ifnet TLS should be denied.
793 	 *
794 	 * - Always permit 'force' requests.
795 	 * - ktls_ifnet_permitted == 0: always deny.
796 	 */
797 	if (!force && ktls_ifnet_permitted == 0) {
798 		INP_RUNLOCK(inp);
799 		return (ENXIO);
800 	}
801 
802 	/*
803 	 * XXX: Use the cached route in the inpcb to find the
804 	 * interface.  This should perhaps instead use
805 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
806 	 * enabled after a connection has completed key negotiation in
807 	 * userland, the cached route will be present in practice.
808 	 */
809 	nh = inp->inp_route.ro_nh;
810 	if (nh == NULL) {
811 		INP_RUNLOCK(inp);
812 		return (ENXIO);
813 	}
814 	ifp = nh->nh_ifp;
815 	if_ref(ifp);
816 
817 	params.hdr.type = IF_SND_TAG_TYPE_TLS;
818 	params.hdr.flowid = inp->inp_flowid;
819 	params.hdr.flowtype = inp->inp_flowtype;
820 	params.hdr.numa_domain = inp->inp_numa_domain;
821 	params.tls.inp = inp;
822 	params.tls.tls = tls;
823 	INP_RUNLOCK(inp);
824 
825 	if (ifp->if_snd_tag_alloc == NULL) {
826 		error = EOPNOTSUPP;
827 		goto out;
828 	}
829 	if ((ifp->if_capenable & IFCAP_NOMAP) == 0) {
830 		error = EOPNOTSUPP;
831 		goto out;
832 	}
833 	if (inp->inp_vflag & INP_IPV6) {
834 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
835 			error = EOPNOTSUPP;
836 			goto out;
837 		}
838 	} else {
839 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
840 			error = EOPNOTSUPP;
841 			goto out;
842 		}
843 	}
844 	error = ifp->if_snd_tag_alloc(ifp, &params, mstp);
845 out:
846 	if_rele(ifp);
847 	return (error);
848 }
849 
850 static int
851 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
852 {
853 	struct m_snd_tag *mst;
854 	int error;
855 
856 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
857 	if (error == 0) {
858 		tls->mode = TCP_TLS_MODE_IFNET;
859 		tls->snd_tag = mst;
860 		switch (tls->params.cipher_algorithm) {
861 		case CRYPTO_AES_CBC:
862 			counter_u64_add(ktls_ifnet_cbc, 1);
863 			break;
864 		case CRYPTO_AES_NIST_GCM_16:
865 			counter_u64_add(ktls_ifnet_gcm, 1);
866 			break;
867 		}
868 	}
869 	return (error);
870 }
871 
872 static int
873 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
874 {
875 	struct rm_priotracker prio;
876 	struct ktls_crypto_backend *be;
877 
878 	/*
879 	 * Choose the best software crypto backend.  Backends are
880 	 * stored in sorted priority order (larget value == most
881 	 * important at the head of the list), so this just stops on
882 	 * the first backend that claims the session by returning
883 	 * success.
884 	 */
885 	if (ktls_allow_unload)
886 		rm_rlock(&ktls_backends_lock, &prio);
887 	LIST_FOREACH(be, &ktls_backends, next) {
888 		if (be->try(so, tls, direction) == 0)
889 			break;
890 		KASSERT(tls->cipher == NULL,
891 		    ("ktls backend leaked a cipher pointer"));
892 	}
893 	if (be != NULL) {
894 		if (ktls_allow_unload)
895 			be->use_count++;
896 		tls->be = be;
897 	}
898 	if (ktls_allow_unload)
899 		rm_runlock(&ktls_backends_lock, &prio);
900 	if (be == NULL)
901 		return (EOPNOTSUPP);
902 	tls->mode = TCP_TLS_MODE_SW;
903 	switch (tls->params.cipher_algorithm) {
904 	case CRYPTO_AES_CBC:
905 		counter_u64_add(ktls_sw_cbc, 1);
906 		break;
907 	case CRYPTO_AES_NIST_GCM_16:
908 		counter_u64_add(ktls_sw_gcm, 1);
909 		break;
910 	}
911 	return (0);
912 }
913 
914 /*
915  * KTLS RX stores data in the socket buffer as a list of TLS records,
916  * where each record is stored as a control message containg the TLS
917  * header followed by data mbufs containing the decrypted data.  This
918  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
919  * both encrypted and decrypted data.  TLS records decrypted by a NIC
920  * should be queued to the socket buffer as records, but encrypted
921  * data which needs to be decrypted by software arrives as a stream of
922  * regular mbufs which need to be converted.  In addition, there may
923  * already be pending encrypted data in the socket buffer when KTLS RX
924  * is enabled.
925  *
926  * To manage not-yet-decrypted data for KTLS RX, the following scheme
927  * is used:
928  *
929  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
930  *
931  * - ktls_check_rx checks this chain of mbufs reading the TLS header
932  *   from the first mbuf.  Once all of the data for that TLS record is
933  *   queued, the socket is queued to a worker thread.
934  *
935  * - The worker thread calls ktls_decrypt to decrypt TLS records in
936  *   the TLS chain.  Each TLS record is detached from the TLS chain,
937  *   decrypted, and inserted into the regular socket buffer chain as
938  *   record starting with a control message holding the TLS header and
939  *   a chain of mbufs holding the encrypted data.
940  */
941 
942 static void
943 sb_mark_notready(struct sockbuf *sb)
944 {
945 	struct mbuf *m;
946 
947 	m = sb->sb_mb;
948 	sb->sb_mtls = m;
949 	sb->sb_mb = NULL;
950 	sb->sb_mbtail = NULL;
951 	sb->sb_lastrecord = NULL;
952 	for (; m != NULL; m = m->m_next) {
953 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
954 		    __func__));
955 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
956 		    __func__));
957 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
958 		    __func__));
959 		m->m_flags |= M_NOTREADY;
960 		sb->sb_acc -= m->m_len;
961 		sb->sb_tlscc += m->m_len;
962 		sb->sb_mtlstail = m;
963 	}
964 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
965 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
966 	    sb->sb_ccc));
967 }
968 
969 int
970 ktls_enable_rx(struct socket *so, struct tls_enable *en)
971 {
972 	struct ktls_session *tls;
973 	int error;
974 
975 	if (!ktls_offload_enable)
976 		return (ENOTSUP);
977 
978 	counter_u64_add(ktls_offload_enable_calls, 1);
979 
980 	/*
981 	 * This should always be true since only the TCP socket option
982 	 * invokes this function.
983 	 */
984 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
985 		return (EINVAL);
986 
987 	/*
988 	 * XXX: Don't overwrite existing sessions.  We should permit
989 	 * this to support rekeying in the future.
990 	 */
991 	if (so->so_rcv.sb_tls_info != NULL)
992 		return (EALREADY);
993 
994 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
995 		return (ENOTSUP);
996 
997 	/* TLS 1.3 is not yet supported. */
998 	if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
999 	    en->tls_vminor == TLS_MINOR_VER_THREE)
1000 		return (ENOTSUP);
1001 
1002 	error = ktls_create_session(so, en, &tls);
1003 	if (error)
1004 		return (error);
1005 
1006 #ifdef TCP_OFFLOAD
1007 	error = ktls_try_toe(so, tls, KTLS_RX);
1008 	if (error)
1009 #endif
1010 		error = ktls_try_sw(so, tls, KTLS_RX);
1011 
1012 	if (error) {
1013 		ktls_cleanup(tls);
1014 		return (error);
1015 	}
1016 
1017 	/* Mark the socket as using TLS offload. */
1018 	SOCKBUF_LOCK(&so->so_rcv);
1019 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1020 	so->so_rcv.sb_tls_info = tls;
1021 	so->so_rcv.sb_flags |= SB_TLS_RX;
1022 
1023 	/* Mark existing data as not ready until it can be decrypted. */
1024 	sb_mark_notready(&so->so_rcv);
1025 	ktls_check_rx(&so->so_rcv);
1026 	SOCKBUF_UNLOCK(&so->so_rcv);
1027 
1028 	counter_u64_add(ktls_offload_total, 1);
1029 
1030 	return (0);
1031 }
1032 
1033 int
1034 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1035 {
1036 	struct ktls_session *tls;
1037 	int error;
1038 
1039 	if (!ktls_offload_enable)
1040 		return (ENOTSUP);
1041 
1042 	counter_u64_add(ktls_offload_enable_calls, 1);
1043 
1044 	/*
1045 	 * This should always be true since only the TCP socket option
1046 	 * invokes this function.
1047 	 */
1048 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1049 		return (EINVAL);
1050 
1051 	/*
1052 	 * XXX: Don't overwrite existing sessions.  We should permit
1053 	 * this to support rekeying in the future.
1054 	 */
1055 	if (so->so_snd.sb_tls_info != NULL)
1056 		return (EALREADY);
1057 
1058 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1059 		return (ENOTSUP);
1060 
1061 	/* TLS requires ext pgs */
1062 	if (mb_use_ext_pgs == 0)
1063 		return (ENXIO);
1064 
1065 	error = ktls_create_session(so, en, &tls);
1066 	if (error)
1067 		return (error);
1068 
1069 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1070 #ifdef TCP_OFFLOAD
1071 	error = ktls_try_toe(so, tls, KTLS_TX);
1072 	if (error)
1073 #endif
1074 		error = ktls_try_ifnet(so, tls, false);
1075 	if (error)
1076 		error = ktls_try_sw(so, tls, KTLS_TX);
1077 
1078 	if (error) {
1079 		ktls_cleanup(tls);
1080 		return (error);
1081 	}
1082 
1083 	error = sblock(&so->so_snd, SBL_WAIT);
1084 	if (error) {
1085 		ktls_cleanup(tls);
1086 		return (error);
1087 	}
1088 
1089 	SOCKBUF_LOCK(&so->so_snd);
1090 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1091 	so->so_snd.sb_tls_info = tls;
1092 	if (tls->mode != TCP_TLS_MODE_SW)
1093 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1094 	SOCKBUF_UNLOCK(&so->so_snd);
1095 	sbunlock(&so->so_snd);
1096 
1097 	counter_u64_add(ktls_offload_total, 1);
1098 
1099 	return (0);
1100 }
1101 
1102 int
1103 ktls_get_rx_mode(struct socket *so)
1104 {
1105 	struct ktls_session *tls;
1106 	struct inpcb *inp;
1107 	int mode;
1108 
1109 	inp = so->so_pcb;
1110 	INP_WLOCK_ASSERT(inp);
1111 	SOCKBUF_LOCK(&so->so_rcv);
1112 	tls = so->so_rcv.sb_tls_info;
1113 	if (tls == NULL)
1114 		mode = TCP_TLS_MODE_NONE;
1115 	else
1116 		mode = tls->mode;
1117 	SOCKBUF_UNLOCK(&so->so_rcv);
1118 	return (mode);
1119 }
1120 
1121 int
1122 ktls_get_tx_mode(struct socket *so)
1123 {
1124 	struct ktls_session *tls;
1125 	struct inpcb *inp;
1126 	int mode;
1127 
1128 	inp = so->so_pcb;
1129 	INP_WLOCK_ASSERT(inp);
1130 	SOCKBUF_LOCK(&so->so_snd);
1131 	tls = so->so_snd.sb_tls_info;
1132 	if (tls == NULL)
1133 		mode = TCP_TLS_MODE_NONE;
1134 	else
1135 		mode = tls->mode;
1136 	SOCKBUF_UNLOCK(&so->so_snd);
1137 	return (mode);
1138 }
1139 
1140 /*
1141  * Switch between SW and ifnet TLS sessions as requested.
1142  */
1143 int
1144 ktls_set_tx_mode(struct socket *so, int mode)
1145 {
1146 	struct ktls_session *tls, *tls_new;
1147 	struct inpcb *inp;
1148 	int error;
1149 
1150 	switch (mode) {
1151 	case TCP_TLS_MODE_SW:
1152 	case TCP_TLS_MODE_IFNET:
1153 		break;
1154 	default:
1155 		return (EINVAL);
1156 	}
1157 
1158 	inp = so->so_pcb;
1159 	INP_WLOCK_ASSERT(inp);
1160 	SOCKBUF_LOCK(&so->so_snd);
1161 	tls = so->so_snd.sb_tls_info;
1162 	if (tls == NULL) {
1163 		SOCKBUF_UNLOCK(&so->so_snd);
1164 		return (0);
1165 	}
1166 
1167 	if (tls->mode == mode) {
1168 		SOCKBUF_UNLOCK(&so->so_snd);
1169 		return (0);
1170 	}
1171 
1172 	tls = ktls_hold(tls);
1173 	SOCKBUF_UNLOCK(&so->so_snd);
1174 	INP_WUNLOCK(inp);
1175 
1176 	tls_new = ktls_clone_session(tls);
1177 
1178 	if (mode == TCP_TLS_MODE_IFNET)
1179 		error = ktls_try_ifnet(so, tls_new, true);
1180 	else
1181 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1182 	if (error) {
1183 		counter_u64_add(ktls_switch_failed, 1);
1184 		ktls_free(tls_new);
1185 		ktls_free(tls);
1186 		INP_WLOCK(inp);
1187 		return (error);
1188 	}
1189 
1190 	error = sblock(&so->so_snd, SBL_WAIT);
1191 	if (error) {
1192 		counter_u64_add(ktls_switch_failed, 1);
1193 		ktls_free(tls_new);
1194 		ktls_free(tls);
1195 		INP_WLOCK(inp);
1196 		return (error);
1197 	}
1198 
1199 	/*
1200 	 * If we raced with another session change, keep the existing
1201 	 * session.
1202 	 */
1203 	if (tls != so->so_snd.sb_tls_info) {
1204 		counter_u64_add(ktls_switch_failed, 1);
1205 		sbunlock(&so->so_snd);
1206 		ktls_free(tls_new);
1207 		ktls_free(tls);
1208 		INP_WLOCK(inp);
1209 		return (EBUSY);
1210 	}
1211 
1212 	SOCKBUF_LOCK(&so->so_snd);
1213 	so->so_snd.sb_tls_info = tls_new;
1214 	if (tls_new->mode != TCP_TLS_MODE_SW)
1215 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1216 	SOCKBUF_UNLOCK(&so->so_snd);
1217 	sbunlock(&so->so_snd);
1218 
1219 	/*
1220 	 * Drop two references on 'tls'.  The first is for the
1221 	 * ktls_hold() above.  The second drops the reference from the
1222 	 * socket buffer.
1223 	 */
1224 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1225 	ktls_free(tls);
1226 	ktls_free(tls);
1227 
1228 	if (mode == TCP_TLS_MODE_IFNET)
1229 		counter_u64_add(ktls_switch_to_ifnet, 1);
1230 	else
1231 		counter_u64_add(ktls_switch_to_sw, 1);
1232 
1233 	INP_WLOCK(inp);
1234 	return (0);
1235 }
1236 
1237 /*
1238  * Try to allocate a new TLS send tag.  This task is scheduled when
1239  * ip_output detects a route change while trying to transmit a packet
1240  * holding a TLS record.  If a new tag is allocated, replace the tag
1241  * in the TLS session.  Subsequent packets on the connection will use
1242  * the new tag.  If a new tag cannot be allocated, drop the
1243  * connection.
1244  */
1245 static void
1246 ktls_reset_send_tag(void *context, int pending)
1247 {
1248 	struct epoch_tracker et;
1249 	struct ktls_session *tls;
1250 	struct m_snd_tag *old, *new;
1251 	struct inpcb *inp;
1252 	struct tcpcb *tp;
1253 	int error;
1254 
1255 	MPASS(pending == 1);
1256 
1257 	tls = context;
1258 	inp = tls->inp;
1259 
1260 	/*
1261 	 * Free the old tag first before allocating a new one.
1262 	 * ip[6]_output_send() will treat a NULL send tag the same as
1263 	 * an ifp mismatch and drop packets until a new tag is
1264 	 * allocated.
1265 	 *
1266 	 * Write-lock the INP when changing tls->snd_tag since
1267 	 * ip[6]_output_send() holds a read-lock when reading the
1268 	 * pointer.
1269 	 */
1270 	INP_WLOCK(inp);
1271 	old = tls->snd_tag;
1272 	tls->snd_tag = NULL;
1273 	INP_WUNLOCK(inp);
1274 	if (old != NULL)
1275 		m_snd_tag_rele(old);
1276 
1277 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1278 
1279 	if (error == 0) {
1280 		INP_WLOCK(inp);
1281 		tls->snd_tag = new;
1282 		mtx_pool_lock(mtxpool_sleep, tls);
1283 		tls->reset_pending = false;
1284 		mtx_pool_unlock(mtxpool_sleep, tls);
1285 		if (!in_pcbrele_wlocked(inp))
1286 			INP_WUNLOCK(inp);
1287 
1288 		counter_u64_add(ktls_ifnet_reset, 1);
1289 
1290 		/*
1291 		 * XXX: Should we kick tcp_output explicitly now that
1292 		 * the send tag is fixed or just rely on timers?
1293 		 */
1294 	} else {
1295 		NET_EPOCH_ENTER(et);
1296 		INP_WLOCK(inp);
1297 		if (!in_pcbrele_wlocked(inp)) {
1298 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1299 			    !(inp->inp_flags & INP_DROPPED)) {
1300 				tp = intotcpcb(inp);
1301 				CURVNET_SET(tp->t_vnet);
1302 				tp = tcp_drop(tp, ECONNABORTED);
1303 				CURVNET_RESTORE();
1304 				if (tp != NULL)
1305 					INP_WUNLOCK(inp);
1306 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1307 			} else
1308 				INP_WUNLOCK(inp);
1309 		}
1310 		NET_EPOCH_EXIT(et);
1311 
1312 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1313 
1314 		/*
1315 		 * Leave reset_pending true to avoid future tasks while
1316 		 * the socket goes away.
1317 		 */
1318 	}
1319 
1320 	ktls_free(tls);
1321 }
1322 
1323 int
1324 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1325 {
1326 
1327 	if (inp == NULL)
1328 		return (ENOBUFS);
1329 
1330 	INP_LOCK_ASSERT(inp);
1331 
1332 	/*
1333 	 * See if we should schedule a task to update the send tag for
1334 	 * this session.
1335 	 */
1336 	mtx_pool_lock(mtxpool_sleep, tls);
1337 	if (!tls->reset_pending) {
1338 		(void) ktls_hold(tls);
1339 		in_pcbref(inp);
1340 		tls->inp = inp;
1341 		tls->reset_pending = true;
1342 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1343 	}
1344 	mtx_pool_unlock(mtxpool_sleep, tls);
1345 	return (ENOBUFS);
1346 }
1347 #endif
1348 
1349 void
1350 ktls_destroy(struct ktls_session *tls)
1351 {
1352 	struct rm_priotracker prio;
1353 
1354 	ktls_cleanup(tls);
1355 	if (tls->be != NULL && ktls_allow_unload) {
1356 		rm_rlock(&ktls_backends_lock, &prio);
1357 		tls->be->use_count--;
1358 		rm_runlock(&ktls_backends_lock, &prio);
1359 	}
1360 	uma_zfree(ktls_session_zone, tls);
1361 }
1362 
1363 void
1364 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1365 {
1366 
1367 	for (; m != NULL; m = m->m_next) {
1368 		KASSERT((m->m_flags & M_EXTPG) != 0,
1369 		    ("ktls_seq: mapped mbuf %p", m));
1370 
1371 		m->m_epg_seqno = sb->sb_tls_seqno;
1372 		sb->sb_tls_seqno++;
1373 	}
1374 }
1375 
1376 /*
1377  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1378  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1379  * mbuf must be populated with the payload of each TLS record.
1380  *
1381  * The record_type argument specifies the TLS record type used when
1382  * populating the TLS header.
1383  *
1384  * The enq_count argument on return is set to the number of pages of
1385  * payload data for this entire chain that need to be encrypted via SW
1386  * encryption.  The returned value should be passed to ktls_enqueue
1387  * when scheduling encryption of this chain of mbufs.  To handle the
1388  * special case of empty fragments for TLS 1.0 sessions, an empty
1389  * fragment counts as one page.
1390  */
1391 void
1392 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1393     uint8_t record_type)
1394 {
1395 	struct tls_record_layer *tlshdr;
1396 	struct mbuf *m;
1397 	uint64_t *noncep;
1398 	uint16_t tls_len;
1399 	int maxlen;
1400 
1401 	maxlen = tls->params.max_frame_len;
1402 	*enq_cnt = 0;
1403 	for (m = top; m != NULL; m = m->m_next) {
1404 		/*
1405 		 * All mbufs in the chain should be TLS records whose
1406 		 * payload does not exceed the maximum frame length.
1407 		 *
1408 		 * Empty TLS records are permitted when using CBC.
1409 		 */
1410 		KASSERT(m->m_len <= maxlen &&
1411 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1412 		    m->m_len >= 0 : m->m_len > 0),
1413 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1414 
1415 		/*
1416 		 * TLS frames require unmapped mbufs to store session
1417 		 * info.
1418 		 */
1419 		KASSERT((m->m_flags & M_EXTPG) != 0,
1420 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1421 
1422 		tls_len = m->m_len;
1423 
1424 		/* Save a reference to the session. */
1425 		m->m_epg_tls = ktls_hold(tls);
1426 
1427 		m->m_epg_hdrlen = tls->params.tls_hlen;
1428 		m->m_epg_trllen = tls->params.tls_tlen;
1429 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1430 			int bs, delta;
1431 
1432 			/*
1433 			 * AES-CBC pads messages to a multiple of the
1434 			 * block size.  Note that the padding is
1435 			 * applied after the digest and the encryption
1436 			 * is done on the "plaintext || mac || padding".
1437 			 * At least one byte of padding is always
1438 			 * present.
1439 			 *
1440 			 * Compute the final trailer length assuming
1441 			 * at most one block of padding.
1442 			 * tls->params.sb_tls_tlen is the maximum
1443 			 * possible trailer length (padding + digest).
1444 			 * delta holds the number of excess padding
1445 			 * bytes if the maximum were used.  Those
1446 			 * extra bytes are removed.
1447 			 */
1448 			bs = tls->params.tls_bs;
1449 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1450 			m->m_epg_trllen -= delta;
1451 		}
1452 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1453 
1454 		/* Populate the TLS header. */
1455 		tlshdr = (void *)m->m_epg_hdr;
1456 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1457 
1458 		/*
1459 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1460 		 * of TLS_RLTYPE_APP.
1461 		 */
1462 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1463 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1464 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1465 			tlshdr->tls_type = TLS_RLTYPE_APP;
1466 			/* save the real record type for later */
1467 			m->m_epg_record_type = record_type;
1468 			m->m_epg_trail[0] = record_type;
1469 		} else {
1470 			tlshdr->tls_vminor = tls->params.tls_vminor;
1471 			tlshdr->tls_type = record_type;
1472 		}
1473 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1474 
1475 		/*
1476 		 * Store nonces / explicit IVs after the end of the
1477 		 * TLS header.
1478 		 *
1479 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1480 		 * from the end of the IV.  The nonce is then
1481 		 * incremented for use by the next record.
1482 		 *
1483 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1484 		 */
1485 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1486 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1487 			noncep = (uint64_t *)(tls->params.iv + 8);
1488 			be64enc(tlshdr + 1, *noncep);
1489 			(*noncep)++;
1490 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1491 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1492 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1493 
1494 		/*
1495 		 * When using SW encryption, mark the mbuf not ready.
1496 		 * It will be marked ready via sbready() after the
1497 		 * record has been encrypted.
1498 		 *
1499 		 * When using ifnet TLS, unencrypted TLS records are
1500 		 * sent down the stack to the NIC.
1501 		 */
1502 		if (tls->mode == TCP_TLS_MODE_SW) {
1503 			m->m_flags |= M_NOTREADY;
1504 			m->m_epg_nrdy = m->m_epg_npgs;
1505 			if (__predict_false(tls_len == 0)) {
1506 				/* TLS 1.0 empty fragment. */
1507 				*enq_cnt += 1;
1508 			} else
1509 				*enq_cnt += m->m_epg_npgs;
1510 		}
1511 	}
1512 }
1513 
1514 void
1515 ktls_check_rx(struct sockbuf *sb)
1516 {
1517 	struct tls_record_layer hdr;
1518 	struct ktls_wq *wq;
1519 	struct socket *so;
1520 	bool running;
1521 
1522 	SOCKBUF_LOCK_ASSERT(sb);
1523 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1524 	    __func__, sb));
1525 	so = __containerof(sb, struct socket, so_rcv);
1526 
1527 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1528 		return;
1529 
1530 	/* Is there enough queued for a TLS header? */
1531 	if (sb->sb_tlscc < sizeof(hdr)) {
1532 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1533 			so->so_error = EMSGSIZE;
1534 		return;
1535 	}
1536 
1537 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1538 
1539 	/* Is the entire record queued? */
1540 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1541 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1542 			so->so_error = EMSGSIZE;
1543 		return;
1544 	}
1545 
1546 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1547 
1548 	soref(so);
1549 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1550 	mtx_lock(&wq->mtx);
1551 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1552 	running = wq->running;
1553 	mtx_unlock(&wq->mtx);
1554 	if (!running)
1555 		wakeup(wq);
1556 	counter_u64_add(ktls_cnt_rx_queued, 1);
1557 }
1558 
1559 static struct mbuf *
1560 ktls_detach_record(struct sockbuf *sb, int len)
1561 {
1562 	struct mbuf *m, *n, *top;
1563 	int remain;
1564 
1565 	SOCKBUF_LOCK_ASSERT(sb);
1566 	MPASS(len <= sb->sb_tlscc);
1567 
1568 	/*
1569 	 * If TLS chain is the exact size of the record,
1570 	 * just grab the whole record.
1571 	 */
1572 	top = sb->sb_mtls;
1573 	if (sb->sb_tlscc == len) {
1574 		sb->sb_mtls = NULL;
1575 		sb->sb_mtlstail = NULL;
1576 		goto out;
1577 	}
1578 
1579 	/*
1580 	 * While it would be nice to use m_split() here, we need
1581 	 * to know exactly what m_split() allocates to update the
1582 	 * accounting, so do it inline instead.
1583 	 */
1584 	remain = len;
1585 	for (m = top; remain > m->m_len; m = m->m_next)
1586 		remain -= m->m_len;
1587 
1588 	/* Easy case: don't have to split 'm'. */
1589 	if (remain == m->m_len) {
1590 		sb->sb_mtls = m->m_next;
1591 		if (sb->sb_mtls == NULL)
1592 			sb->sb_mtlstail = NULL;
1593 		m->m_next = NULL;
1594 		goto out;
1595 	}
1596 
1597 	/*
1598 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1599 	 * with M_NOWAIT first.
1600 	 */
1601 	n = m_get(M_NOWAIT, MT_DATA);
1602 	if (n == NULL) {
1603 		/*
1604 		 * Use M_WAITOK with socket buffer unlocked.  If
1605 		 * 'sb_mtls' changes while the lock is dropped, return
1606 		 * NULL to force the caller to retry.
1607 		 */
1608 		SOCKBUF_UNLOCK(sb);
1609 
1610 		n = m_get(M_WAITOK, MT_DATA);
1611 
1612 		SOCKBUF_LOCK(sb);
1613 		if (sb->sb_mtls != top) {
1614 			m_free(n);
1615 			return (NULL);
1616 		}
1617 	}
1618 	n->m_flags |= M_NOTREADY;
1619 
1620 	/* Store remainder in 'n'. */
1621 	n->m_len = m->m_len - remain;
1622 	if (m->m_flags & M_EXT) {
1623 		n->m_data = m->m_data + remain;
1624 		mb_dupcl(n, m);
1625 	} else {
1626 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1627 	}
1628 
1629 	/* Trim 'm' and update accounting. */
1630 	m->m_len -= n->m_len;
1631 	sb->sb_tlscc -= n->m_len;
1632 	sb->sb_ccc -= n->m_len;
1633 
1634 	/* Account for 'n'. */
1635 	sballoc_ktls_rx(sb, n);
1636 
1637 	/* Insert 'n' into the TLS chain. */
1638 	sb->sb_mtls = n;
1639 	n->m_next = m->m_next;
1640 	if (sb->sb_mtlstail == m)
1641 		sb->sb_mtlstail = n;
1642 
1643 	/* Detach the record from the TLS chain. */
1644 	m->m_next = NULL;
1645 
1646 out:
1647 	MPASS(m_length(top, NULL) == len);
1648 	for (m = top; m != NULL; m = m->m_next)
1649 		sbfree_ktls_rx(sb, m);
1650 	sb->sb_tlsdcc = len;
1651 	sb->sb_ccc += len;
1652 	SBCHECK(sb);
1653 	return (top);
1654 }
1655 
1656 static void
1657 ktls_decrypt(struct socket *so)
1658 {
1659 	char tls_header[MBUF_PEXT_HDR_LEN];
1660 	struct ktls_session *tls;
1661 	struct sockbuf *sb;
1662 	struct tls_record_layer *hdr;
1663 	struct tls_get_record tgr;
1664 	struct mbuf *control, *data, *m;
1665 	uint64_t seqno;
1666 	int error, remain, tls_len, trail_len;
1667 
1668 	hdr = (struct tls_record_layer *)tls_header;
1669 	sb = &so->so_rcv;
1670 	SOCKBUF_LOCK(sb);
1671 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1672 	    ("%s: socket %p not running", __func__, so));
1673 
1674 	tls = sb->sb_tls_info;
1675 	MPASS(tls != NULL);
1676 
1677 	for (;;) {
1678 		/* Is there enough queued for a TLS header? */
1679 		if (sb->sb_tlscc < tls->params.tls_hlen)
1680 			break;
1681 
1682 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1683 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1684 
1685 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1686 		    hdr->tls_vminor != tls->params.tls_vminor)
1687 			error = EINVAL;
1688 		else if (tls_len < tls->params.tls_hlen || tls_len >
1689 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1690 		    tls->params.tls_tlen)
1691 			error = EMSGSIZE;
1692 		else
1693 			error = 0;
1694 		if (__predict_false(error != 0)) {
1695 			/*
1696 			 * We have a corrupted record and are likely
1697 			 * out of sync.  The connection isn't
1698 			 * recoverable at this point, so abort it.
1699 			 */
1700 			SOCKBUF_UNLOCK(sb);
1701 			counter_u64_add(ktls_offload_corrupted_records, 1);
1702 
1703 			CURVNET_SET(so->so_vnet);
1704 			so->so_proto->pr_usrreqs->pru_abort(so);
1705 			so->so_error = error;
1706 			CURVNET_RESTORE();
1707 			goto deref;
1708 		}
1709 
1710 		/* Is the entire record queued? */
1711 		if (sb->sb_tlscc < tls_len)
1712 			break;
1713 
1714 		/*
1715 		 * Split out the portion of the mbuf chain containing
1716 		 * this TLS record.
1717 		 */
1718 		data = ktls_detach_record(sb, tls_len);
1719 		if (data == NULL)
1720 			continue;
1721 		MPASS(sb->sb_tlsdcc == tls_len);
1722 
1723 		seqno = sb->sb_tls_seqno;
1724 		sb->sb_tls_seqno++;
1725 		SBCHECK(sb);
1726 		SOCKBUF_UNLOCK(sb);
1727 
1728 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1729 		if (error) {
1730 			counter_u64_add(ktls_offload_failed_crypto, 1);
1731 
1732 			SOCKBUF_LOCK(sb);
1733 			if (sb->sb_tlsdcc == 0) {
1734 				/*
1735 				 * sbcut/drop/flush discarded these
1736 				 * mbufs.
1737 				 */
1738 				m_freem(data);
1739 				break;
1740 			}
1741 
1742 			/*
1743 			 * Drop this TLS record's data, but keep
1744 			 * decrypting subsequent records.
1745 			 */
1746 			sb->sb_ccc -= tls_len;
1747 			sb->sb_tlsdcc = 0;
1748 
1749 			CURVNET_SET(so->so_vnet);
1750 			so->so_error = EBADMSG;
1751 			sorwakeup_locked(so);
1752 			CURVNET_RESTORE();
1753 
1754 			m_freem(data);
1755 
1756 			SOCKBUF_LOCK(sb);
1757 			continue;
1758 		}
1759 
1760 		/* Allocate the control mbuf. */
1761 		tgr.tls_type = hdr->tls_type;
1762 		tgr.tls_vmajor = hdr->tls_vmajor;
1763 		tgr.tls_vminor = hdr->tls_vminor;
1764 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1765 		    trail_len);
1766 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1767 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1768 
1769 		SOCKBUF_LOCK(sb);
1770 		if (sb->sb_tlsdcc == 0) {
1771 			/* sbcut/drop/flush discarded these mbufs. */
1772 			MPASS(sb->sb_tlscc == 0);
1773 			m_freem(data);
1774 			m_freem(control);
1775 			break;
1776 		}
1777 
1778 		/*
1779 		 * Clear the 'dcc' accounting in preparation for
1780 		 * adding the decrypted record.
1781 		 */
1782 		sb->sb_ccc -= tls_len;
1783 		sb->sb_tlsdcc = 0;
1784 		SBCHECK(sb);
1785 
1786 		/* If there is no payload, drop all of the data. */
1787 		if (tgr.tls_length == htobe16(0)) {
1788 			m_freem(data);
1789 			data = NULL;
1790 		} else {
1791 			/* Trim header. */
1792 			remain = tls->params.tls_hlen;
1793 			while (remain > 0) {
1794 				if (data->m_len > remain) {
1795 					data->m_data += remain;
1796 					data->m_len -= remain;
1797 					break;
1798 				}
1799 				remain -= data->m_len;
1800 				data = m_free(data);
1801 			}
1802 
1803 			/* Trim trailer and clear M_NOTREADY. */
1804 			remain = be16toh(tgr.tls_length);
1805 			m = data;
1806 			for (m = data; remain > m->m_len; m = m->m_next) {
1807 				m->m_flags &= ~M_NOTREADY;
1808 				remain -= m->m_len;
1809 			}
1810 			m->m_len = remain;
1811 			m_freem(m->m_next);
1812 			m->m_next = NULL;
1813 			m->m_flags &= ~M_NOTREADY;
1814 
1815 			/* Set EOR on the final mbuf. */
1816 			m->m_flags |= M_EOR;
1817 		}
1818 
1819 		sbappendcontrol_locked(sb, data, control, 0);
1820 	}
1821 
1822 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1823 
1824 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1825 		so->so_error = EMSGSIZE;
1826 
1827 	sorwakeup_locked(so);
1828 
1829 deref:
1830 	SOCKBUF_UNLOCK_ASSERT(sb);
1831 
1832 	CURVNET_SET(so->so_vnet);
1833 	SOCK_LOCK(so);
1834 	sorele(so);
1835 	CURVNET_RESTORE();
1836 }
1837 
1838 void
1839 ktls_enqueue_to_free(struct mbuf *m)
1840 {
1841 	struct ktls_wq *wq;
1842 	bool running;
1843 
1844 	/* Mark it for freeing. */
1845 	m->m_epg_flags |= EPG_FLAG_2FREE;
1846 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1847 	mtx_lock(&wq->mtx);
1848 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1849 	running = wq->running;
1850 	mtx_unlock(&wq->mtx);
1851 	if (!running)
1852 		wakeup(wq);
1853 }
1854 
1855 void
1856 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1857 {
1858 	struct ktls_wq *wq;
1859 	bool running;
1860 
1861 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1862 	    (M_EXTPG | M_NOTREADY)),
1863 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1864 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1865 
1866 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1867 
1868 	m->m_epg_enc_cnt = page_count;
1869 
1870 	/*
1871 	 * Save a pointer to the socket.  The caller is responsible
1872 	 * for taking an additional reference via soref().
1873 	 */
1874 	m->m_epg_so = so;
1875 
1876 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1877 	mtx_lock(&wq->mtx);
1878 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1879 	running = wq->running;
1880 	mtx_unlock(&wq->mtx);
1881 	if (!running)
1882 		wakeup(wq);
1883 	counter_u64_add(ktls_cnt_tx_queued, 1);
1884 }
1885 
1886 static __noinline void
1887 ktls_encrypt(struct mbuf *top)
1888 {
1889 	struct ktls_session *tls;
1890 	struct socket *so;
1891 	struct mbuf *m;
1892 	vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1893 	struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1894 	struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1895 	vm_page_t pg;
1896 	int error, i, len, npages, off, total_pages;
1897 	bool is_anon;
1898 
1899 	so = top->m_epg_so;
1900 	tls = top->m_epg_tls;
1901 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
1902 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
1903 #ifdef INVARIANTS
1904 	top->m_epg_so = NULL;
1905 #endif
1906 	total_pages = top->m_epg_enc_cnt;
1907 	npages = 0;
1908 
1909 	/*
1910 	 * Encrypt the TLS records in the chain of mbufs starting with
1911 	 * 'top'.  'total_pages' gives us a total count of pages and is
1912 	 * used to know when we have finished encrypting the TLS
1913 	 * records originally queued with 'top'.
1914 	 *
1915 	 * NB: These mbufs are queued in the socket buffer and
1916 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
1917 	 * socket buffer lock is not held while traversing this chain.
1918 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1919 	 * pointers should be stable.  However, the 'm_next' of the
1920 	 * last mbuf encrypted is not necessarily NULL.  It can point
1921 	 * to other mbufs appended while 'top' was on the TLS work
1922 	 * queue.
1923 	 *
1924 	 * Each mbuf holds an entire TLS record.
1925 	 */
1926 	error = 0;
1927 	for (m = top; npages != total_pages; m = m->m_next) {
1928 		KASSERT(m->m_epg_tls == tls,
1929 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
1930 		    tls, m->m_epg_tls));
1931 		KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1932 		    (M_EXTPG | M_NOTREADY),
1933 		    ("%p not unready & nomap mbuf (top = %p)\n", m, top));
1934 		KASSERT(npages + m->m_epg_npgs <= total_pages,
1935 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
1936 		    total_pages, m));
1937 
1938 		/*
1939 		 * Generate source and destination ivoecs to pass to
1940 		 * the SW encryption backend.  For writable mbufs, the
1941 		 * destination iovec is a copy of the source and
1942 		 * encryption is done in place.  For file-backed mbufs
1943 		 * (from sendfile), anonymous wired pages are
1944 		 * allocated and assigned to the destination iovec.
1945 		 */
1946 		is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
1947 
1948 		off = m->m_epg_1st_off;
1949 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
1950 			len = m_epg_pagelen(m, i, off);
1951 			src_iov[i].iov_len = len;
1952 			src_iov[i].iov_base =
1953 			    (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
1954 				off;
1955 
1956 			if (is_anon) {
1957 				dst_iov[i].iov_base = src_iov[i].iov_base;
1958 				dst_iov[i].iov_len = src_iov[i].iov_len;
1959 				continue;
1960 			}
1961 retry_page:
1962 			pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
1963 			    VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
1964 			if (pg == NULL) {
1965 				vm_wait(NULL);
1966 				goto retry_page;
1967 			}
1968 			parray[i] = VM_PAGE_TO_PHYS(pg);
1969 			dst_iov[i].iov_base =
1970 			    (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
1971 			dst_iov[i].iov_len = len;
1972 		}
1973 
1974 		if (__predict_false(m->m_epg_npgs == 0)) {
1975 			/* TLS 1.0 empty fragment. */
1976 			npages++;
1977 		} else
1978 			npages += i;
1979 
1980 		error = (*tls->sw_encrypt)(tls,
1981 		    (const struct tls_record_layer *)m->m_epg_hdr,
1982 		    m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
1983 		    m->m_epg_record_type);
1984 		if (error) {
1985 			counter_u64_add(ktls_offload_failed_crypto, 1);
1986 			break;
1987 		}
1988 
1989 		/*
1990 		 * For file-backed mbufs, release the file-backed
1991 		 * pages and replace them in the ext_pgs array with
1992 		 * the anonymous wired pages allocated above.
1993 		 */
1994 		if (!is_anon) {
1995 			/* Free the old pages. */
1996 			m->m_ext.ext_free(m);
1997 
1998 			/* Replace them with the new pages. */
1999 			for (i = 0; i < m->m_epg_npgs; i++)
2000 				m->m_epg_pa[i] = parray[i];
2001 
2002 			/* Use the basic free routine. */
2003 			m->m_ext.ext_free = mb_free_mext_pgs;
2004 
2005 			/* Pages are now writable. */
2006 			m->m_epg_flags |= EPG_FLAG_ANON;
2007 		}
2008 
2009 		/*
2010 		 * Drop a reference to the session now that it is no
2011 		 * longer needed.  Existing code depends on encrypted
2012 		 * records having no associated session vs
2013 		 * yet-to-be-encrypted records having an associated
2014 		 * session.
2015 		 */
2016 		m->m_epg_tls = NULL;
2017 		ktls_free(tls);
2018 	}
2019 
2020 	CURVNET_SET(so->so_vnet);
2021 	if (error == 0) {
2022 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2023 	} else {
2024 		so->so_proto->pr_usrreqs->pru_abort(so);
2025 		so->so_error = EIO;
2026 		mb_free_notready(top, total_pages);
2027 	}
2028 
2029 	SOCK_LOCK(so);
2030 	sorele(so);
2031 	CURVNET_RESTORE();
2032 }
2033 
2034 static void
2035 ktls_work_thread(void *ctx)
2036 {
2037 	struct ktls_wq *wq = ctx;
2038 	struct mbuf *m, *n;
2039 	struct socket *so, *son;
2040 	STAILQ_HEAD(, mbuf) local_m_head;
2041 	STAILQ_HEAD(, socket) local_so_head;
2042 
2043 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2044 	fpu_kern_thread(0);
2045 #endif
2046 	for (;;) {
2047 		mtx_lock(&wq->mtx);
2048 		while (STAILQ_EMPTY(&wq->m_head) &&
2049 		    STAILQ_EMPTY(&wq->so_head)) {
2050 			wq->running = false;
2051 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2052 			wq->running = true;
2053 		}
2054 
2055 		STAILQ_INIT(&local_m_head);
2056 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2057 		STAILQ_INIT(&local_so_head);
2058 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2059 		mtx_unlock(&wq->mtx);
2060 
2061 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2062 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2063 				ktls_free(m->m_epg_tls);
2064 				uma_zfree(zone_mbuf, m);
2065 			} else {
2066 				ktls_encrypt(m);
2067 				counter_u64_add(ktls_cnt_tx_queued, -1);
2068 			}
2069 		}
2070 
2071 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2072 			ktls_decrypt(so);
2073 			counter_u64_add(ktls_cnt_rx_queued, -1);
2074 		}
2075 	}
2076 }
2077