xref: /freebsd/sys/kern/uipc_ktls.c (revision 3a3deb00a5e449c9478156b162dfa10ec82a2a3f)
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 #ifdef RSS
58 #include <net/netisr.h>
59 #include <net/rss_config.h>
60 #endif
61 #if defined(INET) || defined(INET6)
62 #include <netinet/in.h>
63 #include <netinet/in_pcb.h>
64 #endif
65 #include <netinet/tcp_var.h>
66 #ifdef TCP_OFFLOAD
67 #include <netinet/tcp_offload.h>
68 #endif
69 #include <opencrypto/xform.h>
70 #include <vm/uma_dbg.h>
71 #include <vm/vm.h>
72 #include <vm/vm_pageout.h>
73 #include <vm/vm_page.h>
74 
75 struct ktls_wq {
76 	struct mtx	mtx;
77 	STAILQ_HEAD(, mbuf_ext_pgs) head;
78 	bool		running;
79 } __aligned(CACHE_LINE_SIZE);
80 
81 static struct ktls_wq *ktls_wq;
82 static struct proc *ktls_proc;
83 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
84 static struct rmlock ktls_backends_lock;
85 static uma_zone_t ktls_session_zone;
86 static uint16_t ktls_cpuid_lookup[MAXCPU];
87 
88 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW, 0,
89     "Kernel TLS offload");
90 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW, 0,
91     "Kernel TLS offload stats");
92 
93 static int ktls_allow_unload;
94 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
95     &ktls_allow_unload, 0, "Allow software crypto modules to unload");
96 
97 #ifdef RSS
98 static int ktls_bind_threads = 1;
99 #else
100 static int ktls_bind_threads;
101 #endif
102 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
103     &ktls_bind_threads, 0,
104     "Bind crypto threads to cores or domains at boot");
105 
106 static u_int ktls_maxlen = 16384;
107 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
108     &ktls_maxlen, 0, "Maximum TLS record size");
109 
110 static int ktls_number_threads;
111 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
112     &ktls_number_threads, 0,
113     "Number of TLS threads in thread-pool");
114 
115 static bool ktls_offload_enable;
116 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW,
117     &ktls_offload_enable, 0,
118     "Enable support for kernel TLS offload");
119 
120 static bool ktls_cbc_enable = true;
121 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW,
122     &ktls_cbc_enable, 1,
123     "Enable Support of AES-CBC crypto for kernel TLS");
124 
125 static counter_u64_t ktls_tasks_active;
126 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
127     &ktls_tasks_active, "Number of active tasks");
128 
129 static counter_u64_t ktls_cnt_on;
130 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD,
131     &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto");
132 
133 static counter_u64_t ktls_offload_total;
134 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
135     CTLFLAG_RD, &ktls_offload_total,
136     "Total successful TLS setups (parameters set)");
137 
138 static counter_u64_t ktls_offload_enable_calls;
139 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
140     CTLFLAG_RD, &ktls_offload_enable_calls,
141     "Total number of TLS enable calls made");
142 
143 static counter_u64_t ktls_offload_active;
144 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
145     &ktls_offload_active, "Total Active TLS sessions");
146 
147 static counter_u64_t ktls_offload_failed_crypto;
148 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
149     &ktls_offload_failed_crypto, "Total TLS crypto failures");
150 
151 static counter_u64_t ktls_switch_to_ifnet;
152 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
153     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
154 
155 static counter_u64_t ktls_switch_to_sw;
156 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
157     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
158 
159 static counter_u64_t ktls_switch_failed;
160 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
161     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
162 
163 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD, 0,
164     "Software TLS session stats");
165 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD, 0,
166     "Hardware (ifnet) TLS session stats");
167 #ifdef TCP_OFFLOAD
168 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD, 0,
169     "TOE TLS session stats");
170 #endif
171 
172 static counter_u64_t ktls_sw_cbc;
173 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
174     "Active number of software TLS sessions using AES-CBC");
175 
176 static counter_u64_t ktls_sw_gcm;
177 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
178     "Active number of software TLS sessions using AES-GCM");
179 
180 static counter_u64_t ktls_ifnet_cbc;
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
182     &ktls_ifnet_cbc,
183     "Active number of ifnet TLS sessions using AES-CBC");
184 
185 static counter_u64_t ktls_ifnet_gcm;
186 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
187     &ktls_ifnet_gcm,
188     "Active number of ifnet TLS sessions using AES-GCM");
189 
190 static counter_u64_t ktls_ifnet_reset;
191 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
192     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
193 
194 static counter_u64_t ktls_ifnet_reset_dropped;
195 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
196     &ktls_ifnet_reset_dropped,
197     "TLS sessions dropped after failing to update ifnet send tag");
198 
199 static counter_u64_t ktls_ifnet_reset_failed;
200 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
201     &ktls_ifnet_reset_failed,
202     "TLS sessions that failed to allocate a new ifnet send tag");
203 
204 static int ktls_ifnet_permitted;
205 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
206     &ktls_ifnet_permitted, 1,
207     "Whether to permit hardware (ifnet) TLS sessions");
208 
209 #ifdef TCP_OFFLOAD
210 static counter_u64_t ktls_toe_cbc;
211 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
212     &ktls_toe_cbc,
213     "Active number of TOE TLS sessions using AES-CBC");
214 
215 static counter_u64_t ktls_toe_gcm;
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
217     &ktls_toe_gcm,
218     "Active number of TOE TLS sessions using AES-GCM");
219 #endif
220 
221 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
222 
223 static void ktls_cleanup(struct ktls_session *tls);
224 #if defined(INET) || defined(INET6)
225 static void ktls_reset_send_tag(void *context, int pending);
226 #endif
227 static void ktls_work_thread(void *ctx);
228 
229 int
230 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
231 {
232 	struct ktls_crypto_backend *curr_be, *tmp;
233 
234 	if (be->api_version != KTLS_API_VERSION) {
235 		printf("KTLS: API version mismatch (%d vs %d) for %s\n",
236 		    be->api_version, KTLS_API_VERSION,
237 		    be->name);
238 		return (EINVAL);
239 	}
240 
241 	rm_wlock(&ktls_backends_lock);
242 	printf("KTLS: Registering crypto method %s with prio %d\n",
243 	       be->name, be->prio);
244 	if (LIST_EMPTY(&ktls_backends)) {
245 		LIST_INSERT_HEAD(&ktls_backends, be, next);
246 	} else {
247 		LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
248 			if (curr_be->prio < be->prio) {
249 				LIST_INSERT_BEFORE(curr_be, be, next);
250 				break;
251 			}
252 			if (LIST_NEXT(curr_be, next) == NULL) {
253 				LIST_INSERT_AFTER(curr_be, be, next);
254 				break;
255 			}
256 		}
257 	}
258 	rm_wunlock(&ktls_backends_lock);
259 	return (0);
260 }
261 
262 int
263 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
264 {
265 	struct ktls_crypto_backend *tmp;
266 
267 	/*
268 	 * Don't error if the backend isn't registered.  This permits
269 	 * MOD_UNLOAD handlers to use this function unconditionally.
270 	 */
271 	rm_wlock(&ktls_backends_lock);
272 	LIST_FOREACH(tmp, &ktls_backends, next) {
273 		if (tmp == be)
274 			break;
275 	}
276 	if (tmp == NULL) {
277 		rm_wunlock(&ktls_backends_lock);
278 		return (0);
279 	}
280 
281 	if (!ktls_allow_unload) {
282 		rm_wunlock(&ktls_backends_lock);
283 		printf(
284 		    "KTLS: Deregistering crypto method %s is not supported\n",
285 		    be->name);
286 		return (EBUSY);
287 	}
288 
289 	if (be->use_count) {
290 		rm_wunlock(&ktls_backends_lock);
291 		return (EBUSY);
292 	}
293 
294 	LIST_REMOVE(be, next);
295 	rm_wunlock(&ktls_backends_lock);
296 	return (0);
297 }
298 
299 #if defined(INET) || defined(INET6)
300 static uint16_t
301 ktls_get_cpu(struct socket *so)
302 {
303 	struct inpcb *inp;
304 	uint16_t cpuid;
305 
306 	inp = sotoinpcb(so);
307 #ifdef RSS
308 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
309 	if (cpuid != NETISR_CPUID_NONE)
310 		return (cpuid);
311 #endif
312 	/*
313 	 * Just use the flowid to shard connections in a repeatable
314 	 * fashion.  Note that some crypto backends rely on the
315 	 * serialization provided by having the same connection use
316 	 * the same queue.
317 	 */
318 	cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
319 	return (cpuid);
320 }
321 #endif
322 
323 static void
324 ktls_init(void *dummy __unused)
325 {
326 	struct thread *td;
327 	struct pcpu *pc;
328 	cpuset_t mask;
329 	int error, i;
330 
331 	ktls_tasks_active = counter_u64_alloc(M_WAITOK);
332 	ktls_cnt_on = counter_u64_alloc(M_WAITOK);
333 	ktls_offload_total = counter_u64_alloc(M_WAITOK);
334 	ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
335 	ktls_offload_active = counter_u64_alloc(M_WAITOK);
336 	ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
337 	ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
338 	ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
339 	ktls_switch_failed = counter_u64_alloc(M_WAITOK);
340 	ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
341 	ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
342 	ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
343 	ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
344 	ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
345 	ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
346 	ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
347 #ifdef TCP_OFFLOAD
348 	ktls_toe_cbc = counter_u64_alloc(M_WAITOK);
349 	ktls_toe_gcm = counter_u64_alloc(M_WAITOK);
350 #endif
351 
352 	rm_init(&ktls_backends_lock, "ktls backends");
353 	LIST_INIT(&ktls_backends);
354 
355 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
356 	    M_WAITOK | M_ZERO);
357 
358 	ktls_session_zone = uma_zcreate("ktls_session",
359 	    sizeof(struct ktls_session),
360 	    NULL, NULL, NULL, NULL,
361 	    UMA_ALIGN_CACHE, 0);
362 
363 	/*
364 	 * Initialize the workqueues to run the TLS work.  We create a
365 	 * work queue for each CPU.
366 	 */
367 	CPU_FOREACH(i) {
368 		STAILQ_INIT(&ktls_wq[i].head);
369 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
370 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
371 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
372 		if (error)
373 			panic("Can't add KTLS thread %d error %d", i, error);
374 
375 		/*
376 		 * Bind threads to cores.  If ktls_bind_threads is >
377 		 * 1, then we bind to the NUMA domain.
378 		 */
379 		if (ktls_bind_threads) {
380 			if (ktls_bind_threads > 1) {
381 				pc = pcpu_find(i);
382 				CPU_COPY(&cpuset_domain[pc->pc_domain], &mask);
383 			} else {
384 				CPU_SETOF(i, &mask);
385 			}
386 			error = cpuset_setthread(td->td_tid, &mask);
387 			if (error)
388 				panic(
389 			    "Unable to bind KTLS thread for CPU %d error %d",
390 				     i, error);
391 		}
392 		ktls_cpuid_lookup[ktls_number_threads] = i;
393 		ktls_number_threads++;
394 	}
395 	printf("KTLS: Initialized %d threads\n", ktls_number_threads);
396 }
397 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
398 
399 #if defined(INET) || defined(INET6)
400 static int
401 ktls_create_session(struct socket *so, struct tls_enable *en,
402     struct ktls_session **tlsp)
403 {
404 	struct ktls_session *tls;
405 	int error;
406 
407 	/* Only TLS 1.0 - 1.3 are supported. */
408 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
409 		return (EINVAL);
410 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
411 	    en->tls_vminor > TLS_MINOR_VER_THREE)
412 		return (EINVAL);
413 
414 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
415 		return (EINVAL);
416 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
417 		return (EINVAL);
418 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
419 		return (EINVAL);
420 
421 	/* All supported algorithms require a cipher key. */
422 	if (en->cipher_key_len == 0)
423 		return (EINVAL);
424 
425 	/* No flags are currently supported. */
426 	if (en->flags != 0)
427 		return (EINVAL);
428 
429 	/* Common checks for supported algorithms. */
430 	switch (en->cipher_algorithm) {
431 	case CRYPTO_AES_NIST_GCM_16:
432 		/*
433 		 * auth_algorithm isn't used, but permit GMAC values
434 		 * for compatibility.
435 		 */
436 		switch (en->auth_algorithm) {
437 		case 0:
438 		case CRYPTO_AES_128_NIST_GMAC:
439 		case CRYPTO_AES_192_NIST_GMAC:
440 		case CRYPTO_AES_256_NIST_GMAC:
441 			break;
442 		default:
443 			return (EINVAL);
444 		}
445 		if (en->auth_key_len != 0)
446 			return (EINVAL);
447 		if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
448 			en->iv_len != TLS_AEAD_GCM_LEN) ||
449 		    (en->tls_vminor == TLS_MINOR_VER_THREE &&
450 			en->iv_len != TLS_1_3_GCM_IV_LEN))
451 			return (EINVAL);
452 		break;
453 	case CRYPTO_AES_CBC:
454 		switch (en->auth_algorithm) {
455 		case CRYPTO_SHA1_HMAC:
456 			/*
457 			 * TLS 1.0 requires an implicit IV.  TLS 1.1+
458 			 * all use explicit IVs.
459 			 */
460 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
461 				if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
462 					return (EINVAL);
463 				break;
464 			}
465 
466 			/* FALLTHROUGH */
467 		case CRYPTO_SHA2_256_HMAC:
468 		case CRYPTO_SHA2_384_HMAC:
469 			/* Ignore any supplied IV. */
470 			en->iv_len = 0;
471 			break;
472 		default:
473 			return (EINVAL);
474 		}
475 		if (en->auth_key_len == 0)
476 			return (EINVAL);
477 		break;
478 	default:
479 		return (EINVAL);
480 	}
481 
482 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
483 
484 	counter_u64_add(ktls_offload_active, 1);
485 
486 	refcount_init(&tls->refcount, 1);
487 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
488 
489 	tls->wq_index = ktls_get_cpu(so);
490 
491 	tls->params.cipher_algorithm = en->cipher_algorithm;
492 	tls->params.auth_algorithm = en->auth_algorithm;
493 	tls->params.tls_vmajor = en->tls_vmajor;
494 	tls->params.tls_vminor = en->tls_vminor;
495 	tls->params.flags = en->flags;
496 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
497 
498 	/* Set the header and trailer lengths. */
499 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
500 	switch (en->cipher_algorithm) {
501 	case CRYPTO_AES_NIST_GCM_16:
502 		/*
503 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
504 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
505 		 */
506 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
507 			tls->params.tls_hlen += sizeof(uint64_t);
508 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
509 
510 		/*
511 		 * TLS 1.3 includes optional padding which we
512 		 * do not support, and also puts the "real" record
513 		 * type at the end of the encrypted data.
514 		 */
515 		if (en->tls_vminor == TLS_MINOR_VER_THREE)
516 			tls->params.tls_tlen += sizeof(uint8_t);
517 
518 		tls->params.tls_bs = 1;
519 		break;
520 	case CRYPTO_AES_CBC:
521 		switch (en->auth_algorithm) {
522 		case CRYPTO_SHA1_HMAC:
523 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
524 				/* Implicit IV, no nonce. */
525 			} else {
526 				tls->params.tls_hlen += AES_BLOCK_LEN;
527 			}
528 			tls->params.tls_tlen = AES_BLOCK_LEN +
529 			    SHA1_HASH_LEN;
530 			break;
531 		case CRYPTO_SHA2_256_HMAC:
532 			tls->params.tls_hlen += AES_BLOCK_LEN;
533 			tls->params.tls_tlen = AES_BLOCK_LEN +
534 			    SHA2_256_HASH_LEN;
535 			break;
536 		case CRYPTO_SHA2_384_HMAC:
537 			tls->params.tls_hlen += AES_BLOCK_LEN;
538 			tls->params.tls_tlen = AES_BLOCK_LEN +
539 			    SHA2_384_HASH_LEN;
540 			break;
541 		default:
542 			panic("invalid hmac");
543 		}
544 		tls->params.tls_bs = AES_BLOCK_LEN;
545 		break;
546 	default:
547 		panic("invalid cipher");
548 	}
549 
550 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
551 	    ("TLS header length too long: %d", tls->params.tls_hlen));
552 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
553 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
554 
555 	if (en->auth_key_len != 0) {
556 		tls->params.auth_key_len = en->auth_key_len;
557 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
558 		    M_WAITOK);
559 		error = copyin(en->auth_key, tls->params.auth_key,
560 		    en->auth_key_len);
561 		if (error)
562 			goto out;
563 	}
564 
565 	tls->params.cipher_key_len = en->cipher_key_len;
566 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
567 	error = copyin(en->cipher_key, tls->params.cipher_key,
568 	    en->cipher_key_len);
569 	if (error)
570 		goto out;
571 
572 	/*
573 	 * This holds the implicit portion of the nonce for GCM and
574 	 * the initial implicit IV for TLS 1.0.  The explicit portions
575 	 * of the IV are generated in ktls_frame().
576 	 */
577 	if (en->iv_len != 0) {
578 		tls->params.iv_len = en->iv_len;
579 		error = copyin(en->iv, tls->params.iv, en->iv_len);
580 		if (error)
581 			goto out;
582 
583 		/*
584 		 * For TLS 1.2, generate an 8-byte nonce as a counter
585 		 * to generate unique explicit IVs.
586 		 *
587 		 * Store this counter in the last 8 bytes of the IV
588 		 * array so that it is 8-byte aligned.
589 		 */
590 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
591 		    en->tls_vminor == TLS_MINOR_VER_TWO)
592 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
593 	}
594 
595 	*tlsp = tls;
596 	return (0);
597 
598 out:
599 	ktls_cleanup(tls);
600 	return (error);
601 }
602 
603 static struct ktls_session *
604 ktls_clone_session(struct ktls_session *tls)
605 {
606 	struct ktls_session *tls_new;
607 
608 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
609 
610 	counter_u64_add(ktls_offload_active, 1);
611 
612 	refcount_init(&tls_new->refcount, 1);
613 
614 	/* Copy fields from existing session. */
615 	tls_new->params = tls->params;
616 	tls_new->wq_index = tls->wq_index;
617 
618 	/* Deep copy keys. */
619 	if (tls_new->params.auth_key != NULL) {
620 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
621 		    M_KTLS, M_WAITOK);
622 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
623 		    tls->params.auth_key_len);
624 	}
625 
626 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
627 	    M_WAITOK);
628 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
629 	    tls->params.cipher_key_len);
630 
631 	return (tls_new);
632 }
633 #endif
634 
635 static void
636 ktls_cleanup(struct ktls_session *tls)
637 {
638 
639 	counter_u64_add(ktls_offload_active, -1);
640 	switch (tls->mode) {
641 	case TCP_TLS_MODE_SW:
642 		MPASS(tls->be != NULL);
643 		switch (tls->params.cipher_algorithm) {
644 		case CRYPTO_AES_CBC:
645 			counter_u64_add(ktls_sw_cbc, -1);
646 			break;
647 		case CRYPTO_AES_NIST_GCM_16:
648 			counter_u64_add(ktls_sw_gcm, -1);
649 			break;
650 		}
651 		tls->free(tls);
652 		break;
653 	case TCP_TLS_MODE_IFNET:
654 		switch (tls->params.cipher_algorithm) {
655 		case CRYPTO_AES_CBC:
656 			counter_u64_add(ktls_ifnet_cbc, -1);
657 			break;
658 		case CRYPTO_AES_NIST_GCM_16:
659 			counter_u64_add(ktls_ifnet_gcm, -1);
660 			break;
661 		}
662 		m_snd_tag_rele(tls->snd_tag);
663 		break;
664 #ifdef TCP_OFFLOAD
665 	case TCP_TLS_MODE_TOE:
666 		switch (tls->params.cipher_algorithm) {
667 		case CRYPTO_AES_CBC:
668 			counter_u64_add(ktls_toe_cbc, -1);
669 			break;
670 		case CRYPTO_AES_NIST_GCM_16:
671 			counter_u64_add(ktls_toe_gcm, -1);
672 			break;
673 		}
674 		break;
675 #endif
676 	}
677 	if (tls->params.auth_key != NULL) {
678 		explicit_bzero(tls->params.auth_key, tls->params.auth_key_len);
679 		free(tls->params.auth_key, M_KTLS);
680 		tls->params.auth_key = NULL;
681 		tls->params.auth_key_len = 0;
682 	}
683 	if (tls->params.cipher_key != NULL) {
684 		explicit_bzero(tls->params.cipher_key,
685 		    tls->params.cipher_key_len);
686 		free(tls->params.cipher_key, M_KTLS);
687 		tls->params.cipher_key = NULL;
688 		tls->params.cipher_key_len = 0;
689 	}
690 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
691 }
692 
693 #if defined(INET) || defined(INET6)
694 
695 #ifdef TCP_OFFLOAD
696 static int
697 ktls_try_toe(struct socket *so, struct ktls_session *tls)
698 {
699 	struct inpcb *inp;
700 	struct tcpcb *tp;
701 	int error;
702 
703 	inp = so->so_pcb;
704 	INP_WLOCK(inp);
705 	if (inp->inp_flags2 & INP_FREED) {
706 		INP_WUNLOCK(inp);
707 		return (ECONNRESET);
708 	}
709 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
710 		INP_WUNLOCK(inp);
711 		return (ECONNRESET);
712 	}
713 	if (inp->inp_socket == NULL) {
714 		INP_WUNLOCK(inp);
715 		return (ECONNRESET);
716 	}
717 	tp = intotcpcb(inp);
718 	if (tp->tod == NULL) {
719 		INP_WUNLOCK(inp);
720 		return (EOPNOTSUPP);
721 	}
722 
723 	error = tcp_offload_alloc_tls_session(tp, tls);
724 	INP_WUNLOCK(inp);
725 	if (error == 0) {
726 		tls->mode = TCP_TLS_MODE_TOE;
727 		switch (tls->params.cipher_algorithm) {
728 		case CRYPTO_AES_CBC:
729 			counter_u64_add(ktls_toe_cbc, 1);
730 			break;
731 		case CRYPTO_AES_NIST_GCM_16:
732 			counter_u64_add(ktls_toe_gcm, 1);
733 			break;
734 		}
735 	}
736 	return (error);
737 }
738 #endif
739 
740 /*
741  * Common code used when first enabling ifnet TLS on a connection or
742  * when allocating a new ifnet TLS session due to a routing change.
743  * This function allocates a new TLS send tag on whatever interface
744  * the connection is currently routed over.
745  */
746 static int
747 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
748     struct m_snd_tag **mstp)
749 {
750 	union if_snd_tag_alloc_params params;
751 	struct ifnet *ifp;
752 	struct rtentry *rt;
753 	struct tcpcb *tp;
754 	int error;
755 
756 	INP_RLOCK(inp);
757 	if (inp->inp_flags2 & INP_FREED) {
758 		INP_RUNLOCK(inp);
759 		return (ECONNRESET);
760 	}
761 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
762 		INP_RUNLOCK(inp);
763 		return (ECONNRESET);
764 	}
765 	if (inp->inp_socket == NULL) {
766 		INP_RUNLOCK(inp);
767 		return (ECONNRESET);
768 	}
769 	tp = intotcpcb(inp);
770 
771 	/*
772 	 * Check administrative controls on ifnet TLS to determine if
773 	 * ifnet TLS should be denied.
774 	 *
775 	 * - Always permit 'force' requests.
776 	 * - ktls_ifnet_permitted == 0: always deny.
777 	 */
778 	if (!force && ktls_ifnet_permitted == 0) {
779 		INP_RUNLOCK(inp);
780 		return (ENXIO);
781 	}
782 
783 	/*
784 	 * XXX: Use the cached route in the inpcb to find the
785 	 * interface.  This should perhaps instead use
786 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
787 	 * enabled after a connection has completed key negotiation in
788 	 * userland, the cached route will be present in practice.
789 	 */
790 	rt = inp->inp_route.ro_rt;
791 	if (rt == NULL || rt->rt_ifp == NULL) {
792 		INP_RUNLOCK(inp);
793 		return (ENXIO);
794 	}
795 	ifp = rt->rt_ifp;
796 	if_ref(ifp);
797 
798 	params.hdr.type = IF_SND_TAG_TYPE_TLS;
799 	params.hdr.flowid = inp->inp_flowid;
800 	params.hdr.flowtype = inp->inp_flowtype;
801 	params.tls.inp = inp;
802 	params.tls.tls = tls;
803 	INP_RUNLOCK(inp);
804 
805 	if (ifp->if_snd_tag_alloc == NULL) {
806 		error = EOPNOTSUPP;
807 		goto out;
808 	}
809 	if ((ifp->if_capenable & IFCAP_NOMAP) == 0) {
810 		error = EOPNOTSUPP;
811 		goto out;
812 	}
813 	if (inp->inp_vflag & INP_IPV6) {
814 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
815 			error = EOPNOTSUPP;
816 			goto out;
817 		}
818 	} else {
819 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
820 			error = EOPNOTSUPP;
821 			goto out;
822 		}
823 	}
824 	error = ifp->if_snd_tag_alloc(ifp, &params, mstp);
825 out:
826 	if_rele(ifp);
827 	return (error);
828 }
829 
830 static int
831 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
832 {
833 	struct m_snd_tag *mst;
834 	int error;
835 
836 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
837 	if (error == 0) {
838 		tls->mode = TCP_TLS_MODE_IFNET;
839 		tls->snd_tag = mst;
840 		switch (tls->params.cipher_algorithm) {
841 		case CRYPTO_AES_CBC:
842 			counter_u64_add(ktls_ifnet_cbc, 1);
843 			break;
844 		case CRYPTO_AES_NIST_GCM_16:
845 			counter_u64_add(ktls_ifnet_gcm, 1);
846 			break;
847 		}
848 	}
849 	return (error);
850 }
851 
852 static int
853 ktls_try_sw(struct socket *so, struct ktls_session *tls)
854 {
855 	struct rm_priotracker prio;
856 	struct ktls_crypto_backend *be;
857 
858 	/*
859 	 * Choose the best software crypto backend.  Backends are
860 	 * stored in sorted priority order (larget value == most
861 	 * important at the head of the list), so this just stops on
862 	 * the first backend that claims the session by returning
863 	 * success.
864 	 */
865 	if (ktls_allow_unload)
866 		rm_rlock(&ktls_backends_lock, &prio);
867 	LIST_FOREACH(be, &ktls_backends, next) {
868 		if (be->try(so, tls) == 0)
869 			break;
870 		KASSERT(tls->cipher == NULL,
871 		    ("ktls backend leaked a cipher pointer"));
872 	}
873 	if (be != NULL) {
874 		if (ktls_allow_unload)
875 			be->use_count++;
876 		tls->be = be;
877 	}
878 	if (ktls_allow_unload)
879 		rm_runlock(&ktls_backends_lock, &prio);
880 	if (be == NULL)
881 		return (EOPNOTSUPP);
882 	tls->mode = TCP_TLS_MODE_SW;
883 	switch (tls->params.cipher_algorithm) {
884 	case CRYPTO_AES_CBC:
885 		counter_u64_add(ktls_sw_cbc, 1);
886 		break;
887 	case CRYPTO_AES_NIST_GCM_16:
888 		counter_u64_add(ktls_sw_gcm, 1);
889 		break;
890 	}
891 	return (0);
892 }
893 
894 int
895 ktls_enable_tx(struct socket *so, struct tls_enable *en)
896 {
897 	struct ktls_session *tls;
898 	int error;
899 
900 	if (!ktls_offload_enable)
901 		return (ENOTSUP);
902 
903 	counter_u64_add(ktls_offload_enable_calls, 1);
904 
905 	/*
906 	 * This should always be true since only the TCP socket option
907 	 * invokes this function.
908 	 */
909 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
910 		return (EINVAL);
911 
912 	/*
913 	 * XXX: Don't overwrite existing sessions.  We should permit
914 	 * this to support rekeying in the future.
915 	 */
916 	if (so->so_snd.sb_tls_info != NULL)
917 		return (EALREADY);
918 
919 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
920 		return (ENOTSUP);
921 
922 	/* TLS requires ext pgs */
923 	if (mb_use_ext_pgs == 0)
924 		return (ENXIO);
925 
926 	error = ktls_create_session(so, en, &tls);
927 	if (error)
928 		return (error);
929 
930 	/* Prefer TOE -> ifnet TLS -> software TLS. */
931 #ifdef TCP_OFFLOAD
932 	error = ktls_try_toe(so, tls);
933 	if (error)
934 #endif
935 		error = ktls_try_ifnet(so, tls, false);
936 	if (error)
937 		error = ktls_try_sw(so, tls);
938 
939 	if (error) {
940 		ktls_cleanup(tls);
941 		return (error);
942 	}
943 
944 	error = sblock(&so->so_snd, SBL_WAIT);
945 	if (error) {
946 		ktls_cleanup(tls);
947 		return (error);
948 	}
949 
950 	SOCKBUF_LOCK(&so->so_snd);
951 	so->so_snd.sb_tls_info = tls;
952 	if (tls->mode != TCP_TLS_MODE_SW)
953 		so->so_snd.sb_flags |= SB_TLS_IFNET;
954 	SOCKBUF_UNLOCK(&so->so_snd);
955 	sbunlock(&so->so_snd);
956 
957 	counter_u64_add(ktls_offload_total, 1);
958 
959 	return (0);
960 }
961 
962 int
963 ktls_get_tx_mode(struct socket *so)
964 {
965 	struct ktls_session *tls;
966 	struct inpcb *inp;
967 	int mode;
968 
969 	inp = so->so_pcb;
970 	INP_WLOCK_ASSERT(inp);
971 	SOCKBUF_LOCK(&so->so_snd);
972 	tls = so->so_snd.sb_tls_info;
973 	if (tls == NULL)
974 		mode = TCP_TLS_MODE_NONE;
975 	else
976 		mode = tls->mode;
977 	SOCKBUF_UNLOCK(&so->so_snd);
978 	return (mode);
979 }
980 
981 /*
982  * Switch between SW and ifnet TLS sessions as requested.
983  */
984 int
985 ktls_set_tx_mode(struct socket *so, int mode)
986 {
987 	struct ktls_session *tls, *tls_new;
988 	struct inpcb *inp;
989 	int error;
990 
991 	switch (mode) {
992 	case TCP_TLS_MODE_SW:
993 	case TCP_TLS_MODE_IFNET:
994 		break;
995 	default:
996 		return (EINVAL);
997 	}
998 
999 	inp = so->so_pcb;
1000 	INP_WLOCK_ASSERT(inp);
1001 	SOCKBUF_LOCK(&so->so_snd);
1002 	tls = so->so_snd.sb_tls_info;
1003 	if (tls == NULL) {
1004 		SOCKBUF_UNLOCK(&so->so_snd);
1005 		return (0);
1006 	}
1007 
1008 	if (tls->mode == mode) {
1009 		SOCKBUF_UNLOCK(&so->so_snd);
1010 		return (0);
1011 	}
1012 
1013 	tls = ktls_hold(tls);
1014 	SOCKBUF_UNLOCK(&so->so_snd);
1015 	INP_WUNLOCK(inp);
1016 
1017 	tls_new = ktls_clone_session(tls);
1018 
1019 	if (mode == TCP_TLS_MODE_IFNET)
1020 		error = ktls_try_ifnet(so, tls_new, true);
1021 	else
1022 		error = ktls_try_sw(so, tls_new);
1023 	if (error) {
1024 		counter_u64_add(ktls_switch_failed, 1);
1025 		ktls_free(tls_new);
1026 		ktls_free(tls);
1027 		INP_WLOCK(inp);
1028 		return (error);
1029 	}
1030 
1031 	error = sblock(&so->so_snd, SBL_WAIT);
1032 	if (error) {
1033 		counter_u64_add(ktls_switch_failed, 1);
1034 		ktls_free(tls_new);
1035 		ktls_free(tls);
1036 		INP_WLOCK(inp);
1037 		return (error);
1038 	}
1039 
1040 	/*
1041 	 * If we raced with another session change, keep the existing
1042 	 * session.
1043 	 */
1044 	if (tls != so->so_snd.sb_tls_info) {
1045 		counter_u64_add(ktls_switch_failed, 1);
1046 		sbunlock(&so->so_snd);
1047 		ktls_free(tls_new);
1048 		ktls_free(tls);
1049 		INP_WLOCK(inp);
1050 		return (EBUSY);
1051 	}
1052 
1053 	SOCKBUF_LOCK(&so->so_snd);
1054 	so->so_snd.sb_tls_info = tls_new;
1055 	if (tls_new->mode != TCP_TLS_MODE_SW)
1056 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1057 	SOCKBUF_UNLOCK(&so->so_snd);
1058 	sbunlock(&so->so_snd);
1059 
1060 	/*
1061 	 * Drop two references on 'tls'.  The first is for the
1062 	 * ktls_hold() above.  The second drops the reference from the
1063 	 * socket buffer.
1064 	 */
1065 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1066 	ktls_free(tls);
1067 	ktls_free(tls);
1068 
1069 	if (mode == TCP_TLS_MODE_IFNET)
1070 		counter_u64_add(ktls_switch_to_ifnet, 1);
1071 	else
1072 		counter_u64_add(ktls_switch_to_sw, 1);
1073 
1074 	INP_WLOCK(inp);
1075 	return (0);
1076 }
1077 
1078 /*
1079  * Try to allocate a new TLS send tag.  This task is scheduled when
1080  * ip_output detects a route change while trying to transmit a packet
1081  * holding a TLS record.  If a new tag is allocated, replace the tag
1082  * in the TLS session.  Subsequent packets on the connection will use
1083  * the new tag.  If a new tag cannot be allocated, drop the
1084  * connection.
1085  */
1086 static void
1087 ktls_reset_send_tag(void *context, int pending)
1088 {
1089 	struct epoch_tracker et;
1090 	struct ktls_session *tls;
1091 	struct m_snd_tag *old, *new;
1092 	struct inpcb *inp;
1093 	struct tcpcb *tp;
1094 	int error;
1095 
1096 	MPASS(pending == 1);
1097 
1098 	tls = context;
1099 	inp = tls->inp;
1100 
1101 	/*
1102 	 * Free the old tag first before allocating a new one.
1103 	 * ip[6]_output_send() will treat a NULL send tag the same as
1104 	 * an ifp mismatch and drop packets until a new tag is
1105 	 * allocated.
1106 	 *
1107 	 * Write-lock the INP when changing tls->snd_tag since
1108 	 * ip[6]_output_send() holds a read-lock when reading the
1109 	 * pointer.
1110 	 */
1111 	INP_WLOCK(inp);
1112 	old = tls->snd_tag;
1113 	tls->snd_tag = NULL;
1114 	INP_WUNLOCK(inp);
1115 	if (old != NULL)
1116 		m_snd_tag_rele(old);
1117 
1118 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1119 
1120 	if (error == 0) {
1121 		INP_WLOCK(inp);
1122 		tls->snd_tag = new;
1123 		mtx_pool_lock(mtxpool_sleep, tls);
1124 		tls->reset_pending = false;
1125 		mtx_pool_unlock(mtxpool_sleep, tls);
1126 		if (!in_pcbrele_wlocked(inp))
1127 			INP_WUNLOCK(inp);
1128 
1129 		counter_u64_add(ktls_ifnet_reset, 1);
1130 
1131 		/*
1132 		 * XXX: Should we kick tcp_output explicitly now that
1133 		 * the send tag is fixed or just rely on timers?
1134 		 */
1135 	} else {
1136 		NET_EPOCH_ENTER(et);
1137 		INP_WLOCK(inp);
1138 		if (!in_pcbrele_wlocked(inp)) {
1139 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1140 			    !(inp->inp_flags & INP_DROPPED)) {
1141 				tp = intotcpcb(inp);
1142 				tp = tcp_drop(tp, ECONNABORTED);
1143 				if (tp != NULL)
1144 					INP_WUNLOCK(inp);
1145 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1146 			} else
1147 				INP_WUNLOCK(inp);
1148 		}
1149 		NET_EPOCH_EXIT(et);
1150 
1151 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1152 
1153 		/*
1154 		 * Leave reset_pending true to avoid future tasks while
1155 		 * the socket goes away.
1156 		 */
1157 	}
1158 
1159 	ktls_free(tls);
1160 }
1161 
1162 int
1163 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1164 {
1165 
1166 	if (inp == NULL)
1167 		return (ENOBUFS);
1168 
1169 	INP_LOCK_ASSERT(inp);
1170 
1171 	/*
1172 	 * See if we should schedule a task to update the send tag for
1173 	 * this session.
1174 	 */
1175 	mtx_pool_lock(mtxpool_sleep, tls);
1176 	if (!tls->reset_pending) {
1177 		(void) ktls_hold(tls);
1178 		in_pcbref(inp);
1179 		tls->inp = inp;
1180 		tls->reset_pending = true;
1181 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1182 	}
1183 	mtx_pool_unlock(mtxpool_sleep, tls);
1184 	return (ENOBUFS);
1185 }
1186 #endif
1187 
1188 void
1189 ktls_destroy(struct ktls_session *tls)
1190 {
1191 	struct rm_priotracker prio;
1192 
1193 	ktls_cleanup(tls);
1194 	if (tls->be != NULL && ktls_allow_unload) {
1195 		rm_rlock(&ktls_backends_lock, &prio);
1196 		tls->be->use_count--;
1197 		rm_runlock(&ktls_backends_lock, &prio);
1198 	}
1199 	uma_zfree(ktls_session_zone, tls);
1200 }
1201 
1202 void
1203 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1204 {
1205 	struct mbuf_ext_pgs *pgs;
1206 
1207 	for (; m != NULL; m = m->m_next) {
1208 		KASSERT((m->m_flags & M_NOMAP) != 0,
1209 		    ("ktls_seq: mapped mbuf %p", m));
1210 
1211 		pgs = m->m_ext.ext_pgs;
1212 		pgs->seqno = sb->sb_tls_seqno;
1213 		sb->sb_tls_seqno++;
1214 	}
1215 }
1216 
1217 /*
1218  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1219  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1220  * mbuf must be populated with the payload of each TLS record.
1221  *
1222  * The record_type argument specifies the TLS record type used when
1223  * populating the TLS header.
1224  *
1225  * The enq_count argument on return is set to the number of pages of
1226  * payload data for this entire chain that need to be encrypted via SW
1227  * encryption.  The returned value should be passed to ktls_enqueue
1228  * when scheduling encryption of this chain of mbufs.
1229  */
1230 int
1231 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1232     uint8_t record_type)
1233 {
1234 	struct tls_record_layer *tlshdr;
1235 	struct mbuf *m;
1236 	struct mbuf_ext_pgs *pgs;
1237 	uint64_t *noncep;
1238 	uint16_t tls_len;
1239 	int maxlen;
1240 
1241 	maxlen = tls->params.max_frame_len;
1242 	*enq_cnt = 0;
1243 	for (m = top; m != NULL; m = m->m_next) {
1244 		/*
1245 		 * All mbufs in the chain should be non-empty TLS
1246 		 * records whose payload does not exceed the maximum
1247 		 * frame length.
1248 		 */
1249 		if (m->m_len > maxlen || m->m_len == 0)
1250 			return (EINVAL);
1251 		tls_len = m->m_len;
1252 
1253 		/*
1254 		 * TLS frames require unmapped mbufs to store session
1255 		 * info.
1256 		 */
1257 		KASSERT((m->m_flags & M_NOMAP) != 0,
1258 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1259 
1260 		pgs = m->m_ext.ext_pgs;
1261 
1262 		/* Save a reference to the session. */
1263 		pgs->tls = ktls_hold(tls);
1264 
1265 		pgs->hdr_len = tls->params.tls_hlen;
1266 		pgs->trail_len = tls->params.tls_tlen;
1267 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1268 			int bs, delta;
1269 
1270 			/*
1271 			 * AES-CBC pads messages to a multiple of the
1272 			 * block size.  Note that the padding is
1273 			 * applied after the digest and the encryption
1274 			 * is done on the "plaintext || mac || padding".
1275 			 * At least one byte of padding is always
1276 			 * present.
1277 			 *
1278 			 * Compute the final trailer length assuming
1279 			 * at most one block of padding.
1280 			 * tls->params.sb_tls_tlen is the maximum
1281 			 * possible trailer length (padding + digest).
1282 			 * delta holds the number of excess padding
1283 			 * bytes if the maximum were used.  Those
1284 			 * extra bytes are removed.
1285 			 */
1286 			bs = tls->params.tls_bs;
1287 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1288 			pgs->trail_len -= delta;
1289 		}
1290 		m->m_len += pgs->hdr_len + pgs->trail_len;
1291 
1292 		/* Populate the TLS header. */
1293 		tlshdr = (void *)pgs->hdr;
1294 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1295 
1296 		/*
1297 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1298 		 * of TLS_RLTYPE_APP.
1299 		 */
1300 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1301 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1302 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1303 			tlshdr->tls_type = TLS_RLTYPE_APP;
1304 			/* save the real record type for later */
1305 			pgs->record_type = record_type;
1306 		} else {
1307 			tlshdr->tls_vminor = tls->params.tls_vminor;
1308 			tlshdr->tls_type = record_type;
1309 		}
1310 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1311 
1312 		/*
1313 		 * Store nonces / explicit IVs after the end of the
1314 		 * TLS header.
1315 		 *
1316 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1317 		 * from the end of the IV.  The nonce is then
1318 		 * incremented for use by the next record.
1319 		 *
1320 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1321 		 */
1322 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1323 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1324 			noncep = (uint64_t *)(tls->params.iv + 8);
1325 			be64enc(tlshdr + 1, *noncep);
1326 			(*noncep)++;
1327 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1328 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1329 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1330 
1331 		/*
1332 		 * When using SW encryption, mark the mbuf not ready.
1333 		 * It will be marked ready via sbready() after the
1334 		 * record has been encrypted.
1335 		 *
1336 		 * When using ifnet TLS, unencrypted TLS records are
1337 		 * sent down the stack to the NIC.
1338 		 */
1339 		if (tls->mode == TCP_TLS_MODE_SW) {
1340 			m->m_flags |= M_NOTREADY;
1341 			pgs->nrdy = pgs->npgs;
1342 			*enq_cnt += pgs->npgs;
1343 		}
1344 	}
1345 	return (0);
1346 }
1347 
1348 void
1349 ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs)
1350 {
1351 	struct ktls_wq *wq;
1352 	bool running;
1353 
1354 	/* Mark it for freeing. */
1355 	pgs->mbuf = NULL;
1356 	wq = &ktls_wq[pgs->tls->wq_index];
1357 	mtx_lock(&wq->mtx);
1358 	STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1359 	running = wq->running;
1360 	mtx_unlock(&wq->mtx);
1361 	if (!running)
1362 		wakeup(wq);
1363 }
1364 
1365 void
1366 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1367 {
1368 	struct mbuf_ext_pgs *pgs;
1369 	struct ktls_wq *wq;
1370 	bool running;
1371 
1372 	KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1373 	    (M_NOMAP | M_NOTREADY)),
1374 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1375 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1376 
1377 	pgs = m->m_ext.ext_pgs;
1378 
1379 	KASSERT(pgs->tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1380 
1381 	pgs->enc_cnt = page_count;
1382 	pgs->mbuf = m;
1383 
1384 	/*
1385 	 * Save a pointer to the socket.  The caller is responsible
1386 	 * for taking an additional reference via soref().
1387 	 */
1388 	pgs->so = so;
1389 
1390 	wq = &ktls_wq[pgs->tls->wq_index];
1391 	mtx_lock(&wq->mtx);
1392 	STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1393 	running = wq->running;
1394 	mtx_unlock(&wq->mtx);
1395 	if (!running)
1396 		wakeup(wq);
1397 	counter_u64_add(ktls_cnt_on, 1);
1398 }
1399 
1400 static __noinline void
1401 ktls_encrypt(struct mbuf_ext_pgs *pgs)
1402 {
1403 	struct ktls_session *tls;
1404 	struct socket *so;
1405 	struct mbuf *m, *top;
1406 	vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1407 	struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1408 	struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1409 	vm_page_t pg;
1410 	int error, i, len, npages, off, total_pages;
1411 	bool is_anon;
1412 
1413 	so = pgs->so;
1414 	tls = pgs->tls;
1415 	top = pgs->mbuf;
1416 	KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs));
1417 	KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs));
1418 #ifdef INVARIANTS
1419 	pgs->so = NULL;
1420 	pgs->mbuf = NULL;
1421 #endif
1422 	total_pages = pgs->enc_cnt;
1423 	npages = 0;
1424 
1425 	/*
1426 	 * Encrypt the TLS records in the chain of mbufs starting with
1427 	 * 'top'.  'total_pages' gives us a total count of pages and is
1428 	 * used to know when we have finished encrypting the TLS
1429 	 * records originally queued with 'top'.
1430 	 *
1431 	 * NB: These mbufs are queued in the socket buffer and
1432 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
1433 	 * socket buffer lock is not held while traversing this chain.
1434 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1435 	 * pointers should be stable.  However, the 'm_next' of the
1436 	 * last mbuf encrypted is not necessarily NULL.  It can point
1437 	 * to other mbufs appended while 'top' was on the TLS work
1438 	 * queue.
1439 	 *
1440 	 * Each mbuf holds an entire TLS record.
1441 	 */
1442 	error = 0;
1443 	for (m = top; npages != total_pages; m = m->m_next) {
1444 		pgs = m->m_ext.ext_pgs;
1445 
1446 		KASSERT(pgs->tls == tls,
1447 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
1448 		    tls, pgs->tls));
1449 		KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1450 		    (M_NOMAP | M_NOTREADY),
1451 		    ("%p not unready & nomap mbuf (top = %p)\n", m, top));
1452 		KASSERT(npages + pgs->npgs <= total_pages,
1453 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
1454 		    total_pages, m));
1455 
1456 		/*
1457 		 * Generate source and destination ivoecs to pass to
1458 		 * the SW encryption backend.  For writable mbufs, the
1459 		 * destination iovec is a copy of the source and
1460 		 * encryption is done in place.  For file-backed mbufs
1461 		 * (from sendfile), anonymous wired pages are
1462 		 * allocated and assigned to the destination iovec.
1463 		 */
1464 		is_anon = (pgs->flags & MBUF_PEXT_FLAG_ANON) != 0;
1465 
1466 		off = pgs->first_pg_off;
1467 		for (i = 0; i < pgs->npgs; i++, off = 0) {
1468 			len = mbuf_ext_pg_len(pgs, i, off);
1469 			src_iov[i].iov_len = len;
1470 			src_iov[i].iov_base =
1471 			    (char *)(void *)PHYS_TO_DMAP(pgs->pa[i]) + off;
1472 
1473 			if (is_anon) {
1474 				dst_iov[i].iov_base = src_iov[i].iov_base;
1475 				dst_iov[i].iov_len = src_iov[i].iov_len;
1476 				continue;
1477 			}
1478 retry_page:
1479 			pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
1480 			    VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
1481 			if (pg == NULL) {
1482 				vm_wait(NULL);
1483 				goto retry_page;
1484 			}
1485 			parray[i] = VM_PAGE_TO_PHYS(pg);
1486 			dst_iov[i].iov_base =
1487 			    (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
1488 			dst_iov[i].iov_len = len;
1489 		}
1490 
1491 		npages += i;
1492 
1493 		error = (*tls->sw_encrypt)(tls,
1494 		    (const struct tls_record_layer *)pgs->hdr,
1495 		    pgs->trail, src_iov, dst_iov, i, pgs->seqno,
1496 		    pgs->record_type);
1497 		if (error) {
1498 			counter_u64_add(ktls_offload_failed_crypto, 1);
1499 			break;
1500 		}
1501 
1502 		/*
1503 		 * For file-backed mbufs, release the file-backed
1504 		 * pages and replace them in the ext_pgs array with
1505 		 * the anonymous wired pages allocated above.
1506 		 */
1507 		if (!is_anon) {
1508 			/* Free the old pages. */
1509 			m->m_ext.ext_free(m);
1510 
1511 			/* Replace them with the new pages. */
1512 			for (i = 0; i < pgs->npgs; i++)
1513 				pgs->pa[i] = parray[i];
1514 
1515 			/* Use the basic free routine. */
1516 			m->m_ext.ext_free = mb_free_mext_pgs;
1517 
1518 			/* Pages are now writable. */
1519 			pgs->flags |= MBUF_PEXT_FLAG_ANON;
1520 		}
1521 
1522 		/*
1523 		 * Drop a reference to the session now that it is no
1524 		 * longer needed.  Existing code depends on encrypted
1525 		 * records having no associated session vs
1526 		 * yet-to-be-encrypted records having an associated
1527 		 * session.
1528 		 */
1529 		pgs->tls = NULL;
1530 		ktls_free(tls);
1531 	}
1532 
1533 	CURVNET_SET(so->so_vnet);
1534 	if (error == 0) {
1535 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
1536 	} else {
1537 		so->so_proto->pr_usrreqs->pru_abort(so);
1538 		so->so_error = EIO;
1539 		mb_free_notready(top, total_pages);
1540 	}
1541 
1542 	SOCK_LOCK(so);
1543 	sorele(so);
1544 	CURVNET_RESTORE();
1545 }
1546 
1547 static void
1548 ktls_work_thread(void *ctx)
1549 {
1550 	struct ktls_wq *wq = ctx;
1551 	struct mbuf_ext_pgs *p, *n;
1552 	struct ktls_session *tls;
1553 	STAILQ_HEAD(, mbuf_ext_pgs) local_head;
1554 
1555 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
1556 	fpu_kern_thread(0);
1557 #endif
1558 	for (;;) {
1559 		mtx_lock(&wq->mtx);
1560 		while (STAILQ_EMPTY(&wq->head)) {
1561 			wq->running = false;
1562 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
1563 			wq->running = true;
1564 		}
1565 
1566 		STAILQ_INIT(&local_head);
1567 		STAILQ_CONCAT(&local_head, &wq->head);
1568 		mtx_unlock(&wq->mtx);
1569 
1570 		STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) {
1571 			if (p->mbuf != NULL) {
1572 				ktls_encrypt(p);
1573 				counter_u64_add(ktls_cnt_on, -1);
1574 			} else {
1575 				tls = p->tls;
1576 				ktls_free(tls);
1577 				uma_zfree(zone_extpgs, p);
1578 			}
1579 		}
1580 	}
1581 }
1582