xref: /freebsd/sys/kern/uipc_ktls.c (revision e2eeea75eb8b6dd50c1298067a0655880d186734)
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 	/*
818 	 * Allocate a TLS + ratelimit tag if the connection has an
819 	 * existing pacing rate.
820 	 */
821 	if (tp->t_pacing_rate != -1 &&
822 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
823 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
824 		params.tls_rate_limit.inp = inp;
825 		params.tls_rate_limit.tls = tls;
826 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
827 	} else {
828 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
829 		params.tls.inp = inp;
830 		params.tls.tls = tls;
831 	}
832 	params.hdr.flowid = inp->inp_flowid;
833 	params.hdr.flowtype = inp->inp_flowtype;
834 	params.hdr.numa_domain = inp->inp_numa_domain;
835 	INP_RUNLOCK(inp);
836 
837 	if ((ifp->if_capenable & IFCAP_NOMAP) == 0) {
838 		error = EOPNOTSUPP;
839 		goto out;
840 	}
841 	if (inp->inp_vflag & INP_IPV6) {
842 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
843 			error = EOPNOTSUPP;
844 			goto out;
845 		}
846 	} else {
847 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
848 			error = EOPNOTSUPP;
849 			goto out;
850 		}
851 	}
852 	error = m_snd_tag_alloc(ifp, &params, mstp);
853 out:
854 	if_rele(ifp);
855 	return (error);
856 }
857 
858 static int
859 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
860 {
861 	struct m_snd_tag *mst;
862 	int error;
863 
864 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
865 	if (error == 0) {
866 		tls->mode = TCP_TLS_MODE_IFNET;
867 		tls->snd_tag = mst;
868 		switch (tls->params.cipher_algorithm) {
869 		case CRYPTO_AES_CBC:
870 			counter_u64_add(ktls_ifnet_cbc, 1);
871 			break;
872 		case CRYPTO_AES_NIST_GCM_16:
873 			counter_u64_add(ktls_ifnet_gcm, 1);
874 			break;
875 		}
876 	}
877 	return (error);
878 }
879 
880 static int
881 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
882 {
883 	struct rm_priotracker prio;
884 	struct ktls_crypto_backend *be;
885 
886 	/*
887 	 * Choose the best software crypto backend.  Backends are
888 	 * stored in sorted priority order (larget value == most
889 	 * important at the head of the list), so this just stops on
890 	 * the first backend that claims the session by returning
891 	 * success.
892 	 */
893 	if (ktls_allow_unload)
894 		rm_rlock(&ktls_backends_lock, &prio);
895 	LIST_FOREACH(be, &ktls_backends, next) {
896 		if (be->try(so, tls, direction) == 0)
897 			break;
898 		KASSERT(tls->cipher == NULL,
899 		    ("ktls backend leaked a cipher pointer"));
900 	}
901 	if (be != NULL) {
902 		if (ktls_allow_unload)
903 			be->use_count++;
904 		tls->be = be;
905 	}
906 	if (ktls_allow_unload)
907 		rm_runlock(&ktls_backends_lock, &prio);
908 	if (be == NULL)
909 		return (EOPNOTSUPP);
910 	tls->mode = TCP_TLS_MODE_SW;
911 	switch (tls->params.cipher_algorithm) {
912 	case CRYPTO_AES_CBC:
913 		counter_u64_add(ktls_sw_cbc, 1);
914 		break;
915 	case CRYPTO_AES_NIST_GCM_16:
916 		counter_u64_add(ktls_sw_gcm, 1);
917 		break;
918 	}
919 	return (0);
920 }
921 
922 /*
923  * KTLS RX stores data in the socket buffer as a list of TLS records,
924  * where each record is stored as a control message containg the TLS
925  * header followed by data mbufs containing the decrypted data.  This
926  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
927  * both encrypted and decrypted data.  TLS records decrypted by a NIC
928  * should be queued to the socket buffer as records, but encrypted
929  * data which needs to be decrypted by software arrives as a stream of
930  * regular mbufs which need to be converted.  In addition, there may
931  * already be pending encrypted data in the socket buffer when KTLS RX
932  * is enabled.
933  *
934  * To manage not-yet-decrypted data for KTLS RX, the following scheme
935  * is used:
936  *
937  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
938  *
939  * - ktls_check_rx checks this chain of mbufs reading the TLS header
940  *   from the first mbuf.  Once all of the data for that TLS record is
941  *   queued, the socket is queued to a worker thread.
942  *
943  * - The worker thread calls ktls_decrypt to decrypt TLS records in
944  *   the TLS chain.  Each TLS record is detached from the TLS chain,
945  *   decrypted, and inserted into the regular socket buffer chain as
946  *   record starting with a control message holding the TLS header and
947  *   a chain of mbufs holding the encrypted data.
948  */
949 
950 static void
951 sb_mark_notready(struct sockbuf *sb)
952 {
953 	struct mbuf *m;
954 
955 	m = sb->sb_mb;
956 	sb->sb_mtls = m;
957 	sb->sb_mb = NULL;
958 	sb->sb_mbtail = NULL;
959 	sb->sb_lastrecord = NULL;
960 	for (; m != NULL; m = m->m_next) {
961 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
962 		    __func__));
963 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
964 		    __func__));
965 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
966 		    __func__));
967 		m->m_flags |= M_NOTREADY;
968 		sb->sb_acc -= m->m_len;
969 		sb->sb_tlscc += m->m_len;
970 		sb->sb_mtlstail = m;
971 	}
972 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
973 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
974 	    sb->sb_ccc));
975 }
976 
977 int
978 ktls_enable_rx(struct socket *so, struct tls_enable *en)
979 {
980 	struct ktls_session *tls;
981 	int error;
982 
983 	if (!ktls_offload_enable)
984 		return (ENOTSUP);
985 
986 	counter_u64_add(ktls_offload_enable_calls, 1);
987 
988 	/*
989 	 * This should always be true since only the TCP socket option
990 	 * invokes this function.
991 	 */
992 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
993 		return (EINVAL);
994 
995 	/*
996 	 * XXX: Don't overwrite existing sessions.  We should permit
997 	 * this to support rekeying in the future.
998 	 */
999 	if (so->so_rcv.sb_tls_info != NULL)
1000 		return (EALREADY);
1001 
1002 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1003 		return (ENOTSUP);
1004 
1005 	/* TLS 1.3 is not yet supported. */
1006 	if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1007 	    en->tls_vminor == TLS_MINOR_VER_THREE)
1008 		return (ENOTSUP);
1009 
1010 	error = ktls_create_session(so, en, &tls);
1011 	if (error)
1012 		return (error);
1013 
1014 #ifdef TCP_OFFLOAD
1015 	error = ktls_try_toe(so, tls, KTLS_RX);
1016 	if (error)
1017 #endif
1018 		error = ktls_try_sw(so, tls, KTLS_RX);
1019 
1020 	if (error) {
1021 		ktls_cleanup(tls);
1022 		return (error);
1023 	}
1024 
1025 	/* Mark the socket as using TLS offload. */
1026 	SOCKBUF_LOCK(&so->so_rcv);
1027 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1028 	so->so_rcv.sb_tls_info = tls;
1029 	so->so_rcv.sb_flags |= SB_TLS_RX;
1030 
1031 	/* Mark existing data as not ready until it can be decrypted. */
1032 	sb_mark_notready(&so->so_rcv);
1033 	ktls_check_rx(&so->so_rcv);
1034 	SOCKBUF_UNLOCK(&so->so_rcv);
1035 
1036 	counter_u64_add(ktls_offload_total, 1);
1037 
1038 	return (0);
1039 }
1040 
1041 int
1042 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1043 {
1044 	struct ktls_session *tls;
1045 	struct inpcb *inp;
1046 	int error;
1047 
1048 	if (!ktls_offload_enable)
1049 		return (ENOTSUP);
1050 
1051 	counter_u64_add(ktls_offload_enable_calls, 1);
1052 
1053 	/*
1054 	 * This should always be true since only the TCP socket option
1055 	 * invokes this function.
1056 	 */
1057 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1058 		return (EINVAL);
1059 
1060 	/*
1061 	 * XXX: Don't overwrite existing sessions.  We should permit
1062 	 * this to support rekeying in the future.
1063 	 */
1064 	if (so->so_snd.sb_tls_info != NULL)
1065 		return (EALREADY);
1066 
1067 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1068 		return (ENOTSUP);
1069 
1070 	/* TLS requires ext pgs */
1071 	if (mb_use_ext_pgs == 0)
1072 		return (ENXIO);
1073 
1074 	error = ktls_create_session(so, en, &tls);
1075 	if (error)
1076 		return (error);
1077 
1078 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1079 #ifdef TCP_OFFLOAD
1080 	error = ktls_try_toe(so, tls, KTLS_TX);
1081 	if (error)
1082 #endif
1083 		error = ktls_try_ifnet(so, tls, false);
1084 	if (error)
1085 		error = ktls_try_sw(so, tls, KTLS_TX);
1086 
1087 	if (error) {
1088 		ktls_cleanup(tls);
1089 		return (error);
1090 	}
1091 
1092 	error = sblock(&so->so_snd, SBL_WAIT);
1093 	if (error) {
1094 		ktls_cleanup(tls);
1095 		return (error);
1096 	}
1097 
1098 	/*
1099 	 * Write lock the INP when setting sb_tls_info so that
1100 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1101 	 * holding the INP lock.
1102 	 */
1103 	inp = so->so_pcb;
1104 	INP_WLOCK(inp);
1105 	SOCKBUF_LOCK(&so->so_snd);
1106 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1107 	so->so_snd.sb_tls_info = tls;
1108 	if (tls->mode != TCP_TLS_MODE_SW)
1109 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1110 	SOCKBUF_UNLOCK(&so->so_snd);
1111 	INP_WUNLOCK(inp);
1112 	sbunlock(&so->so_snd);
1113 
1114 	counter_u64_add(ktls_offload_total, 1);
1115 
1116 	return (0);
1117 }
1118 
1119 int
1120 ktls_get_rx_mode(struct socket *so)
1121 {
1122 	struct ktls_session *tls;
1123 	struct inpcb *inp;
1124 	int mode;
1125 
1126 	inp = so->so_pcb;
1127 	INP_WLOCK_ASSERT(inp);
1128 	SOCKBUF_LOCK(&so->so_rcv);
1129 	tls = so->so_rcv.sb_tls_info;
1130 	if (tls == NULL)
1131 		mode = TCP_TLS_MODE_NONE;
1132 	else
1133 		mode = tls->mode;
1134 	SOCKBUF_UNLOCK(&so->so_rcv);
1135 	return (mode);
1136 }
1137 
1138 int
1139 ktls_get_tx_mode(struct socket *so)
1140 {
1141 	struct ktls_session *tls;
1142 	struct inpcb *inp;
1143 	int mode;
1144 
1145 	inp = so->so_pcb;
1146 	INP_WLOCK_ASSERT(inp);
1147 	SOCKBUF_LOCK(&so->so_snd);
1148 	tls = so->so_snd.sb_tls_info;
1149 	if (tls == NULL)
1150 		mode = TCP_TLS_MODE_NONE;
1151 	else
1152 		mode = tls->mode;
1153 	SOCKBUF_UNLOCK(&so->so_snd);
1154 	return (mode);
1155 }
1156 
1157 /*
1158  * Switch between SW and ifnet TLS sessions as requested.
1159  */
1160 int
1161 ktls_set_tx_mode(struct socket *so, int mode)
1162 {
1163 	struct ktls_session *tls, *tls_new;
1164 	struct inpcb *inp;
1165 	int error;
1166 
1167 	switch (mode) {
1168 	case TCP_TLS_MODE_SW:
1169 	case TCP_TLS_MODE_IFNET:
1170 		break;
1171 	default:
1172 		return (EINVAL);
1173 	}
1174 
1175 	inp = so->so_pcb;
1176 	INP_WLOCK_ASSERT(inp);
1177 	SOCKBUF_LOCK(&so->so_snd);
1178 	tls = so->so_snd.sb_tls_info;
1179 	if (tls == NULL) {
1180 		SOCKBUF_UNLOCK(&so->so_snd);
1181 		return (0);
1182 	}
1183 
1184 	if (tls->mode == mode) {
1185 		SOCKBUF_UNLOCK(&so->so_snd);
1186 		return (0);
1187 	}
1188 
1189 	tls = ktls_hold(tls);
1190 	SOCKBUF_UNLOCK(&so->so_snd);
1191 	INP_WUNLOCK(inp);
1192 
1193 	tls_new = ktls_clone_session(tls);
1194 
1195 	if (mode == TCP_TLS_MODE_IFNET)
1196 		error = ktls_try_ifnet(so, tls_new, true);
1197 	else
1198 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1199 	if (error) {
1200 		counter_u64_add(ktls_switch_failed, 1);
1201 		ktls_free(tls_new);
1202 		ktls_free(tls);
1203 		INP_WLOCK(inp);
1204 		return (error);
1205 	}
1206 
1207 	error = sblock(&so->so_snd, SBL_WAIT);
1208 	if (error) {
1209 		counter_u64_add(ktls_switch_failed, 1);
1210 		ktls_free(tls_new);
1211 		ktls_free(tls);
1212 		INP_WLOCK(inp);
1213 		return (error);
1214 	}
1215 
1216 	/*
1217 	 * If we raced with another session change, keep the existing
1218 	 * session.
1219 	 */
1220 	if (tls != so->so_snd.sb_tls_info) {
1221 		counter_u64_add(ktls_switch_failed, 1);
1222 		sbunlock(&so->so_snd);
1223 		ktls_free(tls_new);
1224 		ktls_free(tls);
1225 		INP_WLOCK(inp);
1226 		return (EBUSY);
1227 	}
1228 
1229 	SOCKBUF_LOCK(&so->so_snd);
1230 	so->so_snd.sb_tls_info = tls_new;
1231 	if (tls_new->mode != TCP_TLS_MODE_SW)
1232 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1233 	SOCKBUF_UNLOCK(&so->so_snd);
1234 	sbunlock(&so->so_snd);
1235 
1236 	/*
1237 	 * Drop two references on 'tls'.  The first is for the
1238 	 * ktls_hold() above.  The second drops the reference from the
1239 	 * socket buffer.
1240 	 */
1241 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1242 	ktls_free(tls);
1243 	ktls_free(tls);
1244 
1245 	if (mode == TCP_TLS_MODE_IFNET)
1246 		counter_u64_add(ktls_switch_to_ifnet, 1);
1247 	else
1248 		counter_u64_add(ktls_switch_to_sw, 1);
1249 
1250 	INP_WLOCK(inp);
1251 	return (0);
1252 }
1253 
1254 /*
1255  * Try to allocate a new TLS send tag.  This task is scheduled when
1256  * ip_output detects a route change while trying to transmit a packet
1257  * holding a TLS record.  If a new tag is allocated, replace the tag
1258  * in the TLS session.  Subsequent packets on the connection will use
1259  * the new tag.  If a new tag cannot be allocated, drop the
1260  * connection.
1261  */
1262 static void
1263 ktls_reset_send_tag(void *context, int pending)
1264 {
1265 	struct epoch_tracker et;
1266 	struct ktls_session *tls;
1267 	struct m_snd_tag *old, *new;
1268 	struct inpcb *inp;
1269 	struct tcpcb *tp;
1270 	int error;
1271 
1272 	MPASS(pending == 1);
1273 
1274 	tls = context;
1275 	inp = tls->inp;
1276 
1277 	/*
1278 	 * Free the old tag first before allocating a new one.
1279 	 * ip[6]_output_send() will treat a NULL send tag the same as
1280 	 * an ifp mismatch and drop packets until a new tag is
1281 	 * allocated.
1282 	 *
1283 	 * Write-lock the INP when changing tls->snd_tag since
1284 	 * ip[6]_output_send() holds a read-lock when reading the
1285 	 * pointer.
1286 	 */
1287 	INP_WLOCK(inp);
1288 	old = tls->snd_tag;
1289 	tls->snd_tag = NULL;
1290 	INP_WUNLOCK(inp);
1291 	if (old != NULL)
1292 		m_snd_tag_rele(old);
1293 
1294 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1295 
1296 	if (error == 0) {
1297 		INP_WLOCK(inp);
1298 		tls->snd_tag = new;
1299 		mtx_pool_lock(mtxpool_sleep, tls);
1300 		tls->reset_pending = false;
1301 		mtx_pool_unlock(mtxpool_sleep, tls);
1302 		if (!in_pcbrele_wlocked(inp))
1303 			INP_WUNLOCK(inp);
1304 
1305 		counter_u64_add(ktls_ifnet_reset, 1);
1306 
1307 		/*
1308 		 * XXX: Should we kick tcp_output explicitly now that
1309 		 * the send tag is fixed or just rely on timers?
1310 		 */
1311 	} else {
1312 		NET_EPOCH_ENTER(et);
1313 		INP_WLOCK(inp);
1314 		if (!in_pcbrele_wlocked(inp)) {
1315 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1316 			    !(inp->inp_flags & INP_DROPPED)) {
1317 				tp = intotcpcb(inp);
1318 				CURVNET_SET(tp->t_vnet);
1319 				tp = tcp_drop(tp, ECONNABORTED);
1320 				CURVNET_RESTORE();
1321 				if (tp != NULL)
1322 					INP_WUNLOCK(inp);
1323 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1324 			} else
1325 				INP_WUNLOCK(inp);
1326 		}
1327 		NET_EPOCH_EXIT(et);
1328 
1329 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1330 
1331 		/*
1332 		 * Leave reset_pending true to avoid future tasks while
1333 		 * the socket goes away.
1334 		 */
1335 	}
1336 
1337 	ktls_free(tls);
1338 }
1339 
1340 int
1341 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1342 {
1343 
1344 	if (inp == NULL)
1345 		return (ENOBUFS);
1346 
1347 	INP_LOCK_ASSERT(inp);
1348 
1349 	/*
1350 	 * See if we should schedule a task to update the send tag for
1351 	 * this session.
1352 	 */
1353 	mtx_pool_lock(mtxpool_sleep, tls);
1354 	if (!tls->reset_pending) {
1355 		(void) ktls_hold(tls);
1356 		in_pcbref(inp);
1357 		tls->inp = inp;
1358 		tls->reset_pending = true;
1359 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1360 	}
1361 	mtx_pool_unlock(mtxpool_sleep, tls);
1362 	return (ENOBUFS);
1363 }
1364 
1365 #ifdef RATELIMIT
1366 int
1367 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1368 {
1369 	union if_snd_tag_modify_params params = {
1370 		.rate_limit.max_rate = max_pacing_rate,
1371 		.rate_limit.flags = M_NOWAIT,
1372 	};
1373 	struct m_snd_tag *mst;
1374 	struct ifnet *ifp;
1375 	int error;
1376 
1377 	/* Can't get to the inp, but it should be locked. */
1378 	/* INP_LOCK_ASSERT(inp); */
1379 
1380 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1381 
1382 	if (tls->snd_tag == NULL) {
1383 		/*
1384 		 * Resetting send tag, ignore this change.  The
1385 		 * pending reset may or may not see this updated rate
1386 		 * in the tcpcb.  If it doesn't, we will just lose
1387 		 * this rate change.
1388 		 */
1389 		return (0);
1390 	}
1391 
1392 	MPASS(tls->snd_tag != NULL);
1393 	MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1394 
1395 	mst = tls->snd_tag;
1396 	ifp = mst->ifp;
1397 	return (ifp->if_snd_tag_modify(mst, &params));
1398 }
1399 #endif
1400 #endif
1401 
1402 void
1403 ktls_destroy(struct ktls_session *tls)
1404 {
1405 	struct rm_priotracker prio;
1406 
1407 	ktls_cleanup(tls);
1408 	if (tls->be != NULL && ktls_allow_unload) {
1409 		rm_rlock(&ktls_backends_lock, &prio);
1410 		tls->be->use_count--;
1411 		rm_runlock(&ktls_backends_lock, &prio);
1412 	}
1413 	uma_zfree(ktls_session_zone, tls);
1414 }
1415 
1416 void
1417 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1418 {
1419 
1420 	for (; m != NULL; m = m->m_next) {
1421 		KASSERT((m->m_flags & M_EXTPG) != 0,
1422 		    ("ktls_seq: mapped mbuf %p", m));
1423 
1424 		m->m_epg_seqno = sb->sb_tls_seqno;
1425 		sb->sb_tls_seqno++;
1426 	}
1427 }
1428 
1429 /*
1430  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1431  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1432  * mbuf must be populated with the payload of each TLS record.
1433  *
1434  * The record_type argument specifies the TLS record type used when
1435  * populating the TLS header.
1436  *
1437  * The enq_count argument on return is set to the number of pages of
1438  * payload data for this entire chain that need to be encrypted via SW
1439  * encryption.  The returned value should be passed to ktls_enqueue
1440  * when scheduling encryption of this chain of mbufs.  To handle the
1441  * special case of empty fragments for TLS 1.0 sessions, an empty
1442  * fragment counts as one page.
1443  */
1444 void
1445 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1446     uint8_t record_type)
1447 {
1448 	struct tls_record_layer *tlshdr;
1449 	struct mbuf *m;
1450 	uint64_t *noncep;
1451 	uint16_t tls_len;
1452 	int maxlen;
1453 
1454 	maxlen = tls->params.max_frame_len;
1455 	*enq_cnt = 0;
1456 	for (m = top; m != NULL; m = m->m_next) {
1457 		/*
1458 		 * All mbufs in the chain should be TLS records whose
1459 		 * payload does not exceed the maximum frame length.
1460 		 *
1461 		 * Empty TLS records are permitted when using CBC.
1462 		 */
1463 		KASSERT(m->m_len <= maxlen &&
1464 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1465 		    m->m_len >= 0 : m->m_len > 0),
1466 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1467 
1468 		/*
1469 		 * TLS frames require unmapped mbufs to store session
1470 		 * info.
1471 		 */
1472 		KASSERT((m->m_flags & M_EXTPG) != 0,
1473 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1474 
1475 		tls_len = m->m_len;
1476 
1477 		/* Save a reference to the session. */
1478 		m->m_epg_tls = ktls_hold(tls);
1479 
1480 		m->m_epg_hdrlen = tls->params.tls_hlen;
1481 		m->m_epg_trllen = tls->params.tls_tlen;
1482 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1483 			int bs, delta;
1484 
1485 			/*
1486 			 * AES-CBC pads messages to a multiple of the
1487 			 * block size.  Note that the padding is
1488 			 * applied after the digest and the encryption
1489 			 * is done on the "plaintext || mac || padding".
1490 			 * At least one byte of padding is always
1491 			 * present.
1492 			 *
1493 			 * Compute the final trailer length assuming
1494 			 * at most one block of padding.
1495 			 * tls->params.sb_tls_tlen is the maximum
1496 			 * possible trailer length (padding + digest).
1497 			 * delta holds the number of excess padding
1498 			 * bytes if the maximum were used.  Those
1499 			 * extra bytes are removed.
1500 			 */
1501 			bs = tls->params.tls_bs;
1502 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1503 			m->m_epg_trllen -= delta;
1504 		}
1505 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1506 
1507 		/* Populate the TLS header. */
1508 		tlshdr = (void *)m->m_epg_hdr;
1509 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1510 
1511 		/*
1512 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1513 		 * of TLS_RLTYPE_APP.
1514 		 */
1515 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1516 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1517 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1518 			tlshdr->tls_type = TLS_RLTYPE_APP;
1519 			/* save the real record type for later */
1520 			m->m_epg_record_type = record_type;
1521 			m->m_epg_trail[0] = record_type;
1522 		} else {
1523 			tlshdr->tls_vminor = tls->params.tls_vminor;
1524 			tlshdr->tls_type = record_type;
1525 		}
1526 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1527 
1528 		/*
1529 		 * Store nonces / explicit IVs after the end of the
1530 		 * TLS header.
1531 		 *
1532 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1533 		 * from the end of the IV.  The nonce is then
1534 		 * incremented for use by the next record.
1535 		 *
1536 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1537 		 */
1538 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1539 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1540 			noncep = (uint64_t *)(tls->params.iv + 8);
1541 			be64enc(tlshdr + 1, *noncep);
1542 			(*noncep)++;
1543 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1544 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1545 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1546 
1547 		/*
1548 		 * When using SW encryption, mark the mbuf not ready.
1549 		 * It will be marked ready via sbready() after the
1550 		 * record has been encrypted.
1551 		 *
1552 		 * When using ifnet TLS, unencrypted TLS records are
1553 		 * sent down the stack to the NIC.
1554 		 */
1555 		if (tls->mode == TCP_TLS_MODE_SW) {
1556 			m->m_flags |= M_NOTREADY;
1557 			m->m_epg_nrdy = m->m_epg_npgs;
1558 			if (__predict_false(tls_len == 0)) {
1559 				/* TLS 1.0 empty fragment. */
1560 				*enq_cnt += 1;
1561 			} else
1562 				*enq_cnt += m->m_epg_npgs;
1563 		}
1564 	}
1565 }
1566 
1567 void
1568 ktls_check_rx(struct sockbuf *sb)
1569 {
1570 	struct tls_record_layer hdr;
1571 	struct ktls_wq *wq;
1572 	struct socket *so;
1573 	bool running;
1574 
1575 	SOCKBUF_LOCK_ASSERT(sb);
1576 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1577 	    __func__, sb));
1578 	so = __containerof(sb, struct socket, so_rcv);
1579 
1580 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1581 		return;
1582 
1583 	/* Is there enough queued for a TLS header? */
1584 	if (sb->sb_tlscc < sizeof(hdr)) {
1585 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1586 			so->so_error = EMSGSIZE;
1587 		return;
1588 	}
1589 
1590 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1591 
1592 	/* Is the entire record queued? */
1593 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1594 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1595 			so->so_error = EMSGSIZE;
1596 		return;
1597 	}
1598 
1599 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1600 
1601 	soref(so);
1602 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1603 	mtx_lock(&wq->mtx);
1604 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1605 	running = wq->running;
1606 	mtx_unlock(&wq->mtx);
1607 	if (!running)
1608 		wakeup(wq);
1609 	counter_u64_add(ktls_cnt_rx_queued, 1);
1610 }
1611 
1612 static struct mbuf *
1613 ktls_detach_record(struct sockbuf *sb, int len)
1614 {
1615 	struct mbuf *m, *n, *top;
1616 	int remain;
1617 
1618 	SOCKBUF_LOCK_ASSERT(sb);
1619 	MPASS(len <= sb->sb_tlscc);
1620 
1621 	/*
1622 	 * If TLS chain is the exact size of the record,
1623 	 * just grab the whole record.
1624 	 */
1625 	top = sb->sb_mtls;
1626 	if (sb->sb_tlscc == len) {
1627 		sb->sb_mtls = NULL;
1628 		sb->sb_mtlstail = NULL;
1629 		goto out;
1630 	}
1631 
1632 	/*
1633 	 * While it would be nice to use m_split() here, we need
1634 	 * to know exactly what m_split() allocates to update the
1635 	 * accounting, so do it inline instead.
1636 	 */
1637 	remain = len;
1638 	for (m = top; remain > m->m_len; m = m->m_next)
1639 		remain -= m->m_len;
1640 
1641 	/* Easy case: don't have to split 'm'. */
1642 	if (remain == m->m_len) {
1643 		sb->sb_mtls = m->m_next;
1644 		if (sb->sb_mtls == NULL)
1645 			sb->sb_mtlstail = NULL;
1646 		m->m_next = NULL;
1647 		goto out;
1648 	}
1649 
1650 	/*
1651 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1652 	 * with M_NOWAIT first.
1653 	 */
1654 	n = m_get(M_NOWAIT, MT_DATA);
1655 	if (n == NULL) {
1656 		/*
1657 		 * Use M_WAITOK with socket buffer unlocked.  If
1658 		 * 'sb_mtls' changes while the lock is dropped, return
1659 		 * NULL to force the caller to retry.
1660 		 */
1661 		SOCKBUF_UNLOCK(sb);
1662 
1663 		n = m_get(M_WAITOK, MT_DATA);
1664 
1665 		SOCKBUF_LOCK(sb);
1666 		if (sb->sb_mtls != top) {
1667 			m_free(n);
1668 			return (NULL);
1669 		}
1670 	}
1671 	n->m_flags |= M_NOTREADY;
1672 
1673 	/* Store remainder in 'n'. */
1674 	n->m_len = m->m_len - remain;
1675 	if (m->m_flags & M_EXT) {
1676 		n->m_data = m->m_data + remain;
1677 		mb_dupcl(n, m);
1678 	} else {
1679 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1680 	}
1681 
1682 	/* Trim 'm' and update accounting. */
1683 	m->m_len -= n->m_len;
1684 	sb->sb_tlscc -= n->m_len;
1685 	sb->sb_ccc -= n->m_len;
1686 
1687 	/* Account for 'n'. */
1688 	sballoc_ktls_rx(sb, n);
1689 
1690 	/* Insert 'n' into the TLS chain. */
1691 	sb->sb_mtls = n;
1692 	n->m_next = m->m_next;
1693 	if (sb->sb_mtlstail == m)
1694 		sb->sb_mtlstail = n;
1695 
1696 	/* Detach the record from the TLS chain. */
1697 	m->m_next = NULL;
1698 
1699 out:
1700 	MPASS(m_length(top, NULL) == len);
1701 	for (m = top; m != NULL; m = m->m_next)
1702 		sbfree_ktls_rx(sb, m);
1703 	sb->sb_tlsdcc = len;
1704 	sb->sb_ccc += len;
1705 	SBCHECK(sb);
1706 	return (top);
1707 }
1708 
1709 static void
1710 ktls_decrypt(struct socket *so)
1711 {
1712 	char tls_header[MBUF_PEXT_HDR_LEN];
1713 	struct ktls_session *tls;
1714 	struct sockbuf *sb;
1715 	struct tls_record_layer *hdr;
1716 	struct tls_get_record tgr;
1717 	struct mbuf *control, *data, *m;
1718 	uint64_t seqno;
1719 	int error, remain, tls_len, trail_len;
1720 
1721 	hdr = (struct tls_record_layer *)tls_header;
1722 	sb = &so->so_rcv;
1723 	SOCKBUF_LOCK(sb);
1724 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1725 	    ("%s: socket %p not running", __func__, so));
1726 
1727 	tls = sb->sb_tls_info;
1728 	MPASS(tls != NULL);
1729 
1730 	for (;;) {
1731 		/* Is there enough queued for a TLS header? */
1732 		if (sb->sb_tlscc < tls->params.tls_hlen)
1733 			break;
1734 
1735 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1736 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1737 
1738 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1739 		    hdr->tls_vminor != tls->params.tls_vminor)
1740 			error = EINVAL;
1741 		else if (tls_len < tls->params.tls_hlen || tls_len >
1742 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1743 		    tls->params.tls_tlen)
1744 			error = EMSGSIZE;
1745 		else
1746 			error = 0;
1747 		if (__predict_false(error != 0)) {
1748 			/*
1749 			 * We have a corrupted record and are likely
1750 			 * out of sync.  The connection isn't
1751 			 * recoverable at this point, so abort it.
1752 			 */
1753 			SOCKBUF_UNLOCK(sb);
1754 			counter_u64_add(ktls_offload_corrupted_records, 1);
1755 
1756 			CURVNET_SET(so->so_vnet);
1757 			so->so_proto->pr_usrreqs->pru_abort(so);
1758 			so->so_error = error;
1759 			CURVNET_RESTORE();
1760 			goto deref;
1761 		}
1762 
1763 		/* Is the entire record queued? */
1764 		if (sb->sb_tlscc < tls_len)
1765 			break;
1766 
1767 		/*
1768 		 * Split out the portion of the mbuf chain containing
1769 		 * this TLS record.
1770 		 */
1771 		data = ktls_detach_record(sb, tls_len);
1772 		if (data == NULL)
1773 			continue;
1774 		MPASS(sb->sb_tlsdcc == tls_len);
1775 
1776 		seqno = sb->sb_tls_seqno;
1777 		sb->sb_tls_seqno++;
1778 		SBCHECK(sb);
1779 		SOCKBUF_UNLOCK(sb);
1780 
1781 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1782 		if (error) {
1783 			counter_u64_add(ktls_offload_failed_crypto, 1);
1784 
1785 			SOCKBUF_LOCK(sb);
1786 			if (sb->sb_tlsdcc == 0) {
1787 				/*
1788 				 * sbcut/drop/flush discarded these
1789 				 * mbufs.
1790 				 */
1791 				m_freem(data);
1792 				break;
1793 			}
1794 
1795 			/*
1796 			 * Drop this TLS record's data, but keep
1797 			 * decrypting subsequent records.
1798 			 */
1799 			sb->sb_ccc -= tls_len;
1800 			sb->sb_tlsdcc = 0;
1801 
1802 			CURVNET_SET(so->so_vnet);
1803 			so->so_error = EBADMSG;
1804 			sorwakeup_locked(so);
1805 			CURVNET_RESTORE();
1806 
1807 			m_freem(data);
1808 
1809 			SOCKBUF_LOCK(sb);
1810 			continue;
1811 		}
1812 
1813 		/* Allocate the control mbuf. */
1814 		tgr.tls_type = hdr->tls_type;
1815 		tgr.tls_vmajor = hdr->tls_vmajor;
1816 		tgr.tls_vminor = hdr->tls_vminor;
1817 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1818 		    trail_len);
1819 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1820 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1821 
1822 		SOCKBUF_LOCK(sb);
1823 		if (sb->sb_tlsdcc == 0) {
1824 			/* sbcut/drop/flush discarded these mbufs. */
1825 			MPASS(sb->sb_tlscc == 0);
1826 			m_freem(data);
1827 			m_freem(control);
1828 			break;
1829 		}
1830 
1831 		/*
1832 		 * Clear the 'dcc' accounting in preparation for
1833 		 * adding the decrypted record.
1834 		 */
1835 		sb->sb_ccc -= tls_len;
1836 		sb->sb_tlsdcc = 0;
1837 		SBCHECK(sb);
1838 
1839 		/* If there is no payload, drop all of the data. */
1840 		if (tgr.tls_length == htobe16(0)) {
1841 			m_freem(data);
1842 			data = NULL;
1843 		} else {
1844 			/* Trim header. */
1845 			remain = tls->params.tls_hlen;
1846 			while (remain > 0) {
1847 				if (data->m_len > remain) {
1848 					data->m_data += remain;
1849 					data->m_len -= remain;
1850 					break;
1851 				}
1852 				remain -= data->m_len;
1853 				data = m_free(data);
1854 			}
1855 
1856 			/* Trim trailer and clear M_NOTREADY. */
1857 			remain = be16toh(tgr.tls_length);
1858 			m = data;
1859 			for (m = data; remain > m->m_len; m = m->m_next) {
1860 				m->m_flags &= ~M_NOTREADY;
1861 				remain -= m->m_len;
1862 			}
1863 			m->m_len = remain;
1864 			m_freem(m->m_next);
1865 			m->m_next = NULL;
1866 			m->m_flags &= ~M_NOTREADY;
1867 
1868 			/* Set EOR on the final mbuf. */
1869 			m->m_flags |= M_EOR;
1870 		}
1871 
1872 		sbappendcontrol_locked(sb, data, control, 0);
1873 	}
1874 
1875 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1876 
1877 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1878 		so->so_error = EMSGSIZE;
1879 
1880 	sorwakeup_locked(so);
1881 
1882 deref:
1883 	SOCKBUF_UNLOCK_ASSERT(sb);
1884 
1885 	CURVNET_SET(so->so_vnet);
1886 	SOCK_LOCK(so);
1887 	sorele(so);
1888 	CURVNET_RESTORE();
1889 }
1890 
1891 void
1892 ktls_enqueue_to_free(struct mbuf *m)
1893 {
1894 	struct ktls_wq *wq;
1895 	bool running;
1896 
1897 	/* Mark it for freeing. */
1898 	m->m_epg_flags |= EPG_FLAG_2FREE;
1899 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1900 	mtx_lock(&wq->mtx);
1901 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1902 	running = wq->running;
1903 	mtx_unlock(&wq->mtx);
1904 	if (!running)
1905 		wakeup(wq);
1906 }
1907 
1908 void
1909 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1910 {
1911 	struct ktls_wq *wq;
1912 	bool running;
1913 
1914 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1915 	    (M_EXTPG | M_NOTREADY)),
1916 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1917 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1918 
1919 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1920 
1921 	m->m_epg_enc_cnt = page_count;
1922 
1923 	/*
1924 	 * Save a pointer to the socket.  The caller is responsible
1925 	 * for taking an additional reference via soref().
1926 	 */
1927 	m->m_epg_so = so;
1928 
1929 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1930 	mtx_lock(&wq->mtx);
1931 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1932 	running = wq->running;
1933 	mtx_unlock(&wq->mtx);
1934 	if (!running)
1935 		wakeup(wq);
1936 	counter_u64_add(ktls_cnt_tx_queued, 1);
1937 }
1938 
1939 static __noinline void
1940 ktls_encrypt(struct mbuf *top)
1941 {
1942 	struct ktls_session *tls;
1943 	struct socket *so;
1944 	struct mbuf *m;
1945 	vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1946 	struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1947 	struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1948 	vm_page_t pg;
1949 	int error, i, len, npages, off, total_pages;
1950 	bool is_anon;
1951 
1952 	so = top->m_epg_so;
1953 	tls = top->m_epg_tls;
1954 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
1955 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
1956 #ifdef INVARIANTS
1957 	top->m_epg_so = NULL;
1958 #endif
1959 	total_pages = top->m_epg_enc_cnt;
1960 	npages = 0;
1961 
1962 	/*
1963 	 * Encrypt the TLS records in the chain of mbufs starting with
1964 	 * 'top'.  'total_pages' gives us a total count of pages and is
1965 	 * used to know when we have finished encrypting the TLS
1966 	 * records originally queued with 'top'.
1967 	 *
1968 	 * NB: These mbufs are queued in the socket buffer and
1969 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
1970 	 * socket buffer lock is not held while traversing this chain.
1971 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1972 	 * pointers should be stable.  However, the 'm_next' of the
1973 	 * last mbuf encrypted is not necessarily NULL.  It can point
1974 	 * to other mbufs appended while 'top' was on the TLS work
1975 	 * queue.
1976 	 *
1977 	 * Each mbuf holds an entire TLS record.
1978 	 */
1979 	error = 0;
1980 	for (m = top; npages != total_pages; m = m->m_next) {
1981 		KASSERT(m->m_epg_tls == tls,
1982 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
1983 		    tls, m->m_epg_tls));
1984 		KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1985 		    (M_EXTPG | M_NOTREADY),
1986 		    ("%p not unready & nomap mbuf (top = %p)\n", m, top));
1987 		KASSERT(npages + m->m_epg_npgs <= total_pages,
1988 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
1989 		    total_pages, m));
1990 
1991 		/*
1992 		 * Generate source and destination ivoecs to pass to
1993 		 * the SW encryption backend.  For writable mbufs, the
1994 		 * destination iovec is a copy of the source and
1995 		 * encryption is done in place.  For file-backed mbufs
1996 		 * (from sendfile), anonymous wired pages are
1997 		 * allocated and assigned to the destination iovec.
1998 		 */
1999 		is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
2000 
2001 		off = m->m_epg_1st_off;
2002 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2003 			len = m_epg_pagelen(m, i, off);
2004 			src_iov[i].iov_len = len;
2005 			src_iov[i].iov_base =
2006 			    (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
2007 				off;
2008 
2009 			if (is_anon) {
2010 				dst_iov[i].iov_base = src_iov[i].iov_base;
2011 				dst_iov[i].iov_len = src_iov[i].iov_len;
2012 				continue;
2013 			}
2014 retry_page:
2015 			pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2016 			    VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
2017 			if (pg == NULL) {
2018 				vm_wait(NULL);
2019 				goto retry_page;
2020 			}
2021 			parray[i] = VM_PAGE_TO_PHYS(pg);
2022 			dst_iov[i].iov_base =
2023 			    (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
2024 			dst_iov[i].iov_len = len;
2025 		}
2026 
2027 		if (__predict_false(m->m_epg_npgs == 0)) {
2028 			/* TLS 1.0 empty fragment. */
2029 			npages++;
2030 		} else
2031 			npages += i;
2032 
2033 		error = (*tls->sw_encrypt)(tls,
2034 		    (const struct tls_record_layer *)m->m_epg_hdr,
2035 		    m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
2036 		    m->m_epg_record_type);
2037 		if (error) {
2038 			counter_u64_add(ktls_offload_failed_crypto, 1);
2039 			break;
2040 		}
2041 
2042 		/*
2043 		 * For file-backed mbufs, release the file-backed
2044 		 * pages and replace them in the ext_pgs array with
2045 		 * the anonymous wired pages allocated above.
2046 		 */
2047 		if (!is_anon) {
2048 			/* Free the old pages. */
2049 			m->m_ext.ext_free(m);
2050 
2051 			/* Replace them with the new pages. */
2052 			for (i = 0; i < m->m_epg_npgs; i++)
2053 				m->m_epg_pa[i] = parray[i];
2054 
2055 			/* Use the basic free routine. */
2056 			m->m_ext.ext_free = mb_free_mext_pgs;
2057 
2058 			/* Pages are now writable. */
2059 			m->m_epg_flags |= EPG_FLAG_ANON;
2060 		}
2061 
2062 		/*
2063 		 * Drop a reference to the session now that it is no
2064 		 * longer needed.  Existing code depends on encrypted
2065 		 * records having no associated session vs
2066 		 * yet-to-be-encrypted records having an associated
2067 		 * session.
2068 		 */
2069 		m->m_epg_tls = NULL;
2070 		ktls_free(tls);
2071 	}
2072 
2073 	CURVNET_SET(so->so_vnet);
2074 	if (error == 0) {
2075 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2076 	} else {
2077 		so->so_proto->pr_usrreqs->pru_abort(so);
2078 		so->so_error = EIO;
2079 		mb_free_notready(top, total_pages);
2080 	}
2081 
2082 	SOCK_LOCK(so);
2083 	sorele(so);
2084 	CURVNET_RESTORE();
2085 }
2086 
2087 static void
2088 ktls_work_thread(void *ctx)
2089 {
2090 	struct ktls_wq *wq = ctx;
2091 	struct mbuf *m, *n;
2092 	struct socket *so, *son;
2093 	STAILQ_HEAD(, mbuf) local_m_head;
2094 	STAILQ_HEAD(, socket) local_so_head;
2095 
2096 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2097 	fpu_kern_thread(0);
2098 #endif
2099 	for (;;) {
2100 		mtx_lock(&wq->mtx);
2101 		while (STAILQ_EMPTY(&wq->m_head) &&
2102 		    STAILQ_EMPTY(&wq->so_head)) {
2103 			wq->running = false;
2104 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2105 			wq->running = true;
2106 		}
2107 
2108 		STAILQ_INIT(&local_m_head);
2109 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2110 		STAILQ_INIT(&local_so_head);
2111 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2112 		mtx_unlock(&wq->mtx);
2113 
2114 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2115 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2116 				ktls_free(m->m_epg_tls);
2117 				uma_zfree(zone_mbuf, m);
2118 			} else {
2119 				ktls_encrypt(m);
2120 				counter_u64_add(ktls_cnt_tx_queued, -1);
2121 			}
2122 		}
2123 
2124 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2125 			ktls_decrypt(so);
2126 			counter_u64_add(ktls_cnt_rx_queued, -1);
2127 		}
2128 	}
2129 }
2130