xref: /freebsd/sys/kern/uipc_ktls.c (revision 69c5fa5cd1ec9b09ed88a086607a8a0993818db9)
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/domainset.h>
38 #include <sys/ktls.h>
39 #include <sys/lock.h>
40 #include <sys/mbuf.h>
41 #include <sys/mutex.h>
42 #include <sys/rmlock.h>
43 #include <sys/proc.h>
44 #include <sys/protosw.h>
45 #include <sys/refcount.h>
46 #include <sys/smp.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/taskqueue.h>
51 #include <sys/kthread.h>
52 #include <sys/uio.h>
53 #include <sys/vmmeter.h>
54 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
55 #include <machine/pcb.h>
56 #endif
57 #include <machine/vmparam.h>
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #ifdef RSS
61 #include <net/netisr.h>
62 #include <net/rss_config.h>
63 #endif
64 #include <net/route.h>
65 #include <net/route/nhop.h>
66 #if defined(INET) || defined(INET6)
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #endif
70 #include <netinet/tcp_var.h>
71 #ifdef TCP_OFFLOAD
72 #include <netinet/tcp_offload.h>
73 #endif
74 #include <opencrypto/xform.h>
75 #include <vm/uma_dbg.h>
76 #include <vm/vm.h>
77 #include <vm/vm_pageout.h>
78 #include <vm/vm_page.h>
79 
80 struct ktls_wq {
81 	struct mtx	mtx;
82 	STAILQ_HEAD(, mbuf) m_head;
83 	STAILQ_HEAD(, socket) so_head;
84 	bool		running;
85 } __aligned(CACHE_LINE_SIZE);
86 
87 struct ktls_domain_info {
88 	int count;
89 	int cpu[MAXCPU];
90 };
91 
92 struct ktls_domain_info ktls_domains[MAXMEMDOM];
93 static struct ktls_wq *ktls_wq;
94 static struct proc *ktls_proc;
95 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
96 static struct rmlock ktls_backends_lock;
97 static uma_zone_t ktls_session_zone;
98 static uint16_t ktls_cpuid_lookup[MAXCPU];
99 
100 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101     "Kernel TLS offload");
102 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
103     "Kernel TLS offload stats");
104 
105 static int ktls_allow_unload;
106 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
107     &ktls_allow_unload, 0, "Allow software crypto modules to unload");
108 
109 #ifdef RSS
110 static int ktls_bind_threads = 1;
111 #else
112 static int ktls_bind_threads;
113 #endif
114 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
115     &ktls_bind_threads, 0,
116     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
117 
118 static u_int ktls_maxlen = 16384;
119 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
120     &ktls_maxlen, 0, "Maximum TLS record size");
121 
122 static int ktls_number_threads;
123 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
124     &ktls_number_threads, 0,
125     "Number of TLS threads in thread-pool");
126 
127 static bool ktls_offload_enable;
128 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW,
129     &ktls_offload_enable, 0,
130     "Enable support for kernel TLS offload");
131 
132 static bool ktls_cbc_enable = true;
133 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW,
134     &ktls_cbc_enable, 1,
135     "Enable Support of AES-CBC crypto for kernel TLS");
136 
137 static counter_u64_t ktls_tasks_active;
138 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
139     &ktls_tasks_active, "Number of active tasks");
140 
141 static counter_u64_t ktls_cnt_tx_queued;
142 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
143     &ktls_cnt_tx_queued,
144     "Number of TLS records in queue to tasks for SW encryption");
145 
146 static counter_u64_t ktls_cnt_rx_queued;
147 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
148     &ktls_cnt_rx_queued,
149     "Number of TLS sockets in queue to tasks for SW decryption");
150 
151 static counter_u64_t ktls_offload_total;
152 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
153     CTLFLAG_RD, &ktls_offload_total,
154     "Total successful TLS setups (parameters set)");
155 
156 static counter_u64_t ktls_offload_enable_calls;
157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
158     CTLFLAG_RD, &ktls_offload_enable_calls,
159     "Total number of TLS enable calls made");
160 
161 static counter_u64_t ktls_offload_active;
162 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
163     &ktls_offload_active, "Total Active TLS sessions");
164 
165 static counter_u64_t ktls_offload_corrupted_records;
166 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
167     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
168 
169 static counter_u64_t ktls_offload_failed_crypto;
170 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
171     &ktls_offload_failed_crypto, "Total TLS crypto failures");
172 
173 static counter_u64_t ktls_switch_to_ifnet;
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
175     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
176 
177 static counter_u64_t ktls_switch_to_sw;
178 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
179     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
180 
181 static counter_u64_t ktls_switch_failed;
182 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
183     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
184 
185 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
186     "Software TLS session stats");
187 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
188     "Hardware (ifnet) TLS session stats");
189 #ifdef TCP_OFFLOAD
190 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
191     "TOE TLS session stats");
192 #endif
193 
194 static counter_u64_t ktls_sw_cbc;
195 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
196     "Active number of software TLS sessions using AES-CBC");
197 
198 static counter_u64_t ktls_sw_gcm;
199 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
200     "Active number of software TLS sessions using AES-GCM");
201 
202 static counter_u64_t ktls_ifnet_cbc;
203 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
204     &ktls_ifnet_cbc,
205     "Active number of ifnet TLS sessions using AES-CBC");
206 
207 static counter_u64_t ktls_ifnet_gcm;
208 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
209     &ktls_ifnet_gcm,
210     "Active number of ifnet TLS sessions using AES-GCM");
211 
212 static counter_u64_t ktls_ifnet_reset;
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
214     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
215 
216 static counter_u64_t ktls_ifnet_reset_dropped;
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
218     &ktls_ifnet_reset_dropped,
219     "TLS sessions dropped after failing to update ifnet send tag");
220 
221 static counter_u64_t ktls_ifnet_reset_failed;
222 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
223     &ktls_ifnet_reset_failed,
224     "TLS sessions that failed to allocate a new ifnet send tag");
225 
226 static int ktls_ifnet_permitted;
227 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
228     &ktls_ifnet_permitted, 1,
229     "Whether to permit hardware (ifnet) TLS sessions");
230 
231 #ifdef TCP_OFFLOAD
232 static counter_u64_t ktls_toe_cbc;
233 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
234     &ktls_toe_cbc,
235     "Active number of TOE TLS sessions using AES-CBC");
236 
237 static counter_u64_t ktls_toe_gcm;
238 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
239     &ktls_toe_gcm,
240     "Active number of TOE TLS sessions using AES-GCM");
241 #endif
242 
243 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
244 
245 static void ktls_cleanup(struct ktls_session *tls);
246 #if defined(INET) || defined(INET6)
247 static void ktls_reset_send_tag(void *context, int pending);
248 #endif
249 static void ktls_work_thread(void *ctx);
250 
251 int
252 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
253 {
254 	struct ktls_crypto_backend *curr_be, *tmp;
255 
256 	if (be->api_version != KTLS_API_VERSION) {
257 		printf("KTLS: API version mismatch (%d vs %d) for %s\n",
258 		    be->api_version, KTLS_API_VERSION,
259 		    be->name);
260 		return (EINVAL);
261 	}
262 
263 	rm_wlock(&ktls_backends_lock);
264 	printf("KTLS: Registering crypto method %s with prio %d\n",
265 	       be->name, be->prio);
266 	if (LIST_EMPTY(&ktls_backends)) {
267 		LIST_INSERT_HEAD(&ktls_backends, be, next);
268 	} else {
269 		LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
270 			if (curr_be->prio < be->prio) {
271 				LIST_INSERT_BEFORE(curr_be, be, next);
272 				break;
273 			}
274 			if (LIST_NEXT(curr_be, next) == NULL) {
275 				LIST_INSERT_AFTER(curr_be, be, next);
276 				break;
277 			}
278 		}
279 	}
280 	rm_wunlock(&ktls_backends_lock);
281 	return (0);
282 }
283 
284 int
285 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
286 {
287 	struct ktls_crypto_backend *tmp;
288 
289 	/*
290 	 * Don't error if the backend isn't registered.  This permits
291 	 * MOD_UNLOAD handlers to use this function unconditionally.
292 	 */
293 	rm_wlock(&ktls_backends_lock);
294 	LIST_FOREACH(tmp, &ktls_backends, next) {
295 		if (tmp == be)
296 			break;
297 	}
298 	if (tmp == NULL) {
299 		rm_wunlock(&ktls_backends_lock);
300 		return (0);
301 	}
302 
303 	if (!ktls_allow_unload) {
304 		rm_wunlock(&ktls_backends_lock);
305 		printf(
306 		    "KTLS: Deregistering crypto method %s is not supported\n",
307 		    be->name);
308 		return (EBUSY);
309 	}
310 
311 	if (be->use_count) {
312 		rm_wunlock(&ktls_backends_lock);
313 		return (EBUSY);
314 	}
315 
316 	LIST_REMOVE(be, next);
317 	rm_wunlock(&ktls_backends_lock);
318 	return (0);
319 }
320 
321 #if defined(INET) || defined(INET6)
322 static u_int
323 ktls_get_cpu(struct socket *so)
324 {
325 	struct inpcb *inp;
326 #ifdef NUMA
327 	struct ktls_domain_info *di;
328 #endif
329 	u_int cpuid;
330 
331 	inp = sotoinpcb(so);
332 #ifdef RSS
333 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
334 	if (cpuid != NETISR_CPUID_NONE)
335 		return (cpuid);
336 #endif
337 	/*
338 	 * Just use the flowid to shard connections in a repeatable
339 	 * fashion.  Note that some crypto backends rely on the
340 	 * serialization provided by having the same connection use
341 	 * the same queue.
342 	 */
343 #ifdef NUMA
344 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
345 		di = &ktls_domains[inp->inp_numa_domain];
346 		cpuid = di->cpu[inp->inp_flowid % di->count];
347 	} else
348 #endif
349 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
350 	return (cpuid);
351 }
352 #endif
353 
354 static void
355 ktls_init(void *dummy __unused)
356 {
357 	struct thread *td;
358 	struct pcpu *pc;
359 	cpuset_t mask;
360 	int count, domain, error, i;
361 
362 	ktls_tasks_active = counter_u64_alloc(M_WAITOK);
363 	ktls_cnt_tx_queued = counter_u64_alloc(M_WAITOK);
364 	ktls_cnt_rx_queued = counter_u64_alloc(M_WAITOK);
365 	ktls_offload_total = counter_u64_alloc(M_WAITOK);
366 	ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
367 	ktls_offload_active = counter_u64_alloc(M_WAITOK);
368 	ktls_offload_corrupted_records = counter_u64_alloc(M_WAITOK);
369 	ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
370 	ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
371 	ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
372 	ktls_switch_failed = counter_u64_alloc(M_WAITOK);
373 	ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
374 	ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
375 	ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
376 	ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
377 	ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
378 	ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
379 	ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
380 #ifdef TCP_OFFLOAD
381 	ktls_toe_cbc = counter_u64_alloc(M_WAITOK);
382 	ktls_toe_gcm = counter_u64_alloc(M_WAITOK);
383 #endif
384 
385 	rm_init(&ktls_backends_lock, "ktls backends");
386 	LIST_INIT(&ktls_backends);
387 
388 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
389 	    M_WAITOK | M_ZERO);
390 
391 	ktls_session_zone = uma_zcreate("ktls_session",
392 	    sizeof(struct ktls_session),
393 	    NULL, NULL, NULL, NULL,
394 	    UMA_ALIGN_CACHE, 0);
395 
396 	/*
397 	 * Initialize the workqueues to run the TLS work.  We create a
398 	 * work queue for each CPU.
399 	 */
400 	CPU_FOREACH(i) {
401 		STAILQ_INIT(&ktls_wq[i].m_head);
402 		STAILQ_INIT(&ktls_wq[i].so_head);
403 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
404 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
405 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
406 		if (error)
407 			panic("Can't add KTLS thread %d error %d", i, error);
408 
409 		/*
410 		 * Bind threads to cores.  If ktls_bind_threads is >
411 		 * 1, then we bind to the NUMA domain.
412 		 */
413 		if (ktls_bind_threads) {
414 			if (ktls_bind_threads > 1) {
415 				pc = pcpu_find(i);
416 				domain = pc->pc_domain;
417 				CPU_COPY(&cpuset_domain[domain], &mask);
418 				count = ktls_domains[domain].count;
419 				ktls_domains[domain].cpu[count] = i;
420 				ktls_domains[domain].count++;
421 			} else {
422 				CPU_SETOF(i, &mask);
423 			}
424 			error = cpuset_setthread(td->td_tid, &mask);
425 			if (error)
426 				panic(
427 			    "Unable to bind KTLS thread for CPU %d error %d",
428 				     i, error);
429 		}
430 		ktls_cpuid_lookup[ktls_number_threads] = i;
431 		ktls_number_threads++;
432 	}
433 
434 	/*
435 	 * If we somehow have an empty domain, fall back to choosing
436 	 * among all KTLS threads.
437 	 */
438 	if (ktls_bind_threads > 1) {
439 		for (i = 0; i < vm_ndomains; i++) {
440 			if (ktls_domains[i].count == 0) {
441 				ktls_bind_threads = 1;
442 				break;
443 			}
444 		}
445 	}
446 
447 	printf("KTLS: Initialized %d threads\n", ktls_number_threads);
448 }
449 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
450 
451 #if defined(INET) || defined(INET6)
452 static int
453 ktls_create_session(struct socket *so, struct tls_enable *en,
454     struct ktls_session **tlsp)
455 {
456 	struct ktls_session *tls;
457 	int error;
458 
459 	/* Only TLS 1.0 - 1.3 are supported. */
460 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
461 		return (EINVAL);
462 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
463 	    en->tls_vminor > TLS_MINOR_VER_THREE)
464 		return (EINVAL);
465 
466 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
467 		return (EINVAL);
468 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
469 		return (EINVAL);
470 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
471 		return (EINVAL);
472 
473 	/* All supported algorithms require a cipher key. */
474 	if (en->cipher_key_len == 0)
475 		return (EINVAL);
476 
477 	/* No flags are currently supported. */
478 	if (en->flags != 0)
479 		return (EINVAL);
480 
481 	/* Common checks for supported algorithms. */
482 	switch (en->cipher_algorithm) {
483 	case CRYPTO_AES_NIST_GCM_16:
484 		/*
485 		 * auth_algorithm isn't used, but permit GMAC values
486 		 * for compatibility.
487 		 */
488 		switch (en->auth_algorithm) {
489 		case 0:
490 #ifdef COMPAT_FREEBSD12
491 		/* XXX: Really 13.0-current COMPAT. */
492 		case CRYPTO_AES_128_NIST_GMAC:
493 		case CRYPTO_AES_192_NIST_GMAC:
494 		case CRYPTO_AES_256_NIST_GMAC:
495 #endif
496 			break;
497 		default:
498 			return (EINVAL);
499 		}
500 		if (en->auth_key_len != 0)
501 			return (EINVAL);
502 		if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
503 			en->iv_len != TLS_AEAD_GCM_LEN) ||
504 		    (en->tls_vminor == TLS_MINOR_VER_THREE &&
505 			en->iv_len != TLS_1_3_GCM_IV_LEN))
506 			return (EINVAL);
507 		break;
508 	case CRYPTO_AES_CBC:
509 		switch (en->auth_algorithm) {
510 		case CRYPTO_SHA1_HMAC:
511 			/*
512 			 * TLS 1.0 requires an implicit IV.  TLS 1.1+
513 			 * all use explicit IVs.
514 			 */
515 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
516 				if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
517 					return (EINVAL);
518 				break;
519 			}
520 
521 			/* FALLTHROUGH */
522 		case CRYPTO_SHA2_256_HMAC:
523 		case CRYPTO_SHA2_384_HMAC:
524 			/* Ignore any supplied IV. */
525 			en->iv_len = 0;
526 			break;
527 		default:
528 			return (EINVAL);
529 		}
530 		if (en->auth_key_len == 0)
531 			return (EINVAL);
532 		break;
533 	default:
534 		return (EINVAL);
535 	}
536 
537 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
538 
539 	counter_u64_add(ktls_offload_active, 1);
540 
541 	refcount_init(&tls->refcount, 1);
542 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
543 
544 	tls->wq_index = ktls_get_cpu(so);
545 
546 	tls->params.cipher_algorithm = en->cipher_algorithm;
547 	tls->params.auth_algorithm = en->auth_algorithm;
548 	tls->params.tls_vmajor = en->tls_vmajor;
549 	tls->params.tls_vminor = en->tls_vminor;
550 	tls->params.flags = en->flags;
551 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
552 
553 	/* Set the header and trailer lengths. */
554 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
555 	switch (en->cipher_algorithm) {
556 	case CRYPTO_AES_NIST_GCM_16:
557 		/*
558 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
559 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
560 		 */
561 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
562 			tls->params.tls_hlen += sizeof(uint64_t);
563 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
564 
565 		/*
566 		 * TLS 1.3 includes optional padding which we
567 		 * do not support, and also puts the "real" record
568 		 * type at the end of the encrypted data.
569 		 */
570 		if (en->tls_vminor == TLS_MINOR_VER_THREE)
571 			tls->params.tls_tlen += sizeof(uint8_t);
572 
573 		tls->params.tls_bs = 1;
574 		break;
575 	case CRYPTO_AES_CBC:
576 		switch (en->auth_algorithm) {
577 		case CRYPTO_SHA1_HMAC:
578 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
579 				/* Implicit IV, no nonce. */
580 			} else {
581 				tls->params.tls_hlen += AES_BLOCK_LEN;
582 			}
583 			tls->params.tls_tlen = AES_BLOCK_LEN +
584 			    SHA1_HASH_LEN;
585 			break;
586 		case CRYPTO_SHA2_256_HMAC:
587 			tls->params.tls_hlen += AES_BLOCK_LEN;
588 			tls->params.tls_tlen = AES_BLOCK_LEN +
589 			    SHA2_256_HASH_LEN;
590 			break;
591 		case CRYPTO_SHA2_384_HMAC:
592 			tls->params.tls_hlen += AES_BLOCK_LEN;
593 			tls->params.tls_tlen = AES_BLOCK_LEN +
594 			    SHA2_384_HASH_LEN;
595 			break;
596 		default:
597 			panic("invalid hmac");
598 		}
599 		tls->params.tls_bs = AES_BLOCK_LEN;
600 		break;
601 	default:
602 		panic("invalid cipher");
603 	}
604 
605 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
606 	    ("TLS header length too long: %d", tls->params.tls_hlen));
607 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
608 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
609 
610 	if (en->auth_key_len != 0) {
611 		tls->params.auth_key_len = en->auth_key_len;
612 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
613 		    M_WAITOK);
614 		error = copyin(en->auth_key, tls->params.auth_key,
615 		    en->auth_key_len);
616 		if (error)
617 			goto out;
618 	}
619 
620 	tls->params.cipher_key_len = en->cipher_key_len;
621 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
622 	error = copyin(en->cipher_key, tls->params.cipher_key,
623 	    en->cipher_key_len);
624 	if (error)
625 		goto out;
626 
627 	/*
628 	 * This holds the implicit portion of the nonce for GCM and
629 	 * the initial implicit IV for TLS 1.0.  The explicit portions
630 	 * of the IV are generated in ktls_frame().
631 	 */
632 	if (en->iv_len != 0) {
633 		tls->params.iv_len = en->iv_len;
634 		error = copyin(en->iv, tls->params.iv, en->iv_len);
635 		if (error)
636 			goto out;
637 
638 		/*
639 		 * For TLS 1.2, generate an 8-byte nonce as a counter
640 		 * to generate unique explicit IVs.
641 		 *
642 		 * Store this counter in the last 8 bytes of the IV
643 		 * array so that it is 8-byte aligned.
644 		 */
645 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
646 		    en->tls_vminor == TLS_MINOR_VER_TWO)
647 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
648 	}
649 
650 	*tlsp = tls;
651 	return (0);
652 
653 out:
654 	ktls_cleanup(tls);
655 	return (error);
656 }
657 
658 static struct ktls_session *
659 ktls_clone_session(struct ktls_session *tls)
660 {
661 	struct ktls_session *tls_new;
662 
663 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
664 
665 	counter_u64_add(ktls_offload_active, 1);
666 
667 	refcount_init(&tls_new->refcount, 1);
668 
669 	/* Copy fields from existing session. */
670 	tls_new->params = tls->params;
671 	tls_new->wq_index = tls->wq_index;
672 
673 	/* Deep copy keys. */
674 	if (tls_new->params.auth_key != NULL) {
675 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
676 		    M_KTLS, M_WAITOK);
677 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
678 		    tls->params.auth_key_len);
679 	}
680 
681 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
682 	    M_WAITOK);
683 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
684 	    tls->params.cipher_key_len);
685 
686 	return (tls_new);
687 }
688 #endif
689 
690 static void
691 ktls_cleanup(struct ktls_session *tls)
692 {
693 
694 	counter_u64_add(ktls_offload_active, -1);
695 	switch (tls->mode) {
696 	case TCP_TLS_MODE_SW:
697 		MPASS(tls->be != NULL);
698 		switch (tls->params.cipher_algorithm) {
699 		case CRYPTO_AES_CBC:
700 			counter_u64_add(ktls_sw_cbc, -1);
701 			break;
702 		case CRYPTO_AES_NIST_GCM_16:
703 			counter_u64_add(ktls_sw_gcm, -1);
704 			break;
705 		}
706 		tls->free(tls);
707 		break;
708 	case TCP_TLS_MODE_IFNET:
709 		switch (tls->params.cipher_algorithm) {
710 		case CRYPTO_AES_CBC:
711 			counter_u64_add(ktls_ifnet_cbc, -1);
712 			break;
713 		case CRYPTO_AES_NIST_GCM_16:
714 			counter_u64_add(ktls_ifnet_gcm, -1);
715 			break;
716 		}
717 		if (tls->snd_tag != NULL)
718 			m_snd_tag_rele(tls->snd_tag);
719 		break;
720 #ifdef TCP_OFFLOAD
721 	case TCP_TLS_MODE_TOE:
722 		switch (tls->params.cipher_algorithm) {
723 		case CRYPTO_AES_CBC:
724 			counter_u64_add(ktls_toe_cbc, -1);
725 			break;
726 		case CRYPTO_AES_NIST_GCM_16:
727 			counter_u64_add(ktls_toe_gcm, -1);
728 			break;
729 		}
730 		break;
731 #endif
732 	}
733 	if (tls->params.auth_key != NULL) {
734 		zfree(tls->params.auth_key, M_KTLS);
735 		tls->params.auth_key = NULL;
736 		tls->params.auth_key_len = 0;
737 	}
738 	if (tls->params.cipher_key != NULL) {
739 		zfree(tls->params.cipher_key, M_KTLS);
740 		tls->params.cipher_key = NULL;
741 		tls->params.cipher_key_len = 0;
742 	}
743 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
744 }
745 
746 #if defined(INET) || defined(INET6)
747 
748 #ifdef TCP_OFFLOAD
749 static int
750 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
751 {
752 	struct inpcb *inp;
753 	struct tcpcb *tp;
754 	int error;
755 
756 	inp = so->so_pcb;
757 	INP_WLOCK(inp);
758 	if (inp->inp_flags2 & INP_FREED) {
759 		INP_WUNLOCK(inp);
760 		return (ECONNRESET);
761 	}
762 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
763 		INP_WUNLOCK(inp);
764 		return (ECONNRESET);
765 	}
766 	if (inp->inp_socket == NULL) {
767 		INP_WUNLOCK(inp);
768 		return (ECONNRESET);
769 	}
770 	tp = intotcpcb(inp);
771 	if (!(tp->t_flags & TF_TOE)) {
772 		INP_WUNLOCK(inp);
773 		return (EOPNOTSUPP);
774 	}
775 
776 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
777 	INP_WUNLOCK(inp);
778 	if (error == 0) {
779 		tls->mode = TCP_TLS_MODE_TOE;
780 		switch (tls->params.cipher_algorithm) {
781 		case CRYPTO_AES_CBC:
782 			counter_u64_add(ktls_toe_cbc, 1);
783 			break;
784 		case CRYPTO_AES_NIST_GCM_16:
785 			counter_u64_add(ktls_toe_gcm, 1);
786 			break;
787 		}
788 	}
789 	return (error);
790 }
791 #endif
792 
793 /*
794  * Common code used when first enabling ifnet TLS on a connection or
795  * when allocating a new ifnet TLS session due to a routing change.
796  * This function allocates a new TLS send tag on whatever interface
797  * the connection is currently routed over.
798  */
799 static int
800 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
801     struct m_snd_tag **mstp)
802 {
803 	union if_snd_tag_alloc_params params;
804 	struct ifnet *ifp;
805 	struct nhop_object *nh;
806 	struct tcpcb *tp;
807 	int error;
808 
809 	INP_RLOCK(inp);
810 	if (inp->inp_flags2 & INP_FREED) {
811 		INP_RUNLOCK(inp);
812 		return (ECONNRESET);
813 	}
814 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
815 		INP_RUNLOCK(inp);
816 		return (ECONNRESET);
817 	}
818 	if (inp->inp_socket == NULL) {
819 		INP_RUNLOCK(inp);
820 		return (ECONNRESET);
821 	}
822 	tp = intotcpcb(inp);
823 
824 	/*
825 	 * Check administrative controls on ifnet TLS to determine if
826 	 * ifnet TLS should be denied.
827 	 *
828 	 * - Always permit 'force' requests.
829 	 * - ktls_ifnet_permitted == 0: always deny.
830 	 */
831 	if (!force && ktls_ifnet_permitted == 0) {
832 		INP_RUNLOCK(inp);
833 		return (ENXIO);
834 	}
835 
836 	/*
837 	 * XXX: Use the cached route in the inpcb to find the
838 	 * interface.  This should perhaps instead use
839 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
840 	 * enabled after a connection has completed key negotiation in
841 	 * userland, the cached route will be present in practice.
842 	 */
843 	nh = inp->inp_route.ro_nh;
844 	if (nh == NULL) {
845 		INP_RUNLOCK(inp);
846 		return (ENXIO);
847 	}
848 	ifp = nh->nh_ifp;
849 	if_ref(ifp);
850 
851 	/*
852 	 * Allocate a TLS + ratelimit tag if the connection has an
853 	 * existing pacing rate.
854 	 */
855 	if (tp->t_pacing_rate != -1 &&
856 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
857 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
858 		params.tls_rate_limit.inp = inp;
859 		params.tls_rate_limit.tls = tls;
860 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
861 	} else {
862 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
863 		params.tls.inp = inp;
864 		params.tls.tls = tls;
865 	}
866 	params.hdr.flowid = inp->inp_flowid;
867 	params.hdr.flowtype = inp->inp_flowtype;
868 	params.hdr.numa_domain = inp->inp_numa_domain;
869 	INP_RUNLOCK(inp);
870 
871 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
872 		error = EOPNOTSUPP;
873 		goto out;
874 	}
875 	if (inp->inp_vflag & INP_IPV6) {
876 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
877 			error = EOPNOTSUPP;
878 			goto out;
879 		}
880 	} else {
881 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
882 			error = EOPNOTSUPP;
883 			goto out;
884 		}
885 	}
886 	error = m_snd_tag_alloc(ifp, &params, mstp);
887 out:
888 	if_rele(ifp);
889 	return (error);
890 }
891 
892 static int
893 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
894 {
895 	struct m_snd_tag *mst;
896 	int error;
897 
898 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
899 	if (error == 0) {
900 		tls->mode = TCP_TLS_MODE_IFNET;
901 		tls->snd_tag = mst;
902 		switch (tls->params.cipher_algorithm) {
903 		case CRYPTO_AES_CBC:
904 			counter_u64_add(ktls_ifnet_cbc, 1);
905 			break;
906 		case CRYPTO_AES_NIST_GCM_16:
907 			counter_u64_add(ktls_ifnet_gcm, 1);
908 			break;
909 		}
910 	}
911 	return (error);
912 }
913 
914 static int
915 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
916 {
917 	struct rm_priotracker prio;
918 	struct ktls_crypto_backend *be;
919 
920 	/*
921 	 * Choose the best software crypto backend.  Backends are
922 	 * stored in sorted priority order (larget value == most
923 	 * important at the head of the list), so this just stops on
924 	 * the first backend that claims the session by returning
925 	 * success.
926 	 */
927 	if (ktls_allow_unload)
928 		rm_rlock(&ktls_backends_lock, &prio);
929 	LIST_FOREACH(be, &ktls_backends, next) {
930 		if (be->try(so, tls, direction) == 0)
931 			break;
932 		KASSERT(tls->cipher == NULL,
933 		    ("ktls backend leaked a cipher pointer"));
934 	}
935 	if (be != NULL) {
936 		if (ktls_allow_unload)
937 			be->use_count++;
938 		tls->be = be;
939 	}
940 	if (ktls_allow_unload)
941 		rm_runlock(&ktls_backends_lock, &prio);
942 	if (be == NULL)
943 		return (EOPNOTSUPP);
944 	tls->mode = TCP_TLS_MODE_SW;
945 	switch (tls->params.cipher_algorithm) {
946 	case CRYPTO_AES_CBC:
947 		counter_u64_add(ktls_sw_cbc, 1);
948 		break;
949 	case CRYPTO_AES_NIST_GCM_16:
950 		counter_u64_add(ktls_sw_gcm, 1);
951 		break;
952 	}
953 	return (0);
954 }
955 
956 /*
957  * KTLS RX stores data in the socket buffer as a list of TLS records,
958  * where each record is stored as a control message containg the TLS
959  * header followed by data mbufs containing the decrypted data.  This
960  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
961  * both encrypted and decrypted data.  TLS records decrypted by a NIC
962  * should be queued to the socket buffer as records, but encrypted
963  * data which needs to be decrypted by software arrives as a stream of
964  * regular mbufs which need to be converted.  In addition, there may
965  * already be pending encrypted data in the socket buffer when KTLS RX
966  * is enabled.
967  *
968  * To manage not-yet-decrypted data for KTLS RX, the following scheme
969  * is used:
970  *
971  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
972  *
973  * - ktls_check_rx checks this chain of mbufs reading the TLS header
974  *   from the first mbuf.  Once all of the data for that TLS record is
975  *   queued, the socket is queued to a worker thread.
976  *
977  * - The worker thread calls ktls_decrypt to decrypt TLS records in
978  *   the TLS chain.  Each TLS record is detached from the TLS chain,
979  *   decrypted, and inserted into the regular socket buffer chain as
980  *   record starting with a control message holding the TLS header and
981  *   a chain of mbufs holding the encrypted data.
982  */
983 
984 static void
985 sb_mark_notready(struct sockbuf *sb)
986 {
987 	struct mbuf *m;
988 
989 	m = sb->sb_mb;
990 	sb->sb_mtls = m;
991 	sb->sb_mb = NULL;
992 	sb->sb_mbtail = NULL;
993 	sb->sb_lastrecord = NULL;
994 	for (; m != NULL; m = m->m_next) {
995 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
996 		    __func__));
997 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
998 		    __func__));
999 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1000 		    __func__));
1001 		m->m_flags |= M_NOTREADY;
1002 		sb->sb_acc -= m->m_len;
1003 		sb->sb_tlscc += m->m_len;
1004 		sb->sb_mtlstail = m;
1005 	}
1006 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1007 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1008 	    sb->sb_ccc));
1009 }
1010 
1011 int
1012 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1013 {
1014 	struct ktls_session *tls;
1015 	int error;
1016 
1017 	if (!ktls_offload_enable)
1018 		return (ENOTSUP);
1019 	if (SOLISTENING(so))
1020 		return (EINVAL);
1021 
1022 	counter_u64_add(ktls_offload_enable_calls, 1);
1023 
1024 	/*
1025 	 * This should always be true since only the TCP socket option
1026 	 * invokes this function.
1027 	 */
1028 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1029 		return (EINVAL);
1030 
1031 	/*
1032 	 * XXX: Don't overwrite existing sessions.  We should permit
1033 	 * this to support rekeying in the future.
1034 	 */
1035 	if (so->so_rcv.sb_tls_info != NULL)
1036 		return (EALREADY);
1037 
1038 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1039 		return (ENOTSUP);
1040 
1041 	/* TLS 1.3 is not yet supported. */
1042 	if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1043 	    en->tls_vminor == TLS_MINOR_VER_THREE)
1044 		return (ENOTSUP);
1045 
1046 	error = ktls_create_session(so, en, &tls);
1047 	if (error)
1048 		return (error);
1049 
1050 #ifdef TCP_OFFLOAD
1051 	error = ktls_try_toe(so, tls, KTLS_RX);
1052 	if (error)
1053 #endif
1054 		error = ktls_try_sw(so, tls, KTLS_RX);
1055 
1056 	if (error) {
1057 		ktls_cleanup(tls);
1058 		return (error);
1059 	}
1060 
1061 	/* Mark the socket as using TLS offload. */
1062 	SOCKBUF_LOCK(&so->so_rcv);
1063 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1064 	so->so_rcv.sb_tls_info = tls;
1065 	so->so_rcv.sb_flags |= SB_TLS_RX;
1066 
1067 	/* Mark existing data as not ready until it can be decrypted. */
1068 	sb_mark_notready(&so->so_rcv);
1069 	ktls_check_rx(&so->so_rcv);
1070 	SOCKBUF_UNLOCK(&so->so_rcv);
1071 
1072 	counter_u64_add(ktls_offload_total, 1);
1073 
1074 	return (0);
1075 }
1076 
1077 int
1078 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1079 {
1080 	struct ktls_session *tls;
1081 	struct inpcb *inp;
1082 	int error;
1083 
1084 	if (!ktls_offload_enable)
1085 		return (ENOTSUP);
1086 	if (SOLISTENING(so))
1087 		return (EINVAL);
1088 
1089 	counter_u64_add(ktls_offload_enable_calls, 1);
1090 
1091 	/*
1092 	 * This should always be true since only the TCP socket option
1093 	 * invokes this function.
1094 	 */
1095 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1096 		return (EINVAL);
1097 
1098 	/*
1099 	 * XXX: Don't overwrite existing sessions.  We should permit
1100 	 * this to support rekeying in the future.
1101 	 */
1102 	if (so->so_snd.sb_tls_info != NULL)
1103 		return (EALREADY);
1104 
1105 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1106 		return (ENOTSUP);
1107 
1108 	/* TLS requires ext pgs */
1109 	if (mb_use_ext_pgs == 0)
1110 		return (ENXIO);
1111 
1112 	error = ktls_create_session(so, en, &tls);
1113 	if (error)
1114 		return (error);
1115 
1116 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1117 #ifdef TCP_OFFLOAD
1118 	error = ktls_try_toe(so, tls, KTLS_TX);
1119 	if (error)
1120 #endif
1121 		error = ktls_try_ifnet(so, tls, false);
1122 	if (error)
1123 		error = ktls_try_sw(so, tls, KTLS_TX);
1124 
1125 	if (error) {
1126 		ktls_cleanup(tls);
1127 		return (error);
1128 	}
1129 
1130 	error = sblock(&so->so_snd, SBL_WAIT);
1131 	if (error) {
1132 		ktls_cleanup(tls);
1133 		return (error);
1134 	}
1135 
1136 	/*
1137 	 * Write lock the INP when setting sb_tls_info so that
1138 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1139 	 * holding the INP lock.
1140 	 */
1141 	inp = so->so_pcb;
1142 	INP_WLOCK(inp);
1143 	SOCKBUF_LOCK(&so->so_snd);
1144 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1145 	so->so_snd.sb_tls_info = tls;
1146 	if (tls->mode != TCP_TLS_MODE_SW)
1147 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1148 	SOCKBUF_UNLOCK(&so->so_snd);
1149 	INP_WUNLOCK(inp);
1150 	sbunlock(&so->so_snd);
1151 
1152 	counter_u64_add(ktls_offload_total, 1);
1153 
1154 	return (0);
1155 }
1156 
1157 int
1158 ktls_get_rx_mode(struct socket *so)
1159 {
1160 	struct ktls_session *tls;
1161 	struct inpcb *inp;
1162 	int mode;
1163 
1164 	if (SOLISTENING(so))
1165 		return (EINVAL);
1166 	inp = so->so_pcb;
1167 	INP_WLOCK_ASSERT(inp);
1168 	SOCKBUF_LOCK(&so->so_rcv);
1169 	tls = so->so_rcv.sb_tls_info;
1170 	if (tls == NULL)
1171 		mode = TCP_TLS_MODE_NONE;
1172 	else
1173 		mode = tls->mode;
1174 	SOCKBUF_UNLOCK(&so->so_rcv);
1175 	return (mode);
1176 }
1177 
1178 int
1179 ktls_get_tx_mode(struct socket *so)
1180 {
1181 	struct ktls_session *tls;
1182 	struct inpcb *inp;
1183 	int mode;
1184 
1185 	if (SOLISTENING(so))
1186 		return (EINVAL);
1187 	inp = so->so_pcb;
1188 	INP_WLOCK_ASSERT(inp);
1189 	SOCKBUF_LOCK(&so->so_snd);
1190 	tls = so->so_snd.sb_tls_info;
1191 	if (tls == NULL)
1192 		mode = TCP_TLS_MODE_NONE;
1193 	else
1194 		mode = tls->mode;
1195 	SOCKBUF_UNLOCK(&so->so_snd);
1196 	return (mode);
1197 }
1198 
1199 /*
1200  * Switch between SW and ifnet TLS sessions as requested.
1201  */
1202 int
1203 ktls_set_tx_mode(struct socket *so, int mode)
1204 {
1205 	struct ktls_session *tls, *tls_new;
1206 	struct inpcb *inp;
1207 	int error;
1208 
1209 	if (SOLISTENING(so))
1210 		return (EINVAL);
1211 	switch (mode) {
1212 	case TCP_TLS_MODE_SW:
1213 	case TCP_TLS_MODE_IFNET:
1214 		break;
1215 	default:
1216 		return (EINVAL);
1217 	}
1218 
1219 	inp = so->so_pcb;
1220 	INP_WLOCK_ASSERT(inp);
1221 	SOCKBUF_LOCK(&so->so_snd);
1222 	tls = so->so_snd.sb_tls_info;
1223 	if (tls == NULL) {
1224 		SOCKBUF_UNLOCK(&so->so_snd);
1225 		return (0);
1226 	}
1227 
1228 	if (tls->mode == mode) {
1229 		SOCKBUF_UNLOCK(&so->so_snd);
1230 		return (0);
1231 	}
1232 
1233 	tls = ktls_hold(tls);
1234 	SOCKBUF_UNLOCK(&so->so_snd);
1235 	INP_WUNLOCK(inp);
1236 
1237 	tls_new = ktls_clone_session(tls);
1238 
1239 	if (mode == TCP_TLS_MODE_IFNET)
1240 		error = ktls_try_ifnet(so, tls_new, true);
1241 	else
1242 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1243 	if (error) {
1244 		counter_u64_add(ktls_switch_failed, 1);
1245 		ktls_free(tls_new);
1246 		ktls_free(tls);
1247 		INP_WLOCK(inp);
1248 		return (error);
1249 	}
1250 
1251 	error = sblock(&so->so_snd, SBL_WAIT);
1252 	if (error) {
1253 		counter_u64_add(ktls_switch_failed, 1);
1254 		ktls_free(tls_new);
1255 		ktls_free(tls);
1256 		INP_WLOCK(inp);
1257 		return (error);
1258 	}
1259 
1260 	/*
1261 	 * If we raced with another session change, keep the existing
1262 	 * session.
1263 	 */
1264 	if (tls != so->so_snd.sb_tls_info) {
1265 		counter_u64_add(ktls_switch_failed, 1);
1266 		sbunlock(&so->so_snd);
1267 		ktls_free(tls_new);
1268 		ktls_free(tls);
1269 		INP_WLOCK(inp);
1270 		return (EBUSY);
1271 	}
1272 
1273 	SOCKBUF_LOCK(&so->so_snd);
1274 	so->so_snd.sb_tls_info = tls_new;
1275 	if (tls_new->mode != TCP_TLS_MODE_SW)
1276 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1277 	SOCKBUF_UNLOCK(&so->so_snd);
1278 	sbunlock(&so->so_snd);
1279 
1280 	/*
1281 	 * Drop two references on 'tls'.  The first is for the
1282 	 * ktls_hold() above.  The second drops the reference from the
1283 	 * socket buffer.
1284 	 */
1285 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1286 	ktls_free(tls);
1287 	ktls_free(tls);
1288 
1289 	if (mode == TCP_TLS_MODE_IFNET)
1290 		counter_u64_add(ktls_switch_to_ifnet, 1);
1291 	else
1292 		counter_u64_add(ktls_switch_to_sw, 1);
1293 
1294 	INP_WLOCK(inp);
1295 	return (0);
1296 }
1297 
1298 /*
1299  * Try to allocate a new TLS send tag.  This task is scheduled when
1300  * ip_output detects a route change while trying to transmit a packet
1301  * holding a TLS record.  If a new tag is allocated, replace the tag
1302  * in the TLS session.  Subsequent packets on the connection will use
1303  * the new tag.  If a new tag cannot be allocated, drop the
1304  * connection.
1305  */
1306 static void
1307 ktls_reset_send_tag(void *context, int pending)
1308 {
1309 	struct epoch_tracker et;
1310 	struct ktls_session *tls;
1311 	struct m_snd_tag *old, *new;
1312 	struct inpcb *inp;
1313 	struct tcpcb *tp;
1314 	int error;
1315 
1316 	MPASS(pending == 1);
1317 
1318 	tls = context;
1319 	inp = tls->inp;
1320 
1321 	/*
1322 	 * Free the old tag first before allocating a new one.
1323 	 * ip[6]_output_send() will treat a NULL send tag the same as
1324 	 * an ifp mismatch and drop packets until a new tag is
1325 	 * allocated.
1326 	 *
1327 	 * Write-lock the INP when changing tls->snd_tag since
1328 	 * ip[6]_output_send() holds a read-lock when reading the
1329 	 * pointer.
1330 	 */
1331 	INP_WLOCK(inp);
1332 	old = tls->snd_tag;
1333 	tls->snd_tag = NULL;
1334 	INP_WUNLOCK(inp);
1335 	if (old != NULL)
1336 		m_snd_tag_rele(old);
1337 
1338 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1339 
1340 	if (error == 0) {
1341 		INP_WLOCK(inp);
1342 		tls->snd_tag = new;
1343 		mtx_pool_lock(mtxpool_sleep, tls);
1344 		tls->reset_pending = false;
1345 		mtx_pool_unlock(mtxpool_sleep, tls);
1346 		if (!in_pcbrele_wlocked(inp))
1347 			INP_WUNLOCK(inp);
1348 
1349 		counter_u64_add(ktls_ifnet_reset, 1);
1350 
1351 		/*
1352 		 * XXX: Should we kick tcp_output explicitly now that
1353 		 * the send tag is fixed or just rely on timers?
1354 		 */
1355 	} else {
1356 		NET_EPOCH_ENTER(et);
1357 		INP_WLOCK(inp);
1358 		if (!in_pcbrele_wlocked(inp)) {
1359 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1360 			    !(inp->inp_flags & INP_DROPPED)) {
1361 				tp = intotcpcb(inp);
1362 				CURVNET_SET(tp->t_vnet);
1363 				tp = tcp_drop(tp, ECONNABORTED);
1364 				CURVNET_RESTORE();
1365 				if (tp != NULL)
1366 					INP_WUNLOCK(inp);
1367 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1368 			} else
1369 				INP_WUNLOCK(inp);
1370 		}
1371 		NET_EPOCH_EXIT(et);
1372 
1373 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1374 
1375 		/*
1376 		 * Leave reset_pending true to avoid future tasks while
1377 		 * the socket goes away.
1378 		 */
1379 	}
1380 
1381 	ktls_free(tls);
1382 }
1383 
1384 int
1385 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1386 {
1387 
1388 	if (inp == NULL)
1389 		return (ENOBUFS);
1390 
1391 	INP_LOCK_ASSERT(inp);
1392 
1393 	/*
1394 	 * See if we should schedule a task to update the send tag for
1395 	 * this session.
1396 	 */
1397 	mtx_pool_lock(mtxpool_sleep, tls);
1398 	if (!tls->reset_pending) {
1399 		(void) ktls_hold(tls);
1400 		in_pcbref(inp);
1401 		tls->inp = inp;
1402 		tls->reset_pending = true;
1403 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1404 	}
1405 	mtx_pool_unlock(mtxpool_sleep, tls);
1406 	return (ENOBUFS);
1407 }
1408 
1409 #ifdef RATELIMIT
1410 int
1411 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1412 {
1413 	union if_snd_tag_modify_params params = {
1414 		.rate_limit.max_rate = max_pacing_rate,
1415 		.rate_limit.flags = M_NOWAIT,
1416 	};
1417 	struct m_snd_tag *mst;
1418 	struct ifnet *ifp;
1419 	int error;
1420 
1421 	/* Can't get to the inp, but it should be locked. */
1422 	/* INP_LOCK_ASSERT(inp); */
1423 
1424 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1425 
1426 	if (tls->snd_tag == NULL) {
1427 		/*
1428 		 * Resetting send tag, ignore this change.  The
1429 		 * pending reset may or may not see this updated rate
1430 		 * in the tcpcb.  If it doesn't, we will just lose
1431 		 * this rate change.
1432 		 */
1433 		return (0);
1434 	}
1435 
1436 	MPASS(tls->snd_tag != NULL);
1437 	MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1438 
1439 	mst = tls->snd_tag;
1440 	ifp = mst->ifp;
1441 	return (ifp->if_snd_tag_modify(mst, &params));
1442 }
1443 #endif
1444 #endif
1445 
1446 void
1447 ktls_destroy(struct ktls_session *tls)
1448 {
1449 	struct rm_priotracker prio;
1450 
1451 	ktls_cleanup(tls);
1452 	if (tls->be != NULL && ktls_allow_unload) {
1453 		rm_rlock(&ktls_backends_lock, &prio);
1454 		tls->be->use_count--;
1455 		rm_runlock(&ktls_backends_lock, &prio);
1456 	}
1457 	uma_zfree(ktls_session_zone, tls);
1458 }
1459 
1460 void
1461 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1462 {
1463 
1464 	for (; m != NULL; m = m->m_next) {
1465 		KASSERT((m->m_flags & M_EXTPG) != 0,
1466 		    ("ktls_seq: mapped mbuf %p", m));
1467 
1468 		m->m_epg_seqno = sb->sb_tls_seqno;
1469 		sb->sb_tls_seqno++;
1470 	}
1471 }
1472 
1473 /*
1474  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1475  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1476  * mbuf must be populated with the payload of each TLS record.
1477  *
1478  * The record_type argument specifies the TLS record type used when
1479  * populating the TLS header.
1480  *
1481  * The enq_count argument on return is set to the number of pages of
1482  * payload data for this entire chain that need to be encrypted via SW
1483  * encryption.  The returned value should be passed to ktls_enqueue
1484  * when scheduling encryption of this chain of mbufs.  To handle the
1485  * special case of empty fragments for TLS 1.0 sessions, an empty
1486  * fragment counts as one page.
1487  */
1488 void
1489 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1490     uint8_t record_type)
1491 {
1492 	struct tls_record_layer *tlshdr;
1493 	struct mbuf *m;
1494 	uint64_t *noncep;
1495 	uint16_t tls_len;
1496 	int maxlen;
1497 
1498 	maxlen = tls->params.max_frame_len;
1499 	*enq_cnt = 0;
1500 	for (m = top; m != NULL; m = m->m_next) {
1501 		/*
1502 		 * All mbufs in the chain should be TLS records whose
1503 		 * payload does not exceed the maximum frame length.
1504 		 *
1505 		 * Empty TLS records are permitted when using CBC.
1506 		 */
1507 		KASSERT(m->m_len <= maxlen &&
1508 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1509 		    m->m_len >= 0 : m->m_len > 0),
1510 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1511 
1512 		/*
1513 		 * TLS frames require unmapped mbufs to store session
1514 		 * info.
1515 		 */
1516 		KASSERT((m->m_flags & M_EXTPG) != 0,
1517 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1518 
1519 		tls_len = m->m_len;
1520 
1521 		/* Save a reference to the session. */
1522 		m->m_epg_tls = ktls_hold(tls);
1523 
1524 		m->m_epg_hdrlen = tls->params.tls_hlen;
1525 		m->m_epg_trllen = tls->params.tls_tlen;
1526 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1527 			int bs, delta;
1528 
1529 			/*
1530 			 * AES-CBC pads messages to a multiple of the
1531 			 * block size.  Note that the padding is
1532 			 * applied after the digest and the encryption
1533 			 * is done on the "plaintext || mac || padding".
1534 			 * At least one byte of padding is always
1535 			 * present.
1536 			 *
1537 			 * Compute the final trailer length assuming
1538 			 * at most one block of padding.
1539 			 * tls->params.sb_tls_tlen is the maximum
1540 			 * possible trailer length (padding + digest).
1541 			 * delta holds the number of excess padding
1542 			 * bytes if the maximum were used.  Those
1543 			 * extra bytes are removed.
1544 			 */
1545 			bs = tls->params.tls_bs;
1546 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1547 			m->m_epg_trllen -= delta;
1548 		}
1549 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1550 
1551 		/* Populate the TLS header. */
1552 		tlshdr = (void *)m->m_epg_hdr;
1553 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1554 
1555 		/*
1556 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1557 		 * of TLS_RLTYPE_APP.
1558 		 */
1559 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1560 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1561 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1562 			tlshdr->tls_type = TLS_RLTYPE_APP;
1563 			/* save the real record type for later */
1564 			m->m_epg_record_type = record_type;
1565 			m->m_epg_trail[0] = record_type;
1566 		} else {
1567 			tlshdr->tls_vminor = tls->params.tls_vminor;
1568 			tlshdr->tls_type = record_type;
1569 		}
1570 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1571 
1572 		/*
1573 		 * Store nonces / explicit IVs after the end of the
1574 		 * TLS header.
1575 		 *
1576 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1577 		 * from the end of the IV.  The nonce is then
1578 		 * incremented for use by the next record.
1579 		 *
1580 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1581 		 */
1582 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1583 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1584 			noncep = (uint64_t *)(tls->params.iv + 8);
1585 			be64enc(tlshdr + 1, *noncep);
1586 			(*noncep)++;
1587 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1588 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1589 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1590 
1591 		/*
1592 		 * When using SW encryption, mark the mbuf not ready.
1593 		 * It will be marked ready via sbready() after the
1594 		 * record has been encrypted.
1595 		 *
1596 		 * When using ifnet TLS, unencrypted TLS records are
1597 		 * sent down the stack to the NIC.
1598 		 */
1599 		if (tls->mode == TCP_TLS_MODE_SW) {
1600 			m->m_flags |= M_NOTREADY;
1601 			m->m_epg_nrdy = m->m_epg_npgs;
1602 			if (__predict_false(tls_len == 0)) {
1603 				/* TLS 1.0 empty fragment. */
1604 				*enq_cnt += 1;
1605 			} else
1606 				*enq_cnt += m->m_epg_npgs;
1607 		}
1608 	}
1609 }
1610 
1611 void
1612 ktls_check_rx(struct sockbuf *sb)
1613 {
1614 	struct tls_record_layer hdr;
1615 	struct ktls_wq *wq;
1616 	struct socket *so;
1617 	bool running;
1618 
1619 	SOCKBUF_LOCK_ASSERT(sb);
1620 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1621 	    __func__, sb));
1622 	so = __containerof(sb, struct socket, so_rcv);
1623 
1624 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1625 		return;
1626 
1627 	/* Is there enough queued for a TLS header? */
1628 	if (sb->sb_tlscc < sizeof(hdr)) {
1629 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1630 			so->so_error = EMSGSIZE;
1631 		return;
1632 	}
1633 
1634 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1635 
1636 	/* Is the entire record queued? */
1637 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1638 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1639 			so->so_error = EMSGSIZE;
1640 		return;
1641 	}
1642 
1643 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1644 
1645 	soref(so);
1646 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1647 	mtx_lock(&wq->mtx);
1648 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1649 	running = wq->running;
1650 	mtx_unlock(&wq->mtx);
1651 	if (!running)
1652 		wakeup(wq);
1653 	counter_u64_add(ktls_cnt_rx_queued, 1);
1654 }
1655 
1656 static struct mbuf *
1657 ktls_detach_record(struct sockbuf *sb, int len)
1658 {
1659 	struct mbuf *m, *n, *top;
1660 	int remain;
1661 
1662 	SOCKBUF_LOCK_ASSERT(sb);
1663 	MPASS(len <= sb->sb_tlscc);
1664 
1665 	/*
1666 	 * If TLS chain is the exact size of the record,
1667 	 * just grab the whole record.
1668 	 */
1669 	top = sb->sb_mtls;
1670 	if (sb->sb_tlscc == len) {
1671 		sb->sb_mtls = NULL;
1672 		sb->sb_mtlstail = NULL;
1673 		goto out;
1674 	}
1675 
1676 	/*
1677 	 * While it would be nice to use m_split() here, we need
1678 	 * to know exactly what m_split() allocates to update the
1679 	 * accounting, so do it inline instead.
1680 	 */
1681 	remain = len;
1682 	for (m = top; remain > m->m_len; m = m->m_next)
1683 		remain -= m->m_len;
1684 
1685 	/* Easy case: don't have to split 'm'. */
1686 	if (remain == m->m_len) {
1687 		sb->sb_mtls = m->m_next;
1688 		if (sb->sb_mtls == NULL)
1689 			sb->sb_mtlstail = NULL;
1690 		m->m_next = NULL;
1691 		goto out;
1692 	}
1693 
1694 	/*
1695 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1696 	 * with M_NOWAIT first.
1697 	 */
1698 	n = m_get(M_NOWAIT, MT_DATA);
1699 	if (n == NULL) {
1700 		/*
1701 		 * Use M_WAITOK with socket buffer unlocked.  If
1702 		 * 'sb_mtls' changes while the lock is dropped, return
1703 		 * NULL to force the caller to retry.
1704 		 */
1705 		SOCKBUF_UNLOCK(sb);
1706 
1707 		n = m_get(M_WAITOK, MT_DATA);
1708 
1709 		SOCKBUF_LOCK(sb);
1710 		if (sb->sb_mtls != top) {
1711 			m_free(n);
1712 			return (NULL);
1713 		}
1714 	}
1715 	n->m_flags |= M_NOTREADY;
1716 
1717 	/* Store remainder in 'n'. */
1718 	n->m_len = m->m_len - remain;
1719 	if (m->m_flags & M_EXT) {
1720 		n->m_data = m->m_data + remain;
1721 		mb_dupcl(n, m);
1722 	} else {
1723 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1724 	}
1725 
1726 	/* Trim 'm' and update accounting. */
1727 	m->m_len -= n->m_len;
1728 	sb->sb_tlscc -= n->m_len;
1729 	sb->sb_ccc -= n->m_len;
1730 
1731 	/* Account for 'n'. */
1732 	sballoc_ktls_rx(sb, n);
1733 
1734 	/* Insert 'n' into the TLS chain. */
1735 	sb->sb_mtls = n;
1736 	n->m_next = m->m_next;
1737 	if (sb->sb_mtlstail == m)
1738 		sb->sb_mtlstail = n;
1739 
1740 	/* Detach the record from the TLS chain. */
1741 	m->m_next = NULL;
1742 
1743 out:
1744 	MPASS(m_length(top, NULL) == len);
1745 	for (m = top; m != NULL; m = m->m_next)
1746 		sbfree_ktls_rx(sb, m);
1747 	sb->sb_tlsdcc = len;
1748 	sb->sb_ccc += len;
1749 	SBCHECK(sb);
1750 	return (top);
1751 }
1752 
1753 static void
1754 ktls_decrypt(struct socket *so)
1755 {
1756 	char tls_header[MBUF_PEXT_HDR_LEN];
1757 	struct ktls_session *tls;
1758 	struct sockbuf *sb;
1759 	struct tls_record_layer *hdr;
1760 	struct tls_get_record tgr;
1761 	struct mbuf *control, *data, *m;
1762 	uint64_t seqno;
1763 	int error, remain, tls_len, trail_len;
1764 
1765 	hdr = (struct tls_record_layer *)tls_header;
1766 	sb = &so->so_rcv;
1767 	SOCKBUF_LOCK(sb);
1768 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1769 	    ("%s: socket %p not running", __func__, so));
1770 
1771 	tls = sb->sb_tls_info;
1772 	MPASS(tls != NULL);
1773 
1774 	for (;;) {
1775 		/* Is there enough queued for a TLS header? */
1776 		if (sb->sb_tlscc < tls->params.tls_hlen)
1777 			break;
1778 
1779 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1780 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1781 
1782 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1783 		    hdr->tls_vminor != tls->params.tls_vminor)
1784 			error = EINVAL;
1785 		else if (tls_len < tls->params.tls_hlen || tls_len >
1786 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1787 		    tls->params.tls_tlen)
1788 			error = EMSGSIZE;
1789 		else
1790 			error = 0;
1791 		if (__predict_false(error != 0)) {
1792 			/*
1793 			 * We have a corrupted record and are likely
1794 			 * out of sync.  The connection isn't
1795 			 * recoverable at this point, so abort it.
1796 			 */
1797 			SOCKBUF_UNLOCK(sb);
1798 			counter_u64_add(ktls_offload_corrupted_records, 1);
1799 
1800 			CURVNET_SET(so->so_vnet);
1801 			so->so_proto->pr_usrreqs->pru_abort(so);
1802 			so->so_error = error;
1803 			CURVNET_RESTORE();
1804 			goto deref;
1805 		}
1806 
1807 		/* Is the entire record queued? */
1808 		if (sb->sb_tlscc < tls_len)
1809 			break;
1810 
1811 		/*
1812 		 * Split out the portion of the mbuf chain containing
1813 		 * this TLS record.
1814 		 */
1815 		data = ktls_detach_record(sb, tls_len);
1816 		if (data == NULL)
1817 			continue;
1818 		MPASS(sb->sb_tlsdcc == tls_len);
1819 
1820 		seqno = sb->sb_tls_seqno;
1821 		sb->sb_tls_seqno++;
1822 		SBCHECK(sb);
1823 		SOCKBUF_UNLOCK(sb);
1824 
1825 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1826 		if (error) {
1827 			counter_u64_add(ktls_offload_failed_crypto, 1);
1828 
1829 			SOCKBUF_LOCK(sb);
1830 			if (sb->sb_tlsdcc == 0) {
1831 				/*
1832 				 * sbcut/drop/flush discarded these
1833 				 * mbufs.
1834 				 */
1835 				m_freem(data);
1836 				break;
1837 			}
1838 
1839 			/*
1840 			 * Drop this TLS record's data, but keep
1841 			 * decrypting subsequent records.
1842 			 */
1843 			sb->sb_ccc -= tls_len;
1844 			sb->sb_tlsdcc = 0;
1845 
1846 			CURVNET_SET(so->so_vnet);
1847 			so->so_error = EBADMSG;
1848 			sorwakeup_locked(so);
1849 			CURVNET_RESTORE();
1850 
1851 			m_freem(data);
1852 
1853 			SOCKBUF_LOCK(sb);
1854 			continue;
1855 		}
1856 
1857 		/* Allocate the control mbuf. */
1858 		tgr.tls_type = hdr->tls_type;
1859 		tgr.tls_vmajor = hdr->tls_vmajor;
1860 		tgr.tls_vminor = hdr->tls_vminor;
1861 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1862 		    trail_len);
1863 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1864 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1865 
1866 		SOCKBUF_LOCK(sb);
1867 		if (sb->sb_tlsdcc == 0) {
1868 			/* sbcut/drop/flush discarded these mbufs. */
1869 			MPASS(sb->sb_tlscc == 0);
1870 			m_freem(data);
1871 			m_freem(control);
1872 			break;
1873 		}
1874 
1875 		/*
1876 		 * Clear the 'dcc' accounting in preparation for
1877 		 * adding the decrypted record.
1878 		 */
1879 		sb->sb_ccc -= tls_len;
1880 		sb->sb_tlsdcc = 0;
1881 		SBCHECK(sb);
1882 
1883 		/* If there is no payload, drop all of the data. */
1884 		if (tgr.tls_length == htobe16(0)) {
1885 			m_freem(data);
1886 			data = NULL;
1887 		} else {
1888 			/* Trim header. */
1889 			remain = tls->params.tls_hlen;
1890 			while (remain > 0) {
1891 				if (data->m_len > remain) {
1892 					data->m_data += remain;
1893 					data->m_len -= remain;
1894 					break;
1895 				}
1896 				remain -= data->m_len;
1897 				data = m_free(data);
1898 			}
1899 
1900 			/* Trim trailer and clear M_NOTREADY. */
1901 			remain = be16toh(tgr.tls_length);
1902 			m = data;
1903 			for (m = data; remain > m->m_len; m = m->m_next) {
1904 				m->m_flags &= ~M_NOTREADY;
1905 				remain -= m->m_len;
1906 			}
1907 			m->m_len = remain;
1908 			m_freem(m->m_next);
1909 			m->m_next = NULL;
1910 			m->m_flags &= ~M_NOTREADY;
1911 
1912 			/* Set EOR on the final mbuf. */
1913 			m->m_flags |= M_EOR;
1914 		}
1915 
1916 		sbappendcontrol_locked(sb, data, control, 0);
1917 	}
1918 
1919 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1920 
1921 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1922 		so->so_error = EMSGSIZE;
1923 
1924 	sorwakeup_locked(so);
1925 
1926 deref:
1927 	SOCKBUF_UNLOCK_ASSERT(sb);
1928 
1929 	CURVNET_SET(so->so_vnet);
1930 	SOCK_LOCK(so);
1931 	sorele(so);
1932 	CURVNET_RESTORE();
1933 }
1934 
1935 void
1936 ktls_enqueue_to_free(struct mbuf *m)
1937 {
1938 	struct ktls_wq *wq;
1939 	bool running;
1940 
1941 	/* Mark it for freeing. */
1942 	m->m_epg_flags |= EPG_FLAG_2FREE;
1943 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1944 	mtx_lock(&wq->mtx);
1945 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1946 	running = wq->running;
1947 	mtx_unlock(&wq->mtx);
1948 	if (!running)
1949 		wakeup(wq);
1950 }
1951 
1952 void
1953 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1954 {
1955 	struct ktls_wq *wq;
1956 	bool running;
1957 
1958 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1959 	    (M_EXTPG | M_NOTREADY)),
1960 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1961 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1962 
1963 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1964 
1965 	m->m_epg_enc_cnt = page_count;
1966 
1967 	/*
1968 	 * Save a pointer to the socket.  The caller is responsible
1969 	 * for taking an additional reference via soref().
1970 	 */
1971 	m->m_epg_so = so;
1972 
1973 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1974 	mtx_lock(&wq->mtx);
1975 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1976 	running = wq->running;
1977 	mtx_unlock(&wq->mtx);
1978 	if (!running)
1979 		wakeup(wq);
1980 	counter_u64_add(ktls_cnt_tx_queued, 1);
1981 }
1982 
1983 static __noinline void
1984 ktls_encrypt(struct mbuf *top)
1985 {
1986 	struct ktls_session *tls;
1987 	struct socket *so;
1988 	struct mbuf *m;
1989 	vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1990 	struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1991 	struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1992 	vm_page_t pg;
1993 	int error, i, len, npages, off, total_pages;
1994 	bool is_anon;
1995 
1996 	so = top->m_epg_so;
1997 	tls = top->m_epg_tls;
1998 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
1999 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2000 #ifdef INVARIANTS
2001 	top->m_epg_so = NULL;
2002 #endif
2003 	total_pages = top->m_epg_enc_cnt;
2004 	npages = 0;
2005 
2006 	/*
2007 	 * Encrypt the TLS records in the chain of mbufs starting with
2008 	 * 'top'.  'total_pages' gives us a total count of pages and is
2009 	 * used to know when we have finished encrypting the TLS
2010 	 * records originally queued with 'top'.
2011 	 *
2012 	 * NB: These mbufs are queued in the socket buffer and
2013 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2014 	 * socket buffer lock is not held while traversing this chain.
2015 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2016 	 * pointers should be stable.  However, the 'm_next' of the
2017 	 * last mbuf encrypted is not necessarily NULL.  It can point
2018 	 * to other mbufs appended while 'top' was on the TLS work
2019 	 * queue.
2020 	 *
2021 	 * Each mbuf holds an entire TLS record.
2022 	 */
2023 	error = 0;
2024 	for (m = top; npages != total_pages; m = m->m_next) {
2025 		KASSERT(m->m_epg_tls == tls,
2026 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2027 		    tls, m->m_epg_tls));
2028 		KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2029 		    (M_EXTPG | M_NOTREADY),
2030 		    ("%p not unready & nomap mbuf (top = %p)\n", m, top));
2031 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2032 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2033 		    total_pages, m));
2034 
2035 		/*
2036 		 * Generate source and destination ivoecs to pass to
2037 		 * the SW encryption backend.  For writable mbufs, the
2038 		 * destination iovec is a copy of the source and
2039 		 * encryption is done in place.  For file-backed mbufs
2040 		 * (from sendfile), anonymous wired pages are
2041 		 * allocated and assigned to the destination iovec.
2042 		 */
2043 		is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
2044 
2045 		off = m->m_epg_1st_off;
2046 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2047 			len = m_epg_pagelen(m, i, off);
2048 			src_iov[i].iov_len = len;
2049 			src_iov[i].iov_base =
2050 			    (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
2051 				off;
2052 
2053 			if (is_anon) {
2054 				dst_iov[i].iov_base = src_iov[i].iov_base;
2055 				dst_iov[i].iov_len = src_iov[i].iov_len;
2056 				continue;
2057 			}
2058 retry_page:
2059 			pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2060 			    VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
2061 			if (pg == NULL) {
2062 				vm_wait(NULL);
2063 				goto retry_page;
2064 			}
2065 			parray[i] = VM_PAGE_TO_PHYS(pg);
2066 			dst_iov[i].iov_base =
2067 			    (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
2068 			dst_iov[i].iov_len = len;
2069 		}
2070 
2071 		if (__predict_false(m->m_epg_npgs == 0)) {
2072 			/* TLS 1.0 empty fragment. */
2073 			npages++;
2074 		} else
2075 			npages += i;
2076 
2077 		error = (*tls->sw_encrypt)(tls,
2078 		    (const struct tls_record_layer *)m->m_epg_hdr,
2079 		    m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
2080 		    m->m_epg_record_type);
2081 		if (error) {
2082 			counter_u64_add(ktls_offload_failed_crypto, 1);
2083 			break;
2084 		}
2085 
2086 		/*
2087 		 * For file-backed mbufs, release the file-backed
2088 		 * pages and replace them in the ext_pgs array with
2089 		 * the anonymous wired pages allocated above.
2090 		 */
2091 		if (!is_anon) {
2092 			/* Free the old pages. */
2093 			m->m_ext.ext_free(m);
2094 
2095 			/* Replace them with the new pages. */
2096 			for (i = 0; i < m->m_epg_npgs; i++)
2097 				m->m_epg_pa[i] = parray[i];
2098 
2099 			/* Use the basic free routine. */
2100 			m->m_ext.ext_free = mb_free_mext_pgs;
2101 
2102 			/* Pages are now writable. */
2103 			m->m_epg_flags |= EPG_FLAG_ANON;
2104 		}
2105 
2106 		/*
2107 		 * Drop a reference to the session now that it is no
2108 		 * longer needed.  Existing code depends on encrypted
2109 		 * records having no associated session vs
2110 		 * yet-to-be-encrypted records having an associated
2111 		 * session.
2112 		 */
2113 		m->m_epg_tls = NULL;
2114 		ktls_free(tls);
2115 	}
2116 
2117 	CURVNET_SET(so->so_vnet);
2118 	if (error == 0) {
2119 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2120 	} else {
2121 		so->so_proto->pr_usrreqs->pru_abort(so);
2122 		so->so_error = EIO;
2123 		mb_free_notready(top, total_pages);
2124 	}
2125 
2126 	SOCK_LOCK(so);
2127 	sorele(so);
2128 	CURVNET_RESTORE();
2129 }
2130 
2131 static void
2132 ktls_work_thread(void *ctx)
2133 {
2134 	struct ktls_wq *wq = ctx;
2135 	struct mbuf *m, *n;
2136 	struct socket *so, *son;
2137 	STAILQ_HEAD(, mbuf) local_m_head;
2138 	STAILQ_HEAD(, socket) local_so_head;
2139 
2140 	if (ktls_bind_threads > 1) {
2141 		curthread->td_domain.dr_policy =
2142 			DOMAINSET_PREF(PCPU_GET(domain));
2143 	}
2144 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2145 	fpu_kern_thread(0);
2146 #endif
2147 	for (;;) {
2148 		mtx_lock(&wq->mtx);
2149 		while (STAILQ_EMPTY(&wq->m_head) &&
2150 		    STAILQ_EMPTY(&wq->so_head)) {
2151 			wq->running = false;
2152 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2153 			wq->running = true;
2154 		}
2155 
2156 		STAILQ_INIT(&local_m_head);
2157 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2158 		STAILQ_INIT(&local_so_head);
2159 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2160 		mtx_unlock(&wq->mtx);
2161 
2162 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2163 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2164 				ktls_free(m->m_epg_tls);
2165 				uma_zfree(zone_mbuf, m);
2166 			} else {
2167 				ktls_encrypt(m);
2168 				counter_u64_add(ktls_cnt_tx_queued, -1);
2169 			}
2170 		}
2171 
2172 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2173 			ktls_decrypt(so);
2174 			counter_u64_add(ktls_cnt_rx_queued, -1);
2175 		}
2176 	}
2177 }
2178