xref: /freebsd/sys/kern/uipc_ktls.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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 	int		lastallocfail;
86 } __aligned(CACHE_LINE_SIZE);
87 
88 struct ktls_domain_info {
89 	int count;
90 	int cpu[MAXCPU];
91 };
92 
93 struct ktls_domain_info ktls_domains[MAXMEMDOM];
94 static struct ktls_wq *ktls_wq;
95 static struct proc *ktls_proc;
96 static uma_zone_t ktls_session_zone;
97 static uma_zone_t ktls_buffer_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 #ifdef RSS
106 static int ktls_bind_threads = 1;
107 #else
108 static int ktls_bind_threads;
109 #endif
110 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
111     &ktls_bind_threads, 0,
112     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
113 
114 static u_int ktls_maxlen = 16384;
115 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
116     &ktls_maxlen, 0, "Maximum TLS record size");
117 
118 static int ktls_number_threads;
119 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
120     &ktls_number_threads, 0,
121     "Number of TLS threads in thread-pool");
122 
123 static bool ktls_offload_enable;
124 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
125     &ktls_offload_enable, 0,
126     "Enable support for kernel TLS offload");
127 
128 static bool ktls_cbc_enable = true;
129 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
130     &ktls_cbc_enable, 1,
131     "Enable Support of AES-CBC crypto for kernel TLS");
132 
133 static bool ktls_sw_buffer_cache = true;
134 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
135     &ktls_sw_buffer_cache, 1,
136     "Enable caching of output buffers for SW encryption");
137 
138 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
139 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
140     &ktls_tasks_active, "Number of active tasks");
141 
142 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
143 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
144     &ktls_cnt_tx_queued,
145     "Number of TLS records in queue to tasks for SW encryption");
146 
147 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
148 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
149     &ktls_cnt_rx_queued,
150     "Number of TLS sockets in queue to tasks for SW decryption");
151 
152 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
153 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
154     CTLFLAG_RD, &ktls_offload_total,
155     "Total successful TLS setups (parameters set)");
156 
157 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
158 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
159     CTLFLAG_RD, &ktls_offload_enable_calls,
160     "Total number of TLS enable calls made");
161 
162 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
163 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
164     &ktls_offload_active, "Total Active TLS sessions");
165 
166 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
167 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
168     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
169 
170 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
171 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
172     &ktls_offload_failed_crypto, "Total TLS crypto failures");
173 
174 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
175 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
176     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
180     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
181 
182 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
183 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
184     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
185 
186 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
187     "Software TLS session stats");
188 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
189     "Hardware (ifnet) TLS session stats");
190 #ifdef TCP_OFFLOAD
191 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
192     "TOE TLS session stats");
193 #endif
194 
195 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
197     "Active number of software TLS sessions using AES-CBC");
198 
199 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
200 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
201     "Active number of software TLS sessions using AES-GCM");
202 
203 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
204 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
205     &ktls_sw_chacha20,
206     "Active number of software TLS sessions using Chacha20-Poly1305");
207 
208 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
209 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
210     &ktls_ifnet_cbc,
211     "Active number of ifnet TLS sessions using AES-CBC");
212 
213 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
215     &ktls_ifnet_gcm,
216     "Active number of ifnet TLS sessions using AES-GCM");
217 
218 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
219 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
220     &ktls_ifnet_chacha20,
221     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
222 
223 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
224 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
225     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
226 
227 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
228 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
229     &ktls_ifnet_reset_dropped,
230     "TLS sessions dropped after failing to update ifnet send tag");
231 
232 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
233 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
234     &ktls_ifnet_reset_failed,
235     "TLS sessions that failed to allocate a new ifnet send tag");
236 
237 static int ktls_ifnet_permitted;
238 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
239     &ktls_ifnet_permitted, 1,
240     "Whether to permit hardware (ifnet) TLS sessions");
241 
242 #ifdef TCP_OFFLOAD
243 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
244 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
245     &ktls_toe_cbc,
246     "Active number of TOE TLS sessions using AES-CBC");
247 
248 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
249 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
250     &ktls_toe_gcm,
251     "Active number of TOE TLS sessions using AES-GCM");
252 
253 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
254 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
255     &ktls_toe_chacha20,
256     "Active number of TOE TLS sessions using Chacha20-Poly1305");
257 #endif
258 
259 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
260 
261 static void ktls_cleanup(struct ktls_session *tls);
262 #if defined(INET) || defined(INET6)
263 static void ktls_reset_send_tag(void *context, int pending);
264 #endif
265 static void ktls_work_thread(void *ctx);
266 
267 #if defined(INET) || defined(INET6)
268 static u_int
269 ktls_get_cpu(struct socket *so)
270 {
271 	struct inpcb *inp;
272 #ifdef NUMA
273 	struct ktls_domain_info *di;
274 #endif
275 	u_int cpuid;
276 
277 	inp = sotoinpcb(so);
278 #ifdef RSS
279 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
280 	if (cpuid != NETISR_CPUID_NONE)
281 		return (cpuid);
282 #endif
283 	/*
284 	 * Just use the flowid to shard connections in a repeatable
285 	 * fashion.  Note that TLS 1.0 sessions rely on the
286 	 * serialization provided by having the same connection use
287 	 * the same queue.
288 	 */
289 #ifdef NUMA
290 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
291 		di = &ktls_domains[inp->inp_numa_domain];
292 		cpuid = di->cpu[inp->inp_flowid % di->count];
293 	} else
294 #endif
295 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
296 	return (cpuid);
297 }
298 #endif
299 
300 static int
301 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
302 {
303 	vm_page_t m;
304 	int i;
305 
306 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
307 	    ("%s: ktls max length %d is not page size-aligned",
308 	    __func__, ktls_maxlen));
309 
310 	for (i = 0; i < count; i++) {
311 		m = vm_page_alloc_contig_domain(NULL, 0, domain,
312 		    VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
313 		    VM_ALLOC_NODUMP | malloc2vm_flags(flags),
314 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
315 		    VM_MEMATTR_DEFAULT);
316 		if (m == NULL)
317 			break;
318 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
319 	}
320 	return (i);
321 }
322 
323 static void
324 ktls_buffer_release(void *arg __unused, void **store, int count)
325 {
326 	vm_page_t m;
327 	int i, j;
328 
329 	for (i = 0; i < count; i++) {
330 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
331 		for (j = 0; j < atop(ktls_maxlen); j++) {
332 			(void)vm_page_unwire_noq(m + j);
333 			vm_page_free(m + j);
334 		}
335 	}
336 }
337 
338 static void
339 ktls_free_mext_contig(struct mbuf *m)
340 {
341 	M_ASSERTEXTPG(m);
342 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
343 }
344 
345 static void
346 ktls_init(void *dummy __unused)
347 {
348 	struct thread *td;
349 	struct pcpu *pc;
350 	cpuset_t mask;
351 	int count, domain, error, i;
352 
353 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
354 	    M_WAITOK | M_ZERO);
355 
356 	ktls_session_zone = uma_zcreate("ktls_session",
357 	    sizeof(struct ktls_session),
358 	    NULL, NULL, NULL, NULL,
359 	    UMA_ALIGN_CACHE, 0);
360 
361 	if (ktls_sw_buffer_cache) {
362 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
363 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
364 		    ktls_buffer_import, ktls_buffer_release, NULL,
365 		    UMA_ZONE_FIRSTTOUCH);
366 	}
367 
368 	/*
369 	 * Initialize the workqueues to run the TLS work.  We create a
370 	 * work queue for each CPU.
371 	 */
372 	CPU_FOREACH(i) {
373 		STAILQ_INIT(&ktls_wq[i].m_head);
374 		STAILQ_INIT(&ktls_wq[i].so_head);
375 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
376 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
377 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
378 		if (error)
379 			panic("Can't add KTLS thread %d error %d", i, error);
380 
381 		/*
382 		 * Bind threads to cores.  If ktls_bind_threads is >
383 		 * 1, then we bind to the NUMA domain.
384 		 */
385 		if (ktls_bind_threads) {
386 			if (ktls_bind_threads > 1) {
387 				pc = pcpu_find(i);
388 				domain = pc->pc_domain;
389 				CPU_COPY(&cpuset_domain[domain], &mask);
390 				count = ktls_domains[domain].count;
391 				ktls_domains[domain].cpu[count] = i;
392 				ktls_domains[domain].count++;
393 			} else {
394 				CPU_SETOF(i, &mask);
395 			}
396 			error = cpuset_setthread(td->td_tid, &mask);
397 			if (error)
398 				panic(
399 			    "Unable to bind KTLS thread for CPU %d error %d",
400 				     i, error);
401 		}
402 		ktls_cpuid_lookup[ktls_number_threads] = i;
403 		ktls_number_threads++;
404 	}
405 
406 	/*
407 	 * If we somehow have an empty domain, fall back to choosing
408 	 * among all KTLS threads.
409 	 */
410 	if (ktls_bind_threads > 1) {
411 		for (i = 0; i < vm_ndomains; i++) {
412 			if (ktls_domains[i].count == 0) {
413 				ktls_bind_threads = 1;
414 				break;
415 			}
416 		}
417 	}
418 
419 	if (bootverbose)
420 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
421 }
422 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
423 
424 #if defined(INET) || defined(INET6)
425 static int
426 ktls_create_session(struct socket *so, struct tls_enable *en,
427     struct ktls_session **tlsp)
428 {
429 	struct ktls_session *tls;
430 	int error;
431 
432 	/* Only TLS 1.0 - 1.3 are supported. */
433 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
434 		return (EINVAL);
435 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
436 	    en->tls_vminor > TLS_MINOR_VER_THREE)
437 		return (EINVAL);
438 
439 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
440 		return (EINVAL);
441 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
442 		return (EINVAL);
443 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
444 		return (EINVAL);
445 
446 	/* All supported algorithms require a cipher key. */
447 	if (en->cipher_key_len == 0)
448 		return (EINVAL);
449 
450 	/* No flags are currently supported. */
451 	if (en->flags != 0)
452 		return (EINVAL);
453 
454 	/* Common checks for supported algorithms. */
455 	switch (en->cipher_algorithm) {
456 	case CRYPTO_AES_NIST_GCM_16:
457 		/*
458 		 * auth_algorithm isn't used, but permit GMAC values
459 		 * for compatibility.
460 		 */
461 		switch (en->auth_algorithm) {
462 		case 0:
463 #ifdef COMPAT_FREEBSD12
464 		/* XXX: Really 13.0-current COMPAT. */
465 		case CRYPTO_AES_128_NIST_GMAC:
466 		case CRYPTO_AES_192_NIST_GMAC:
467 		case CRYPTO_AES_256_NIST_GMAC:
468 #endif
469 			break;
470 		default:
471 			return (EINVAL);
472 		}
473 		if (en->auth_key_len != 0)
474 			return (EINVAL);
475 		if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
476 			en->iv_len != TLS_AEAD_GCM_LEN) ||
477 		    (en->tls_vminor == TLS_MINOR_VER_THREE &&
478 			en->iv_len != TLS_1_3_GCM_IV_LEN))
479 			return (EINVAL);
480 		break;
481 	case CRYPTO_AES_CBC:
482 		switch (en->auth_algorithm) {
483 		case CRYPTO_SHA1_HMAC:
484 			/*
485 			 * TLS 1.0 requires an implicit IV.  TLS 1.1+
486 			 * all use explicit IVs.
487 			 */
488 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
489 				if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
490 					return (EINVAL);
491 				break;
492 			}
493 
494 			/* FALLTHROUGH */
495 		case CRYPTO_SHA2_256_HMAC:
496 		case CRYPTO_SHA2_384_HMAC:
497 			/* Ignore any supplied IV. */
498 			en->iv_len = 0;
499 			break;
500 		default:
501 			return (EINVAL);
502 		}
503 		if (en->auth_key_len == 0)
504 			return (EINVAL);
505 		break;
506 	case CRYPTO_CHACHA20_POLY1305:
507 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
508 			return (EINVAL);
509 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
510 		    en->tls_vminor != TLS_MINOR_VER_THREE)
511 			return (EINVAL);
512 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
513 			return (EINVAL);
514 		break;
515 	default:
516 		return (EINVAL);
517 	}
518 
519 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
520 
521 	counter_u64_add(ktls_offload_active, 1);
522 
523 	refcount_init(&tls->refcount, 1);
524 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
525 
526 	tls->wq_index = ktls_get_cpu(so);
527 
528 	tls->params.cipher_algorithm = en->cipher_algorithm;
529 	tls->params.auth_algorithm = en->auth_algorithm;
530 	tls->params.tls_vmajor = en->tls_vmajor;
531 	tls->params.tls_vminor = en->tls_vminor;
532 	tls->params.flags = en->flags;
533 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
534 
535 	/* Set the header and trailer lengths. */
536 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
537 	switch (en->cipher_algorithm) {
538 	case CRYPTO_AES_NIST_GCM_16:
539 		/*
540 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
541 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
542 		 */
543 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
544 			tls->params.tls_hlen += sizeof(uint64_t);
545 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
546 		tls->params.tls_bs = 1;
547 		break;
548 	case CRYPTO_AES_CBC:
549 		switch (en->auth_algorithm) {
550 		case CRYPTO_SHA1_HMAC:
551 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
552 				/* Implicit IV, no nonce. */
553 			} else {
554 				tls->params.tls_hlen += AES_BLOCK_LEN;
555 			}
556 			tls->params.tls_tlen = AES_BLOCK_LEN +
557 			    SHA1_HASH_LEN;
558 			break;
559 		case CRYPTO_SHA2_256_HMAC:
560 			tls->params.tls_hlen += AES_BLOCK_LEN;
561 			tls->params.tls_tlen = AES_BLOCK_LEN +
562 			    SHA2_256_HASH_LEN;
563 			break;
564 		case CRYPTO_SHA2_384_HMAC:
565 			tls->params.tls_hlen += AES_BLOCK_LEN;
566 			tls->params.tls_tlen = AES_BLOCK_LEN +
567 			    SHA2_384_HASH_LEN;
568 			break;
569 		default:
570 			panic("invalid hmac");
571 		}
572 		tls->params.tls_bs = AES_BLOCK_LEN;
573 		break;
574 	case CRYPTO_CHACHA20_POLY1305:
575 		/*
576 		 * Chacha20 uses a 12 byte implicit IV.
577 		 */
578 		tls->params.tls_tlen = POLY1305_HASH_LEN;
579 		tls->params.tls_bs = 1;
580 		break;
581 	default:
582 		panic("invalid cipher");
583 	}
584 
585 	/*
586 	 * TLS 1.3 includes optional padding which we do not support,
587 	 * and also puts the "real" record type at the end of the
588 	 * encrypted data.
589 	 */
590 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
591 		tls->params.tls_tlen += sizeof(uint8_t);
592 
593 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
594 	    ("TLS header length too long: %d", tls->params.tls_hlen));
595 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
596 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
597 
598 	if (en->auth_key_len != 0) {
599 		tls->params.auth_key_len = en->auth_key_len;
600 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
601 		    M_WAITOK);
602 		error = copyin(en->auth_key, tls->params.auth_key,
603 		    en->auth_key_len);
604 		if (error)
605 			goto out;
606 	}
607 
608 	tls->params.cipher_key_len = en->cipher_key_len;
609 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
610 	error = copyin(en->cipher_key, tls->params.cipher_key,
611 	    en->cipher_key_len);
612 	if (error)
613 		goto out;
614 
615 	/*
616 	 * This holds the implicit portion of the nonce for AEAD
617 	 * ciphers and the initial implicit IV for TLS 1.0.  The
618 	 * explicit portions of the IV are generated in ktls_frame().
619 	 */
620 	if (en->iv_len != 0) {
621 		tls->params.iv_len = en->iv_len;
622 		error = copyin(en->iv, tls->params.iv, en->iv_len);
623 		if (error)
624 			goto out;
625 
626 		/*
627 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
628 		 * counter to generate unique explicit IVs.
629 		 *
630 		 * Store this counter in the last 8 bytes of the IV
631 		 * array so that it is 8-byte aligned.
632 		 */
633 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
634 		    en->tls_vminor == TLS_MINOR_VER_TWO)
635 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
636 	}
637 
638 	*tlsp = tls;
639 	return (0);
640 
641 out:
642 	ktls_cleanup(tls);
643 	return (error);
644 }
645 
646 static struct ktls_session *
647 ktls_clone_session(struct ktls_session *tls)
648 {
649 	struct ktls_session *tls_new;
650 
651 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
652 
653 	counter_u64_add(ktls_offload_active, 1);
654 
655 	refcount_init(&tls_new->refcount, 1);
656 
657 	/* Copy fields from existing session. */
658 	tls_new->params = tls->params;
659 	tls_new->wq_index = tls->wq_index;
660 
661 	/* Deep copy keys. */
662 	if (tls_new->params.auth_key != NULL) {
663 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
664 		    M_KTLS, M_WAITOK);
665 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
666 		    tls->params.auth_key_len);
667 	}
668 
669 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
670 	    M_WAITOK);
671 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
672 	    tls->params.cipher_key_len);
673 
674 	return (tls_new);
675 }
676 #endif
677 
678 static void
679 ktls_cleanup(struct ktls_session *tls)
680 {
681 
682 	counter_u64_add(ktls_offload_active, -1);
683 	switch (tls->mode) {
684 	case TCP_TLS_MODE_SW:
685 		switch (tls->params.cipher_algorithm) {
686 		case CRYPTO_AES_CBC:
687 			counter_u64_add(ktls_sw_cbc, -1);
688 			break;
689 		case CRYPTO_AES_NIST_GCM_16:
690 			counter_u64_add(ktls_sw_gcm, -1);
691 			break;
692 		case CRYPTO_CHACHA20_POLY1305:
693 			counter_u64_add(ktls_sw_chacha20, -1);
694 			break;
695 		}
696 		ktls_ocf_free(tls);
697 		break;
698 	case TCP_TLS_MODE_IFNET:
699 		switch (tls->params.cipher_algorithm) {
700 		case CRYPTO_AES_CBC:
701 			counter_u64_add(ktls_ifnet_cbc, -1);
702 			break;
703 		case CRYPTO_AES_NIST_GCM_16:
704 			counter_u64_add(ktls_ifnet_gcm, -1);
705 			break;
706 		case CRYPTO_CHACHA20_POLY1305:
707 			counter_u64_add(ktls_ifnet_chacha20, -1);
708 			break;
709 		}
710 		if (tls->snd_tag != NULL)
711 			m_snd_tag_rele(tls->snd_tag);
712 		break;
713 #ifdef TCP_OFFLOAD
714 	case TCP_TLS_MODE_TOE:
715 		switch (tls->params.cipher_algorithm) {
716 		case CRYPTO_AES_CBC:
717 			counter_u64_add(ktls_toe_cbc, -1);
718 			break;
719 		case CRYPTO_AES_NIST_GCM_16:
720 			counter_u64_add(ktls_toe_gcm, -1);
721 			break;
722 		case CRYPTO_CHACHA20_POLY1305:
723 			counter_u64_add(ktls_toe_chacha20, -1);
724 			break;
725 		}
726 		break;
727 #endif
728 	}
729 	if (tls->params.auth_key != NULL) {
730 		zfree(tls->params.auth_key, M_KTLS);
731 		tls->params.auth_key = NULL;
732 		tls->params.auth_key_len = 0;
733 	}
734 	if (tls->params.cipher_key != NULL) {
735 		zfree(tls->params.cipher_key, M_KTLS);
736 		tls->params.cipher_key = NULL;
737 		tls->params.cipher_key_len = 0;
738 	}
739 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
740 }
741 
742 #if defined(INET) || defined(INET6)
743 
744 #ifdef TCP_OFFLOAD
745 static int
746 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
747 {
748 	struct inpcb *inp;
749 	struct tcpcb *tp;
750 	int error;
751 
752 	inp = so->so_pcb;
753 	INP_WLOCK(inp);
754 	if (inp->inp_flags2 & INP_FREED) {
755 		INP_WUNLOCK(inp);
756 		return (ECONNRESET);
757 	}
758 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
759 		INP_WUNLOCK(inp);
760 		return (ECONNRESET);
761 	}
762 	if (inp->inp_socket == NULL) {
763 		INP_WUNLOCK(inp);
764 		return (ECONNRESET);
765 	}
766 	tp = intotcpcb(inp);
767 	if (!(tp->t_flags & TF_TOE)) {
768 		INP_WUNLOCK(inp);
769 		return (EOPNOTSUPP);
770 	}
771 
772 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
773 	INP_WUNLOCK(inp);
774 	if (error == 0) {
775 		tls->mode = TCP_TLS_MODE_TOE;
776 		switch (tls->params.cipher_algorithm) {
777 		case CRYPTO_AES_CBC:
778 			counter_u64_add(ktls_toe_cbc, 1);
779 			break;
780 		case CRYPTO_AES_NIST_GCM_16:
781 			counter_u64_add(ktls_toe_gcm, 1);
782 			break;
783 		case CRYPTO_CHACHA20_POLY1305:
784 			counter_u64_add(ktls_toe_chacha20, 1);
785 			break;
786 		}
787 	}
788 	return (error);
789 }
790 #endif
791 
792 /*
793  * Common code used when first enabling ifnet TLS on a connection or
794  * when allocating a new ifnet TLS session due to a routing change.
795  * This function allocates a new TLS send tag on whatever interface
796  * the connection is currently routed over.
797  */
798 static int
799 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
800     struct m_snd_tag **mstp)
801 {
802 	union if_snd_tag_alloc_params params;
803 	struct ifnet *ifp;
804 	struct nhop_object *nh;
805 	struct tcpcb *tp;
806 	int error;
807 
808 	INP_RLOCK(inp);
809 	if (inp->inp_flags2 & INP_FREED) {
810 		INP_RUNLOCK(inp);
811 		return (ECONNRESET);
812 	}
813 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
814 		INP_RUNLOCK(inp);
815 		return (ECONNRESET);
816 	}
817 	if (inp->inp_socket == NULL) {
818 		INP_RUNLOCK(inp);
819 		return (ECONNRESET);
820 	}
821 	tp = intotcpcb(inp);
822 
823 	/*
824 	 * Check administrative controls on ifnet TLS to determine if
825 	 * ifnet TLS should be denied.
826 	 *
827 	 * - Always permit 'force' requests.
828 	 * - ktls_ifnet_permitted == 0: always deny.
829 	 */
830 	if (!force && ktls_ifnet_permitted == 0) {
831 		INP_RUNLOCK(inp);
832 		return (ENXIO);
833 	}
834 
835 	/*
836 	 * XXX: Use the cached route in the inpcb to find the
837 	 * interface.  This should perhaps instead use
838 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
839 	 * enabled after a connection has completed key negotiation in
840 	 * userland, the cached route will be present in practice.
841 	 */
842 	nh = inp->inp_route.ro_nh;
843 	if (nh == NULL) {
844 		INP_RUNLOCK(inp);
845 		return (ENXIO);
846 	}
847 	ifp = nh->nh_ifp;
848 	if_ref(ifp);
849 
850 	/*
851 	 * Allocate a TLS + ratelimit tag if the connection has an
852 	 * existing pacing rate.
853 	 */
854 	if (tp->t_pacing_rate != -1 &&
855 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
856 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
857 		params.tls_rate_limit.inp = inp;
858 		params.tls_rate_limit.tls = tls;
859 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
860 	} else {
861 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
862 		params.tls.inp = inp;
863 		params.tls.tls = tls;
864 	}
865 	params.hdr.flowid = inp->inp_flowid;
866 	params.hdr.flowtype = inp->inp_flowtype;
867 	params.hdr.numa_domain = inp->inp_numa_domain;
868 	INP_RUNLOCK(inp);
869 
870 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
871 		error = EOPNOTSUPP;
872 		goto out;
873 	}
874 	if (inp->inp_vflag & INP_IPV6) {
875 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
876 			error = EOPNOTSUPP;
877 			goto out;
878 		}
879 	} else {
880 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
881 			error = EOPNOTSUPP;
882 			goto out;
883 		}
884 	}
885 	error = m_snd_tag_alloc(ifp, &params, mstp);
886 out:
887 	if_rele(ifp);
888 	return (error);
889 }
890 
891 static int
892 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
893 {
894 	struct m_snd_tag *mst;
895 	int error;
896 
897 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
898 	if (error == 0) {
899 		tls->mode = TCP_TLS_MODE_IFNET;
900 		tls->snd_tag = mst;
901 		switch (tls->params.cipher_algorithm) {
902 		case CRYPTO_AES_CBC:
903 			counter_u64_add(ktls_ifnet_cbc, 1);
904 			break;
905 		case CRYPTO_AES_NIST_GCM_16:
906 			counter_u64_add(ktls_ifnet_gcm, 1);
907 			break;
908 		case CRYPTO_CHACHA20_POLY1305:
909 			counter_u64_add(ktls_ifnet_chacha20, 1);
910 			break;
911 		}
912 	}
913 	return (error);
914 }
915 
916 static int
917 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
918 {
919 	int error;
920 
921 	error = ktls_ocf_try(so, tls, direction);
922 	if (error)
923 		return (error);
924 	tls->mode = TCP_TLS_MODE_SW;
925 	switch (tls->params.cipher_algorithm) {
926 	case CRYPTO_AES_CBC:
927 		counter_u64_add(ktls_sw_cbc, 1);
928 		break;
929 	case CRYPTO_AES_NIST_GCM_16:
930 		counter_u64_add(ktls_sw_gcm, 1);
931 		break;
932 	case CRYPTO_CHACHA20_POLY1305:
933 		counter_u64_add(ktls_sw_chacha20, 1);
934 		break;
935 	}
936 	return (0);
937 }
938 
939 /*
940  * KTLS RX stores data in the socket buffer as a list of TLS records,
941  * where each record is stored as a control message containg the TLS
942  * header followed by data mbufs containing the decrypted data.  This
943  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
944  * both encrypted and decrypted data.  TLS records decrypted by a NIC
945  * should be queued to the socket buffer as records, but encrypted
946  * data which needs to be decrypted by software arrives as a stream of
947  * regular mbufs which need to be converted.  In addition, there may
948  * already be pending encrypted data in the socket buffer when KTLS RX
949  * is enabled.
950  *
951  * To manage not-yet-decrypted data for KTLS RX, the following scheme
952  * is used:
953  *
954  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
955  *
956  * - ktls_check_rx checks this chain of mbufs reading the TLS header
957  *   from the first mbuf.  Once all of the data for that TLS record is
958  *   queued, the socket is queued to a worker thread.
959  *
960  * - The worker thread calls ktls_decrypt to decrypt TLS records in
961  *   the TLS chain.  Each TLS record is detached from the TLS chain,
962  *   decrypted, and inserted into the regular socket buffer chain as
963  *   record starting with a control message holding the TLS header and
964  *   a chain of mbufs holding the encrypted data.
965  */
966 
967 static void
968 sb_mark_notready(struct sockbuf *sb)
969 {
970 	struct mbuf *m;
971 
972 	m = sb->sb_mb;
973 	sb->sb_mtls = m;
974 	sb->sb_mb = NULL;
975 	sb->sb_mbtail = NULL;
976 	sb->sb_lastrecord = NULL;
977 	for (; m != NULL; m = m->m_next) {
978 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
979 		    __func__));
980 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
981 		    __func__));
982 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
983 		    __func__));
984 		m->m_flags |= M_NOTREADY;
985 		sb->sb_acc -= m->m_len;
986 		sb->sb_tlscc += m->m_len;
987 		sb->sb_mtlstail = m;
988 	}
989 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
990 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
991 	    sb->sb_ccc));
992 }
993 
994 int
995 ktls_enable_rx(struct socket *so, struct tls_enable *en)
996 {
997 	struct ktls_session *tls;
998 	int error;
999 
1000 	if (!ktls_offload_enable)
1001 		return (ENOTSUP);
1002 	if (SOLISTENING(so))
1003 		return (EINVAL);
1004 
1005 	counter_u64_add(ktls_offload_enable_calls, 1);
1006 
1007 	/*
1008 	 * This should always be true since only the TCP socket option
1009 	 * invokes this function.
1010 	 */
1011 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1012 		return (EINVAL);
1013 
1014 	/*
1015 	 * XXX: Don't overwrite existing sessions.  We should permit
1016 	 * this to support rekeying in the future.
1017 	 */
1018 	if (so->so_rcv.sb_tls_info != NULL)
1019 		return (EALREADY);
1020 
1021 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1022 		return (ENOTSUP);
1023 
1024 	/* TLS 1.3 is not yet supported. */
1025 	if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1026 	    en->tls_vminor == TLS_MINOR_VER_THREE)
1027 		return (ENOTSUP);
1028 
1029 	error = ktls_create_session(so, en, &tls);
1030 	if (error)
1031 		return (error);
1032 
1033 #ifdef TCP_OFFLOAD
1034 	error = ktls_try_toe(so, tls, KTLS_RX);
1035 	if (error)
1036 #endif
1037 		error = ktls_try_sw(so, tls, KTLS_RX);
1038 
1039 	if (error) {
1040 		ktls_cleanup(tls);
1041 		return (error);
1042 	}
1043 
1044 	/* Mark the socket as using TLS offload. */
1045 	SOCKBUF_LOCK(&so->so_rcv);
1046 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1047 	so->so_rcv.sb_tls_info = tls;
1048 	so->so_rcv.sb_flags |= SB_TLS_RX;
1049 
1050 	/* Mark existing data as not ready until it can be decrypted. */
1051 	sb_mark_notready(&so->so_rcv);
1052 	ktls_check_rx(&so->so_rcv);
1053 	SOCKBUF_UNLOCK(&so->so_rcv);
1054 
1055 	counter_u64_add(ktls_offload_total, 1);
1056 
1057 	return (0);
1058 }
1059 
1060 int
1061 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1062 {
1063 	struct ktls_session *tls;
1064 	struct inpcb *inp;
1065 	int error;
1066 
1067 	if (!ktls_offload_enable)
1068 		return (ENOTSUP);
1069 	if (SOLISTENING(so))
1070 		return (EINVAL);
1071 
1072 	counter_u64_add(ktls_offload_enable_calls, 1);
1073 
1074 	/*
1075 	 * This should always be true since only the TCP socket option
1076 	 * invokes this function.
1077 	 */
1078 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1079 		return (EINVAL);
1080 
1081 	/*
1082 	 * XXX: Don't overwrite existing sessions.  We should permit
1083 	 * this to support rekeying in the future.
1084 	 */
1085 	if (so->so_snd.sb_tls_info != NULL)
1086 		return (EALREADY);
1087 
1088 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1089 		return (ENOTSUP);
1090 
1091 	/* TLS requires ext pgs */
1092 	if (mb_use_ext_pgs == 0)
1093 		return (ENXIO);
1094 
1095 	error = ktls_create_session(so, en, &tls);
1096 	if (error)
1097 		return (error);
1098 
1099 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1100 #ifdef TCP_OFFLOAD
1101 	error = ktls_try_toe(so, tls, KTLS_TX);
1102 	if (error)
1103 #endif
1104 		error = ktls_try_ifnet(so, tls, false);
1105 	if (error)
1106 		error = ktls_try_sw(so, tls, KTLS_TX);
1107 
1108 	if (error) {
1109 		ktls_cleanup(tls);
1110 		return (error);
1111 	}
1112 
1113 	error = sblock(&so->so_snd, SBL_WAIT);
1114 	if (error) {
1115 		ktls_cleanup(tls);
1116 		return (error);
1117 	}
1118 
1119 	/*
1120 	 * Write lock the INP when setting sb_tls_info so that
1121 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1122 	 * holding the INP lock.
1123 	 */
1124 	inp = so->so_pcb;
1125 	INP_WLOCK(inp);
1126 	SOCKBUF_LOCK(&so->so_snd);
1127 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1128 	so->so_snd.sb_tls_info = tls;
1129 	if (tls->mode != TCP_TLS_MODE_SW)
1130 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1131 	SOCKBUF_UNLOCK(&so->so_snd);
1132 	INP_WUNLOCK(inp);
1133 	sbunlock(&so->so_snd);
1134 
1135 	counter_u64_add(ktls_offload_total, 1);
1136 
1137 	return (0);
1138 }
1139 
1140 int
1141 ktls_get_rx_mode(struct socket *so)
1142 {
1143 	struct ktls_session *tls;
1144 	struct inpcb *inp;
1145 	int mode;
1146 
1147 	if (SOLISTENING(so))
1148 		return (EINVAL);
1149 	inp = so->so_pcb;
1150 	INP_WLOCK_ASSERT(inp);
1151 	SOCKBUF_LOCK(&so->so_rcv);
1152 	tls = so->so_rcv.sb_tls_info;
1153 	if (tls == NULL)
1154 		mode = TCP_TLS_MODE_NONE;
1155 	else
1156 		mode = tls->mode;
1157 	SOCKBUF_UNLOCK(&so->so_rcv);
1158 	return (mode);
1159 }
1160 
1161 int
1162 ktls_get_tx_mode(struct socket *so)
1163 {
1164 	struct ktls_session *tls;
1165 	struct inpcb *inp;
1166 	int mode;
1167 
1168 	if (SOLISTENING(so))
1169 		return (EINVAL);
1170 	inp = so->so_pcb;
1171 	INP_WLOCK_ASSERT(inp);
1172 	SOCKBUF_LOCK(&so->so_snd);
1173 	tls = so->so_snd.sb_tls_info;
1174 	if (tls == NULL)
1175 		mode = TCP_TLS_MODE_NONE;
1176 	else
1177 		mode = tls->mode;
1178 	SOCKBUF_UNLOCK(&so->so_snd);
1179 	return (mode);
1180 }
1181 
1182 /*
1183  * Switch between SW and ifnet TLS sessions as requested.
1184  */
1185 int
1186 ktls_set_tx_mode(struct socket *so, int mode)
1187 {
1188 	struct ktls_session *tls, *tls_new;
1189 	struct inpcb *inp;
1190 	int error;
1191 
1192 	if (SOLISTENING(so))
1193 		return (EINVAL);
1194 	switch (mode) {
1195 	case TCP_TLS_MODE_SW:
1196 	case TCP_TLS_MODE_IFNET:
1197 		break;
1198 	default:
1199 		return (EINVAL);
1200 	}
1201 
1202 	inp = so->so_pcb;
1203 	INP_WLOCK_ASSERT(inp);
1204 	SOCKBUF_LOCK(&so->so_snd);
1205 	tls = so->so_snd.sb_tls_info;
1206 	if (tls == NULL) {
1207 		SOCKBUF_UNLOCK(&so->so_snd);
1208 		return (0);
1209 	}
1210 
1211 	if (tls->mode == mode) {
1212 		SOCKBUF_UNLOCK(&so->so_snd);
1213 		return (0);
1214 	}
1215 
1216 	tls = ktls_hold(tls);
1217 	SOCKBUF_UNLOCK(&so->so_snd);
1218 	INP_WUNLOCK(inp);
1219 
1220 	tls_new = ktls_clone_session(tls);
1221 
1222 	if (mode == TCP_TLS_MODE_IFNET)
1223 		error = ktls_try_ifnet(so, tls_new, true);
1224 	else
1225 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1226 	if (error) {
1227 		counter_u64_add(ktls_switch_failed, 1);
1228 		ktls_free(tls_new);
1229 		ktls_free(tls);
1230 		INP_WLOCK(inp);
1231 		return (error);
1232 	}
1233 
1234 	error = sblock(&so->so_snd, SBL_WAIT);
1235 	if (error) {
1236 		counter_u64_add(ktls_switch_failed, 1);
1237 		ktls_free(tls_new);
1238 		ktls_free(tls);
1239 		INP_WLOCK(inp);
1240 		return (error);
1241 	}
1242 
1243 	/*
1244 	 * If we raced with another session change, keep the existing
1245 	 * session.
1246 	 */
1247 	if (tls != so->so_snd.sb_tls_info) {
1248 		counter_u64_add(ktls_switch_failed, 1);
1249 		sbunlock(&so->so_snd);
1250 		ktls_free(tls_new);
1251 		ktls_free(tls);
1252 		INP_WLOCK(inp);
1253 		return (EBUSY);
1254 	}
1255 
1256 	SOCKBUF_LOCK(&so->so_snd);
1257 	so->so_snd.sb_tls_info = tls_new;
1258 	if (tls_new->mode != TCP_TLS_MODE_SW)
1259 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1260 	SOCKBUF_UNLOCK(&so->so_snd);
1261 	sbunlock(&so->so_snd);
1262 
1263 	/*
1264 	 * Drop two references on 'tls'.  The first is for the
1265 	 * ktls_hold() above.  The second drops the reference from the
1266 	 * socket buffer.
1267 	 */
1268 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1269 	ktls_free(tls);
1270 	ktls_free(tls);
1271 
1272 	if (mode == TCP_TLS_MODE_IFNET)
1273 		counter_u64_add(ktls_switch_to_ifnet, 1);
1274 	else
1275 		counter_u64_add(ktls_switch_to_sw, 1);
1276 
1277 	INP_WLOCK(inp);
1278 	return (0);
1279 }
1280 
1281 /*
1282  * Try to allocate a new TLS send tag.  This task is scheduled when
1283  * ip_output detects a route change while trying to transmit a packet
1284  * holding a TLS record.  If a new tag is allocated, replace the tag
1285  * in the TLS session.  Subsequent packets on the connection will use
1286  * the new tag.  If a new tag cannot be allocated, drop the
1287  * connection.
1288  */
1289 static void
1290 ktls_reset_send_tag(void *context, int pending)
1291 {
1292 	struct epoch_tracker et;
1293 	struct ktls_session *tls;
1294 	struct m_snd_tag *old, *new;
1295 	struct inpcb *inp;
1296 	struct tcpcb *tp;
1297 	int error;
1298 
1299 	MPASS(pending == 1);
1300 
1301 	tls = context;
1302 	inp = tls->inp;
1303 
1304 	/*
1305 	 * Free the old tag first before allocating a new one.
1306 	 * ip[6]_output_send() will treat a NULL send tag the same as
1307 	 * an ifp mismatch and drop packets until a new tag is
1308 	 * allocated.
1309 	 *
1310 	 * Write-lock the INP when changing tls->snd_tag since
1311 	 * ip[6]_output_send() holds a read-lock when reading the
1312 	 * pointer.
1313 	 */
1314 	INP_WLOCK(inp);
1315 	old = tls->snd_tag;
1316 	tls->snd_tag = NULL;
1317 	INP_WUNLOCK(inp);
1318 	if (old != NULL)
1319 		m_snd_tag_rele(old);
1320 
1321 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1322 
1323 	if (error == 0) {
1324 		INP_WLOCK(inp);
1325 		tls->snd_tag = new;
1326 		mtx_pool_lock(mtxpool_sleep, tls);
1327 		tls->reset_pending = false;
1328 		mtx_pool_unlock(mtxpool_sleep, tls);
1329 		if (!in_pcbrele_wlocked(inp))
1330 			INP_WUNLOCK(inp);
1331 
1332 		counter_u64_add(ktls_ifnet_reset, 1);
1333 
1334 		/*
1335 		 * XXX: Should we kick tcp_output explicitly now that
1336 		 * the send tag is fixed or just rely on timers?
1337 		 */
1338 	} else {
1339 		NET_EPOCH_ENTER(et);
1340 		INP_WLOCK(inp);
1341 		if (!in_pcbrele_wlocked(inp)) {
1342 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1343 			    !(inp->inp_flags & INP_DROPPED)) {
1344 				tp = intotcpcb(inp);
1345 				CURVNET_SET(tp->t_vnet);
1346 				tp = tcp_drop(tp, ECONNABORTED);
1347 				CURVNET_RESTORE();
1348 				if (tp != NULL)
1349 					INP_WUNLOCK(inp);
1350 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1351 			} else
1352 				INP_WUNLOCK(inp);
1353 		}
1354 		NET_EPOCH_EXIT(et);
1355 
1356 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1357 
1358 		/*
1359 		 * Leave reset_pending true to avoid future tasks while
1360 		 * the socket goes away.
1361 		 */
1362 	}
1363 
1364 	ktls_free(tls);
1365 }
1366 
1367 int
1368 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1369 {
1370 
1371 	if (inp == NULL)
1372 		return (ENOBUFS);
1373 
1374 	INP_LOCK_ASSERT(inp);
1375 
1376 	/*
1377 	 * See if we should schedule a task to update the send tag for
1378 	 * this session.
1379 	 */
1380 	mtx_pool_lock(mtxpool_sleep, tls);
1381 	if (!tls->reset_pending) {
1382 		(void) ktls_hold(tls);
1383 		in_pcbref(inp);
1384 		tls->inp = inp;
1385 		tls->reset_pending = true;
1386 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1387 	}
1388 	mtx_pool_unlock(mtxpool_sleep, tls);
1389 	return (ENOBUFS);
1390 }
1391 
1392 #ifdef RATELIMIT
1393 int
1394 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1395 {
1396 	union if_snd_tag_modify_params params = {
1397 		.rate_limit.max_rate = max_pacing_rate,
1398 		.rate_limit.flags = M_NOWAIT,
1399 	};
1400 	struct m_snd_tag *mst;
1401 	struct ifnet *ifp;
1402 	int error;
1403 
1404 	/* Can't get to the inp, but it should be locked. */
1405 	/* INP_LOCK_ASSERT(inp); */
1406 
1407 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1408 
1409 	if (tls->snd_tag == NULL) {
1410 		/*
1411 		 * Resetting send tag, ignore this change.  The
1412 		 * pending reset may or may not see this updated rate
1413 		 * in the tcpcb.  If it doesn't, we will just lose
1414 		 * this rate change.
1415 		 */
1416 		return (0);
1417 	}
1418 
1419 	MPASS(tls->snd_tag != NULL);
1420 	MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1421 
1422 	mst = tls->snd_tag;
1423 	ifp = mst->ifp;
1424 	return (ifp->if_snd_tag_modify(mst, &params));
1425 }
1426 #endif
1427 #endif
1428 
1429 void
1430 ktls_destroy(struct ktls_session *tls)
1431 {
1432 
1433 	ktls_cleanup(tls);
1434 	uma_zfree(ktls_session_zone, tls);
1435 }
1436 
1437 void
1438 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1439 {
1440 
1441 	for (; m != NULL; m = m->m_next) {
1442 		KASSERT((m->m_flags & M_EXTPG) != 0,
1443 		    ("ktls_seq: mapped mbuf %p", m));
1444 
1445 		m->m_epg_seqno = sb->sb_tls_seqno;
1446 		sb->sb_tls_seqno++;
1447 	}
1448 }
1449 
1450 /*
1451  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1452  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1453  * mbuf must be populated with the payload of each TLS record.
1454  *
1455  * The record_type argument specifies the TLS record type used when
1456  * populating the TLS header.
1457  *
1458  * The enq_count argument on return is set to the number of pages of
1459  * payload data for this entire chain that need to be encrypted via SW
1460  * encryption.  The returned value should be passed to ktls_enqueue
1461  * when scheduling encryption of this chain of mbufs.  To handle the
1462  * special case of empty fragments for TLS 1.0 sessions, an empty
1463  * fragment counts as one page.
1464  */
1465 void
1466 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1467     uint8_t record_type)
1468 {
1469 	struct tls_record_layer *tlshdr;
1470 	struct mbuf *m;
1471 	uint64_t *noncep;
1472 	uint16_t tls_len;
1473 	int maxlen;
1474 
1475 	maxlen = tls->params.max_frame_len;
1476 	*enq_cnt = 0;
1477 	for (m = top; m != NULL; m = m->m_next) {
1478 		/*
1479 		 * All mbufs in the chain should be TLS records whose
1480 		 * payload does not exceed the maximum frame length.
1481 		 *
1482 		 * Empty TLS records are permitted when using CBC.
1483 		 */
1484 		KASSERT(m->m_len <= maxlen &&
1485 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1486 		    m->m_len >= 0 : m->m_len > 0),
1487 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1488 
1489 		/*
1490 		 * TLS frames require unmapped mbufs to store session
1491 		 * info.
1492 		 */
1493 		KASSERT((m->m_flags & M_EXTPG) != 0,
1494 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1495 
1496 		tls_len = m->m_len;
1497 
1498 		/* Save a reference to the session. */
1499 		m->m_epg_tls = ktls_hold(tls);
1500 
1501 		m->m_epg_hdrlen = tls->params.tls_hlen;
1502 		m->m_epg_trllen = tls->params.tls_tlen;
1503 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1504 			int bs, delta;
1505 
1506 			/*
1507 			 * AES-CBC pads messages to a multiple of the
1508 			 * block size.  Note that the padding is
1509 			 * applied after the digest and the encryption
1510 			 * is done on the "plaintext || mac || padding".
1511 			 * At least one byte of padding is always
1512 			 * present.
1513 			 *
1514 			 * Compute the final trailer length assuming
1515 			 * at most one block of padding.
1516 			 * tls->params.tls_tlen is the maximum
1517 			 * possible trailer length (padding + digest).
1518 			 * delta holds the number of excess padding
1519 			 * bytes if the maximum were used.  Those
1520 			 * extra bytes are removed.
1521 			 */
1522 			bs = tls->params.tls_bs;
1523 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1524 			m->m_epg_trllen -= delta;
1525 		}
1526 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1527 
1528 		/* Populate the TLS header. */
1529 		tlshdr = (void *)m->m_epg_hdr;
1530 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1531 
1532 		/*
1533 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1534 		 * of TLS_RLTYPE_APP.
1535 		 */
1536 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1537 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1538 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1539 			tlshdr->tls_type = TLS_RLTYPE_APP;
1540 			/* save the real record type for later */
1541 			m->m_epg_record_type = record_type;
1542 			m->m_epg_trail[0] = record_type;
1543 		} else {
1544 			tlshdr->tls_vminor = tls->params.tls_vminor;
1545 			tlshdr->tls_type = record_type;
1546 		}
1547 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1548 
1549 		/*
1550 		 * Store nonces / explicit IVs after the end of the
1551 		 * TLS header.
1552 		 *
1553 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1554 		 * from the end of the IV.  The nonce is then
1555 		 * incremented for use by the next record.
1556 		 *
1557 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1558 		 */
1559 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1560 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1561 			noncep = (uint64_t *)(tls->params.iv + 8);
1562 			be64enc(tlshdr + 1, *noncep);
1563 			(*noncep)++;
1564 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1565 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1566 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1567 
1568 		/*
1569 		 * When using SW encryption, mark the mbuf not ready.
1570 		 * It will be marked ready via sbready() after the
1571 		 * record has been encrypted.
1572 		 *
1573 		 * When using ifnet TLS, unencrypted TLS records are
1574 		 * sent down the stack to the NIC.
1575 		 */
1576 		if (tls->mode == TCP_TLS_MODE_SW) {
1577 			m->m_flags |= M_NOTREADY;
1578 			m->m_epg_nrdy = m->m_epg_npgs;
1579 			if (__predict_false(tls_len == 0)) {
1580 				/* TLS 1.0 empty fragment. */
1581 				*enq_cnt += 1;
1582 			} else
1583 				*enq_cnt += m->m_epg_npgs;
1584 		}
1585 	}
1586 }
1587 
1588 void
1589 ktls_check_rx(struct sockbuf *sb)
1590 {
1591 	struct tls_record_layer hdr;
1592 	struct ktls_wq *wq;
1593 	struct socket *so;
1594 	bool running;
1595 
1596 	SOCKBUF_LOCK_ASSERT(sb);
1597 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1598 	    __func__, sb));
1599 	so = __containerof(sb, struct socket, so_rcv);
1600 
1601 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1602 		return;
1603 
1604 	/* Is there enough queued for a TLS header? */
1605 	if (sb->sb_tlscc < sizeof(hdr)) {
1606 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1607 			so->so_error = EMSGSIZE;
1608 		return;
1609 	}
1610 
1611 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1612 
1613 	/* Is the entire record queued? */
1614 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1615 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1616 			so->so_error = EMSGSIZE;
1617 		return;
1618 	}
1619 
1620 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1621 
1622 	soref(so);
1623 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1624 	mtx_lock(&wq->mtx);
1625 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1626 	running = wq->running;
1627 	mtx_unlock(&wq->mtx);
1628 	if (!running)
1629 		wakeup(wq);
1630 	counter_u64_add(ktls_cnt_rx_queued, 1);
1631 }
1632 
1633 static struct mbuf *
1634 ktls_detach_record(struct sockbuf *sb, int len)
1635 {
1636 	struct mbuf *m, *n, *top;
1637 	int remain;
1638 
1639 	SOCKBUF_LOCK_ASSERT(sb);
1640 	MPASS(len <= sb->sb_tlscc);
1641 
1642 	/*
1643 	 * If TLS chain is the exact size of the record,
1644 	 * just grab the whole record.
1645 	 */
1646 	top = sb->sb_mtls;
1647 	if (sb->sb_tlscc == len) {
1648 		sb->sb_mtls = NULL;
1649 		sb->sb_mtlstail = NULL;
1650 		goto out;
1651 	}
1652 
1653 	/*
1654 	 * While it would be nice to use m_split() here, we need
1655 	 * to know exactly what m_split() allocates to update the
1656 	 * accounting, so do it inline instead.
1657 	 */
1658 	remain = len;
1659 	for (m = top; remain > m->m_len; m = m->m_next)
1660 		remain -= m->m_len;
1661 
1662 	/* Easy case: don't have to split 'm'. */
1663 	if (remain == m->m_len) {
1664 		sb->sb_mtls = m->m_next;
1665 		if (sb->sb_mtls == NULL)
1666 			sb->sb_mtlstail = NULL;
1667 		m->m_next = NULL;
1668 		goto out;
1669 	}
1670 
1671 	/*
1672 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1673 	 * with M_NOWAIT first.
1674 	 */
1675 	n = m_get(M_NOWAIT, MT_DATA);
1676 	if (n == NULL) {
1677 		/*
1678 		 * Use M_WAITOK with socket buffer unlocked.  If
1679 		 * 'sb_mtls' changes while the lock is dropped, return
1680 		 * NULL to force the caller to retry.
1681 		 */
1682 		SOCKBUF_UNLOCK(sb);
1683 
1684 		n = m_get(M_WAITOK, MT_DATA);
1685 
1686 		SOCKBUF_LOCK(sb);
1687 		if (sb->sb_mtls != top) {
1688 			m_free(n);
1689 			return (NULL);
1690 		}
1691 	}
1692 	n->m_flags |= M_NOTREADY;
1693 
1694 	/* Store remainder in 'n'. */
1695 	n->m_len = m->m_len - remain;
1696 	if (m->m_flags & M_EXT) {
1697 		n->m_data = m->m_data + remain;
1698 		mb_dupcl(n, m);
1699 	} else {
1700 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1701 	}
1702 
1703 	/* Trim 'm' and update accounting. */
1704 	m->m_len -= n->m_len;
1705 	sb->sb_tlscc -= n->m_len;
1706 	sb->sb_ccc -= n->m_len;
1707 
1708 	/* Account for 'n'. */
1709 	sballoc_ktls_rx(sb, n);
1710 
1711 	/* Insert 'n' into the TLS chain. */
1712 	sb->sb_mtls = n;
1713 	n->m_next = m->m_next;
1714 	if (sb->sb_mtlstail == m)
1715 		sb->sb_mtlstail = n;
1716 
1717 	/* Detach the record from the TLS chain. */
1718 	m->m_next = NULL;
1719 
1720 out:
1721 	MPASS(m_length(top, NULL) == len);
1722 	for (m = top; m != NULL; m = m->m_next)
1723 		sbfree_ktls_rx(sb, m);
1724 	sb->sb_tlsdcc = len;
1725 	sb->sb_ccc += len;
1726 	SBCHECK(sb);
1727 	return (top);
1728 }
1729 
1730 static void
1731 ktls_decrypt(struct socket *so)
1732 {
1733 	char tls_header[MBUF_PEXT_HDR_LEN];
1734 	struct ktls_session *tls;
1735 	struct sockbuf *sb;
1736 	struct tls_record_layer *hdr;
1737 	struct tls_get_record tgr;
1738 	struct mbuf *control, *data, *m;
1739 	uint64_t seqno;
1740 	int error, remain, tls_len, trail_len;
1741 
1742 	hdr = (struct tls_record_layer *)tls_header;
1743 	sb = &so->so_rcv;
1744 	SOCKBUF_LOCK(sb);
1745 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1746 	    ("%s: socket %p not running", __func__, so));
1747 
1748 	tls = sb->sb_tls_info;
1749 	MPASS(tls != NULL);
1750 
1751 	for (;;) {
1752 		/* Is there enough queued for a TLS header? */
1753 		if (sb->sb_tlscc < tls->params.tls_hlen)
1754 			break;
1755 
1756 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1757 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1758 
1759 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1760 		    hdr->tls_vminor != tls->params.tls_vminor)
1761 			error = EINVAL;
1762 		else if (tls_len < tls->params.tls_hlen || tls_len >
1763 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1764 		    tls->params.tls_tlen)
1765 			error = EMSGSIZE;
1766 		else
1767 			error = 0;
1768 		if (__predict_false(error != 0)) {
1769 			/*
1770 			 * We have a corrupted record and are likely
1771 			 * out of sync.  The connection isn't
1772 			 * recoverable at this point, so abort it.
1773 			 */
1774 			SOCKBUF_UNLOCK(sb);
1775 			counter_u64_add(ktls_offload_corrupted_records, 1);
1776 
1777 			CURVNET_SET(so->so_vnet);
1778 			so->so_proto->pr_usrreqs->pru_abort(so);
1779 			so->so_error = error;
1780 			CURVNET_RESTORE();
1781 			goto deref;
1782 		}
1783 
1784 		/* Is the entire record queued? */
1785 		if (sb->sb_tlscc < tls_len)
1786 			break;
1787 
1788 		/*
1789 		 * Split out the portion of the mbuf chain containing
1790 		 * this TLS record.
1791 		 */
1792 		data = ktls_detach_record(sb, tls_len);
1793 		if (data == NULL)
1794 			continue;
1795 		MPASS(sb->sb_tlsdcc == tls_len);
1796 
1797 		seqno = sb->sb_tls_seqno;
1798 		sb->sb_tls_seqno++;
1799 		SBCHECK(sb);
1800 		SOCKBUF_UNLOCK(sb);
1801 
1802 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1803 		if (error) {
1804 			counter_u64_add(ktls_offload_failed_crypto, 1);
1805 
1806 			SOCKBUF_LOCK(sb);
1807 			if (sb->sb_tlsdcc == 0) {
1808 				/*
1809 				 * sbcut/drop/flush discarded these
1810 				 * mbufs.
1811 				 */
1812 				m_freem(data);
1813 				break;
1814 			}
1815 
1816 			/*
1817 			 * Drop this TLS record's data, but keep
1818 			 * decrypting subsequent records.
1819 			 */
1820 			sb->sb_ccc -= tls_len;
1821 			sb->sb_tlsdcc = 0;
1822 
1823 			CURVNET_SET(so->so_vnet);
1824 			so->so_error = EBADMSG;
1825 			sorwakeup_locked(so);
1826 			CURVNET_RESTORE();
1827 
1828 			m_freem(data);
1829 
1830 			SOCKBUF_LOCK(sb);
1831 			continue;
1832 		}
1833 
1834 		/* Allocate the control mbuf. */
1835 		tgr.tls_type = hdr->tls_type;
1836 		tgr.tls_vmajor = hdr->tls_vmajor;
1837 		tgr.tls_vminor = hdr->tls_vminor;
1838 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1839 		    trail_len);
1840 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1841 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1842 
1843 		SOCKBUF_LOCK(sb);
1844 		if (sb->sb_tlsdcc == 0) {
1845 			/* sbcut/drop/flush discarded these mbufs. */
1846 			MPASS(sb->sb_tlscc == 0);
1847 			m_freem(data);
1848 			m_freem(control);
1849 			break;
1850 		}
1851 
1852 		/*
1853 		 * Clear the 'dcc' accounting in preparation for
1854 		 * adding the decrypted record.
1855 		 */
1856 		sb->sb_ccc -= tls_len;
1857 		sb->sb_tlsdcc = 0;
1858 		SBCHECK(sb);
1859 
1860 		/* If there is no payload, drop all of the data. */
1861 		if (tgr.tls_length == htobe16(0)) {
1862 			m_freem(data);
1863 			data = NULL;
1864 		} else {
1865 			/* Trim header. */
1866 			remain = tls->params.tls_hlen;
1867 			while (remain > 0) {
1868 				if (data->m_len > remain) {
1869 					data->m_data += remain;
1870 					data->m_len -= remain;
1871 					break;
1872 				}
1873 				remain -= data->m_len;
1874 				data = m_free(data);
1875 			}
1876 
1877 			/* Trim trailer and clear M_NOTREADY. */
1878 			remain = be16toh(tgr.tls_length);
1879 			m = data;
1880 			for (m = data; remain > m->m_len; m = m->m_next) {
1881 				m->m_flags &= ~M_NOTREADY;
1882 				remain -= m->m_len;
1883 			}
1884 			m->m_len = remain;
1885 			m_freem(m->m_next);
1886 			m->m_next = NULL;
1887 			m->m_flags &= ~M_NOTREADY;
1888 
1889 			/* Set EOR on the final mbuf. */
1890 			m->m_flags |= M_EOR;
1891 		}
1892 
1893 		sbappendcontrol_locked(sb, data, control, 0);
1894 	}
1895 
1896 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1897 
1898 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1899 		so->so_error = EMSGSIZE;
1900 
1901 	sorwakeup_locked(so);
1902 
1903 deref:
1904 	SOCKBUF_UNLOCK_ASSERT(sb);
1905 
1906 	CURVNET_SET(so->so_vnet);
1907 	SOCK_LOCK(so);
1908 	sorele(so);
1909 	CURVNET_RESTORE();
1910 }
1911 
1912 void
1913 ktls_enqueue_to_free(struct mbuf *m)
1914 {
1915 	struct ktls_wq *wq;
1916 	bool running;
1917 
1918 	/* Mark it for freeing. */
1919 	m->m_epg_flags |= EPG_FLAG_2FREE;
1920 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1921 	mtx_lock(&wq->mtx);
1922 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1923 	running = wq->running;
1924 	mtx_unlock(&wq->mtx);
1925 	if (!running)
1926 		wakeup(wq);
1927 }
1928 
1929 static void *
1930 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
1931 {
1932 	void *buf;
1933 
1934 	if (m->m_epg_npgs <= 2)
1935 		return (NULL);
1936 	if (ktls_buffer_zone == NULL)
1937 		return (NULL);
1938 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
1939 		/*
1940 		 * Rate-limit allocation attempts after a failure.
1941 		 * ktls_buffer_import() will acquire a per-domain mutex to check
1942 		 * the free page queues and may fail consistently if memory is
1943 		 * fragmented.
1944 		 */
1945 		return (NULL);
1946 	}
1947 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
1948 	if (buf == NULL)
1949 		wq->lastallocfail = ticks;
1950 	return (buf);
1951 }
1952 
1953 void
1954 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1955 {
1956 	struct ktls_wq *wq;
1957 	bool running;
1958 
1959 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1960 	    (M_EXTPG | M_NOTREADY)),
1961 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1962 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1963 
1964 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1965 
1966 	m->m_epg_enc_cnt = page_count;
1967 
1968 	/*
1969 	 * Save a pointer to the socket.  The caller is responsible
1970 	 * for taking an additional reference via soref().
1971 	 */
1972 	m->m_epg_so = so;
1973 
1974 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1975 	mtx_lock(&wq->mtx);
1976 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1977 	running = wq->running;
1978 	mtx_unlock(&wq->mtx);
1979 	if (!running)
1980 		wakeup(wq);
1981 	counter_u64_add(ktls_cnt_tx_queued, 1);
1982 }
1983 
1984 #define	MAX_TLS_PAGES	(1 + btoc(TLS_MAX_MSG_SIZE_V10_2))
1985 
1986 static __noinline void
1987 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
1988 {
1989 	struct ktls_session *tls;
1990 	struct socket *so;
1991 	struct mbuf *m;
1992 	vm_paddr_t parray[MAX_TLS_PAGES + 1];
1993 	struct iovec dst_iov[MAX_TLS_PAGES + 2];
1994 	vm_page_t pg;
1995 	void *cbuf;
1996 	int error, i, len, npages, off, total_pages;
1997 
1998 	so = top->m_epg_so;
1999 	tls = top->m_epg_tls;
2000 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2001 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2002 #ifdef INVARIANTS
2003 	top->m_epg_so = NULL;
2004 #endif
2005 	total_pages = top->m_epg_enc_cnt;
2006 	npages = 0;
2007 
2008 	/*
2009 	 * Encrypt the TLS records in the chain of mbufs starting with
2010 	 * 'top'.  'total_pages' gives us a total count of pages and is
2011 	 * used to know when we have finished encrypting the TLS
2012 	 * records originally queued with 'top'.
2013 	 *
2014 	 * NB: These mbufs are queued in the socket buffer and
2015 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2016 	 * socket buffer lock is not held while traversing this chain.
2017 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2018 	 * pointers should be stable.  However, the 'm_next' of the
2019 	 * last mbuf encrypted is not necessarily NULL.  It can point
2020 	 * to other mbufs appended while 'top' was on the TLS work
2021 	 * queue.
2022 	 *
2023 	 * Each mbuf holds an entire TLS record.
2024 	 */
2025 	error = 0;
2026 	for (m = top; npages != total_pages; m = m->m_next) {
2027 		KASSERT(m->m_epg_tls == tls,
2028 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2029 		    tls, m->m_epg_tls));
2030 		KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2031 		    (M_EXTPG | M_NOTREADY),
2032 		    ("%p not unready & nomap mbuf (top = %p)\n", m, top));
2033 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2034 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2035 		    total_pages, m));
2036 		KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2037 		    ("page count %d larger than maximum frame length %d",
2038 		    m->m_epg_npgs, ktls_maxlen));
2039 
2040 		/*
2041 		 * For anonymous mbufs, encryption is done in place.
2042 		 * For file-backed mbufs (from sendfile), anonymous
2043 		 * wired pages are allocated and used as the
2044 		 * encryption destination.
2045 		 */
2046 		if ((m->m_epg_flags & EPG_FLAG_ANON) != 0) {
2047 			error = (*tls->sw_encrypt)(tls, m, NULL, 0);
2048 		} else {
2049 			if ((cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2050 				len = ptoa(m->m_epg_npgs - 1) +
2051 				    m->m_epg_last_len - m->m_epg_1st_off;
2052 				dst_iov[0].iov_base = (char *)cbuf +
2053 				    m->m_epg_1st_off;
2054 				dst_iov[0].iov_len = len;
2055 				parray[0] = DMAP_TO_PHYS((vm_offset_t)cbuf);
2056 				i = 1;
2057 			} else {
2058 				off = m->m_epg_1st_off;
2059 				for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2060 					do {
2061 						pg = vm_page_alloc(NULL, 0,
2062 						    VM_ALLOC_NORMAL |
2063 						    VM_ALLOC_NOOBJ |
2064 						    VM_ALLOC_NODUMP |
2065 						    VM_ALLOC_WIRED |
2066 						    VM_ALLOC_WAITFAIL);
2067 					} while (pg == NULL);
2068 
2069 					len = m_epg_pagelen(m, i, off);
2070 					parray[i] = VM_PAGE_TO_PHYS(pg);
2071 					dst_iov[i].iov_base =
2072 					    (char *)(void *)PHYS_TO_DMAP(
2073 					    parray[i]) + off;
2074 					dst_iov[i].iov_len = len;
2075 				}
2076 			}
2077 			KASSERT(i + 1 <= nitems(dst_iov),
2078 			    ("dst_iov is too small"));
2079 			dst_iov[i].iov_base = m->m_epg_trail;
2080 			dst_iov[i].iov_len = m->m_epg_trllen;
2081 
2082 			error = (*tls->sw_encrypt)(tls, m, dst_iov, i + 1);
2083 
2084 			/* Free the old pages. */
2085 			m->m_ext.ext_free(m);
2086 
2087 			/* Replace them with the new pages. */
2088 			if (cbuf != NULL) {
2089 				for (i = 0; i < m->m_epg_npgs; i++)
2090 					m->m_epg_pa[i] = parray[0] + ptoa(i);
2091 
2092 				/* Contig pages should go back to the cache. */
2093 				m->m_ext.ext_free = ktls_free_mext_contig;
2094 			} else {
2095 				for (i = 0; i < m->m_epg_npgs; i++)
2096 					m->m_epg_pa[i] = parray[i];
2097 
2098 				/* Use the basic free routine. */
2099 				m->m_ext.ext_free = mb_free_mext_pgs;
2100 			}
2101 
2102 			/* Pages are now writable. */
2103 			m->m_epg_flags |= EPG_FLAG_ANON;
2104 		}
2105 		if (error) {
2106 			counter_u64_add(ktls_offload_failed_crypto, 1);
2107 			break;
2108 		}
2109 
2110 		if (__predict_false(m->m_epg_npgs == 0)) {
2111 			/* TLS 1.0 empty fragment. */
2112 			npages++;
2113 		} else
2114 			npages += m->m_epg_npgs;
2115 
2116 		/*
2117 		 * Drop a reference to the session now that it is no
2118 		 * longer needed.  Existing code depends on encrypted
2119 		 * records having no associated session vs
2120 		 * yet-to-be-encrypted records having an associated
2121 		 * session.
2122 		 */
2123 		m->m_epg_tls = NULL;
2124 		ktls_free(tls);
2125 	}
2126 
2127 	CURVNET_SET(so->so_vnet);
2128 	if (error == 0) {
2129 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2130 	} else {
2131 		so->so_proto->pr_usrreqs->pru_abort(so);
2132 		so->so_error = EIO;
2133 		mb_free_notready(top, total_pages);
2134 	}
2135 
2136 	SOCK_LOCK(so);
2137 	sorele(so);
2138 	CURVNET_RESTORE();
2139 }
2140 
2141 static void
2142 ktls_work_thread(void *ctx)
2143 {
2144 	struct ktls_wq *wq = ctx;
2145 	struct mbuf *m, *n;
2146 	struct socket *so, *son;
2147 	STAILQ_HEAD(, mbuf) local_m_head;
2148 	STAILQ_HEAD(, socket) local_so_head;
2149 
2150 	if (ktls_bind_threads > 1) {
2151 		curthread->td_domain.dr_policy =
2152 			DOMAINSET_PREF(PCPU_GET(domain));
2153 	}
2154 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2155 	fpu_kern_thread(0);
2156 #endif
2157 	for (;;) {
2158 		mtx_lock(&wq->mtx);
2159 		while (STAILQ_EMPTY(&wq->m_head) &&
2160 		    STAILQ_EMPTY(&wq->so_head)) {
2161 			wq->running = false;
2162 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2163 			wq->running = true;
2164 		}
2165 
2166 		STAILQ_INIT(&local_m_head);
2167 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2168 		STAILQ_INIT(&local_so_head);
2169 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2170 		mtx_unlock(&wq->mtx);
2171 
2172 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2173 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2174 				ktls_free(m->m_epg_tls);
2175 				uma_zfree(zone_mbuf, m);
2176 			} else {
2177 				ktls_encrypt(wq, m);
2178 				counter_u64_add(ktls_cnt_tx_queued, -1);
2179 			}
2180 		}
2181 
2182 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2183 			ktls_decrypt(so);
2184 			counter_u64_add(ktls_cnt_rx_queued, -1);
2185 		}
2186 	}
2187 }
2188