xref: /freebsd-13-stable/sys/kern/uipc_ktls.c (revision 3bc80996974a61a4223eae4c1ccd47b6ee32a48a)
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 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_rss.h"
32 
33 #include <sys/param.h>
34 #include <sys/kernel.h>
35 #include <sys/domainset.h>
36 #include <sys/ktls.h>
37 #include <sys/lock.h>
38 #include <sys/mbuf.h>
39 #include <sys/mutex.h>
40 #include <sys/rmlock.h>
41 #include <sys/proc.h>
42 #include <sys/protosw.h>
43 #include <sys/refcount.h>
44 #include <sys/smp.h>
45 #include <sys/socket.h>
46 #include <sys/socketvar.h>
47 #include <sys/sysctl.h>
48 #include <sys/taskqueue.h>
49 #include <sys/kthread.h>
50 #include <sys/uio.h>
51 #include <sys/vmmeter.h>
52 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
53 #include <machine/pcb.h>
54 #endif
55 #include <machine/vmparam.h>
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #ifdef RSS
59 #include <net/netisr.h>
60 #include <net/rss_config.h>
61 #endif
62 #include <net/route.h>
63 #include <net/route/nhop.h>
64 #if defined(INET) || defined(INET6)
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #endif
68 #include <netinet/tcp_var.h>
69 #ifdef TCP_OFFLOAD
70 #include <netinet/tcp_offload.h>
71 #endif
72 #include <opencrypto/xform.h>
73 #include <vm/uma_dbg.h>
74 #include <vm/vm.h>
75 #include <vm/vm_pageout.h>
76 #include <vm/vm_page.h>
77 
78 struct ktls_wq {
79 	struct mtx	mtx;
80 	STAILQ_HEAD(, mbuf) m_head;
81 	STAILQ_HEAD(, socket) so_head;
82 	bool		running;
83 } __aligned(CACHE_LINE_SIZE);
84 
85 struct ktls_domain_info {
86 	int count;
87 	int cpu[MAXCPU];
88 };
89 
90 struct ktls_domain_info ktls_domains[MAXMEMDOM];
91 static struct ktls_wq *ktls_wq;
92 static struct proc *ktls_proc;
93 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
94 static struct rmlock ktls_backends_lock;
95 static uma_zone_t ktls_session_zone;
96 static uint16_t ktls_cpuid_lookup[MAXCPU];
97 
98 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
99     "Kernel TLS offload");
100 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101     "Kernel TLS offload stats");
102 
103 static int ktls_allow_unload;
104 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
105     &ktls_allow_unload, 0, "Allow software crypto modules to unload");
106 
107 #ifdef RSS
108 static int ktls_bind_threads = 1;
109 #else
110 static int ktls_bind_threads;
111 #endif
112 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
113     &ktls_bind_threads, 0,
114     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
115 
116 static u_int ktls_maxlen = 16384;
117 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
118     &ktls_maxlen, 0, "Maximum TLS record size");
119 
120 static int ktls_number_threads;
121 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
122     &ktls_number_threads, 0,
123     "Number of TLS threads in thread-pool");
124 
125 static bool ktls_offload_enable;
126 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
127     &ktls_offload_enable, 0,
128     "Enable support for kernel TLS offload");
129 
130 static bool ktls_cbc_enable = true;
131 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
132     &ktls_cbc_enable, 1,
133     "Enable Support of AES-CBC crypto for kernel TLS");
134 
135 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
136 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
137     &ktls_tasks_active, "Number of active tasks");
138 
139 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
140 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
141     &ktls_cnt_tx_pending,
142     "Number of TLS 1.0 records waiting for earlier TLS records");
143 
144 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
145 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
146     &ktls_cnt_tx_queued,
147     "Number of TLS records in queue to tasks for SW encryption");
148 
149 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
150 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
151     &ktls_cnt_rx_queued,
152     "Number of TLS sockets in queue to tasks for SW decryption");
153 
154 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
155 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
156     CTLFLAG_RD, &ktls_offload_total,
157     "Total successful TLS setups (parameters set)");
158 
159 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
160 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
161     CTLFLAG_RD, &ktls_offload_enable_calls,
162     "Total number of TLS enable calls made");
163 
164 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
165 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
166     &ktls_offload_active, "Total Active TLS sessions");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
170     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
171 
172 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
173 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
174     &ktls_offload_failed_crypto, "Total TLS crypto failures");
175 
176 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
177 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
178     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
179 
180 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
182     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
183 
184 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
185 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
186     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
187 
188 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
189     "Software TLS session stats");
190 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
191     "Hardware (ifnet) TLS session stats");
192 #ifdef TCP_OFFLOAD
193 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
194     "TOE TLS session stats");
195 #endif
196 
197 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
199     "Active number of software TLS sessions using AES-CBC");
200 
201 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
203     "Active number of software TLS sessions using AES-GCM");
204 
205 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
207     &ktls_sw_chacha20,
208     "Active number of software TLS sessions using Chacha20-Poly1305");
209 
210 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
211 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
212     &ktls_ifnet_cbc,
213     "Active number of ifnet TLS sessions using AES-CBC");
214 
215 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
217     &ktls_ifnet_gcm,
218     "Active number of ifnet TLS sessions using AES-GCM");
219 
220 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
222     &ktls_ifnet_chacha20,
223     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
224 
225 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
226 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
227     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
228 
229 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
230 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
231     &ktls_ifnet_reset_dropped,
232     "TLS sessions dropped after failing to update ifnet send tag");
233 
234 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
236     &ktls_ifnet_reset_failed,
237     "TLS sessions that failed to allocate a new ifnet send tag");
238 
239 static int ktls_ifnet_permitted;
240 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
241     &ktls_ifnet_permitted, 1,
242     "Whether to permit hardware (ifnet) TLS sessions");
243 
244 #ifdef TCP_OFFLOAD
245 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
246 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
247     &ktls_toe_cbc,
248     "Active number of TOE TLS sessions using AES-CBC");
249 
250 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
251 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
252     &ktls_toe_gcm,
253     "Active number of TOE TLS sessions using AES-GCM");
254 
255 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
256 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
257     &ktls_toe_chacha20,
258     "Active number of TOE TLS sessions using Chacha20-Poly1305");
259 #endif
260 
261 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
262 
263 static void ktls_cleanup(struct ktls_session *tls);
264 #if defined(INET) || defined(INET6)
265 static void ktls_reset_send_tag(void *context, int pending);
266 #endif
267 static void ktls_work_thread(void *ctx);
268 
269 int
ktls_crypto_backend_register(struct ktls_crypto_backend * be)270 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
271 {
272 	struct ktls_crypto_backend *curr_be, *tmp;
273 
274 	if (be->api_version != KTLS_API_VERSION) {
275 		printf("KTLS: API version mismatch (%d vs %d) for %s\n",
276 		    be->api_version, KTLS_API_VERSION,
277 		    be->name);
278 		return (EINVAL);
279 	}
280 
281 	rm_wlock(&ktls_backends_lock);
282 	printf("KTLS: Registering crypto method %s with prio %d\n",
283 	       be->name, be->prio);
284 	if (LIST_EMPTY(&ktls_backends)) {
285 		LIST_INSERT_HEAD(&ktls_backends, be, next);
286 	} else {
287 		LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
288 			if (curr_be->prio < be->prio) {
289 				LIST_INSERT_BEFORE(curr_be, be, next);
290 				break;
291 			}
292 			if (LIST_NEXT(curr_be, next) == NULL) {
293 				LIST_INSERT_AFTER(curr_be, be, next);
294 				break;
295 			}
296 		}
297 	}
298 	rm_wunlock(&ktls_backends_lock);
299 	return (0);
300 }
301 
302 int
ktls_crypto_backend_deregister(struct ktls_crypto_backend * be)303 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
304 {
305 	struct ktls_crypto_backend *tmp;
306 
307 	/*
308 	 * Don't error if the backend isn't registered.  This permits
309 	 * MOD_UNLOAD handlers to use this function unconditionally.
310 	 */
311 	rm_wlock(&ktls_backends_lock);
312 	LIST_FOREACH(tmp, &ktls_backends, next) {
313 		if (tmp == be)
314 			break;
315 	}
316 	if (tmp == NULL) {
317 		rm_wunlock(&ktls_backends_lock);
318 		return (0);
319 	}
320 
321 	if (!ktls_allow_unload) {
322 		rm_wunlock(&ktls_backends_lock);
323 		printf(
324 		    "KTLS: Deregistering crypto method %s is not supported\n",
325 		    be->name);
326 		return (EBUSY);
327 	}
328 
329 	if (be->use_count) {
330 		rm_wunlock(&ktls_backends_lock);
331 		return (EBUSY);
332 	}
333 
334 	LIST_REMOVE(be, next);
335 	rm_wunlock(&ktls_backends_lock);
336 	return (0);
337 }
338 
339 #if defined(INET) || defined(INET6)
340 static u_int
ktls_get_cpu(struct socket * so)341 ktls_get_cpu(struct socket *so)
342 {
343 	struct inpcb *inp;
344 #ifdef NUMA
345 	struct ktls_domain_info *di;
346 #endif
347 	u_int cpuid;
348 
349 	inp = sotoinpcb(so);
350 #ifdef RSS
351 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
352 	if (cpuid != NETISR_CPUID_NONE)
353 		return (cpuid);
354 #endif
355 	/*
356 	 * Just use the flowid to shard connections in a repeatable
357 	 * fashion.  Note that some crypto backends rely on the
358 	 * serialization provided by having the same connection use
359 	 * the same queue.
360 	 */
361 #ifdef NUMA
362 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
363 		di = &ktls_domains[inp->inp_numa_domain];
364 		cpuid = di->cpu[inp->inp_flowid % di->count];
365 	} else
366 #endif
367 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
368 	return (cpuid);
369 }
370 #endif
371 
372 static void
ktls_init(void * dummy __unused)373 ktls_init(void *dummy __unused)
374 {
375 	struct thread *td;
376 	struct pcpu *pc;
377 	cpuset_t mask;
378 	int count, domain, error, i;
379 
380 	rm_init(&ktls_backends_lock, "ktls backends");
381 	LIST_INIT(&ktls_backends);
382 
383 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
384 	    M_WAITOK | M_ZERO);
385 
386 	ktls_session_zone = uma_zcreate("ktls_session",
387 	    sizeof(struct ktls_session),
388 	    NULL, NULL, NULL, NULL,
389 	    UMA_ALIGN_CACHE, 0);
390 
391 	/*
392 	 * Initialize the workqueues to run the TLS work.  We create a
393 	 * work queue for each CPU.
394 	 */
395 	CPU_FOREACH(i) {
396 		STAILQ_INIT(&ktls_wq[i].m_head);
397 		STAILQ_INIT(&ktls_wq[i].so_head);
398 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
399 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
400 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
401 		if (error)
402 			panic("Can't add KTLS thread %d error %d", i, error);
403 
404 		/*
405 		 * Bind threads to cores.  If ktls_bind_threads is >
406 		 * 1, then we bind to the NUMA domain.
407 		 */
408 		if (ktls_bind_threads) {
409 			if (ktls_bind_threads > 1) {
410 				pc = pcpu_find(i);
411 				domain = pc->pc_domain;
412 				CPU_COPY(&cpuset_domain[domain], &mask);
413 				count = ktls_domains[domain].count;
414 				ktls_domains[domain].cpu[count] = i;
415 				ktls_domains[domain].count++;
416 			} else {
417 				CPU_SETOF(i, &mask);
418 			}
419 			error = cpuset_setthread(td->td_tid, &mask);
420 			if (error)
421 				panic(
422 			    "Unable to bind KTLS thread for CPU %d error %d",
423 				     i, error);
424 		}
425 		ktls_cpuid_lookup[ktls_number_threads] = i;
426 		ktls_number_threads++;
427 	}
428 
429 	/*
430 	 * If we somehow have an empty domain, fall back to choosing
431 	 * among all KTLS threads.
432 	 */
433 	if (ktls_bind_threads > 1) {
434 		for (i = 0; i < vm_ndomains; i++) {
435 			if (ktls_domains[i].count == 0) {
436 				ktls_bind_threads = 1;
437 				break;
438 			}
439 		}
440 	}
441 
442 	if (bootverbose)
443 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
444 }
445 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
446 
447 #if defined(INET) || defined(INET6)
448 static int
ktls_create_session(struct socket * so,struct tls_enable * en,struct ktls_session ** tlsp)449 ktls_create_session(struct socket *so, struct tls_enable *en,
450     struct ktls_session **tlsp)
451 {
452 	struct ktls_session *tls;
453 	int error;
454 
455 	/* Only TLS 1.0 - 1.3 are supported. */
456 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
457 		return (EINVAL);
458 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
459 	    en->tls_vminor > TLS_MINOR_VER_THREE)
460 		return (EINVAL);
461 
462 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
463 		return (EINVAL);
464 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
465 		return (EINVAL);
466 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
467 		return (EINVAL);
468 
469 	/* All supported algorithms require a cipher key. */
470 	if (en->cipher_key_len == 0)
471 		return (EINVAL);
472 
473 	/* No flags are currently supported. */
474 	if (en->flags != 0)
475 		return (EINVAL);
476 
477 	/* Common checks for supported algorithms. */
478 	switch (en->cipher_algorithm) {
479 	case CRYPTO_AES_NIST_GCM_16:
480 		/*
481 		 * auth_algorithm isn't used, but permit GMAC values
482 		 * for compatibility.
483 		 */
484 		switch (en->auth_algorithm) {
485 		case 0:
486 #ifdef COMPAT_FREEBSD12
487 		/* XXX: Really 13.0-current COMPAT. */
488 		case CRYPTO_AES_128_NIST_GMAC:
489 		case CRYPTO_AES_192_NIST_GMAC:
490 		case CRYPTO_AES_256_NIST_GMAC:
491 #endif
492 			break;
493 		default:
494 			return (EINVAL);
495 		}
496 		if (en->auth_key_len != 0)
497 			return (EINVAL);
498 		switch (en->tls_vminor) {
499 		case TLS_MINOR_VER_TWO:
500 			if (en->iv_len != TLS_AEAD_GCM_LEN)
501 				return (EINVAL);
502 			break;
503 		case TLS_MINOR_VER_THREE:
504 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
505 				return (EINVAL);
506 			break;
507 		default:
508 			return (EINVAL);
509 		}
510 		break;
511 	case CRYPTO_AES_CBC:
512 		switch (en->auth_algorithm) {
513 		case CRYPTO_SHA1_HMAC:
514 			break;
515 		case CRYPTO_SHA2_256_HMAC:
516 		case CRYPTO_SHA2_384_HMAC:
517 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
518 				return (EINVAL);
519 			break;
520 		default:
521 			return (EINVAL);
522 		}
523 		if (en->auth_key_len == 0)
524 			return (EINVAL);
525 
526 		/*
527 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
528 		 * use explicit IVs.
529 		 */
530 		switch (en->tls_vminor) {
531 		case TLS_MINOR_VER_ZERO:
532 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
533 				return (EINVAL);
534 			break;
535 		case TLS_MINOR_VER_ONE:
536 		case TLS_MINOR_VER_TWO:
537 			/* Ignore any supplied IV. */
538 			en->iv_len = 0;
539 			break;
540 		default:
541 			return (EINVAL);
542 		}
543 		break;
544 	case CRYPTO_CHACHA20_POLY1305:
545 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
546 			return (EINVAL);
547 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
548 		    en->tls_vminor != TLS_MINOR_VER_THREE)
549 			return (EINVAL);
550 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
551 			return (EINVAL);
552 		break;
553 	default:
554 		return (EINVAL);
555 	}
556 
557 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
558 
559 	counter_u64_add(ktls_offload_active, 1);
560 
561 	refcount_init(&tls->refcount, 1);
562 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
563 
564 	tls->wq_index = ktls_get_cpu(so);
565 
566 	tls->params.cipher_algorithm = en->cipher_algorithm;
567 	tls->params.auth_algorithm = en->auth_algorithm;
568 	tls->params.tls_vmajor = en->tls_vmajor;
569 	tls->params.tls_vminor = en->tls_vminor;
570 	tls->params.flags = en->flags;
571 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
572 
573 	/* Set the header and trailer lengths. */
574 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
575 	switch (en->cipher_algorithm) {
576 	case CRYPTO_AES_NIST_GCM_16:
577 		/*
578 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
579 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
580 		 */
581 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
582 			tls->params.tls_hlen += sizeof(uint64_t);
583 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
584 		tls->params.tls_bs = 1;
585 		break;
586 	case CRYPTO_AES_CBC:
587 		switch (en->auth_algorithm) {
588 		case CRYPTO_SHA1_HMAC:
589 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
590 				/* Implicit IV, no nonce. */
591 				tls->sequential_records = true;
592 				tls->next_seqno = be64dec(en->rec_seq);
593 				STAILQ_INIT(&tls->pending_records);
594 			} else {
595 				tls->params.tls_hlen += AES_BLOCK_LEN;
596 			}
597 			tls->params.tls_tlen = AES_BLOCK_LEN +
598 			    SHA1_HASH_LEN;
599 			break;
600 		case CRYPTO_SHA2_256_HMAC:
601 			tls->params.tls_hlen += AES_BLOCK_LEN;
602 			tls->params.tls_tlen = AES_BLOCK_LEN +
603 			    SHA2_256_HASH_LEN;
604 			break;
605 		case CRYPTO_SHA2_384_HMAC:
606 			tls->params.tls_hlen += AES_BLOCK_LEN;
607 			tls->params.tls_tlen = AES_BLOCK_LEN +
608 			    SHA2_384_HASH_LEN;
609 			break;
610 		default:
611 			panic("invalid hmac");
612 		}
613 		tls->params.tls_bs = AES_BLOCK_LEN;
614 		break;
615 	case CRYPTO_CHACHA20_POLY1305:
616 		/*
617 		 * Chacha20 uses a 12 byte implicit IV.
618 		 */
619 		tls->params.tls_tlen = POLY1305_HASH_LEN;
620 		tls->params.tls_bs = 1;
621 		break;
622 	default:
623 		panic("invalid cipher");
624 	}
625 
626 	/*
627 	 * TLS 1.3 includes optional padding which we do not support,
628 	 * and also puts the "real" record type at the end of the
629 	 * encrypted data.
630 	 */
631 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
632 		tls->params.tls_tlen += sizeof(uint8_t);
633 
634 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
635 	    ("TLS header length too long: %d", tls->params.tls_hlen));
636 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
637 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
638 
639 	if (en->auth_key_len != 0) {
640 		tls->params.auth_key_len = en->auth_key_len;
641 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
642 		    M_WAITOK);
643 		error = copyin(en->auth_key, tls->params.auth_key,
644 		    en->auth_key_len);
645 		if (error)
646 			goto out;
647 	}
648 
649 	tls->params.cipher_key_len = en->cipher_key_len;
650 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
651 	error = copyin(en->cipher_key, tls->params.cipher_key,
652 	    en->cipher_key_len);
653 	if (error)
654 		goto out;
655 
656 	/*
657 	 * This holds the implicit portion of the nonce for AEAD
658 	 * ciphers and the initial implicit IV for TLS 1.0.  The
659 	 * explicit portions of the IV are generated in ktls_frame().
660 	 */
661 	if (en->iv_len != 0) {
662 		tls->params.iv_len = en->iv_len;
663 		error = copyin(en->iv, tls->params.iv, en->iv_len);
664 		if (error)
665 			goto out;
666 
667 		/*
668 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
669 		 * counter to generate unique explicit IVs.
670 		 *
671 		 * Store this counter in the last 8 bytes of the IV
672 		 * array so that it is 8-byte aligned.
673 		 */
674 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
675 		    en->tls_vminor == TLS_MINOR_VER_TWO)
676 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
677 	}
678 
679 	*tlsp = tls;
680 	return (0);
681 
682 out:
683 	ktls_free(tls);
684 	return (error);
685 }
686 
687 static struct ktls_session *
ktls_clone_session(struct ktls_session * tls)688 ktls_clone_session(struct ktls_session *tls)
689 {
690 	struct ktls_session *tls_new;
691 
692 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
693 
694 	counter_u64_add(ktls_offload_active, 1);
695 
696 	refcount_init(&tls_new->refcount, 1);
697 	TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
698 
699 	/* Copy fields from existing session. */
700 	tls_new->params = tls->params;
701 	tls_new->wq_index = tls->wq_index;
702 
703 	/* Deep copy keys. */
704 	if (tls_new->params.auth_key != NULL) {
705 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
706 		    M_KTLS, M_WAITOK);
707 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
708 		    tls->params.auth_key_len);
709 	}
710 
711 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
712 	    M_WAITOK);
713 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
714 	    tls->params.cipher_key_len);
715 
716 	return (tls_new);
717 }
718 #endif
719 
720 static void
ktls_cleanup(struct ktls_session * tls)721 ktls_cleanup(struct ktls_session *tls)
722 {
723 
724 	counter_u64_add(ktls_offload_active, -1);
725 	switch (tls->mode) {
726 	case TCP_TLS_MODE_SW:
727 		MPASS(tls->be != NULL);
728 		switch (tls->params.cipher_algorithm) {
729 		case CRYPTO_AES_CBC:
730 			counter_u64_add(ktls_sw_cbc, -1);
731 			break;
732 		case CRYPTO_AES_NIST_GCM_16:
733 			counter_u64_add(ktls_sw_gcm, -1);
734 			break;
735 		case CRYPTO_CHACHA20_POLY1305:
736 			counter_u64_add(ktls_sw_chacha20, -1);
737 			break;
738 		}
739 		tls->free(tls);
740 		break;
741 	case TCP_TLS_MODE_IFNET:
742 		switch (tls->params.cipher_algorithm) {
743 		case CRYPTO_AES_CBC:
744 			counter_u64_add(ktls_ifnet_cbc, -1);
745 			break;
746 		case CRYPTO_AES_NIST_GCM_16:
747 			counter_u64_add(ktls_ifnet_gcm, -1);
748 			break;
749 		case CRYPTO_CHACHA20_POLY1305:
750 			counter_u64_add(ktls_ifnet_chacha20, -1);
751 			break;
752 		}
753 		if (tls->snd_tag != NULL)
754 			m_snd_tag_rele(tls->snd_tag);
755 		break;
756 #ifdef TCP_OFFLOAD
757 	case TCP_TLS_MODE_TOE:
758 		switch (tls->params.cipher_algorithm) {
759 		case CRYPTO_AES_CBC:
760 			counter_u64_add(ktls_toe_cbc, -1);
761 			break;
762 		case CRYPTO_AES_NIST_GCM_16:
763 			counter_u64_add(ktls_toe_gcm, -1);
764 			break;
765 		case CRYPTO_CHACHA20_POLY1305:
766 			counter_u64_add(ktls_toe_chacha20, -1);
767 			break;
768 		}
769 		break;
770 #endif
771 	}
772 	if (tls->params.auth_key != NULL) {
773 		zfree(tls->params.auth_key, M_KTLS);
774 		tls->params.auth_key = NULL;
775 		tls->params.auth_key_len = 0;
776 	}
777 	if (tls->params.cipher_key != NULL) {
778 		zfree(tls->params.cipher_key, M_KTLS);
779 		tls->params.cipher_key = NULL;
780 		tls->params.cipher_key_len = 0;
781 	}
782 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
783 }
784 
785 #if defined(INET) || defined(INET6)
786 
787 #ifdef TCP_OFFLOAD
788 static int
ktls_try_toe(struct socket * so,struct ktls_session * tls,int direction)789 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
790 {
791 	struct inpcb *inp;
792 	struct tcpcb *tp;
793 	int error;
794 
795 	inp = so->so_pcb;
796 	INP_WLOCK(inp);
797 	if (inp->inp_flags2 & INP_FREED) {
798 		INP_WUNLOCK(inp);
799 		return (ECONNRESET);
800 	}
801 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
802 		INP_WUNLOCK(inp);
803 		return (ECONNRESET);
804 	}
805 	if (inp->inp_socket == NULL) {
806 		INP_WUNLOCK(inp);
807 		return (ECONNRESET);
808 	}
809 	tp = intotcpcb(inp);
810 	if (!(tp->t_flags & TF_TOE)) {
811 		INP_WUNLOCK(inp);
812 		return (EOPNOTSUPP);
813 	}
814 
815 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
816 	INP_WUNLOCK(inp);
817 	if (error == 0) {
818 		tls->mode = TCP_TLS_MODE_TOE;
819 		switch (tls->params.cipher_algorithm) {
820 		case CRYPTO_AES_CBC:
821 			counter_u64_add(ktls_toe_cbc, 1);
822 			break;
823 		case CRYPTO_AES_NIST_GCM_16:
824 			counter_u64_add(ktls_toe_gcm, 1);
825 			break;
826 		case CRYPTO_CHACHA20_POLY1305:
827 			counter_u64_add(ktls_toe_chacha20, 1);
828 			break;
829 		}
830 	}
831 	return (error);
832 }
833 #endif
834 
835 /*
836  * Common code used when first enabling ifnet TLS on a connection or
837  * when allocating a new ifnet TLS session due to a routing change.
838  * This function allocates a new TLS send tag on whatever interface
839  * the connection is currently routed over.
840  */
841 static int
ktls_alloc_snd_tag(struct inpcb * inp,struct ktls_session * tls,bool force,struct m_snd_tag ** mstp)842 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
843     struct m_snd_tag **mstp)
844 {
845 	union if_snd_tag_alloc_params params;
846 	struct ifnet *ifp;
847 	struct nhop_object *nh;
848 	struct tcpcb *tp;
849 	int error;
850 
851 	INP_RLOCK(inp);
852 	if (inp->inp_flags2 & INP_FREED) {
853 		INP_RUNLOCK(inp);
854 		return (ECONNRESET);
855 	}
856 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
857 		INP_RUNLOCK(inp);
858 		return (ECONNRESET);
859 	}
860 	if (inp->inp_socket == NULL) {
861 		INP_RUNLOCK(inp);
862 		return (ECONNRESET);
863 	}
864 	tp = intotcpcb(inp);
865 
866 	/*
867 	 * Check administrative controls on ifnet TLS to determine if
868 	 * ifnet TLS should be denied.
869 	 *
870 	 * - Always permit 'force' requests.
871 	 * - ktls_ifnet_permitted == 0: always deny.
872 	 */
873 	if (!force && ktls_ifnet_permitted == 0) {
874 		INP_RUNLOCK(inp);
875 		return (ENXIO);
876 	}
877 
878 	/*
879 	 * XXX: Use the cached route in the inpcb to find the
880 	 * interface.  This should perhaps instead use
881 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
882 	 * enabled after a connection has completed key negotiation in
883 	 * userland, the cached route will be present in practice.
884 	 */
885 	nh = inp->inp_route.ro_nh;
886 	if (nh == NULL) {
887 		INP_RUNLOCK(inp);
888 		return (ENXIO);
889 	}
890 	ifp = nh->nh_ifp;
891 	if_ref(ifp);
892 
893 	/*
894 	 * Allocate a TLS + ratelimit tag if the connection has an
895 	 * existing pacing rate.
896 	 */
897 	if (tp->t_pacing_rate != -1 &&
898 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
899 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
900 		params.tls_rate_limit.inp = inp;
901 		params.tls_rate_limit.tls = tls;
902 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
903 	} else {
904 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
905 		params.tls.inp = inp;
906 		params.tls.tls = tls;
907 	}
908 	params.hdr.flowid = inp->inp_flowid;
909 	params.hdr.flowtype = inp->inp_flowtype;
910 	params.hdr.numa_domain = inp->inp_numa_domain;
911 	INP_RUNLOCK(inp);
912 
913 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
914 		error = EOPNOTSUPP;
915 		goto out;
916 	}
917 	if (inp->inp_vflag & INP_IPV6) {
918 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
919 			error = EOPNOTSUPP;
920 			goto out;
921 		}
922 	} else {
923 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
924 			error = EOPNOTSUPP;
925 			goto out;
926 		}
927 	}
928 	error = m_snd_tag_alloc(ifp, &params, mstp);
929 out:
930 	if_rele(ifp);
931 	return (error);
932 }
933 
934 static int
ktls_try_ifnet(struct socket * so,struct ktls_session * tls,bool force)935 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
936 {
937 	struct m_snd_tag *mst;
938 	int error;
939 
940 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
941 	if (error == 0) {
942 		tls->mode = TCP_TLS_MODE_IFNET;
943 		tls->snd_tag = mst;
944 		switch (tls->params.cipher_algorithm) {
945 		case CRYPTO_AES_CBC:
946 			counter_u64_add(ktls_ifnet_cbc, 1);
947 			break;
948 		case CRYPTO_AES_NIST_GCM_16:
949 			counter_u64_add(ktls_ifnet_gcm, 1);
950 			break;
951 		case CRYPTO_CHACHA20_POLY1305:
952 			counter_u64_add(ktls_ifnet_chacha20, 1);
953 			break;
954 		}
955 	}
956 	return (error);
957 }
958 
959 static int
ktls_try_sw(struct socket * so,struct ktls_session * tls,int direction)960 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
961 {
962 	struct rm_priotracker prio;
963 	struct ktls_crypto_backend *be;
964 
965 	/*
966 	 * Choose the best software crypto backend.  Backends are
967 	 * stored in sorted priority order (larget value == most
968 	 * important at the head of the list), so this just stops on
969 	 * the first backend that claims the session by returning
970 	 * success.
971 	 */
972 	if (ktls_allow_unload)
973 		rm_rlock(&ktls_backends_lock, &prio);
974 	LIST_FOREACH(be, &ktls_backends, next) {
975 		if (be->try(so, tls, direction) == 0)
976 			break;
977 		KASSERT(tls->cipher == NULL,
978 		    ("ktls backend leaked a cipher pointer"));
979 	}
980 	if (be != NULL) {
981 		if (ktls_allow_unload)
982 			be->use_count++;
983 		tls->be = be;
984 	}
985 	if (ktls_allow_unload)
986 		rm_runlock(&ktls_backends_lock, &prio);
987 	if (be == NULL)
988 		return (EOPNOTSUPP);
989 	tls->mode = TCP_TLS_MODE_SW;
990 	switch (tls->params.cipher_algorithm) {
991 	case CRYPTO_AES_CBC:
992 		counter_u64_add(ktls_sw_cbc, 1);
993 		break;
994 	case CRYPTO_AES_NIST_GCM_16:
995 		counter_u64_add(ktls_sw_gcm, 1);
996 		break;
997 	case CRYPTO_CHACHA20_POLY1305:
998 		counter_u64_add(ktls_sw_chacha20, 1);
999 		break;
1000 	}
1001 	return (0);
1002 }
1003 
1004 /*
1005  * KTLS RX stores data in the socket buffer as a list of TLS records,
1006  * where each record is stored as a control message containg the TLS
1007  * header followed by data mbufs containing the decrypted data.  This
1008  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1009  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1010  * should be queued to the socket buffer as records, but encrypted
1011  * data which needs to be decrypted by software arrives as a stream of
1012  * regular mbufs which need to be converted.  In addition, there may
1013  * already be pending encrypted data in the socket buffer when KTLS RX
1014  * is enabled.
1015  *
1016  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1017  * is used:
1018  *
1019  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1020  *
1021  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1022  *   from the first mbuf.  Once all of the data for that TLS record is
1023  *   queued, the socket is queued to a worker thread.
1024  *
1025  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1026  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1027  *   decrypted, and inserted into the regular socket buffer chain as
1028  *   record starting with a control message holding the TLS header and
1029  *   a chain of mbufs holding the encrypted data.
1030  */
1031 
1032 static void
sb_mark_notready(struct sockbuf * sb)1033 sb_mark_notready(struct sockbuf *sb)
1034 {
1035 	struct mbuf *m;
1036 
1037 	m = sb->sb_mb;
1038 	sb->sb_mtls = m;
1039 	sb->sb_mb = NULL;
1040 	sb->sb_mbtail = NULL;
1041 	sb->sb_lastrecord = NULL;
1042 	for (; m != NULL; m = m->m_next) {
1043 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1044 		    __func__));
1045 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1046 		    __func__));
1047 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1048 		    __func__));
1049 		m->m_flags |= M_NOTREADY;
1050 		sb->sb_acc -= m->m_len;
1051 		sb->sb_tlscc += m->m_len;
1052 		sb->sb_mtlstail = m;
1053 	}
1054 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1055 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1056 	    sb->sb_ccc));
1057 }
1058 
1059 int
ktls_enable_rx(struct socket * so,struct tls_enable * en)1060 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1061 {
1062 	struct ktls_session *tls;
1063 	int error;
1064 
1065 	if (!ktls_offload_enable)
1066 		return (ENOTSUP);
1067 	if (SOLISTENING(so))
1068 		return (EINVAL);
1069 
1070 	counter_u64_add(ktls_offload_enable_calls, 1);
1071 
1072 	/*
1073 	 * This should always be true since only the TCP socket option
1074 	 * invokes this function.
1075 	 */
1076 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1077 		return (EINVAL);
1078 
1079 	/*
1080 	 * XXX: Don't overwrite existing sessions.  We should permit
1081 	 * this to support rekeying in the future.
1082 	 */
1083 	if (so->so_rcv.sb_tls_info != NULL)
1084 		return (EALREADY);
1085 
1086 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1087 		return (ENOTSUP);
1088 
1089 	error = ktls_create_session(so, en, &tls);
1090 	if (error)
1091 		return (error);
1092 
1093 #ifdef TCP_OFFLOAD
1094 	error = ktls_try_toe(so, tls, KTLS_RX);
1095 	if (error)
1096 #endif
1097 		error = ktls_try_sw(so, tls, KTLS_RX);
1098 
1099 	if (error) {
1100 		ktls_free(tls);
1101 		return (error);
1102 	}
1103 
1104 	/* Mark the socket as using TLS offload. */
1105 	SOCKBUF_LOCK(&so->so_rcv);
1106 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1107 	so->so_rcv.sb_tls_info = tls;
1108 	so->so_rcv.sb_flags |= SB_TLS_RX;
1109 
1110 	/* Mark existing data as not ready until it can be decrypted. */
1111 	if (tls->mode != TCP_TLS_MODE_TOE) {
1112 		sb_mark_notready(&so->so_rcv);
1113 		ktls_check_rx(&so->so_rcv);
1114 	}
1115 	SOCKBUF_UNLOCK(&so->so_rcv);
1116 
1117 	counter_u64_add(ktls_offload_total, 1);
1118 
1119 	return (0);
1120 }
1121 
1122 int
ktls_enable_tx(struct socket * so,struct tls_enable * en)1123 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1124 {
1125 	struct ktls_session *tls;
1126 	struct inpcb *inp;
1127 	int error;
1128 
1129 	if (!ktls_offload_enable)
1130 		return (ENOTSUP);
1131 	if (SOLISTENING(so))
1132 		return (EINVAL);
1133 
1134 	counter_u64_add(ktls_offload_enable_calls, 1);
1135 
1136 	/*
1137 	 * This should always be true since only the TCP socket option
1138 	 * invokes this function.
1139 	 */
1140 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1141 		return (EINVAL);
1142 
1143 	/*
1144 	 * XXX: Don't overwrite existing sessions.  We should permit
1145 	 * this to support rekeying in the future.
1146 	 */
1147 	if (so->so_snd.sb_tls_info != NULL)
1148 		return (EALREADY);
1149 
1150 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1151 		return (ENOTSUP);
1152 
1153 	/* TLS requires ext pgs */
1154 	if (mb_use_ext_pgs == 0)
1155 		return (ENXIO);
1156 
1157 	error = ktls_create_session(so, en, &tls);
1158 	if (error)
1159 		return (error);
1160 
1161 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1162 #ifdef TCP_OFFLOAD
1163 	error = ktls_try_toe(so, tls, KTLS_TX);
1164 	if (error)
1165 #endif
1166 		error = ktls_try_ifnet(so, tls, false);
1167 	if (error)
1168 		error = ktls_try_sw(so, tls, KTLS_TX);
1169 
1170 	if (error) {
1171 		ktls_free(tls);
1172 		return (error);
1173 	}
1174 
1175 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1176 	if (error) {
1177 		ktls_free(tls);
1178 		return (error);
1179 	}
1180 
1181 	/*
1182 	 * Write lock the INP when setting sb_tls_info so that
1183 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1184 	 * holding the INP lock.
1185 	 */
1186 	inp = so->so_pcb;
1187 	INP_WLOCK(inp);
1188 	SOCKBUF_LOCK(&so->so_snd);
1189 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1190 	so->so_snd.sb_tls_info = tls;
1191 	if (tls->mode != TCP_TLS_MODE_SW)
1192 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1193 	SOCKBUF_UNLOCK(&so->so_snd);
1194 	INP_WUNLOCK(inp);
1195 	SOCK_IO_SEND_UNLOCK(so);
1196 
1197 	counter_u64_add(ktls_offload_total, 1);
1198 
1199 	return (0);
1200 }
1201 
1202 int
ktls_get_rx_mode(struct socket * so)1203 ktls_get_rx_mode(struct socket *so)
1204 {
1205 	struct ktls_session *tls;
1206 	struct inpcb *inp;
1207 	int mode;
1208 
1209 	if (SOLISTENING(so))
1210 		return (EINVAL);
1211 	inp = so->so_pcb;
1212 	INP_WLOCK_ASSERT(inp);
1213 	SOCKBUF_LOCK(&so->so_rcv);
1214 	tls = so->so_rcv.sb_tls_info;
1215 	if (tls == NULL)
1216 		mode = TCP_TLS_MODE_NONE;
1217 	else
1218 		mode = tls->mode;
1219 	SOCKBUF_UNLOCK(&so->so_rcv);
1220 	return (mode);
1221 }
1222 
1223 int
ktls_get_tx_mode(struct socket * so)1224 ktls_get_tx_mode(struct socket *so)
1225 {
1226 	struct ktls_session *tls;
1227 	struct inpcb *inp;
1228 	int mode;
1229 
1230 	if (SOLISTENING(so))
1231 		return (EINVAL);
1232 	inp = so->so_pcb;
1233 	INP_WLOCK_ASSERT(inp);
1234 	SOCKBUF_LOCK(&so->so_snd);
1235 	tls = so->so_snd.sb_tls_info;
1236 	if (tls == NULL)
1237 		mode = TCP_TLS_MODE_NONE;
1238 	else
1239 		mode = tls->mode;
1240 	SOCKBUF_UNLOCK(&so->so_snd);
1241 	return (mode);
1242 }
1243 
1244 /*
1245  * Switch between SW and ifnet TLS sessions as requested.
1246  */
1247 int
ktls_set_tx_mode(struct socket * so,int mode)1248 ktls_set_tx_mode(struct socket *so, int mode)
1249 {
1250 	struct ktls_session *tls, *tls_new;
1251 	struct inpcb *inp;
1252 	int error;
1253 
1254 	if (SOLISTENING(so))
1255 		return (EINVAL);
1256 	switch (mode) {
1257 	case TCP_TLS_MODE_SW:
1258 	case TCP_TLS_MODE_IFNET:
1259 		break;
1260 	default:
1261 		return (EINVAL);
1262 	}
1263 
1264 	inp = so->so_pcb;
1265 	INP_WLOCK_ASSERT(inp);
1266 	SOCKBUF_LOCK(&so->so_snd);
1267 	tls = so->so_snd.sb_tls_info;
1268 	if (tls == NULL) {
1269 		SOCKBUF_UNLOCK(&so->so_snd);
1270 		return (0);
1271 	}
1272 
1273 	if (tls->mode == mode) {
1274 		SOCKBUF_UNLOCK(&so->so_snd);
1275 		return (0);
1276 	}
1277 
1278 	tls = ktls_hold(tls);
1279 	SOCKBUF_UNLOCK(&so->so_snd);
1280 	INP_WUNLOCK(inp);
1281 
1282 	tls_new = ktls_clone_session(tls);
1283 
1284 	if (mode == TCP_TLS_MODE_IFNET)
1285 		error = ktls_try_ifnet(so, tls_new, true);
1286 	else
1287 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1288 	if (error) {
1289 		counter_u64_add(ktls_switch_failed, 1);
1290 		ktls_free(tls_new);
1291 		ktls_free(tls);
1292 		INP_WLOCK(inp);
1293 		return (error);
1294 	}
1295 
1296 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1297 	if (error) {
1298 		counter_u64_add(ktls_switch_failed, 1);
1299 		ktls_free(tls_new);
1300 		ktls_free(tls);
1301 		INP_WLOCK(inp);
1302 		return (error);
1303 	}
1304 
1305 	/*
1306 	 * If we raced with another session change, keep the existing
1307 	 * session.
1308 	 */
1309 	if (tls != so->so_snd.sb_tls_info) {
1310 		counter_u64_add(ktls_switch_failed, 1);
1311 		SOCK_IO_SEND_UNLOCK(so);
1312 		ktls_free(tls_new);
1313 		ktls_free(tls);
1314 		INP_WLOCK(inp);
1315 		return (EBUSY);
1316 	}
1317 
1318 	INP_WLOCK(inp);
1319 	SOCKBUF_LOCK(&so->so_snd);
1320 	so->so_snd.sb_tls_info = tls_new;
1321 	if (tls_new->mode != TCP_TLS_MODE_SW)
1322 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1323 	SOCKBUF_UNLOCK(&so->so_snd);
1324 	SOCK_IO_SEND_UNLOCK(so);
1325 
1326 	/*
1327 	 * Drop two references on 'tls'.  The first is for the
1328 	 * ktls_hold() above.  The second drops the reference from the
1329 	 * socket buffer.
1330 	 */
1331 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1332 	ktls_free(tls);
1333 	ktls_free(tls);
1334 
1335 	if (mode == TCP_TLS_MODE_IFNET)
1336 		counter_u64_add(ktls_switch_to_ifnet, 1);
1337 	else
1338 		counter_u64_add(ktls_switch_to_sw, 1);
1339 
1340 	return (0);
1341 }
1342 
1343 /*
1344  * Try to allocate a new TLS send tag.  This task is scheduled when
1345  * ip_output detects a route change while trying to transmit a packet
1346  * holding a TLS record.  If a new tag is allocated, replace the tag
1347  * in the TLS session.  Subsequent packets on the connection will use
1348  * the new tag.  If a new tag cannot be allocated, drop the
1349  * connection.
1350  */
1351 static void
ktls_reset_send_tag(void * context,int pending)1352 ktls_reset_send_tag(void *context, int pending)
1353 {
1354 	struct epoch_tracker et;
1355 	struct ktls_session *tls;
1356 	struct m_snd_tag *old, *new;
1357 	struct inpcb *inp;
1358 	struct tcpcb *tp;
1359 	int error;
1360 
1361 	MPASS(pending == 1);
1362 
1363 	tls = context;
1364 	inp = tls->inp;
1365 
1366 	/*
1367 	 * Free the old tag first before allocating a new one.
1368 	 * ip[6]_output_send() will treat a NULL send tag the same as
1369 	 * an ifp mismatch and drop packets until a new tag is
1370 	 * allocated.
1371 	 *
1372 	 * Write-lock the INP when changing tls->snd_tag since
1373 	 * ip[6]_output_send() holds a read-lock when reading the
1374 	 * pointer.
1375 	 */
1376 	INP_WLOCK(inp);
1377 	old = tls->snd_tag;
1378 	tls->snd_tag = NULL;
1379 	INP_WUNLOCK(inp);
1380 	if (old != NULL)
1381 		m_snd_tag_rele(old);
1382 
1383 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1384 
1385 	if (error == 0) {
1386 		INP_WLOCK(inp);
1387 		tls->snd_tag = new;
1388 		mtx_pool_lock(mtxpool_sleep, tls);
1389 		tls->reset_pending = false;
1390 		mtx_pool_unlock(mtxpool_sleep, tls);
1391 		if (!in_pcbrele_wlocked(inp))
1392 			INP_WUNLOCK(inp);
1393 
1394 		counter_u64_add(ktls_ifnet_reset, 1);
1395 
1396 		/*
1397 		 * XXX: Should we kick tcp_output explicitly now that
1398 		 * the send tag is fixed or just rely on timers?
1399 		 */
1400 	} else {
1401 		NET_EPOCH_ENTER(et);
1402 		INP_WLOCK(inp);
1403 		if (!in_pcbrele_wlocked(inp)) {
1404 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1405 			    !(inp->inp_flags & INP_DROPPED)) {
1406 				tp = intotcpcb(inp);
1407 				CURVNET_SET(tp->t_vnet);
1408 				tp = tcp_drop(tp, ECONNABORTED);
1409 				CURVNET_RESTORE();
1410 				if (tp != NULL)
1411 					INP_WUNLOCK(inp);
1412 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1413 			} else
1414 				INP_WUNLOCK(inp);
1415 		}
1416 		NET_EPOCH_EXIT(et);
1417 
1418 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1419 
1420 		/*
1421 		 * Leave reset_pending true to avoid future tasks while
1422 		 * the socket goes away.
1423 		 */
1424 	}
1425 
1426 	ktls_free(tls);
1427 }
1428 
1429 int
ktls_output_eagain(struct inpcb * inp,struct ktls_session * tls)1430 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1431 {
1432 
1433 	if (inp == NULL)
1434 		return (ENOBUFS);
1435 
1436 	INP_LOCK_ASSERT(inp);
1437 
1438 	/*
1439 	 * See if we should schedule a task to update the send tag for
1440 	 * this session.
1441 	 */
1442 	mtx_pool_lock(mtxpool_sleep, tls);
1443 	if (!tls->reset_pending) {
1444 		(void) ktls_hold(tls);
1445 		in_pcbref(inp);
1446 		tls->inp = inp;
1447 		tls->reset_pending = true;
1448 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1449 	}
1450 	mtx_pool_unlock(mtxpool_sleep, tls);
1451 	return (ENOBUFS);
1452 }
1453 
1454 #ifdef RATELIMIT
1455 int
ktls_modify_txrtlmt(struct ktls_session * tls,uint64_t max_pacing_rate)1456 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1457 {
1458 	union if_snd_tag_modify_params params = {
1459 		.rate_limit.max_rate = max_pacing_rate,
1460 		.rate_limit.flags = M_NOWAIT,
1461 	};
1462 	struct m_snd_tag *mst;
1463 	struct ifnet *ifp;
1464 	int error;
1465 
1466 	/* Can't get to the inp, but it should be locked. */
1467 	/* INP_LOCK_ASSERT(inp); */
1468 
1469 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1470 
1471 	if (tls->snd_tag == NULL) {
1472 		/*
1473 		 * Resetting send tag, ignore this change.  The
1474 		 * pending reset may or may not see this updated rate
1475 		 * in the tcpcb.  If it doesn't, we will just lose
1476 		 * this rate change.
1477 		 */
1478 		return (0);
1479 	}
1480 
1481 	mst = tls->snd_tag;
1482 
1483 	MPASS(mst != NULL);
1484 	MPASS(mst->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1485 
1486 	ifp = mst->ifp;
1487 	return (ifp->if_snd_tag_modify(mst, &params));
1488 }
1489 #endif
1490 #endif
1491 
1492 void
ktls_destroy(struct ktls_session * tls)1493 ktls_destroy(struct ktls_session *tls)
1494 {
1495 	struct rm_priotracker prio;
1496 
1497 	if (tls->sequential_records) {
1498 		struct mbuf *m, *n;
1499 		int page_count;
1500 
1501 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1502 			page_count = m->m_epg_enc_cnt;
1503 			while (page_count > 0) {
1504 				KASSERT(page_count >= m->m_epg_nrdy,
1505 				    ("%s: too few pages", __func__));
1506 				page_count -= m->m_epg_nrdy;
1507 				m = m_free(m);
1508 			}
1509 		}
1510 	}
1511 	ktls_cleanup(tls);
1512 	if (tls->be != NULL && ktls_allow_unload) {
1513 		rm_rlock(&ktls_backends_lock, &prio);
1514 		tls->be->use_count--;
1515 		rm_runlock(&ktls_backends_lock, &prio);
1516 	}
1517 	uma_zfree(ktls_session_zone, tls);
1518 }
1519 
1520 void
ktls_seq(struct sockbuf * sb,struct mbuf * m)1521 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1522 {
1523 
1524 	for (; m != NULL; m = m->m_next) {
1525 		KASSERT((m->m_flags & M_EXTPG) != 0,
1526 		    ("ktls_seq: mapped mbuf %p", m));
1527 
1528 		m->m_epg_seqno = sb->sb_tls_seqno;
1529 		sb->sb_tls_seqno++;
1530 	}
1531 }
1532 
1533 /*
1534  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1535  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1536  * mbuf must be populated with the payload of each TLS record.
1537  *
1538  * The record_type argument specifies the TLS record type used when
1539  * populating the TLS header.
1540  *
1541  * The enq_count argument on return is set to the number of pages of
1542  * payload data for this entire chain that need to be encrypted via SW
1543  * encryption.  The returned value should be passed to ktls_enqueue
1544  * when scheduling encryption of this chain of mbufs.  To handle the
1545  * special case of empty fragments for TLS 1.0 sessions, an empty
1546  * fragment counts as one page.
1547  */
1548 void
ktls_frame(struct mbuf * top,struct ktls_session * tls,int * enq_cnt,uint8_t record_type)1549 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1550     uint8_t record_type)
1551 {
1552 	struct tls_record_layer *tlshdr;
1553 	struct mbuf *m;
1554 	uint64_t *noncep;
1555 	uint16_t tls_len;
1556 	int maxlen;
1557 
1558 	maxlen = tls->params.max_frame_len;
1559 	*enq_cnt = 0;
1560 	for (m = top; m != NULL; m = m->m_next) {
1561 		/*
1562 		 * All mbufs in the chain should be TLS records whose
1563 		 * payload does not exceed the maximum frame length.
1564 		 *
1565 		 * Empty TLS 1.0 records are permitted when using CBC.
1566 		 */
1567 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1568 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1569 		    ("ktls_frame: m %p len %d", m, m->m_len));
1570 
1571 		/*
1572 		 * TLS frames require unmapped mbufs to store session
1573 		 * info.
1574 		 */
1575 		KASSERT((m->m_flags & M_EXTPG) != 0,
1576 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1577 
1578 		tls_len = m->m_len;
1579 
1580 		/* Save a reference to the session. */
1581 		m->m_epg_tls = ktls_hold(tls);
1582 
1583 		m->m_epg_hdrlen = tls->params.tls_hlen;
1584 		m->m_epg_trllen = tls->params.tls_tlen;
1585 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1586 			int bs, delta;
1587 
1588 			/*
1589 			 * AES-CBC pads messages to a multiple of the
1590 			 * block size.  Note that the padding is
1591 			 * applied after the digest and the encryption
1592 			 * is done on the "plaintext || mac || padding".
1593 			 * At least one byte of padding is always
1594 			 * present.
1595 			 *
1596 			 * Compute the final trailer length assuming
1597 			 * at most one block of padding.
1598 			 * tls->params.sb_tls_tlen is the maximum
1599 			 * possible trailer length (padding + digest).
1600 			 * delta holds the number of excess padding
1601 			 * bytes if the maximum were used.  Those
1602 			 * extra bytes are removed.
1603 			 */
1604 			bs = tls->params.tls_bs;
1605 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1606 			m->m_epg_trllen -= delta;
1607 		}
1608 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1609 
1610 		/* Populate the TLS header. */
1611 		tlshdr = (void *)m->m_epg_hdr;
1612 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1613 
1614 		/*
1615 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1616 		 * of TLS_RLTYPE_APP.
1617 		 */
1618 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1619 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1620 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1621 			tlshdr->tls_type = TLS_RLTYPE_APP;
1622 			/* save the real record type for later */
1623 			m->m_epg_record_type = record_type;
1624 			m->m_epg_trail[0] = record_type;
1625 		} else {
1626 			tlshdr->tls_vminor = tls->params.tls_vminor;
1627 			tlshdr->tls_type = record_type;
1628 		}
1629 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1630 
1631 		/*
1632 		 * Store nonces / explicit IVs after the end of the
1633 		 * TLS header.
1634 		 *
1635 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1636 		 * from the end of the IV.  The nonce is then
1637 		 * incremented for use by the next record.
1638 		 *
1639 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1640 		 */
1641 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1642 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1643 			noncep = (uint64_t *)(tls->params.iv + 8);
1644 			be64enc(tlshdr + 1, *noncep);
1645 			(*noncep)++;
1646 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1647 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1648 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1649 
1650 		/*
1651 		 * When using SW encryption, mark the mbuf not ready.
1652 		 * It will be marked ready via sbready() after the
1653 		 * record has been encrypted.
1654 		 *
1655 		 * When using ifnet TLS, unencrypted TLS records are
1656 		 * sent down the stack to the NIC.
1657 		 */
1658 		if (tls->mode == TCP_TLS_MODE_SW) {
1659 			m->m_flags |= M_NOTREADY;
1660 			if (__predict_false(tls_len == 0)) {
1661 				/* TLS 1.0 empty fragment. */
1662 				m->m_epg_nrdy = 1;
1663 			} else
1664 				m->m_epg_nrdy = m->m_epg_npgs;
1665 			*enq_cnt += m->m_epg_nrdy;
1666 		}
1667 	}
1668 }
1669 
1670 bool
ktls_permit_empty_frames(struct ktls_session * tls)1671 ktls_permit_empty_frames(struct ktls_session *tls)
1672 {
1673 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1674 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
1675 }
1676 
1677 void
ktls_check_rx(struct sockbuf * sb)1678 ktls_check_rx(struct sockbuf *sb)
1679 {
1680 	struct tls_record_layer hdr;
1681 	struct ktls_wq *wq;
1682 	struct socket *so;
1683 	bool running;
1684 
1685 	SOCKBUF_LOCK_ASSERT(sb);
1686 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1687 	    __func__, sb));
1688 	so = __containerof(sb, struct socket, so_rcv);
1689 
1690 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1691 		return;
1692 
1693 	/* Is there enough queued for a TLS header? */
1694 	if (sb->sb_tlscc < sizeof(hdr)) {
1695 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1696 			so->so_error = EMSGSIZE;
1697 		return;
1698 	}
1699 
1700 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1701 
1702 	/* Is the entire record queued? */
1703 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1704 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1705 			so->so_error = EMSGSIZE;
1706 		return;
1707 	}
1708 
1709 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1710 
1711 	soref(so);
1712 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1713 	mtx_lock(&wq->mtx);
1714 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1715 	running = wq->running;
1716 	mtx_unlock(&wq->mtx);
1717 	if (!running)
1718 		wakeup(wq);
1719 	counter_u64_add(ktls_cnt_rx_queued, 1);
1720 }
1721 
1722 static struct mbuf *
ktls_detach_record(struct sockbuf * sb,int len)1723 ktls_detach_record(struct sockbuf *sb, int len)
1724 {
1725 	struct mbuf *m, *n, *top;
1726 	int remain;
1727 
1728 	SOCKBUF_LOCK_ASSERT(sb);
1729 	MPASS(len <= sb->sb_tlscc);
1730 
1731 	/*
1732 	 * If TLS chain is the exact size of the record,
1733 	 * just grab the whole record.
1734 	 */
1735 	top = sb->sb_mtls;
1736 	if (sb->sb_tlscc == len) {
1737 		sb->sb_mtls = NULL;
1738 		sb->sb_mtlstail = NULL;
1739 		goto out;
1740 	}
1741 
1742 	/*
1743 	 * While it would be nice to use m_split() here, we need
1744 	 * to know exactly what m_split() allocates to update the
1745 	 * accounting, so do it inline instead.
1746 	 */
1747 	remain = len;
1748 	for (m = top; remain > m->m_len; m = m->m_next)
1749 		remain -= m->m_len;
1750 
1751 	/* Easy case: don't have to split 'm'. */
1752 	if (remain == m->m_len) {
1753 		sb->sb_mtls = m->m_next;
1754 		if (sb->sb_mtls == NULL)
1755 			sb->sb_mtlstail = NULL;
1756 		m->m_next = NULL;
1757 		goto out;
1758 	}
1759 
1760 	/*
1761 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1762 	 * with M_NOWAIT first.
1763 	 */
1764 	n = m_get(M_NOWAIT, MT_DATA);
1765 	if (n == NULL) {
1766 		/*
1767 		 * Use M_WAITOK with socket buffer unlocked.  If
1768 		 * 'sb_mtls' changes while the lock is dropped, return
1769 		 * NULL to force the caller to retry.
1770 		 */
1771 		SOCKBUF_UNLOCK(sb);
1772 
1773 		n = m_get(M_WAITOK, MT_DATA);
1774 
1775 		SOCKBUF_LOCK(sb);
1776 		if (sb->sb_mtls != top) {
1777 			m_free(n);
1778 			return (NULL);
1779 		}
1780 	}
1781 	n->m_flags |= M_NOTREADY;
1782 
1783 	/* Store remainder in 'n'. */
1784 	n->m_len = m->m_len - remain;
1785 	if (m->m_flags & M_EXT) {
1786 		n->m_data = m->m_data + remain;
1787 		mb_dupcl(n, m);
1788 	} else {
1789 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1790 	}
1791 
1792 	/* Trim 'm' and update accounting. */
1793 	m->m_len -= n->m_len;
1794 	sb->sb_tlscc -= n->m_len;
1795 	sb->sb_ccc -= n->m_len;
1796 
1797 	/* Account for 'n'. */
1798 	sballoc_ktls_rx(sb, n);
1799 
1800 	/* Insert 'n' into the TLS chain. */
1801 	sb->sb_mtls = n;
1802 	n->m_next = m->m_next;
1803 	if (sb->sb_mtlstail == m)
1804 		sb->sb_mtlstail = n;
1805 
1806 	/* Detach the record from the TLS chain. */
1807 	m->m_next = NULL;
1808 
1809 out:
1810 	MPASS(m_length(top, NULL) == len);
1811 	for (m = top; m != NULL; m = m->m_next)
1812 		sbfree_ktls_rx(sb, m);
1813 	sb->sb_tlsdcc = len;
1814 	sb->sb_ccc += len;
1815 	SBCHECK(sb);
1816 	return (top);
1817 }
1818 
1819 /*
1820  * Determine the length of the trailing zero padding and find the real
1821  * record type in the byte before the padding.
1822  *
1823  * Walking the mbuf chain backwards is clumsy, so another option would
1824  * be to scan forwards remembering the last non-zero byte before the
1825  * trailer.  However, it would be expensive to scan the entire record.
1826  * Instead, find the last non-zero byte of each mbuf in the chain
1827  * keeping track of the relative offset of that nonzero byte.
1828  *
1829  * trail_len is the size of the MAC/tag on input and is set to the
1830  * size of the full trailer including padding and the record type on
1831  * return.
1832  */
1833 static int
tls13_find_record_type(struct ktls_session * tls,struct mbuf * m,int tls_len,int * trailer_len,uint8_t * record_typep)1834 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
1835     int *trailer_len, uint8_t *record_typep)
1836 {
1837 	char *cp;
1838 	u_int digest_start, last_offset, m_len, offset;
1839 	uint8_t record_type;
1840 
1841 	digest_start = tls_len - *trailer_len;
1842 	last_offset = 0;
1843 	offset = 0;
1844 	for (; m != NULL && offset < digest_start;
1845 	     offset += m->m_len, m = m->m_next) {
1846 		/* Don't look for padding in the tag. */
1847 		m_len = min(digest_start - offset, m->m_len);
1848 		cp = mtod(m, char *);
1849 
1850 		/* Find last non-zero byte in this mbuf. */
1851 		while (m_len > 0 && cp[m_len - 1] == 0)
1852 			m_len--;
1853 		if (m_len > 0) {
1854 			record_type = cp[m_len - 1];
1855 			last_offset = offset + m_len;
1856 		}
1857 	}
1858 	if (last_offset < tls->params.tls_hlen)
1859 		return (EBADMSG);
1860 
1861 	*record_typep = record_type;
1862 	*trailer_len = tls_len - last_offset + 1;
1863 	return (0);
1864 }
1865 
1866 static void
ktls_drop(struct socket * so,int error)1867 ktls_drop(struct socket *so, int error)
1868 {
1869 	struct epoch_tracker et;
1870 	struct inpcb *inp = sotoinpcb(so);
1871 	struct tcpcb *tp;
1872 
1873 	NET_EPOCH_ENTER(et);
1874 	INP_WLOCK(inp);
1875 	if (!(inp->inp_flags & INP_DROPPED)) {
1876 		tp = intotcpcb(inp);
1877 		CURVNET_SET(inp->inp_vnet);
1878 		tp = tcp_drop(tp, error);
1879 		CURVNET_RESTORE();
1880 		if (tp != NULL)
1881 			INP_WUNLOCK(inp);
1882 	} else {
1883 		so->so_error = error;
1884 		SOCKBUF_LOCK(&so->so_rcv);
1885 		sorwakeup_locked(so);
1886 		INP_WUNLOCK(inp);
1887 	}
1888 	NET_EPOCH_EXIT(et);
1889 }
1890 
1891 static void
ktls_decrypt(struct socket * so)1892 ktls_decrypt(struct socket *so)
1893 {
1894 	char tls_header[MBUF_PEXT_HDR_LEN];
1895 	struct ktls_session *tls;
1896 	struct sockbuf *sb;
1897 	struct tls_record_layer *hdr;
1898 	struct tls_get_record tgr;
1899 	struct mbuf *control, *data, *m;
1900 	uint64_t seqno;
1901 	int error, remain, tls_len, trail_len;
1902 	bool tls13;
1903 	uint8_t vminor, record_type;
1904 
1905 	hdr = (struct tls_record_layer *)tls_header;
1906 	sb = &so->so_rcv;
1907 	SOCKBUF_LOCK(sb);
1908 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1909 	    ("%s: socket %p not running", __func__, so));
1910 
1911 	tls = sb->sb_tls_info;
1912 	MPASS(tls != NULL);
1913 
1914 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
1915 	if (tls13)
1916 		vminor = TLS_MINOR_VER_TWO;
1917 	else
1918 		vminor = tls->params.tls_vminor;
1919 	for (;;) {
1920 		/* Is there enough queued for a TLS header? */
1921 		if (sb->sb_tlscc < tls->params.tls_hlen)
1922 			break;
1923 
1924 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1925 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1926 
1927 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1928 		    hdr->tls_vminor != vminor)
1929 			error = EINVAL;
1930 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
1931 			error = EINVAL;
1932 		else if (tls_len < tls->params.tls_hlen || tls_len >
1933 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1934 		    tls->params.tls_tlen)
1935 			error = EMSGSIZE;
1936 		else
1937 			error = 0;
1938 		if (__predict_false(error != 0)) {
1939 			/*
1940 			 * We have a corrupted record and are likely
1941 			 * out of sync.  The connection isn't
1942 			 * recoverable at this point, so abort it.
1943 			 */
1944 			SOCKBUF_UNLOCK(sb);
1945 			counter_u64_add(ktls_offload_corrupted_records, 1);
1946 
1947 			ktls_drop(so, error);
1948 			goto deref;
1949 		}
1950 
1951 		/* Is the entire record queued? */
1952 		if (sb->sb_tlscc < tls_len)
1953 			break;
1954 
1955 		/*
1956 		 * Split out the portion of the mbuf chain containing
1957 		 * this TLS record.
1958 		 */
1959 		data = ktls_detach_record(sb, tls_len);
1960 		if (data == NULL)
1961 			continue;
1962 		MPASS(sb->sb_tlsdcc == tls_len);
1963 
1964 		seqno = sb->sb_tls_seqno;
1965 		sb->sb_tls_seqno++;
1966 		SBCHECK(sb);
1967 		SOCKBUF_UNLOCK(sb);
1968 
1969 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1970 		if (error == 0) {
1971 			if (tls13)
1972 				error = tls13_find_record_type(tls, data,
1973 				    tls_len, &trail_len, &record_type);
1974 			else
1975 				record_type = hdr->tls_type;
1976 		}
1977 		if (error) {
1978 			counter_u64_add(ktls_offload_failed_crypto, 1);
1979 
1980 			SOCKBUF_LOCK(sb);
1981 			if (sb->sb_tlsdcc == 0) {
1982 				/*
1983 				 * sbcut/drop/flush discarded these
1984 				 * mbufs.
1985 				 */
1986 				m_freem(data);
1987 				break;
1988 			}
1989 
1990 			/*
1991 			 * Drop this TLS record's data, but keep
1992 			 * decrypting subsequent records.
1993 			 */
1994 			sb->sb_ccc -= tls_len;
1995 			sb->sb_tlsdcc = 0;
1996 
1997 			CURVNET_SET(so->so_vnet);
1998 			so->so_error = EBADMSG;
1999 			sorwakeup_locked(so);
2000 			CURVNET_RESTORE();
2001 
2002 			m_freem(data);
2003 
2004 			SOCKBUF_LOCK(sb);
2005 			continue;
2006 		}
2007 
2008 		/* Allocate the control mbuf. */
2009 		memset(&tgr, 0, sizeof(tgr));
2010 		tgr.tls_type = record_type;
2011 		tgr.tls_vmajor = hdr->tls_vmajor;
2012 		tgr.tls_vminor = hdr->tls_vminor;
2013 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2014 		    trail_len);
2015 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
2016 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2017 
2018 		SOCKBUF_LOCK(sb);
2019 		if (sb->sb_tlsdcc == 0) {
2020 			/* sbcut/drop/flush discarded these mbufs. */
2021 			MPASS(sb->sb_tlscc == 0);
2022 			m_freem(data);
2023 			m_freem(control);
2024 			break;
2025 		}
2026 
2027 		/*
2028 		 * Clear the 'dcc' accounting in preparation for
2029 		 * adding the decrypted record.
2030 		 */
2031 		sb->sb_ccc -= tls_len;
2032 		sb->sb_tlsdcc = 0;
2033 		SBCHECK(sb);
2034 
2035 		/* If there is no payload, drop all of the data. */
2036 		if (tgr.tls_length == htobe16(0)) {
2037 			m_freem(data);
2038 			data = NULL;
2039 		} else {
2040 			/* Trim header. */
2041 			remain = tls->params.tls_hlen;
2042 			while (remain > 0) {
2043 				if (data->m_len > remain) {
2044 					data->m_data += remain;
2045 					data->m_len -= remain;
2046 					break;
2047 				}
2048 				remain -= data->m_len;
2049 				data = m_free(data);
2050 			}
2051 
2052 			/* Trim trailer and clear M_NOTREADY. */
2053 			remain = be16toh(tgr.tls_length);
2054 			m = data;
2055 			for (m = data; remain > m->m_len; m = m->m_next) {
2056 				m->m_flags &= ~M_NOTREADY;
2057 				remain -= m->m_len;
2058 			}
2059 			m->m_len = remain;
2060 			m_freem(m->m_next);
2061 			m->m_next = NULL;
2062 			m->m_flags &= ~M_NOTREADY;
2063 
2064 			/* Set EOR on the final mbuf. */
2065 			m->m_flags |= M_EOR;
2066 		}
2067 
2068 		sbappendcontrol_locked(sb, data, control, 0);
2069 	}
2070 
2071 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2072 
2073 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2074 		so->so_error = EMSGSIZE;
2075 
2076 	sorwakeup_locked(so);
2077 
2078 deref:
2079 	SOCKBUF_UNLOCK_ASSERT(sb);
2080 
2081 	CURVNET_SET(so->so_vnet);
2082 	SOCK_LOCK(so);
2083 	sorele(so);
2084 	CURVNET_RESTORE();
2085 }
2086 
2087 void
ktls_enqueue_to_free(struct mbuf * m)2088 ktls_enqueue_to_free(struct mbuf *m)
2089 {
2090 	struct ktls_wq *wq;
2091 	bool running;
2092 
2093 	/* Mark it for freeing. */
2094 	m->m_epg_flags |= EPG_FLAG_2FREE;
2095 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2096 	mtx_lock(&wq->mtx);
2097 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2098 	running = wq->running;
2099 	mtx_unlock(&wq->mtx);
2100 	if (!running)
2101 		wakeup(wq);
2102 }
2103 
2104 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2105 static u_int
ktls_batched_records(struct mbuf * m)2106 ktls_batched_records(struct mbuf *m)
2107 {
2108 	int page_count, records;
2109 
2110 	records = 0;
2111 	page_count = m->m_epg_enc_cnt;
2112 	while (page_count > 0) {
2113 		records++;
2114 		page_count -= m->m_epg_nrdy;
2115 		m = m->m_next;
2116 	}
2117 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2118 	return (records);
2119 }
2120 
2121 void
ktls_enqueue(struct mbuf * m,struct socket * so,int page_count)2122 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2123 {
2124 	struct ktls_session *tls;
2125 	struct ktls_wq *wq;
2126 	int queued;
2127 	bool running;
2128 
2129 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2130 	    (M_EXTPG | M_NOTREADY)),
2131 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2132 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2133 
2134 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2135 
2136 	m->m_epg_enc_cnt = page_count;
2137 
2138 	/*
2139 	 * Save a pointer to the socket.  The caller is responsible
2140 	 * for taking an additional reference via soref().
2141 	 */
2142 	m->m_epg_so = so;
2143 
2144 	queued = 1;
2145 	tls = m->m_epg_tls;
2146 	wq = &ktls_wq[tls->wq_index];
2147 	mtx_lock(&wq->mtx);
2148 	if (__predict_false(tls->sequential_records)) {
2149 		/*
2150 		 * For TLS 1.0, records must be encrypted
2151 		 * sequentially.  For a given connection, all records
2152 		 * queued to the associated work queue are processed
2153 		 * sequentially.  However, sendfile(2) might complete
2154 		 * I/O requests spanning multiple TLS records out of
2155 		 * order.  Here we ensure TLS records are enqueued to
2156 		 * the work queue in FIFO order.
2157 		 *
2158 		 * tls->next_seqno holds the sequence number of the
2159 		 * next TLS record that should be enqueued to the work
2160 		 * queue.  If this next record is not tls->next_seqno,
2161 		 * it must be a future record, so insert it, sorted by
2162 		 * TLS sequence number, into tls->pending_records and
2163 		 * return.
2164 		 *
2165 		 * If this TLS record matches tls->next_seqno, place
2166 		 * it in the work queue and then check
2167 		 * tls->pending_records to see if any
2168 		 * previously-queued records are now ready for
2169 		 * encryption.
2170 		 */
2171 		if (m->m_epg_seqno != tls->next_seqno) {
2172 			struct mbuf *n, *p;
2173 
2174 			p = NULL;
2175 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2176 				if (n->m_epg_seqno > m->m_epg_seqno)
2177 					break;
2178 				p = n;
2179 			}
2180 			if (n == NULL)
2181 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2182 				    m_epg_stailq);
2183 			else if (p == NULL)
2184 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2185 				    m_epg_stailq);
2186 			else
2187 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2188 				    m_epg_stailq);
2189 			mtx_unlock(&wq->mtx);
2190 			counter_u64_add(ktls_cnt_tx_pending, 1);
2191 			return;
2192 		}
2193 
2194 		tls->next_seqno += ktls_batched_records(m);
2195 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2196 
2197 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2198 			struct mbuf *n;
2199 
2200 			n = STAILQ_FIRST(&tls->pending_records);
2201 			if (n->m_epg_seqno != tls->next_seqno)
2202 				break;
2203 
2204 			queued++;
2205 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2206 			tls->next_seqno += ktls_batched_records(n);
2207 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2208 		}
2209 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2210 	} else
2211 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2212 
2213 	running = wq->running;
2214 	mtx_unlock(&wq->mtx);
2215 	if (!running)
2216 		wakeup(wq);
2217 	counter_u64_add(ktls_cnt_tx_queued, queued);
2218 }
2219 
2220 static __noinline void
ktls_encrypt(struct mbuf * top)2221 ktls_encrypt(struct mbuf *top)
2222 {
2223 	struct ktls_session *tls;
2224 	struct socket *so;
2225 	struct mbuf *m;
2226 	vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
2227 	struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
2228 	struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
2229 	vm_page_t pg;
2230 	int error, i, len, npages, off, total_pages;
2231 	bool is_anon;
2232 
2233 	so = top->m_epg_so;
2234 	tls = top->m_epg_tls;
2235 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2236 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2237 #ifdef INVARIANTS
2238 	top->m_epg_so = NULL;
2239 #endif
2240 	total_pages = top->m_epg_enc_cnt;
2241 	npages = 0;
2242 
2243 	/*
2244 	 * Encrypt the TLS records in the chain of mbufs starting with
2245 	 * 'top'.  'total_pages' gives us a total count of pages and is
2246 	 * used to know when we have finished encrypting the TLS
2247 	 * records originally queued with 'top'.
2248 	 *
2249 	 * NB: These mbufs are queued in the socket buffer and
2250 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2251 	 * socket buffer lock is not held while traversing this chain.
2252 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2253 	 * pointers should be stable.  However, the 'm_next' of the
2254 	 * last mbuf encrypted is not necessarily NULL.  It can point
2255 	 * to other mbufs appended while 'top' was on the TLS work
2256 	 * queue.
2257 	 *
2258 	 * Each mbuf holds an entire TLS record.
2259 	 */
2260 	error = 0;
2261 	for (m = top; npages != total_pages; m = m->m_next) {
2262 		KASSERT(m->m_epg_tls == tls,
2263 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2264 		    tls, m->m_epg_tls));
2265 		KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2266 		    (M_EXTPG | M_NOTREADY),
2267 		    ("%p not unready & nomap mbuf (top = %p)\n", m, top));
2268 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2269 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2270 		    total_pages, m));
2271 
2272 		/*
2273 		 * Generate source and destination ivoecs to pass to
2274 		 * the SW encryption backend.  For writable mbufs, the
2275 		 * destination iovec is a copy of the source and
2276 		 * encryption is done in place.  For file-backed mbufs
2277 		 * (from sendfile), anonymous wired pages are
2278 		 * allocated and assigned to the destination iovec.
2279 		 */
2280 		is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
2281 
2282 		off = m->m_epg_1st_off;
2283 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2284 			len = m_epg_pagelen(m, i, off);
2285 			src_iov[i].iov_len = len;
2286 			src_iov[i].iov_base =
2287 			    (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
2288 				off;
2289 
2290 			if (is_anon) {
2291 				dst_iov[i].iov_base = src_iov[i].iov_base;
2292 				dst_iov[i].iov_len = src_iov[i].iov_len;
2293 				continue;
2294 			}
2295 retry_page:
2296 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2297 			    VM_ALLOC_WIRED);
2298 			if (pg == NULL) {
2299 				vm_wait(NULL);
2300 				goto retry_page;
2301 			}
2302 			parray[i] = VM_PAGE_TO_PHYS(pg);
2303 			dst_iov[i].iov_base =
2304 			    (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
2305 			dst_iov[i].iov_len = len;
2306 		}
2307 
2308 		npages += m->m_epg_nrdy;
2309 
2310 		error = (*tls->sw_encrypt)(tls,
2311 		    (const struct tls_record_layer *)m->m_epg_hdr,
2312 		    m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
2313 		    m->m_epg_record_type);
2314 		if (error) {
2315 			counter_u64_add(ktls_offload_failed_crypto, 1);
2316 			break;
2317 		}
2318 
2319 		/*
2320 		 * For file-backed mbufs, release the file-backed
2321 		 * pages and replace them in the ext_pgs array with
2322 		 * the anonymous wired pages allocated above.
2323 		 */
2324 		if (!is_anon) {
2325 			/* Free the old pages. */
2326 			m->m_ext.ext_free(m);
2327 
2328 			/* Replace them with the new pages. */
2329 			for (i = 0; i < m->m_epg_npgs; i++)
2330 				m->m_epg_pa[i] = parray[i];
2331 
2332 			/* Use the basic free routine. */
2333 			m->m_ext.ext_free = mb_free_mext_pgs;
2334 
2335 			/* Pages are now writable. */
2336 			m->m_epg_flags |= EPG_FLAG_ANON;
2337 		}
2338 
2339 		/*
2340 		 * Drop a reference to the session now that it is no
2341 		 * longer needed.  Existing code depends on encrypted
2342 		 * records having no associated session vs
2343 		 * yet-to-be-encrypted records having an associated
2344 		 * session.
2345 		 */
2346 		m->m_epg_tls = NULL;
2347 		ktls_free(tls);
2348 	}
2349 
2350 	CURVNET_SET(so->so_vnet);
2351 	if (error == 0) {
2352 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2353 	} else {
2354 		ktls_drop(so, EIO);
2355 		mb_free_notready(top, total_pages);
2356 	}
2357 
2358 	SOCK_LOCK(so);
2359 	sorele(so);
2360 	CURVNET_RESTORE();
2361 }
2362 
2363 static void
ktls_work_thread(void * ctx)2364 ktls_work_thread(void *ctx)
2365 {
2366 	struct ktls_wq *wq = ctx;
2367 	struct mbuf *m, *n;
2368 	struct socket *so, *son;
2369 	STAILQ_HEAD(, mbuf) local_m_head;
2370 	STAILQ_HEAD(, socket) local_so_head;
2371 
2372 	if (ktls_bind_threads > 1) {
2373 		curthread->td_domain.dr_policy =
2374 			DOMAINSET_PREF(PCPU_GET(domain));
2375 	}
2376 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2377 	fpu_kern_thread(0);
2378 #endif
2379 	for (;;) {
2380 		mtx_lock(&wq->mtx);
2381 		while (STAILQ_EMPTY(&wq->m_head) &&
2382 		    STAILQ_EMPTY(&wq->so_head)) {
2383 			wq->running = false;
2384 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2385 			wq->running = true;
2386 		}
2387 
2388 		STAILQ_INIT(&local_m_head);
2389 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2390 		STAILQ_INIT(&local_so_head);
2391 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2392 		mtx_unlock(&wq->mtx);
2393 
2394 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2395 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2396 				ktls_free(m->m_epg_tls);
2397 				m_free_raw(m);
2398 			} else {
2399 				ktls_encrypt(m);
2400 				counter_u64_add(ktls_cnt_tx_queued, -1);
2401 			}
2402 		}
2403 
2404 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2405 			ktls_decrypt(so);
2406 			counter_u64_add(ktls_cnt_rx_queued, -1);
2407 		}
2408 	}
2409 }
2410