xref: /freebsd/sys/kern/uipc_ktls.c (revision 9768746b)
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_kern_tls.h"
34 #include "opt_ratelimit.h"
35 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/domainset.h>
40 #include <sys/endian.h>
41 #include <sys/ktls.h>
42 #include <sys/lock.h>
43 #include <sys/mbuf.h>
44 #include <sys/mutex.h>
45 #include <sys/rmlock.h>
46 #include <sys/proc.h>
47 #include <sys/protosw.h>
48 #include <sys/refcount.h>
49 #include <sys/smp.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/kthread.h>
55 #include <sys/uio.h>
56 #include <sys/vmmeter.h>
57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58 #include <machine/pcb.h>
59 #endif
60 #include <machine/vmparam.h>
61 #include <net/if.h>
62 #include <net/if_var.h>
63 #ifdef RSS
64 #include <net/netisr.h>
65 #include <net/rss_config.h>
66 #endif
67 #include <net/route.h>
68 #include <net/route/nhop.h>
69 #if defined(INET) || defined(INET6)
70 #include <netinet/in.h>
71 #include <netinet/in_pcb.h>
72 #endif
73 #include <netinet/tcp_var.h>
74 #ifdef TCP_OFFLOAD
75 #include <netinet/tcp_offload.h>
76 #endif
77 #include <opencrypto/cryptodev.h>
78 #include <opencrypto/ktls.h>
79 #include <vm/uma_dbg.h>
80 #include <vm/vm.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pagequeue.h>
84 
85 struct ktls_wq {
86 	struct mtx	mtx;
87 	STAILQ_HEAD(, mbuf) m_head;
88 	STAILQ_HEAD(, socket) so_head;
89 	bool		running;
90 	int		lastallocfail;
91 } __aligned(CACHE_LINE_SIZE);
92 
93 struct ktls_alloc_thread {
94 	uint64_t wakeups;
95 	uint64_t allocs;
96 	struct thread *td;
97 	int running;
98 };
99 
100 struct ktls_domain_info {
101 	int count;
102 	int cpu[MAXCPU];
103 	struct ktls_alloc_thread alloc_td;
104 };
105 
106 struct ktls_domain_info ktls_domains[MAXMEMDOM];
107 static struct ktls_wq *ktls_wq;
108 static struct proc *ktls_proc;
109 static uma_zone_t ktls_session_zone;
110 static uma_zone_t ktls_buffer_zone;
111 static uint16_t ktls_cpuid_lookup[MAXCPU];
112 static int ktls_init_state;
113 static struct sx ktls_init_lock;
114 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
115 
116 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117     "Kernel TLS offload");
118 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
119     "Kernel TLS offload stats");
120 
121 #ifdef RSS
122 static int ktls_bind_threads = 1;
123 #else
124 static int ktls_bind_threads;
125 #endif
126 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
127     &ktls_bind_threads, 0,
128     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
129 
130 static u_int ktls_maxlen = 16384;
131 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
132     &ktls_maxlen, 0, "Maximum TLS record size");
133 
134 static int ktls_number_threads;
135 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
136     &ktls_number_threads, 0,
137     "Number of TLS threads in thread-pool");
138 
139 unsigned int ktls_ifnet_max_rexmit_pct = 2;
140 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
141     &ktls_ifnet_max_rexmit_pct, 2,
142     "Max percent bytes retransmitted before ifnet TLS is disabled");
143 
144 static bool ktls_offload_enable;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
146     &ktls_offload_enable, 0,
147     "Enable support for kernel TLS offload");
148 
149 static bool ktls_cbc_enable = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
151     &ktls_cbc_enable, 1,
152     "Enable support of AES-CBC crypto for kernel TLS");
153 
154 static bool ktls_sw_buffer_cache = true;
155 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
156     &ktls_sw_buffer_cache, 1,
157     "Enable caching of output buffers for SW encryption");
158 
159 static int ktls_max_alloc = 128;
160 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
161     &ktls_max_alloc, 128,
162     "Max number of 16k buffers to allocate in thread context");
163 
164 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
165 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
166     &ktls_tasks_active, "Number of active tasks");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
170     &ktls_cnt_tx_pending,
171     "Number of TLS 1.0 records waiting for earlier TLS records");
172 
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
175     &ktls_cnt_tx_queued,
176     "Number of TLS records in queue to tasks for SW encryption");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
180     &ktls_cnt_rx_queued,
181     "Number of TLS sockets in queue to tasks for SW decryption");
182 
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
185     CTLFLAG_RD, &ktls_offload_total,
186     "Total successful TLS setups (parameters set)");
187 
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
190     CTLFLAG_RD, &ktls_offload_enable_calls,
191     "Total number of TLS enable calls made");
192 
193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
195     &ktls_offload_active, "Total Active TLS sessions");
196 
197 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
199     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
200 
201 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
203     &ktls_offload_failed_crypto, "Total TLS crypto failures");
204 
205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
207     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
208 
209 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
211     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
212 
213 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
215     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
216 
217 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
218 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
219     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
220 
221 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
222 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
223     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
224 
225 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
226 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
227     &ktls_destroy_task,
228     "Number of times ktls session was destroyed via taskqueue");
229 
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "Software TLS session stats");
232 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
233     "Hardware (ifnet) TLS session stats");
234 #ifdef TCP_OFFLOAD
235 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
236     "TOE TLS session stats");
237 #endif
238 
239 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
240 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
241     "Active number of software TLS sessions using AES-CBC");
242 
243 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
244 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
245     "Active number of software TLS sessions using AES-GCM");
246 
247 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
249     &ktls_sw_chacha20,
250     "Active number of software TLS sessions using Chacha20-Poly1305");
251 
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
254     &ktls_ifnet_cbc,
255     "Active number of ifnet TLS sessions using AES-CBC");
256 
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
259     &ktls_ifnet_gcm,
260     "Active number of ifnet TLS sessions using AES-GCM");
261 
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
264     &ktls_ifnet_chacha20,
265     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
266 
267 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
268 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
269     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
270 
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
273     &ktls_ifnet_reset_dropped,
274     "TLS sessions dropped after failing to update ifnet send tag");
275 
276 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
277 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
278     &ktls_ifnet_reset_failed,
279     "TLS sessions that failed to allocate a new ifnet send tag");
280 
281 static int ktls_ifnet_permitted;
282 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
283     &ktls_ifnet_permitted, 1,
284     "Whether to permit hardware (ifnet) TLS sessions");
285 
286 #ifdef TCP_OFFLOAD
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
289     &ktls_toe_cbc,
290     "Active number of TOE TLS sessions using AES-CBC");
291 
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
294     &ktls_toe_gcm,
295     "Active number of TOE TLS sessions using AES-GCM");
296 
297 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
298 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
299     &ktls_toe_chacha20,
300     "Active number of TOE TLS sessions using Chacha20-Poly1305");
301 #endif
302 
303 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
304 
305 #if defined(INET) || defined(INET6)
306 static void ktls_reset_receive_tag(void *context, int pending);
307 static void ktls_reset_send_tag(void *context, int pending);
308 #endif
309 static void ktls_work_thread(void *ctx);
310 static void ktls_alloc_thread(void *ctx);
311 
312 #if defined(INET) || defined(INET6)
313 static u_int
314 ktls_get_cpu(struct socket *so)
315 {
316 	struct inpcb *inp;
317 #ifdef NUMA
318 	struct ktls_domain_info *di;
319 #endif
320 	u_int cpuid;
321 
322 	inp = sotoinpcb(so);
323 #ifdef RSS
324 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
325 	if (cpuid != NETISR_CPUID_NONE)
326 		return (cpuid);
327 #endif
328 	/*
329 	 * Just use the flowid to shard connections in a repeatable
330 	 * fashion.  Note that TLS 1.0 sessions rely on the
331 	 * serialization provided by having the same connection use
332 	 * the same queue.
333 	 */
334 #ifdef NUMA
335 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
336 		di = &ktls_domains[inp->inp_numa_domain];
337 		cpuid = di->cpu[inp->inp_flowid % di->count];
338 	} else
339 #endif
340 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
341 	return (cpuid);
342 }
343 #endif
344 
345 static int
346 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
347 {
348 	vm_page_t m;
349 	int i, req;
350 
351 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
352 	    ("%s: ktls max length %d is not page size-aligned",
353 	    __func__, ktls_maxlen));
354 
355 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
356 	for (i = 0; i < count; i++) {
357 		m = vm_page_alloc_noobj_contig_domain(domain, req,
358 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
359 		    VM_MEMATTR_DEFAULT);
360 		if (m == NULL)
361 			break;
362 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
363 	}
364 	return (i);
365 }
366 
367 static void
368 ktls_buffer_release(void *arg __unused, void **store, int count)
369 {
370 	vm_page_t m;
371 	int i, j;
372 
373 	for (i = 0; i < count; i++) {
374 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
375 		for (j = 0; j < atop(ktls_maxlen); j++) {
376 			(void)vm_page_unwire_noq(m + j);
377 			vm_page_free(m + j);
378 		}
379 	}
380 }
381 
382 static void
383 ktls_free_mext_contig(struct mbuf *m)
384 {
385 	M_ASSERTEXTPG(m);
386 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
387 }
388 
389 static int
390 ktls_init(void)
391 {
392 	struct thread *td;
393 	struct pcpu *pc;
394 	int count, domain, error, i;
395 
396 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
397 	    M_WAITOK | M_ZERO);
398 
399 	ktls_session_zone = uma_zcreate("ktls_session",
400 	    sizeof(struct ktls_session),
401 	    NULL, NULL, NULL, NULL,
402 	    UMA_ALIGN_CACHE, 0);
403 
404 	if (ktls_sw_buffer_cache) {
405 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
406 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
407 		    ktls_buffer_import, ktls_buffer_release, NULL,
408 		    UMA_ZONE_FIRSTTOUCH);
409 	}
410 
411 	/*
412 	 * Initialize the workqueues to run the TLS work.  We create a
413 	 * work queue for each CPU.
414 	 */
415 	CPU_FOREACH(i) {
416 		STAILQ_INIT(&ktls_wq[i].m_head);
417 		STAILQ_INIT(&ktls_wq[i].so_head);
418 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
419 		if (ktls_bind_threads > 1) {
420 			pc = pcpu_find(i);
421 			domain = pc->pc_domain;
422 			count = ktls_domains[domain].count;
423 			ktls_domains[domain].cpu[count] = i;
424 			ktls_domains[domain].count++;
425 		}
426 		ktls_cpuid_lookup[ktls_number_threads] = i;
427 		ktls_number_threads++;
428 	}
429 
430 	/*
431 	 * If we somehow have an empty domain, fall back to choosing
432 	 * among all KTLS threads.
433 	 */
434 	if (ktls_bind_threads > 1) {
435 		for (i = 0; i < vm_ndomains; i++) {
436 			if (ktls_domains[i].count == 0) {
437 				ktls_bind_threads = 1;
438 				break;
439 			}
440 		}
441 	}
442 
443 	/* Start kthreads for each workqueue. */
444 	CPU_FOREACH(i) {
445 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
446 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
447 		if (error) {
448 			printf("Can't add KTLS thread %d error %d\n", i, error);
449 			return (error);
450 		}
451 	}
452 
453 	/*
454 	 * Start an allocation thread per-domain to perform blocking allocations
455 	 * of 16k physically contiguous TLS crypto destination buffers.
456 	 */
457 	if (ktls_sw_buffer_cache) {
458 		for (domain = 0; domain < vm_ndomains; domain++) {
459 			if (VM_DOMAIN_EMPTY(domain))
460 				continue;
461 			if (CPU_EMPTY(&cpuset_domain[domain]))
462 				continue;
463 			error = kproc_kthread_add(ktls_alloc_thread,
464 			    &ktls_domains[domain], &ktls_proc,
465 			    &ktls_domains[domain].alloc_td.td,
466 			    0, 0, "KTLS", "alloc_%d", domain);
467 			if (error) {
468 				printf("Can't add KTLS alloc thread %d error %d\n",
469 				    domain, error);
470 				return (error);
471 			}
472 		}
473 	}
474 
475 	if (bootverbose)
476 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
477 	return (0);
478 }
479 
480 static int
481 ktls_start_kthreads(void)
482 {
483 	int error, state;
484 
485 start:
486 	state = atomic_load_acq_int(&ktls_init_state);
487 	if (__predict_true(state > 0))
488 		return (0);
489 	if (state < 0)
490 		return (ENXIO);
491 
492 	sx_xlock(&ktls_init_lock);
493 	if (ktls_init_state != 0) {
494 		sx_xunlock(&ktls_init_lock);
495 		goto start;
496 	}
497 
498 	error = ktls_init();
499 	if (error == 0)
500 		state = 1;
501 	else
502 		state = -1;
503 	atomic_store_rel_int(&ktls_init_state, state);
504 	sx_xunlock(&ktls_init_lock);
505 	return (error);
506 }
507 
508 #if defined(INET) || defined(INET6)
509 static int
510 ktls_create_session(struct socket *so, struct tls_enable *en,
511     struct ktls_session **tlsp, int direction)
512 {
513 	struct ktls_session *tls;
514 	int error;
515 
516 	/* Only TLS 1.0 - 1.3 are supported. */
517 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
518 		return (EINVAL);
519 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
520 	    en->tls_vminor > TLS_MINOR_VER_THREE)
521 		return (EINVAL);
522 
523 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
524 		return (EINVAL);
525 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
526 		return (EINVAL);
527 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
528 		return (EINVAL);
529 
530 	/* All supported algorithms require a cipher key. */
531 	if (en->cipher_key_len == 0)
532 		return (EINVAL);
533 
534 	/* No flags are currently supported. */
535 	if (en->flags != 0)
536 		return (EINVAL);
537 
538 	/* Common checks for supported algorithms. */
539 	switch (en->cipher_algorithm) {
540 	case CRYPTO_AES_NIST_GCM_16:
541 		/*
542 		 * auth_algorithm isn't used, but permit GMAC values
543 		 * for compatibility.
544 		 */
545 		switch (en->auth_algorithm) {
546 		case 0:
547 #ifdef COMPAT_FREEBSD12
548 		/* XXX: Really 13.0-current COMPAT. */
549 		case CRYPTO_AES_128_NIST_GMAC:
550 		case CRYPTO_AES_192_NIST_GMAC:
551 		case CRYPTO_AES_256_NIST_GMAC:
552 #endif
553 			break;
554 		default:
555 			return (EINVAL);
556 		}
557 		if (en->auth_key_len != 0)
558 			return (EINVAL);
559 		switch (en->tls_vminor) {
560 		case TLS_MINOR_VER_TWO:
561 			if (en->iv_len != TLS_AEAD_GCM_LEN)
562 				return (EINVAL);
563 			break;
564 		case TLS_MINOR_VER_THREE:
565 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
566 				return (EINVAL);
567 			break;
568 		default:
569 			return (EINVAL);
570 		}
571 		break;
572 	case CRYPTO_AES_CBC:
573 		switch (en->auth_algorithm) {
574 		case CRYPTO_SHA1_HMAC:
575 			break;
576 		case CRYPTO_SHA2_256_HMAC:
577 		case CRYPTO_SHA2_384_HMAC:
578 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
579 				return (EINVAL);
580 			break;
581 		default:
582 			return (EINVAL);
583 		}
584 		if (en->auth_key_len == 0)
585 			return (EINVAL);
586 
587 		/*
588 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
589 		 * use explicit IVs.
590 		 */
591 		switch (en->tls_vminor) {
592 		case TLS_MINOR_VER_ZERO:
593 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
594 				return (EINVAL);
595 			break;
596 		case TLS_MINOR_VER_ONE:
597 		case TLS_MINOR_VER_TWO:
598 			/* Ignore any supplied IV. */
599 			en->iv_len = 0;
600 			break;
601 		default:
602 			return (EINVAL);
603 		}
604 		break;
605 	case CRYPTO_CHACHA20_POLY1305:
606 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
607 			return (EINVAL);
608 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
609 		    en->tls_vminor != TLS_MINOR_VER_THREE)
610 			return (EINVAL);
611 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
612 			return (EINVAL);
613 		break;
614 	default:
615 		return (EINVAL);
616 	}
617 
618 	error = ktls_start_kthreads();
619 	if (error != 0)
620 		return (error);
621 
622 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
623 
624 	counter_u64_add(ktls_offload_active, 1);
625 
626 	refcount_init(&tls->refcount, 1);
627 	if (direction == KTLS_RX) {
628 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
629 	} else {
630 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
631 		tls->inp = so->so_pcb;
632 		in_pcbref(tls->inp);
633 		tls->tx = true;
634 	}
635 
636 	tls->wq_index = ktls_get_cpu(so);
637 
638 	tls->params.cipher_algorithm = en->cipher_algorithm;
639 	tls->params.auth_algorithm = en->auth_algorithm;
640 	tls->params.tls_vmajor = en->tls_vmajor;
641 	tls->params.tls_vminor = en->tls_vminor;
642 	tls->params.flags = en->flags;
643 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
644 
645 	/* Set the header and trailer lengths. */
646 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
647 	switch (en->cipher_algorithm) {
648 	case CRYPTO_AES_NIST_GCM_16:
649 		/*
650 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
651 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
652 		 */
653 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
654 			tls->params.tls_hlen += sizeof(uint64_t);
655 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
656 		tls->params.tls_bs = 1;
657 		break;
658 	case CRYPTO_AES_CBC:
659 		switch (en->auth_algorithm) {
660 		case CRYPTO_SHA1_HMAC:
661 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
662 				/* Implicit IV, no nonce. */
663 				tls->sequential_records = true;
664 				tls->next_seqno = be64dec(en->rec_seq);
665 				STAILQ_INIT(&tls->pending_records);
666 			} else {
667 				tls->params.tls_hlen += AES_BLOCK_LEN;
668 			}
669 			tls->params.tls_tlen = AES_BLOCK_LEN +
670 			    SHA1_HASH_LEN;
671 			break;
672 		case CRYPTO_SHA2_256_HMAC:
673 			tls->params.tls_hlen += AES_BLOCK_LEN;
674 			tls->params.tls_tlen = AES_BLOCK_LEN +
675 			    SHA2_256_HASH_LEN;
676 			break;
677 		case CRYPTO_SHA2_384_HMAC:
678 			tls->params.tls_hlen += AES_BLOCK_LEN;
679 			tls->params.tls_tlen = AES_BLOCK_LEN +
680 			    SHA2_384_HASH_LEN;
681 			break;
682 		default:
683 			panic("invalid hmac");
684 		}
685 		tls->params.tls_bs = AES_BLOCK_LEN;
686 		break;
687 	case CRYPTO_CHACHA20_POLY1305:
688 		/*
689 		 * Chacha20 uses a 12 byte implicit IV.
690 		 */
691 		tls->params.tls_tlen = POLY1305_HASH_LEN;
692 		tls->params.tls_bs = 1;
693 		break;
694 	default:
695 		panic("invalid cipher");
696 	}
697 
698 	/*
699 	 * TLS 1.3 includes optional padding which we do not support,
700 	 * and also puts the "real" record type at the end of the
701 	 * encrypted data.
702 	 */
703 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
704 		tls->params.tls_tlen += sizeof(uint8_t);
705 
706 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
707 	    ("TLS header length too long: %d", tls->params.tls_hlen));
708 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
709 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
710 
711 	if (en->auth_key_len != 0) {
712 		tls->params.auth_key_len = en->auth_key_len;
713 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
714 		    M_WAITOK);
715 		error = copyin(en->auth_key, tls->params.auth_key,
716 		    en->auth_key_len);
717 		if (error)
718 			goto out;
719 	}
720 
721 	tls->params.cipher_key_len = en->cipher_key_len;
722 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
723 	error = copyin(en->cipher_key, tls->params.cipher_key,
724 	    en->cipher_key_len);
725 	if (error)
726 		goto out;
727 
728 	/*
729 	 * This holds the implicit portion of the nonce for AEAD
730 	 * ciphers and the initial implicit IV for TLS 1.0.  The
731 	 * explicit portions of the IV are generated in ktls_frame().
732 	 */
733 	if (en->iv_len != 0) {
734 		tls->params.iv_len = en->iv_len;
735 		error = copyin(en->iv, tls->params.iv, en->iv_len);
736 		if (error)
737 			goto out;
738 
739 		/*
740 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
741 		 * counter to generate unique explicit IVs.
742 		 *
743 		 * Store this counter in the last 8 bytes of the IV
744 		 * array so that it is 8-byte aligned.
745 		 */
746 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
747 		    en->tls_vminor == TLS_MINOR_VER_TWO)
748 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
749 	}
750 
751 	*tlsp = tls;
752 	return (0);
753 
754 out:
755 	ktls_free(tls);
756 	return (error);
757 }
758 
759 static struct ktls_session *
760 ktls_clone_session(struct ktls_session *tls, int direction)
761 {
762 	struct ktls_session *tls_new;
763 
764 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
765 
766 	counter_u64_add(ktls_offload_active, 1);
767 
768 	refcount_init(&tls_new->refcount, 1);
769 	if (direction == KTLS_RX) {
770 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
771 		    tls_new);
772 	} else {
773 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
774 		    tls_new);
775 		tls_new->inp = tls->inp;
776 		tls_new->tx = true;
777 		in_pcbref(tls_new->inp);
778 	}
779 
780 	/* Copy fields from existing session. */
781 	tls_new->params = tls->params;
782 	tls_new->wq_index = tls->wq_index;
783 
784 	/* Deep copy keys. */
785 	if (tls_new->params.auth_key != NULL) {
786 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
787 		    M_KTLS, M_WAITOK);
788 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
789 		    tls->params.auth_key_len);
790 	}
791 
792 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
793 	    M_WAITOK);
794 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
795 	    tls->params.cipher_key_len);
796 
797 	return (tls_new);
798 }
799 
800 #ifdef TCP_OFFLOAD
801 static int
802 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
803 {
804 	struct inpcb *inp;
805 	struct tcpcb *tp;
806 	int error;
807 
808 	inp = so->so_pcb;
809 	INP_WLOCK(inp);
810 	if (inp->inp_flags & INP_DROPPED) {
811 		INP_WUNLOCK(inp);
812 		return (ECONNRESET);
813 	}
814 	if (inp->inp_socket == NULL) {
815 		INP_WUNLOCK(inp);
816 		return (ECONNRESET);
817 	}
818 	tp = intotcpcb(inp);
819 	if (!(tp->t_flags & TF_TOE)) {
820 		INP_WUNLOCK(inp);
821 		return (EOPNOTSUPP);
822 	}
823 
824 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
825 	INP_WUNLOCK(inp);
826 	if (error == 0) {
827 		tls->mode = TCP_TLS_MODE_TOE;
828 		switch (tls->params.cipher_algorithm) {
829 		case CRYPTO_AES_CBC:
830 			counter_u64_add(ktls_toe_cbc, 1);
831 			break;
832 		case CRYPTO_AES_NIST_GCM_16:
833 			counter_u64_add(ktls_toe_gcm, 1);
834 			break;
835 		case CRYPTO_CHACHA20_POLY1305:
836 			counter_u64_add(ktls_toe_chacha20, 1);
837 			break;
838 		}
839 	}
840 	return (error);
841 }
842 #endif
843 
844 /*
845  * Common code used when first enabling ifnet TLS on a connection or
846  * when allocating a new ifnet TLS session due to a routing change.
847  * This function allocates a new TLS send tag on whatever interface
848  * the connection is currently routed over.
849  */
850 static int
851 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
852     struct m_snd_tag **mstp)
853 {
854 	union if_snd_tag_alloc_params params;
855 	struct ifnet *ifp;
856 	struct nhop_object *nh;
857 	struct tcpcb *tp;
858 	int error;
859 
860 	INP_RLOCK(inp);
861 	if (inp->inp_flags & INP_DROPPED) {
862 		INP_RUNLOCK(inp);
863 		return (ECONNRESET);
864 	}
865 	if (inp->inp_socket == NULL) {
866 		INP_RUNLOCK(inp);
867 		return (ECONNRESET);
868 	}
869 	tp = intotcpcb(inp);
870 
871 	/*
872 	 * Check administrative controls on ifnet TLS to determine if
873 	 * ifnet TLS should be denied.
874 	 *
875 	 * - Always permit 'force' requests.
876 	 * - ktls_ifnet_permitted == 0: always deny.
877 	 */
878 	if (!force && ktls_ifnet_permitted == 0) {
879 		INP_RUNLOCK(inp);
880 		return (ENXIO);
881 	}
882 
883 	/*
884 	 * XXX: Use the cached route in the inpcb to find the
885 	 * interface.  This should perhaps instead use
886 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
887 	 * enabled after a connection has completed key negotiation in
888 	 * userland, the cached route will be present in practice.
889 	 */
890 	nh = inp->inp_route.ro_nh;
891 	if (nh == NULL) {
892 		INP_RUNLOCK(inp);
893 		return (ENXIO);
894 	}
895 	ifp = nh->nh_ifp;
896 	if_ref(ifp);
897 
898 	/*
899 	 * Allocate a TLS + ratelimit tag if the connection has an
900 	 * existing pacing rate.
901 	 */
902 	if (tp->t_pacing_rate != -1 &&
903 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
904 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
905 		params.tls_rate_limit.inp = inp;
906 		params.tls_rate_limit.tls = tls;
907 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
908 	} else {
909 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
910 		params.tls.inp = inp;
911 		params.tls.tls = tls;
912 	}
913 	params.hdr.flowid = inp->inp_flowid;
914 	params.hdr.flowtype = inp->inp_flowtype;
915 	params.hdr.numa_domain = inp->inp_numa_domain;
916 	INP_RUNLOCK(inp);
917 
918 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
919 		error = EOPNOTSUPP;
920 		goto out;
921 	}
922 	if (inp->inp_vflag & INP_IPV6) {
923 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
924 			error = EOPNOTSUPP;
925 			goto out;
926 		}
927 	} else {
928 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
929 			error = EOPNOTSUPP;
930 			goto out;
931 		}
932 	}
933 	error = m_snd_tag_alloc(ifp, &params, mstp);
934 out:
935 	if_rele(ifp);
936 	return (error);
937 }
938 
939 /*
940  * Allocate an initial TLS receive tag for doing HW decryption of TLS
941  * data.
942  *
943  * This function allocates a new TLS receive tag on whatever interface
944  * the connection is currently routed over.  If the connection ends up
945  * using a different interface for receive this will get fixed up via
946  * ktls_input_ifp_mismatch as future packets arrive.
947  */
948 static int
949 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
950     struct m_snd_tag **mstp)
951 {
952 	union if_snd_tag_alloc_params params;
953 	struct ifnet *ifp;
954 	struct nhop_object *nh;
955 	int error;
956 
957 	if (!ktls_ocf_recrypt_supported(tls))
958 		return (ENXIO);
959 
960 	INP_RLOCK(inp);
961 	if (inp->inp_flags & INP_DROPPED) {
962 		INP_RUNLOCK(inp);
963 		return (ECONNRESET);
964 	}
965 	if (inp->inp_socket == NULL) {
966 		INP_RUNLOCK(inp);
967 		return (ECONNRESET);
968 	}
969 
970 	/*
971 	 * Check administrative controls on ifnet TLS to determine if
972 	 * ifnet TLS should be denied.
973 	 */
974 	if (ktls_ifnet_permitted == 0) {
975 		INP_RUNLOCK(inp);
976 		return (ENXIO);
977 	}
978 
979 	/*
980 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
981 	 * the inpcb to find the interface.
982 	 */
983 	nh = inp->inp_route.ro_nh;
984 	if (nh == NULL) {
985 		INP_RUNLOCK(inp);
986 		return (ENXIO);
987 	}
988 	ifp = nh->nh_ifp;
989 	if_ref(ifp);
990 	tls->rx_ifp = ifp;
991 
992 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
993 	params.hdr.flowid = inp->inp_flowid;
994 	params.hdr.flowtype = inp->inp_flowtype;
995 	params.hdr.numa_domain = inp->inp_numa_domain;
996 	params.tls_rx.inp = inp;
997 	params.tls_rx.tls = tls;
998 	params.tls_rx.vlan_id = 0;
999 
1000 	INP_RUNLOCK(inp);
1001 
1002 	if (inp->inp_vflag & INP_IPV6) {
1003 		if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) {
1004 			error = EOPNOTSUPP;
1005 			goto out;
1006 		}
1007 	} else {
1008 		if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) {
1009 			error = EOPNOTSUPP;
1010 			goto out;
1011 		}
1012 	}
1013 	error = m_snd_tag_alloc(ifp, &params, mstp);
1014 
1015 	/*
1016 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1017 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1018 	 * ID for use in ktls_reset_receive_tag which allocates new
1019 	 * receive tags directly from the leaf interface bypassing
1020 	 * if_vlan.
1021 	 */
1022 	if (error == 0)
1023 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1024 out:
1025 	return (error);
1026 }
1027 
1028 static int
1029 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1030     bool force)
1031 {
1032 	struct m_snd_tag *mst;
1033 	int error;
1034 
1035 	switch (direction) {
1036 	case KTLS_TX:
1037 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1038 		if (__predict_false(error != 0))
1039 			goto done;
1040 		break;
1041 	case KTLS_RX:
1042 		KASSERT(!force, ("%s: forced receive tag", __func__));
1043 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1044 		if (__predict_false(error != 0))
1045 			goto done;
1046 		break;
1047 	default:
1048 		__assert_unreachable();
1049 	}
1050 
1051 	tls->mode = TCP_TLS_MODE_IFNET;
1052 	tls->snd_tag = mst;
1053 
1054 	switch (tls->params.cipher_algorithm) {
1055 	case CRYPTO_AES_CBC:
1056 		counter_u64_add(ktls_ifnet_cbc, 1);
1057 		break;
1058 	case CRYPTO_AES_NIST_GCM_16:
1059 		counter_u64_add(ktls_ifnet_gcm, 1);
1060 		break;
1061 	case CRYPTO_CHACHA20_POLY1305:
1062 		counter_u64_add(ktls_ifnet_chacha20, 1);
1063 		break;
1064 	default:
1065 		break;
1066 	}
1067 done:
1068 	return (error);
1069 }
1070 
1071 static void
1072 ktls_use_sw(struct ktls_session *tls)
1073 {
1074 	tls->mode = TCP_TLS_MODE_SW;
1075 	switch (tls->params.cipher_algorithm) {
1076 	case CRYPTO_AES_CBC:
1077 		counter_u64_add(ktls_sw_cbc, 1);
1078 		break;
1079 	case CRYPTO_AES_NIST_GCM_16:
1080 		counter_u64_add(ktls_sw_gcm, 1);
1081 		break;
1082 	case CRYPTO_CHACHA20_POLY1305:
1083 		counter_u64_add(ktls_sw_chacha20, 1);
1084 		break;
1085 	}
1086 }
1087 
1088 static int
1089 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1090 {
1091 	int error;
1092 
1093 	error = ktls_ocf_try(so, tls, direction);
1094 	if (error)
1095 		return (error);
1096 	ktls_use_sw(tls);
1097 	return (0);
1098 }
1099 
1100 /*
1101  * KTLS RX stores data in the socket buffer as a list of TLS records,
1102  * where each record is stored as a control message containg the TLS
1103  * header followed by data mbufs containing the decrypted data.  This
1104  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1105  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1106  * should be queued to the socket buffer as records, but encrypted
1107  * data which needs to be decrypted by software arrives as a stream of
1108  * regular mbufs which need to be converted.  In addition, there may
1109  * already be pending encrypted data in the socket buffer when KTLS RX
1110  * is enabled.
1111  *
1112  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1113  * is used:
1114  *
1115  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1116  *
1117  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1118  *   from the first mbuf.  Once all of the data for that TLS record is
1119  *   queued, the socket is queued to a worker thread.
1120  *
1121  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1122  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1123  *   decrypted, and inserted into the regular socket buffer chain as
1124  *   record starting with a control message holding the TLS header and
1125  *   a chain of mbufs holding the encrypted data.
1126  */
1127 
1128 static void
1129 sb_mark_notready(struct sockbuf *sb)
1130 {
1131 	struct mbuf *m;
1132 
1133 	m = sb->sb_mb;
1134 	sb->sb_mtls = m;
1135 	sb->sb_mb = NULL;
1136 	sb->sb_mbtail = NULL;
1137 	sb->sb_lastrecord = NULL;
1138 	for (; m != NULL; m = m->m_next) {
1139 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1140 		    __func__));
1141 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1142 		    __func__));
1143 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1144 		    __func__));
1145 		m->m_flags |= M_NOTREADY;
1146 		sb->sb_acc -= m->m_len;
1147 		sb->sb_tlscc += m->m_len;
1148 		sb->sb_mtlstail = m;
1149 	}
1150 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1151 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1152 	    sb->sb_ccc));
1153 }
1154 
1155 /*
1156  * Return information about the pending TLS data in a socket
1157  * buffer.  On return, 'seqno' is set to the sequence number
1158  * of the next TLS record to be received, 'resid' is set to
1159  * the amount of bytes still needed for the last pending
1160  * record.  The function returns 'false' if the last pending
1161  * record contains a partial TLS header.  In that case, 'resid'
1162  * is the number of bytes needed to complete the TLS header.
1163  */
1164 bool
1165 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1166 {
1167 	struct tls_record_layer hdr;
1168 	struct mbuf *m;
1169 	uint64_t seqno;
1170 	size_t resid;
1171 	u_int offset, record_len;
1172 
1173 	SOCKBUF_LOCK_ASSERT(sb);
1174 	MPASS(sb->sb_flags & SB_TLS_RX);
1175 	seqno = sb->sb_tls_seqno;
1176 	resid = sb->sb_tlscc;
1177 	m = sb->sb_mtls;
1178 	offset = 0;
1179 
1180 	if (resid == 0) {
1181 		*seqnop = seqno;
1182 		*residp = 0;
1183 		return (true);
1184 	}
1185 
1186 	for (;;) {
1187 		seqno++;
1188 
1189 		if (resid < sizeof(hdr)) {
1190 			*seqnop = seqno;
1191 			*residp = sizeof(hdr) - resid;
1192 			return (false);
1193 		}
1194 
1195 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1196 
1197 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1198 		if (resid <= record_len) {
1199 			*seqnop = seqno;
1200 			*residp = record_len - resid;
1201 			return (true);
1202 		}
1203 		resid -= record_len;
1204 
1205 		while (record_len != 0) {
1206 			if (m->m_len - offset > record_len) {
1207 				offset += record_len;
1208 				break;
1209 			}
1210 
1211 			record_len -= (m->m_len - offset);
1212 			offset = 0;
1213 			m = m->m_next;
1214 		}
1215 	}
1216 }
1217 
1218 int
1219 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1220 {
1221 	struct ktls_session *tls;
1222 	int error;
1223 
1224 	if (!ktls_offload_enable)
1225 		return (ENOTSUP);
1226 	if (SOLISTENING(so))
1227 		return (EINVAL);
1228 
1229 	counter_u64_add(ktls_offload_enable_calls, 1);
1230 
1231 	/*
1232 	 * This should always be true since only the TCP socket option
1233 	 * invokes this function.
1234 	 */
1235 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1236 		return (EINVAL);
1237 
1238 	/*
1239 	 * XXX: Don't overwrite existing sessions.  We should permit
1240 	 * this to support rekeying in the future.
1241 	 */
1242 	if (so->so_rcv.sb_tls_info != NULL)
1243 		return (EALREADY);
1244 
1245 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1246 		return (ENOTSUP);
1247 
1248 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1249 	if (error)
1250 		return (error);
1251 
1252 	error = ktls_ocf_try(so, tls, KTLS_RX);
1253 	if (error) {
1254 		ktls_free(tls);
1255 		return (error);
1256 	}
1257 
1258 	/* Mark the socket as using TLS offload. */
1259 	SOCKBUF_LOCK(&so->so_rcv);
1260 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1261 	so->so_rcv.sb_tls_info = tls;
1262 	so->so_rcv.sb_flags |= SB_TLS_RX;
1263 
1264 	/* Mark existing data as not ready until it can be decrypted. */
1265 	sb_mark_notready(&so->so_rcv);
1266 	ktls_check_rx(&so->so_rcv);
1267 	SOCKBUF_UNLOCK(&so->so_rcv);
1268 
1269 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1270 #ifdef TCP_OFFLOAD
1271 	error = ktls_try_toe(so, tls, KTLS_RX);
1272 	if (error)
1273 #endif
1274 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1275 	if (error)
1276 		ktls_use_sw(tls);
1277 
1278 	counter_u64_add(ktls_offload_total, 1);
1279 
1280 	return (0);
1281 }
1282 
1283 int
1284 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1285 {
1286 	struct ktls_session *tls;
1287 	struct inpcb *inp;
1288 	struct tcpcb *tp;
1289 	int error;
1290 
1291 	if (!ktls_offload_enable)
1292 		return (ENOTSUP);
1293 	if (SOLISTENING(so))
1294 		return (EINVAL);
1295 
1296 	counter_u64_add(ktls_offload_enable_calls, 1);
1297 
1298 	/*
1299 	 * This should always be true since only the TCP socket option
1300 	 * invokes this function.
1301 	 */
1302 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1303 		return (EINVAL);
1304 
1305 	/*
1306 	 * XXX: Don't overwrite existing sessions.  We should permit
1307 	 * this to support rekeying in the future.
1308 	 */
1309 	if (so->so_snd.sb_tls_info != NULL)
1310 		return (EALREADY);
1311 
1312 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1313 		return (ENOTSUP);
1314 
1315 	/* TLS requires ext pgs */
1316 	if (mb_use_ext_pgs == 0)
1317 		return (ENXIO);
1318 
1319 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1320 	if (error)
1321 		return (error);
1322 
1323 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1324 #ifdef TCP_OFFLOAD
1325 	error = ktls_try_toe(so, tls, KTLS_TX);
1326 	if (error)
1327 #endif
1328 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1329 	if (error)
1330 		error = ktls_try_sw(so, tls, KTLS_TX);
1331 
1332 	if (error) {
1333 		ktls_free(tls);
1334 		return (error);
1335 	}
1336 
1337 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1338 	if (error) {
1339 		ktls_free(tls);
1340 		return (error);
1341 	}
1342 
1343 	/*
1344 	 * Write lock the INP when setting sb_tls_info so that
1345 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1346 	 * holding the INP lock.
1347 	 */
1348 	inp = so->so_pcb;
1349 	INP_WLOCK(inp);
1350 	SOCKBUF_LOCK(&so->so_snd);
1351 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1352 	so->so_snd.sb_tls_info = tls;
1353 	if (tls->mode != TCP_TLS_MODE_SW) {
1354 		tp = intotcpcb(inp);
1355 		MPASS(tp->t_nic_ktls_xmit == 0);
1356 		tp->t_nic_ktls_xmit = 1;
1357 		if (tp->t_fb->tfb_hwtls_change != NULL)
1358 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1359 	}
1360 	SOCKBUF_UNLOCK(&so->so_snd);
1361 	INP_WUNLOCK(inp);
1362 	SOCK_IO_SEND_UNLOCK(so);
1363 
1364 	counter_u64_add(ktls_offload_total, 1);
1365 
1366 	return (0);
1367 }
1368 
1369 int
1370 ktls_get_rx_mode(struct socket *so, int *modep)
1371 {
1372 	struct ktls_session *tls;
1373 	struct inpcb *inp __diagused;
1374 
1375 	if (SOLISTENING(so))
1376 		return (EINVAL);
1377 	inp = so->so_pcb;
1378 	INP_WLOCK_ASSERT(inp);
1379 	SOCK_RECVBUF_LOCK(so);
1380 	tls = so->so_rcv.sb_tls_info;
1381 	if (tls == NULL)
1382 		*modep = TCP_TLS_MODE_NONE;
1383 	else
1384 		*modep = tls->mode;
1385 	SOCK_RECVBUF_UNLOCK(so);
1386 	return (0);
1387 }
1388 
1389 /*
1390  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1391  *
1392  * This function gets information about the next TCP- and TLS-
1393  * sequence number to be processed by the TLS receive worker
1394  * thread. The information is extracted from the given "inpcb"
1395  * structure. The values are stored in host endian format at the two
1396  * given output pointer locations. The TCP sequence number points to
1397  * the beginning of the TLS header.
1398  *
1399  * This function returns zero on success, else a non-zero error code
1400  * is returned.
1401  */
1402 int
1403 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1404 {
1405 	struct socket *so;
1406 	struct tcpcb *tp;
1407 
1408 	INP_RLOCK(inp);
1409 	so = inp->inp_socket;
1410 	if (__predict_false(so == NULL)) {
1411 		INP_RUNLOCK(inp);
1412 		return (EINVAL);
1413 	}
1414 	if (inp->inp_flags & INP_DROPPED) {
1415 		INP_RUNLOCK(inp);
1416 		return (ECONNRESET);
1417 	}
1418 
1419 	tp = intotcpcb(inp);
1420 	MPASS(tp != NULL);
1421 
1422 	SOCKBUF_LOCK(&so->so_rcv);
1423 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1424 	*tlsseq = so->so_rcv.sb_tls_seqno;
1425 	SOCKBUF_UNLOCK(&so->so_rcv);
1426 
1427 	INP_RUNLOCK(inp);
1428 
1429 	return (0);
1430 }
1431 
1432 int
1433 ktls_get_tx_mode(struct socket *so, int *modep)
1434 {
1435 	struct ktls_session *tls;
1436 	struct inpcb *inp __diagused;
1437 
1438 	if (SOLISTENING(so))
1439 		return (EINVAL);
1440 	inp = so->so_pcb;
1441 	INP_WLOCK_ASSERT(inp);
1442 	SOCK_SENDBUF_LOCK(so);
1443 	tls = so->so_snd.sb_tls_info;
1444 	if (tls == NULL)
1445 		*modep = TCP_TLS_MODE_NONE;
1446 	else
1447 		*modep = tls->mode;
1448 	SOCK_SENDBUF_UNLOCK(so);
1449 	return (0);
1450 }
1451 
1452 /*
1453  * Switch between SW and ifnet TLS sessions as requested.
1454  */
1455 int
1456 ktls_set_tx_mode(struct socket *so, int mode)
1457 {
1458 	struct ktls_session *tls, *tls_new;
1459 	struct inpcb *inp;
1460 	struct tcpcb *tp;
1461 	int error;
1462 
1463 	if (SOLISTENING(so))
1464 		return (EINVAL);
1465 	switch (mode) {
1466 	case TCP_TLS_MODE_SW:
1467 	case TCP_TLS_MODE_IFNET:
1468 		break;
1469 	default:
1470 		return (EINVAL);
1471 	}
1472 
1473 	inp = so->so_pcb;
1474 	INP_WLOCK_ASSERT(inp);
1475 	tp = intotcpcb(inp);
1476 
1477 	if (mode == TCP_TLS_MODE_IFNET) {
1478 		/* Don't allow enabling ifnet ktls multiple times */
1479 		if (tp->t_nic_ktls_xmit)
1480 			return (EALREADY);
1481 
1482 		/*
1483 		 * Don't enable ifnet ktls if we disabled it due to an
1484 		 * excessive retransmission rate
1485 		 */
1486 		if (tp->t_nic_ktls_xmit_dis)
1487 			return (ENXIO);
1488 	}
1489 
1490 	SOCKBUF_LOCK(&so->so_snd);
1491 	tls = so->so_snd.sb_tls_info;
1492 	if (tls == NULL) {
1493 		SOCKBUF_UNLOCK(&so->so_snd);
1494 		return (0);
1495 	}
1496 
1497 	if (tls->mode == mode) {
1498 		SOCKBUF_UNLOCK(&so->so_snd);
1499 		return (0);
1500 	}
1501 
1502 	tls = ktls_hold(tls);
1503 	SOCKBUF_UNLOCK(&so->so_snd);
1504 	INP_WUNLOCK(inp);
1505 
1506 	tls_new = ktls_clone_session(tls, KTLS_TX);
1507 
1508 	if (mode == TCP_TLS_MODE_IFNET)
1509 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1510 	else
1511 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1512 	if (error) {
1513 		counter_u64_add(ktls_switch_failed, 1);
1514 		ktls_free(tls_new);
1515 		ktls_free(tls);
1516 		INP_WLOCK(inp);
1517 		return (error);
1518 	}
1519 
1520 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1521 	if (error) {
1522 		counter_u64_add(ktls_switch_failed, 1);
1523 		ktls_free(tls_new);
1524 		ktls_free(tls);
1525 		INP_WLOCK(inp);
1526 		return (error);
1527 	}
1528 
1529 	/*
1530 	 * If we raced with another session change, keep the existing
1531 	 * session.
1532 	 */
1533 	if (tls != so->so_snd.sb_tls_info) {
1534 		counter_u64_add(ktls_switch_failed, 1);
1535 		SOCK_IO_SEND_UNLOCK(so);
1536 		ktls_free(tls_new);
1537 		ktls_free(tls);
1538 		INP_WLOCK(inp);
1539 		return (EBUSY);
1540 	}
1541 
1542 	INP_WLOCK(inp);
1543 	SOCKBUF_LOCK(&so->so_snd);
1544 	so->so_snd.sb_tls_info = tls_new;
1545 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1546 		MPASS(tp->t_nic_ktls_xmit == 0);
1547 		tp->t_nic_ktls_xmit = 1;
1548 		if (tp->t_fb->tfb_hwtls_change != NULL)
1549 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1550 	}
1551 	SOCKBUF_UNLOCK(&so->so_snd);
1552 	SOCK_IO_SEND_UNLOCK(so);
1553 
1554 	/*
1555 	 * Drop two references on 'tls'.  The first is for the
1556 	 * ktls_hold() above.  The second drops the reference from the
1557 	 * socket buffer.
1558 	 */
1559 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1560 	ktls_free(tls);
1561 	ktls_free(tls);
1562 
1563 	if (mode == TCP_TLS_MODE_IFNET)
1564 		counter_u64_add(ktls_switch_to_ifnet, 1);
1565 	else
1566 		counter_u64_add(ktls_switch_to_sw, 1);
1567 
1568 	return (0);
1569 }
1570 
1571 /*
1572  * Try to allocate a new TLS receive tag.  This task is scheduled when
1573  * sbappend_ktls_rx detects an input path change.  If a new tag is
1574  * allocated, replace the tag in the TLS session.  If a new tag cannot
1575  * be allocated, let the session fall back to software decryption.
1576  */
1577 static void
1578 ktls_reset_receive_tag(void *context, int pending)
1579 {
1580 	union if_snd_tag_alloc_params params;
1581 	struct ktls_session *tls;
1582 	struct m_snd_tag *mst;
1583 	struct inpcb *inp;
1584 	struct ifnet *ifp;
1585 	struct socket *so;
1586 	int error;
1587 
1588 	MPASS(pending == 1);
1589 
1590 	tls = context;
1591 	so = tls->so;
1592 	inp = so->so_pcb;
1593 	ifp = NULL;
1594 
1595 	INP_RLOCK(inp);
1596 	if (inp->inp_flags & INP_DROPPED) {
1597 		INP_RUNLOCK(inp);
1598 		goto out;
1599 	}
1600 
1601 	SOCKBUF_LOCK(&so->so_rcv);
1602 	mst = tls->snd_tag;
1603 	tls->snd_tag = NULL;
1604 	if (mst != NULL)
1605 		m_snd_tag_rele(mst);
1606 
1607 	ifp = tls->rx_ifp;
1608 	if_ref(ifp);
1609 	SOCKBUF_UNLOCK(&so->so_rcv);
1610 
1611 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1612 	params.hdr.flowid = inp->inp_flowid;
1613 	params.hdr.flowtype = inp->inp_flowtype;
1614 	params.hdr.numa_domain = inp->inp_numa_domain;
1615 	params.tls_rx.inp = inp;
1616 	params.tls_rx.tls = tls;
1617 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1618 	INP_RUNLOCK(inp);
1619 
1620 	if (inp->inp_vflag & INP_IPV6) {
1621 		if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0)
1622 			goto out;
1623 	} else {
1624 		if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0)
1625 			goto out;
1626 	}
1627 
1628 	error = m_snd_tag_alloc(ifp, &params, &mst);
1629 	if (error == 0) {
1630 		SOCKBUF_LOCK(&so->so_rcv);
1631 		tls->snd_tag = mst;
1632 		SOCKBUF_UNLOCK(&so->so_rcv);
1633 
1634 		counter_u64_add(ktls_ifnet_reset, 1);
1635 	} else {
1636 		/*
1637 		 * Just fall back to software decryption if a tag
1638 		 * cannot be allocated leaving the connection intact.
1639 		 * If a future input path change switches to another
1640 		 * interface this connection will resume ifnet TLS.
1641 		 */
1642 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1643 	}
1644 
1645 out:
1646 	mtx_pool_lock(mtxpool_sleep, tls);
1647 	tls->reset_pending = false;
1648 	mtx_pool_unlock(mtxpool_sleep, tls);
1649 
1650 	if (ifp != NULL)
1651 		if_rele(ifp);
1652 	sorele(so);
1653 	ktls_free(tls);
1654 }
1655 
1656 /*
1657  * Try to allocate a new TLS send tag.  This task is scheduled when
1658  * ip_output detects a route change while trying to transmit a packet
1659  * holding a TLS record.  If a new tag is allocated, replace the tag
1660  * in the TLS session.  Subsequent packets on the connection will use
1661  * the new tag.  If a new tag cannot be allocated, drop the
1662  * connection.
1663  */
1664 static void
1665 ktls_reset_send_tag(void *context, int pending)
1666 {
1667 	struct epoch_tracker et;
1668 	struct ktls_session *tls;
1669 	struct m_snd_tag *old, *new;
1670 	struct inpcb *inp;
1671 	struct tcpcb *tp;
1672 	int error;
1673 
1674 	MPASS(pending == 1);
1675 
1676 	tls = context;
1677 	inp = tls->inp;
1678 
1679 	/*
1680 	 * Free the old tag first before allocating a new one.
1681 	 * ip[6]_output_send() will treat a NULL send tag the same as
1682 	 * an ifp mismatch and drop packets until a new tag is
1683 	 * allocated.
1684 	 *
1685 	 * Write-lock the INP when changing tls->snd_tag since
1686 	 * ip[6]_output_send() holds a read-lock when reading the
1687 	 * pointer.
1688 	 */
1689 	INP_WLOCK(inp);
1690 	old = tls->snd_tag;
1691 	tls->snd_tag = NULL;
1692 	INP_WUNLOCK(inp);
1693 	if (old != NULL)
1694 		m_snd_tag_rele(old);
1695 
1696 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1697 
1698 	if (error == 0) {
1699 		INP_WLOCK(inp);
1700 		tls->snd_tag = new;
1701 		mtx_pool_lock(mtxpool_sleep, tls);
1702 		tls->reset_pending = false;
1703 		mtx_pool_unlock(mtxpool_sleep, tls);
1704 		INP_WUNLOCK(inp);
1705 
1706 		counter_u64_add(ktls_ifnet_reset, 1);
1707 
1708 		/*
1709 		 * XXX: Should we kick tcp_output explicitly now that
1710 		 * the send tag is fixed or just rely on timers?
1711 		 */
1712 	} else {
1713 		NET_EPOCH_ENTER(et);
1714 		INP_WLOCK(inp);
1715 		if (!(inp->inp_flags & INP_DROPPED)) {
1716 			tp = intotcpcb(inp);
1717 			CURVNET_SET(inp->inp_vnet);
1718 			tp = tcp_drop(tp, ECONNABORTED);
1719 			CURVNET_RESTORE();
1720 			if (tp != NULL)
1721 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1722 		}
1723 		INP_WUNLOCK(inp);
1724 		NET_EPOCH_EXIT(et);
1725 
1726 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1727 
1728 		/*
1729 		 * Leave reset_pending true to avoid future tasks while
1730 		 * the socket goes away.
1731 		 */
1732 	}
1733 
1734 	ktls_free(tls);
1735 }
1736 
1737 void
1738 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1739 {
1740 	struct ktls_session *tls;
1741 	struct socket *so;
1742 
1743 	SOCKBUF_LOCK_ASSERT(sb);
1744 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1745 	    __func__, sb));
1746 	so = __containerof(sb, struct socket, so_rcv);
1747 
1748 	tls = sb->sb_tls_info;
1749 	if_rele(tls->rx_ifp);
1750 	if_ref(ifp);
1751 	tls->rx_ifp = ifp;
1752 
1753 	/*
1754 	 * See if we should schedule a task to update the receive tag for
1755 	 * this session.
1756 	 */
1757 	mtx_pool_lock(mtxpool_sleep, tls);
1758 	if (!tls->reset_pending) {
1759 		(void) ktls_hold(tls);
1760 		soref(so);
1761 		tls->so = so;
1762 		tls->reset_pending = true;
1763 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1764 	}
1765 	mtx_pool_unlock(mtxpool_sleep, tls);
1766 }
1767 
1768 int
1769 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1770 {
1771 
1772 	if (inp == NULL)
1773 		return (ENOBUFS);
1774 
1775 	INP_LOCK_ASSERT(inp);
1776 
1777 	/*
1778 	 * See if we should schedule a task to update the send tag for
1779 	 * this session.
1780 	 */
1781 	mtx_pool_lock(mtxpool_sleep, tls);
1782 	if (!tls->reset_pending) {
1783 		(void) ktls_hold(tls);
1784 		tls->reset_pending = true;
1785 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1786 	}
1787 	mtx_pool_unlock(mtxpool_sleep, tls);
1788 	return (ENOBUFS);
1789 }
1790 
1791 #ifdef RATELIMIT
1792 int
1793 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1794 {
1795 	union if_snd_tag_modify_params params = {
1796 		.rate_limit.max_rate = max_pacing_rate,
1797 		.rate_limit.flags = M_NOWAIT,
1798 	};
1799 	struct m_snd_tag *mst;
1800 
1801 	/* Can't get to the inp, but it should be locked. */
1802 	/* INP_LOCK_ASSERT(inp); */
1803 
1804 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1805 
1806 	if (tls->snd_tag == NULL) {
1807 		/*
1808 		 * Resetting send tag, ignore this change.  The
1809 		 * pending reset may or may not see this updated rate
1810 		 * in the tcpcb.  If it doesn't, we will just lose
1811 		 * this rate change.
1812 		 */
1813 		return (0);
1814 	}
1815 
1816 	mst = tls->snd_tag;
1817 
1818 	MPASS(mst != NULL);
1819 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1820 
1821 	return (mst->sw->snd_tag_modify(mst, &params));
1822 }
1823 #endif
1824 #endif
1825 
1826 static void
1827 ktls_destroy_help(void *context, int pending __unused)
1828 {
1829 	ktls_destroy(context);
1830 }
1831 
1832 void
1833 ktls_destroy(struct ktls_session *tls)
1834 {
1835 	struct inpcb *inp;
1836 	struct tcpcb *tp;
1837 	bool wlocked;
1838 
1839 	MPASS(tls->refcount == 0);
1840 
1841 	inp = tls->inp;
1842 	if (tls->tx) {
1843 		wlocked = INP_WLOCKED(inp);
1844 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1845 			/*
1846 			 * rwlocks read locks are anonymous, and there
1847 			 * is no way to know if our current thread
1848 			 * holds an rlock on the inp.  As a rough
1849 			 * estimate, check to see if the thread holds
1850 			 * *any* rlocks at all.  If it does not, then we
1851 			 * know that we don't hold the inp rlock, and
1852 			 * can safely take the wlock
1853 			 */
1854 			if (curthread->td_rw_rlocks == 0) {
1855 				INP_WLOCK(inp);
1856 			} else {
1857 				/*
1858 				 * We might hold the rlock, so let's
1859 				 * do the destroy in a taskqueue
1860 				 * context to avoid a potential
1861 				 * deadlock.  This should be very
1862 				 * rare.
1863 				 */
1864 				counter_u64_add(ktls_destroy_task, 1);
1865 				TASK_INIT(&tls->destroy_task, 0,
1866 				    ktls_destroy_help, tls);
1867 				(void)taskqueue_enqueue(taskqueue_thread,
1868 				    &tls->destroy_task);
1869 				return;
1870 			}
1871 		}
1872 	}
1873 
1874 	if (tls->sequential_records) {
1875 		struct mbuf *m, *n;
1876 		int page_count;
1877 
1878 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1879 			page_count = m->m_epg_enc_cnt;
1880 			while (page_count > 0) {
1881 				KASSERT(page_count >= m->m_epg_nrdy,
1882 				    ("%s: too few pages", __func__));
1883 				page_count -= m->m_epg_nrdy;
1884 				m = m_free(m);
1885 			}
1886 		}
1887 	}
1888 
1889 	counter_u64_add(ktls_offload_active, -1);
1890 	switch (tls->mode) {
1891 	case TCP_TLS_MODE_SW:
1892 		switch (tls->params.cipher_algorithm) {
1893 		case CRYPTO_AES_CBC:
1894 			counter_u64_add(ktls_sw_cbc, -1);
1895 			break;
1896 		case CRYPTO_AES_NIST_GCM_16:
1897 			counter_u64_add(ktls_sw_gcm, -1);
1898 			break;
1899 		case CRYPTO_CHACHA20_POLY1305:
1900 			counter_u64_add(ktls_sw_chacha20, -1);
1901 			break;
1902 		}
1903 		break;
1904 	case TCP_TLS_MODE_IFNET:
1905 		switch (tls->params.cipher_algorithm) {
1906 		case CRYPTO_AES_CBC:
1907 			counter_u64_add(ktls_ifnet_cbc, -1);
1908 			break;
1909 		case CRYPTO_AES_NIST_GCM_16:
1910 			counter_u64_add(ktls_ifnet_gcm, -1);
1911 			break;
1912 		case CRYPTO_CHACHA20_POLY1305:
1913 			counter_u64_add(ktls_ifnet_chacha20, -1);
1914 			break;
1915 		}
1916 		if (tls->snd_tag != NULL)
1917 			m_snd_tag_rele(tls->snd_tag);
1918 		if (tls->rx_ifp != NULL)
1919 			if_rele(tls->rx_ifp);
1920 		if (tls->tx) {
1921 			INP_WLOCK_ASSERT(inp);
1922 			tp = intotcpcb(inp);
1923 			MPASS(tp->t_nic_ktls_xmit == 1);
1924 			tp->t_nic_ktls_xmit = 0;
1925 		}
1926 		break;
1927 #ifdef TCP_OFFLOAD
1928 	case TCP_TLS_MODE_TOE:
1929 		switch (tls->params.cipher_algorithm) {
1930 		case CRYPTO_AES_CBC:
1931 			counter_u64_add(ktls_toe_cbc, -1);
1932 			break;
1933 		case CRYPTO_AES_NIST_GCM_16:
1934 			counter_u64_add(ktls_toe_gcm, -1);
1935 			break;
1936 		case CRYPTO_CHACHA20_POLY1305:
1937 			counter_u64_add(ktls_toe_chacha20, -1);
1938 			break;
1939 		}
1940 		break;
1941 #endif
1942 	}
1943 	if (tls->ocf_session != NULL)
1944 		ktls_ocf_free(tls);
1945 	if (tls->params.auth_key != NULL) {
1946 		zfree(tls->params.auth_key, M_KTLS);
1947 		tls->params.auth_key = NULL;
1948 		tls->params.auth_key_len = 0;
1949 	}
1950 	if (tls->params.cipher_key != NULL) {
1951 		zfree(tls->params.cipher_key, M_KTLS);
1952 		tls->params.cipher_key = NULL;
1953 		tls->params.cipher_key_len = 0;
1954 	}
1955 	if (tls->tx) {
1956 		INP_WLOCK_ASSERT(inp);
1957 		if (!in_pcbrele_wlocked(inp) && !wlocked)
1958 			INP_WUNLOCK(inp);
1959 	}
1960 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1961 
1962 	uma_zfree(ktls_session_zone, tls);
1963 }
1964 
1965 void
1966 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1967 {
1968 
1969 	for (; m != NULL; m = m->m_next) {
1970 		KASSERT((m->m_flags & M_EXTPG) != 0,
1971 		    ("ktls_seq: mapped mbuf %p", m));
1972 
1973 		m->m_epg_seqno = sb->sb_tls_seqno;
1974 		sb->sb_tls_seqno++;
1975 	}
1976 }
1977 
1978 /*
1979  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1980  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1981  * mbuf must be populated with the payload of each TLS record.
1982  *
1983  * The record_type argument specifies the TLS record type used when
1984  * populating the TLS header.
1985  *
1986  * The enq_count argument on return is set to the number of pages of
1987  * payload data for this entire chain that need to be encrypted via SW
1988  * encryption.  The returned value should be passed to ktls_enqueue
1989  * when scheduling encryption of this chain of mbufs.  To handle the
1990  * special case of empty fragments for TLS 1.0 sessions, an empty
1991  * fragment counts as one page.
1992  */
1993 void
1994 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1995     uint8_t record_type)
1996 {
1997 	struct tls_record_layer *tlshdr;
1998 	struct mbuf *m;
1999 	uint64_t *noncep;
2000 	uint16_t tls_len;
2001 	int maxlen __diagused;
2002 
2003 	maxlen = tls->params.max_frame_len;
2004 	*enq_cnt = 0;
2005 	for (m = top; m != NULL; m = m->m_next) {
2006 		/*
2007 		 * All mbufs in the chain should be TLS records whose
2008 		 * payload does not exceed the maximum frame length.
2009 		 *
2010 		 * Empty TLS 1.0 records are permitted when using CBC.
2011 		 */
2012 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2013 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2014 		    ("ktls_frame: m %p len %d", m, m->m_len));
2015 
2016 		/*
2017 		 * TLS frames require unmapped mbufs to store session
2018 		 * info.
2019 		 */
2020 		KASSERT((m->m_flags & M_EXTPG) != 0,
2021 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2022 
2023 		tls_len = m->m_len;
2024 
2025 		/* Save a reference to the session. */
2026 		m->m_epg_tls = ktls_hold(tls);
2027 
2028 		m->m_epg_hdrlen = tls->params.tls_hlen;
2029 		m->m_epg_trllen = tls->params.tls_tlen;
2030 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2031 			int bs, delta;
2032 
2033 			/*
2034 			 * AES-CBC pads messages to a multiple of the
2035 			 * block size.  Note that the padding is
2036 			 * applied after the digest and the encryption
2037 			 * is done on the "plaintext || mac || padding".
2038 			 * At least one byte of padding is always
2039 			 * present.
2040 			 *
2041 			 * Compute the final trailer length assuming
2042 			 * at most one block of padding.
2043 			 * tls->params.tls_tlen is the maximum
2044 			 * possible trailer length (padding + digest).
2045 			 * delta holds the number of excess padding
2046 			 * bytes if the maximum were used.  Those
2047 			 * extra bytes are removed.
2048 			 */
2049 			bs = tls->params.tls_bs;
2050 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2051 			m->m_epg_trllen -= delta;
2052 		}
2053 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2054 
2055 		/* Populate the TLS header. */
2056 		tlshdr = (void *)m->m_epg_hdr;
2057 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2058 
2059 		/*
2060 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2061 		 * of TLS_RLTYPE_APP.
2062 		 */
2063 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2064 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2065 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2066 			tlshdr->tls_type = TLS_RLTYPE_APP;
2067 			/* save the real record type for later */
2068 			m->m_epg_record_type = record_type;
2069 			m->m_epg_trail[0] = record_type;
2070 		} else {
2071 			tlshdr->tls_vminor = tls->params.tls_vminor;
2072 			tlshdr->tls_type = record_type;
2073 		}
2074 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2075 
2076 		/*
2077 		 * Store nonces / explicit IVs after the end of the
2078 		 * TLS header.
2079 		 *
2080 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2081 		 * from the end of the IV.  The nonce is then
2082 		 * incremented for use by the next record.
2083 		 *
2084 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2085 		 */
2086 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2087 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2088 			noncep = (uint64_t *)(tls->params.iv + 8);
2089 			be64enc(tlshdr + 1, *noncep);
2090 			(*noncep)++;
2091 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2092 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2093 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2094 
2095 		/*
2096 		 * When using SW encryption, mark the mbuf not ready.
2097 		 * It will be marked ready via sbready() after the
2098 		 * record has been encrypted.
2099 		 *
2100 		 * When using ifnet TLS, unencrypted TLS records are
2101 		 * sent down the stack to the NIC.
2102 		 */
2103 		if (tls->mode == TCP_TLS_MODE_SW) {
2104 			m->m_flags |= M_NOTREADY;
2105 			if (__predict_false(tls_len == 0)) {
2106 				/* TLS 1.0 empty fragment. */
2107 				m->m_epg_nrdy = 1;
2108 			} else
2109 				m->m_epg_nrdy = m->m_epg_npgs;
2110 			*enq_cnt += m->m_epg_nrdy;
2111 		}
2112 	}
2113 }
2114 
2115 bool
2116 ktls_permit_empty_frames(struct ktls_session *tls)
2117 {
2118 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2119 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2120 }
2121 
2122 void
2123 ktls_check_rx(struct sockbuf *sb)
2124 {
2125 	struct tls_record_layer hdr;
2126 	struct ktls_wq *wq;
2127 	struct socket *so;
2128 	bool running;
2129 
2130 	SOCKBUF_LOCK_ASSERT(sb);
2131 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2132 	    __func__, sb));
2133 	so = __containerof(sb, struct socket, so_rcv);
2134 
2135 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2136 		return;
2137 
2138 	/* Is there enough queued for a TLS header? */
2139 	if (sb->sb_tlscc < sizeof(hdr)) {
2140 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2141 			so->so_error = EMSGSIZE;
2142 		return;
2143 	}
2144 
2145 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2146 
2147 	/* Is the entire record queued? */
2148 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2149 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2150 			so->so_error = EMSGSIZE;
2151 		return;
2152 	}
2153 
2154 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2155 
2156 	soref(so);
2157 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2158 	mtx_lock(&wq->mtx);
2159 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2160 	running = wq->running;
2161 	mtx_unlock(&wq->mtx);
2162 	if (!running)
2163 		wakeup(wq);
2164 	counter_u64_add(ktls_cnt_rx_queued, 1);
2165 }
2166 
2167 static struct mbuf *
2168 ktls_detach_record(struct sockbuf *sb, int len)
2169 {
2170 	struct mbuf *m, *n, *top;
2171 	int remain;
2172 
2173 	SOCKBUF_LOCK_ASSERT(sb);
2174 	MPASS(len <= sb->sb_tlscc);
2175 
2176 	/*
2177 	 * If TLS chain is the exact size of the record,
2178 	 * just grab the whole record.
2179 	 */
2180 	top = sb->sb_mtls;
2181 	if (sb->sb_tlscc == len) {
2182 		sb->sb_mtls = NULL;
2183 		sb->sb_mtlstail = NULL;
2184 		goto out;
2185 	}
2186 
2187 	/*
2188 	 * While it would be nice to use m_split() here, we need
2189 	 * to know exactly what m_split() allocates to update the
2190 	 * accounting, so do it inline instead.
2191 	 */
2192 	remain = len;
2193 	for (m = top; remain > m->m_len; m = m->m_next)
2194 		remain -= m->m_len;
2195 
2196 	/* Easy case: don't have to split 'm'. */
2197 	if (remain == m->m_len) {
2198 		sb->sb_mtls = m->m_next;
2199 		if (sb->sb_mtls == NULL)
2200 			sb->sb_mtlstail = NULL;
2201 		m->m_next = NULL;
2202 		goto out;
2203 	}
2204 
2205 	/*
2206 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2207 	 * with M_NOWAIT first.
2208 	 */
2209 	n = m_get(M_NOWAIT, MT_DATA);
2210 	if (n == NULL) {
2211 		/*
2212 		 * Use M_WAITOK with socket buffer unlocked.  If
2213 		 * 'sb_mtls' changes while the lock is dropped, return
2214 		 * NULL to force the caller to retry.
2215 		 */
2216 		SOCKBUF_UNLOCK(sb);
2217 
2218 		n = m_get(M_WAITOK, MT_DATA);
2219 
2220 		SOCKBUF_LOCK(sb);
2221 		if (sb->sb_mtls != top) {
2222 			m_free(n);
2223 			return (NULL);
2224 		}
2225 	}
2226 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2227 
2228 	/* Store remainder in 'n'. */
2229 	n->m_len = m->m_len - remain;
2230 	if (m->m_flags & M_EXT) {
2231 		n->m_data = m->m_data + remain;
2232 		mb_dupcl(n, m);
2233 	} else {
2234 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2235 	}
2236 
2237 	/* Trim 'm' and update accounting. */
2238 	m->m_len -= n->m_len;
2239 	sb->sb_tlscc -= n->m_len;
2240 	sb->sb_ccc -= n->m_len;
2241 
2242 	/* Account for 'n'. */
2243 	sballoc_ktls_rx(sb, n);
2244 
2245 	/* Insert 'n' into the TLS chain. */
2246 	sb->sb_mtls = n;
2247 	n->m_next = m->m_next;
2248 	if (sb->sb_mtlstail == m)
2249 		sb->sb_mtlstail = n;
2250 
2251 	/* Detach the record from the TLS chain. */
2252 	m->m_next = NULL;
2253 
2254 out:
2255 	MPASS(m_length(top, NULL) == len);
2256 	for (m = top; m != NULL; m = m->m_next)
2257 		sbfree_ktls_rx(sb, m);
2258 	sb->sb_tlsdcc = len;
2259 	sb->sb_ccc += len;
2260 	SBCHECK(sb);
2261 	return (top);
2262 }
2263 
2264 /*
2265  * Determine the length of the trailing zero padding and find the real
2266  * record type in the byte before the padding.
2267  *
2268  * Walking the mbuf chain backwards is clumsy, so another option would
2269  * be to scan forwards remembering the last non-zero byte before the
2270  * trailer.  However, it would be expensive to scan the entire record.
2271  * Instead, find the last non-zero byte of each mbuf in the chain
2272  * keeping track of the relative offset of that nonzero byte.
2273  *
2274  * trail_len is the size of the MAC/tag on input and is set to the
2275  * size of the full trailer including padding and the record type on
2276  * return.
2277  */
2278 static int
2279 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2280     int *trailer_len, uint8_t *record_typep)
2281 {
2282 	char *cp;
2283 	u_int digest_start, last_offset, m_len, offset;
2284 	uint8_t record_type;
2285 
2286 	digest_start = tls_len - *trailer_len;
2287 	last_offset = 0;
2288 	offset = 0;
2289 	for (; m != NULL && offset < digest_start;
2290 	     offset += m->m_len, m = m->m_next) {
2291 		/* Don't look for padding in the tag. */
2292 		m_len = min(digest_start - offset, m->m_len);
2293 		cp = mtod(m, char *);
2294 
2295 		/* Find last non-zero byte in this mbuf. */
2296 		while (m_len > 0 && cp[m_len - 1] == 0)
2297 			m_len--;
2298 		if (m_len > 0) {
2299 			record_type = cp[m_len - 1];
2300 			last_offset = offset + m_len;
2301 		}
2302 	}
2303 	if (last_offset < tls->params.tls_hlen)
2304 		return (EBADMSG);
2305 
2306 	*record_typep = record_type;
2307 	*trailer_len = tls_len - last_offset + 1;
2308 	return (0);
2309 }
2310 
2311 /*
2312  * Check if a mbuf chain is fully decrypted at the given offset and
2313  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2314  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2315  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2316  * is encrypted.
2317  */
2318 ktls_mbuf_crypto_st_t
2319 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2320 {
2321 	int m_flags_ored = 0;
2322 	int m_flags_anded = -1;
2323 
2324 	for (; mb != NULL; mb = mb->m_next) {
2325 		if (offset < mb->m_len)
2326 			break;
2327 		offset -= mb->m_len;
2328 	}
2329 	offset += len;
2330 
2331 	for (; mb != NULL; mb = mb->m_next) {
2332 		m_flags_ored |= mb->m_flags;
2333 		m_flags_anded &= mb->m_flags;
2334 
2335 		if (offset <= mb->m_len)
2336 			break;
2337 		offset -= mb->m_len;
2338 	}
2339 	MPASS(mb != NULL || offset == 0);
2340 
2341 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2342 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2343 	else
2344 		return ((m_flags_ored & M_DECRYPTED) ?
2345 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2346 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2347 }
2348 
2349 /*
2350  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2351  */
2352 static int
2353 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2354 {
2355 	union if_snd_tag_modify_params params;
2356 	struct m_snd_tag *mst;
2357 	struct inpcb *inp;
2358 	struct tcpcb *tp;
2359 
2360 	mst = so->so_rcv.sb_tls_info->snd_tag;
2361 	if (__predict_false(mst == NULL))
2362 		return (EINVAL);
2363 
2364 	inp = sotoinpcb(so);
2365 	if (__predict_false(inp == NULL))
2366 		return (EINVAL);
2367 
2368 	INP_RLOCK(inp);
2369 	if (inp->inp_flags & INP_DROPPED) {
2370 		INP_RUNLOCK(inp);
2371 		return (ECONNRESET);
2372 	}
2373 
2374 	tp = intotcpcb(inp);
2375 	MPASS(tp != NULL);
2376 
2377 	/* Get the TCP sequence number of the next valid TLS header. */
2378 	SOCKBUF_LOCK(&so->so_rcv);
2379 	params.tls_rx.tls_hdr_tcp_sn =
2380 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2381 	params.tls_rx.tls_rec_length = tls_len;
2382 	params.tls_rx.tls_seq_number = tls_rcd_num;
2383 	SOCKBUF_UNLOCK(&so->so_rcv);
2384 
2385 	INP_RUNLOCK(inp);
2386 
2387 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2388 	return (mst->sw->snd_tag_modify(mst, &params));
2389 }
2390 
2391 static void
2392 ktls_drop(struct socket *so, int error)
2393 {
2394 	struct epoch_tracker et;
2395 	struct inpcb *inp = sotoinpcb(so);
2396 	struct tcpcb *tp;
2397 
2398 	NET_EPOCH_ENTER(et);
2399 	INP_WLOCK(inp);
2400 	if (!(inp->inp_flags & INP_DROPPED)) {
2401 		tp = intotcpcb(inp);
2402 		CURVNET_SET(inp->inp_vnet);
2403 		tp = tcp_drop(tp, error);
2404 		CURVNET_RESTORE();
2405 		if (tp != NULL)
2406 			INP_WUNLOCK(inp);
2407 	} else {
2408 		so->so_error = error;
2409 		SOCK_RECVBUF_LOCK(so);
2410 		sorwakeup_locked(so);
2411 		INP_WUNLOCK(inp);
2412 	}
2413 	NET_EPOCH_EXIT(et);
2414 }
2415 
2416 static void
2417 ktls_decrypt(struct socket *so)
2418 {
2419 	char tls_header[MBUF_PEXT_HDR_LEN];
2420 	struct ktls_session *tls;
2421 	struct sockbuf *sb;
2422 	struct tls_record_layer *hdr;
2423 	struct tls_get_record tgr;
2424 	struct mbuf *control, *data, *m;
2425 	ktls_mbuf_crypto_st_t state;
2426 	uint64_t seqno;
2427 	int error, remain, tls_len, trail_len;
2428 	bool tls13;
2429 	uint8_t vminor, record_type;
2430 
2431 	hdr = (struct tls_record_layer *)tls_header;
2432 	sb = &so->so_rcv;
2433 	SOCKBUF_LOCK(sb);
2434 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2435 	    ("%s: socket %p not running", __func__, so));
2436 
2437 	tls = sb->sb_tls_info;
2438 	MPASS(tls != NULL);
2439 
2440 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2441 	if (tls13)
2442 		vminor = TLS_MINOR_VER_TWO;
2443 	else
2444 		vminor = tls->params.tls_vminor;
2445 	for (;;) {
2446 		/* Is there enough queued for a TLS header? */
2447 		if (sb->sb_tlscc < tls->params.tls_hlen)
2448 			break;
2449 
2450 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2451 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2452 
2453 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2454 		    hdr->tls_vminor != vminor)
2455 			error = EINVAL;
2456 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2457 			error = EINVAL;
2458 		else if (tls_len < tls->params.tls_hlen || tls_len >
2459 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2460 		    tls->params.tls_tlen)
2461 			error = EMSGSIZE;
2462 		else
2463 			error = 0;
2464 		if (__predict_false(error != 0)) {
2465 			/*
2466 			 * We have a corrupted record and are likely
2467 			 * out of sync.  The connection isn't
2468 			 * recoverable at this point, so abort it.
2469 			 */
2470 			SOCKBUF_UNLOCK(sb);
2471 			counter_u64_add(ktls_offload_corrupted_records, 1);
2472 
2473 			ktls_drop(so, error);
2474 			goto deref;
2475 		}
2476 
2477 		/* Is the entire record queued? */
2478 		if (sb->sb_tlscc < tls_len)
2479 			break;
2480 
2481 		/*
2482 		 * Split out the portion of the mbuf chain containing
2483 		 * this TLS record.
2484 		 */
2485 		data = ktls_detach_record(sb, tls_len);
2486 		if (data == NULL)
2487 			continue;
2488 		MPASS(sb->sb_tlsdcc == tls_len);
2489 
2490 		seqno = sb->sb_tls_seqno;
2491 		sb->sb_tls_seqno++;
2492 		SBCHECK(sb);
2493 		SOCKBUF_UNLOCK(sb);
2494 
2495 		/* get crypto state for this TLS record */
2496 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2497 
2498 		switch (state) {
2499 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2500 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2501 			if (error)
2502 				break;
2503 			/* FALLTHROUGH */
2504 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2505 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2506 			    &trail_len);
2507 			if (__predict_true(error == 0)) {
2508 				if (tls13) {
2509 					error = tls13_find_record_type(tls, data,
2510 					    tls_len, &trail_len, &record_type);
2511 				} else {
2512 					record_type = hdr->tls_type;
2513 				}
2514 			}
2515 			break;
2516 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2517 			/*
2518 			 * NIC TLS is only supported for AEAD
2519 			 * ciphersuites which used a fixed sized
2520 			 * trailer.
2521 			 */
2522 			if (tls13) {
2523 				trail_len = tls->params.tls_tlen - 1;
2524 				error = tls13_find_record_type(tls, data,
2525 				    tls_len, &trail_len, &record_type);
2526 			} else {
2527 				trail_len = tls->params.tls_tlen;
2528 				error = 0;
2529 				record_type = hdr->tls_type;
2530 			}
2531 			break;
2532 		default:
2533 			error = EINVAL;
2534 			break;
2535 		}
2536 		if (error) {
2537 			counter_u64_add(ktls_offload_failed_crypto, 1);
2538 
2539 			SOCKBUF_LOCK(sb);
2540 			if (sb->sb_tlsdcc == 0) {
2541 				/*
2542 				 * sbcut/drop/flush discarded these
2543 				 * mbufs.
2544 				 */
2545 				m_freem(data);
2546 				break;
2547 			}
2548 
2549 			/*
2550 			 * Drop this TLS record's data, but keep
2551 			 * decrypting subsequent records.
2552 			 */
2553 			sb->sb_ccc -= tls_len;
2554 			sb->sb_tlsdcc = 0;
2555 
2556 			if (error != EMSGSIZE)
2557 				error = EBADMSG;
2558 			CURVNET_SET(so->so_vnet);
2559 			so->so_error = error;
2560 			sorwakeup_locked(so);
2561 			CURVNET_RESTORE();
2562 
2563 			m_freem(data);
2564 
2565 			SOCKBUF_LOCK(sb);
2566 			continue;
2567 		}
2568 
2569 		/* Allocate the control mbuf. */
2570 		memset(&tgr, 0, sizeof(tgr));
2571 		tgr.tls_type = record_type;
2572 		tgr.tls_vmajor = hdr->tls_vmajor;
2573 		tgr.tls_vminor = hdr->tls_vminor;
2574 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2575 		    trail_len);
2576 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2577 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2578 
2579 		SOCKBUF_LOCK(sb);
2580 		if (sb->sb_tlsdcc == 0) {
2581 			/* sbcut/drop/flush discarded these mbufs. */
2582 			MPASS(sb->sb_tlscc == 0);
2583 			m_freem(data);
2584 			m_freem(control);
2585 			break;
2586 		}
2587 
2588 		/*
2589 		 * Clear the 'dcc' accounting in preparation for
2590 		 * adding the decrypted record.
2591 		 */
2592 		sb->sb_ccc -= tls_len;
2593 		sb->sb_tlsdcc = 0;
2594 		SBCHECK(sb);
2595 
2596 		/* If there is no payload, drop all of the data. */
2597 		if (tgr.tls_length == htobe16(0)) {
2598 			m_freem(data);
2599 			data = NULL;
2600 		} else {
2601 			/* Trim header. */
2602 			remain = tls->params.tls_hlen;
2603 			while (remain > 0) {
2604 				if (data->m_len > remain) {
2605 					data->m_data += remain;
2606 					data->m_len -= remain;
2607 					break;
2608 				}
2609 				remain -= data->m_len;
2610 				data = m_free(data);
2611 			}
2612 
2613 			/* Trim trailer and clear M_NOTREADY. */
2614 			remain = be16toh(tgr.tls_length);
2615 			m = data;
2616 			for (m = data; remain > m->m_len; m = m->m_next) {
2617 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2618 				remain -= m->m_len;
2619 			}
2620 			m->m_len = remain;
2621 			m_freem(m->m_next);
2622 			m->m_next = NULL;
2623 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2624 
2625 			/* Set EOR on the final mbuf. */
2626 			m->m_flags |= M_EOR;
2627 		}
2628 
2629 		sbappendcontrol_locked(sb, data, control, 0);
2630 
2631 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2632 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2633 			SOCKBUF_UNLOCK(sb);
2634 			ktls_resync_ifnet(so, tls_len, seqno);
2635 			SOCKBUF_LOCK(sb);
2636 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2637 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2638 			SOCKBUF_UNLOCK(sb);
2639 			ktls_resync_ifnet(so, 0, seqno);
2640 			SOCKBUF_LOCK(sb);
2641 		}
2642 	}
2643 
2644 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2645 
2646 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2647 		so->so_error = EMSGSIZE;
2648 
2649 	sorwakeup_locked(so);
2650 
2651 deref:
2652 	SOCKBUF_UNLOCK_ASSERT(sb);
2653 
2654 	CURVNET_SET(so->so_vnet);
2655 	sorele(so);
2656 	CURVNET_RESTORE();
2657 }
2658 
2659 void
2660 ktls_enqueue_to_free(struct mbuf *m)
2661 {
2662 	struct ktls_wq *wq;
2663 	bool running;
2664 
2665 	/* Mark it for freeing. */
2666 	m->m_epg_flags |= EPG_FLAG_2FREE;
2667 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2668 	mtx_lock(&wq->mtx);
2669 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2670 	running = wq->running;
2671 	mtx_unlock(&wq->mtx);
2672 	if (!running)
2673 		wakeup(wq);
2674 }
2675 
2676 static void *
2677 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2678 {
2679 	void *buf;
2680 	int domain, running;
2681 
2682 	if (m->m_epg_npgs <= 2)
2683 		return (NULL);
2684 	if (ktls_buffer_zone == NULL)
2685 		return (NULL);
2686 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2687 		/*
2688 		 * Rate-limit allocation attempts after a failure.
2689 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2690 		 * the free page queues and may fail consistently if memory is
2691 		 * fragmented.
2692 		 */
2693 		return (NULL);
2694 	}
2695 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2696 	if (buf == NULL) {
2697 		domain = PCPU_GET(domain);
2698 		wq->lastallocfail = ticks;
2699 
2700 		/*
2701 		 * Note that this check is "racy", but the races are
2702 		 * harmless, and are either a spurious wakeup if
2703 		 * multiple threads fail allocations before the alloc
2704 		 * thread wakes, or waiting an extra second in case we
2705 		 * see an old value of running == true.
2706 		 */
2707 		if (!VM_DOMAIN_EMPTY(domain)) {
2708 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2709 			if (!running)
2710 				wakeup(&ktls_domains[domain].alloc_td);
2711 		}
2712 	}
2713 	return (buf);
2714 }
2715 
2716 static int
2717 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2718     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2719 {
2720 	vm_page_t pg;
2721 	int error, i, len, off;
2722 
2723 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2724 	    ("%p not unready & nomap mbuf\n", m));
2725 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2726 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2727 	    ktls_maxlen));
2728 
2729 	/* Anonymous mbufs are encrypted in place. */
2730 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2731 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2732 
2733 	/*
2734 	 * For file-backed mbufs (from sendfile), anonymous wired
2735 	 * pages are allocated and used as the encryption destination.
2736 	 */
2737 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2738 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2739 		    m->m_epg_1st_off;
2740 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2741 		    m->m_epg_1st_off;
2742 		state->dst_iov[0].iov_len = len;
2743 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2744 		i = 1;
2745 	} else {
2746 		off = m->m_epg_1st_off;
2747 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2748 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2749 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2750 			len = m_epg_pagelen(m, i, off);
2751 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2752 			state->dst_iov[i].iov_base =
2753 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2754 			state->dst_iov[i].iov_len = len;
2755 		}
2756 	}
2757 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2758 	state->dst_iov[i].iov_base = m->m_epg_trail;
2759 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2760 
2761 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2762 
2763 	if (__predict_false(error != 0)) {
2764 		/* Free the anonymous pages. */
2765 		if (state->cbuf != NULL)
2766 			uma_zfree(ktls_buffer_zone, state->cbuf);
2767 		else {
2768 			for (i = 0; i < m->m_epg_npgs; i++) {
2769 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2770 				(void)vm_page_unwire_noq(pg);
2771 				vm_page_free(pg);
2772 			}
2773 		}
2774 	}
2775 	return (error);
2776 }
2777 
2778 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2779 static u_int
2780 ktls_batched_records(struct mbuf *m)
2781 {
2782 	int page_count, records;
2783 
2784 	records = 0;
2785 	page_count = m->m_epg_enc_cnt;
2786 	while (page_count > 0) {
2787 		records++;
2788 		page_count -= m->m_epg_nrdy;
2789 		m = m->m_next;
2790 	}
2791 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2792 	return (records);
2793 }
2794 
2795 void
2796 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2797 {
2798 	struct ktls_session *tls;
2799 	struct ktls_wq *wq;
2800 	int queued;
2801 	bool running;
2802 
2803 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2804 	    (M_EXTPG | M_NOTREADY)),
2805 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2806 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2807 
2808 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2809 
2810 	m->m_epg_enc_cnt = page_count;
2811 
2812 	/*
2813 	 * Save a pointer to the socket.  The caller is responsible
2814 	 * for taking an additional reference via soref().
2815 	 */
2816 	m->m_epg_so = so;
2817 
2818 	queued = 1;
2819 	tls = m->m_epg_tls;
2820 	wq = &ktls_wq[tls->wq_index];
2821 	mtx_lock(&wq->mtx);
2822 	if (__predict_false(tls->sequential_records)) {
2823 		/*
2824 		 * For TLS 1.0, records must be encrypted
2825 		 * sequentially.  For a given connection, all records
2826 		 * queued to the associated work queue are processed
2827 		 * sequentially.  However, sendfile(2) might complete
2828 		 * I/O requests spanning multiple TLS records out of
2829 		 * order.  Here we ensure TLS records are enqueued to
2830 		 * the work queue in FIFO order.
2831 		 *
2832 		 * tls->next_seqno holds the sequence number of the
2833 		 * next TLS record that should be enqueued to the work
2834 		 * queue.  If this next record is not tls->next_seqno,
2835 		 * it must be a future record, so insert it, sorted by
2836 		 * TLS sequence number, into tls->pending_records and
2837 		 * return.
2838 		 *
2839 		 * If this TLS record matches tls->next_seqno, place
2840 		 * it in the work queue and then check
2841 		 * tls->pending_records to see if any
2842 		 * previously-queued records are now ready for
2843 		 * encryption.
2844 		 */
2845 		if (m->m_epg_seqno != tls->next_seqno) {
2846 			struct mbuf *n, *p;
2847 
2848 			p = NULL;
2849 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2850 				if (n->m_epg_seqno > m->m_epg_seqno)
2851 					break;
2852 				p = n;
2853 			}
2854 			if (n == NULL)
2855 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2856 				    m_epg_stailq);
2857 			else if (p == NULL)
2858 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2859 				    m_epg_stailq);
2860 			else
2861 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2862 				    m_epg_stailq);
2863 			mtx_unlock(&wq->mtx);
2864 			counter_u64_add(ktls_cnt_tx_pending, 1);
2865 			return;
2866 		}
2867 
2868 		tls->next_seqno += ktls_batched_records(m);
2869 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2870 
2871 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2872 			struct mbuf *n;
2873 
2874 			n = STAILQ_FIRST(&tls->pending_records);
2875 			if (n->m_epg_seqno != tls->next_seqno)
2876 				break;
2877 
2878 			queued++;
2879 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2880 			tls->next_seqno += ktls_batched_records(n);
2881 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2882 		}
2883 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2884 	} else
2885 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2886 
2887 	running = wq->running;
2888 	mtx_unlock(&wq->mtx);
2889 	if (!running)
2890 		wakeup(wq);
2891 	counter_u64_add(ktls_cnt_tx_queued, queued);
2892 }
2893 
2894 /*
2895  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2896  * the pages from the file and replace them with the anonymous pages
2897  * allocated in ktls_encrypt_record().
2898  */
2899 static void
2900 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2901 {
2902 	int i;
2903 
2904 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2905 
2906 	/* Free the old pages. */
2907 	m->m_ext.ext_free(m);
2908 
2909 	/* Replace them with the new pages. */
2910 	if (state->cbuf != NULL) {
2911 		for (i = 0; i < m->m_epg_npgs; i++)
2912 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2913 
2914 		/* Contig pages should go back to the cache. */
2915 		m->m_ext.ext_free = ktls_free_mext_contig;
2916 	} else {
2917 		for (i = 0; i < m->m_epg_npgs; i++)
2918 			m->m_epg_pa[i] = state->parray[i];
2919 
2920 		/* Use the basic free routine. */
2921 		m->m_ext.ext_free = mb_free_mext_pgs;
2922 	}
2923 
2924 	/* Pages are now writable. */
2925 	m->m_epg_flags |= EPG_FLAG_ANON;
2926 }
2927 
2928 static __noinline void
2929 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2930 {
2931 	struct ktls_ocf_encrypt_state state;
2932 	struct ktls_session *tls;
2933 	struct socket *so;
2934 	struct mbuf *m;
2935 	int error, npages, total_pages;
2936 
2937 	so = top->m_epg_so;
2938 	tls = top->m_epg_tls;
2939 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2940 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2941 #ifdef INVARIANTS
2942 	top->m_epg_so = NULL;
2943 #endif
2944 	total_pages = top->m_epg_enc_cnt;
2945 	npages = 0;
2946 
2947 	/*
2948 	 * Encrypt the TLS records in the chain of mbufs starting with
2949 	 * 'top'.  'total_pages' gives us a total count of pages and is
2950 	 * used to know when we have finished encrypting the TLS
2951 	 * records originally queued with 'top'.
2952 	 *
2953 	 * NB: These mbufs are queued in the socket buffer and
2954 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2955 	 * socket buffer lock is not held while traversing this chain.
2956 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2957 	 * pointers should be stable.  However, the 'm_next' of the
2958 	 * last mbuf encrypted is not necessarily NULL.  It can point
2959 	 * to other mbufs appended while 'top' was on the TLS work
2960 	 * queue.
2961 	 *
2962 	 * Each mbuf holds an entire TLS record.
2963 	 */
2964 	error = 0;
2965 	for (m = top; npages != total_pages; m = m->m_next) {
2966 		KASSERT(m->m_epg_tls == tls,
2967 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2968 		    tls, m->m_epg_tls));
2969 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2970 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2971 		    total_pages, m));
2972 
2973 		error = ktls_encrypt_record(wq, m, tls, &state);
2974 		if (error) {
2975 			counter_u64_add(ktls_offload_failed_crypto, 1);
2976 			break;
2977 		}
2978 
2979 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2980 			ktls_finish_nonanon(m, &state);
2981 
2982 		npages += m->m_epg_nrdy;
2983 
2984 		/*
2985 		 * Drop a reference to the session now that it is no
2986 		 * longer needed.  Existing code depends on encrypted
2987 		 * records having no associated session vs
2988 		 * yet-to-be-encrypted records having an associated
2989 		 * session.
2990 		 */
2991 		m->m_epg_tls = NULL;
2992 		ktls_free(tls);
2993 	}
2994 
2995 	CURVNET_SET(so->so_vnet);
2996 	if (error == 0) {
2997 		(void)so->so_proto->pr_ready(so, top, npages);
2998 	} else {
2999 		ktls_drop(so, EIO);
3000 		mb_free_notready(top, total_pages);
3001 	}
3002 
3003 	sorele(so);
3004 	CURVNET_RESTORE();
3005 }
3006 
3007 void
3008 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3009 {
3010 	struct ktls_session *tls;
3011 	struct socket *so;
3012 	struct mbuf *m;
3013 	int npages;
3014 
3015 	m = state->m;
3016 
3017 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3018 		ktls_finish_nonanon(m, state);
3019 
3020 	so = state->so;
3021 	free(state, M_KTLS);
3022 
3023 	/*
3024 	 * Drop a reference to the session now that it is no longer
3025 	 * needed.  Existing code depends on encrypted records having
3026 	 * no associated session vs yet-to-be-encrypted records having
3027 	 * an associated session.
3028 	 */
3029 	tls = m->m_epg_tls;
3030 	m->m_epg_tls = NULL;
3031 	ktls_free(tls);
3032 
3033 	if (error != 0)
3034 		counter_u64_add(ktls_offload_failed_crypto, 1);
3035 
3036 	CURVNET_SET(so->so_vnet);
3037 	npages = m->m_epg_nrdy;
3038 
3039 	if (error == 0) {
3040 		(void)so->so_proto->pr_ready(so, m, npages);
3041 	} else {
3042 		ktls_drop(so, EIO);
3043 		mb_free_notready(m, npages);
3044 	}
3045 
3046 	sorele(so);
3047 	CURVNET_RESTORE();
3048 }
3049 
3050 /*
3051  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3052  * (coprocessors) where encryption does not use host CPU resources and
3053  * it can be beneficial to queue more requests than CPUs.
3054  */
3055 static __noinline void
3056 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3057 {
3058 	struct ktls_ocf_encrypt_state *state;
3059 	struct ktls_session *tls;
3060 	struct socket *so;
3061 	struct mbuf *m, *n;
3062 	int error, mpages, npages, total_pages;
3063 
3064 	so = top->m_epg_so;
3065 	tls = top->m_epg_tls;
3066 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3067 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3068 #ifdef INVARIANTS
3069 	top->m_epg_so = NULL;
3070 #endif
3071 	total_pages = top->m_epg_enc_cnt;
3072 	npages = 0;
3073 
3074 	error = 0;
3075 	for (m = top; npages != total_pages; m = n) {
3076 		KASSERT(m->m_epg_tls == tls,
3077 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3078 		    tls, m->m_epg_tls));
3079 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3080 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3081 		    total_pages, m));
3082 
3083 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3084 		soref(so);
3085 		state->so = so;
3086 		state->m = m;
3087 
3088 		mpages = m->m_epg_nrdy;
3089 		n = m->m_next;
3090 
3091 		error = ktls_encrypt_record(wq, m, tls, state);
3092 		if (error) {
3093 			counter_u64_add(ktls_offload_failed_crypto, 1);
3094 			free(state, M_KTLS);
3095 			CURVNET_SET(so->so_vnet);
3096 			sorele(so);
3097 			CURVNET_RESTORE();
3098 			break;
3099 		}
3100 
3101 		npages += mpages;
3102 	}
3103 
3104 	CURVNET_SET(so->so_vnet);
3105 	if (error != 0) {
3106 		ktls_drop(so, EIO);
3107 		mb_free_notready(m, total_pages - npages);
3108 	}
3109 
3110 	sorele(so);
3111 	CURVNET_RESTORE();
3112 }
3113 
3114 static int
3115 ktls_bind_domain(int domain)
3116 {
3117 	int error;
3118 
3119 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3120 	if (error != 0)
3121 		return (error);
3122 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3123 	return (0);
3124 }
3125 
3126 static void
3127 ktls_alloc_thread(void *ctx)
3128 {
3129 	struct ktls_domain_info *ktls_domain = ctx;
3130 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3131 	void **buf;
3132 	struct sysctl_oid *oid;
3133 	char name[80];
3134 	int domain, error, i, nbufs;
3135 
3136 	domain = ktls_domain - ktls_domains;
3137 	if (bootverbose)
3138 		printf("Starting KTLS alloc thread for domain %d\n", domain);
3139 	error = ktls_bind_domain(domain);
3140 	if (error)
3141 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3142 		    domain, error);
3143 	snprintf(name, sizeof(name), "domain%d", domain);
3144 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3145 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3146 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3147 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
3148 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3149 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3150 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3151 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3152 
3153 	buf = NULL;
3154 	nbufs = 0;
3155 	for (;;) {
3156 		atomic_store_int(&sc->running, 0);
3157 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3158 		atomic_store_int(&sc->running, 1);
3159 		sc->wakeups++;
3160 		if (nbufs != ktls_max_alloc) {
3161 			free(buf, M_KTLS);
3162 			nbufs = atomic_load_int(&ktls_max_alloc);
3163 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3164 			    M_WAITOK | M_ZERO);
3165 		}
3166 		/*
3167 		 * Below we allocate nbufs with different allocation
3168 		 * flags than we use when allocating normally during
3169 		 * encryption in the ktls worker thread.  We specify
3170 		 * M_NORECLAIM in the worker thread. However, we omit
3171 		 * that flag here and add M_WAITOK so that the VM
3172 		 * system is permitted to perform expensive work to
3173 		 * defragment memory.  We do this here, as it does not
3174 		 * matter if this thread blocks.  If we block a ktls
3175 		 * worker thread, we risk developing backlogs of
3176 		 * buffers to be encrypted, leading to surges of
3177 		 * traffic and potential NIC output drops.
3178 		 */
3179 		for (i = 0; i < nbufs; i++) {
3180 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3181 			sc->allocs++;
3182 		}
3183 		for (i = 0; i < nbufs; i++) {
3184 			uma_zfree(ktls_buffer_zone, buf[i]);
3185 			buf[i] = NULL;
3186 		}
3187 	}
3188 }
3189 
3190 static void
3191 ktls_work_thread(void *ctx)
3192 {
3193 	struct ktls_wq *wq = ctx;
3194 	struct mbuf *m, *n;
3195 	struct socket *so, *son;
3196 	STAILQ_HEAD(, mbuf) local_m_head;
3197 	STAILQ_HEAD(, socket) local_so_head;
3198 	int cpu;
3199 
3200 	cpu = wq - ktls_wq;
3201 	if (bootverbose)
3202 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3203 
3204 	/*
3205 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3206 	 * we bind to the NUMA domain instead.
3207 	 */
3208 	if (ktls_bind_threads) {
3209 		int error;
3210 
3211 		if (ktls_bind_threads > 1) {
3212 			struct pcpu *pc = pcpu_find(cpu);
3213 
3214 			error = ktls_bind_domain(pc->pc_domain);
3215 		} else {
3216 			cpuset_t mask;
3217 
3218 			CPU_SETOF(cpu, &mask);
3219 			error = cpuset_setthread(curthread->td_tid, &mask);
3220 		}
3221 		if (error)
3222 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3223 				cpu, error);
3224 	}
3225 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3226 	fpu_kern_thread(0);
3227 #endif
3228 	for (;;) {
3229 		mtx_lock(&wq->mtx);
3230 		while (STAILQ_EMPTY(&wq->m_head) &&
3231 		    STAILQ_EMPTY(&wq->so_head)) {
3232 			wq->running = false;
3233 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3234 			wq->running = true;
3235 		}
3236 
3237 		STAILQ_INIT(&local_m_head);
3238 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3239 		STAILQ_INIT(&local_so_head);
3240 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3241 		mtx_unlock(&wq->mtx);
3242 
3243 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3244 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3245 				ktls_free(m->m_epg_tls);
3246 				m_free_raw(m);
3247 			} else {
3248 				if (m->m_epg_tls->sync_dispatch)
3249 					ktls_encrypt(wq, m);
3250 				else
3251 					ktls_encrypt_async(wq, m);
3252 				counter_u64_add(ktls_cnt_tx_queued, -1);
3253 			}
3254 		}
3255 
3256 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3257 			ktls_decrypt(so);
3258 			counter_u64_add(ktls_cnt_rx_queued, -1);
3259 		}
3260 	}
3261 }
3262 
3263 #if defined(INET) || defined(INET6)
3264 static void
3265 ktls_disable_ifnet_help(void *context, int pending __unused)
3266 {
3267 	struct ktls_session *tls;
3268 	struct inpcb *inp;
3269 	struct tcpcb *tp;
3270 	struct socket *so;
3271 	int err;
3272 
3273 	tls = context;
3274 	inp = tls->inp;
3275 	if (inp == NULL)
3276 		return;
3277 	INP_WLOCK(inp);
3278 	so = inp->inp_socket;
3279 	MPASS(so != NULL);
3280 	if (inp->inp_flags & INP_DROPPED) {
3281 		goto out;
3282 	}
3283 
3284 	if (so->so_snd.sb_tls_info != NULL)
3285 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3286 	else
3287 		err = ENXIO;
3288 	if (err == 0) {
3289 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3290 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3291 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3292 		    (tp = intotcpcb(inp)) != NULL &&
3293 		    tp->t_fb->tfb_hwtls_change != NULL)
3294 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3295 	} else {
3296 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3297 	}
3298 
3299 out:
3300 	CURVNET_SET(so->so_vnet);
3301 	sorele(so);
3302 	CURVNET_RESTORE();
3303 	INP_WUNLOCK(inp);
3304 	ktls_free(tls);
3305 }
3306 
3307 /*
3308  * Called when re-transmits are becoming a substantial portion of the
3309  * sends on this connection.  When this happens, we transition the
3310  * connection to software TLS.  This is needed because most inline TLS
3311  * NICs keep crypto state only for in-order transmits.  This means
3312  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3313  * re-DMA the entire TLS record up to and including the current
3314  * segment.  This means that when re-transmitting the last ~1448 byte
3315  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3316  * of magnitude more data than we are sending.  This can cause the
3317  * PCIe link to saturate well before the network, which can cause
3318  * output drops, and a general loss of capacity.
3319  */
3320 void
3321 ktls_disable_ifnet(void *arg)
3322 {
3323 	struct tcpcb *tp;
3324 	struct inpcb *inp;
3325 	struct socket *so;
3326 	struct ktls_session *tls;
3327 
3328 	tp = arg;
3329 	inp = tptoinpcb(tp);
3330 	INP_WLOCK_ASSERT(inp);
3331 	so = inp->inp_socket;
3332 	SOCK_LOCK(so);
3333 	tls = so->so_snd.sb_tls_info;
3334 	if (tp->t_nic_ktls_xmit_dis == 1) {
3335 		SOCK_UNLOCK(so);
3336 		return;
3337 	}
3338 
3339 	/*
3340 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3341 	 * ifnet can only be done once per connection, so we never want
3342 	 * to do it again
3343 	 */
3344 
3345 	(void)ktls_hold(tls);
3346 	soref(so);
3347 	tp->t_nic_ktls_xmit_dis = 1;
3348 	SOCK_UNLOCK(so);
3349 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3350 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3351 }
3352 #endif
3353