xref: /freebsd/sys/kern/uipc_ktls.c (revision 657729a89dd578d8cfc70d6616f5c65a48a8b33a)
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 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226     "Software TLS session stats");
227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228     "Hardware (ifnet) TLS session stats");
229 #ifdef TCP_OFFLOAD
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "TOE TLS session stats");
232 #endif
233 
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236     "Active number of software TLS sessions using AES-CBC");
237 
238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240     "Active number of software TLS sessions using AES-GCM");
241 
242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
244     &ktls_sw_chacha20,
245     "Active number of software TLS sessions using Chacha20-Poly1305");
246 
247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
249     &ktls_ifnet_cbc,
250     "Active number of ifnet TLS sessions using AES-CBC");
251 
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
254     &ktls_ifnet_gcm,
255     "Active number of ifnet TLS sessions using AES-GCM");
256 
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259     &ktls_ifnet_chacha20,
260     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
261 
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
265 
266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268     &ktls_ifnet_reset_dropped,
269     "TLS sessions dropped after failing to update ifnet send tag");
270 
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273     &ktls_ifnet_reset_failed,
274     "TLS sessions that failed to allocate a new ifnet send tag");
275 
276 static int ktls_ifnet_permitted;
277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278     &ktls_ifnet_permitted, 1,
279     "Whether to permit hardware (ifnet) TLS sessions");
280 
281 #ifdef TCP_OFFLOAD
282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
284     &ktls_toe_cbc,
285     "Active number of TOE TLS sessions using AES-CBC");
286 
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
289     &ktls_toe_gcm,
290     "Active number of TOE TLS sessions using AES-GCM");
291 
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
294     &ktls_toe_chacha20,
295     "Active number of TOE TLS sessions using Chacha20-Poly1305");
296 #endif
297 
298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
299 
300 #if defined(INET) || defined(INET6)
301 static void ktls_reset_receive_tag(void *context, int pending);
302 static void ktls_reset_send_tag(void *context, int pending);
303 #endif
304 static void ktls_work_thread(void *ctx);
305 static void ktls_alloc_thread(void *ctx);
306 
307 #if defined(INET) || defined(INET6)
308 static u_int
309 ktls_get_cpu(struct socket *so)
310 {
311 	struct inpcb *inp;
312 #ifdef NUMA
313 	struct ktls_domain_info *di;
314 #endif
315 	u_int cpuid;
316 
317 	inp = sotoinpcb(so);
318 #ifdef RSS
319 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
320 	if (cpuid != NETISR_CPUID_NONE)
321 		return (cpuid);
322 #endif
323 	/*
324 	 * Just use the flowid to shard connections in a repeatable
325 	 * fashion.  Note that TLS 1.0 sessions rely on the
326 	 * serialization provided by having the same connection use
327 	 * the same queue.
328 	 */
329 #ifdef NUMA
330 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
331 		di = &ktls_domains[inp->inp_numa_domain];
332 		cpuid = di->cpu[inp->inp_flowid % di->count];
333 	} else
334 #endif
335 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
336 	return (cpuid);
337 }
338 #endif
339 
340 static int
341 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
342 {
343 	vm_page_t m;
344 	int i, req;
345 
346 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
347 	    ("%s: ktls max length %d is not page size-aligned",
348 	    __func__, ktls_maxlen));
349 
350 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
351 	for (i = 0; i < count; i++) {
352 		m = vm_page_alloc_noobj_contig_domain(domain, req,
353 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
354 		    VM_MEMATTR_DEFAULT);
355 		if (m == NULL)
356 			break;
357 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
358 	}
359 	return (i);
360 }
361 
362 static void
363 ktls_buffer_release(void *arg __unused, void **store, int count)
364 {
365 	vm_page_t m;
366 	int i, j;
367 
368 	for (i = 0; i < count; i++) {
369 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
370 		for (j = 0; j < atop(ktls_maxlen); j++) {
371 			(void)vm_page_unwire_noq(m + j);
372 			vm_page_free(m + j);
373 		}
374 	}
375 }
376 
377 static void
378 ktls_free_mext_contig(struct mbuf *m)
379 {
380 	M_ASSERTEXTPG(m);
381 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
382 }
383 
384 static int
385 ktls_init(void)
386 {
387 	struct thread *td;
388 	struct pcpu *pc;
389 	int count, domain, error, i;
390 
391 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
392 	    M_WAITOK | M_ZERO);
393 
394 	ktls_session_zone = uma_zcreate("ktls_session",
395 	    sizeof(struct ktls_session),
396 	    NULL, NULL, NULL, NULL,
397 	    UMA_ALIGN_CACHE, 0);
398 
399 	if (ktls_sw_buffer_cache) {
400 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
401 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
402 		    ktls_buffer_import, ktls_buffer_release, NULL,
403 		    UMA_ZONE_FIRSTTOUCH);
404 	}
405 
406 	/*
407 	 * Initialize the workqueues to run the TLS work.  We create a
408 	 * work queue for each CPU.
409 	 */
410 	CPU_FOREACH(i) {
411 		STAILQ_INIT(&ktls_wq[i].m_head);
412 		STAILQ_INIT(&ktls_wq[i].so_head);
413 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
414 		if (ktls_bind_threads > 1) {
415 			pc = pcpu_find(i);
416 			domain = pc->pc_domain;
417 			count = ktls_domains[domain].count;
418 			ktls_domains[domain].cpu[count] = i;
419 			ktls_domains[domain].count++;
420 		}
421 		ktls_cpuid_lookup[ktls_number_threads] = i;
422 		ktls_number_threads++;
423 	}
424 
425 	/*
426 	 * If we somehow have an empty domain, fall back to choosing
427 	 * among all KTLS threads.
428 	 */
429 	if (ktls_bind_threads > 1) {
430 		for (i = 0; i < vm_ndomains; i++) {
431 			if (ktls_domains[i].count == 0) {
432 				ktls_bind_threads = 1;
433 				break;
434 			}
435 		}
436 	}
437 
438 	/* Start kthreads for each workqueue. */
439 	CPU_FOREACH(i) {
440 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
441 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
442 		if (error) {
443 			printf("Can't add KTLS thread %d error %d\n", i, error);
444 			return (error);
445 		}
446 	}
447 
448 	/*
449 	 * Start an allocation thread per-domain to perform blocking allocations
450 	 * of 16k physically contiguous TLS crypto destination buffers.
451 	 */
452 	if (ktls_sw_buffer_cache) {
453 		for (domain = 0; domain < vm_ndomains; domain++) {
454 			if (VM_DOMAIN_EMPTY(domain))
455 				continue;
456 			if (CPU_EMPTY(&cpuset_domain[domain]))
457 				continue;
458 			error = kproc_kthread_add(ktls_alloc_thread,
459 			    &ktls_domains[domain], &ktls_proc,
460 			    &ktls_domains[domain].alloc_td.td,
461 			    0, 0, "KTLS", "alloc_%d", domain);
462 			if (error) {
463 				printf("Can't add KTLS alloc thread %d error %d\n",
464 				    domain, error);
465 				return (error);
466 			}
467 		}
468 	}
469 
470 	if (bootverbose)
471 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
472 	return (0);
473 }
474 
475 static int
476 ktls_start_kthreads(void)
477 {
478 	int error, state;
479 
480 start:
481 	state = atomic_load_acq_int(&ktls_init_state);
482 	if (__predict_true(state > 0))
483 		return (0);
484 	if (state < 0)
485 		return (ENXIO);
486 
487 	sx_xlock(&ktls_init_lock);
488 	if (ktls_init_state != 0) {
489 		sx_xunlock(&ktls_init_lock);
490 		goto start;
491 	}
492 
493 	error = ktls_init();
494 	if (error == 0)
495 		state = 1;
496 	else
497 		state = -1;
498 	atomic_store_rel_int(&ktls_init_state, state);
499 	sx_xunlock(&ktls_init_lock);
500 	return (error);
501 }
502 
503 #if defined(INET) || defined(INET6)
504 static int
505 ktls_create_session(struct socket *so, struct tls_enable *en,
506     struct ktls_session **tlsp, int direction)
507 {
508 	struct ktls_session *tls;
509 	int error;
510 
511 	/* Only TLS 1.0 - 1.3 are supported. */
512 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
513 		return (EINVAL);
514 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
515 	    en->tls_vminor > TLS_MINOR_VER_THREE)
516 		return (EINVAL);
517 
518 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
519 		return (EINVAL);
520 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
521 		return (EINVAL);
522 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
523 		return (EINVAL);
524 
525 	/* All supported algorithms require a cipher key. */
526 	if (en->cipher_key_len == 0)
527 		return (EINVAL);
528 
529 	/* No flags are currently supported. */
530 	if (en->flags != 0)
531 		return (EINVAL);
532 
533 	/* Common checks for supported algorithms. */
534 	switch (en->cipher_algorithm) {
535 	case CRYPTO_AES_NIST_GCM_16:
536 		/*
537 		 * auth_algorithm isn't used, but permit GMAC values
538 		 * for compatibility.
539 		 */
540 		switch (en->auth_algorithm) {
541 		case 0:
542 #ifdef COMPAT_FREEBSD12
543 		/* XXX: Really 13.0-current COMPAT. */
544 		case CRYPTO_AES_128_NIST_GMAC:
545 		case CRYPTO_AES_192_NIST_GMAC:
546 		case CRYPTO_AES_256_NIST_GMAC:
547 #endif
548 			break;
549 		default:
550 			return (EINVAL);
551 		}
552 		if (en->auth_key_len != 0)
553 			return (EINVAL);
554 		switch (en->tls_vminor) {
555 		case TLS_MINOR_VER_TWO:
556 			if (en->iv_len != TLS_AEAD_GCM_LEN)
557 				return (EINVAL);
558 			break;
559 		case TLS_MINOR_VER_THREE:
560 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
561 				return (EINVAL);
562 			break;
563 		default:
564 			return (EINVAL);
565 		}
566 		break;
567 	case CRYPTO_AES_CBC:
568 		switch (en->auth_algorithm) {
569 		case CRYPTO_SHA1_HMAC:
570 			break;
571 		case CRYPTO_SHA2_256_HMAC:
572 		case CRYPTO_SHA2_384_HMAC:
573 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
574 				return (EINVAL);
575 			break;
576 		default:
577 			return (EINVAL);
578 		}
579 		if (en->auth_key_len == 0)
580 			return (EINVAL);
581 
582 		/*
583 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
584 		 * use explicit IVs.
585 		 */
586 		switch (en->tls_vminor) {
587 		case TLS_MINOR_VER_ZERO:
588 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
589 				return (EINVAL);
590 			break;
591 		case TLS_MINOR_VER_ONE:
592 		case TLS_MINOR_VER_TWO:
593 			/* Ignore any supplied IV. */
594 			en->iv_len = 0;
595 			break;
596 		default:
597 			return (EINVAL);
598 		}
599 		break;
600 	case CRYPTO_CHACHA20_POLY1305:
601 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
602 			return (EINVAL);
603 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
604 		    en->tls_vminor != TLS_MINOR_VER_THREE)
605 			return (EINVAL);
606 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
607 			return (EINVAL);
608 		break;
609 	default:
610 		return (EINVAL);
611 	}
612 
613 	error = ktls_start_kthreads();
614 	if (error != 0)
615 		return (error);
616 
617 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
618 
619 	counter_u64_add(ktls_offload_active, 1);
620 
621 	refcount_init(&tls->refcount, 1);
622 	if (direction == KTLS_RX)
623 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
624 	else
625 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
626 
627 	tls->wq_index = ktls_get_cpu(so);
628 
629 	tls->params.cipher_algorithm = en->cipher_algorithm;
630 	tls->params.auth_algorithm = en->auth_algorithm;
631 	tls->params.tls_vmajor = en->tls_vmajor;
632 	tls->params.tls_vminor = en->tls_vminor;
633 	tls->params.flags = en->flags;
634 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
635 
636 	/* Set the header and trailer lengths. */
637 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
638 	switch (en->cipher_algorithm) {
639 	case CRYPTO_AES_NIST_GCM_16:
640 		/*
641 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
642 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
643 		 */
644 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
645 			tls->params.tls_hlen += sizeof(uint64_t);
646 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
647 		tls->params.tls_bs = 1;
648 		break;
649 	case CRYPTO_AES_CBC:
650 		switch (en->auth_algorithm) {
651 		case CRYPTO_SHA1_HMAC:
652 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
653 				/* Implicit IV, no nonce. */
654 				tls->sequential_records = true;
655 				tls->next_seqno = be64dec(en->rec_seq);
656 				STAILQ_INIT(&tls->pending_records);
657 			} else {
658 				tls->params.tls_hlen += AES_BLOCK_LEN;
659 			}
660 			tls->params.tls_tlen = AES_BLOCK_LEN +
661 			    SHA1_HASH_LEN;
662 			break;
663 		case CRYPTO_SHA2_256_HMAC:
664 			tls->params.tls_hlen += AES_BLOCK_LEN;
665 			tls->params.tls_tlen = AES_BLOCK_LEN +
666 			    SHA2_256_HASH_LEN;
667 			break;
668 		case CRYPTO_SHA2_384_HMAC:
669 			tls->params.tls_hlen += AES_BLOCK_LEN;
670 			tls->params.tls_tlen = AES_BLOCK_LEN +
671 			    SHA2_384_HASH_LEN;
672 			break;
673 		default:
674 			panic("invalid hmac");
675 		}
676 		tls->params.tls_bs = AES_BLOCK_LEN;
677 		break;
678 	case CRYPTO_CHACHA20_POLY1305:
679 		/*
680 		 * Chacha20 uses a 12 byte implicit IV.
681 		 */
682 		tls->params.tls_tlen = POLY1305_HASH_LEN;
683 		tls->params.tls_bs = 1;
684 		break;
685 	default:
686 		panic("invalid cipher");
687 	}
688 
689 	/*
690 	 * TLS 1.3 includes optional padding which we do not support,
691 	 * and also puts the "real" record type at the end of the
692 	 * encrypted data.
693 	 */
694 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
695 		tls->params.tls_tlen += sizeof(uint8_t);
696 
697 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
698 	    ("TLS header length too long: %d", tls->params.tls_hlen));
699 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
700 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
701 
702 	if (en->auth_key_len != 0) {
703 		tls->params.auth_key_len = en->auth_key_len;
704 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
705 		    M_WAITOK);
706 		error = copyin(en->auth_key, tls->params.auth_key,
707 		    en->auth_key_len);
708 		if (error)
709 			goto out;
710 	}
711 
712 	tls->params.cipher_key_len = en->cipher_key_len;
713 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
714 	error = copyin(en->cipher_key, tls->params.cipher_key,
715 	    en->cipher_key_len);
716 	if (error)
717 		goto out;
718 
719 	/*
720 	 * This holds the implicit portion of the nonce for AEAD
721 	 * ciphers and the initial implicit IV for TLS 1.0.  The
722 	 * explicit portions of the IV are generated in ktls_frame().
723 	 */
724 	if (en->iv_len != 0) {
725 		tls->params.iv_len = en->iv_len;
726 		error = copyin(en->iv, tls->params.iv, en->iv_len);
727 		if (error)
728 			goto out;
729 
730 		/*
731 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
732 		 * counter to generate unique explicit IVs.
733 		 *
734 		 * Store this counter in the last 8 bytes of the IV
735 		 * array so that it is 8-byte aligned.
736 		 */
737 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
738 		    en->tls_vminor == TLS_MINOR_VER_TWO)
739 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
740 	}
741 
742 	*tlsp = tls;
743 	return (0);
744 
745 out:
746 	ktls_free(tls);
747 	return (error);
748 }
749 
750 static struct ktls_session *
751 ktls_clone_session(struct ktls_session *tls, int direction)
752 {
753 	struct ktls_session *tls_new;
754 
755 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
756 
757 	counter_u64_add(ktls_offload_active, 1);
758 
759 	refcount_init(&tls_new->refcount, 1);
760 	if (direction == KTLS_RX)
761 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
762 		    tls_new);
763 	else
764 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
765 		    tls_new);
766 
767 	/* Copy fields from existing session. */
768 	tls_new->params = tls->params;
769 	tls_new->wq_index = tls->wq_index;
770 
771 	/* Deep copy keys. */
772 	if (tls_new->params.auth_key != NULL) {
773 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
774 		    M_KTLS, M_WAITOK);
775 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
776 		    tls->params.auth_key_len);
777 	}
778 
779 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
780 	    M_WAITOK);
781 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
782 	    tls->params.cipher_key_len);
783 
784 	return (tls_new);
785 }
786 
787 #ifdef TCP_OFFLOAD
788 static int
789 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
790 {
791 	struct inpcb *inp;
792 	struct tcpcb *tp;
793 	int error;
794 
795 	inp = so->so_pcb;
796 	INP_WLOCK(inp);
797 	if (inp->inp_flags & INP_DROPPED) {
798 		INP_WUNLOCK(inp);
799 		return (ECONNRESET);
800 	}
801 	if (inp->inp_socket == NULL) {
802 		INP_WUNLOCK(inp);
803 		return (ECONNRESET);
804 	}
805 	tp = intotcpcb(inp);
806 	if (!(tp->t_flags & TF_TOE)) {
807 		INP_WUNLOCK(inp);
808 		return (EOPNOTSUPP);
809 	}
810 
811 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
812 	INP_WUNLOCK(inp);
813 	if (error == 0) {
814 		tls->mode = TCP_TLS_MODE_TOE;
815 		switch (tls->params.cipher_algorithm) {
816 		case CRYPTO_AES_CBC:
817 			counter_u64_add(ktls_toe_cbc, 1);
818 			break;
819 		case CRYPTO_AES_NIST_GCM_16:
820 			counter_u64_add(ktls_toe_gcm, 1);
821 			break;
822 		case CRYPTO_CHACHA20_POLY1305:
823 			counter_u64_add(ktls_toe_chacha20, 1);
824 			break;
825 		}
826 	}
827 	return (error);
828 }
829 #endif
830 
831 /*
832  * Common code used when first enabling ifnet TLS on a connection or
833  * when allocating a new ifnet TLS session due to a routing change.
834  * This function allocates a new TLS send tag on whatever interface
835  * the connection is currently routed over.
836  */
837 static int
838 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
839     struct m_snd_tag **mstp)
840 {
841 	union if_snd_tag_alloc_params params;
842 	struct ifnet *ifp;
843 	struct nhop_object *nh;
844 	struct tcpcb *tp;
845 	int error;
846 
847 	INP_RLOCK(inp);
848 	if (inp->inp_flags & INP_DROPPED) {
849 		INP_RUNLOCK(inp);
850 		return (ECONNRESET);
851 	}
852 	if (inp->inp_socket == NULL) {
853 		INP_RUNLOCK(inp);
854 		return (ECONNRESET);
855 	}
856 	tp = intotcpcb(inp);
857 
858 	/*
859 	 * Check administrative controls on ifnet TLS to determine if
860 	 * ifnet TLS should be denied.
861 	 *
862 	 * - Always permit 'force' requests.
863 	 * - ktls_ifnet_permitted == 0: always deny.
864 	 */
865 	if (!force && ktls_ifnet_permitted == 0) {
866 		INP_RUNLOCK(inp);
867 		return (ENXIO);
868 	}
869 
870 	/*
871 	 * XXX: Use the cached route in the inpcb to find the
872 	 * interface.  This should perhaps instead use
873 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
874 	 * enabled after a connection has completed key negotiation in
875 	 * userland, the cached route will be present in practice.
876 	 */
877 	nh = inp->inp_route.ro_nh;
878 	if (nh == NULL) {
879 		INP_RUNLOCK(inp);
880 		return (ENXIO);
881 	}
882 	ifp = nh->nh_ifp;
883 	if_ref(ifp);
884 
885 	/*
886 	 * Allocate a TLS + ratelimit tag if the connection has an
887 	 * existing pacing rate.
888 	 */
889 	if (tp->t_pacing_rate != -1 &&
890 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
891 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
892 		params.tls_rate_limit.inp = inp;
893 		params.tls_rate_limit.tls = tls;
894 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
895 	} else {
896 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
897 		params.tls.inp = inp;
898 		params.tls.tls = tls;
899 	}
900 	params.hdr.flowid = inp->inp_flowid;
901 	params.hdr.flowtype = inp->inp_flowtype;
902 	params.hdr.numa_domain = inp->inp_numa_domain;
903 	INP_RUNLOCK(inp);
904 
905 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
906 		error = EOPNOTSUPP;
907 		goto out;
908 	}
909 	if (inp->inp_vflag & INP_IPV6) {
910 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
911 			error = EOPNOTSUPP;
912 			goto out;
913 		}
914 	} else {
915 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
916 			error = EOPNOTSUPP;
917 			goto out;
918 		}
919 	}
920 	error = m_snd_tag_alloc(ifp, &params, mstp);
921 out:
922 	if_rele(ifp);
923 	return (error);
924 }
925 
926 /*
927  * Allocate an initial TLS receive tag for doing HW decryption of TLS
928  * data.
929  *
930  * This function allocates a new TLS receive tag on whatever interface
931  * the connection is currently routed over.  If the connection ends up
932  * using a different interface for receive this will get fixed up via
933  * ktls_input_ifp_mismatch as future packets arrive.
934  */
935 static int
936 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
937     struct m_snd_tag **mstp)
938 {
939 	union if_snd_tag_alloc_params params;
940 	struct ifnet *ifp;
941 	struct nhop_object *nh;
942 	int error;
943 
944 	if (!ktls_ocf_recrypt_supported(tls))
945 		return (ENXIO);
946 
947 	INP_RLOCK(inp);
948 	if (inp->inp_flags & INP_DROPPED) {
949 		INP_RUNLOCK(inp);
950 		return (ECONNRESET);
951 	}
952 	if (inp->inp_socket == NULL) {
953 		INP_RUNLOCK(inp);
954 		return (ECONNRESET);
955 	}
956 
957 	/*
958 	 * Check administrative controls on ifnet TLS to determine if
959 	 * ifnet TLS should be denied.
960 	 */
961 	if (ktls_ifnet_permitted == 0) {
962 		INP_RUNLOCK(inp);
963 		return (ENXIO);
964 	}
965 
966 	/*
967 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
968 	 * the inpcb to find the interface.
969 	 */
970 	nh = inp->inp_route.ro_nh;
971 	if (nh == NULL) {
972 		INP_RUNLOCK(inp);
973 		return (ENXIO);
974 	}
975 	ifp = nh->nh_ifp;
976 	if_ref(ifp);
977 	tls->rx_ifp = ifp;
978 
979 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
980 	params.hdr.flowid = inp->inp_flowid;
981 	params.hdr.flowtype = inp->inp_flowtype;
982 	params.hdr.numa_domain = inp->inp_numa_domain;
983 	params.tls_rx.inp = inp;
984 	params.tls_rx.tls = tls;
985 	params.tls_rx.vlan_id = 0;
986 
987 	INP_RUNLOCK(inp);
988 
989 	if (inp->inp_vflag & INP_IPV6) {
990 		if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) {
991 			error = EOPNOTSUPP;
992 			goto out;
993 		}
994 	} else {
995 		if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) {
996 			error = EOPNOTSUPP;
997 			goto out;
998 		}
999 	}
1000 	error = m_snd_tag_alloc(ifp, &params, mstp);
1001 
1002 	/*
1003 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1004 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1005 	 * ID for use in ktls_reset_receive_tag which allocates new
1006 	 * receive tags directly from the leaf interface bypassing
1007 	 * if_vlan.
1008 	 */
1009 	if (error == 0)
1010 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1011 out:
1012 	return (error);
1013 }
1014 
1015 static int
1016 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1017     bool force)
1018 {
1019 	struct m_snd_tag *mst;
1020 	int error;
1021 
1022 	switch (direction) {
1023 	case KTLS_TX:
1024 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1025 		if (__predict_false(error != 0))
1026 			goto done;
1027 		break;
1028 	case KTLS_RX:
1029 		KASSERT(!force, ("%s: forced receive tag", __func__));
1030 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1031 		if (__predict_false(error != 0))
1032 			goto done;
1033 		break;
1034 	default:
1035 		__assert_unreachable();
1036 	}
1037 
1038 	tls->mode = TCP_TLS_MODE_IFNET;
1039 	tls->snd_tag = mst;
1040 
1041 	switch (tls->params.cipher_algorithm) {
1042 	case CRYPTO_AES_CBC:
1043 		counter_u64_add(ktls_ifnet_cbc, 1);
1044 		break;
1045 	case CRYPTO_AES_NIST_GCM_16:
1046 		counter_u64_add(ktls_ifnet_gcm, 1);
1047 		break;
1048 	case CRYPTO_CHACHA20_POLY1305:
1049 		counter_u64_add(ktls_ifnet_chacha20, 1);
1050 		break;
1051 	default:
1052 		break;
1053 	}
1054 done:
1055 	return (error);
1056 }
1057 
1058 static void
1059 ktls_use_sw(struct ktls_session *tls)
1060 {
1061 	tls->mode = TCP_TLS_MODE_SW;
1062 	switch (tls->params.cipher_algorithm) {
1063 	case CRYPTO_AES_CBC:
1064 		counter_u64_add(ktls_sw_cbc, 1);
1065 		break;
1066 	case CRYPTO_AES_NIST_GCM_16:
1067 		counter_u64_add(ktls_sw_gcm, 1);
1068 		break;
1069 	case CRYPTO_CHACHA20_POLY1305:
1070 		counter_u64_add(ktls_sw_chacha20, 1);
1071 		break;
1072 	}
1073 }
1074 
1075 static int
1076 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1077 {
1078 	int error;
1079 
1080 	error = ktls_ocf_try(so, tls, direction);
1081 	if (error)
1082 		return (error);
1083 	ktls_use_sw(tls);
1084 	return (0);
1085 }
1086 
1087 /*
1088  * KTLS RX stores data in the socket buffer as a list of TLS records,
1089  * where each record is stored as a control message containg the TLS
1090  * header followed by data mbufs containing the decrypted data.  This
1091  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1092  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1093  * should be queued to the socket buffer as records, but encrypted
1094  * data which needs to be decrypted by software arrives as a stream of
1095  * regular mbufs which need to be converted.  In addition, there may
1096  * already be pending encrypted data in the socket buffer when KTLS RX
1097  * is enabled.
1098  *
1099  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1100  * is used:
1101  *
1102  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1103  *
1104  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1105  *   from the first mbuf.  Once all of the data for that TLS record is
1106  *   queued, the socket is queued to a worker thread.
1107  *
1108  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1109  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1110  *   decrypted, and inserted into the regular socket buffer chain as
1111  *   record starting with a control message holding the TLS header and
1112  *   a chain of mbufs holding the encrypted data.
1113  */
1114 
1115 static void
1116 sb_mark_notready(struct sockbuf *sb)
1117 {
1118 	struct mbuf *m;
1119 
1120 	m = sb->sb_mb;
1121 	sb->sb_mtls = m;
1122 	sb->sb_mb = NULL;
1123 	sb->sb_mbtail = NULL;
1124 	sb->sb_lastrecord = NULL;
1125 	for (; m != NULL; m = m->m_next) {
1126 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1127 		    __func__));
1128 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1129 		    __func__));
1130 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1131 		    __func__));
1132 		m->m_flags |= M_NOTREADY;
1133 		sb->sb_acc -= m->m_len;
1134 		sb->sb_tlscc += m->m_len;
1135 		sb->sb_mtlstail = m;
1136 	}
1137 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1138 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1139 	    sb->sb_ccc));
1140 }
1141 
1142 /*
1143  * Return information about the pending TLS data in a socket
1144  * buffer.  On return, 'seqno' is set to the sequence number
1145  * of the next TLS record to be received, 'resid' is set to
1146  * the amount of bytes still needed for the last pending
1147  * record.  The function returns 'false' if the last pending
1148  * record contains a partial TLS header.  In that case, 'resid'
1149  * is the number of bytes needed to complete the TLS header.
1150  */
1151 bool
1152 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1153 {
1154 	struct tls_record_layer hdr;
1155 	struct mbuf *m;
1156 	uint64_t seqno;
1157 	size_t resid;
1158 	u_int offset, record_len;
1159 
1160 	SOCKBUF_LOCK_ASSERT(sb);
1161 	MPASS(sb->sb_flags & SB_TLS_RX);
1162 	seqno = sb->sb_tls_seqno;
1163 	resid = sb->sb_tlscc;
1164 	m = sb->sb_mtls;
1165 	offset = 0;
1166 
1167 	if (resid == 0) {
1168 		*seqnop = seqno;
1169 		*residp = 0;
1170 		return (true);
1171 	}
1172 
1173 	for (;;) {
1174 		seqno++;
1175 
1176 		if (resid < sizeof(hdr)) {
1177 			*seqnop = seqno;
1178 			*residp = sizeof(hdr) - resid;
1179 			return (false);
1180 		}
1181 
1182 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1183 
1184 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1185 		if (resid <= record_len) {
1186 			*seqnop = seqno;
1187 			*residp = record_len - resid;
1188 			return (true);
1189 		}
1190 		resid -= record_len;
1191 
1192 		while (record_len != 0) {
1193 			if (m->m_len - offset > record_len) {
1194 				offset += record_len;
1195 				break;
1196 			}
1197 
1198 			record_len -= (m->m_len - offset);
1199 			offset = 0;
1200 			m = m->m_next;
1201 		}
1202 	}
1203 }
1204 
1205 int
1206 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1207 {
1208 	struct ktls_session *tls;
1209 	int error;
1210 
1211 	if (!ktls_offload_enable)
1212 		return (ENOTSUP);
1213 	if (SOLISTENING(so))
1214 		return (EINVAL);
1215 
1216 	counter_u64_add(ktls_offload_enable_calls, 1);
1217 
1218 	/*
1219 	 * This should always be true since only the TCP socket option
1220 	 * invokes this function.
1221 	 */
1222 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1223 		return (EINVAL);
1224 
1225 	/*
1226 	 * XXX: Don't overwrite existing sessions.  We should permit
1227 	 * this to support rekeying in the future.
1228 	 */
1229 	if (so->so_rcv.sb_tls_info != NULL)
1230 		return (EALREADY);
1231 
1232 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1233 		return (ENOTSUP);
1234 
1235 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1236 	if (error)
1237 		return (error);
1238 
1239 	error = ktls_ocf_try(so, tls, KTLS_RX);
1240 	if (error) {
1241 		ktls_free(tls);
1242 		return (error);
1243 	}
1244 
1245 	/* Mark the socket as using TLS offload. */
1246 	SOCKBUF_LOCK(&so->so_rcv);
1247 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1248 	so->so_rcv.sb_tls_info = tls;
1249 	so->so_rcv.sb_flags |= SB_TLS_RX;
1250 
1251 	/* Mark existing data as not ready until it can be decrypted. */
1252 	sb_mark_notready(&so->so_rcv);
1253 	ktls_check_rx(&so->so_rcv);
1254 	SOCKBUF_UNLOCK(&so->so_rcv);
1255 
1256 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1257 #ifdef TCP_OFFLOAD
1258 	error = ktls_try_toe(so, tls, KTLS_RX);
1259 	if (error)
1260 #endif
1261 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1262 	if (error)
1263 		ktls_use_sw(tls);
1264 
1265 	counter_u64_add(ktls_offload_total, 1);
1266 
1267 	return (0);
1268 }
1269 
1270 int
1271 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1272 {
1273 	struct ktls_session *tls;
1274 	struct inpcb *inp;
1275 	int error;
1276 
1277 	if (!ktls_offload_enable)
1278 		return (ENOTSUP);
1279 	if (SOLISTENING(so))
1280 		return (EINVAL);
1281 
1282 	counter_u64_add(ktls_offload_enable_calls, 1);
1283 
1284 	/*
1285 	 * This should always be true since only the TCP socket option
1286 	 * invokes this function.
1287 	 */
1288 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1289 		return (EINVAL);
1290 
1291 	/*
1292 	 * XXX: Don't overwrite existing sessions.  We should permit
1293 	 * this to support rekeying in the future.
1294 	 */
1295 	if (so->so_snd.sb_tls_info != NULL)
1296 		return (EALREADY);
1297 
1298 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1299 		return (ENOTSUP);
1300 
1301 	/* TLS requires ext pgs */
1302 	if (mb_use_ext_pgs == 0)
1303 		return (ENXIO);
1304 
1305 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1306 	if (error)
1307 		return (error);
1308 
1309 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1310 #ifdef TCP_OFFLOAD
1311 	error = ktls_try_toe(so, tls, KTLS_TX);
1312 	if (error)
1313 #endif
1314 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1315 	if (error)
1316 		error = ktls_try_sw(so, tls, KTLS_TX);
1317 
1318 	if (error) {
1319 		ktls_free(tls);
1320 		return (error);
1321 	}
1322 
1323 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1324 	if (error) {
1325 		ktls_free(tls);
1326 		return (error);
1327 	}
1328 
1329 	/*
1330 	 * Write lock the INP when setting sb_tls_info so that
1331 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1332 	 * holding the INP lock.
1333 	 */
1334 	inp = so->so_pcb;
1335 	INP_WLOCK(inp);
1336 	SOCKBUF_LOCK(&so->so_snd);
1337 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1338 	so->so_snd.sb_tls_info = tls;
1339 	if (tls->mode != TCP_TLS_MODE_SW)
1340 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1341 	SOCKBUF_UNLOCK(&so->so_snd);
1342 	INP_WUNLOCK(inp);
1343 	SOCK_IO_SEND_UNLOCK(so);
1344 
1345 	counter_u64_add(ktls_offload_total, 1);
1346 
1347 	return (0);
1348 }
1349 
1350 int
1351 ktls_get_rx_mode(struct socket *so, int *modep)
1352 {
1353 	struct ktls_session *tls;
1354 	struct inpcb *inp __diagused;
1355 
1356 	if (SOLISTENING(so))
1357 		return (EINVAL);
1358 	inp = so->so_pcb;
1359 	INP_WLOCK_ASSERT(inp);
1360 	SOCK_RECVBUF_LOCK(so);
1361 	tls = so->so_rcv.sb_tls_info;
1362 	if (tls == NULL)
1363 		*modep = TCP_TLS_MODE_NONE;
1364 	else
1365 		*modep = tls->mode;
1366 	SOCK_RECVBUF_UNLOCK(so);
1367 	return (0);
1368 }
1369 
1370 /*
1371  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1372  *
1373  * This function gets information about the next TCP- and TLS-
1374  * sequence number to be processed by the TLS receive worker
1375  * thread. The information is extracted from the given "inpcb"
1376  * structure. The values are stored in host endian format at the two
1377  * given output pointer locations. The TCP sequence number points to
1378  * the beginning of the TLS header.
1379  *
1380  * This function returns zero on success, else a non-zero error code
1381  * is returned.
1382  */
1383 int
1384 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1385 {
1386 	struct socket *so;
1387 	struct tcpcb *tp;
1388 
1389 	INP_RLOCK(inp);
1390 	so = inp->inp_socket;
1391 	if (__predict_false(so == NULL)) {
1392 		INP_RUNLOCK(inp);
1393 		return (EINVAL);
1394 	}
1395 	if (inp->inp_flags & INP_DROPPED) {
1396 		INP_RUNLOCK(inp);
1397 		return (ECONNRESET);
1398 	}
1399 
1400 	tp = intotcpcb(inp);
1401 	MPASS(tp != NULL);
1402 
1403 	SOCKBUF_LOCK(&so->so_rcv);
1404 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1405 	*tlsseq = so->so_rcv.sb_tls_seqno;
1406 	SOCKBUF_UNLOCK(&so->so_rcv);
1407 
1408 	INP_RUNLOCK(inp);
1409 
1410 	return (0);
1411 }
1412 
1413 int
1414 ktls_get_tx_mode(struct socket *so, int *modep)
1415 {
1416 	struct ktls_session *tls;
1417 	struct inpcb *inp __diagused;
1418 
1419 	if (SOLISTENING(so))
1420 		return (EINVAL);
1421 	inp = so->so_pcb;
1422 	INP_WLOCK_ASSERT(inp);
1423 	SOCK_SENDBUF_LOCK(so);
1424 	tls = so->so_snd.sb_tls_info;
1425 	if (tls == NULL)
1426 		*modep = TCP_TLS_MODE_NONE;
1427 	else
1428 		*modep = tls->mode;
1429 	SOCK_SENDBUF_UNLOCK(so);
1430 	return (0);
1431 }
1432 
1433 /*
1434  * Switch between SW and ifnet TLS sessions as requested.
1435  */
1436 int
1437 ktls_set_tx_mode(struct socket *so, int mode)
1438 {
1439 	struct ktls_session *tls, *tls_new;
1440 	struct inpcb *inp;
1441 	int error;
1442 
1443 	if (SOLISTENING(so))
1444 		return (EINVAL);
1445 	switch (mode) {
1446 	case TCP_TLS_MODE_SW:
1447 	case TCP_TLS_MODE_IFNET:
1448 		break;
1449 	default:
1450 		return (EINVAL);
1451 	}
1452 
1453 	inp = so->so_pcb;
1454 	INP_WLOCK_ASSERT(inp);
1455 	SOCKBUF_LOCK(&so->so_snd);
1456 	tls = so->so_snd.sb_tls_info;
1457 	if (tls == NULL) {
1458 		SOCKBUF_UNLOCK(&so->so_snd);
1459 		return (0);
1460 	}
1461 
1462 	if (tls->mode == mode) {
1463 		SOCKBUF_UNLOCK(&so->so_snd);
1464 		return (0);
1465 	}
1466 
1467 	tls = ktls_hold(tls);
1468 	SOCKBUF_UNLOCK(&so->so_snd);
1469 	INP_WUNLOCK(inp);
1470 
1471 	tls_new = ktls_clone_session(tls, KTLS_TX);
1472 
1473 	if (mode == TCP_TLS_MODE_IFNET)
1474 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1475 	else
1476 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1477 	if (error) {
1478 		counter_u64_add(ktls_switch_failed, 1);
1479 		ktls_free(tls_new);
1480 		ktls_free(tls);
1481 		INP_WLOCK(inp);
1482 		return (error);
1483 	}
1484 
1485 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1486 	if (error) {
1487 		counter_u64_add(ktls_switch_failed, 1);
1488 		ktls_free(tls_new);
1489 		ktls_free(tls);
1490 		INP_WLOCK(inp);
1491 		return (error);
1492 	}
1493 
1494 	/*
1495 	 * If we raced with another session change, keep the existing
1496 	 * session.
1497 	 */
1498 	if (tls != so->so_snd.sb_tls_info) {
1499 		counter_u64_add(ktls_switch_failed, 1);
1500 		SOCK_IO_SEND_UNLOCK(so);
1501 		ktls_free(tls_new);
1502 		ktls_free(tls);
1503 		INP_WLOCK(inp);
1504 		return (EBUSY);
1505 	}
1506 
1507 	INP_WLOCK(inp);
1508 	SOCKBUF_LOCK(&so->so_snd);
1509 	so->so_snd.sb_tls_info = tls_new;
1510 	if (tls_new->mode != TCP_TLS_MODE_SW)
1511 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1512 	SOCKBUF_UNLOCK(&so->so_snd);
1513 	SOCK_IO_SEND_UNLOCK(so);
1514 
1515 	/*
1516 	 * Drop two references on 'tls'.  The first is for the
1517 	 * ktls_hold() above.  The second drops the reference from the
1518 	 * socket buffer.
1519 	 */
1520 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1521 	ktls_free(tls);
1522 	ktls_free(tls);
1523 
1524 	if (mode == TCP_TLS_MODE_IFNET)
1525 		counter_u64_add(ktls_switch_to_ifnet, 1);
1526 	else
1527 		counter_u64_add(ktls_switch_to_sw, 1);
1528 
1529 	return (0);
1530 }
1531 
1532 /*
1533  * Try to allocate a new TLS receive tag.  This task is scheduled when
1534  * sbappend_ktls_rx detects an input path change.  If a new tag is
1535  * allocated, replace the tag in the TLS session.  If a new tag cannot
1536  * be allocated, let the session fall back to software decryption.
1537  */
1538 static void
1539 ktls_reset_receive_tag(void *context, int pending)
1540 {
1541 	union if_snd_tag_alloc_params params;
1542 	struct ktls_session *tls;
1543 	struct m_snd_tag *mst;
1544 	struct inpcb *inp;
1545 	struct ifnet *ifp;
1546 	struct socket *so;
1547 	int error;
1548 
1549 	MPASS(pending == 1);
1550 
1551 	tls = context;
1552 	so = tls->so;
1553 	inp = so->so_pcb;
1554 	ifp = NULL;
1555 
1556 	INP_RLOCK(inp);
1557 	if (inp->inp_flags & INP_DROPPED) {
1558 		INP_RUNLOCK(inp);
1559 		goto out;
1560 	}
1561 
1562 	SOCKBUF_LOCK(&so->so_rcv);
1563 	mst = tls->snd_tag;
1564 	tls->snd_tag = NULL;
1565 	if (mst != NULL)
1566 		m_snd_tag_rele(mst);
1567 
1568 	ifp = tls->rx_ifp;
1569 	if_ref(ifp);
1570 	SOCKBUF_UNLOCK(&so->so_rcv);
1571 
1572 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1573 	params.hdr.flowid = inp->inp_flowid;
1574 	params.hdr.flowtype = inp->inp_flowtype;
1575 	params.hdr.numa_domain = inp->inp_numa_domain;
1576 	params.tls_rx.inp = inp;
1577 	params.tls_rx.tls = tls;
1578 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1579 	INP_RUNLOCK(inp);
1580 
1581 	if (inp->inp_vflag & INP_IPV6) {
1582 		if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0)
1583 			goto out;
1584 	} else {
1585 		if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0)
1586 			goto out;
1587 	}
1588 
1589 	error = m_snd_tag_alloc(ifp, &params, &mst);
1590 	if (error == 0) {
1591 		SOCKBUF_LOCK(&so->so_rcv);
1592 		tls->snd_tag = mst;
1593 		SOCKBUF_UNLOCK(&so->so_rcv);
1594 
1595 		counter_u64_add(ktls_ifnet_reset, 1);
1596 	} else {
1597 		/*
1598 		 * Just fall back to software decryption if a tag
1599 		 * cannot be allocated leaving the connection intact.
1600 		 * If a future input path change switches to another
1601 		 * interface this connection will resume ifnet TLS.
1602 		 */
1603 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1604 	}
1605 
1606 out:
1607 	mtx_pool_lock(mtxpool_sleep, tls);
1608 	tls->reset_pending = false;
1609 	mtx_pool_unlock(mtxpool_sleep, tls);
1610 
1611 	if (ifp != NULL)
1612 		if_rele(ifp);
1613 	sorele(so);
1614 	ktls_free(tls);
1615 }
1616 
1617 /*
1618  * Try to allocate a new TLS send tag.  This task is scheduled when
1619  * ip_output detects a route change while trying to transmit a packet
1620  * holding a TLS record.  If a new tag is allocated, replace the tag
1621  * in the TLS session.  Subsequent packets on the connection will use
1622  * the new tag.  If a new tag cannot be allocated, drop the
1623  * connection.
1624  */
1625 static void
1626 ktls_reset_send_tag(void *context, int pending)
1627 {
1628 	struct epoch_tracker et;
1629 	struct ktls_session *tls;
1630 	struct m_snd_tag *old, *new;
1631 	struct inpcb *inp;
1632 	struct tcpcb *tp;
1633 	int error;
1634 
1635 	MPASS(pending == 1);
1636 
1637 	tls = context;
1638 	inp = tls->inp;
1639 
1640 	/*
1641 	 * Free the old tag first before allocating a new one.
1642 	 * ip[6]_output_send() will treat a NULL send tag the same as
1643 	 * an ifp mismatch and drop packets until a new tag is
1644 	 * allocated.
1645 	 *
1646 	 * Write-lock the INP when changing tls->snd_tag since
1647 	 * ip[6]_output_send() holds a read-lock when reading the
1648 	 * pointer.
1649 	 */
1650 	INP_WLOCK(inp);
1651 	old = tls->snd_tag;
1652 	tls->snd_tag = NULL;
1653 	INP_WUNLOCK(inp);
1654 	if (old != NULL)
1655 		m_snd_tag_rele(old);
1656 
1657 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1658 
1659 	if (error == 0) {
1660 		INP_WLOCK(inp);
1661 		tls->snd_tag = new;
1662 		mtx_pool_lock(mtxpool_sleep, tls);
1663 		tls->reset_pending = false;
1664 		mtx_pool_unlock(mtxpool_sleep, tls);
1665 		if (!in_pcbrele_wlocked(inp))
1666 			INP_WUNLOCK(inp);
1667 
1668 		counter_u64_add(ktls_ifnet_reset, 1);
1669 
1670 		/*
1671 		 * XXX: Should we kick tcp_output explicitly now that
1672 		 * the send tag is fixed or just rely on timers?
1673 		 */
1674 	} else {
1675 		NET_EPOCH_ENTER(et);
1676 		INP_WLOCK(inp);
1677 		if (!in_pcbrele_wlocked(inp)) {
1678 			if (!(inp->inp_flags & INP_DROPPED)) {
1679 				tp = intotcpcb(inp);
1680 				CURVNET_SET(inp->inp_vnet);
1681 				tp = tcp_drop(tp, ECONNABORTED);
1682 				CURVNET_RESTORE();
1683 				if (tp != NULL)
1684 					INP_WUNLOCK(inp);
1685 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1686 			} else
1687 				INP_WUNLOCK(inp);
1688 		}
1689 		NET_EPOCH_EXIT(et);
1690 
1691 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1692 
1693 		/*
1694 		 * Leave reset_pending true to avoid future tasks while
1695 		 * the socket goes away.
1696 		 */
1697 	}
1698 
1699 	ktls_free(tls);
1700 }
1701 
1702 void
1703 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1704 {
1705 	struct ktls_session *tls;
1706 	struct socket *so;
1707 
1708 	SOCKBUF_LOCK_ASSERT(sb);
1709 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1710 	    __func__, sb));
1711 	so = __containerof(sb, struct socket, so_rcv);
1712 
1713 	tls = sb->sb_tls_info;
1714 	if_rele(tls->rx_ifp);
1715 	if_ref(ifp);
1716 	tls->rx_ifp = ifp;
1717 
1718 	/*
1719 	 * See if we should schedule a task to update the receive tag for
1720 	 * this session.
1721 	 */
1722 	mtx_pool_lock(mtxpool_sleep, tls);
1723 	if (!tls->reset_pending) {
1724 		(void) ktls_hold(tls);
1725 		soref(so);
1726 		tls->so = so;
1727 		tls->reset_pending = true;
1728 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1729 	}
1730 	mtx_pool_unlock(mtxpool_sleep, tls);
1731 }
1732 
1733 int
1734 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1735 {
1736 
1737 	if (inp == NULL)
1738 		return (ENOBUFS);
1739 
1740 	INP_LOCK_ASSERT(inp);
1741 
1742 	/*
1743 	 * See if we should schedule a task to update the send tag for
1744 	 * this session.
1745 	 */
1746 	mtx_pool_lock(mtxpool_sleep, tls);
1747 	if (!tls->reset_pending) {
1748 		(void) ktls_hold(tls);
1749 		in_pcbref(inp);
1750 		tls->inp = inp;
1751 		tls->reset_pending = true;
1752 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1753 	}
1754 	mtx_pool_unlock(mtxpool_sleep, tls);
1755 	return (ENOBUFS);
1756 }
1757 
1758 #ifdef RATELIMIT
1759 int
1760 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1761 {
1762 	union if_snd_tag_modify_params params = {
1763 		.rate_limit.max_rate = max_pacing_rate,
1764 		.rate_limit.flags = M_NOWAIT,
1765 	};
1766 	struct m_snd_tag *mst;
1767 
1768 	/* Can't get to the inp, but it should be locked. */
1769 	/* INP_LOCK_ASSERT(inp); */
1770 
1771 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1772 
1773 	if (tls->snd_tag == NULL) {
1774 		/*
1775 		 * Resetting send tag, ignore this change.  The
1776 		 * pending reset may or may not see this updated rate
1777 		 * in the tcpcb.  If it doesn't, we will just lose
1778 		 * this rate change.
1779 		 */
1780 		return (0);
1781 	}
1782 
1783 	mst = tls->snd_tag;
1784 
1785 	MPASS(mst != NULL);
1786 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1787 
1788 	return (mst->sw->snd_tag_modify(mst, &params));
1789 }
1790 #endif
1791 #endif
1792 
1793 void
1794 ktls_destroy(struct ktls_session *tls)
1795 {
1796 	MPASS(tls->refcount == 0);
1797 
1798 	if (tls->sequential_records) {
1799 		struct mbuf *m, *n;
1800 		int page_count;
1801 
1802 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1803 			page_count = m->m_epg_enc_cnt;
1804 			while (page_count > 0) {
1805 				KASSERT(page_count >= m->m_epg_nrdy,
1806 				    ("%s: too few pages", __func__));
1807 				page_count -= m->m_epg_nrdy;
1808 				m = m_free(m);
1809 			}
1810 		}
1811 	}
1812 
1813 	counter_u64_add(ktls_offload_active, -1);
1814 	switch (tls->mode) {
1815 	case TCP_TLS_MODE_SW:
1816 		switch (tls->params.cipher_algorithm) {
1817 		case CRYPTO_AES_CBC:
1818 			counter_u64_add(ktls_sw_cbc, -1);
1819 			break;
1820 		case CRYPTO_AES_NIST_GCM_16:
1821 			counter_u64_add(ktls_sw_gcm, -1);
1822 			break;
1823 		case CRYPTO_CHACHA20_POLY1305:
1824 			counter_u64_add(ktls_sw_chacha20, -1);
1825 			break;
1826 		}
1827 		break;
1828 	case TCP_TLS_MODE_IFNET:
1829 		switch (tls->params.cipher_algorithm) {
1830 		case CRYPTO_AES_CBC:
1831 			counter_u64_add(ktls_ifnet_cbc, -1);
1832 			break;
1833 		case CRYPTO_AES_NIST_GCM_16:
1834 			counter_u64_add(ktls_ifnet_gcm, -1);
1835 			break;
1836 		case CRYPTO_CHACHA20_POLY1305:
1837 			counter_u64_add(ktls_ifnet_chacha20, -1);
1838 			break;
1839 		}
1840 		if (tls->snd_tag != NULL)
1841 			m_snd_tag_rele(tls->snd_tag);
1842 		if (tls->rx_ifp != NULL)
1843 			if_rele(tls->rx_ifp);
1844 		break;
1845 #ifdef TCP_OFFLOAD
1846 	case TCP_TLS_MODE_TOE:
1847 		switch (tls->params.cipher_algorithm) {
1848 		case CRYPTO_AES_CBC:
1849 			counter_u64_add(ktls_toe_cbc, -1);
1850 			break;
1851 		case CRYPTO_AES_NIST_GCM_16:
1852 			counter_u64_add(ktls_toe_gcm, -1);
1853 			break;
1854 		case CRYPTO_CHACHA20_POLY1305:
1855 			counter_u64_add(ktls_toe_chacha20, -1);
1856 			break;
1857 		}
1858 		break;
1859 #endif
1860 	}
1861 	if (tls->ocf_session != NULL)
1862 		ktls_ocf_free(tls);
1863 	if (tls->params.auth_key != NULL) {
1864 		zfree(tls->params.auth_key, M_KTLS);
1865 		tls->params.auth_key = NULL;
1866 		tls->params.auth_key_len = 0;
1867 	}
1868 	if (tls->params.cipher_key != NULL) {
1869 		zfree(tls->params.cipher_key, M_KTLS);
1870 		tls->params.cipher_key = NULL;
1871 		tls->params.cipher_key_len = 0;
1872 	}
1873 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1874 
1875 	uma_zfree(ktls_session_zone, tls);
1876 }
1877 
1878 void
1879 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1880 {
1881 
1882 	for (; m != NULL; m = m->m_next) {
1883 		KASSERT((m->m_flags & M_EXTPG) != 0,
1884 		    ("ktls_seq: mapped mbuf %p", m));
1885 
1886 		m->m_epg_seqno = sb->sb_tls_seqno;
1887 		sb->sb_tls_seqno++;
1888 	}
1889 }
1890 
1891 /*
1892  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1893  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1894  * mbuf must be populated with the payload of each TLS record.
1895  *
1896  * The record_type argument specifies the TLS record type used when
1897  * populating the TLS header.
1898  *
1899  * The enq_count argument on return is set to the number of pages of
1900  * payload data for this entire chain that need to be encrypted via SW
1901  * encryption.  The returned value should be passed to ktls_enqueue
1902  * when scheduling encryption of this chain of mbufs.  To handle the
1903  * special case of empty fragments for TLS 1.0 sessions, an empty
1904  * fragment counts as one page.
1905  */
1906 void
1907 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1908     uint8_t record_type)
1909 {
1910 	struct tls_record_layer *tlshdr;
1911 	struct mbuf *m;
1912 	uint64_t *noncep;
1913 	uint16_t tls_len;
1914 	int maxlen __diagused;
1915 
1916 	maxlen = tls->params.max_frame_len;
1917 	*enq_cnt = 0;
1918 	for (m = top; m != NULL; m = m->m_next) {
1919 		/*
1920 		 * All mbufs in the chain should be TLS records whose
1921 		 * payload does not exceed the maximum frame length.
1922 		 *
1923 		 * Empty TLS 1.0 records are permitted when using CBC.
1924 		 */
1925 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1926 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1927 		    ("ktls_frame: m %p len %d", m, m->m_len));
1928 
1929 		/*
1930 		 * TLS frames require unmapped mbufs to store session
1931 		 * info.
1932 		 */
1933 		KASSERT((m->m_flags & M_EXTPG) != 0,
1934 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1935 
1936 		tls_len = m->m_len;
1937 
1938 		/* Save a reference to the session. */
1939 		m->m_epg_tls = ktls_hold(tls);
1940 
1941 		m->m_epg_hdrlen = tls->params.tls_hlen;
1942 		m->m_epg_trllen = tls->params.tls_tlen;
1943 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1944 			int bs, delta;
1945 
1946 			/*
1947 			 * AES-CBC pads messages to a multiple of the
1948 			 * block size.  Note that the padding is
1949 			 * applied after the digest and the encryption
1950 			 * is done on the "plaintext || mac || padding".
1951 			 * At least one byte of padding is always
1952 			 * present.
1953 			 *
1954 			 * Compute the final trailer length assuming
1955 			 * at most one block of padding.
1956 			 * tls->params.tls_tlen is the maximum
1957 			 * possible trailer length (padding + digest).
1958 			 * delta holds the number of excess padding
1959 			 * bytes if the maximum were used.  Those
1960 			 * extra bytes are removed.
1961 			 */
1962 			bs = tls->params.tls_bs;
1963 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1964 			m->m_epg_trllen -= delta;
1965 		}
1966 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1967 
1968 		/* Populate the TLS header. */
1969 		tlshdr = (void *)m->m_epg_hdr;
1970 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1971 
1972 		/*
1973 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1974 		 * of TLS_RLTYPE_APP.
1975 		 */
1976 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1977 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1978 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1979 			tlshdr->tls_type = TLS_RLTYPE_APP;
1980 			/* save the real record type for later */
1981 			m->m_epg_record_type = record_type;
1982 			m->m_epg_trail[0] = record_type;
1983 		} else {
1984 			tlshdr->tls_vminor = tls->params.tls_vminor;
1985 			tlshdr->tls_type = record_type;
1986 		}
1987 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1988 
1989 		/*
1990 		 * Store nonces / explicit IVs after the end of the
1991 		 * TLS header.
1992 		 *
1993 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1994 		 * from the end of the IV.  The nonce is then
1995 		 * incremented for use by the next record.
1996 		 *
1997 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1998 		 */
1999 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2000 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2001 			noncep = (uint64_t *)(tls->params.iv + 8);
2002 			be64enc(tlshdr + 1, *noncep);
2003 			(*noncep)++;
2004 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2005 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2006 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2007 
2008 		/*
2009 		 * When using SW encryption, mark the mbuf not ready.
2010 		 * It will be marked ready via sbready() after the
2011 		 * record has been encrypted.
2012 		 *
2013 		 * When using ifnet TLS, unencrypted TLS records are
2014 		 * sent down the stack to the NIC.
2015 		 */
2016 		if (tls->mode == TCP_TLS_MODE_SW) {
2017 			m->m_flags |= M_NOTREADY;
2018 			if (__predict_false(tls_len == 0)) {
2019 				/* TLS 1.0 empty fragment. */
2020 				m->m_epg_nrdy = 1;
2021 			} else
2022 				m->m_epg_nrdy = m->m_epg_npgs;
2023 			*enq_cnt += m->m_epg_nrdy;
2024 		}
2025 	}
2026 }
2027 
2028 bool
2029 ktls_permit_empty_frames(struct ktls_session *tls)
2030 {
2031 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2032 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2033 }
2034 
2035 void
2036 ktls_check_rx(struct sockbuf *sb)
2037 {
2038 	struct tls_record_layer hdr;
2039 	struct ktls_wq *wq;
2040 	struct socket *so;
2041 	bool running;
2042 
2043 	SOCKBUF_LOCK_ASSERT(sb);
2044 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2045 	    __func__, sb));
2046 	so = __containerof(sb, struct socket, so_rcv);
2047 
2048 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2049 		return;
2050 
2051 	/* Is there enough queued for a TLS header? */
2052 	if (sb->sb_tlscc < sizeof(hdr)) {
2053 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2054 			so->so_error = EMSGSIZE;
2055 		return;
2056 	}
2057 
2058 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2059 
2060 	/* Is the entire record queued? */
2061 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2062 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2063 			so->so_error = EMSGSIZE;
2064 		return;
2065 	}
2066 
2067 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2068 
2069 	soref(so);
2070 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2071 	mtx_lock(&wq->mtx);
2072 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2073 	running = wq->running;
2074 	mtx_unlock(&wq->mtx);
2075 	if (!running)
2076 		wakeup(wq);
2077 	counter_u64_add(ktls_cnt_rx_queued, 1);
2078 }
2079 
2080 static struct mbuf *
2081 ktls_detach_record(struct sockbuf *sb, int len)
2082 {
2083 	struct mbuf *m, *n, *top;
2084 	int remain;
2085 
2086 	SOCKBUF_LOCK_ASSERT(sb);
2087 	MPASS(len <= sb->sb_tlscc);
2088 
2089 	/*
2090 	 * If TLS chain is the exact size of the record,
2091 	 * just grab the whole record.
2092 	 */
2093 	top = sb->sb_mtls;
2094 	if (sb->sb_tlscc == len) {
2095 		sb->sb_mtls = NULL;
2096 		sb->sb_mtlstail = NULL;
2097 		goto out;
2098 	}
2099 
2100 	/*
2101 	 * While it would be nice to use m_split() here, we need
2102 	 * to know exactly what m_split() allocates to update the
2103 	 * accounting, so do it inline instead.
2104 	 */
2105 	remain = len;
2106 	for (m = top; remain > m->m_len; m = m->m_next)
2107 		remain -= m->m_len;
2108 
2109 	/* Easy case: don't have to split 'm'. */
2110 	if (remain == m->m_len) {
2111 		sb->sb_mtls = m->m_next;
2112 		if (sb->sb_mtls == NULL)
2113 			sb->sb_mtlstail = NULL;
2114 		m->m_next = NULL;
2115 		goto out;
2116 	}
2117 
2118 	/*
2119 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2120 	 * with M_NOWAIT first.
2121 	 */
2122 	n = m_get(M_NOWAIT, MT_DATA);
2123 	if (n == NULL) {
2124 		/*
2125 		 * Use M_WAITOK with socket buffer unlocked.  If
2126 		 * 'sb_mtls' changes while the lock is dropped, return
2127 		 * NULL to force the caller to retry.
2128 		 */
2129 		SOCKBUF_UNLOCK(sb);
2130 
2131 		n = m_get(M_WAITOK, MT_DATA);
2132 
2133 		SOCKBUF_LOCK(sb);
2134 		if (sb->sb_mtls != top) {
2135 			m_free(n);
2136 			return (NULL);
2137 		}
2138 	}
2139 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2140 
2141 	/* Store remainder in 'n'. */
2142 	n->m_len = m->m_len - remain;
2143 	if (m->m_flags & M_EXT) {
2144 		n->m_data = m->m_data + remain;
2145 		mb_dupcl(n, m);
2146 	} else {
2147 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2148 	}
2149 
2150 	/* Trim 'm' and update accounting. */
2151 	m->m_len -= n->m_len;
2152 	sb->sb_tlscc -= n->m_len;
2153 	sb->sb_ccc -= n->m_len;
2154 
2155 	/* Account for 'n'. */
2156 	sballoc_ktls_rx(sb, n);
2157 
2158 	/* Insert 'n' into the TLS chain. */
2159 	sb->sb_mtls = n;
2160 	n->m_next = m->m_next;
2161 	if (sb->sb_mtlstail == m)
2162 		sb->sb_mtlstail = n;
2163 
2164 	/* Detach the record from the TLS chain. */
2165 	m->m_next = NULL;
2166 
2167 out:
2168 	MPASS(m_length(top, NULL) == len);
2169 	for (m = top; m != NULL; m = m->m_next)
2170 		sbfree_ktls_rx(sb, m);
2171 	sb->sb_tlsdcc = len;
2172 	sb->sb_ccc += len;
2173 	SBCHECK(sb);
2174 	return (top);
2175 }
2176 
2177 /*
2178  * Determine the length of the trailing zero padding and find the real
2179  * record type in the byte before the padding.
2180  *
2181  * Walking the mbuf chain backwards is clumsy, so another option would
2182  * be to scan forwards remembering the last non-zero byte before the
2183  * trailer.  However, it would be expensive to scan the entire record.
2184  * Instead, find the last non-zero byte of each mbuf in the chain
2185  * keeping track of the relative offset of that nonzero byte.
2186  *
2187  * trail_len is the size of the MAC/tag on input and is set to the
2188  * size of the full trailer including padding and the record type on
2189  * return.
2190  */
2191 static int
2192 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2193     int *trailer_len, uint8_t *record_typep)
2194 {
2195 	char *cp;
2196 	u_int digest_start, last_offset, m_len, offset;
2197 	uint8_t record_type;
2198 
2199 	digest_start = tls_len - *trailer_len;
2200 	last_offset = 0;
2201 	offset = 0;
2202 	for (; m != NULL && offset < digest_start;
2203 	     offset += m->m_len, m = m->m_next) {
2204 		/* Don't look for padding in the tag. */
2205 		m_len = min(digest_start - offset, m->m_len);
2206 		cp = mtod(m, char *);
2207 
2208 		/* Find last non-zero byte in this mbuf. */
2209 		while (m_len > 0 && cp[m_len - 1] == 0)
2210 			m_len--;
2211 		if (m_len > 0) {
2212 			record_type = cp[m_len - 1];
2213 			last_offset = offset + m_len;
2214 		}
2215 	}
2216 	if (last_offset < tls->params.tls_hlen)
2217 		return (EBADMSG);
2218 
2219 	*record_typep = record_type;
2220 	*trailer_len = tls_len - last_offset + 1;
2221 	return (0);
2222 }
2223 
2224 /*
2225  * Check if a mbuf chain is fully decrypted at the given offset and
2226  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2227  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2228  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2229  * is encrypted.
2230  */
2231 ktls_mbuf_crypto_st_t
2232 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2233 {
2234 	int m_flags_ored = 0;
2235 	int m_flags_anded = -1;
2236 
2237 	for (; mb != NULL; mb = mb->m_next) {
2238 		if (offset < mb->m_len)
2239 			break;
2240 		offset -= mb->m_len;
2241 	}
2242 	offset += len;
2243 
2244 	for (; mb != NULL; mb = mb->m_next) {
2245 		m_flags_ored |= mb->m_flags;
2246 		m_flags_anded &= mb->m_flags;
2247 
2248 		if (offset <= mb->m_len)
2249 			break;
2250 		offset -= mb->m_len;
2251 	}
2252 	MPASS(mb != NULL || offset == 0);
2253 
2254 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2255 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2256 	else
2257 		return ((m_flags_ored & M_DECRYPTED) ?
2258 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2259 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2260 }
2261 
2262 /*
2263  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2264  */
2265 static int
2266 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2267 {
2268 	union if_snd_tag_modify_params params;
2269 	struct m_snd_tag *mst;
2270 	struct inpcb *inp;
2271 	struct tcpcb *tp;
2272 
2273 	mst = so->so_rcv.sb_tls_info->snd_tag;
2274 	if (__predict_false(mst == NULL))
2275 		return (EINVAL);
2276 
2277 	inp = sotoinpcb(so);
2278 	if (__predict_false(inp == NULL))
2279 		return (EINVAL);
2280 
2281 	INP_RLOCK(inp);
2282 	if (inp->inp_flags & INP_DROPPED) {
2283 		INP_RUNLOCK(inp);
2284 		return (ECONNRESET);
2285 	}
2286 
2287 	tp = intotcpcb(inp);
2288 	MPASS(tp != NULL);
2289 
2290 	/* Get the TCP sequence number of the next valid TLS header. */
2291 	SOCKBUF_LOCK(&so->so_rcv);
2292 	params.tls_rx.tls_hdr_tcp_sn =
2293 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2294 	params.tls_rx.tls_rec_length = tls_len;
2295 	params.tls_rx.tls_seq_number = tls_rcd_num;
2296 	SOCKBUF_UNLOCK(&so->so_rcv);
2297 
2298 	INP_RUNLOCK(inp);
2299 
2300 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2301 	return (mst->sw->snd_tag_modify(mst, &params));
2302 }
2303 
2304 static void
2305 ktls_decrypt(struct socket *so)
2306 {
2307 	char tls_header[MBUF_PEXT_HDR_LEN];
2308 	struct ktls_session *tls;
2309 	struct sockbuf *sb;
2310 	struct tls_record_layer *hdr;
2311 	struct tls_get_record tgr;
2312 	struct mbuf *control, *data, *m;
2313 	ktls_mbuf_crypto_st_t state;
2314 	uint64_t seqno;
2315 	int error, remain, tls_len, trail_len;
2316 	bool tls13;
2317 	uint8_t vminor, record_type;
2318 
2319 	hdr = (struct tls_record_layer *)tls_header;
2320 	sb = &so->so_rcv;
2321 	SOCKBUF_LOCK(sb);
2322 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2323 	    ("%s: socket %p not running", __func__, so));
2324 
2325 	tls = sb->sb_tls_info;
2326 	MPASS(tls != NULL);
2327 
2328 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2329 	if (tls13)
2330 		vminor = TLS_MINOR_VER_TWO;
2331 	else
2332 		vminor = tls->params.tls_vminor;
2333 	for (;;) {
2334 		/* Is there enough queued for a TLS header? */
2335 		if (sb->sb_tlscc < tls->params.tls_hlen)
2336 			break;
2337 
2338 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2339 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2340 
2341 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2342 		    hdr->tls_vminor != vminor)
2343 			error = EINVAL;
2344 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2345 			error = EINVAL;
2346 		else if (tls_len < tls->params.tls_hlen || tls_len >
2347 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2348 		    tls->params.tls_tlen)
2349 			error = EMSGSIZE;
2350 		else
2351 			error = 0;
2352 		if (__predict_false(error != 0)) {
2353 			/*
2354 			 * We have a corrupted record and are likely
2355 			 * out of sync.  The connection isn't
2356 			 * recoverable at this point, so abort it.
2357 			 */
2358 			SOCKBUF_UNLOCK(sb);
2359 			counter_u64_add(ktls_offload_corrupted_records, 1);
2360 
2361 			CURVNET_SET(so->so_vnet);
2362 			so->so_proto->pr_abort(so);
2363 			so->so_error = error;
2364 			CURVNET_RESTORE();
2365 			goto deref;
2366 		}
2367 
2368 		/* Is the entire record queued? */
2369 		if (sb->sb_tlscc < tls_len)
2370 			break;
2371 
2372 		/*
2373 		 * Split out the portion of the mbuf chain containing
2374 		 * this TLS record.
2375 		 */
2376 		data = ktls_detach_record(sb, tls_len);
2377 		if (data == NULL)
2378 			continue;
2379 		MPASS(sb->sb_tlsdcc == tls_len);
2380 
2381 		seqno = sb->sb_tls_seqno;
2382 		sb->sb_tls_seqno++;
2383 		SBCHECK(sb);
2384 		SOCKBUF_UNLOCK(sb);
2385 
2386 		/* get crypto state for this TLS record */
2387 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2388 
2389 		switch (state) {
2390 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2391 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2392 			if (error)
2393 				break;
2394 			/* FALLTHROUGH */
2395 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2396 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2397 			    &trail_len);
2398 			if (__predict_true(error == 0)) {
2399 				if (tls13) {
2400 					error = tls13_find_record_type(tls, data,
2401 					    tls_len, &trail_len, &record_type);
2402 				} else {
2403 					record_type = hdr->tls_type;
2404 				}
2405 			}
2406 			break;
2407 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2408 			/*
2409 			 * NIC TLS is only supported for AEAD
2410 			 * ciphersuites which used a fixed sized
2411 			 * trailer.
2412 			 */
2413 			if (tls13) {
2414 				trail_len = tls->params.tls_tlen - 1;
2415 				error = tls13_find_record_type(tls, data,
2416 				    tls_len, &trail_len, &record_type);
2417 			} else {
2418 				trail_len = tls->params.tls_tlen;
2419 				error = 0;
2420 				record_type = hdr->tls_type;
2421 			}
2422 			break;
2423 		default:
2424 			error = EINVAL;
2425 			break;
2426 		}
2427 		if (error) {
2428 			counter_u64_add(ktls_offload_failed_crypto, 1);
2429 
2430 			SOCKBUF_LOCK(sb);
2431 			if (sb->sb_tlsdcc == 0) {
2432 				/*
2433 				 * sbcut/drop/flush discarded these
2434 				 * mbufs.
2435 				 */
2436 				m_freem(data);
2437 				break;
2438 			}
2439 
2440 			/*
2441 			 * Drop this TLS record's data, but keep
2442 			 * decrypting subsequent records.
2443 			 */
2444 			sb->sb_ccc -= tls_len;
2445 			sb->sb_tlsdcc = 0;
2446 
2447 			if (error != EMSGSIZE)
2448 				error = EBADMSG;
2449 			CURVNET_SET(so->so_vnet);
2450 			so->so_error = error;
2451 			sorwakeup_locked(so);
2452 			CURVNET_RESTORE();
2453 
2454 			m_freem(data);
2455 
2456 			SOCKBUF_LOCK(sb);
2457 			continue;
2458 		}
2459 
2460 		/* Allocate the control mbuf. */
2461 		memset(&tgr, 0, sizeof(tgr));
2462 		tgr.tls_type = record_type;
2463 		tgr.tls_vmajor = hdr->tls_vmajor;
2464 		tgr.tls_vminor = hdr->tls_vminor;
2465 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2466 		    trail_len);
2467 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2468 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2469 
2470 		SOCKBUF_LOCK(sb);
2471 		if (sb->sb_tlsdcc == 0) {
2472 			/* sbcut/drop/flush discarded these mbufs. */
2473 			MPASS(sb->sb_tlscc == 0);
2474 			m_freem(data);
2475 			m_freem(control);
2476 			break;
2477 		}
2478 
2479 		/*
2480 		 * Clear the 'dcc' accounting in preparation for
2481 		 * adding the decrypted record.
2482 		 */
2483 		sb->sb_ccc -= tls_len;
2484 		sb->sb_tlsdcc = 0;
2485 		SBCHECK(sb);
2486 
2487 		/* If there is no payload, drop all of the data. */
2488 		if (tgr.tls_length == htobe16(0)) {
2489 			m_freem(data);
2490 			data = NULL;
2491 		} else {
2492 			/* Trim header. */
2493 			remain = tls->params.tls_hlen;
2494 			while (remain > 0) {
2495 				if (data->m_len > remain) {
2496 					data->m_data += remain;
2497 					data->m_len -= remain;
2498 					break;
2499 				}
2500 				remain -= data->m_len;
2501 				data = m_free(data);
2502 			}
2503 
2504 			/* Trim trailer and clear M_NOTREADY. */
2505 			remain = be16toh(tgr.tls_length);
2506 			m = data;
2507 			for (m = data; remain > m->m_len; m = m->m_next) {
2508 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2509 				remain -= m->m_len;
2510 			}
2511 			m->m_len = remain;
2512 			m_freem(m->m_next);
2513 			m->m_next = NULL;
2514 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2515 
2516 			/* Set EOR on the final mbuf. */
2517 			m->m_flags |= M_EOR;
2518 		}
2519 
2520 		sbappendcontrol_locked(sb, data, control, 0);
2521 
2522 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2523 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2524 			SOCKBUF_UNLOCK(sb);
2525 			ktls_resync_ifnet(so, tls_len, seqno);
2526 			SOCKBUF_LOCK(sb);
2527 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2528 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2529 			SOCKBUF_UNLOCK(sb);
2530 			ktls_resync_ifnet(so, 0, seqno);
2531 			SOCKBUF_LOCK(sb);
2532 		}
2533 	}
2534 
2535 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2536 
2537 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2538 		so->so_error = EMSGSIZE;
2539 
2540 	sorwakeup_locked(so);
2541 
2542 deref:
2543 	SOCKBUF_UNLOCK_ASSERT(sb);
2544 
2545 	CURVNET_SET(so->so_vnet);
2546 	sorele(so);
2547 	CURVNET_RESTORE();
2548 }
2549 
2550 void
2551 ktls_enqueue_to_free(struct mbuf *m)
2552 {
2553 	struct ktls_wq *wq;
2554 	bool running;
2555 
2556 	/* Mark it for freeing. */
2557 	m->m_epg_flags |= EPG_FLAG_2FREE;
2558 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2559 	mtx_lock(&wq->mtx);
2560 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2561 	running = wq->running;
2562 	mtx_unlock(&wq->mtx);
2563 	if (!running)
2564 		wakeup(wq);
2565 }
2566 
2567 static void *
2568 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2569 {
2570 	void *buf;
2571 	int domain, running;
2572 
2573 	if (m->m_epg_npgs <= 2)
2574 		return (NULL);
2575 	if (ktls_buffer_zone == NULL)
2576 		return (NULL);
2577 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2578 		/*
2579 		 * Rate-limit allocation attempts after a failure.
2580 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2581 		 * the free page queues and may fail consistently if memory is
2582 		 * fragmented.
2583 		 */
2584 		return (NULL);
2585 	}
2586 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2587 	if (buf == NULL) {
2588 		domain = PCPU_GET(domain);
2589 		wq->lastallocfail = ticks;
2590 
2591 		/*
2592 		 * Note that this check is "racy", but the races are
2593 		 * harmless, and are either a spurious wakeup if
2594 		 * multiple threads fail allocations before the alloc
2595 		 * thread wakes, or waiting an extra second in case we
2596 		 * see an old value of running == true.
2597 		 */
2598 		if (!VM_DOMAIN_EMPTY(domain)) {
2599 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2600 			if (!running)
2601 				wakeup(&ktls_domains[domain].alloc_td);
2602 		}
2603 	}
2604 	return (buf);
2605 }
2606 
2607 static int
2608 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2609     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2610 {
2611 	vm_page_t pg;
2612 	int error, i, len, off;
2613 
2614 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2615 	    ("%p not unready & nomap mbuf\n", m));
2616 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2617 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2618 	    ktls_maxlen));
2619 
2620 	/* Anonymous mbufs are encrypted in place. */
2621 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2622 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2623 
2624 	/*
2625 	 * For file-backed mbufs (from sendfile), anonymous wired
2626 	 * pages are allocated and used as the encryption destination.
2627 	 */
2628 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2629 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2630 		    m->m_epg_1st_off;
2631 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2632 		    m->m_epg_1st_off;
2633 		state->dst_iov[0].iov_len = len;
2634 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2635 		i = 1;
2636 	} else {
2637 		off = m->m_epg_1st_off;
2638 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2639 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2640 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2641 			len = m_epg_pagelen(m, i, off);
2642 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2643 			state->dst_iov[i].iov_base =
2644 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2645 			state->dst_iov[i].iov_len = len;
2646 		}
2647 	}
2648 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2649 	state->dst_iov[i].iov_base = m->m_epg_trail;
2650 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2651 
2652 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2653 
2654 	if (__predict_false(error != 0)) {
2655 		/* Free the anonymous pages. */
2656 		if (state->cbuf != NULL)
2657 			uma_zfree(ktls_buffer_zone, state->cbuf);
2658 		else {
2659 			for (i = 0; i < m->m_epg_npgs; i++) {
2660 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2661 				(void)vm_page_unwire_noq(pg);
2662 				vm_page_free(pg);
2663 			}
2664 		}
2665 	}
2666 	return (error);
2667 }
2668 
2669 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2670 static u_int
2671 ktls_batched_records(struct mbuf *m)
2672 {
2673 	int page_count, records;
2674 
2675 	records = 0;
2676 	page_count = m->m_epg_enc_cnt;
2677 	while (page_count > 0) {
2678 		records++;
2679 		page_count -= m->m_epg_nrdy;
2680 		m = m->m_next;
2681 	}
2682 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2683 	return (records);
2684 }
2685 
2686 void
2687 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2688 {
2689 	struct ktls_session *tls;
2690 	struct ktls_wq *wq;
2691 	int queued;
2692 	bool running;
2693 
2694 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2695 	    (M_EXTPG | M_NOTREADY)),
2696 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2697 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2698 
2699 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2700 
2701 	m->m_epg_enc_cnt = page_count;
2702 
2703 	/*
2704 	 * Save a pointer to the socket.  The caller is responsible
2705 	 * for taking an additional reference via soref().
2706 	 */
2707 	m->m_epg_so = so;
2708 
2709 	queued = 1;
2710 	tls = m->m_epg_tls;
2711 	wq = &ktls_wq[tls->wq_index];
2712 	mtx_lock(&wq->mtx);
2713 	if (__predict_false(tls->sequential_records)) {
2714 		/*
2715 		 * For TLS 1.0, records must be encrypted
2716 		 * sequentially.  For a given connection, all records
2717 		 * queued to the associated work queue are processed
2718 		 * sequentially.  However, sendfile(2) might complete
2719 		 * I/O requests spanning multiple TLS records out of
2720 		 * order.  Here we ensure TLS records are enqueued to
2721 		 * the work queue in FIFO order.
2722 		 *
2723 		 * tls->next_seqno holds the sequence number of the
2724 		 * next TLS record that should be enqueued to the work
2725 		 * queue.  If this next record is not tls->next_seqno,
2726 		 * it must be a future record, so insert it, sorted by
2727 		 * TLS sequence number, into tls->pending_records and
2728 		 * return.
2729 		 *
2730 		 * If this TLS record matches tls->next_seqno, place
2731 		 * it in the work queue and then check
2732 		 * tls->pending_records to see if any
2733 		 * previously-queued records are now ready for
2734 		 * encryption.
2735 		 */
2736 		if (m->m_epg_seqno != tls->next_seqno) {
2737 			struct mbuf *n, *p;
2738 
2739 			p = NULL;
2740 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2741 				if (n->m_epg_seqno > m->m_epg_seqno)
2742 					break;
2743 				p = n;
2744 			}
2745 			if (n == NULL)
2746 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2747 				    m_epg_stailq);
2748 			else if (p == NULL)
2749 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2750 				    m_epg_stailq);
2751 			else
2752 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2753 				    m_epg_stailq);
2754 			mtx_unlock(&wq->mtx);
2755 			counter_u64_add(ktls_cnt_tx_pending, 1);
2756 			return;
2757 		}
2758 
2759 		tls->next_seqno += ktls_batched_records(m);
2760 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2761 
2762 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2763 			struct mbuf *n;
2764 
2765 			n = STAILQ_FIRST(&tls->pending_records);
2766 			if (n->m_epg_seqno != tls->next_seqno)
2767 				break;
2768 
2769 			queued++;
2770 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2771 			tls->next_seqno += ktls_batched_records(n);
2772 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2773 		}
2774 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2775 	} else
2776 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2777 
2778 	running = wq->running;
2779 	mtx_unlock(&wq->mtx);
2780 	if (!running)
2781 		wakeup(wq);
2782 	counter_u64_add(ktls_cnt_tx_queued, queued);
2783 }
2784 
2785 /*
2786  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2787  * the pages from the file and replace them with the anonymous pages
2788  * allocated in ktls_encrypt_record().
2789  */
2790 static void
2791 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2792 {
2793 	int i;
2794 
2795 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2796 
2797 	/* Free the old pages. */
2798 	m->m_ext.ext_free(m);
2799 
2800 	/* Replace them with the new pages. */
2801 	if (state->cbuf != NULL) {
2802 		for (i = 0; i < m->m_epg_npgs; i++)
2803 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2804 
2805 		/* Contig pages should go back to the cache. */
2806 		m->m_ext.ext_free = ktls_free_mext_contig;
2807 	} else {
2808 		for (i = 0; i < m->m_epg_npgs; i++)
2809 			m->m_epg_pa[i] = state->parray[i];
2810 
2811 		/* Use the basic free routine. */
2812 		m->m_ext.ext_free = mb_free_mext_pgs;
2813 	}
2814 
2815 	/* Pages are now writable. */
2816 	m->m_epg_flags |= EPG_FLAG_ANON;
2817 }
2818 
2819 static __noinline void
2820 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2821 {
2822 	struct ktls_ocf_encrypt_state state;
2823 	struct ktls_session *tls;
2824 	struct socket *so;
2825 	struct mbuf *m;
2826 	int error, npages, total_pages;
2827 
2828 	so = top->m_epg_so;
2829 	tls = top->m_epg_tls;
2830 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2831 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2832 #ifdef INVARIANTS
2833 	top->m_epg_so = NULL;
2834 #endif
2835 	total_pages = top->m_epg_enc_cnt;
2836 	npages = 0;
2837 
2838 	/*
2839 	 * Encrypt the TLS records in the chain of mbufs starting with
2840 	 * 'top'.  'total_pages' gives us a total count of pages and is
2841 	 * used to know when we have finished encrypting the TLS
2842 	 * records originally queued with 'top'.
2843 	 *
2844 	 * NB: These mbufs are queued in the socket buffer and
2845 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2846 	 * socket buffer lock is not held while traversing this chain.
2847 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2848 	 * pointers should be stable.  However, the 'm_next' of the
2849 	 * last mbuf encrypted is not necessarily NULL.  It can point
2850 	 * to other mbufs appended while 'top' was on the TLS work
2851 	 * queue.
2852 	 *
2853 	 * Each mbuf holds an entire TLS record.
2854 	 */
2855 	error = 0;
2856 	for (m = top; npages != total_pages; m = m->m_next) {
2857 		KASSERT(m->m_epg_tls == tls,
2858 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2859 		    tls, m->m_epg_tls));
2860 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2861 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2862 		    total_pages, m));
2863 
2864 		error = ktls_encrypt_record(wq, m, tls, &state);
2865 		if (error) {
2866 			counter_u64_add(ktls_offload_failed_crypto, 1);
2867 			break;
2868 		}
2869 
2870 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2871 			ktls_finish_nonanon(m, &state);
2872 
2873 		npages += m->m_epg_nrdy;
2874 
2875 		/*
2876 		 * Drop a reference to the session now that it is no
2877 		 * longer needed.  Existing code depends on encrypted
2878 		 * records having no associated session vs
2879 		 * yet-to-be-encrypted records having an associated
2880 		 * session.
2881 		 */
2882 		m->m_epg_tls = NULL;
2883 		ktls_free(tls);
2884 	}
2885 
2886 	CURVNET_SET(so->so_vnet);
2887 	if (error == 0) {
2888 		(void)so->so_proto->pr_ready(so, top, npages);
2889 	} else {
2890 		so->so_proto->pr_abort(so);
2891 		so->so_error = EIO;
2892 		mb_free_notready(top, total_pages);
2893 	}
2894 
2895 	sorele(so);
2896 	CURVNET_RESTORE();
2897 }
2898 
2899 void
2900 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2901 {
2902 	struct ktls_session *tls;
2903 	struct socket *so;
2904 	struct mbuf *m;
2905 	int npages;
2906 
2907 	m = state->m;
2908 
2909 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2910 		ktls_finish_nonanon(m, state);
2911 
2912 	so = state->so;
2913 	free(state, M_KTLS);
2914 
2915 	/*
2916 	 * Drop a reference to the session now that it is no longer
2917 	 * needed.  Existing code depends on encrypted records having
2918 	 * no associated session vs yet-to-be-encrypted records having
2919 	 * an associated session.
2920 	 */
2921 	tls = m->m_epg_tls;
2922 	m->m_epg_tls = NULL;
2923 	ktls_free(tls);
2924 
2925 	if (error != 0)
2926 		counter_u64_add(ktls_offload_failed_crypto, 1);
2927 
2928 	CURVNET_SET(so->so_vnet);
2929 	npages = m->m_epg_nrdy;
2930 
2931 	if (error == 0) {
2932 		(void)so->so_proto->pr_ready(so, m, npages);
2933 	} else {
2934 		so->so_proto->pr_abort(so);
2935 		so->so_error = EIO;
2936 		mb_free_notready(m, npages);
2937 	}
2938 
2939 	sorele(so);
2940 	CURVNET_RESTORE();
2941 }
2942 
2943 /*
2944  * Similar to ktls_encrypt, but used with asynchronous OCF backends
2945  * (coprocessors) where encryption does not use host CPU resources and
2946  * it can be beneficial to queue more requests than CPUs.
2947  */
2948 static __noinline void
2949 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2950 {
2951 	struct ktls_ocf_encrypt_state *state;
2952 	struct ktls_session *tls;
2953 	struct socket *so;
2954 	struct mbuf *m, *n;
2955 	int error, mpages, npages, total_pages;
2956 
2957 	so = top->m_epg_so;
2958 	tls = top->m_epg_tls;
2959 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2960 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2961 #ifdef INVARIANTS
2962 	top->m_epg_so = NULL;
2963 #endif
2964 	total_pages = top->m_epg_enc_cnt;
2965 	npages = 0;
2966 
2967 	error = 0;
2968 	for (m = top; npages != total_pages; m = n) {
2969 		KASSERT(m->m_epg_tls == tls,
2970 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2971 		    tls, m->m_epg_tls));
2972 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2973 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2974 		    total_pages, m));
2975 
2976 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2977 		soref(so);
2978 		state->so = so;
2979 		state->m = m;
2980 
2981 		mpages = m->m_epg_nrdy;
2982 		n = m->m_next;
2983 
2984 		error = ktls_encrypt_record(wq, m, tls, state);
2985 		if (error) {
2986 			counter_u64_add(ktls_offload_failed_crypto, 1);
2987 			free(state, M_KTLS);
2988 			CURVNET_SET(so->so_vnet);
2989 			sorele(so);
2990 			CURVNET_RESTORE();
2991 			break;
2992 		}
2993 
2994 		npages += mpages;
2995 	}
2996 
2997 	CURVNET_SET(so->so_vnet);
2998 	if (error != 0) {
2999 		so->so_proto->pr_abort(so);
3000 		so->so_error = EIO;
3001 		mb_free_notready(m, total_pages - npages);
3002 	}
3003 
3004 	sorele(so);
3005 	CURVNET_RESTORE();
3006 }
3007 
3008 static int
3009 ktls_bind_domain(int domain)
3010 {
3011 	int error;
3012 
3013 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3014 	if (error != 0)
3015 		return (error);
3016 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3017 	return (0);
3018 }
3019 
3020 static void
3021 ktls_alloc_thread(void *ctx)
3022 {
3023 	struct ktls_domain_info *ktls_domain = ctx;
3024 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3025 	void **buf;
3026 	struct sysctl_oid *oid;
3027 	char name[80];
3028 	int domain, error, i, nbufs;
3029 
3030 	domain = ktls_domain - ktls_domains;
3031 	if (bootverbose)
3032 		printf("Starting KTLS alloc thread for domain %d\n", domain);
3033 	error = ktls_bind_domain(domain);
3034 	if (error)
3035 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3036 		    domain, error);
3037 	snprintf(name, sizeof(name), "domain%d", domain);
3038 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3039 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3040 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3041 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
3042 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3043 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3044 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3045 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3046 
3047 	buf = NULL;
3048 	nbufs = 0;
3049 	for (;;) {
3050 		atomic_store_int(&sc->running, 0);
3051 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3052 		atomic_store_int(&sc->running, 1);
3053 		sc->wakeups++;
3054 		if (nbufs != ktls_max_alloc) {
3055 			free(buf, M_KTLS);
3056 			nbufs = atomic_load_int(&ktls_max_alloc);
3057 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3058 			    M_WAITOK | M_ZERO);
3059 		}
3060 		/*
3061 		 * Below we allocate nbufs with different allocation
3062 		 * flags than we use when allocating normally during
3063 		 * encryption in the ktls worker thread.  We specify
3064 		 * M_NORECLAIM in the worker thread. However, we omit
3065 		 * that flag here and add M_WAITOK so that the VM
3066 		 * system is permitted to perform expensive work to
3067 		 * defragment memory.  We do this here, as it does not
3068 		 * matter if this thread blocks.  If we block a ktls
3069 		 * worker thread, we risk developing backlogs of
3070 		 * buffers to be encrypted, leading to surges of
3071 		 * traffic and potential NIC output drops.
3072 		 */
3073 		for (i = 0; i < nbufs; i++) {
3074 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3075 			sc->allocs++;
3076 		}
3077 		for (i = 0; i < nbufs; i++) {
3078 			uma_zfree(ktls_buffer_zone, buf[i]);
3079 			buf[i] = NULL;
3080 		}
3081 	}
3082 }
3083 
3084 static void
3085 ktls_work_thread(void *ctx)
3086 {
3087 	struct ktls_wq *wq = ctx;
3088 	struct mbuf *m, *n;
3089 	struct socket *so, *son;
3090 	STAILQ_HEAD(, mbuf) local_m_head;
3091 	STAILQ_HEAD(, socket) local_so_head;
3092 	int cpu;
3093 
3094 	cpu = wq - ktls_wq;
3095 	if (bootverbose)
3096 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3097 
3098 	/*
3099 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3100 	 * we bind to the NUMA domain instead.
3101 	 */
3102 	if (ktls_bind_threads) {
3103 		int error;
3104 
3105 		if (ktls_bind_threads > 1) {
3106 			struct pcpu *pc = pcpu_find(cpu);
3107 
3108 			error = ktls_bind_domain(pc->pc_domain);
3109 		} else {
3110 			cpuset_t mask;
3111 
3112 			CPU_SETOF(cpu, &mask);
3113 			error = cpuset_setthread(curthread->td_tid, &mask);
3114 		}
3115 		if (error)
3116 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3117 				cpu, error);
3118 	}
3119 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3120 	fpu_kern_thread(0);
3121 #endif
3122 	for (;;) {
3123 		mtx_lock(&wq->mtx);
3124 		while (STAILQ_EMPTY(&wq->m_head) &&
3125 		    STAILQ_EMPTY(&wq->so_head)) {
3126 			wq->running = false;
3127 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3128 			wq->running = true;
3129 		}
3130 
3131 		STAILQ_INIT(&local_m_head);
3132 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3133 		STAILQ_INIT(&local_so_head);
3134 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3135 		mtx_unlock(&wq->mtx);
3136 
3137 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3138 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3139 				ktls_free(m->m_epg_tls);
3140 				m_free_raw(m);
3141 			} else {
3142 				if (m->m_epg_tls->sync_dispatch)
3143 					ktls_encrypt(wq, m);
3144 				else
3145 					ktls_encrypt_async(wq, m);
3146 				counter_u64_add(ktls_cnt_tx_queued, -1);
3147 			}
3148 		}
3149 
3150 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3151 			ktls_decrypt(so);
3152 			counter_u64_add(ktls_cnt_rx_queued, -1);
3153 		}
3154 	}
3155 }
3156 
3157 #if defined(INET) || defined(INET6)
3158 static void
3159 ktls_disable_ifnet_help(void *context, int pending __unused)
3160 {
3161 	struct ktls_session *tls;
3162 	struct inpcb *inp;
3163 	struct tcpcb *tp;
3164 	struct socket *so;
3165 	int err;
3166 
3167 	tls = context;
3168 	inp = tls->inp;
3169 	if (inp == NULL)
3170 		return;
3171 	INP_WLOCK(inp);
3172 	so = inp->inp_socket;
3173 	MPASS(so != NULL);
3174 	if (inp->inp_flags & INP_DROPPED) {
3175 		goto out;
3176 	}
3177 
3178 	if (so->so_snd.sb_tls_info != NULL)
3179 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3180 	else
3181 		err = ENXIO;
3182 	if (err == 0) {
3183 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3184 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3185 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3186 		    (tp = intotcpcb(inp)) != NULL &&
3187 		    tp->t_fb->tfb_hwtls_change != NULL)
3188 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3189 	} else {
3190 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3191 	}
3192 
3193 out:
3194 	sorele(so);
3195 	if (!in_pcbrele_wlocked(inp))
3196 		INP_WUNLOCK(inp);
3197 	ktls_free(tls);
3198 }
3199 
3200 /*
3201  * Called when re-transmits are becoming a substantial portion of the
3202  * sends on this connection.  When this happens, we transition the
3203  * connection to software TLS.  This is needed because most inline TLS
3204  * NICs keep crypto state only for in-order transmits.  This means
3205  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3206  * re-DMA the entire TLS record up to and including the current
3207  * segment.  This means that when re-transmitting the last ~1448 byte
3208  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3209  * of magnitude more data than we are sending.  This can cause the
3210  * PCIe link to saturate well before the network, which can cause
3211  * output drops, and a general loss of capacity.
3212  */
3213 void
3214 ktls_disable_ifnet(void *arg)
3215 {
3216 	struct tcpcb *tp;
3217 	struct inpcb *inp;
3218 	struct socket *so;
3219 	struct ktls_session *tls;
3220 
3221 	tp = arg;
3222 	inp = tptoinpcb(tp);
3223 	INP_WLOCK_ASSERT(inp);
3224 	so = inp->inp_socket;
3225 	SOCK_LOCK(so);
3226 	tls = so->so_snd.sb_tls_info;
3227 	if (tls->disable_ifnet_pending) {
3228 		SOCK_UNLOCK(so);
3229 		return;
3230 	}
3231 
3232 	/*
3233 	 * note that disable_ifnet_pending is never cleared; disabling
3234 	 * ifnet can only be done once per session, so we never want
3235 	 * to do it again
3236 	 */
3237 
3238 	(void)ktls_hold(tls);
3239 	in_pcbref(inp);
3240 	soref(so);
3241 	tls->disable_ifnet_pending = true;
3242 	tls->inp = inp;
3243 	SOCK_UNLOCK(so);
3244 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3245 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3246 }
3247 #endif
3248