xref: /freebsd/sys/kern/uipc_ktls.c (revision 3332f1b444d4a73238e9f59cca27bfc95fe936bd)
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 static void ktls_cleanup(struct ktls_session *tls);
301 #if defined(INET) || defined(INET6)
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)
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 		if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
555 			en->iv_len != TLS_AEAD_GCM_LEN) ||
556 		    (en->tls_vminor == TLS_MINOR_VER_THREE &&
557 			en->iv_len != TLS_1_3_GCM_IV_LEN))
558 			return (EINVAL);
559 		break;
560 	case CRYPTO_AES_CBC:
561 		switch (en->auth_algorithm) {
562 		case CRYPTO_SHA1_HMAC:
563 			/*
564 			 * TLS 1.0 requires an implicit IV.  TLS 1.1+
565 			 * all use explicit IVs.
566 			 */
567 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
568 				if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
569 					return (EINVAL);
570 				break;
571 			}
572 
573 			/* FALLTHROUGH */
574 		case CRYPTO_SHA2_256_HMAC:
575 		case CRYPTO_SHA2_384_HMAC:
576 			/* Ignore any supplied IV. */
577 			en->iv_len = 0;
578 			break;
579 		default:
580 			return (EINVAL);
581 		}
582 		if (en->auth_key_len == 0)
583 			return (EINVAL);
584 		if (en->tls_vminor != TLS_MINOR_VER_ZERO &&
585 		    en->tls_vminor != TLS_MINOR_VER_ONE &&
586 		    en->tls_vminor != TLS_MINOR_VER_TWO)
587 			return (EINVAL);
588 		break;
589 	case CRYPTO_CHACHA20_POLY1305:
590 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
591 			return (EINVAL);
592 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
593 		    en->tls_vminor != TLS_MINOR_VER_THREE)
594 			return (EINVAL);
595 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
596 			return (EINVAL);
597 		break;
598 	default:
599 		return (EINVAL);
600 	}
601 
602 	error = ktls_start_kthreads();
603 	if (error != 0)
604 		return (error);
605 
606 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
607 
608 	counter_u64_add(ktls_offload_active, 1);
609 
610 	refcount_init(&tls->refcount, 1);
611 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
612 
613 	tls->wq_index = ktls_get_cpu(so);
614 
615 	tls->params.cipher_algorithm = en->cipher_algorithm;
616 	tls->params.auth_algorithm = en->auth_algorithm;
617 	tls->params.tls_vmajor = en->tls_vmajor;
618 	tls->params.tls_vminor = en->tls_vminor;
619 	tls->params.flags = en->flags;
620 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
621 
622 	/* Set the header and trailer lengths. */
623 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
624 	switch (en->cipher_algorithm) {
625 	case CRYPTO_AES_NIST_GCM_16:
626 		/*
627 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
628 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
629 		 */
630 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
631 			tls->params.tls_hlen += sizeof(uint64_t);
632 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
633 		tls->params.tls_bs = 1;
634 		break;
635 	case CRYPTO_AES_CBC:
636 		switch (en->auth_algorithm) {
637 		case CRYPTO_SHA1_HMAC:
638 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
639 				/* Implicit IV, no nonce. */
640 				tls->sequential_records = true;
641 				tls->next_seqno = be64dec(en->rec_seq);
642 				STAILQ_INIT(&tls->pending_records);
643 			} else {
644 				tls->params.tls_hlen += AES_BLOCK_LEN;
645 			}
646 			tls->params.tls_tlen = AES_BLOCK_LEN +
647 			    SHA1_HASH_LEN;
648 			break;
649 		case CRYPTO_SHA2_256_HMAC:
650 			tls->params.tls_hlen += AES_BLOCK_LEN;
651 			tls->params.tls_tlen = AES_BLOCK_LEN +
652 			    SHA2_256_HASH_LEN;
653 			break;
654 		case CRYPTO_SHA2_384_HMAC:
655 			tls->params.tls_hlen += AES_BLOCK_LEN;
656 			tls->params.tls_tlen = AES_BLOCK_LEN +
657 			    SHA2_384_HASH_LEN;
658 			break;
659 		default:
660 			panic("invalid hmac");
661 		}
662 		tls->params.tls_bs = AES_BLOCK_LEN;
663 		break;
664 	case CRYPTO_CHACHA20_POLY1305:
665 		/*
666 		 * Chacha20 uses a 12 byte implicit IV.
667 		 */
668 		tls->params.tls_tlen = POLY1305_HASH_LEN;
669 		tls->params.tls_bs = 1;
670 		break;
671 	default:
672 		panic("invalid cipher");
673 	}
674 
675 	/*
676 	 * TLS 1.3 includes optional padding which we do not support,
677 	 * and also puts the "real" record type at the end of the
678 	 * encrypted data.
679 	 */
680 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
681 		tls->params.tls_tlen += sizeof(uint8_t);
682 
683 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
684 	    ("TLS header length too long: %d", tls->params.tls_hlen));
685 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
686 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
687 
688 	if (en->auth_key_len != 0) {
689 		tls->params.auth_key_len = en->auth_key_len;
690 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
691 		    M_WAITOK);
692 		error = copyin(en->auth_key, tls->params.auth_key,
693 		    en->auth_key_len);
694 		if (error)
695 			goto out;
696 	}
697 
698 	tls->params.cipher_key_len = en->cipher_key_len;
699 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
700 	error = copyin(en->cipher_key, tls->params.cipher_key,
701 	    en->cipher_key_len);
702 	if (error)
703 		goto out;
704 
705 	/*
706 	 * This holds the implicit portion of the nonce for AEAD
707 	 * ciphers and the initial implicit IV for TLS 1.0.  The
708 	 * explicit portions of the IV are generated in ktls_frame().
709 	 */
710 	if (en->iv_len != 0) {
711 		tls->params.iv_len = en->iv_len;
712 		error = copyin(en->iv, tls->params.iv, en->iv_len);
713 		if (error)
714 			goto out;
715 
716 		/*
717 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
718 		 * counter to generate unique explicit IVs.
719 		 *
720 		 * Store this counter in the last 8 bytes of the IV
721 		 * array so that it is 8-byte aligned.
722 		 */
723 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
724 		    en->tls_vminor == TLS_MINOR_VER_TWO)
725 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
726 	}
727 
728 	*tlsp = tls;
729 	return (0);
730 
731 out:
732 	ktls_cleanup(tls);
733 	return (error);
734 }
735 
736 static struct ktls_session *
737 ktls_clone_session(struct ktls_session *tls)
738 {
739 	struct ktls_session *tls_new;
740 
741 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
742 
743 	counter_u64_add(ktls_offload_active, 1);
744 
745 	refcount_init(&tls_new->refcount, 1);
746 	TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
747 
748 	/* Copy fields from existing session. */
749 	tls_new->params = tls->params;
750 	tls_new->wq_index = tls->wq_index;
751 
752 	/* Deep copy keys. */
753 	if (tls_new->params.auth_key != NULL) {
754 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
755 		    M_KTLS, M_WAITOK);
756 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
757 		    tls->params.auth_key_len);
758 	}
759 
760 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
761 	    M_WAITOK);
762 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
763 	    tls->params.cipher_key_len);
764 
765 	return (tls_new);
766 }
767 #endif
768 
769 static void
770 ktls_cleanup(struct ktls_session *tls)
771 {
772 
773 	counter_u64_add(ktls_offload_active, -1);
774 	switch (tls->mode) {
775 	case TCP_TLS_MODE_SW:
776 		switch (tls->params.cipher_algorithm) {
777 		case CRYPTO_AES_CBC:
778 			counter_u64_add(ktls_sw_cbc, -1);
779 			break;
780 		case CRYPTO_AES_NIST_GCM_16:
781 			counter_u64_add(ktls_sw_gcm, -1);
782 			break;
783 		case CRYPTO_CHACHA20_POLY1305:
784 			counter_u64_add(ktls_sw_chacha20, -1);
785 			break;
786 		}
787 		break;
788 	case TCP_TLS_MODE_IFNET:
789 		switch (tls->params.cipher_algorithm) {
790 		case CRYPTO_AES_CBC:
791 			counter_u64_add(ktls_ifnet_cbc, -1);
792 			break;
793 		case CRYPTO_AES_NIST_GCM_16:
794 			counter_u64_add(ktls_ifnet_gcm, -1);
795 			break;
796 		case CRYPTO_CHACHA20_POLY1305:
797 			counter_u64_add(ktls_ifnet_chacha20, -1);
798 			break;
799 		}
800 		if (tls->snd_tag != NULL)
801 			m_snd_tag_rele(tls->snd_tag);
802 		break;
803 #ifdef TCP_OFFLOAD
804 	case TCP_TLS_MODE_TOE:
805 		switch (tls->params.cipher_algorithm) {
806 		case CRYPTO_AES_CBC:
807 			counter_u64_add(ktls_toe_cbc, -1);
808 			break;
809 		case CRYPTO_AES_NIST_GCM_16:
810 			counter_u64_add(ktls_toe_gcm, -1);
811 			break;
812 		case CRYPTO_CHACHA20_POLY1305:
813 			counter_u64_add(ktls_toe_chacha20, -1);
814 			break;
815 		}
816 		break;
817 #endif
818 	}
819 	if (tls->ocf_session != NULL)
820 		ktls_ocf_free(tls);
821 	if (tls->params.auth_key != NULL) {
822 		zfree(tls->params.auth_key, M_KTLS);
823 		tls->params.auth_key = NULL;
824 		tls->params.auth_key_len = 0;
825 	}
826 	if (tls->params.cipher_key != NULL) {
827 		zfree(tls->params.cipher_key, M_KTLS);
828 		tls->params.cipher_key = NULL;
829 		tls->params.cipher_key_len = 0;
830 	}
831 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
832 }
833 
834 #if defined(INET) || defined(INET6)
835 
836 #ifdef TCP_OFFLOAD
837 static int
838 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
839 {
840 	struct inpcb *inp;
841 	struct tcpcb *tp;
842 	int error;
843 
844 	inp = so->so_pcb;
845 	INP_WLOCK(inp);
846 	if (inp->inp_flags2 & INP_FREED) {
847 		INP_WUNLOCK(inp);
848 		return (ECONNRESET);
849 	}
850 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
851 		INP_WUNLOCK(inp);
852 		return (ECONNRESET);
853 	}
854 	if (inp->inp_socket == NULL) {
855 		INP_WUNLOCK(inp);
856 		return (ECONNRESET);
857 	}
858 	tp = intotcpcb(inp);
859 	if (!(tp->t_flags & TF_TOE)) {
860 		INP_WUNLOCK(inp);
861 		return (EOPNOTSUPP);
862 	}
863 
864 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
865 	INP_WUNLOCK(inp);
866 	if (error == 0) {
867 		tls->mode = TCP_TLS_MODE_TOE;
868 		switch (tls->params.cipher_algorithm) {
869 		case CRYPTO_AES_CBC:
870 			counter_u64_add(ktls_toe_cbc, 1);
871 			break;
872 		case CRYPTO_AES_NIST_GCM_16:
873 			counter_u64_add(ktls_toe_gcm, 1);
874 			break;
875 		case CRYPTO_CHACHA20_POLY1305:
876 			counter_u64_add(ktls_toe_chacha20, 1);
877 			break;
878 		}
879 	}
880 	return (error);
881 }
882 #endif
883 
884 /*
885  * Common code used when first enabling ifnet TLS on a connection or
886  * when allocating a new ifnet TLS session due to a routing change.
887  * This function allocates a new TLS send tag on whatever interface
888  * the connection is currently routed over.
889  */
890 static int
891 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
892     struct m_snd_tag **mstp)
893 {
894 	union if_snd_tag_alloc_params params;
895 	struct ifnet *ifp;
896 	struct nhop_object *nh;
897 	struct tcpcb *tp;
898 	int error;
899 
900 	INP_RLOCK(inp);
901 	if (inp->inp_flags2 & INP_FREED) {
902 		INP_RUNLOCK(inp);
903 		return (ECONNRESET);
904 	}
905 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
906 		INP_RUNLOCK(inp);
907 		return (ECONNRESET);
908 	}
909 	if (inp->inp_socket == NULL) {
910 		INP_RUNLOCK(inp);
911 		return (ECONNRESET);
912 	}
913 	tp = intotcpcb(inp);
914 
915 	/*
916 	 * Check administrative controls on ifnet TLS to determine if
917 	 * ifnet TLS should be denied.
918 	 *
919 	 * - Always permit 'force' requests.
920 	 * - ktls_ifnet_permitted == 0: always deny.
921 	 */
922 	if (!force && ktls_ifnet_permitted == 0) {
923 		INP_RUNLOCK(inp);
924 		return (ENXIO);
925 	}
926 
927 	/*
928 	 * XXX: Use the cached route in the inpcb to find the
929 	 * interface.  This should perhaps instead use
930 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
931 	 * enabled after a connection has completed key negotiation in
932 	 * userland, the cached route will be present in practice.
933 	 */
934 	nh = inp->inp_route.ro_nh;
935 	if (nh == NULL) {
936 		INP_RUNLOCK(inp);
937 		return (ENXIO);
938 	}
939 	ifp = nh->nh_ifp;
940 	if_ref(ifp);
941 
942 	/*
943 	 * Allocate a TLS + ratelimit tag if the connection has an
944 	 * existing pacing rate.
945 	 */
946 	if (tp->t_pacing_rate != -1 &&
947 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
948 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
949 		params.tls_rate_limit.inp = inp;
950 		params.tls_rate_limit.tls = tls;
951 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
952 	} else {
953 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
954 		params.tls.inp = inp;
955 		params.tls.tls = tls;
956 	}
957 	params.hdr.flowid = inp->inp_flowid;
958 	params.hdr.flowtype = inp->inp_flowtype;
959 	params.hdr.numa_domain = inp->inp_numa_domain;
960 	INP_RUNLOCK(inp);
961 
962 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
963 		error = EOPNOTSUPP;
964 		goto out;
965 	}
966 	if (inp->inp_vflag & INP_IPV6) {
967 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
968 			error = EOPNOTSUPP;
969 			goto out;
970 		}
971 	} else {
972 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
973 			error = EOPNOTSUPP;
974 			goto out;
975 		}
976 	}
977 	error = m_snd_tag_alloc(ifp, &params, mstp);
978 out:
979 	if_rele(ifp);
980 	return (error);
981 }
982 
983 static int
984 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
985 {
986 	struct m_snd_tag *mst;
987 	int error;
988 
989 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
990 	if (error == 0) {
991 		tls->mode = TCP_TLS_MODE_IFNET;
992 		tls->snd_tag = mst;
993 		switch (tls->params.cipher_algorithm) {
994 		case CRYPTO_AES_CBC:
995 			counter_u64_add(ktls_ifnet_cbc, 1);
996 			break;
997 		case CRYPTO_AES_NIST_GCM_16:
998 			counter_u64_add(ktls_ifnet_gcm, 1);
999 			break;
1000 		case CRYPTO_CHACHA20_POLY1305:
1001 			counter_u64_add(ktls_ifnet_chacha20, 1);
1002 			break;
1003 		}
1004 	}
1005 	return (error);
1006 }
1007 
1008 static void
1009 ktls_use_sw(struct ktls_session *tls)
1010 {
1011 	tls->mode = TCP_TLS_MODE_SW;
1012 	switch (tls->params.cipher_algorithm) {
1013 	case CRYPTO_AES_CBC:
1014 		counter_u64_add(ktls_sw_cbc, 1);
1015 		break;
1016 	case CRYPTO_AES_NIST_GCM_16:
1017 		counter_u64_add(ktls_sw_gcm, 1);
1018 		break;
1019 	case CRYPTO_CHACHA20_POLY1305:
1020 		counter_u64_add(ktls_sw_chacha20, 1);
1021 		break;
1022 	}
1023 }
1024 
1025 static int
1026 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1027 {
1028 	int error;
1029 
1030 	error = ktls_ocf_try(so, tls, direction);
1031 	if (error)
1032 		return (error);
1033 	ktls_use_sw(tls);
1034 	return (0);
1035 }
1036 
1037 /*
1038  * KTLS RX stores data in the socket buffer as a list of TLS records,
1039  * where each record is stored as a control message containg the TLS
1040  * header followed by data mbufs containing the decrypted data.  This
1041  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1042  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1043  * should be queued to the socket buffer as records, but encrypted
1044  * data which needs to be decrypted by software arrives as a stream of
1045  * regular mbufs which need to be converted.  In addition, there may
1046  * already be pending encrypted data in the socket buffer when KTLS RX
1047  * is enabled.
1048  *
1049  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1050  * is used:
1051  *
1052  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1053  *
1054  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1055  *   from the first mbuf.  Once all of the data for that TLS record is
1056  *   queued, the socket is queued to a worker thread.
1057  *
1058  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1059  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1060  *   decrypted, and inserted into the regular socket buffer chain as
1061  *   record starting with a control message holding the TLS header and
1062  *   a chain of mbufs holding the encrypted data.
1063  */
1064 
1065 static void
1066 sb_mark_notready(struct sockbuf *sb)
1067 {
1068 	struct mbuf *m;
1069 
1070 	m = sb->sb_mb;
1071 	sb->sb_mtls = m;
1072 	sb->sb_mb = NULL;
1073 	sb->sb_mbtail = NULL;
1074 	sb->sb_lastrecord = NULL;
1075 	for (; m != NULL; m = m->m_next) {
1076 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1077 		    __func__));
1078 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1079 		    __func__));
1080 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1081 		    __func__));
1082 		m->m_flags |= M_NOTREADY;
1083 		sb->sb_acc -= m->m_len;
1084 		sb->sb_tlscc += m->m_len;
1085 		sb->sb_mtlstail = m;
1086 	}
1087 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1088 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1089 	    sb->sb_ccc));
1090 }
1091 
1092 /*
1093  * Return information about the pending TLS data in a socket
1094  * buffer.  On return, 'seqno' is set to the sequence number
1095  * of the next TLS record to be received, 'resid' is set to
1096  * the amount of bytes still needed for the last pending
1097  * record.  The function returns 'false' if the last pending
1098  * record contains a partial TLS header.  In that case, 'resid'
1099  * is the number of bytes needed to complete the TLS header.
1100  */
1101 bool
1102 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1103 {
1104 	struct tls_record_layer hdr;
1105 	struct mbuf *m;
1106 	uint64_t seqno;
1107 	size_t resid;
1108 	u_int offset, record_len;
1109 
1110 	SOCKBUF_LOCK_ASSERT(sb);
1111 	MPASS(sb->sb_flags & SB_TLS_RX);
1112 	seqno = sb->sb_tls_seqno;
1113 	resid = sb->sb_tlscc;
1114 	m = sb->sb_mtls;
1115 	offset = 0;
1116 
1117 	if (resid == 0) {
1118 		*seqnop = seqno;
1119 		*residp = 0;
1120 		return (true);
1121 	}
1122 
1123 	for (;;) {
1124 		seqno++;
1125 
1126 		if (resid < sizeof(hdr)) {
1127 			*seqnop = seqno;
1128 			*residp = sizeof(hdr) - resid;
1129 			return (false);
1130 		}
1131 
1132 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1133 
1134 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1135 		if (resid <= record_len) {
1136 			*seqnop = seqno;
1137 			*residp = record_len - resid;
1138 			return (true);
1139 		}
1140 		resid -= record_len;
1141 
1142 		while (record_len != 0) {
1143 			if (m->m_len - offset > record_len) {
1144 				offset += record_len;
1145 				break;
1146 			}
1147 
1148 			record_len -= (m->m_len - offset);
1149 			offset = 0;
1150 			m = m->m_next;
1151 		}
1152 	}
1153 }
1154 
1155 int
1156 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1157 {
1158 	struct ktls_session *tls;
1159 	int error;
1160 
1161 	if (!ktls_offload_enable)
1162 		return (ENOTSUP);
1163 	if (SOLISTENING(so))
1164 		return (EINVAL);
1165 
1166 	counter_u64_add(ktls_offload_enable_calls, 1);
1167 
1168 	/*
1169 	 * This should always be true since only the TCP socket option
1170 	 * invokes this function.
1171 	 */
1172 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1173 		return (EINVAL);
1174 
1175 	/*
1176 	 * XXX: Don't overwrite existing sessions.  We should permit
1177 	 * this to support rekeying in the future.
1178 	 */
1179 	if (so->so_rcv.sb_tls_info != NULL)
1180 		return (EALREADY);
1181 
1182 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1183 		return (ENOTSUP);
1184 
1185 	/* TLS 1.3 is not yet supported. */
1186 	if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1187 	    en->tls_vminor == TLS_MINOR_VER_THREE)
1188 		return (ENOTSUP);
1189 
1190 	error = ktls_create_session(so, en, &tls);
1191 	if (error)
1192 		return (error);
1193 
1194 	error = ktls_ocf_try(so, tls, KTLS_RX);
1195 	if (error) {
1196 		ktls_cleanup(tls);
1197 		return (error);
1198 	}
1199 
1200 #ifdef TCP_OFFLOAD
1201 	error = ktls_try_toe(so, tls, KTLS_RX);
1202 	if (error)
1203 #endif
1204 		ktls_use_sw(tls);
1205 
1206 	/* Mark the socket as using TLS offload. */
1207 	SOCKBUF_LOCK(&so->so_rcv);
1208 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1209 	so->so_rcv.sb_tls_info = tls;
1210 	so->so_rcv.sb_flags |= SB_TLS_RX;
1211 
1212 	/* Mark existing data as not ready until it can be decrypted. */
1213 	if (tls->mode != TCP_TLS_MODE_TOE) {
1214 		sb_mark_notready(&so->so_rcv);
1215 		ktls_check_rx(&so->so_rcv);
1216 	}
1217 	SOCKBUF_UNLOCK(&so->so_rcv);
1218 
1219 	counter_u64_add(ktls_offload_total, 1);
1220 
1221 	return (0);
1222 }
1223 
1224 int
1225 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1226 {
1227 	struct ktls_session *tls;
1228 	struct inpcb *inp;
1229 	int error;
1230 
1231 	if (!ktls_offload_enable)
1232 		return (ENOTSUP);
1233 	if (SOLISTENING(so))
1234 		return (EINVAL);
1235 
1236 	counter_u64_add(ktls_offload_enable_calls, 1);
1237 
1238 	/*
1239 	 * This should always be true since only the TCP socket option
1240 	 * invokes this function.
1241 	 */
1242 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1243 		return (EINVAL);
1244 
1245 	/*
1246 	 * XXX: Don't overwrite existing sessions.  We should permit
1247 	 * this to support rekeying in the future.
1248 	 */
1249 	if (so->so_snd.sb_tls_info != NULL)
1250 		return (EALREADY);
1251 
1252 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1253 		return (ENOTSUP);
1254 
1255 	/* TLS requires ext pgs */
1256 	if (mb_use_ext_pgs == 0)
1257 		return (ENXIO);
1258 
1259 	error = ktls_create_session(so, en, &tls);
1260 	if (error)
1261 		return (error);
1262 
1263 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1264 #ifdef TCP_OFFLOAD
1265 	error = ktls_try_toe(so, tls, KTLS_TX);
1266 	if (error)
1267 #endif
1268 		error = ktls_try_ifnet(so, tls, false);
1269 	if (error)
1270 		error = ktls_try_sw(so, tls, KTLS_TX);
1271 
1272 	if (error) {
1273 		ktls_cleanup(tls);
1274 		return (error);
1275 	}
1276 
1277 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1278 	if (error) {
1279 		ktls_cleanup(tls);
1280 		return (error);
1281 	}
1282 
1283 	/*
1284 	 * Write lock the INP when setting sb_tls_info so that
1285 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1286 	 * holding the INP lock.
1287 	 */
1288 	inp = so->so_pcb;
1289 	INP_WLOCK(inp);
1290 	SOCKBUF_LOCK(&so->so_snd);
1291 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1292 	so->so_snd.sb_tls_info = tls;
1293 	if (tls->mode != TCP_TLS_MODE_SW)
1294 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1295 	SOCKBUF_UNLOCK(&so->so_snd);
1296 	INP_WUNLOCK(inp);
1297 	SOCK_IO_SEND_UNLOCK(so);
1298 
1299 	counter_u64_add(ktls_offload_total, 1);
1300 
1301 	return (0);
1302 }
1303 
1304 int
1305 ktls_get_rx_mode(struct socket *so, int *modep)
1306 {
1307 	struct ktls_session *tls;
1308 	struct inpcb *inp;
1309 
1310 	if (SOLISTENING(so))
1311 		return (EINVAL);
1312 	inp = so->so_pcb;
1313 	INP_WLOCK_ASSERT(inp);
1314 	SOCK_RECVBUF_LOCK(so);
1315 	tls = so->so_rcv.sb_tls_info;
1316 	if (tls == NULL)
1317 		*modep = TCP_TLS_MODE_NONE;
1318 	else
1319 		*modep = tls->mode;
1320 	SOCK_RECVBUF_UNLOCK(so);
1321 	return (0);
1322 }
1323 
1324 int
1325 ktls_get_tx_mode(struct socket *so, int *modep)
1326 {
1327 	struct ktls_session *tls;
1328 	struct inpcb *inp;
1329 
1330 	if (SOLISTENING(so))
1331 		return (EINVAL);
1332 	inp = so->so_pcb;
1333 	INP_WLOCK_ASSERT(inp);
1334 	SOCK_SENDBUF_LOCK(so);
1335 	tls = so->so_snd.sb_tls_info;
1336 	if (tls == NULL)
1337 		*modep = TCP_TLS_MODE_NONE;
1338 	else
1339 		*modep = tls->mode;
1340 	SOCK_SENDBUF_UNLOCK(so);
1341 	return (0);
1342 }
1343 
1344 /*
1345  * Switch between SW and ifnet TLS sessions as requested.
1346  */
1347 int
1348 ktls_set_tx_mode(struct socket *so, int mode)
1349 {
1350 	struct ktls_session *tls, *tls_new;
1351 	struct inpcb *inp;
1352 	int error;
1353 
1354 	if (SOLISTENING(so))
1355 		return (EINVAL);
1356 	switch (mode) {
1357 	case TCP_TLS_MODE_SW:
1358 	case TCP_TLS_MODE_IFNET:
1359 		break;
1360 	default:
1361 		return (EINVAL);
1362 	}
1363 
1364 	inp = so->so_pcb;
1365 	INP_WLOCK_ASSERT(inp);
1366 	SOCKBUF_LOCK(&so->so_snd);
1367 	tls = so->so_snd.sb_tls_info;
1368 	if (tls == NULL) {
1369 		SOCKBUF_UNLOCK(&so->so_snd);
1370 		return (0);
1371 	}
1372 
1373 	if (tls->mode == mode) {
1374 		SOCKBUF_UNLOCK(&so->so_snd);
1375 		return (0);
1376 	}
1377 
1378 	tls = ktls_hold(tls);
1379 	SOCKBUF_UNLOCK(&so->so_snd);
1380 	INP_WUNLOCK(inp);
1381 
1382 	tls_new = ktls_clone_session(tls);
1383 
1384 	if (mode == TCP_TLS_MODE_IFNET)
1385 		error = ktls_try_ifnet(so, tls_new, true);
1386 	else
1387 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1388 	if (error) {
1389 		counter_u64_add(ktls_switch_failed, 1);
1390 		ktls_free(tls_new);
1391 		ktls_free(tls);
1392 		INP_WLOCK(inp);
1393 		return (error);
1394 	}
1395 
1396 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1397 	if (error) {
1398 		counter_u64_add(ktls_switch_failed, 1);
1399 		ktls_free(tls_new);
1400 		ktls_free(tls);
1401 		INP_WLOCK(inp);
1402 		return (error);
1403 	}
1404 
1405 	/*
1406 	 * If we raced with another session change, keep the existing
1407 	 * session.
1408 	 */
1409 	if (tls != so->so_snd.sb_tls_info) {
1410 		counter_u64_add(ktls_switch_failed, 1);
1411 		SOCK_IO_SEND_UNLOCK(so);
1412 		ktls_free(tls_new);
1413 		ktls_free(tls);
1414 		INP_WLOCK(inp);
1415 		return (EBUSY);
1416 	}
1417 
1418 	SOCKBUF_LOCK(&so->so_snd);
1419 	so->so_snd.sb_tls_info = tls_new;
1420 	if (tls_new->mode != TCP_TLS_MODE_SW)
1421 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1422 	SOCKBUF_UNLOCK(&so->so_snd);
1423 	SOCK_IO_SEND_UNLOCK(so);
1424 
1425 	/*
1426 	 * Drop two references on 'tls'.  The first is for the
1427 	 * ktls_hold() above.  The second drops the reference from the
1428 	 * socket buffer.
1429 	 */
1430 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1431 	ktls_free(tls);
1432 	ktls_free(tls);
1433 
1434 	if (mode == TCP_TLS_MODE_IFNET)
1435 		counter_u64_add(ktls_switch_to_ifnet, 1);
1436 	else
1437 		counter_u64_add(ktls_switch_to_sw, 1);
1438 
1439 	INP_WLOCK(inp);
1440 	return (0);
1441 }
1442 
1443 /*
1444  * Try to allocate a new TLS send tag.  This task is scheduled when
1445  * ip_output detects a route change while trying to transmit a packet
1446  * holding a TLS record.  If a new tag is allocated, replace the tag
1447  * in the TLS session.  Subsequent packets on the connection will use
1448  * the new tag.  If a new tag cannot be allocated, drop the
1449  * connection.
1450  */
1451 static void
1452 ktls_reset_send_tag(void *context, int pending)
1453 {
1454 	struct epoch_tracker et;
1455 	struct ktls_session *tls;
1456 	struct m_snd_tag *old, *new;
1457 	struct inpcb *inp;
1458 	struct tcpcb *tp;
1459 	int error;
1460 
1461 	MPASS(pending == 1);
1462 
1463 	tls = context;
1464 	inp = tls->inp;
1465 
1466 	/*
1467 	 * Free the old tag first before allocating a new one.
1468 	 * ip[6]_output_send() will treat a NULL send tag the same as
1469 	 * an ifp mismatch and drop packets until a new tag is
1470 	 * allocated.
1471 	 *
1472 	 * Write-lock the INP when changing tls->snd_tag since
1473 	 * ip[6]_output_send() holds a read-lock when reading the
1474 	 * pointer.
1475 	 */
1476 	INP_WLOCK(inp);
1477 	old = tls->snd_tag;
1478 	tls->snd_tag = NULL;
1479 	INP_WUNLOCK(inp);
1480 	if (old != NULL)
1481 		m_snd_tag_rele(old);
1482 
1483 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1484 
1485 	if (error == 0) {
1486 		INP_WLOCK(inp);
1487 		tls->snd_tag = new;
1488 		mtx_pool_lock(mtxpool_sleep, tls);
1489 		tls->reset_pending = false;
1490 		mtx_pool_unlock(mtxpool_sleep, tls);
1491 		if (!in_pcbrele_wlocked(inp))
1492 			INP_WUNLOCK(inp);
1493 
1494 		counter_u64_add(ktls_ifnet_reset, 1);
1495 
1496 		/*
1497 		 * XXX: Should we kick tcp_output explicitly now that
1498 		 * the send tag is fixed or just rely on timers?
1499 		 */
1500 	} else {
1501 		NET_EPOCH_ENTER(et);
1502 		INP_WLOCK(inp);
1503 		if (!in_pcbrele_wlocked(inp)) {
1504 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1505 			    !(inp->inp_flags & INP_DROPPED)) {
1506 				tp = intotcpcb(inp);
1507 				CURVNET_SET(tp->t_vnet);
1508 				tp = tcp_drop(tp, ECONNABORTED);
1509 				CURVNET_RESTORE();
1510 				if (tp != NULL)
1511 					INP_WUNLOCK(inp);
1512 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1513 			} else
1514 				INP_WUNLOCK(inp);
1515 		}
1516 		NET_EPOCH_EXIT(et);
1517 
1518 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1519 
1520 		/*
1521 		 * Leave reset_pending true to avoid future tasks while
1522 		 * the socket goes away.
1523 		 */
1524 	}
1525 
1526 	ktls_free(tls);
1527 }
1528 
1529 int
1530 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1531 {
1532 
1533 	if (inp == NULL)
1534 		return (ENOBUFS);
1535 
1536 	INP_LOCK_ASSERT(inp);
1537 
1538 	/*
1539 	 * See if we should schedule a task to update the send tag for
1540 	 * this session.
1541 	 */
1542 	mtx_pool_lock(mtxpool_sleep, tls);
1543 	if (!tls->reset_pending) {
1544 		(void) ktls_hold(tls);
1545 		in_pcbref(inp);
1546 		tls->inp = inp;
1547 		tls->reset_pending = true;
1548 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1549 	}
1550 	mtx_pool_unlock(mtxpool_sleep, tls);
1551 	return (ENOBUFS);
1552 }
1553 
1554 #ifdef RATELIMIT
1555 int
1556 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1557 {
1558 	union if_snd_tag_modify_params params = {
1559 		.rate_limit.max_rate = max_pacing_rate,
1560 		.rate_limit.flags = M_NOWAIT,
1561 	};
1562 	struct m_snd_tag *mst;
1563 
1564 	/* Can't get to the inp, but it should be locked. */
1565 	/* INP_LOCK_ASSERT(inp); */
1566 
1567 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1568 
1569 	if (tls->snd_tag == NULL) {
1570 		/*
1571 		 * Resetting send tag, ignore this change.  The
1572 		 * pending reset may or may not see this updated rate
1573 		 * in the tcpcb.  If it doesn't, we will just lose
1574 		 * this rate change.
1575 		 */
1576 		return (0);
1577 	}
1578 
1579 	MPASS(tls->snd_tag != NULL);
1580 	MPASS(tls->snd_tag->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1581 
1582 	mst = tls->snd_tag;
1583 	return (mst->sw->snd_tag_modify(mst, &params));
1584 }
1585 #endif
1586 #endif
1587 
1588 void
1589 ktls_destroy(struct ktls_session *tls)
1590 {
1591 
1592 	if (tls->sequential_records) {
1593 		struct mbuf *m, *n;
1594 		int page_count;
1595 
1596 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1597 			page_count = m->m_epg_enc_cnt;
1598 			while (page_count > 0) {
1599 				KASSERT(page_count >= m->m_epg_nrdy,
1600 				    ("%s: too few pages", __func__));
1601 				page_count -= m->m_epg_nrdy;
1602 				m = m_free(m);
1603 			}
1604 		}
1605 	}
1606 	ktls_cleanup(tls);
1607 	uma_zfree(ktls_session_zone, tls);
1608 }
1609 
1610 void
1611 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1612 {
1613 
1614 	for (; m != NULL; m = m->m_next) {
1615 		KASSERT((m->m_flags & M_EXTPG) != 0,
1616 		    ("ktls_seq: mapped mbuf %p", m));
1617 
1618 		m->m_epg_seqno = sb->sb_tls_seqno;
1619 		sb->sb_tls_seqno++;
1620 	}
1621 }
1622 
1623 /*
1624  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1625  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1626  * mbuf must be populated with the payload of each TLS record.
1627  *
1628  * The record_type argument specifies the TLS record type used when
1629  * populating the TLS header.
1630  *
1631  * The enq_count argument on return is set to the number of pages of
1632  * payload data for this entire chain that need to be encrypted via SW
1633  * encryption.  The returned value should be passed to ktls_enqueue
1634  * when scheduling encryption of this chain of mbufs.  To handle the
1635  * special case of empty fragments for TLS 1.0 sessions, an empty
1636  * fragment counts as one page.
1637  */
1638 void
1639 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1640     uint8_t record_type)
1641 {
1642 	struct tls_record_layer *tlshdr;
1643 	struct mbuf *m;
1644 	uint64_t *noncep;
1645 	uint16_t tls_len;
1646 	int maxlen;
1647 
1648 	maxlen = tls->params.max_frame_len;
1649 	*enq_cnt = 0;
1650 	for (m = top; m != NULL; m = m->m_next) {
1651 		/*
1652 		 * All mbufs in the chain should be TLS records whose
1653 		 * payload does not exceed the maximum frame length.
1654 		 *
1655 		 * Empty TLS records are permitted when using CBC.
1656 		 */
1657 		KASSERT(m->m_len <= maxlen &&
1658 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1659 		    m->m_len >= 0 : m->m_len > 0),
1660 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1661 
1662 		/*
1663 		 * TLS frames require unmapped mbufs to store session
1664 		 * info.
1665 		 */
1666 		KASSERT((m->m_flags & M_EXTPG) != 0,
1667 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1668 
1669 		tls_len = m->m_len;
1670 
1671 		/* Save a reference to the session. */
1672 		m->m_epg_tls = ktls_hold(tls);
1673 
1674 		m->m_epg_hdrlen = tls->params.tls_hlen;
1675 		m->m_epg_trllen = tls->params.tls_tlen;
1676 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1677 			int bs, delta;
1678 
1679 			/*
1680 			 * AES-CBC pads messages to a multiple of the
1681 			 * block size.  Note that the padding is
1682 			 * applied after the digest and the encryption
1683 			 * is done on the "plaintext || mac || padding".
1684 			 * At least one byte of padding is always
1685 			 * present.
1686 			 *
1687 			 * Compute the final trailer length assuming
1688 			 * at most one block of padding.
1689 			 * tls->params.tls_tlen is the maximum
1690 			 * possible trailer length (padding + digest).
1691 			 * delta holds the number of excess padding
1692 			 * bytes if the maximum were used.  Those
1693 			 * extra bytes are removed.
1694 			 */
1695 			bs = tls->params.tls_bs;
1696 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1697 			m->m_epg_trllen -= delta;
1698 		}
1699 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1700 
1701 		/* Populate the TLS header. */
1702 		tlshdr = (void *)m->m_epg_hdr;
1703 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1704 
1705 		/*
1706 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1707 		 * of TLS_RLTYPE_APP.
1708 		 */
1709 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1710 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1711 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1712 			tlshdr->tls_type = TLS_RLTYPE_APP;
1713 			/* save the real record type for later */
1714 			m->m_epg_record_type = record_type;
1715 			m->m_epg_trail[0] = record_type;
1716 		} else {
1717 			tlshdr->tls_vminor = tls->params.tls_vminor;
1718 			tlshdr->tls_type = record_type;
1719 		}
1720 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1721 
1722 		/*
1723 		 * Store nonces / explicit IVs after the end of the
1724 		 * TLS header.
1725 		 *
1726 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1727 		 * from the end of the IV.  The nonce is then
1728 		 * incremented for use by the next record.
1729 		 *
1730 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1731 		 */
1732 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1733 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1734 			noncep = (uint64_t *)(tls->params.iv + 8);
1735 			be64enc(tlshdr + 1, *noncep);
1736 			(*noncep)++;
1737 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1738 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1739 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1740 
1741 		/*
1742 		 * When using SW encryption, mark the mbuf not ready.
1743 		 * It will be marked ready via sbready() after the
1744 		 * record has been encrypted.
1745 		 *
1746 		 * When using ifnet TLS, unencrypted TLS records are
1747 		 * sent down the stack to the NIC.
1748 		 */
1749 		if (tls->mode == TCP_TLS_MODE_SW) {
1750 			m->m_flags |= M_NOTREADY;
1751 			if (__predict_false(tls_len == 0)) {
1752 				/* TLS 1.0 empty fragment. */
1753 				m->m_epg_nrdy = 1;
1754 			} else
1755 				m->m_epg_nrdy = m->m_epg_npgs;
1756 			*enq_cnt += m->m_epg_nrdy;
1757 		}
1758 	}
1759 }
1760 
1761 void
1762 ktls_check_rx(struct sockbuf *sb)
1763 {
1764 	struct tls_record_layer hdr;
1765 	struct ktls_wq *wq;
1766 	struct socket *so;
1767 	bool running;
1768 
1769 	SOCKBUF_LOCK_ASSERT(sb);
1770 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1771 	    __func__, sb));
1772 	so = __containerof(sb, struct socket, so_rcv);
1773 
1774 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1775 		return;
1776 
1777 	/* Is there enough queued for a TLS header? */
1778 	if (sb->sb_tlscc < sizeof(hdr)) {
1779 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1780 			so->so_error = EMSGSIZE;
1781 		return;
1782 	}
1783 
1784 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1785 
1786 	/* Is the entire record queued? */
1787 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1788 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1789 			so->so_error = EMSGSIZE;
1790 		return;
1791 	}
1792 
1793 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1794 
1795 	soref(so);
1796 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1797 	mtx_lock(&wq->mtx);
1798 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1799 	running = wq->running;
1800 	mtx_unlock(&wq->mtx);
1801 	if (!running)
1802 		wakeup(wq);
1803 	counter_u64_add(ktls_cnt_rx_queued, 1);
1804 }
1805 
1806 static struct mbuf *
1807 ktls_detach_record(struct sockbuf *sb, int len)
1808 {
1809 	struct mbuf *m, *n, *top;
1810 	int remain;
1811 
1812 	SOCKBUF_LOCK_ASSERT(sb);
1813 	MPASS(len <= sb->sb_tlscc);
1814 
1815 	/*
1816 	 * If TLS chain is the exact size of the record,
1817 	 * just grab the whole record.
1818 	 */
1819 	top = sb->sb_mtls;
1820 	if (sb->sb_tlscc == len) {
1821 		sb->sb_mtls = NULL;
1822 		sb->sb_mtlstail = NULL;
1823 		goto out;
1824 	}
1825 
1826 	/*
1827 	 * While it would be nice to use m_split() here, we need
1828 	 * to know exactly what m_split() allocates to update the
1829 	 * accounting, so do it inline instead.
1830 	 */
1831 	remain = len;
1832 	for (m = top; remain > m->m_len; m = m->m_next)
1833 		remain -= m->m_len;
1834 
1835 	/* Easy case: don't have to split 'm'. */
1836 	if (remain == m->m_len) {
1837 		sb->sb_mtls = m->m_next;
1838 		if (sb->sb_mtls == NULL)
1839 			sb->sb_mtlstail = NULL;
1840 		m->m_next = NULL;
1841 		goto out;
1842 	}
1843 
1844 	/*
1845 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1846 	 * with M_NOWAIT first.
1847 	 */
1848 	n = m_get(M_NOWAIT, MT_DATA);
1849 	if (n == NULL) {
1850 		/*
1851 		 * Use M_WAITOK with socket buffer unlocked.  If
1852 		 * 'sb_mtls' changes while the lock is dropped, return
1853 		 * NULL to force the caller to retry.
1854 		 */
1855 		SOCKBUF_UNLOCK(sb);
1856 
1857 		n = m_get(M_WAITOK, MT_DATA);
1858 
1859 		SOCKBUF_LOCK(sb);
1860 		if (sb->sb_mtls != top) {
1861 			m_free(n);
1862 			return (NULL);
1863 		}
1864 	}
1865 	n->m_flags |= M_NOTREADY;
1866 
1867 	/* Store remainder in 'n'. */
1868 	n->m_len = m->m_len - remain;
1869 	if (m->m_flags & M_EXT) {
1870 		n->m_data = m->m_data + remain;
1871 		mb_dupcl(n, m);
1872 	} else {
1873 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1874 	}
1875 
1876 	/* Trim 'm' and update accounting. */
1877 	m->m_len -= n->m_len;
1878 	sb->sb_tlscc -= n->m_len;
1879 	sb->sb_ccc -= n->m_len;
1880 
1881 	/* Account for 'n'. */
1882 	sballoc_ktls_rx(sb, n);
1883 
1884 	/* Insert 'n' into the TLS chain. */
1885 	sb->sb_mtls = n;
1886 	n->m_next = m->m_next;
1887 	if (sb->sb_mtlstail == m)
1888 		sb->sb_mtlstail = n;
1889 
1890 	/* Detach the record from the TLS chain. */
1891 	m->m_next = NULL;
1892 
1893 out:
1894 	MPASS(m_length(top, NULL) == len);
1895 	for (m = top; m != NULL; m = m->m_next)
1896 		sbfree_ktls_rx(sb, m);
1897 	sb->sb_tlsdcc = len;
1898 	sb->sb_ccc += len;
1899 	SBCHECK(sb);
1900 	return (top);
1901 }
1902 
1903 static void
1904 ktls_decrypt(struct socket *so)
1905 {
1906 	char tls_header[MBUF_PEXT_HDR_LEN];
1907 	struct ktls_session *tls;
1908 	struct sockbuf *sb;
1909 	struct tls_record_layer *hdr;
1910 	struct tls_get_record tgr;
1911 	struct mbuf *control, *data, *m;
1912 	uint64_t seqno;
1913 	int error, remain, tls_len, trail_len;
1914 
1915 	hdr = (struct tls_record_layer *)tls_header;
1916 	sb = &so->so_rcv;
1917 	SOCKBUF_LOCK(sb);
1918 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1919 	    ("%s: socket %p not running", __func__, so));
1920 
1921 	tls = sb->sb_tls_info;
1922 	MPASS(tls != NULL);
1923 
1924 	for (;;) {
1925 		/* Is there enough queued for a TLS header? */
1926 		if (sb->sb_tlscc < tls->params.tls_hlen)
1927 			break;
1928 
1929 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1930 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1931 
1932 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1933 		    hdr->tls_vminor != tls->params.tls_vminor)
1934 			error = EINVAL;
1935 		else if (tls_len < tls->params.tls_hlen || tls_len >
1936 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1937 		    tls->params.tls_tlen)
1938 			error = EMSGSIZE;
1939 		else
1940 			error = 0;
1941 		if (__predict_false(error != 0)) {
1942 			/*
1943 			 * We have a corrupted record and are likely
1944 			 * out of sync.  The connection isn't
1945 			 * recoverable at this point, so abort it.
1946 			 */
1947 			SOCKBUF_UNLOCK(sb);
1948 			counter_u64_add(ktls_offload_corrupted_records, 1);
1949 
1950 			CURVNET_SET(so->so_vnet);
1951 			so->so_proto->pr_usrreqs->pru_abort(so);
1952 			so->so_error = error;
1953 			CURVNET_RESTORE();
1954 			goto deref;
1955 		}
1956 
1957 		/* Is the entire record queued? */
1958 		if (sb->sb_tlscc < tls_len)
1959 			break;
1960 
1961 		/*
1962 		 * Split out the portion of the mbuf chain containing
1963 		 * this TLS record.
1964 		 */
1965 		data = ktls_detach_record(sb, tls_len);
1966 		if (data == NULL)
1967 			continue;
1968 		MPASS(sb->sb_tlsdcc == tls_len);
1969 
1970 		seqno = sb->sb_tls_seqno;
1971 		sb->sb_tls_seqno++;
1972 		SBCHECK(sb);
1973 		SOCKBUF_UNLOCK(sb);
1974 
1975 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1976 		if (error) {
1977 			counter_u64_add(ktls_offload_failed_crypto, 1);
1978 
1979 			SOCKBUF_LOCK(sb);
1980 			if (sb->sb_tlsdcc == 0) {
1981 				/*
1982 				 * sbcut/drop/flush discarded these
1983 				 * mbufs.
1984 				 */
1985 				m_freem(data);
1986 				break;
1987 			}
1988 
1989 			/*
1990 			 * Drop this TLS record's data, but keep
1991 			 * decrypting subsequent records.
1992 			 */
1993 			sb->sb_ccc -= tls_len;
1994 			sb->sb_tlsdcc = 0;
1995 
1996 			CURVNET_SET(so->so_vnet);
1997 			so->so_error = EBADMSG;
1998 			sorwakeup_locked(so);
1999 			CURVNET_RESTORE();
2000 
2001 			m_freem(data);
2002 
2003 			SOCKBUF_LOCK(sb);
2004 			continue;
2005 		}
2006 
2007 		/* Allocate the control mbuf. */
2008 		tgr.tls_type = hdr->tls_type;
2009 		tgr.tls_vmajor = hdr->tls_vmajor;
2010 		tgr.tls_vminor = hdr->tls_vminor;
2011 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2012 		    trail_len);
2013 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
2014 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2015 
2016 		SOCKBUF_LOCK(sb);
2017 		if (sb->sb_tlsdcc == 0) {
2018 			/* sbcut/drop/flush discarded these mbufs. */
2019 			MPASS(sb->sb_tlscc == 0);
2020 			m_freem(data);
2021 			m_freem(control);
2022 			break;
2023 		}
2024 
2025 		/*
2026 		 * Clear the 'dcc' accounting in preparation for
2027 		 * adding the decrypted record.
2028 		 */
2029 		sb->sb_ccc -= tls_len;
2030 		sb->sb_tlsdcc = 0;
2031 		SBCHECK(sb);
2032 
2033 		/* If there is no payload, drop all of the data. */
2034 		if (tgr.tls_length == htobe16(0)) {
2035 			m_freem(data);
2036 			data = NULL;
2037 		} else {
2038 			/* Trim header. */
2039 			remain = tls->params.tls_hlen;
2040 			while (remain > 0) {
2041 				if (data->m_len > remain) {
2042 					data->m_data += remain;
2043 					data->m_len -= remain;
2044 					break;
2045 				}
2046 				remain -= data->m_len;
2047 				data = m_free(data);
2048 			}
2049 
2050 			/* Trim trailer and clear M_NOTREADY. */
2051 			remain = be16toh(tgr.tls_length);
2052 			m = data;
2053 			for (m = data; remain > m->m_len; m = m->m_next) {
2054 				m->m_flags &= ~M_NOTREADY;
2055 				remain -= m->m_len;
2056 			}
2057 			m->m_len = remain;
2058 			m_freem(m->m_next);
2059 			m->m_next = NULL;
2060 			m->m_flags &= ~M_NOTREADY;
2061 
2062 			/* Set EOR on the final mbuf. */
2063 			m->m_flags |= M_EOR;
2064 		}
2065 
2066 		sbappendcontrol_locked(sb, data, control, 0);
2067 	}
2068 
2069 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2070 
2071 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2072 		so->so_error = EMSGSIZE;
2073 
2074 	sorwakeup_locked(so);
2075 
2076 deref:
2077 	SOCKBUF_UNLOCK_ASSERT(sb);
2078 
2079 	CURVNET_SET(so->so_vnet);
2080 	sorele(so);
2081 	CURVNET_RESTORE();
2082 }
2083 
2084 void
2085 ktls_enqueue_to_free(struct mbuf *m)
2086 {
2087 	struct ktls_wq *wq;
2088 	bool running;
2089 
2090 	/* Mark it for freeing. */
2091 	m->m_epg_flags |= EPG_FLAG_2FREE;
2092 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2093 	mtx_lock(&wq->mtx);
2094 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2095 	running = wq->running;
2096 	mtx_unlock(&wq->mtx);
2097 	if (!running)
2098 		wakeup(wq);
2099 }
2100 
2101 static void *
2102 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2103 {
2104 	void *buf;
2105 	int domain, running;
2106 
2107 	if (m->m_epg_npgs <= 2)
2108 		return (NULL);
2109 	if (ktls_buffer_zone == NULL)
2110 		return (NULL);
2111 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2112 		/*
2113 		 * Rate-limit allocation attempts after a failure.
2114 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2115 		 * the free page queues and may fail consistently if memory is
2116 		 * fragmented.
2117 		 */
2118 		return (NULL);
2119 	}
2120 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2121 	if (buf == NULL) {
2122 		domain = PCPU_GET(domain);
2123 		wq->lastallocfail = ticks;
2124 
2125 		/*
2126 		 * Note that this check is "racy", but the races are
2127 		 * harmless, and are either a spurious wakeup if
2128 		 * multiple threads fail allocations before the alloc
2129 		 * thread wakes, or waiting an extra second in case we
2130 		 * see an old value of running == true.
2131 		 */
2132 		if (!VM_DOMAIN_EMPTY(domain)) {
2133 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2134 			if (!running)
2135 				wakeup(&ktls_domains[domain].alloc_td);
2136 		}
2137 	}
2138 	return (buf);
2139 }
2140 
2141 static int
2142 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2143     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2144 {
2145 	vm_page_t pg;
2146 	int error, i, len, off;
2147 
2148 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2149 	    ("%p not unready & nomap mbuf\n", m));
2150 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2151 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2152 	    ktls_maxlen));
2153 
2154 	/* Anonymous mbufs are encrypted in place. */
2155 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2156 		return (tls->sw_encrypt(state, tls, m, NULL, 0));
2157 
2158 	/*
2159 	 * For file-backed mbufs (from sendfile), anonymous wired
2160 	 * pages are allocated and used as the encryption destination.
2161 	 */
2162 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2163 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2164 		    m->m_epg_1st_off;
2165 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2166 		    m->m_epg_1st_off;
2167 		state->dst_iov[0].iov_len = len;
2168 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2169 		i = 1;
2170 	} else {
2171 		off = m->m_epg_1st_off;
2172 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2173 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2174 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2175 			len = m_epg_pagelen(m, i, off);
2176 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2177 			state->dst_iov[i].iov_base =
2178 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2179 			state->dst_iov[i].iov_len = len;
2180 		}
2181 	}
2182 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2183 	state->dst_iov[i].iov_base = m->m_epg_trail;
2184 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2185 
2186 	error = tls->sw_encrypt(state, tls, m, state->dst_iov, i + 1);
2187 
2188 	if (__predict_false(error != 0)) {
2189 		/* Free the anonymous pages. */
2190 		if (state->cbuf != NULL)
2191 			uma_zfree(ktls_buffer_zone, state->cbuf);
2192 		else {
2193 			for (i = 0; i < m->m_epg_npgs; i++) {
2194 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2195 				(void)vm_page_unwire_noq(pg);
2196 				vm_page_free(pg);
2197 			}
2198 		}
2199 	}
2200 	return (error);
2201 }
2202 
2203 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2204 static u_int
2205 ktls_batched_records(struct mbuf *m)
2206 {
2207 	int page_count, records;
2208 
2209 	records = 0;
2210 	page_count = m->m_epg_enc_cnt;
2211 	while (page_count > 0) {
2212 		records++;
2213 		page_count -= m->m_epg_nrdy;
2214 		m = m->m_next;
2215 	}
2216 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2217 	return (records);
2218 }
2219 
2220 void
2221 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2222 {
2223 	struct ktls_session *tls;
2224 	struct ktls_wq *wq;
2225 	int queued;
2226 	bool running;
2227 
2228 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2229 	    (M_EXTPG | M_NOTREADY)),
2230 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2231 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2232 
2233 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2234 
2235 	m->m_epg_enc_cnt = page_count;
2236 
2237 	/*
2238 	 * Save a pointer to the socket.  The caller is responsible
2239 	 * for taking an additional reference via soref().
2240 	 */
2241 	m->m_epg_so = so;
2242 
2243 	queued = 1;
2244 	tls = m->m_epg_tls;
2245 	wq = &ktls_wq[tls->wq_index];
2246 	mtx_lock(&wq->mtx);
2247 	if (__predict_false(tls->sequential_records)) {
2248 		/*
2249 		 * For TLS 1.0, records must be encrypted
2250 		 * sequentially.  For a given connection, all records
2251 		 * queued to the associated work queue are processed
2252 		 * sequentially.  However, sendfile(2) might complete
2253 		 * I/O requests spanning multiple TLS records out of
2254 		 * order.  Here we ensure TLS records are enqueued to
2255 		 * the work queue in FIFO order.
2256 		 *
2257 		 * tls->next_seqno holds the sequence number of the
2258 		 * next TLS record that should be enqueued to the work
2259 		 * queue.  If this next record is not tls->next_seqno,
2260 		 * it must be a future record, so insert it, sorted by
2261 		 * TLS sequence number, into tls->pending_records and
2262 		 * return.
2263 		 *
2264 		 * If this TLS record matches tls->next_seqno, place
2265 		 * it in the work queue and then check
2266 		 * tls->pending_records to see if any
2267 		 * previously-queued records are now ready for
2268 		 * encryption.
2269 		 */
2270 		if (m->m_epg_seqno != tls->next_seqno) {
2271 			struct mbuf *n, *p;
2272 
2273 			p = NULL;
2274 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2275 				if (n->m_epg_seqno > m->m_epg_seqno)
2276 					break;
2277 				p = n;
2278 			}
2279 			if (n == NULL)
2280 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2281 				    m_epg_stailq);
2282 			else if (p == NULL)
2283 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2284 				    m_epg_stailq);
2285 			else
2286 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2287 				    m_epg_stailq);
2288 			mtx_unlock(&wq->mtx);
2289 			counter_u64_add(ktls_cnt_tx_pending, 1);
2290 			return;
2291 		}
2292 
2293 		tls->next_seqno += ktls_batched_records(m);
2294 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2295 
2296 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2297 			struct mbuf *n;
2298 
2299 			n = STAILQ_FIRST(&tls->pending_records);
2300 			if (n->m_epg_seqno != tls->next_seqno)
2301 				break;
2302 
2303 			queued++;
2304 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2305 			tls->next_seqno += ktls_batched_records(n);
2306 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2307 		}
2308 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2309 	} else
2310 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2311 
2312 	running = wq->running;
2313 	mtx_unlock(&wq->mtx);
2314 	if (!running)
2315 		wakeup(wq);
2316 	counter_u64_add(ktls_cnt_tx_queued, queued);
2317 }
2318 
2319 /*
2320  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2321  * the pages from the file and replace them with the anonymous pages
2322  * allocated in ktls_encrypt_record().
2323  */
2324 static void
2325 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2326 {
2327 	int i;
2328 
2329 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2330 
2331 	/* Free the old pages. */
2332 	m->m_ext.ext_free(m);
2333 
2334 	/* Replace them with the new pages. */
2335 	if (state->cbuf != NULL) {
2336 		for (i = 0; i < m->m_epg_npgs; i++)
2337 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2338 
2339 		/* Contig pages should go back to the cache. */
2340 		m->m_ext.ext_free = ktls_free_mext_contig;
2341 	} else {
2342 		for (i = 0; i < m->m_epg_npgs; i++)
2343 			m->m_epg_pa[i] = state->parray[i];
2344 
2345 		/* Use the basic free routine. */
2346 		m->m_ext.ext_free = mb_free_mext_pgs;
2347 	}
2348 
2349 	/* Pages are now writable. */
2350 	m->m_epg_flags |= EPG_FLAG_ANON;
2351 }
2352 
2353 static __noinline void
2354 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2355 {
2356 	struct ktls_ocf_encrypt_state state;
2357 	struct ktls_session *tls;
2358 	struct socket *so;
2359 	struct mbuf *m;
2360 	int error, npages, total_pages;
2361 
2362 	so = top->m_epg_so;
2363 	tls = top->m_epg_tls;
2364 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2365 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2366 #ifdef INVARIANTS
2367 	top->m_epg_so = NULL;
2368 #endif
2369 	total_pages = top->m_epg_enc_cnt;
2370 	npages = 0;
2371 
2372 	/*
2373 	 * Encrypt the TLS records in the chain of mbufs starting with
2374 	 * 'top'.  'total_pages' gives us a total count of pages and is
2375 	 * used to know when we have finished encrypting the TLS
2376 	 * records originally queued with 'top'.
2377 	 *
2378 	 * NB: These mbufs are queued in the socket buffer and
2379 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2380 	 * socket buffer lock is not held while traversing this chain.
2381 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2382 	 * pointers should be stable.  However, the 'm_next' of the
2383 	 * last mbuf encrypted is not necessarily NULL.  It can point
2384 	 * to other mbufs appended while 'top' was on the TLS work
2385 	 * queue.
2386 	 *
2387 	 * Each mbuf holds an entire TLS record.
2388 	 */
2389 	error = 0;
2390 	for (m = top; npages != total_pages; m = m->m_next) {
2391 		KASSERT(m->m_epg_tls == tls,
2392 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2393 		    tls, m->m_epg_tls));
2394 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2395 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2396 		    total_pages, m));
2397 
2398 		error = ktls_encrypt_record(wq, m, tls, &state);
2399 		if (error) {
2400 			counter_u64_add(ktls_offload_failed_crypto, 1);
2401 			break;
2402 		}
2403 
2404 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2405 			ktls_finish_nonanon(m, &state);
2406 
2407 		npages += m->m_epg_nrdy;
2408 
2409 		/*
2410 		 * Drop a reference to the session now that it is no
2411 		 * longer needed.  Existing code depends on encrypted
2412 		 * records having no associated session vs
2413 		 * yet-to-be-encrypted records having an associated
2414 		 * session.
2415 		 */
2416 		m->m_epg_tls = NULL;
2417 		ktls_free(tls);
2418 	}
2419 
2420 	CURVNET_SET(so->so_vnet);
2421 	if (error == 0) {
2422 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2423 	} else {
2424 		so->so_proto->pr_usrreqs->pru_abort(so);
2425 		so->so_error = EIO;
2426 		mb_free_notready(top, total_pages);
2427 	}
2428 
2429 	sorele(so);
2430 	CURVNET_RESTORE();
2431 }
2432 
2433 void
2434 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2435 {
2436 	struct ktls_session *tls;
2437 	struct socket *so;
2438 	struct mbuf *m;
2439 	int npages;
2440 
2441 	m = state->m;
2442 
2443 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2444 		ktls_finish_nonanon(m, state);
2445 
2446 	so = state->so;
2447 	free(state, M_KTLS);
2448 
2449 	/*
2450 	 * Drop a reference to the session now that it is no longer
2451 	 * needed.  Existing code depends on encrypted records having
2452 	 * no associated session vs yet-to-be-encrypted records having
2453 	 * an associated session.
2454 	 */
2455 	tls = m->m_epg_tls;
2456 	m->m_epg_tls = NULL;
2457 	ktls_free(tls);
2458 
2459 	if (error != 0)
2460 		counter_u64_add(ktls_offload_failed_crypto, 1);
2461 
2462 	CURVNET_SET(so->so_vnet);
2463 	npages = m->m_epg_nrdy;
2464 
2465 	if (error == 0) {
2466 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, m, npages);
2467 	} else {
2468 		so->so_proto->pr_usrreqs->pru_abort(so);
2469 		so->so_error = EIO;
2470 		mb_free_notready(m, npages);
2471 	}
2472 
2473 	sorele(so);
2474 	CURVNET_RESTORE();
2475 }
2476 
2477 /*
2478  * Similar to ktls_encrypt, but used with asynchronous OCF backends
2479  * (coprocessors) where encryption does not use host CPU resources and
2480  * it can be beneficial to queue more requests than CPUs.
2481  */
2482 static __noinline void
2483 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2484 {
2485 	struct ktls_ocf_encrypt_state *state;
2486 	struct ktls_session *tls;
2487 	struct socket *so;
2488 	struct mbuf *m, *n;
2489 	int error, mpages, npages, total_pages;
2490 
2491 	so = top->m_epg_so;
2492 	tls = top->m_epg_tls;
2493 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2494 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2495 #ifdef INVARIANTS
2496 	top->m_epg_so = NULL;
2497 #endif
2498 	total_pages = top->m_epg_enc_cnt;
2499 	npages = 0;
2500 
2501 	error = 0;
2502 	for (m = top; npages != total_pages; m = n) {
2503 		KASSERT(m->m_epg_tls == tls,
2504 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2505 		    tls, m->m_epg_tls));
2506 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2507 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2508 		    total_pages, m));
2509 
2510 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2511 		soref(so);
2512 		state->so = so;
2513 		state->m = m;
2514 
2515 		mpages = m->m_epg_nrdy;
2516 		n = m->m_next;
2517 
2518 		error = ktls_encrypt_record(wq, m, tls, state);
2519 		if (error) {
2520 			counter_u64_add(ktls_offload_failed_crypto, 1);
2521 			free(state, M_KTLS);
2522 			CURVNET_SET(so->so_vnet);
2523 			sorele(so);
2524 			CURVNET_RESTORE();
2525 			break;
2526 		}
2527 
2528 		npages += mpages;
2529 	}
2530 
2531 	CURVNET_SET(so->so_vnet);
2532 	if (error != 0) {
2533 		so->so_proto->pr_usrreqs->pru_abort(so);
2534 		so->so_error = EIO;
2535 		mb_free_notready(m, total_pages - npages);
2536 	}
2537 
2538 	sorele(so);
2539 	CURVNET_RESTORE();
2540 }
2541 
2542 static int
2543 ktls_bind_domain(int domain)
2544 {
2545 	int error;
2546 
2547 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
2548 	if (error != 0)
2549 		return (error);
2550 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
2551 	return (0);
2552 }
2553 
2554 static void
2555 ktls_alloc_thread(void *ctx)
2556 {
2557 	struct ktls_domain_info *ktls_domain = ctx;
2558 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
2559 	void **buf;
2560 	struct sysctl_oid *oid;
2561 	char name[80];
2562 	int domain, error, i, nbufs;
2563 
2564 	domain = ktls_domain - ktls_domains;
2565 	if (bootverbose)
2566 		printf("Starting KTLS alloc thread for domain %d\n", domain);
2567 	error = ktls_bind_domain(domain);
2568 	if (error)
2569 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
2570 		    domain, error);
2571 	snprintf(name, sizeof(name), "domain%d", domain);
2572 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
2573 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2574 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
2575 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
2576 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
2577 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
2578 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
2579 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
2580 
2581 	buf = NULL;
2582 	nbufs = 0;
2583 	for (;;) {
2584 		atomic_store_int(&sc->running, 0);
2585 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
2586 		atomic_store_int(&sc->running, 1);
2587 		sc->wakeups++;
2588 		if (nbufs != ktls_max_alloc) {
2589 			free(buf, M_KTLS);
2590 			nbufs = atomic_load_int(&ktls_max_alloc);
2591 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
2592 			    M_WAITOK | M_ZERO);
2593 		}
2594 		/*
2595 		 * Below we allocate nbufs with different allocation
2596 		 * flags than we use when allocating normally during
2597 		 * encryption in the ktls worker thread.  We specify
2598 		 * M_NORECLAIM in the worker thread. However, we omit
2599 		 * that flag here and add M_WAITOK so that the VM
2600 		 * system is permitted to perform expensive work to
2601 		 * defragment memory.  We do this here, as it does not
2602 		 * matter if this thread blocks.  If we block a ktls
2603 		 * worker thread, we risk developing backlogs of
2604 		 * buffers to be encrypted, leading to surges of
2605 		 * traffic and potential NIC output drops.
2606 		 */
2607 		for (i = 0; i < nbufs; i++) {
2608 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
2609 			sc->allocs++;
2610 		}
2611 		for (i = 0; i < nbufs; i++) {
2612 			uma_zfree(ktls_buffer_zone, buf[i]);
2613 			buf[i] = NULL;
2614 		}
2615 	}
2616 }
2617 
2618 static void
2619 ktls_work_thread(void *ctx)
2620 {
2621 	struct ktls_wq *wq = ctx;
2622 	struct mbuf *m, *n;
2623 	struct socket *so, *son;
2624 	STAILQ_HEAD(, mbuf) local_m_head;
2625 	STAILQ_HEAD(, socket) local_so_head;
2626 	int cpu;
2627 
2628 	cpu = wq - ktls_wq;
2629 	if (bootverbose)
2630 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
2631 
2632 	/*
2633 	 * Bind to a core.  If ktls_bind_threads is > 1, then
2634 	 * we bind to the NUMA domain instead.
2635 	 */
2636 	if (ktls_bind_threads) {
2637 		int error;
2638 
2639 		if (ktls_bind_threads > 1) {
2640 			struct pcpu *pc = pcpu_find(cpu);
2641 
2642 			error = ktls_bind_domain(pc->pc_domain);
2643 		} else {
2644 			cpuset_t mask;
2645 
2646 			CPU_SETOF(cpu, &mask);
2647 			error = cpuset_setthread(curthread->td_tid, &mask);
2648 		}
2649 		if (error)
2650 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
2651 				cpu, error);
2652 	}
2653 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2654 	fpu_kern_thread(0);
2655 #endif
2656 	for (;;) {
2657 		mtx_lock(&wq->mtx);
2658 		while (STAILQ_EMPTY(&wq->m_head) &&
2659 		    STAILQ_EMPTY(&wq->so_head)) {
2660 			wq->running = false;
2661 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2662 			wq->running = true;
2663 		}
2664 
2665 		STAILQ_INIT(&local_m_head);
2666 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2667 		STAILQ_INIT(&local_so_head);
2668 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2669 		mtx_unlock(&wq->mtx);
2670 
2671 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2672 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2673 				ktls_free(m->m_epg_tls);
2674 				m_free_raw(m);
2675 			} else {
2676 				if (m->m_epg_tls->sync_dispatch)
2677 					ktls_encrypt(wq, m);
2678 				else
2679 					ktls_encrypt_async(wq, m);
2680 				counter_u64_add(ktls_cnt_tx_queued, -1);
2681 			}
2682 		}
2683 
2684 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2685 			ktls_decrypt(so);
2686 			counter_u64_add(ktls_cnt_rx_queued, -1);
2687 		}
2688 	}
2689 }
2690 
2691 #if defined(INET) || defined(INET6)
2692 static void
2693 ktls_disable_ifnet_help(void *context, int pending __unused)
2694 {
2695 	struct ktls_session *tls;
2696 	struct inpcb *inp;
2697 	struct tcpcb *tp;
2698 	struct socket *so;
2699 	int err;
2700 
2701 	tls = context;
2702 	inp = tls->inp;
2703 	if (inp == NULL)
2704 		return;
2705 	INP_WLOCK(inp);
2706 	so = inp->inp_socket;
2707 	MPASS(so != NULL);
2708 	if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
2709 	    (inp->inp_flags2 & INP_FREED)) {
2710 		goto out;
2711 	}
2712 
2713 	if (so->so_snd.sb_tls_info != NULL)
2714 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
2715 	else
2716 		err = ENXIO;
2717 	if (err == 0) {
2718 		counter_u64_add(ktls_ifnet_disable_ok, 1);
2719 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
2720 		if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 &&
2721 		    (inp->inp_flags2 & INP_FREED) == 0 &&
2722 		    (tp = intotcpcb(inp)) != NULL &&
2723 		    tp->t_fb->tfb_hwtls_change != NULL)
2724 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
2725 	} else {
2726 		counter_u64_add(ktls_ifnet_disable_fail, 1);
2727 	}
2728 
2729 out:
2730 	sorele(so);
2731 	if (!in_pcbrele_wlocked(inp))
2732 		INP_WUNLOCK(inp);
2733 	ktls_free(tls);
2734 }
2735 
2736 /*
2737  * Called when re-transmits are becoming a substantial portion of the
2738  * sends on this connection.  When this happens, we transition the
2739  * connection to software TLS.  This is needed because most inline TLS
2740  * NICs keep crypto state only for in-order transmits.  This means
2741  * that to handle a TCP rexmit (which is out-of-order), the NIC must
2742  * re-DMA the entire TLS record up to and including the current
2743  * segment.  This means that when re-transmitting the last ~1448 byte
2744  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
2745  * of magnitude more data than we are sending.  This can cause the
2746  * PCIe link to saturate well before the network, which can cause
2747  * output drops, and a general loss of capacity.
2748  */
2749 void
2750 ktls_disable_ifnet(void *arg)
2751 {
2752 	struct tcpcb *tp;
2753 	struct inpcb *inp;
2754 	struct socket *so;
2755 	struct ktls_session *tls;
2756 
2757 	tp = arg;
2758 	inp = tp->t_inpcb;
2759 	INP_WLOCK_ASSERT(inp);
2760 	so = inp->inp_socket;
2761 	SOCK_LOCK(so);
2762 	tls = so->so_snd.sb_tls_info;
2763 	if (tls->disable_ifnet_pending) {
2764 		SOCK_UNLOCK(so);
2765 		return;
2766 	}
2767 
2768 	/*
2769 	 * note that disable_ifnet_pending is never cleared; disabling
2770 	 * ifnet can only be done once per session, so we never want
2771 	 * to do it again
2772 	 */
2773 
2774 	(void)ktls_hold(tls);
2775 	in_pcbref(inp);
2776 	soref(so);
2777 	tls->disable_ifnet_pending = true;
2778 	tls->inp = inp;
2779 	SOCK_UNLOCK(so);
2780 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
2781 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
2782 }
2783 #endif
2784