xref: /freebsd/sys/kern/uipc_ktls.c (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
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_receive_tag(void *context, int pending);
303 static void ktls_reset_send_tag(void *context, int pending);
304 #endif
305 static void ktls_work_thread(void *ctx);
306 static void ktls_alloc_thread(void *ctx);
307 
308 #if defined(INET) || defined(INET6)
309 static u_int
310 ktls_get_cpu(struct socket *so)
311 {
312 	struct inpcb *inp;
313 #ifdef NUMA
314 	struct ktls_domain_info *di;
315 #endif
316 	u_int cpuid;
317 
318 	inp = sotoinpcb(so);
319 #ifdef RSS
320 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
321 	if (cpuid != NETISR_CPUID_NONE)
322 		return (cpuid);
323 #endif
324 	/*
325 	 * Just use the flowid to shard connections in a repeatable
326 	 * fashion.  Note that TLS 1.0 sessions rely on the
327 	 * serialization provided by having the same connection use
328 	 * the same queue.
329 	 */
330 #ifdef NUMA
331 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
332 		di = &ktls_domains[inp->inp_numa_domain];
333 		cpuid = di->cpu[inp->inp_flowid % di->count];
334 	} else
335 #endif
336 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
337 	return (cpuid);
338 }
339 #endif
340 
341 static int
342 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
343 {
344 	vm_page_t m;
345 	int i, req;
346 
347 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
348 	    ("%s: ktls max length %d is not page size-aligned",
349 	    __func__, ktls_maxlen));
350 
351 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
352 	for (i = 0; i < count; i++) {
353 		m = vm_page_alloc_noobj_contig_domain(domain, req,
354 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
355 		    VM_MEMATTR_DEFAULT);
356 		if (m == NULL)
357 			break;
358 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
359 	}
360 	return (i);
361 }
362 
363 static void
364 ktls_buffer_release(void *arg __unused, void **store, int count)
365 {
366 	vm_page_t m;
367 	int i, j;
368 
369 	for (i = 0; i < count; i++) {
370 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
371 		for (j = 0; j < atop(ktls_maxlen); j++) {
372 			(void)vm_page_unwire_noq(m + j);
373 			vm_page_free(m + j);
374 		}
375 	}
376 }
377 
378 static void
379 ktls_free_mext_contig(struct mbuf *m)
380 {
381 	M_ASSERTEXTPG(m);
382 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
383 }
384 
385 static int
386 ktls_init(void)
387 {
388 	struct thread *td;
389 	struct pcpu *pc;
390 	int count, domain, error, i;
391 
392 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
393 	    M_WAITOK | M_ZERO);
394 
395 	ktls_session_zone = uma_zcreate("ktls_session",
396 	    sizeof(struct ktls_session),
397 	    NULL, NULL, NULL, NULL,
398 	    UMA_ALIGN_CACHE, 0);
399 
400 	if (ktls_sw_buffer_cache) {
401 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
402 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
403 		    ktls_buffer_import, ktls_buffer_release, NULL,
404 		    UMA_ZONE_FIRSTTOUCH);
405 	}
406 
407 	/*
408 	 * Initialize the workqueues to run the TLS work.  We create a
409 	 * work queue for each CPU.
410 	 */
411 	CPU_FOREACH(i) {
412 		STAILQ_INIT(&ktls_wq[i].m_head);
413 		STAILQ_INIT(&ktls_wq[i].so_head);
414 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
415 		if (ktls_bind_threads > 1) {
416 			pc = pcpu_find(i);
417 			domain = pc->pc_domain;
418 			count = ktls_domains[domain].count;
419 			ktls_domains[domain].cpu[count] = i;
420 			ktls_domains[domain].count++;
421 		}
422 		ktls_cpuid_lookup[ktls_number_threads] = i;
423 		ktls_number_threads++;
424 	}
425 
426 	/*
427 	 * If we somehow have an empty domain, fall back to choosing
428 	 * among all KTLS threads.
429 	 */
430 	if (ktls_bind_threads > 1) {
431 		for (i = 0; i < vm_ndomains; i++) {
432 			if (ktls_domains[i].count == 0) {
433 				ktls_bind_threads = 1;
434 				break;
435 			}
436 		}
437 	}
438 
439 	/* Start kthreads for each workqueue. */
440 	CPU_FOREACH(i) {
441 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
442 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
443 		if (error) {
444 			printf("Can't add KTLS thread %d error %d\n", i, error);
445 			return (error);
446 		}
447 	}
448 
449 	/*
450 	 * Start an allocation thread per-domain to perform blocking allocations
451 	 * of 16k physically contiguous TLS crypto destination buffers.
452 	 */
453 	if (ktls_sw_buffer_cache) {
454 		for (domain = 0; domain < vm_ndomains; domain++) {
455 			if (VM_DOMAIN_EMPTY(domain))
456 				continue;
457 			if (CPU_EMPTY(&cpuset_domain[domain]))
458 				continue;
459 			error = kproc_kthread_add(ktls_alloc_thread,
460 			    &ktls_domains[domain], &ktls_proc,
461 			    &ktls_domains[domain].alloc_td.td,
462 			    0, 0, "KTLS", "alloc_%d", domain);
463 			if (error) {
464 				printf("Can't add KTLS alloc thread %d error %d\n",
465 				    domain, error);
466 				return (error);
467 			}
468 		}
469 	}
470 
471 	if (bootverbose)
472 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
473 	return (0);
474 }
475 
476 static int
477 ktls_start_kthreads(void)
478 {
479 	int error, state;
480 
481 start:
482 	state = atomic_load_acq_int(&ktls_init_state);
483 	if (__predict_true(state > 0))
484 		return (0);
485 	if (state < 0)
486 		return (ENXIO);
487 
488 	sx_xlock(&ktls_init_lock);
489 	if (ktls_init_state != 0) {
490 		sx_xunlock(&ktls_init_lock);
491 		goto start;
492 	}
493 
494 	error = ktls_init();
495 	if (error == 0)
496 		state = 1;
497 	else
498 		state = -1;
499 	atomic_store_rel_int(&ktls_init_state, state);
500 	sx_xunlock(&ktls_init_lock);
501 	return (error);
502 }
503 
504 #if defined(INET) || defined(INET6)
505 static int
506 ktls_create_session(struct socket *so, struct tls_enable *en,
507     struct ktls_session **tlsp, int direction)
508 {
509 	struct ktls_session *tls;
510 	int error;
511 
512 	/* Only TLS 1.0 - 1.3 are supported. */
513 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
514 		return (EINVAL);
515 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
516 	    en->tls_vminor > TLS_MINOR_VER_THREE)
517 		return (EINVAL);
518 
519 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
520 		return (EINVAL);
521 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
522 		return (EINVAL);
523 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
524 		return (EINVAL);
525 
526 	/* All supported algorithms require a cipher key. */
527 	if (en->cipher_key_len == 0)
528 		return (EINVAL);
529 
530 	/* No flags are currently supported. */
531 	if (en->flags != 0)
532 		return (EINVAL);
533 
534 	/* Common checks for supported algorithms. */
535 	switch (en->cipher_algorithm) {
536 	case CRYPTO_AES_NIST_GCM_16:
537 		/*
538 		 * auth_algorithm isn't used, but permit GMAC values
539 		 * for compatibility.
540 		 */
541 		switch (en->auth_algorithm) {
542 		case 0:
543 #ifdef COMPAT_FREEBSD12
544 		/* XXX: Really 13.0-current COMPAT. */
545 		case CRYPTO_AES_128_NIST_GMAC:
546 		case CRYPTO_AES_192_NIST_GMAC:
547 		case CRYPTO_AES_256_NIST_GMAC:
548 #endif
549 			break;
550 		default:
551 			return (EINVAL);
552 		}
553 		if (en->auth_key_len != 0)
554 			return (EINVAL);
555 		switch (en->tls_vminor) {
556 		case TLS_MINOR_VER_TWO:
557 			if (en->iv_len != TLS_AEAD_GCM_LEN)
558 				return (EINVAL);
559 			break;
560 		case TLS_MINOR_VER_THREE:
561 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
562 				return (EINVAL);
563 			break;
564 		default:
565 			return (EINVAL);
566 		}
567 		break;
568 	case CRYPTO_AES_CBC:
569 		switch (en->auth_algorithm) {
570 		case CRYPTO_SHA1_HMAC:
571 			break;
572 		case CRYPTO_SHA2_256_HMAC:
573 		case CRYPTO_SHA2_384_HMAC:
574 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
575 				return (EINVAL);
576 			break;
577 		default:
578 			return (EINVAL);
579 		}
580 		if (en->auth_key_len == 0)
581 			return (EINVAL);
582 
583 		/*
584 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
585 		 * use explicit IVs.
586 		 */
587 		switch (en->tls_vminor) {
588 		case TLS_MINOR_VER_ZERO:
589 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
590 				return (EINVAL);
591 			break;
592 		case TLS_MINOR_VER_ONE:
593 		case TLS_MINOR_VER_TWO:
594 			/* Ignore any supplied IV. */
595 			en->iv_len = 0;
596 			break;
597 		default:
598 			return (EINVAL);
599 		}
600 		break;
601 	case CRYPTO_CHACHA20_POLY1305:
602 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
603 			return (EINVAL);
604 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
605 		    en->tls_vminor != TLS_MINOR_VER_THREE)
606 			return (EINVAL);
607 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
608 			return (EINVAL);
609 		break;
610 	default:
611 		return (EINVAL);
612 	}
613 
614 	error = ktls_start_kthreads();
615 	if (error != 0)
616 		return (error);
617 
618 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
619 
620 	counter_u64_add(ktls_offload_active, 1);
621 
622 	refcount_init(&tls->refcount, 1);
623 	if (direction == KTLS_RX)
624 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
625 	else
626 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
627 
628 	tls->wq_index = ktls_get_cpu(so);
629 
630 	tls->params.cipher_algorithm = en->cipher_algorithm;
631 	tls->params.auth_algorithm = en->auth_algorithm;
632 	tls->params.tls_vmajor = en->tls_vmajor;
633 	tls->params.tls_vminor = en->tls_vminor;
634 	tls->params.flags = en->flags;
635 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
636 
637 	/* Set the header and trailer lengths. */
638 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
639 	switch (en->cipher_algorithm) {
640 	case CRYPTO_AES_NIST_GCM_16:
641 		/*
642 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
643 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
644 		 */
645 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
646 			tls->params.tls_hlen += sizeof(uint64_t);
647 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
648 		tls->params.tls_bs = 1;
649 		break;
650 	case CRYPTO_AES_CBC:
651 		switch (en->auth_algorithm) {
652 		case CRYPTO_SHA1_HMAC:
653 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
654 				/* Implicit IV, no nonce. */
655 				tls->sequential_records = true;
656 				tls->next_seqno = be64dec(en->rec_seq);
657 				STAILQ_INIT(&tls->pending_records);
658 			} else {
659 				tls->params.tls_hlen += AES_BLOCK_LEN;
660 			}
661 			tls->params.tls_tlen = AES_BLOCK_LEN +
662 			    SHA1_HASH_LEN;
663 			break;
664 		case CRYPTO_SHA2_256_HMAC:
665 			tls->params.tls_hlen += AES_BLOCK_LEN;
666 			tls->params.tls_tlen = AES_BLOCK_LEN +
667 			    SHA2_256_HASH_LEN;
668 			break;
669 		case CRYPTO_SHA2_384_HMAC:
670 			tls->params.tls_hlen += AES_BLOCK_LEN;
671 			tls->params.tls_tlen = AES_BLOCK_LEN +
672 			    SHA2_384_HASH_LEN;
673 			break;
674 		default:
675 			panic("invalid hmac");
676 		}
677 		tls->params.tls_bs = AES_BLOCK_LEN;
678 		break;
679 	case CRYPTO_CHACHA20_POLY1305:
680 		/*
681 		 * Chacha20 uses a 12 byte implicit IV.
682 		 */
683 		tls->params.tls_tlen = POLY1305_HASH_LEN;
684 		tls->params.tls_bs = 1;
685 		break;
686 	default:
687 		panic("invalid cipher");
688 	}
689 
690 	/*
691 	 * TLS 1.3 includes optional padding which we do not support,
692 	 * and also puts the "real" record type at the end of the
693 	 * encrypted data.
694 	 */
695 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
696 		tls->params.tls_tlen += sizeof(uint8_t);
697 
698 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
699 	    ("TLS header length too long: %d", tls->params.tls_hlen));
700 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
701 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
702 
703 	if (en->auth_key_len != 0) {
704 		tls->params.auth_key_len = en->auth_key_len;
705 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
706 		    M_WAITOK);
707 		error = copyin(en->auth_key, tls->params.auth_key,
708 		    en->auth_key_len);
709 		if (error)
710 			goto out;
711 	}
712 
713 	tls->params.cipher_key_len = en->cipher_key_len;
714 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
715 	error = copyin(en->cipher_key, tls->params.cipher_key,
716 	    en->cipher_key_len);
717 	if (error)
718 		goto out;
719 
720 	/*
721 	 * This holds the implicit portion of the nonce for AEAD
722 	 * ciphers and the initial implicit IV for TLS 1.0.  The
723 	 * explicit portions of the IV are generated in ktls_frame().
724 	 */
725 	if (en->iv_len != 0) {
726 		tls->params.iv_len = en->iv_len;
727 		error = copyin(en->iv, tls->params.iv, en->iv_len);
728 		if (error)
729 			goto out;
730 
731 		/*
732 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
733 		 * counter to generate unique explicit IVs.
734 		 *
735 		 * Store this counter in the last 8 bytes of the IV
736 		 * array so that it is 8-byte aligned.
737 		 */
738 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
739 		    en->tls_vminor == TLS_MINOR_VER_TWO)
740 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
741 	}
742 
743 	*tlsp = tls;
744 	return (0);
745 
746 out:
747 	ktls_cleanup(tls);
748 	return (error);
749 }
750 
751 static struct ktls_session *
752 ktls_clone_session(struct ktls_session *tls, int direction)
753 {
754 	struct ktls_session *tls_new;
755 
756 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
757 
758 	counter_u64_add(ktls_offload_active, 1);
759 
760 	refcount_init(&tls_new->refcount, 1);
761 	if (direction == KTLS_RX)
762 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
763 		    tls_new);
764 	else
765 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
766 		    tls_new);
767 
768 	/* Copy fields from existing session. */
769 	tls_new->params = tls->params;
770 	tls_new->wq_index = tls->wq_index;
771 
772 	/* Deep copy keys. */
773 	if (tls_new->params.auth_key != NULL) {
774 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
775 		    M_KTLS, M_WAITOK);
776 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
777 		    tls->params.auth_key_len);
778 	}
779 
780 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
781 	    M_WAITOK);
782 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
783 	    tls->params.cipher_key_len);
784 
785 	return (tls_new);
786 }
787 #endif
788 
789 static void
790 ktls_cleanup(struct ktls_session *tls)
791 {
792 
793 	counter_u64_add(ktls_offload_active, -1);
794 	switch (tls->mode) {
795 	case TCP_TLS_MODE_SW:
796 		switch (tls->params.cipher_algorithm) {
797 		case CRYPTO_AES_CBC:
798 			counter_u64_add(ktls_sw_cbc, -1);
799 			break;
800 		case CRYPTO_AES_NIST_GCM_16:
801 			counter_u64_add(ktls_sw_gcm, -1);
802 			break;
803 		case CRYPTO_CHACHA20_POLY1305:
804 			counter_u64_add(ktls_sw_chacha20, -1);
805 			break;
806 		}
807 		break;
808 	case TCP_TLS_MODE_IFNET:
809 		switch (tls->params.cipher_algorithm) {
810 		case CRYPTO_AES_CBC:
811 			counter_u64_add(ktls_ifnet_cbc, -1);
812 			break;
813 		case CRYPTO_AES_NIST_GCM_16:
814 			counter_u64_add(ktls_ifnet_gcm, -1);
815 			break;
816 		case CRYPTO_CHACHA20_POLY1305:
817 			counter_u64_add(ktls_ifnet_chacha20, -1);
818 			break;
819 		}
820 		if (tls->snd_tag != NULL)
821 			m_snd_tag_rele(tls->snd_tag);
822 		if (tls->rx_ifp != NULL)
823 			if_rele(tls->rx_ifp);
824 		break;
825 #ifdef TCP_OFFLOAD
826 	case TCP_TLS_MODE_TOE:
827 		switch (tls->params.cipher_algorithm) {
828 		case CRYPTO_AES_CBC:
829 			counter_u64_add(ktls_toe_cbc, -1);
830 			break;
831 		case CRYPTO_AES_NIST_GCM_16:
832 			counter_u64_add(ktls_toe_gcm, -1);
833 			break;
834 		case CRYPTO_CHACHA20_POLY1305:
835 			counter_u64_add(ktls_toe_chacha20, -1);
836 			break;
837 		}
838 		break;
839 #endif
840 	}
841 	if (tls->ocf_session != NULL)
842 		ktls_ocf_free(tls);
843 	if (tls->params.auth_key != NULL) {
844 		zfree(tls->params.auth_key, M_KTLS);
845 		tls->params.auth_key = NULL;
846 		tls->params.auth_key_len = 0;
847 	}
848 	if (tls->params.cipher_key != NULL) {
849 		zfree(tls->params.cipher_key, M_KTLS);
850 		tls->params.cipher_key = NULL;
851 		tls->params.cipher_key_len = 0;
852 	}
853 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
854 }
855 
856 #if defined(INET) || defined(INET6)
857 
858 #ifdef TCP_OFFLOAD
859 static int
860 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
861 {
862 	struct inpcb *inp;
863 	struct tcpcb *tp;
864 	int error;
865 
866 	inp = so->so_pcb;
867 	INP_WLOCK(inp);
868 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
869 		INP_WUNLOCK(inp);
870 		return (ECONNRESET);
871 	}
872 	if (inp->inp_socket == NULL) {
873 		INP_WUNLOCK(inp);
874 		return (ECONNRESET);
875 	}
876 	tp = intotcpcb(inp);
877 	if (!(tp->t_flags & TF_TOE)) {
878 		INP_WUNLOCK(inp);
879 		return (EOPNOTSUPP);
880 	}
881 
882 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
883 	INP_WUNLOCK(inp);
884 	if (error == 0) {
885 		tls->mode = TCP_TLS_MODE_TOE;
886 		switch (tls->params.cipher_algorithm) {
887 		case CRYPTO_AES_CBC:
888 			counter_u64_add(ktls_toe_cbc, 1);
889 			break;
890 		case CRYPTO_AES_NIST_GCM_16:
891 			counter_u64_add(ktls_toe_gcm, 1);
892 			break;
893 		case CRYPTO_CHACHA20_POLY1305:
894 			counter_u64_add(ktls_toe_chacha20, 1);
895 			break;
896 		}
897 	}
898 	return (error);
899 }
900 #endif
901 
902 /*
903  * Common code used when first enabling ifnet TLS on a connection or
904  * when allocating a new ifnet TLS session due to a routing change.
905  * This function allocates a new TLS send tag on whatever interface
906  * the connection is currently routed over.
907  */
908 static int
909 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
910     struct m_snd_tag **mstp)
911 {
912 	union if_snd_tag_alloc_params params;
913 	struct ifnet *ifp;
914 	struct nhop_object *nh;
915 	struct tcpcb *tp;
916 	int error;
917 
918 	INP_RLOCK(inp);
919 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
920 		INP_RUNLOCK(inp);
921 		return (ECONNRESET);
922 	}
923 	if (inp->inp_socket == NULL) {
924 		INP_RUNLOCK(inp);
925 		return (ECONNRESET);
926 	}
927 	tp = intotcpcb(inp);
928 
929 	/*
930 	 * Check administrative controls on ifnet TLS to determine if
931 	 * ifnet TLS should be denied.
932 	 *
933 	 * - Always permit 'force' requests.
934 	 * - ktls_ifnet_permitted == 0: always deny.
935 	 */
936 	if (!force && ktls_ifnet_permitted == 0) {
937 		INP_RUNLOCK(inp);
938 		return (ENXIO);
939 	}
940 
941 	/*
942 	 * XXX: Use the cached route in the inpcb to find the
943 	 * interface.  This should perhaps instead use
944 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
945 	 * enabled after a connection has completed key negotiation in
946 	 * userland, the cached route will be present in practice.
947 	 */
948 	nh = inp->inp_route.ro_nh;
949 	if (nh == NULL) {
950 		INP_RUNLOCK(inp);
951 		return (ENXIO);
952 	}
953 	ifp = nh->nh_ifp;
954 	if_ref(ifp);
955 
956 	/*
957 	 * Allocate a TLS + ratelimit tag if the connection has an
958 	 * existing pacing rate.
959 	 */
960 	if (tp->t_pacing_rate != -1 &&
961 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
962 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
963 		params.tls_rate_limit.inp = inp;
964 		params.tls_rate_limit.tls = tls;
965 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
966 	} else {
967 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
968 		params.tls.inp = inp;
969 		params.tls.tls = tls;
970 	}
971 	params.hdr.flowid = inp->inp_flowid;
972 	params.hdr.flowtype = inp->inp_flowtype;
973 	params.hdr.numa_domain = inp->inp_numa_domain;
974 	INP_RUNLOCK(inp);
975 
976 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
977 		error = EOPNOTSUPP;
978 		goto out;
979 	}
980 	if (inp->inp_vflag & INP_IPV6) {
981 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
982 			error = EOPNOTSUPP;
983 			goto out;
984 		}
985 	} else {
986 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
987 			error = EOPNOTSUPP;
988 			goto out;
989 		}
990 	}
991 	error = m_snd_tag_alloc(ifp, &params, mstp);
992 out:
993 	if_rele(ifp);
994 	return (error);
995 }
996 
997 /*
998  * Allocate an initial TLS receive tag for doing HW decryption of TLS
999  * data.
1000  *
1001  * This function allocates a new TLS receive tag on whatever interface
1002  * the connection is currently routed over.  If the connection ends up
1003  * using a different interface for receive this will get fixed up via
1004  * ktls_input_ifp_mismatch as future packets arrive.
1005  */
1006 static int
1007 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1008     struct m_snd_tag **mstp)
1009 {
1010 	union if_snd_tag_alloc_params params;
1011 	struct ifnet *ifp;
1012 	struct nhop_object *nh;
1013 	int error;
1014 
1015 	if (!ktls_ocf_recrypt_supported(tls))
1016 		return (ENXIO);
1017 
1018 	INP_RLOCK(inp);
1019 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
1020 		INP_RUNLOCK(inp);
1021 		return (ECONNRESET);
1022 	}
1023 	if (inp->inp_socket == NULL) {
1024 		INP_RUNLOCK(inp);
1025 		return (ECONNRESET);
1026 	}
1027 
1028 	/*
1029 	 * Check administrative controls on ifnet TLS to determine if
1030 	 * ifnet TLS should be denied.
1031 	 */
1032 	if (ktls_ifnet_permitted == 0) {
1033 		INP_RUNLOCK(inp);
1034 		return (ENXIO);
1035 	}
1036 
1037 	/*
1038 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1039 	 * the inpcb to find the interface.
1040 	 */
1041 	nh = inp->inp_route.ro_nh;
1042 	if (nh == NULL) {
1043 		INP_RUNLOCK(inp);
1044 		return (ENXIO);
1045 	}
1046 	ifp = nh->nh_ifp;
1047 	if_ref(ifp);
1048 	tls->rx_ifp = ifp;
1049 
1050 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1051 	params.hdr.flowid = inp->inp_flowid;
1052 	params.hdr.flowtype = inp->inp_flowtype;
1053 	params.hdr.numa_domain = inp->inp_numa_domain;
1054 	params.tls_rx.inp = inp;
1055 	params.tls_rx.tls = tls;
1056 	params.tls_rx.vlan_id = 0;
1057 
1058 	INP_RUNLOCK(inp);
1059 
1060 	if (inp->inp_vflag & INP_IPV6) {
1061 		if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) {
1062 			error = EOPNOTSUPP;
1063 			goto out;
1064 		}
1065 	} else {
1066 		if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) {
1067 			error = EOPNOTSUPP;
1068 			goto out;
1069 		}
1070 	}
1071 	error = m_snd_tag_alloc(ifp, &params, mstp);
1072 
1073 	/*
1074 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1075 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1076 	 * ID for use in ktls_reset_receive_tag which allocates new
1077 	 * receive tags directly from the leaf interface bypassing
1078 	 * if_vlan.
1079 	 */
1080 	if (error == 0)
1081 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1082 out:
1083 	return (error);
1084 }
1085 
1086 static int
1087 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1088     bool force)
1089 {
1090 	struct m_snd_tag *mst;
1091 	int error;
1092 
1093 	switch (direction) {
1094 	case KTLS_TX:
1095 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1096 		if (__predict_false(error != 0))
1097 			goto done;
1098 		break;
1099 	case KTLS_RX:
1100 		KASSERT(!force, ("%s: forced receive tag", __func__));
1101 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1102 		if (__predict_false(error != 0))
1103 			goto done;
1104 		break;
1105 	default:
1106 		__assert_unreachable();
1107 	}
1108 
1109 	tls->mode = TCP_TLS_MODE_IFNET;
1110 	tls->snd_tag = mst;
1111 
1112 	switch (tls->params.cipher_algorithm) {
1113 	case CRYPTO_AES_CBC:
1114 		counter_u64_add(ktls_ifnet_cbc, 1);
1115 		break;
1116 	case CRYPTO_AES_NIST_GCM_16:
1117 		counter_u64_add(ktls_ifnet_gcm, 1);
1118 		break;
1119 	case CRYPTO_CHACHA20_POLY1305:
1120 		counter_u64_add(ktls_ifnet_chacha20, 1);
1121 		break;
1122 	default:
1123 		break;
1124 	}
1125 done:
1126 	return (error);
1127 }
1128 
1129 static void
1130 ktls_use_sw(struct ktls_session *tls)
1131 {
1132 	tls->mode = TCP_TLS_MODE_SW;
1133 	switch (tls->params.cipher_algorithm) {
1134 	case CRYPTO_AES_CBC:
1135 		counter_u64_add(ktls_sw_cbc, 1);
1136 		break;
1137 	case CRYPTO_AES_NIST_GCM_16:
1138 		counter_u64_add(ktls_sw_gcm, 1);
1139 		break;
1140 	case CRYPTO_CHACHA20_POLY1305:
1141 		counter_u64_add(ktls_sw_chacha20, 1);
1142 		break;
1143 	}
1144 }
1145 
1146 static int
1147 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1148 {
1149 	int error;
1150 
1151 	error = ktls_ocf_try(so, tls, direction);
1152 	if (error)
1153 		return (error);
1154 	ktls_use_sw(tls);
1155 	return (0);
1156 }
1157 
1158 /*
1159  * KTLS RX stores data in the socket buffer as a list of TLS records,
1160  * where each record is stored as a control message containg the TLS
1161  * header followed by data mbufs containing the decrypted data.  This
1162  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1163  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1164  * should be queued to the socket buffer as records, but encrypted
1165  * data which needs to be decrypted by software arrives as a stream of
1166  * regular mbufs which need to be converted.  In addition, there may
1167  * already be pending encrypted data in the socket buffer when KTLS RX
1168  * is enabled.
1169  *
1170  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1171  * is used:
1172  *
1173  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1174  *
1175  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1176  *   from the first mbuf.  Once all of the data for that TLS record is
1177  *   queued, the socket is queued to a worker thread.
1178  *
1179  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1180  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1181  *   decrypted, and inserted into the regular socket buffer chain as
1182  *   record starting with a control message holding the TLS header and
1183  *   a chain of mbufs holding the encrypted data.
1184  */
1185 
1186 static void
1187 sb_mark_notready(struct sockbuf *sb)
1188 {
1189 	struct mbuf *m;
1190 
1191 	m = sb->sb_mb;
1192 	sb->sb_mtls = m;
1193 	sb->sb_mb = NULL;
1194 	sb->sb_mbtail = NULL;
1195 	sb->sb_lastrecord = NULL;
1196 	for (; m != NULL; m = m->m_next) {
1197 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1198 		    __func__));
1199 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1200 		    __func__));
1201 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1202 		    __func__));
1203 		m->m_flags |= M_NOTREADY;
1204 		sb->sb_acc -= m->m_len;
1205 		sb->sb_tlscc += m->m_len;
1206 		sb->sb_mtlstail = m;
1207 	}
1208 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1209 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1210 	    sb->sb_ccc));
1211 }
1212 
1213 /*
1214  * Return information about the pending TLS data in a socket
1215  * buffer.  On return, 'seqno' is set to the sequence number
1216  * of the next TLS record to be received, 'resid' is set to
1217  * the amount of bytes still needed for the last pending
1218  * record.  The function returns 'false' if the last pending
1219  * record contains a partial TLS header.  In that case, 'resid'
1220  * is the number of bytes needed to complete the TLS header.
1221  */
1222 bool
1223 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1224 {
1225 	struct tls_record_layer hdr;
1226 	struct mbuf *m;
1227 	uint64_t seqno;
1228 	size_t resid;
1229 	u_int offset, record_len;
1230 
1231 	SOCKBUF_LOCK_ASSERT(sb);
1232 	MPASS(sb->sb_flags & SB_TLS_RX);
1233 	seqno = sb->sb_tls_seqno;
1234 	resid = sb->sb_tlscc;
1235 	m = sb->sb_mtls;
1236 	offset = 0;
1237 
1238 	if (resid == 0) {
1239 		*seqnop = seqno;
1240 		*residp = 0;
1241 		return (true);
1242 	}
1243 
1244 	for (;;) {
1245 		seqno++;
1246 
1247 		if (resid < sizeof(hdr)) {
1248 			*seqnop = seqno;
1249 			*residp = sizeof(hdr) - resid;
1250 			return (false);
1251 		}
1252 
1253 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1254 
1255 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1256 		if (resid <= record_len) {
1257 			*seqnop = seqno;
1258 			*residp = record_len - resid;
1259 			return (true);
1260 		}
1261 		resid -= record_len;
1262 
1263 		while (record_len != 0) {
1264 			if (m->m_len - offset > record_len) {
1265 				offset += record_len;
1266 				break;
1267 			}
1268 
1269 			record_len -= (m->m_len - offset);
1270 			offset = 0;
1271 			m = m->m_next;
1272 		}
1273 	}
1274 }
1275 
1276 int
1277 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1278 {
1279 	struct ktls_session *tls;
1280 	int error;
1281 
1282 	if (!ktls_offload_enable)
1283 		return (ENOTSUP);
1284 	if (SOLISTENING(so))
1285 		return (EINVAL);
1286 
1287 	counter_u64_add(ktls_offload_enable_calls, 1);
1288 
1289 	/*
1290 	 * This should always be true since only the TCP socket option
1291 	 * invokes this function.
1292 	 */
1293 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1294 		return (EINVAL);
1295 
1296 	/*
1297 	 * XXX: Don't overwrite existing sessions.  We should permit
1298 	 * this to support rekeying in the future.
1299 	 */
1300 	if (so->so_rcv.sb_tls_info != NULL)
1301 		return (EALREADY);
1302 
1303 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1304 		return (ENOTSUP);
1305 
1306 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1307 	if (error)
1308 		return (error);
1309 
1310 	error = ktls_ocf_try(so, tls, KTLS_RX);
1311 	if (error) {
1312 		ktls_cleanup(tls);
1313 		return (error);
1314 	}
1315 
1316 	/* Mark the socket as using TLS offload. */
1317 	SOCKBUF_LOCK(&so->so_rcv);
1318 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1319 	so->so_rcv.sb_tls_info = tls;
1320 	so->so_rcv.sb_flags |= SB_TLS_RX;
1321 
1322 	/* Mark existing data as not ready until it can be decrypted. */
1323 	sb_mark_notready(&so->so_rcv);
1324 	ktls_check_rx(&so->so_rcv);
1325 	SOCKBUF_UNLOCK(&so->so_rcv);
1326 
1327 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1328 #ifdef TCP_OFFLOAD
1329 	error = ktls_try_toe(so, tls, KTLS_RX);
1330 	if (error)
1331 #endif
1332 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1333 	if (error)
1334 		ktls_use_sw(tls);
1335 
1336 	counter_u64_add(ktls_offload_total, 1);
1337 
1338 	return (0);
1339 }
1340 
1341 int
1342 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1343 {
1344 	struct ktls_session *tls;
1345 	struct inpcb *inp;
1346 	int error;
1347 
1348 	if (!ktls_offload_enable)
1349 		return (ENOTSUP);
1350 	if (SOLISTENING(so))
1351 		return (EINVAL);
1352 
1353 	counter_u64_add(ktls_offload_enable_calls, 1);
1354 
1355 	/*
1356 	 * This should always be true since only the TCP socket option
1357 	 * invokes this function.
1358 	 */
1359 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1360 		return (EINVAL);
1361 
1362 	/*
1363 	 * XXX: Don't overwrite existing sessions.  We should permit
1364 	 * this to support rekeying in the future.
1365 	 */
1366 	if (so->so_snd.sb_tls_info != NULL)
1367 		return (EALREADY);
1368 
1369 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1370 		return (ENOTSUP);
1371 
1372 	/* TLS requires ext pgs */
1373 	if (mb_use_ext_pgs == 0)
1374 		return (ENXIO);
1375 
1376 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1377 	if (error)
1378 		return (error);
1379 
1380 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1381 #ifdef TCP_OFFLOAD
1382 	error = ktls_try_toe(so, tls, KTLS_TX);
1383 	if (error)
1384 #endif
1385 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1386 	if (error)
1387 		error = ktls_try_sw(so, tls, KTLS_TX);
1388 
1389 	if (error) {
1390 		ktls_cleanup(tls);
1391 		return (error);
1392 	}
1393 
1394 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1395 	if (error) {
1396 		ktls_cleanup(tls);
1397 		return (error);
1398 	}
1399 
1400 	/*
1401 	 * Write lock the INP when setting sb_tls_info so that
1402 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1403 	 * holding the INP lock.
1404 	 */
1405 	inp = so->so_pcb;
1406 	INP_WLOCK(inp);
1407 	SOCKBUF_LOCK(&so->so_snd);
1408 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1409 	so->so_snd.sb_tls_info = tls;
1410 	if (tls->mode != TCP_TLS_MODE_SW)
1411 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1412 	SOCKBUF_UNLOCK(&so->so_snd);
1413 	INP_WUNLOCK(inp);
1414 	SOCK_IO_SEND_UNLOCK(so);
1415 
1416 	counter_u64_add(ktls_offload_total, 1);
1417 
1418 	return (0);
1419 }
1420 
1421 int
1422 ktls_get_rx_mode(struct socket *so, int *modep)
1423 {
1424 	struct ktls_session *tls;
1425 	struct inpcb *inp __diagused;
1426 
1427 	if (SOLISTENING(so))
1428 		return (EINVAL);
1429 	inp = so->so_pcb;
1430 	INP_WLOCK_ASSERT(inp);
1431 	SOCK_RECVBUF_LOCK(so);
1432 	tls = so->so_rcv.sb_tls_info;
1433 	if (tls == NULL)
1434 		*modep = TCP_TLS_MODE_NONE;
1435 	else
1436 		*modep = tls->mode;
1437 	SOCK_RECVBUF_UNLOCK(so);
1438 	return (0);
1439 }
1440 
1441 /*
1442  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1443  *
1444  * This function gets information about the next TCP- and TLS-
1445  * sequence number to be processed by the TLS receive worker
1446  * thread. The information is extracted from the given "inpcb"
1447  * structure. The values are stored in host endian format at the two
1448  * given output pointer locations. The TCP sequence number points to
1449  * the beginning of the TLS header.
1450  *
1451  * This function returns zero on success, else a non-zero error code
1452  * is returned.
1453  */
1454 int
1455 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1456 {
1457 	struct socket *so;
1458 	struct tcpcb *tp;
1459 
1460 	INP_RLOCK(inp);
1461 	so = inp->inp_socket;
1462 	if (__predict_false(so == NULL)) {
1463 		INP_RUNLOCK(inp);
1464 		return (EINVAL);
1465 	}
1466 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
1467 		INP_RUNLOCK(inp);
1468 		return (ECONNRESET);
1469 	}
1470 
1471 	tp = intotcpcb(inp);
1472 	MPASS(tp != NULL);
1473 
1474 	SOCKBUF_LOCK(&so->so_rcv);
1475 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1476 	*tlsseq = so->so_rcv.sb_tls_seqno;
1477 	SOCKBUF_UNLOCK(&so->so_rcv);
1478 
1479 	INP_RUNLOCK(inp);
1480 
1481 	return (0);
1482 }
1483 
1484 int
1485 ktls_get_tx_mode(struct socket *so, int *modep)
1486 {
1487 	struct ktls_session *tls;
1488 	struct inpcb *inp __diagused;
1489 
1490 	if (SOLISTENING(so))
1491 		return (EINVAL);
1492 	inp = so->so_pcb;
1493 	INP_WLOCK_ASSERT(inp);
1494 	SOCK_SENDBUF_LOCK(so);
1495 	tls = so->so_snd.sb_tls_info;
1496 	if (tls == NULL)
1497 		*modep = TCP_TLS_MODE_NONE;
1498 	else
1499 		*modep = tls->mode;
1500 	SOCK_SENDBUF_UNLOCK(so);
1501 	return (0);
1502 }
1503 
1504 /*
1505  * Switch between SW and ifnet TLS sessions as requested.
1506  */
1507 int
1508 ktls_set_tx_mode(struct socket *so, int mode)
1509 {
1510 	struct ktls_session *tls, *tls_new;
1511 	struct inpcb *inp;
1512 	int error;
1513 
1514 	if (SOLISTENING(so))
1515 		return (EINVAL);
1516 	switch (mode) {
1517 	case TCP_TLS_MODE_SW:
1518 	case TCP_TLS_MODE_IFNET:
1519 		break;
1520 	default:
1521 		return (EINVAL);
1522 	}
1523 
1524 	inp = so->so_pcb;
1525 	INP_WLOCK_ASSERT(inp);
1526 	SOCKBUF_LOCK(&so->so_snd);
1527 	tls = so->so_snd.sb_tls_info;
1528 	if (tls == NULL) {
1529 		SOCKBUF_UNLOCK(&so->so_snd);
1530 		return (0);
1531 	}
1532 
1533 	if (tls->mode == mode) {
1534 		SOCKBUF_UNLOCK(&so->so_snd);
1535 		return (0);
1536 	}
1537 
1538 	tls = ktls_hold(tls);
1539 	SOCKBUF_UNLOCK(&so->so_snd);
1540 	INP_WUNLOCK(inp);
1541 
1542 	tls_new = ktls_clone_session(tls, KTLS_TX);
1543 
1544 	if (mode == TCP_TLS_MODE_IFNET)
1545 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1546 	else
1547 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1548 	if (error) {
1549 		counter_u64_add(ktls_switch_failed, 1);
1550 		ktls_free(tls_new);
1551 		ktls_free(tls);
1552 		INP_WLOCK(inp);
1553 		return (error);
1554 	}
1555 
1556 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1557 	if (error) {
1558 		counter_u64_add(ktls_switch_failed, 1);
1559 		ktls_free(tls_new);
1560 		ktls_free(tls);
1561 		INP_WLOCK(inp);
1562 		return (error);
1563 	}
1564 
1565 	/*
1566 	 * If we raced with another session change, keep the existing
1567 	 * session.
1568 	 */
1569 	if (tls != so->so_snd.sb_tls_info) {
1570 		counter_u64_add(ktls_switch_failed, 1);
1571 		SOCK_IO_SEND_UNLOCK(so);
1572 		ktls_free(tls_new);
1573 		ktls_free(tls);
1574 		INP_WLOCK(inp);
1575 		return (EBUSY);
1576 	}
1577 
1578 	INP_WLOCK(inp);
1579 	SOCKBUF_LOCK(&so->so_snd);
1580 	so->so_snd.sb_tls_info = tls_new;
1581 	if (tls_new->mode != TCP_TLS_MODE_SW)
1582 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1583 	SOCKBUF_UNLOCK(&so->so_snd);
1584 	SOCK_IO_SEND_UNLOCK(so);
1585 
1586 	/*
1587 	 * Drop two references on 'tls'.  The first is for the
1588 	 * ktls_hold() above.  The second drops the reference from the
1589 	 * socket buffer.
1590 	 */
1591 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1592 	ktls_free(tls);
1593 	ktls_free(tls);
1594 
1595 	if (mode == TCP_TLS_MODE_IFNET)
1596 		counter_u64_add(ktls_switch_to_ifnet, 1);
1597 	else
1598 		counter_u64_add(ktls_switch_to_sw, 1);
1599 
1600 	return (0);
1601 }
1602 
1603 /*
1604  * Try to allocate a new TLS receive tag.  This task is scheduled when
1605  * sbappend_ktls_rx detects an input path change.  If a new tag is
1606  * allocated, replace the tag in the TLS session.  If a new tag cannot
1607  * be allocated, let the session fall back to software decryption.
1608  */
1609 static void
1610 ktls_reset_receive_tag(void *context, int pending)
1611 {
1612 	union if_snd_tag_alloc_params params;
1613 	struct ktls_session *tls;
1614 	struct m_snd_tag *mst;
1615 	struct inpcb *inp;
1616 	struct ifnet *ifp;
1617 	struct socket *so;
1618 	int error;
1619 
1620 	MPASS(pending == 1);
1621 
1622 	tls = context;
1623 	so = tls->so;
1624 	inp = so->so_pcb;
1625 	ifp = NULL;
1626 
1627 	INP_RLOCK(inp);
1628 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
1629 		INP_RUNLOCK(inp);
1630 		goto out;
1631 	}
1632 
1633 	SOCKBUF_LOCK(&so->so_rcv);
1634 	m_snd_tag_rele(tls->snd_tag);
1635 	tls->snd_tag = NULL;
1636 
1637 	ifp = tls->rx_ifp;
1638 	if_ref(ifp);
1639 	SOCKBUF_UNLOCK(&so->so_rcv);
1640 
1641 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1642 	params.hdr.flowid = inp->inp_flowid;
1643 	params.hdr.flowtype = inp->inp_flowtype;
1644 	params.hdr.numa_domain = inp->inp_numa_domain;
1645 	params.tls_rx.inp = inp;
1646 	params.tls_rx.tls = tls;
1647 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1648 	INP_RUNLOCK(inp);
1649 
1650 	if (inp->inp_vflag & INP_IPV6) {
1651 		if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0)
1652 			goto out;
1653 	} else {
1654 		if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0)
1655 			goto out;
1656 	}
1657 
1658 	error = m_snd_tag_alloc(ifp, &params, &mst);
1659 	if (error == 0) {
1660 		SOCKBUF_LOCK(&so->so_rcv);
1661 		tls->snd_tag = mst;
1662 		SOCKBUF_UNLOCK(&so->so_rcv);
1663 
1664 		counter_u64_add(ktls_ifnet_reset, 1);
1665 	} else {
1666 		/*
1667 		 * Just fall back to software decryption if a tag
1668 		 * cannot be allocated leaving the connection intact.
1669 		 * If a future input path change switches to another
1670 		 * interface this connection will resume ifnet TLS.
1671 		 */
1672 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1673 	}
1674 
1675 out:
1676 	mtx_pool_lock(mtxpool_sleep, tls);
1677 	tls->reset_pending = false;
1678 	mtx_pool_unlock(mtxpool_sleep, tls);
1679 
1680 	if (ifp != NULL)
1681 		if_rele(ifp);
1682 	sorele(so);
1683 	ktls_free(tls);
1684 }
1685 
1686 /*
1687  * Try to allocate a new TLS send tag.  This task is scheduled when
1688  * ip_output detects a route change while trying to transmit a packet
1689  * holding a TLS record.  If a new tag is allocated, replace the tag
1690  * in the TLS session.  Subsequent packets on the connection will use
1691  * the new tag.  If a new tag cannot be allocated, drop the
1692  * connection.
1693  */
1694 static void
1695 ktls_reset_send_tag(void *context, int pending)
1696 {
1697 	struct epoch_tracker et;
1698 	struct ktls_session *tls;
1699 	struct m_snd_tag *old, *new;
1700 	struct inpcb *inp;
1701 	struct tcpcb *tp;
1702 	int error;
1703 
1704 	MPASS(pending == 1);
1705 
1706 	tls = context;
1707 	inp = tls->inp;
1708 
1709 	/*
1710 	 * Free the old tag first before allocating a new one.
1711 	 * ip[6]_output_send() will treat a NULL send tag the same as
1712 	 * an ifp mismatch and drop packets until a new tag is
1713 	 * allocated.
1714 	 *
1715 	 * Write-lock the INP when changing tls->snd_tag since
1716 	 * ip[6]_output_send() holds a read-lock when reading the
1717 	 * pointer.
1718 	 */
1719 	INP_WLOCK(inp);
1720 	old = tls->snd_tag;
1721 	tls->snd_tag = NULL;
1722 	INP_WUNLOCK(inp);
1723 	if (old != NULL)
1724 		m_snd_tag_rele(old);
1725 
1726 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1727 
1728 	if (error == 0) {
1729 		INP_WLOCK(inp);
1730 		tls->snd_tag = new;
1731 		mtx_pool_lock(mtxpool_sleep, tls);
1732 		tls->reset_pending = false;
1733 		mtx_pool_unlock(mtxpool_sleep, tls);
1734 		if (!in_pcbrele_wlocked(inp))
1735 			INP_WUNLOCK(inp);
1736 
1737 		counter_u64_add(ktls_ifnet_reset, 1);
1738 
1739 		/*
1740 		 * XXX: Should we kick tcp_output explicitly now that
1741 		 * the send tag is fixed or just rely on timers?
1742 		 */
1743 	} else {
1744 		NET_EPOCH_ENTER(et);
1745 		INP_WLOCK(inp);
1746 		if (!in_pcbrele_wlocked(inp)) {
1747 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1748 			    !(inp->inp_flags & INP_DROPPED)) {
1749 				tp = intotcpcb(inp);
1750 				CURVNET_SET(tp->t_vnet);
1751 				tp = tcp_drop(tp, ECONNABORTED);
1752 				CURVNET_RESTORE();
1753 				if (tp != NULL)
1754 					INP_WUNLOCK(inp);
1755 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1756 			} else
1757 				INP_WUNLOCK(inp);
1758 		}
1759 		NET_EPOCH_EXIT(et);
1760 
1761 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1762 
1763 		/*
1764 		 * Leave reset_pending true to avoid future tasks while
1765 		 * the socket goes away.
1766 		 */
1767 	}
1768 
1769 	ktls_free(tls);
1770 }
1771 
1772 void
1773 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1774 {
1775 	struct ktls_session *tls;
1776 	struct socket *so;
1777 
1778 	SOCKBUF_LOCK_ASSERT(sb);
1779 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1780 	    __func__, sb));
1781 	so = __containerof(sb, struct socket, so_rcv);
1782 
1783 	tls = sb->sb_tls_info;
1784 	if_rele(tls->rx_ifp);
1785 	if_ref(ifp);
1786 	tls->rx_ifp = ifp;
1787 
1788 	/*
1789 	 * See if we should schedule a task to update the receive tag for
1790 	 * this session.
1791 	 */
1792 	mtx_pool_lock(mtxpool_sleep, tls);
1793 	if (!tls->reset_pending) {
1794 		(void) ktls_hold(tls);
1795 		soref(so);
1796 		tls->so = so;
1797 		tls->reset_pending = true;
1798 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1799 	}
1800 	mtx_pool_unlock(mtxpool_sleep, tls);
1801 }
1802 
1803 int
1804 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1805 {
1806 
1807 	if (inp == NULL)
1808 		return (ENOBUFS);
1809 
1810 	INP_LOCK_ASSERT(inp);
1811 
1812 	/*
1813 	 * See if we should schedule a task to update the send tag for
1814 	 * this session.
1815 	 */
1816 	mtx_pool_lock(mtxpool_sleep, tls);
1817 	if (!tls->reset_pending) {
1818 		(void) ktls_hold(tls);
1819 		in_pcbref(inp);
1820 		tls->inp = inp;
1821 		tls->reset_pending = true;
1822 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1823 	}
1824 	mtx_pool_unlock(mtxpool_sleep, tls);
1825 	return (ENOBUFS);
1826 }
1827 
1828 #ifdef RATELIMIT
1829 int
1830 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1831 {
1832 	union if_snd_tag_modify_params params = {
1833 		.rate_limit.max_rate = max_pacing_rate,
1834 		.rate_limit.flags = M_NOWAIT,
1835 	};
1836 	struct m_snd_tag *mst;
1837 
1838 	/* Can't get to the inp, but it should be locked. */
1839 	/* INP_LOCK_ASSERT(inp); */
1840 
1841 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1842 
1843 	if (tls->snd_tag == NULL) {
1844 		/*
1845 		 * Resetting send tag, ignore this change.  The
1846 		 * pending reset may or may not see this updated rate
1847 		 * in the tcpcb.  If it doesn't, we will just lose
1848 		 * this rate change.
1849 		 */
1850 		return (0);
1851 	}
1852 
1853 	mst = tls->snd_tag;
1854 
1855 	MPASS(mst != NULL);
1856 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1857 
1858 	return (mst->sw->snd_tag_modify(mst, &params));
1859 }
1860 #endif
1861 #endif
1862 
1863 void
1864 ktls_destroy(struct ktls_session *tls)
1865 {
1866 
1867 	if (tls->sequential_records) {
1868 		struct mbuf *m, *n;
1869 		int page_count;
1870 
1871 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1872 			page_count = m->m_epg_enc_cnt;
1873 			while (page_count > 0) {
1874 				KASSERT(page_count >= m->m_epg_nrdy,
1875 				    ("%s: too few pages", __func__));
1876 				page_count -= m->m_epg_nrdy;
1877 				m = m_free(m);
1878 			}
1879 		}
1880 	}
1881 	ktls_cleanup(tls);
1882 	uma_zfree(ktls_session_zone, tls);
1883 }
1884 
1885 void
1886 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1887 {
1888 
1889 	for (; m != NULL; m = m->m_next) {
1890 		KASSERT((m->m_flags & M_EXTPG) != 0,
1891 		    ("ktls_seq: mapped mbuf %p", m));
1892 
1893 		m->m_epg_seqno = sb->sb_tls_seqno;
1894 		sb->sb_tls_seqno++;
1895 	}
1896 }
1897 
1898 /*
1899  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1900  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1901  * mbuf must be populated with the payload of each TLS record.
1902  *
1903  * The record_type argument specifies the TLS record type used when
1904  * populating the TLS header.
1905  *
1906  * The enq_count argument on return is set to the number of pages of
1907  * payload data for this entire chain that need to be encrypted via SW
1908  * encryption.  The returned value should be passed to ktls_enqueue
1909  * when scheduling encryption of this chain of mbufs.  To handle the
1910  * special case of empty fragments for TLS 1.0 sessions, an empty
1911  * fragment counts as one page.
1912  */
1913 void
1914 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1915     uint8_t record_type)
1916 {
1917 	struct tls_record_layer *tlshdr;
1918 	struct mbuf *m;
1919 	uint64_t *noncep;
1920 	uint16_t tls_len;
1921 	int maxlen __diagused;
1922 
1923 	maxlen = tls->params.max_frame_len;
1924 	*enq_cnt = 0;
1925 	for (m = top; m != NULL; m = m->m_next) {
1926 		/*
1927 		 * All mbufs in the chain should be TLS records whose
1928 		 * payload does not exceed the maximum frame length.
1929 		 *
1930 		 * Empty TLS 1.0 records are permitted when using CBC.
1931 		 */
1932 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1933 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1934 		    ("ktls_frame: m %p len %d", m, m->m_len));
1935 
1936 		/*
1937 		 * TLS frames require unmapped mbufs to store session
1938 		 * info.
1939 		 */
1940 		KASSERT((m->m_flags & M_EXTPG) != 0,
1941 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1942 
1943 		tls_len = m->m_len;
1944 
1945 		/* Save a reference to the session. */
1946 		m->m_epg_tls = ktls_hold(tls);
1947 
1948 		m->m_epg_hdrlen = tls->params.tls_hlen;
1949 		m->m_epg_trllen = tls->params.tls_tlen;
1950 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1951 			int bs, delta;
1952 
1953 			/*
1954 			 * AES-CBC pads messages to a multiple of the
1955 			 * block size.  Note that the padding is
1956 			 * applied after the digest and the encryption
1957 			 * is done on the "plaintext || mac || padding".
1958 			 * At least one byte of padding is always
1959 			 * present.
1960 			 *
1961 			 * Compute the final trailer length assuming
1962 			 * at most one block of padding.
1963 			 * tls->params.tls_tlen is the maximum
1964 			 * possible trailer length (padding + digest).
1965 			 * delta holds the number of excess padding
1966 			 * bytes if the maximum were used.  Those
1967 			 * extra bytes are removed.
1968 			 */
1969 			bs = tls->params.tls_bs;
1970 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1971 			m->m_epg_trllen -= delta;
1972 		}
1973 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1974 
1975 		/* Populate the TLS header. */
1976 		tlshdr = (void *)m->m_epg_hdr;
1977 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1978 
1979 		/*
1980 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1981 		 * of TLS_RLTYPE_APP.
1982 		 */
1983 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1984 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1985 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1986 			tlshdr->tls_type = TLS_RLTYPE_APP;
1987 			/* save the real record type for later */
1988 			m->m_epg_record_type = record_type;
1989 			m->m_epg_trail[0] = record_type;
1990 		} else {
1991 			tlshdr->tls_vminor = tls->params.tls_vminor;
1992 			tlshdr->tls_type = record_type;
1993 		}
1994 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1995 
1996 		/*
1997 		 * Store nonces / explicit IVs after the end of the
1998 		 * TLS header.
1999 		 *
2000 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2001 		 * from the end of the IV.  The nonce is then
2002 		 * incremented for use by the next record.
2003 		 *
2004 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2005 		 */
2006 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2007 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2008 			noncep = (uint64_t *)(tls->params.iv + 8);
2009 			be64enc(tlshdr + 1, *noncep);
2010 			(*noncep)++;
2011 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2012 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2013 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2014 
2015 		/*
2016 		 * When using SW encryption, mark the mbuf not ready.
2017 		 * It will be marked ready via sbready() after the
2018 		 * record has been encrypted.
2019 		 *
2020 		 * When using ifnet TLS, unencrypted TLS records are
2021 		 * sent down the stack to the NIC.
2022 		 */
2023 		if (tls->mode == TCP_TLS_MODE_SW) {
2024 			m->m_flags |= M_NOTREADY;
2025 			if (__predict_false(tls_len == 0)) {
2026 				/* TLS 1.0 empty fragment. */
2027 				m->m_epg_nrdy = 1;
2028 			} else
2029 				m->m_epg_nrdy = m->m_epg_npgs;
2030 			*enq_cnt += m->m_epg_nrdy;
2031 		}
2032 	}
2033 }
2034 
2035 bool
2036 ktls_permit_empty_frames(struct ktls_session *tls)
2037 {
2038 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2039 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2040 }
2041 
2042 void
2043 ktls_check_rx(struct sockbuf *sb)
2044 {
2045 	struct tls_record_layer hdr;
2046 	struct ktls_wq *wq;
2047 	struct socket *so;
2048 	bool running;
2049 
2050 	SOCKBUF_LOCK_ASSERT(sb);
2051 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2052 	    __func__, sb));
2053 	so = __containerof(sb, struct socket, so_rcv);
2054 
2055 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2056 		return;
2057 
2058 	/* Is there enough queued for a TLS header? */
2059 	if (sb->sb_tlscc < sizeof(hdr)) {
2060 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2061 			so->so_error = EMSGSIZE;
2062 		return;
2063 	}
2064 
2065 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2066 
2067 	/* Is the entire record queued? */
2068 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2069 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2070 			so->so_error = EMSGSIZE;
2071 		return;
2072 	}
2073 
2074 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2075 
2076 	soref(so);
2077 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2078 	mtx_lock(&wq->mtx);
2079 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2080 	running = wq->running;
2081 	mtx_unlock(&wq->mtx);
2082 	if (!running)
2083 		wakeup(wq);
2084 	counter_u64_add(ktls_cnt_rx_queued, 1);
2085 }
2086 
2087 static struct mbuf *
2088 ktls_detach_record(struct sockbuf *sb, int len)
2089 {
2090 	struct mbuf *m, *n, *top;
2091 	int remain;
2092 
2093 	SOCKBUF_LOCK_ASSERT(sb);
2094 	MPASS(len <= sb->sb_tlscc);
2095 
2096 	/*
2097 	 * If TLS chain is the exact size of the record,
2098 	 * just grab the whole record.
2099 	 */
2100 	top = sb->sb_mtls;
2101 	if (sb->sb_tlscc == len) {
2102 		sb->sb_mtls = NULL;
2103 		sb->sb_mtlstail = NULL;
2104 		goto out;
2105 	}
2106 
2107 	/*
2108 	 * While it would be nice to use m_split() here, we need
2109 	 * to know exactly what m_split() allocates to update the
2110 	 * accounting, so do it inline instead.
2111 	 */
2112 	remain = len;
2113 	for (m = top; remain > m->m_len; m = m->m_next)
2114 		remain -= m->m_len;
2115 
2116 	/* Easy case: don't have to split 'm'. */
2117 	if (remain == m->m_len) {
2118 		sb->sb_mtls = m->m_next;
2119 		if (sb->sb_mtls == NULL)
2120 			sb->sb_mtlstail = NULL;
2121 		m->m_next = NULL;
2122 		goto out;
2123 	}
2124 
2125 	/*
2126 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2127 	 * with M_NOWAIT first.
2128 	 */
2129 	n = m_get(M_NOWAIT, MT_DATA);
2130 	if (n == NULL) {
2131 		/*
2132 		 * Use M_WAITOK with socket buffer unlocked.  If
2133 		 * 'sb_mtls' changes while the lock is dropped, return
2134 		 * NULL to force the caller to retry.
2135 		 */
2136 		SOCKBUF_UNLOCK(sb);
2137 
2138 		n = m_get(M_WAITOK, MT_DATA);
2139 
2140 		SOCKBUF_LOCK(sb);
2141 		if (sb->sb_mtls != top) {
2142 			m_free(n);
2143 			return (NULL);
2144 		}
2145 	}
2146 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2147 
2148 	/* Store remainder in 'n'. */
2149 	n->m_len = m->m_len - remain;
2150 	if (m->m_flags & M_EXT) {
2151 		n->m_data = m->m_data + remain;
2152 		mb_dupcl(n, m);
2153 	} else {
2154 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2155 	}
2156 
2157 	/* Trim 'm' and update accounting. */
2158 	m->m_len -= n->m_len;
2159 	sb->sb_tlscc -= n->m_len;
2160 	sb->sb_ccc -= n->m_len;
2161 
2162 	/* Account for 'n'. */
2163 	sballoc_ktls_rx(sb, n);
2164 
2165 	/* Insert 'n' into the TLS chain. */
2166 	sb->sb_mtls = n;
2167 	n->m_next = m->m_next;
2168 	if (sb->sb_mtlstail == m)
2169 		sb->sb_mtlstail = n;
2170 
2171 	/* Detach the record from the TLS chain. */
2172 	m->m_next = NULL;
2173 
2174 out:
2175 	MPASS(m_length(top, NULL) == len);
2176 	for (m = top; m != NULL; m = m->m_next)
2177 		sbfree_ktls_rx(sb, m);
2178 	sb->sb_tlsdcc = len;
2179 	sb->sb_ccc += len;
2180 	SBCHECK(sb);
2181 	return (top);
2182 }
2183 
2184 /*
2185  * Determine the length of the trailing zero padding and find the real
2186  * record type in the byte before the padding.
2187  *
2188  * Walking the mbuf chain backwards is clumsy, so another option would
2189  * be to scan forwards remembering the last non-zero byte before the
2190  * trailer.  However, it would be expensive to scan the entire record.
2191  * Instead, find the last non-zero byte of each mbuf in the chain
2192  * keeping track of the relative offset of that nonzero byte.
2193  *
2194  * trail_len is the size of the MAC/tag on input and is set to the
2195  * size of the full trailer including padding and the record type on
2196  * return.
2197  */
2198 static int
2199 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2200     int *trailer_len, uint8_t *record_typep)
2201 {
2202 	char *cp;
2203 	u_int digest_start, last_offset, m_len, offset;
2204 	uint8_t record_type;
2205 
2206 	digest_start = tls_len - *trailer_len;
2207 	last_offset = 0;
2208 	offset = 0;
2209 	for (; m != NULL && offset < digest_start;
2210 	     offset += m->m_len, m = m->m_next) {
2211 		/* Don't look for padding in the tag. */
2212 		m_len = min(digest_start - offset, m->m_len);
2213 		cp = mtod(m, char *);
2214 
2215 		/* Find last non-zero byte in this mbuf. */
2216 		while (m_len > 0 && cp[m_len - 1] == 0)
2217 			m_len--;
2218 		if (m_len > 0) {
2219 			record_type = cp[m_len - 1];
2220 			last_offset = offset + m_len;
2221 		}
2222 	}
2223 	if (last_offset < tls->params.tls_hlen)
2224 		return (EBADMSG);
2225 
2226 	*record_typep = record_type;
2227 	*trailer_len = tls_len - last_offset + 1;
2228 	return (0);
2229 }
2230 
2231 /*
2232  * Check if a mbuf chain is fully decrypted at the given offset and
2233  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2234  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2235  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2236  * is encrypted.
2237  */
2238 ktls_mbuf_crypto_st_t
2239 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2240 {
2241 	int m_flags_ored = 0;
2242 	int m_flags_anded = -1;
2243 
2244 	for (; mb != NULL; mb = mb->m_next) {
2245 		if (offset < mb->m_len)
2246 			break;
2247 		offset -= mb->m_len;
2248 	}
2249 	offset += len;
2250 
2251 	for (; mb != NULL; mb = mb->m_next) {
2252 		m_flags_ored |= mb->m_flags;
2253 		m_flags_anded &= mb->m_flags;
2254 
2255 		if (offset <= mb->m_len)
2256 			break;
2257 		offset -= mb->m_len;
2258 	}
2259 	MPASS(mb != NULL || offset == 0);
2260 
2261 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2262 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2263 	else
2264 		return ((m_flags_ored & M_DECRYPTED) ?
2265 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2266 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2267 }
2268 
2269 /*
2270  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2271  */
2272 static int
2273 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2274 {
2275 	union if_snd_tag_modify_params params;
2276 	struct m_snd_tag *mst;
2277 	struct inpcb *inp;
2278 	struct tcpcb *tp;
2279 
2280 	mst = so->so_rcv.sb_tls_info->snd_tag;
2281 	if (__predict_false(mst == NULL))
2282 		return (EINVAL);
2283 
2284 	inp = sotoinpcb(so);
2285 	if (__predict_false(inp == NULL))
2286 		return (EINVAL);
2287 
2288 	INP_RLOCK(inp);
2289 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2290 		INP_RUNLOCK(inp);
2291 		return (ECONNRESET);
2292 	}
2293 
2294 	tp = intotcpcb(inp);
2295 	MPASS(tp != NULL);
2296 
2297 	/* Get the TCP sequence number of the next valid TLS header. */
2298 	SOCKBUF_LOCK(&so->so_rcv);
2299 	params.tls_rx.tls_hdr_tcp_sn =
2300 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2301 	params.tls_rx.tls_rec_length = tls_len;
2302 	params.tls_rx.tls_seq_number = tls_rcd_num;
2303 	SOCKBUF_UNLOCK(&so->so_rcv);
2304 
2305 	INP_RUNLOCK(inp);
2306 
2307 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2308 	return (mst->sw->snd_tag_modify(mst, &params));
2309 }
2310 
2311 static void
2312 ktls_decrypt(struct socket *so)
2313 {
2314 	char tls_header[MBUF_PEXT_HDR_LEN];
2315 	struct ktls_session *tls;
2316 	struct sockbuf *sb;
2317 	struct tls_record_layer *hdr;
2318 	struct tls_get_record tgr;
2319 	struct mbuf *control, *data, *m;
2320 	ktls_mbuf_crypto_st_t state;
2321 	uint64_t seqno;
2322 	int error, remain, tls_len, trail_len;
2323 	bool tls13;
2324 	uint8_t vminor, record_type;
2325 
2326 	hdr = (struct tls_record_layer *)tls_header;
2327 	sb = &so->so_rcv;
2328 	SOCKBUF_LOCK(sb);
2329 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2330 	    ("%s: socket %p not running", __func__, so));
2331 
2332 	tls = sb->sb_tls_info;
2333 	MPASS(tls != NULL);
2334 
2335 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2336 	if (tls13)
2337 		vminor = TLS_MINOR_VER_TWO;
2338 	else
2339 		vminor = tls->params.tls_vminor;
2340 	for (;;) {
2341 		/* Is there enough queued for a TLS header? */
2342 		if (sb->sb_tlscc < tls->params.tls_hlen)
2343 			break;
2344 
2345 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2346 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2347 
2348 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2349 		    hdr->tls_vminor != vminor)
2350 			error = EINVAL;
2351 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2352 			error = EINVAL;
2353 		else if (tls_len < tls->params.tls_hlen || tls_len >
2354 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2355 		    tls->params.tls_tlen)
2356 			error = EMSGSIZE;
2357 		else
2358 			error = 0;
2359 		if (__predict_false(error != 0)) {
2360 			/*
2361 			 * We have a corrupted record and are likely
2362 			 * out of sync.  The connection isn't
2363 			 * recoverable at this point, so abort it.
2364 			 */
2365 			SOCKBUF_UNLOCK(sb);
2366 			counter_u64_add(ktls_offload_corrupted_records, 1);
2367 
2368 			CURVNET_SET(so->so_vnet);
2369 			so->so_proto->pr_usrreqs->pru_abort(so);
2370 			so->so_error = error;
2371 			CURVNET_RESTORE();
2372 			goto deref;
2373 		}
2374 
2375 		/* Is the entire record queued? */
2376 		if (sb->sb_tlscc < tls_len)
2377 			break;
2378 
2379 		/*
2380 		 * Split out the portion of the mbuf chain containing
2381 		 * this TLS record.
2382 		 */
2383 		data = ktls_detach_record(sb, tls_len);
2384 		if (data == NULL)
2385 			continue;
2386 		MPASS(sb->sb_tlsdcc == tls_len);
2387 
2388 		seqno = sb->sb_tls_seqno;
2389 		sb->sb_tls_seqno++;
2390 		SBCHECK(sb);
2391 		SOCKBUF_UNLOCK(sb);
2392 
2393 		/* get crypto state for this TLS record */
2394 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2395 
2396 		switch (state) {
2397 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2398 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2399 			if (error)
2400 				break;
2401 			/* FALLTHROUGH */
2402 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2403 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2404 			    &trail_len);
2405 			if (__predict_true(error == 0)) {
2406 				if (tls13) {
2407 					error = tls13_find_record_type(tls, data,
2408 					    tls_len, &trail_len, &record_type);
2409 				} else {
2410 					record_type = hdr->tls_type;
2411 				}
2412 			}
2413 			break;
2414 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2415 			/*
2416 			 * NIC TLS is only supported for AEAD
2417 			 * ciphersuites which used a fixed sized
2418 			 * trailer.
2419 			 */
2420 			if (tls13) {
2421 				trail_len = tls->params.tls_tlen - 1;
2422 				error = tls13_find_record_type(tls, data,
2423 				    tls_len, &trail_len, &record_type);
2424 			} else {
2425 				trail_len = tls->params.tls_tlen;
2426 				error = 0;
2427 				record_type = hdr->tls_type;
2428 			}
2429 			break;
2430 		default:
2431 			error = EINVAL;
2432 			break;
2433 		}
2434 		if (error) {
2435 			counter_u64_add(ktls_offload_failed_crypto, 1);
2436 
2437 			SOCKBUF_LOCK(sb);
2438 			if (sb->sb_tlsdcc == 0) {
2439 				/*
2440 				 * sbcut/drop/flush discarded these
2441 				 * mbufs.
2442 				 */
2443 				m_freem(data);
2444 				break;
2445 			}
2446 
2447 			/*
2448 			 * Drop this TLS record's data, but keep
2449 			 * decrypting subsequent records.
2450 			 */
2451 			sb->sb_ccc -= tls_len;
2452 			sb->sb_tlsdcc = 0;
2453 
2454 			CURVNET_SET(so->so_vnet);
2455 			so->so_error = EBADMSG;
2456 			sorwakeup_locked(so);
2457 			CURVNET_RESTORE();
2458 
2459 			m_freem(data);
2460 
2461 			SOCKBUF_LOCK(sb);
2462 			continue;
2463 		}
2464 
2465 		/* Allocate the control mbuf. */
2466 		memset(&tgr, 0, sizeof(tgr));
2467 		tgr.tls_type = record_type;
2468 		tgr.tls_vmajor = hdr->tls_vmajor;
2469 		tgr.tls_vminor = hdr->tls_vminor;
2470 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2471 		    trail_len);
2472 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2473 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2474 
2475 		SOCKBUF_LOCK(sb);
2476 		if (sb->sb_tlsdcc == 0) {
2477 			/* sbcut/drop/flush discarded these mbufs. */
2478 			MPASS(sb->sb_tlscc == 0);
2479 			m_freem(data);
2480 			m_freem(control);
2481 			break;
2482 		}
2483 
2484 		/*
2485 		 * Clear the 'dcc' accounting in preparation for
2486 		 * adding the decrypted record.
2487 		 */
2488 		sb->sb_ccc -= tls_len;
2489 		sb->sb_tlsdcc = 0;
2490 		SBCHECK(sb);
2491 
2492 		/* If there is no payload, drop all of the data. */
2493 		if (tgr.tls_length == htobe16(0)) {
2494 			m_freem(data);
2495 			data = NULL;
2496 		} else {
2497 			/* Trim header. */
2498 			remain = tls->params.tls_hlen;
2499 			while (remain > 0) {
2500 				if (data->m_len > remain) {
2501 					data->m_data += remain;
2502 					data->m_len -= remain;
2503 					break;
2504 				}
2505 				remain -= data->m_len;
2506 				data = m_free(data);
2507 			}
2508 
2509 			/* Trim trailer and clear M_NOTREADY. */
2510 			remain = be16toh(tgr.tls_length);
2511 			m = data;
2512 			for (m = data; remain > m->m_len; m = m->m_next) {
2513 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2514 				remain -= m->m_len;
2515 			}
2516 			m->m_len = remain;
2517 			m_freem(m->m_next);
2518 			m->m_next = NULL;
2519 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2520 
2521 			/* Set EOR on the final mbuf. */
2522 			m->m_flags |= M_EOR;
2523 		}
2524 
2525 		sbappendcontrol_locked(sb, data, control, 0);
2526 
2527 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2528 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2529 			SOCKBUF_UNLOCK(sb);
2530 			ktls_resync_ifnet(so, tls_len, seqno);
2531 			SOCKBUF_LOCK(sb);
2532 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2533 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2534 			SOCKBUF_UNLOCK(sb);
2535 			ktls_resync_ifnet(so, 0, seqno);
2536 			SOCKBUF_LOCK(sb);
2537 		}
2538 	}
2539 
2540 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2541 
2542 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2543 		so->so_error = EMSGSIZE;
2544 
2545 	sorwakeup_locked(so);
2546 
2547 deref:
2548 	SOCKBUF_UNLOCK_ASSERT(sb);
2549 
2550 	CURVNET_SET(so->so_vnet);
2551 	sorele(so);
2552 	CURVNET_RESTORE();
2553 }
2554 
2555 void
2556 ktls_enqueue_to_free(struct mbuf *m)
2557 {
2558 	struct ktls_wq *wq;
2559 	bool running;
2560 
2561 	/* Mark it for freeing. */
2562 	m->m_epg_flags |= EPG_FLAG_2FREE;
2563 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2564 	mtx_lock(&wq->mtx);
2565 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2566 	running = wq->running;
2567 	mtx_unlock(&wq->mtx);
2568 	if (!running)
2569 		wakeup(wq);
2570 }
2571 
2572 static void *
2573 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2574 {
2575 	void *buf;
2576 	int domain, running;
2577 
2578 	if (m->m_epg_npgs <= 2)
2579 		return (NULL);
2580 	if (ktls_buffer_zone == NULL)
2581 		return (NULL);
2582 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2583 		/*
2584 		 * Rate-limit allocation attempts after a failure.
2585 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2586 		 * the free page queues and may fail consistently if memory is
2587 		 * fragmented.
2588 		 */
2589 		return (NULL);
2590 	}
2591 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2592 	if (buf == NULL) {
2593 		domain = PCPU_GET(domain);
2594 		wq->lastallocfail = ticks;
2595 
2596 		/*
2597 		 * Note that this check is "racy", but the races are
2598 		 * harmless, and are either a spurious wakeup if
2599 		 * multiple threads fail allocations before the alloc
2600 		 * thread wakes, or waiting an extra second in case we
2601 		 * see an old value of running == true.
2602 		 */
2603 		if (!VM_DOMAIN_EMPTY(domain)) {
2604 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2605 			if (!running)
2606 				wakeup(&ktls_domains[domain].alloc_td);
2607 		}
2608 	}
2609 	return (buf);
2610 }
2611 
2612 static int
2613 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2614     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2615 {
2616 	vm_page_t pg;
2617 	int error, i, len, off;
2618 
2619 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2620 	    ("%p not unready & nomap mbuf\n", m));
2621 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2622 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2623 	    ktls_maxlen));
2624 
2625 	/* Anonymous mbufs are encrypted in place. */
2626 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2627 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2628 
2629 	/*
2630 	 * For file-backed mbufs (from sendfile), anonymous wired
2631 	 * pages are allocated and used as the encryption destination.
2632 	 */
2633 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2634 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2635 		    m->m_epg_1st_off;
2636 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2637 		    m->m_epg_1st_off;
2638 		state->dst_iov[0].iov_len = len;
2639 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2640 		i = 1;
2641 	} else {
2642 		off = m->m_epg_1st_off;
2643 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2644 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2645 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2646 			len = m_epg_pagelen(m, i, off);
2647 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2648 			state->dst_iov[i].iov_base =
2649 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2650 			state->dst_iov[i].iov_len = len;
2651 		}
2652 	}
2653 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2654 	state->dst_iov[i].iov_base = m->m_epg_trail;
2655 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2656 
2657 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2658 
2659 	if (__predict_false(error != 0)) {
2660 		/* Free the anonymous pages. */
2661 		if (state->cbuf != NULL)
2662 			uma_zfree(ktls_buffer_zone, state->cbuf);
2663 		else {
2664 			for (i = 0; i < m->m_epg_npgs; i++) {
2665 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2666 				(void)vm_page_unwire_noq(pg);
2667 				vm_page_free(pg);
2668 			}
2669 		}
2670 	}
2671 	return (error);
2672 }
2673 
2674 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2675 static u_int
2676 ktls_batched_records(struct mbuf *m)
2677 {
2678 	int page_count, records;
2679 
2680 	records = 0;
2681 	page_count = m->m_epg_enc_cnt;
2682 	while (page_count > 0) {
2683 		records++;
2684 		page_count -= m->m_epg_nrdy;
2685 		m = m->m_next;
2686 	}
2687 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2688 	return (records);
2689 }
2690 
2691 void
2692 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2693 {
2694 	struct ktls_session *tls;
2695 	struct ktls_wq *wq;
2696 	int queued;
2697 	bool running;
2698 
2699 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2700 	    (M_EXTPG | M_NOTREADY)),
2701 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2702 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2703 
2704 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2705 
2706 	m->m_epg_enc_cnt = page_count;
2707 
2708 	/*
2709 	 * Save a pointer to the socket.  The caller is responsible
2710 	 * for taking an additional reference via soref().
2711 	 */
2712 	m->m_epg_so = so;
2713 
2714 	queued = 1;
2715 	tls = m->m_epg_tls;
2716 	wq = &ktls_wq[tls->wq_index];
2717 	mtx_lock(&wq->mtx);
2718 	if (__predict_false(tls->sequential_records)) {
2719 		/*
2720 		 * For TLS 1.0, records must be encrypted
2721 		 * sequentially.  For a given connection, all records
2722 		 * queued to the associated work queue are processed
2723 		 * sequentially.  However, sendfile(2) might complete
2724 		 * I/O requests spanning multiple TLS records out of
2725 		 * order.  Here we ensure TLS records are enqueued to
2726 		 * the work queue in FIFO order.
2727 		 *
2728 		 * tls->next_seqno holds the sequence number of the
2729 		 * next TLS record that should be enqueued to the work
2730 		 * queue.  If this next record is not tls->next_seqno,
2731 		 * it must be a future record, so insert it, sorted by
2732 		 * TLS sequence number, into tls->pending_records and
2733 		 * return.
2734 		 *
2735 		 * If this TLS record matches tls->next_seqno, place
2736 		 * it in the work queue and then check
2737 		 * tls->pending_records to see if any
2738 		 * previously-queued records are now ready for
2739 		 * encryption.
2740 		 */
2741 		if (m->m_epg_seqno != tls->next_seqno) {
2742 			struct mbuf *n, *p;
2743 
2744 			p = NULL;
2745 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2746 				if (n->m_epg_seqno > m->m_epg_seqno)
2747 					break;
2748 				p = n;
2749 			}
2750 			if (n == NULL)
2751 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2752 				    m_epg_stailq);
2753 			else if (p == NULL)
2754 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2755 				    m_epg_stailq);
2756 			else
2757 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2758 				    m_epg_stailq);
2759 			mtx_unlock(&wq->mtx);
2760 			counter_u64_add(ktls_cnt_tx_pending, 1);
2761 			return;
2762 		}
2763 
2764 		tls->next_seqno += ktls_batched_records(m);
2765 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2766 
2767 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2768 			struct mbuf *n;
2769 
2770 			n = STAILQ_FIRST(&tls->pending_records);
2771 			if (n->m_epg_seqno != tls->next_seqno)
2772 				break;
2773 
2774 			queued++;
2775 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2776 			tls->next_seqno += ktls_batched_records(n);
2777 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2778 		}
2779 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2780 	} else
2781 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2782 
2783 	running = wq->running;
2784 	mtx_unlock(&wq->mtx);
2785 	if (!running)
2786 		wakeup(wq);
2787 	counter_u64_add(ktls_cnt_tx_queued, queued);
2788 }
2789 
2790 /*
2791  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2792  * the pages from the file and replace them with the anonymous pages
2793  * allocated in ktls_encrypt_record().
2794  */
2795 static void
2796 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2797 {
2798 	int i;
2799 
2800 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2801 
2802 	/* Free the old pages. */
2803 	m->m_ext.ext_free(m);
2804 
2805 	/* Replace them with the new pages. */
2806 	if (state->cbuf != NULL) {
2807 		for (i = 0; i < m->m_epg_npgs; i++)
2808 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2809 
2810 		/* Contig pages should go back to the cache. */
2811 		m->m_ext.ext_free = ktls_free_mext_contig;
2812 	} else {
2813 		for (i = 0; i < m->m_epg_npgs; i++)
2814 			m->m_epg_pa[i] = state->parray[i];
2815 
2816 		/* Use the basic free routine. */
2817 		m->m_ext.ext_free = mb_free_mext_pgs;
2818 	}
2819 
2820 	/* Pages are now writable. */
2821 	m->m_epg_flags |= EPG_FLAG_ANON;
2822 }
2823 
2824 static __noinline void
2825 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2826 {
2827 	struct ktls_ocf_encrypt_state state;
2828 	struct ktls_session *tls;
2829 	struct socket *so;
2830 	struct mbuf *m;
2831 	int error, npages, total_pages;
2832 
2833 	so = top->m_epg_so;
2834 	tls = top->m_epg_tls;
2835 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2836 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2837 #ifdef INVARIANTS
2838 	top->m_epg_so = NULL;
2839 #endif
2840 	total_pages = top->m_epg_enc_cnt;
2841 	npages = 0;
2842 
2843 	/*
2844 	 * Encrypt the TLS records in the chain of mbufs starting with
2845 	 * 'top'.  'total_pages' gives us a total count of pages and is
2846 	 * used to know when we have finished encrypting the TLS
2847 	 * records originally queued with 'top'.
2848 	 *
2849 	 * NB: These mbufs are queued in the socket buffer and
2850 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2851 	 * socket buffer lock is not held while traversing this chain.
2852 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2853 	 * pointers should be stable.  However, the 'm_next' of the
2854 	 * last mbuf encrypted is not necessarily NULL.  It can point
2855 	 * to other mbufs appended while 'top' was on the TLS work
2856 	 * queue.
2857 	 *
2858 	 * Each mbuf holds an entire TLS record.
2859 	 */
2860 	error = 0;
2861 	for (m = top; npages != total_pages; m = m->m_next) {
2862 		KASSERT(m->m_epg_tls == tls,
2863 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2864 		    tls, m->m_epg_tls));
2865 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2866 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2867 		    total_pages, m));
2868 
2869 		error = ktls_encrypt_record(wq, m, tls, &state);
2870 		if (error) {
2871 			counter_u64_add(ktls_offload_failed_crypto, 1);
2872 			break;
2873 		}
2874 
2875 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2876 			ktls_finish_nonanon(m, &state);
2877 
2878 		npages += m->m_epg_nrdy;
2879 
2880 		/*
2881 		 * Drop a reference to the session now that it is no
2882 		 * longer needed.  Existing code depends on encrypted
2883 		 * records having no associated session vs
2884 		 * yet-to-be-encrypted records having an associated
2885 		 * session.
2886 		 */
2887 		m->m_epg_tls = NULL;
2888 		ktls_free(tls);
2889 	}
2890 
2891 	CURVNET_SET(so->so_vnet);
2892 	if (error == 0) {
2893 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2894 	} else {
2895 		so->so_proto->pr_usrreqs->pru_abort(so);
2896 		so->so_error = EIO;
2897 		mb_free_notready(top, total_pages);
2898 	}
2899 
2900 	sorele(so);
2901 	CURVNET_RESTORE();
2902 }
2903 
2904 void
2905 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2906 {
2907 	struct ktls_session *tls;
2908 	struct socket *so;
2909 	struct mbuf *m;
2910 	int npages;
2911 
2912 	m = state->m;
2913 
2914 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2915 		ktls_finish_nonanon(m, state);
2916 
2917 	so = state->so;
2918 	free(state, M_KTLS);
2919 
2920 	/*
2921 	 * Drop a reference to the session now that it is no longer
2922 	 * needed.  Existing code depends on encrypted records having
2923 	 * no associated session vs yet-to-be-encrypted records having
2924 	 * an associated session.
2925 	 */
2926 	tls = m->m_epg_tls;
2927 	m->m_epg_tls = NULL;
2928 	ktls_free(tls);
2929 
2930 	if (error != 0)
2931 		counter_u64_add(ktls_offload_failed_crypto, 1);
2932 
2933 	CURVNET_SET(so->so_vnet);
2934 	npages = m->m_epg_nrdy;
2935 
2936 	if (error == 0) {
2937 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, m, npages);
2938 	} else {
2939 		so->so_proto->pr_usrreqs->pru_abort(so);
2940 		so->so_error = EIO;
2941 		mb_free_notready(m, npages);
2942 	}
2943 
2944 	sorele(so);
2945 	CURVNET_RESTORE();
2946 }
2947 
2948 /*
2949  * Similar to ktls_encrypt, but used with asynchronous OCF backends
2950  * (coprocessors) where encryption does not use host CPU resources and
2951  * it can be beneficial to queue more requests than CPUs.
2952  */
2953 static __noinline void
2954 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2955 {
2956 	struct ktls_ocf_encrypt_state *state;
2957 	struct ktls_session *tls;
2958 	struct socket *so;
2959 	struct mbuf *m, *n;
2960 	int error, mpages, npages, total_pages;
2961 
2962 	so = top->m_epg_so;
2963 	tls = top->m_epg_tls;
2964 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2965 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2966 #ifdef INVARIANTS
2967 	top->m_epg_so = NULL;
2968 #endif
2969 	total_pages = top->m_epg_enc_cnt;
2970 	npages = 0;
2971 
2972 	error = 0;
2973 	for (m = top; npages != total_pages; m = n) {
2974 		KASSERT(m->m_epg_tls == tls,
2975 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2976 		    tls, m->m_epg_tls));
2977 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2978 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2979 		    total_pages, m));
2980 
2981 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2982 		soref(so);
2983 		state->so = so;
2984 		state->m = m;
2985 
2986 		mpages = m->m_epg_nrdy;
2987 		n = m->m_next;
2988 
2989 		error = ktls_encrypt_record(wq, m, tls, state);
2990 		if (error) {
2991 			counter_u64_add(ktls_offload_failed_crypto, 1);
2992 			free(state, M_KTLS);
2993 			CURVNET_SET(so->so_vnet);
2994 			sorele(so);
2995 			CURVNET_RESTORE();
2996 			break;
2997 		}
2998 
2999 		npages += mpages;
3000 	}
3001 
3002 	CURVNET_SET(so->so_vnet);
3003 	if (error != 0) {
3004 		so->so_proto->pr_usrreqs->pru_abort(so);
3005 		so->so_error = EIO;
3006 		mb_free_notready(m, total_pages - npages);
3007 	}
3008 
3009 	sorele(so);
3010 	CURVNET_RESTORE();
3011 }
3012 
3013 static int
3014 ktls_bind_domain(int domain)
3015 {
3016 	int error;
3017 
3018 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3019 	if (error != 0)
3020 		return (error);
3021 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3022 	return (0);
3023 }
3024 
3025 static void
3026 ktls_alloc_thread(void *ctx)
3027 {
3028 	struct ktls_domain_info *ktls_domain = ctx;
3029 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3030 	void **buf;
3031 	struct sysctl_oid *oid;
3032 	char name[80];
3033 	int domain, error, i, nbufs;
3034 
3035 	domain = ktls_domain - ktls_domains;
3036 	if (bootverbose)
3037 		printf("Starting KTLS alloc thread for domain %d\n", domain);
3038 	error = ktls_bind_domain(domain);
3039 	if (error)
3040 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3041 		    domain, error);
3042 	snprintf(name, sizeof(name), "domain%d", domain);
3043 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3044 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3045 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3046 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
3047 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3048 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3049 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3050 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3051 
3052 	buf = NULL;
3053 	nbufs = 0;
3054 	for (;;) {
3055 		atomic_store_int(&sc->running, 0);
3056 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3057 		atomic_store_int(&sc->running, 1);
3058 		sc->wakeups++;
3059 		if (nbufs != ktls_max_alloc) {
3060 			free(buf, M_KTLS);
3061 			nbufs = atomic_load_int(&ktls_max_alloc);
3062 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3063 			    M_WAITOK | M_ZERO);
3064 		}
3065 		/*
3066 		 * Below we allocate nbufs with different allocation
3067 		 * flags than we use when allocating normally during
3068 		 * encryption in the ktls worker thread.  We specify
3069 		 * M_NORECLAIM in the worker thread. However, we omit
3070 		 * that flag here and add M_WAITOK so that the VM
3071 		 * system is permitted to perform expensive work to
3072 		 * defragment memory.  We do this here, as it does not
3073 		 * matter if this thread blocks.  If we block a ktls
3074 		 * worker thread, we risk developing backlogs of
3075 		 * buffers to be encrypted, leading to surges of
3076 		 * traffic and potential NIC output drops.
3077 		 */
3078 		for (i = 0; i < nbufs; i++) {
3079 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3080 			sc->allocs++;
3081 		}
3082 		for (i = 0; i < nbufs; i++) {
3083 			uma_zfree(ktls_buffer_zone, buf[i]);
3084 			buf[i] = NULL;
3085 		}
3086 	}
3087 }
3088 
3089 static void
3090 ktls_work_thread(void *ctx)
3091 {
3092 	struct ktls_wq *wq = ctx;
3093 	struct mbuf *m, *n;
3094 	struct socket *so, *son;
3095 	STAILQ_HEAD(, mbuf) local_m_head;
3096 	STAILQ_HEAD(, socket) local_so_head;
3097 	int cpu;
3098 
3099 	cpu = wq - ktls_wq;
3100 	if (bootverbose)
3101 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3102 
3103 	/*
3104 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3105 	 * we bind to the NUMA domain instead.
3106 	 */
3107 	if (ktls_bind_threads) {
3108 		int error;
3109 
3110 		if (ktls_bind_threads > 1) {
3111 			struct pcpu *pc = pcpu_find(cpu);
3112 
3113 			error = ktls_bind_domain(pc->pc_domain);
3114 		} else {
3115 			cpuset_t mask;
3116 
3117 			CPU_SETOF(cpu, &mask);
3118 			error = cpuset_setthread(curthread->td_tid, &mask);
3119 		}
3120 		if (error)
3121 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3122 				cpu, error);
3123 	}
3124 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3125 	fpu_kern_thread(0);
3126 #endif
3127 	for (;;) {
3128 		mtx_lock(&wq->mtx);
3129 		while (STAILQ_EMPTY(&wq->m_head) &&
3130 		    STAILQ_EMPTY(&wq->so_head)) {
3131 			wq->running = false;
3132 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3133 			wq->running = true;
3134 		}
3135 
3136 		STAILQ_INIT(&local_m_head);
3137 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3138 		STAILQ_INIT(&local_so_head);
3139 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3140 		mtx_unlock(&wq->mtx);
3141 
3142 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3143 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3144 				ktls_free(m->m_epg_tls);
3145 				m_free_raw(m);
3146 			} else {
3147 				if (m->m_epg_tls->sync_dispatch)
3148 					ktls_encrypt(wq, m);
3149 				else
3150 					ktls_encrypt_async(wq, m);
3151 				counter_u64_add(ktls_cnt_tx_queued, -1);
3152 			}
3153 		}
3154 
3155 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3156 			ktls_decrypt(so);
3157 			counter_u64_add(ktls_cnt_rx_queued, -1);
3158 		}
3159 	}
3160 }
3161 
3162 #if defined(INET) || defined(INET6)
3163 static void
3164 ktls_disable_ifnet_help(void *context, int pending __unused)
3165 {
3166 	struct ktls_session *tls;
3167 	struct inpcb *inp;
3168 	struct tcpcb *tp;
3169 	struct socket *so;
3170 	int err;
3171 
3172 	tls = context;
3173 	inp = tls->inp;
3174 	if (inp == NULL)
3175 		return;
3176 	INP_WLOCK(inp);
3177 	so = inp->inp_socket;
3178 	MPASS(so != NULL);
3179 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
3180 		goto out;
3181 	}
3182 
3183 	if (so->so_snd.sb_tls_info != NULL)
3184 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3185 	else
3186 		err = ENXIO;
3187 	if (err == 0) {
3188 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3189 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3190 		if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 &&
3191 		    (tp = intotcpcb(inp)) != NULL &&
3192 		    tp->t_fb->tfb_hwtls_change != NULL)
3193 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3194 	} else {
3195 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3196 	}
3197 
3198 out:
3199 	sorele(so);
3200 	if (!in_pcbrele_wlocked(inp))
3201 		INP_WUNLOCK(inp);
3202 	ktls_free(tls);
3203 }
3204 
3205 /*
3206  * Called when re-transmits are becoming a substantial portion of the
3207  * sends on this connection.  When this happens, we transition the
3208  * connection to software TLS.  This is needed because most inline TLS
3209  * NICs keep crypto state only for in-order transmits.  This means
3210  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3211  * re-DMA the entire TLS record up to and including the current
3212  * segment.  This means that when re-transmitting the last ~1448 byte
3213  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3214  * of magnitude more data than we are sending.  This can cause the
3215  * PCIe link to saturate well before the network, which can cause
3216  * output drops, and a general loss of capacity.
3217  */
3218 void
3219 ktls_disable_ifnet(void *arg)
3220 {
3221 	struct tcpcb *tp;
3222 	struct inpcb *inp;
3223 	struct socket *so;
3224 	struct ktls_session *tls;
3225 
3226 	tp = arg;
3227 	inp = tp->t_inpcb;
3228 	INP_WLOCK_ASSERT(inp);
3229 	so = inp->inp_socket;
3230 	SOCK_LOCK(so);
3231 	tls = so->so_snd.sb_tls_info;
3232 	if (tls->disable_ifnet_pending) {
3233 		SOCK_UNLOCK(so);
3234 		return;
3235 	}
3236 
3237 	/*
3238 	 * note that disable_ifnet_pending is never cleared; disabling
3239 	 * ifnet can only be done once per session, so we never want
3240 	 * to do it again
3241 	 */
3242 
3243 	(void)ktls_hold(tls);
3244 	in_pcbref(inp);
3245 	soref(so);
3246 	tls->disable_ifnet_pending = true;
3247 	tls->inp = inp;
3248 	SOCK_UNLOCK(so);
3249 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3250 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3251 }
3252 #endif
3253