xref: /freebsd/sys/kern/uipc_ktls.c (revision b4b33821fa3d970121d886467d10126679876905)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/domainset.h>
40 #include <sys/endian.h>
41 #include <sys/ktls.h>
42 #include <sys/lock.h>
43 #include <sys/mbuf.h>
44 #include <sys/mutex.h>
45 #include <sys/rmlock.h>
46 #include <sys/proc.h>
47 #include <sys/protosw.h>
48 #include <sys/refcount.h>
49 #include <sys/smp.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/kthread.h>
55 #include <sys/uio.h>
56 #include <sys/vmmeter.h>
57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58 #include <machine/pcb.h>
59 #endif
60 #include <machine/vmparam.h>
61 #include <net/if.h>
62 #include <net/if_var.h>
63 #ifdef RSS
64 #include <net/netisr.h>
65 #include <net/rss_config.h>
66 #endif
67 #include <net/route.h>
68 #include <net/route/nhop.h>
69 #if defined(INET) || defined(INET6)
70 #include <netinet/in.h>
71 #include <netinet/in_pcb.h>
72 #endif
73 #include <netinet/tcp_var.h>
74 #ifdef TCP_OFFLOAD
75 #include <netinet/tcp_offload.h>
76 #endif
77 #include <opencrypto/cryptodev.h>
78 #include <opencrypto/ktls.h>
79 #include <vm/uma_dbg.h>
80 #include <vm/vm.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pagequeue.h>
84 
85 struct ktls_wq {
86 	struct mtx	mtx;
87 	STAILQ_HEAD(, mbuf) m_head;
88 	STAILQ_HEAD(, socket) so_head;
89 	bool		running;
90 	int		lastallocfail;
91 } __aligned(CACHE_LINE_SIZE);
92 
93 struct ktls_alloc_thread {
94 	uint64_t wakeups;
95 	uint64_t allocs;
96 	struct thread *td;
97 	int running;
98 };
99 
100 struct ktls_domain_info {
101 	int count;
102 	int cpu[MAXCPU];
103 	struct ktls_alloc_thread alloc_td;
104 };
105 
106 struct ktls_domain_info ktls_domains[MAXMEMDOM];
107 static struct ktls_wq *ktls_wq;
108 static struct proc *ktls_proc;
109 static uma_zone_t ktls_session_zone;
110 static uma_zone_t ktls_buffer_zone;
111 static uint16_t ktls_cpuid_lookup[MAXCPU];
112 static int ktls_init_state;
113 static struct sx ktls_init_lock;
114 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
115 
116 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117     "Kernel TLS offload");
118 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
119     "Kernel TLS offload stats");
120 
121 #ifdef RSS
122 static int ktls_bind_threads = 1;
123 #else
124 static int ktls_bind_threads;
125 #endif
126 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
127     &ktls_bind_threads, 0,
128     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
129 
130 static u_int ktls_maxlen = 16384;
131 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
132     &ktls_maxlen, 0, "Maximum TLS record size");
133 
134 static int ktls_number_threads;
135 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
136     &ktls_number_threads, 0,
137     "Number of TLS threads in thread-pool");
138 
139 unsigned int ktls_ifnet_max_rexmit_pct = 2;
140 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
141     &ktls_ifnet_max_rexmit_pct, 2,
142     "Max percent bytes retransmitted before ifnet TLS is disabled");
143 
144 static bool ktls_offload_enable;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
146     &ktls_offload_enable, 0,
147     "Enable support for kernel TLS offload");
148 
149 static bool ktls_cbc_enable = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
151     &ktls_cbc_enable, 1,
152     "Enable support of AES-CBC crypto for kernel TLS");
153 
154 static bool ktls_sw_buffer_cache = true;
155 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
156     &ktls_sw_buffer_cache, 1,
157     "Enable caching of output buffers for SW encryption");
158 
159 static int ktls_max_alloc = 128;
160 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
161     &ktls_max_alloc, 128,
162     "Max number of 16k buffers to allocate in thread context");
163 
164 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
165 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
166     &ktls_tasks_active, "Number of active tasks");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
170     &ktls_cnt_tx_pending,
171     "Number of TLS 1.0 records waiting for earlier TLS records");
172 
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
175     &ktls_cnt_tx_queued,
176     "Number of TLS records in queue to tasks for SW encryption");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
180     &ktls_cnt_rx_queued,
181     "Number of TLS sockets in queue to tasks for SW decryption");
182 
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
185     CTLFLAG_RD, &ktls_offload_total,
186     "Total successful TLS setups (parameters set)");
187 
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
190     CTLFLAG_RD, &ktls_offload_enable_calls,
191     "Total number of TLS enable calls made");
192 
193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
195     &ktls_offload_active, "Total Active TLS sessions");
196 
197 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
199     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
200 
201 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
203     &ktls_offload_failed_crypto, "Total TLS crypto failures");
204 
205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
207     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
208 
209 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
211     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
212 
213 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
215     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
216 
217 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
218 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
219     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
220 
221 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
222 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
223     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
224 
225 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
226 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
227     &ktls_destroy_task,
228     "Number of times ktls session was destroyed via taskqueue");
229 
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "Software TLS session stats");
232 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
233     "Hardware (ifnet) TLS session stats");
234 #ifdef TCP_OFFLOAD
235 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
236     "TOE TLS session stats");
237 #endif
238 
239 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
240 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
241     "Active number of software TLS sessions using AES-CBC");
242 
243 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
244 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
245     "Active number of software TLS sessions using AES-GCM");
246 
247 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
249     &ktls_sw_chacha20,
250     "Active number of software TLS sessions using Chacha20-Poly1305");
251 
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
254     &ktls_ifnet_cbc,
255     "Active number of ifnet TLS sessions using AES-CBC");
256 
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
259     &ktls_ifnet_gcm,
260     "Active number of ifnet TLS sessions using AES-GCM");
261 
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
264     &ktls_ifnet_chacha20,
265     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
266 
267 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
268 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
269     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
270 
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
273     &ktls_ifnet_reset_dropped,
274     "TLS sessions dropped after failing to update ifnet send tag");
275 
276 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
277 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
278     &ktls_ifnet_reset_failed,
279     "TLS sessions that failed to allocate a new ifnet send tag");
280 
281 static int ktls_ifnet_permitted;
282 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
283     &ktls_ifnet_permitted, 1,
284     "Whether to permit hardware (ifnet) TLS sessions");
285 
286 #ifdef TCP_OFFLOAD
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
289     &ktls_toe_cbc,
290     "Active number of TOE TLS sessions using AES-CBC");
291 
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
294     &ktls_toe_gcm,
295     "Active number of TOE TLS sessions using AES-GCM");
296 
297 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
298 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
299     &ktls_toe_chacha20,
300     "Active number of TOE TLS sessions using Chacha20-Poly1305");
301 #endif
302 
303 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
304 
305 #if defined(INET) || defined(INET6)
306 static void ktls_reset_receive_tag(void *context, int pending);
307 static void ktls_reset_send_tag(void *context, int pending);
308 #endif
309 static void ktls_work_thread(void *ctx);
310 static void ktls_alloc_thread(void *ctx);
311 
312 #if defined(INET) || defined(INET6)
313 static u_int
314 ktls_get_cpu(struct socket *so)
315 {
316 	struct inpcb *inp;
317 #ifdef NUMA
318 	struct ktls_domain_info *di;
319 #endif
320 	u_int cpuid;
321 
322 	inp = sotoinpcb(so);
323 #ifdef RSS
324 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
325 	if (cpuid != NETISR_CPUID_NONE)
326 		return (cpuid);
327 #endif
328 	/*
329 	 * Just use the flowid to shard connections in a repeatable
330 	 * fashion.  Note that TLS 1.0 sessions rely on the
331 	 * serialization provided by having the same connection use
332 	 * the same queue.
333 	 */
334 #ifdef NUMA
335 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
336 		di = &ktls_domains[inp->inp_numa_domain];
337 		cpuid = di->cpu[inp->inp_flowid % di->count];
338 	} else
339 #endif
340 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
341 	return (cpuid);
342 }
343 #endif
344 
345 static int
346 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
347 {
348 	vm_page_t m;
349 	int i, req;
350 
351 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
352 	    ("%s: ktls max length %d is not page size-aligned",
353 	    __func__, ktls_maxlen));
354 
355 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
356 	for (i = 0; i < count; i++) {
357 		m = vm_page_alloc_noobj_contig_domain(domain, req,
358 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
359 		    VM_MEMATTR_DEFAULT);
360 		if (m == NULL)
361 			break;
362 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
363 	}
364 	return (i);
365 }
366 
367 static void
368 ktls_buffer_release(void *arg __unused, void **store, int count)
369 {
370 	vm_page_t m;
371 	int i, j;
372 
373 	for (i = 0; i < count; i++) {
374 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
375 		for (j = 0; j < atop(ktls_maxlen); j++) {
376 			(void)vm_page_unwire_noq(m + j);
377 			vm_page_free(m + j);
378 		}
379 	}
380 }
381 
382 static void
383 ktls_free_mext_contig(struct mbuf *m)
384 {
385 	M_ASSERTEXTPG(m);
386 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
387 }
388 
389 static int
390 ktls_init(void)
391 {
392 	struct thread *td;
393 	struct pcpu *pc;
394 	int count, domain, error, i;
395 
396 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
397 	    M_WAITOK | M_ZERO);
398 
399 	ktls_session_zone = uma_zcreate("ktls_session",
400 	    sizeof(struct ktls_session),
401 	    NULL, NULL, NULL, NULL,
402 	    UMA_ALIGN_CACHE, 0);
403 
404 	if (ktls_sw_buffer_cache) {
405 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
406 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
407 		    ktls_buffer_import, ktls_buffer_release, NULL,
408 		    UMA_ZONE_FIRSTTOUCH);
409 	}
410 
411 	/*
412 	 * Initialize the workqueues to run the TLS work.  We create a
413 	 * work queue for each CPU.
414 	 */
415 	CPU_FOREACH(i) {
416 		STAILQ_INIT(&ktls_wq[i].m_head);
417 		STAILQ_INIT(&ktls_wq[i].so_head);
418 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
419 		if (ktls_bind_threads > 1) {
420 			pc = pcpu_find(i);
421 			domain = pc->pc_domain;
422 			count = ktls_domains[domain].count;
423 			ktls_domains[domain].cpu[count] = i;
424 			ktls_domains[domain].count++;
425 		}
426 		ktls_cpuid_lookup[ktls_number_threads] = i;
427 		ktls_number_threads++;
428 	}
429 
430 	/*
431 	 * If we somehow have an empty domain, fall back to choosing
432 	 * among all KTLS threads.
433 	 */
434 	if (ktls_bind_threads > 1) {
435 		for (i = 0; i < vm_ndomains; i++) {
436 			if (ktls_domains[i].count == 0) {
437 				ktls_bind_threads = 1;
438 				break;
439 			}
440 		}
441 	}
442 
443 	/* Start kthreads for each workqueue. */
444 	CPU_FOREACH(i) {
445 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
446 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
447 		if (error) {
448 			printf("Can't add KTLS thread %d error %d\n", i, error);
449 			return (error);
450 		}
451 	}
452 
453 	/*
454 	 * Start an allocation thread per-domain to perform blocking allocations
455 	 * of 16k physically contiguous TLS crypto destination buffers.
456 	 */
457 	if (ktls_sw_buffer_cache) {
458 		for (domain = 0; domain < vm_ndomains; domain++) {
459 			if (VM_DOMAIN_EMPTY(domain))
460 				continue;
461 			if (CPU_EMPTY(&cpuset_domain[domain]))
462 				continue;
463 			error = kproc_kthread_add(ktls_alloc_thread,
464 			    &ktls_domains[domain], &ktls_proc,
465 			    &ktls_domains[domain].alloc_td.td,
466 			    0, 0, "KTLS", "alloc_%d", domain);
467 			if (error) {
468 				printf("Can't add KTLS alloc thread %d error %d\n",
469 				    domain, error);
470 				return (error);
471 			}
472 		}
473 	}
474 
475 	if (bootverbose)
476 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
477 	return (0);
478 }
479 
480 static int
481 ktls_start_kthreads(void)
482 {
483 	int error, state;
484 
485 start:
486 	state = atomic_load_acq_int(&ktls_init_state);
487 	if (__predict_true(state > 0))
488 		return (0);
489 	if (state < 0)
490 		return (ENXIO);
491 
492 	sx_xlock(&ktls_init_lock);
493 	if (ktls_init_state != 0) {
494 		sx_xunlock(&ktls_init_lock);
495 		goto start;
496 	}
497 
498 	error = ktls_init();
499 	if (error == 0)
500 		state = 1;
501 	else
502 		state = -1;
503 	atomic_store_rel_int(&ktls_init_state, state);
504 	sx_xunlock(&ktls_init_lock);
505 	return (error);
506 }
507 
508 #if defined(INET) || defined(INET6)
509 static int
510 ktls_create_session(struct socket *so, struct tls_enable *en,
511     struct ktls_session **tlsp, int direction)
512 {
513 	struct ktls_session *tls;
514 	int error;
515 
516 	/* Only TLS 1.0 - 1.3 are supported. */
517 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
518 		return (EINVAL);
519 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
520 	    en->tls_vminor > TLS_MINOR_VER_THREE)
521 		return (EINVAL);
522 
523 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
524 		return (EINVAL);
525 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
526 		return (EINVAL);
527 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
528 		return (EINVAL);
529 
530 	/* All supported algorithms require a cipher key. */
531 	if (en->cipher_key_len == 0)
532 		return (EINVAL);
533 
534 	/* No flags are currently supported. */
535 	if (en->flags != 0)
536 		return (EINVAL);
537 
538 	/* Common checks for supported algorithms. */
539 	switch (en->cipher_algorithm) {
540 	case CRYPTO_AES_NIST_GCM_16:
541 		/*
542 		 * auth_algorithm isn't used, but permit GMAC values
543 		 * for compatibility.
544 		 */
545 		switch (en->auth_algorithm) {
546 		case 0:
547 #ifdef COMPAT_FREEBSD12
548 		/* XXX: Really 13.0-current COMPAT. */
549 		case CRYPTO_AES_128_NIST_GMAC:
550 		case CRYPTO_AES_192_NIST_GMAC:
551 		case CRYPTO_AES_256_NIST_GMAC:
552 #endif
553 			break;
554 		default:
555 			return (EINVAL);
556 		}
557 		if (en->auth_key_len != 0)
558 			return (EINVAL);
559 		switch (en->tls_vminor) {
560 		case TLS_MINOR_VER_TWO:
561 			if (en->iv_len != TLS_AEAD_GCM_LEN)
562 				return (EINVAL);
563 			break;
564 		case TLS_MINOR_VER_THREE:
565 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
566 				return (EINVAL);
567 			break;
568 		default:
569 			return (EINVAL);
570 		}
571 		break;
572 	case CRYPTO_AES_CBC:
573 		switch (en->auth_algorithm) {
574 		case CRYPTO_SHA1_HMAC:
575 			break;
576 		case CRYPTO_SHA2_256_HMAC:
577 		case CRYPTO_SHA2_384_HMAC:
578 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
579 				return (EINVAL);
580 			break;
581 		default:
582 			return (EINVAL);
583 		}
584 		if (en->auth_key_len == 0)
585 			return (EINVAL);
586 
587 		/*
588 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
589 		 * use explicit IVs.
590 		 */
591 		switch (en->tls_vminor) {
592 		case TLS_MINOR_VER_ZERO:
593 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
594 				return (EINVAL);
595 			break;
596 		case TLS_MINOR_VER_ONE:
597 		case TLS_MINOR_VER_TWO:
598 			/* Ignore any supplied IV. */
599 			en->iv_len = 0;
600 			break;
601 		default:
602 			return (EINVAL);
603 		}
604 		break;
605 	case CRYPTO_CHACHA20_POLY1305:
606 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
607 			return (EINVAL);
608 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
609 		    en->tls_vminor != TLS_MINOR_VER_THREE)
610 			return (EINVAL);
611 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
612 			return (EINVAL);
613 		break;
614 	default:
615 		return (EINVAL);
616 	}
617 
618 	error = ktls_start_kthreads();
619 	if (error != 0)
620 		return (error);
621 
622 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
623 
624 	counter_u64_add(ktls_offload_active, 1);
625 
626 	refcount_init(&tls->refcount, 1);
627 	if (direction == KTLS_RX) {
628 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
629 	} else {
630 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
631 		tls->inp = so->so_pcb;
632 		in_pcbref(tls->inp);
633 		tls->tx = true;
634 	}
635 
636 	tls->wq_index = ktls_get_cpu(so);
637 
638 	tls->params.cipher_algorithm = en->cipher_algorithm;
639 	tls->params.auth_algorithm = en->auth_algorithm;
640 	tls->params.tls_vmajor = en->tls_vmajor;
641 	tls->params.tls_vminor = en->tls_vminor;
642 	tls->params.flags = en->flags;
643 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
644 
645 	/* Set the header and trailer lengths. */
646 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
647 	switch (en->cipher_algorithm) {
648 	case CRYPTO_AES_NIST_GCM_16:
649 		/*
650 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
651 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
652 		 */
653 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
654 			tls->params.tls_hlen += sizeof(uint64_t);
655 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
656 		tls->params.tls_bs = 1;
657 		break;
658 	case CRYPTO_AES_CBC:
659 		switch (en->auth_algorithm) {
660 		case CRYPTO_SHA1_HMAC:
661 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
662 				/* Implicit IV, no nonce. */
663 				tls->sequential_records = true;
664 				tls->next_seqno = be64dec(en->rec_seq);
665 				STAILQ_INIT(&tls->pending_records);
666 			} else {
667 				tls->params.tls_hlen += AES_BLOCK_LEN;
668 			}
669 			tls->params.tls_tlen = AES_BLOCK_LEN +
670 			    SHA1_HASH_LEN;
671 			break;
672 		case CRYPTO_SHA2_256_HMAC:
673 			tls->params.tls_hlen += AES_BLOCK_LEN;
674 			tls->params.tls_tlen = AES_BLOCK_LEN +
675 			    SHA2_256_HASH_LEN;
676 			break;
677 		case CRYPTO_SHA2_384_HMAC:
678 			tls->params.tls_hlen += AES_BLOCK_LEN;
679 			tls->params.tls_tlen = AES_BLOCK_LEN +
680 			    SHA2_384_HASH_LEN;
681 			break;
682 		default:
683 			panic("invalid hmac");
684 		}
685 		tls->params.tls_bs = AES_BLOCK_LEN;
686 		break;
687 	case CRYPTO_CHACHA20_POLY1305:
688 		/*
689 		 * Chacha20 uses a 12 byte implicit IV.
690 		 */
691 		tls->params.tls_tlen = POLY1305_HASH_LEN;
692 		tls->params.tls_bs = 1;
693 		break;
694 	default:
695 		panic("invalid cipher");
696 	}
697 
698 	/*
699 	 * TLS 1.3 includes optional padding which we do not support,
700 	 * and also puts the "real" record type at the end of the
701 	 * encrypted data.
702 	 */
703 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
704 		tls->params.tls_tlen += sizeof(uint8_t);
705 
706 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
707 	    ("TLS header length too long: %d", tls->params.tls_hlen));
708 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
709 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
710 
711 	if (en->auth_key_len != 0) {
712 		tls->params.auth_key_len = en->auth_key_len;
713 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
714 		    M_WAITOK);
715 		error = copyin(en->auth_key, tls->params.auth_key,
716 		    en->auth_key_len);
717 		if (error)
718 			goto out;
719 	}
720 
721 	tls->params.cipher_key_len = en->cipher_key_len;
722 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
723 	error = copyin(en->cipher_key, tls->params.cipher_key,
724 	    en->cipher_key_len);
725 	if (error)
726 		goto out;
727 
728 	/*
729 	 * This holds the implicit portion of the nonce for AEAD
730 	 * ciphers and the initial implicit IV for TLS 1.0.  The
731 	 * explicit portions of the IV are generated in ktls_frame().
732 	 */
733 	if (en->iv_len != 0) {
734 		tls->params.iv_len = en->iv_len;
735 		error = copyin(en->iv, tls->params.iv, en->iv_len);
736 		if (error)
737 			goto out;
738 
739 		/*
740 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
741 		 * counter to generate unique explicit IVs.
742 		 *
743 		 * Store this counter in the last 8 bytes of the IV
744 		 * array so that it is 8-byte aligned.
745 		 */
746 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
747 		    en->tls_vminor == TLS_MINOR_VER_TWO)
748 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
749 	}
750 
751 	*tlsp = tls;
752 	return (0);
753 
754 out:
755 	ktls_free(tls);
756 	return (error);
757 }
758 
759 static struct ktls_session *
760 ktls_clone_session(struct ktls_session *tls, int direction)
761 {
762 	struct ktls_session *tls_new;
763 
764 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
765 
766 	counter_u64_add(ktls_offload_active, 1);
767 
768 	refcount_init(&tls_new->refcount, 1);
769 	if (direction == KTLS_RX) {
770 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
771 		    tls_new);
772 	} else {
773 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
774 		    tls_new);
775 		tls_new->inp = tls->inp;
776 		tls_new->tx = true;
777 		in_pcbref(tls_new->inp);
778 	}
779 
780 	/* Copy fields from existing session. */
781 	tls_new->params = tls->params;
782 	tls_new->wq_index = tls->wq_index;
783 
784 	/* Deep copy keys. */
785 	if (tls_new->params.auth_key != NULL) {
786 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
787 		    M_KTLS, M_WAITOK);
788 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
789 		    tls->params.auth_key_len);
790 	}
791 
792 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
793 	    M_WAITOK);
794 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
795 	    tls->params.cipher_key_len);
796 
797 	return (tls_new);
798 }
799 
800 #ifdef TCP_OFFLOAD
801 static int
802 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
803 {
804 	struct inpcb *inp;
805 	struct tcpcb *tp;
806 	int error;
807 
808 	inp = so->so_pcb;
809 	INP_WLOCK(inp);
810 	if (inp->inp_flags & INP_DROPPED) {
811 		INP_WUNLOCK(inp);
812 		return (ECONNRESET);
813 	}
814 	if (inp->inp_socket == NULL) {
815 		INP_WUNLOCK(inp);
816 		return (ECONNRESET);
817 	}
818 	tp = intotcpcb(inp);
819 	if (!(tp->t_flags & TF_TOE)) {
820 		INP_WUNLOCK(inp);
821 		return (EOPNOTSUPP);
822 	}
823 
824 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
825 	INP_WUNLOCK(inp);
826 	if (error == 0) {
827 		tls->mode = TCP_TLS_MODE_TOE;
828 		switch (tls->params.cipher_algorithm) {
829 		case CRYPTO_AES_CBC:
830 			counter_u64_add(ktls_toe_cbc, 1);
831 			break;
832 		case CRYPTO_AES_NIST_GCM_16:
833 			counter_u64_add(ktls_toe_gcm, 1);
834 			break;
835 		case CRYPTO_CHACHA20_POLY1305:
836 			counter_u64_add(ktls_toe_chacha20, 1);
837 			break;
838 		}
839 	}
840 	return (error);
841 }
842 #endif
843 
844 /*
845  * Common code used when first enabling ifnet TLS on a connection or
846  * when allocating a new ifnet TLS session due to a routing change.
847  * This function allocates a new TLS send tag on whatever interface
848  * the connection is currently routed over.
849  */
850 static int
851 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
852     struct m_snd_tag **mstp)
853 {
854 	union if_snd_tag_alloc_params params;
855 	struct ifnet *ifp;
856 	struct nhop_object *nh;
857 	struct tcpcb *tp;
858 	int error;
859 
860 	INP_RLOCK(inp);
861 	if (inp->inp_flags & INP_DROPPED) {
862 		INP_RUNLOCK(inp);
863 		return (ECONNRESET);
864 	}
865 	if (inp->inp_socket == NULL) {
866 		INP_RUNLOCK(inp);
867 		return (ECONNRESET);
868 	}
869 	tp = intotcpcb(inp);
870 
871 	/*
872 	 * Check administrative controls on ifnet TLS to determine if
873 	 * ifnet TLS should be denied.
874 	 *
875 	 * - Always permit 'force' requests.
876 	 * - ktls_ifnet_permitted == 0: always deny.
877 	 */
878 	if (!force && ktls_ifnet_permitted == 0) {
879 		INP_RUNLOCK(inp);
880 		return (ENXIO);
881 	}
882 
883 	/*
884 	 * XXX: Use the cached route in the inpcb to find the
885 	 * interface.  This should perhaps instead use
886 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
887 	 * enabled after a connection has completed key negotiation in
888 	 * userland, the cached route will be present in practice.
889 	 */
890 	nh = inp->inp_route.ro_nh;
891 	if (nh == NULL) {
892 		INP_RUNLOCK(inp);
893 		return (ENXIO);
894 	}
895 	ifp = nh->nh_ifp;
896 	if_ref(ifp);
897 
898 	/*
899 	 * Allocate a TLS + ratelimit tag if the connection has an
900 	 * existing pacing rate.
901 	 */
902 	if (tp->t_pacing_rate != -1 &&
903 	    (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
904 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
905 		params.tls_rate_limit.inp = inp;
906 		params.tls_rate_limit.tls = tls;
907 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
908 	} else {
909 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
910 		params.tls.inp = inp;
911 		params.tls.tls = tls;
912 	}
913 	params.hdr.flowid = inp->inp_flowid;
914 	params.hdr.flowtype = inp->inp_flowtype;
915 	params.hdr.numa_domain = inp->inp_numa_domain;
916 	INP_RUNLOCK(inp);
917 
918 	if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
919 		error = EOPNOTSUPP;
920 		goto out;
921 	}
922 	if (inp->inp_vflag & INP_IPV6) {
923 		if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
924 			error = EOPNOTSUPP;
925 			goto out;
926 		}
927 	} else {
928 		if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
929 			error = EOPNOTSUPP;
930 			goto out;
931 		}
932 	}
933 	error = m_snd_tag_alloc(ifp, &params, mstp);
934 out:
935 	if_rele(ifp);
936 	return (error);
937 }
938 
939 /*
940  * Allocate an initial TLS receive tag for doing HW decryption of TLS
941  * data.
942  *
943  * This function allocates a new TLS receive tag on whatever interface
944  * the connection is currently routed over.  If the connection ends up
945  * using a different interface for receive this will get fixed up via
946  * ktls_input_ifp_mismatch as future packets arrive.
947  */
948 static int
949 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
950     struct m_snd_tag **mstp)
951 {
952 	union if_snd_tag_alloc_params params;
953 	struct ifnet *ifp;
954 	struct nhop_object *nh;
955 	int error;
956 
957 	if (!ktls_ocf_recrypt_supported(tls))
958 		return (ENXIO);
959 
960 	INP_RLOCK(inp);
961 	if (inp->inp_flags & INP_DROPPED) {
962 		INP_RUNLOCK(inp);
963 		return (ECONNRESET);
964 	}
965 	if (inp->inp_socket == NULL) {
966 		INP_RUNLOCK(inp);
967 		return (ECONNRESET);
968 	}
969 
970 	/*
971 	 * Check administrative controls on ifnet TLS to determine if
972 	 * ifnet TLS should be denied.
973 	 */
974 	if (ktls_ifnet_permitted == 0) {
975 		INP_RUNLOCK(inp);
976 		return (ENXIO);
977 	}
978 
979 	/*
980 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
981 	 * the inpcb to find the interface.
982 	 */
983 	nh = inp->inp_route.ro_nh;
984 	if (nh == NULL) {
985 		INP_RUNLOCK(inp);
986 		return (ENXIO);
987 	}
988 	ifp = nh->nh_ifp;
989 	if_ref(ifp);
990 	tls->rx_ifp = ifp;
991 
992 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
993 	params.hdr.flowid = inp->inp_flowid;
994 	params.hdr.flowtype = inp->inp_flowtype;
995 	params.hdr.numa_domain = inp->inp_numa_domain;
996 	params.tls_rx.inp = inp;
997 	params.tls_rx.tls = tls;
998 	params.tls_rx.vlan_id = 0;
999 
1000 	INP_RUNLOCK(inp);
1001 
1002 	if (inp->inp_vflag & INP_IPV6) {
1003 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0) {
1004 			error = EOPNOTSUPP;
1005 			goto out;
1006 		}
1007 	} else {
1008 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0) {
1009 			error = EOPNOTSUPP;
1010 			goto out;
1011 		}
1012 	}
1013 	error = m_snd_tag_alloc(ifp, &params, mstp);
1014 
1015 	/*
1016 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1017 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1018 	 * ID for use in ktls_reset_receive_tag which allocates new
1019 	 * receive tags directly from the leaf interface bypassing
1020 	 * if_vlan.
1021 	 */
1022 	if (error == 0)
1023 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1024 out:
1025 	return (error);
1026 }
1027 
1028 static int
1029 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1030     bool force)
1031 {
1032 	struct m_snd_tag *mst;
1033 	int error;
1034 
1035 	switch (direction) {
1036 	case KTLS_TX:
1037 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1038 		if (__predict_false(error != 0))
1039 			goto done;
1040 		break;
1041 	case KTLS_RX:
1042 		KASSERT(!force, ("%s: forced receive tag", __func__));
1043 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1044 		if (__predict_false(error != 0))
1045 			goto done;
1046 		break;
1047 	default:
1048 		__assert_unreachable();
1049 	}
1050 
1051 	tls->mode = TCP_TLS_MODE_IFNET;
1052 	tls->snd_tag = mst;
1053 
1054 	switch (tls->params.cipher_algorithm) {
1055 	case CRYPTO_AES_CBC:
1056 		counter_u64_add(ktls_ifnet_cbc, 1);
1057 		break;
1058 	case CRYPTO_AES_NIST_GCM_16:
1059 		counter_u64_add(ktls_ifnet_gcm, 1);
1060 		break;
1061 	case CRYPTO_CHACHA20_POLY1305:
1062 		counter_u64_add(ktls_ifnet_chacha20, 1);
1063 		break;
1064 	default:
1065 		break;
1066 	}
1067 done:
1068 	return (error);
1069 }
1070 
1071 static void
1072 ktls_use_sw(struct ktls_session *tls)
1073 {
1074 	tls->mode = TCP_TLS_MODE_SW;
1075 	switch (tls->params.cipher_algorithm) {
1076 	case CRYPTO_AES_CBC:
1077 		counter_u64_add(ktls_sw_cbc, 1);
1078 		break;
1079 	case CRYPTO_AES_NIST_GCM_16:
1080 		counter_u64_add(ktls_sw_gcm, 1);
1081 		break;
1082 	case CRYPTO_CHACHA20_POLY1305:
1083 		counter_u64_add(ktls_sw_chacha20, 1);
1084 		break;
1085 	}
1086 }
1087 
1088 static int
1089 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1090 {
1091 	int error;
1092 
1093 	error = ktls_ocf_try(so, tls, direction);
1094 	if (error)
1095 		return (error);
1096 	ktls_use_sw(tls);
1097 	return (0);
1098 }
1099 
1100 /*
1101  * KTLS RX stores data in the socket buffer as a list of TLS records,
1102  * where each record is stored as a control message containg the TLS
1103  * header followed by data mbufs containing the decrypted data.  This
1104  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1105  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1106  * should be queued to the socket buffer as records, but encrypted
1107  * data which needs to be decrypted by software arrives as a stream of
1108  * regular mbufs which need to be converted.  In addition, there may
1109  * already be pending encrypted data in the socket buffer when KTLS RX
1110  * is enabled.
1111  *
1112  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1113  * is used:
1114  *
1115  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1116  *
1117  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1118  *   from the first mbuf.  Once all of the data for that TLS record is
1119  *   queued, the socket is queued to a worker thread.
1120  *
1121  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1122  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1123  *   decrypted, and inserted into the regular socket buffer chain as
1124  *   record starting with a control message holding the TLS header and
1125  *   a chain of mbufs holding the encrypted data.
1126  */
1127 
1128 static void
1129 sb_mark_notready(struct sockbuf *sb)
1130 {
1131 	struct mbuf *m;
1132 
1133 	m = sb->sb_mb;
1134 	sb->sb_mtls = m;
1135 	sb->sb_mb = NULL;
1136 	sb->sb_mbtail = NULL;
1137 	sb->sb_lastrecord = NULL;
1138 	for (; m != NULL; m = m->m_next) {
1139 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1140 		    __func__));
1141 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1142 		    __func__));
1143 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1144 		    __func__));
1145 		m->m_flags |= M_NOTREADY;
1146 		sb->sb_acc -= m->m_len;
1147 		sb->sb_tlscc += m->m_len;
1148 		sb->sb_mtlstail = m;
1149 	}
1150 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1151 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1152 	    sb->sb_ccc));
1153 }
1154 
1155 /*
1156  * Return information about the pending TLS data in a socket
1157  * buffer.  On return, 'seqno' is set to the sequence number
1158  * of the next TLS record to be received, 'resid' is set to
1159  * the amount of bytes still needed for the last pending
1160  * record.  The function returns 'false' if the last pending
1161  * record contains a partial TLS header.  In that case, 'resid'
1162  * is the number of bytes needed to complete the TLS header.
1163  */
1164 bool
1165 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1166 {
1167 	struct tls_record_layer hdr;
1168 	struct mbuf *m;
1169 	uint64_t seqno;
1170 	size_t resid;
1171 	u_int offset, record_len;
1172 
1173 	SOCKBUF_LOCK_ASSERT(sb);
1174 	MPASS(sb->sb_flags & SB_TLS_RX);
1175 	seqno = sb->sb_tls_seqno;
1176 	resid = sb->sb_tlscc;
1177 	m = sb->sb_mtls;
1178 	offset = 0;
1179 
1180 	if (resid == 0) {
1181 		*seqnop = seqno;
1182 		*residp = 0;
1183 		return (true);
1184 	}
1185 
1186 	for (;;) {
1187 		seqno++;
1188 
1189 		if (resid < sizeof(hdr)) {
1190 			*seqnop = seqno;
1191 			*residp = sizeof(hdr) - resid;
1192 			return (false);
1193 		}
1194 
1195 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1196 
1197 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1198 		if (resid <= record_len) {
1199 			*seqnop = seqno;
1200 			*residp = record_len - resid;
1201 			return (true);
1202 		}
1203 		resid -= record_len;
1204 
1205 		while (record_len != 0) {
1206 			if (m->m_len - offset > record_len) {
1207 				offset += record_len;
1208 				break;
1209 			}
1210 
1211 			record_len -= (m->m_len - offset);
1212 			offset = 0;
1213 			m = m->m_next;
1214 		}
1215 	}
1216 }
1217 
1218 int
1219 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1220 {
1221 	struct ktls_session *tls;
1222 	int error;
1223 
1224 	if (!ktls_offload_enable)
1225 		return (ENOTSUP);
1226 
1227 	counter_u64_add(ktls_offload_enable_calls, 1);
1228 
1229 	/*
1230 	 * This should always be true since only the TCP socket option
1231 	 * invokes this function.
1232 	 */
1233 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1234 		return (EINVAL);
1235 
1236 	/*
1237 	 * XXX: Don't overwrite existing sessions.  We should permit
1238 	 * this to support rekeying in the future.
1239 	 */
1240 	if (so->so_rcv.sb_tls_info != NULL)
1241 		return (EALREADY);
1242 
1243 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1244 		return (ENOTSUP);
1245 
1246 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1247 	if (error)
1248 		return (error);
1249 
1250 	error = ktls_ocf_try(so, tls, KTLS_RX);
1251 	if (error) {
1252 		ktls_free(tls);
1253 		return (error);
1254 	}
1255 
1256 	/* Mark the socket as using TLS offload. */
1257 	SOCK_RECVBUF_LOCK(so);
1258 	if (SOLISTENING(so)) {
1259 		SOCK_RECVBUF_UNLOCK(so);
1260 		ktls_free(tls);
1261 		return (EINVAL);
1262 	}
1263 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1264 	so->so_rcv.sb_tls_info = tls;
1265 	so->so_rcv.sb_flags |= SB_TLS_RX;
1266 
1267 	/* Mark existing data as not ready until it can be decrypted. */
1268 	sb_mark_notready(&so->so_rcv);
1269 	ktls_check_rx(&so->so_rcv);
1270 	SOCK_RECVBUF_UNLOCK(so);
1271 
1272 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1273 #ifdef TCP_OFFLOAD
1274 	error = ktls_try_toe(so, tls, KTLS_RX);
1275 	if (error)
1276 #endif
1277 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1278 	if (error)
1279 		ktls_use_sw(tls);
1280 
1281 	counter_u64_add(ktls_offload_total, 1);
1282 
1283 	return (0);
1284 }
1285 
1286 int
1287 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1288 {
1289 	struct ktls_session *tls;
1290 	struct inpcb *inp;
1291 	struct tcpcb *tp;
1292 	int error;
1293 
1294 	if (!ktls_offload_enable)
1295 		return (ENOTSUP);
1296 
1297 	counter_u64_add(ktls_offload_enable_calls, 1);
1298 
1299 	/*
1300 	 * This should always be true since only the TCP socket option
1301 	 * invokes this function.
1302 	 */
1303 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1304 		return (EINVAL);
1305 
1306 	/*
1307 	 * XXX: Don't overwrite existing sessions.  We should permit
1308 	 * this to support rekeying in the future.
1309 	 */
1310 	if (so->so_snd.sb_tls_info != NULL)
1311 		return (EALREADY);
1312 
1313 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1314 		return (ENOTSUP);
1315 
1316 	/* TLS requires ext pgs */
1317 	if (mb_use_ext_pgs == 0)
1318 		return (ENXIO);
1319 
1320 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1321 	if (error)
1322 		return (error);
1323 
1324 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1325 #ifdef TCP_OFFLOAD
1326 	error = ktls_try_toe(so, tls, KTLS_TX);
1327 	if (error)
1328 #endif
1329 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1330 	if (error)
1331 		error = ktls_try_sw(so, tls, KTLS_TX);
1332 
1333 	if (error) {
1334 		ktls_free(tls);
1335 		return (error);
1336 	}
1337 
1338 	/*
1339 	 * Serialize with sosend_generic() and make sure that we're not
1340 	 * operating on a listening socket.
1341 	 */
1342 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1343 	if (error) {
1344 		ktls_free(tls);
1345 		return (error);
1346 	}
1347 
1348 	/*
1349 	 * Write lock the INP when setting sb_tls_info so that
1350 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1351 	 * holding the INP lock.
1352 	 */
1353 	inp = so->so_pcb;
1354 	INP_WLOCK(inp);
1355 	SOCK_SENDBUF_LOCK(so);
1356 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1357 	so->so_snd.sb_tls_info = tls;
1358 	if (tls->mode != TCP_TLS_MODE_SW) {
1359 		tp = intotcpcb(inp);
1360 		MPASS(tp->t_nic_ktls_xmit == 0);
1361 		tp->t_nic_ktls_xmit = 1;
1362 		if (tp->t_fb->tfb_hwtls_change != NULL)
1363 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1364 	}
1365 	SOCK_SENDBUF_UNLOCK(so);
1366 	INP_WUNLOCK(inp);
1367 	SOCK_IO_SEND_UNLOCK(so);
1368 
1369 	counter_u64_add(ktls_offload_total, 1);
1370 
1371 	return (0);
1372 }
1373 
1374 int
1375 ktls_get_rx_mode(struct socket *so, int *modep)
1376 {
1377 	struct ktls_session *tls;
1378 	struct inpcb *inp __diagused;
1379 
1380 	if (SOLISTENING(so))
1381 		return (EINVAL);
1382 	inp = so->so_pcb;
1383 	INP_WLOCK_ASSERT(inp);
1384 	SOCK_RECVBUF_LOCK(so);
1385 	tls = so->so_rcv.sb_tls_info;
1386 	if (tls == NULL)
1387 		*modep = TCP_TLS_MODE_NONE;
1388 	else
1389 		*modep = tls->mode;
1390 	SOCK_RECVBUF_UNLOCK(so);
1391 	return (0);
1392 }
1393 
1394 /*
1395  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1396  *
1397  * This function gets information about the next TCP- and TLS-
1398  * sequence number to be processed by the TLS receive worker
1399  * thread. The information is extracted from the given "inpcb"
1400  * structure. The values are stored in host endian format at the two
1401  * given output pointer locations. The TCP sequence number points to
1402  * the beginning of the TLS header.
1403  *
1404  * This function returns zero on success, else a non-zero error code
1405  * is returned.
1406  */
1407 int
1408 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1409 {
1410 	struct socket *so;
1411 	struct tcpcb *tp;
1412 
1413 	INP_RLOCK(inp);
1414 	so = inp->inp_socket;
1415 	if (__predict_false(so == NULL)) {
1416 		INP_RUNLOCK(inp);
1417 		return (EINVAL);
1418 	}
1419 	if (inp->inp_flags & INP_DROPPED) {
1420 		INP_RUNLOCK(inp);
1421 		return (ECONNRESET);
1422 	}
1423 
1424 	tp = intotcpcb(inp);
1425 	MPASS(tp != NULL);
1426 
1427 	SOCKBUF_LOCK(&so->so_rcv);
1428 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1429 	*tlsseq = so->so_rcv.sb_tls_seqno;
1430 	SOCKBUF_UNLOCK(&so->so_rcv);
1431 
1432 	INP_RUNLOCK(inp);
1433 
1434 	return (0);
1435 }
1436 
1437 int
1438 ktls_get_tx_mode(struct socket *so, int *modep)
1439 {
1440 	struct ktls_session *tls;
1441 	struct inpcb *inp __diagused;
1442 
1443 	if (SOLISTENING(so))
1444 		return (EINVAL);
1445 	inp = so->so_pcb;
1446 	INP_WLOCK_ASSERT(inp);
1447 	SOCK_SENDBUF_LOCK(so);
1448 	tls = so->so_snd.sb_tls_info;
1449 	if (tls == NULL)
1450 		*modep = TCP_TLS_MODE_NONE;
1451 	else
1452 		*modep = tls->mode;
1453 	SOCK_SENDBUF_UNLOCK(so);
1454 	return (0);
1455 }
1456 
1457 /*
1458  * Switch between SW and ifnet TLS sessions as requested.
1459  */
1460 int
1461 ktls_set_tx_mode(struct socket *so, int mode)
1462 {
1463 	struct ktls_session *tls, *tls_new;
1464 	struct inpcb *inp;
1465 	struct tcpcb *tp;
1466 	int error;
1467 
1468 	if (SOLISTENING(so))
1469 		return (EINVAL);
1470 	switch (mode) {
1471 	case TCP_TLS_MODE_SW:
1472 	case TCP_TLS_MODE_IFNET:
1473 		break;
1474 	default:
1475 		return (EINVAL);
1476 	}
1477 
1478 	inp = so->so_pcb;
1479 	INP_WLOCK_ASSERT(inp);
1480 	tp = intotcpcb(inp);
1481 
1482 	if (mode == TCP_TLS_MODE_IFNET) {
1483 		/* Don't allow enabling ifnet ktls multiple times */
1484 		if (tp->t_nic_ktls_xmit)
1485 			return (EALREADY);
1486 
1487 		/*
1488 		 * Don't enable ifnet ktls if we disabled it due to an
1489 		 * excessive retransmission rate
1490 		 */
1491 		if (tp->t_nic_ktls_xmit_dis)
1492 			return (ENXIO);
1493 	}
1494 
1495 	SOCKBUF_LOCK(&so->so_snd);
1496 	tls = so->so_snd.sb_tls_info;
1497 	if (tls == NULL) {
1498 		SOCKBUF_UNLOCK(&so->so_snd);
1499 		return (0);
1500 	}
1501 
1502 	if (tls->mode == mode) {
1503 		SOCKBUF_UNLOCK(&so->so_snd);
1504 		return (0);
1505 	}
1506 
1507 	tls = ktls_hold(tls);
1508 	SOCKBUF_UNLOCK(&so->so_snd);
1509 	INP_WUNLOCK(inp);
1510 
1511 	tls_new = ktls_clone_session(tls, KTLS_TX);
1512 
1513 	if (mode == TCP_TLS_MODE_IFNET)
1514 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1515 	else
1516 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1517 	if (error) {
1518 		counter_u64_add(ktls_switch_failed, 1);
1519 		ktls_free(tls_new);
1520 		ktls_free(tls);
1521 		INP_WLOCK(inp);
1522 		return (error);
1523 	}
1524 
1525 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1526 	if (error) {
1527 		counter_u64_add(ktls_switch_failed, 1);
1528 		ktls_free(tls_new);
1529 		ktls_free(tls);
1530 		INP_WLOCK(inp);
1531 		return (error);
1532 	}
1533 
1534 	/*
1535 	 * If we raced with another session change, keep the existing
1536 	 * session.
1537 	 */
1538 	if (tls != so->so_snd.sb_tls_info) {
1539 		counter_u64_add(ktls_switch_failed, 1);
1540 		SOCK_IO_SEND_UNLOCK(so);
1541 		ktls_free(tls_new);
1542 		ktls_free(tls);
1543 		INP_WLOCK(inp);
1544 		return (EBUSY);
1545 	}
1546 
1547 	INP_WLOCK(inp);
1548 	SOCKBUF_LOCK(&so->so_snd);
1549 	so->so_snd.sb_tls_info = tls_new;
1550 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1551 		MPASS(tp->t_nic_ktls_xmit == 0);
1552 		tp->t_nic_ktls_xmit = 1;
1553 		if (tp->t_fb->tfb_hwtls_change != NULL)
1554 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1555 	}
1556 	SOCKBUF_UNLOCK(&so->so_snd);
1557 	SOCK_IO_SEND_UNLOCK(so);
1558 
1559 	/*
1560 	 * Drop two references on 'tls'.  The first is for the
1561 	 * ktls_hold() above.  The second drops the reference from the
1562 	 * socket buffer.
1563 	 */
1564 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1565 	ktls_free(tls);
1566 	ktls_free(tls);
1567 
1568 	if (mode == TCP_TLS_MODE_IFNET)
1569 		counter_u64_add(ktls_switch_to_ifnet, 1);
1570 	else
1571 		counter_u64_add(ktls_switch_to_sw, 1);
1572 
1573 	return (0);
1574 }
1575 
1576 /*
1577  * Try to allocate a new TLS receive tag.  This task is scheduled when
1578  * sbappend_ktls_rx detects an input path change.  If a new tag is
1579  * allocated, replace the tag in the TLS session.  If a new tag cannot
1580  * be allocated, let the session fall back to software decryption.
1581  */
1582 static void
1583 ktls_reset_receive_tag(void *context, int pending)
1584 {
1585 	union if_snd_tag_alloc_params params;
1586 	struct ktls_session *tls;
1587 	struct m_snd_tag *mst;
1588 	struct inpcb *inp;
1589 	struct ifnet *ifp;
1590 	struct socket *so;
1591 	int error;
1592 
1593 	MPASS(pending == 1);
1594 
1595 	tls = context;
1596 	so = tls->so;
1597 	inp = so->so_pcb;
1598 	ifp = NULL;
1599 
1600 	INP_RLOCK(inp);
1601 	if (inp->inp_flags & INP_DROPPED) {
1602 		INP_RUNLOCK(inp);
1603 		goto out;
1604 	}
1605 
1606 	SOCKBUF_LOCK(&so->so_rcv);
1607 	mst = tls->snd_tag;
1608 	tls->snd_tag = NULL;
1609 	if (mst != NULL)
1610 		m_snd_tag_rele(mst);
1611 
1612 	ifp = tls->rx_ifp;
1613 	if_ref(ifp);
1614 	SOCKBUF_UNLOCK(&so->so_rcv);
1615 
1616 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1617 	params.hdr.flowid = inp->inp_flowid;
1618 	params.hdr.flowtype = inp->inp_flowtype;
1619 	params.hdr.numa_domain = inp->inp_numa_domain;
1620 	params.tls_rx.inp = inp;
1621 	params.tls_rx.tls = tls;
1622 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1623 	INP_RUNLOCK(inp);
1624 
1625 	if (inp->inp_vflag & INP_IPV6) {
1626 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1627 			goto out;
1628 	} else {
1629 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1630 			goto out;
1631 	}
1632 
1633 	error = m_snd_tag_alloc(ifp, &params, &mst);
1634 	if (error == 0) {
1635 		SOCKBUF_LOCK(&so->so_rcv);
1636 		tls->snd_tag = mst;
1637 		SOCKBUF_UNLOCK(&so->so_rcv);
1638 
1639 		counter_u64_add(ktls_ifnet_reset, 1);
1640 	} else {
1641 		/*
1642 		 * Just fall back to software decryption if a tag
1643 		 * cannot be allocated leaving the connection intact.
1644 		 * If a future input path change switches to another
1645 		 * interface this connection will resume ifnet TLS.
1646 		 */
1647 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1648 	}
1649 
1650 out:
1651 	mtx_pool_lock(mtxpool_sleep, tls);
1652 	tls->reset_pending = false;
1653 	mtx_pool_unlock(mtxpool_sleep, tls);
1654 
1655 	if (ifp != NULL)
1656 		if_rele(ifp);
1657 	sorele(so);
1658 	ktls_free(tls);
1659 }
1660 
1661 /*
1662  * Try to allocate a new TLS send tag.  This task is scheduled when
1663  * ip_output detects a route change while trying to transmit a packet
1664  * holding a TLS record.  If a new tag is allocated, replace the tag
1665  * in the TLS session.  Subsequent packets on the connection will use
1666  * the new tag.  If a new tag cannot be allocated, drop the
1667  * connection.
1668  */
1669 static void
1670 ktls_reset_send_tag(void *context, int pending)
1671 {
1672 	struct epoch_tracker et;
1673 	struct ktls_session *tls;
1674 	struct m_snd_tag *old, *new;
1675 	struct inpcb *inp;
1676 	struct tcpcb *tp;
1677 	int error;
1678 
1679 	MPASS(pending == 1);
1680 
1681 	tls = context;
1682 	inp = tls->inp;
1683 
1684 	/*
1685 	 * Free the old tag first before allocating a new one.
1686 	 * ip[6]_output_send() will treat a NULL send tag the same as
1687 	 * an ifp mismatch and drop packets until a new tag is
1688 	 * allocated.
1689 	 *
1690 	 * Write-lock the INP when changing tls->snd_tag since
1691 	 * ip[6]_output_send() holds a read-lock when reading the
1692 	 * pointer.
1693 	 */
1694 	INP_WLOCK(inp);
1695 	old = tls->snd_tag;
1696 	tls->snd_tag = NULL;
1697 	INP_WUNLOCK(inp);
1698 	if (old != NULL)
1699 		m_snd_tag_rele(old);
1700 
1701 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1702 
1703 	if (error == 0) {
1704 		INP_WLOCK(inp);
1705 		tls->snd_tag = new;
1706 		mtx_pool_lock(mtxpool_sleep, tls);
1707 		tls->reset_pending = false;
1708 		mtx_pool_unlock(mtxpool_sleep, tls);
1709 		INP_WUNLOCK(inp);
1710 
1711 		counter_u64_add(ktls_ifnet_reset, 1);
1712 
1713 		/*
1714 		 * XXX: Should we kick tcp_output explicitly now that
1715 		 * the send tag is fixed or just rely on timers?
1716 		 */
1717 	} else {
1718 		NET_EPOCH_ENTER(et);
1719 		INP_WLOCK(inp);
1720 		if (!(inp->inp_flags & INP_DROPPED)) {
1721 			tp = intotcpcb(inp);
1722 			CURVNET_SET(inp->inp_vnet);
1723 			tp = tcp_drop(tp, ECONNABORTED);
1724 			CURVNET_RESTORE();
1725 			if (tp != NULL)
1726 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1727 		}
1728 		INP_WUNLOCK(inp);
1729 		NET_EPOCH_EXIT(et);
1730 
1731 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1732 
1733 		/*
1734 		 * Leave reset_pending true to avoid future tasks while
1735 		 * the socket goes away.
1736 		 */
1737 	}
1738 
1739 	ktls_free(tls);
1740 }
1741 
1742 void
1743 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1744 {
1745 	struct ktls_session *tls;
1746 	struct socket *so;
1747 
1748 	SOCKBUF_LOCK_ASSERT(sb);
1749 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1750 	    __func__, sb));
1751 	so = __containerof(sb, struct socket, so_rcv);
1752 
1753 	tls = sb->sb_tls_info;
1754 	if_rele(tls->rx_ifp);
1755 	if_ref(ifp);
1756 	tls->rx_ifp = ifp;
1757 
1758 	/*
1759 	 * See if we should schedule a task to update the receive tag for
1760 	 * this session.
1761 	 */
1762 	mtx_pool_lock(mtxpool_sleep, tls);
1763 	if (!tls->reset_pending) {
1764 		(void) ktls_hold(tls);
1765 		soref(so);
1766 		tls->so = so;
1767 		tls->reset_pending = true;
1768 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1769 	}
1770 	mtx_pool_unlock(mtxpool_sleep, tls);
1771 }
1772 
1773 int
1774 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1775 {
1776 
1777 	if (inp == NULL)
1778 		return (ENOBUFS);
1779 
1780 	INP_LOCK_ASSERT(inp);
1781 
1782 	/*
1783 	 * See if we should schedule a task to update the send tag for
1784 	 * this session.
1785 	 */
1786 	mtx_pool_lock(mtxpool_sleep, tls);
1787 	if (!tls->reset_pending) {
1788 		(void) ktls_hold(tls);
1789 		tls->reset_pending = true;
1790 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1791 	}
1792 	mtx_pool_unlock(mtxpool_sleep, tls);
1793 	return (ENOBUFS);
1794 }
1795 
1796 #ifdef RATELIMIT
1797 int
1798 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1799 {
1800 	union if_snd_tag_modify_params params = {
1801 		.rate_limit.max_rate = max_pacing_rate,
1802 		.rate_limit.flags = M_NOWAIT,
1803 	};
1804 	struct m_snd_tag *mst;
1805 
1806 	/* Can't get to the inp, but it should be locked. */
1807 	/* INP_LOCK_ASSERT(inp); */
1808 
1809 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1810 
1811 	if (tls->snd_tag == NULL) {
1812 		/*
1813 		 * Resetting send tag, ignore this change.  The
1814 		 * pending reset may or may not see this updated rate
1815 		 * in the tcpcb.  If it doesn't, we will just lose
1816 		 * this rate change.
1817 		 */
1818 		return (0);
1819 	}
1820 
1821 	mst = tls->snd_tag;
1822 
1823 	MPASS(mst != NULL);
1824 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1825 
1826 	return (mst->sw->snd_tag_modify(mst, &params));
1827 }
1828 #endif
1829 #endif
1830 
1831 static void
1832 ktls_destroy_help(void *context, int pending __unused)
1833 {
1834 	ktls_destroy(context);
1835 }
1836 
1837 void
1838 ktls_destroy(struct ktls_session *tls)
1839 {
1840 	struct inpcb *inp;
1841 	struct tcpcb *tp;
1842 	bool wlocked;
1843 
1844 	MPASS(tls->refcount == 0);
1845 
1846 	inp = tls->inp;
1847 	if (tls->tx) {
1848 		wlocked = INP_WLOCKED(inp);
1849 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1850 			/*
1851 			 * rwlocks read locks are anonymous, and there
1852 			 * is no way to know if our current thread
1853 			 * holds an rlock on the inp.  As a rough
1854 			 * estimate, check to see if the thread holds
1855 			 * *any* rlocks at all.  If it does not, then we
1856 			 * know that we don't hold the inp rlock, and
1857 			 * can safely take the wlock
1858 			 */
1859 			if (curthread->td_rw_rlocks == 0) {
1860 				INP_WLOCK(inp);
1861 			} else {
1862 				/*
1863 				 * We might hold the rlock, so let's
1864 				 * do the destroy in a taskqueue
1865 				 * context to avoid a potential
1866 				 * deadlock.  This should be very
1867 				 * rare.
1868 				 */
1869 				counter_u64_add(ktls_destroy_task, 1);
1870 				TASK_INIT(&tls->destroy_task, 0,
1871 				    ktls_destroy_help, tls);
1872 				(void)taskqueue_enqueue(taskqueue_thread,
1873 				    &tls->destroy_task);
1874 				return;
1875 			}
1876 		}
1877 	}
1878 
1879 	if (tls->sequential_records) {
1880 		struct mbuf *m, *n;
1881 		int page_count;
1882 
1883 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1884 			page_count = m->m_epg_enc_cnt;
1885 			while (page_count > 0) {
1886 				KASSERT(page_count >= m->m_epg_nrdy,
1887 				    ("%s: too few pages", __func__));
1888 				page_count -= m->m_epg_nrdy;
1889 				m = m_free(m);
1890 			}
1891 		}
1892 	}
1893 
1894 	counter_u64_add(ktls_offload_active, -1);
1895 	switch (tls->mode) {
1896 	case TCP_TLS_MODE_SW:
1897 		switch (tls->params.cipher_algorithm) {
1898 		case CRYPTO_AES_CBC:
1899 			counter_u64_add(ktls_sw_cbc, -1);
1900 			break;
1901 		case CRYPTO_AES_NIST_GCM_16:
1902 			counter_u64_add(ktls_sw_gcm, -1);
1903 			break;
1904 		case CRYPTO_CHACHA20_POLY1305:
1905 			counter_u64_add(ktls_sw_chacha20, -1);
1906 			break;
1907 		}
1908 		break;
1909 	case TCP_TLS_MODE_IFNET:
1910 		switch (tls->params.cipher_algorithm) {
1911 		case CRYPTO_AES_CBC:
1912 			counter_u64_add(ktls_ifnet_cbc, -1);
1913 			break;
1914 		case CRYPTO_AES_NIST_GCM_16:
1915 			counter_u64_add(ktls_ifnet_gcm, -1);
1916 			break;
1917 		case CRYPTO_CHACHA20_POLY1305:
1918 			counter_u64_add(ktls_ifnet_chacha20, -1);
1919 			break;
1920 		}
1921 		if (tls->snd_tag != NULL)
1922 			m_snd_tag_rele(tls->snd_tag);
1923 		if (tls->rx_ifp != NULL)
1924 			if_rele(tls->rx_ifp);
1925 		if (tls->tx) {
1926 			INP_WLOCK_ASSERT(inp);
1927 			tp = intotcpcb(inp);
1928 			MPASS(tp->t_nic_ktls_xmit == 1);
1929 			tp->t_nic_ktls_xmit = 0;
1930 		}
1931 		break;
1932 #ifdef TCP_OFFLOAD
1933 	case TCP_TLS_MODE_TOE:
1934 		switch (tls->params.cipher_algorithm) {
1935 		case CRYPTO_AES_CBC:
1936 			counter_u64_add(ktls_toe_cbc, -1);
1937 			break;
1938 		case CRYPTO_AES_NIST_GCM_16:
1939 			counter_u64_add(ktls_toe_gcm, -1);
1940 			break;
1941 		case CRYPTO_CHACHA20_POLY1305:
1942 			counter_u64_add(ktls_toe_chacha20, -1);
1943 			break;
1944 		}
1945 		break;
1946 #endif
1947 	}
1948 	if (tls->ocf_session != NULL)
1949 		ktls_ocf_free(tls);
1950 	if (tls->params.auth_key != NULL) {
1951 		zfree(tls->params.auth_key, M_KTLS);
1952 		tls->params.auth_key = NULL;
1953 		tls->params.auth_key_len = 0;
1954 	}
1955 	if (tls->params.cipher_key != NULL) {
1956 		zfree(tls->params.cipher_key, M_KTLS);
1957 		tls->params.cipher_key = NULL;
1958 		tls->params.cipher_key_len = 0;
1959 	}
1960 	if (tls->tx) {
1961 		INP_WLOCK_ASSERT(inp);
1962 		if (!in_pcbrele_wlocked(inp) && !wlocked)
1963 			INP_WUNLOCK(inp);
1964 	}
1965 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1966 
1967 	uma_zfree(ktls_session_zone, tls);
1968 }
1969 
1970 void
1971 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1972 {
1973 
1974 	for (; m != NULL; m = m->m_next) {
1975 		KASSERT((m->m_flags & M_EXTPG) != 0,
1976 		    ("ktls_seq: mapped mbuf %p", m));
1977 
1978 		m->m_epg_seqno = sb->sb_tls_seqno;
1979 		sb->sb_tls_seqno++;
1980 	}
1981 }
1982 
1983 /*
1984  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1985  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1986  * mbuf must be populated with the payload of each TLS record.
1987  *
1988  * The record_type argument specifies the TLS record type used when
1989  * populating the TLS header.
1990  *
1991  * The enq_count argument on return is set to the number of pages of
1992  * payload data for this entire chain that need to be encrypted via SW
1993  * encryption.  The returned value should be passed to ktls_enqueue
1994  * when scheduling encryption of this chain of mbufs.  To handle the
1995  * special case of empty fragments for TLS 1.0 sessions, an empty
1996  * fragment counts as one page.
1997  */
1998 void
1999 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2000     uint8_t record_type)
2001 {
2002 	struct tls_record_layer *tlshdr;
2003 	struct mbuf *m;
2004 	uint64_t *noncep;
2005 	uint16_t tls_len;
2006 	int maxlen __diagused;
2007 
2008 	maxlen = tls->params.max_frame_len;
2009 	*enq_cnt = 0;
2010 	for (m = top; m != NULL; m = m->m_next) {
2011 		/*
2012 		 * All mbufs in the chain should be TLS records whose
2013 		 * payload does not exceed the maximum frame length.
2014 		 *
2015 		 * Empty TLS 1.0 records are permitted when using CBC.
2016 		 */
2017 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2018 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2019 		    ("ktls_frame: m %p len %d", m, m->m_len));
2020 
2021 		/*
2022 		 * TLS frames require unmapped mbufs to store session
2023 		 * info.
2024 		 */
2025 		KASSERT((m->m_flags & M_EXTPG) != 0,
2026 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2027 
2028 		tls_len = m->m_len;
2029 
2030 		/* Save a reference to the session. */
2031 		m->m_epg_tls = ktls_hold(tls);
2032 
2033 		m->m_epg_hdrlen = tls->params.tls_hlen;
2034 		m->m_epg_trllen = tls->params.tls_tlen;
2035 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2036 			int bs, delta;
2037 
2038 			/*
2039 			 * AES-CBC pads messages to a multiple of the
2040 			 * block size.  Note that the padding is
2041 			 * applied after the digest and the encryption
2042 			 * is done on the "plaintext || mac || padding".
2043 			 * At least one byte of padding is always
2044 			 * present.
2045 			 *
2046 			 * Compute the final trailer length assuming
2047 			 * at most one block of padding.
2048 			 * tls->params.tls_tlen is the maximum
2049 			 * possible trailer length (padding + digest).
2050 			 * delta holds the number of excess padding
2051 			 * bytes if the maximum were used.  Those
2052 			 * extra bytes are removed.
2053 			 */
2054 			bs = tls->params.tls_bs;
2055 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2056 			m->m_epg_trllen -= delta;
2057 		}
2058 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2059 
2060 		/* Populate the TLS header. */
2061 		tlshdr = (void *)m->m_epg_hdr;
2062 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2063 
2064 		/*
2065 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2066 		 * of TLS_RLTYPE_APP.
2067 		 */
2068 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2069 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2070 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2071 			tlshdr->tls_type = TLS_RLTYPE_APP;
2072 			/* save the real record type for later */
2073 			m->m_epg_record_type = record_type;
2074 			m->m_epg_trail[0] = record_type;
2075 		} else {
2076 			tlshdr->tls_vminor = tls->params.tls_vminor;
2077 			tlshdr->tls_type = record_type;
2078 		}
2079 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2080 
2081 		/*
2082 		 * Store nonces / explicit IVs after the end of the
2083 		 * TLS header.
2084 		 *
2085 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2086 		 * from the end of the IV.  The nonce is then
2087 		 * incremented for use by the next record.
2088 		 *
2089 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2090 		 */
2091 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2092 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2093 			noncep = (uint64_t *)(tls->params.iv + 8);
2094 			be64enc(tlshdr + 1, *noncep);
2095 			(*noncep)++;
2096 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2097 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2098 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2099 
2100 		/*
2101 		 * When using SW encryption, mark the mbuf not ready.
2102 		 * It will be marked ready via sbready() after the
2103 		 * record has been encrypted.
2104 		 *
2105 		 * When using ifnet TLS, unencrypted TLS records are
2106 		 * sent down the stack to the NIC.
2107 		 */
2108 		if (tls->mode == TCP_TLS_MODE_SW) {
2109 			m->m_flags |= M_NOTREADY;
2110 			if (__predict_false(tls_len == 0)) {
2111 				/* TLS 1.0 empty fragment. */
2112 				m->m_epg_nrdy = 1;
2113 			} else
2114 				m->m_epg_nrdy = m->m_epg_npgs;
2115 			*enq_cnt += m->m_epg_nrdy;
2116 		}
2117 	}
2118 }
2119 
2120 bool
2121 ktls_permit_empty_frames(struct ktls_session *tls)
2122 {
2123 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2124 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2125 }
2126 
2127 void
2128 ktls_check_rx(struct sockbuf *sb)
2129 {
2130 	struct tls_record_layer hdr;
2131 	struct ktls_wq *wq;
2132 	struct socket *so;
2133 	bool running;
2134 
2135 	SOCKBUF_LOCK_ASSERT(sb);
2136 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2137 	    __func__, sb));
2138 	so = __containerof(sb, struct socket, so_rcv);
2139 
2140 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2141 		return;
2142 
2143 	/* Is there enough queued for a TLS header? */
2144 	if (sb->sb_tlscc < sizeof(hdr)) {
2145 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2146 			so->so_error = EMSGSIZE;
2147 		return;
2148 	}
2149 
2150 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2151 
2152 	/* Is the entire record queued? */
2153 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2154 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2155 			so->so_error = EMSGSIZE;
2156 		return;
2157 	}
2158 
2159 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2160 
2161 	soref(so);
2162 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2163 	mtx_lock(&wq->mtx);
2164 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2165 	running = wq->running;
2166 	mtx_unlock(&wq->mtx);
2167 	if (!running)
2168 		wakeup(wq);
2169 	counter_u64_add(ktls_cnt_rx_queued, 1);
2170 }
2171 
2172 static struct mbuf *
2173 ktls_detach_record(struct sockbuf *sb, int len)
2174 {
2175 	struct mbuf *m, *n, *top;
2176 	int remain;
2177 
2178 	SOCKBUF_LOCK_ASSERT(sb);
2179 	MPASS(len <= sb->sb_tlscc);
2180 
2181 	/*
2182 	 * If TLS chain is the exact size of the record,
2183 	 * just grab the whole record.
2184 	 */
2185 	top = sb->sb_mtls;
2186 	if (sb->sb_tlscc == len) {
2187 		sb->sb_mtls = NULL;
2188 		sb->sb_mtlstail = NULL;
2189 		goto out;
2190 	}
2191 
2192 	/*
2193 	 * While it would be nice to use m_split() here, we need
2194 	 * to know exactly what m_split() allocates to update the
2195 	 * accounting, so do it inline instead.
2196 	 */
2197 	remain = len;
2198 	for (m = top; remain > m->m_len; m = m->m_next)
2199 		remain -= m->m_len;
2200 
2201 	/* Easy case: don't have to split 'm'. */
2202 	if (remain == m->m_len) {
2203 		sb->sb_mtls = m->m_next;
2204 		if (sb->sb_mtls == NULL)
2205 			sb->sb_mtlstail = NULL;
2206 		m->m_next = NULL;
2207 		goto out;
2208 	}
2209 
2210 	/*
2211 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2212 	 * with M_NOWAIT first.
2213 	 */
2214 	n = m_get(M_NOWAIT, MT_DATA);
2215 	if (n == NULL) {
2216 		/*
2217 		 * Use M_WAITOK with socket buffer unlocked.  If
2218 		 * 'sb_mtls' changes while the lock is dropped, return
2219 		 * NULL to force the caller to retry.
2220 		 */
2221 		SOCKBUF_UNLOCK(sb);
2222 
2223 		n = m_get(M_WAITOK, MT_DATA);
2224 
2225 		SOCKBUF_LOCK(sb);
2226 		if (sb->sb_mtls != top) {
2227 			m_free(n);
2228 			return (NULL);
2229 		}
2230 	}
2231 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2232 
2233 	/* Store remainder in 'n'. */
2234 	n->m_len = m->m_len - remain;
2235 	if (m->m_flags & M_EXT) {
2236 		n->m_data = m->m_data + remain;
2237 		mb_dupcl(n, m);
2238 	} else {
2239 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2240 	}
2241 
2242 	/* Trim 'm' and update accounting. */
2243 	m->m_len -= n->m_len;
2244 	sb->sb_tlscc -= n->m_len;
2245 	sb->sb_ccc -= n->m_len;
2246 
2247 	/* Account for 'n'. */
2248 	sballoc_ktls_rx(sb, n);
2249 
2250 	/* Insert 'n' into the TLS chain. */
2251 	sb->sb_mtls = n;
2252 	n->m_next = m->m_next;
2253 	if (sb->sb_mtlstail == m)
2254 		sb->sb_mtlstail = n;
2255 
2256 	/* Detach the record from the TLS chain. */
2257 	m->m_next = NULL;
2258 
2259 out:
2260 	MPASS(m_length(top, NULL) == len);
2261 	for (m = top; m != NULL; m = m->m_next)
2262 		sbfree_ktls_rx(sb, m);
2263 	sb->sb_tlsdcc = len;
2264 	sb->sb_ccc += len;
2265 	SBCHECK(sb);
2266 	return (top);
2267 }
2268 
2269 /*
2270  * Determine the length of the trailing zero padding and find the real
2271  * record type in the byte before the padding.
2272  *
2273  * Walking the mbuf chain backwards is clumsy, so another option would
2274  * be to scan forwards remembering the last non-zero byte before the
2275  * trailer.  However, it would be expensive to scan the entire record.
2276  * Instead, find the last non-zero byte of each mbuf in the chain
2277  * keeping track of the relative offset of that nonzero byte.
2278  *
2279  * trail_len is the size of the MAC/tag on input and is set to the
2280  * size of the full trailer including padding and the record type on
2281  * return.
2282  */
2283 static int
2284 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2285     int *trailer_len, uint8_t *record_typep)
2286 {
2287 	char *cp;
2288 	u_int digest_start, last_offset, m_len, offset;
2289 	uint8_t record_type;
2290 
2291 	digest_start = tls_len - *trailer_len;
2292 	last_offset = 0;
2293 	offset = 0;
2294 	for (; m != NULL && offset < digest_start;
2295 	     offset += m->m_len, m = m->m_next) {
2296 		/* Don't look for padding in the tag. */
2297 		m_len = min(digest_start - offset, m->m_len);
2298 		cp = mtod(m, char *);
2299 
2300 		/* Find last non-zero byte in this mbuf. */
2301 		while (m_len > 0 && cp[m_len - 1] == 0)
2302 			m_len--;
2303 		if (m_len > 0) {
2304 			record_type = cp[m_len - 1];
2305 			last_offset = offset + m_len;
2306 		}
2307 	}
2308 	if (last_offset < tls->params.tls_hlen)
2309 		return (EBADMSG);
2310 
2311 	*record_typep = record_type;
2312 	*trailer_len = tls_len - last_offset + 1;
2313 	return (0);
2314 }
2315 
2316 /*
2317  * Check if a mbuf chain is fully decrypted at the given offset and
2318  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2319  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2320  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2321  * is encrypted.
2322  */
2323 ktls_mbuf_crypto_st_t
2324 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2325 {
2326 	int m_flags_ored = 0;
2327 	int m_flags_anded = -1;
2328 
2329 	for (; mb != NULL; mb = mb->m_next) {
2330 		if (offset < mb->m_len)
2331 			break;
2332 		offset -= mb->m_len;
2333 	}
2334 	offset += len;
2335 
2336 	for (; mb != NULL; mb = mb->m_next) {
2337 		m_flags_ored |= mb->m_flags;
2338 		m_flags_anded &= mb->m_flags;
2339 
2340 		if (offset <= mb->m_len)
2341 			break;
2342 		offset -= mb->m_len;
2343 	}
2344 	MPASS(mb != NULL || offset == 0);
2345 
2346 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2347 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2348 	else
2349 		return ((m_flags_ored & M_DECRYPTED) ?
2350 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2351 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2352 }
2353 
2354 /*
2355  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2356  */
2357 static int
2358 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2359 {
2360 	union if_snd_tag_modify_params params;
2361 	struct m_snd_tag *mst;
2362 	struct inpcb *inp;
2363 	struct tcpcb *tp;
2364 
2365 	mst = so->so_rcv.sb_tls_info->snd_tag;
2366 	if (__predict_false(mst == NULL))
2367 		return (EINVAL);
2368 
2369 	inp = sotoinpcb(so);
2370 	if (__predict_false(inp == NULL))
2371 		return (EINVAL);
2372 
2373 	INP_RLOCK(inp);
2374 	if (inp->inp_flags & INP_DROPPED) {
2375 		INP_RUNLOCK(inp);
2376 		return (ECONNRESET);
2377 	}
2378 
2379 	tp = intotcpcb(inp);
2380 	MPASS(tp != NULL);
2381 
2382 	/* Get the TCP sequence number of the next valid TLS header. */
2383 	SOCKBUF_LOCK(&so->so_rcv);
2384 	params.tls_rx.tls_hdr_tcp_sn =
2385 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2386 	params.tls_rx.tls_rec_length = tls_len;
2387 	params.tls_rx.tls_seq_number = tls_rcd_num;
2388 	SOCKBUF_UNLOCK(&so->so_rcv);
2389 
2390 	INP_RUNLOCK(inp);
2391 
2392 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2393 	return (mst->sw->snd_tag_modify(mst, &params));
2394 }
2395 
2396 static void
2397 ktls_drop(struct socket *so, int error)
2398 {
2399 	struct epoch_tracker et;
2400 	struct inpcb *inp = sotoinpcb(so);
2401 	struct tcpcb *tp;
2402 
2403 	NET_EPOCH_ENTER(et);
2404 	INP_WLOCK(inp);
2405 	if (!(inp->inp_flags & INP_DROPPED)) {
2406 		tp = intotcpcb(inp);
2407 		CURVNET_SET(inp->inp_vnet);
2408 		tp = tcp_drop(tp, error);
2409 		CURVNET_RESTORE();
2410 		if (tp != NULL)
2411 			INP_WUNLOCK(inp);
2412 	} else {
2413 		so->so_error = error;
2414 		SOCK_RECVBUF_LOCK(so);
2415 		sorwakeup_locked(so);
2416 		INP_WUNLOCK(inp);
2417 	}
2418 	NET_EPOCH_EXIT(et);
2419 }
2420 
2421 static void
2422 ktls_decrypt(struct socket *so)
2423 {
2424 	char tls_header[MBUF_PEXT_HDR_LEN];
2425 	struct ktls_session *tls;
2426 	struct sockbuf *sb;
2427 	struct tls_record_layer *hdr;
2428 	struct tls_get_record tgr;
2429 	struct mbuf *control, *data, *m;
2430 	ktls_mbuf_crypto_st_t state;
2431 	uint64_t seqno;
2432 	int error, remain, tls_len, trail_len;
2433 	bool tls13;
2434 	uint8_t vminor, record_type;
2435 
2436 	hdr = (struct tls_record_layer *)tls_header;
2437 	sb = &so->so_rcv;
2438 	SOCKBUF_LOCK(sb);
2439 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2440 	    ("%s: socket %p not running", __func__, so));
2441 
2442 	tls = sb->sb_tls_info;
2443 	MPASS(tls != NULL);
2444 
2445 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2446 	if (tls13)
2447 		vminor = TLS_MINOR_VER_TWO;
2448 	else
2449 		vminor = tls->params.tls_vminor;
2450 	for (;;) {
2451 		/* Is there enough queued for a TLS header? */
2452 		if (sb->sb_tlscc < tls->params.tls_hlen)
2453 			break;
2454 
2455 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2456 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2457 
2458 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2459 		    hdr->tls_vminor != vminor)
2460 			error = EINVAL;
2461 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2462 			error = EINVAL;
2463 		else if (tls_len < tls->params.tls_hlen || tls_len >
2464 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2465 		    tls->params.tls_tlen)
2466 			error = EMSGSIZE;
2467 		else
2468 			error = 0;
2469 		if (__predict_false(error != 0)) {
2470 			/*
2471 			 * We have a corrupted record and are likely
2472 			 * out of sync.  The connection isn't
2473 			 * recoverable at this point, so abort it.
2474 			 */
2475 			SOCKBUF_UNLOCK(sb);
2476 			counter_u64_add(ktls_offload_corrupted_records, 1);
2477 
2478 			ktls_drop(so, error);
2479 			goto deref;
2480 		}
2481 
2482 		/* Is the entire record queued? */
2483 		if (sb->sb_tlscc < tls_len)
2484 			break;
2485 
2486 		/*
2487 		 * Split out the portion of the mbuf chain containing
2488 		 * this TLS record.
2489 		 */
2490 		data = ktls_detach_record(sb, tls_len);
2491 		if (data == NULL)
2492 			continue;
2493 		MPASS(sb->sb_tlsdcc == tls_len);
2494 
2495 		seqno = sb->sb_tls_seqno;
2496 		sb->sb_tls_seqno++;
2497 		SBCHECK(sb);
2498 		SOCKBUF_UNLOCK(sb);
2499 
2500 		/* get crypto state for this TLS record */
2501 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2502 
2503 		switch (state) {
2504 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2505 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2506 			if (error)
2507 				break;
2508 			/* FALLTHROUGH */
2509 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2510 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2511 			    &trail_len);
2512 			if (__predict_true(error == 0)) {
2513 				if (tls13) {
2514 					error = tls13_find_record_type(tls, data,
2515 					    tls_len, &trail_len, &record_type);
2516 				} else {
2517 					record_type = hdr->tls_type;
2518 				}
2519 			}
2520 			break;
2521 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2522 			/*
2523 			 * NIC TLS is only supported for AEAD
2524 			 * ciphersuites which used a fixed sized
2525 			 * trailer.
2526 			 */
2527 			if (tls13) {
2528 				trail_len = tls->params.tls_tlen - 1;
2529 				error = tls13_find_record_type(tls, data,
2530 				    tls_len, &trail_len, &record_type);
2531 			} else {
2532 				trail_len = tls->params.tls_tlen;
2533 				error = 0;
2534 				record_type = hdr->tls_type;
2535 			}
2536 			break;
2537 		default:
2538 			error = EINVAL;
2539 			break;
2540 		}
2541 		if (error) {
2542 			counter_u64_add(ktls_offload_failed_crypto, 1);
2543 
2544 			SOCKBUF_LOCK(sb);
2545 			if (sb->sb_tlsdcc == 0) {
2546 				/*
2547 				 * sbcut/drop/flush discarded these
2548 				 * mbufs.
2549 				 */
2550 				m_freem(data);
2551 				break;
2552 			}
2553 
2554 			/*
2555 			 * Drop this TLS record's data, but keep
2556 			 * decrypting subsequent records.
2557 			 */
2558 			sb->sb_ccc -= tls_len;
2559 			sb->sb_tlsdcc = 0;
2560 
2561 			if (error != EMSGSIZE)
2562 				error = EBADMSG;
2563 			CURVNET_SET(so->so_vnet);
2564 			so->so_error = error;
2565 			sorwakeup_locked(so);
2566 			CURVNET_RESTORE();
2567 
2568 			m_freem(data);
2569 
2570 			SOCKBUF_LOCK(sb);
2571 			continue;
2572 		}
2573 
2574 		/* Allocate the control mbuf. */
2575 		memset(&tgr, 0, sizeof(tgr));
2576 		tgr.tls_type = record_type;
2577 		tgr.tls_vmajor = hdr->tls_vmajor;
2578 		tgr.tls_vminor = hdr->tls_vminor;
2579 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2580 		    trail_len);
2581 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2582 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2583 
2584 		SOCKBUF_LOCK(sb);
2585 		if (sb->sb_tlsdcc == 0) {
2586 			/* sbcut/drop/flush discarded these mbufs. */
2587 			MPASS(sb->sb_tlscc == 0);
2588 			m_freem(data);
2589 			m_freem(control);
2590 			break;
2591 		}
2592 
2593 		/*
2594 		 * Clear the 'dcc' accounting in preparation for
2595 		 * adding the decrypted record.
2596 		 */
2597 		sb->sb_ccc -= tls_len;
2598 		sb->sb_tlsdcc = 0;
2599 		SBCHECK(sb);
2600 
2601 		/* If there is no payload, drop all of the data. */
2602 		if (tgr.tls_length == htobe16(0)) {
2603 			m_freem(data);
2604 			data = NULL;
2605 		} else {
2606 			/* Trim header. */
2607 			remain = tls->params.tls_hlen;
2608 			while (remain > 0) {
2609 				if (data->m_len > remain) {
2610 					data->m_data += remain;
2611 					data->m_len -= remain;
2612 					break;
2613 				}
2614 				remain -= data->m_len;
2615 				data = m_free(data);
2616 			}
2617 
2618 			/* Trim trailer and clear M_NOTREADY. */
2619 			remain = be16toh(tgr.tls_length);
2620 			m = data;
2621 			for (m = data; remain > m->m_len; m = m->m_next) {
2622 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2623 				remain -= m->m_len;
2624 			}
2625 			m->m_len = remain;
2626 			m_freem(m->m_next);
2627 			m->m_next = NULL;
2628 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2629 
2630 			/* Set EOR on the final mbuf. */
2631 			m->m_flags |= M_EOR;
2632 		}
2633 
2634 		sbappendcontrol_locked(sb, data, control, 0);
2635 
2636 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2637 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2638 			SOCKBUF_UNLOCK(sb);
2639 			ktls_resync_ifnet(so, tls_len, seqno);
2640 			SOCKBUF_LOCK(sb);
2641 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2642 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2643 			SOCKBUF_UNLOCK(sb);
2644 			ktls_resync_ifnet(so, 0, seqno);
2645 			SOCKBUF_LOCK(sb);
2646 		}
2647 	}
2648 
2649 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2650 
2651 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2652 		so->so_error = EMSGSIZE;
2653 
2654 	sorwakeup_locked(so);
2655 
2656 deref:
2657 	SOCKBUF_UNLOCK_ASSERT(sb);
2658 
2659 	CURVNET_SET(so->so_vnet);
2660 	sorele(so);
2661 	CURVNET_RESTORE();
2662 }
2663 
2664 void
2665 ktls_enqueue_to_free(struct mbuf *m)
2666 {
2667 	struct ktls_wq *wq;
2668 	bool running;
2669 
2670 	/* Mark it for freeing. */
2671 	m->m_epg_flags |= EPG_FLAG_2FREE;
2672 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2673 	mtx_lock(&wq->mtx);
2674 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2675 	running = wq->running;
2676 	mtx_unlock(&wq->mtx);
2677 	if (!running)
2678 		wakeup(wq);
2679 }
2680 
2681 static void *
2682 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2683 {
2684 	void *buf;
2685 	int domain, running;
2686 
2687 	if (m->m_epg_npgs <= 2)
2688 		return (NULL);
2689 	if (ktls_buffer_zone == NULL)
2690 		return (NULL);
2691 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2692 		/*
2693 		 * Rate-limit allocation attempts after a failure.
2694 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2695 		 * the free page queues and may fail consistently if memory is
2696 		 * fragmented.
2697 		 */
2698 		return (NULL);
2699 	}
2700 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2701 	if (buf == NULL) {
2702 		domain = PCPU_GET(domain);
2703 		wq->lastallocfail = ticks;
2704 
2705 		/*
2706 		 * Note that this check is "racy", but the races are
2707 		 * harmless, and are either a spurious wakeup if
2708 		 * multiple threads fail allocations before the alloc
2709 		 * thread wakes, or waiting an extra second in case we
2710 		 * see an old value of running == true.
2711 		 */
2712 		if (!VM_DOMAIN_EMPTY(domain)) {
2713 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2714 			if (!running)
2715 				wakeup(&ktls_domains[domain].alloc_td);
2716 		}
2717 	}
2718 	return (buf);
2719 }
2720 
2721 static int
2722 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2723     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2724 {
2725 	vm_page_t pg;
2726 	int error, i, len, off;
2727 
2728 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2729 	    ("%p not unready & nomap mbuf\n", m));
2730 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2731 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2732 	    ktls_maxlen));
2733 
2734 	/* Anonymous mbufs are encrypted in place. */
2735 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2736 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2737 
2738 	/*
2739 	 * For file-backed mbufs (from sendfile), anonymous wired
2740 	 * pages are allocated and used as the encryption destination.
2741 	 */
2742 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2743 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2744 		    m->m_epg_1st_off;
2745 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2746 		    m->m_epg_1st_off;
2747 		state->dst_iov[0].iov_len = len;
2748 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2749 		i = 1;
2750 	} else {
2751 		off = m->m_epg_1st_off;
2752 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2753 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2754 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2755 			len = m_epg_pagelen(m, i, off);
2756 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2757 			state->dst_iov[i].iov_base =
2758 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2759 			state->dst_iov[i].iov_len = len;
2760 		}
2761 	}
2762 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2763 	state->dst_iov[i].iov_base = m->m_epg_trail;
2764 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2765 
2766 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2767 
2768 	if (__predict_false(error != 0)) {
2769 		/* Free the anonymous pages. */
2770 		if (state->cbuf != NULL)
2771 			uma_zfree(ktls_buffer_zone, state->cbuf);
2772 		else {
2773 			for (i = 0; i < m->m_epg_npgs; i++) {
2774 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2775 				(void)vm_page_unwire_noq(pg);
2776 				vm_page_free(pg);
2777 			}
2778 		}
2779 	}
2780 	return (error);
2781 }
2782 
2783 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2784 static u_int
2785 ktls_batched_records(struct mbuf *m)
2786 {
2787 	int page_count, records;
2788 
2789 	records = 0;
2790 	page_count = m->m_epg_enc_cnt;
2791 	while (page_count > 0) {
2792 		records++;
2793 		page_count -= m->m_epg_nrdy;
2794 		m = m->m_next;
2795 	}
2796 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2797 	return (records);
2798 }
2799 
2800 void
2801 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2802 {
2803 	struct ktls_session *tls;
2804 	struct ktls_wq *wq;
2805 	int queued;
2806 	bool running;
2807 
2808 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2809 	    (M_EXTPG | M_NOTREADY)),
2810 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2811 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2812 
2813 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2814 
2815 	m->m_epg_enc_cnt = page_count;
2816 
2817 	/*
2818 	 * Save a pointer to the socket.  The caller is responsible
2819 	 * for taking an additional reference via soref().
2820 	 */
2821 	m->m_epg_so = so;
2822 
2823 	queued = 1;
2824 	tls = m->m_epg_tls;
2825 	wq = &ktls_wq[tls->wq_index];
2826 	mtx_lock(&wq->mtx);
2827 	if (__predict_false(tls->sequential_records)) {
2828 		/*
2829 		 * For TLS 1.0, records must be encrypted
2830 		 * sequentially.  For a given connection, all records
2831 		 * queued to the associated work queue are processed
2832 		 * sequentially.  However, sendfile(2) might complete
2833 		 * I/O requests spanning multiple TLS records out of
2834 		 * order.  Here we ensure TLS records are enqueued to
2835 		 * the work queue in FIFO order.
2836 		 *
2837 		 * tls->next_seqno holds the sequence number of the
2838 		 * next TLS record that should be enqueued to the work
2839 		 * queue.  If this next record is not tls->next_seqno,
2840 		 * it must be a future record, so insert it, sorted by
2841 		 * TLS sequence number, into tls->pending_records and
2842 		 * return.
2843 		 *
2844 		 * If this TLS record matches tls->next_seqno, place
2845 		 * it in the work queue and then check
2846 		 * tls->pending_records to see if any
2847 		 * previously-queued records are now ready for
2848 		 * encryption.
2849 		 */
2850 		if (m->m_epg_seqno != tls->next_seqno) {
2851 			struct mbuf *n, *p;
2852 
2853 			p = NULL;
2854 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2855 				if (n->m_epg_seqno > m->m_epg_seqno)
2856 					break;
2857 				p = n;
2858 			}
2859 			if (n == NULL)
2860 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2861 				    m_epg_stailq);
2862 			else if (p == NULL)
2863 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2864 				    m_epg_stailq);
2865 			else
2866 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2867 				    m_epg_stailq);
2868 			mtx_unlock(&wq->mtx);
2869 			counter_u64_add(ktls_cnt_tx_pending, 1);
2870 			return;
2871 		}
2872 
2873 		tls->next_seqno += ktls_batched_records(m);
2874 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2875 
2876 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2877 			struct mbuf *n;
2878 
2879 			n = STAILQ_FIRST(&tls->pending_records);
2880 			if (n->m_epg_seqno != tls->next_seqno)
2881 				break;
2882 
2883 			queued++;
2884 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2885 			tls->next_seqno += ktls_batched_records(n);
2886 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2887 		}
2888 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2889 	} else
2890 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2891 
2892 	running = wq->running;
2893 	mtx_unlock(&wq->mtx);
2894 	if (!running)
2895 		wakeup(wq);
2896 	counter_u64_add(ktls_cnt_tx_queued, queued);
2897 }
2898 
2899 /*
2900  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2901  * the pages from the file and replace them with the anonymous pages
2902  * allocated in ktls_encrypt_record().
2903  */
2904 static void
2905 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2906 {
2907 	int i;
2908 
2909 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2910 
2911 	/* Free the old pages. */
2912 	m->m_ext.ext_free(m);
2913 
2914 	/* Replace them with the new pages. */
2915 	if (state->cbuf != NULL) {
2916 		for (i = 0; i < m->m_epg_npgs; i++)
2917 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2918 
2919 		/* Contig pages should go back to the cache. */
2920 		m->m_ext.ext_free = ktls_free_mext_contig;
2921 	} else {
2922 		for (i = 0; i < m->m_epg_npgs; i++)
2923 			m->m_epg_pa[i] = state->parray[i];
2924 
2925 		/* Use the basic free routine. */
2926 		m->m_ext.ext_free = mb_free_mext_pgs;
2927 	}
2928 
2929 	/* Pages are now writable. */
2930 	m->m_epg_flags |= EPG_FLAG_ANON;
2931 }
2932 
2933 static __noinline void
2934 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2935 {
2936 	struct ktls_ocf_encrypt_state state;
2937 	struct ktls_session *tls;
2938 	struct socket *so;
2939 	struct mbuf *m;
2940 	int error, npages, total_pages;
2941 
2942 	so = top->m_epg_so;
2943 	tls = top->m_epg_tls;
2944 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2945 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2946 #ifdef INVARIANTS
2947 	top->m_epg_so = NULL;
2948 #endif
2949 	total_pages = top->m_epg_enc_cnt;
2950 	npages = 0;
2951 
2952 	/*
2953 	 * Encrypt the TLS records in the chain of mbufs starting with
2954 	 * 'top'.  'total_pages' gives us a total count of pages and is
2955 	 * used to know when we have finished encrypting the TLS
2956 	 * records originally queued with 'top'.
2957 	 *
2958 	 * NB: These mbufs are queued in the socket buffer and
2959 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2960 	 * socket buffer lock is not held while traversing this chain.
2961 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2962 	 * pointers should be stable.  However, the 'm_next' of the
2963 	 * last mbuf encrypted is not necessarily NULL.  It can point
2964 	 * to other mbufs appended while 'top' was on the TLS work
2965 	 * queue.
2966 	 *
2967 	 * Each mbuf holds an entire TLS record.
2968 	 */
2969 	error = 0;
2970 	for (m = top; npages != total_pages; m = m->m_next) {
2971 		KASSERT(m->m_epg_tls == tls,
2972 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2973 		    tls, m->m_epg_tls));
2974 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2975 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2976 		    total_pages, m));
2977 
2978 		error = ktls_encrypt_record(wq, m, tls, &state);
2979 		if (error) {
2980 			counter_u64_add(ktls_offload_failed_crypto, 1);
2981 			break;
2982 		}
2983 
2984 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2985 			ktls_finish_nonanon(m, &state);
2986 
2987 		npages += m->m_epg_nrdy;
2988 
2989 		/*
2990 		 * Drop a reference to the session now that it is no
2991 		 * longer needed.  Existing code depends on encrypted
2992 		 * records having no associated session vs
2993 		 * yet-to-be-encrypted records having an associated
2994 		 * session.
2995 		 */
2996 		m->m_epg_tls = NULL;
2997 		ktls_free(tls);
2998 	}
2999 
3000 	CURVNET_SET(so->so_vnet);
3001 	if (error == 0) {
3002 		(void)so->so_proto->pr_ready(so, top, npages);
3003 	} else {
3004 		ktls_drop(so, EIO);
3005 		mb_free_notready(top, total_pages);
3006 	}
3007 
3008 	sorele(so);
3009 	CURVNET_RESTORE();
3010 }
3011 
3012 void
3013 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3014 {
3015 	struct ktls_session *tls;
3016 	struct socket *so;
3017 	struct mbuf *m;
3018 	int npages;
3019 
3020 	m = state->m;
3021 
3022 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3023 		ktls_finish_nonanon(m, state);
3024 
3025 	so = state->so;
3026 	free(state, M_KTLS);
3027 
3028 	/*
3029 	 * Drop a reference to the session now that it is no longer
3030 	 * needed.  Existing code depends on encrypted records having
3031 	 * no associated session vs yet-to-be-encrypted records having
3032 	 * an associated session.
3033 	 */
3034 	tls = m->m_epg_tls;
3035 	m->m_epg_tls = NULL;
3036 	ktls_free(tls);
3037 
3038 	if (error != 0)
3039 		counter_u64_add(ktls_offload_failed_crypto, 1);
3040 
3041 	CURVNET_SET(so->so_vnet);
3042 	npages = m->m_epg_nrdy;
3043 
3044 	if (error == 0) {
3045 		(void)so->so_proto->pr_ready(so, m, npages);
3046 	} else {
3047 		ktls_drop(so, EIO);
3048 		mb_free_notready(m, npages);
3049 	}
3050 
3051 	sorele(so);
3052 	CURVNET_RESTORE();
3053 }
3054 
3055 /*
3056  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3057  * (coprocessors) where encryption does not use host CPU resources and
3058  * it can be beneficial to queue more requests than CPUs.
3059  */
3060 static __noinline void
3061 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3062 {
3063 	struct ktls_ocf_encrypt_state *state;
3064 	struct ktls_session *tls;
3065 	struct socket *so;
3066 	struct mbuf *m, *n;
3067 	int error, mpages, npages, total_pages;
3068 
3069 	so = top->m_epg_so;
3070 	tls = top->m_epg_tls;
3071 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3072 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3073 #ifdef INVARIANTS
3074 	top->m_epg_so = NULL;
3075 #endif
3076 	total_pages = top->m_epg_enc_cnt;
3077 	npages = 0;
3078 
3079 	error = 0;
3080 	for (m = top; npages != total_pages; m = n) {
3081 		KASSERT(m->m_epg_tls == tls,
3082 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3083 		    tls, m->m_epg_tls));
3084 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3085 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3086 		    total_pages, m));
3087 
3088 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3089 		soref(so);
3090 		state->so = so;
3091 		state->m = m;
3092 
3093 		mpages = m->m_epg_nrdy;
3094 		n = m->m_next;
3095 
3096 		error = ktls_encrypt_record(wq, m, tls, state);
3097 		if (error) {
3098 			counter_u64_add(ktls_offload_failed_crypto, 1);
3099 			free(state, M_KTLS);
3100 			CURVNET_SET(so->so_vnet);
3101 			sorele(so);
3102 			CURVNET_RESTORE();
3103 			break;
3104 		}
3105 
3106 		npages += mpages;
3107 	}
3108 
3109 	CURVNET_SET(so->so_vnet);
3110 	if (error != 0) {
3111 		ktls_drop(so, EIO);
3112 		mb_free_notready(m, total_pages - npages);
3113 	}
3114 
3115 	sorele(so);
3116 	CURVNET_RESTORE();
3117 }
3118 
3119 static int
3120 ktls_bind_domain(int domain)
3121 {
3122 	int error;
3123 
3124 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3125 	if (error != 0)
3126 		return (error);
3127 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3128 	return (0);
3129 }
3130 
3131 static void
3132 ktls_alloc_thread(void *ctx)
3133 {
3134 	struct ktls_domain_info *ktls_domain = ctx;
3135 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3136 	void **buf;
3137 	struct sysctl_oid *oid;
3138 	char name[80];
3139 	int domain, error, i, nbufs;
3140 
3141 	domain = ktls_domain - ktls_domains;
3142 	if (bootverbose)
3143 		printf("Starting KTLS alloc thread for domain %d\n", domain);
3144 	error = ktls_bind_domain(domain);
3145 	if (error)
3146 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3147 		    domain, error);
3148 	snprintf(name, sizeof(name), "domain%d", domain);
3149 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3150 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3151 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3152 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
3153 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3154 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3155 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3156 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3157 
3158 	buf = NULL;
3159 	nbufs = 0;
3160 	for (;;) {
3161 		atomic_store_int(&sc->running, 0);
3162 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3163 		atomic_store_int(&sc->running, 1);
3164 		sc->wakeups++;
3165 		if (nbufs != ktls_max_alloc) {
3166 			free(buf, M_KTLS);
3167 			nbufs = atomic_load_int(&ktls_max_alloc);
3168 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3169 			    M_WAITOK | M_ZERO);
3170 		}
3171 		/*
3172 		 * Below we allocate nbufs with different allocation
3173 		 * flags than we use when allocating normally during
3174 		 * encryption in the ktls worker thread.  We specify
3175 		 * M_NORECLAIM in the worker thread. However, we omit
3176 		 * that flag here and add M_WAITOK so that the VM
3177 		 * system is permitted to perform expensive work to
3178 		 * defragment memory.  We do this here, as it does not
3179 		 * matter if this thread blocks.  If we block a ktls
3180 		 * worker thread, we risk developing backlogs of
3181 		 * buffers to be encrypted, leading to surges of
3182 		 * traffic and potential NIC output drops.
3183 		 */
3184 		for (i = 0; i < nbufs; i++) {
3185 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3186 			sc->allocs++;
3187 		}
3188 		for (i = 0; i < nbufs; i++) {
3189 			uma_zfree(ktls_buffer_zone, buf[i]);
3190 			buf[i] = NULL;
3191 		}
3192 	}
3193 }
3194 
3195 static void
3196 ktls_work_thread(void *ctx)
3197 {
3198 	struct ktls_wq *wq = ctx;
3199 	struct mbuf *m, *n;
3200 	struct socket *so, *son;
3201 	STAILQ_HEAD(, mbuf) local_m_head;
3202 	STAILQ_HEAD(, socket) local_so_head;
3203 	int cpu;
3204 
3205 	cpu = wq - ktls_wq;
3206 	if (bootverbose)
3207 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3208 
3209 	/*
3210 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3211 	 * we bind to the NUMA domain instead.
3212 	 */
3213 	if (ktls_bind_threads) {
3214 		int error;
3215 
3216 		if (ktls_bind_threads > 1) {
3217 			struct pcpu *pc = pcpu_find(cpu);
3218 
3219 			error = ktls_bind_domain(pc->pc_domain);
3220 		} else {
3221 			cpuset_t mask;
3222 
3223 			CPU_SETOF(cpu, &mask);
3224 			error = cpuset_setthread(curthread->td_tid, &mask);
3225 		}
3226 		if (error)
3227 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3228 				cpu, error);
3229 	}
3230 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3231 	fpu_kern_thread(0);
3232 #endif
3233 	for (;;) {
3234 		mtx_lock(&wq->mtx);
3235 		while (STAILQ_EMPTY(&wq->m_head) &&
3236 		    STAILQ_EMPTY(&wq->so_head)) {
3237 			wq->running = false;
3238 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3239 			wq->running = true;
3240 		}
3241 
3242 		STAILQ_INIT(&local_m_head);
3243 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3244 		STAILQ_INIT(&local_so_head);
3245 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3246 		mtx_unlock(&wq->mtx);
3247 
3248 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3249 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3250 				ktls_free(m->m_epg_tls);
3251 				m_free_raw(m);
3252 			} else {
3253 				if (m->m_epg_tls->sync_dispatch)
3254 					ktls_encrypt(wq, m);
3255 				else
3256 					ktls_encrypt_async(wq, m);
3257 				counter_u64_add(ktls_cnt_tx_queued, -1);
3258 			}
3259 		}
3260 
3261 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3262 			ktls_decrypt(so);
3263 			counter_u64_add(ktls_cnt_rx_queued, -1);
3264 		}
3265 	}
3266 }
3267 
3268 #if defined(INET) || defined(INET6)
3269 static void
3270 ktls_disable_ifnet_help(void *context, int pending __unused)
3271 {
3272 	struct ktls_session *tls;
3273 	struct inpcb *inp;
3274 	struct tcpcb *tp;
3275 	struct socket *so;
3276 	int err;
3277 
3278 	tls = context;
3279 	inp = tls->inp;
3280 	if (inp == NULL)
3281 		return;
3282 	INP_WLOCK(inp);
3283 	so = inp->inp_socket;
3284 	MPASS(so != NULL);
3285 	if (inp->inp_flags & INP_DROPPED) {
3286 		goto out;
3287 	}
3288 
3289 	if (so->so_snd.sb_tls_info != NULL)
3290 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3291 	else
3292 		err = ENXIO;
3293 	if (err == 0) {
3294 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3295 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3296 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3297 		    (tp = intotcpcb(inp)) != NULL &&
3298 		    tp->t_fb->tfb_hwtls_change != NULL)
3299 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3300 	} else {
3301 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3302 	}
3303 
3304 out:
3305 	CURVNET_SET(so->so_vnet);
3306 	sorele(so);
3307 	CURVNET_RESTORE();
3308 	INP_WUNLOCK(inp);
3309 	ktls_free(tls);
3310 }
3311 
3312 /*
3313  * Called when re-transmits are becoming a substantial portion of the
3314  * sends on this connection.  When this happens, we transition the
3315  * connection to software TLS.  This is needed because most inline TLS
3316  * NICs keep crypto state only for in-order transmits.  This means
3317  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3318  * re-DMA the entire TLS record up to and including the current
3319  * segment.  This means that when re-transmitting the last ~1448 byte
3320  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3321  * of magnitude more data than we are sending.  This can cause the
3322  * PCIe link to saturate well before the network, which can cause
3323  * output drops, and a general loss of capacity.
3324  */
3325 void
3326 ktls_disable_ifnet(void *arg)
3327 {
3328 	struct tcpcb *tp;
3329 	struct inpcb *inp;
3330 	struct socket *so;
3331 	struct ktls_session *tls;
3332 
3333 	tp = arg;
3334 	inp = tptoinpcb(tp);
3335 	INP_WLOCK_ASSERT(inp);
3336 	so = inp->inp_socket;
3337 	SOCK_LOCK(so);
3338 	tls = so->so_snd.sb_tls_info;
3339 	if (tp->t_nic_ktls_xmit_dis == 1) {
3340 		SOCK_UNLOCK(so);
3341 		return;
3342 	}
3343 
3344 	/*
3345 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3346 	 * ifnet can only be done once per connection, so we never want
3347 	 * to do it again
3348 	 */
3349 
3350 	(void)ktls_hold(tls);
3351 	soref(so);
3352 	tp->t_nic_ktls_xmit_dis = 1;
3353 	SOCK_UNLOCK(so);
3354 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3355 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3356 }
3357 #endif
3358