xref: /freebsd/sys/kern/uipc_ktls.c (revision 43a5ec4eb41567cc92586503212743d89686d78f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/domainset.h>
40 #include <sys/endian.h>
41 #include <sys/ktls.h>
42 #include <sys/lock.h>
43 #include <sys/mbuf.h>
44 #include <sys/mutex.h>
45 #include <sys/rmlock.h>
46 #include <sys/proc.h>
47 #include <sys/protosw.h>
48 #include <sys/refcount.h>
49 #include <sys/smp.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/kthread.h>
55 #include <sys/uio.h>
56 #include <sys/vmmeter.h>
57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58 #include <machine/pcb.h>
59 #endif
60 #include <machine/vmparam.h>
61 #include <net/if.h>
62 #include <net/if_var.h>
63 #ifdef RSS
64 #include <net/netisr.h>
65 #include <net/rss_config.h>
66 #endif
67 #include <net/route.h>
68 #include <net/route/nhop.h>
69 #if defined(INET) || defined(INET6)
70 #include <netinet/in.h>
71 #include <netinet/in_pcb.h>
72 #endif
73 #include <netinet/tcp_var.h>
74 #ifdef TCP_OFFLOAD
75 #include <netinet/tcp_offload.h>
76 #endif
77 #include <opencrypto/cryptodev.h>
78 #include <opencrypto/ktls.h>
79 #include <vm/uma_dbg.h>
80 #include <vm/vm.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pagequeue.h>
84 
85 struct ktls_wq {
86 	struct mtx	mtx;
87 	STAILQ_HEAD(, mbuf) m_head;
88 	STAILQ_HEAD(, socket) so_head;
89 	bool		running;
90 	int		lastallocfail;
91 } __aligned(CACHE_LINE_SIZE);
92 
93 struct ktls_alloc_thread {
94 	uint64_t wakeups;
95 	uint64_t allocs;
96 	struct thread *td;
97 	int running;
98 };
99 
100 struct ktls_domain_info {
101 	int count;
102 	int cpu[MAXCPU];
103 	struct ktls_alloc_thread alloc_td;
104 };
105 
106 struct ktls_domain_info ktls_domains[MAXMEMDOM];
107 static struct ktls_wq *ktls_wq;
108 static struct proc *ktls_proc;
109 static uma_zone_t ktls_session_zone;
110 static uma_zone_t ktls_buffer_zone;
111 static uint16_t ktls_cpuid_lookup[MAXCPU];
112 static int ktls_init_state;
113 static struct sx ktls_init_lock;
114 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
115 
116 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117     "Kernel TLS offload");
118 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
119     "Kernel TLS offload stats");
120 
121 #ifdef RSS
122 static int ktls_bind_threads = 1;
123 #else
124 static int ktls_bind_threads;
125 #endif
126 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
127     &ktls_bind_threads, 0,
128     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
129 
130 static u_int ktls_maxlen = 16384;
131 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
132     &ktls_maxlen, 0, "Maximum TLS record size");
133 
134 static int ktls_number_threads;
135 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
136     &ktls_number_threads, 0,
137     "Number of TLS threads in thread-pool");
138 
139 unsigned int ktls_ifnet_max_rexmit_pct = 2;
140 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
141     &ktls_ifnet_max_rexmit_pct, 2,
142     "Max percent bytes retransmitted before ifnet TLS is disabled");
143 
144 static bool ktls_offload_enable;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
146     &ktls_offload_enable, 0,
147     "Enable support for kernel TLS offload");
148 
149 static bool ktls_cbc_enable = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
151     &ktls_cbc_enable, 1,
152     "Enable Support of AES-CBC crypto for kernel TLS");
153 
154 static bool ktls_sw_buffer_cache = true;
155 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
156     &ktls_sw_buffer_cache, 1,
157     "Enable caching of output buffers for SW encryption");
158 
159 static int ktls_max_alloc = 128;
160 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
161     &ktls_max_alloc, 128,
162     "Max number of 16k buffers to allocate in thread context");
163 
164 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
165 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
166     &ktls_tasks_active, "Number of active tasks");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
170     &ktls_cnt_tx_pending,
171     "Number of TLS 1.0 records waiting for earlier TLS records");
172 
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
175     &ktls_cnt_tx_queued,
176     "Number of TLS records in queue to tasks for SW encryption");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
180     &ktls_cnt_rx_queued,
181     "Number of TLS sockets in queue to tasks for SW decryption");
182 
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
185     CTLFLAG_RD, &ktls_offload_total,
186     "Total successful TLS setups (parameters set)");
187 
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
190     CTLFLAG_RD, &ktls_offload_enable_calls,
191     "Total number of TLS enable calls made");
192 
193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
195     &ktls_offload_active, "Total Active TLS sessions");
196 
197 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
199     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
200 
201 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
203     &ktls_offload_failed_crypto, "Total TLS crypto failures");
204 
205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
207     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
208 
209 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
211     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
212 
213 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
215     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
216 
217 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
218 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
219     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
220 
221 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
222 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
223     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
224 
225 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226     "Software TLS session stats");
227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228     "Hardware (ifnet) TLS session stats");
229 #ifdef TCP_OFFLOAD
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "TOE TLS session stats");
232 #endif
233 
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236     "Active number of software TLS sessions using AES-CBC");
237 
238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240     "Active number of software TLS sessions using AES-GCM");
241 
242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
244     &ktls_sw_chacha20,
245     "Active number of software TLS sessions using Chacha20-Poly1305");
246 
247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
249     &ktls_ifnet_cbc,
250     "Active number of ifnet TLS sessions using AES-CBC");
251 
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
254     &ktls_ifnet_gcm,
255     "Active number of ifnet TLS sessions using AES-GCM");
256 
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259     &ktls_ifnet_chacha20,
260     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
261 
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
265 
266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268     &ktls_ifnet_reset_dropped,
269     "TLS sessions dropped after failing to update ifnet send tag");
270 
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273     &ktls_ifnet_reset_failed,
274     "TLS sessions that failed to allocate a new ifnet send tag");
275 
276 static int ktls_ifnet_permitted;
277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278     &ktls_ifnet_permitted, 1,
279     "Whether to permit hardware (ifnet) TLS sessions");
280 
281 #ifdef TCP_OFFLOAD
282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
284     &ktls_toe_cbc,
285     "Active number of TOE TLS sessions using AES-CBC");
286 
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
289     &ktls_toe_gcm,
290     "Active number of TOE TLS sessions using AES-GCM");
291 
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
294     &ktls_toe_chacha20,
295     "Active number of TOE TLS sessions using Chacha20-Poly1305");
296 #endif
297 
298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
299 
300 static void ktls_cleanup(struct ktls_session *tls);
301 #if defined(INET) || defined(INET6)
302 static void ktls_reset_send_tag(void *context, int pending);
303 #endif
304 static void ktls_work_thread(void *ctx);
305 static void ktls_alloc_thread(void *ctx);
306 
307 #if defined(INET) || defined(INET6)
308 static u_int
309 ktls_get_cpu(struct socket *so)
310 {
311 	struct inpcb *inp;
312 #ifdef NUMA
313 	struct ktls_domain_info *di;
314 #endif
315 	u_int cpuid;
316 
317 	inp = sotoinpcb(so);
318 #ifdef RSS
319 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
320 	if (cpuid != NETISR_CPUID_NONE)
321 		return (cpuid);
322 #endif
323 	/*
324 	 * Just use the flowid to shard connections in a repeatable
325 	 * fashion.  Note that TLS 1.0 sessions rely on the
326 	 * serialization provided by having the same connection use
327 	 * the same queue.
328 	 */
329 #ifdef NUMA
330 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
331 		di = &ktls_domains[inp->inp_numa_domain];
332 		cpuid = di->cpu[inp->inp_flowid % di->count];
333 	} else
334 #endif
335 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
336 	return (cpuid);
337 }
338 #endif
339 
340 static int
341 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
342 {
343 	vm_page_t m;
344 	int i, req;
345 
346 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
347 	    ("%s: ktls max length %d is not page size-aligned",
348 	    __func__, ktls_maxlen));
349 
350 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
351 	for (i = 0; i < count; i++) {
352 		m = vm_page_alloc_noobj_contig_domain(domain, req,
353 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
354 		    VM_MEMATTR_DEFAULT);
355 		if (m == NULL)
356 			break;
357 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
358 	}
359 	return (i);
360 }
361 
362 static void
363 ktls_buffer_release(void *arg __unused, void **store, int count)
364 {
365 	vm_page_t m;
366 	int i, j;
367 
368 	for (i = 0; i < count; i++) {
369 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
370 		for (j = 0; j < atop(ktls_maxlen); j++) {
371 			(void)vm_page_unwire_noq(m + j);
372 			vm_page_free(m + j);
373 		}
374 	}
375 }
376 
377 static void
378 ktls_free_mext_contig(struct mbuf *m)
379 {
380 	M_ASSERTEXTPG(m);
381 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
382 }
383 
384 static int
385 ktls_init(void)
386 {
387 	struct thread *td;
388 	struct pcpu *pc;
389 	int count, domain, error, i;
390 
391 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
392 	    M_WAITOK | M_ZERO);
393 
394 	ktls_session_zone = uma_zcreate("ktls_session",
395 	    sizeof(struct ktls_session),
396 	    NULL, NULL, NULL, NULL,
397 	    UMA_ALIGN_CACHE, 0);
398 
399 	if (ktls_sw_buffer_cache) {
400 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
401 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
402 		    ktls_buffer_import, ktls_buffer_release, NULL,
403 		    UMA_ZONE_FIRSTTOUCH);
404 	}
405 
406 	/*
407 	 * Initialize the workqueues to run the TLS work.  We create a
408 	 * work queue for each CPU.
409 	 */
410 	CPU_FOREACH(i) {
411 		STAILQ_INIT(&ktls_wq[i].m_head);
412 		STAILQ_INIT(&ktls_wq[i].so_head);
413 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
414 		if (ktls_bind_threads > 1) {
415 			pc = pcpu_find(i);
416 			domain = pc->pc_domain;
417 			count = ktls_domains[domain].count;
418 			ktls_domains[domain].cpu[count] = i;
419 			ktls_domains[domain].count++;
420 		}
421 		ktls_cpuid_lookup[ktls_number_threads] = i;
422 		ktls_number_threads++;
423 	}
424 
425 	/*
426 	 * If we somehow have an empty domain, fall back to choosing
427 	 * among all KTLS threads.
428 	 */
429 	if (ktls_bind_threads > 1) {
430 		for (i = 0; i < vm_ndomains; i++) {
431 			if (ktls_domains[i].count == 0) {
432 				ktls_bind_threads = 1;
433 				break;
434 			}
435 		}
436 	}
437 
438 	/* Start kthreads for each workqueue. */
439 	CPU_FOREACH(i) {
440 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
441 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
442 		if (error) {
443 			printf("Can't add KTLS thread %d error %d\n", i, error);
444 			return (error);
445 		}
446 	}
447 
448 	/*
449 	 * Start an allocation thread per-domain to perform blocking allocations
450 	 * of 16k physically contiguous TLS crypto destination buffers.
451 	 */
452 	if (ktls_sw_buffer_cache) {
453 		for (domain = 0; domain < vm_ndomains; domain++) {
454 			if (VM_DOMAIN_EMPTY(domain))
455 				continue;
456 			if (CPU_EMPTY(&cpuset_domain[domain]))
457 				continue;
458 			error = kproc_kthread_add(ktls_alloc_thread,
459 			    &ktls_domains[domain], &ktls_proc,
460 			    &ktls_domains[domain].alloc_td.td,
461 			    0, 0, "KTLS", "alloc_%d", domain);
462 			if (error) {
463 				printf("Can't add KTLS alloc thread %d error %d\n",
464 				    domain, error);
465 				return (error);
466 			}
467 		}
468 	}
469 
470 	if (bootverbose)
471 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
472 	return (0);
473 }
474 
475 static int
476 ktls_start_kthreads(void)
477 {
478 	int error, state;
479 
480 start:
481 	state = atomic_load_acq_int(&ktls_init_state);
482 	if (__predict_true(state > 0))
483 		return (0);
484 	if (state < 0)
485 		return (ENXIO);
486 
487 	sx_xlock(&ktls_init_lock);
488 	if (ktls_init_state != 0) {
489 		sx_xunlock(&ktls_init_lock);
490 		goto start;
491 	}
492 
493 	error = ktls_init();
494 	if (error == 0)
495 		state = 1;
496 	else
497 		state = -1;
498 	atomic_store_rel_int(&ktls_init_state, state);
499 	sx_xunlock(&ktls_init_lock);
500 	return (error);
501 }
502 
503 #if defined(INET) || defined(INET6)
504 static int
505 ktls_create_session(struct socket *so, struct tls_enable *en,
506     struct ktls_session **tlsp)
507 {
508 	struct ktls_session *tls;
509 	int error;
510 
511 	/* Only TLS 1.0 - 1.3 are supported. */
512 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
513 		return (EINVAL);
514 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
515 	    en->tls_vminor > TLS_MINOR_VER_THREE)
516 		return (EINVAL);
517 
518 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
519 		return (EINVAL);
520 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
521 		return (EINVAL);
522 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
523 		return (EINVAL);
524 
525 	/* All supported algorithms require a cipher key. */
526 	if (en->cipher_key_len == 0)
527 		return (EINVAL);
528 
529 	/* No flags are currently supported. */
530 	if (en->flags != 0)
531 		return (EINVAL);
532 
533 	/* Common checks for supported algorithms. */
534 	switch (en->cipher_algorithm) {
535 	case CRYPTO_AES_NIST_GCM_16:
536 		/*
537 		 * auth_algorithm isn't used, but permit GMAC values
538 		 * for compatibility.
539 		 */
540 		switch (en->auth_algorithm) {
541 		case 0:
542 #ifdef COMPAT_FREEBSD12
543 		/* XXX: Really 13.0-current COMPAT. */
544 		case CRYPTO_AES_128_NIST_GMAC:
545 		case CRYPTO_AES_192_NIST_GMAC:
546 		case CRYPTO_AES_256_NIST_GMAC:
547 #endif
548 			break;
549 		default:
550 			return (EINVAL);
551 		}
552 		if (en->auth_key_len != 0)
553 			return (EINVAL);
554 		switch (en->tls_vminor) {
555 		case TLS_MINOR_VER_TWO:
556 			if (en->iv_len != TLS_AEAD_GCM_LEN)
557 				return (EINVAL);
558 			break;
559 		case TLS_MINOR_VER_THREE:
560 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
561 				return (EINVAL);
562 			break;
563 		default:
564 			return (EINVAL);
565 		}
566 		break;
567 	case CRYPTO_AES_CBC:
568 		switch (en->auth_algorithm) {
569 		case CRYPTO_SHA1_HMAC:
570 			break;
571 		case CRYPTO_SHA2_256_HMAC:
572 		case CRYPTO_SHA2_384_HMAC:
573 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
574 				return (EINVAL);
575 			break;
576 		default:
577 			return (EINVAL);
578 		}
579 		if (en->auth_key_len == 0)
580 			return (EINVAL);
581 
582 		/*
583 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
584 		 * use explicit IVs.
585 		 */
586 		switch (en->tls_vminor) {
587 		case TLS_MINOR_VER_ZERO:
588 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
589 				return (EINVAL);
590 			break;
591 		case TLS_MINOR_VER_ONE:
592 		case TLS_MINOR_VER_TWO:
593 			/* Ignore any supplied IV. */
594 			en->iv_len = 0;
595 			break;
596 		default:
597 			return (EINVAL);
598 		}
599 		break;
600 	case CRYPTO_CHACHA20_POLY1305:
601 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
602 			return (EINVAL);
603 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
604 		    en->tls_vminor != TLS_MINOR_VER_THREE)
605 			return (EINVAL);
606 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
607 			return (EINVAL);
608 		break;
609 	default:
610 		return (EINVAL);
611 	}
612 
613 	error = ktls_start_kthreads();
614 	if (error != 0)
615 		return (error);
616 
617 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
618 
619 	counter_u64_add(ktls_offload_active, 1);
620 
621 	refcount_init(&tls->refcount, 1);
622 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
623 
624 	tls->wq_index = ktls_get_cpu(so);
625 
626 	tls->params.cipher_algorithm = en->cipher_algorithm;
627 	tls->params.auth_algorithm = en->auth_algorithm;
628 	tls->params.tls_vmajor = en->tls_vmajor;
629 	tls->params.tls_vminor = en->tls_vminor;
630 	tls->params.flags = en->flags;
631 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
632 
633 	/* Set the header and trailer lengths. */
634 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
635 	switch (en->cipher_algorithm) {
636 	case CRYPTO_AES_NIST_GCM_16:
637 		/*
638 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
639 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
640 		 */
641 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
642 			tls->params.tls_hlen += sizeof(uint64_t);
643 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
644 		tls->params.tls_bs = 1;
645 		break;
646 	case CRYPTO_AES_CBC:
647 		switch (en->auth_algorithm) {
648 		case CRYPTO_SHA1_HMAC:
649 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
650 				/* Implicit IV, no nonce. */
651 				tls->sequential_records = true;
652 				tls->next_seqno = be64dec(en->rec_seq);
653 				STAILQ_INIT(&tls->pending_records);
654 			} else {
655 				tls->params.tls_hlen += AES_BLOCK_LEN;
656 			}
657 			tls->params.tls_tlen = AES_BLOCK_LEN +
658 			    SHA1_HASH_LEN;
659 			break;
660 		case CRYPTO_SHA2_256_HMAC:
661 			tls->params.tls_hlen += AES_BLOCK_LEN;
662 			tls->params.tls_tlen = AES_BLOCK_LEN +
663 			    SHA2_256_HASH_LEN;
664 			break;
665 		case CRYPTO_SHA2_384_HMAC:
666 			tls->params.tls_hlen += AES_BLOCK_LEN;
667 			tls->params.tls_tlen = AES_BLOCK_LEN +
668 			    SHA2_384_HASH_LEN;
669 			break;
670 		default:
671 			panic("invalid hmac");
672 		}
673 		tls->params.tls_bs = AES_BLOCK_LEN;
674 		break;
675 	case CRYPTO_CHACHA20_POLY1305:
676 		/*
677 		 * Chacha20 uses a 12 byte implicit IV.
678 		 */
679 		tls->params.tls_tlen = POLY1305_HASH_LEN;
680 		tls->params.tls_bs = 1;
681 		break;
682 	default:
683 		panic("invalid cipher");
684 	}
685 
686 	/*
687 	 * TLS 1.3 includes optional padding which we do not support,
688 	 * and also puts the "real" record type at the end of the
689 	 * encrypted data.
690 	 */
691 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
692 		tls->params.tls_tlen += sizeof(uint8_t);
693 
694 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
695 	    ("TLS header length too long: %d", tls->params.tls_hlen));
696 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
697 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
698 
699 	if (en->auth_key_len != 0) {
700 		tls->params.auth_key_len = en->auth_key_len;
701 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
702 		    M_WAITOK);
703 		error = copyin(en->auth_key, tls->params.auth_key,
704 		    en->auth_key_len);
705 		if (error)
706 			goto out;
707 	}
708 
709 	tls->params.cipher_key_len = en->cipher_key_len;
710 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
711 	error = copyin(en->cipher_key, tls->params.cipher_key,
712 	    en->cipher_key_len);
713 	if (error)
714 		goto out;
715 
716 	/*
717 	 * This holds the implicit portion of the nonce for AEAD
718 	 * ciphers and the initial implicit IV for TLS 1.0.  The
719 	 * explicit portions of the IV are generated in ktls_frame().
720 	 */
721 	if (en->iv_len != 0) {
722 		tls->params.iv_len = en->iv_len;
723 		error = copyin(en->iv, tls->params.iv, en->iv_len);
724 		if (error)
725 			goto out;
726 
727 		/*
728 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
729 		 * counter to generate unique explicit IVs.
730 		 *
731 		 * Store this counter in the last 8 bytes of the IV
732 		 * array so that it is 8-byte aligned.
733 		 */
734 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
735 		    en->tls_vminor == TLS_MINOR_VER_TWO)
736 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
737 	}
738 
739 	*tlsp = tls;
740 	return (0);
741 
742 out:
743 	ktls_cleanup(tls);
744 	return (error);
745 }
746 
747 static struct ktls_session *
748 ktls_clone_session(struct ktls_session *tls)
749 {
750 	struct ktls_session *tls_new;
751 
752 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
753 
754 	counter_u64_add(ktls_offload_active, 1);
755 
756 	refcount_init(&tls_new->refcount, 1);
757 	TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
758 
759 	/* Copy fields from existing session. */
760 	tls_new->params = tls->params;
761 	tls_new->wq_index = tls->wq_index;
762 
763 	/* Deep copy keys. */
764 	if (tls_new->params.auth_key != NULL) {
765 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
766 		    M_KTLS, M_WAITOK);
767 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
768 		    tls->params.auth_key_len);
769 	}
770 
771 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
772 	    M_WAITOK);
773 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
774 	    tls->params.cipher_key_len);
775 
776 	return (tls_new);
777 }
778 #endif
779 
780 static void
781 ktls_cleanup(struct ktls_session *tls)
782 {
783 
784 	counter_u64_add(ktls_offload_active, -1);
785 	switch (tls->mode) {
786 	case TCP_TLS_MODE_SW:
787 		switch (tls->params.cipher_algorithm) {
788 		case CRYPTO_AES_CBC:
789 			counter_u64_add(ktls_sw_cbc, -1);
790 			break;
791 		case CRYPTO_AES_NIST_GCM_16:
792 			counter_u64_add(ktls_sw_gcm, -1);
793 			break;
794 		case CRYPTO_CHACHA20_POLY1305:
795 			counter_u64_add(ktls_sw_chacha20, -1);
796 			break;
797 		}
798 		break;
799 	case TCP_TLS_MODE_IFNET:
800 		switch (tls->params.cipher_algorithm) {
801 		case CRYPTO_AES_CBC:
802 			counter_u64_add(ktls_ifnet_cbc, -1);
803 			break;
804 		case CRYPTO_AES_NIST_GCM_16:
805 			counter_u64_add(ktls_ifnet_gcm, -1);
806 			break;
807 		case CRYPTO_CHACHA20_POLY1305:
808 			counter_u64_add(ktls_ifnet_chacha20, -1);
809 			break;
810 		}
811 		if (tls->snd_tag != NULL)
812 			m_snd_tag_rele(tls->snd_tag);
813 		break;
814 #ifdef TCP_OFFLOAD
815 	case TCP_TLS_MODE_TOE:
816 		switch (tls->params.cipher_algorithm) {
817 		case CRYPTO_AES_CBC:
818 			counter_u64_add(ktls_toe_cbc, -1);
819 			break;
820 		case CRYPTO_AES_NIST_GCM_16:
821 			counter_u64_add(ktls_toe_gcm, -1);
822 			break;
823 		case CRYPTO_CHACHA20_POLY1305:
824 			counter_u64_add(ktls_toe_chacha20, -1);
825 			break;
826 		}
827 		break;
828 #endif
829 	}
830 	if (tls->ocf_session != NULL)
831 		ktls_ocf_free(tls);
832 	if (tls->params.auth_key != NULL) {
833 		zfree(tls->params.auth_key, M_KTLS);
834 		tls->params.auth_key = NULL;
835 		tls->params.auth_key_len = 0;
836 	}
837 	if (tls->params.cipher_key != NULL) {
838 		zfree(tls->params.cipher_key, M_KTLS);
839 		tls->params.cipher_key = NULL;
840 		tls->params.cipher_key_len = 0;
841 	}
842 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
843 }
844 
845 #if defined(INET) || defined(INET6)
846 
847 #ifdef TCP_OFFLOAD
848 static int
849 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
850 {
851 	struct inpcb *inp;
852 	struct tcpcb *tp;
853 	int error;
854 
855 	inp = so->so_pcb;
856 	INP_WLOCK(inp);
857 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
858 		INP_WUNLOCK(inp);
859 		return (ECONNRESET);
860 	}
861 	if (inp->inp_socket == NULL) {
862 		INP_WUNLOCK(inp);
863 		return (ECONNRESET);
864 	}
865 	tp = intotcpcb(inp);
866 	if (!(tp->t_flags & TF_TOE)) {
867 		INP_WUNLOCK(inp);
868 		return (EOPNOTSUPP);
869 	}
870 
871 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
872 	INP_WUNLOCK(inp);
873 	if (error == 0) {
874 		tls->mode = TCP_TLS_MODE_TOE;
875 		switch (tls->params.cipher_algorithm) {
876 		case CRYPTO_AES_CBC:
877 			counter_u64_add(ktls_toe_cbc, 1);
878 			break;
879 		case CRYPTO_AES_NIST_GCM_16:
880 			counter_u64_add(ktls_toe_gcm, 1);
881 			break;
882 		case CRYPTO_CHACHA20_POLY1305:
883 			counter_u64_add(ktls_toe_chacha20, 1);
884 			break;
885 		}
886 	}
887 	return (error);
888 }
889 #endif
890 
891 /*
892  * Common code used when first enabling ifnet TLS on a connection or
893  * when allocating a new ifnet TLS session due to a routing change.
894  * This function allocates a new TLS send tag on whatever interface
895  * the connection is currently routed over.
896  */
897 static int
898 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
899     struct m_snd_tag **mstp)
900 {
901 	union if_snd_tag_alloc_params params;
902 	struct ifnet *ifp;
903 	struct nhop_object *nh;
904 	struct tcpcb *tp;
905 	int error;
906 
907 	INP_RLOCK(inp);
908 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
909 		INP_RUNLOCK(inp);
910 		return (ECONNRESET);
911 	}
912 	if (inp->inp_socket == NULL) {
913 		INP_RUNLOCK(inp);
914 		return (ECONNRESET);
915 	}
916 	tp = intotcpcb(inp);
917 
918 	/*
919 	 * Check administrative controls on ifnet TLS to determine if
920 	 * ifnet TLS should be denied.
921 	 *
922 	 * - Always permit 'force' requests.
923 	 * - ktls_ifnet_permitted == 0: always deny.
924 	 */
925 	if (!force && ktls_ifnet_permitted == 0) {
926 		INP_RUNLOCK(inp);
927 		return (ENXIO);
928 	}
929 
930 	/*
931 	 * XXX: Use the cached route in the inpcb to find the
932 	 * interface.  This should perhaps instead use
933 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
934 	 * enabled after a connection has completed key negotiation in
935 	 * userland, the cached route will be present in practice.
936 	 */
937 	nh = inp->inp_route.ro_nh;
938 	if (nh == NULL) {
939 		INP_RUNLOCK(inp);
940 		return (ENXIO);
941 	}
942 	ifp = nh->nh_ifp;
943 	if_ref(ifp);
944 
945 	/*
946 	 * Allocate a TLS + ratelimit tag if the connection has an
947 	 * existing pacing rate.
948 	 */
949 	if (tp->t_pacing_rate != -1 &&
950 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
951 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
952 		params.tls_rate_limit.inp = inp;
953 		params.tls_rate_limit.tls = tls;
954 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
955 	} else {
956 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
957 		params.tls.inp = inp;
958 		params.tls.tls = tls;
959 	}
960 	params.hdr.flowid = inp->inp_flowid;
961 	params.hdr.flowtype = inp->inp_flowtype;
962 	params.hdr.numa_domain = inp->inp_numa_domain;
963 	INP_RUNLOCK(inp);
964 
965 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
966 		error = EOPNOTSUPP;
967 		goto out;
968 	}
969 	if (inp->inp_vflag & INP_IPV6) {
970 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
971 			error = EOPNOTSUPP;
972 			goto out;
973 		}
974 	} else {
975 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
976 			error = EOPNOTSUPP;
977 			goto out;
978 		}
979 	}
980 	error = m_snd_tag_alloc(ifp, &params, mstp);
981 out:
982 	if_rele(ifp);
983 	return (error);
984 }
985 
986 static int
987 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
988 {
989 	struct m_snd_tag *mst;
990 	int error;
991 
992 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
993 	if (error == 0) {
994 		tls->mode = TCP_TLS_MODE_IFNET;
995 		tls->snd_tag = mst;
996 		switch (tls->params.cipher_algorithm) {
997 		case CRYPTO_AES_CBC:
998 			counter_u64_add(ktls_ifnet_cbc, 1);
999 			break;
1000 		case CRYPTO_AES_NIST_GCM_16:
1001 			counter_u64_add(ktls_ifnet_gcm, 1);
1002 			break;
1003 		case CRYPTO_CHACHA20_POLY1305:
1004 			counter_u64_add(ktls_ifnet_chacha20, 1);
1005 			break;
1006 		}
1007 	}
1008 	return (error);
1009 }
1010 
1011 static void
1012 ktls_use_sw(struct ktls_session *tls)
1013 {
1014 	tls->mode = TCP_TLS_MODE_SW;
1015 	switch (tls->params.cipher_algorithm) {
1016 	case CRYPTO_AES_CBC:
1017 		counter_u64_add(ktls_sw_cbc, 1);
1018 		break;
1019 	case CRYPTO_AES_NIST_GCM_16:
1020 		counter_u64_add(ktls_sw_gcm, 1);
1021 		break;
1022 	case CRYPTO_CHACHA20_POLY1305:
1023 		counter_u64_add(ktls_sw_chacha20, 1);
1024 		break;
1025 	}
1026 }
1027 
1028 static int
1029 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1030 {
1031 	int error;
1032 
1033 	error = ktls_ocf_try(so, tls, direction);
1034 	if (error)
1035 		return (error);
1036 	ktls_use_sw(tls);
1037 	return (0);
1038 }
1039 
1040 /*
1041  * KTLS RX stores data in the socket buffer as a list of TLS records,
1042  * where each record is stored as a control message containg the TLS
1043  * header followed by data mbufs containing the decrypted data.  This
1044  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1045  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1046  * should be queued to the socket buffer as records, but encrypted
1047  * data which needs to be decrypted by software arrives as a stream of
1048  * regular mbufs which need to be converted.  In addition, there may
1049  * already be pending encrypted data in the socket buffer when KTLS RX
1050  * is enabled.
1051  *
1052  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1053  * is used:
1054  *
1055  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1056  *
1057  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1058  *   from the first mbuf.  Once all of the data for that TLS record is
1059  *   queued, the socket is queued to a worker thread.
1060  *
1061  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1062  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1063  *   decrypted, and inserted into the regular socket buffer chain as
1064  *   record starting with a control message holding the TLS header and
1065  *   a chain of mbufs holding the encrypted data.
1066  */
1067 
1068 static void
1069 sb_mark_notready(struct sockbuf *sb)
1070 {
1071 	struct mbuf *m;
1072 
1073 	m = sb->sb_mb;
1074 	sb->sb_mtls = m;
1075 	sb->sb_mb = NULL;
1076 	sb->sb_mbtail = NULL;
1077 	sb->sb_lastrecord = NULL;
1078 	for (; m != NULL; m = m->m_next) {
1079 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1080 		    __func__));
1081 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1082 		    __func__));
1083 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1084 		    __func__));
1085 		m->m_flags |= M_NOTREADY;
1086 		sb->sb_acc -= m->m_len;
1087 		sb->sb_tlscc += m->m_len;
1088 		sb->sb_mtlstail = m;
1089 	}
1090 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1091 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1092 	    sb->sb_ccc));
1093 }
1094 
1095 /*
1096  * Return information about the pending TLS data in a socket
1097  * buffer.  On return, 'seqno' is set to the sequence number
1098  * of the next TLS record to be received, 'resid' is set to
1099  * the amount of bytes still needed for the last pending
1100  * record.  The function returns 'false' if the last pending
1101  * record contains a partial TLS header.  In that case, 'resid'
1102  * is the number of bytes needed to complete the TLS header.
1103  */
1104 bool
1105 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1106 {
1107 	struct tls_record_layer hdr;
1108 	struct mbuf *m;
1109 	uint64_t seqno;
1110 	size_t resid;
1111 	u_int offset, record_len;
1112 
1113 	SOCKBUF_LOCK_ASSERT(sb);
1114 	MPASS(sb->sb_flags & SB_TLS_RX);
1115 	seqno = sb->sb_tls_seqno;
1116 	resid = sb->sb_tlscc;
1117 	m = sb->sb_mtls;
1118 	offset = 0;
1119 
1120 	if (resid == 0) {
1121 		*seqnop = seqno;
1122 		*residp = 0;
1123 		return (true);
1124 	}
1125 
1126 	for (;;) {
1127 		seqno++;
1128 
1129 		if (resid < sizeof(hdr)) {
1130 			*seqnop = seqno;
1131 			*residp = sizeof(hdr) - resid;
1132 			return (false);
1133 		}
1134 
1135 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1136 
1137 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1138 		if (resid <= record_len) {
1139 			*seqnop = seqno;
1140 			*residp = record_len - resid;
1141 			return (true);
1142 		}
1143 		resid -= record_len;
1144 
1145 		while (record_len != 0) {
1146 			if (m->m_len - offset > record_len) {
1147 				offset += record_len;
1148 				break;
1149 			}
1150 
1151 			record_len -= (m->m_len - offset);
1152 			offset = 0;
1153 			m = m->m_next;
1154 		}
1155 	}
1156 }
1157 
1158 int
1159 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1160 {
1161 	struct ktls_session *tls;
1162 	int error;
1163 
1164 	if (!ktls_offload_enable)
1165 		return (ENOTSUP);
1166 	if (SOLISTENING(so))
1167 		return (EINVAL);
1168 
1169 	counter_u64_add(ktls_offload_enable_calls, 1);
1170 
1171 	/*
1172 	 * This should always be true since only the TCP socket option
1173 	 * invokes this function.
1174 	 */
1175 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1176 		return (EINVAL);
1177 
1178 	/*
1179 	 * XXX: Don't overwrite existing sessions.  We should permit
1180 	 * this to support rekeying in the future.
1181 	 */
1182 	if (so->so_rcv.sb_tls_info != NULL)
1183 		return (EALREADY);
1184 
1185 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1186 		return (ENOTSUP);
1187 
1188 	error = ktls_create_session(so, en, &tls);
1189 	if (error)
1190 		return (error);
1191 
1192 	error = ktls_ocf_try(so, tls, KTLS_RX);
1193 	if (error) {
1194 		ktls_cleanup(tls);
1195 		return (error);
1196 	}
1197 
1198 #ifdef TCP_OFFLOAD
1199 	error = ktls_try_toe(so, tls, KTLS_RX);
1200 	if (error)
1201 #endif
1202 		ktls_use_sw(tls);
1203 
1204 	/* Mark the socket as using TLS offload. */
1205 	SOCKBUF_LOCK(&so->so_rcv);
1206 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1207 	so->so_rcv.sb_tls_info = tls;
1208 	so->so_rcv.sb_flags |= SB_TLS_RX;
1209 
1210 	/* Mark existing data as not ready until it can be decrypted. */
1211 	if (tls->mode != TCP_TLS_MODE_TOE) {
1212 		sb_mark_notready(&so->so_rcv);
1213 		ktls_check_rx(&so->so_rcv);
1214 	}
1215 	SOCKBUF_UNLOCK(&so->so_rcv);
1216 
1217 	counter_u64_add(ktls_offload_total, 1);
1218 
1219 	return (0);
1220 }
1221 
1222 int
1223 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1224 {
1225 	struct ktls_session *tls;
1226 	struct inpcb *inp;
1227 	int error;
1228 
1229 	if (!ktls_offload_enable)
1230 		return (ENOTSUP);
1231 	if (SOLISTENING(so))
1232 		return (EINVAL);
1233 
1234 	counter_u64_add(ktls_offload_enable_calls, 1);
1235 
1236 	/*
1237 	 * This should always be true since only the TCP socket option
1238 	 * invokes this function.
1239 	 */
1240 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1241 		return (EINVAL);
1242 
1243 	/*
1244 	 * XXX: Don't overwrite existing sessions.  We should permit
1245 	 * this to support rekeying in the future.
1246 	 */
1247 	if (so->so_snd.sb_tls_info != NULL)
1248 		return (EALREADY);
1249 
1250 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1251 		return (ENOTSUP);
1252 
1253 	/* TLS requires ext pgs */
1254 	if (mb_use_ext_pgs == 0)
1255 		return (ENXIO);
1256 
1257 	error = ktls_create_session(so, en, &tls);
1258 	if (error)
1259 		return (error);
1260 
1261 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1262 #ifdef TCP_OFFLOAD
1263 	error = ktls_try_toe(so, tls, KTLS_TX);
1264 	if (error)
1265 #endif
1266 		error = ktls_try_ifnet(so, tls, false);
1267 	if (error)
1268 		error = ktls_try_sw(so, tls, KTLS_TX);
1269 
1270 	if (error) {
1271 		ktls_cleanup(tls);
1272 		return (error);
1273 	}
1274 
1275 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1276 	if (error) {
1277 		ktls_cleanup(tls);
1278 		return (error);
1279 	}
1280 
1281 	/*
1282 	 * Write lock the INP when setting sb_tls_info so that
1283 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1284 	 * holding the INP lock.
1285 	 */
1286 	inp = so->so_pcb;
1287 	INP_WLOCK(inp);
1288 	SOCKBUF_LOCK(&so->so_snd);
1289 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1290 	so->so_snd.sb_tls_info = tls;
1291 	if (tls->mode != TCP_TLS_MODE_SW)
1292 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1293 	SOCKBUF_UNLOCK(&so->so_snd);
1294 	INP_WUNLOCK(inp);
1295 	SOCK_IO_SEND_UNLOCK(so);
1296 
1297 	counter_u64_add(ktls_offload_total, 1);
1298 
1299 	return (0);
1300 }
1301 
1302 int
1303 ktls_get_rx_mode(struct socket *so, int *modep)
1304 {
1305 	struct ktls_session *tls;
1306 	struct inpcb *inp __diagused;
1307 
1308 	if (SOLISTENING(so))
1309 		return (EINVAL);
1310 	inp = so->so_pcb;
1311 	INP_WLOCK_ASSERT(inp);
1312 	SOCK_RECVBUF_LOCK(so);
1313 	tls = so->so_rcv.sb_tls_info;
1314 	if (tls == NULL)
1315 		*modep = TCP_TLS_MODE_NONE;
1316 	else
1317 		*modep = tls->mode;
1318 	SOCK_RECVBUF_UNLOCK(so);
1319 	return (0);
1320 }
1321 
1322 int
1323 ktls_get_tx_mode(struct socket *so, int *modep)
1324 {
1325 	struct ktls_session *tls;
1326 	struct inpcb *inp __diagused;
1327 
1328 	if (SOLISTENING(so))
1329 		return (EINVAL);
1330 	inp = so->so_pcb;
1331 	INP_WLOCK_ASSERT(inp);
1332 	SOCK_SENDBUF_LOCK(so);
1333 	tls = so->so_snd.sb_tls_info;
1334 	if (tls == NULL)
1335 		*modep = TCP_TLS_MODE_NONE;
1336 	else
1337 		*modep = tls->mode;
1338 	SOCK_SENDBUF_UNLOCK(so);
1339 	return (0);
1340 }
1341 
1342 /*
1343  * Switch between SW and ifnet TLS sessions as requested.
1344  */
1345 int
1346 ktls_set_tx_mode(struct socket *so, int mode)
1347 {
1348 	struct ktls_session *tls, *tls_new;
1349 	struct inpcb *inp;
1350 	int error;
1351 
1352 	if (SOLISTENING(so))
1353 		return (EINVAL);
1354 	switch (mode) {
1355 	case TCP_TLS_MODE_SW:
1356 	case TCP_TLS_MODE_IFNET:
1357 		break;
1358 	default:
1359 		return (EINVAL);
1360 	}
1361 
1362 	inp = so->so_pcb;
1363 	INP_WLOCK_ASSERT(inp);
1364 	SOCKBUF_LOCK(&so->so_snd);
1365 	tls = so->so_snd.sb_tls_info;
1366 	if (tls == NULL) {
1367 		SOCKBUF_UNLOCK(&so->so_snd);
1368 		return (0);
1369 	}
1370 
1371 	if (tls->mode == mode) {
1372 		SOCKBUF_UNLOCK(&so->so_snd);
1373 		return (0);
1374 	}
1375 
1376 	tls = ktls_hold(tls);
1377 	SOCKBUF_UNLOCK(&so->so_snd);
1378 	INP_WUNLOCK(inp);
1379 
1380 	tls_new = ktls_clone_session(tls);
1381 
1382 	if (mode == TCP_TLS_MODE_IFNET)
1383 		error = ktls_try_ifnet(so, tls_new, true);
1384 	else
1385 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1386 	if (error) {
1387 		counter_u64_add(ktls_switch_failed, 1);
1388 		ktls_free(tls_new);
1389 		ktls_free(tls);
1390 		INP_WLOCK(inp);
1391 		return (error);
1392 	}
1393 
1394 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1395 	if (error) {
1396 		counter_u64_add(ktls_switch_failed, 1);
1397 		ktls_free(tls_new);
1398 		ktls_free(tls);
1399 		INP_WLOCK(inp);
1400 		return (error);
1401 	}
1402 
1403 	/*
1404 	 * If we raced with another session change, keep the existing
1405 	 * session.
1406 	 */
1407 	if (tls != so->so_snd.sb_tls_info) {
1408 		counter_u64_add(ktls_switch_failed, 1);
1409 		SOCK_IO_SEND_UNLOCK(so);
1410 		ktls_free(tls_new);
1411 		ktls_free(tls);
1412 		INP_WLOCK(inp);
1413 		return (EBUSY);
1414 	}
1415 
1416 	SOCKBUF_LOCK(&so->so_snd);
1417 	so->so_snd.sb_tls_info = tls_new;
1418 	if (tls_new->mode != TCP_TLS_MODE_SW)
1419 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1420 	SOCKBUF_UNLOCK(&so->so_snd);
1421 	SOCK_IO_SEND_UNLOCK(so);
1422 
1423 	/*
1424 	 * Drop two references on 'tls'.  The first is for the
1425 	 * ktls_hold() above.  The second drops the reference from the
1426 	 * socket buffer.
1427 	 */
1428 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1429 	ktls_free(tls);
1430 	ktls_free(tls);
1431 
1432 	if (mode == TCP_TLS_MODE_IFNET)
1433 		counter_u64_add(ktls_switch_to_ifnet, 1);
1434 	else
1435 		counter_u64_add(ktls_switch_to_sw, 1);
1436 
1437 	INP_WLOCK(inp);
1438 	return (0);
1439 }
1440 
1441 /*
1442  * Try to allocate a new TLS send tag.  This task is scheduled when
1443  * ip_output detects a route change while trying to transmit a packet
1444  * holding a TLS record.  If a new tag is allocated, replace the tag
1445  * in the TLS session.  Subsequent packets on the connection will use
1446  * the new tag.  If a new tag cannot be allocated, drop the
1447  * connection.
1448  */
1449 static void
1450 ktls_reset_send_tag(void *context, int pending)
1451 {
1452 	struct epoch_tracker et;
1453 	struct ktls_session *tls;
1454 	struct m_snd_tag *old, *new;
1455 	struct inpcb *inp;
1456 	struct tcpcb *tp;
1457 	int error;
1458 
1459 	MPASS(pending == 1);
1460 
1461 	tls = context;
1462 	inp = tls->inp;
1463 
1464 	/*
1465 	 * Free the old tag first before allocating a new one.
1466 	 * ip[6]_output_send() will treat a NULL send tag the same as
1467 	 * an ifp mismatch and drop packets until a new tag is
1468 	 * allocated.
1469 	 *
1470 	 * Write-lock the INP when changing tls->snd_tag since
1471 	 * ip[6]_output_send() holds a read-lock when reading the
1472 	 * pointer.
1473 	 */
1474 	INP_WLOCK(inp);
1475 	old = tls->snd_tag;
1476 	tls->snd_tag = NULL;
1477 	INP_WUNLOCK(inp);
1478 	if (old != NULL)
1479 		m_snd_tag_rele(old);
1480 
1481 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1482 
1483 	if (error == 0) {
1484 		INP_WLOCK(inp);
1485 		tls->snd_tag = new;
1486 		mtx_pool_lock(mtxpool_sleep, tls);
1487 		tls->reset_pending = false;
1488 		mtx_pool_unlock(mtxpool_sleep, tls);
1489 		if (!in_pcbrele_wlocked(inp))
1490 			INP_WUNLOCK(inp);
1491 
1492 		counter_u64_add(ktls_ifnet_reset, 1);
1493 
1494 		/*
1495 		 * XXX: Should we kick tcp_output explicitly now that
1496 		 * the send tag is fixed or just rely on timers?
1497 		 */
1498 	} else {
1499 		NET_EPOCH_ENTER(et);
1500 		INP_WLOCK(inp);
1501 		if (!in_pcbrele_wlocked(inp)) {
1502 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1503 			    !(inp->inp_flags & INP_DROPPED)) {
1504 				tp = intotcpcb(inp);
1505 				CURVNET_SET(tp->t_vnet);
1506 				tp = tcp_drop(tp, ECONNABORTED);
1507 				CURVNET_RESTORE();
1508 				if (tp != NULL)
1509 					INP_WUNLOCK(inp);
1510 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1511 			} else
1512 				INP_WUNLOCK(inp);
1513 		}
1514 		NET_EPOCH_EXIT(et);
1515 
1516 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1517 
1518 		/*
1519 		 * Leave reset_pending true to avoid future tasks while
1520 		 * the socket goes away.
1521 		 */
1522 	}
1523 
1524 	ktls_free(tls);
1525 }
1526 
1527 int
1528 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1529 {
1530 
1531 	if (inp == NULL)
1532 		return (ENOBUFS);
1533 
1534 	INP_LOCK_ASSERT(inp);
1535 
1536 	/*
1537 	 * See if we should schedule a task to update the send tag for
1538 	 * this session.
1539 	 */
1540 	mtx_pool_lock(mtxpool_sleep, tls);
1541 	if (!tls->reset_pending) {
1542 		(void) ktls_hold(tls);
1543 		in_pcbref(inp);
1544 		tls->inp = inp;
1545 		tls->reset_pending = true;
1546 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1547 	}
1548 	mtx_pool_unlock(mtxpool_sleep, tls);
1549 	return (ENOBUFS);
1550 }
1551 
1552 #ifdef RATELIMIT
1553 int
1554 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1555 {
1556 	union if_snd_tag_modify_params params = {
1557 		.rate_limit.max_rate = max_pacing_rate,
1558 		.rate_limit.flags = M_NOWAIT,
1559 	};
1560 	struct m_snd_tag *mst;
1561 
1562 	/* Can't get to the inp, but it should be locked. */
1563 	/* INP_LOCK_ASSERT(inp); */
1564 
1565 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1566 
1567 	if (tls->snd_tag == NULL) {
1568 		/*
1569 		 * Resetting send tag, ignore this change.  The
1570 		 * pending reset may or may not see this updated rate
1571 		 * in the tcpcb.  If it doesn't, we will just lose
1572 		 * this rate change.
1573 		 */
1574 		return (0);
1575 	}
1576 
1577 	MPASS(tls->snd_tag != NULL);
1578 	MPASS(tls->snd_tag->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1579 
1580 	mst = tls->snd_tag;
1581 	return (mst->sw->snd_tag_modify(mst, &params));
1582 }
1583 #endif
1584 #endif
1585 
1586 void
1587 ktls_destroy(struct ktls_session *tls)
1588 {
1589 
1590 	if (tls->sequential_records) {
1591 		struct mbuf *m, *n;
1592 		int page_count;
1593 
1594 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1595 			page_count = m->m_epg_enc_cnt;
1596 			while (page_count > 0) {
1597 				KASSERT(page_count >= m->m_epg_nrdy,
1598 				    ("%s: too few pages", __func__));
1599 				page_count -= m->m_epg_nrdy;
1600 				m = m_free(m);
1601 			}
1602 		}
1603 	}
1604 	ktls_cleanup(tls);
1605 	uma_zfree(ktls_session_zone, tls);
1606 }
1607 
1608 void
1609 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1610 {
1611 
1612 	for (; m != NULL; m = m->m_next) {
1613 		KASSERT((m->m_flags & M_EXTPG) != 0,
1614 		    ("ktls_seq: mapped mbuf %p", m));
1615 
1616 		m->m_epg_seqno = sb->sb_tls_seqno;
1617 		sb->sb_tls_seqno++;
1618 	}
1619 }
1620 
1621 /*
1622  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1623  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1624  * mbuf must be populated with the payload of each TLS record.
1625  *
1626  * The record_type argument specifies the TLS record type used when
1627  * populating the TLS header.
1628  *
1629  * The enq_count argument on return is set to the number of pages of
1630  * payload data for this entire chain that need to be encrypted via SW
1631  * encryption.  The returned value should be passed to ktls_enqueue
1632  * when scheduling encryption of this chain of mbufs.  To handle the
1633  * special case of empty fragments for TLS 1.0 sessions, an empty
1634  * fragment counts as one page.
1635  */
1636 void
1637 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1638     uint8_t record_type)
1639 {
1640 	struct tls_record_layer *tlshdr;
1641 	struct mbuf *m;
1642 	uint64_t *noncep;
1643 	uint16_t tls_len;
1644 	int maxlen __diagused;
1645 
1646 	maxlen = tls->params.max_frame_len;
1647 	*enq_cnt = 0;
1648 	for (m = top; m != NULL; m = m->m_next) {
1649 		/*
1650 		 * All mbufs in the chain should be TLS records whose
1651 		 * payload does not exceed the maximum frame length.
1652 		 *
1653 		 * Empty TLS records are permitted when using CBC.
1654 		 */
1655 		KASSERT(m->m_len <= maxlen &&
1656 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1657 		    m->m_len >= 0 : m->m_len > 0),
1658 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1659 
1660 		/*
1661 		 * TLS frames require unmapped mbufs to store session
1662 		 * info.
1663 		 */
1664 		KASSERT((m->m_flags & M_EXTPG) != 0,
1665 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1666 
1667 		tls_len = m->m_len;
1668 
1669 		/* Save a reference to the session. */
1670 		m->m_epg_tls = ktls_hold(tls);
1671 
1672 		m->m_epg_hdrlen = tls->params.tls_hlen;
1673 		m->m_epg_trllen = tls->params.tls_tlen;
1674 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1675 			int bs, delta;
1676 
1677 			/*
1678 			 * AES-CBC pads messages to a multiple of the
1679 			 * block size.  Note that the padding is
1680 			 * applied after the digest and the encryption
1681 			 * is done on the "plaintext || mac || padding".
1682 			 * At least one byte of padding is always
1683 			 * present.
1684 			 *
1685 			 * Compute the final trailer length assuming
1686 			 * at most one block of padding.
1687 			 * tls->params.tls_tlen is the maximum
1688 			 * possible trailer length (padding + digest).
1689 			 * delta holds the number of excess padding
1690 			 * bytes if the maximum were used.  Those
1691 			 * extra bytes are removed.
1692 			 */
1693 			bs = tls->params.tls_bs;
1694 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1695 			m->m_epg_trllen -= delta;
1696 		}
1697 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1698 
1699 		/* Populate the TLS header. */
1700 		tlshdr = (void *)m->m_epg_hdr;
1701 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1702 
1703 		/*
1704 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1705 		 * of TLS_RLTYPE_APP.
1706 		 */
1707 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1708 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1709 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1710 			tlshdr->tls_type = TLS_RLTYPE_APP;
1711 			/* save the real record type for later */
1712 			m->m_epg_record_type = record_type;
1713 			m->m_epg_trail[0] = record_type;
1714 		} else {
1715 			tlshdr->tls_vminor = tls->params.tls_vminor;
1716 			tlshdr->tls_type = record_type;
1717 		}
1718 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1719 
1720 		/*
1721 		 * Store nonces / explicit IVs after the end of the
1722 		 * TLS header.
1723 		 *
1724 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1725 		 * from the end of the IV.  The nonce is then
1726 		 * incremented for use by the next record.
1727 		 *
1728 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1729 		 */
1730 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1731 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1732 			noncep = (uint64_t *)(tls->params.iv + 8);
1733 			be64enc(tlshdr + 1, *noncep);
1734 			(*noncep)++;
1735 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1736 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1737 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1738 
1739 		/*
1740 		 * When using SW encryption, mark the mbuf not ready.
1741 		 * It will be marked ready via sbready() after the
1742 		 * record has been encrypted.
1743 		 *
1744 		 * When using ifnet TLS, unencrypted TLS records are
1745 		 * sent down the stack to the NIC.
1746 		 */
1747 		if (tls->mode == TCP_TLS_MODE_SW) {
1748 			m->m_flags |= M_NOTREADY;
1749 			if (__predict_false(tls_len == 0)) {
1750 				/* TLS 1.0 empty fragment. */
1751 				m->m_epg_nrdy = 1;
1752 			} else
1753 				m->m_epg_nrdy = m->m_epg_npgs;
1754 			*enq_cnt += m->m_epg_nrdy;
1755 		}
1756 	}
1757 }
1758 
1759 void
1760 ktls_check_rx(struct sockbuf *sb)
1761 {
1762 	struct tls_record_layer hdr;
1763 	struct ktls_wq *wq;
1764 	struct socket *so;
1765 	bool running;
1766 
1767 	SOCKBUF_LOCK_ASSERT(sb);
1768 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1769 	    __func__, sb));
1770 	so = __containerof(sb, struct socket, so_rcv);
1771 
1772 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1773 		return;
1774 
1775 	/* Is there enough queued for a TLS header? */
1776 	if (sb->sb_tlscc < sizeof(hdr)) {
1777 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1778 			so->so_error = EMSGSIZE;
1779 		return;
1780 	}
1781 
1782 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1783 
1784 	/* Is the entire record queued? */
1785 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1786 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1787 			so->so_error = EMSGSIZE;
1788 		return;
1789 	}
1790 
1791 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1792 
1793 	soref(so);
1794 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1795 	mtx_lock(&wq->mtx);
1796 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1797 	running = wq->running;
1798 	mtx_unlock(&wq->mtx);
1799 	if (!running)
1800 		wakeup(wq);
1801 	counter_u64_add(ktls_cnt_rx_queued, 1);
1802 }
1803 
1804 static struct mbuf *
1805 ktls_detach_record(struct sockbuf *sb, int len)
1806 {
1807 	struct mbuf *m, *n, *top;
1808 	int remain;
1809 
1810 	SOCKBUF_LOCK_ASSERT(sb);
1811 	MPASS(len <= sb->sb_tlscc);
1812 
1813 	/*
1814 	 * If TLS chain is the exact size of the record,
1815 	 * just grab the whole record.
1816 	 */
1817 	top = sb->sb_mtls;
1818 	if (sb->sb_tlscc == len) {
1819 		sb->sb_mtls = NULL;
1820 		sb->sb_mtlstail = NULL;
1821 		goto out;
1822 	}
1823 
1824 	/*
1825 	 * While it would be nice to use m_split() here, we need
1826 	 * to know exactly what m_split() allocates to update the
1827 	 * accounting, so do it inline instead.
1828 	 */
1829 	remain = len;
1830 	for (m = top; remain > m->m_len; m = m->m_next)
1831 		remain -= m->m_len;
1832 
1833 	/* Easy case: don't have to split 'm'. */
1834 	if (remain == m->m_len) {
1835 		sb->sb_mtls = m->m_next;
1836 		if (sb->sb_mtls == NULL)
1837 			sb->sb_mtlstail = NULL;
1838 		m->m_next = NULL;
1839 		goto out;
1840 	}
1841 
1842 	/*
1843 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1844 	 * with M_NOWAIT first.
1845 	 */
1846 	n = m_get(M_NOWAIT, MT_DATA);
1847 	if (n == NULL) {
1848 		/*
1849 		 * Use M_WAITOK with socket buffer unlocked.  If
1850 		 * 'sb_mtls' changes while the lock is dropped, return
1851 		 * NULL to force the caller to retry.
1852 		 */
1853 		SOCKBUF_UNLOCK(sb);
1854 
1855 		n = m_get(M_WAITOK, MT_DATA);
1856 
1857 		SOCKBUF_LOCK(sb);
1858 		if (sb->sb_mtls != top) {
1859 			m_free(n);
1860 			return (NULL);
1861 		}
1862 	}
1863 	n->m_flags |= M_NOTREADY;
1864 
1865 	/* Store remainder in 'n'. */
1866 	n->m_len = m->m_len - remain;
1867 	if (m->m_flags & M_EXT) {
1868 		n->m_data = m->m_data + remain;
1869 		mb_dupcl(n, m);
1870 	} else {
1871 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1872 	}
1873 
1874 	/* Trim 'm' and update accounting. */
1875 	m->m_len -= n->m_len;
1876 	sb->sb_tlscc -= n->m_len;
1877 	sb->sb_ccc -= n->m_len;
1878 
1879 	/* Account for 'n'. */
1880 	sballoc_ktls_rx(sb, n);
1881 
1882 	/* Insert 'n' into the TLS chain. */
1883 	sb->sb_mtls = n;
1884 	n->m_next = m->m_next;
1885 	if (sb->sb_mtlstail == m)
1886 		sb->sb_mtlstail = n;
1887 
1888 	/* Detach the record from the TLS chain. */
1889 	m->m_next = NULL;
1890 
1891 out:
1892 	MPASS(m_length(top, NULL) == len);
1893 	for (m = top; m != NULL; m = m->m_next)
1894 		sbfree_ktls_rx(sb, m);
1895 	sb->sb_tlsdcc = len;
1896 	sb->sb_ccc += len;
1897 	SBCHECK(sb);
1898 	return (top);
1899 }
1900 
1901 /*
1902  * Determine the length of the trailing zero padding and find the real
1903  * record type in the byte before the padding.
1904  *
1905  * Walking the mbuf chain backwards is clumsy, so another option would
1906  * be to scan forwards remembering the last non-zero byte before the
1907  * trailer.  However, it would be expensive to scan the entire record.
1908  * Instead, find the last non-zero byte of each mbuf in the chain
1909  * keeping track of the relative offset of that nonzero byte.
1910  *
1911  * trail_len is the size of the MAC/tag on input and is set to the
1912  * size of the full trailer including padding and the record type on
1913  * return.
1914  */
1915 static int
1916 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
1917     int *trailer_len, uint8_t *record_typep)
1918 {
1919 	char *cp;
1920 	u_int digest_start, last_offset, m_len, offset;
1921 	uint8_t record_type;
1922 
1923 	digest_start = tls_len - *trailer_len;
1924 	last_offset = 0;
1925 	offset = 0;
1926 	for (; m != NULL && offset < digest_start;
1927 	     offset += m->m_len, m = m->m_next) {
1928 		/* Don't look for padding in the tag. */
1929 		m_len = min(digest_start - offset, m->m_len);
1930 		cp = mtod(m, char *);
1931 
1932 		/* Find last non-zero byte in this mbuf. */
1933 		while (m_len > 0 && cp[m_len - 1] == 0)
1934 			m_len--;
1935 		if (m_len > 0) {
1936 			record_type = cp[m_len - 1];
1937 			last_offset = offset + m_len;
1938 		}
1939 	}
1940 	if (last_offset < tls->params.tls_hlen)
1941 		return (EBADMSG);
1942 
1943 	*record_typep = record_type;
1944 	*trailer_len = tls_len - last_offset + 1;
1945 	return (0);
1946 }
1947 
1948 static void
1949 ktls_decrypt(struct socket *so)
1950 {
1951 	char tls_header[MBUF_PEXT_HDR_LEN];
1952 	struct ktls_session *tls;
1953 	struct sockbuf *sb;
1954 	struct tls_record_layer *hdr;
1955 	struct tls_get_record tgr;
1956 	struct mbuf *control, *data, *m;
1957 	uint64_t seqno;
1958 	int error, remain, tls_len, trail_len;
1959 	bool tls13;
1960 	uint8_t vminor, record_type;
1961 
1962 	hdr = (struct tls_record_layer *)tls_header;
1963 	sb = &so->so_rcv;
1964 	SOCKBUF_LOCK(sb);
1965 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1966 	    ("%s: socket %p not running", __func__, so));
1967 
1968 	tls = sb->sb_tls_info;
1969 	MPASS(tls != NULL);
1970 
1971 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
1972 	if (tls13)
1973 		vminor = TLS_MINOR_VER_TWO;
1974 	else
1975 		vminor = tls->params.tls_vminor;
1976 	for (;;) {
1977 		/* Is there enough queued for a TLS header? */
1978 		if (sb->sb_tlscc < tls->params.tls_hlen)
1979 			break;
1980 
1981 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1982 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1983 
1984 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1985 		    hdr->tls_vminor != vminor)
1986 			error = EINVAL;
1987 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
1988 			error = EINVAL;
1989 		else if (tls_len < tls->params.tls_hlen || tls_len >
1990 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1991 		    tls->params.tls_tlen)
1992 			error = EMSGSIZE;
1993 		else
1994 			error = 0;
1995 		if (__predict_false(error != 0)) {
1996 			/*
1997 			 * We have a corrupted record and are likely
1998 			 * out of sync.  The connection isn't
1999 			 * recoverable at this point, so abort it.
2000 			 */
2001 			SOCKBUF_UNLOCK(sb);
2002 			counter_u64_add(ktls_offload_corrupted_records, 1);
2003 
2004 			CURVNET_SET(so->so_vnet);
2005 			so->so_proto->pr_usrreqs->pru_abort(so);
2006 			so->so_error = error;
2007 			CURVNET_RESTORE();
2008 			goto deref;
2009 		}
2010 
2011 		/* Is the entire record queued? */
2012 		if (sb->sb_tlscc < tls_len)
2013 			break;
2014 
2015 		/*
2016 		 * Split out the portion of the mbuf chain containing
2017 		 * this TLS record.
2018 		 */
2019 		data = ktls_detach_record(sb, tls_len);
2020 		if (data == NULL)
2021 			continue;
2022 		MPASS(sb->sb_tlsdcc == tls_len);
2023 
2024 		seqno = sb->sb_tls_seqno;
2025 		sb->sb_tls_seqno++;
2026 		SBCHECK(sb);
2027 		SOCKBUF_UNLOCK(sb);
2028 
2029 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
2030 		if (error == 0) {
2031 			if (tls13)
2032 				error = tls13_find_record_type(tls, data,
2033 				    tls_len, &trail_len, &record_type);
2034 			else
2035 				record_type = hdr->tls_type;
2036 		}
2037 		if (error) {
2038 			counter_u64_add(ktls_offload_failed_crypto, 1);
2039 
2040 			SOCKBUF_LOCK(sb);
2041 			if (sb->sb_tlsdcc == 0) {
2042 				/*
2043 				 * sbcut/drop/flush discarded these
2044 				 * mbufs.
2045 				 */
2046 				m_freem(data);
2047 				break;
2048 			}
2049 
2050 			/*
2051 			 * Drop this TLS record's data, but keep
2052 			 * decrypting subsequent records.
2053 			 */
2054 			sb->sb_ccc -= tls_len;
2055 			sb->sb_tlsdcc = 0;
2056 
2057 			CURVNET_SET(so->so_vnet);
2058 			so->so_error = EBADMSG;
2059 			sorwakeup_locked(so);
2060 			CURVNET_RESTORE();
2061 
2062 			m_freem(data);
2063 
2064 			SOCKBUF_LOCK(sb);
2065 			continue;
2066 		}
2067 
2068 		/* Allocate the control mbuf. */
2069 		tgr.tls_type = record_type;
2070 		tgr.tls_vmajor = hdr->tls_vmajor;
2071 		tgr.tls_vminor = hdr->tls_vminor;
2072 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2073 		    trail_len);
2074 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
2075 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2076 
2077 		SOCKBUF_LOCK(sb);
2078 		if (sb->sb_tlsdcc == 0) {
2079 			/* sbcut/drop/flush discarded these mbufs. */
2080 			MPASS(sb->sb_tlscc == 0);
2081 			m_freem(data);
2082 			m_freem(control);
2083 			break;
2084 		}
2085 
2086 		/*
2087 		 * Clear the 'dcc' accounting in preparation for
2088 		 * adding the decrypted record.
2089 		 */
2090 		sb->sb_ccc -= tls_len;
2091 		sb->sb_tlsdcc = 0;
2092 		SBCHECK(sb);
2093 
2094 		/* If there is no payload, drop all of the data. */
2095 		if (tgr.tls_length == htobe16(0)) {
2096 			m_freem(data);
2097 			data = NULL;
2098 		} else {
2099 			/* Trim header. */
2100 			remain = tls->params.tls_hlen;
2101 			while (remain > 0) {
2102 				if (data->m_len > remain) {
2103 					data->m_data += remain;
2104 					data->m_len -= remain;
2105 					break;
2106 				}
2107 				remain -= data->m_len;
2108 				data = m_free(data);
2109 			}
2110 
2111 			/* Trim trailer and clear M_NOTREADY. */
2112 			remain = be16toh(tgr.tls_length);
2113 			m = data;
2114 			for (m = data; remain > m->m_len; m = m->m_next) {
2115 				m->m_flags &= ~M_NOTREADY;
2116 				remain -= m->m_len;
2117 			}
2118 			m->m_len = remain;
2119 			m_freem(m->m_next);
2120 			m->m_next = NULL;
2121 			m->m_flags &= ~M_NOTREADY;
2122 
2123 			/* Set EOR on the final mbuf. */
2124 			m->m_flags |= M_EOR;
2125 		}
2126 
2127 		sbappendcontrol_locked(sb, data, control, 0);
2128 	}
2129 
2130 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2131 
2132 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2133 		so->so_error = EMSGSIZE;
2134 
2135 	sorwakeup_locked(so);
2136 
2137 deref:
2138 	SOCKBUF_UNLOCK_ASSERT(sb);
2139 
2140 	CURVNET_SET(so->so_vnet);
2141 	sorele(so);
2142 	CURVNET_RESTORE();
2143 }
2144 
2145 void
2146 ktls_enqueue_to_free(struct mbuf *m)
2147 {
2148 	struct ktls_wq *wq;
2149 	bool running;
2150 
2151 	/* Mark it for freeing. */
2152 	m->m_epg_flags |= EPG_FLAG_2FREE;
2153 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2154 	mtx_lock(&wq->mtx);
2155 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2156 	running = wq->running;
2157 	mtx_unlock(&wq->mtx);
2158 	if (!running)
2159 		wakeup(wq);
2160 }
2161 
2162 static void *
2163 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2164 {
2165 	void *buf;
2166 	int domain, running;
2167 
2168 	if (m->m_epg_npgs <= 2)
2169 		return (NULL);
2170 	if (ktls_buffer_zone == NULL)
2171 		return (NULL);
2172 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2173 		/*
2174 		 * Rate-limit allocation attempts after a failure.
2175 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2176 		 * the free page queues and may fail consistently if memory is
2177 		 * fragmented.
2178 		 */
2179 		return (NULL);
2180 	}
2181 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2182 	if (buf == NULL) {
2183 		domain = PCPU_GET(domain);
2184 		wq->lastallocfail = ticks;
2185 
2186 		/*
2187 		 * Note that this check is "racy", but the races are
2188 		 * harmless, and are either a spurious wakeup if
2189 		 * multiple threads fail allocations before the alloc
2190 		 * thread wakes, or waiting an extra second in case we
2191 		 * see an old value of running == true.
2192 		 */
2193 		if (!VM_DOMAIN_EMPTY(domain)) {
2194 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2195 			if (!running)
2196 				wakeup(&ktls_domains[domain].alloc_td);
2197 		}
2198 	}
2199 	return (buf);
2200 }
2201 
2202 static int
2203 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2204     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2205 {
2206 	vm_page_t pg;
2207 	int error, i, len, off;
2208 
2209 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2210 	    ("%p not unready & nomap mbuf\n", m));
2211 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2212 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2213 	    ktls_maxlen));
2214 
2215 	/* Anonymous mbufs are encrypted in place. */
2216 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2217 		return (tls->sw_encrypt(state, tls, m, NULL, 0));
2218 
2219 	/*
2220 	 * For file-backed mbufs (from sendfile), anonymous wired
2221 	 * pages are allocated and used as the encryption destination.
2222 	 */
2223 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2224 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2225 		    m->m_epg_1st_off;
2226 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2227 		    m->m_epg_1st_off;
2228 		state->dst_iov[0].iov_len = len;
2229 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2230 		i = 1;
2231 	} else {
2232 		off = m->m_epg_1st_off;
2233 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2234 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2235 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2236 			len = m_epg_pagelen(m, i, off);
2237 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2238 			state->dst_iov[i].iov_base =
2239 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2240 			state->dst_iov[i].iov_len = len;
2241 		}
2242 	}
2243 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2244 	state->dst_iov[i].iov_base = m->m_epg_trail;
2245 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2246 
2247 	error = tls->sw_encrypt(state, tls, m, state->dst_iov, i + 1);
2248 
2249 	if (__predict_false(error != 0)) {
2250 		/* Free the anonymous pages. */
2251 		if (state->cbuf != NULL)
2252 			uma_zfree(ktls_buffer_zone, state->cbuf);
2253 		else {
2254 			for (i = 0; i < m->m_epg_npgs; i++) {
2255 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2256 				(void)vm_page_unwire_noq(pg);
2257 				vm_page_free(pg);
2258 			}
2259 		}
2260 	}
2261 	return (error);
2262 }
2263 
2264 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2265 static u_int
2266 ktls_batched_records(struct mbuf *m)
2267 {
2268 	int page_count, records;
2269 
2270 	records = 0;
2271 	page_count = m->m_epg_enc_cnt;
2272 	while (page_count > 0) {
2273 		records++;
2274 		page_count -= m->m_epg_nrdy;
2275 		m = m->m_next;
2276 	}
2277 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2278 	return (records);
2279 }
2280 
2281 void
2282 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2283 {
2284 	struct ktls_session *tls;
2285 	struct ktls_wq *wq;
2286 	int queued;
2287 	bool running;
2288 
2289 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2290 	    (M_EXTPG | M_NOTREADY)),
2291 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2292 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2293 
2294 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2295 
2296 	m->m_epg_enc_cnt = page_count;
2297 
2298 	/*
2299 	 * Save a pointer to the socket.  The caller is responsible
2300 	 * for taking an additional reference via soref().
2301 	 */
2302 	m->m_epg_so = so;
2303 
2304 	queued = 1;
2305 	tls = m->m_epg_tls;
2306 	wq = &ktls_wq[tls->wq_index];
2307 	mtx_lock(&wq->mtx);
2308 	if (__predict_false(tls->sequential_records)) {
2309 		/*
2310 		 * For TLS 1.0, records must be encrypted
2311 		 * sequentially.  For a given connection, all records
2312 		 * queued to the associated work queue are processed
2313 		 * sequentially.  However, sendfile(2) might complete
2314 		 * I/O requests spanning multiple TLS records out of
2315 		 * order.  Here we ensure TLS records are enqueued to
2316 		 * the work queue in FIFO order.
2317 		 *
2318 		 * tls->next_seqno holds the sequence number of the
2319 		 * next TLS record that should be enqueued to the work
2320 		 * queue.  If this next record is not tls->next_seqno,
2321 		 * it must be a future record, so insert it, sorted by
2322 		 * TLS sequence number, into tls->pending_records and
2323 		 * return.
2324 		 *
2325 		 * If this TLS record matches tls->next_seqno, place
2326 		 * it in the work queue and then check
2327 		 * tls->pending_records to see if any
2328 		 * previously-queued records are now ready for
2329 		 * encryption.
2330 		 */
2331 		if (m->m_epg_seqno != tls->next_seqno) {
2332 			struct mbuf *n, *p;
2333 
2334 			p = NULL;
2335 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2336 				if (n->m_epg_seqno > m->m_epg_seqno)
2337 					break;
2338 				p = n;
2339 			}
2340 			if (n == NULL)
2341 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2342 				    m_epg_stailq);
2343 			else if (p == NULL)
2344 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2345 				    m_epg_stailq);
2346 			else
2347 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2348 				    m_epg_stailq);
2349 			mtx_unlock(&wq->mtx);
2350 			counter_u64_add(ktls_cnt_tx_pending, 1);
2351 			return;
2352 		}
2353 
2354 		tls->next_seqno += ktls_batched_records(m);
2355 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2356 
2357 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2358 			struct mbuf *n;
2359 
2360 			n = STAILQ_FIRST(&tls->pending_records);
2361 			if (n->m_epg_seqno != tls->next_seqno)
2362 				break;
2363 
2364 			queued++;
2365 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2366 			tls->next_seqno += ktls_batched_records(n);
2367 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2368 		}
2369 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2370 	} else
2371 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2372 
2373 	running = wq->running;
2374 	mtx_unlock(&wq->mtx);
2375 	if (!running)
2376 		wakeup(wq);
2377 	counter_u64_add(ktls_cnt_tx_queued, queued);
2378 }
2379 
2380 /*
2381  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2382  * the pages from the file and replace them with the anonymous pages
2383  * allocated in ktls_encrypt_record().
2384  */
2385 static void
2386 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2387 {
2388 	int i;
2389 
2390 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2391 
2392 	/* Free the old pages. */
2393 	m->m_ext.ext_free(m);
2394 
2395 	/* Replace them with the new pages. */
2396 	if (state->cbuf != NULL) {
2397 		for (i = 0; i < m->m_epg_npgs; i++)
2398 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2399 
2400 		/* Contig pages should go back to the cache. */
2401 		m->m_ext.ext_free = ktls_free_mext_contig;
2402 	} else {
2403 		for (i = 0; i < m->m_epg_npgs; i++)
2404 			m->m_epg_pa[i] = state->parray[i];
2405 
2406 		/* Use the basic free routine. */
2407 		m->m_ext.ext_free = mb_free_mext_pgs;
2408 	}
2409 
2410 	/* Pages are now writable. */
2411 	m->m_epg_flags |= EPG_FLAG_ANON;
2412 }
2413 
2414 static __noinline void
2415 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2416 {
2417 	struct ktls_ocf_encrypt_state state;
2418 	struct ktls_session *tls;
2419 	struct socket *so;
2420 	struct mbuf *m;
2421 	int error, npages, total_pages;
2422 
2423 	so = top->m_epg_so;
2424 	tls = top->m_epg_tls;
2425 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2426 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2427 #ifdef INVARIANTS
2428 	top->m_epg_so = NULL;
2429 #endif
2430 	total_pages = top->m_epg_enc_cnt;
2431 	npages = 0;
2432 
2433 	/*
2434 	 * Encrypt the TLS records in the chain of mbufs starting with
2435 	 * 'top'.  'total_pages' gives us a total count of pages and is
2436 	 * used to know when we have finished encrypting the TLS
2437 	 * records originally queued with 'top'.
2438 	 *
2439 	 * NB: These mbufs are queued in the socket buffer and
2440 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2441 	 * socket buffer lock is not held while traversing this chain.
2442 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2443 	 * pointers should be stable.  However, the 'm_next' of the
2444 	 * last mbuf encrypted is not necessarily NULL.  It can point
2445 	 * to other mbufs appended while 'top' was on the TLS work
2446 	 * queue.
2447 	 *
2448 	 * Each mbuf holds an entire TLS record.
2449 	 */
2450 	error = 0;
2451 	for (m = top; npages != total_pages; m = m->m_next) {
2452 		KASSERT(m->m_epg_tls == tls,
2453 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2454 		    tls, m->m_epg_tls));
2455 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2456 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2457 		    total_pages, m));
2458 
2459 		error = ktls_encrypt_record(wq, m, tls, &state);
2460 		if (error) {
2461 			counter_u64_add(ktls_offload_failed_crypto, 1);
2462 			break;
2463 		}
2464 
2465 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2466 			ktls_finish_nonanon(m, &state);
2467 
2468 		npages += m->m_epg_nrdy;
2469 
2470 		/*
2471 		 * Drop a reference to the session now that it is no
2472 		 * longer needed.  Existing code depends on encrypted
2473 		 * records having no associated session vs
2474 		 * yet-to-be-encrypted records having an associated
2475 		 * session.
2476 		 */
2477 		m->m_epg_tls = NULL;
2478 		ktls_free(tls);
2479 	}
2480 
2481 	CURVNET_SET(so->so_vnet);
2482 	if (error == 0) {
2483 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2484 	} else {
2485 		so->so_proto->pr_usrreqs->pru_abort(so);
2486 		so->so_error = EIO;
2487 		mb_free_notready(top, total_pages);
2488 	}
2489 
2490 	sorele(so);
2491 	CURVNET_RESTORE();
2492 }
2493 
2494 void
2495 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2496 {
2497 	struct ktls_session *tls;
2498 	struct socket *so;
2499 	struct mbuf *m;
2500 	int npages;
2501 
2502 	m = state->m;
2503 
2504 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2505 		ktls_finish_nonanon(m, state);
2506 
2507 	so = state->so;
2508 	free(state, M_KTLS);
2509 
2510 	/*
2511 	 * Drop a reference to the session now that it is no longer
2512 	 * needed.  Existing code depends on encrypted records having
2513 	 * no associated session vs yet-to-be-encrypted records having
2514 	 * an associated session.
2515 	 */
2516 	tls = m->m_epg_tls;
2517 	m->m_epg_tls = NULL;
2518 	ktls_free(tls);
2519 
2520 	if (error != 0)
2521 		counter_u64_add(ktls_offload_failed_crypto, 1);
2522 
2523 	CURVNET_SET(so->so_vnet);
2524 	npages = m->m_epg_nrdy;
2525 
2526 	if (error == 0) {
2527 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, m, npages);
2528 	} else {
2529 		so->so_proto->pr_usrreqs->pru_abort(so);
2530 		so->so_error = EIO;
2531 		mb_free_notready(m, npages);
2532 	}
2533 
2534 	sorele(so);
2535 	CURVNET_RESTORE();
2536 }
2537 
2538 /*
2539  * Similar to ktls_encrypt, but used with asynchronous OCF backends
2540  * (coprocessors) where encryption does not use host CPU resources and
2541  * it can be beneficial to queue more requests than CPUs.
2542  */
2543 static __noinline void
2544 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2545 {
2546 	struct ktls_ocf_encrypt_state *state;
2547 	struct ktls_session *tls;
2548 	struct socket *so;
2549 	struct mbuf *m, *n;
2550 	int error, mpages, npages, total_pages;
2551 
2552 	so = top->m_epg_so;
2553 	tls = top->m_epg_tls;
2554 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2555 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2556 #ifdef INVARIANTS
2557 	top->m_epg_so = NULL;
2558 #endif
2559 	total_pages = top->m_epg_enc_cnt;
2560 	npages = 0;
2561 
2562 	error = 0;
2563 	for (m = top; npages != total_pages; m = n) {
2564 		KASSERT(m->m_epg_tls == tls,
2565 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2566 		    tls, m->m_epg_tls));
2567 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2568 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2569 		    total_pages, m));
2570 
2571 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2572 		soref(so);
2573 		state->so = so;
2574 		state->m = m;
2575 
2576 		mpages = m->m_epg_nrdy;
2577 		n = m->m_next;
2578 
2579 		error = ktls_encrypt_record(wq, m, tls, state);
2580 		if (error) {
2581 			counter_u64_add(ktls_offload_failed_crypto, 1);
2582 			free(state, M_KTLS);
2583 			CURVNET_SET(so->so_vnet);
2584 			sorele(so);
2585 			CURVNET_RESTORE();
2586 			break;
2587 		}
2588 
2589 		npages += mpages;
2590 	}
2591 
2592 	CURVNET_SET(so->so_vnet);
2593 	if (error != 0) {
2594 		so->so_proto->pr_usrreqs->pru_abort(so);
2595 		so->so_error = EIO;
2596 		mb_free_notready(m, total_pages - npages);
2597 	}
2598 
2599 	sorele(so);
2600 	CURVNET_RESTORE();
2601 }
2602 
2603 static int
2604 ktls_bind_domain(int domain)
2605 {
2606 	int error;
2607 
2608 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
2609 	if (error != 0)
2610 		return (error);
2611 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
2612 	return (0);
2613 }
2614 
2615 static void
2616 ktls_alloc_thread(void *ctx)
2617 {
2618 	struct ktls_domain_info *ktls_domain = ctx;
2619 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
2620 	void **buf;
2621 	struct sysctl_oid *oid;
2622 	char name[80];
2623 	int domain, error, i, nbufs;
2624 
2625 	domain = ktls_domain - ktls_domains;
2626 	if (bootverbose)
2627 		printf("Starting KTLS alloc thread for domain %d\n", domain);
2628 	error = ktls_bind_domain(domain);
2629 	if (error)
2630 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
2631 		    domain, error);
2632 	snprintf(name, sizeof(name), "domain%d", domain);
2633 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
2634 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2635 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
2636 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
2637 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
2638 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
2639 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
2640 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
2641 
2642 	buf = NULL;
2643 	nbufs = 0;
2644 	for (;;) {
2645 		atomic_store_int(&sc->running, 0);
2646 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
2647 		atomic_store_int(&sc->running, 1);
2648 		sc->wakeups++;
2649 		if (nbufs != ktls_max_alloc) {
2650 			free(buf, M_KTLS);
2651 			nbufs = atomic_load_int(&ktls_max_alloc);
2652 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
2653 			    M_WAITOK | M_ZERO);
2654 		}
2655 		/*
2656 		 * Below we allocate nbufs with different allocation
2657 		 * flags than we use when allocating normally during
2658 		 * encryption in the ktls worker thread.  We specify
2659 		 * M_NORECLAIM in the worker thread. However, we omit
2660 		 * that flag here and add M_WAITOK so that the VM
2661 		 * system is permitted to perform expensive work to
2662 		 * defragment memory.  We do this here, as it does not
2663 		 * matter if this thread blocks.  If we block a ktls
2664 		 * worker thread, we risk developing backlogs of
2665 		 * buffers to be encrypted, leading to surges of
2666 		 * traffic and potential NIC output drops.
2667 		 */
2668 		for (i = 0; i < nbufs; i++) {
2669 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
2670 			sc->allocs++;
2671 		}
2672 		for (i = 0; i < nbufs; i++) {
2673 			uma_zfree(ktls_buffer_zone, buf[i]);
2674 			buf[i] = NULL;
2675 		}
2676 	}
2677 }
2678 
2679 static void
2680 ktls_work_thread(void *ctx)
2681 {
2682 	struct ktls_wq *wq = ctx;
2683 	struct mbuf *m, *n;
2684 	struct socket *so, *son;
2685 	STAILQ_HEAD(, mbuf) local_m_head;
2686 	STAILQ_HEAD(, socket) local_so_head;
2687 	int cpu;
2688 
2689 	cpu = wq - ktls_wq;
2690 	if (bootverbose)
2691 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
2692 
2693 	/*
2694 	 * Bind to a core.  If ktls_bind_threads is > 1, then
2695 	 * we bind to the NUMA domain instead.
2696 	 */
2697 	if (ktls_bind_threads) {
2698 		int error;
2699 
2700 		if (ktls_bind_threads > 1) {
2701 			struct pcpu *pc = pcpu_find(cpu);
2702 
2703 			error = ktls_bind_domain(pc->pc_domain);
2704 		} else {
2705 			cpuset_t mask;
2706 
2707 			CPU_SETOF(cpu, &mask);
2708 			error = cpuset_setthread(curthread->td_tid, &mask);
2709 		}
2710 		if (error)
2711 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
2712 				cpu, error);
2713 	}
2714 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2715 	fpu_kern_thread(0);
2716 #endif
2717 	for (;;) {
2718 		mtx_lock(&wq->mtx);
2719 		while (STAILQ_EMPTY(&wq->m_head) &&
2720 		    STAILQ_EMPTY(&wq->so_head)) {
2721 			wq->running = false;
2722 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2723 			wq->running = true;
2724 		}
2725 
2726 		STAILQ_INIT(&local_m_head);
2727 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2728 		STAILQ_INIT(&local_so_head);
2729 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2730 		mtx_unlock(&wq->mtx);
2731 
2732 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2733 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2734 				ktls_free(m->m_epg_tls);
2735 				m_free_raw(m);
2736 			} else {
2737 				if (m->m_epg_tls->sync_dispatch)
2738 					ktls_encrypt(wq, m);
2739 				else
2740 					ktls_encrypt_async(wq, m);
2741 				counter_u64_add(ktls_cnt_tx_queued, -1);
2742 			}
2743 		}
2744 
2745 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2746 			ktls_decrypt(so);
2747 			counter_u64_add(ktls_cnt_rx_queued, -1);
2748 		}
2749 	}
2750 }
2751 
2752 #if defined(INET) || defined(INET6)
2753 static void
2754 ktls_disable_ifnet_help(void *context, int pending __unused)
2755 {
2756 	struct ktls_session *tls;
2757 	struct inpcb *inp;
2758 	struct tcpcb *tp;
2759 	struct socket *so;
2760 	int err;
2761 
2762 	tls = context;
2763 	inp = tls->inp;
2764 	if (inp == NULL)
2765 		return;
2766 	INP_WLOCK(inp);
2767 	so = inp->inp_socket;
2768 	MPASS(so != NULL);
2769 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2770 		goto out;
2771 	}
2772 
2773 	if (so->so_snd.sb_tls_info != NULL)
2774 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
2775 	else
2776 		err = ENXIO;
2777 	if (err == 0) {
2778 		counter_u64_add(ktls_ifnet_disable_ok, 1);
2779 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
2780 		if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 &&
2781 		    (tp = intotcpcb(inp)) != NULL &&
2782 		    tp->t_fb->tfb_hwtls_change != NULL)
2783 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
2784 	} else {
2785 		counter_u64_add(ktls_ifnet_disable_fail, 1);
2786 	}
2787 
2788 out:
2789 	sorele(so);
2790 	if (!in_pcbrele_wlocked(inp))
2791 		INP_WUNLOCK(inp);
2792 	ktls_free(tls);
2793 }
2794 
2795 /*
2796  * Called when re-transmits are becoming a substantial portion of the
2797  * sends on this connection.  When this happens, we transition the
2798  * connection to software TLS.  This is needed because most inline TLS
2799  * NICs keep crypto state only for in-order transmits.  This means
2800  * that to handle a TCP rexmit (which is out-of-order), the NIC must
2801  * re-DMA the entire TLS record up to and including the current
2802  * segment.  This means that when re-transmitting the last ~1448 byte
2803  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
2804  * of magnitude more data than we are sending.  This can cause the
2805  * PCIe link to saturate well before the network, which can cause
2806  * output drops, and a general loss of capacity.
2807  */
2808 void
2809 ktls_disable_ifnet(void *arg)
2810 {
2811 	struct tcpcb *tp;
2812 	struct inpcb *inp;
2813 	struct socket *so;
2814 	struct ktls_session *tls;
2815 
2816 	tp = arg;
2817 	inp = tp->t_inpcb;
2818 	INP_WLOCK_ASSERT(inp);
2819 	so = inp->inp_socket;
2820 	SOCK_LOCK(so);
2821 	tls = so->so_snd.sb_tls_info;
2822 	if (tls->disable_ifnet_pending) {
2823 		SOCK_UNLOCK(so);
2824 		return;
2825 	}
2826 
2827 	/*
2828 	 * note that disable_ifnet_pending is never cleared; disabling
2829 	 * ifnet can only be done once per session, so we never want
2830 	 * to do it again
2831 	 */
2832 
2833 	(void)ktls_hold(tls);
2834 	in_pcbref(inp);
2835 	soref(so);
2836 	tls->disable_ifnet_pending = true;
2837 	tls->inp = inp;
2838 	SOCK_UNLOCK(so);
2839 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
2840 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
2841 }
2842 #endif
2843