xref: /freebsd/sys/kern/uipc_ktls.c (revision aa339f1d5df9e38f36a34eb522355c4eebcae6c4)
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 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_kern_tls.h"
32 #include "opt_ratelimit.h"
33 #include "opt_rss.h"
34 
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/domainset.h>
38 #include <sys/endian.h>
39 #include <sys/ktls.h>
40 #include <sys/lock.h>
41 #include <sys/mbuf.h>
42 #include <sys/mutex.h>
43 #include <sys/rmlock.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/refcount.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/taskqueue.h>
52 #include <sys/kthread.h>
53 #include <sys/uio.h>
54 #include <sys/vmmeter.h>
55 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
56 #include <machine/pcb.h>
57 #endif
58 #include <machine/vmparam.h>
59 #include <net/if.h>
60 #include <net/if_var.h>
61 #ifdef RSS
62 #include <net/netisr.h>
63 #include <net/rss_config.h>
64 #endif
65 #include <net/route.h>
66 #include <net/route/nhop.h>
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_var.h>
70 #ifdef TCP_OFFLOAD
71 #include <netinet/tcp_offload.h>
72 #endif
73 #include <opencrypto/cryptodev.h>
74 #include <opencrypto/ktls.h>
75 #include <vm/vm.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_pagequeue.h>
79 
80 struct ktls_wq {
81 	struct mtx	mtx;
82 	STAILQ_HEAD(, mbuf) m_head;
83 	STAILQ_HEAD(, socket) so_head;
84 	bool		running;
85 	int		lastallocfail;
86 } __aligned(CACHE_LINE_SIZE);
87 
88 struct ktls_reclaim_thread {
89 	uint64_t wakeups;
90 	uint64_t reclaims;
91 	struct thread *td;
92 	int running;
93 };
94 
95 struct ktls_domain_info {
96 	int count;
97 	int cpu[MAXCPU];
98 	struct ktls_reclaim_thread reclaim_td;
99 };
100 
101 struct ktls_domain_info ktls_domains[MAXMEMDOM];
102 static struct ktls_wq *ktls_wq;
103 static struct proc *ktls_proc;
104 static uma_zone_t ktls_session_zone;
105 static uma_zone_t ktls_buffer_zone;
106 static uint16_t ktls_cpuid_lookup[MAXCPU];
107 static int ktls_init_state;
108 static struct sx ktls_init_lock;
109 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
110 
111 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
112     "Kernel TLS offload");
113 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114     "Kernel TLS offload stats");
115 
116 #ifdef RSS
117 static int ktls_bind_threads = 1;
118 #else
119 static int ktls_bind_threads;
120 #endif
121 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
122     &ktls_bind_threads, 0,
123     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
124 
125 static u_int ktls_maxlen = 16384;
126 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
127     &ktls_maxlen, 0, "Maximum TLS record size");
128 
129 static int ktls_number_threads;
130 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
131     &ktls_number_threads, 0,
132     "Number of TLS threads in thread-pool");
133 
134 unsigned int ktls_ifnet_max_rexmit_pct = 2;
135 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
136     &ktls_ifnet_max_rexmit_pct, 2,
137     "Max percent bytes retransmitted before ifnet TLS is disabled");
138 
139 static bool ktls_offload_enable;
140 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
141     &ktls_offload_enable, 0,
142     "Enable support for kernel TLS offload");
143 
144 static bool ktls_cbc_enable = true;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
146     &ktls_cbc_enable, 1,
147     "Enable support of AES-CBC crypto for kernel TLS");
148 
149 static bool ktls_sw_buffer_cache = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
151     &ktls_sw_buffer_cache, 1,
152     "Enable caching of output buffers for SW encryption");
153 
154 static int ktls_max_reclaim = 1024;
155 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN,
156     &ktls_max_reclaim, 128,
157     "Max number of 16k buffers to reclaim in thread context");
158 
159 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
160 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
161     &ktls_tasks_active, "Number of active tasks");
162 
163 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
164 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
165     &ktls_cnt_tx_pending,
166     "Number of TLS 1.0 records waiting for earlier TLS records");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
170     &ktls_cnt_tx_queued,
171     "Number of TLS records in queue to tasks for SW encryption");
172 
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
175     &ktls_cnt_rx_queued,
176     "Number of TLS sockets in queue to tasks for SW decryption");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
180     CTLFLAG_RD, &ktls_offload_total,
181     "Total successful TLS setups (parameters set)");
182 
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
185     CTLFLAG_RD, &ktls_offload_enable_calls,
186     "Total number of TLS enable calls made");
187 
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
190     &ktls_offload_active, "Total Active TLS sessions");
191 
192 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
193 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
194     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
195 
196 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
197 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
198     &ktls_offload_failed_crypto, "Total TLS crypto failures");
199 
200 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
201 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
202     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
203 
204 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
205 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
206     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
207 
208 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
209 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
210     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
211 
212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
214     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
215 
216 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
218     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
219 
220 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
222     &ktls_destroy_task,
223     "Number of times ktls session was destroyed via taskqueue");
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_reset_receive_tag(void *context, int pending);
301 static void ktls_reset_send_tag(void *context, int pending);
302 static void ktls_work_thread(void *ctx);
303 static void ktls_reclaim_thread(void *ctx);
304 
305 static u_int
306 ktls_get_cpu(struct socket *so)
307 {
308 	struct inpcb *inp;
309 #ifdef NUMA
310 	struct ktls_domain_info *di;
311 #endif
312 	u_int cpuid;
313 
314 	inp = sotoinpcb(so);
315 #ifdef RSS
316 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
317 	if (cpuid != NETISR_CPUID_NONE)
318 		return (cpuid);
319 #endif
320 	/*
321 	 * Just use the flowid to shard connections in a repeatable
322 	 * fashion.  Note that TLS 1.0 sessions rely on the
323 	 * serialization provided by having the same connection use
324 	 * the same queue.
325 	 */
326 #ifdef NUMA
327 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
328 		di = &ktls_domains[inp->inp_numa_domain];
329 		cpuid = di->cpu[inp->inp_flowid % di->count];
330 	} else
331 #endif
332 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
333 	return (cpuid);
334 }
335 
336 static int
337 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
338 {
339 	vm_page_t m;
340 	int i, req;
341 
342 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
343 	    ("%s: ktls max length %d is not page size-aligned",
344 	    __func__, ktls_maxlen));
345 
346 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
347 	for (i = 0; i < count; i++) {
348 		m = vm_page_alloc_noobj_contig_domain(domain, req,
349 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
350 		    VM_MEMATTR_DEFAULT);
351 		if (m == NULL)
352 			break;
353 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
354 	}
355 	return (i);
356 }
357 
358 static void
359 ktls_buffer_release(void *arg __unused, void **store, int count)
360 {
361 	vm_page_t m;
362 	int i, j;
363 
364 	for (i = 0; i < count; i++) {
365 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
366 		for (j = 0; j < atop(ktls_maxlen); j++) {
367 			(void)vm_page_unwire_noq(m + j);
368 			vm_page_free(m + j);
369 		}
370 	}
371 }
372 
373 static void
374 ktls_free_mext_contig(struct mbuf *m)
375 {
376 	M_ASSERTEXTPG(m);
377 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
378 }
379 
380 static int
381 ktls_init(void)
382 {
383 	struct thread *td;
384 	struct pcpu *pc;
385 	int count, domain, error, i;
386 
387 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
388 	    M_WAITOK | M_ZERO);
389 
390 	ktls_session_zone = uma_zcreate("ktls_session",
391 	    sizeof(struct ktls_session),
392 	    NULL, NULL, NULL, NULL,
393 	    UMA_ALIGN_CACHE, 0);
394 
395 	if (ktls_sw_buffer_cache) {
396 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
397 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
398 		    ktls_buffer_import, ktls_buffer_release, NULL,
399 		    UMA_ZONE_FIRSTTOUCH);
400 	}
401 
402 	/*
403 	 * Initialize the workqueues to run the TLS work.  We create a
404 	 * work queue for each CPU.
405 	 */
406 	CPU_FOREACH(i) {
407 		STAILQ_INIT(&ktls_wq[i].m_head);
408 		STAILQ_INIT(&ktls_wq[i].so_head);
409 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
410 		if (ktls_bind_threads > 1) {
411 			pc = pcpu_find(i);
412 			domain = pc->pc_domain;
413 			count = ktls_domains[domain].count;
414 			ktls_domains[domain].cpu[count] = i;
415 			ktls_domains[domain].count++;
416 		}
417 		ktls_cpuid_lookup[ktls_number_threads] = i;
418 		ktls_number_threads++;
419 	}
420 
421 	/*
422 	 * If we somehow have an empty domain, fall back to choosing
423 	 * among all KTLS threads.
424 	 */
425 	if (ktls_bind_threads > 1) {
426 		for (i = 0; i < vm_ndomains; i++) {
427 			if (ktls_domains[i].count == 0) {
428 				ktls_bind_threads = 1;
429 				break;
430 			}
431 		}
432 	}
433 
434 	/* Start kthreads for each workqueue. */
435 	CPU_FOREACH(i) {
436 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
437 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
438 		if (error) {
439 			printf("Can't add KTLS thread %d error %d\n", i, error);
440 			return (error);
441 		}
442 	}
443 
444 	/*
445 	 * Start an allocation thread per-domain to perform blocking allocations
446 	 * of 16k physically contiguous TLS crypto destination buffers.
447 	 */
448 	if (ktls_sw_buffer_cache) {
449 		for (domain = 0; domain < vm_ndomains; domain++) {
450 			if (VM_DOMAIN_EMPTY(domain))
451 				continue;
452 			if (CPU_EMPTY(&cpuset_domain[domain]))
453 				continue;
454 			error = kproc_kthread_add(ktls_reclaim_thread,
455 			    &ktls_domains[domain], &ktls_proc,
456 			    &ktls_domains[domain].reclaim_td.td,
457 			    0, 0, "KTLS", "reclaim_%d", domain);
458 			if (error) {
459 				printf("Can't add KTLS reclaim thread %d error %d\n",
460 				    domain, error);
461 				return (error);
462 			}
463 		}
464 	}
465 
466 	if (bootverbose)
467 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
468 	return (0);
469 }
470 
471 static int
472 ktls_start_kthreads(void)
473 {
474 	int error, state;
475 
476 start:
477 	state = atomic_load_acq_int(&ktls_init_state);
478 	if (__predict_true(state > 0))
479 		return (0);
480 	if (state < 0)
481 		return (ENXIO);
482 
483 	sx_xlock(&ktls_init_lock);
484 	if (ktls_init_state != 0) {
485 		sx_xunlock(&ktls_init_lock);
486 		goto start;
487 	}
488 
489 	error = ktls_init();
490 	if (error == 0)
491 		state = 1;
492 	else
493 		state = -1;
494 	atomic_store_rel_int(&ktls_init_state, state);
495 	sx_xunlock(&ktls_init_lock);
496 	return (error);
497 }
498 
499 static int
500 ktls_create_session(struct socket *so, struct tls_enable *en,
501     struct ktls_session **tlsp, int direction)
502 {
503 	struct ktls_session *tls;
504 	int error;
505 
506 	/* Only TLS 1.0 - 1.3 are supported. */
507 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
508 		return (EINVAL);
509 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
510 	    en->tls_vminor > TLS_MINOR_VER_THREE)
511 		return (EINVAL);
512 
513 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
514 		return (EINVAL);
515 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
516 		return (EINVAL);
517 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
518 		return (EINVAL);
519 
520 	/* All supported algorithms require a cipher key. */
521 	if (en->cipher_key_len == 0)
522 		return (EINVAL);
523 
524 	/* No flags are currently supported. */
525 	if (en->flags != 0)
526 		return (EINVAL);
527 
528 	/* Common checks for supported algorithms. */
529 	switch (en->cipher_algorithm) {
530 	case CRYPTO_AES_NIST_GCM_16:
531 		/*
532 		 * auth_algorithm isn't used, but permit GMAC values
533 		 * for compatibility.
534 		 */
535 		switch (en->auth_algorithm) {
536 		case 0:
537 #ifdef COMPAT_FREEBSD12
538 		/* XXX: Really 13.0-current COMPAT. */
539 		case CRYPTO_AES_128_NIST_GMAC:
540 		case CRYPTO_AES_192_NIST_GMAC:
541 		case CRYPTO_AES_256_NIST_GMAC:
542 #endif
543 			break;
544 		default:
545 			return (EINVAL);
546 		}
547 		if (en->auth_key_len != 0)
548 			return (EINVAL);
549 		switch (en->tls_vminor) {
550 		case TLS_MINOR_VER_TWO:
551 			if (en->iv_len != TLS_AEAD_GCM_LEN)
552 				return (EINVAL);
553 			break;
554 		case TLS_MINOR_VER_THREE:
555 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
556 				return (EINVAL);
557 			break;
558 		default:
559 			return (EINVAL);
560 		}
561 		break;
562 	case CRYPTO_AES_CBC:
563 		switch (en->auth_algorithm) {
564 		case CRYPTO_SHA1_HMAC:
565 			break;
566 		case CRYPTO_SHA2_256_HMAC:
567 		case CRYPTO_SHA2_384_HMAC:
568 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
569 				return (EINVAL);
570 			break;
571 		default:
572 			return (EINVAL);
573 		}
574 		if (en->auth_key_len == 0)
575 			return (EINVAL);
576 
577 		/*
578 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
579 		 * use explicit IVs.
580 		 */
581 		switch (en->tls_vminor) {
582 		case TLS_MINOR_VER_ZERO:
583 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
584 				return (EINVAL);
585 			break;
586 		case TLS_MINOR_VER_ONE:
587 		case TLS_MINOR_VER_TWO:
588 			/* Ignore any supplied IV. */
589 			en->iv_len = 0;
590 			break;
591 		default:
592 			return (EINVAL);
593 		}
594 		break;
595 	case CRYPTO_CHACHA20_POLY1305:
596 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
597 			return (EINVAL);
598 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
599 		    en->tls_vminor != TLS_MINOR_VER_THREE)
600 			return (EINVAL);
601 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
602 			return (EINVAL);
603 		break;
604 	default:
605 		return (EINVAL);
606 	}
607 
608 	error = ktls_start_kthreads();
609 	if (error != 0)
610 		return (error);
611 
612 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
613 
614 	counter_u64_add(ktls_offload_active, 1);
615 
616 	refcount_init(&tls->refcount, 1);
617 	if (direction == KTLS_RX) {
618 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
619 	} else {
620 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
621 		tls->inp = so->so_pcb;
622 		in_pcbref(tls->inp);
623 		tls->tx = true;
624 	}
625 
626 	tls->wq_index = ktls_get_cpu(so);
627 
628 	tls->params.cipher_algorithm = en->cipher_algorithm;
629 	tls->params.auth_algorithm = en->auth_algorithm;
630 	tls->params.tls_vmajor = en->tls_vmajor;
631 	tls->params.tls_vminor = en->tls_vminor;
632 	tls->params.flags = en->flags;
633 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
634 
635 	/* Set the header and trailer lengths. */
636 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
637 	switch (en->cipher_algorithm) {
638 	case CRYPTO_AES_NIST_GCM_16:
639 		/*
640 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
641 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
642 		 */
643 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
644 			tls->params.tls_hlen += sizeof(uint64_t);
645 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
646 		tls->params.tls_bs = 1;
647 		break;
648 	case CRYPTO_AES_CBC:
649 		switch (en->auth_algorithm) {
650 		case CRYPTO_SHA1_HMAC:
651 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
652 				/* Implicit IV, no nonce. */
653 				tls->sequential_records = true;
654 				tls->next_seqno = be64dec(en->rec_seq);
655 				STAILQ_INIT(&tls->pending_records);
656 			} else {
657 				tls->params.tls_hlen += AES_BLOCK_LEN;
658 			}
659 			tls->params.tls_tlen = AES_BLOCK_LEN +
660 			    SHA1_HASH_LEN;
661 			break;
662 		case CRYPTO_SHA2_256_HMAC:
663 			tls->params.tls_hlen += AES_BLOCK_LEN;
664 			tls->params.tls_tlen = AES_BLOCK_LEN +
665 			    SHA2_256_HASH_LEN;
666 			break;
667 		case CRYPTO_SHA2_384_HMAC:
668 			tls->params.tls_hlen += AES_BLOCK_LEN;
669 			tls->params.tls_tlen = AES_BLOCK_LEN +
670 			    SHA2_384_HASH_LEN;
671 			break;
672 		default:
673 			panic("invalid hmac");
674 		}
675 		tls->params.tls_bs = AES_BLOCK_LEN;
676 		break;
677 	case CRYPTO_CHACHA20_POLY1305:
678 		/*
679 		 * Chacha20 uses a 12 byte implicit IV.
680 		 */
681 		tls->params.tls_tlen = POLY1305_HASH_LEN;
682 		tls->params.tls_bs = 1;
683 		break;
684 	default:
685 		panic("invalid cipher");
686 	}
687 
688 	/*
689 	 * TLS 1.3 includes optional padding which we do not support,
690 	 * and also puts the "real" record type at the end of the
691 	 * encrypted data.
692 	 */
693 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
694 		tls->params.tls_tlen += sizeof(uint8_t);
695 
696 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
697 	    ("TLS header length too long: %d", tls->params.tls_hlen));
698 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
699 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
700 
701 	if (en->auth_key_len != 0) {
702 		tls->params.auth_key_len = en->auth_key_len;
703 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
704 		    M_WAITOK);
705 		error = copyin(en->auth_key, tls->params.auth_key,
706 		    en->auth_key_len);
707 		if (error)
708 			goto out;
709 	}
710 
711 	tls->params.cipher_key_len = en->cipher_key_len;
712 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
713 	error = copyin(en->cipher_key, tls->params.cipher_key,
714 	    en->cipher_key_len);
715 	if (error)
716 		goto out;
717 
718 	/*
719 	 * This holds the implicit portion of the nonce for AEAD
720 	 * ciphers and the initial implicit IV for TLS 1.0.  The
721 	 * explicit portions of the IV are generated in ktls_frame().
722 	 */
723 	if (en->iv_len != 0) {
724 		tls->params.iv_len = en->iv_len;
725 		error = copyin(en->iv, tls->params.iv, en->iv_len);
726 		if (error)
727 			goto out;
728 
729 		/*
730 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
731 		 * counter to generate unique explicit IVs.
732 		 *
733 		 * Store this counter in the last 8 bytes of the IV
734 		 * array so that it is 8-byte aligned.
735 		 */
736 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
737 		    en->tls_vminor == TLS_MINOR_VER_TWO)
738 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
739 	}
740 
741 	*tlsp = tls;
742 	return (0);
743 
744 out:
745 	ktls_free(tls);
746 	return (error);
747 }
748 
749 static struct ktls_session *
750 ktls_clone_session(struct ktls_session *tls, int direction)
751 {
752 	struct ktls_session *tls_new;
753 
754 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
755 
756 	counter_u64_add(ktls_offload_active, 1);
757 
758 	refcount_init(&tls_new->refcount, 1);
759 	if (direction == KTLS_RX) {
760 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
761 		    tls_new);
762 	} else {
763 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
764 		    tls_new);
765 		tls_new->inp = tls->inp;
766 		tls_new->tx = true;
767 		in_pcbref(tls_new->inp);
768 	}
769 
770 	/* Copy fields from existing session. */
771 	tls_new->params = tls->params;
772 	tls_new->wq_index = tls->wq_index;
773 
774 	/* Deep copy keys. */
775 	if (tls_new->params.auth_key != NULL) {
776 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
777 		    M_KTLS, M_WAITOK);
778 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
779 		    tls->params.auth_key_len);
780 	}
781 
782 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
783 	    M_WAITOK);
784 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
785 	    tls->params.cipher_key_len);
786 
787 	return (tls_new);
788 }
789 
790 #ifdef TCP_OFFLOAD
791 static int
792 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
793 {
794 	struct inpcb *inp;
795 	struct tcpcb *tp;
796 	int error;
797 
798 	inp = so->so_pcb;
799 	INP_WLOCK(inp);
800 	if (inp->inp_flags & INP_DROPPED) {
801 		INP_WUNLOCK(inp);
802 		return (ECONNRESET);
803 	}
804 	if (inp->inp_socket == NULL) {
805 		INP_WUNLOCK(inp);
806 		return (ECONNRESET);
807 	}
808 	tp = intotcpcb(inp);
809 	if (!(tp->t_flags & TF_TOE)) {
810 		INP_WUNLOCK(inp);
811 		return (EOPNOTSUPP);
812 	}
813 
814 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
815 	INP_WUNLOCK(inp);
816 	if (error == 0) {
817 		tls->mode = TCP_TLS_MODE_TOE;
818 		switch (tls->params.cipher_algorithm) {
819 		case CRYPTO_AES_CBC:
820 			counter_u64_add(ktls_toe_cbc, 1);
821 			break;
822 		case CRYPTO_AES_NIST_GCM_16:
823 			counter_u64_add(ktls_toe_gcm, 1);
824 			break;
825 		case CRYPTO_CHACHA20_POLY1305:
826 			counter_u64_add(ktls_toe_chacha20, 1);
827 			break;
828 		}
829 	}
830 	return (error);
831 }
832 #endif
833 
834 /*
835  * Common code used when first enabling ifnet TLS on a connection or
836  * when allocating a new ifnet TLS session due to a routing change.
837  * This function allocates a new TLS send tag on whatever interface
838  * the connection is currently routed over.
839  */
840 static int
841 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
842     struct m_snd_tag **mstp)
843 {
844 	union if_snd_tag_alloc_params params;
845 	struct ifnet *ifp;
846 	struct nhop_object *nh;
847 	struct tcpcb *tp;
848 	int error;
849 
850 	INP_RLOCK(inp);
851 	if (inp->inp_flags & INP_DROPPED) {
852 		INP_RUNLOCK(inp);
853 		return (ECONNRESET);
854 	}
855 	if (inp->inp_socket == NULL) {
856 		INP_RUNLOCK(inp);
857 		return (ECONNRESET);
858 	}
859 	tp = intotcpcb(inp);
860 
861 	/*
862 	 * Check administrative controls on ifnet TLS to determine if
863 	 * ifnet TLS should be denied.
864 	 *
865 	 * - Always permit 'force' requests.
866 	 * - ktls_ifnet_permitted == 0: always deny.
867 	 */
868 	if (!force && ktls_ifnet_permitted == 0) {
869 		INP_RUNLOCK(inp);
870 		return (ENXIO);
871 	}
872 
873 	/*
874 	 * XXX: Use the cached route in the inpcb to find the
875 	 * interface.  This should perhaps instead use
876 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
877 	 * enabled after a connection has completed key negotiation in
878 	 * userland, the cached route will be present in practice.
879 	 */
880 	nh = inp->inp_route.ro_nh;
881 	if (nh == NULL) {
882 		INP_RUNLOCK(inp);
883 		return (ENXIO);
884 	}
885 	ifp = nh->nh_ifp;
886 	if_ref(ifp);
887 
888 	/*
889 	 * Allocate a TLS + ratelimit tag if the connection has an
890 	 * existing pacing rate.
891 	 */
892 	if (tp->t_pacing_rate != -1 &&
893 	    (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
894 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
895 		params.tls_rate_limit.inp = inp;
896 		params.tls_rate_limit.tls = tls;
897 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
898 	} else {
899 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
900 		params.tls.inp = inp;
901 		params.tls.tls = tls;
902 	}
903 	params.hdr.flowid = inp->inp_flowid;
904 	params.hdr.flowtype = inp->inp_flowtype;
905 	params.hdr.numa_domain = inp->inp_numa_domain;
906 	INP_RUNLOCK(inp);
907 
908 	if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
909 		error = EOPNOTSUPP;
910 		goto out;
911 	}
912 	if (inp->inp_vflag & INP_IPV6) {
913 		if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
914 			error = EOPNOTSUPP;
915 			goto out;
916 		}
917 	} else {
918 		if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
919 			error = EOPNOTSUPP;
920 			goto out;
921 		}
922 	}
923 	error = m_snd_tag_alloc(ifp, &params, mstp);
924 out:
925 	if_rele(ifp);
926 	return (error);
927 }
928 
929 /*
930  * Allocate an initial TLS receive tag for doing HW decryption of TLS
931  * data.
932  *
933  * This function allocates a new TLS receive tag on whatever interface
934  * the connection is currently routed over.  If the connection ends up
935  * using a different interface for receive this will get fixed up via
936  * ktls_input_ifp_mismatch as future packets arrive.
937  */
938 static int
939 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
940     struct m_snd_tag **mstp)
941 {
942 	union if_snd_tag_alloc_params params;
943 	struct ifnet *ifp;
944 	struct nhop_object *nh;
945 	int error;
946 
947 	if (!ktls_ocf_recrypt_supported(tls))
948 		return (ENXIO);
949 
950 	INP_RLOCK(inp);
951 	if (inp->inp_flags & INP_DROPPED) {
952 		INP_RUNLOCK(inp);
953 		return (ECONNRESET);
954 	}
955 	if (inp->inp_socket == NULL) {
956 		INP_RUNLOCK(inp);
957 		return (ECONNRESET);
958 	}
959 
960 	/*
961 	 * Check administrative controls on ifnet TLS to determine if
962 	 * ifnet TLS should be denied.
963 	 */
964 	if (ktls_ifnet_permitted == 0) {
965 		INP_RUNLOCK(inp);
966 		return (ENXIO);
967 	}
968 
969 	/*
970 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
971 	 * the inpcb to find the interface.
972 	 */
973 	nh = inp->inp_route.ro_nh;
974 	if (nh == NULL) {
975 		INP_RUNLOCK(inp);
976 		return (ENXIO);
977 	}
978 	ifp = nh->nh_ifp;
979 	if_ref(ifp);
980 	tls->rx_ifp = ifp;
981 
982 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
983 	params.hdr.flowid = inp->inp_flowid;
984 	params.hdr.flowtype = inp->inp_flowtype;
985 	params.hdr.numa_domain = inp->inp_numa_domain;
986 	params.tls_rx.inp = inp;
987 	params.tls_rx.tls = tls;
988 	params.tls_rx.vlan_id = 0;
989 
990 	INP_RUNLOCK(inp);
991 
992 	if (inp->inp_vflag & INP_IPV6) {
993 		if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
994 			error = EOPNOTSUPP;
995 			goto out;
996 		}
997 	} else {
998 		if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
999 			error = EOPNOTSUPP;
1000 			goto out;
1001 		}
1002 	}
1003 	error = m_snd_tag_alloc(ifp, &params, mstp);
1004 
1005 	/*
1006 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1007 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1008 	 * ID for use in ktls_reset_receive_tag which allocates new
1009 	 * receive tags directly from the leaf interface bypassing
1010 	 * if_vlan.
1011 	 */
1012 	if (error == 0)
1013 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1014 out:
1015 	return (error);
1016 }
1017 
1018 static int
1019 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1020     bool force)
1021 {
1022 	struct m_snd_tag *mst;
1023 	int error;
1024 
1025 	switch (direction) {
1026 	case KTLS_TX:
1027 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1028 		if (__predict_false(error != 0))
1029 			goto done;
1030 		break;
1031 	case KTLS_RX:
1032 		KASSERT(!force, ("%s: forced receive tag", __func__));
1033 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1034 		if (__predict_false(error != 0))
1035 			goto done;
1036 		break;
1037 	default:
1038 		__assert_unreachable();
1039 	}
1040 
1041 	tls->mode = TCP_TLS_MODE_IFNET;
1042 	tls->snd_tag = mst;
1043 
1044 	switch (tls->params.cipher_algorithm) {
1045 	case CRYPTO_AES_CBC:
1046 		counter_u64_add(ktls_ifnet_cbc, 1);
1047 		break;
1048 	case CRYPTO_AES_NIST_GCM_16:
1049 		counter_u64_add(ktls_ifnet_gcm, 1);
1050 		break;
1051 	case CRYPTO_CHACHA20_POLY1305:
1052 		counter_u64_add(ktls_ifnet_chacha20, 1);
1053 		break;
1054 	default:
1055 		break;
1056 	}
1057 done:
1058 	return (error);
1059 }
1060 
1061 static void
1062 ktls_use_sw(struct ktls_session *tls)
1063 {
1064 	tls->mode = TCP_TLS_MODE_SW;
1065 	switch (tls->params.cipher_algorithm) {
1066 	case CRYPTO_AES_CBC:
1067 		counter_u64_add(ktls_sw_cbc, 1);
1068 		break;
1069 	case CRYPTO_AES_NIST_GCM_16:
1070 		counter_u64_add(ktls_sw_gcm, 1);
1071 		break;
1072 	case CRYPTO_CHACHA20_POLY1305:
1073 		counter_u64_add(ktls_sw_chacha20, 1);
1074 		break;
1075 	}
1076 }
1077 
1078 static int
1079 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1080 {
1081 	int error;
1082 
1083 	error = ktls_ocf_try(so, tls, direction);
1084 	if (error)
1085 		return (error);
1086 	ktls_use_sw(tls);
1087 	return (0);
1088 }
1089 
1090 /*
1091  * KTLS RX stores data in the socket buffer as a list of TLS records,
1092  * where each record is stored as a control message containg the TLS
1093  * header followed by data mbufs containing the decrypted data.  This
1094  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1095  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1096  * should be queued to the socket buffer as records, but encrypted
1097  * data which needs to be decrypted by software arrives as a stream of
1098  * regular mbufs which need to be converted.  In addition, there may
1099  * already be pending encrypted data in the socket buffer when KTLS RX
1100  * is enabled.
1101  *
1102  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1103  * is used:
1104  *
1105  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1106  *
1107  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1108  *   from the first mbuf.  Once all of the data for that TLS record is
1109  *   queued, the socket is queued to a worker thread.
1110  *
1111  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1112  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1113  *   decrypted, and inserted into the regular socket buffer chain as
1114  *   record starting with a control message holding the TLS header and
1115  *   a chain of mbufs holding the encrypted data.
1116  */
1117 
1118 static void
1119 sb_mark_notready(struct sockbuf *sb)
1120 {
1121 	struct mbuf *m;
1122 
1123 	m = sb->sb_mb;
1124 	sb->sb_mtls = m;
1125 	sb->sb_mb = NULL;
1126 	sb->sb_mbtail = NULL;
1127 	sb->sb_lastrecord = NULL;
1128 	for (; m != NULL; m = m->m_next) {
1129 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1130 		    __func__));
1131 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1132 		    __func__));
1133 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1134 		    __func__));
1135 		m->m_flags |= M_NOTREADY;
1136 		sb->sb_acc -= m->m_len;
1137 		sb->sb_tlscc += m->m_len;
1138 		sb->sb_mtlstail = m;
1139 	}
1140 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1141 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1142 	    sb->sb_ccc));
1143 }
1144 
1145 /*
1146  * Return information about the pending TLS data in a socket
1147  * buffer.  On return, 'seqno' is set to the sequence number
1148  * of the next TLS record to be received, 'resid' is set to
1149  * the amount of bytes still needed for the last pending
1150  * record.  The function returns 'false' if the last pending
1151  * record contains a partial TLS header.  In that case, 'resid'
1152  * is the number of bytes needed to complete the TLS header.
1153  */
1154 bool
1155 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1156 {
1157 	struct tls_record_layer hdr;
1158 	struct mbuf *m;
1159 	uint64_t seqno;
1160 	size_t resid;
1161 	u_int offset, record_len;
1162 
1163 	SOCKBUF_LOCK_ASSERT(sb);
1164 	MPASS(sb->sb_flags & SB_TLS_RX);
1165 	seqno = sb->sb_tls_seqno;
1166 	resid = sb->sb_tlscc;
1167 	m = sb->sb_mtls;
1168 	offset = 0;
1169 
1170 	if (resid == 0) {
1171 		*seqnop = seqno;
1172 		*residp = 0;
1173 		return (true);
1174 	}
1175 
1176 	for (;;) {
1177 		seqno++;
1178 
1179 		if (resid < sizeof(hdr)) {
1180 			*seqnop = seqno;
1181 			*residp = sizeof(hdr) - resid;
1182 			return (false);
1183 		}
1184 
1185 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1186 
1187 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1188 		if (resid <= record_len) {
1189 			*seqnop = seqno;
1190 			*residp = record_len - resid;
1191 			return (true);
1192 		}
1193 		resid -= record_len;
1194 
1195 		while (record_len != 0) {
1196 			if (m->m_len - offset > record_len) {
1197 				offset += record_len;
1198 				break;
1199 			}
1200 
1201 			record_len -= (m->m_len - offset);
1202 			offset = 0;
1203 			m = m->m_next;
1204 		}
1205 	}
1206 }
1207 
1208 int
1209 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1210 {
1211 	struct ktls_session *tls;
1212 	int error;
1213 
1214 	if (!ktls_offload_enable)
1215 		return (ENOTSUP);
1216 
1217 	counter_u64_add(ktls_offload_enable_calls, 1);
1218 
1219 	/*
1220 	 * This should always be true since only the TCP socket option
1221 	 * invokes this function.
1222 	 */
1223 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1224 		return (EINVAL);
1225 
1226 	/*
1227 	 * XXX: Don't overwrite existing sessions.  We should permit
1228 	 * this to support rekeying in the future.
1229 	 */
1230 	if (so->so_rcv.sb_tls_info != NULL)
1231 		return (EALREADY);
1232 
1233 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1234 		return (ENOTSUP);
1235 
1236 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1237 	if (error)
1238 		return (error);
1239 
1240 	error = ktls_ocf_try(so, tls, KTLS_RX);
1241 	if (error) {
1242 		ktls_free(tls);
1243 		return (error);
1244 	}
1245 
1246 	/* Mark the socket as using TLS offload. */
1247 	SOCK_RECVBUF_LOCK(so);
1248 	if (SOLISTENING(so)) {
1249 		SOCK_RECVBUF_UNLOCK(so);
1250 		ktls_free(tls);
1251 		return (EINVAL);
1252 	}
1253 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1254 	so->so_rcv.sb_tls_info = tls;
1255 	so->so_rcv.sb_flags |= SB_TLS_RX;
1256 
1257 	/* Mark existing data as not ready until it can be decrypted. */
1258 	sb_mark_notready(&so->so_rcv);
1259 	ktls_check_rx(&so->so_rcv);
1260 	SOCK_RECVBUF_UNLOCK(so);
1261 
1262 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1263 #ifdef TCP_OFFLOAD
1264 	error = ktls_try_toe(so, tls, KTLS_RX);
1265 	if (error)
1266 #endif
1267 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1268 	if (error)
1269 		ktls_use_sw(tls);
1270 
1271 	counter_u64_add(ktls_offload_total, 1);
1272 
1273 	return (0);
1274 }
1275 
1276 int
1277 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1278 {
1279 	struct ktls_session *tls;
1280 	struct inpcb *inp;
1281 	struct tcpcb *tp;
1282 	int error;
1283 
1284 	if (!ktls_offload_enable)
1285 		return (ENOTSUP);
1286 
1287 	counter_u64_add(ktls_offload_enable_calls, 1);
1288 
1289 	/*
1290 	 * This should always be true since only the TCP socket option
1291 	 * invokes this function.
1292 	 */
1293 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1294 		return (EINVAL);
1295 
1296 	/*
1297 	 * XXX: Don't overwrite existing sessions.  We should permit
1298 	 * this to support rekeying in the future.
1299 	 */
1300 	if (so->so_snd.sb_tls_info != NULL)
1301 		return (EALREADY);
1302 
1303 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1304 		return (ENOTSUP);
1305 
1306 	/* TLS requires ext pgs */
1307 	if (mb_use_ext_pgs == 0)
1308 		return (ENXIO);
1309 
1310 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1311 	if (error)
1312 		return (error);
1313 
1314 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1315 #ifdef TCP_OFFLOAD
1316 	error = ktls_try_toe(so, tls, KTLS_TX);
1317 	if (error)
1318 #endif
1319 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1320 	if (error)
1321 		error = ktls_try_sw(so, tls, KTLS_TX);
1322 
1323 	if (error) {
1324 		ktls_free(tls);
1325 		return (error);
1326 	}
1327 
1328 	/*
1329 	 * Serialize with sosend_generic() and make sure that we're not
1330 	 * operating on a listening socket.
1331 	 */
1332 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1333 	if (error) {
1334 		ktls_free(tls);
1335 		return (error);
1336 	}
1337 
1338 	/*
1339 	 * Write lock the INP when setting sb_tls_info so that
1340 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1341 	 * holding the INP lock.
1342 	 */
1343 	inp = so->so_pcb;
1344 	INP_WLOCK(inp);
1345 	SOCK_SENDBUF_LOCK(so);
1346 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1347 	so->so_snd.sb_tls_info = tls;
1348 	if (tls->mode != TCP_TLS_MODE_SW) {
1349 		tp = intotcpcb(inp);
1350 		MPASS(tp->t_nic_ktls_xmit == 0);
1351 		tp->t_nic_ktls_xmit = 1;
1352 		if (tp->t_fb->tfb_hwtls_change != NULL)
1353 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1354 	}
1355 	SOCK_SENDBUF_UNLOCK(so);
1356 	INP_WUNLOCK(inp);
1357 	SOCK_IO_SEND_UNLOCK(so);
1358 
1359 	counter_u64_add(ktls_offload_total, 1);
1360 
1361 	return (0);
1362 }
1363 
1364 int
1365 ktls_get_rx_mode(struct socket *so, int *modep)
1366 {
1367 	struct ktls_session *tls;
1368 	struct inpcb *inp __diagused;
1369 
1370 	if (SOLISTENING(so))
1371 		return (EINVAL);
1372 	inp = so->so_pcb;
1373 	INP_WLOCK_ASSERT(inp);
1374 	SOCK_RECVBUF_LOCK(so);
1375 	tls = so->so_rcv.sb_tls_info;
1376 	if (tls == NULL)
1377 		*modep = TCP_TLS_MODE_NONE;
1378 	else
1379 		*modep = tls->mode;
1380 	SOCK_RECVBUF_UNLOCK(so);
1381 	return (0);
1382 }
1383 
1384 /*
1385  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1386  *
1387  * This function gets information about the next TCP- and TLS-
1388  * sequence number to be processed by the TLS receive worker
1389  * thread. The information is extracted from the given "inpcb"
1390  * structure. The values are stored in host endian format at the two
1391  * given output pointer locations. The TCP sequence number points to
1392  * the beginning of the TLS header.
1393  *
1394  * This function returns zero on success, else a non-zero error code
1395  * is returned.
1396  */
1397 int
1398 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1399 {
1400 	struct socket *so;
1401 	struct tcpcb *tp;
1402 
1403 	INP_RLOCK(inp);
1404 	so = inp->inp_socket;
1405 	if (__predict_false(so == NULL)) {
1406 		INP_RUNLOCK(inp);
1407 		return (EINVAL);
1408 	}
1409 	if (inp->inp_flags & INP_DROPPED) {
1410 		INP_RUNLOCK(inp);
1411 		return (ECONNRESET);
1412 	}
1413 
1414 	tp = intotcpcb(inp);
1415 	MPASS(tp != NULL);
1416 
1417 	SOCKBUF_LOCK(&so->so_rcv);
1418 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1419 	*tlsseq = so->so_rcv.sb_tls_seqno;
1420 	SOCKBUF_UNLOCK(&so->so_rcv);
1421 
1422 	INP_RUNLOCK(inp);
1423 
1424 	return (0);
1425 }
1426 
1427 int
1428 ktls_get_tx_mode(struct socket *so, int *modep)
1429 {
1430 	struct ktls_session *tls;
1431 	struct inpcb *inp __diagused;
1432 
1433 	if (SOLISTENING(so))
1434 		return (EINVAL);
1435 	inp = so->so_pcb;
1436 	INP_WLOCK_ASSERT(inp);
1437 	SOCK_SENDBUF_LOCK(so);
1438 	tls = so->so_snd.sb_tls_info;
1439 	if (tls == NULL)
1440 		*modep = TCP_TLS_MODE_NONE;
1441 	else
1442 		*modep = tls->mode;
1443 	SOCK_SENDBUF_UNLOCK(so);
1444 	return (0);
1445 }
1446 
1447 /*
1448  * Switch between SW and ifnet TLS sessions as requested.
1449  */
1450 int
1451 ktls_set_tx_mode(struct socket *so, int mode)
1452 {
1453 	struct ktls_session *tls, *tls_new;
1454 	struct inpcb *inp;
1455 	struct tcpcb *tp;
1456 	int error;
1457 
1458 	if (SOLISTENING(so))
1459 		return (EINVAL);
1460 	switch (mode) {
1461 	case TCP_TLS_MODE_SW:
1462 	case TCP_TLS_MODE_IFNET:
1463 		break;
1464 	default:
1465 		return (EINVAL);
1466 	}
1467 
1468 	inp = so->so_pcb;
1469 	INP_WLOCK_ASSERT(inp);
1470 	tp = intotcpcb(inp);
1471 
1472 	if (mode == TCP_TLS_MODE_IFNET) {
1473 		/* Don't allow enabling ifnet ktls multiple times */
1474 		if (tp->t_nic_ktls_xmit)
1475 			return (EALREADY);
1476 
1477 		/*
1478 		 * Don't enable ifnet ktls if we disabled it due to an
1479 		 * excessive retransmission rate
1480 		 */
1481 		if (tp->t_nic_ktls_xmit_dis)
1482 			return (ENXIO);
1483 	}
1484 
1485 	SOCKBUF_LOCK(&so->so_snd);
1486 	tls = so->so_snd.sb_tls_info;
1487 	if (tls == NULL) {
1488 		SOCKBUF_UNLOCK(&so->so_snd);
1489 		return (0);
1490 	}
1491 
1492 	if (tls->mode == mode) {
1493 		SOCKBUF_UNLOCK(&so->so_snd);
1494 		return (0);
1495 	}
1496 
1497 	tls = ktls_hold(tls);
1498 	SOCKBUF_UNLOCK(&so->so_snd);
1499 	INP_WUNLOCK(inp);
1500 
1501 	tls_new = ktls_clone_session(tls, KTLS_TX);
1502 
1503 	if (mode == TCP_TLS_MODE_IFNET)
1504 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1505 	else
1506 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1507 	if (error) {
1508 		counter_u64_add(ktls_switch_failed, 1);
1509 		ktls_free(tls_new);
1510 		ktls_free(tls);
1511 		INP_WLOCK(inp);
1512 		return (error);
1513 	}
1514 
1515 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1516 	if (error) {
1517 		counter_u64_add(ktls_switch_failed, 1);
1518 		ktls_free(tls_new);
1519 		ktls_free(tls);
1520 		INP_WLOCK(inp);
1521 		return (error);
1522 	}
1523 
1524 	/*
1525 	 * If we raced with another session change, keep the existing
1526 	 * session.
1527 	 */
1528 	if (tls != so->so_snd.sb_tls_info) {
1529 		counter_u64_add(ktls_switch_failed, 1);
1530 		SOCK_IO_SEND_UNLOCK(so);
1531 		ktls_free(tls_new);
1532 		ktls_free(tls);
1533 		INP_WLOCK(inp);
1534 		return (EBUSY);
1535 	}
1536 
1537 	INP_WLOCK(inp);
1538 	SOCKBUF_LOCK(&so->so_snd);
1539 	so->so_snd.sb_tls_info = tls_new;
1540 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1541 		MPASS(tp->t_nic_ktls_xmit == 0);
1542 		tp->t_nic_ktls_xmit = 1;
1543 		if (tp->t_fb->tfb_hwtls_change != NULL)
1544 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1545 	}
1546 	SOCKBUF_UNLOCK(&so->so_snd);
1547 	SOCK_IO_SEND_UNLOCK(so);
1548 
1549 	/*
1550 	 * Drop two references on 'tls'.  The first is for the
1551 	 * ktls_hold() above.  The second drops the reference from the
1552 	 * socket buffer.
1553 	 */
1554 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1555 	ktls_free(tls);
1556 	ktls_free(tls);
1557 
1558 	if (mode == TCP_TLS_MODE_IFNET)
1559 		counter_u64_add(ktls_switch_to_ifnet, 1);
1560 	else
1561 		counter_u64_add(ktls_switch_to_sw, 1);
1562 
1563 	return (0);
1564 }
1565 
1566 /*
1567  * Try to allocate a new TLS receive tag.  This task is scheduled when
1568  * sbappend_ktls_rx detects an input path change.  If a new tag is
1569  * allocated, replace the tag in the TLS session.  If a new tag cannot
1570  * be allocated, let the session fall back to software decryption.
1571  */
1572 static void
1573 ktls_reset_receive_tag(void *context, int pending)
1574 {
1575 	union if_snd_tag_alloc_params params;
1576 	struct ktls_session *tls;
1577 	struct m_snd_tag *mst;
1578 	struct inpcb *inp;
1579 	struct ifnet *ifp;
1580 	struct socket *so;
1581 	int error;
1582 
1583 	MPASS(pending == 1);
1584 
1585 	tls = context;
1586 	so = tls->so;
1587 	inp = so->so_pcb;
1588 	ifp = NULL;
1589 
1590 	INP_RLOCK(inp);
1591 	if (inp->inp_flags & INP_DROPPED) {
1592 		INP_RUNLOCK(inp);
1593 		goto out;
1594 	}
1595 
1596 	SOCKBUF_LOCK(&so->so_rcv);
1597 	mst = tls->snd_tag;
1598 	tls->snd_tag = NULL;
1599 	if (mst != NULL)
1600 		m_snd_tag_rele(mst);
1601 
1602 	ifp = tls->rx_ifp;
1603 	if_ref(ifp);
1604 	SOCKBUF_UNLOCK(&so->so_rcv);
1605 
1606 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1607 	params.hdr.flowid = inp->inp_flowid;
1608 	params.hdr.flowtype = inp->inp_flowtype;
1609 	params.hdr.numa_domain = inp->inp_numa_domain;
1610 	params.tls_rx.inp = inp;
1611 	params.tls_rx.tls = tls;
1612 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1613 	INP_RUNLOCK(inp);
1614 
1615 	if (inp->inp_vflag & INP_IPV6) {
1616 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1617 			goto out;
1618 	} else {
1619 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1620 			goto out;
1621 	}
1622 
1623 	error = m_snd_tag_alloc(ifp, &params, &mst);
1624 	if (error == 0) {
1625 		SOCKBUF_LOCK(&so->so_rcv);
1626 		tls->snd_tag = mst;
1627 		SOCKBUF_UNLOCK(&so->so_rcv);
1628 
1629 		counter_u64_add(ktls_ifnet_reset, 1);
1630 	} else {
1631 		/*
1632 		 * Just fall back to software decryption if a tag
1633 		 * cannot be allocated leaving the connection intact.
1634 		 * If a future input path change switches to another
1635 		 * interface this connection will resume ifnet TLS.
1636 		 */
1637 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1638 	}
1639 
1640 out:
1641 	mtx_pool_lock(mtxpool_sleep, tls);
1642 	tls->reset_pending = false;
1643 	mtx_pool_unlock(mtxpool_sleep, tls);
1644 
1645 	if (ifp != NULL)
1646 		if_rele(ifp);
1647 	sorele(so);
1648 	ktls_free(tls);
1649 }
1650 
1651 /*
1652  * Try to allocate a new TLS send tag.  This task is scheduled when
1653  * ip_output detects a route change while trying to transmit a packet
1654  * holding a TLS record.  If a new tag is allocated, replace the tag
1655  * in the TLS session.  Subsequent packets on the connection will use
1656  * the new tag.  If a new tag cannot be allocated, drop the
1657  * connection.
1658  */
1659 static void
1660 ktls_reset_send_tag(void *context, int pending)
1661 {
1662 	struct epoch_tracker et;
1663 	struct ktls_session *tls;
1664 	struct m_snd_tag *old, *new;
1665 	struct inpcb *inp;
1666 	struct tcpcb *tp;
1667 	int error;
1668 
1669 	MPASS(pending == 1);
1670 
1671 	tls = context;
1672 	inp = tls->inp;
1673 
1674 	/*
1675 	 * Free the old tag first before allocating a new one.
1676 	 * ip[6]_output_send() will treat a NULL send tag the same as
1677 	 * an ifp mismatch and drop packets until a new tag is
1678 	 * allocated.
1679 	 *
1680 	 * Write-lock the INP when changing tls->snd_tag since
1681 	 * ip[6]_output_send() holds a read-lock when reading the
1682 	 * pointer.
1683 	 */
1684 	INP_WLOCK(inp);
1685 	old = tls->snd_tag;
1686 	tls->snd_tag = NULL;
1687 	INP_WUNLOCK(inp);
1688 	if (old != NULL)
1689 		m_snd_tag_rele(old);
1690 
1691 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1692 
1693 	if (error == 0) {
1694 		INP_WLOCK(inp);
1695 		tls->snd_tag = new;
1696 		mtx_pool_lock(mtxpool_sleep, tls);
1697 		tls->reset_pending = false;
1698 		mtx_pool_unlock(mtxpool_sleep, tls);
1699 		INP_WUNLOCK(inp);
1700 
1701 		counter_u64_add(ktls_ifnet_reset, 1);
1702 
1703 		/*
1704 		 * XXX: Should we kick tcp_output explicitly now that
1705 		 * the send tag is fixed or just rely on timers?
1706 		 */
1707 	} else {
1708 		NET_EPOCH_ENTER(et);
1709 		INP_WLOCK(inp);
1710 		if (!(inp->inp_flags & INP_DROPPED)) {
1711 			tp = intotcpcb(inp);
1712 			CURVNET_SET(inp->inp_vnet);
1713 			tp = tcp_drop(tp, ECONNABORTED);
1714 			CURVNET_RESTORE();
1715 			if (tp != NULL) {
1716 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1717 				INP_WUNLOCK(inp);
1718 			}
1719 		} else
1720 			INP_WUNLOCK(inp);
1721 		NET_EPOCH_EXIT(et);
1722 
1723 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1724 
1725 		/*
1726 		 * Leave reset_pending true to avoid future tasks while
1727 		 * the socket goes away.
1728 		 */
1729 	}
1730 
1731 	ktls_free(tls);
1732 }
1733 
1734 void
1735 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1736 {
1737 	struct ktls_session *tls;
1738 	struct socket *so;
1739 
1740 	SOCKBUF_LOCK_ASSERT(sb);
1741 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1742 	    __func__, sb));
1743 	so = __containerof(sb, struct socket, so_rcv);
1744 
1745 	tls = sb->sb_tls_info;
1746 	if_rele(tls->rx_ifp);
1747 	if_ref(ifp);
1748 	tls->rx_ifp = ifp;
1749 
1750 	/*
1751 	 * See if we should schedule a task to update the receive tag for
1752 	 * this session.
1753 	 */
1754 	mtx_pool_lock(mtxpool_sleep, tls);
1755 	if (!tls->reset_pending) {
1756 		(void) ktls_hold(tls);
1757 		soref(so);
1758 		tls->so = so;
1759 		tls->reset_pending = true;
1760 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1761 	}
1762 	mtx_pool_unlock(mtxpool_sleep, tls);
1763 }
1764 
1765 int
1766 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1767 {
1768 
1769 	if (inp == NULL)
1770 		return (ENOBUFS);
1771 
1772 	INP_LOCK_ASSERT(inp);
1773 
1774 	/*
1775 	 * See if we should schedule a task to update the send tag for
1776 	 * this session.
1777 	 */
1778 	mtx_pool_lock(mtxpool_sleep, tls);
1779 	if (!tls->reset_pending) {
1780 		(void) ktls_hold(tls);
1781 		tls->reset_pending = true;
1782 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1783 	}
1784 	mtx_pool_unlock(mtxpool_sleep, tls);
1785 	return (ENOBUFS);
1786 }
1787 
1788 #ifdef RATELIMIT
1789 int
1790 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1791 {
1792 	union if_snd_tag_modify_params params = {
1793 		.rate_limit.max_rate = max_pacing_rate,
1794 		.rate_limit.flags = M_NOWAIT,
1795 	};
1796 	struct m_snd_tag *mst;
1797 
1798 	/* Can't get to the inp, but it should be locked. */
1799 	/* INP_LOCK_ASSERT(inp); */
1800 
1801 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1802 
1803 	if (tls->snd_tag == NULL) {
1804 		/*
1805 		 * Resetting send tag, ignore this change.  The
1806 		 * pending reset may or may not see this updated rate
1807 		 * in the tcpcb.  If it doesn't, we will just lose
1808 		 * this rate change.
1809 		 */
1810 		return (0);
1811 	}
1812 
1813 	mst = tls->snd_tag;
1814 
1815 	MPASS(mst != NULL);
1816 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1817 
1818 	return (mst->sw->snd_tag_modify(mst, &params));
1819 }
1820 #endif
1821 
1822 static void
1823 ktls_destroy_help(void *context, int pending __unused)
1824 {
1825 	ktls_destroy(context);
1826 }
1827 
1828 void
1829 ktls_destroy(struct ktls_session *tls)
1830 {
1831 	struct inpcb *inp;
1832 	struct tcpcb *tp;
1833 	bool wlocked;
1834 
1835 	MPASS(tls->refcount == 0);
1836 
1837 	inp = tls->inp;
1838 	if (tls->tx) {
1839 		wlocked = INP_WLOCKED(inp);
1840 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1841 			/*
1842 			 * rwlocks read locks are anonymous, and there
1843 			 * is no way to know if our current thread
1844 			 * holds an rlock on the inp.  As a rough
1845 			 * estimate, check to see if the thread holds
1846 			 * *any* rlocks at all.  If it does not, then we
1847 			 * know that we don't hold the inp rlock, and
1848 			 * can safely take the wlock
1849 			 */
1850 			if (curthread->td_rw_rlocks == 0) {
1851 				INP_WLOCK(inp);
1852 			} else {
1853 				/*
1854 				 * We might hold the rlock, so let's
1855 				 * do the destroy in a taskqueue
1856 				 * context to avoid a potential
1857 				 * deadlock.  This should be very
1858 				 * rare.
1859 				 */
1860 				counter_u64_add(ktls_destroy_task, 1);
1861 				TASK_INIT(&tls->destroy_task, 0,
1862 				    ktls_destroy_help, tls);
1863 				(void)taskqueue_enqueue(taskqueue_thread,
1864 				    &tls->destroy_task);
1865 				return;
1866 			}
1867 		}
1868 	}
1869 
1870 	if (tls->sequential_records) {
1871 		struct mbuf *m, *n;
1872 		int page_count;
1873 
1874 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1875 			page_count = m->m_epg_enc_cnt;
1876 			while (page_count > 0) {
1877 				KASSERT(page_count >= m->m_epg_nrdy,
1878 				    ("%s: too few pages", __func__));
1879 				page_count -= m->m_epg_nrdy;
1880 				m = m_free(m);
1881 			}
1882 		}
1883 	}
1884 
1885 	counter_u64_add(ktls_offload_active, -1);
1886 	switch (tls->mode) {
1887 	case TCP_TLS_MODE_SW:
1888 		switch (tls->params.cipher_algorithm) {
1889 		case CRYPTO_AES_CBC:
1890 			counter_u64_add(ktls_sw_cbc, -1);
1891 			break;
1892 		case CRYPTO_AES_NIST_GCM_16:
1893 			counter_u64_add(ktls_sw_gcm, -1);
1894 			break;
1895 		case CRYPTO_CHACHA20_POLY1305:
1896 			counter_u64_add(ktls_sw_chacha20, -1);
1897 			break;
1898 		}
1899 		break;
1900 	case TCP_TLS_MODE_IFNET:
1901 		switch (tls->params.cipher_algorithm) {
1902 		case CRYPTO_AES_CBC:
1903 			counter_u64_add(ktls_ifnet_cbc, -1);
1904 			break;
1905 		case CRYPTO_AES_NIST_GCM_16:
1906 			counter_u64_add(ktls_ifnet_gcm, -1);
1907 			break;
1908 		case CRYPTO_CHACHA20_POLY1305:
1909 			counter_u64_add(ktls_ifnet_chacha20, -1);
1910 			break;
1911 		}
1912 		if (tls->snd_tag != NULL)
1913 			m_snd_tag_rele(tls->snd_tag);
1914 		if (tls->rx_ifp != NULL)
1915 			if_rele(tls->rx_ifp);
1916 		if (tls->tx) {
1917 			INP_WLOCK_ASSERT(inp);
1918 			tp = intotcpcb(inp);
1919 			MPASS(tp->t_nic_ktls_xmit == 1);
1920 			tp->t_nic_ktls_xmit = 0;
1921 		}
1922 		break;
1923 #ifdef TCP_OFFLOAD
1924 	case TCP_TLS_MODE_TOE:
1925 		switch (tls->params.cipher_algorithm) {
1926 		case CRYPTO_AES_CBC:
1927 			counter_u64_add(ktls_toe_cbc, -1);
1928 			break;
1929 		case CRYPTO_AES_NIST_GCM_16:
1930 			counter_u64_add(ktls_toe_gcm, -1);
1931 			break;
1932 		case CRYPTO_CHACHA20_POLY1305:
1933 			counter_u64_add(ktls_toe_chacha20, -1);
1934 			break;
1935 		}
1936 		break;
1937 #endif
1938 	}
1939 	if (tls->ocf_session != NULL)
1940 		ktls_ocf_free(tls);
1941 	if (tls->params.auth_key != NULL) {
1942 		zfree(tls->params.auth_key, M_KTLS);
1943 		tls->params.auth_key = NULL;
1944 		tls->params.auth_key_len = 0;
1945 	}
1946 	if (tls->params.cipher_key != NULL) {
1947 		zfree(tls->params.cipher_key, M_KTLS);
1948 		tls->params.cipher_key = NULL;
1949 		tls->params.cipher_key_len = 0;
1950 	}
1951 	if (tls->tx) {
1952 		INP_WLOCK_ASSERT(inp);
1953 		if (!in_pcbrele_wlocked(inp) && !wlocked)
1954 			INP_WUNLOCK(inp);
1955 	}
1956 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1957 
1958 	uma_zfree(ktls_session_zone, tls);
1959 }
1960 
1961 void
1962 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1963 {
1964 
1965 	for (; m != NULL; m = m->m_next) {
1966 		KASSERT((m->m_flags & M_EXTPG) != 0,
1967 		    ("ktls_seq: mapped mbuf %p", m));
1968 
1969 		m->m_epg_seqno = sb->sb_tls_seqno;
1970 		sb->sb_tls_seqno++;
1971 	}
1972 }
1973 
1974 /*
1975  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1976  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1977  * mbuf must be populated with the payload of each TLS record.
1978  *
1979  * The record_type argument specifies the TLS record type used when
1980  * populating the TLS header.
1981  *
1982  * The enq_count argument on return is set to the number of pages of
1983  * payload data for this entire chain that need to be encrypted via SW
1984  * encryption.  The returned value should be passed to ktls_enqueue
1985  * when scheduling encryption of this chain of mbufs.  To handle the
1986  * special case of empty fragments for TLS 1.0 sessions, an empty
1987  * fragment counts as one page.
1988  */
1989 void
1990 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1991     uint8_t record_type)
1992 {
1993 	struct tls_record_layer *tlshdr;
1994 	struct mbuf *m;
1995 	uint64_t *noncep;
1996 	uint16_t tls_len;
1997 	int maxlen __diagused;
1998 
1999 	maxlen = tls->params.max_frame_len;
2000 	*enq_cnt = 0;
2001 	for (m = top; m != NULL; m = m->m_next) {
2002 		/*
2003 		 * All mbufs in the chain should be TLS records whose
2004 		 * payload does not exceed the maximum frame length.
2005 		 *
2006 		 * Empty TLS 1.0 records are permitted when using CBC.
2007 		 */
2008 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2009 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2010 		    ("ktls_frame: m %p len %d", m, m->m_len));
2011 
2012 		/*
2013 		 * TLS frames require unmapped mbufs to store session
2014 		 * info.
2015 		 */
2016 		KASSERT((m->m_flags & M_EXTPG) != 0,
2017 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2018 
2019 		tls_len = m->m_len;
2020 
2021 		/* Save a reference to the session. */
2022 		m->m_epg_tls = ktls_hold(tls);
2023 
2024 		m->m_epg_hdrlen = tls->params.tls_hlen;
2025 		m->m_epg_trllen = tls->params.tls_tlen;
2026 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2027 			int bs, delta;
2028 
2029 			/*
2030 			 * AES-CBC pads messages to a multiple of the
2031 			 * block size.  Note that the padding is
2032 			 * applied after the digest and the encryption
2033 			 * is done on the "plaintext || mac || padding".
2034 			 * At least one byte of padding is always
2035 			 * present.
2036 			 *
2037 			 * Compute the final trailer length assuming
2038 			 * at most one block of padding.
2039 			 * tls->params.tls_tlen is the maximum
2040 			 * possible trailer length (padding + digest).
2041 			 * delta holds the number of excess padding
2042 			 * bytes if the maximum were used.  Those
2043 			 * extra bytes are removed.
2044 			 */
2045 			bs = tls->params.tls_bs;
2046 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2047 			m->m_epg_trllen -= delta;
2048 		}
2049 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2050 
2051 		/* Populate the TLS header. */
2052 		tlshdr = (void *)m->m_epg_hdr;
2053 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2054 
2055 		/*
2056 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2057 		 * of TLS_RLTYPE_APP.
2058 		 */
2059 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2060 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2061 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2062 			tlshdr->tls_type = TLS_RLTYPE_APP;
2063 			/* save the real record type for later */
2064 			m->m_epg_record_type = record_type;
2065 			m->m_epg_trail[0] = record_type;
2066 		} else {
2067 			tlshdr->tls_vminor = tls->params.tls_vminor;
2068 			tlshdr->tls_type = record_type;
2069 		}
2070 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2071 
2072 		/*
2073 		 * Store nonces / explicit IVs after the end of the
2074 		 * TLS header.
2075 		 *
2076 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2077 		 * from the end of the IV.  The nonce is then
2078 		 * incremented for use by the next record.
2079 		 *
2080 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2081 		 */
2082 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2083 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2084 			noncep = (uint64_t *)(tls->params.iv + 8);
2085 			be64enc(tlshdr + 1, *noncep);
2086 			(*noncep)++;
2087 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2088 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2089 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2090 
2091 		/*
2092 		 * When using SW encryption, mark the mbuf not ready.
2093 		 * It will be marked ready via sbready() after the
2094 		 * record has been encrypted.
2095 		 *
2096 		 * When using ifnet TLS, unencrypted TLS records are
2097 		 * sent down the stack to the NIC.
2098 		 */
2099 		if (tls->mode == TCP_TLS_MODE_SW) {
2100 			m->m_flags |= M_NOTREADY;
2101 			if (__predict_false(tls_len == 0)) {
2102 				/* TLS 1.0 empty fragment. */
2103 				m->m_epg_nrdy = 1;
2104 			} else
2105 				m->m_epg_nrdy = m->m_epg_npgs;
2106 			*enq_cnt += m->m_epg_nrdy;
2107 		}
2108 	}
2109 }
2110 
2111 bool
2112 ktls_permit_empty_frames(struct ktls_session *tls)
2113 {
2114 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2115 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2116 }
2117 
2118 void
2119 ktls_check_rx(struct sockbuf *sb)
2120 {
2121 	struct tls_record_layer hdr;
2122 	struct ktls_wq *wq;
2123 	struct socket *so;
2124 	bool running;
2125 
2126 	SOCKBUF_LOCK_ASSERT(sb);
2127 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2128 	    __func__, sb));
2129 	so = __containerof(sb, struct socket, so_rcv);
2130 
2131 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2132 		return;
2133 
2134 	/* Is there enough queued for a TLS header? */
2135 	if (sb->sb_tlscc < sizeof(hdr)) {
2136 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2137 			so->so_error = EMSGSIZE;
2138 		return;
2139 	}
2140 
2141 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2142 
2143 	/* Is the entire record queued? */
2144 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2145 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2146 			so->so_error = EMSGSIZE;
2147 		return;
2148 	}
2149 
2150 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2151 
2152 	soref(so);
2153 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2154 	mtx_lock(&wq->mtx);
2155 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2156 	running = wq->running;
2157 	mtx_unlock(&wq->mtx);
2158 	if (!running)
2159 		wakeup(wq);
2160 	counter_u64_add(ktls_cnt_rx_queued, 1);
2161 }
2162 
2163 static struct mbuf *
2164 ktls_detach_record(struct sockbuf *sb, int len)
2165 {
2166 	struct mbuf *m, *n, *top;
2167 	int remain;
2168 
2169 	SOCKBUF_LOCK_ASSERT(sb);
2170 	MPASS(len <= sb->sb_tlscc);
2171 
2172 	/*
2173 	 * If TLS chain is the exact size of the record,
2174 	 * just grab the whole record.
2175 	 */
2176 	top = sb->sb_mtls;
2177 	if (sb->sb_tlscc == len) {
2178 		sb->sb_mtls = NULL;
2179 		sb->sb_mtlstail = NULL;
2180 		goto out;
2181 	}
2182 
2183 	/*
2184 	 * While it would be nice to use m_split() here, we need
2185 	 * to know exactly what m_split() allocates to update the
2186 	 * accounting, so do it inline instead.
2187 	 */
2188 	remain = len;
2189 	for (m = top; remain > m->m_len; m = m->m_next)
2190 		remain -= m->m_len;
2191 
2192 	/* Easy case: don't have to split 'm'. */
2193 	if (remain == m->m_len) {
2194 		sb->sb_mtls = m->m_next;
2195 		if (sb->sb_mtls == NULL)
2196 			sb->sb_mtlstail = NULL;
2197 		m->m_next = NULL;
2198 		goto out;
2199 	}
2200 
2201 	/*
2202 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2203 	 * with M_NOWAIT first.
2204 	 */
2205 	n = m_get(M_NOWAIT, MT_DATA);
2206 	if (n == NULL) {
2207 		/*
2208 		 * Use M_WAITOK with socket buffer unlocked.  If
2209 		 * 'sb_mtls' changes while the lock is dropped, return
2210 		 * NULL to force the caller to retry.
2211 		 */
2212 		SOCKBUF_UNLOCK(sb);
2213 
2214 		n = m_get(M_WAITOK, MT_DATA);
2215 
2216 		SOCKBUF_LOCK(sb);
2217 		if (sb->sb_mtls != top) {
2218 			m_free(n);
2219 			return (NULL);
2220 		}
2221 	}
2222 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2223 
2224 	/* Store remainder in 'n'. */
2225 	n->m_len = m->m_len - remain;
2226 	if (m->m_flags & M_EXT) {
2227 		n->m_data = m->m_data + remain;
2228 		mb_dupcl(n, m);
2229 	} else {
2230 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2231 	}
2232 
2233 	/* Trim 'm' and update accounting. */
2234 	m->m_len -= n->m_len;
2235 	sb->sb_tlscc -= n->m_len;
2236 	sb->sb_ccc -= n->m_len;
2237 
2238 	/* Account for 'n'. */
2239 	sballoc_ktls_rx(sb, n);
2240 
2241 	/* Insert 'n' into the TLS chain. */
2242 	sb->sb_mtls = n;
2243 	n->m_next = m->m_next;
2244 	if (sb->sb_mtlstail == m)
2245 		sb->sb_mtlstail = n;
2246 
2247 	/* Detach the record from the TLS chain. */
2248 	m->m_next = NULL;
2249 
2250 out:
2251 	MPASS(m_length(top, NULL) == len);
2252 	for (m = top; m != NULL; m = m->m_next)
2253 		sbfree_ktls_rx(sb, m);
2254 	sb->sb_tlsdcc = len;
2255 	sb->sb_ccc += len;
2256 	SBCHECK(sb);
2257 	return (top);
2258 }
2259 
2260 /*
2261  * Determine the length of the trailing zero padding and find the real
2262  * record type in the byte before the padding.
2263  *
2264  * Walking the mbuf chain backwards is clumsy, so another option would
2265  * be to scan forwards remembering the last non-zero byte before the
2266  * trailer.  However, it would be expensive to scan the entire record.
2267  * Instead, find the last non-zero byte of each mbuf in the chain
2268  * keeping track of the relative offset of that nonzero byte.
2269  *
2270  * trail_len is the size of the MAC/tag on input and is set to the
2271  * size of the full trailer including padding and the record type on
2272  * return.
2273  */
2274 static int
2275 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2276     int *trailer_len, uint8_t *record_typep)
2277 {
2278 	char *cp;
2279 	u_int digest_start, last_offset, m_len, offset;
2280 	uint8_t record_type;
2281 
2282 	digest_start = tls_len - *trailer_len;
2283 	last_offset = 0;
2284 	offset = 0;
2285 	for (; m != NULL && offset < digest_start;
2286 	     offset += m->m_len, m = m->m_next) {
2287 		/* Don't look for padding in the tag. */
2288 		m_len = min(digest_start - offset, m->m_len);
2289 		cp = mtod(m, char *);
2290 
2291 		/* Find last non-zero byte in this mbuf. */
2292 		while (m_len > 0 && cp[m_len - 1] == 0)
2293 			m_len--;
2294 		if (m_len > 0) {
2295 			record_type = cp[m_len - 1];
2296 			last_offset = offset + m_len;
2297 		}
2298 	}
2299 	if (last_offset < tls->params.tls_hlen)
2300 		return (EBADMSG);
2301 
2302 	*record_typep = record_type;
2303 	*trailer_len = tls_len - last_offset + 1;
2304 	return (0);
2305 }
2306 
2307 /*
2308  * Check if a mbuf chain is fully decrypted at the given offset and
2309  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2310  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2311  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2312  * is encrypted.
2313  */
2314 ktls_mbuf_crypto_st_t
2315 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2316 {
2317 	int m_flags_ored = 0;
2318 	int m_flags_anded = -1;
2319 
2320 	for (; mb != NULL; mb = mb->m_next) {
2321 		if (offset < mb->m_len)
2322 			break;
2323 		offset -= mb->m_len;
2324 	}
2325 	offset += len;
2326 
2327 	for (; mb != NULL; mb = mb->m_next) {
2328 		m_flags_ored |= mb->m_flags;
2329 		m_flags_anded &= mb->m_flags;
2330 
2331 		if (offset <= mb->m_len)
2332 			break;
2333 		offset -= mb->m_len;
2334 	}
2335 	MPASS(mb != NULL || offset == 0);
2336 
2337 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2338 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2339 	else
2340 		return ((m_flags_ored & M_DECRYPTED) ?
2341 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2342 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2343 }
2344 
2345 /*
2346  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2347  */
2348 static int
2349 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2350 {
2351 	union if_snd_tag_modify_params params;
2352 	struct m_snd_tag *mst;
2353 	struct inpcb *inp;
2354 	struct tcpcb *tp;
2355 
2356 	mst = so->so_rcv.sb_tls_info->snd_tag;
2357 	if (__predict_false(mst == NULL))
2358 		return (EINVAL);
2359 
2360 	inp = sotoinpcb(so);
2361 	if (__predict_false(inp == NULL))
2362 		return (EINVAL);
2363 
2364 	INP_RLOCK(inp);
2365 	if (inp->inp_flags & INP_DROPPED) {
2366 		INP_RUNLOCK(inp);
2367 		return (ECONNRESET);
2368 	}
2369 
2370 	tp = intotcpcb(inp);
2371 	MPASS(tp != NULL);
2372 
2373 	/* Get the TCP sequence number of the next valid TLS header. */
2374 	SOCKBUF_LOCK(&so->so_rcv);
2375 	params.tls_rx.tls_hdr_tcp_sn =
2376 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2377 	params.tls_rx.tls_rec_length = tls_len;
2378 	params.tls_rx.tls_seq_number = tls_rcd_num;
2379 	SOCKBUF_UNLOCK(&so->so_rcv);
2380 
2381 	INP_RUNLOCK(inp);
2382 
2383 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2384 	return (mst->sw->snd_tag_modify(mst, &params));
2385 }
2386 
2387 static void
2388 ktls_drop(struct socket *so, int error)
2389 {
2390 	struct epoch_tracker et;
2391 	struct inpcb *inp = sotoinpcb(so);
2392 	struct tcpcb *tp;
2393 
2394 	NET_EPOCH_ENTER(et);
2395 	INP_WLOCK(inp);
2396 	if (!(inp->inp_flags & INP_DROPPED)) {
2397 		tp = intotcpcb(inp);
2398 		CURVNET_SET(inp->inp_vnet);
2399 		tp = tcp_drop(tp, error);
2400 		CURVNET_RESTORE();
2401 		if (tp != NULL)
2402 			INP_WUNLOCK(inp);
2403 	} else {
2404 		so->so_error = error;
2405 		SOCK_RECVBUF_LOCK(so);
2406 		sorwakeup_locked(so);
2407 		INP_WUNLOCK(inp);
2408 	}
2409 	NET_EPOCH_EXIT(et);
2410 }
2411 
2412 static void
2413 ktls_decrypt(struct socket *so)
2414 {
2415 	char tls_header[MBUF_PEXT_HDR_LEN];
2416 	struct ktls_session *tls;
2417 	struct sockbuf *sb;
2418 	struct tls_record_layer *hdr;
2419 	struct tls_get_record tgr;
2420 	struct mbuf *control, *data, *m;
2421 	ktls_mbuf_crypto_st_t state;
2422 	uint64_t seqno;
2423 	int error, remain, tls_len, trail_len;
2424 	bool tls13;
2425 	uint8_t vminor, record_type;
2426 
2427 	hdr = (struct tls_record_layer *)tls_header;
2428 	sb = &so->so_rcv;
2429 	SOCKBUF_LOCK(sb);
2430 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2431 	    ("%s: socket %p not running", __func__, so));
2432 
2433 	tls = sb->sb_tls_info;
2434 	MPASS(tls != NULL);
2435 
2436 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2437 	if (tls13)
2438 		vminor = TLS_MINOR_VER_TWO;
2439 	else
2440 		vminor = tls->params.tls_vminor;
2441 	for (;;) {
2442 		/* Is there enough queued for a TLS header? */
2443 		if (sb->sb_tlscc < tls->params.tls_hlen)
2444 			break;
2445 
2446 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2447 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2448 
2449 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2450 		    hdr->tls_vminor != vminor)
2451 			error = EINVAL;
2452 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2453 			error = EINVAL;
2454 		else if (tls_len < tls->params.tls_hlen || tls_len >
2455 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2456 		    tls->params.tls_tlen)
2457 			error = EMSGSIZE;
2458 		else
2459 			error = 0;
2460 		if (__predict_false(error != 0)) {
2461 			/*
2462 			 * We have a corrupted record and are likely
2463 			 * out of sync.  The connection isn't
2464 			 * recoverable at this point, so abort it.
2465 			 */
2466 			SOCKBUF_UNLOCK(sb);
2467 			counter_u64_add(ktls_offload_corrupted_records, 1);
2468 
2469 			ktls_drop(so, error);
2470 			goto deref;
2471 		}
2472 
2473 		/* Is the entire record queued? */
2474 		if (sb->sb_tlscc < tls_len)
2475 			break;
2476 
2477 		/*
2478 		 * Split out the portion of the mbuf chain containing
2479 		 * this TLS record.
2480 		 */
2481 		data = ktls_detach_record(sb, tls_len);
2482 		if (data == NULL)
2483 			continue;
2484 		MPASS(sb->sb_tlsdcc == tls_len);
2485 
2486 		seqno = sb->sb_tls_seqno;
2487 		sb->sb_tls_seqno++;
2488 		SBCHECK(sb);
2489 		SOCKBUF_UNLOCK(sb);
2490 
2491 		/* get crypto state for this TLS record */
2492 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2493 
2494 		switch (state) {
2495 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2496 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2497 			if (error)
2498 				break;
2499 			/* FALLTHROUGH */
2500 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2501 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2502 			    &trail_len);
2503 			if (__predict_true(error == 0)) {
2504 				if (tls13) {
2505 					error = tls13_find_record_type(tls, data,
2506 					    tls_len, &trail_len, &record_type);
2507 				} else {
2508 					record_type = hdr->tls_type;
2509 				}
2510 			}
2511 			break;
2512 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2513 			/*
2514 			 * NIC TLS is only supported for AEAD
2515 			 * ciphersuites which used a fixed sized
2516 			 * trailer.
2517 			 */
2518 			if (tls13) {
2519 				trail_len = tls->params.tls_tlen - 1;
2520 				error = tls13_find_record_type(tls, data,
2521 				    tls_len, &trail_len, &record_type);
2522 			} else {
2523 				trail_len = tls->params.tls_tlen;
2524 				error = 0;
2525 				record_type = hdr->tls_type;
2526 			}
2527 			break;
2528 		default:
2529 			error = EINVAL;
2530 			break;
2531 		}
2532 		if (error) {
2533 			counter_u64_add(ktls_offload_failed_crypto, 1);
2534 
2535 			SOCKBUF_LOCK(sb);
2536 			if (sb->sb_tlsdcc == 0) {
2537 				/*
2538 				 * sbcut/drop/flush discarded these
2539 				 * mbufs.
2540 				 */
2541 				m_freem(data);
2542 				break;
2543 			}
2544 
2545 			/*
2546 			 * Drop this TLS record's data, but keep
2547 			 * decrypting subsequent records.
2548 			 */
2549 			sb->sb_ccc -= tls_len;
2550 			sb->sb_tlsdcc = 0;
2551 
2552 			if (error != EMSGSIZE)
2553 				error = EBADMSG;
2554 			CURVNET_SET(so->so_vnet);
2555 			so->so_error = error;
2556 			sorwakeup_locked(so);
2557 			CURVNET_RESTORE();
2558 
2559 			m_freem(data);
2560 
2561 			SOCKBUF_LOCK(sb);
2562 			continue;
2563 		}
2564 
2565 		/* Allocate the control mbuf. */
2566 		memset(&tgr, 0, sizeof(tgr));
2567 		tgr.tls_type = record_type;
2568 		tgr.tls_vmajor = hdr->tls_vmajor;
2569 		tgr.tls_vminor = hdr->tls_vminor;
2570 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2571 		    trail_len);
2572 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2573 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2574 
2575 		SOCKBUF_LOCK(sb);
2576 		if (sb->sb_tlsdcc == 0) {
2577 			/* sbcut/drop/flush discarded these mbufs. */
2578 			MPASS(sb->sb_tlscc == 0);
2579 			m_freem(data);
2580 			m_freem(control);
2581 			break;
2582 		}
2583 
2584 		/*
2585 		 * Clear the 'dcc' accounting in preparation for
2586 		 * adding the decrypted record.
2587 		 */
2588 		sb->sb_ccc -= tls_len;
2589 		sb->sb_tlsdcc = 0;
2590 		SBCHECK(sb);
2591 
2592 		/* If there is no payload, drop all of the data. */
2593 		if (tgr.tls_length == htobe16(0)) {
2594 			m_freem(data);
2595 			data = NULL;
2596 		} else {
2597 			/* Trim header. */
2598 			remain = tls->params.tls_hlen;
2599 			while (remain > 0) {
2600 				if (data->m_len > remain) {
2601 					data->m_data += remain;
2602 					data->m_len -= remain;
2603 					break;
2604 				}
2605 				remain -= data->m_len;
2606 				data = m_free(data);
2607 			}
2608 
2609 			/* Trim trailer and clear M_NOTREADY. */
2610 			remain = be16toh(tgr.tls_length);
2611 			m = data;
2612 			for (m = data; remain > m->m_len; m = m->m_next) {
2613 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2614 				remain -= m->m_len;
2615 			}
2616 			m->m_len = remain;
2617 			m_freem(m->m_next);
2618 			m->m_next = NULL;
2619 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2620 
2621 			/* Set EOR on the final mbuf. */
2622 			m->m_flags |= M_EOR;
2623 		}
2624 
2625 		sbappendcontrol_locked(sb, data, control, 0);
2626 
2627 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2628 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2629 			SOCKBUF_UNLOCK(sb);
2630 			ktls_resync_ifnet(so, tls_len, seqno);
2631 			SOCKBUF_LOCK(sb);
2632 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2633 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2634 			SOCKBUF_UNLOCK(sb);
2635 			ktls_resync_ifnet(so, 0, seqno);
2636 			SOCKBUF_LOCK(sb);
2637 		}
2638 	}
2639 
2640 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2641 
2642 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2643 		so->so_error = EMSGSIZE;
2644 
2645 	sorwakeup_locked(so);
2646 
2647 deref:
2648 	SOCKBUF_UNLOCK_ASSERT(sb);
2649 
2650 	CURVNET_SET(so->so_vnet);
2651 	sorele(so);
2652 	CURVNET_RESTORE();
2653 }
2654 
2655 void
2656 ktls_enqueue_to_free(struct mbuf *m)
2657 {
2658 	struct ktls_wq *wq;
2659 	bool running;
2660 
2661 	/* Mark it for freeing. */
2662 	m->m_epg_flags |= EPG_FLAG_2FREE;
2663 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2664 	mtx_lock(&wq->mtx);
2665 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2666 	running = wq->running;
2667 	mtx_unlock(&wq->mtx);
2668 	if (!running)
2669 		wakeup(wq);
2670 }
2671 
2672 static void *
2673 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2674 {
2675 	void *buf;
2676 	int domain, running;
2677 
2678 	if (m->m_epg_npgs <= 2)
2679 		return (NULL);
2680 	if (ktls_buffer_zone == NULL)
2681 		return (NULL);
2682 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2683 		/*
2684 		 * Rate-limit allocation attempts after a failure.
2685 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2686 		 * the free page queues and may fail consistently if memory is
2687 		 * fragmented.
2688 		 */
2689 		return (NULL);
2690 	}
2691 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2692 	if (buf == NULL) {
2693 		domain = PCPU_GET(domain);
2694 		wq->lastallocfail = ticks;
2695 
2696 		/*
2697 		 * Note that this check is "racy", but the races are
2698 		 * harmless, and are either a spurious wakeup if
2699 		 * multiple threads fail allocations before the alloc
2700 		 * thread wakes, or waiting an extra second in case we
2701 		 * see an old value of running == true.
2702 		 */
2703 		if (!VM_DOMAIN_EMPTY(domain)) {
2704 			running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2705 			if (!running)
2706 				wakeup(&ktls_domains[domain].reclaim_td);
2707 		}
2708 	}
2709 	return (buf);
2710 }
2711 
2712 static int
2713 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2714     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2715 {
2716 	vm_page_t pg;
2717 	int error, i, len, off;
2718 
2719 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2720 	    ("%p not unready & nomap mbuf\n", m));
2721 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2722 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2723 	    ktls_maxlen));
2724 
2725 	/* Anonymous mbufs are encrypted in place. */
2726 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2727 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2728 
2729 	/*
2730 	 * For file-backed mbufs (from sendfile), anonymous wired
2731 	 * pages are allocated and used as the encryption destination.
2732 	 */
2733 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2734 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2735 		    m->m_epg_1st_off;
2736 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2737 		    m->m_epg_1st_off;
2738 		state->dst_iov[0].iov_len = len;
2739 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2740 		i = 1;
2741 	} else {
2742 		off = m->m_epg_1st_off;
2743 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2744 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2745 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2746 			len = m_epg_pagelen(m, i, off);
2747 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2748 			state->dst_iov[i].iov_base =
2749 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2750 			state->dst_iov[i].iov_len = len;
2751 		}
2752 	}
2753 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2754 	state->dst_iov[i].iov_base = m->m_epg_trail;
2755 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2756 
2757 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2758 
2759 	if (__predict_false(error != 0)) {
2760 		/* Free the anonymous pages. */
2761 		if (state->cbuf != NULL)
2762 			uma_zfree(ktls_buffer_zone, state->cbuf);
2763 		else {
2764 			for (i = 0; i < m->m_epg_npgs; i++) {
2765 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2766 				(void)vm_page_unwire_noq(pg);
2767 				vm_page_free(pg);
2768 			}
2769 		}
2770 	}
2771 	return (error);
2772 }
2773 
2774 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2775 static u_int
2776 ktls_batched_records(struct mbuf *m)
2777 {
2778 	int page_count, records;
2779 
2780 	records = 0;
2781 	page_count = m->m_epg_enc_cnt;
2782 	while (page_count > 0) {
2783 		records++;
2784 		page_count -= m->m_epg_nrdy;
2785 		m = m->m_next;
2786 	}
2787 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2788 	return (records);
2789 }
2790 
2791 void
2792 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2793 {
2794 	struct ktls_session *tls;
2795 	struct ktls_wq *wq;
2796 	int queued;
2797 	bool running;
2798 
2799 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2800 	    (M_EXTPG | M_NOTREADY)),
2801 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2802 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2803 
2804 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2805 
2806 	m->m_epg_enc_cnt = page_count;
2807 
2808 	/*
2809 	 * Save a pointer to the socket.  The caller is responsible
2810 	 * for taking an additional reference via soref().
2811 	 */
2812 	m->m_epg_so = so;
2813 
2814 	queued = 1;
2815 	tls = m->m_epg_tls;
2816 	wq = &ktls_wq[tls->wq_index];
2817 	mtx_lock(&wq->mtx);
2818 	if (__predict_false(tls->sequential_records)) {
2819 		/*
2820 		 * For TLS 1.0, records must be encrypted
2821 		 * sequentially.  For a given connection, all records
2822 		 * queued to the associated work queue are processed
2823 		 * sequentially.  However, sendfile(2) might complete
2824 		 * I/O requests spanning multiple TLS records out of
2825 		 * order.  Here we ensure TLS records are enqueued to
2826 		 * the work queue in FIFO order.
2827 		 *
2828 		 * tls->next_seqno holds the sequence number of the
2829 		 * next TLS record that should be enqueued to the work
2830 		 * queue.  If this next record is not tls->next_seqno,
2831 		 * it must be a future record, so insert it, sorted by
2832 		 * TLS sequence number, into tls->pending_records and
2833 		 * return.
2834 		 *
2835 		 * If this TLS record matches tls->next_seqno, place
2836 		 * it in the work queue and then check
2837 		 * tls->pending_records to see if any
2838 		 * previously-queued records are now ready for
2839 		 * encryption.
2840 		 */
2841 		if (m->m_epg_seqno != tls->next_seqno) {
2842 			struct mbuf *n, *p;
2843 
2844 			p = NULL;
2845 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2846 				if (n->m_epg_seqno > m->m_epg_seqno)
2847 					break;
2848 				p = n;
2849 			}
2850 			if (n == NULL)
2851 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2852 				    m_epg_stailq);
2853 			else if (p == NULL)
2854 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2855 				    m_epg_stailq);
2856 			else
2857 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2858 				    m_epg_stailq);
2859 			mtx_unlock(&wq->mtx);
2860 			counter_u64_add(ktls_cnt_tx_pending, 1);
2861 			return;
2862 		}
2863 
2864 		tls->next_seqno += ktls_batched_records(m);
2865 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2866 
2867 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2868 			struct mbuf *n;
2869 
2870 			n = STAILQ_FIRST(&tls->pending_records);
2871 			if (n->m_epg_seqno != tls->next_seqno)
2872 				break;
2873 
2874 			queued++;
2875 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2876 			tls->next_seqno += ktls_batched_records(n);
2877 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2878 		}
2879 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2880 	} else
2881 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2882 
2883 	running = wq->running;
2884 	mtx_unlock(&wq->mtx);
2885 	if (!running)
2886 		wakeup(wq);
2887 	counter_u64_add(ktls_cnt_tx_queued, queued);
2888 }
2889 
2890 /*
2891  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2892  * the pages from the file and replace them with the anonymous pages
2893  * allocated in ktls_encrypt_record().
2894  */
2895 static void
2896 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2897 {
2898 	int i;
2899 
2900 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2901 
2902 	/* Free the old pages. */
2903 	m->m_ext.ext_free(m);
2904 
2905 	/* Replace them with the new pages. */
2906 	if (state->cbuf != NULL) {
2907 		for (i = 0; i < m->m_epg_npgs; i++)
2908 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2909 
2910 		/* Contig pages should go back to the cache. */
2911 		m->m_ext.ext_free = ktls_free_mext_contig;
2912 	} else {
2913 		for (i = 0; i < m->m_epg_npgs; i++)
2914 			m->m_epg_pa[i] = state->parray[i];
2915 
2916 		/* Use the basic free routine. */
2917 		m->m_ext.ext_free = mb_free_mext_pgs;
2918 	}
2919 
2920 	/* Pages are now writable. */
2921 	m->m_epg_flags |= EPG_FLAG_ANON;
2922 }
2923 
2924 static __noinline void
2925 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2926 {
2927 	struct ktls_ocf_encrypt_state state;
2928 	struct ktls_session *tls;
2929 	struct socket *so;
2930 	struct mbuf *m;
2931 	int error, npages, total_pages;
2932 
2933 	so = top->m_epg_so;
2934 	tls = top->m_epg_tls;
2935 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2936 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2937 #ifdef INVARIANTS
2938 	top->m_epg_so = NULL;
2939 #endif
2940 	total_pages = top->m_epg_enc_cnt;
2941 	npages = 0;
2942 
2943 	/*
2944 	 * Encrypt the TLS records in the chain of mbufs starting with
2945 	 * 'top'.  'total_pages' gives us a total count of pages and is
2946 	 * used to know when we have finished encrypting the TLS
2947 	 * records originally queued with 'top'.
2948 	 *
2949 	 * NB: These mbufs are queued in the socket buffer and
2950 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2951 	 * socket buffer lock is not held while traversing this chain.
2952 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2953 	 * pointers should be stable.  However, the 'm_next' of the
2954 	 * last mbuf encrypted is not necessarily NULL.  It can point
2955 	 * to other mbufs appended while 'top' was on the TLS work
2956 	 * queue.
2957 	 *
2958 	 * Each mbuf holds an entire TLS record.
2959 	 */
2960 	error = 0;
2961 	for (m = top; npages != total_pages; m = m->m_next) {
2962 		KASSERT(m->m_epg_tls == tls,
2963 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2964 		    tls, m->m_epg_tls));
2965 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2966 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2967 		    total_pages, m));
2968 
2969 		error = ktls_encrypt_record(wq, m, tls, &state);
2970 		if (error) {
2971 			counter_u64_add(ktls_offload_failed_crypto, 1);
2972 			break;
2973 		}
2974 
2975 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2976 			ktls_finish_nonanon(m, &state);
2977 
2978 		npages += m->m_epg_nrdy;
2979 
2980 		/*
2981 		 * Drop a reference to the session now that it is no
2982 		 * longer needed.  Existing code depends on encrypted
2983 		 * records having no associated session vs
2984 		 * yet-to-be-encrypted records having an associated
2985 		 * session.
2986 		 */
2987 		m->m_epg_tls = NULL;
2988 		ktls_free(tls);
2989 	}
2990 
2991 	CURVNET_SET(so->so_vnet);
2992 	if (error == 0) {
2993 		(void)so->so_proto->pr_ready(so, top, npages);
2994 	} else {
2995 		ktls_drop(so, EIO);
2996 		mb_free_notready(top, total_pages);
2997 	}
2998 
2999 	sorele(so);
3000 	CURVNET_RESTORE();
3001 }
3002 
3003 void
3004 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3005 {
3006 	struct ktls_session *tls;
3007 	struct socket *so;
3008 	struct mbuf *m;
3009 	int npages;
3010 
3011 	m = state->m;
3012 
3013 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3014 		ktls_finish_nonanon(m, state);
3015 
3016 	so = state->so;
3017 	free(state, M_KTLS);
3018 
3019 	/*
3020 	 * Drop a reference to the session now that it is no longer
3021 	 * needed.  Existing code depends on encrypted records having
3022 	 * no associated session vs yet-to-be-encrypted records having
3023 	 * an associated session.
3024 	 */
3025 	tls = m->m_epg_tls;
3026 	m->m_epg_tls = NULL;
3027 	ktls_free(tls);
3028 
3029 	if (error != 0)
3030 		counter_u64_add(ktls_offload_failed_crypto, 1);
3031 
3032 	CURVNET_SET(so->so_vnet);
3033 	npages = m->m_epg_nrdy;
3034 
3035 	if (error == 0) {
3036 		(void)so->so_proto->pr_ready(so, m, npages);
3037 	} else {
3038 		ktls_drop(so, EIO);
3039 		mb_free_notready(m, npages);
3040 	}
3041 
3042 	sorele(so);
3043 	CURVNET_RESTORE();
3044 }
3045 
3046 /*
3047  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3048  * (coprocessors) where encryption does not use host CPU resources and
3049  * it can be beneficial to queue more requests than CPUs.
3050  */
3051 static __noinline void
3052 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3053 {
3054 	struct ktls_ocf_encrypt_state *state;
3055 	struct ktls_session *tls;
3056 	struct socket *so;
3057 	struct mbuf *m, *n;
3058 	int error, mpages, npages, total_pages;
3059 
3060 	so = top->m_epg_so;
3061 	tls = top->m_epg_tls;
3062 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3063 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3064 #ifdef INVARIANTS
3065 	top->m_epg_so = NULL;
3066 #endif
3067 	total_pages = top->m_epg_enc_cnt;
3068 	npages = 0;
3069 
3070 	error = 0;
3071 	for (m = top; npages != total_pages; m = n) {
3072 		KASSERT(m->m_epg_tls == tls,
3073 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3074 		    tls, m->m_epg_tls));
3075 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3076 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3077 		    total_pages, m));
3078 
3079 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3080 		soref(so);
3081 		state->so = so;
3082 		state->m = m;
3083 
3084 		mpages = m->m_epg_nrdy;
3085 		n = m->m_next;
3086 
3087 		error = ktls_encrypt_record(wq, m, tls, state);
3088 		if (error) {
3089 			counter_u64_add(ktls_offload_failed_crypto, 1);
3090 			free(state, M_KTLS);
3091 			CURVNET_SET(so->so_vnet);
3092 			sorele(so);
3093 			CURVNET_RESTORE();
3094 			break;
3095 		}
3096 
3097 		npages += mpages;
3098 	}
3099 
3100 	CURVNET_SET(so->so_vnet);
3101 	if (error != 0) {
3102 		ktls_drop(so, EIO);
3103 		mb_free_notready(m, total_pages - npages);
3104 	}
3105 
3106 	sorele(so);
3107 	CURVNET_RESTORE();
3108 }
3109 
3110 static int
3111 ktls_bind_domain(int domain)
3112 {
3113 	int error;
3114 
3115 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3116 	if (error != 0)
3117 		return (error);
3118 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3119 	return (0);
3120 }
3121 
3122 static void
3123 ktls_reclaim_thread(void *ctx)
3124 {
3125 	struct ktls_domain_info *ktls_domain = ctx;
3126 	struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3127 	struct sysctl_oid *oid;
3128 	char name[80];
3129 	int error, domain;
3130 
3131 	domain = ktls_domain - ktls_domains;
3132 	if (bootverbose)
3133 		printf("Starting KTLS reclaim thread for domain %d\n", domain);
3134 	error = ktls_bind_domain(domain);
3135 	if (error)
3136 		printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3137 		    domain, error);
3138 	snprintf(name, sizeof(name), "domain%d", domain);
3139 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3140 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3141 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3142 	    CTLFLAG_RD,  &sc->reclaims, 0, "buffers reclaimed");
3143 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3144 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3145 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3146 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3147 
3148 	for (;;) {
3149 		atomic_store_int(&sc->running, 0);
3150 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3151 		atomic_store_int(&sc->running, 1);
3152 		sc->wakeups++;
3153 		/*
3154 		 * Below we attempt to reclaim ktls_max_reclaim
3155 		 * buffers using vm_page_reclaim_contig_domain_ext().
3156 		 * We do this here, as this function can take several
3157 		 * seconds to scan all of memory and it does not
3158 		 * matter if this thread pauses for a while.  If we
3159 		 * block a ktls worker thread, we risk developing
3160 		 * backlogs of buffers to be encrypted, leading to
3161 		 * surges of traffic and potential NIC output drops.
3162 		 */
3163 		if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3164 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3165 		    ktls_max_reclaim) != 0) {
3166 			vm_wait_domain(domain);
3167 		} else {
3168 			sc->reclaims += ktls_max_reclaim;
3169 		}
3170 	}
3171 }
3172 
3173 static void
3174 ktls_work_thread(void *ctx)
3175 {
3176 	struct ktls_wq *wq = ctx;
3177 	struct mbuf *m, *n;
3178 	struct socket *so, *son;
3179 	STAILQ_HEAD(, mbuf) local_m_head;
3180 	STAILQ_HEAD(, socket) local_so_head;
3181 	int cpu;
3182 
3183 	cpu = wq - ktls_wq;
3184 	if (bootverbose)
3185 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3186 
3187 	/*
3188 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3189 	 * we bind to the NUMA domain instead.
3190 	 */
3191 	if (ktls_bind_threads) {
3192 		int error;
3193 
3194 		if (ktls_bind_threads > 1) {
3195 			struct pcpu *pc = pcpu_find(cpu);
3196 
3197 			error = ktls_bind_domain(pc->pc_domain);
3198 		} else {
3199 			cpuset_t mask;
3200 
3201 			CPU_SETOF(cpu, &mask);
3202 			error = cpuset_setthread(curthread->td_tid, &mask);
3203 		}
3204 		if (error)
3205 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3206 				cpu, error);
3207 	}
3208 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3209 	fpu_kern_thread(0);
3210 #endif
3211 	for (;;) {
3212 		mtx_lock(&wq->mtx);
3213 		while (STAILQ_EMPTY(&wq->m_head) &&
3214 		    STAILQ_EMPTY(&wq->so_head)) {
3215 			wq->running = false;
3216 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3217 			wq->running = true;
3218 		}
3219 
3220 		STAILQ_INIT(&local_m_head);
3221 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3222 		STAILQ_INIT(&local_so_head);
3223 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3224 		mtx_unlock(&wq->mtx);
3225 
3226 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3227 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3228 				ktls_free(m->m_epg_tls);
3229 				m_free_raw(m);
3230 			} else {
3231 				if (m->m_epg_tls->sync_dispatch)
3232 					ktls_encrypt(wq, m);
3233 				else
3234 					ktls_encrypt_async(wq, m);
3235 				counter_u64_add(ktls_cnt_tx_queued, -1);
3236 			}
3237 		}
3238 
3239 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3240 			ktls_decrypt(so);
3241 			counter_u64_add(ktls_cnt_rx_queued, -1);
3242 		}
3243 	}
3244 }
3245 
3246 static void
3247 ktls_disable_ifnet_help(void *context, int pending __unused)
3248 {
3249 	struct ktls_session *tls;
3250 	struct inpcb *inp;
3251 	struct tcpcb *tp;
3252 	struct socket *so;
3253 	int err;
3254 
3255 	tls = context;
3256 	inp = tls->inp;
3257 	if (inp == NULL)
3258 		return;
3259 	INP_WLOCK(inp);
3260 	so = inp->inp_socket;
3261 	MPASS(so != NULL);
3262 	if (inp->inp_flags & INP_DROPPED) {
3263 		goto out;
3264 	}
3265 
3266 	if (so->so_snd.sb_tls_info != NULL)
3267 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3268 	else
3269 		err = ENXIO;
3270 	if (err == 0) {
3271 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3272 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3273 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3274 		    (tp = intotcpcb(inp)) != NULL &&
3275 		    tp->t_fb->tfb_hwtls_change != NULL)
3276 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3277 	} else {
3278 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3279 	}
3280 
3281 out:
3282 	CURVNET_SET(so->so_vnet);
3283 	sorele(so);
3284 	CURVNET_RESTORE();
3285 	INP_WUNLOCK(inp);
3286 	ktls_free(tls);
3287 }
3288 
3289 /*
3290  * Called when re-transmits are becoming a substantial portion of the
3291  * sends on this connection.  When this happens, we transition the
3292  * connection to software TLS.  This is needed because most inline TLS
3293  * NICs keep crypto state only for in-order transmits.  This means
3294  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3295  * re-DMA the entire TLS record up to and including the current
3296  * segment.  This means that when re-transmitting the last ~1448 byte
3297  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3298  * of magnitude more data than we are sending.  This can cause the
3299  * PCIe link to saturate well before the network, which can cause
3300  * output drops, and a general loss of capacity.
3301  */
3302 void
3303 ktls_disable_ifnet(void *arg)
3304 {
3305 	struct tcpcb *tp;
3306 	struct inpcb *inp;
3307 	struct socket *so;
3308 	struct ktls_session *tls;
3309 
3310 	tp = arg;
3311 	inp = tptoinpcb(tp);
3312 	INP_WLOCK_ASSERT(inp);
3313 	so = inp->inp_socket;
3314 	SOCK_LOCK(so);
3315 	tls = so->so_snd.sb_tls_info;
3316 	if (tp->t_nic_ktls_xmit_dis == 1) {
3317 		SOCK_UNLOCK(so);
3318 		return;
3319 	}
3320 
3321 	/*
3322 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3323 	 * ifnet can only be done once per connection, so we never want
3324 	 * to do it again
3325 	 */
3326 
3327 	(void)ktls_hold(tls);
3328 	soref(so);
3329 	tp->t_nic_ktls_xmit_dis = 1;
3330 	SOCK_UNLOCK(so);
3331 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3332 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3333 }
3334