xref: /freebsd/sys/kern/uipc_ktls.c (revision 1ca12bd927d76f5a1b9aad2c72f673a3b01919c0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/domainset.h>
40 #include <sys/endian.h>
41 #include <sys/ktls.h>
42 #include <sys/lock.h>
43 #include <sys/mbuf.h>
44 #include <sys/mutex.h>
45 #include <sys/rmlock.h>
46 #include <sys/proc.h>
47 #include <sys/protosw.h>
48 #include <sys/refcount.h>
49 #include <sys/smp.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/kthread.h>
55 #include <sys/uio.h>
56 #include <sys/vmmeter.h>
57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58 #include <machine/pcb.h>
59 #endif
60 #include <machine/vmparam.h>
61 #include <net/if.h>
62 #include <net/if_var.h>
63 #ifdef RSS
64 #include <net/netisr.h>
65 #include <net/rss_config.h>
66 #endif
67 #include <net/route.h>
68 #include <net/route/nhop.h>
69 #include <netinet/in.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/tcp_var.h>
72 #ifdef TCP_OFFLOAD
73 #include <netinet/tcp_offload.h>
74 #endif
75 #include <opencrypto/cryptodev.h>
76 #include <opencrypto/ktls.h>
77 #include <vm/uma_dbg.h>
78 #include <vm/vm.h>
79 #include <vm/vm_pageout.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pagequeue.h>
82 
83 struct ktls_wq {
84 	struct mtx	mtx;
85 	STAILQ_HEAD(, mbuf) m_head;
86 	STAILQ_HEAD(, socket) so_head;
87 	bool		running;
88 	int		lastallocfail;
89 } __aligned(CACHE_LINE_SIZE);
90 
91 struct ktls_alloc_thread {
92 	uint64_t wakeups;
93 	uint64_t allocs;
94 	struct thread *td;
95 	int running;
96 };
97 
98 struct ktls_domain_info {
99 	int count;
100 	int cpu[MAXCPU];
101 	struct ktls_alloc_thread alloc_td;
102 };
103 
104 struct ktls_domain_info ktls_domains[MAXMEMDOM];
105 static struct ktls_wq *ktls_wq;
106 static struct proc *ktls_proc;
107 static uma_zone_t ktls_session_zone;
108 static uma_zone_t ktls_buffer_zone;
109 static uint16_t ktls_cpuid_lookup[MAXCPU];
110 static int ktls_init_state;
111 static struct sx ktls_init_lock;
112 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
113 
114 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
115     "Kernel TLS offload");
116 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117     "Kernel TLS offload stats");
118 
119 #ifdef RSS
120 static int ktls_bind_threads = 1;
121 #else
122 static int ktls_bind_threads;
123 #endif
124 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
125     &ktls_bind_threads, 0,
126     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
127 
128 static u_int ktls_maxlen = 16384;
129 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
130     &ktls_maxlen, 0, "Maximum TLS record size");
131 
132 static int ktls_number_threads;
133 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
134     &ktls_number_threads, 0,
135     "Number of TLS threads in thread-pool");
136 
137 unsigned int ktls_ifnet_max_rexmit_pct = 2;
138 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
139     &ktls_ifnet_max_rexmit_pct, 2,
140     "Max percent bytes retransmitted before ifnet TLS is disabled");
141 
142 static bool ktls_offload_enable;
143 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
144     &ktls_offload_enable, 0,
145     "Enable support for kernel TLS offload");
146 
147 static bool ktls_cbc_enable = true;
148 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
149     &ktls_cbc_enable, 1,
150     "Enable support of AES-CBC crypto for kernel TLS");
151 
152 static bool ktls_sw_buffer_cache = true;
153 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
154     &ktls_sw_buffer_cache, 1,
155     "Enable caching of output buffers for SW encryption");
156 
157 static int ktls_max_alloc = 128;
158 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
159     &ktls_max_alloc, 128,
160     "Max number of 16k buffers to allocate in thread context");
161 
162 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
163 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
164     &ktls_tasks_active, "Number of active tasks");
165 
166 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
167 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
168     &ktls_cnt_tx_pending,
169     "Number of TLS 1.0 records waiting for earlier TLS records");
170 
171 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
172 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
173     &ktls_cnt_tx_queued,
174     "Number of TLS records in queue to tasks for SW encryption");
175 
176 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
177 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
178     &ktls_cnt_rx_queued,
179     "Number of TLS sockets in queue to tasks for SW decryption");
180 
181 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
182 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
183     CTLFLAG_RD, &ktls_offload_total,
184     "Total successful TLS setups (parameters set)");
185 
186 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
187 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
188     CTLFLAG_RD, &ktls_offload_enable_calls,
189     "Total number of TLS enable calls made");
190 
191 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
192 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
193     &ktls_offload_active, "Total Active TLS sessions");
194 
195 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
197     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
198 
199 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
200 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
201     &ktls_offload_failed_crypto, "Total TLS crypto failures");
202 
203 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
204 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
205     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
206 
207 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
208 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
209     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
210 
211 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
212 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
213     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
214 
215 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
217     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
218 
219 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
220 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
221     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
222 
223 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
224 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
225     &ktls_destroy_task,
226     "Number of times ktls session was destroyed via taskqueue");
227 
228 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
229     "Software TLS session stats");
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "Hardware (ifnet) TLS session stats");
232 #ifdef TCP_OFFLOAD
233 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
234     "TOE TLS session stats");
235 #endif
236 
237 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
238 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
239     "Active number of software TLS sessions using AES-CBC");
240 
241 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
242 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
243     "Active number of software TLS sessions using AES-GCM");
244 
245 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
246 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
247     &ktls_sw_chacha20,
248     "Active number of software TLS sessions using Chacha20-Poly1305");
249 
250 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
251 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
252     &ktls_ifnet_cbc,
253     "Active number of ifnet TLS sessions using AES-CBC");
254 
255 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
256 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
257     &ktls_ifnet_gcm,
258     "Active number of ifnet TLS sessions using AES-GCM");
259 
260 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
261 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
262     &ktls_ifnet_chacha20,
263     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
264 
265 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
266 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
267     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
268 
269 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
270 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
271     &ktls_ifnet_reset_dropped,
272     "TLS sessions dropped after failing to update ifnet send tag");
273 
274 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
275 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
276     &ktls_ifnet_reset_failed,
277     "TLS sessions that failed to allocate a new ifnet send tag");
278 
279 static int ktls_ifnet_permitted;
280 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
281     &ktls_ifnet_permitted, 1,
282     "Whether to permit hardware (ifnet) TLS sessions");
283 
284 #ifdef TCP_OFFLOAD
285 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
286 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
287     &ktls_toe_cbc,
288     "Active number of TOE TLS sessions using AES-CBC");
289 
290 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
291 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
292     &ktls_toe_gcm,
293     "Active number of TOE TLS sessions using AES-GCM");
294 
295 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
296 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
297     &ktls_toe_chacha20,
298     "Active number of TOE TLS sessions using Chacha20-Poly1305");
299 #endif
300 
301 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
302 
303 static void ktls_reset_receive_tag(void *context, int pending);
304 static void ktls_reset_send_tag(void *context, int pending);
305 static void ktls_work_thread(void *ctx);
306 static void ktls_alloc_thread(void *ctx);
307 
308 static u_int
309 ktls_get_cpu(struct socket *so)
310 {
311 	struct inpcb *inp;
312 #ifdef NUMA
313 	struct ktls_domain_info *di;
314 #endif
315 	u_int cpuid;
316 
317 	inp = sotoinpcb(so);
318 #ifdef RSS
319 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
320 	if (cpuid != NETISR_CPUID_NONE)
321 		return (cpuid);
322 #endif
323 	/*
324 	 * Just use the flowid to shard connections in a repeatable
325 	 * fashion.  Note that TLS 1.0 sessions rely on the
326 	 * serialization provided by having the same connection use
327 	 * the same queue.
328 	 */
329 #ifdef NUMA
330 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
331 		di = &ktls_domains[inp->inp_numa_domain];
332 		cpuid = di->cpu[inp->inp_flowid % di->count];
333 	} else
334 #endif
335 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
336 	return (cpuid);
337 }
338 
339 static int
340 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
341 {
342 	vm_page_t m;
343 	int i, req;
344 
345 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
346 	    ("%s: ktls max length %d is not page size-aligned",
347 	    __func__, ktls_maxlen));
348 
349 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
350 	for (i = 0; i < count; i++) {
351 		m = vm_page_alloc_noobj_contig_domain(domain, req,
352 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
353 		    VM_MEMATTR_DEFAULT);
354 		if (m == NULL)
355 			break;
356 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
357 	}
358 	return (i);
359 }
360 
361 static void
362 ktls_buffer_release(void *arg __unused, void **store, int count)
363 {
364 	vm_page_t m;
365 	int i, j;
366 
367 	for (i = 0; i < count; i++) {
368 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
369 		for (j = 0; j < atop(ktls_maxlen); j++) {
370 			(void)vm_page_unwire_noq(m + j);
371 			vm_page_free(m + j);
372 		}
373 	}
374 }
375 
376 static void
377 ktls_free_mext_contig(struct mbuf *m)
378 {
379 	M_ASSERTEXTPG(m);
380 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
381 }
382 
383 static int
384 ktls_init(void)
385 {
386 	struct thread *td;
387 	struct pcpu *pc;
388 	int count, domain, error, i;
389 
390 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
391 	    M_WAITOK | M_ZERO);
392 
393 	ktls_session_zone = uma_zcreate("ktls_session",
394 	    sizeof(struct ktls_session),
395 	    NULL, NULL, NULL, NULL,
396 	    UMA_ALIGN_CACHE, 0);
397 
398 	if (ktls_sw_buffer_cache) {
399 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
400 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
401 		    ktls_buffer_import, ktls_buffer_release, NULL,
402 		    UMA_ZONE_FIRSTTOUCH);
403 	}
404 
405 	/*
406 	 * Initialize the workqueues to run the TLS work.  We create a
407 	 * work queue for each CPU.
408 	 */
409 	CPU_FOREACH(i) {
410 		STAILQ_INIT(&ktls_wq[i].m_head);
411 		STAILQ_INIT(&ktls_wq[i].so_head);
412 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
413 		if (ktls_bind_threads > 1) {
414 			pc = pcpu_find(i);
415 			domain = pc->pc_domain;
416 			count = ktls_domains[domain].count;
417 			ktls_domains[domain].cpu[count] = i;
418 			ktls_domains[domain].count++;
419 		}
420 		ktls_cpuid_lookup[ktls_number_threads] = i;
421 		ktls_number_threads++;
422 	}
423 
424 	/*
425 	 * If we somehow have an empty domain, fall back to choosing
426 	 * among all KTLS threads.
427 	 */
428 	if (ktls_bind_threads > 1) {
429 		for (i = 0; i < vm_ndomains; i++) {
430 			if (ktls_domains[i].count == 0) {
431 				ktls_bind_threads = 1;
432 				break;
433 			}
434 		}
435 	}
436 
437 	/* Start kthreads for each workqueue. */
438 	CPU_FOREACH(i) {
439 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
440 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
441 		if (error) {
442 			printf("Can't add KTLS thread %d error %d\n", i, error);
443 			return (error);
444 		}
445 	}
446 
447 	/*
448 	 * Start an allocation thread per-domain to perform blocking allocations
449 	 * of 16k physically contiguous TLS crypto destination buffers.
450 	 */
451 	if (ktls_sw_buffer_cache) {
452 		for (domain = 0; domain < vm_ndomains; domain++) {
453 			if (VM_DOMAIN_EMPTY(domain))
454 				continue;
455 			if (CPU_EMPTY(&cpuset_domain[domain]))
456 				continue;
457 			error = kproc_kthread_add(ktls_alloc_thread,
458 			    &ktls_domains[domain], &ktls_proc,
459 			    &ktls_domains[domain].alloc_td.td,
460 			    0, 0, "KTLS", "alloc_%d", domain);
461 			if (error) {
462 				printf("Can't add KTLS alloc thread %d error %d\n",
463 				    domain, error);
464 				return (error);
465 			}
466 		}
467 	}
468 
469 	if (bootverbose)
470 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
471 	return (0);
472 }
473 
474 static int
475 ktls_start_kthreads(void)
476 {
477 	int error, state;
478 
479 start:
480 	state = atomic_load_acq_int(&ktls_init_state);
481 	if (__predict_true(state > 0))
482 		return (0);
483 	if (state < 0)
484 		return (ENXIO);
485 
486 	sx_xlock(&ktls_init_lock);
487 	if (ktls_init_state != 0) {
488 		sx_xunlock(&ktls_init_lock);
489 		goto start;
490 	}
491 
492 	error = ktls_init();
493 	if (error == 0)
494 		state = 1;
495 	else
496 		state = -1;
497 	atomic_store_rel_int(&ktls_init_state, state);
498 	sx_xunlock(&ktls_init_lock);
499 	return (error);
500 }
501 
502 static int
503 ktls_create_session(struct socket *so, struct tls_enable *en,
504     struct ktls_session **tlsp, int direction)
505 {
506 	struct ktls_session *tls;
507 	int error;
508 
509 	/* Only TLS 1.0 - 1.3 are supported. */
510 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
511 		return (EINVAL);
512 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
513 	    en->tls_vminor > TLS_MINOR_VER_THREE)
514 		return (EINVAL);
515 
516 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
517 		return (EINVAL);
518 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
519 		return (EINVAL);
520 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
521 		return (EINVAL);
522 
523 	/* All supported algorithms require a cipher key. */
524 	if (en->cipher_key_len == 0)
525 		return (EINVAL);
526 
527 	/* No flags are currently supported. */
528 	if (en->flags != 0)
529 		return (EINVAL);
530 
531 	/* Common checks for supported algorithms. */
532 	switch (en->cipher_algorithm) {
533 	case CRYPTO_AES_NIST_GCM_16:
534 		/*
535 		 * auth_algorithm isn't used, but permit GMAC values
536 		 * for compatibility.
537 		 */
538 		switch (en->auth_algorithm) {
539 		case 0:
540 #ifdef COMPAT_FREEBSD12
541 		/* XXX: Really 13.0-current COMPAT. */
542 		case CRYPTO_AES_128_NIST_GMAC:
543 		case CRYPTO_AES_192_NIST_GMAC:
544 		case CRYPTO_AES_256_NIST_GMAC:
545 #endif
546 			break;
547 		default:
548 			return (EINVAL);
549 		}
550 		if (en->auth_key_len != 0)
551 			return (EINVAL);
552 		switch (en->tls_vminor) {
553 		case TLS_MINOR_VER_TWO:
554 			if (en->iv_len != TLS_AEAD_GCM_LEN)
555 				return (EINVAL);
556 			break;
557 		case TLS_MINOR_VER_THREE:
558 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
559 				return (EINVAL);
560 			break;
561 		default:
562 			return (EINVAL);
563 		}
564 		break;
565 	case CRYPTO_AES_CBC:
566 		switch (en->auth_algorithm) {
567 		case CRYPTO_SHA1_HMAC:
568 			break;
569 		case CRYPTO_SHA2_256_HMAC:
570 		case CRYPTO_SHA2_384_HMAC:
571 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
572 				return (EINVAL);
573 			break;
574 		default:
575 			return (EINVAL);
576 		}
577 		if (en->auth_key_len == 0)
578 			return (EINVAL);
579 
580 		/*
581 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
582 		 * use explicit IVs.
583 		 */
584 		switch (en->tls_vminor) {
585 		case TLS_MINOR_VER_ZERO:
586 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
587 				return (EINVAL);
588 			break;
589 		case TLS_MINOR_VER_ONE:
590 		case TLS_MINOR_VER_TWO:
591 			/* Ignore any supplied IV. */
592 			en->iv_len = 0;
593 			break;
594 		default:
595 			return (EINVAL);
596 		}
597 		break;
598 	case CRYPTO_CHACHA20_POLY1305:
599 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
600 			return (EINVAL);
601 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
602 		    en->tls_vminor != TLS_MINOR_VER_THREE)
603 			return (EINVAL);
604 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
605 			return (EINVAL);
606 		break;
607 	default:
608 		return (EINVAL);
609 	}
610 
611 	error = ktls_start_kthreads();
612 	if (error != 0)
613 		return (error);
614 
615 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
616 
617 	counter_u64_add(ktls_offload_active, 1);
618 
619 	refcount_init(&tls->refcount, 1);
620 	if (direction == KTLS_RX) {
621 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
622 	} else {
623 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
624 		tls->inp = so->so_pcb;
625 		in_pcbref(tls->inp);
626 		tls->tx = true;
627 	}
628 
629 	tls->wq_index = ktls_get_cpu(so);
630 
631 	tls->params.cipher_algorithm = en->cipher_algorithm;
632 	tls->params.auth_algorithm = en->auth_algorithm;
633 	tls->params.tls_vmajor = en->tls_vmajor;
634 	tls->params.tls_vminor = en->tls_vminor;
635 	tls->params.flags = en->flags;
636 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
637 
638 	/* Set the header and trailer lengths. */
639 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
640 	switch (en->cipher_algorithm) {
641 	case CRYPTO_AES_NIST_GCM_16:
642 		/*
643 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
644 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
645 		 */
646 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
647 			tls->params.tls_hlen += sizeof(uint64_t);
648 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
649 		tls->params.tls_bs = 1;
650 		break;
651 	case CRYPTO_AES_CBC:
652 		switch (en->auth_algorithm) {
653 		case CRYPTO_SHA1_HMAC:
654 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
655 				/* Implicit IV, no nonce. */
656 				tls->sequential_records = true;
657 				tls->next_seqno = be64dec(en->rec_seq);
658 				STAILQ_INIT(&tls->pending_records);
659 			} else {
660 				tls->params.tls_hlen += AES_BLOCK_LEN;
661 			}
662 			tls->params.tls_tlen = AES_BLOCK_LEN +
663 			    SHA1_HASH_LEN;
664 			break;
665 		case CRYPTO_SHA2_256_HMAC:
666 			tls->params.tls_hlen += AES_BLOCK_LEN;
667 			tls->params.tls_tlen = AES_BLOCK_LEN +
668 			    SHA2_256_HASH_LEN;
669 			break;
670 		case CRYPTO_SHA2_384_HMAC:
671 			tls->params.tls_hlen += AES_BLOCK_LEN;
672 			tls->params.tls_tlen = AES_BLOCK_LEN +
673 			    SHA2_384_HASH_LEN;
674 			break;
675 		default:
676 			panic("invalid hmac");
677 		}
678 		tls->params.tls_bs = AES_BLOCK_LEN;
679 		break;
680 	case CRYPTO_CHACHA20_POLY1305:
681 		/*
682 		 * Chacha20 uses a 12 byte implicit IV.
683 		 */
684 		tls->params.tls_tlen = POLY1305_HASH_LEN;
685 		tls->params.tls_bs = 1;
686 		break;
687 	default:
688 		panic("invalid cipher");
689 	}
690 
691 	/*
692 	 * TLS 1.3 includes optional padding which we do not support,
693 	 * and also puts the "real" record type at the end of the
694 	 * encrypted data.
695 	 */
696 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
697 		tls->params.tls_tlen += sizeof(uint8_t);
698 
699 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
700 	    ("TLS header length too long: %d", tls->params.tls_hlen));
701 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
702 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
703 
704 	if (en->auth_key_len != 0) {
705 		tls->params.auth_key_len = en->auth_key_len;
706 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
707 		    M_WAITOK);
708 		error = copyin(en->auth_key, tls->params.auth_key,
709 		    en->auth_key_len);
710 		if (error)
711 			goto out;
712 	}
713 
714 	tls->params.cipher_key_len = en->cipher_key_len;
715 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
716 	error = copyin(en->cipher_key, tls->params.cipher_key,
717 	    en->cipher_key_len);
718 	if (error)
719 		goto out;
720 
721 	/*
722 	 * This holds the implicit portion of the nonce for AEAD
723 	 * ciphers and the initial implicit IV for TLS 1.0.  The
724 	 * explicit portions of the IV are generated in ktls_frame().
725 	 */
726 	if (en->iv_len != 0) {
727 		tls->params.iv_len = en->iv_len;
728 		error = copyin(en->iv, tls->params.iv, en->iv_len);
729 		if (error)
730 			goto out;
731 
732 		/*
733 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
734 		 * counter to generate unique explicit IVs.
735 		 *
736 		 * Store this counter in the last 8 bytes of the IV
737 		 * array so that it is 8-byte aligned.
738 		 */
739 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
740 		    en->tls_vminor == TLS_MINOR_VER_TWO)
741 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
742 	}
743 
744 	*tlsp = tls;
745 	return (0);
746 
747 out:
748 	ktls_free(tls);
749 	return (error);
750 }
751 
752 static struct ktls_session *
753 ktls_clone_session(struct ktls_session *tls, int direction)
754 {
755 	struct ktls_session *tls_new;
756 
757 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
758 
759 	counter_u64_add(ktls_offload_active, 1);
760 
761 	refcount_init(&tls_new->refcount, 1);
762 	if (direction == KTLS_RX) {
763 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
764 		    tls_new);
765 	} else {
766 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
767 		    tls_new);
768 		tls_new->inp = tls->inp;
769 		tls_new->tx = true;
770 		in_pcbref(tls_new->inp);
771 	}
772 
773 	/* Copy fields from existing session. */
774 	tls_new->params = tls->params;
775 	tls_new->wq_index = tls->wq_index;
776 
777 	/* Deep copy keys. */
778 	if (tls_new->params.auth_key != NULL) {
779 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
780 		    M_KTLS, M_WAITOK);
781 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
782 		    tls->params.auth_key_len);
783 	}
784 
785 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
786 	    M_WAITOK);
787 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
788 	    tls->params.cipher_key_len);
789 
790 	return (tls_new);
791 }
792 
793 #ifdef TCP_OFFLOAD
794 static int
795 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
796 {
797 	struct inpcb *inp;
798 	struct tcpcb *tp;
799 	int error;
800 
801 	inp = so->so_pcb;
802 	INP_WLOCK(inp);
803 	if (inp->inp_flags & INP_DROPPED) {
804 		INP_WUNLOCK(inp);
805 		return (ECONNRESET);
806 	}
807 	if (inp->inp_socket == NULL) {
808 		INP_WUNLOCK(inp);
809 		return (ECONNRESET);
810 	}
811 	tp = intotcpcb(inp);
812 	if (!(tp->t_flags & TF_TOE)) {
813 		INP_WUNLOCK(inp);
814 		return (EOPNOTSUPP);
815 	}
816 
817 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
818 	INP_WUNLOCK(inp);
819 	if (error == 0) {
820 		tls->mode = TCP_TLS_MODE_TOE;
821 		switch (tls->params.cipher_algorithm) {
822 		case CRYPTO_AES_CBC:
823 			counter_u64_add(ktls_toe_cbc, 1);
824 			break;
825 		case CRYPTO_AES_NIST_GCM_16:
826 			counter_u64_add(ktls_toe_gcm, 1);
827 			break;
828 		case CRYPTO_CHACHA20_POLY1305:
829 			counter_u64_add(ktls_toe_chacha20, 1);
830 			break;
831 		}
832 	}
833 	return (error);
834 }
835 #endif
836 
837 /*
838  * Common code used when first enabling ifnet TLS on a connection or
839  * when allocating a new ifnet TLS session due to a routing change.
840  * This function allocates a new TLS send tag on whatever interface
841  * the connection is currently routed over.
842  */
843 static int
844 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
845     struct m_snd_tag **mstp)
846 {
847 	union if_snd_tag_alloc_params params;
848 	struct ifnet *ifp;
849 	struct nhop_object *nh;
850 	struct tcpcb *tp;
851 	int error;
852 
853 	INP_RLOCK(inp);
854 	if (inp->inp_flags & INP_DROPPED) {
855 		INP_RUNLOCK(inp);
856 		return (ECONNRESET);
857 	}
858 	if (inp->inp_socket == NULL) {
859 		INP_RUNLOCK(inp);
860 		return (ECONNRESET);
861 	}
862 	tp = intotcpcb(inp);
863 
864 	/*
865 	 * Check administrative controls on ifnet TLS to determine if
866 	 * ifnet TLS should be denied.
867 	 *
868 	 * - Always permit 'force' requests.
869 	 * - ktls_ifnet_permitted == 0: always deny.
870 	 */
871 	if (!force && ktls_ifnet_permitted == 0) {
872 		INP_RUNLOCK(inp);
873 		return (ENXIO);
874 	}
875 
876 	/*
877 	 * XXX: Use the cached route in the inpcb to find the
878 	 * interface.  This should perhaps instead use
879 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
880 	 * enabled after a connection has completed key negotiation in
881 	 * userland, the cached route will be present in practice.
882 	 */
883 	nh = inp->inp_route.ro_nh;
884 	if (nh == NULL) {
885 		INP_RUNLOCK(inp);
886 		return (ENXIO);
887 	}
888 	ifp = nh->nh_ifp;
889 	if_ref(ifp);
890 
891 	/*
892 	 * Allocate a TLS + ratelimit tag if the connection has an
893 	 * existing pacing rate.
894 	 */
895 	if (tp->t_pacing_rate != -1 &&
896 	    (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
897 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
898 		params.tls_rate_limit.inp = inp;
899 		params.tls_rate_limit.tls = tls;
900 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
901 	} else {
902 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
903 		params.tls.inp = inp;
904 		params.tls.tls = tls;
905 	}
906 	params.hdr.flowid = inp->inp_flowid;
907 	params.hdr.flowtype = inp->inp_flowtype;
908 	params.hdr.numa_domain = inp->inp_numa_domain;
909 	INP_RUNLOCK(inp);
910 
911 	if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
912 		error = EOPNOTSUPP;
913 		goto out;
914 	}
915 	if (inp->inp_vflag & INP_IPV6) {
916 		if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
917 			error = EOPNOTSUPP;
918 			goto out;
919 		}
920 	} else {
921 		if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
922 			error = EOPNOTSUPP;
923 			goto out;
924 		}
925 	}
926 	error = m_snd_tag_alloc(ifp, &params, mstp);
927 out:
928 	if_rele(ifp);
929 	return (error);
930 }
931 
932 /*
933  * Allocate an initial TLS receive tag for doing HW decryption of TLS
934  * data.
935  *
936  * This function allocates a new TLS receive tag on whatever interface
937  * the connection is currently routed over.  If the connection ends up
938  * using a different interface for receive this will get fixed up via
939  * ktls_input_ifp_mismatch as future packets arrive.
940  */
941 static int
942 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
943     struct m_snd_tag **mstp)
944 {
945 	union if_snd_tag_alloc_params params;
946 	struct ifnet *ifp;
947 	struct nhop_object *nh;
948 	int error;
949 
950 	if (!ktls_ocf_recrypt_supported(tls))
951 		return (ENXIO);
952 
953 	INP_RLOCK(inp);
954 	if (inp->inp_flags & INP_DROPPED) {
955 		INP_RUNLOCK(inp);
956 		return (ECONNRESET);
957 	}
958 	if (inp->inp_socket == NULL) {
959 		INP_RUNLOCK(inp);
960 		return (ECONNRESET);
961 	}
962 
963 	/*
964 	 * Check administrative controls on ifnet TLS to determine if
965 	 * ifnet TLS should be denied.
966 	 */
967 	if (ktls_ifnet_permitted == 0) {
968 		INP_RUNLOCK(inp);
969 		return (ENXIO);
970 	}
971 
972 	/*
973 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
974 	 * the inpcb to find the interface.
975 	 */
976 	nh = inp->inp_route.ro_nh;
977 	if (nh == NULL) {
978 		INP_RUNLOCK(inp);
979 		return (ENXIO);
980 	}
981 	ifp = nh->nh_ifp;
982 	if_ref(ifp);
983 	tls->rx_ifp = ifp;
984 
985 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
986 	params.hdr.flowid = inp->inp_flowid;
987 	params.hdr.flowtype = inp->inp_flowtype;
988 	params.hdr.numa_domain = inp->inp_numa_domain;
989 	params.tls_rx.inp = inp;
990 	params.tls_rx.tls = tls;
991 	params.tls_rx.vlan_id = 0;
992 
993 	INP_RUNLOCK(inp);
994 
995 	if (inp->inp_vflag & INP_IPV6) {
996 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0) {
997 			error = EOPNOTSUPP;
998 			goto out;
999 		}
1000 	} else {
1001 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0) {
1002 			error = EOPNOTSUPP;
1003 			goto out;
1004 		}
1005 	}
1006 	error = m_snd_tag_alloc(ifp, &params, mstp);
1007 
1008 	/*
1009 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1010 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1011 	 * ID for use in ktls_reset_receive_tag which allocates new
1012 	 * receive tags directly from the leaf interface bypassing
1013 	 * if_vlan.
1014 	 */
1015 	if (error == 0)
1016 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1017 out:
1018 	return (error);
1019 }
1020 
1021 static int
1022 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1023     bool force)
1024 {
1025 	struct m_snd_tag *mst;
1026 	int error;
1027 
1028 	switch (direction) {
1029 	case KTLS_TX:
1030 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1031 		if (__predict_false(error != 0))
1032 			goto done;
1033 		break;
1034 	case KTLS_RX:
1035 		KASSERT(!force, ("%s: forced receive tag", __func__));
1036 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1037 		if (__predict_false(error != 0))
1038 			goto done;
1039 		break;
1040 	default:
1041 		__assert_unreachable();
1042 	}
1043 
1044 	tls->mode = TCP_TLS_MODE_IFNET;
1045 	tls->snd_tag = mst;
1046 
1047 	switch (tls->params.cipher_algorithm) {
1048 	case CRYPTO_AES_CBC:
1049 		counter_u64_add(ktls_ifnet_cbc, 1);
1050 		break;
1051 	case CRYPTO_AES_NIST_GCM_16:
1052 		counter_u64_add(ktls_ifnet_gcm, 1);
1053 		break;
1054 	case CRYPTO_CHACHA20_POLY1305:
1055 		counter_u64_add(ktls_ifnet_chacha20, 1);
1056 		break;
1057 	default:
1058 		break;
1059 	}
1060 done:
1061 	return (error);
1062 }
1063 
1064 static void
1065 ktls_use_sw(struct ktls_session *tls)
1066 {
1067 	tls->mode = TCP_TLS_MODE_SW;
1068 	switch (tls->params.cipher_algorithm) {
1069 	case CRYPTO_AES_CBC:
1070 		counter_u64_add(ktls_sw_cbc, 1);
1071 		break;
1072 	case CRYPTO_AES_NIST_GCM_16:
1073 		counter_u64_add(ktls_sw_gcm, 1);
1074 		break;
1075 	case CRYPTO_CHACHA20_POLY1305:
1076 		counter_u64_add(ktls_sw_chacha20, 1);
1077 		break;
1078 	}
1079 }
1080 
1081 static int
1082 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1083 {
1084 	int error;
1085 
1086 	error = ktls_ocf_try(so, tls, direction);
1087 	if (error)
1088 		return (error);
1089 	ktls_use_sw(tls);
1090 	return (0);
1091 }
1092 
1093 /*
1094  * KTLS RX stores data in the socket buffer as a list of TLS records,
1095  * where each record is stored as a control message containg the TLS
1096  * header followed by data mbufs containing the decrypted data.  This
1097  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1098  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1099  * should be queued to the socket buffer as records, but encrypted
1100  * data which needs to be decrypted by software arrives as a stream of
1101  * regular mbufs which need to be converted.  In addition, there may
1102  * already be pending encrypted data in the socket buffer when KTLS RX
1103  * is enabled.
1104  *
1105  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1106  * is used:
1107  *
1108  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1109  *
1110  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1111  *   from the first mbuf.  Once all of the data for that TLS record is
1112  *   queued, the socket is queued to a worker thread.
1113  *
1114  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1115  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1116  *   decrypted, and inserted into the regular socket buffer chain as
1117  *   record starting with a control message holding the TLS header and
1118  *   a chain of mbufs holding the encrypted data.
1119  */
1120 
1121 static void
1122 sb_mark_notready(struct sockbuf *sb)
1123 {
1124 	struct mbuf *m;
1125 
1126 	m = sb->sb_mb;
1127 	sb->sb_mtls = m;
1128 	sb->sb_mb = NULL;
1129 	sb->sb_mbtail = NULL;
1130 	sb->sb_lastrecord = NULL;
1131 	for (; m != NULL; m = m->m_next) {
1132 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1133 		    __func__));
1134 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1135 		    __func__));
1136 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1137 		    __func__));
1138 		m->m_flags |= M_NOTREADY;
1139 		sb->sb_acc -= m->m_len;
1140 		sb->sb_tlscc += m->m_len;
1141 		sb->sb_mtlstail = m;
1142 	}
1143 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1144 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1145 	    sb->sb_ccc));
1146 }
1147 
1148 /*
1149  * Return information about the pending TLS data in a socket
1150  * buffer.  On return, 'seqno' is set to the sequence number
1151  * of the next TLS record to be received, 'resid' is set to
1152  * the amount of bytes still needed for the last pending
1153  * record.  The function returns 'false' if the last pending
1154  * record contains a partial TLS header.  In that case, 'resid'
1155  * is the number of bytes needed to complete the TLS header.
1156  */
1157 bool
1158 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1159 {
1160 	struct tls_record_layer hdr;
1161 	struct mbuf *m;
1162 	uint64_t seqno;
1163 	size_t resid;
1164 	u_int offset, record_len;
1165 
1166 	SOCKBUF_LOCK_ASSERT(sb);
1167 	MPASS(sb->sb_flags & SB_TLS_RX);
1168 	seqno = sb->sb_tls_seqno;
1169 	resid = sb->sb_tlscc;
1170 	m = sb->sb_mtls;
1171 	offset = 0;
1172 
1173 	if (resid == 0) {
1174 		*seqnop = seqno;
1175 		*residp = 0;
1176 		return (true);
1177 	}
1178 
1179 	for (;;) {
1180 		seqno++;
1181 
1182 		if (resid < sizeof(hdr)) {
1183 			*seqnop = seqno;
1184 			*residp = sizeof(hdr) - resid;
1185 			return (false);
1186 		}
1187 
1188 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1189 
1190 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1191 		if (resid <= record_len) {
1192 			*seqnop = seqno;
1193 			*residp = record_len - resid;
1194 			return (true);
1195 		}
1196 		resid -= record_len;
1197 
1198 		while (record_len != 0) {
1199 			if (m->m_len - offset > record_len) {
1200 				offset += record_len;
1201 				break;
1202 			}
1203 
1204 			record_len -= (m->m_len - offset);
1205 			offset = 0;
1206 			m = m->m_next;
1207 		}
1208 	}
1209 }
1210 
1211 int
1212 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1213 {
1214 	struct ktls_session *tls;
1215 	int error;
1216 
1217 	if (!ktls_offload_enable)
1218 		return (ENOTSUP);
1219 
1220 	counter_u64_add(ktls_offload_enable_calls, 1);
1221 
1222 	/*
1223 	 * This should always be true since only the TCP socket option
1224 	 * invokes this function.
1225 	 */
1226 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1227 		return (EINVAL);
1228 
1229 	/*
1230 	 * XXX: Don't overwrite existing sessions.  We should permit
1231 	 * this to support rekeying in the future.
1232 	 */
1233 	if (so->so_rcv.sb_tls_info != NULL)
1234 		return (EALREADY);
1235 
1236 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1237 		return (ENOTSUP);
1238 
1239 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1240 	if (error)
1241 		return (error);
1242 
1243 	error = ktls_ocf_try(so, tls, KTLS_RX);
1244 	if (error) {
1245 		ktls_free(tls);
1246 		return (error);
1247 	}
1248 
1249 	/* Mark the socket as using TLS offload. */
1250 	SOCK_RECVBUF_LOCK(so);
1251 	if (SOLISTENING(so)) {
1252 		SOCK_RECVBUF_UNLOCK(so);
1253 		ktls_free(tls);
1254 		return (EINVAL);
1255 	}
1256 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1257 	so->so_rcv.sb_tls_info = tls;
1258 	so->so_rcv.sb_flags |= SB_TLS_RX;
1259 
1260 	/* Mark existing data as not ready until it can be decrypted. */
1261 	sb_mark_notready(&so->so_rcv);
1262 	ktls_check_rx(&so->so_rcv);
1263 	SOCK_RECVBUF_UNLOCK(so);
1264 
1265 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1266 #ifdef TCP_OFFLOAD
1267 	error = ktls_try_toe(so, tls, KTLS_RX);
1268 	if (error)
1269 #endif
1270 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1271 	if (error)
1272 		ktls_use_sw(tls);
1273 
1274 	counter_u64_add(ktls_offload_total, 1);
1275 
1276 	return (0);
1277 }
1278 
1279 int
1280 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1281 {
1282 	struct ktls_session *tls;
1283 	struct inpcb *inp;
1284 	struct tcpcb *tp;
1285 	int error;
1286 
1287 	if (!ktls_offload_enable)
1288 		return (ENOTSUP);
1289 
1290 	counter_u64_add(ktls_offload_enable_calls, 1);
1291 
1292 	/*
1293 	 * This should always be true since only the TCP socket option
1294 	 * invokes this function.
1295 	 */
1296 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1297 		return (EINVAL);
1298 
1299 	/*
1300 	 * XXX: Don't overwrite existing sessions.  We should permit
1301 	 * this to support rekeying in the future.
1302 	 */
1303 	if (so->so_snd.sb_tls_info != NULL)
1304 		return (EALREADY);
1305 
1306 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1307 		return (ENOTSUP);
1308 
1309 	/* TLS requires ext pgs */
1310 	if (mb_use_ext_pgs == 0)
1311 		return (ENXIO);
1312 
1313 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1314 	if (error)
1315 		return (error);
1316 
1317 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1318 #ifdef TCP_OFFLOAD
1319 	error = ktls_try_toe(so, tls, KTLS_TX);
1320 	if (error)
1321 #endif
1322 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1323 	if (error)
1324 		error = ktls_try_sw(so, tls, KTLS_TX);
1325 
1326 	if (error) {
1327 		ktls_free(tls);
1328 		return (error);
1329 	}
1330 
1331 	/*
1332 	 * Serialize with sosend_generic() and make sure that we're not
1333 	 * operating on a listening socket.
1334 	 */
1335 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1336 	if (error) {
1337 		ktls_free(tls);
1338 		return (error);
1339 	}
1340 
1341 	/*
1342 	 * Write lock the INP when setting sb_tls_info so that
1343 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1344 	 * holding the INP lock.
1345 	 */
1346 	inp = so->so_pcb;
1347 	INP_WLOCK(inp);
1348 	SOCK_SENDBUF_LOCK(so);
1349 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1350 	so->so_snd.sb_tls_info = tls;
1351 	if (tls->mode != TCP_TLS_MODE_SW) {
1352 		tp = intotcpcb(inp);
1353 		MPASS(tp->t_nic_ktls_xmit == 0);
1354 		tp->t_nic_ktls_xmit = 1;
1355 		if (tp->t_fb->tfb_hwtls_change != NULL)
1356 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1357 	}
1358 	SOCK_SENDBUF_UNLOCK(so);
1359 	INP_WUNLOCK(inp);
1360 	SOCK_IO_SEND_UNLOCK(so);
1361 
1362 	counter_u64_add(ktls_offload_total, 1);
1363 
1364 	return (0);
1365 }
1366 
1367 int
1368 ktls_get_rx_mode(struct socket *so, int *modep)
1369 {
1370 	struct ktls_session *tls;
1371 	struct inpcb *inp __diagused;
1372 
1373 	if (SOLISTENING(so))
1374 		return (EINVAL);
1375 	inp = so->so_pcb;
1376 	INP_WLOCK_ASSERT(inp);
1377 	SOCK_RECVBUF_LOCK(so);
1378 	tls = so->so_rcv.sb_tls_info;
1379 	if (tls == NULL)
1380 		*modep = TCP_TLS_MODE_NONE;
1381 	else
1382 		*modep = tls->mode;
1383 	SOCK_RECVBUF_UNLOCK(so);
1384 	return (0);
1385 }
1386 
1387 /*
1388  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1389  *
1390  * This function gets information about the next TCP- and TLS-
1391  * sequence number to be processed by the TLS receive worker
1392  * thread. The information is extracted from the given "inpcb"
1393  * structure. The values are stored in host endian format at the two
1394  * given output pointer locations. The TCP sequence number points to
1395  * the beginning of the TLS header.
1396  *
1397  * This function returns zero on success, else a non-zero error code
1398  * is returned.
1399  */
1400 int
1401 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1402 {
1403 	struct socket *so;
1404 	struct tcpcb *tp;
1405 
1406 	INP_RLOCK(inp);
1407 	so = inp->inp_socket;
1408 	if (__predict_false(so == NULL)) {
1409 		INP_RUNLOCK(inp);
1410 		return (EINVAL);
1411 	}
1412 	if (inp->inp_flags & INP_DROPPED) {
1413 		INP_RUNLOCK(inp);
1414 		return (ECONNRESET);
1415 	}
1416 
1417 	tp = intotcpcb(inp);
1418 	MPASS(tp != NULL);
1419 
1420 	SOCKBUF_LOCK(&so->so_rcv);
1421 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1422 	*tlsseq = so->so_rcv.sb_tls_seqno;
1423 	SOCKBUF_UNLOCK(&so->so_rcv);
1424 
1425 	INP_RUNLOCK(inp);
1426 
1427 	return (0);
1428 }
1429 
1430 int
1431 ktls_get_tx_mode(struct socket *so, int *modep)
1432 {
1433 	struct ktls_session *tls;
1434 	struct inpcb *inp __diagused;
1435 
1436 	if (SOLISTENING(so))
1437 		return (EINVAL);
1438 	inp = so->so_pcb;
1439 	INP_WLOCK_ASSERT(inp);
1440 	SOCK_SENDBUF_LOCK(so);
1441 	tls = so->so_snd.sb_tls_info;
1442 	if (tls == NULL)
1443 		*modep = TCP_TLS_MODE_NONE;
1444 	else
1445 		*modep = tls->mode;
1446 	SOCK_SENDBUF_UNLOCK(so);
1447 	return (0);
1448 }
1449 
1450 /*
1451  * Switch between SW and ifnet TLS sessions as requested.
1452  */
1453 int
1454 ktls_set_tx_mode(struct socket *so, int mode)
1455 {
1456 	struct ktls_session *tls, *tls_new;
1457 	struct inpcb *inp;
1458 	struct tcpcb *tp;
1459 	int error;
1460 
1461 	if (SOLISTENING(so))
1462 		return (EINVAL);
1463 	switch (mode) {
1464 	case TCP_TLS_MODE_SW:
1465 	case TCP_TLS_MODE_IFNET:
1466 		break;
1467 	default:
1468 		return (EINVAL);
1469 	}
1470 
1471 	inp = so->so_pcb;
1472 	INP_WLOCK_ASSERT(inp);
1473 	tp = intotcpcb(inp);
1474 
1475 	if (mode == TCP_TLS_MODE_IFNET) {
1476 		/* Don't allow enabling ifnet ktls multiple times */
1477 		if (tp->t_nic_ktls_xmit)
1478 			return (EALREADY);
1479 
1480 		/*
1481 		 * Don't enable ifnet ktls if we disabled it due to an
1482 		 * excessive retransmission rate
1483 		 */
1484 		if (tp->t_nic_ktls_xmit_dis)
1485 			return (ENXIO);
1486 	}
1487 
1488 	SOCKBUF_LOCK(&so->so_snd);
1489 	tls = so->so_snd.sb_tls_info;
1490 	if (tls == NULL) {
1491 		SOCKBUF_UNLOCK(&so->so_snd);
1492 		return (0);
1493 	}
1494 
1495 	if (tls->mode == mode) {
1496 		SOCKBUF_UNLOCK(&so->so_snd);
1497 		return (0);
1498 	}
1499 
1500 	tls = ktls_hold(tls);
1501 	SOCKBUF_UNLOCK(&so->so_snd);
1502 	INP_WUNLOCK(inp);
1503 
1504 	tls_new = ktls_clone_session(tls, KTLS_TX);
1505 
1506 	if (mode == TCP_TLS_MODE_IFNET)
1507 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1508 	else
1509 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1510 	if (error) {
1511 		counter_u64_add(ktls_switch_failed, 1);
1512 		ktls_free(tls_new);
1513 		ktls_free(tls);
1514 		INP_WLOCK(inp);
1515 		return (error);
1516 	}
1517 
1518 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1519 	if (error) {
1520 		counter_u64_add(ktls_switch_failed, 1);
1521 		ktls_free(tls_new);
1522 		ktls_free(tls);
1523 		INP_WLOCK(inp);
1524 		return (error);
1525 	}
1526 
1527 	/*
1528 	 * If we raced with another session change, keep the existing
1529 	 * session.
1530 	 */
1531 	if (tls != so->so_snd.sb_tls_info) {
1532 		counter_u64_add(ktls_switch_failed, 1);
1533 		SOCK_IO_SEND_UNLOCK(so);
1534 		ktls_free(tls_new);
1535 		ktls_free(tls);
1536 		INP_WLOCK(inp);
1537 		return (EBUSY);
1538 	}
1539 
1540 	INP_WLOCK(inp);
1541 	SOCKBUF_LOCK(&so->so_snd);
1542 	so->so_snd.sb_tls_info = tls_new;
1543 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1544 		MPASS(tp->t_nic_ktls_xmit == 0);
1545 		tp->t_nic_ktls_xmit = 1;
1546 		if (tp->t_fb->tfb_hwtls_change != NULL)
1547 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1548 	}
1549 	SOCKBUF_UNLOCK(&so->so_snd);
1550 	SOCK_IO_SEND_UNLOCK(so);
1551 
1552 	/*
1553 	 * Drop two references on 'tls'.  The first is for the
1554 	 * ktls_hold() above.  The second drops the reference from the
1555 	 * socket buffer.
1556 	 */
1557 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1558 	ktls_free(tls);
1559 	ktls_free(tls);
1560 
1561 	if (mode == TCP_TLS_MODE_IFNET)
1562 		counter_u64_add(ktls_switch_to_ifnet, 1);
1563 	else
1564 		counter_u64_add(ktls_switch_to_sw, 1);
1565 
1566 	return (0);
1567 }
1568 
1569 /*
1570  * Try to allocate a new TLS receive tag.  This task is scheduled when
1571  * sbappend_ktls_rx detects an input path change.  If a new tag is
1572  * allocated, replace the tag in the TLS session.  If a new tag cannot
1573  * be allocated, let the session fall back to software decryption.
1574  */
1575 static void
1576 ktls_reset_receive_tag(void *context, int pending)
1577 {
1578 	union if_snd_tag_alloc_params params;
1579 	struct ktls_session *tls;
1580 	struct m_snd_tag *mst;
1581 	struct inpcb *inp;
1582 	struct ifnet *ifp;
1583 	struct socket *so;
1584 	int error;
1585 
1586 	MPASS(pending == 1);
1587 
1588 	tls = context;
1589 	so = tls->so;
1590 	inp = so->so_pcb;
1591 	ifp = NULL;
1592 
1593 	INP_RLOCK(inp);
1594 	if (inp->inp_flags & INP_DROPPED) {
1595 		INP_RUNLOCK(inp);
1596 		goto out;
1597 	}
1598 
1599 	SOCKBUF_LOCK(&so->so_rcv);
1600 	mst = tls->snd_tag;
1601 	tls->snd_tag = NULL;
1602 	if (mst != NULL)
1603 		m_snd_tag_rele(mst);
1604 
1605 	ifp = tls->rx_ifp;
1606 	if_ref(ifp);
1607 	SOCKBUF_UNLOCK(&so->so_rcv);
1608 
1609 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1610 	params.hdr.flowid = inp->inp_flowid;
1611 	params.hdr.flowtype = inp->inp_flowtype;
1612 	params.hdr.numa_domain = inp->inp_numa_domain;
1613 	params.tls_rx.inp = inp;
1614 	params.tls_rx.tls = tls;
1615 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1616 	INP_RUNLOCK(inp);
1617 
1618 	if (inp->inp_vflag & INP_IPV6) {
1619 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1620 			goto out;
1621 	} else {
1622 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1623 			goto out;
1624 	}
1625 
1626 	error = m_snd_tag_alloc(ifp, &params, &mst);
1627 	if (error == 0) {
1628 		SOCKBUF_LOCK(&so->so_rcv);
1629 		tls->snd_tag = mst;
1630 		SOCKBUF_UNLOCK(&so->so_rcv);
1631 
1632 		counter_u64_add(ktls_ifnet_reset, 1);
1633 	} else {
1634 		/*
1635 		 * Just fall back to software decryption if a tag
1636 		 * cannot be allocated leaving the connection intact.
1637 		 * If a future input path change switches to another
1638 		 * interface this connection will resume ifnet TLS.
1639 		 */
1640 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1641 	}
1642 
1643 out:
1644 	mtx_pool_lock(mtxpool_sleep, tls);
1645 	tls->reset_pending = false;
1646 	mtx_pool_unlock(mtxpool_sleep, tls);
1647 
1648 	if (ifp != NULL)
1649 		if_rele(ifp);
1650 	sorele(so);
1651 	ktls_free(tls);
1652 }
1653 
1654 /*
1655  * Try to allocate a new TLS send tag.  This task is scheduled when
1656  * ip_output detects a route change while trying to transmit a packet
1657  * holding a TLS record.  If a new tag is allocated, replace the tag
1658  * in the TLS session.  Subsequent packets on the connection will use
1659  * the new tag.  If a new tag cannot be allocated, drop the
1660  * connection.
1661  */
1662 static void
1663 ktls_reset_send_tag(void *context, int pending)
1664 {
1665 	struct epoch_tracker et;
1666 	struct ktls_session *tls;
1667 	struct m_snd_tag *old, *new;
1668 	struct inpcb *inp;
1669 	struct tcpcb *tp;
1670 	int error;
1671 
1672 	MPASS(pending == 1);
1673 
1674 	tls = context;
1675 	inp = tls->inp;
1676 
1677 	/*
1678 	 * Free the old tag first before allocating a new one.
1679 	 * ip[6]_output_send() will treat a NULL send tag the same as
1680 	 * an ifp mismatch and drop packets until a new tag is
1681 	 * allocated.
1682 	 *
1683 	 * Write-lock the INP when changing tls->snd_tag since
1684 	 * ip[6]_output_send() holds a read-lock when reading the
1685 	 * pointer.
1686 	 */
1687 	INP_WLOCK(inp);
1688 	old = tls->snd_tag;
1689 	tls->snd_tag = NULL;
1690 	INP_WUNLOCK(inp);
1691 	if (old != NULL)
1692 		m_snd_tag_rele(old);
1693 
1694 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1695 
1696 	if (error == 0) {
1697 		INP_WLOCK(inp);
1698 		tls->snd_tag = new;
1699 		mtx_pool_lock(mtxpool_sleep, tls);
1700 		tls->reset_pending = false;
1701 		mtx_pool_unlock(mtxpool_sleep, tls);
1702 		INP_WUNLOCK(inp);
1703 
1704 		counter_u64_add(ktls_ifnet_reset, 1);
1705 
1706 		/*
1707 		 * XXX: Should we kick tcp_output explicitly now that
1708 		 * the send tag is fixed or just rely on timers?
1709 		 */
1710 	} else {
1711 		NET_EPOCH_ENTER(et);
1712 		INP_WLOCK(inp);
1713 		if (!(inp->inp_flags & INP_DROPPED)) {
1714 			tp = intotcpcb(inp);
1715 			CURVNET_SET(inp->inp_vnet);
1716 			tp = tcp_drop(tp, ECONNABORTED);
1717 			CURVNET_RESTORE();
1718 			if (tp != NULL)
1719 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1720 		}
1721 		INP_WUNLOCK(inp);
1722 		NET_EPOCH_EXIT(et);
1723 
1724 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1725 
1726 		/*
1727 		 * Leave reset_pending true to avoid future tasks while
1728 		 * the socket goes away.
1729 		 */
1730 	}
1731 
1732 	ktls_free(tls);
1733 }
1734 
1735 void
1736 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1737 {
1738 	struct ktls_session *tls;
1739 	struct socket *so;
1740 
1741 	SOCKBUF_LOCK_ASSERT(sb);
1742 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1743 	    __func__, sb));
1744 	so = __containerof(sb, struct socket, so_rcv);
1745 
1746 	tls = sb->sb_tls_info;
1747 	if_rele(tls->rx_ifp);
1748 	if_ref(ifp);
1749 	tls->rx_ifp = ifp;
1750 
1751 	/*
1752 	 * See if we should schedule a task to update the receive tag for
1753 	 * this session.
1754 	 */
1755 	mtx_pool_lock(mtxpool_sleep, tls);
1756 	if (!tls->reset_pending) {
1757 		(void) ktls_hold(tls);
1758 		soref(so);
1759 		tls->so = so;
1760 		tls->reset_pending = true;
1761 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1762 	}
1763 	mtx_pool_unlock(mtxpool_sleep, tls);
1764 }
1765 
1766 int
1767 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1768 {
1769 
1770 	if (inp == NULL)
1771 		return (ENOBUFS);
1772 
1773 	INP_LOCK_ASSERT(inp);
1774 
1775 	/*
1776 	 * See if we should schedule a task to update the send tag for
1777 	 * this session.
1778 	 */
1779 	mtx_pool_lock(mtxpool_sleep, tls);
1780 	if (!tls->reset_pending) {
1781 		(void) ktls_hold(tls);
1782 		tls->reset_pending = true;
1783 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1784 	}
1785 	mtx_pool_unlock(mtxpool_sleep, tls);
1786 	return (ENOBUFS);
1787 }
1788 
1789 #ifdef RATELIMIT
1790 int
1791 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1792 {
1793 	union if_snd_tag_modify_params params = {
1794 		.rate_limit.max_rate = max_pacing_rate,
1795 		.rate_limit.flags = M_NOWAIT,
1796 	};
1797 	struct m_snd_tag *mst;
1798 
1799 	/* Can't get to the inp, but it should be locked. */
1800 	/* INP_LOCK_ASSERT(inp); */
1801 
1802 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1803 
1804 	if (tls->snd_tag == NULL) {
1805 		/*
1806 		 * Resetting send tag, ignore this change.  The
1807 		 * pending reset may or may not see this updated rate
1808 		 * in the tcpcb.  If it doesn't, we will just lose
1809 		 * this rate change.
1810 		 */
1811 		return (0);
1812 	}
1813 
1814 	mst = tls->snd_tag;
1815 
1816 	MPASS(mst != NULL);
1817 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1818 
1819 	return (mst->sw->snd_tag_modify(mst, &params));
1820 }
1821 #endif
1822 
1823 static void
1824 ktls_destroy_help(void *context, int pending __unused)
1825 {
1826 	ktls_destroy(context);
1827 }
1828 
1829 void
1830 ktls_destroy(struct ktls_session *tls)
1831 {
1832 	struct inpcb *inp;
1833 	struct tcpcb *tp;
1834 	bool wlocked;
1835 
1836 	MPASS(tls->refcount == 0);
1837 
1838 	inp = tls->inp;
1839 	if (tls->tx) {
1840 		wlocked = INP_WLOCKED(inp);
1841 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1842 			/*
1843 			 * rwlocks read locks are anonymous, and there
1844 			 * is no way to know if our current thread
1845 			 * holds an rlock on the inp.  As a rough
1846 			 * estimate, check to see if the thread holds
1847 			 * *any* rlocks at all.  If it does not, then we
1848 			 * know that we don't hold the inp rlock, and
1849 			 * can safely take the wlock
1850 			 */
1851 			if (curthread->td_rw_rlocks == 0) {
1852 				INP_WLOCK(inp);
1853 			} else {
1854 				/*
1855 				 * We might hold the rlock, so let's
1856 				 * do the destroy in a taskqueue
1857 				 * context to avoid a potential
1858 				 * deadlock.  This should be very
1859 				 * rare.
1860 				 */
1861 				counter_u64_add(ktls_destroy_task, 1);
1862 				TASK_INIT(&tls->destroy_task, 0,
1863 				    ktls_destroy_help, tls);
1864 				(void)taskqueue_enqueue(taskqueue_thread,
1865 				    &tls->destroy_task);
1866 				return;
1867 			}
1868 		}
1869 	}
1870 
1871 	if (tls->sequential_records) {
1872 		struct mbuf *m, *n;
1873 		int page_count;
1874 
1875 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1876 			page_count = m->m_epg_enc_cnt;
1877 			while (page_count > 0) {
1878 				KASSERT(page_count >= m->m_epg_nrdy,
1879 				    ("%s: too few pages", __func__));
1880 				page_count -= m->m_epg_nrdy;
1881 				m = m_free(m);
1882 			}
1883 		}
1884 	}
1885 
1886 	counter_u64_add(ktls_offload_active, -1);
1887 	switch (tls->mode) {
1888 	case TCP_TLS_MODE_SW:
1889 		switch (tls->params.cipher_algorithm) {
1890 		case CRYPTO_AES_CBC:
1891 			counter_u64_add(ktls_sw_cbc, -1);
1892 			break;
1893 		case CRYPTO_AES_NIST_GCM_16:
1894 			counter_u64_add(ktls_sw_gcm, -1);
1895 			break;
1896 		case CRYPTO_CHACHA20_POLY1305:
1897 			counter_u64_add(ktls_sw_chacha20, -1);
1898 			break;
1899 		}
1900 		break;
1901 	case TCP_TLS_MODE_IFNET:
1902 		switch (tls->params.cipher_algorithm) {
1903 		case CRYPTO_AES_CBC:
1904 			counter_u64_add(ktls_ifnet_cbc, -1);
1905 			break;
1906 		case CRYPTO_AES_NIST_GCM_16:
1907 			counter_u64_add(ktls_ifnet_gcm, -1);
1908 			break;
1909 		case CRYPTO_CHACHA20_POLY1305:
1910 			counter_u64_add(ktls_ifnet_chacha20, -1);
1911 			break;
1912 		}
1913 		if (tls->snd_tag != NULL)
1914 			m_snd_tag_rele(tls->snd_tag);
1915 		if (tls->rx_ifp != NULL)
1916 			if_rele(tls->rx_ifp);
1917 		if (tls->tx) {
1918 			INP_WLOCK_ASSERT(inp);
1919 			tp = intotcpcb(inp);
1920 			MPASS(tp->t_nic_ktls_xmit == 1);
1921 			tp->t_nic_ktls_xmit = 0;
1922 		}
1923 		break;
1924 #ifdef TCP_OFFLOAD
1925 	case TCP_TLS_MODE_TOE:
1926 		switch (tls->params.cipher_algorithm) {
1927 		case CRYPTO_AES_CBC:
1928 			counter_u64_add(ktls_toe_cbc, -1);
1929 			break;
1930 		case CRYPTO_AES_NIST_GCM_16:
1931 			counter_u64_add(ktls_toe_gcm, -1);
1932 			break;
1933 		case CRYPTO_CHACHA20_POLY1305:
1934 			counter_u64_add(ktls_toe_chacha20, -1);
1935 			break;
1936 		}
1937 		break;
1938 #endif
1939 	}
1940 	if (tls->ocf_session != NULL)
1941 		ktls_ocf_free(tls);
1942 	if (tls->params.auth_key != NULL) {
1943 		zfree(tls->params.auth_key, M_KTLS);
1944 		tls->params.auth_key = NULL;
1945 		tls->params.auth_key_len = 0;
1946 	}
1947 	if (tls->params.cipher_key != NULL) {
1948 		zfree(tls->params.cipher_key, M_KTLS);
1949 		tls->params.cipher_key = NULL;
1950 		tls->params.cipher_key_len = 0;
1951 	}
1952 	if (tls->tx) {
1953 		INP_WLOCK_ASSERT(inp);
1954 		if (!in_pcbrele_wlocked(inp) && !wlocked)
1955 			INP_WUNLOCK(inp);
1956 	}
1957 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1958 
1959 	uma_zfree(ktls_session_zone, tls);
1960 }
1961 
1962 void
1963 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1964 {
1965 
1966 	for (; m != NULL; m = m->m_next) {
1967 		KASSERT((m->m_flags & M_EXTPG) != 0,
1968 		    ("ktls_seq: mapped mbuf %p", m));
1969 
1970 		m->m_epg_seqno = sb->sb_tls_seqno;
1971 		sb->sb_tls_seqno++;
1972 	}
1973 }
1974 
1975 /*
1976  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1977  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1978  * mbuf must be populated with the payload of each TLS record.
1979  *
1980  * The record_type argument specifies the TLS record type used when
1981  * populating the TLS header.
1982  *
1983  * The enq_count argument on return is set to the number of pages of
1984  * payload data for this entire chain that need to be encrypted via SW
1985  * encryption.  The returned value should be passed to ktls_enqueue
1986  * when scheduling encryption of this chain of mbufs.  To handle the
1987  * special case of empty fragments for TLS 1.0 sessions, an empty
1988  * fragment counts as one page.
1989  */
1990 void
1991 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1992     uint8_t record_type)
1993 {
1994 	struct tls_record_layer *tlshdr;
1995 	struct mbuf *m;
1996 	uint64_t *noncep;
1997 	uint16_t tls_len;
1998 	int maxlen __diagused;
1999 
2000 	maxlen = tls->params.max_frame_len;
2001 	*enq_cnt = 0;
2002 	for (m = top; m != NULL; m = m->m_next) {
2003 		/*
2004 		 * All mbufs in the chain should be TLS records whose
2005 		 * payload does not exceed the maximum frame length.
2006 		 *
2007 		 * Empty TLS 1.0 records are permitted when using CBC.
2008 		 */
2009 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2010 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2011 		    ("ktls_frame: m %p len %d", m, m->m_len));
2012 
2013 		/*
2014 		 * TLS frames require unmapped mbufs to store session
2015 		 * info.
2016 		 */
2017 		KASSERT((m->m_flags & M_EXTPG) != 0,
2018 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2019 
2020 		tls_len = m->m_len;
2021 
2022 		/* Save a reference to the session. */
2023 		m->m_epg_tls = ktls_hold(tls);
2024 
2025 		m->m_epg_hdrlen = tls->params.tls_hlen;
2026 		m->m_epg_trllen = tls->params.tls_tlen;
2027 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2028 			int bs, delta;
2029 
2030 			/*
2031 			 * AES-CBC pads messages to a multiple of the
2032 			 * block size.  Note that the padding is
2033 			 * applied after the digest and the encryption
2034 			 * is done on the "plaintext || mac || padding".
2035 			 * At least one byte of padding is always
2036 			 * present.
2037 			 *
2038 			 * Compute the final trailer length assuming
2039 			 * at most one block of padding.
2040 			 * tls->params.tls_tlen is the maximum
2041 			 * possible trailer length (padding + digest).
2042 			 * delta holds the number of excess padding
2043 			 * bytes if the maximum were used.  Those
2044 			 * extra bytes are removed.
2045 			 */
2046 			bs = tls->params.tls_bs;
2047 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2048 			m->m_epg_trllen -= delta;
2049 		}
2050 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2051 
2052 		/* Populate the TLS header. */
2053 		tlshdr = (void *)m->m_epg_hdr;
2054 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2055 
2056 		/*
2057 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2058 		 * of TLS_RLTYPE_APP.
2059 		 */
2060 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2061 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2062 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2063 			tlshdr->tls_type = TLS_RLTYPE_APP;
2064 			/* save the real record type for later */
2065 			m->m_epg_record_type = record_type;
2066 			m->m_epg_trail[0] = record_type;
2067 		} else {
2068 			tlshdr->tls_vminor = tls->params.tls_vminor;
2069 			tlshdr->tls_type = record_type;
2070 		}
2071 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2072 
2073 		/*
2074 		 * Store nonces / explicit IVs after the end of the
2075 		 * TLS header.
2076 		 *
2077 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2078 		 * from the end of the IV.  The nonce is then
2079 		 * incremented for use by the next record.
2080 		 *
2081 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2082 		 */
2083 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2084 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2085 			noncep = (uint64_t *)(tls->params.iv + 8);
2086 			be64enc(tlshdr + 1, *noncep);
2087 			(*noncep)++;
2088 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2089 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2090 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2091 
2092 		/*
2093 		 * When using SW encryption, mark the mbuf not ready.
2094 		 * It will be marked ready via sbready() after the
2095 		 * record has been encrypted.
2096 		 *
2097 		 * When using ifnet TLS, unencrypted TLS records are
2098 		 * sent down the stack to the NIC.
2099 		 */
2100 		if (tls->mode == TCP_TLS_MODE_SW) {
2101 			m->m_flags |= M_NOTREADY;
2102 			if (__predict_false(tls_len == 0)) {
2103 				/* TLS 1.0 empty fragment. */
2104 				m->m_epg_nrdy = 1;
2105 			} else
2106 				m->m_epg_nrdy = m->m_epg_npgs;
2107 			*enq_cnt += m->m_epg_nrdy;
2108 		}
2109 	}
2110 }
2111 
2112 bool
2113 ktls_permit_empty_frames(struct ktls_session *tls)
2114 {
2115 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2116 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2117 }
2118 
2119 void
2120 ktls_check_rx(struct sockbuf *sb)
2121 {
2122 	struct tls_record_layer hdr;
2123 	struct ktls_wq *wq;
2124 	struct socket *so;
2125 	bool running;
2126 
2127 	SOCKBUF_LOCK_ASSERT(sb);
2128 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2129 	    __func__, sb));
2130 	so = __containerof(sb, struct socket, so_rcv);
2131 
2132 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2133 		return;
2134 
2135 	/* Is there enough queued for a TLS header? */
2136 	if (sb->sb_tlscc < sizeof(hdr)) {
2137 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2138 			so->so_error = EMSGSIZE;
2139 		return;
2140 	}
2141 
2142 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2143 
2144 	/* Is the entire record queued? */
2145 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2146 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2147 			so->so_error = EMSGSIZE;
2148 		return;
2149 	}
2150 
2151 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2152 
2153 	soref(so);
2154 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2155 	mtx_lock(&wq->mtx);
2156 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2157 	running = wq->running;
2158 	mtx_unlock(&wq->mtx);
2159 	if (!running)
2160 		wakeup(wq);
2161 	counter_u64_add(ktls_cnt_rx_queued, 1);
2162 }
2163 
2164 static struct mbuf *
2165 ktls_detach_record(struct sockbuf *sb, int len)
2166 {
2167 	struct mbuf *m, *n, *top;
2168 	int remain;
2169 
2170 	SOCKBUF_LOCK_ASSERT(sb);
2171 	MPASS(len <= sb->sb_tlscc);
2172 
2173 	/*
2174 	 * If TLS chain is the exact size of the record,
2175 	 * just grab the whole record.
2176 	 */
2177 	top = sb->sb_mtls;
2178 	if (sb->sb_tlscc == len) {
2179 		sb->sb_mtls = NULL;
2180 		sb->sb_mtlstail = NULL;
2181 		goto out;
2182 	}
2183 
2184 	/*
2185 	 * While it would be nice to use m_split() here, we need
2186 	 * to know exactly what m_split() allocates to update the
2187 	 * accounting, so do it inline instead.
2188 	 */
2189 	remain = len;
2190 	for (m = top; remain > m->m_len; m = m->m_next)
2191 		remain -= m->m_len;
2192 
2193 	/* Easy case: don't have to split 'm'. */
2194 	if (remain == m->m_len) {
2195 		sb->sb_mtls = m->m_next;
2196 		if (sb->sb_mtls == NULL)
2197 			sb->sb_mtlstail = NULL;
2198 		m->m_next = NULL;
2199 		goto out;
2200 	}
2201 
2202 	/*
2203 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2204 	 * with M_NOWAIT first.
2205 	 */
2206 	n = m_get(M_NOWAIT, MT_DATA);
2207 	if (n == NULL) {
2208 		/*
2209 		 * Use M_WAITOK with socket buffer unlocked.  If
2210 		 * 'sb_mtls' changes while the lock is dropped, return
2211 		 * NULL to force the caller to retry.
2212 		 */
2213 		SOCKBUF_UNLOCK(sb);
2214 
2215 		n = m_get(M_WAITOK, MT_DATA);
2216 
2217 		SOCKBUF_LOCK(sb);
2218 		if (sb->sb_mtls != top) {
2219 			m_free(n);
2220 			return (NULL);
2221 		}
2222 	}
2223 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2224 
2225 	/* Store remainder in 'n'. */
2226 	n->m_len = m->m_len - remain;
2227 	if (m->m_flags & M_EXT) {
2228 		n->m_data = m->m_data + remain;
2229 		mb_dupcl(n, m);
2230 	} else {
2231 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2232 	}
2233 
2234 	/* Trim 'm' and update accounting. */
2235 	m->m_len -= n->m_len;
2236 	sb->sb_tlscc -= n->m_len;
2237 	sb->sb_ccc -= n->m_len;
2238 
2239 	/* Account for 'n'. */
2240 	sballoc_ktls_rx(sb, n);
2241 
2242 	/* Insert 'n' into the TLS chain. */
2243 	sb->sb_mtls = n;
2244 	n->m_next = m->m_next;
2245 	if (sb->sb_mtlstail == m)
2246 		sb->sb_mtlstail = n;
2247 
2248 	/* Detach the record from the TLS chain. */
2249 	m->m_next = NULL;
2250 
2251 out:
2252 	MPASS(m_length(top, NULL) == len);
2253 	for (m = top; m != NULL; m = m->m_next)
2254 		sbfree_ktls_rx(sb, m);
2255 	sb->sb_tlsdcc = len;
2256 	sb->sb_ccc += len;
2257 	SBCHECK(sb);
2258 	return (top);
2259 }
2260 
2261 /*
2262  * Determine the length of the trailing zero padding and find the real
2263  * record type in the byte before the padding.
2264  *
2265  * Walking the mbuf chain backwards is clumsy, so another option would
2266  * be to scan forwards remembering the last non-zero byte before the
2267  * trailer.  However, it would be expensive to scan the entire record.
2268  * Instead, find the last non-zero byte of each mbuf in the chain
2269  * keeping track of the relative offset of that nonzero byte.
2270  *
2271  * trail_len is the size of the MAC/tag on input and is set to the
2272  * size of the full trailer including padding and the record type on
2273  * return.
2274  */
2275 static int
2276 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2277     int *trailer_len, uint8_t *record_typep)
2278 {
2279 	char *cp;
2280 	u_int digest_start, last_offset, m_len, offset;
2281 	uint8_t record_type;
2282 
2283 	digest_start = tls_len - *trailer_len;
2284 	last_offset = 0;
2285 	offset = 0;
2286 	for (; m != NULL && offset < digest_start;
2287 	     offset += m->m_len, m = m->m_next) {
2288 		/* Don't look for padding in the tag. */
2289 		m_len = min(digest_start - offset, m->m_len);
2290 		cp = mtod(m, char *);
2291 
2292 		/* Find last non-zero byte in this mbuf. */
2293 		while (m_len > 0 && cp[m_len - 1] == 0)
2294 			m_len--;
2295 		if (m_len > 0) {
2296 			record_type = cp[m_len - 1];
2297 			last_offset = offset + m_len;
2298 		}
2299 	}
2300 	if (last_offset < tls->params.tls_hlen)
2301 		return (EBADMSG);
2302 
2303 	*record_typep = record_type;
2304 	*trailer_len = tls_len - last_offset + 1;
2305 	return (0);
2306 }
2307 
2308 /*
2309  * Check if a mbuf chain is fully decrypted at the given offset and
2310  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2311  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2312  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2313  * is encrypted.
2314  */
2315 ktls_mbuf_crypto_st_t
2316 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2317 {
2318 	int m_flags_ored = 0;
2319 	int m_flags_anded = -1;
2320 
2321 	for (; mb != NULL; mb = mb->m_next) {
2322 		if (offset < mb->m_len)
2323 			break;
2324 		offset -= mb->m_len;
2325 	}
2326 	offset += len;
2327 
2328 	for (; mb != NULL; mb = mb->m_next) {
2329 		m_flags_ored |= mb->m_flags;
2330 		m_flags_anded &= mb->m_flags;
2331 
2332 		if (offset <= mb->m_len)
2333 			break;
2334 		offset -= mb->m_len;
2335 	}
2336 	MPASS(mb != NULL || offset == 0);
2337 
2338 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2339 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2340 	else
2341 		return ((m_flags_ored & M_DECRYPTED) ?
2342 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2343 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2344 }
2345 
2346 /*
2347  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2348  */
2349 static int
2350 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2351 {
2352 	union if_snd_tag_modify_params params;
2353 	struct m_snd_tag *mst;
2354 	struct inpcb *inp;
2355 	struct tcpcb *tp;
2356 
2357 	mst = so->so_rcv.sb_tls_info->snd_tag;
2358 	if (__predict_false(mst == NULL))
2359 		return (EINVAL);
2360 
2361 	inp = sotoinpcb(so);
2362 	if (__predict_false(inp == NULL))
2363 		return (EINVAL);
2364 
2365 	INP_RLOCK(inp);
2366 	if (inp->inp_flags & INP_DROPPED) {
2367 		INP_RUNLOCK(inp);
2368 		return (ECONNRESET);
2369 	}
2370 
2371 	tp = intotcpcb(inp);
2372 	MPASS(tp != NULL);
2373 
2374 	/* Get the TCP sequence number of the next valid TLS header. */
2375 	SOCKBUF_LOCK(&so->so_rcv);
2376 	params.tls_rx.tls_hdr_tcp_sn =
2377 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2378 	params.tls_rx.tls_rec_length = tls_len;
2379 	params.tls_rx.tls_seq_number = tls_rcd_num;
2380 	SOCKBUF_UNLOCK(&so->so_rcv);
2381 
2382 	INP_RUNLOCK(inp);
2383 
2384 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2385 	return (mst->sw->snd_tag_modify(mst, &params));
2386 }
2387 
2388 static void
2389 ktls_drop(struct socket *so, int error)
2390 {
2391 	struct epoch_tracker et;
2392 	struct inpcb *inp = sotoinpcb(so);
2393 	struct tcpcb *tp;
2394 
2395 	NET_EPOCH_ENTER(et);
2396 	INP_WLOCK(inp);
2397 	if (!(inp->inp_flags & INP_DROPPED)) {
2398 		tp = intotcpcb(inp);
2399 		CURVNET_SET(inp->inp_vnet);
2400 		tp = tcp_drop(tp, error);
2401 		CURVNET_RESTORE();
2402 		if (tp != NULL)
2403 			INP_WUNLOCK(inp);
2404 	} else {
2405 		so->so_error = error;
2406 		SOCK_RECVBUF_LOCK(so);
2407 		sorwakeup_locked(so);
2408 		INP_WUNLOCK(inp);
2409 	}
2410 	NET_EPOCH_EXIT(et);
2411 }
2412 
2413 static void
2414 ktls_decrypt(struct socket *so)
2415 {
2416 	char tls_header[MBUF_PEXT_HDR_LEN];
2417 	struct ktls_session *tls;
2418 	struct sockbuf *sb;
2419 	struct tls_record_layer *hdr;
2420 	struct tls_get_record tgr;
2421 	struct mbuf *control, *data, *m;
2422 	ktls_mbuf_crypto_st_t state;
2423 	uint64_t seqno;
2424 	int error, remain, tls_len, trail_len;
2425 	bool tls13;
2426 	uint8_t vminor, record_type;
2427 
2428 	hdr = (struct tls_record_layer *)tls_header;
2429 	sb = &so->so_rcv;
2430 	SOCKBUF_LOCK(sb);
2431 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2432 	    ("%s: socket %p not running", __func__, so));
2433 
2434 	tls = sb->sb_tls_info;
2435 	MPASS(tls != NULL);
2436 
2437 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2438 	if (tls13)
2439 		vminor = TLS_MINOR_VER_TWO;
2440 	else
2441 		vminor = tls->params.tls_vminor;
2442 	for (;;) {
2443 		/* Is there enough queued for a TLS header? */
2444 		if (sb->sb_tlscc < tls->params.tls_hlen)
2445 			break;
2446 
2447 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2448 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2449 
2450 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2451 		    hdr->tls_vminor != vminor)
2452 			error = EINVAL;
2453 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2454 			error = EINVAL;
2455 		else if (tls_len < tls->params.tls_hlen || tls_len >
2456 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2457 		    tls->params.tls_tlen)
2458 			error = EMSGSIZE;
2459 		else
2460 			error = 0;
2461 		if (__predict_false(error != 0)) {
2462 			/*
2463 			 * We have a corrupted record and are likely
2464 			 * out of sync.  The connection isn't
2465 			 * recoverable at this point, so abort it.
2466 			 */
2467 			SOCKBUF_UNLOCK(sb);
2468 			counter_u64_add(ktls_offload_corrupted_records, 1);
2469 
2470 			ktls_drop(so, error);
2471 			goto deref;
2472 		}
2473 
2474 		/* Is the entire record queued? */
2475 		if (sb->sb_tlscc < tls_len)
2476 			break;
2477 
2478 		/*
2479 		 * Split out the portion of the mbuf chain containing
2480 		 * this TLS record.
2481 		 */
2482 		data = ktls_detach_record(sb, tls_len);
2483 		if (data == NULL)
2484 			continue;
2485 		MPASS(sb->sb_tlsdcc == tls_len);
2486 
2487 		seqno = sb->sb_tls_seqno;
2488 		sb->sb_tls_seqno++;
2489 		SBCHECK(sb);
2490 		SOCKBUF_UNLOCK(sb);
2491 
2492 		/* get crypto state for this TLS record */
2493 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2494 
2495 		switch (state) {
2496 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2497 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2498 			if (error)
2499 				break;
2500 			/* FALLTHROUGH */
2501 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2502 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2503 			    &trail_len);
2504 			if (__predict_true(error == 0)) {
2505 				if (tls13) {
2506 					error = tls13_find_record_type(tls, data,
2507 					    tls_len, &trail_len, &record_type);
2508 				} else {
2509 					record_type = hdr->tls_type;
2510 				}
2511 			}
2512 			break;
2513 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2514 			/*
2515 			 * NIC TLS is only supported for AEAD
2516 			 * ciphersuites which used a fixed sized
2517 			 * trailer.
2518 			 */
2519 			if (tls13) {
2520 				trail_len = tls->params.tls_tlen - 1;
2521 				error = tls13_find_record_type(tls, data,
2522 				    tls_len, &trail_len, &record_type);
2523 			} else {
2524 				trail_len = tls->params.tls_tlen;
2525 				error = 0;
2526 				record_type = hdr->tls_type;
2527 			}
2528 			break;
2529 		default:
2530 			error = EINVAL;
2531 			break;
2532 		}
2533 		if (error) {
2534 			counter_u64_add(ktls_offload_failed_crypto, 1);
2535 
2536 			SOCKBUF_LOCK(sb);
2537 			if (sb->sb_tlsdcc == 0) {
2538 				/*
2539 				 * sbcut/drop/flush discarded these
2540 				 * mbufs.
2541 				 */
2542 				m_freem(data);
2543 				break;
2544 			}
2545 
2546 			/*
2547 			 * Drop this TLS record's data, but keep
2548 			 * decrypting subsequent records.
2549 			 */
2550 			sb->sb_ccc -= tls_len;
2551 			sb->sb_tlsdcc = 0;
2552 
2553 			if (error != EMSGSIZE)
2554 				error = EBADMSG;
2555 			CURVNET_SET(so->so_vnet);
2556 			so->so_error = error;
2557 			sorwakeup_locked(so);
2558 			CURVNET_RESTORE();
2559 
2560 			m_freem(data);
2561 
2562 			SOCKBUF_LOCK(sb);
2563 			continue;
2564 		}
2565 
2566 		/* Allocate the control mbuf. */
2567 		memset(&tgr, 0, sizeof(tgr));
2568 		tgr.tls_type = record_type;
2569 		tgr.tls_vmajor = hdr->tls_vmajor;
2570 		tgr.tls_vminor = hdr->tls_vminor;
2571 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2572 		    trail_len);
2573 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2574 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2575 
2576 		SOCKBUF_LOCK(sb);
2577 		if (sb->sb_tlsdcc == 0) {
2578 			/* sbcut/drop/flush discarded these mbufs. */
2579 			MPASS(sb->sb_tlscc == 0);
2580 			m_freem(data);
2581 			m_freem(control);
2582 			break;
2583 		}
2584 
2585 		/*
2586 		 * Clear the 'dcc' accounting in preparation for
2587 		 * adding the decrypted record.
2588 		 */
2589 		sb->sb_ccc -= tls_len;
2590 		sb->sb_tlsdcc = 0;
2591 		SBCHECK(sb);
2592 
2593 		/* If there is no payload, drop all of the data. */
2594 		if (tgr.tls_length == htobe16(0)) {
2595 			m_freem(data);
2596 			data = NULL;
2597 		} else {
2598 			/* Trim header. */
2599 			remain = tls->params.tls_hlen;
2600 			while (remain > 0) {
2601 				if (data->m_len > remain) {
2602 					data->m_data += remain;
2603 					data->m_len -= remain;
2604 					break;
2605 				}
2606 				remain -= data->m_len;
2607 				data = m_free(data);
2608 			}
2609 
2610 			/* Trim trailer and clear M_NOTREADY. */
2611 			remain = be16toh(tgr.tls_length);
2612 			m = data;
2613 			for (m = data; remain > m->m_len; m = m->m_next) {
2614 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2615 				remain -= m->m_len;
2616 			}
2617 			m->m_len = remain;
2618 			m_freem(m->m_next);
2619 			m->m_next = NULL;
2620 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2621 
2622 			/* Set EOR on the final mbuf. */
2623 			m->m_flags |= M_EOR;
2624 		}
2625 
2626 		sbappendcontrol_locked(sb, data, control, 0);
2627 
2628 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2629 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2630 			SOCKBUF_UNLOCK(sb);
2631 			ktls_resync_ifnet(so, tls_len, seqno);
2632 			SOCKBUF_LOCK(sb);
2633 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2634 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2635 			SOCKBUF_UNLOCK(sb);
2636 			ktls_resync_ifnet(so, 0, seqno);
2637 			SOCKBUF_LOCK(sb);
2638 		}
2639 	}
2640 
2641 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2642 
2643 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2644 		so->so_error = EMSGSIZE;
2645 
2646 	sorwakeup_locked(so);
2647 
2648 deref:
2649 	SOCKBUF_UNLOCK_ASSERT(sb);
2650 
2651 	CURVNET_SET(so->so_vnet);
2652 	sorele(so);
2653 	CURVNET_RESTORE();
2654 }
2655 
2656 void
2657 ktls_enqueue_to_free(struct mbuf *m)
2658 {
2659 	struct ktls_wq *wq;
2660 	bool running;
2661 
2662 	/* Mark it for freeing. */
2663 	m->m_epg_flags |= EPG_FLAG_2FREE;
2664 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2665 	mtx_lock(&wq->mtx);
2666 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2667 	running = wq->running;
2668 	mtx_unlock(&wq->mtx);
2669 	if (!running)
2670 		wakeup(wq);
2671 }
2672 
2673 static void *
2674 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2675 {
2676 	void *buf;
2677 	int domain, running;
2678 
2679 	if (m->m_epg_npgs <= 2)
2680 		return (NULL);
2681 	if (ktls_buffer_zone == NULL)
2682 		return (NULL);
2683 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2684 		/*
2685 		 * Rate-limit allocation attempts after a failure.
2686 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2687 		 * the free page queues and may fail consistently if memory is
2688 		 * fragmented.
2689 		 */
2690 		return (NULL);
2691 	}
2692 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2693 	if (buf == NULL) {
2694 		domain = PCPU_GET(domain);
2695 		wq->lastallocfail = ticks;
2696 
2697 		/*
2698 		 * Note that this check is "racy", but the races are
2699 		 * harmless, and are either a spurious wakeup if
2700 		 * multiple threads fail allocations before the alloc
2701 		 * thread wakes, or waiting an extra second in case we
2702 		 * see an old value of running == true.
2703 		 */
2704 		if (!VM_DOMAIN_EMPTY(domain)) {
2705 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2706 			if (!running)
2707 				wakeup(&ktls_domains[domain].alloc_td);
2708 		}
2709 	}
2710 	return (buf);
2711 }
2712 
2713 static int
2714 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2715     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2716 {
2717 	vm_page_t pg;
2718 	int error, i, len, off;
2719 
2720 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2721 	    ("%p not unready & nomap mbuf\n", m));
2722 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2723 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2724 	    ktls_maxlen));
2725 
2726 	/* Anonymous mbufs are encrypted in place. */
2727 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2728 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2729 
2730 	/*
2731 	 * For file-backed mbufs (from sendfile), anonymous wired
2732 	 * pages are allocated and used as the encryption destination.
2733 	 */
2734 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2735 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2736 		    m->m_epg_1st_off;
2737 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2738 		    m->m_epg_1st_off;
2739 		state->dst_iov[0].iov_len = len;
2740 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2741 		i = 1;
2742 	} else {
2743 		off = m->m_epg_1st_off;
2744 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2745 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2746 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2747 			len = m_epg_pagelen(m, i, off);
2748 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2749 			state->dst_iov[i].iov_base =
2750 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2751 			state->dst_iov[i].iov_len = len;
2752 		}
2753 	}
2754 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2755 	state->dst_iov[i].iov_base = m->m_epg_trail;
2756 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2757 
2758 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2759 
2760 	if (__predict_false(error != 0)) {
2761 		/* Free the anonymous pages. */
2762 		if (state->cbuf != NULL)
2763 			uma_zfree(ktls_buffer_zone, state->cbuf);
2764 		else {
2765 			for (i = 0; i < m->m_epg_npgs; i++) {
2766 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2767 				(void)vm_page_unwire_noq(pg);
2768 				vm_page_free(pg);
2769 			}
2770 		}
2771 	}
2772 	return (error);
2773 }
2774 
2775 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2776 static u_int
2777 ktls_batched_records(struct mbuf *m)
2778 {
2779 	int page_count, records;
2780 
2781 	records = 0;
2782 	page_count = m->m_epg_enc_cnt;
2783 	while (page_count > 0) {
2784 		records++;
2785 		page_count -= m->m_epg_nrdy;
2786 		m = m->m_next;
2787 	}
2788 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2789 	return (records);
2790 }
2791 
2792 void
2793 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2794 {
2795 	struct ktls_session *tls;
2796 	struct ktls_wq *wq;
2797 	int queued;
2798 	bool running;
2799 
2800 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2801 	    (M_EXTPG | M_NOTREADY)),
2802 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2803 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2804 
2805 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2806 
2807 	m->m_epg_enc_cnt = page_count;
2808 
2809 	/*
2810 	 * Save a pointer to the socket.  The caller is responsible
2811 	 * for taking an additional reference via soref().
2812 	 */
2813 	m->m_epg_so = so;
2814 
2815 	queued = 1;
2816 	tls = m->m_epg_tls;
2817 	wq = &ktls_wq[tls->wq_index];
2818 	mtx_lock(&wq->mtx);
2819 	if (__predict_false(tls->sequential_records)) {
2820 		/*
2821 		 * For TLS 1.0, records must be encrypted
2822 		 * sequentially.  For a given connection, all records
2823 		 * queued to the associated work queue are processed
2824 		 * sequentially.  However, sendfile(2) might complete
2825 		 * I/O requests spanning multiple TLS records out of
2826 		 * order.  Here we ensure TLS records are enqueued to
2827 		 * the work queue in FIFO order.
2828 		 *
2829 		 * tls->next_seqno holds the sequence number of the
2830 		 * next TLS record that should be enqueued to the work
2831 		 * queue.  If this next record is not tls->next_seqno,
2832 		 * it must be a future record, so insert it, sorted by
2833 		 * TLS sequence number, into tls->pending_records and
2834 		 * return.
2835 		 *
2836 		 * If this TLS record matches tls->next_seqno, place
2837 		 * it in the work queue and then check
2838 		 * tls->pending_records to see if any
2839 		 * previously-queued records are now ready for
2840 		 * encryption.
2841 		 */
2842 		if (m->m_epg_seqno != tls->next_seqno) {
2843 			struct mbuf *n, *p;
2844 
2845 			p = NULL;
2846 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2847 				if (n->m_epg_seqno > m->m_epg_seqno)
2848 					break;
2849 				p = n;
2850 			}
2851 			if (n == NULL)
2852 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2853 				    m_epg_stailq);
2854 			else if (p == NULL)
2855 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2856 				    m_epg_stailq);
2857 			else
2858 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2859 				    m_epg_stailq);
2860 			mtx_unlock(&wq->mtx);
2861 			counter_u64_add(ktls_cnt_tx_pending, 1);
2862 			return;
2863 		}
2864 
2865 		tls->next_seqno += ktls_batched_records(m);
2866 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2867 
2868 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2869 			struct mbuf *n;
2870 
2871 			n = STAILQ_FIRST(&tls->pending_records);
2872 			if (n->m_epg_seqno != tls->next_seqno)
2873 				break;
2874 
2875 			queued++;
2876 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2877 			tls->next_seqno += ktls_batched_records(n);
2878 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2879 		}
2880 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2881 	} else
2882 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2883 
2884 	running = wq->running;
2885 	mtx_unlock(&wq->mtx);
2886 	if (!running)
2887 		wakeup(wq);
2888 	counter_u64_add(ktls_cnt_tx_queued, queued);
2889 }
2890 
2891 /*
2892  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2893  * the pages from the file and replace them with the anonymous pages
2894  * allocated in ktls_encrypt_record().
2895  */
2896 static void
2897 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2898 {
2899 	int i;
2900 
2901 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2902 
2903 	/* Free the old pages. */
2904 	m->m_ext.ext_free(m);
2905 
2906 	/* Replace them with the new pages. */
2907 	if (state->cbuf != NULL) {
2908 		for (i = 0; i < m->m_epg_npgs; i++)
2909 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2910 
2911 		/* Contig pages should go back to the cache. */
2912 		m->m_ext.ext_free = ktls_free_mext_contig;
2913 	} else {
2914 		for (i = 0; i < m->m_epg_npgs; i++)
2915 			m->m_epg_pa[i] = state->parray[i];
2916 
2917 		/* Use the basic free routine. */
2918 		m->m_ext.ext_free = mb_free_mext_pgs;
2919 	}
2920 
2921 	/* Pages are now writable. */
2922 	m->m_epg_flags |= EPG_FLAG_ANON;
2923 }
2924 
2925 static __noinline void
2926 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2927 {
2928 	struct ktls_ocf_encrypt_state state;
2929 	struct ktls_session *tls;
2930 	struct socket *so;
2931 	struct mbuf *m;
2932 	int error, npages, total_pages;
2933 
2934 	so = top->m_epg_so;
2935 	tls = top->m_epg_tls;
2936 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2937 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2938 #ifdef INVARIANTS
2939 	top->m_epg_so = NULL;
2940 #endif
2941 	total_pages = top->m_epg_enc_cnt;
2942 	npages = 0;
2943 
2944 	/*
2945 	 * Encrypt the TLS records in the chain of mbufs starting with
2946 	 * 'top'.  'total_pages' gives us a total count of pages and is
2947 	 * used to know when we have finished encrypting the TLS
2948 	 * records originally queued with 'top'.
2949 	 *
2950 	 * NB: These mbufs are queued in the socket buffer and
2951 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2952 	 * socket buffer lock is not held while traversing this chain.
2953 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2954 	 * pointers should be stable.  However, the 'm_next' of the
2955 	 * last mbuf encrypted is not necessarily NULL.  It can point
2956 	 * to other mbufs appended while 'top' was on the TLS work
2957 	 * queue.
2958 	 *
2959 	 * Each mbuf holds an entire TLS record.
2960 	 */
2961 	error = 0;
2962 	for (m = top; npages != total_pages; m = m->m_next) {
2963 		KASSERT(m->m_epg_tls == tls,
2964 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2965 		    tls, m->m_epg_tls));
2966 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2967 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2968 		    total_pages, m));
2969 
2970 		error = ktls_encrypt_record(wq, m, tls, &state);
2971 		if (error) {
2972 			counter_u64_add(ktls_offload_failed_crypto, 1);
2973 			break;
2974 		}
2975 
2976 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2977 			ktls_finish_nonanon(m, &state);
2978 
2979 		npages += m->m_epg_nrdy;
2980 
2981 		/*
2982 		 * Drop a reference to the session now that it is no
2983 		 * longer needed.  Existing code depends on encrypted
2984 		 * records having no associated session vs
2985 		 * yet-to-be-encrypted records having an associated
2986 		 * session.
2987 		 */
2988 		m->m_epg_tls = NULL;
2989 		ktls_free(tls);
2990 	}
2991 
2992 	CURVNET_SET(so->so_vnet);
2993 	if (error == 0) {
2994 		(void)so->so_proto->pr_ready(so, top, npages);
2995 	} else {
2996 		ktls_drop(so, EIO);
2997 		mb_free_notready(top, total_pages);
2998 	}
2999 
3000 	sorele(so);
3001 	CURVNET_RESTORE();
3002 }
3003 
3004 void
3005 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3006 {
3007 	struct ktls_session *tls;
3008 	struct socket *so;
3009 	struct mbuf *m;
3010 	int npages;
3011 
3012 	m = state->m;
3013 
3014 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3015 		ktls_finish_nonanon(m, state);
3016 
3017 	so = state->so;
3018 	free(state, M_KTLS);
3019 
3020 	/*
3021 	 * Drop a reference to the session now that it is no longer
3022 	 * needed.  Existing code depends on encrypted records having
3023 	 * no associated session vs yet-to-be-encrypted records having
3024 	 * an associated session.
3025 	 */
3026 	tls = m->m_epg_tls;
3027 	m->m_epg_tls = NULL;
3028 	ktls_free(tls);
3029 
3030 	if (error != 0)
3031 		counter_u64_add(ktls_offload_failed_crypto, 1);
3032 
3033 	CURVNET_SET(so->so_vnet);
3034 	npages = m->m_epg_nrdy;
3035 
3036 	if (error == 0) {
3037 		(void)so->so_proto->pr_ready(so, m, npages);
3038 	} else {
3039 		ktls_drop(so, EIO);
3040 		mb_free_notready(m, npages);
3041 	}
3042 
3043 	sorele(so);
3044 	CURVNET_RESTORE();
3045 }
3046 
3047 /*
3048  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3049  * (coprocessors) where encryption does not use host CPU resources and
3050  * it can be beneficial to queue more requests than CPUs.
3051  */
3052 static __noinline void
3053 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3054 {
3055 	struct ktls_ocf_encrypt_state *state;
3056 	struct ktls_session *tls;
3057 	struct socket *so;
3058 	struct mbuf *m, *n;
3059 	int error, mpages, npages, total_pages;
3060 
3061 	so = top->m_epg_so;
3062 	tls = top->m_epg_tls;
3063 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3064 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3065 #ifdef INVARIANTS
3066 	top->m_epg_so = NULL;
3067 #endif
3068 	total_pages = top->m_epg_enc_cnt;
3069 	npages = 0;
3070 
3071 	error = 0;
3072 	for (m = top; npages != total_pages; m = n) {
3073 		KASSERT(m->m_epg_tls == tls,
3074 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3075 		    tls, m->m_epg_tls));
3076 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3077 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3078 		    total_pages, m));
3079 
3080 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3081 		soref(so);
3082 		state->so = so;
3083 		state->m = m;
3084 
3085 		mpages = m->m_epg_nrdy;
3086 		n = m->m_next;
3087 
3088 		error = ktls_encrypt_record(wq, m, tls, state);
3089 		if (error) {
3090 			counter_u64_add(ktls_offload_failed_crypto, 1);
3091 			free(state, M_KTLS);
3092 			CURVNET_SET(so->so_vnet);
3093 			sorele(so);
3094 			CURVNET_RESTORE();
3095 			break;
3096 		}
3097 
3098 		npages += mpages;
3099 	}
3100 
3101 	CURVNET_SET(so->so_vnet);
3102 	if (error != 0) {
3103 		ktls_drop(so, EIO);
3104 		mb_free_notready(m, total_pages - npages);
3105 	}
3106 
3107 	sorele(so);
3108 	CURVNET_RESTORE();
3109 }
3110 
3111 static int
3112 ktls_bind_domain(int domain)
3113 {
3114 	int error;
3115 
3116 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3117 	if (error != 0)
3118 		return (error);
3119 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3120 	return (0);
3121 }
3122 
3123 static void
3124 ktls_alloc_thread(void *ctx)
3125 {
3126 	struct ktls_domain_info *ktls_domain = ctx;
3127 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3128 	void **buf;
3129 	struct sysctl_oid *oid;
3130 	char name[80];
3131 	int domain, error, i, nbufs;
3132 
3133 	domain = ktls_domain - ktls_domains;
3134 	if (bootverbose)
3135 		printf("Starting KTLS alloc thread for domain %d\n", domain);
3136 	error = ktls_bind_domain(domain);
3137 	if (error)
3138 		printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3139 		    domain, error);
3140 	snprintf(name, sizeof(name), "domain%d", domain);
3141 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3142 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3143 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3144 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
3145 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3146 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3147 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3148 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3149 
3150 	buf = NULL;
3151 	nbufs = 0;
3152 	for (;;) {
3153 		atomic_store_int(&sc->running, 0);
3154 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3155 		atomic_store_int(&sc->running, 1);
3156 		sc->wakeups++;
3157 		if (nbufs != ktls_max_alloc) {
3158 			free(buf, M_KTLS);
3159 			nbufs = atomic_load_int(&ktls_max_alloc);
3160 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3161 			    M_WAITOK | M_ZERO);
3162 		}
3163 		/*
3164 		 * Below we allocate nbufs with different allocation
3165 		 * flags than we use when allocating normally during
3166 		 * encryption in the ktls worker thread.  We specify
3167 		 * M_NORECLAIM in the worker thread. However, we omit
3168 		 * that flag here and add M_WAITOK so that the VM
3169 		 * system is permitted to perform expensive work to
3170 		 * defragment memory.  We do this here, as it does not
3171 		 * matter if this thread blocks.  If we block a ktls
3172 		 * worker thread, we risk developing backlogs of
3173 		 * buffers to be encrypted, leading to surges of
3174 		 * traffic and potential NIC output drops.
3175 		 */
3176 		for (i = 0; i < nbufs; i++) {
3177 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3178 			sc->allocs++;
3179 		}
3180 		for (i = 0; i < nbufs; i++) {
3181 			uma_zfree(ktls_buffer_zone, buf[i]);
3182 			buf[i] = NULL;
3183 		}
3184 	}
3185 }
3186 
3187 static void
3188 ktls_work_thread(void *ctx)
3189 {
3190 	struct ktls_wq *wq = ctx;
3191 	struct mbuf *m, *n;
3192 	struct socket *so, *son;
3193 	STAILQ_HEAD(, mbuf) local_m_head;
3194 	STAILQ_HEAD(, socket) local_so_head;
3195 	int cpu;
3196 
3197 	cpu = wq - ktls_wq;
3198 	if (bootverbose)
3199 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3200 
3201 	/*
3202 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3203 	 * we bind to the NUMA domain instead.
3204 	 */
3205 	if (ktls_bind_threads) {
3206 		int error;
3207 
3208 		if (ktls_bind_threads > 1) {
3209 			struct pcpu *pc = pcpu_find(cpu);
3210 
3211 			error = ktls_bind_domain(pc->pc_domain);
3212 		} else {
3213 			cpuset_t mask;
3214 
3215 			CPU_SETOF(cpu, &mask);
3216 			error = cpuset_setthread(curthread->td_tid, &mask);
3217 		}
3218 		if (error)
3219 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3220 				cpu, error);
3221 	}
3222 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3223 	fpu_kern_thread(0);
3224 #endif
3225 	for (;;) {
3226 		mtx_lock(&wq->mtx);
3227 		while (STAILQ_EMPTY(&wq->m_head) &&
3228 		    STAILQ_EMPTY(&wq->so_head)) {
3229 			wq->running = false;
3230 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3231 			wq->running = true;
3232 		}
3233 
3234 		STAILQ_INIT(&local_m_head);
3235 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3236 		STAILQ_INIT(&local_so_head);
3237 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3238 		mtx_unlock(&wq->mtx);
3239 
3240 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3241 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3242 				ktls_free(m->m_epg_tls);
3243 				m_free_raw(m);
3244 			} else {
3245 				if (m->m_epg_tls->sync_dispatch)
3246 					ktls_encrypt(wq, m);
3247 				else
3248 					ktls_encrypt_async(wq, m);
3249 				counter_u64_add(ktls_cnt_tx_queued, -1);
3250 			}
3251 		}
3252 
3253 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3254 			ktls_decrypt(so);
3255 			counter_u64_add(ktls_cnt_rx_queued, -1);
3256 		}
3257 	}
3258 }
3259 
3260 static void
3261 ktls_disable_ifnet_help(void *context, int pending __unused)
3262 {
3263 	struct ktls_session *tls;
3264 	struct inpcb *inp;
3265 	struct tcpcb *tp;
3266 	struct socket *so;
3267 	int err;
3268 
3269 	tls = context;
3270 	inp = tls->inp;
3271 	if (inp == NULL)
3272 		return;
3273 	INP_WLOCK(inp);
3274 	so = inp->inp_socket;
3275 	MPASS(so != NULL);
3276 	if (inp->inp_flags & INP_DROPPED) {
3277 		goto out;
3278 	}
3279 
3280 	if (so->so_snd.sb_tls_info != NULL)
3281 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3282 	else
3283 		err = ENXIO;
3284 	if (err == 0) {
3285 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3286 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3287 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3288 		    (tp = intotcpcb(inp)) != NULL &&
3289 		    tp->t_fb->tfb_hwtls_change != NULL)
3290 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3291 	} else {
3292 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3293 	}
3294 
3295 out:
3296 	CURVNET_SET(so->so_vnet);
3297 	sorele(so);
3298 	CURVNET_RESTORE();
3299 	INP_WUNLOCK(inp);
3300 	ktls_free(tls);
3301 }
3302 
3303 /*
3304  * Called when re-transmits are becoming a substantial portion of the
3305  * sends on this connection.  When this happens, we transition the
3306  * connection to software TLS.  This is needed because most inline TLS
3307  * NICs keep crypto state only for in-order transmits.  This means
3308  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3309  * re-DMA the entire TLS record up to and including the current
3310  * segment.  This means that when re-transmitting the last ~1448 byte
3311  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3312  * of magnitude more data than we are sending.  This can cause the
3313  * PCIe link to saturate well before the network, which can cause
3314  * output drops, and a general loss of capacity.
3315  */
3316 void
3317 ktls_disable_ifnet(void *arg)
3318 {
3319 	struct tcpcb *tp;
3320 	struct inpcb *inp;
3321 	struct socket *so;
3322 	struct ktls_session *tls;
3323 
3324 	tp = arg;
3325 	inp = tptoinpcb(tp);
3326 	INP_WLOCK_ASSERT(inp);
3327 	so = inp->inp_socket;
3328 	SOCK_LOCK(so);
3329 	tls = so->so_snd.sb_tls_info;
3330 	if (tp->t_nic_ktls_xmit_dis == 1) {
3331 		SOCK_UNLOCK(so);
3332 		return;
3333 	}
3334 
3335 	/*
3336 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3337 	 * ifnet can only be done once per connection, so we never want
3338 	 * to do it again
3339 	 */
3340 
3341 	(void)ktls_hold(tls);
3342 	soref(so);
3343 	tp->t_nic_ktls_xmit_dis = 1;
3344 	SOCK_UNLOCK(so);
3345 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3346 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3347 }
3348