xref: /freebsd/sys/kern/uipc_ktls.c (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_kern_tls.h"
32 #include "opt_ratelimit.h"
33 #include "opt_rss.h"
34 
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/domainset.h>
38 #include <sys/endian.h>
39 #include <sys/ktls.h>
40 #include <sys/lock.h>
41 #include <sys/mbuf.h>
42 #include <sys/mutex.h>
43 #include <sys/rmlock.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/refcount.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/taskqueue.h>
52 #include <sys/kthread.h>
53 #include <sys/uio.h>
54 #include <sys/vmmeter.h>
55 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
56 #include <machine/pcb.h>
57 #endif
58 #include <machine/vmparam.h>
59 #include <net/if.h>
60 #include <net/if_var.h>
61 #ifdef RSS
62 #include <net/netisr.h>
63 #include <net/rss_config.h>
64 #endif
65 #include <net/route.h>
66 #include <net/route/nhop.h>
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_var.h>
70 #ifdef TCP_OFFLOAD
71 #include <netinet/tcp_offload.h>
72 #endif
73 #include <opencrypto/cryptodev.h>
74 #include <opencrypto/ktls.h>
75 #include <vm/vm.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_pagequeue.h>
79 
80 struct ktls_wq {
81 	struct mtx	mtx;
82 	STAILQ_HEAD(, mbuf) m_head;
83 	STAILQ_HEAD(, socket) so_head;
84 	bool		running;
85 	int		lastallocfail;
86 } __aligned(CACHE_LINE_SIZE);
87 
88 struct ktls_reclaim_thread {
89 	uint64_t wakeups;
90 	uint64_t reclaims;
91 	struct thread *td;
92 	int running;
93 };
94 
95 struct ktls_domain_info {
96 	int count;
97 	int cpu[MAXCPU];
98 	struct ktls_reclaim_thread reclaim_td;
99 };
100 
101 struct ktls_domain_info ktls_domains[MAXMEMDOM];
102 static struct ktls_wq *ktls_wq;
103 static struct proc *ktls_proc;
104 static uma_zone_t ktls_session_zone;
105 static uma_zone_t ktls_buffer_zone;
106 static uint16_t ktls_cpuid_lookup[MAXCPU];
107 static int ktls_init_state;
108 static struct sx ktls_init_lock;
109 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
110 
111 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
112     "Kernel TLS offload");
113 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114     "Kernel TLS offload stats");
115 
116 #ifdef RSS
117 static int ktls_bind_threads = 1;
118 #else
119 static int ktls_bind_threads;
120 #endif
121 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
122     &ktls_bind_threads, 0,
123     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
124 
125 static u_int ktls_maxlen = 16384;
126 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
127     &ktls_maxlen, 0, "Maximum TLS record size");
128 
129 static int ktls_number_threads;
130 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
131     &ktls_number_threads, 0,
132     "Number of TLS threads in thread-pool");
133 
134 unsigned int ktls_ifnet_max_rexmit_pct = 2;
135 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
136     &ktls_ifnet_max_rexmit_pct, 2,
137     "Max percent bytes retransmitted before ifnet TLS is disabled");
138 
139 static bool ktls_offload_enable;
140 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
141     &ktls_offload_enable, 0,
142     "Enable support for kernel TLS offload");
143 
144 static bool ktls_cbc_enable = true;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
146     &ktls_cbc_enable, 1,
147     "Enable support of AES-CBC crypto for kernel TLS");
148 
149 static bool ktls_sw_buffer_cache = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
151     &ktls_sw_buffer_cache, 1,
152     "Enable caching of output buffers for SW encryption");
153 
154 static int ktls_max_reclaim = 1024;
155 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN,
156     &ktls_max_reclaim, 128,
157     "Max number of 16k buffers to reclaim in thread context");
158 
159 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
160 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
161     &ktls_tasks_active, "Number of active tasks");
162 
163 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
164 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
165     &ktls_cnt_tx_pending,
166     "Number of TLS 1.0 records waiting for earlier TLS records");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
170     &ktls_cnt_tx_queued,
171     "Number of TLS records in queue to tasks for SW encryption");
172 
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
175     &ktls_cnt_rx_queued,
176     "Number of TLS sockets in queue to tasks for SW decryption");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
180     CTLFLAG_RD, &ktls_offload_total,
181     "Total successful TLS setups (parameters set)");
182 
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
185     CTLFLAG_RD, &ktls_offload_enable_calls,
186     "Total number of TLS enable calls made");
187 
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
190     &ktls_offload_active, "Total Active TLS sessions");
191 
192 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
193 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
194     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
195 
196 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
197 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
198     &ktls_offload_failed_crypto, "Total TLS crypto failures");
199 
200 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
201 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
202     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
203 
204 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
205 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
206     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
207 
208 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
209 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
210     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
211 
212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
214     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
215 
216 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
218     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
219 
220 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
222     &ktls_destroy_task,
223     "Number of times ktls session was destroyed via taskqueue");
224 
225 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226     "Software TLS session stats");
227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228     "Hardware (ifnet) TLS session stats");
229 #ifdef TCP_OFFLOAD
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "TOE TLS session stats");
232 #endif
233 
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236     "Active number of software TLS sessions using AES-CBC");
237 
238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240     "Active number of software TLS sessions using AES-GCM");
241 
242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
244     &ktls_sw_chacha20,
245     "Active number of software TLS sessions using Chacha20-Poly1305");
246 
247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
249     &ktls_ifnet_cbc,
250     "Active number of ifnet TLS sessions using AES-CBC");
251 
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
254     &ktls_ifnet_gcm,
255     "Active number of ifnet TLS sessions using AES-GCM");
256 
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259     &ktls_ifnet_chacha20,
260     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
261 
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
265 
266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268     &ktls_ifnet_reset_dropped,
269     "TLS sessions dropped after failing to update ifnet send tag");
270 
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273     &ktls_ifnet_reset_failed,
274     "TLS sessions that failed to allocate a new ifnet send tag");
275 
276 static int ktls_ifnet_permitted;
277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278     &ktls_ifnet_permitted, 1,
279     "Whether to permit hardware (ifnet) TLS sessions");
280 
281 #ifdef TCP_OFFLOAD
282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
284     &ktls_toe_cbc,
285     "Active number of TOE TLS sessions using AES-CBC");
286 
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
289     &ktls_toe_gcm,
290     "Active number of TOE TLS sessions using AES-GCM");
291 
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
294     &ktls_toe_chacha20,
295     "Active number of TOE TLS sessions using Chacha20-Poly1305");
296 #endif
297 
298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
299 
300 static void ktls_reclaim_thread(void *ctx);
301 static void ktls_reset_receive_tag(void *context, int pending);
302 static void ktls_reset_send_tag(void *context, int pending);
303 static void ktls_work_thread(void *ctx);
304 
305 int
306 ktls_copyin_tls_enable(struct sockopt *sopt, struct tls_enable *tls)
307 {
308 	struct tls_enable_v0 tls_v0;
309 	int error;
310 	uint8_t *cipher_key = NULL, *iv = NULL, *auth_key = NULL;
311 
312 	if (sopt->sopt_valsize == sizeof(tls_v0)) {
313 		error = sooptcopyin(sopt, &tls_v0, sizeof(tls_v0), sizeof(tls_v0));
314 		if (error != 0)
315 			goto done;
316 		memset(tls, 0, sizeof(*tls));
317 		tls->cipher_key = tls_v0.cipher_key;
318 		tls->iv = tls_v0.iv;
319 		tls->auth_key = tls_v0.auth_key;
320 		tls->cipher_algorithm = tls_v0.cipher_algorithm;
321 		tls->cipher_key_len = tls_v0.cipher_key_len;
322 		tls->iv_len = tls_v0.iv_len;
323 		tls->auth_algorithm = tls_v0.auth_algorithm;
324 		tls->auth_key_len = tls_v0.auth_key_len;
325 		tls->flags = tls_v0.flags;
326 		tls->tls_vmajor = tls_v0.tls_vmajor;
327 		tls->tls_vminor = tls_v0.tls_vminor;
328 	} else
329 		error = sooptcopyin(sopt, tls, sizeof(*tls), sizeof(*tls));
330 
331 	if (error != 0)
332 		return (error);
333 
334 	if (tls->cipher_key_len < 0 || tls->cipher_key_len > TLS_MAX_PARAM_SIZE)
335 		return (EINVAL);
336 	if (tls->iv_len < 0 || tls->iv_len > sizeof(((struct ktls_session *)NULL)->params.iv))
337 		return (EINVAL);
338 	if (tls->auth_key_len < 0 || tls->auth_key_len > TLS_MAX_PARAM_SIZE)
339 		return (EINVAL);
340 
341 	/* All supported algorithms require a cipher key. */
342 	if (tls->cipher_key_len == 0)
343 		return (EINVAL);
344 
345 	/*
346 	 * Now do a deep copy of the variable-length arrays in the struct, so that
347 	 * subsequent consumers of it can reliably assume kernel memory. This
348 	 * requires doing our own allocations, which we will free in the
349 	 * error paths so that our caller need only worry about outstanding
350 	 * allocations existing on successful return.
351 	 */
352 	if (tls->cipher_key_len != 0) {
353 		cipher_key = malloc(tls->cipher_key_len, M_KTLS, M_WAITOK);
354 		if (sopt->sopt_td != NULL) {
355 			error = copyin(tls->cipher_key, cipher_key, tls->cipher_key_len);
356 			if (error != 0)
357 				goto done;
358 		} else {
359 			bcopy(tls->cipher_key, cipher_key, tls->cipher_key_len);
360 		}
361 	}
362 	if (tls->iv_len != 0) {
363 		iv = malloc(tls->iv_len, M_KTLS, M_WAITOK);
364 		if (sopt->sopt_td != NULL) {
365 			error = copyin(tls->iv, iv, tls->iv_len);
366 			if (error != 0)
367 				goto done;
368 		} else {
369 			bcopy(tls->iv, iv, tls->iv_len);
370 		}
371 	}
372 	if (tls->auth_key_len != 0) {
373 		auth_key = malloc(tls->auth_key_len, M_KTLS, M_WAITOK);
374 		if (sopt->sopt_td != NULL) {
375 			error = copyin(tls->auth_key, auth_key, tls->auth_key_len);
376 			if (error != 0)
377 				goto done;
378 		} else {
379 			bcopy(tls->auth_key, auth_key, tls->auth_key_len);
380 		}
381 	}
382 	tls->cipher_key = cipher_key;
383 	tls->iv = iv;
384 	tls->auth_key = auth_key;
385 
386 done:
387 	if (error != 0) {
388 		zfree(cipher_key, M_KTLS);
389 		zfree(iv, M_KTLS);
390 		zfree(auth_key, M_KTLS);
391 	}
392 
393 	return (error);
394 }
395 
396 void
397 ktls_cleanup_tls_enable(struct tls_enable *tls)
398 {
399 	zfree(__DECONST(void *, tls->cipher_key), M_KTLS);
400 	zfree(__DECONST(void *, tls->iv), M_KTLS);
401 	zfree(__DECONST(void *, tls->auth_key), M_KTLS);
402 }
403 
404 static u_int
405 ktls_get_cpu(struct socket *so)
406 {
407 	struct inpcb *inp;
408 #ifdef NUMA
409 	struct ktls_domain_info *di;
410 #endif
411 	u_int cpuid;
412 
413 	inp = sotoinpcb(so);
414 #ifdef RSS
415 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
416 	if (cpuid != NETISR_CPUID_NONE)
417 		return (cpuid);
418 #endif
419 	/*
420 	 * Just use the flowid to shard connections in a repeatable
421 	 * fashion.  Note that TLS 1.0 sessions rely on the
422 	 * serialization provided by having the same connection use
423 	 * the same queue.
424 	 */
425 #ifdef NUMA
426 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
427 		di = &ktls_domains[inp->inp_numa_domain];
428 		cpuid = di->cpu[inp->inp_flowid % di->count];
429 	} else
430 #endif
431 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
432 	return (cpuid);
433 }
434 
435 static int
436 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
437 {
438 	vm_page_t m;
439 	int i, req;
440 
441 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
442 	    ("%s: ktls max length %d is not page size-aligned",
443 	    __func__, ktls_maxlen));
444 
445 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
446 	for (i = 0; i < count; i++) {
447 		m = vm_page_alloc_noobj_contig_domain(domain, req,
448 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
449 		    VM_MEMATTR_DEFAULT);
450 		if (m == NULL)
451 			break;
452 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
453 	}
454 	return (i);
455 }
456 
457 static void
458 ktls_buffer_release(void *arg __unused, void **store, int count)
459 {
460 	vm_page_t m;
461 	int i, j;
462 
463 	for (i = 0; i < count; i++) {
464 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
465 		for (j = 0; j < atop(ktls_maxlen); j++) {
466 			(void)vm_page_unwire_noq(m + j);
467 			vm_page_free(m + j);
468 		}
469 	}
470 }
471 
472 static void
473 ktls_free_mext_contig(struct mbuf *m)
474 {
475 	M_ASSERTEXTPG(m);
476 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
477 }
478 
479 static int
480 ktls_init(void)
481 {
482 	struct thread *td;
483 	struct pcpu *pc;
484 	int count, domain, error, i;
485 
486 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
487 	    M_WAITOK | M_ZERO);
488 
489 	ktls_session_zone = uma_zcreate("ktls_session",
490 	    sizeof(struct ktls_session),
491 	    NULL, NULL, NULL, NULL,
492 	    UMA_ALIGN_CACHE, 0);
493 
494 	if (ktls_sw_buffer_cache) {
495 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
496 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
497 		    ktls_buffer_import, ktls_buffer_release, NULL,
498 		    UMA_ZONE_FIRSTTOUCH);
499 	}
500 
501 	/*
502 	 * Initialize the workqueues to run the TLS work.  We create a
503 	 * work queue for each CPU.
504 	 */
505 	CPU_FOREACH(i) {
506 		STAILQ_INIT(&ktls_wq[i].m_head);
507 		STAILQ_INIT(&ktls_wq[i].so_head);
508 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
509 		if (ktls_bind_threads > 1) {
510 			pc = pcpu_find(i);
511 			domain = pc->pc_domain;
512 			count = ktls_domains[domain].count;
513 			ktls_domains[domain].cpu[count] = i;
514 			ktls_domains[domain].count++;
515 		}
516 		ktls_cpuid_lookup[ktls_number_threads] = i;
517 		ktls_number_threads++;
518 	}
519 
520 	/*
521 	 * If we somehow have an empty domain, fall back to choosing
522 	 * among all KTLS threads.
523 	 */
524 	if (ktls_bind_threads > 1) {
525 		for (i = 0; i < vm_ndomains; i++) {
526 			if (ktls_domains[i].count == 0) {
527 				ktls_bind_threads = 1;
528 				break;
529 			}
530 		}
531 	}
532 
533 	/* Start kthreads for each workqueue. */
534 	CPU_FOREACH(i) {
535 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
536 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
537 		if (error) {
538 			printf("Can't add KTLS thread %d error %d\n", i, error);
539 			return (error);
540 		}
541 	}
542 
543 	/*
544 	 * Start an allocation thread per-domain to perform blocking allocations
545 	 * of 16k physically contiguous TLS crypto destination buffers.
546 	 */
547 	if (ktls_sw_buffer_cache) {
548 		for (domain = 0; domain < vm_ndomains; domain++) {
549 			if (VM_DOMAIN_EMPTY(domain))
550 				continue;
551 			if (CPU_EMPTY(&cpuset_domain[domain]))
552 				continue;
553 			error = kproc_kthread_add(ktls_reclaim_thread,
554 			    &ktls_domains[domain], &ktls_proc,
555 			    &ktls_domains[domain].reclaim_td.td,
556 			    0, 0, "KTLS", "reclaim_%d", domain);
557 			if (error) {
558 				printf("Can't add KTLS reclaim thread %d error %d\n",
559 				    domain, error);
560 				return (error);
561 			}
562 		}
563 	}
564 
565 	if (bootverbose)
566 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
567 	return (0);
568 }
569 
570 static int
571 ktls_start_kthreads(void)
572 {
573 	int error, state;
574 
575 start:
576 	state = atomic_load_acq_int(&ktls_init_state);
577 	if (__predict_true(state > 0))
578 		return (0);
579 	if (state < 0)
580 		return (ENXIO);
581 
582 	sx_xlock(&ktls_init_lock);
583 	if (ktls_init_state != 0) {
584 		sx_xunlock(&ktls_init_lock);
585 		goto start;
586 	}
587 
588 	error = ktls_init();
589 	if (error == 0)
590 		state = 1;
591 	else
592 		state = -1;
593 	atomic_store_rel_int(&ktls_init_state, state);
594 	sx_xunlock(&ktls_init_lock);
595 	return (error);
596 }
597 
598 static int
599 ktls_create_session(struct socket *so, struct tls_enable *en,
600     struct ktls_session **tlsp, int direction)
601 {
602 	struct ktls_session *tls;
603 	int error;
604 
605 	/* Only TLS 1.0 - 1.3 are supported. */
606 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
607 		return (EINVAL);
608 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
609 	    en->tls_vminor > TLS_MINOR_VER_THREE)
610 		return (EINVAL);
611 
612 
613 	/* No flags are currently supported. */
614 	if (en->flags != 0)
615 		return (EINVAL);
616 
617 	/* Common checks for supported algorithms. */
618 	switch (en->cipher_algorithm) {
619 	case CRYPTO_AES_NIST_GCM_16:
620 		/*
621 		 * auth_algorithm isn't used, but permit GMAC values
622 		 * for compatibility.
623 		 */
624 		switch (en->auth_algorithm) {
625 		case 0:
626 #ifdef COMPAT_FREEBSD12
627 		/* XXX: Really 13.0-current COMPAT. */
628 		case CRYPTO_AES_128_NIST_GMAC:
629 		case CRYPTO_AES_192_NIST_GMAC:
630 		case CRYPTO_AES_256_NIST_GMAC:
631 #endif
632 			break;
633 		default:
634 			return (EINVAL);
635 		}
636 		if (en->auth_key_len != 0)
637 			return (EINVAL);
638 		switch (en->tls_vminor) {
639 		case TLS_MINOR_VER_TWO:
640 			if (en->iv_len != TLS_AEAD_GCM_LEN)
641 				return (EINVAL);
642 			break;
643 		case TLS_MINOR_VER_THREE:
644 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
645 				return (EINVAL);
646 			break;
647 		default:
648 			return (EINVAL);
649 		}
650 		break;
651 	case CRYPTO_AES_CBC:
652 		switch (en->auth_algorithm) {
653 		case CRYPTO_SHA1_HMAC:
654 			break;
655 		case CRYPTO_SHA2_256_HMAC:
656 		case CRYPTO_SHA2_384_HMAC:
657 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
658 				return (EINVAL);
659 			break;
660 		default:
661 			return (EINVAL);
662 		}
663 		if (en->auth_key_len == 0)
664 			return (EINVAL);
665 
666 		/*
667 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
668 		 * use explicit IVs.
669 		 */
670 		switch (en->tls_vminor) {
671 		case TLS_MINOR_VER_ZERO:
672 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
673 				return (EINVAL);
674 			break;
675 		case TLS_MINOR_VER_ONE:
676 		case TLS_MINOR_VER_TWO:
677 			/* Ignore any supplied IV. */
678 			en->iv_len = 0;
679 			break;
680 		default:
681 			return (EINVAL);
682 		}
683 		break;
684 	case CRYPTO_CHACHA20_POLY1305:
685 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
686 			return (EINVAL);
687 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
688 		    en->tls_vminor != TLS_MINOR_VER_THREE)
689 			return (EINVAL);
690 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
691 			return (EINVAL);
692 		break;
693 	default:
694 		return (EINVAL);
695 	}
696 
697 	error = ktls_start_kthreads();
698 	if (error != 0)
699 		return (error);
700 
701 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
702 
703 	counter_u64_add(ktls_offload_active, 1);
704 
705 	refcount_init(&tls->refcount, 1);
706 	if (direction == KTLS_RX) {
707 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
708 	} else {
709 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
710 		tls->inp = so->so_pcb;
711 		in_pcbref(tls->inp);
712 		tls->tx = true;
713 	}
714 
715 	tls->wq_index = ktls_get_cpu(so);
716 
717 	tls->params.cipher_algorithm = en->cipher_algorithm;
718 	tls->params.auth_algorithm = en->auth_algorithm;
719 	tls->params.tls_vmajor = en->tls_vmajor;
720 	tls->params.tls_vminor = en->tls_vminor;
721 	tls->params.flags = en->flags;
722 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
723 
724 	/* Set the header and trailer lengths. */
725 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
726 	switch (en->cipher_algorithm) {
727 	case CRYPTO_AES_NIST_GCM_16:
728 		/*
729 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
730 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
731 		 */
732 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
733 			tls->params.tls_hlen += sizeof(uint64_t);
734 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
735 		tls->params.tls_bs = 1;
736 		break;
737 	case CRYPTO_AES_CBC:
738 		switch (en->auth_algorithm) {
739 		case CRYPTO_SHA1_HMAC:
740 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
741 				/* Implicit IV, no nonce. */
742 				tls->sequential_records = true;
743 				tls->next_seqno = be64dec(en->rec_seq);
744 				STAILQ_INIT(&tls->pending_records);
745 			} else {
746 				tls->params.tls_hlen += AES_BLOCK_LEN;
747 			}
748 			tls->params.tls_tlen = AES_BLOCK_LEN +
749 			    SHA1_HASH_LEN;
750 			break;
751 		case CRYPTO_SHA2_256_HMAC:
752 			tls->params.tls_hlen += AES_BLOCK_LEN;
753 			tls->params.tls_tlen = AES_BLOCK_LEN +
754 			    SHA2_256_HASH_LEN;
755 			break;
756 		case CRYPTO_SHA2_384_HMAC:
757 			tls->params.tls_hlen += AES_BLOCK_LEN;
758 			tls->params.tls_tlen = AES_BLOCK_LEN +
759 			    SHA2_384_HASH_LEN;
760 			break;
761 		default:
762 			panic("invalid hmac");
763 		}
764 		tls->params.tls_bs = AES_BLOCK_LEN;
765 		break;
766 	case CRYPTO_CHACHA20_POLY1305:
767 		/*
768 		 * Chacha20 uses a 12 byte implicit IV.
769 		 */
770 		tls->params.tls_tlen = POLY1305_HASH_LEN;
771 		tls->params.tls_bs = 1;
772 		break;
773 	default:
774 		panic("invalid cipher");
775 	}
776 
777 	/*
778 	 * TLS 1.3 includes optional padding which we do not support,
779 	 * and also puts the "real" record type at the end of the
780 	 * encrypted data.
781 	 */
782 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
783 		tls->params.tls_tlen += sizeof(uint8_t);
784 
785 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
786 	    ("TLS header length too long: %d", tls->params.tls_hlen));
787 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
788 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
789 
790 	if (en->auth_key_len != 0) {
791 		tls->params.auth_key_len = en->auth_key_len;
792 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
793 		    M_WAITOK);
794 		bcopy(en->auth_key, tls->params.auth_key, en->auth_key_len);
795 	}
796 
797 	tls->params.cipher_key_len = en->cipher_key_len;
798 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
799 	bcopy(en->cipher_key, tls->params.cipher_key, en->cipher_key_len);
800 
801 	/*
802 	 * This holds the implicit portion of the nonce for AEAD
803 	 * ciphers and the initial implicit IV for TLS 1.0.  The
804 	 * explicit portions of the IV are generated in ktls_frame().
805 	 */
806 	if (en->iv_len != 0) {
807 		tls->params.iv_len = en->iv_len;
808 		bcopy(en->iv, tls->params.iv, en->iv_len);
809 
810 		/*
811 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
812 		 * counter to generate unique explicit IVs.
813 		 *
814 		 * Store this counter in the last 8 bytes of the IV
815 		 * array so that it is 8-byte aligned.
816 		 */
817 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
818 		    en->tls_vminor == TLS_MINOR_VER_TWO)
819 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
820 	}
821 
822 	*tlsp = tls;
823 	return (0);
824 }
825 
826 static struct ktls_session *
827 ktls_clone_session(struct ktls_session *tls, int direction)
828 {
829 	struct ktls_session *tls_new;
830 
831 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
832 
833 	counter_u64_add(ktls_offload_active, 1);
834 
835 	refcount_init(&tls_new->refcount, 1);
836 	if (direction == KTLS_RX) {
837 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
838 		    tls_new);
839 	} else {
840 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
841 		    tls_new);
842 		tls_new->inp = tls->inp;
843 		tls_new->tx = true;
844 		in_pcbref(tls_new->inp);
845 	}
846 
847 	/* Copy fields from existing session. */
848 	tls_new->params = tls->params;
849 	tls_new->wq_index = tls->wq_index;
850 
851 	/* Deep copy keys. */
852 	if (tls_new->params.auth_key != NULL) {
853 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
854 		    M_KTLS, M_WAITOK);
855 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
856 		    tls->params.auth_key_len);
857 	}
858 
859 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
860 	    M_WAITOK);
861 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
862 	    tls->params.cipher_key_len);
863 
864 	return (tls_new);
865 }
866 
867 #ifdef TCP_OFFLOAD
868 static int
869 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
870 {
871 	struct inpcb *inp;
872 	struct tcpcb *tp;
873 	int error;
874 
875 	inp = so->so_pcb;
876 	INP_WLOCK(inp);
877 	if (inp->inp_flags & INP_DROPPED) {
878 		INP_WUNLOCK(inp);
879 		return (ECONNRESET);
880 	}
881 	if (inp->inp_socket == NULL) {
882 		INP_WUNLOCK(inp);
883 		return (ECONNRESET);
884 	}
885 	tp = intotcpcb(inp);
886 	if (!(tp->t_flags & TF_TOE)) {
887 		INP_WUNLOCK(inp);
888 		return (EOPNOTSUPP);
889 	}
890 
891 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
892 	INP_WUNLOCK(inp);
893 	if (error == 0) {
894 		tls->mode = TCP_TLS_MODE_TOE;
895 		switch (tls->params.cipher_algorithm) {
896 		case CRYPTO_AES_CBC:
897 			counter_u64_add(ktls_toe_cbc, 1);
898 			break;
899 		case CRYPTO_AES_NIST_GCM_16:
900 			counter_u64_add(ktls_toe_gcm, 1);
901 			break;
902 		case CRYPTO_CHACHA20_POLY1305:
903 			counter_u64_add(ktls_toe_chacha20, 1);
904 			break;
905 		}
906 	}
907 	return (error);
908 }
909 #endif
910 
911 /*
912  * Common code used when first enabling ifnet TLS on a connection or
913  * when allocating a new ifnet TLS session due to a routing change.
914  * This function allocates a new TLS send tag on whatever interface
915  * the connection is currently routed over.
916  */
917 static int
918 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
919     struct m_snd_tag **mstp)
920 {
921 	union if_snd_tag_alloc_params params;
922 	struct ifnet *ifp;
923 	struct nhop_object *nh;
924 	struct tcpcb *tp;
925 	int error;
926 
927 	INP_RLOCK(inp);
928 	if (inp->inp_flags & INP_DROPPED) {
929 		INP_RUNLOCK(inp);
930 		return (ECONNRESET);
931 	}
932 	if (inp->inp_socket == NULL) {
933 		INP_RUNLOCK(inp);
934 		return (ECONNRESET);
935 	}
936 	tp = intotcpcb(inp);
937 
938 	/*
939 	 * Check administrative controls on ifnet TLS to determine if
940 	 * ifnet TLS should be denied.
941 	 *
942 	 * - Always permit 'force' requests.
943 	 * - ktls_ifnet_permitted == 0: always deny.
944 	 */
945 	if (!force && ktls_ifnet_permitted == 0) {
946 		INP_RUNLOCK(inp);
947 		return (ENXIO);
948 	}
949 
950 	/*
951 	 * XXX: Use the cached route in the inpcb to find the
952 	 * interface.  This should perhaps instead use
953 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
954 	 * enabled after a connection has completed key negotiation in
955 	 * userland, the cached route will be present in practice.
956 	 */
957 	nh = inp->inp_route.ro_nh;
958 	if (nh == NULL) {
959 		INP_RUNLOCK(inp);
960 		return (ENXIO);
961 	}
962 	ifp = nh->nh_ifp;
963 	if_ref(ifp);
964 
965 	/*
966 	 * Allocate a TLS + ratelimit tag if the connection has an
967 	 * existing pacing rate.
968 	 */
969 	if (tp->t_pacing_rate != -1 &&
970 	    (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
971 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
972 		params.tls_rate_limit.inp = inp;
973 		params.tls_rate_limit.tls = tls;
974 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
975 	} else {
976 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
977 		params.tls.inp = inp;
978 		params.tls.tls = tls;
979 	}
980 	params.hdr.flowid = inp->inp_flowid;
981 	params.hdr.flowtype = inp->inp_flowtype;
982 	params.hdr.numa_domain = inp->inp_numa_domain;
983 	INP_RUNLOCK(inp);
984 
985 	if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
986 		error = EOPNOTSUPP;
987 		goto out;
988 	}
989 	if (inp->inp_vflag & INP_IPV6) {
990 		if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
991 			error = EOPNOTSUPP;
992 			goto out;
993 		}
994 	} else {
995 		if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
996 			error = EOPNOTSUPP;
997 			goto out;
998 		}
999 	}
1000 	error = m_snd_tag_alloc(ifp, &params, mstp);
1001 out:
1002 	if_rele(ifp);
1003 	return (error);
1004 }
1005 
1006 /*
1007  * Allocate an initial TLS receive tag for doing HW decryption of TLS
1008  * data.
1009  *
1010  * This function allocates a new TLS receive tag on whatever interface
1011  * the connection is currently routed over.  If the connection ends up
1012  * using a different interface for receive this will get fixed up via
1013  * ktls_input_ifp_mismatch as future packets arrive.
1014  */
1015 static int
1016 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1017     struct m_snd_tag **mstp)
1018 {
1019 	union if_snd_tag_alloc_params params;
1020 	struct ifnet *ifp;
1021 	struct nhop_object *nh;
1022 	int error;
1023 
1024 	if (!ktls_ocf_recrypt_supported(tls))
1025 		return (ENXIO);
1026 
1027 	INP_RLOCK(inp);
1028 	if (inp->inp_flags & INP_DROPPED) {
1029 		INP_RUNLOCK(inp);
1030 		return (ECONNRESET);
1031 	}
1032 	if (inp->inp_socket == NULL) {
1033 		INP_RUNLOCK(inp);
1034 		return (ECONNRESET);
1035 	}
1036 
1037 	/*
1038 	 * Check administrative controls on ifnet TLS to determine if
1039 	 * ifnet TLS should be denied.
1040 	 */
1041 	if (ktls_ifnet_permitted == 0) {
1042 		INP_RUNLOCK(inp);
1043 		return (ENXIO);
1044 	}
1045 
1046 	/*
1047 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1048 	 * the inpcb to find the interface.
1049 	 */
1050 	nh = inp->inp_route.ro_nh;
1051 	if (nh == NULL) {
1052 		INP_RUNLOCK(inp);
1053 		return (ENXIO);
1054 	}
1055 	ifp = nh->nh_ifp;
1056 	if_ref(ifp);
1057 	tls->rx_ifp = ifp;
1058 
1059 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1060 	params.hdr.flowid = inp->inp_flowid;
1061 	params.hdr.flowtype = inp->inp_flowtype;
1062 	params.hdr.numa_domain = inp->inp_numa_domain;
1063 	params.tls_rx.inp = inp;
1064 	params.tls_rx.tls = tls;
1065 	params.tls_rx.vlan_id = 0;
1066 
1067 	INP_RUNLOCK(inp);
1068 
1069 	if (inp->inp_vflag & INP_IPV6) {
1070 		if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
1071 			error = EOPNOTSUPP;
1072 			goto out;
1073 		}
1074 	} else {
1075 		if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
1076 			error = EOPNOTSUPP;
1077 			goto out;
1078 		}
1079 	}
1080 	error = m_snd_tag_alloc(ifp, &params, mstp);
1081 
1082 	/*
1083 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1084 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1085 	 * ID for use in ktls_reset_receive_tag which allocates new
1086 	 * receive tags directly from the leaf interface bypassing
1087 	 * if_vlan.
1088 	 */
1089 	if (error == 0)
1090 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1091 out:
1092 	return (error);
1093 }
1094 
1095 static int
1096 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1097     bool force)
1098 {
1099 	struct m_snd_tag *mst;
1100 	int error;
1101 
1102 	switch (direction) {
1103 	case KTLS_TX:
1104 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1105 		if (__predict_false(error != 0))
1106 			goto done;
1107 		break;
1108 	case KTLS_RX:
1109 		KASSERT(!force, ("%s: forced receive tag", __func__));
1110 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1111 		if (__predict_false(error != 0))
1112 			goto done;
1113 		break;
1114 	default:
1115 		__assert_unreachable();
1116 	}
1117 
1118 	tls->mode = TCP_TLS_MODE_IFNET;
1119 	tls->snd_tag = mst;
1120 
1121 	switch (tls->params.cipher_algorithm) {
1122 	case CRYPTO_AES_CBC:
1123 		counter_u64_add(ktls_ifnet_cbc, 1);
1124 		break;
1125 	case CRYPTO_AES_NIST_GCM_16:
1126 		counter_u64_add(ktls_ifnet_gcm, 1);
1127 		break;
1128 	case CRYPTO_CHACHA20_POLY1305:
1129 		counter_u64_add(ktls_ifnet_chacha20, 1);
1130 		break;
1131 	default:
1132 		break;
1133 	}
1134 done:
1135 	return (error);
1136 }
1137 
1138 static void
1139 ktls_use_sw(struct ktls_session *tls)
1140 {
1141 	tls->mode = TCP_TLS_MODE_SW;
1142 	switch (tls->params.cipher_algorithm) {
1143 	case CRYPTO_AES_CBC:
1144 		counter_u64_add(ktls_sw_cbc, 1);
1145 		break;
1146 	case CRYPTO_AES_NIST_GCM_16:
1147 		counter_u64_add(ktls_sw_gcm, 1);
1148 		break;
1149 	case CRYPTO_CHACHA20_POLY1305:
1150 		counter_u64_add(ktls_sw_chacha20, 1);
1151 		break;
1152 	}
1153 }
1154 
1155 static int
1156 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1157 {
1158 	int error;
1159 
1160 	error = ktls_ocf_try(so, tls, direction);
1161 	if (error)
1162 		return (error);
1163 	ktls_use_sw(tls);
1164 	return (0);
1165 }
1166 
1167 /*
1168  * KTLS RX stores data in the socket buffer as a list of TLS records,
1169  * where each record is stored as a control message containg the TLS
1170  * header followed by data mbufs containing the decrypted data.  This
1171  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1172  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1173  * should be queued to the socket buffer as records, but encrypted
1174  * data which needs to be decrypted by software arrives as a stream of
1175  * regular mbufs which need to be converted.  In addition, there may
1176  * already be pending encrypted data in the socket buffer when KTLS RX
1177  * is enabled.
1178  *
1179  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1180  * is used:
1181  *
1182  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1183  *
1184  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1185  *   from the first mbuf.  Once all of the data for that TLS record is
1186  *   queued, the socket is queued to a worker thread.
1187  *
1188  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1189  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1190  *   decrypted, and inserted into the regular socket buffer chain as
1191  *   record starting with a control message holding the TLS header and
1192  *   a chain of mbufs holding the encrypted data.
1193  */
1194 
1195 static void
1196 sb_mark_notready(struct sockbuf *sb)
1197 {
1198 	struct mbuf *m;
1199 
1200 	m = sb->sb_mb;
1201 	sb->sb_mtls = m;
1202 	sb->sb_mb = NULL;
1203 	sb->sb_mbtail = NULL;
1204 	sb->sb_lastrecord = NULL;
1205 	for (; m != NULL; m = m->m_next) {
1206 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1207 		    __func__));
1208 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1209 		    __func__));
1210 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1211 		    __func__));
1212 		m->m_flags |= M_NOTREADY;
1213 		sb->sb_acc -= m->m_len;
1214 		sb->sb_tlscc += m->m_len;
1215 		sb->sb_mtlstail = m;
1216 	}
1217 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1218 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1219 	    sb->sb_ccc));
1220 }
1221 
1222 /*
1223  * Return information about the pending TLS data in a socket
1224  * buffer.  On return, 'seqno' is set to the sequence number
1225  * of the next TLS record to be received, 'resid' is set to
1226  * the amount of bytes still needed for the last pending
1227  * record.  The function returns 'false' if the last pending
1228  * record contains a partial TLS header.  In that case, 'resid'
1229  * is the number of bytes needed to complete the TLS header.
1230  */
1231 bool
1232 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1233 {
1234 	struct tls_record_layer hdr;
1235 	struct mbuf *m;
1236 	uint64_t seqno;
1237 	size_t resid;
1238 	u_int offset, record_len;
1239 
1240 	SOCKBUF_LOCK_ASSERT(sb);
1241 	MPASS(sb->sb_flags & SB_TLS_RX);
1242 	seqno = sb->sb_tls_seqno;
1243 	resid = sb->sb_tlscc;
1244 	m = sb->sb_mtls;
1245 	offset = 0;
1246 
1247 	if (resid == 0) {
1248 		*seqnop = seqno;
1249 		*residp = 0;
1250 		return (true);
1251 	}
1252 
1253 	for (;;) {
1254 		seqno++;
1255 
1256 		if (resid < sizeof(hdr)) {
1257 			*seqnop = seqno;
1258 			*residp = sizeof(hdr) - resid;
1259 			return (false);
1260 		}
1261 
1262 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1263 
1264 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1265 		if (resid <= record_len) {
1266 			*seqnop = seqno;
1267 			*residp = record_len - resid;
1268 			return (true);
1269 		}
1270 		resid -= record_len;
1271 
1272 		while (record_len != 0) {
1273 			if (m->m_len - offset > record_len) {
1274 				offset += record_len;
1275 				break;
1276 			}
1277 
1278 			record_len -= (m->m_len - offset);
1279 			offset = 0;
1280 			m = m->m_next;
1281 		}
1282 	}
1283 }
1284 
1285 int
1286 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1287 {
1288 	struct ktls_session *tls;
1289 	int error;
1290 
1291 	if (!ktls_offload_enable)
1292 		return (ENOTSUP);
1293 
1294 	counter_u64_add(ktls_offload_enable_calls, 1);
1295 
1296 	/*
1297 	 * This should always be true since only the TCP socket option
1298 	 * invokes this function.
1299 	 */
1300 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1301 		return (EINVAL);
1302 
1303 	/*
1304 	 * XXX: Don't overwrite existing sessions.  We should permit
1305 	 * this to support rekeying in the future.
1306 	 */
1307 	if (so->so_rcv.sb_tls_info != NULL)
1308 		return (EALREADY);
1309 
1310 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1311 		return (ENOTSUP);
1312 
1313 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1314 	if (error)
1315 		return (error);
1316 
1317 	error = ktls_ocf_try(so, tls, KTLS_RX);
1318 	if (error) {
1319 		ktls_free(tls);
1320 		return (error);
1321 	}
1322 
1323 	/* Mark the socket as using TLS offload. */
1324 	SOCK_RECVBUF_LOCK(so);
1325 	if (SOLISTENING(so)) {
1326 		SOCK_RECVBUF_UNLOCK(so);
1327 		ktls_free(tls);
1328 		return (EINVAL);
1329 	}
1330 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1331 	so->so_rcv.sb_tls_info = tls;
1332 	so->so_rcv.sb_flags |= SB_TLS_RX;
1333 
1334 	/* Mark existing data as not ready until it can be decrypted. */
1335 	sb_mark_notready(&so->so_rcv);
1336 	ktls_check_rx(&so->so_rcv);
1337 	SOCK_RECVBUF_UNLOCK(so);
1338 
1339 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1340 #ifdef TCP_OFFLOAD
1341 	error = ktls_try_toe(so, tls, KTLS_RX);
1342 	if (error)
1343 #endif
1344 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1345 	if (error)
1346 		ktls_use_sw(tls);
1347 
1348 	counter_u64_add(ktls_offload_total, 1);
1349 
1350 	return (0);
1351 }
1352 
1353 int
1354 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1355 {
1356 	struct ktls_session *tls;
1357 	struct inpcb *inp;
1358 	struct tcpcb *tp;
1359 	int error;
1360 
1361 	if (!ktls_offload_enable)
1362 		return (ENOTSUP);
1363 
1364 	counter_u64_add(ktls_offload_enable_calls, 1);
1365 
1366 	/*
1367 	 * This should always be true since only the TCP socket option
1368 	 * invokes this function.
1369 	 */
1370 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1371 		return (EINVAL);
1372 
1373 	/*
1374 	 * XXX: Don't overwrite existing sessions.  We should permit
1375 	 * this to support rekeying in the future.
1376 	 */
1377 	if (so->so_snd.sb_tls_info != NULL)
1378 		return (EALREADY);
1379 
1380 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1381 		return (ENOTSUP);
1382 
1383 	/* TLS requires ext pgs */
1384 	if (mb_use_ext_pgs == 0)
1385 		return (ENXIO);
1386 
1387 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1388 	if (error)
1389 		return (error);
1390 
1391 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1392 #ifdef TCP_OFFLOAD
1393 	error = ktls_try_toe(so, tls, KTLS_TX);
1394 	if (error)
1395 #endif
1396 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1397 	if (error)
1398 		error = ktls_try_sw(so, tls, KTLS_TX);
1399 
1400 	if (error) {
1401 		ktls_free(tls);
1402 		return (error);
1403 	}
1404 
1405 	/*
1406 	 * Serialize with sosend_generic() and make sure that we're not
1407 	 * operating on a listening socket.
1408 	 */
1409 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1410 	if (error) {
1411 		ktls_free(tls);
1412 		return (error);
1413 	}
1414 
1415 	/*
1416 	 * Write lock the INP when setting sb_tls_info so that
1417 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1418 	 * holding the INP lock.
1419 	 */
1420 	inp = so->so_pcb;
1421 	INP_WLOCK(inp);
1422 	SOCK_SENDBUF_LOCK(so);
1423 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1424 	so->so_snd.sb_tls_info = tls;
1425 	if (tls->mode != TCP_TLS_MODE_SW) {
1426 		tp = intotcpcb(inp);
1427 		MPASS(tp->t_nic_ktls_xmit == 0);
1428 		tp->t_nic_ktls_xmit = 1;
1429 		if (tp->t_fb->tfb_hwtls_change != NULL)
1430 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1431 	}
1432 	SOCK_SENDBUF_UNLOCK(so);
1433 	INP_WUNLOCK(inp);
1434 	SOCK_IO_SEND_UNLOCK(so);
1435 
1436 	counter_u64_add(ktls_offload_total, 1);
1437 
1438 	return (0);
1439 }
1440 
1441 int
1442 ktls_get_rx_mode(struct socket *so, int *modep)
1443 {
1444 	struct ktls_session *tls;
1445 	struct inpcb *inp __diagused;
1446 
1447 	if (SOLISTENING(so))
1448 		return (EINVAL);
1449 	inp = so->so_pcb;
1450 	INP_WLOCK_ASSERT(inp);
1451 	SOCK_RECVBUF_LOCK(so);
1452 	tls = so->so_rcv.sb_tls_info;
1453 	if (tls == NULL)
1454 		*modep = TCP_TLS_MODE_NONE;
1455 	else
1456 		*modep = tls->mode;
1457 	SOCK_RECVBUF_UNLOCK(so);
1458 	return (0);
1459 }
1460 
1461 /*
1462  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1463  *
1464  * This function gets information about the next TCP- and TLS-
1465  * sequence number to be processed by the TLS receive worker
1466  * thread. The information is extracted from the given "inpcb"
1467  * structure. The values are stored in host endian format at the two
1468  * given output pointer locations. The TCP sequence number points to
1469  * the beginning of the TLS header.
1470  *
1471  * This function returns zero on success, else a non-zero error code
1472  * is returned.
1473  */
1474 int
1475 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1476 {
1477 	struct socket *so;
1478 	struct tcpcb *tp;
1479 
1480 	INP_RLOCK(inp);
1481 	so = inp->inp_socket;
1482 	if (__predict_false(so == NULL)) {
1483 		INP_RUNLOCK(inp);
1484 		return (EINVAL);
1485 	}
1486 	if (inp->inp_flags & INP_DROPPED) {
1487 		INP_RUNLOCK(inp);
1488 		return (ECONNRESET);
1489 	}
1490 
1491 	tp = intotcpcb(inp);
1492 	MPASS(tp != NULL);
1493 
1494 	SOCKBUF_LOCK(&so->so_rcv);
1495 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1496 	*tlsseq = so->so_rcv.sb_tls_seqno;
1497 	SOCKBUF_UNLOCK(&so->so_rcv);
1498 
1499 	INP_RUNLOCK(inp);
1500 
1501 	return (0);
1502 }
1503 
1504 int
1505 ktls_get_tx_mode(struct socket *so, int *modep)
1506 {
1507 	struct ktls_session *tls;
1508 	struct inpcb *inp __diagused;
1509 
1510 	if (SOLISTENING(so))
1511 		return (EINVAL);
1512 	inp = so->so_pcb;
1513 	INP_WLOCK_ASSERT(inp);
1514 	SOCK_SENDBUF_LOCK(so);
1515 	tls = so->so_snd.sb_tls_info;
1516 	if (tls == NULL)
1517 		*modep = TCP_TLS_MODE_NONE;
1518 	else
1519 		*modep = tls->mode;
1520 	SOCK_SENDBUF_UNLOCK(so);
1521 	return (0);
1522 }
1523 
1524 /*
1525  * Switch between SW and ifnet TLS sessions as requested.
1526  */
1527 int
1528 ktls_set_tx_mode(struct socket *so, int mode)
1529 {
1530 	struct ktls_session *tls, *tls_new;
1531 	struct inpcb *inp;
1532 	struct tcpcb *tp;
1533 	int error;
1534 
1535 	if (SOLISTENING(so))
1536 		return (EINVAL);
1537 	switch (mode) {
1538 	case TCP_TLS_MODE_SW:
1539 	case TCP_TLS_MODE_IFNET:
1540 		break;
1541 	default:
1542 		return (EINVAL);
1543 	}
1544 
1545 	inp = so->so_pcb;
1546 	INP_WLOCK_ASSERT(inp);
1547 	tp = intotcpcb(inp);
1548 
1549 	if (mode == TCP_TLS_MODE_IFNET) {
1550 		/* Don't allow enabling ifnet ktls multiple times */
1551 		if (tp->t_nic_ktls_xmit)
1552 			return (EALREADY);
1553 
1554 		/*
1555 		 * Don't enable ifnet ktls if we disabled it due to an
1556 		 * excessive retransmission rate
1557 		 */
1558 		if (tp->t_nic_ktls_xmit_dis)
1559 			return (ENXIO);
1560 	}
1561 
1562 	SOCKBUF_LOCK(&so->so_snd);
1563 	tls = so->so_snd.sb_tls_info;
1564 	if (tls == NULL) {
1565 		SOCKBUF_UNLOCK(&so->so_snd);
1566 		return (0);
1567 	}
1568 
1569 	if (tls->mode == mode) {
1570 		SOCKBUF_UNLOCK(&so->so_snd);
1571 		return (0);
1572 	}
1573 
1574 	tls = ktls_hold(tls);
1575 	SOCKBUF_UNLOCK(&so->so_snd);
1576 	INP_WUNLOCK(inp);
1577 
1578 	tls_new = ktls_clone_session(tls, KTLS_TX);
1579 
1580 	if (mode == TCP_TLS_MODE_IFNET)
1581 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1582 	else
1583 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1584 	if (error) {
1585 		counter_u64_add(ktls_switch_failed, 1);
1586 		ktls_free(tls_new);
1587 		ktls_free(tls);
1588 		INP_WLOCK(inp);
1589 		return (error);
1590 	}
1591 
1592 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1593 	if (error) {
1594 		counter_u64_add(ktls_switch_failed, 1);
1595 		ktls_free(tls_new);
1596 		ktls_free(tls);
1597 		INP_WLOCK(inp);
1598 		return (error);
1599 	}
1600 
1601 	/*
1602 	 * If we raced with another session change, keep the existing
1603 	 * session.
1604 	 */
1605 	if (tls != so->so_snd.sb_tls_info) {
1606 		counter_u64_add(ktls_switch_failed, 1);
1607 		SOCK_IO_SEND_UNLOCK(so);
1608 		ktls_free(tls_new);
1609 		ktls_free(tls);
1610 		INP_WLOCK(inp);
1611 		return (EBUSY);
1612 	}
1613 
1614 	INP_WLOCK(inp);
1615 	SOCKBUF_LOCK(&so->so_snd);
1616 	so->so_snd.sb_tls_info = tls_new;
1617 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1618 		MPASS(tp->t_nic_ktls_xmit == 0);
1619 		tp->t_nic_ktls_xmit = 1;
1620 		if (tp->t_fb->tfb_hwtls_change != NULL)
1621 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1622 	}
1623 	SOCKBUF_UNLOCK(&so->so_snd);
1624 	SOCK_IO_SEND_UNLOCK(so);
1625 
1626 	/*
1627 	 * Drop two references on 'tls'.  The first is for the
1628 	 * ktls_hold() above.  The second drops the reference from the
1629 	 * socket buffer.
1630 	 */
1631 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1632 	ktls_free(tls);
1633 	ktls_free(tls);
1634 
1635 	if (mode == TCP_TLS_MODE_IFNET)
1636 		counter_u64_add(ktls_switch_to_ifnet, 1);
1637 	else
1638 		counter_u64_add(ktls_switch_to_sw, 1);
1639 
1640 	return (0);
1641 }
1642 
1643 /*
1644  * Try to allocate a new TLS receive tag.  This task is scheduled when
1645  * sbappend_ktls_rx detects an input path change.  If a new tag is
1646  * allocated, replace the tag in the TLS session.  If a new tag cannot
1647  * be allocated, let the session fall back to software decryption.
1648  */
1649 static void
1650 ktls_reset_receive_tag(void *context, int pending)
1651 {
1652 	union if_snd_tag_alloc_params params;
1653 	struct ktls_session *tls;
1654 	struct m_snd_tag *mst;
1655 	struct inpcb *inp;
1656 	struct ifnet *ifp;
1657 	struct socket *so;
1658 	int error;
1659 
1660 	MPASS(pending == 1);
1661 
1662 	tls = context;
1663 	so = tls->so;
1664 	inp = so->so_pcb;
1665 	ifp = NULL;
1666 
1667 	INP_RLOCK(inp);
1668 	if (inp->inp_flags & INP_DROPPED) {
1669 		INP_RUNLOCK(inp);
1670 		goto out;
1671 	}
1672 
1673 	SOCKBUF_LOCK(&so->so_rcv);
1674 	mst = tls->snd_tag;
1675 	tls->snd_tag = NULL;
1676 	if (mst != NULL)
1677 		m_snd_tag_rele(mst);
1678 
1679 	ifp = tls->rx_ifp;
1680 	if_ref(ifp);
1681 	SOCKBUF_UNLOCK(&so->so_rcv);
1682 
1683 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1684 	params.hdr.flowid = inp->inp_flowid;
1685 	params.hdr.flowtype = inp->inp_flowtype;
1686 	params.hdr.numa_domain = inp->inp_numa_domain;
1687 	params.tls_rx.inp = inp;
1688 	params.tls_rx.tls = tls;
1689 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1690 	INP_RUNLOCK(inp);
1691 
1692 	if (inp->inp_vflag & INP_IPV6) {
1693 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1694 			goto out;
1695 	} else {
1696 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1697 			goto out;
1698 	}
1699 
1700 	error = m_snd_tag_alloc(ifp, &params, &mst);
1701 	if (error == 0) {
1702 		SOCKBUF_LOCK(&so->so_rcv);
1703 		tls->snd_tag = mst;
1704 		SOCKBUF_UNLOCK(&so->so_rcv);
1705 
1706 		counter_u64_add(ktls_ifnet_reset, 1);
1707 	} else {
1708 		/*
1709 		 * Just fall back to software decryption if a tag
1710 		 * cannot be allocated leaving the connection intact.
1711 		 * If a future input path change switches to another
1712 		 * interface this connection will resume ifnet TLS.
1713 		 */
1714 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1715 	}
1716 
1717 out:
1718 	mtx_pool_lock(mtxpool_sleep, tls);
1719 	tls->reset_pending = false;
1720 	mtx_pool_unlock(mtxpool_sleep, tls);
1721 
1722 	if (ifp != NULL)
1723 		if_rele(ifp);
1724 	CURVNET_SET(so->so_vnet);
1725 	sorele(so);
1726 	CURVNET_RESTORE();
1727 	ktls_free(tls);
1728 }
1729 
1730 /*
1731  * Try to allocate a new TLS send tag.  This task is scheduled when
1732  * ip_output detects a route change while trying to transmit a packet
1733  * holding a TLS record.  If a new tag is allocated, replace the tag
1734  * in the TLS session.  Subsequent packets on the connection will use
1735  * the new tag.  If a new tag cannot be allocated, drop the
1736  * connection.
1737  */
1738 static void
1739 ktls_reset_send_tag(void *context, int pending)
1740 {
1741 	struct epoch_tracker et;
1742 	struct ktls_session *tls;
1743 	struct m_snd_tag *old, *new;
1744 	struct inpcb *inp;
1745 	struct tcpcb *tp;
1746 	int error;
1747 
1748 	MPASS(pending == 1);
1749 
1750 	tls = context;
1751 	inp = tls->inp;
1752 
1753 	/*
1754 	 * Free the old tag first before allocating a new one.
1755 	 * ip[6]_output_send() will treat a NULL send tag the same as
1756 	 * an ifp mismatch and drop packets until a new tag is
1757 	 * allocated.
1758 	 *
1759 	 * Write-lock the INP when changing tls->snd_tag since
1760 	 * ip[6]_output_send() holds a read-lock when reading the
1761 	 * pointer.
1762 	 */
1763 	INP_WLOCK(inp);
1764 	old = tls->snd_tag;
1765 	tls->snd_tag = NULL;
1766 	INP_WUNLOCK(inp);
1767 	if (old != NULL)
1768 		m_snd_tag_rele(old);
1769 
1770 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1771 
1772 	if (error == 0) {
1773 		INP_WLOCK(inp);
1774 		tls->snd_tag = new;
1775 		mtx_pool_lock(mtxpool_sleep, tls);
1776 		tls->reset_pending = false;
1777 		mtx_pool_unlock(mtxpool_sleep, tls);
1778 		INP_WUNLOCK(inp);
1779 
1780 		counter_u64_add(ktls_ifnet_reset, 1);
1781 
1782 		/*
1783 		 * XXX: Should we kick tcp_output explicitly now that
1784 		 * the send tag is fixed or just rely on timers?
1785 		 */
1786 	} else {
1787 		NET_EPOCH_ENTER(et);
1788 		INP_WLOCK(inp);
1789 		if (!(inp->inp_flags & INP_DROPPED)) {
1790 			tp = intotcpcb(inp);
1791 			CURVNET_SET(inp->inp_vnet);
1792 			tp = tcp_drop(tp, ECONNABORTED);
1793 			CURVNET_RESTORE();
1794 			if (tp != NULL) {
1795 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1796 				INP_WUNLOCK(inp);
1797 			}
1798 		} else
1799 			INP_WUNLOCK(inp);
1800 		NET_EPOCH_EXIT(et);
1801 
1802 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1803 
1804 		/*
1805 		 * Leave reset_pending true to avoid future tasks while
1806 		 * the socket goes away.
1807 		 */
1808 	}
1809 
1810 	ktls_free(tls);
1811 }
1812 
1813 void
1814 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1815 {
1816 	struct ktls_session *tls;
1817 	struct socket *so;
1818 
1819 	SOCKBUF_LOCK_ASSERT(sb);
1820 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1821 	    __func__, sb));
1822 	so = __containerof(sb, struct socket, so_rcv);
1823 
1824 	tls = sb->sb_tls_info;
1825 	if_rele(tls->rx_ifp);
1826 	if_ref(ifp);
1827 	tls->rx_ifp = ifp;
1828 
1829 	/*
1830 	 * See if we should schedule a task to update the receive tag for
1831 	 * this session.
1832 	 */
1833 	mtx_pool_lock(mtxpool_sleep, tls);
1834 	if (!tls->reset_pending) {
1835 		(void) ktls_hold(tls);
1836 		soref(so);
1837 		tls->so = so;
1838 		tls->reset_pending = true;
1839 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1840 	}
1841 	mtx_pool_unlock(mtxpool_sleep, tls);
1842 }
1843 
1844 int
1845 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1846 {
1847 
1848 	if (inp == NULL)
1849 		return (ENOBUFS);
1850 
1851 	INP_LOCK_ASSERT(inp);
1852 
1853 	/*
1854 	 * See if we should schedule a task to update the send tag for
1855 	 * this session.
1856 	 */
1857 	mtx_pool_lock(mtxpool_sleep, tls);
1858 	if (!tls->reset_pending) {
1859 		(void) ktls_hold(tls);
1860 		tls->reset_pending = true;
1861 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1862 	}
1863 	mtx_pool_unlock(mtxpool_sleep, tls);
1864 	return (ENOBUFS);
1865 }
1866 
1867 #ifdef RATELIMIT
1868 int
1869 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1870 {
1871 	union if_snd_tag_modify_params params = {
1872 		.rate_limit.max_rate = max_pacing_rate,
1873 		.rate_limit.flags = M_NOWAIT,
1874 	};
1875 	struct m_snd_tag *mst;
1876 
1877 	/* Can't get to the inp, but it should be locked. */
1878 	/* INP_LOCK_ASSERT(inp); */
1879 
1880 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1881 
1882 	if (tls->snd_tag == NULL) {
1883 		/*
1884 		 * Resetting send tag, ignore this change.  The
1885 		 * pending reset may or may not see this updated rate
1886 		 * in the tcpcb.  If it doesn't, we will just lose
1887 		 * this rate change.
1888 		 */
1889 		return (0);
1890 	}
1891 
1892 	mst = tls->snd_tag;
1893 
1894 	MPASS(mst != NULL);
1895 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1896 
1897 	return (mst->sw->snd_tag_modify(mst, &params));
1898 }
1899 #endif
1900 
1901 static void
1902 ktls_destroy_help(void *context, int pending __unused)
1903 {
1904 	ktls_destroy(context);
1905 }
1906 
1907 void
1908 ktls_destroy(struct ktls_session *tls)
1909 {
1910 	struct inpcb *inp;
1911 	struct tcpcb *tp;
1912 	bool wlocked;
1913 
1914 	MPASS(tls->refcount == 0);
1915 
1916 	inp = tls->inp;
1917 	if (tls->tx) {
1918 		wlocked = INP_WLOCKED(inp);
1919 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1920 			/*
1921 			 * rwlocks read locks are anonymous, and there
1922 			 * is no way to know if our current thread
1923 			 * holds an rlock on the inp.  As a rough
1924 			 * estimate, check to see if the thread holds
1925 			 * *any* rlocks at all.  If it does not, then we
1926 			 * know that we don't hold the inp rlock, and
1927 			 * can safely take the wlock
1928 			 */
1929 			if (curthread->td_rw_rlocks == 0) {
1930 				INP_WLOCK(inp);
1931 			} else {
1932 				/*
1933 				 * We might hold the rlock, so let's
1934 				 * do the destroy in a taskqueue
1935 				 * context to avoid a potential
1936 				 * deadlock.  This should be very
1937 				 * rare.
1938 				 */
1939 				counter_u64_add(ktls_destroy_task, 1);
1940 				TASK_INIT(&tls->destroy_task, 0,
1941 				    ktls_destroy_help, tls);
1942 				(void)taskqueue_enqueue(taskqueue_thread,
1943 				    &tls->destroy_task);
1944 				return;
1945 			}
1946 		}
1947 	}
1948 
1949 	if (tls->sequential_records) {
1950 		struct mbuf *m, *n;
1951 		int page_count;
1952 
1953 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1954 			page_count = m->m_epg_enc_cnt;
1955 			while (page_count > 0) {
1956 				KASSERT(page_count >= m->m_epg_nrdy,
1957 				    ("%s: too few pages", __func__));
1958 				page_count -= m->m_epg_nrdy;
1959 				m = m_free(m);
1960 			}
1961 		}
1962 	}
1963 
1964 	counter_u64_add(ktls_offload_active, -1);
1965 	switch (tls->mode) {
1966 	case TCP_TLS_MODE_SW:
1967 		switch (tls->params.cipher_algorithm) {
1968 		case CRYPTO_AES_CBC:
1969 			counter_u64_add(ktls_sw_cbc, -1);
1970 			break;
1971 		case CRYPTO_AES_NIST_GCM_16:
1972 			counter_u64_add(ktls_sw_gcm, -1);
1973 			break;
1974 		case CRYPTO_CHACHA20_POLY1305:
1975 			counter_u64_add(ktls_sw_chacha20, -1);
1976 			break;
1977 		}
1978 		break;
1979 	case TCP_TLS_MODE_IFNET:
1980 		switch (tls->params.cipher_algorithm) {
1981 		case CRYPTO_AES_CBC:
1982 			counter_u64_add(ktls_ifnet_cbc, -1);
1983 			break;
1984 		case CRYPTO_AES_NIST_GCM_16:
1985 			counter_u64_add(ktls_ifnet_gcm, -1);
1986 			break;
1987 		case CRYPTO_CHACHA20_POLY1305:
1988 			counter_u64_add(ktls_ifnet_chacha20, -1);
1989 			break;
1990 		}
1991 		if (tls->snd_tag != NULL)
1992 			m_snd_tag_rele(tls->snd_tag);
1993 		if (tls->rx_ifp != NULL)
1994 			if_rele(tls->rx_ifp);
1995 		if (tls->tx) {
1996 			INP_WLOCK_ASSERT(inp);
1997 			tp = intotcpcb(inp);
1998 			MPASS(tp->t_nic_ktls_xmit == 1);
1999 			tp->t_nic_ktls_xmit = 0;
2000 		}
2001 		break;
2002 #ifdef TCP_OFFLOAD
2003 	case TCP_TLS_MODE_TOE:
2004 		switch (tls->params.cipher_algorithm) {
2005 		case CRYPTO_AES_CBC:
2006 			counter_u64_add(ktls_toe_cbc, -1);
2007 			break;
2008 		case CRYPTO_AES_NIST_GCM_16:
2009 			counter_u64_add(ktls_toe_gcm, -1);
2010 			break;
2011 		case CRYPTO_CHACHA20_POLY1305:
2012 			counter_u64_add(ktls_toe_chacha20, -1);
2013 			break;
2014 		}
2015 		break;
2016 #endif
2017 	}
2018 	if (tls->ocf_session != NULL)
2019 		ktls_ocf_free(tls);
2020 	if (tls->params.auth_key != NULL) {
2021 		zfree(tls->params.auth_key, M_KTLS);
2022 		tls->params.auth_key = NULL;
2023 		tls->params.auth_key_len = 0;
2024 	}
2025 	if (tls->params.cipher_key != NULL) {
2026 		zfree(tls->params.cipher_key, M_KTLS);
2027 		tls->params.cipher_key = NULL;
2028 		tls->params.cipher_key_len = 0;
2029 	}
2030 	if (tls->tx) {
2031 		INP_WLOCK_ASSERT(inp);
2032 		if (!in_pcbrele_wlocked(inp) && !wlocked)
2033 			INP_WUNLOCK(inp);
2034 	}
2035 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
2036 
2037 	uma_zfree(ktls_session_zone, tls);
2038 }
2039 
2040 void
2041 ktls_seq(struct sockbuf *sb, struct mbuf *m)
2042 {
2043 
2044 	for (; m != NULL; m = m->m_next) {
2045 		KASSERT((m->m_flags & M_EXTPG) != 0,
2046 		    ("ktls_seq: mapped mbuf %p", m));
2047 
2048 		m->m_epg_seqno = sb->sb_tls_seqno;
2049 		sb->sb_tls_seqno++;
2050 	}
2051 }
2052 
2053 /*
2054  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
2055  * mbuf in the chain must be an unmapped mbuf.  The payload of the
2056  * mbuf must be populated with the payload of each TLS record.
2057  *
2058  * The record_type argument specifies the TLS record type used when
2059  * populating the TLS header.
2060  *
2061  * The enq_count argument on return is set to the number of pages of
2062  * payload data for this entire chain that need to be encrypted via SW
2063  * encryption.  The returned value should be passed to ktls_enqueue
2064  * when scheduling encryption of this chain of mbufs.  To handle the
2065  * special case of empty fragments for TLS 1.0 sessions, an empty
2066  * fragment counts as one page.
2067  */
2068 void
2069 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2070     uint8_t record_type)
2071 {
2072 	struct tls_record_layer *tlshdr;
2073 	struct mbuf *m;
2074 	uint64_t *noncep;
2075 	uint16_t tls_len;
2076 	int maxlen __diagused;
2077 
2078 	maxlen = tls->params.max_frame_len;
2079 	*enq_cnt = 0;
2080 	for (m = top; m != NULL; m = m->m_next) {
2081 		/*
2082 		 * All mbufs in the chain should be TLS records whose
2083 		 * payload does not exceed the maximum frame length.
2084 		 *
2085 		 * Empty TLS 1.0 records are permitted when using CBC.
2086 		 */
2087 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2088 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2089 		    ("ktls_frame: m %p len %d", m, m->m_len));
2090 
2091 		/*
2092 		 * TLS frames require unmapped mbufs to store session
2093 		 * info.
2094 		 */
2095 		KASSERT((m->m_flags & M_EXTPG) != 0,
2096 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2097 
2098 		tls_len = m->m_len;
2099 
2100 		/* Save a reference to the session. */
2101 		m->m_epg_tls = ktls_hold(tls);
2102 
2103 		m->m_epg_hdrlen = tls->params.tls_hlen;
2104 		m->m_epg_trllen = tls->params.tls_tlen;
2105 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2106 			int bs, delta;
2107 
2108 			/*
2109 			 * AES-CBC pads messages to a multiple of the
2110 			 * block size.  Note that the padding is
2111 			 * applied after the digest and the encryption
2112 			 * is done on the "plaintext || mac || padding".
2113 			 * At least one byte of padding is always
2114 			 * present.
2115 			 *
2116 			 * Compute the final trailer length assuming
2117 			 * at most one block of padding.
2118 			 * tls->params.tls_tlen is the maximum
2119 			 * possible trailer length (padding + digest).
2120 			 * delta holds the number of excess padding
2121 			 * bytes if the maximum were used.  Those
2122 			 * extra bytes are removed.
2123 			 */
2124 			bs = tls->params.tls_bs;
2125 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2126 			m->m_epg_trllen -= delta;
2127 		}
2128 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2129 
2130 		/* Populate the TLS header. */
2131 		tlshdr = (void *)m->m_epg_hdr;
2132 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2133 
2134 		/*
2135 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2136 		 * of TLS_RLTYPE_APP.
2137 		 */
2138 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2139 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2140 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2141 			tlshdr->tls_type = TLS_RLTYPE_APP;
2142 			/* save the real record type for later */
2143 			m->m_epg_record_type = record_type;
2144 			m->m_epg_trail[0] = record_type;
2145 		} else {
2146 			tlshdr->tls_vminor = tls->params.tls_vminor;
2147 			tlshdr->tls_type = record_type;
2148 		}
2149 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2150 
2151 		/*
2152 		 * Store nonces / explicit IVs after the end of the
2153 		 * TLS header.
2154 		 *
2155 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2156 		 * from the end of the IV.  The nonce is then
2157 		 * incremented for use by the next record.
2158 		 *
2159 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2160 		 */
2161 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2162 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2163 			noncep = (uint64_t *)(tls->params.iv + 8);
2164 			be64enc(tlshdr + 1, *noncep);
2165 			(*noncep)++;
2166 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2167 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2168 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2169 
2170 		/*
2171 		 * When using SW encryption, mark the mbuf not ready.
2172 		 * It will be marked ready via sbready() after the
2173 		 * record has been encrypted.
2174 		 *
2175 		 * When using ifnet TLS, unencrypted TLS records are
2176 		 * sent down the stack to the NIC.
2177 		 */
2178 		if (tls->mode == TCP_TLS_MODE_SW) {
2179 			m->m_flags |= M_NOTREADY;
2180 			if (__predict_false(tls_len == 0)) {
2181 				/* TLS 1.0 empty fragment. */
2182 				m->m_epg_nrdy = 1;
2183 			} else
2184 				m->m_epg_nrdy = m->m_epg_npgs;
2185 			*enq_cnt += m->m_epg_nrdy;
2186 		}
2187 	}
2188 }
2189 
2190 bool
2191 ktls_permit_empty_frames(struct ktls_session *tls)
2192 {
2193 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2194 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2195 }
2196 
2197 void
2198 ktls_check_rx(struct sockbuf *sb)
2199 {
2200 	struct tls_record_layer hdr;
2201 	struct ktls_wq *wq;
2202 	struct socket *so;
2203 	bool running;
2204 
2205 	SOCKBUF_LOCK_ASSERT(sb);
2206 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2207 	    __func__, sb));
2208 	so = __containerof(sb, struct socket, so_rcv);
2209 
2210 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2211 		return;
2212 
2213 	/* Is there enough queued for a TLS header? */
2214 	if (sb->sb_tlscc < sizeof(hdr)) {
2215 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2216 			so->so_error = EMSGSIZE;
2217 		return;
2218 	}
2219 
2220 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2221 
2222 	/* Is the entire record queued? */
2223 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2224 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2225 			so->so_error = EMSGSIZE;
2226 		return;
2227 	}
2228 
2229 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2230 
2231 	soref(so);
2232 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2233 	mtx_lock(&wq->mtx);
2234 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2235 	running = wq->running;
2236 	mtx_unlock(&wq->mtx);
2237 	if (!running)
2238 		wakeup(wq);
2239 	counter_u64_add(ktls_cnt_rx_queued, 1);
2240 }
2241 
2242 static struct mbuf *
2243 ktls_detach_record(struct sockbuf *sb, int len)
2244 {
2245 	struct mbuf *m, *n, *top;
2246 	int remain;
2247 
2248 	SOCKBUF_LOCK_ASSERT(sb);
2249 	MPASS(len <= sb->sb_tlscc);
2250 
2251 	/*
2252 	 * If TLS chain is the exact size of the record,
2253 	 * just grab the whole record.
2254 	 */
2255 	top = sb->sb_mtls;
2256 	if (sb->sb_tlscc == len) {
2257 		sb->sb_mtls = NULL;
2258 		sb->sb_mtlstail = NULL;
2259 		goto out;
2260 	}
2261 
2262 	/*
2263 	 * While it would be nice to use m_split() here, we need
2264 	 * to know exactly what m_split() allocates to update the
2265 	 * accounting, so do it inline instead.
2266 	 */
2267 	remain = len;
2268 	for (m = top; remain > m->m_len; m = m->m_next)
2269 		remain -= m->m_len;
2270 
2271 	/* Easy case: don't have to split 'm'. */
2272 	if (remain == m->m_len) {
2273 		sb->sb_mtls = m->m_next;
2274 		if (sb->sb_mtls == NULL)
2275 			sb->sb_mtlstail = NULL;
2276 		m->m_next = NULL;
2277 		goto out;
2278 	}
2279 
2280 	/*
2281 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2282 	 * with M_NOWAIT first.
2283 	 */
2284 	n = m_get(M_NOWAIT, MT_DATA);
2285 	if (n == NULL) {
2286 		/*
2287 		 * Use M_WAITOK with socket buffer unlocked.  If
2288 		 * 'sb_mtls' changes while the lock is dropped, return
2289 		 * NULL to force the caller to retry.
2290 		 */
2291 		SOCKBUF_UNLOCK(sb);
2292 
2293 		n = m_get(M_WAITOK, MT_DATA);
2294 
2295 		SOCKBUF_LOCK(sb);
2296 		if (sb->sb_mtls != top) {
2297 			m_free(n);
2298 			return (NULL);
2299 		}
2300 	}
2301 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2302 
2303 	/* Store remainder in 'n'. */
2304 	n->m_len = m->m_len - remain;
2305 	if (m->m_flags & M_EXT) {
2306 		n->m_data = m->m_data + remain;
2307 		mb_dupcl(n, m);
2308 	} else {
2309 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2310 	}
2311 
2312 	/* Trim 'm' and update accounting. */
2313 	m->m_len -= n->m_len;
2314 	sb->sb_tlscc -= n->m_len;
2315 	sb->sb_ccc -= n->m_len;
2316 
2317 	/* Account for 'n'. */
2318 	sballoc_ktls_rx(sb, n);
2319 
2320 	/* Insert 'n' into the TLS chain. */
2321 	sb->sb_mtls = n;
2322 	n->m_next = m->m_next;
2323 	if (sb->sb_mtlstail == m)
2324 		sb->sb_mtlstail = n;
2325 
2326 	/* Detach the record from the TLS chain. */
2327 	m->m_next = NULL;
2328 
2329 out:
2330 	MPASS(m_length(top, NULL) == len);
2331 	for (m = top; m != NULL; m = m->m_next)
2332 		sbfree_ktls_rx(sb, m);
2333 	sb->sb_tlsdcc = len;
2334 	sb->sb_ccc += len;
2335 	SBCHECK(sb);
2336 	return (top);
2337 }
2338 
2339 /*
2340  * Determine the length of the trailing zero padding and find the real
2341  * record type in the byte before the padding.
2342  *
2343  * Walking the mbuf chain backwards is clumsy, so another option would
2344  * be to scan forwards remembering the last non-zero byte before the
2345  * trailer.  However, it would be expensive to scan the entire record.
2346  * Instead, find the last non-zero byte of each mbuf in the chain
2347  * keeping track of the relative offset of that nonzero byte.
2348  *
2349  * trail_len is the size of the MAC/tag on input and is set to the
2350  * size of the full trailer including padding and the record type on
2351  * return.
2352  */
2353 static int
2354 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2355     int *trailer_len, uint8_t *record_typep)
2356 {
2357 	char *cp;
2358 	u_int digest_start, last_offset, m_len, offset;
2359 	uint8_t record_type;
2360 
2361 	digest_start = tls_len - *trailer_len;
2362 	last_offset = 0;
2363 	offset = 0;
2364 	for (; m != NULL && offset < digest_start;
2365 	     offset += m->m_len, m = m->m_next) {
2366 		/* Don't look for padding in the tag. */
2367 		m_len = min(digest_start - offset, m->m_len);
2368 		cp = mtod(m, char *);
2369 
2370 		/* Find last non-zero byte in this mbuf. */
2371 		while (m_len > 0 && cp[m_len - 1] == 0)
2372 			m_len--;
2373 		if (m_len > 0) {
2374 			record_type = cp[m_len - 1];
2375 			last_offset = offset + m_len;
2376 		}
2377 	}
2378 	if (last_offset < tls->params.tls_hlen)
2379 		return (EBADMSG);
2380 
2381 	*record_typep = record_type;
2382 	*trailer_len = tls_len - last_offset + 1;
2383 	return (0);
2384 }
2385 
2386 /*
2387  * Check if a mbuf chain is fully decrypted at the given offset and
2388  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2389  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2390  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2391  * is encrypted.
2392  */
2393 ktls_mbuf_crypto_st_t
2394 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2395 {
2396 	int m_flags_ored = 0;
2397 	int m_flags_anded = -1;
2398 
2399 	for (; mb != NULL; mb = mb->m_next) {
2400 		if (offset < mb->m_len)
2401 			break;
2402 		offset -= mb->m_len;
2403 	}
2404 	offset += len;
2405 
2406 	for (; mb != NULL; mb = mb->m_next) {
2407 		m_flags_ored |= mb->m_flags;
2408 		m_flags_anded &= mb->m_flags;
2409 
2410 		if (offset <= mb->m_len)
2411 			break;
2412 		offset -= mb->m_len;
2413 	}
2414 	MPASS(mb != NULL || offset == 0);
2415 
2416 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2417 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2418 	else
2419 		return ((m_flags_ored & M_DECRYPTED) ?
2420 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2421 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2422 }
2423 
2424 /*
2425  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2426  */
2427 static int
2428 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2429 {
2430 	union if_snd_tag_modify_params params;
2431 	struct m_snd_tag *mst;
2432 	struct inpcb *inp;
2433 	struct tcpcb *tp;
2434 
2435 	mst = so->so_rcv.sb_tls_info->snd_tag;
2436 	if (__predict_false(mst == NULL))
2437 		return (EINVAL);
2438 
2439 	inp = sotoinpcb(so);
2440 	if (__predict_false(inp == NULL))
2441 		return (EINVAL);
2442 
2443 	INP_RLOCK(inp);
2444 	if (inp->inp_flags & INP_DROPPED) {
2445 		INP_RUNLOCK(inp);
2446 		return (ECONNRESET);
2447 	}
2448 
2449 	tp = intotcpcb(inp);
2450 	MPASS(tp != NULL);
2451 
2452 	/* Get the TCP sequence number of the next valid TLS header. */
2453 	SOCKBUF_LOCK(&so->so_rcv);
2454 	params.tls_rx.tls_hdr_tcp_sn =
2455 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2456 	params.tls_rx.tls_rec_length = tls_len;
2457 	params.tls_rx.tls_seq_number = tls_rcd_num;
2458 	SOCKBUF_UNLOCK(&so->so_rcv);
2459 
2460 	INP_RUNLOCK(inp);
2461 
2462 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2463 	return (mst->sw->snd_tag_modify(mst, &params));
2464 }
2465 
2466 static void
2467 ktls_drop(struct socket *so, int error)
2468 {
2469 	struct epoch_tracker et;
2470 	struct inpcb *inp = sotoinpcb(so);
2471 	struct tcpcb *tp;
2472 
2473 	NET_EPOCH_ENTER(et);
2474 	INP_WLOCK(inp);
2475 	if (!(inp->inp_flags & INP_DROPPED)) {
2476 		tp = intotcpcb(inp);
2477 		CURVNET_SET(inp->inp_vnet);
2478 		tp = tcp_drop(tp, error);
2479 		CURVNET_RESTORE();
2480 		if (tp != NULL)
2481 			INP_WUNLOCK(inp);
2482 	} else {
2483 		so->so_error = error;
2484 		SOCK_RECVBUF_LOCK(so);
2485 		sorwakeup_locked(so);
2486 		INP_WUNLOCK(inp);
2487 	}
2488 	NET_EPOCH_EXIT(et);
2489 }
2490 
2491 static void
2492 ktls_decrypt(struct socket *so)
2493 {
2494 	char tls_header[MBUF_PEXT_HDR_LEN];
2495 	struct ktls_session *tls;
2496 	struct sockbuf *sb;
2497 	struct tls_record_layer *hdr;
2498 	struct tls_get_record tgr;
2499 	struct mbuf *control, *data, *m;
2500 	ktls_mbuf_crypto_st_t state;
2501 	uint64_t seqno;
2502 	int error, remain, tls_len, trail_len;
2503 	bool tls13;
2504 	uint8_t vminor, record_type;
2505 
2506 	hdr = (struct tls_record_layer *)tls_header;
2507 	sb = &so->so_rcv;
2508 	SOCKBUF_LOCK(sb);
2509 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2510 	    ("%s: socket %p not running", __func__, so));
2511 
2512 	tls = sb->sb_tls_info;
2513 	MPASS(tls != NULL);
2514 
2515 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2516 	if (tls13)
2517 		vminor = TLS_MINOR_VER_TWO;
2518 	else
2519 		vminor = tls->params.tls_vminor;
2520 	for (;;) {
2521 		/* Is there enough queued for a TLS header? */
2522 		if (sb->sb_tlscc < tls->params.tls_hlen)
2523 			break;
2524 
2525 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2526 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2527 
2528 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2529 		    hdr->tls_vminor != vminor)
2530 			error = EINVAL;
2531 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2532 			error = EINVAL;
2533 		else if (tls_len < tls->params.tls_hlen || tls_len >
2534 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2535 		    tls->params.tls_tlen)
2536 			error = EMSGSIZE;
2537 		else
2538 			error = 0;
2539 		if (__predict_false(error != 0)) {
2540 			/*
2541 			 * We have a corrupted record and are likely
2542 			 * out of sync.  The connection isn't
2543 			 * recoverable at this point, so abort it.
2544 			 */
2545 			SOCKBUF_UNLOCK(sb);
2546 			counter_u64_add(ktls_offload_corrupted_records, 1);
2547 
2548 			ktls_drop(so, error);
2549 			goto deref;
2550 		}
2551 
2552 		/* Is the entire record queued? */
2553 		if (sb->sb_tlscc < tls_len)
2554 			break;
2555 
2556 		/*
2557 		 * Split out the portion of the mbuf chain containing
2558 		 * this TLS record.
2559 		 */
2560 		data = ktls_detach_record(sb, tls_len);
2561 		if (data == NULL)
2562 			continue;
2563 		MPASS(sb->sb_tlsdcc == tls_len);
2564 
2565 		seqno = sb->sb_tls_seqno;
2566 		sb->sb_tls_seqno++;
2567 		SBCHECK(sb);
2568 		SOCKBUF_UNLOCK(sb);
2569 
2570 		/* get crypto state for this TLS record */
2571 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2572 
2573 		switch (state) {
2574 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2575 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2576 			if (error)
2577 				break;
2578 			/* FALLTHROUGH */
2579 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2580 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2581 			    &trail_len);
2582 			if (__predict_true(error == 0)) {
2583 				if (tls13) {
2584 					error = tls13_find_record_type(tls, data,
2585 					    tls_len, &trail_len, &record_type);
2586 				} else {
2587 					record_type = hdr->tls_type;
2588 				}
2589 			}
2590 			break;
2591 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2592 			/*
2593 			 * NIC TLS is only supported for AEAD
2594 			 * ciphersuites which used a fixed sized
2595 			 * trailer.
2596 			 */
2597 			if (tls13) {
2598 				trail_len = tls->params.tls_tlen - 1;
2599 				error = tls13_find_record_type(tls, data,
2600 				    tls_len, &trail_len, &record_type);
2601 			} else {
2602 				trail_len = tls->params.tls_tlen;
2603 				error = 0;
2604 				record_type = hdr->tls_type;
2605 			}
2606 			break;
2607 		default:
2608 			error = EINVAL;
2609 			break;
2610 		}
2611 		if (error) {
2612 			counter_u64_add(ktls_offload_failed_crypto, 1);
2613 
2614 			SOCKBUF_LOCK(sb);
2615 			if (sb->sb_tlsdcc == 0) {
2616 				/*
2617 				 * sbcut/drop/flush discarded these
2618 				 * mbufs.
2619 				 */
2620 				m_freem(data);
2621 				break;
2622 			}
2623 
2624 			/*
2625 			 * Drop this TLS record's data, but keep
2626 			 * decrypting subsequent records.
2627 			 */
2628 			sb->sb_ccc -= tls_len;
2629 			sb->sb_tlsdcc = 0;
2630 
2631 			if (error != EMSGSIZE)
2632 				error = EBADMSG;
2633 			CURVNET_SET(so->so_vnet);
2634 			so->so_error = error;
2635 			sorwakeup_locked(so);
2636 			CURVNET_RESTORE();
2637 
2638 			m_freem(data);
2639 
2640 			SOCKBUF_LOCK(sb);
2641 			continue;
2642 		}
2643 
2644 		/* Allocate the control mbuf. */
2645 		memset(&tgr, 0, sizeof(tgr));
2646 		tgr.tls_type = record_type;
2647 		tgr.tls_vmajor = hdr->tls_vmajor;
2648 		tgr.tls_vminor = hdr->tls_vminor;
2649 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2650 		    trail_len);
2651 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2652 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2653 
2654 		SOCKBUF_LOCK(sb);
2655 		if (sb->sb_tlsdcc == 0) {
2656 			/* sbcut/drop/flush discarded these mbufs. */
2657 			MPASS(sb->sb_tlscc == 0);
2658 			m_freem(data);
2659 			m_freem(control);
2660 			break;
2661 		}
2662 
2663 		/*
2664 		 * Clear the 'dcc' accounting in preparation for
2665 		 * adding the decrypted record.
2666 		 */
2667 		sb->sb_ccc -= tls_len;
2668 		sb->sb_tlsdcc = 0;
2669 		SBCHECK(sb);
2670 
2671 		/* If there is no payload, drop all of the data. */
2672 		if (tgr.tls_length == htobe16(0)) {
2673 			m_freem(data);
2674 			data = NULL;
2675 		} else {
2676 			/* Trim header. */
2677 			remain = tls->params.tls_hlen;
2678 			while (remain > 0) {
2679 				if (data->m_len > remain) {
2680 					data->m_data += remain;
2681 					data->m_len -= remain;
2682 					break;
2683 				}
2684 				remain -= data->m_len;
2685 				data = m_free(data);
2686 			}
2687 
2688 			/* Trim trailer and clear M_NOTREADY. */
2689 			remain = be16toh(tgr.tls_length);
2690 			m = data;
2691 			for (m = data; remain > m->m_len; m = m->m_next) {
2692 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2693 				remain -= m->m_len;
2694 			}
2695 			m->m_len = remain;
2696 			m_freem(m->m_next);
2697 			m->m_next = NULL;
2698 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2699 
2700 			/* Set EOR on the final mbuf. */
2701 			m->m_flags |= M_EOR;
2702 		}
2703 
2704 		sbappendcontrol_locked(sb, data, control, 0);
2705 
2706 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2707 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2708 			SOCKBUF_UNLOCK(sb);
2709 			ktls_resync_ifnet(so, tls_len, seqno);
2710 			SOCKBUF_LOCK(sb);
2711 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2712 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2713 			SOCKBUF_UNLOCK(sb);
2714 			ktls_resync_ifnet(so, 0, seqno);
2715 			SOCKBUF_LOCK(sb);
2716 		}
2717 	}
2718 
2719 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2720 
2721 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2722 		so->so_error = EMSGSIZE;
2723 
2724 	sorwakeup_locked(so);
2725 
2726 deref:
2727 	SOCKBUF_UNLOCK_ASSERT(sb);
2728 
2729 	CURVNET_SET(so->so_vnet);
2730 	sorele(so);
2731 	CURVNET_RESTORE();
2732 }
2733 
2734 void
2735 ktls_enqueue_to_free(struct mbuf *m)
2736 {
2737 	struct ktls_wq *wq;
2738 	bool running;
2739 
2740 	/* Mark it for freeing. */
2741 	m->m_epg_flags |= EPG_FLAG_2FREE;
2742 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2743 	mtx_lock(&wq->mtx);
2744 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2745 	running = wq->running;
2746 	mtx_unlock(&wq->mtx);
2747 	if (!running)
2748 		wakeup(wq);
2749 }
2750 
2751 static void *
2752 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2753 {
2754 	void *buf;
2755 	int domain, running;
2756 
2757 	if (m->m_epg_npgs <= 2)
2758 		return (NULL);
2759 	if (ktls_buffer_zone == NULL)
2760 		return (NULL);
2761 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2762 		/*
2763 		 * Rate-limit allocation attempts after a failure.
2764 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2765 		 * the free page queues and may fail consistently if memory is
2766 		 * fragmented.
2767 		 */
2768 		return (NULL);
2769 	}
2770 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2771 	if (buf == NULL) {
2772 		domain = PCPU_GET(domain);
2773 		wq->lastallocfail = ticks;
2774 
2775 		/*
2776 		 * Note that this check is "racy", but the races are
2777 		 * harmless, and are either a spurious wakeup if
2778 		 * multiple threads fail allocations before the alloc
2779 		 * thread wakes, or waiting an extra second in case we
2780 		 * see an old value of running == true.
2781 		 */
2782 		if (!VM_DOMAIN_EMPTY(domain)) {
2783 			running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2784 			if (!running)
2785 				wakeup(&ktls_domains[domain].reclaim_td);
2786 		}
2787 	}
2788 	return (buf);
2789 }
2790 
2791 static int
2792 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2793     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2794 {
2795 	vm_page_t pg;
2796 	int error, i, len, off;
2797 
2798 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2799 	    ("%p not unready & nomap mbuf\n", m));
2800 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2801 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2802 	    ktls_maxlen));
2803 
2804 	/* Anonymous mbufs are encrypted in place. */
2805 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2806 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2807 
2808 	/*
2809 	 * For file-backed mbufs (from sendfile), anonymous wired
2810 	 * pages are allocated and used as the encryption destination.
2811 	 */
2812 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2813 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2814 		    m->m_epg_1st_off;
2815 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2816 		    m->m_epg_1st_off;
2817 		state->dst_iov[0].iov_len = len;
2818 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2819 		i = 1;
2820 	} else {
2821 		off = m->m_epg_1st_off;
2822 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2823 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2824 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2825 			len = m_epg_pagelen(m, i, off);
2826 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2827 			state->dst_iov[i].iov_base =
2828 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2829 			state->dst_iov[i].iov_len = len;
2830 		}
2831 	}
2832 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2833 	state->dst_iov[i].iov_base = m->m_epg_trail;
2834 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2835 
2836 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2837 
2838 	if (__predict_false(error != 0)) {
2839 		/* Free the anonymous pages. */
2840 		if (state->cbuf != NULL)
2841 			uma_zfree(ktls_buffer_zone, state->cbuf);
2842 		else {
2843 			for (i = 0; i < m->m_epg_npgs; i++) {
2844 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2845 				(void)vm_page_unwire_noq(pg);
2846 				vm_page_free(pg);
2847 			}
2848 		}
2849 	}
2850 	return (error);
2851 }
2852 
2853 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2854 static u_int
2855 ktls_batched_records(struct mbuf *m)
2856 {
2857 	int page_count, records;
2858 
2859 	records = 0;
2860 	page_count = m->m_epg_enc_cnt;
2861 	while (page_count > 0) {
2862 		records++;
2863 		page_count -= m->m_epg_nrdy;
2864 		m = m->m_next;
2865 	}
2866 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2867 	return (records);
2868 }
2869 
2870 void
2871 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2872 {
2873 	struct ktls_session *tls;
2874 	struct ktls_wq *wq;
2875 	int queued;
2876 	bool running;
2877 
2878 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2879 	    (M_EXTPG | M_NOTREADY)),
2880 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2881 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2882 
2883 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2884 
2885 	m->m_epg_enc_cnt = page_count;
2886 
2887 	/*
2888 	 * Save a pointer to the socket.  The caller is responsible
2889 	 * for taking an additional reference via soref().
2890 	 */
2891 	m->m_epg_so = so;
2892 
2893 	queued = 1;
2894 	tls = m->m_epg_tls;
2895 	wq = &ktls_wq[tls->wq_index];
2896 	mtx_lock(&wq->mtx);
2897 	if (__predict_false(tls->sequential_records)) {
2898 		/*
2899 		 * For TLS 1.0, records must be encrypted
2900 		 * sequentially.  For a given connection, all records
2901 		 * queued to the associated work queue are processed
2902 		 * sequentially.  However, sendfile(2) might complete
2903 		 * I/O requests spanning multiple TLS records out of
2904 		 * order.  Here we ensure TLS records are enqueued to
2905 		 * the work queue in FIFO order.
2906 		 *
2907 		 * tls->next_seqno holds the sequence number of the
2908 		 * next TLS record that should be enqueued to the work
2909 		 * queue.  If this next record is not tls->next_seqno,
2910 		 * it must be a future record, so insert it, sorted by
2911 		 * TLS sequence number, into tls->pending_records and
2912 		 * return.
2913 		 *
2914 		 * If this TLS record matches tls->next_seqno, place
2915 		 * it in the work queue and then check
2916 		 * tls->pending_records to see if any
2917 		 * previously-queued records are now ready for
2918 		 * encryption.
2919 		 */
2920 		if (m->m_epg_seqno != tls->next_seqno) {
2921 			struct mbuf *n, *p;
2922 
2923 			p = NULL;
2924 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2925 				if (n->m_epg_seqno > m->m_epg_seqno)
2926 					break;
2927 				p = n;
2928 			}
2929 			if (n == NULL)
2930 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2931 				    m_epg_stailq);
2932 			else if (p == NULL)
2933 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2934 				    m_epg_stailq);
2935 			else
2936 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2937 				    m_epg_stailq);
2938 			mtx_unlock(&wq->mtx);
2939 			counter_u64_add(ktls_cnt_tx_pending, 1);
2940 			return;
2941 		}
2942 
2943 		tls->next_seqno += ktls_batched_records(m);
2944 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2945 
2946 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2947 			struct mbuf *n;
2948 
2949 			n = STAILQ_FIRST(&tls->pending_records);
2950 			if (n->m_epg_seqno != tls->next_seqno)
2951 				break;
2952 
2953 			queued++;
2954 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2955 			tls->next_seqno += ktls_batched_records(n);
2956 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2957 		}
2958 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2959 	} else
2960 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2961 
2962 	running = wq->running;
2963 	mtx_unlock(&wq->mtx);
2964 	if (!running)
2965 		wakeup(wq);
2966 	counter_u64_add(ktls_cnt_tx_queued, queued);
2967 }
2968 
2969 /*
2970  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2971  * the pages from the file and replace them with the anonymous pages
2972  * allocated in ktls_encrypt_record().
2973  */
2974 static void
2975 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2976 {
2977 	int i;
2978 
2979 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2980 
2981 	/* Free the old pages. */
2982 	m->m_ext.ext_free(m);
2983 
2984 	/* Replace them with the new pages. */
2985 	if (state->cbuf != NULL) {
2986 		for (i = 0; i < m->m_epg_npgs; i++)
2987 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2988 
2989 		/* Contig pages should go back to the cache. */
2990 		m->m_ext.ext_free = ktls_free_mext_contig;
2991 	} else {
2992 		for (i = 0; i < m->m_epg_npgs; i++)
2993 			m->m_epg_pa[i] = state->parray[i];
2994 
2995 		/* Use the basic free routine. */
2996 		m->m_ext.ext_free = mb_free_mext_pgs;
2997 	}
2998 
2999 	/* Pages are now writable. */
3000 	m->m_epg_flags |= EPG_FLAG_ANON;
3001 }
3002 
3003 static __noinline void
3004 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
3005 {
3006 	struct ktls_ocf_encrypt_state state;
3007 	struct ktls_session *tls;
3008 	struct socket *so;
3009 	struct mbuf *m;
3010 	int error, npages, total_pages;
3011 
3012 	so = top->m_epg_so;
3013 	tls = top->m_epg_tls;
3014 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3015 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3016 #ifdef INVARIANTS
3017 	top->m_epg_so = NULL;
3018 #endif
3019 	total_pages = top->m_epg_enc_cnt;
3020 	npages = 0;
3021 
3022 	/*
3023 	 * Encrypt the TLS records in the chain of mbufs starting with
3024 	 * 'top'.  'total_pages' gives us a total count of pages and is
3025 	 * used to know when we have finished encrypting the TLS
3026 	 * records originally queued with 'top'.
3027 	 *
3028 	 * NB: These mbufs are queued in the socket buffer and
3029 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
3030 	 * socket buffer lock is not held while traversing this chain.
3031 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
3032 	 * pointers should be stable.  However, the 'm_next' of the
3033 	 * last mbuf encrypted is not necessarily NULL.  It can point
3034 	 * to other mbufs appended while 'top' was on the TLS work
3035 	 * queue.
3036 	 *
3037 	 * Each mbuf holds an entire TLS record.
3038 	 */
3039 	error = 0;
3040 	for (m = top; npages != total_pages; m = m->m_next) {
3041 		KASSERT(m->m_epg_tls == tls,
3042 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3043 		    tls, m->m_epg_tls));
3044 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3045 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3046 		    total_pages, m));
3047 
3048 		error = ktls_encrypt_record(wq, m, tls, &state);
3049 		if (error) {
3050 			counter_u64_add(ktls_offload_failed_crypto, 1);
3051 			break;
3052 		}
3053 
3054 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3055 			ktls_finish_nonanon(m, &state);
3056 
3057 		npages += m->m_epg_nrdy;
3058 
3059 		/*
3060 		 * Drop a reference to the session now that it is no
3061 		 * longer needed.  Existing code depends on encrypted
3062 		 * records having no associated session vs
3063 		 * yet-to-be-encrypted records having an associated
3064 		 * session.
3065 		 */
3066 		m->m_epg_tls = NULL;
3067 		ktls_free(tls);
3068 	}
3069 
3070 	CURVNET_SET(so->so_vnet);
3071 	if (error == 0) {
3072 		(void)so->so_proto->pr_ready(so, top, npages);
3073 	} else {
3074 		ktls_drop(so, EIO);
3075 		mb_free_notready(top, total_pages);
3076 	}
3077 
3078 	sorele(so);
3079 	CURVNET_RESTORE();
3080 }
3081 
3082 void
3083 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3084 {
3085 	struct ktls_session *tls;
3086 	struct socket *so;
3087 	struct mbuf *m;
3088 	int npages;
3089 
3090 	m = state->m;
3091 
3092 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3093 		ktls_finish_nonanon(m, state);
3094 
3095 	so = state->so;
3096 	free(state, M_KTLS);
3097 
3098 	/*
3099 	 * Drop a reference to the session now that it is no longer
3100 	 * needed.  Existing code depends on encrypted records having
3101 	 * no associated session vs yet-to-be-encrypted records having
3102 	 * an associated session.
3103 	 */
3104 	tls = m->m_epg_tls;
3105 	m->m_epg_tls = NULL;
3106 	ktls_free(tls);
3107 
3108 	if (error != 0)
3109 		counter_u64_add(ktls_offload_failed_crypto, 1);
3110 
3111 	CURVNET_SET(so->so_vnet);
3112 	npages = m->m_epg_nrdy;
3113 
3114 	if (error == 0) {
3115 		(void)so->so_proto->pr_ready(so, m, npages);
3116 	} else {
3117 		ktls_drop(so, EIO);
3118 		mb_free_notready(m, npages);
3119 	}
3120 
3121 	sorele(so);
3122 	CURVNET_RESTORE();
3123 }
3124 
3125 /*
3126  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3127  * (coprocessors) where encryption does not use host CPU resources and
3128  * it can be beneficial to queue more requests than CPUs.
3129  */
3130 static __noinline void
3131 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3132 {
3133 	struct ktls_ocf_encrypt_state *state;
3134 	struct ktls_session *tls;
3135 	struct socket *so;
3136 	struct mbuf *m, *n;
3137 	int error, mpages, npages, total_pages;
3138 
3139 	so = top->m_epg_so;
3140 	tls = top->m_epg_tls;
3141 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3142 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3143 #ifdef INVARIANTS
3144 	top->m_epg_so = NULL;
3145 #endif
3146 	total_pages = top->m_epg_enc_cnt;
3147 	npages = 0;
3148 
3149 	error = 0;
3150 	for (m = top; npages != total_pages; m = n) {
3151 		KASSERT(m->m_epg_tls == tls,
3152 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3153 		    tls, m->m_epg_tls));
3154 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3155 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3156 		    total_pages, m));
3157 
3158 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3159 		soref(so);
3160 		state->so = so;
3161 		state->m = m;
3162 
3163 		mpages = m->m_epg_nrdy;
3164 		n = m->m_next;
3165 
3166 		error = ktls_encrypt_record(wq, m, tls, state);
3167 		if (error) {
3168 			counter_u64_add(ktls_offload_failed_crypto, 1);
3169 			free(state, M_KTLS);
3170 			CURVNET_SET(so->so_vnet);
3171 			sorele(so);
3172 			CURVNET_RESTORE();
3173 			break;
3174 		}
3175 
3176 		npages += mpages;
3177 	}
3178 
3179 	CURVNET_SET(so->so_vnet);
3180 	if (error != 0) {
3181 		ktls_drop(so, EIO);
3182 		mb_free_notready(m, total_pages - npages);
3183 	}
3184 
3185 	sorele(so);
3186 	CURVNET_RESTORE();
3187 }
3188 
3189 static int
3190 ktls_bind_domain(int domain)
3191 {
3192 	int error;
3193 
3194 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3195 	if (error != 0)
3196 		return (error);
3197 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3198 	return (0);
3199 }
3200 
3201 static void
3202 ktls_reclaim_thread(void *ctx)
3203 {
3204 	struct ktls_domain_info *ktls_domain = ctx;
3205 	struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3206 	struct sysctl_oid *oid;
3207 	char name[80];
3208 	int error, domain;
3209 
3210 	domain = ktls_domain - ktls_domains;
3211 	if (bootverbose)
3212 		printf("Starting KTLS reclaim thread for domain %d\n", domain);
3213 	error = ktls_bind_domain(domain);
3214 	if (error)
3215 		printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3216 		    domain, error);
3217 	snprintf(name, sizeof(name), "domain%d", domain);
3218 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3219 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3220 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3221 	    CTLFLAG_RD,  &sc->reclaims, 0, "buffers reclaimed");
3222 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3223 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3224 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3225 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3226 
3227 	for (;;) {
3228 		atomic_store_int(&sc->running, 0);
3229 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3230 		atomic_store_int(&sc->running, 1);
3231 		sc->wakeups++;
3232 		/*
3233 		 * Below we attempt to reclaim ktls_max_reclaim
3234 		 * buffers using vm_page_reclaim_contig_domain_ext().
3235 		 * We do this here, as this function can take several
3236 		 * seconds to scan all of memory and it does not
3237 		 * matter if this thread pauses for a while.  If we
3238 		 * block a ktls worker thread, we risk developing
3239 		 * backlogs of buffers to be encrypted, leading to
3240 		 * surges of traffic and potential NIC output drops.
3241 		 */
3242 		if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3243 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3244 		    ktls_max_reclaim) != 0) {
3245 			vm_wait_domain(domain);
3246 		} else {
3247 			sc->reclaims += ktls_max_reclaim;
3248 		}
3249 	}
3250 }
3251 
3252 static void
3253 ktls_work_thread(void *ctx)
3254 {
3255 	struct ktls_wq *wq = ctx;
3256 	struct mbuf *m, *n;
3257 	struct socket *so, *son;
3258 	STAILQ_HEAD(, mbuf) local_m_head;
3259 	STAILQ_HEAD(, socket) local_so_head;
3260 	int cpu;
3261 
3262 	cpu = wq - ktls_wq;
3263 	if (bootverbose)
3264 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3265 
3266 	/*
3267 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3268 	 * we bind to the NUMA domain instead.
3269 	 */
3270 	if (ktls_bind_threads) {
3271 		int error;
3272 
3273 		if (ktls_bind_threads > 1) {
3274 			struct pcpu *pc = pcpu_find(cpu);
3275 
3276 			error = ktls_bind_domain(pc->pc_domain);
3277 		} else {
3278 			cpuset_t mask;
3279 
3280 			CPU_SETOF(cpu, &mask);
3281 			error = cpuset_setthread(curthread->td_tid, &mask);
3282 		}
3283 		if (error)
3284 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3285 				cpu, error);
3286 	}
3287 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3288 	fpu_kern_thread(0);
3289 #endif
3290 	for (;;) {
3291 		mtx_lock(&wq->mtx);
3292 		while (STAILQ_EMPTY(&wq->m_head) &&
3293 		    STAILQ_EMPTY(&wq->so_head)) {
3294 			wq->running = false;
3295 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3296 			wq->running = true;
3297 		}
3298 
3299 		STAILQ_INIT(&local_m_head);
3300 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3301 		STAILQ_INIT(&local_so_head);
3302 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3303 		mtx_unlock(&wq->mtx);
3304 
3305 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3306 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3307 				ktls_free(m->m_epg_tls);
3308 				m_free_raw(m);
3309 			} else {
3310 				if (m->m_epg_tls->sync_dispatch)
3311 					ktls_encrypt(wq, m);
3312 				else
3313 					ktls_encrypt_async(wq, m);
3314 				counter_u64_add(ktls_cnt_tx_queued, -1);
3315 			}
3316 		}
3317 
3318 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3319 			ktls_decrypt(so);
3320 			counter_u64_add(ktls_cnt_rx_queued, -1);
3321 		}
3322 	}
3323 }
3324 
3325 static void
3326 ktls_disable_ifnet_help(void *context, int pending __unused)
3327 {
3328 	struct ktls_session *tls;
3329 	struct inpcb *inp;
3330 	struct tcpcb *tp;
3331 	struct socket *so;
3332 	int err;
3333 
3334 	tls = context;
3335 	inp = tls->inp;
3336 	if (inp == NULL)
3337 		return;
3338 	INP_WLOCK(inp);
3339 	so = inp->inp_socket;
3340 	MPASS(so != NULL);
3341 	if (inp->inp_flags & INP_DROPPED) {
3342 		goto out;
3343 	}
3344 
3345 	if (so->so_snd.sb_tls_info != NULL)
3346 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3347 	else
3348 		err = ENXIO;
3349 	if (err == 0) {
3350 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3351 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3352 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3353 		    (tp = intotcpcb(inp)) != NULL &&
3354 		    tp->t_fb->tfb_hwtls_change != NULL)
3355 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3356 	} else {
3357 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3358 	}
3359 
3360 out:
3361 	CURVNET_SET(so->so_vnet);
3362 	sorele(so);
3363 	CURVNET_RESTORE();
3364 	INP_WUNLOCK(inp);
3365 	ktls_free(tls);
3366 }
3367 
3368 /*
3369  * Called when re-transmits are becoming a substantial portion of the
3370  * sends on this connection.  When this happens, we transition the
3371  * connection to software TLS.  This is needed because most inline TLS
3372  * NICs keep crypto state only for in-order transmits.  This means
3373  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3374  * re-DMA the entire TLS record up to and including the current
3375  * segment.  This means that when re-transmitting the last ~1448 byte
3376  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3377  * of magnitude more data than we are sending.  This can cause the
3378  * PCIe link to saturate well before the network, which can cause
3379  * output drops, and a general loss of capacity.
3380  */
3381 void
3382 ktls_disable_ifnet(void *arg)
3383 {
3384 	struct tcpcb *tp;
3385 	struct inpcb *inp;
3386 	struct socket *so;
3387 	struct ktls_session *tls;
3388 
3389 	tp = arg;
3390 	inp = tptoinpcb(tp);
3391 	INP_WLOCK_ASSERT(inp);
3392 	so = inp->inp_socket;
3393 	SOCK_LOCK(so);
3394 	tls = so->so_snd.sb_tls_info;
3395 	if (tp->t_nic_ktls_xmit_dis == 1) {
3396 		SOCK_UNLOCK(so);
3397 		return;
3398 	}
3399 
3400 	/*
3401 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3402 	 * ifnet can only be done once per connection, so we never want
3403 	 * to do it again
3404 	 */
3405 
3406 	(void)ktls_hold(tls);
3407 	soref(so);
3408 	tp->t_nic_ktls_xmit_dis = 1;
3409 	SOCK_UNLOCK(so);
3410 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3411 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3412 }
3413