xref: /freebsd/sys/kern/uipc_ktls.c (revision 21b492ed51aa6ff8008a8aa83333b1de30288a15)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/domainset.h>
40 #include <sys/endian.h>
41 #include <sys/ktls.h>
42 #include <sys/lock.h>
43 #include <sys/mbuf.h>
44 #include <sys/mutex.h>
45 #include <sys/rmlock.h>
46 #include <sys/proc.h>
47 #include <sys/protosw.h>
48 #include <sys/refcount.h>
49 #include <sys/smp.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/kthread.h>
55 #include <sys/uio.h>
56 #include <sys/vmmeter.h>
57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58 #include <machine/pcb.h>
59 #endif
60 #include <machine/vmparam.h>
61 #include <net/if.h>
62 #include <net/if_var.h>
63 #ifdef RSS
64 #include <net/netisr.h>
65 #include <net/rss_config.h>
66 #endif
67 #include <net/route.h>
68 #include <net/route/nhop.h>
69 #if defined(INET) || defined(INET6)
70 #include <netinet/in.h>
71 #include <netinet/in_pcb.h>
72 #endif
73 #include <netinet/tcp_var.h>
74 #ifdef TCP_OFFLOAD
75 #include <netinet/tcp_offload.h>
76 #endif
77 #include <opencrypto/cryptodev.h>
78 #include <opencrypto/ktls.h>
79 #include <vm/uma_dbg.h>
80 #include <vm/vm.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pagequeue.h>
84 
85 struct ktls_wq {
86 	struct mtx	mtx;
87 	STAILQ_HEAD(, mbuf) m_head;
88 	STAILQ_HEAD(, socket) so_head;
89 	bool		running;
90 	int		lastallocfail;
91 } __aligned(CACHE_LINE_SIZE);
92 
93 struct ktls_alloc_thread {
94 	uint64_t wakeups;
95 	uint64_t allocs;
96 	struct thread *td;
97 	int running;
98 };
99 
100 struct ktls_domain_info {
101 	int count;
102 	int cpu[MAXCPU];
103 	struct ktls_alloc_thread alloc_td;
104 };
105 
106 struct ktls_domain_info ktls_domains[MAXMEMDOM];
107 static struct ktls_wq *ktls_wq;
108 static struct proc *ktls_proc;
109 static uma_zone_t ktls_session_zone;
110 static uma_zone_t ktls_buffer_zone;
111 static uint16_t ktls_cpuid_lookup[MAXCPU];
112 
113 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114     "Kernel TLS offload");
115 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
116     "Kernel TLS offload stats");
117 
118 #ifdef RSS
119 static int ktls_bind_threads = 1;
120 #else
121 static int ktls_bind_threads;
122 #endif
123 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
124     &ktls_bind_threads, 0,
125     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
126 
127 static u_int ktls_maxlen = 16384;
128 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
129     &ktls_maxlen, 0, "Maximum TLS record size");
130 
131 static int ktls_number_threads;
132 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
133     &ktls_number_threads, 0,
134     "Number of TLS threads in thread-pool");
135 
136 unsigned int ktls_ifnet_max_rexmit_pct = 2;
137 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
138     &ktls_ifnet_max_rexmit_pct, 2,
139     "Max percent bytes retransmitted before ifnet TLS is disabled");
140 
141 static bool ktls_offload_enable;
142 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
143     &ktls_offload_enable, 0,
144     "Enable support for kernel TLS offload");
145 
146 static bool ktls_cbc_enable = true;
147 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
148     &ktls_cbc_enable, 1,
149     "Enable Support of AES-CBC crypto for kernel TLS");
150 
151 static bool ktls_sw_buffer_cache = true;
152 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
153     &ktls_sw_buffer_cache, 1,
154     "Enable caching of output buffers for SW encryption");
155 
156 static int ktls_max_alloc = 128;
157 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
158     &ktls_max_alloc, 128,
159     "Max number of 16k buffers to allocate in thread context");
160 
161 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
162 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
163     &ktls_tasks_active, "Number of active tasks");
164 
165 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
166 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
167     &ktls_cnt_tx_queued,
168     "Number of TLS records in queue to tasks for SW encryption");
169 
170 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
171 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
172     &ktls_cnt_rx_queued,
173     "Number of TLS sockets in queue to tasks for SW decryption");
174 
175 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
176 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
177     CTLFLAG_RD, &ktls_offload_total,
178     "Total successful TLS setups (parameters set)");
179 
180 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
182     CTLFLAG_RD, &ktls_offload_enable_calls,
183     "Total number of TLS enable calls made");
184 
185 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
186 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
187     &ktls_offload_active, "Total Active TLS sessions");
188 
189 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
190 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
191     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
192 
193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
195     &ktls_offload_failed_crypto, "Total TLS crypto failures");
196 
197 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
199     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
200 
201 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
203     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
204 
205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
207     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
208 
209 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
211     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
212 
213 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
215     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
216 
217 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
218     "Software TLS session stats");
219 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
220     "Hardware (ifnet) TLS session stats");
221 #ifdef TCP_OFFLOAD
222 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
223     "TOE TLS session stats");
224 #endif
225 
226 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
227 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
228     "Active number of software TLS sessions using AES-CBC");
229 
230 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
231 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
232     "Active number of software TLS sessions using AES-GCM");
233 
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
236     &ktls_sw_chacha20,
237     "Active number of software TLS sessions using Chacha20-Poly1305");
238 
239 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
240 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
241     &ktls_ifnet_cbc,
242     "Active number of ifnet TLS sessions using AES-CBC");
243 
244 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
245 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
246     &ktls_ifnet_gcm,
247     "Active number of ifnet TLS sessions using AES-GCM");
248 
249 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
250 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
251     &ktls_ifnet_chacha20,
252     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
253 
254 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
255 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
256     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
257 
258 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
259 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
260     &ktls_ifnet_reset_dropped,
261     "TLS sessions dropped after failing to update ifnet send tag");
262 
263 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
264 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
265     &ktls_ifnet_reset_failed,
266     "TLS sessions that failed to allocate a new ifnet send tag");
267 
268 static int ktls_ifnet_permitted;
269 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
270     &ktls_ifnet_permitted, 1,
271     "Whether to permit hardware (ifnet) TLS sessions");
272 
273 #ifdef TCP_OFFLOAD
274 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
275 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
276     &ktls_toe_cbc,
277     "Active number of TOE TLS sessions using AES-CBC");
278 
279 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
280 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
281     &ktls_toe_gcm,
282     "Active number of TOE TLS sessions using AES-GCM");
283 
284 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
285 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
286     &ktls_toe_chacha20,
287     "Active number of TOE TLS sessions using Chacha20-Poly1305");
288 #endif
289 
290 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
291 
292 static void ktls_cleanup(struct ktls_session *tls);
293 #if defined(INET) || defined(INET6)
294 static void ktls_reset_send_tag(void *context, int pending);
295 #endif
296 static void ktls_work_thread(void *ctx);
297 static void ktls_alloc_thread(void *ctx);
298 
299 #if defined(INET) || defined(INET6)
300 static u_int
301 ktls_get_cpu(struct socket *so)
302 {
303 	struct inpcb *inp;
304 #ifdef NUMA
305 	struct ktls_domain_info *di;
306 #endif
307 	u_int cpuid;
308 
309 	inp = sotoinpcb(so);
310 #ifdef RSS
311 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
312 	if (cpuid != NETISR_CPUID_NONE)
313 		return (cpuid);
314 #endif
315 	/*
316 	 * Just use the flowid to shard connections in a repeatable
317 	 * fashion.  Note that TLS 1.0 sessions rely on the
318 	 * serialization provided by having the same connection use
319 	 * the same queue.
320 	 */
321 #ifdef NUMA
322 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
323 		di = &ktls_domains[inp->inp_numa_domain];
324 		cpuid = di->cpu[inp->inp_flowid % di->count];
325 	} else
326 #endif
327 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
328 	return (cpuid);
329 }
330 #endif
331 
332 static int
333 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
334 {
335 	vm_page_t m;
336 	int i;
337 
338 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
339 	    ("%s: ktls max length %d is not page size-aligned",
340 	    __func__, ktls_maxlen));
341 
342 	for (i = 0; i < count; i++) {
343 		m = vm_page_alloc_contig_domain(NULL, 0, domain,
344 		    VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
345 		    VM_ALLOC_NODUMP | malloc2vm_flags(flags),
346 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
347 		    VM_MEMATTR_DEFAULT);
348 		if (m == NULL)
349 			break;
350 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
351 	}
352 	return (i);
353 }
354 
355 static void
356 ktls_buffer_release(void *arg __unused, void **store, int count)
357 {
358 	vm_page_t m;
359 	int i, j;
360 
361 	for (i = 0; i < count; i++) {
362 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
363 		for (j = 0; j < atop(ktls_maxlen); j++) {
364 			(void)vm_page_unwire_noq(m + j);
365 			vm_page_free(m + j);
366 		}
367 	}
368 }
369 
370 static void
371 ktls_free_mext_contig(struct mbuf *m)
372 {
373 	M_ASSERTEXTPG(m);
374 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
375 }
376 
377 static void
378 ktls_init(void *dummy __unused)
379 {
380 	struct thread *td;
381 	struct pcpu *pc;
382 	cpuset_t mask;
383 	int count, domain, error, i;
384 
385 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
386 	    M_WAITOK | M_ZERO);
387 
388 	ktls_session_zone = uma_zcreate("ktls_session",
389 	    sizeof(struct ktls_session),
390 	    NULL, NULL, NULL, NULL,
391 	    UMA_ALIGN_CACHE, 0);
392 
393 	if (ktls_sw_buffer_cache) {
394 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
395 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
396 		    ktls_buffer_import, ktls_buffer_release, NULL,
397 		    UMA_ZONE_FIRSTTOUCH);
398 	}
399 
400 	/*
401 	 * Initialize the workqueues to run the TLS work.  We create a
402 	 * work queue for each CPU.
403 	 */
404 	CPU_FOREACH(i) {
405 		STAILQ_INIT(&ktls_wq[i].m_head);
406 		STAILQ_INIT(&ktls_wq[i].so_head);
407 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
408 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
409 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
410 		if (error)
411 			panic("Can't add KTLS thread %d error %d", i, error);
412 
413 		/*
414 		 * Bind threads to cores.  If ktls_bind_threads is >
415 		 * 1, then we bind to the NUMA domain.
416 		 */
417 		if (ktls_bind_threads) {
418 			if (ktls_bind_threads > 1) {
419 				pc = pcpu_find(i);
420 				domain = pc->pc_domain;
421 				CPU_COPY(&cpuset_domain[domain], &mask);
422 				count = ktls_domains[domain].count;
423 				ktls_domains[domain].cpu[count] = i;
424 				ktls_domains[domain].count++;
425 			} else {
426 				CPU_SETOF(i, &mask);
427 			}
428 			error = cpuset_setthread(td->td_tid, &mask);
429 			if (error)
430 				panic(
431 			    "Unable to bind KTLS thread for CPU %d error %d",
432 				     i, error);
433 		}
434 		ktls_cpuid_lookup[ktls_number_threads] = i;
435 		ktls_number_threads++;
436 	}
437 
438 	/*
439 	 * Start an allocation thread per-domain to perform blocking allocations
440 	 * of 16k physically contiguous TLS crypto destination buffers.
441 	 */
442 	if (ktls_sw_buffer_cache) {
443 		for (domain = 0; domain < vm_ndomains; domain++) {
444 			if (VM_DOMAIN_EMPTY(domain))
445 				continue;
446 			if (CPU_EMPTY(&cpuset_domain[domain]))
447 				continue;
448 			error = kproc_kthread_add(ktls_alloc_thread,
449 			    &ktls_domains[domain], &ktls_proc,
450 			    &ktls_domains[domain].alloc_td.td,
451 			    0, 0, "KTLS", "alloc_%d", domain);
452 			if (error)
453 				panic("Can't add KTLS alloc thread %d error %d",
454 				    domain, error);
455 			CPU_COPY(&cpuset_domain[domain], &mask);
456 			error = cpuset_setthread(ktls_domains[domain].alloc_td.td->td_tid,
457 			    &mask);
458 			if (error)
459 				panic("Unable to bind KTLS alloc %d error %d",
460 				    domain, error);
461 		}
462 	}
463 
464 	/*
465 	 * If we somehow have an empty domain, fall back to choosing
466 	 * among all KTLS threads.
467 	 */
468 	if (ktls_bind_threads > 1) {
469 		for (i = 0; i < vm_ndomains; i++) {
470 			if (ktls_domains[i].count == 0) {
471 				ktls_bind_threads = 1;
472 				break;
473 			}
474 		}
475 	}
476 
477 	if (bootverbose)
478 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
479 }
480 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
481 
482 #if defined(INET) || defined(INET6)
483 static int
484 ktls_create_session(struct socket *so, struct tls_enable *en,
485     struct ktls_session **tlsp)
486 {
487 	struct ktls_session *tls;
488 	int error;
489 
490 	/* Only TLS 1.0 - 1.3 are supported. */
491 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
492 		return (EINVAL);
493 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
494 	    en->tls_vminor > TLS_MINOR_VER_THREE)
495 		return (EINVAL);
496 
497 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
498 		return (EINVAL);
499 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
500 		return (EINVAL);
501 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
502 		return (EINVAL);
503 
504 	/* All supported algorithms require a cipher key. */
505 	if (en->cipher_key_len == 0)
506 		return (EINVAL);
507 
508 	/* No flags are currently supported. */
509 	if (en->flags != 0)
510 		return (EINVAL);
511 
512 	/* Common checks for supported algorithms. */
513 	switch (en->cipher_algorithm) {
514 	case CRYPTO_AES_NIST_GCM_16:
515 		/*
516 		 * auth_algorithm isn't used, but permit GMAC values
517 		 * for compatibility.
518 		 */
519 		switch (en->auth_algorithm) {
520 		case 0:
521 #ifdef COMPAT_FREEBSD12
522 		/* XXX: Really 13.0-current COMPAT. */
523 		case CRYPTO_AES_128_NIST_GMAC:
524 		case CRYPTO_AES_192_NIST_GMAC:
525 		case CRYPTO_AES_256_NIST_GMAC:
526 #endif
527 			break;
528 		default:
529 			return (EINVAL);
530 		}
531 		if (en->auth_key_len != 0)
532 			return (EINVAL);
533 		if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
534 			en->iv_len != TLS_AEAD_GCM_LEN) ||
535 		    (en->tls_vminor == TLS_MINOR_VER_THREE &&
536 			en->iv_len != TLS_1_3_GCM_IV_LEN))
537 			return (EINVAL);
538 		break;
539 	case CRYPTO_AES_CBC:
540 		switch (en->auth_algorithm) {
541 		case CRYPTO_SHA1_HMAC:
542 			/*
543 			 * TLS 1.0 requires an implicit IV.  TLS 1.1+
544 			 * all use explicit IVs.
545 			 */
546 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
547 				if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
548 					return (EINVAL);
549 				break;
550 			}
551 
552 			/* FALLTHROUGH */
553 		case CRYPTO_SHA2_256_HMAC:
554 		case CRYPTO_SHA2_384_HMAC:
555 			/* Ignore any supplied IV. */
556 			en->iv_len = 0;
557 			break;
558 		default:
559 			return (EINVAL);
560 		}
561 		if (en->auth_key_len == 0)
562 			return (EINVAL);
563 		break;
564 	case CRYPTO_CHACHA20_POLY1305:
565 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
566 			return (EINVAL);
567 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
568 		    en->tls_vminor != TLS_MINOR_VER_THREE)
569 			return (EINVAL);
570 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
571 			return (EINVAL);
572 		break;
573 	default:
574 		return (EINVAL);
575 	}
576 
577 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
578 
579 	counter_u64_add(ktls_offload_active, 1);
580 
581 	refcount_init(&tls->refcount, 1);
582 	TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
583 
584 	tls->wq_index = ktls_get_cpu(so);
585 
586 	tls->params.cipher_algorithm = en->cipher_algorithm;
587 	tls->params.auth_algorithm = en->auth_algorithm;
588 	tls->params.tls_vmajor = en->tls_vmajor;
589 	tls->params.tls_vminor = en->tls_vminor;
590 	tls->params.flags = en->flags;
591 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
592 
593 	/* Set the header and trailer lengths. */
594 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
595 	switch (en->cipher_algorithm) {
596 	case CRYPTO_AES_NIST_GCM_16:
597 		/*
598 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
599 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
600 		 */
601 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
602 			tls->params.tls_hlen += sizeof(uint64_t);
603 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
604 		tls->params.tls_bs = 1;
605 		break;
606 	case CRYPTO_AES_CBC:
607 		switch (en->auth_algorithm) {
608 		case CRYPTO_SHA1_HMAC:
609 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
610 				/* Implicit IV, no nonce. */
611 			} else {
612 				tls->params.tls_hlen += AES_BLOCK_LEN;
613 			}
614 			tls->params.tls_tlen = AES_BLOCK_LEN +
615 			    SHA1_HASH_LEN;
616 			break;
617 		case CRYPTO_SHA2_256_HMAC:
618 			tls->params.tls_hlen += AES_BLOCK_LEN;
619 			tls->params.tls_tlen = AES_BLOCK_LEN +
620 			    SHA2_256_HASH_LEN;
621 			break;
622 		case CRYPTO_SHA2_384_HMAC:
623 			tls->params.tls_hlen += AES_BLOCK_LEN;
624 			tls->params.tls_tlen = AES_BLOCK_LEN +
625 			    SHA2_384_HASH_LEN;
626 			break;
627 		default:
628 			panic("invalid hmac");
629 		}
630 		tls->params.tls_bs = AES_BLOCK_LEN;
631 		break;
632 	case CRYPTO_CHACHA20_POLY1305:
633 		/*
634 		 * Chacha20 uses a 12 byte implicit IV.
635 		 */
636 		tls->params.tls_tlen = POLY1305_HASH_LEN;
637 		tls->params.tls_bs = 1;
638 		break;
639 	default:
640 		panic("invalid cipher");
641 	}
642 
643 	/*
644 	 * TLS 1.3 includes optional padding which we do not support,
645 	 * and also puts the "real" record type at the end of the
646 	 * encrypted data.
647 	 */
648 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
649 		tls->params.tls_tlen += sizeof(uint8_t);
650 
651 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
652 	    ("TLS header length too long: %d", tls->params.tls_hlen));
653 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
654 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
655 
656 	if (en->auth_key_len != 0) {
657 		tls->params.auth_key_len = en->auth_key_len;
658 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
659 		    M_WAITOK);
660 		error = copyin(en->auth_key, tls->params.auth_key,
661 		    en->auth_key_len);
662 		if (error)
663 			goto out;
664 	}
665 
666 	tls->params.cipher_key_len = en->cipher_key_len;
667 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
668 	error = copyin(en->cipher_key, tls->params.cipher_key,
669 	    en->cipher_key_len);
670 	if (error)
671 		goto out;
672 
673 	/*
674 	 * This holds the implicit portion of the nonce for AEAD
675 	 * ciphers and the initial implicit IV for TLS 1.0.  The
676 	 * explicit portions of the IV are generated in ktls_frame().
677 	 */
678 	if (en->iv_len != 0) {
679 		tls->params.iv_len = en->iv_len;
680 		error = copyin(en->iv, tls->params.iv, en->iv_len);
681 		if (error)
682 			goto out;
683 
684 		/*
685 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
686 		 * counter to generate unique explicit IVs.
687 		 *
688 		 * Store this counter in the last 8 bytes of the IV
689 		 * array so that it is 8-byte aligned.
690 		 */
691 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
692 		    en->tls_vminor == TLS_MINOR_VER_TWO)
693 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
694 	}
695 
696 	*tlsp = tls;
697 	return (0);
698 
699 out:
700 	ktls_cleanup(tls);
701 	return (error);
702 }
703 
704 static struct ktls_session *
705 ktls_clone_session(struct ktls_session *tls)
706 {
707 	struct ktls_session *tls_new;
708 
709 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
710 
711 	counter_u64_add(ktls_offload_active, 1);
712 
713 	refcount_init(&tls_new->refcount, 1);
714 	TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
715 
716 	/* Copy fields from existing session. */
717 	tls_new->params = tls->params;
718 	tls_new->wq_index = tls->wq_index;
719 
720 	/* Deep copy keys. */
721 	if (tls_new->params.auth_key != NULL) {
722 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
723 		    M_KTLS, M_WAITOK);
724 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
725 		    tls->params.auth_key_len);
726 	}
727 
728 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
729 	    M_WAITOK);
730 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
731 	    tls->params.cipher_key_len);
732 
733 	return (tls_new);
734 }
735 #endif
736 
737 static void
738 ktls_cleanup(struct ktls_session *tls)
739 {
740 
741 	counter_u64_add(ktls_offload_active, -1);
742 	switch (tls->mode) {
743 	case TCP_TLS_MODE_SW:
744 		switch (tls->params.cipher_algorithm) {
745 		case CRYPTO_AES_CBC:
746 			counter_u64_add(ktls_sw_cbc, -1);
747 			break;
748 		case CRYPTO_AES_NIST_GCM_16:
749 			counter_u64_add(ktls_sw_gcm, -1);
750 			break;
751 		case CRYPTO_CHACHA20_POLY1305:
752 			counter_u64_add(ktls_sw_chacha20, -1);
753 			break;
754 		}
755 		ktls_ocf_free(tls);
756 		break;
757 	case TCP_TLS_MODE_IFNET:
758 		switch (tls->params.cipher_algorithm) {
759 		case CRYPTO_AES_CBC:
760 			counter_u64_add(ktls_ifnet_cbc, -1);
761 			break;
762 		case CRYPTO_AES_NIST_GCM_16:
763 			counter_u64_add(ktls_ifnet_gcm, -1);
764 			break;
765 		case CRYPTO_CHACHA20_POLY1305:
766 			counter_u64_add(ktls_ifnet_chacha20, -1);
767 			break;
768 		}
769 		if (tls->snd_tag != NULL)
770 			m_snd_tag_rele(tls->snd_tag);
771 		break;
772 #ifdef TCP_OFFLOAD
773 	case TCP_TLS_MODE_TOE:
774 		switch (tls->params.cipher_algorithm) {
775 		case CRYPTO_AES_CBC:
776 			counter_u64_add(ktls_toe_cbc, -1);
777 			break;
778 		case CRYPTO_AES_NIST_GCM_16:
779 			counter_u64_add(ktls_toe_gcm, -1);
780 			break;
781 		case CRYPTO_CHACHA20_POLY1305:
782 			counter_u64_add(ktls_toe_chacha20, -1);
783 			break;
784 		}
785 		break;
786 #endif
787 	}
788 	if (tls->params.auth_key != NULL) {
789 		zfree(tls->params.auth_key, M_KTLS);
790 		tls->params.auth_key = NULL;
791 		tls->params.auth_key_len = 0;
792 	}
793 	if (tls->params.cipher_key != NULL) {
794 		zfree(tls->params.cipher_key, M_KTLS);
795 		tls->params.cipher_key = NULL;
796 		tls->params.cipher_key_len = 0;
797 	}
798 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
799 }
800 
801 #if defined(INET) || defined(INET6)
802 
803 #ifdef TCP_OFFLOAD
804 static int
805 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
806 {
807 	struct inpcb *inp;
808 	struct tcpcb *tp;
809 	int error;
810 
811 	inp = so->so_pcb;
812 	INP_WLOCK(inp);
813 	if (inp->inp_flags2 & INP_FREED) {
814 		INP_WUNLOCK(inp);
815 		return (ECONNRESET);
816 	}
817 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
818 		INP_WUNLOCK(inp);
819 		return (ECONNRESET);
820 	}
821 	if (inp->inp_socket == NULL) {
822 		INP_WUNLOCK(inp);
823 		return (ECONNRESET);
824 	}
825 	tp = intotcpcb(inp);
826 	if (!(tp->t_flags & TF_TOE)) {
827 		INP_WUNLOCK(inp);
828 		return (EOPNOTSUPP);
829 	}
830 
831 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
832 	INP_WUNLOCK(inp);
833 	if (error == 0) {
834 		tls->mode = TCP_TLS_MODE_TOE;
835 		switch (tls->params.cipher_algorithm) {
836 		case CRYPTO_AES_CBC:
837 			counter_u64_add(ktls_toe_cbc, 1);
838 			break;
839 		case CRYPTO_AES_NIST_GCM_16:
840 			counter_u64_add(ktls_toe_gcm, 1);
841 			break;
842 		case CRYPTO_CHACHA20_POLY1305:
843 			counter_u64_add(ktls_toe_chacha20, 1);
844 			break;
845 		}
846 	}
847 	return (error);
848 }
849 #endif
850 
851 /*
852  * Common code used when first enabling ifnet TLS on a connection or
853  * when allocating a new ifnet TLS session due to a routing change.
854  * This function allocates a new TLS send tag on whatever interface
855  * the connection is currently routed over.
856  */
857 static int
858 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
859     struct m_snd_tag **mstp)
860 {
861 	union if_snd_tag_alloc_params params;
862 	struct ifnet *ifp;
863 	struct nhop_object *nh;
864 	struct tcpcb *tp;
865 	int error;
866 
867 	INP_RLOCK(inp);
868 	if (inp->inp_flags2 & INP_FREED) {
869 		INP_RUNLOCK(inp);
870 		return (ECONNRESET);
871 	}
872 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
873 		INP_RUNLOCK(inp);
874 		return (ECONNRESET);
875 	}
876 	if (inp->inp_socket == NULL) {
877 		INP_RUNLOCK(inp);
878 		return (ECONNRESET);
879 	}
880 	tp = intotcpcb(inp);
881 
882 	/*
883 	 * Check administrative controls on ifnet TLS to determine if
884 	 * ifnet TLS should be denied.
885 	 *
886 	 * - Always permit 'force' requests.
887 	 * - ktls_ifnet_permitted == 0: always deny.
888 	 */
889 	if (!force && ktls_ifnet_permitted == 0) {
890 		INP_RUNLOCK(inp);
891 		return (ENXIO);
892 	}
893 
894 	/*
895 	 * XXX: Use the cached route in the inpcb to find the
896 	 * interface.  This should perhaps instead use
897 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
898 	 * enabled after a connection has completed key negotiation in
899 	 * userland, the cached route will be present in practice.
900 	 */
901 	nh = inp->inp_route.ro_nh;
902 	if (nh == NULL) {
903 		INP_RUNLOCK(inp);
904 		return (ENXIO);
905 	}
906 	ifp = nh->nh_ifp;
907 	if_ref(ifp);
908 
909 	/*
910 	 * Allocate a TLS + ratelimit tag if the connection has an
911 	 * existing pacing rate.
912 	 */
913 	if (tp->t_pacing_rate != -1 &&
914 	    (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
915 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
916 		params.tls_rate_limit.inp = inp;
917 		params.tls_rate_limit.tls = tls;
918 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
919 	} else {
920 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
921 		params.tls.inp = inp;
922 		params.tls.tls = tls;
923 	}
924 	params.hdr.flowid = inp->inp_flowid;
925 	params.hdr.flowtype = inp->inp_flowtype;
926 	params.hdr.numa_domain = inp->inp_numa_domain;
927 	INP_RUNLOCK(inp);
928 
929 	if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
930 		error = EOPNOTSUPP;
931 		goto out;
932 	}
933 	if (inp->inp_vflag & INP_IPV6) {
934 		if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
935 			error = EOPNOTSUPP;
936 			goto out;
937 		}
938 	} else {
939 		if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
940 			error = EOPNOTSUPP;
941 			goto out;
942 		}
943 	}
944 	error = m_snd_tag_alloc(ifp, &params, mstp);
945 out:
946 	if_rele(ifp);
947 	return (error);
948 }
949 
950 static int
951 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
952 {
953 	struct m_snd_tag *mst;
954 	int error;
955 
956 	error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
957 	if (error == 0) {
958 		tls->mode = TCP_TLS_MODE_IFNET;
959 		tls->snd_tag = mst;
960 		switch (tls->params.cipher_algorithm) {
961 		case CRYPTO_AES_CBC:
962 			counter_u64_add(ktls_ifnet_cbc, 1);
963 			break;
964 		case CRYPTO_AES_NIST_GCM_16:
965 			counter_u64_add(ktls_ifnet_gcm, 1);
966 			break;
967 		case CRYPTO_CHACHA20_POLY1305:
968 			counter_u64_add(ktls_ifnet_chacha20, 1);
969 			break;
970 		}
971 	}
972 	return (error);
973 }
974 
975 static int
976 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
977 {
978 	int error;
979 
980 	error = ktls_ocf_try(so, tls, direction);
981 	if (error)
982 		return (error);
983 	tls->mode = TCP_TLS_MODE_SW;
984 	switch (tls->params.cipher_algorithm) {
985 	case CRYPTO_AES_CBC:
986 		counter_u64_add(ktls_sw_cbc, 1);
987 		break;
988 	case CRYPTO_AES_NIST_GCM_16:
989 		counter_u64_add(ktls_sw_gcm, 1);
990 		break;
991 	case CRYPTO_CHACHA20_POLY1305:
992 		counter_u64_add(ktls_sw_chacha20, 1);
993 		break;
994 	}
995 	return (0);
996 }
997 
998 /*
999  * KTLS RX stores data in the socket buffer as a list of TLS records,
1000  * where each record is stored as a control message containg the TLS
1001  * header followed by data mbufs containing the decrypted data.  This
1002  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1003  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1004  * should be queued to the socket buffer as records, but encrypted
1005  * data which needs to be decrypted by software arrives as a stream of
1006  * regular mbufs which need to be converted.  In addition, there may
1007  * already be pending encrypted data in the socket buffer when KTLS RX
1008  * is enabled.
1009  *
1010  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1011  * is used:
1012  *
1013  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1014  *
1015  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1016  *   from the first mbuf.  Once all of the data for that TLS record is
1017  *   queued, the socket is queued to a worker thread.
1018  *
1019  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1020  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1021  *   decrypted, and inserted into the regular socket buffer chain as
1022  *   record starting with a control message holding the TLS header and
1023  *   a chain of mbufs holding the encrypted data.
1024  */
1025 
1026 static void
1027 sb_mark_notready(struct sockbuf *sb)
1028 {
1029 	struct mbuf *m;
1030 
1031 	m = sb->sb_mb;
1032 	sb->sb_mtls = m;
1033 	sb->sb_mb = NULL;
1034 	sb->sb_mbtail = NULL;
1035 	sb->sb_lastrecord = NULL;
1036 	for (; m != NULL; m = m->m_next) {
1037 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1038 		    __func__));
1039 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1040 		    __func__));
1041 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1042 		    __func__));
1043 		m->m_flags |= M_NOTREADY;
1044 		sb->sb_acc -= m->m_len;
1045 		sb->sb_tlscc += m->m_len;
1046 		sb->sb_mtlstail = m;
1047 	}
1048 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1049 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1050 	    sb->sb_ccc));
1051 }
1052 
1053 int
1054 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1055 {
1056 	struct ktls_session *tls;
1057 	int error;
1058 
1059 	if (!ktls_offload_enable)
1060 		return (ENOTSUP);
1061 	if (SOLISTENING(so))
1062 		return (EINVAL);
1063 
1064 	counter_u64_add(ktls_offload_enable_calls, 1);
1065 
1066 	/*
1067 	 * This should always be true since only the TCP socket option
1068 	 * invokes this function.
1069 	 */
1070 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1071 		return (EINVAL);
1072 
1073 	/*
1074 	 * XXX: Don't overwrite existing sessions.  We should permit
1075 	 * this to support rekeying in the future.
1076 	 */
1077 	if (so->so_rcv.sb_tls_info != NULL)
1078 		return (EALREADY);
1079 
1080 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1081 		return (ENOTSUP);
1082 
1083 	/* TLS 1.3 is not yet supported. */
1084 	if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1085 	    en->tls_vminor == TLS_MINOR_VER_THREE)
1086 		return (ENOTSUP);
1087 
1088 	error = ktls_create_session(so, en, &tls);
1089 	if (error)
1090 		return (error);
1091 
1092 #ifdef TCP_OFFLOAD
1093 	error = ktls_try_toe(so, tls, KTLS_RX);
1094 	if (error)
1095 #endif
1096 		error = ktls_try_sw(so, tls, KTLS_RX);
1097 
1098 	if (error) {
1099 		ktls_cleanup(tls);
1100 		return (error);
1101 	}
1102 
1103 	/* Mark the socket as using TLS offload. */
1104 	SOCKBUF_LOCK(&so->so_rcv);
1105 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1106 	so->so_rcv.sb_tls_info = tls;
1107 	so->so_rcv.sb_flags |= SB_TLS_RX;
1108 
1109 	/* Mark existing data as not ready until it can be decrypted. */
1110 	if (tls->mode != TCP_TLS_MODE_TOE) {
1111 		sb_mark_notready(&so->so_rcv);
1112 		ktls_check_rx(&so->so_rcv);
1113 	}
1114 	SOCKBUF_UNLOCK(&so->so_rcv);
1115 
1116 	counter_u64_add(ktls_offload_total, 1);
1117 
1118 	return (0);
1119 }
1120 
1121 int
1122 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1123 {
1124 	struct ktls_session *tls;
1125 	struct inpcb *inp;
1126 	int error;
1127 
1128 	if (!ktls_offload_enable)
1129 		return (ENOTSUP);
1130 	if (SOLISTENING(so))
1131 		return (EINVAL);
1132 
1133 	counter_u64_add(ktls_offload_enable_calls, 1);
1134 
1135 	/*
1136 	 * This should always be true since only the TCP socket option
1137 	 * invokes this function.
1138 	 */
1139 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1140 		return (EINVAL);
1141 
1142 	/*
1143 	 * XXX: Don't overwrite existing sessions.  We should permit
1144 	 * this to support rekeying in the future.
1145 	 */
1146 	if (so->so_snd.sb_tls_info != NULL)
1147 		return (EALREADY);
1148 
1149 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1150 		return (ENOTSUP);
1151 
1152 	/* TLS requires ext pgs */
1153 	if (mb_use_ext_pgs == 0)
1154 		return (ENXIO);
1155 
1156 	error = ktls_create_session(so, en, &tls);
1157 	if (error)
1158 		return (error);
1159 
1160 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1161 #ifdef TCP_OFFLOAD
1162 	error = ktls_try_toe(so, tls, KTLS_TX);
1163 	if (error)
1164 #endif
1165 		error = ktls_try_ifnet(so, tls, false);
1166 	if (error)
1167 		error = ktls_try_sw(so, tls, KTLS_TX);
1168 
1169 	if (error) {
1170 		ktls_cleanup(tls);
1171 		return (error);
1172 	}
1173 
1174 	error = sblock(&so->so_snd, SBL_WAIT);
1175 	if (error) {
1176 		ktls_cleanup(tls);
1177 		return (error);
1178 	}
1179 
1180 	/*
1181 	 * Write lock the INP when setting sb_tls_info so that
1182 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1183 	 * holding the INP lock.
1184 	 */
1185 	inp = so->so_pcb;
1186 	INP_WLOCK(inp);
1187 	SOCKBUF_LOCK(&so->so_snd);
1188 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1189 	so->so_snd.sb_tls_info = tls;
1190 	if (tls->mode != TCP_TLS_MODE_SW)
1191 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1192 	SOCKBUF_UNLOCK(&so->so_snd);
1193 	INP_WUNLOCK(inp);
1194 	sbunlock(&so->so_snd);
1195 
1196 	counter_u64_add(ktls_offload_total, 1);
1197 
1198 	return (0);
1199 }
1200 
1201 int
1202 ktls_get_rx_mode(struct socket *so)
1203 {
1204 	struct ktls_session *tls;
1205 	struct inpcb *inp;
1206 	int mode;
1207 
1208 	if (SOLISTENING(so))
1209 		return (EINVAL);
1210 	inp = so->so_pcb;
1211 	INP_WLOCK_ASSERT(inp);
1212 	SOCKBUF_LOCK(&so->so_rcv);
1213 	tls = so->so_rcv.sb_tls_info;
1214 	if (tls == NULL)
1215 		mode = TCP_TLS_MODE_NONE;
1216 	else
1217 		mode = tls->mode;
1218 	SOCKBUF_UNLOCK(&so->so_rcv);
1219 	return (mode);
1220 }
1221 
1222 int
1223 ktls_get_tx_mode(struct socket *so)
1224 {
1225 	struct ktls_session *tls;
1226 	struct inpcb *inp;
1227 	int mode;
1228 
1229 	if (SOLISTENING(so))
1230 		return (EINVAL);
1231 	inp = so->so_pcb;
1232 	INP_WLOCK_ASSERT(inp);
1233 	SOCKBUF_LOCK(&so->so_snd);
1234 	tls = so->so_snd.sb_tls_info;
1235 	if (tls == NULL)
1236 		mode = TCP_TLS_MODE_NONE;
1237 	else
1238 		mode = tls->mode;
1239 	SOCKBUF_UNLOCK(&so->so_snd);
1240 	return (mode);
1241 }
1242 
1243 /*
1244  * Switch between SW and ifnet TLS sessions as requested.
1245  */
1246 int
1247 ktls_set_tx_mode(struct socket *so, int mode)
1248 {
1249 	struct ktls_session *tls, *tls_new;
1250 	struct inpcb *inp;
1251 	int error;
1252 
1253 	if (SOLISTENING(so))
1254 		return (EINVAL);
1255 	switch (mode) {
1256 	case TCP_TLS_MODE_SW:
1257 	case TCP_TLS_MODE_IFNET:
1258 		break;
1259 	default:
1260 		return (EINVAL);
1261 	}
1262 
1263 	inp = so->so_pcb;
1264 	INP_WLOCK_ASSERT(inp);
1265 	SOCKBUF_LOCK(&so->so_snd);
1266 	tls = so->so_snd.sb_tls_info;
1267 	if (tls == NULL) {
1268 		SOCKBUF_UNLOCK(&so->so_snd);
1269 		return (0);
1270 	}
1271 
1272 	if (tls->mode == mode) {
1273 		SOCKBUF_UNLOCK(&so->so_snd);
1274 		return (0);
1275 	}
1276 
1277 	tls = ktls_hold(tls);
1278 	SOCKBUF_UNLOCK(&so->so_snd);
1279 	INP_WUNLOCK(inp);
1280 
1281 	tls_new = ktls_clone_session(tls);
1282 
1283 	if (mode == TCP_TLS_MODE_IFNET)
1284 		error = ktls_try_ifnet(so, tls_new, true);
1285 	else
1286 		error = ktls_try_sw(so, tls_new, KTLS_TX);
1287 	if (error) {
1288 		counter_u64_add(ktls_switch_failed, 1);
1289 		ktls_free(tls_new);
1290 		ktls_free(tls);
1291 		INP_WLOCK(inp);
1292 		return (error);
1293 	}
1294 
1295 	error = sblock(&so->so_snd, SBL_WAIT);
1296 	if (error) {
1297 		counter_u64_add(ktls_switch_failed, 1);
1298 		ktls_free(tls_new);
1299 		ktls_free(tls);
1300 		INP_WLOCK(inp);
1301 		return (error);
1302 	}
1303 
1304 	/*
1305 	 * If we raced with another session change, keep the existing
1306 	 * session.
1307 	 */
1308 	if (tls != so->so_snd.sb_tls_info) {
1309 		counter_u64_add(ktls_switch_failed, 1);
1310 		sbunlock(&so->so_snd);
1311 		ktls_free(tls_new);
1312 		ktls_free(tls);
1313 		INP_WLOCK(inp);
1314 		return (EBUSY);
1315 	}
1316 
1317 	SOCKBUF_LOCK(&so->so_snd);
1318 	so->so_snd.sb_tls_info = tls_new;
1319 	if (tls_new->mode != TCP_TLS_MODE_SW)
1320 		so->so_snd.sb_flags |= SB_TLS_IFNET;
1321 	SOCKBUF_UNLOCK(&so->so_snd);
1322 	sbunlock(&so->so_snd);
1323 
1324 	/*
1325 	 * Drop two references on 'tls'.  The first is for the
1326 	 * ktls_hold() above.  The second drops the reference from the
1327 	 * socket buffer.
1328 	 */
1329 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1330 	ktls_free(tls);
1331 	ktls_free(tls);
1332 
1333 	if (mode == TCP_TLS_MODE_IFNET)
1334 		counter_u64_add(ktls_switch_to_ifnet, 1);
1335 	else
1336 		counter_u64_add(ktls_switch_to_sw, 1);
1337 
1338 	INP_WLOCK(inp);
1339 	return (0);
1340 }
1341 
1342 /*
1343  * Try to allocate a new TLS send tag.  This task is scheduled when
1344  * ip_output detects a route change while trying to transmit a packet
1345  * holding a TLS record.  If a new tag is allocated, replace the tag
1346  * in the TLS session.  Subsequent packets on the connection will use
1347  * the new tag.  If a new tag cannot be allocated, drop the
1348  * connection.
1349  */
1350 static void
1351 ktls_reset_send_tag(void *context, int pending)
1352 {
1353 	struct epoch_tracker et;
1354 	struct ktls_session *tls;
1355 	struct m_snd_tag *old, *new;
1356 	struct inpcb *inp;
1357 	struct tcpcb *tp;
1358 	int error;
1359 
1360 	MPASS(pending == 1);
1361 
1362 	tls = context;
1363 	inp = tls->inp;
1364 
1365 	/*
1366 	 * Free the old tag first before allocating a new one.
1367 	 * ip[6]_output_send() will treat a NULL send tag the same as
1368 	 * an ifp mismatch and drop packets until a new tag is
1369 	 * allocated.
1370 	 *
1371 	 * Write-lock the INP when changing tls->snd_tag since
1372 	 * ip[6]_output_send() holds a read-lock when reading the
1373 	 * pointer.
1374 	 */
1375 	INP_WLOCK(inp);
1376 	old = tls->snd_tag;
1377 	tls->snd_tag = NULL;
1378 	INP_WUNLOCK(inp);
1379 	if (old != NULL)
1380 		m_snd_tag_rele(old);
1381 
1382 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1383 
1384 	if (error == 0) {
1385 		INP_WLOCK(inp);
1386 		tls->snd_tag = new;
1387 		mtx_pool_lock(mtxpool_sleep, tls);
1388 		tls->reset_pending = false;
1389 		mtx_pool_unlock(mtxpool_sleep, tls);
1390 		if (!in_pcbrele_wlocked(inp))
1391 			INP_WUNLOCK(inp);
1392 
1393 		counter_u64_add(ktls_ifnet_reset, 1);
1394 
1395 		/*
1396 		 * XXX: Should we kick tcp_output explicitly now that
1397 		 * the send tag is fixed or just rely on timers?
1398 		 */
1399 	} else {
1400 		NET_EPOCH_ENTER(et);
1401 		INP_WLOCK(inp);
1402 		if (!in_pcbrele_wlocked(inp)) {
1403 			if (!(inp->inp_flags & INP_TIMEWAIT) &&
1404 			    !(inp->inp_flags & INP_DROPPED)) {
1405 				tp = intotcpcb(inp);
1406 				CURVNET_SET(tp->t_vnet);
1407 				tp = tcp_drop(tp, ECONNABORTED);
1408 				CURVNET_RESTORE();
1409 				if (tp != NULL)
1410 					INP_WUNLOCK(inp);
1411 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1412 			} else
1413 				INP_WUNLOCK(inp);
1414 		}
1415 		NET_EPOCH_EXIT(et);
1416 
1417 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1418 
1419 		/*
1420 		 * Leave reset_pending true to avoid future tasks while
1421 		 * the socket goes away.
1422 		 */
1423 	}
1424 
1425 	ktls_free(tls);
1426 }
1427 
1428 int
1429 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1430 {
1431 
1432 	if (inp == NULL)
1433 		return (ENOBUFS);
1434 
1435 	INP_LOCK_ASSERT(inp);
1436 
1437 	/*
1438 	 * See if we should schedule a task to update the send tag for
1439 	 * this session.
1440 	 */
1441 	mtx_pool_lock(mtxpool_sleep, tls);
1442 	if (!tls->reset_pending) {
1443 		(void) ktls_hold(tls);
1444 		in_pcbref(inp);
1445 		tls->inp = inp;
1446 		tls->reset_pending = true;
1447 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1448 	}
1449 	mtx_pool_unlock(mtxpool_sleep, tls);
1450 	return (ENOBUFS);
1451 }
1452 
1453 #ifdef RATELIMIT
1454 int
1455 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1456 {
1457 	union if_snd_tag_modify_params params = {
1458 		.rate_limit.max_rate = max_pacing_rate,
1459 		.rate_limit.flags = M_NOWAIT,
1460 	};
1461 	struct m_snd_tag *mst;
1462 	struct ifnet *ifp;
1463 
1464 	/* Can't get to the inp, but it should be locked. */
1465 	/* INP_LOCK_ASSERT(inp); */
1466 
1467 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1468 
1469 	if (tls->snd_tag == NULL) {
1470 		/*
1471 		 * Resetting send tag, ignore this change.  The
1472 		 * pending reset may or may not see this updated rate
1473 		 * in the tcpcb.  If it doesn't, we will just lose
1474 		 * this rate change.
1475 		 */
1476 		return (0);
1477 	}
1478 
1479 	MPASS(tls->snd_tag != NULL);
1480 	MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1481 
1482 	mst = tls->snd_tag;
1483 	ifp = mst->ifp;
1484 	return (ifp->if_snd_tag_modify(mst, &params));
1485 }
1486 #endif
1487 #endif
1488 
1489 void
1490 ktls_destroy(struct ktls_session *tls)
1491 {
1492 
1493 	ktls_cleanup(tls);
1494 	uma_zfree(ktls_session_zone, tls);
1495 }
1496 
1497 void
1498 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1499 {
1500 
1501 	for (; m != NULL; m = m->m_next) {
1502 		KASSERT((m->m_flags & M_EXTPG) != 0,
1503 		    ("ktls_seq: mapped mbuf %p", m));
1504 
1505 		m->m_epg_seqno = sb->sb_tls_seqno;
1506 		sb->sb_tls_seqno++;
1507 	}
1508 }
1509 
1510 /*
1511  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1512  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1513  * mbuf must be populated with the payload of each TLS record.
1514  *
1515  * The record_type argument specifies the TLS record type used when
1516  * populating the TLS header.
1517  *
1518  * The enq_count argument on return is set to the number of pages of
1519  * payload data for this entire chain that need to be encrypted via SW
1520  * encryption.  The returned value should be passed to ktls_enqueue
1521  * when scheduling encryption of this chain of mbufs.  To handle the
1522  * special case of empty fragments for TLS 1.0 sessions, an empty
1523  * fragment counts as one page.
1524  */
1525 void
1526 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1527     uint8_t record_type)
1528 {
1529 	struct tls_record_layer *tlshdr;
1530 	struct mbuf *m;
1531 	uint64_t *noncep;
1532 	uint16_t tls_len;
1533 	int maxlen;
1534 
1535 	maxlen = tls->params.max_frame_len;
1536 	*enq_cnt = 0;
1537 	for (m = top; m != NULL; m = m->m_next) {
1538 		/*
1539 		 * All mbufs in the chain should be TLS records whose
1540 		 * payload does not exceed the maximum frame length.
1541 		 *
1542 		 * Empty TLS records are permitted when using CBC.
1543 		 */
1544 		KASSERT(m->m_len <= maxlen &&
1545 		    (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1546 		    m->m_len >= 0 : m->m_len > 0),
1547 		    ("ktls_frame: m %p len %d\n", m, m->m_len));
1548 
1549 		/*
1550 		 * TLS frames require unmapped mbufs to store session
1551 		 * info.
1552 		 */
1553 		KASSERT((m->m_flags & M_EXTPG) != 0,
1554 		    ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1555 
1556 		tls_len = m->m_len;
1557 
1558 		/* Save a reference to the session. */
1559 		m->m_epg_tls = ktls_hold(tls);
1560 
1561 		m->m_epg_hdrlen = tls->params.tls_hlen;
1562 		m->m_epg_trllen = tls->params.tls_tlen;
1563 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1564 			int bs, delta;
1565 
1566 			/*
1567 			 * AES-CBC pads messages to a multiple of the
1568 			 * block size.  Note that the padding is
1569 			 * applied after the digest and the encryption
1570 			 * is done on the "plaintext || mac || padding".
1571 			 * At least one byte of padding is always
1572 			 * present.
1573 			 *
1574 			 * Compute the final trailer length assuming
1575 			 * at most one block of padding.
1576 			 * tls->params.tls_tlen is the maximum
1577 			 * possible trailer length (padding + digest).
1578 			 * delta holds the number of excess padding
1579 			 * bytes if the maximum were used.  Those
1580 			 * extra bytes are removed.
1581 			 */
1582 			bs = tls->params.tls_bs;
1583 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1584 			m->m_epg_trllen -= delta;
1585 		}
1586 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1587 
1588 		/* Populate the TLS header. */
1589 		tlshdr = (void *)m->m_epg_hdr;
1590 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
1591 
1592 		/*
1593 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
1594 		 * of TLS_RLTYPE_APP.
1595 		 */
1596 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1597 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1598 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1599 			tlshdr->tls_type = TLS_RLTYPE_APP;
1600 			/* save the real record type for later */
1601 			m->m_epg_record_type = record_type;
1602 			m->m_epg_trail[0] = record_type;
1603 		} else {
1604 			tlshdr->tls_vminor = tls->params.tls_vminor;
1605 			tlshdr->tls_type = record_type;
1606 		}
1607 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1608 
1609 		/*
1610 		 * Store nonces / explicit IVs after the end of the
1611 		 * TLS header.
1612 		 *
1613 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
1614 		 * from the end of the IV.  The nonce is then
1615 		 * incremented for use by the next record.
1616 		 *
1617 		 * For CBC, a random nonce is inserted for TLS 1.1+.
1618 		 */
1619 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1620 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1621 			noncep = (uint64_t *)(tls->params.iv + 8);
1622 			be64enc(tlshdr + 1, *noncep);
1623 			(*noncep)++;
1624 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1625 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1626 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1627 
1628 		/*
1629 		 * When using SW encryption, mark the mbuf not ready.
1630 		 * It will be marked ready via sbready() after the
1631 		 * record has been encrypted.
1632 		 *
1633 		 * When using ifnet TLS, unencrypted TLS records are
1634 		 * sent down the stack to the NIC.
1635 		 */
1636 		if (tls->mode == TCP_TLS_MODE_SW) {
1637 			m->m_flags |= M_NOTREADY;
1638 			if (__predict_false(tls_len == 0)) {
1639 				/* TLS 1.0 empty fragment. */
1640 				m->m_epg_nrdy = 1;
1641 			} else
1642 				m->m_epg_nrdy = m->m_epg_npgs;
1643 			*enq_cnt += m->m_epg_nrdy;
1644 		}
1645 	}
1646 }
1647 
1648 void
1649 ktls_check_rx(struct sockbuf *sb)
1650 {
1651 	struct tls_record_layer hdr;
1652 	struct ktls_wq *wq;
1653 	struct socket *so;
1654 	bool running;
1655 
1656 	SOCKBUF_LOCK_ASSERT(sb);
1657 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1658 	    __func__, sb));
1659 	so = __containerof(sb, struct socket, so_rcv);
1660 
1661 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
1662 		return;
1663 
1664 	/* Is there enough queued for a TLS header? */
1665 	if (sb->sb_tlscc < sizeof(hdr)) {
1666 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1667 			so->so_error = EMSGSIZE;
1668 		return;
1669 	}
1670 
1671 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1672 
1673 	/* Is the entire record queued? */
1674 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1675 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1676 			so->so_error = EMSGSIZE;
1677 		return;
1678 	}
1679 
1680 	sb->sb_flags |= SB_TLS_RX_RUNNING;
1681 
1682 	soref(so);
1683 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1684 	mtx_lock(&wq->mtx);
1685 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1686 	running = wq->running;
1687 	mtx_unlock(&wq->mtx);
1688 	if (!running)
1689 		wakeup(wq);
1690 	counter_u64_add(ktls_cnt_rx_queued, 1);
1691 }
1692 
1693 static struct mbuf *
1694 ktls_detach_record(struct sockbuf *sb, int len)
1695 {
1696 	struct mbuf *m, *n, *top;
1697 	int remain;
1698 
1699 	SOCKBUF_LOCK_ASSERT(sb);
1700 	MPASS(len <= sb->sb_tlscc);
1701 
1702 	/*
1703 	 * If TLS chain is the exact size of the record,
1704 	 * just grab the whole record.
1705 	 */
1706 	top = sb->sb_mtls;
1707 	if (sb->sb_tlscc == len) {
1708 		sb->sb_mtls = NULL;
1709 		sb->sb_mtlstail = NULL;
1710 		goto out;
1711 	}
1712 
1713 	/*
1714 	 * While it would be nice to use m_split() here, we need
1715 	 * to know exactly what m_split() allocates to update the
1716 	 * accounting, so do it inline instead.
1717 	 */
1718 	remain = len;
1719 	for (m = top; remain > m->m_len; m = m->m_next)
1720 		remain -= m->m_len;
1721 
1722 	/* Easy case: don't have to split 'm'. */
1723 	if (remain == m->m_len) {
1724 		sb->sb_mtls = m->m_next;
1725 		if (sb->sb_mtls == NULL)
1726 			sb->sb_mtlstail = NULL;
1727 		m->m_next = NULL;
1728 		goto out;
1729 	}
1730 
1731 	/*
1732 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
1733 	 * with M_NOWAIT first.
1734 	 */
1735 	n = m_get(M_NOWAIT, MT_DATA);
1736 	if (n == NULL) {
1737 		/*
1738 		 * Use M_WAITOK with socket buffer unlocked.  If
1739 		 * 'sb_mtls' changes while the lock is dropped, return
1740 		 * NULL to force the caller to retry.
1741 		 */
1742 		SOCKBUF_UNLOCK(sb);
1743 
1744 		n = m_get(M_WAITOK, MT_DATA);
1745 
1746 		SOCKBUF_LOCK(sb);
1747 		if (sb->sb_mtls != top) {
1748 			m_free(n);
1749 			return (NULL);
1750 		}
1751 	}
1752 	n->m_flags |= M_NOTREADY;
1753 
1754 	/* Store remainder in 'n'. */
1755 	n->m_len = m->m_len - remain;
1756 	if (m->m_flags & M_EXT) {
1757 		n->m_data = m->m_data + remain;
1758 		mb_dupcl(n, m);
1759 	} else {
1760 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1761 	}
1762 
1763 	/* Trim 'm' and update accounting. */
1764 	m->m_len -= n->m_len;
1765 	sb->sb_tlscc -= n->m_len;
1766 	sb->sb_ccc -= n->m_len;
1767 
1768 	/* Account for 'n'. */
1769 	sballoc_ktls_rx(sb, n);
1770 
1771 	/* Insert 'n' into the TLS chain. */
1772 	sb->sb_mtls = n;
1773 	n->m_next = m->m_next;
1774 	if (sb->sb_mtlstail == m)
1775 		sb->sb_mtlstail = n;
1776 
1777 	/* Detach the record from the TLS chain. */
1778 	m->m_next = NULL;
1779 
1780 out:
1781 	MPASS(m_length(top, NULL) == len);
1782 	for (m = top; m != NULL; m = m->m_next)
1783 		sbfree_ktls_rx(sb, m);
1784 	sb->sb_tlsdcc = len;
1785 	sb->sb_ccc += len;
1786 	SBCHECK(sb);
1787 	return (top);
1788 }
1789 
1790 static void
1791 ktls_decrypt(struct socket *so)
1792 {
1793 	char tls_header[MBUF_PEXT_HDR_LEN];
1794 	struct ktls_session *tls;
1795 	struct sockbuf *sb;
1796 	struct tls_record_layer *hdr;
1797 	struct tls_get_record tgr;
1798 	struct mbuf *control, *data, *m;
1799 	uint64_t seqno;
1800 	int error, remain, tls_len, trail_len;
1801 
1802 	hdr = (struct tls_record_layer *)tls_header;
1803 	sb = &so->so_rcv;
1804 	SOCKBUF_LOCK(sb);
1805 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1806 	    ("%s: socket %p not running", __func__, so));
1807 
1808 	tls = sb->sb_tls_info;
1809 	MPASS(tls != NULL);
1810 
1811 	for (;;) {
1812 		/* Is there enough queued for a TLS header? */
1813 		if (sb->sb_tlscc < tls->params.tls_hlen)
1814 			break;
1815 
1816 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1817 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1818 
1819 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1820 		    hdr->tls_vminor != tls->params.tls_vminor)
1821 			error = EINVAL;
1822 		else if (tls_len < tls->params.tls_hlen || tls_len >
1823 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1824 		    tls->params.tls_tlen)
1825 			error = EMSGSIZE;
1826 		else
1827 			error = 0;
1828 		if (__predict_false(error != 0)) {
1829 			/*
1830 			 * We have a corrupted record and are likely
1831 			 * out of sync.  The connection isn't
1832 			 * recoverable at this point, so abort it.
1833 			 */
1834 			SOCKBUF_UNLOCK(sb);
1835 			counter_u64_add(ktls_offload_corrupted_records, 1);
1836 
1837 			CURVNET_SET(so->so_vnet);
1838 			so->so_proto->pr_usrreqs->pru_abort(so);
1839 			so->so_error = error;
1840 			CURVNET_RESTORE();
1841 			goto deref;
1842 		}
1843 
1844 		/* Is the entire record queued? */
1845 		if (sb->sb_tlscc < tls_len)
1846 			break;
1847 
1848 		/*
1849 		 * Split out the portion of the mbuf chain containing
1850 		 * this TLS record.
1851 		 */
1852 		data = ktls_detach_record(sb, tls_len);
1853 		if (data == NULL)
1854 			continue;
1855 		MPASS(sb->sb_tlsdcc == tls_len);
1856 
1857 		seqno = sb->sb_tls_seqno;
1858 		sb->sb_tls_seqno++;
1859 		SBCHECK(sb);
1860 		SOCKBUF_UNLOCK(sb);
1861 
1862 		error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1863 		if (error) {
1864 			counter_u64_add(ktls_offload_failed_crypto, 1);
1865 
1866 			SOCKBUF_LOCK(sb);
1867 			if (sb->sb_tlsdcc == 0) {
1868 				/*
1869 				 * sbcut/drop/flush discarded these
1870 				 * mbufs.
1871 				 */
1872 				m_freem(data);
1873 				break;
1874 			}
1875 
1876 			/*
1877 			 * Drop this TLS record's data, but keep
1878 			 * decrypting subsequent records.
1879 			 */
1880 			sb->sb_ccc -= tls_len;
1881 			sb->sb_tlsdcc = 0;
1882 
1883 			CURVNET_SET(so->so_vnet);
1884 			so->so_error = EBADMSG;
1885 			sorwakeup_locked(so);
1886 			CURVNET_RESTORE();
1887 
1888 			m_freem(data);
1889 
1890 			SOCKBUF_LOCK(sb);
1891 			continue;
1892 		}
1893 
1894 		/* Allocate the control mbuf. */
1895 		tgr.tls_type = hdr->tls_type;
1896 		tgr.tls_vmajor = hdr->tls_vmajor;
1897 		tgr.tls_vminor = hdr->tls_vminor;
1898 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1899 		    trail_len);
1900 		control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1901 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1902 
1903 		SOCKBUF_LOCK(sb);
1904 		if (sb->sb_tlsdcc == 0) {
1905 			/* sbcut/drop/flush discarded these mbufs. */
1906 			MPASS(sb->sb_tlscc == 0);
1907 			m_freem(data);
1908 			m_freem(control);
1909 			break;
1910 		}
1911 
1912 		/*
1913 		 * Clear the 'dcc' accounting in preparation for
1914 		 * adding the decrypted record.
1915 		 */
1916 		sb->sb_ccc -= tls_len;
1917 		sb->sb_tlsdcc = 0;
1918 		SBCHECK(sb);
1919 
1920 		/* If there is no payload, drop all of the data. */
1921 		if (tgr.tls_length == htobe16(0)) {
1922 			m_freem(data);
1923 			data = NULL;
1924 		} else {
1925 			/* Trim header. */
1926 			remain = tls->params.tls_hlen;
1927 			while (remain > 0) {
1928 				if (data->m_len > remain) {
1929 					data->m_data += remain;
1930 					data->m_len -= remain;
1931 					break;
1932 				}
1933 				remain -= data->m_len;
1934 				data = m_free(data);
1935 			}
1936 
1937 			/* Trim trailer and clear M_NOTREADY. */
1938 			remain = be16toh(tgr.tls_length);
1939 			m = data;
1940 			for (m = data; remain > m->m_len; m = m->m_next) {
1941 				m->m_flags &= ~M_NOTREADY;
1942 				remain -= m->m_len;
1943 			}
1944 			m->m_len = remain;
1945 			m_freem(m->m_next);
1946 			m->m_next = NULL;
1947 			m->m_flags &= ~M_NOTREADY;
1948 
1949 			/* Set EOR on the final mbuf. */
1950 			m->m_flags |= M_EOR;
1951 		}
1952 
1953 		sbappendcontrol_locked(sb, data, control, 0);
1954 	}
1955 
1956 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1957 
1958 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1959 		so->so_error = EMSGSIZE;
1960 
1961 	sorwakeup_locked(so);
1962 
1963 deref:
1964 	SOCKBUF_UNLOCK_ASSERT(sb);
1965 
1966 	CURVNET_SET(so->so_vnet);
1967 	SOCK_LOCK(so);
1968 	sorele(so);
1969 	CURVNET_RESTORE();
1970 }
1971 
1972 void
1973 ktls_enqueue_to_free(struct mbuf *m)
1974 {
1975 	struct ktls_wq *wq;
1976 	bool running;
1977 
1978 	/* Mark it for freeing. */
1979 	m->m_epg_flags |= EPG_FLAG_2FREE;
1980 	wq = &ktls_wq[m->m_epg_tls->wq_index];
1981 	mtx_lock(&wq->mtx);
1982 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1983 	running = wq->running;
1984 	mtx_unlock(&wq->mtx);
1985 	if (!running)
1986 		wakeup(wq);
1987 }
1988 
1989 static void *
1990 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
1991 {
1992 	void *buf;
1993 	int domain, running;
1994 
1995 	if (m->m_epg_npgs <= 2)
1996 		return (NULL);
1997 	if (ktls_buffer_zone == NULL)
1998 		return (NULL);
1999 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2000 		/*
2001 		 * Rate-limit allocation attempts after a failure.
2002 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2003 		 * the free page queues and may fail consistently if memory is
2004 		 * fragmented.
2005 		 */
2006 		return (NULL);
2007 	}
2008 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2009 	if (buf == NULL) {
2010 		domain = PCPU_GET(domain);
2011 		wq->lastallocfail = ticks;
2012 
2013 		/*
2014 		 * Note that this check is "racy", but the races are
2015 		 * harmless, and are either a spurious wakeup if
2016 		 * multiple threads fail allocations before the alloc
2017 		 * thread wakes, or waiting an extra second in case we
2018 		 * see an old value of running == true.
2019 		 */
2020 		if (!VM_DOMAIN_EMPTY(domain)) {
2021 			running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2022 			if (!running)
2023 				wakeup(&ktls_domains[domain].alloc_td);
2024 		}
2025 	}
2026 	return (buf);
2027 }
2028 
2029 static int
2030 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2031     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2032 {
2033 	vm_page_t pg;
2034 	int error, i, len, off;
2035 
2036 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2037 	    ("%p not unready & nomap mbuf\n", m));
2038 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2039 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2040 	    ktls_maxlen));
2041 
2042 	/* Anonymous mbufs are encrypted in place. */
2043 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2044 		return (tls->sw_encrypt(state, tls, m, NULL, 0));
2045 
2046 	/*
2047 	 * For file-backed mbufs (from sendfile), anonymous wired
2048 	 * pages are allocated and used as the encryption destination.
2049 	 */
2050 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2051 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2052 		    m->m_epg_1st_off;
2053 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2054 		    m->m_epg_1st_off;
2055 		state->dst_iov[0].iov_len = len;
2056 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2057 		i = 1;
2058 	} else {
2059 		off = m->m_epg_1st_off;
2060 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2061 			do {
2062 				pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2063 				    VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP |
2064 				    VM_ALLOC_WIRED | VM_ALLOC_WAITFAIL);
2065 			} while (pg == NULL);
2066 
2067 			len = m_epg_pagelen(m, i, off);
2068 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2069 			state->dst_iov[i].iov_base =
2070 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2071 			state->dst_iov[i].iov_len = len;
2072 		}
2073 	}
2074 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2075 	state->dst_iov[i].iov_base = m->m_epg_trail;
2076 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2077 
2078 	error = tls->sw_encrypt(state, tls, m, state->dst_iov, i + 1);
2079 
2080 	if (__predict_false(error != 0)) {
2081 		/* Free the anonymous pages. */
2082 		if (state->cbuf != NULL)
2083 			uma_zfree(ktls_buffer_zone, state->cbuf);
2084 		else {
2085 			for (i = 0; i < m->m_epg_npgs; i++) {
2086 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2087 				(void)vm_page_unwire_noq(pg);
2088 				vm_page_free(pg);
2089 			}
2090 		}
2091 	}
2092 	return (error);
2093 }
2094 
2095 void
2096 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2097 {
2098 	struct ktls_wq *wq;
2099 	bool running;
2100 
2101 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2102 	    (M_EXTPG | M_NOTREADY)),
2103 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2104 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2105 
2106 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2107 
2108 	m->m_epg_enc_cnt = page_count;
2109 
2110 	/*
2111 	 * Save a pointer to the socket.  The caller is responsible
2112 	 * for taking an additional reference via soref().
2113 	 */
2114 	m->m_epg_so = so;
2115 
2116 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2117 	mtx_lock(&wq->mtx);
2118 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2119 	running = wq->running;
2120 	mtx_unlock(&wq->mtx);
2121 	if (!running)
2122 		wakeup(wq);
2123 	counter_u64_add(ktls_cnt_tx_queued, 1);
2124 }
2125 
2126 /*
2127  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2128  * the pages from the file and replace them with the anonymous pages
2129  * allocated in ktls_encrypt_record().
2130  */
2131 static void
2132 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2133 {
2134 	int i;
2135 
2136 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2137 
2138 	/* Free the old pages. */
2139 	m->m_ext.ext_free(m);
2140 
2141 	/* Replace them with the new pages. */
2142 	if (state->cbuf != NULL) {
2143 		for (i = 0; i < m->m_epg_npgs; i++)
2144 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2145 
2146 		/* Contig pages should go back to the cache. */
2147 		m->m_ext.ext_free = ktls_free_mext_contig;
2148 	} else {
2149 		for (i = 0; i < m->m_epg_npgs; i++)
2150 			m->m_epg_pa[i] = state->parray[i];
2151 
2152 		/* Use the basic free routine. */
2153 		m->m_ext.ext_free = mb_free_mext_pgs;
2154 	}
2155 
2156 	/* Pages are now writable. */
2157 	m->m_epg_flags |= EPG_FLAG_ANON;
2158 }
2159 
2160 static __noinline void
2161 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2162 {
2163 	struct ktls_ocf_encrypt_state state;
2164 	struct ktls_session *tls;
2165 	struct socket *so;
2166 	struct mbuf *m;
2167 	int error, npages, total_pages;
2168 
2169 	so = top->m_epg_so;
2170 	tls = top->m_epg_tls;
2171 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2172 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2173 #ifdef INVARIANTS
2174 	top->m_epg_so = NULL;
2175 #endif
2176 	total_pages = top->m_epg_enc_cnt;
2177 	npages = 0;
2178 
2179 	/*
2180 	 * Encrypt the TLS records in the chain of mbufs starting with
2181 	 * 'top'.  'total_pages' gives us a total count of pages and is
2182 	 * used to know when we have finished encrypting the TLS
2183 	 * records originally queued with 'top'.
2184 	 *
2185 	 * NB: These mbufs are queued in the socket buffer and
2186 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2187 	 * socket buffer lock is not held while traversing this chain.
2188 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2189 	 * pointers should be stable.  However, the 'm_next' of the
2190 	 * last mbuf encrypted is not necessarily NULL.  It can point
2191 	 * to other mbufs appended while 'top' was on the TLS work
2192 	 * queue.
2193 	 *
2194 	 * Each mbuf holds an entire TLS record.
2195 	 */
2196 	error = 0;
2197 	for (m = top; npages != total_pages; m = m->m_next) {
2198 		KASSERT(m->m_epg_tls == tls,
2199 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2200 		    tls, m->m_epg_tls));
2201 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2202 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2203 		    total_pages, m));
2204 
2205 		error = ktls_encrypt_record(wq, m, tls, &state);
2206 		if (error) {
2207 			counter_u64_add(ktls_offload_failed_crypto, 1);
2208 			break;
2209 		}
2210 
2211 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2212 			ktls_finish_nonanon(m, &state);
2213 
2214 		npages += m->m_epg_nrdy;
2215 
2216 		/*
2217 		 * Drop a reference to the session now that it is no
2218 		 * longer needed.  Existing code depends on encrypted
2219 		 * records having no associated session vs
2220 		 * yet-to-be-encrypted records having an associated
2221 		 * session.
2222 		 */
2223 		m->m_epg_tls = NULL;
2224 		ktls_free(tls);
2225 	}
2226 
2227 	CURVNET_SET(so->so_vnet);
2228 	if (error == 0) {
2229 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2230 	} else {
2231 		so->so_proto->pr_usrreqs->pru_abort(so);
2232 		so->so_error = EIO;
2233 		mb_free_notready(top, total_pages);
2234 	}
2235 
2236 	SOCK_LOCK(so);
2237 	sorele(so);
2238 	CURVNET_RESTORE();
2239 }
2240 
2241 void
2242 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2243 {
2244 	struct ktls_session *tls;
2245 	struct socket *so;
2246 	struct mbuf *m;
2247 	int npages;
2248 
2249 	m = state->m;
2250 
2251 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2252 		ktls_finish_nonanon(m, state);
2253 
2254 	so = state->so;
2255 	free(state, M_KTLS);
2256 
2257 	/*
2258 	 * Drop a reference to the session now that it is no longer
2259 	 * needed.  Existing code depends on encrypted records having
2260 	 * no associated session vs yet-to-be-encrypted records having
2261 	 * an associated session.
2262 	 */
2263 	tls = m->m_epg_tls;
2264 	m->m_epg_tls = NULL;
2265 	ktls_free(tls);
2266 
2267 	if (error != 0)
2268 		counter_u64_add(ktls_offload_failed_crypto, 1);
2269 
2270 	CURVNET_SET(so->so_vnet);
2271 	npages = m->m_epg_nrdy;
2272 
2273 	if (error == 0) {
2274 		(void)(*so->so_proto->pr_usrreqs->pru_ready)(so, m, npages);
2275 	} else {
2276 		so->so_proto->pr_usrreqs->pru_abort(so);
2277 		so->so_error = EIO;
2278 		mb_free_notready(m, npages);
2279 	}
2280 
2281 	SOCK_LOCK(so);
2282 	sorele(so);
2283 	CURVNET_RESTORE();
2284 }
2285 
2286 /*
2287  * Similar to ktls_encrypt, but used with asynchronous OCF backends
2288  * (coprocessors) where encryption does not use host CPU resources and
2289  * it can be beneficial to queue more requests than CPUs.
2290  */
2291 static __noinline void
2292 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2293 {
2294 	struct ktls_ocf_encrypt_state *state;
2295 	struct ktls_session *tls;
2296 	struct socket *so;
2297 	struct mbuf *m, *n;
2298 	int error, mpages, npages, total_pages;
2299 
2300 	so = top->m_epg_so;
2301 	tls = top->m_epg_tls;
2302 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2303 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2304 #ifdef INVARIANTS
2305 	top->m_epg_so = NULL;
2306 #endif
2307 	total_pages = top->m_epg_enc_cnt;
2308 	npages = 0;
2309 
2310 	error = 0;
2311 	for (m = top; npages != total_pages; m = n) {
2312 		KASSERT(m->m_epg_tls == tls,
2313 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2314 		    tls, m->m_epg_tls));
2315 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2316 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2317 		    total_pages, m));
2318 
2319 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2320 		soref(so);
2321 		state->so = so;
2322 		state->m = m;
2323 
2324 		mpages = m->m_epg_nrdy;
2325 		n = m->m_next;
2326 
2327 		error = ktls_encrypt_record(wq, m, tls, state);
2328 		if (error) {
2329 			counter_u64_add(ktls_offload_failed_crypto, 1);
2330 			free(state, M_KTLS);
2331 			CURVNET_SET(so->so_vnet);
2332 			SOCK_LOCK(so);
2333 			sorele(so);
2334 			CURVNET_RESTORE();
2335 			break;
2336 		}
2337 
2338 		npages += mpages;
2339 	}
2340 
2341 	CURVNET_SET(so->so_vnet);
2342 	if (error != 0) {
2343 		so->so_proto->pr_usrreqs->pru_abort(so);
2344 		so->so_error = EIO;
2345 		mb_free_notready(m, total_pages - npages);
2346 	}
2347 
2348 	SOCK_LOCK(so);
2349 	sorele(so);
2350 	CURVNET_RESTORE();
2351 }
2352 
2353 static void
2354 ktls_alloc_thread(void *ctx)
2355 {
2356 	struct ktls_domain_info *ktls_domain = ctx;
2357 	struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
2358 	void **buf;
2359 	struct sysctl_oid *oid;
2360 	char name[80];
2361 	int i, nbufs;
2362 
2363 	curthread->td_domain.dr_policy =
2364 	    DOMAINSET_PREF(PCPU_GET(domain));
2365 	snprintf(name, sizeof(name), "domain%d", PCPU_GET(domain));
2366 	if (bootverbose)
2367 		printf("Starting KTLS alloc thread for domain %d\n",
2368 		    PCPU_GET(domain));
2369 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
2370 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2371 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
2372 	    CTLFLAG_RD,  &sc->allocs, 0, "buffers allocated");
2373 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
2374 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
2375 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
2376 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
2377 
2378 	buf = NULL;
2379 	nbufs = 0;
2380 	for (;;) {
2381 		atomic_store_int(&sc->running, 0);
2382 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
2383 		atomic_store_int(&sc->running, 1);
2384 		sc->wakeups++;
2385 		if (nbufs != ktls_max_alloc) {
2386 			free(buf, M_KTLS);
2387 			nbufs = atomic_load_int(&ktls_max_alloc);
2388 			buf = malloc(sizeof(void *) * nbufs, M_KTLS,
2389 			    M_WAITOK | M_ZERO);
2390 		}
2391 		/*
2392 		 * Below we allocate nbufs with different allocation
2393 		 * flags than we use when allocating normally during
2394 		 * encryption in the ktls worker thread.  We specify
2395 		 * M_NORECLAIM in the worker thread. However, we omit
2396 		 * that flag here and add M_WAITOK so that the VM
2397 		 * system is permitted to perform expensive work to
2398 		 * defragment memory.  We do this here, as it does not
2399 		 * matter if this thread blocks.  If we block a ktls
2400 		 * worker thread, we risk developing backlogs of
2401 		 * buffers to be encrypted, leading to surges of
2402 		 * traffic and potential NIC output drops.
2403 		 */
2404 		for (i = 0; i < nbufs; i++) {
2405 			buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
2406 			sc->allocs++;
2407 		}
2408 		for (i = 0; i < nbufs; i++) {
2409 			uma_zfree(ktls_buffer_zone, buf[i]);
2410 			buf[i] = NULL;
2411 		}
2412 	}
2413 }
2414 
2415 static void
2416 ktls_work_thread(void *ctx)
2417 {
2418 	struct ktls_wq *wq = ctx;
2419 	struct mbuf *m, *n;
2420 	struct socket *so, *son;
2421 	STAILQ_HEAD(, mbuf) local_m_head;
2422 	STAILQ_HEAD(, socket) local_so_head;
2423 
2424 	if (ktls_bind_threads > 1) {
2425 		curthread->td_domain.dr_policy =
2426 			DOMAINSET_PREF(PCPU_GET(domain));
2427 	}
2428 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2429 	fpu_kern_thread(0);
2430 #endif
2431 	for (;;) {
2432 		mtx_lock(&wq->mtx);
2433 		while (STAILQ_EMPTY(&wq->m_head) &&
2434 		    STAILQ_EMPTY(&wq->so_head)) {
2435 			wq->running = false;
2436 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2437 			wq->running = true;
2438 		}
2439 
2440 		STAILQ_INIT(&local_m_head);
2441 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
2442 		STAILQ_INIT(&local_so_head);
2443 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
2444 		mtx_unlock(&wq->mtx);
2445 
2446 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2447 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
2448 				ktls_free(m->m_epg_tls);
2449 				m_free_raw(m);
2450 			} else {
2451 				if (m->m_epg_tls->sync_dispatch)
2452 					ktls_encrypt(wq, m);
2453 				else
2454 					ktls_encrypt_async(wq, m);
2455 				counter_u64_add(ktls_cnt_tx_queued, -1);
2456 			}
2457 		}
2458 
2459 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2460 			ktls_decrypt(so);
2461 			counter_u64_add(ktls_cnt_rx_queued, -1);
2462 		}
2463 	}
2464 }
2465 
2466 #if defined(INET) || defined(INET6)
2467 static void
2468 ktls_disable_ifnet_help(void *context, int pending __unused)
2469 {
2470 	struct ktls_session *tls;
2471 	struct inpcb *inp;
2472 	struct tcpcb *tp;
2473 	struct socket *so;
2474 	int err;
2475 
2476 	tls = context;
2477 	inp = tls->inp;
2478 	if (inp == NULL)
2479 		return;
2480 	INP_WLOCK(inp);
2481 	so = inp->inp_socket;
2482 	MPASS(so != NULL);
2483 	if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
2484 	    (inp->inp_flags2 & INP_FREED)) {
2485 		goto out;
2486 	}
2487 
2488 	if (so->so_snd.sb_tls_info != NULL)
2489 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
2490 	else
2491 		err = ENXIO;
2492 	if (err == 0) {
2493 		counter_u64_add(ktls_ifnet_disable_ok, 1);
2494 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
2495 		if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 &&
2496 		    (inp->inp_flags2 & INP_FREED) == 0 &&
2497 		    (tp = intotcpcb(inp)) != NULL &&
2498 		    tp->t_fb->tfb_hwtls_change != NULL)
2499 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
2500 	} else {
2501 		counter_u64_add(ktls_ifnet_disable_fail, 1);
2502 	}
2503 
2504 out:
2505 	SOCK_LOCK(so);
2506 	sorele(so);
2507 	if (!in_pcbrele_wlocked(inp))
2508 		INP_WUNLOCK(inp);
2509 	ktls_free(tls);
2510 }
2511 
2512 /*
2513  * Called when re-transmits are becoming a substantial portion of the
2514  * sends on this connection.  When this happens, we transition the
2515  * connection to software TLS.  This is needed because most inline TLS
2516  * NICs keep crypto state only for in-order transmits.  This means
2517  * that to handle a TCP rexmit (which is out-of-order), the NIC must
2518  * re-DMA the entire TLS record up to and including the current
2519  * segment.  This means that when re-transmitting the last ~1448 byte
2520  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
2521  * of magnitude more data than we are sending.  This can cause the
2522  * PCIe link to saturate well before the network, which can cause
2523  * output drops, and a general loss of capacity.
2524  */
2525 void
2526 ktls_disable_ifnet(void *arg)
2527 {
2528 	struct tcpcb *tp;
2529 	struct inpcb *inp;
2530 	struct socket *so;
2531 	struct ktls_session *tls;
2532 
2533 	tp = arg;
2534 	inp = tp->t_inpcb;
2535 	INP_WLOCK_ASSERT(inp);
2536 	so = inp->inp_socket;
2537 	SOCK_LOCK(so);
2538 	tls = so->so_snd.sb_tls_info;
2539 	if (tls->disable_ifnet_pending) {
2540 		SOCK_UNLOCK(so);
2541 		return;
2542 	}
2543 
2544 	/*
2545 	 * note that disable_ifnet_pending is never cleared; disabling
2546 	 * ifnet can only be done once per session, so we never want
2547 	 * to do it again
2548 	 */
2549 
2550 	(void)ktls_hold(tls);
2551 	in_pcbref(inp);
2552 	soref(so);
2553 	tls->disable_ifnet_pending = true;
2554 	tls->inp = inp;
2555 	SOCK_UNLOCK(so);
2556 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
2557 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
2558 }
2559 #endif
2560