xref: /freebsd/sys/kern/uipc_ktls.c (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
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 ktls_session *tls, int direction)
1157 {
1158 	int error;
1159 
1160 	error = ktls_ocf_try(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(tls, KTLS_RX);
1318 	if (error) {
1319 		ktls_free(tls);
1320 		return (error);
1321 	}
1322 
1323 	/*
1324 	 * Serialize with soreceive_generic() and make sure that we're not
1325 	 * operating on a listening socket.
1326 	 */
1327 	error = SOCK_IO_RECV_LOCK(so, SBL_WAIT);
1328 	if (error) {
1329 		ktls_free(tls);
1330 		return (error);
1331 	}
1332 
1333 	/* Mark the socket as using TLS offload. */
1334 	SOCK_RECVBUF_LOCK(so);
1335 	if (__predict_false(so->so_rcv.sb_tls_info != NULL)) {
1336 		SOCK_RECVBUF_UNLOCK(so);
1337 		SOCK_IO_RECV_UNLOCK(so);
1338 		ktls_free(tls);
1339 		return (EALREADY);
1340 	}
1341 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1342 	so->so_rcv.sb_tls_info = tls;
1343 	so->so_rcv.sb_flags |= SB_TLS_RX;
1344 
1345 	/* Mark existing data as not ready until it can be decrypted. */
1346 	sb_mark_notready(&so->so_rcv);
1347 	ktls_check_rx(&so->so_rcv);
1348 	SOCK_RECVBUF_UNLOCK(so);
1349 	SOCK_IO_RECV_UNLOCK(so);
1350 
1351 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1352 #ifdef TCP_OFFLOAD
1353 	error = ktls_try_toe(so, tls, KTLS_RX);
1354 	if (error)
1355 #endif
1356 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1357 	if (error)
1358 		ktls_use_sw(tls);
1359 
1360 	counter_u64_add(ktls_offload_total, 1);
1361 
1362 	return (0);
1363 }
1364 
1365 int
1366 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1367 {
1368 	struct ktls_session *tls;
1369 	struct inpcb *inp;
1370 	struct tcpcb *tp;
1371 	int error;
1372 
1373 	if (!ktls_offload_enable)
1374 		return (ENOTSUP);
1375 
1376 	counter_u64_add(ktls_offload_enable_calls, 1);
1377 
1378 	/*
1379 	 * This should always be true since only the TCP socket option
1380 	 * invokes this function.
1381 	 */
1382 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1383 		return (EINVAL);
1384 
1385 	/*
1386 	 * XXX: Don't overwrite existing sessions.  We should permit
1387 	 * this to support rekeying in the future.
1388 	 */
1389 	if (so->so_snd.sb_tls_info != NULL)
1390 		return (EALREADY);
1391 
1392 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1393 		return (ENOTSUP);
1394 
1395 	/* TLS requires ext pgs */
1396 	if (mb_use_ext_pgs == 0)
1397 		return (ENXIO);
1398 
1399 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1400 	if (error)
1401 		return (error);
1402 
1403 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1404 #ifdef TCP_OFFLOAD
1405 	error = ktls_try_toe(so, tls, KTLS_TX);
1406 	if (error)
1407 #endif
1408 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1409 	if (error)
1410 		error = ktls_try_sw(tls, KTLS_TX);
1411 
1412 	if (error) {
1413 		ktls_free(tls);
1414 		return (error);
1415 	}
1416 
1417 	/*
1418 	 * Serialize with sosend_generic() and make sure that we're not
1419 	 * operating on a listening socket.
1420 	 */
1421 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1422 	if (error) {
1423 		ktls_free(tls);
1424 		return (error);
1425 	}
1426 
1427 	/*
1428 	 * Write lock the INP when setting sb_tls_info so that
1429 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1430 	 * holding the INP lock.
1431 	 */
1432 	inp = so->so_pcb;
1433 	INP_WLOCK(inp);
1434 	SOCK_SENDBUF_LOCK(so);
1435 	if (__predict_false(so->so_snd.sb_tls_info != NULL)) {
1436 		SOCK_SENDBUF_UNLOCK(so);
1437 		INP_WUNLOCK(inp);
1438 		SOCK_IO_SEND_UNLOCK(so);
1439 		ktls_free(tls);
1440 		return (EALREADY);
1441 	}
1442 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1443 	so->so_snd.sb_tls_info = tls;
1444 	if (tls->mode != TCP_TLS_MODE_SW) {
1445 		tp = intotcpcb(inp);
1446 		MPASS(tp->t_nic_ktls_xmit == 0);
1447 		tp->t_nic_ktls_xmit = 1;
1448 		if (tp->t_fb->tfb_hwtls_change != NULL)
1449 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1450 	}
1451 	SOCK_SENDBUF_UNLOCK(so);
1452 	INP_WUNLOCK(inp);
1453 	SOCK_IO_SEND_UNLOCK(so);
1454 
1455 	counter_u64_add(ktls_offload_total, 1);
1456 
1457 	return (0);
1458 }
1459 
1460 int
1461 ktls_get_rx_mode(struct socket *so, int *modep)
1462 {
1463 	struct ktls_session *tls;
1464 	struct inpcb *inp __diagused;
1465 
1466 	if (SOLISTENING(so))
1467 		return (EINVAL);
1468 	inp = so->so_pcb;
1469 	INP_WLOCK_ASSERT(inp);
1470 	SOCK_RECVBUF_LOCK(so);
1471 	tls = so->so_rcv.sb_tls_info;
1472 	if (tls == NULL)
1473 		*modep = TCP_TLS_MODE_NONE;
1474 	else
1475 		*modep = tls->mode;
1476 	SOCK_RECVBUF_UNLOCK(so);
1477 	return (0);
1478 }
1479 
1480 /*
1481  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1482  *
1483  * This function gets information about the next TCP- and TLS-
1484  * sequence number to be processed by the TLS receive worker
1485  * thread. The information is extracted from the given "inpcb"
1486  * structure. The values are stored in host endian format at the two
1487  * given output pointer locations. The TCP sequence number points to
1488  * the beginning of the TLS header.
1489  *
1490  * This function returns zero on success, else a non-zero error code
1491  * is returned.
1492  */
1493 int
1494 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1495 {
1496 	struct socket *so;
1497 	struct tcpcb *tp;
1498 
1499 	INP_RLOCK(inp);
1500 	so = inp->inp_socket;
1501 	if (__predict_false(so == NULL)) {
1502 		INP_RUNLOCK(inp);
1503 		return (EINVAL);
1504 	}
1505 	if (inp->inp_flags & INP_DROPPED) {
1506 		INP_RUNLOCK(inp);
1507 		return (ECONNRESET);
1508 	}
1509 
1510 	tp = intotcpcb(inp);
1511 	MPASS(tp != NULL);
1512 
1513 	SOCKBUF_LOCK(&so->so_rcv);
1514 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1515 	*tlsseq = so->so_rcv.sb_tls_seqno;
1516 	SOCKBUF_UNLOCK(&so->so_rcv);
1517 
1518 	INP_RUNLOCK(inp);
1519 
1520 	return (0);
1521 }
1522 
1523 int
1524 ktls_get_tx_mode(struct socket *so, int *modep)
1525 {
1526 	struct ktls_session *tls;
1527 	struct inpcb *inp __diagused;
1528 
1529 	if (SOLISTENING(so))
1530 		return (EINVAL);
1531 	inp = so->so_pcb;
1532 	INP_WLOCK_ASSERT(inp);
1533 	SOCK_SENDBUF_LOCK(so);
1534 	tls = so->so_snd.sb_tls_info;
1535 	if (tls == NULL)
1536 		*modep = TCP_TLS_MODE_NONE;
1537 	else
1538 		*modep = tls->mode;
1539 	SOCK_SENDBUF_UNLOCK(so);
1540 	return (0);
1541 }
1542 
1543 /*
1544  * Switch between SW and ifnet TLS sessions as requested.
1545  */
1546 int
1547 ktls_set_tx_mode(struct socket *so, int mode)
1548 {
1549 	struct ktls_session *tls, *tls_new;
1550 	struct inpcb *inp;
1551 	struct tcpcb *tp;
1552 	int error;
1553 
1554 	if (SOLISTENING(so))
1555 		return (EINVAL);
1556 	switch (mode) {
1557 	case TCP_TLS_MODE_SW:
1558 	case TCP_TLS_MODE_IFNET:
1559 		break;
1560 	default:
1561 		return (EINVAL);
1562 	}
1563 
1564 	inp = so->so_pcb;
1565 	INP_WLOCK_ASSERT(inp);
1566 	tp = intotcpcb(inp);
1567 
1568 	if (mode == TCP_TLS_MODE_IFNET) {
1569 		/* Don't allow enabling ifnet ktls multiple times */
1570 		if (tp->t_nic_ktls_xmit)
1571 			return (EALREADY);
1572 
1573 		/*
1574 		 * Don't enable ifnet ktls if we disabled it due to an
1575 		 * excessive retransmission rate
1576 		 */
1577 		if (tp->t_nic_ktls_xmit_dis)
1578 			return (ENXIO);
1579 	}
1580 
1581 	SOCKBUF_LOCK(&so->so_snd);
1582 	tls = so->so_snd.sb_tls_info;
1583 	if (tls == NULL) {
1584 		SOCKBUF_UNLOCK(&so->so_snd);
1585 		return (0);
1586 	}
1587 
1588 	if (tls->mode == mode) {
1589 		SOCKBUF_UNLOCK(&so->so_snd);
1590 		return (0);
1591 	}
1592 
1593 	tls = ktls_hold(tls);
1594 	SOCKBUF_UNLOCK(&so->so_snd);
1595 	INP_WUNLOCK(inp);
1596 
1597 	tls_new = ktls_clone_session(tls, KTLS_TX);
1598 
1599 	if (mode == TCP_TLS_MODE_IFNET)
1600 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1601 	else
1602 		error = ktls_try_sw(tls_new, KTLS_TX);
1603 	if (error) {
1604 		counter_u64_add(ktls_switch_failed, 1);
1605 		ktls_free(tls_new);
1606 		ktls_free(tls);
1607 		INP_WLOCK(inp);
1608 		return (error);
1609 	}
1610 
1611 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1612 	if (error) {
1613 		counter_u64_add(ktls_switch_failed, 1);
1614 		ktls_free(tls_new);
1615 		ktls_free(tls);
1616 		INP_WLOCK(inp);
1617 		return (error);
1618 	}
1619 
1620 	/*
1621 	 * If we raced with another session change, keep the existing
1622 	 * session.
1623 	 */
1624 	if (tls != so->so_snd.sb_tls_info) {
1625 		counter_u64_add(ktls_switch_failed, 1);
1626 		SOCK_IO_SEND_UNLOCK(so);
1627 		ktls_free(tls_new);
1628 		ktls_free(tls);
1629 		INP_WLOCK(inp);
1630 		return (EBUSY);
1631 	}
1632 
1633 	INP_WLOCK(inp);
1634 	SOCKBUF_LOCK(&so->so_snd);
1635 	so->so_snd.sb_tls_info = tls_new;
1636 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1637 		MPASS(tp->t_nic_ktls_xmit == 0);
1638 		tp->t_nic_ktls_xmit = 1;
1639 		if (tp->t_fb->tfb_hwtls_change != NULL)
1640 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1641 	}
1642 	SOCKBUF_UNLOCK(&so->so_snd);
1643 	SOCK_IO_SEND_UNLOCK(so);
1644 
1645 	/*
1646 	 * Drop two references on 'tls'.  The first is for the
1647 	 * ktls_hold() above.  The second drops the reference from the
1648 	 * socket buffer.
1649 	 */
1650 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1651 	ktls_free(tls);
1652 	ktls_free(tls);
1653 
1654 	if (mode == TCP_TLS_MODE_IFNET)
1655 		counter_u64_add(ktls_switch_to_ifnet, 1);
1656 	else
1657 		counter_u64_add(ktls_switch_to_sw, 1);
1658 
1659 	return (0);
1660 }
1661 
1662 /*
1663  * Try to allocate a new TLS receive tag.  This task is scheduled when
1664  * sbappend_ktls_rx detects an input path change.  If a new tag is
1665  * allocated, replace the tag in the TLS session.  If a new tag cannot
1666  * be allocated, let the session fall back to software decryption.
1667  */
1668 static void
1669 ktls_reset_receive_tag(void *context, int pending)
1670 {
1671 	union if_snd_tag_alloc_params params;
1672 	struct ktls_session *tls;
1673 	struct m_snd_tag *mst;
1674 	struct inpcb *inp;
1675 	struct ifnet *ifp;
1676 	struct socket *so;
1677 	int error;
1678 
1679 	MPASS(pending == 1);
1680 
1681 	tls = context;
1682 	so = tls->so;
1683 	inp = so->so_pcb;
1684 	ifp = NULL;
1685 
1686 	INP_RLOCK(inp);
1687 	if (inp->inp_flags & INP_DROPPED) {
1688 		INP_RUNLOCK(inp);
1689 		goto out;
1690 	}
1691 
1692 	SOCKBUF_LOCK(&so->so_rcv);
1693 	mst = tls->snd_tag;
1694 	tls->snd_tag = NULL;
1695 	if (mst != NULL)
1696 		m_snd_tag_rele(mst);
1697 
1698 	ifp = tls->rx_ifp;
1699 	if_ref(ifp);
1700 	SOCKBUF_UNLOCK(&so->so_rcv);
1701 
1702 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1703 	params.hdr.flowid = inp->inp_flowid;
1704 	params.hdr.flowtype = inp->inp_flowtype;
1705 	params.hdr.numa_domain = inp->inp_numa_domain;
1706 	params.tls_rx.inp = inp;
1707 	params.tls_rx.tls = tls;
1708 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1709 	INP_RUNLOCK(inp);
1710 
1711 	if (inp->inp_vflag & INP_IPV6) {
1712 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1713 			goto out;
1714 	} else {
1715 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1716 			goto out;
1717 	}
1718 
1719 	error = m_snd_tag_alloc(ifp, &params, &mst);
1720 	if (error == 0) {
1721 		SOCKBUF_LOCK(&so->so_rcv);
1722 		tls->snd_tag = mst;
1723 		SOCKBUF_UNLOCK(&so->so_rcv);
1724 
1725 		counter_u64_add(ktls_ifnet_reset, 1);
1726 	} else {
1727 		/*
1728 		 * Just fall back to software decryption if a tag
1729 		 * cannot be allocated leaving the connection intact.
1730 		 * If a future input path change switches to another
1731 		 * interface this connection will resume ifnet TLS.
1732 		 */
1733 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1734 	}
1735 
1736 out:
1737 	mtx_pool_lock(mtxpool_sleep, tls);
1738 	tls->reset_pending = false;
1739 	mtx_pool_unlock(mtxpool_sleep, tls);
1740 
1741 	if (ifp != NULL)
1742 		if_rele(ifp);
1743 	CURVNET_SET(so->so_vnet);
1744 	sorele(so);
1745 	CURVNET_RESTORE();
1746 	ktls_free(tls);
1747 }
1748 
1749 /*
1750  * Try to allocate a new TLS send tag.  This task is scheduled when
1751  * ip_output detects a route change while trying to transmit a packet
1752  * holding a TLS record.  If a new tag is allocated, replace the tag
1753  * in the TLS session.  Subsequent packets on the connection will use
1754  * the new tag.  If a new tag cannot be allocated, drop the
1755  * connection.
1756  */
1757 static void
1758 ktls_reset_send_tag(void *context, int pending)
1759 {
1760 	struct epoch_tracker et;
1761 	struct ktls_session *tls;
1762 	struct m_snd_tag *old, *new;
1763 	struct inpcb *inp;
1764 	struct tcpcb *tp;
1765 	int error;
1766 
1767 	MPASS(pending == 1);
1768 
1769 	tls = context;
1770 	inp = tls->inp;
1771 
1772 	/*
1773 	 * Free the old tag first before allocating a new one.
1774 	 * ip[6]_output_send() will treat a NULL send tag the same as
1775 	 * an ifp mismatch and drop packets until a new tag is
1776 	 * allocated.
1777 	 *
1778 	 * Write-lock the INP when changing tls->snd_tag since
1779 	 * ip[6]_output_send() holds a read-lock when reading the
1780 	 * pointer.
1781 	 */
1782 	INP_WLOCK(inp);
1783 	old = tls->snd_tag;
1784 	tls->snd_tag = NULL;
1785 	INP_WUNLOCK(inp);
1786 	if (old != NULL)
1787 		m_snd_tag_rele(old);
1788 
1789 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1790 
1791 	if (error == 0) {
1792 		INP_WLOCK(inp);
1793 		tls->snd_tag = new;
1794 		mtx_pool_lock(mtxpool_sleep, tls);
1795 		tls->reset_pending = false;
1796 		mtx_pool_unlock(mtxpool_sleep, tls);
1797 		INP_WUNLOCK(inp);
1798 
1799 		counter_u64_add(ktls_ifnet_reset, 1);
1800 
1801 		/*
1802 		 * XXX: Should we kick tcp_output explicitly now that
1803 		 * the send tag is fixed or just rely on timers?
1804 		 */
1805 	} else {
1806 		NET_EPOCH_ENTER(et);
1807 		INP_WLOCK(inp);
1808 		if (!(inp->inp_flags & INP_DROPPED)) {
1809 			tp = intotcpcb(inp);
1810 			CURVNET_SET(inp->inp_vnet);
1811 			tp = tcp_drop(tp, ECONNABORTED);
1812 			CURVNET_RESTORE();
1813 			if (tp != NULL) {
1814 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1815 				INP_WUNLOCK(inp);
1816 			}
1817 		} else
1818 			INP_WUNLOCK(inp);
1819 		NET_EPOCH_EXIT(et);
1820 
1821 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1822 
1823 		/*
1824 		 * Leave reset_pending true to avoid future tasks while
1825 		 * the socket goes away.
1826 		 */
1827 	}
1828 
1829 	ktls_free(tls);
1830 }
1831 
1832 void
1833 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1834 {
1835 	struct ktls_session *tls;
1836 	struct socket *so;
1837 
1838 	SOCKBUF_LOCK_ASSERT(sb);
1839 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1840 	    __func__, sb));
1841 	so = __containerof(sb, struct socket, so_rcv);
1842 
1843 	tls = sb->sb_tls_info;
1844 	if_rele(tls->rx_ifp);
1845 	if_ref(ifp);
1846 	tls->rx_ifp = ifp;
1847 
1848 	/*
1849 	 * See if we should schedule a task to update the receive tag for
1850 	 * this session.
1851 	 */
1852 	mtx_pool_lock(mtxpool_sleep, tls);
1853 	if (!tls->reset_pending) {
1854 		(void) ktls_hold(tls);
1855 		soref(so);
1856 		tls->so = so;
1857 		tls->reset_pending = true;
1858 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1859 	}
1860 	mtx_pool_unlock(mtxpool_sleep, tls);
1861 }
1862 
1863 int
1864 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1865 {
1866 
1867 	if (inp == NULL)
1868 		return (ENOBUFS);
1869 
1870 	INP_LOCK_ASSERT(inp);
1871 
1872 	/*
1873 	 * See if we should schedule a task to update the send tag for
1874 	 * this session.
1875 	 */
1876 	mtx_pool_lock(mtxpool_sleep, tls);
1877 	if (!tls->reset_pending) {
1878 		(void) ktls_hold(tls);
1879 		tls->reset_pending = true;
1880 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1881 	}
1882 	mtx_pool_unlock(mtxpool_sleep, tls);
1883 	return (ENOBUFS);
1884 }
1885 
1886 #ifdef RATELIMIT
1887 int
1888 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1889 {
1890 	union if_snd_tag_modify_params params = {
1891 		.rate_limit.max_rate = max_pacing_rate,
1892 		.rate_limit.flags = M_NOWAIT,
1893 	};
1894 	struct m_snd_tag *mst;
1895 
1896 	/* Can't get to the inp, but it should be locked. */
1897 	/* INP_LOCK_ASSERT(inp); */
1898 
1899 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1900 
1901 	if (tls->snd_tag == NULL) {
1902 		/*
1903 		 * Resetting send tag, ignore this change.  The
1904 		 * pending reset may or may not see this updated rate
1905 		 * in the tcpcb.  If it doesn't, we will just lose
1906 		 * this rate change.
1907 		 */
1908 		return (0);
1909 	}
1910 
1911 	mst = tls->snd_tag;
1912 
1913 	MPASS(mst != NULL);
1914 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1915 
1916 	return (mst->sw->snd_tag_modify(mst, &params));
1917 }
1918 #endif
1919 
1920 static void
1921 ktls_destroy_help(void *context, int pending __unused)
1922 {
1923 	ktls_destroy(context);
1924 }
1925 
1926 void
1927 ktls_destroy(struct ktls_session *tls)
1928 {
1929 	struct inpcb *inp;
1930 	struct tcpcb *tp;
1931 	bool wlocked;
1932 
1933 	MPASS(tls->refcount == 0);
1934 
1935 	inp = tls->inp;
1936 	if (tls->tx) {
1937 		wlocked = INP_WLOCKED(inp);
1938 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1939 			/*
1940 			 * rwlocks read locks are anonymous, and there
1941 			 * is no way to know if our current thread
1942 			 * holds an rlock on the inp.  As a rough
1943 			 * estimate, check to see if the thread holds
1944 			 * *any* rlocks at all.  If it does not, then we
1945 			 * know that we don't hold the inp rlock, and
1946 			 * can safely take the wlock
1947 			 */
1948 			if (curthread->td_rw_rlocks == 0) {
1949 				INP_WLOCK(inp);
1950 			} else {
1951 				/*
1952 				 * We might hold the rlock, so let's
1953 				 * do the destroy in a taskqueue
1954 				 * context to avoid a potential
1955 				 * deadlock.  This should be very
1956 				 * rare.
1957 				 */
1958 				counter_u64_add(ktls_destroy_task, 1);
1959 				TASK_INIT(&tls->destroy_task, 0,
1960 				    ktls_destroy_help, tls);
1961 				(void)taskqueue_enqueue(taskqueue_thread,
1962 				    &tls->destroy_task);
1963 				return;
1964 			}
1965 		}
1966 	}
1967 
1968 	if (tls->sequential_records) {
1969 		struct mbuf *m, *n;
1970 		int page_count;
1971 
1972 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1973 			page_count = m->m_epg_enc_cnt;
1974 			while (page_count > 0) {
1975 				KASSERT(page_count >= m->m_epg_nrdy,
1976 				    ("%s: too few pages", __func__));
1977 				page_count -= m->m_epg_nrdy;
1978 				m = m_free(m);
1979 			}
1980 		}
1981 	}
1982 
1983 	counter_u64_add(ktls_offload_active, -1);
1984 	switch (tls->mode) {
1985 	case TCP_TLS_MODE_SW:
1986 		switch (tls->params.cipher_algorithm) {
1987 		case CRYPTO_AES_CBC:
1988 			counter_u64_add(ktls_sw_cbc, -1);
1989 			break;
1990 		case CRYPTO_AES_NIST_GCM_16:
1991 			counter_u64_add(ktls_sw_gcm, -1);
1992 			break;
1993 		case CRYPTO_CHACHA20_POLY1305:
1994 			counter_u64_add(ktls_sw_chacha20, -1);
1995 			break;
1996 		}
1997 		break;
1998 	case TCP_TLS_MODE_IFNET:
1999 		switch (tls->params.cipher_algorithm) {
2000 		case CRYPTO_AES_CBC:
2001 			counter_u64_add(ktls_ifnet_cbc, -1);
2002 			break;
2003 		case CRYPTO_AES_NIST_GCM_16:
2004 			counter_u64_add(ktls_ifnet_gcm, -1);
2005 			break;
2006 		case CRYPTO_CHACHA20_POLY1305:
2007 			counter_u64_add(ktls_ifnet_chacha20, -1);
2008 			break;
2009 		}
2010 		if (tls->snd_tag != NULL)
2011 			m_snd_tag_rele(tls->snd_tag);
2012 		if (tls->rx_ifp != NULL)
2013 			if_rele(tls->rx_ifp);
2014 		if (tls->tx) {
2015 			INP_WLOCK_ASSERT(inp);
2016 			tp = intotcpcb(inp);
2017 			MPASS(tp->t_nic_ktls_xmit == 1);
2018 			tp->t_nic_ktls_xmit = 0;
2019 		}
2020 		break;
2021 #ifdef TCP_OFFLOAD
2022 	case TCP_TLS_MODE_TOE:
2023 		switch (tls->params.cipher_algorithm) {
2024 		case CRYPTO_AES_CBC:
2025 			counter_u64_add(ktls_toe_cbc, -1);
2026 			break;
2027 		case CRYPTO_AES_NIST_GCM_16:
2028 			counter_u64_add(ktls_toe_gcm, -1);
2029 			break;
2030 		case CRYPTO_CHACHA20_POLY1305:
2031 			counter_u64_add(ktls_toe_chacha20, -1);
2032 			break;
2033 		}
2034 		break;
2035 #endif
2036 	}
2037 	if (tls->ocf_session != NULL)
2038 		ktls_ocf_free(tls);
2039 	if (tls->params.auth_key != NULL) {
2040 		zfree(tls->params.auth_key, M_KTLS);
2041 		tls->params.auth_key = NULL;
2042 		tls->params.auth_key_len = 0;
2043 	}
2044 	if (tls->params.cipher_key != NULL) {
2045 		zfree(tls->params.cipher_key, M_KTLS);
2046 		tls->params.cipher_key = NULL;
2047 		tls->params.cipher_key_len = 0;
2048 	}
2049 	if (tls->tx) {
2050 		INP_WLOCK_ASSERT(inp);
2051 		if (!in_pcbrele_wlocked(inp) && !wlocked)
2052 			INP_WUNLOCK(inp);
2053 	}
2054 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
2055 
2056 	uma_zfree(ktls_session_zone, tls);
2057 }
2058 
2059 void
2060 ktls_seq(struct sockbuf *sb, struct mbuf *m)
2061 {
2062 
2063 	for (; m != NULL; m = m->m_next) {
2064 		KASSERT((m->m_flags & M_EXTPG) != 0,
2065 		    ("ktls_seq: mapped mbuf %p", m));
2066 
2067 		m->m_epg_seqno = sb->sb_tls_seqno;
2068 		sb->sb_tls_seqno++;
2069 	}
2070 }
2071 
2072 /*
2073  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
2074  * mbuf in the chain must be an unmapped mbuf.  The payload of the
2075  * mbuf must be populated with the payload of each TLS record.
2076  *
2077  * The record_type argument specifies the TLS record type used when
2078  * populating the TLS header.
2079  *
2080  * The enq_count argument on return is set to the number of pages of
2081  * payload data for this entire chain that need to be encrypted via SW
2082  * encryption.  The returned value should be passed to ktls_enqueue
2083  * when scheduling encryption of this chain of mbufs.  To handle the
2084  * special case of empty fragments for TLS 1.0 sessions, an empty
2085  * fragment counts as one page.
2086  */
2087 void
2088 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2089     uint8_t record_type)
2090 {
2091 	struct tls_record_layer *tlshdr;
2092 	struct mbuf *m;
2093 	uint64_t *noncep;
2094 	uint16_t tls_len;
2095 	int maxlen __diagused;
2096 
2097 	maxlen = tls->params.max_frame_len;
2098 	*enq_cnt = 0;
2099 	for (m = top; m != NULL; m = m->m_next) {
2100 		/*
2101 		 * All mbufs in the chain should be TLS records whose
2102 		 * payload does not exceed the maximum frame length.
2103 		 *
2104 		 * Empty TLS 1.0 records are permitted when using CBC.
2105 		 */
2106 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2107 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2108 		    ("ktls_frame: m %p len %d", m, m->m_len));
2109 
2110 		/*
2111 		 * TLS frames require unmapped mbufs to store session
2112 		 * info.
2113 		 */
2114 		KASSERT((m->m_flags & M_EXTPG) != 0,
2115 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2116 
2117 		tls_len = m->m_len;
2118 
2119 		/* Save a reference to the session. */
2120 		m->m_epg_tls = ktls_hold(tls);
2121 
2122 		m->m_epg_hdrlen = tls->params.tls_hlen;
2123 		m->m_epg_trllen = tls->params.tls_tlen;
2124 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2125 			int bs, delta;
2126 
2127 			/*
2128 			 * AES-CBC pads messages to a multiple of the
2129 			 * block size.  Note that the padding is
2130 			 * applied after the digest and the encryption
2131 			 * is done on the "plaintext || mac || padding".
2132 			 * At least one byte of padding is always
2133 			 * present.
2134 			 *
2135 			 * Compute the final trailer length assuming
2136 			 * at most one block of padding.
2137 			 * tls->params.tls_tlen is the maximum
2138 			 * possible trailer length (padding + digest).
2139 			 * delta holds the number of excess padding
2140 			 * bytes if the maximum were used.  Those
2141 			 * extra bytes are removed.
2142 			 */
2143 			bs = tls->params.tls_bs;
2144 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2145 			m->m_epg_trllen -= delta;
2146 		}
2147 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2148 
2149 		/* Populate the TLS header. */
2150 		tlshdr = (void *)m->m_epg_hdr;
2151 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2152 
2153 		/*
2154 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2155 		 * of TLS_RLTYPE_APP.
2156 		 */
2157 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2158 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2159 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2160 			tlshdr->tls_type = TLS_RLTYPE_APP;
2161 			/* save the real record type for later */
2162 			m->m_epg_record_type = record_type;
2163 			m->m_epg_trail[0] = record_type;
2164 		} else {
2165 			tlshdr->tls_vminor = tls->params.tls_vminor;
2166 			tlshdr->tls_type = record_type;
2167 		}
2168 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2169 
2170 		/*
2171 		 * Store nonces / explicit IVs after the end of the
2172 		 * TLS header.
2173 		 *
2174 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2175 		 * from the end of the IV.  The nonce is then
2176 		 * incremented for use by the next record.
2177 		 *
2178 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2179 		 */
2180 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2181 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2182 			noncep = (uint64_t *)(tls->params.iv + 8);
2183 			be64enc(tlshdr + 1, *noncep);
2184 			(*noncep)++;
2185 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2186 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2187 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2188 
2189 		/*
2190 		 * When using SW encryption, mark the mbuf not ready.
2191 		 * It will be marked ready via sbready() after the
2192 		 * record has been encrypted.
2193 		 *
2194 		 * When using ifnet TLS, unencrypted TLS records are
2195 		 * sent down the stack to the NIC.
2196 		 */
2197 		if (tls->mode == TCP_TLS_MODE_SW) {
2198 			m->m_flags |= M_NOTREADY;
2199 			if (__predict_false(tls_len == 0)) {
2200 				/* TLS 1.0 empty fragment. */
2201 				m->m_epg_nrdy = 1;
2202 			} else
2203 				m->m_epg_nrdy = m->m_epg_npgs;
2204 			*enq_cnt += m->m_epg_nrdy;
2205 		}
2206 	}
2207 }
2208 
2209 bool
2210 ktls_permit_empty_frames(struct ktls_session *tls)
2211 {
2212 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2213 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2214 }
2215 
2216 void
2217 ktls_check_rx(struct sockbuf *sb)
2218 {
2219 	struct tls_record_layer hdr;
2220 	struct ktls_wq *wq;
2221 	struct socket *so;
2222 	bool running;
2223 
2224 	SOCKBUF_LOCK_ASSERT(sb);
2225 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2226 	    __func__, sb));
2227 	so = __containerof(sb, struct socket, so_rcv);
2228 
2229 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2230 		return;
2231 
2232 	/* Is there enough queued for a TLS header? */
2233 	if (sb->sb_tlscc < sizeof(hdr)) {
2234 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2235 			so->so_error = EMSGSIZE;
2236 		return;
2237 	}
2238 
2239 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2240 
2241 	/* Is the entire record queued? */
2242 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2243 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2244 			so->so_error = EMSGSIZE;
2245 		return;
2246 	}
2247 
2248 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2249 
2250 	soref(so);
2251 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2252 	mtx_lock(&wq->mtx);
2253 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2254 	running = wq->running;
2255 	mtx_unlock(&wq->mtx);
2256 	if (!running)
2257 		wakeup(wq);
2258 	counter_u64_add(ktls_cnt_rx_queued, 1);
2259 }
2260 
2261 static struct mbuf *
2262 ktls_detach_record(struct sockbuf *sb, int len)
2263 {
2264 	struct mbuf *m, *n, *top;
2265 	int remain;
2266 
2267 	SOCKBUF_LOCK_ASSERT(sb);
2268 	MPASS(len <= sb->sb_tlscc);
2269 
2270 	/*
2271 	 * If TLS chain is the exact size of the record,
2272 	 * just grab the whole record.
2273 	 */
2274 	top = sb->sb_mtls;
2275 	if (sb->sb_tlscc == len) {
2276 		sb->sb_mtls = NULL;
2277 		sb->sb_mtlstail = NULL;
2278 		goto out;
2279 	}
2280 
2281 	/*
2282 	 * While it would be nice to use m_split() here, we need
2283 	 * to know exactly what m_split() allocates to update the
2284 	 * accounting, so do it inline instead.
2285 	 */
2286 	remain = len;
2287 	for (m = top; remain > m->m_len; m = m->m_next)
2288 		remain -= m->m_len;
2289 
2290 	/* Easy case: don't have to split 'm'. */
2291 	if (remain == m->m_len) {
2292 		sb->sb_mtls = m->m_next;
2293 		if (sb->sb_mtls == NULL)
2294 			sb->sb_mtlstail = NULL;
2295 		m->m_next = NULL;
2296 		goto out;
2297 	}
2298 
2299 	/*
2300 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2301 	 * with M_NOWAIT first.
2302 	 */
2303 	n = m_get(M_NOWAIT, MT_DATA);
2304 	if (n == NULL) {
2305 		/*
2306 		 * Use M_WAITOK with socket buffer unlocked.  If
2307 		 * 'sb_mtls' changes while the lock is dropped, return
2308 		 * NULL to force the caller to retry.
2309 		 */
2310 		SOCKBUF_UNLOCK(sb);
2311 
2312 		n = m_get(M_WAITOK, MT_DATA);
2313 
2314 		SOCKBUF_LOCK(sb);
2315 		if (sb->sb_mtls != top) {
2316 			m_free(n);
2317 			return (NULL);
2318 		}
2319 	}
2320 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2321 
2322 	/* Store remainder in 'n'. */
2323 	n->m_len = m->m_len - remain;
2324 	if (m->m_flags & M_EXT) {
2325 		n->m_data = m->m_data + remain;
2326 		mb_dupcl(n, m);
2327 	} else {
2328 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2329 	}
2330 
2331 	/* Trim 'm' and update accounting. */
2332 	m->m_len -= n->m_len;
2333 	sb->sb_tlscc -= n->m_len;
2334 	sb->sb_ccc -= n->m_len;
2335 
2336 	/* Account for 'n'. */
2337 	sballoc_ktls_rx(sb, n);
2338 
2339 	/* Insert 'n' into the TLS chain. */
2340 	sb->sb_mtls = n;
2341 	n->m_next = m->m_next;
2342 	if (sb->sb_mtlstail == m)
2343 		sb->sb_mtlstail = n;
2344 
2345 	/* Detach the record from the TLS chain. */
2346 	m->m_next = NULL;
2347 
2348 out:
2349 	MPASS(m_length(top, NULL) == len);
2350 	for (m = top; m != NULL; m = m->m_next)
2351 		sbfree_ktls_rx(sb, m);
2352 	sb->sb_tlsdcc = len;
2353 	sb->sb_ccc += len;
2354 	SBCHECK(sb);
2355 	return (top);
2356 }
2357 
2358 /*
2359  * Determine the length of the trailing zero padding and find the real
2360  * record type in the byte before the padding.
2361  *
2362  * Walking the mbuf chain backwards is clumsy, so another option would
2363  * be to scan forwards remembering the last non-zero byte before the
2364  * trailer.  However, it would be expensive to scan the entire record.
2365  * Instead, find the last non-zero byte of each mbuf in the chain
2366  * keeping track of the relative offset of that nonzero byte.
2367  *
2368  * trail_len is the size of the MAC/tag on input and is set to the
2369  * size of the full trailer including padding and the record type on
2370  * return.
2371  */
2372 static int
2373 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2374     int *trailer_len, uint8_t *record_typep)
2375 {
2376 	char *cp;
2377 	u_int digest_start, last_offset, m_len, offset;
2378 	uint8_t record_type;
2379 
2380 	digest_start = tls_len - *trailer_len;
2381 	last_offset = 0;
2382 	offset = 0;
2383 	for (; m != NULL && offset < digest_start;
2384 	     offset += m->m_len, m = m->m_next) {
2385 		/* Don't look for padding in the tag. */
2386 		m_len = min(digest_start - offset, m->m_len);
2387 		cp = mtod(m, char *);
2388 
2389 		/* Find last non-zero byte in this mbuf. */
2390 		while (m_len > 0 && cp[m_len - 1] == 0)
2391 			m_len--;
2392 		if (m_len > 0) {
2393 			record_type = cp[m_len - 1];
2394 			last_offset = offset + m_len;
2395 		}
2396 	}
2397 	if (last_offset < tls->params.tls_hlen)
2398 		return (EBADMSG);
2399 
2400 	*record_typep = record_type;
2401 	*trailer_len = tls_len - last_offset + 1;
2402 	return (0);
2403 }
2404 
2405 /*
2406  * Check if a mbuf chain is fully decrypted at the given offset and
2407  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2408  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2409  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2410  * is encrypted.
2411  */
2412 ktls_mbuf_crypto_st_t
2413 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2414 {
2415 	int m_flags_ored = 0;
2416 	int m_flags_anded = -1;
2417 
2418 	for (; mb != NULL; mb = mb->m_next) {
2419 		if (offset < mb->m_len)
2420 			break;
2421 		offset -= mb->m_len;
2422 	}
2423 	offset += len;
2424 
2425 	for (; mb != NULL; mb = mb->m_next) {
2426 		m_flags_ored |= mb->m_flags;
2427 		m_flags_anded &= mb->m_flags;
2428 
2429 		if (offset <= mb->m_len)
2430 			break;
2431 		offset -= mb->m_len;
2432 	}
2433 	MPASS(mb != NULL || offset == 0);
2434 
2435 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2436 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2437 	else
2438 		return ((m_flags_ored & M_DECRYPTED) ?
2439 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2440 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2441 }
2442 
2443 /*
2444  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2445  */
2446 static int
2447 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2448 {
2449 	union if_snd_tag_modify_params params;
2450 	struct m_snd_tag *mst;
2451 	struct inpcb *inp;
2452 	struct tcpcb *tp;
2453 
2454 	mst = so->so_rcv.sb_tls_info->snd_tag;
2455 	if (__predict_false(mst == NULL))
2456 		return (EINVAL);
2457 
2458 	inp = sotoinpcb(so);
2459 	if (__predict_false(inp == NULL))
2460 		return (EINVAL);
2461 
2462 	INP_RLOCK(inp);
2463 	if (inp->inp_flags & INP_DROPPED) {
2464 		INP_RUNLOCK(inp);
2465 		return (ECONNRESET);
2466 	}
2467 
2468 	tp = intotcpcb(inp);
2469 	MPASS(tp != NULL);
2470 
2471 	/* Get the TCP sequence number of the next valid TLS header. */
2472 	SOCKBUF_LOCK(&so->so_rcv);
2473 	params.tls_rx.tls_hdr_tcp_sn =
2474 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2475 	params.tls_rx.tls_rec_length = tls_len;
2476 	params.tls_rx.tls_seq_number = tls_rcd_num;
2477 	SOCKBUF_UNLOCK(&so->so_rcv);
2478 
2479 	INP_RUNLOCK(inp);
2480 
2481 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2482 	return (mst->sw->snd_tag_modify(mst, &params));
2483 }
2484 
2485 static void
2486 ktls_drop(struct socket *so, int error)
2487 {
2488 	struct epoch_tracker et;
2489 	struct inpcb *inp = sotoinpcb(so);
2490 	struct tcpcb *tp;
2491 
2492 	NET_EPOCH_ENTER(et);
2493 	INP_WLOCK(inp);
2494 	if (!(inp->inp_flags & INP_DROPPED)) {
2495 		tp = intotcpcb(inp);
2496 		CURVNET_SET(inp->inp_vnet);
2497 		tp = tcp_drop(tp, error);
2498 		CURVNET_RESTORE();
2499 		if (tp != NULL)
2500 			INP_WUNLOCK(inp);
2501 	} else {
2502 		so->so_error = error;
2503 		SOCK_RECVBUF_LOCK(so);
2504 		sorwakeup_locked(so);
2505 		INP_WUNLOCK(inp);
2506 	}
2507 	NET_EPOCH_EXIT(et);
2508 }
2509 
2510 static void
2511 ktls_decrypt(struct socket *so)
2512 {
2513 	char tls_header[MBUF_PEXT_HDR_LEN];
2514 	struct ktls_session *tls;
2515 	struct sockbuf *sb;
2516 	struct tls_record_layer *hdr;
2517 	struct tls_get_record tgr;
2518 	struct mbuf *control, *data, *m;
2519 	ktls_mbuf_crypto_st_t state;
2520 	uint64_t seqno;
2521 	int error, remain, tls_len, trail_len;
2522 	bool tls13;
2523 	uint8_t vminor, record_type;
2524 
2525 	hdr = (struct tls_record_layer *)tls_header;
2526 	sb = &so->so_rcv;
2527 	SOCKBUF_LOCK(sb);
2528 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2529 	    ("%s: socket %p not running", __func__, so));
2530 
2531 	tls = sb->sb_tls_info;
2532 	MPASS(tls != NULL);
2533 
2534 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2535 	if (tls13)
2536 		vminor = TLS_MINOR_VER_TWO;
2537 	else
2538 		vminor = tls->params.tls_vminor;
2539 	for (;;) {
2540 		/* Is there enough queued for a TLS header? */
2541 		if (sb->sb_tlscc < tls->params.tls_hlen)
2542 			break;
2543 
2544 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2545 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2546 
2547 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2548 		    hdr->tls_vminor != vminor)
2549 			error = EINVAL;
2550 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2551 			error = EINVAL;
2552 		else if (tls_len < tls->params.tls_hlen || tls_len >
2553 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2554 		    tls->params.tls_tlen)
2555 			error = EMSGSIZE;
2556 		else
2557 			error = 0;
2558 		if (__predict_false(error != 0)) {
2559 			/*
2560 			 * We have a corrupted record and are likely
2561 			 * out of sync.  The connection isn't
2562 			 * recoverable at this point, so abort it.
2563 			 */
2564 			SOCKBUF_UNLOCK(sb);
2565 			counter_u64_add(ktls_offload_corrupted_records, 1);
2566 
2567 			ktls_drop(so, error);
2568 			goto deref;
2569 		}
2570 
2571 		/* Is the entire record queued? */
2572 		if (sb->sb_tlscc < tls_len)
2573 			break;
2574 
2575 		/*
2576 		 * Split out the portion of the mbuf chain containing
2577 		 * this TLS record.
2578 		 */
2579 		data = ktls_detach_record(sb, tls_len);
2580 		if (data == NULL)
2581 			continue;
2582 		MPASS(sb->sb_tlsdcc == tls_len);
2583 
2584 		seqno = sb->sb_tls_seqno;
2585 		sb->sb_tls_seqno++;
2586 		SBCHECK(sb);
2587 		SOCKBUF_UNLOCK(sb);
2588 
2589 		/* get crypto state for this TLS record */
2590 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2591 
2592 		switch (state) {
2593 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2594 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2595 			if (error)
2596 				break;
2597 			/* FALLTHROUGH */
2598 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2599 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2600 			    &trail_len);
2601 			if (__predict_true(error == 0)) {
2602 				if (tls13) {
2603 					error = tls13_find_record_type(tls, data,
2604 					    tls_len, &trail_len, &record_type);
2605 				} else {
2606 					record_type = hdr->tls_type;
2607 				}
2608 			}
2609 			break;
2610 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2611 			/*
2612 			 * NIC TLS is only supported for AEAD
2613 			 * ciphersuites which used a fixed sized
2614 			 * trailer.
2615 			 */
2616 			if (tls13) {
2617 				trail_len = tls->params.tls_tlen - 1;
2618 				error = tls13_find_record_type(tls, data,
2619 				    tls_len, &trail_len, &record_type);
2620 			} else {
2621 				trail_len = tls->params.tls_tlen;
2622 				error = 0;
2623 				record_type = hdr->tls_type;
2624 			}
2625 			break;
2626 		default:
2627 			error = EINVAL;
2628 			break;
2629 		}
2630 		if (error) {
2631 			counter_u64_add(ktls_offload_failed_crypto, 1);
2632 
2633 			SOCKBUF_LOCK(sb);
2634 			if (sb->sb_tlsdcc == 0) {
2635 				/*
2636 				 * sbcut/drop/flush discarded these
2637 				 * mbufs.
2638 				 */
2639 				m_freem(data);
2640 				break;
2641 			}
2642 
2643 			/*
2644 			 * Drop this TLS record's data, but keep
2645 			 * decrypting subsequent records.
2646 			 */
2647 			sb->sb_ccc -= tls_len;
2648 			sb->sb_tlsdcc = 0;
2649 
2650 			if (error != EMSGSIZE)
2651 				error = EBADMSG;
2652 			CURVNET_SET(so->so_vnet);
2653 			so->so_error = error;
2654 			sorwakeup_locked(so);
2655 			CURVNET_RESTORE();
2656 
2657 			m_freem(data);
2658 
2659 			SOCKBUF_LOCK(sb);
2660 			continue;
2661 		}
2662 
2663 		/* Allocate the control mbuf. */
2664 		memset(&tgr, 0, sizeof(tgr));
2665 		tgr.tls_type = record_type;
2666 		tgr.tls_vmajor = hdr->tls_vmajor;
2667 		tgr.tls_vminor = hdr->tls_vminor;
2668 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2669 		    trail_len);
2670 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2671 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2672 
2673 		SOCKBUF_LOCK(sb);
2674 		if (sb->sb_tlsdcc == 0) {
2675 			/* sbcut/drop/flush discarded these mbufs. */
2676 			MPASS(sb->sb_tlscc == 0);
2677 			m_freem(data);
2678 			m_freem(control);
2679 			break;
2680 		}
2681 
2682 		/*
2683 		 * Clear the 'dcc' accounting in preparation for
2684 		 * adding the decrypted record.
2685 		 */
2686 		sb->sb_ccc -= tls_len;
2687 		sb->sb_tlsdcc = 0;
2688 		SBCHECK(sb);
2689 
2690 		/* If there is no payload, drop all of the data. */
2691 		if (tgr.tls_length == htobe16(0)) {
2692 			m_freem(data);
2693 			data = NULL;
2694 		} else {
2695 			/* Trim header. */
2696 			remain = tls->params.tls_hlen;
2697 			while (remain > 0) {
2698 				if (data->m_len > remain) {
2699 					data->m_data += remain;
2700 					data->m_len -= remain;
2701 					break;
2702 				}
2703 				remain -= data->m_len;
2704 				data = m_free(data);
2705 			}
2706 
2707 			/* Trim trailer and clear M_NOTREADY. */
2708 			remain = be16toh(tgr.tls_length);
2709 			m = data;
2710 			for (m = data; remain > m->m_len; m = m->m_next) {
2711 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2712 				remain -= m->m_len;
2713 			}
2714 			m->m_len = remain;
2715 			m_freem(m->m_next);
2716 			m->m_next = NULL;
2717 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2718 
2719 			/* Set EOR on the final mbuf. */
2720 			m->m_flags |= M_EOR;
2721 		}
2722 
2723 		sbappendcontrol_locked(sb, data, control, 0);
2724 
2725 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2726 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2727 			SOCKBUF_UNLOCK(sb);
2728 			ktls_resync_ifnet(so, tls_len, seqno);
2729 			SOCKBUF_LOCK(sb);
2730 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2731 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2732 			SOCKBUF_UNLOCK(sb);
2733 			ktls_resync_ifnet(so, 0, seqno);
2734 			SOCKBUF_LOCK(sb);
2735 		}
2736 	}
2737 
2738 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2739 
2740 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2741 		so->so_error = EMSGSIZE;
2742 
2743 	sorwakeup_locked(so);
2744 
2745 deref:
2746 	SOCKBUF_UNLOCK_ASSERT(sb);
2747 
2748 	CURVNET_SET(so->so_vnet);
2749 	sorele(so);
2750 	CURVNET_RESTORE();
2751 }
2752 
2753 void
2754 ktls_enqueue_to_free(struct mbuf *m)
2755 {
2756 	struct ktls_wq *wq;
2757 	bool running;
2758 
2759 	/* Mark it for freeing. */
2760 	m->m_epg_flags |= EPG_FLAG_2FREE;
2761 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2762 	mtx_lock(&wq->mtx);
2763 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2764 	running = wq->running;
2765 	mtx_unlock(&wq->mtx);
2766 	if (!running)
2767 		wakeup(wq);
2768 }
2769 
2770 static void *
2771 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2772 {
2773 	void *buf;
2774 	int domain, running;
2775 
2776 	if (m->m_epg_npgs <= 2)
2777 		return (NULL);
2778 	if (ktls_buffer_zone == NULL)
2779 		return (NULL);
2780 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2781 		/*
2782 		 * Rate-limit allocation attempts after a failure.
2783 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2784 		 * the free page queues and may fail consistently if memory is
2785 		 * fragmented.
2786 		 */
2787 		return (NULL);
2788 	}
2789 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2790 	if (buf == NULL) {
2791 		domain = PCPU_GET(domain);
2792 		wq->lastallocfail = ticks;
2793 
2794 		/*
2795 		 * Note that this check is "racy", but the races are
2796 		 * harmless, and are either a spurious wakeup if
2797 		 * multiple threads fail allocations before the alloc
2798 		 * thread wakes, or waiting an extra second in case we
2799 		 * see an old value of running == true.
2800 		 */
2801 		if (!VM_DOMAIN_EMPTY(domain)) {
2802 			running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2803 			if (!running)
2804 				wakeup(&ktls_domains[domain].reclaim_td);
2805 		}
2806 	}
2807 	return (buf);
2808 }
2809 
2810 static int
2811 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2812     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2813 {
2814 	vm_page_t pg;
2815 	int error, i, len, off;
2816 
2817 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2818 	    ("%p not unready & nomap mbuf\n", m));
2819 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2820 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2821 	    ktls_maxlen));
2822 
2823 	/* Anonymous mbufs are encrypted in place. */
2824 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2825 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2826 
2827 	/*
2828 	 * For file-backed mbufs (from sendfile), anonymous wired
2829 	 * pages are allocated and used as the encryption destination.
2830 	 */
2831 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2832 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2833 		    m->m_epg_1st_off;
2834 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2835 		    m->m_epg_1st_off;
2836 		state->dst_iov[0].iov_len = len;
2837 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2838 		i = 1;
2839 	} else {
2840 		off = m->m_epg_1st_off;
2841 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2842 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2843 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2844 			len = m_epg_pagelen(m, i, off);
2845 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2846 			state->dst_iov[i].iov_base =
2847 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2848 			state->dst_iov[i].iov_len = len;
2849 		}
2850 	}
2851 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2852 	state->dst_iov[i].iov_base = m->m_epg_trail;
2853 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2854 
2855 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2856 
2857 	if (__predict_false(error != 0)) {
2858 		/* Free the anonymous pages. */
2859 		if (state->cbuf != NULL)
2860 			uma_zfree(ktls_buffer_zone, state->cbuf);
2861 		else {
2862 			for (i = 0; i < m->m_epg_npgs; i++) {
2863 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2864 				(void)vm_page_unwire_noq(pg);
2865 				vm_page_free(pg);
2866 			}
2867 		}
2868 	}
2869 	return (error);
2870 }
2871 
2872 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2873 static u_int
2874 ktls_batched_records(struct mbuf *m)
2875 {
2876 	int page_count, records;
2877 
2878 	records = 0;
2879 	page_count = m->m_epg_enc_cnt;
2880 	while (page_count > 0) {
2881 		records++;
2882 		page_count -= m->m_epg_nrdy;
2883 		m = m->m_next;
2884 	}
2885 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2886 	return (records);
2887 }
2888 
2889 void
2890 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2891 {
2892 	struct ktls_session *tls;
2893 	struct ktls_wq *wq;
2894 	int queued;
2895 	bool running;
2896 
2897 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2898 	    (M_EXTPG | M_NOTREADY)),
2899 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2900 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2901 
2902 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2903 
2904 	m->m_epg_enc_cnt = page_count;
2905 
2906 	/*
2907 	 * Save a pointer to the socket.  The caller is responsible
2908 	 * for taking an additional reference via soref().
2909 	 */
2910 	m->m_epg_so = so;
2911 
2912 	queued = 1;
2913 	tls = m->m_epg_tls;
2914 	wq = &ktls_wq[tls->wq_index];
2915 	mtx_lock(&wq->mtx);
2916 	if (__predict_false(tls->sequential_records)) {
2917 		/*
2918 		 * For TLS 1.0, records must be encrypted
2919 		 * sequentially.  For a given connection, all records
2920 		 * queued to the associated work queue are processed
2921 		 * sequentially.  However, sendfile(2) might complete
2922 		 * I/O requests spanning multiple TLS records out of
2923 		 * order.  Here we ensure TLS records are enqueued to
2924 		 * the work queue in FIFO order.
2925 		 *
2926 		 * tls->next_seqno holds the sequence number of the
2927 		 * next TLS record that should be enqueued to the work
2928 		 * queue.  If this next record is not tls->next_seqno,
2929 		 * it must be a future record, so insert it, sorted by
2930 		 * TLS sequence number, into tls->pending_records and
2931 		 * return.
2932 		 *
2933 		 * If this TLS record matches tls->next_seqno, place
2934 		 * it in the work queue and then check
2935 		 * tls->pending_records to see if any
2936 		 * previously-queued records are now ready for
2937 		 * encryption.
2938 		 */
2939 		if (m->m_epg_seqno != tls->next_seqno) {
2940 			struct mbuf *n, *p;
2941 
2942 			p = NULL;
2943 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2944 				if (n->m_epg_seqno > m->m_epg_seqno)
2945 					break;
2946 				p = n;
2947 			}
2948 			if (n == NULL)
2949 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2950 				    m_epg_stailq);
2951 			else if (p == NULL)
2952 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2953 				    m_epg_stailq);
2954 			else
2955 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2956 				    m_epg_stailq);
2957 			mtx_unlock(&wq->mtx);
2958 			counter_u64_add(ktls_cnt_tx_pending, 1);
2959 			return;
2960 		}
2961 
2962 		tls->next_seqno += ktls_batched_records(m);
2963 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2964 
2965 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2966 			struct mbuf *n;
2967 
2968 			n = STAILQ_FIRST(&tls->pending_records);
2969 			if (n->m_epg_seqno != tls->next_seqno)
2970 				break;
2971 
2972 			queued++;
2973 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2974 			tls->next_seqno += ktls_batched_records(n);
2975 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2976 		}
2977 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2978 	} else
2979 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2980 
2981 	running = wq->running;
2982 	mtx_unlock(&wq->mtx);
2983 	if (!running)
2984 		wakeup(wq);
2985 	counter_u64_add(ktls_cnt_tx_queued, queued);
2986 }
2987 
2988 /*
2989  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2990  * the pages from the file and replace them with the anonymous pages
2991  * allocated in ktls_encrypt_record().
2992  */
2993 static void
2994 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2995 {
2996 	int i;
2997 
2998 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2999 
3000 	/* Free the old pages. */
3001 	m->m_ext.ext_free(m);
3002 
3003 	/* Replace them with the new pages. */
3004 	if (state->cbuf != NULL) {
3005 		for (i = 0; i < m->m_epg_npgs; i++)
3006 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
3007 
3008 		/* Contig pages should go back to the cache. */
3009 		m->m_ext.ext_free = ktls_free_mext_contig;
3010 	} else {
3011 		for (i = 0; i < m->m_epg_npgs; i++)
3012 			m->m_epg_pa[i] = state->parray[i];
3013 
3014 		/* Use the basic free routine. */
3015 		m->m_ext.ext_free = mb_free_mext_pgs;
3016 	}
3017 
3018 	/* Pages are now writable. */
3019 	m->m_epg_flags |= EPG_FLAG_ANON;
3020 }
3021 
3022 static __noinline void
3023 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
3024 {
3025 	struct ktls_ocf_encrypt_state state;
3026 	struct ktls_session *tls;
3027 	struct socket *so;
3028 	struct mbuf *m;
3029 	int error, npages, total_pages;
3030 
3031 	so = top->m_epg_so;
3032 	tls = top->m_epg_tls;
3033 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3034 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3035 #ifdef INVARIANTS
3036 	top->m_epg_so = NULL;
3037 #endif
3038 	total_pages = top->m_epg_enc_cnt;
3039 	npages = 0;
3040 
3041 	/*
3042 	 * Encrypt the TLS records in the chain of mbufs starting with
3043 	 * 'top'.  'total_pages' gives us a total count of pages and is
3044 	 * used to know when we have finished encrypting the TLS
3045 	 * records originally queued with 'top'.
3046 	 *
3047 	 * NB: These mbufs are queued in the socket buffer and
3048 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
3049 	 * socket buffer lock is not held while traversing this chain.
3050 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
3051 	 * pointers should be stable.  However, the 'm_next' of the
3052 	 * last mbuf encrypted is not necessarily NULL.  It can point
3053 	 * to other mbufs appended while 'top' was on the TLS work
3054 	 * queue.
3055 	 *
3056 	 * Each mbuf holds an entire TLS record.
3057 	 */
3058 	error = 0;
3059 	for (m = top; npages != total_pages; m = m->m_next) {
3060 		KASSERT(m->m_epg_tls == tls,
3061 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3062 		    tls, m->m_epg_tls));
3063 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3064 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3065 		    total_pages, m));
3066 
3067 		error = ktls_encrypt_record(wq, m, tls, &state);
3068 		if (error) {
3069 			counter_u64_add(ktls_offload_failed_crypto, 1);
3070 			break;
3071 		}
3072 
3073 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3074 			ktls_finish_nonanon(m, &state);
3075 		m->m_flags |= M_RDONLY;
3076 
3077 		npages += m->m_epg_nrdy;
3078 
3079 		/*
3080 		 * Drop a reference to the session now that it is no
3081 		 * longer needed.  Existing code depends on encrypted
3082 		 * records having no associated session vs
3083 		 * yet-to-be-encrypted records having an associated
3084 		 * session.
3085 		 */
3086 		m->m_epg_tls = NULL;
3087 		ktls_free(tls);
3088 	}
3089 
3090 	CURVNET_SET(so->so_vnet);
3091 	if (error == 0) {
3092 		(void)so->so_proto->pr_ready(so, top, npages);
3093 	} else {
3094 		ktls_drop(so, EIO);
3095 		mb_free_notready(top, total_pages);
3096 	}
3097 
3098 	sorele(so);
3099 	CURVNET_RESTORE();
3100 }
3101 
3102 void
3103 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3104 {
3105 	struct ktls_session *tls;
3106 	struct socket *so;
3107 	struct mbuf *m;
3108 	int npages;
3109 
3110 	m = state->m;
3111 
3112 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3113 		ktls_finish_nonanon(m, state);
3114 	m->m_flags |= M_RDONLY;
3115 
3116 	so = state->so;
3117 	free(state, M_KTLS);
3118 
3119 	/*
3120 	 * Drop a reference to the session now that it is no longer
3121 	 * needed.  Existing code depends on encrypted records having
3122 	 * no associated session vs yet-to-be-encrypted records having
3123 	 * an associated session.
3124 	 */
3125 	tls = m->m_epg_tls;
3126 	m->m_epg_tls = NULL;
3127 	ktls_free(tls);
3128 
3129 	if (error != 0)
3130 		counter_u64_add(ktls_offload_failed_crypto, 1);
3131 
3132 	CURVNET_SET(so->so_vnet);
3133 	npages = m->m_epg_nrdy;
3134 
3135 	if (error == 0) {
3136 		(void)so->so_proto->pr_ready(so, m, npages);
3137 	} else {
3138 		ktls_drop(so, EIO);
3139 		mb_free_notready(m, npages);
3140 	}
3141 
3142 	sorele(so);
3143 	CURVNET_RESTORE();
3144 }
3145 
3146 /*
3147  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3148  * (coprocessors) where encryption does not use host CPU resources and
3149  * it can be beneficial to queue more requests than CPUs.
3150  */
3151 static __noinline void
3152 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3153 {
3154 	struct ktls_ocf_encrypt_state *state;
3155 	struct ktls_session *tls;
3156 	struct socket *so;
3157 	struct mbuf *m, *n;
3158 	int error, mpages, npages, total_pages;
3159 
3160 	so = top->m_epg_so;
3161 	tls = top->m_epg_tls;
3162 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3163 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3164 #ifdef INVARIANTS
3165 	top->m_epg_so = NULL;
3166 #endif
3167 	total_pages = top->m_epg_enc_cnt;
3168 	npages = 0;
3169 
3170 	error = 0;
3171 	for (m = top; npages != total_pages; m = n) {
3172 		KASSERT(m->m_epg_tls == tls,
3173 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3174 		    tls, m->m_epg_tls));
3175 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3176 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3177 		    total_pages, m));
3178 
3179 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3180 		soref(so);
3181 		state->so = so;
3182 		state->m = m;
3183 
3184 		mpages = m->m_epg_nrdy;
3185 		n = m->m_next;
3186 
3187 		error = ktls_encrypt_record(wq, m, tls, state);
3188 		if (error) {
3189 			counter_u64_add(ktls_offload_failed_crypto, 1);
3190 			free(state, M_KTLS);
3191 			CURVNET_SET(so->so_vnet);
3192 			sorele(so);
3193 			CURVNET_RESTORE();
3194 			break;
3195 		}
3196 
3197 		npages += mpages;
3198 	}
3199 
3200 	CURVNET_SET(so->so_vnet);
3201 	if (error != 0) {
3202 		ktls_drop(so, EIO);
3203 		mb_free_notready(m, total_pages - npages);
3204 	}
3205 
3206 	sorele(so);
3207 	CURVNET_RESTORE();
3208 }
3209 
3210 static int
3211 ktls_bind_domain(int domain)
3212 {
3213 	int error;
3214 
3215 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3216 	if (error != 0)
3217 		return (error);
3218 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3219 	return (0);
3220 }
3221 
3222 static void
3223 ktls_reclaim_thread(void *ctx)
3224 {
3225 	struct ktls_domain_info *ktls_domain = ctx;
3226 	struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3227 	struct sysctl_oid *oid;
3228 	char name[80];
3229 	int error, domain;
3230 
3231 	domain = ktls_domain - ktls_domains;
3232 	if (bootverbose)
3233 		printf("Starting KTLS reclaim thread for domain %d\n", domain);
3234 	error = ktls_bind_domain(domain);
3235 	if (error)
3236 		printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3237 		    domain, error);
3238 	snprintf(name, sizeof(name), "domain%d", domain);
3239 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3240 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3241 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3242 	    CTLFLAG_RD,  &sc->reclaims, 0, "buffers reclaimed");
3243 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3244 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3245 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3246 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3247 
3248 	for (;;) {
3249 		atomic_store_int(&sc->running, 0);
3250 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3251 		atomic_store_int(&sc->running, 1);
3252 		sc->wakeups++;
3253 		/*
3254 		 * Below we attempt to reclaim ktls_max_reclaim
3255 		 * buffers using vm_page_reclaim_contig_domain_ext().
3256 		 * We do this here, as this function can take several
3257 		 * seconds to scan all of memory and it does not
3258 		 * matter if this thread pauses for a while.  If we
3259 		 * block a ktls worker thread, we risk developing
3260 		 * backlogs of buffers to be encrypted, leading to
3261 		 * surges of traffic and potential NIC output drops.
3262 		 */
3263 		if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3264 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3265 		    ktls_max_reclaim) != 0) {
3266 			vm_wait_domain(domain);
3267 		} else {
3268 			sc->reclaims += ktls_max_reclaim;
3269 		}
3270 	}
3271 }
3272 
3273 static void
3274 ktls_work_thread(void *ctx)
3275 {
3276 	struct ktls_wq *wq = ctx;
3277 	struct mbuf *m, *n;
3278 	struct socket *so, *son;
3279 	STAILQ_HEAD(, mbuf) local_m_head;
3280 	STAILQ_HEAD(, socket) local_so_head;
3281 	int cpu;
3282 
3283 	cpu = wq - ktls_wq;
3284 	if (bootverbose)
3285 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3286 
3287 	/*
3288 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3289 	 * we bind to the NUMA domain instead.
3290 	 */
3291 	if (ktls_bind_threads) {
3292 		int error;
3293 
3294 		if (ktls_bind_threads > 1) {
3295 			struct pcpu *pc = pcpu_find(cpu);
3296 
3297 			error = ktls_bind_domain(pc->pc_domain);
3298 		} else {
3299 			cpuset_t mask;
3300 
3301 			CPU_SETOF(cpu, &mask);
3302 			error = cpuset_setthread(curthread->td_tid, &mask);
3303 		}
3304 		if (error)
3305 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3306 				cpu, error);
3307 	}
3308 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3309 	fpu_kern_thread(0);
3310 #endif
3311 	for (;;) {
3312 		mtx_lock(&wq->mtx);
3313 		while (STAILQ_EMPTY(&wq->m_head) &&
3314 		    STAILQ_EMPTY(&wq->so_head)) {
3315 			wq->running = false;
3316 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3317 			wq->running = true;
3318 		}
3319 
3320 		STAILQ_INIT(&local_m_head);
3321 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3322 		STAILQ_INIT(&local_so_head);
3323 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3324 		mtx_unlock(&wq->mtx);
3325 
3326 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3327 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3328 				ktls_free(m->m_epg_tls);
3329 				m_free_raw(m);
3330 			} else {
3331 				if (m->m_epg_tls->sync_dispatch)
3332 					ktls_encrypt(wq, m);
3333 				else
3334 					ktls_encrypt_async(wq, m);
3335 				counter_u64_add(ktls_cnt_tx_queued, -1);
3336 			}
3337 		}
3338 
3339 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3340 			ktls_decrypt(so);
3341 			counter_u64_add(ktls_cnt_rx_queued, -1);
3342 		}
3343 	}
3344 }
3345 
3346 static void
3347 ktls_disable_ifnet_help(void *context, int pending __unused)
3348 {
3349 	struct ktls_session *tls;
3350 	struct inpcb *inp;
3351 	struct tcpcb *tp;
3352 	struct socket *so;
3353 	int err;
3354 
3355 	tls = context;
3356 	inp = tls->inp;
3357 	if (inp == NULL)
3358 		return;
3359 	INP_WLOCK(inp);
3360 	so = inp->inp_socket;
3361 	MPASS(so != NULL);
3362 	if (inp->inp_flags & INP_DROPPED) {
3363 		goto out;
3364 	}
3365 
3366 	if (so->so_snd.sb_tls_info != NULL)
3367 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3368 	else
3369 		err = ENXIO;
3370 	if (err == 0) {
3371 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3372 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3373 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3374 		    (tp = intotcpcb(inp)) != NULL &&
3375 		    tp->t_fb->tfb_hwtls_change != NULL)
3376 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3377 	} else {
3378 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3379 	}
3380 
3381 out:
3382 	CURVNET_SET(so->so_vnet);
3383 	sorele(so);
3384 	CURVNET_RESTORE();
3385 	INP_WUNLOCK(inp);
3386 	ktls_free(tls);
3387 }
3388 
3389 /*
3390  * Called when re-transmits are becoming a substantial portion of the
3391  * sends on this connection.  When this happens, we transition the
3392  * connection to software TLS.  This is needed because most inline TLS
3393  * NICs keep crypto state only for in-order transmits.  This means
3394  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3395  * re-DMA the entire TLS record up to and including the current
3396  * segment.  This means that when re-transmitting the last ~1448 byte
3397  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3398  * of magnitude more data than we are sending.  This can cause the
3399  * PCIe link to saturate well before the network, which can cause
3400  * output drops, and a general loss of capacity.
3401  */
3402 void
3403 ktls_disable_ifnet(void *arg)
3404 {
3405 	struct tcpcb *tp;
3406 	struct inpcb *inp;
3407 	struct socket *so;
3408 	struct ktls_session *tls;
3409 
3410 	tp = arg;
3411 	inp = tptoinpcb(tp);
3412 	INP_WLOCK_ASSERT(inp);
3413 	so = inp->inp_socket;
3414 	SOCK_LOCK(so);
3415 	tls = so->so_snd.sb_tls_info;
3416 	if (tp->t_nic_ktls_xmit_dis == 1) {
3417 		SOCK_UNLOCK(so);
3418 		return;
3419 	}
3420 
3421 	/*
3422 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3423 	 * ifnet can only be done once per connection, so we never want
3424 	 * to do it again
3425 	 */
3426 
3427 	(void)ktls_hold(tls);
3428 	soref(so);
3429 	tp->t_nic_ktls_xmit_dis = 1;
3430 	SOCK_UNLOCK(so);
3431 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3432 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3433 }
3434