2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2014-2019 Netflix Inc.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 #include "opt_inet6.h"
35 #include <sys/param.h>
36 #include <sys/kernel.h>
40 #include <sys/mutex.h>
41 #include <sys/rmlock.h>
43 #include <sys/protosw.h>
44 #include <sys/refcount.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/taskqueue.h>
50 #include <sys/kthread.h>
52 #include <sys/vmmeter.h>
53 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
54 #include <machine/pcb.h>
56 #include <machine/vmparam.h>
58 #include <net/if_var.h>
60 #include <net/netisr.h>
62 #include <net/rss_config.h>
64 #include <net/route.h>
65 #include <net/route/nhop.h>
66 #if defined(INET) || defined(INET6)
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
70 #include <netinet/tcp_var.h>
72 #include <netinet/tcp_offload.h>
74 #include <opencrypto/xform.h>
75 #include <vm/uma_dbg.h>
77 #include <vm/vm_pageout.h>
78 #include <vm/vm_page.h>
82 STAILQ_HEAD(, mbuf_ext_pgs) head;
84 } __aligned(CACHE_LINE_SIZE);
86 static struct ktls_wq *ktls_wq;
87 static struct proc *ktls_proc;
88 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
89 static struct rmlock ktls_backends_lock;
90 static uma_zone_t ktls_session_zone;
91 static uint16_t ktls_cpuid_lookup[MAXCPU];
93 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
94 "Kernel TLS offload");
95 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
96 "Kernel TLS offload stats");
98 static int ktls_allow_unload;
99 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
100 &ktls_allow_unload, 0, "Allow software crypto modules to unload");
103 static int ktls_bind_threads = 1;
105 static int ktls_bind_threads;
107 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
108 &ktls_bind_threads, 0,
109 "Bind crypto threads to cores or domains at boot");
111 static u_int ktls_maxlen = 16384;
112 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
113 &ktls_maxlen, 0, "Maximum TLS record size");
115 static int ktls_number_threads;
116 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
117 &ktls_number_threads, 0,
118 "Number of TLS threads in thread-pool");
120 static bool ktls_offload_enable;
121 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW,
122 &ktls_offload_enable, 0,
123 "Enable support for kernel TLS offload");
125 static bool ktls_cbc_enable = true;
126 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW,
128 "Enable Support of AES-CBC crypto for kernel TLS");
130 static counter_u64_t ktls_tasks_active;
131 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
132 &ktls_tasks_active, "Number of active tasks");
134 static counter_u64_t ktls_cnt_on;
135 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD,
136 &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto");
138 static counter_u64_t ktls_offload_total;
139 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
140 CTLFLAG_RD, &ktls_offload_total,
141 "Total successful TLS setups (parameters set)");
143 static counter_u64_t ktls_offload_enable_calls;
144 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
145 CTLFLAG_RD, &ktls_offload_enable_calls,
146 "Total number of TLS enable calls made");
148 static counter_u64_t ktls_offload_active;
149 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
150 &ktls_offload_active, "Total Active TLS sessions");
152 static counter_u64_t ktls_offload_failed_crypto;
153 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
154 &ktls_offload_failed_crypto, "Total TLS crypto failures");
156 static counter_u64_t ktls_switch_to_ifnet;
157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
158 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
160 static counter_u64_t ktls_switch_to_sw;
161 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
162 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
164 static counter_u64_t ktls_switch_failed;
165 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
166 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
168 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
169 "Software TLS session stats");
170 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
171 "Hardware (ifnet) TLS session stats");
173 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
174 "TOE TLS session stats");
177 static counter_u64_t ktls_sw_cbc;
178 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
179 "Active number of software TLS sessions using AES-CBC");
181 static counter_u64_t ktls_sw_gcm;
182 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
183 "Active number of software TLS sessions using AES-GCM");
185 static counter_u64_t ktls_ifnet_cbc;
186 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
188 "Active number of ifnet TLS sessions using AES-CBC");
190 static counter_u64_t ktls_ifnet_gcm;
191 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
193 "Active number of ifnet TLS sessions using AES-GCM");
195 static counter_u64_t ktls_ifnet_reset;
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
197 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
199 static counter_u64_t ktls_ifnet_reset_dropped;
200 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
201 &ktls_ifnet_reset_dropped,
202 "TLS sessions dropped after failing to update ifnet send tag");
204 static counter_u64_t ktls_ifnet_reset_failed;
205 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
206 &ktls_ifnet_reset_failed,
207 "TLS sessions that failed to allocate a new ifnet send tag");
209 static int ktls_ifnet_permitted;
210 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
211 &ktls_ifnet_permitted, 1,
212 "Whether to permit hardware (ifnet) TLS sessions");
215 static counter_u64_t ktls_toe_cbc;
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
218 "Active number of TOE TLS sessions using AES-CBC");
220 static counter_u64_t ktls_toe_gcm;
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
223 "Active number of TOE TLS sessions using AES-GCM");
226 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
228 static void ktls_cleanup(struct ktls_session *tls);
229 #if defined(INET) || defined(INET6)
230 static void ktls_reset_send_tag(void *context, int pending);
232 static void ktls_work_thread(void *ctx);
235 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
237 struct ktls_crypto_backend *curr_be, *tmp;
239 if (be->api_version != KTLS_API_VERSION) {
240 printf("KTLS: API version mismatch (%d vs %d) for %s\n",
241 be->api_version, KTLS_API_VERSION,
246 rm_wlock(&ktls_backends_lock);
247 printf("KTLS: Registering crypto method %s with prio %d\n",
249 if (LIST_EMPTY(&ktls_backends)) {
250 LIST_INSERT_HEAD(&ktls_backends, be, next);
252 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
253 if (curr_be->prio < be->prio) {
254 LIST_INSERT_BEFORE(curr_be, be, next);
257 if (LIST_NEXT(curr_be, next) == NULL) {
258 LIST_INSERT_AFTER(curr_be, be, next);
263 rm_wunlock(&ktls_backends_lock);
268 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
270 struct ktls_crypto_backend *tmp;
273 * Don't error if the backend isn't registered. This permits
274 * MOD_UNLOAD handlers to use this function unconditionally.
276 rm_wlock(&ktls_backends_lock);
277 LIST_FOREACH(tmp, &ktls_backends, next) {
282 rm_wunlock(&ktls_backends_lock);
286 if (!ktls_allow_unload) {
287 rm_wunlock(&ktls_backends_lock);
289 "KTLS: Deregistering crypto method %s is not supported\n",
295 rm_wunlock(&ktls_backends_lock);
299 LIST_REMOVE(be, next);
300 rm_wunlock(&ktls_backends_lock);
304 #if defined(INET) || defined(INET6)
306 ktls_get_cpu(struct socket *so)
313 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
314 if (cpuid != NETISR_CPUID_NONE)
318 * Just use the flowid to shard connections in a repeatable
319 * fashion. Note that some crypto backends rely on the
320 * serialization provided by having the same connection use
323 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
329 ktls_init(void *dummy __unused)
336 ktls_tasks_active = counter_u64_alloc(M_WAITOK);
337 ktls_cnt_on = counter_u64_alloc(M_WAITOK);
338 ktls_offload_total = counter_u64_alloc(M_WAITOK);
339 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
340 ktls_offload_active = counter_u64_alloc(M_WAITOK);
341 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
342 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
343 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
344 ktls_switch_failed = counter_u64_alloc(M_WAITOK);
345 ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
346 ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
347 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
348 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
349 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
350 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
351 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
353 ktls_toe_cbc = counter_u64_alloc(M_WAITOK);
354 ktls_toe_gcm = counter_u64_alloc(M_WAITOK);
357 rm_init(&ktls_backends_lock, "ktls backends");
358 LIST_INIT(&ktls_backends);
360 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
363 ktls_session_zone = uma_zcreate("ktls_session",
364 sizeof(struct ktls_session),
365 NULL, NULL, NULL, NULL,
369 * Initialize the workqueues to run the TLS work. We create a
370 * work queue for each CPU.
373 STAILQ_INIT(&ktls_wq[i].head);
374 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
375 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
376 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
378 panic("Can't add KTLS thread %d error %d", i, error);
381 * Bind threads to cores. If ktls_bind_threads is >
382 * 1, then we bind to the NUMA domain.
384 if (ktls_bind_threads) {
385 if (ktls_bind_threads > 1) {
387 CPU_COPY(&cpuset_domain[pc->pc_domain], &mask);
391 error = cpuset_setthread(td->td_tid, &mask);
394 "Unable to bind KTLS thread for CPU %d error %d",
397 ktls_cpuid_lookup[ktls_number_threads] = i;
398 ktls_number_threads++;
400 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
402 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
404 #if defined(INET) || defined(INET6)
406 ktls_create_session(struct socket *so, struct tls_enable *en,
407 struct ktls_session **tlsp)
409 struct ktls_session *tls;
412 /* Only TLS 1.0 - 1.3 are supported. */
413 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
415 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
416 en->tls_vminor > TLS_MINOR_VER_THREE)
419 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
421 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
423 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
426 /* All supported algorithms require a cipher key. */
427 if (en->cipher_key_len == 0)
430 /* No flags are currently supported. */
434 /* Common checks for supported algorithms. */
435 switch (en->cipher_algorithm) {
436 case CRYPTO_AES_NIST_GCM_16:
438 * auth_algorithm isn't used, but permit GMAC values
441 switch (en->auth_algorithm) {
443 #ifdef COMPAT_FREEBSD12
444 /* XXX: Really 13.0-current COMPAT. */
445 case CRYPTO_AES_128_NIST_GMAC:
446 case CRYPTO_AES_192_NIST_GMAC:
447 case CRYPTO_AES_256_NIST_GMAC:
453 if (en->auth_key_len != 0)
455 if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
456 en->iv_len != TLS_AEAD_GCM_LEN) ||
457 (en->tls_vminor == TLS_MINOR_VER_THREE &&
458 en->iv_len != TLS_1_3_GCM_IV_LEN))
462 switch (en->auth_algorithm) {
463 case CRYPTO_SHA1_HMAC:
465 * TLS 1.0 requires an implicit IV. TLS 1.1+
466 * all use explicit IVs.
468 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
469 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
475 case CRYPTO_SHA2_256_HMAC:
476 case CRYPTO_SHA2_384_HMAC:
477 /* Ignore any supplied IV. */
483 if (en->auth_key_len == 0)
490 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
492 counter_u64_add(ktls_offload_active, 1);
494 refcount_init(&tls->refcount, 1);
495 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
497 tls->wq_index = ktls_get_cpu(so);
499 tls->params.cipher_algorithm = en->cipher_algorithm;
500 tls->params.auth_algorithm = en->auth_algorithm;
501 tls->params.tls_vmajor = en->tls_vmajor;
502 tls->params.tls_vminor = en->tls_vminor;
503 tls->params.flags = en->flags;
504 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
506 /* Set the header and trailer lengths. */
507 tls->params.tls_hlen = sizeof(struct tls_record_layer);
508 switch (en->cipher_algorithm) {
509 case CRYPTO_AES_NIST_GCM_16:
511 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
512 * nonce. TLS 1.3 uses a 12 byte implicit IV.
514 if (en->tls_vminor < TLS_MINOR_VER_THREE)
515 tls->params.tls_hlen += sizeof(uint64_t);
516 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
519 * TLS 1.3 includes optional padding which we
520 * do not support, and also puts the "real" record
521 * type at the end of the encrypted data.
523 if (en->tls_vminor == TLS_MINOR_VER_THREE)
524 tls->params.tls_tlen += sizeof(uint8_t);
526 tls->params.tls_bs = 1;
529 switch (en->auth_algorithm) {
530 case CRYPTO_SHA1_HMAC:
531 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
532 /* Implicit IV, no nonce. */
534 tls->params.tls_hlen += AES_BLOCK_LEN;
536 tls->params.tls_tlen = AES_BLOCK_LEN +
539 case CRYPTO_SHA2_256_HMAC:
540 tls->params.tls_hlen += AES_BLOCK_LEN;
541 tls->params.tls_tlen = AES_BLOCK_LEN +
544 case CRYPTO_SHA2_384_HMAC:
545 tls->params.tls_hlen += AES_BLOCK_LEN;
546 tls->params.tls_tlen = AES_BLOCK_LEN +
550 panic("invalid hmac");
552 tls->params.tls_bs = AES_BLOCK_LEN;
555 panic("invalid cipher");
558 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
559 ("TLS header length too long: %d", tls->params.tls_hlen));
560 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
561 ("TLS trailer length too long: %d", tls->params.tls_tlen));
563 if (en->auth_key_len != 0) {
564 tls->params.auth_key_len = en->auth_key_len;
565 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
567 error = copyin(en->auth_key, tls->params.auth_key,
573 tls->params.cipher_key_len = en->cipher_key_len;
574 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
575 error = copyin(en->cipher_key, tls->params.cipher_key,
581 * This holds the implicit portion of the nonce for GCM and
582 * the initial implicit IV for TLS 1.0. The explicit portions
583 * of the IV are generated in ktls_frame().
585 if (en->iv_len != 0) {
586 tls->params.iv_len = en->iv_len;
587 error = copyin(en->iv, tls->params.iv, en->iv_len);
592 * For TLS 1.2, generate an 8-byte nonce as a counter
593 * to generate unique explicit IVs.
595 * Store this counter in the last 8 bytes of the IV
596 * array so that it is 8-byte aligned.
598 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
599 en->tls_vminor == TLS_MINOR_VER_TWO)
600 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
611 static struct ktls_session *
612 ktls_clone_session(struct ktls_session *tls)
614 struct ktls_session *tls_new;
616 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
618 counter_u64_add(ktls_offload_active, 1);
620 refcount_init(&tls_new->refcount, 1);
622 /* Copy fields from existing session. */
623 tls_new->params = tls->params;
624 tls_new->wq_index = tls->wq_index;
626 /* Deep copy keys. */
627 if (tls_new->params.auth_key != NULL) {
628 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
630 memcpy(tls_new->params.auth_key, tls->params.auth_key,
631 tls->params.auth_key_len);
634 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
636 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
637 tls->params.cipher_key_len);
644 ktls_cleanup(struct ktls_session *tls)
647 counter_u64_add(ktls_offload_active, -1);
649 case TCP_TLS_MODE_SW:
650 MPASS(tls->be != NULL);
651 switch (tls->params.cipher_algorithm) {
653 counter_u64_add(ktls_sw_cbc, -1);
655 case CRYPTO_AES_NIST_GCM_16:
656 counter_u64_add(ktls_sw_gcm, -1);
661 case TCP_TLS_MODE_IFNET:
662 switch (tls->params.cipher_algorithm) {
664 counter_u64_add(ktls_ifnet_cbc, -1);
666 case CRYPTO_AES_NIST_GCM_16:
667 counter_u64_add(ktls_ifnet_gcm, -1);
670 m_snd_tag_rele(tls->snd_tag);
673 case TCP_TLS_MODE_TOE:
674 switch (tls->params.cipher_algorithm) {
676 counter_u64_add(ktls_toe_cbc, -1);
678 case CRYPTO_AES_NIST_GCM_16:
679 counter_u64_add(ktls_toe_gcm, -1);
685 if (tls->params.auth_key != NULL) {
686 explicit_bzero(tls->params.auth_key, tls->params.auth_key_len);
687 free(tls->params.auth_key, M_KTLS);
688 tls->params.auth_key = NULL;
689 tls->params.auth_key_len = 0;
691 if (tls->params.cipher_key != NULL) {
692 explicit_bzero(tls->params.cipher_key,
693 tls->params.cipher_key_len);
694 free(tls->params.cipher_key, M_KTLS);
695 tls->params.cipher_key = NULL;
696 tls->params.cipher_key_len = 0;
698 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
701 #if defined(INET) || defined(INET6)
705 ktls_try_toe(struct socket *so, struct ktls_session *tls)
713 if (inp->inp_flags2 & INP_FREED) {
717 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
721 if (inp->inp_socket == NULL) {
726 if (tp->tod == NULL) {
731 error = tcp_offload_alloc_tls_session(tp, tls);
734 tls->mode = TCP_TLS_MODE_TOE;
735 switch (tls->params.cipher_algorithm) {
737 counter_u64_add(ktls_toe_cbc, 1);
739 case CRYPTO_AES_NIST_GCM_16:
740 counter_u64_add(ktls_toe_gcm, 1);
749 * Common code used when first enabling ifnet TLS on a connection or
750 * when allocating a new ifnet TLS session due to a routing change.
751 * This function allocates a new TLS send tag on whatever interface
752 * the connection is currently routed over.
755 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
756 struct m_snd_tag **mstp)
758 union if_snd_tag_alloc_params params;
760 struct nhop_object *nh;
765 if (inp->inp_flags2 & INP_FREED) {
769 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
773 if (inp->inp_socket == NULL) {
780 * Check administrative controls on ifnet TLS to determine if
781 * ifnet TLS should be denied.
783 * - Always permit 'force' requests.
784 * - ktls_ifnet_permitted == 0: always deny.
786 if (!force && ktls_ifnet_permitted == 0) {
792 * XXX: Use the cached route in the inpcb to find the
793 * interface. This should perhaps instead use
794 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
795 * enabled after a connection has completed key negotiation in
796 * userland, the cached route will be present in practice.
798 nh = inp->inp_route.ro_nh;
806 params.hdr.type = IF_SND_TAG_TYPE_TLS;
807 params.hdr.flowid = inp->inp_flowid;
808 params.hdr.flowtype = inp->inp_flowtype;
809 params.hdr.numa_domain = inp->inp_numa_domain;
810 params.tls.inp = inp;
811 params.tls.tls = tls;
814 if (ifp->if_snd_tag_alloc == NULL) {
818 if ((ifp->if_capenable & IFCAP_NOMAP) == 0) {
822 if (inp->inp_vflag & INP_IPV6) {
823 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
828 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
833 error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp);
840 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
842 struct m_snd_tag *mst;
845 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
847 tls->mode = TCP_TLS_MODE_IFNET;
849 switch (tls->params.cipher_algorithm) {
851 counter_u64_add(ktls_ifnet_cbc, 1);
853 case CRYPTO_AES_NIST_GCM_16:
854 counter_u64_add(ktls_ifnet_gcm, 1);
862 ktls_try_sw(struct socket *so, struct ktls_session *tls)
864 struct rm_priotracker prio;
865 struct ktls_crypto_backend *be;
868 * Choose the best software crypto backend. Backends are
869 * stored in sorted priority order (larget value == most
870 * important at the head of the list), so this just stops on
871 * the first backend that claims the session by returning
874 if (ktls_allow_unload)
875 rm_rlock(&ktls_backends_lock, &prio);
876 LIST_FOREACH(be, &ktls_backends, next) {
877 if (be->try(so, tls) == 0)
879 KASSERT(tls->cipher == NULL,
880 ("ktls backend leaked a cipher pointer"));
883 if (ktls_allow_unload)
887 if (ktls_allow_unload)
888 rm_runlock(&ktls_backends_lock, &prio);
891 tls->mode = TCP_TLS_MODE_SW;
892 switch (tls->params.cipher_algorithm) {
894 counter_u64_add(ktls_sw_cbc, 1);
896 case CRYPTO_AES_NIST_GCM_16:
897 counter_u64_add(ktls_sw_gcm, 1);
904 ktls_enable_tx(struct socket *so, struct tls_enable *en)
906 struct ktls_session *tls;
909 if (!ktls_offload_enable)
912 counter_u64_add(ktls_offload_enable_calls, 1);
915 * This should always be true since only the TCP socket option
916 * invokes this function.
918 if (so->so_proto->pr_protocol != IPPROTO_TCP)
922 * XXX: Don't overwrite existing sessions. We should permit
923 * this to support rekeying in the future.
925 if (so->so_snd.sb_tls_info != NULL)
928 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
931 /* TLS requires ext pgs */
932 if (mb_use_ext_pgs == 0)
935 error = ktls_create_session(so, en, &tls);
939 /* Prefer TOE -> ifnet TLS -> software TLS. */
941 error = ktls_try_toe(so, tls);
944 error = ktls_try_ifnet(so, tls, false);
946 error = ktls_try_sw(so, tls);
953 error = sblock(&so->so_snd, SBL_WAIT);
959 SOCKBUF_LOCK(&so->so_snd);
960 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
961 so->so_snd.sb_tls_info = tls;
962 if (tls->mode != TCP_TLS_MODE_SW)
963 so->so_snd.sb_flags |= SB_TLS_IFNET;
964 SOCKBUF_UNLOCK(&so->so_snd);
965 sbunlock(&so->so_snd);
967 counter_u64_add(ktls_offload_total, 1);
973 ktls_get_tx_mode(struct socket *so)
975 struct ktls_session *tls;
980 INP_WLOCK_ASSERT(inp);
981 SOCKBUF_LOCK(&so->so_snd);
982 tls = so->so_snd.sb_tls_info;
984 mode = TCP_TLS_MODE_NONE;
987 SOCKBUF_UNLOCK(&so->so_snd);
992 * Switch between SW and ifnet TLS sessions as requested.
995 ktls_set_tx_mode(struct socket *so, int mode)
997 struct ktls_session *tls, *tls_new;
1002 case TCP_TLS_MODE_SW:
1003 case TCP_TLS_MODE_IFNET:
1010 INP_WLOCK_ASSERT(inp);
1011 SOCKBUF_LOCK(&so->so_snd);
1012 tls = so->so_snd.sb_tls_info;
1014 SOCKBUF_UNLOCK(&so->so_snd);
1018 if (tls->mode == mode) {
1019 SOCKBUF_UNLOCK(&so->so_snd);
1023 tls = ktls_hold(tls);
1024 SOCKBUF_UNLOCK(&so->so_snd);
1027 tls_new = ktls_clone_session(tls);
1029 if (mode == TCP_TLS_MODE_IFNET)
1030 error = ktls_try_ifnet(so, tls_new, true);
1032 error = ktls_try_sw(so, tls_new);
1034 counter_u64_add(ktls_switch_failed, 1);
1041 error = sblock(&so->so_snd, SBL_WAIT);
1043 counter_u64_add(ktls_switch_failed, 1);
1051 * If we raced with another session change, keep the existing
1054 if (tls != so->so_snd.sb_tls_info) {
1055 counter_u64_add(ktls_switch_failed, 1);
1056 sbunlock(&so->so_snd);
1063 SOCKBUF_LOCK(&so->so_snd);
1064 so->so_snd.sb_tls_info = tls_new;
1065 if (tls_new->mode != TCP_TLS_MODE_SW)
1066 so->so_snd.sb_flags |= SB_TLS_IFNET;
1067 SOCKBUF_UNLOCK(&so->so_snd);
1068 sbunlock(&so->so_snd);
1071 * Drop two references on 'tls'. The first is for the
1072 * ktls_hold() above. The second drops the reference from the
1075 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1079 if (mode == TCP_TLS_MODE_IFNET)
1080 counter_u64_add(ktls_switch_to_ifnet, 1);
1082 counter_u64_add(ktls_switch_to_sw, 1);
1089 * Try to allocate a new TLS send tag. This task is scheduled when
1090 * ip_output detects a route change while trying to transmit a packet
1091 * holding a TLS record. If a new tag is allocated, replace the tag
1092 * in the TLS session. Subsequent packets on the connection will use
1093 * the new tag. If a new tag cannot be allocated, drop the
1097 ktls_reset_send_tag(void *context, int pending)
1099 struct epoch_tracker et;
1100 struct ktls_session *tls;
1101 struct m_snd_tag *old, *new;
1106 MPASS(pending == 1);
1112 * Free the old tag first before allocating a new one.
1113 * ip[6]_output_send() will treat a NULL send tag the same as
1114 * an ifp mismatch and drop packets until a new tag is
1117 * Write-lock the INP when changing tls->snd_tag since
1118 * ip[6]_output_send() holds a read-lock when reading the
1123 tls->snd_tag = NULL;
1126 m_snd_tag_rele(old);
1128 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1133 mtx_pool_lock(mtxpool_sleep, tls);
1134 tls->reset_pending = false;
1135 mtx_pool_unlock(mtxpool_sleep, tls);
1136 if (!in_pcbrele_wlocked(inp))
1139 counter_u64_add(ktls_ifnet_reset, 1);
1142 * XXX: Should we kick tcp_output explicitly now that
1143 * the send tag is fixed or just rely on timers?
1146 NET_EPOCH_ENTER(et);
1148 if (!in_pcbrele_wlocked(inp)) {
1149 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1150 !(inp->inp_flags & INP_DROPPED)) {
1151 tp = intotcpcb(inp);
1152 CURVNET_SET(tp->t_vnet);
1153 tp = tcp_drop(tp, ECONNABORTED);
1157 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1163 counter_u64_add(ktls_ifnet_reset_failed, 1);
1166 * Leave reset_pending true to avoid future tasks while
1167 * the socket goes away.
1175 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1181 INP_LOCK_ASSERT(inp);
1184 * See if we should schedule a task to update the send tag for
1187 mtx_pool_lock(mtxpool_sleep, tls);
1188 if (!tls->reset_pending) {
1189 (void) ktls_hold(tls);
1192 tls->reset_pending = true;
1193 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1195 mtx_pool_unlock(mtxpool_sleep, tls);
1201 ktls_destroy(struct ktls_session *tls)
1203 struct rm_priotracker prio;
1206 if (tls->be != NULL && ktls_allow_unload) {
1207 rm_rlock(&ktls_backends_lock, &prio);
1208 tls->be->use_count--;
1209 rm_runlock(&ktls_backends_lock, &prio);
1211 uma_zfree(ktls_session_zone, tls);
1215 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1217 struct mbuf_ext_pgs *pgs;
1219 for (; m != NULL; m = m->m_next) {
1220 KASSERT((m->m_flags & M_NOMAP) != 0,
1221 ("ktls_seq: mapped mbuf %p", m));
1223 pgs = &m->m_ext_pgs;
1224 pgs->seqno = sb->sb_tls_seqno;
1230 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1231 * mbuf in the chain must be an unmapped mbuf. The payload of the
1232 * mbuf must be populated with the payload of each TLS record.
1234 * The record_type argument specifies the TLS record type used when
1235 * populating the TLS header.
1237 * The enq_count argument on return is set to the number of pages of
1238 * payload data for this entire chain that need to be encrypted via SW
1239 * encryption. The returned value should be passed to ktls_enqueue
1240 * when scheduling encryption of this chain of mbufs.
1243 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1244 uint8_t record_type)
1246 struct tls_record_layer *tlshdr;
1248 struct mbuf_ext_pgs *pgs;
1253 maxlen = tls->params.max_frame_len;
1255 for (m = top; m != NULL; m = m->m_next) {
1257 * All mbufs in the chain should be non-empty TLS
1258 * records whose payload does not exceed the maximum
1261 KASSERT(m->m_len <= maxlen && m->m_len > 0,
1262 ("ktls_frame: m %p len %d\n", m, m->m_len));
1264 * TLS frames require unmapped mbufs to store session
1267 KASSERT((m->m_flags & M_NOMAP) != 0,
1268 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1271 pgs = &m->m_ext_pgs;
1273 /* Save a reference to the session. */
1274 pgs->tls = ktls_hold(tls);
1276 pgs->hdr_len = tls->params.tls_hlen;
1277 pgs->trail_len = tls->params.tls_tlen;
1278 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1282 * AES-CBC pads messages to a multiple of the
1283 * block size. Note that the padding is
1284 * applied after the digest and the encryption
1285 * is done on the "plaintext || mac || padding".
1286 * At least one byte of padding is always
1289 * Compute the final trailer length assuming
1290 * at most one block of padding.
1291 * tls->params.sb_tls_tlen is the maximum
1292 * possible trailer length (padding + digest).
1293 * delta holds the number of excess padding
1294 * bytes if the maximum were used. Those
1295 * extra bytes are removed.
1297 bs = tls->params.tls_bs;
1298 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1299 pgs->trail_len -= delta;
1301 m->m_len += pgs->hdr_len + pgs->trail_len;
1303 /* Populate the TLS header. */
1304 tlshdr = (void *)pgs->m_epg_hdr;
1305 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1308 * TLS 1.3 masquarades as TLS 1.2 with a record type
1309 * of TLS_RLTYPE_APP.
1311 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1312 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1313 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1314 tlshdr->tls_type = TLS_RLTYPE_APP;
1315 /* save the real record type for later */
1316 pgs->record_type = record_type;
1317 pgs->m_epg_trail[0] = record_type;
1319 tlshdr->tls_vminor = tls->params.tls_vminor;
1320 tlshdr->tls_type = record_type;
1322 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1325 * Store nonces / explicit IVs after the end of the
1328 * For GCM with TLS 1.2, an 8 byte nonce is copied
1329 * from the end of the IV. The nonce is then
1330 * incremented for use by the next record.
1332 * For CBC, a random nonce is inserted for TLS 1.1+.
1334 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1335 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1336 noncep = (uint64_t *)(tls->params.iv + 8);
1337 be64enc(tlshdr + 1, *noncep);
1339 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1340 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1341 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1344 * When using SW encryption, mark the mbuf not ready.
1345 * It will be marked ready via sbready() after the
1346 * record has been encrypted.
1348 * When using ifnet TLS, unencrypted TLS records are
1349 * sent down the stack to the NIC.
1351 if (tls->mode == TCP_TLS_MODE_SW) {
1352 m->m_flags |= M_NOTREADY;
1353 pgs->nrdy = pgs->npgs;
1354 *enq_cnt += pgs->npgs;
1360 ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs)
1365 /* Mark it for freeing. */
1367 wq = &ktls_wq[pgs->tls->wq_index];
1369 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1370 running = wq->running;
1371 mtx_unlock(&wq->mtx);
1377 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1379 struct mbuf_ext_pgs *pgs;
1383 KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1384 (M_NOMAP | M_NOTREADY)),
1385 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1386 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1388 pgs = &m->m_ext_pgs;
1390 KASSERT(pgs->tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1392 pgs->enc_cnt = page_count;
1396 * Save a pointer to the socket. The caller is responsible
1397 * for taking an additional reference via soref().
1401 wq = &ktls_wq[pgs->tls->wq_index];
1403 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1404 running = wq->running;
1405 mtx_unlock(&wq->mtx);
1408 counter_u64_add(ktls_cnt_on, 1);
1411 static __noinline void
1412 ktls_encrypt(struct mbuf_ext_pgs *pgs)
1414 struct ktls_session *tls;
1416 struct mbuf *m, *top;
1417 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1418 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1419 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1421 int error, i, len, npages, off, total_pages;
1427 KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs));
1428 KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs));
1433 total_pages = pgs->enc_cnt;
1437 * Encrypt the TLS records in the chain of mbufs starting with
1438 * 'top'. 'total_pages' gives us a total count of pages and is
1439 * used to know when we have finished encrypting the TLS
1440 * records originally queued with 'top'.
1442 * NB: These mbufs are queued in the socket buffer and
1443 * 'm_next' is traversing the mbufs in the socket buffer. The
1444 * socket buffer lock is not held while traversing this chain.
1445 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1446 * pointers should be stable. However, the 'm_next' of the
1447 * last mbuf encrypted is not necessarily NULL. It can point
1448 * to other mbufs appended while 'top' was on the TLS work
1451 * Each mbuf holds an entire TLS record.
1454 for (m = top; npages != total_pages; m = m->m_next) {
1455 pgs = &m->m_ext_pgs;
1457 KASSERT(pgs->tls == tls,
1458 ("different TLS sessions in a single mbuf chain: %p vs %p",
1460 KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1461 (M_NOMAP | M_NOTREADY),
1462 ("%p not unready & nomap mbuf (top = %p)\n", m, top));
1463 KASSERT(npages + pgs->npgs <= total_pages,
1464 ("page count mismatch: top %p, total_pages %d, m %p", top,
1468 * Generate source and destination ivoecs to pass to
1469 * the SW encryption backend. For writable mbufs, the
1470 * destination iovec is a copy of the source and
1471 * encryption is done in place. For file-backed mbufs
1472 * (from sendfile), anonymous wired pages are
1473 * allocated and assigned to the destination iovec.
1475 is_anon = (pgs->flags & MBUF_PEXT_FLAG_ANON) != 0;
1477 off = pgs->first_pg_off;
1478 for (i = 0; i < pgs->npgs; i++, off = 0) {
1479 len = mbuf_ext_pg_len(pgs, i, off);
1480 src_iov[i].iov_len = len;
1481 src_iov[i].iov_base =
1482 (char *)(void *)PHYS_TO_DMAP(pgs->m_epg_pa[i]) +
1486 dst_iov[i].iov_base = src_iov[i].iov_base;
1487 dst_iov[i].iov_len = src_iov[i].iov_len;
1491 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
1492 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
1497 parray[i] = VM_PAGE_TO_PHYS(pg);
1498 dst_iov[i].iov_base =
1499 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
1500 dst_iov[i].iov_len = len;
1505 error = (*tls->sw_encrypt)(tls,
1506 (const struct tls_record_layer *)pgs->m_epg_hdr,
1507 pgs->m_epg_trail, src_iov, dst_iov, i, pgs->seqno,
1510 counter_u64_add(ktls_offload_failed_crypto, 1);
1515 * For file-backed mbufs, release the file-backed
1516 * pages and replace them in the ext_pgs array with
1517 * the anonymous wired pages allocated above.
1520 /* Free the old pages. */
1521 m->m_ext.ext_free(m);
1523 /* Replace them with the new pages. */
1524 for (i = 0; i < pgs->npgs; i++)
1525 pgs->m_epg_pa[i] = parray[i];
1527 /* Use the basic free routine. */
1528 m->m_ext.ext_free = mb_free_mext_pgs;
1530 /* Pages are now writable. */
1531 pgs->flags |= MBUF_PEXT_FLAG_ANON;
1535 * Drop a reference to the session now that it is no
1536 * longer needed. Existing code depends on encrypted
1537 * records having no associated session vs
1538 * yet-to-be-encrypted records having an associated
1545 CURVNET_SET(so->so_vnet);
1547 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
1549 so->so_proto->pr_usrreqs->pru_abort(so);
1551 mb_free_notready(top, total_pages);
1560 ktls_work_thread(void *ctx)
1562 struct ktls_wq *wq = ctx;
1563 struct mbuf_ext_pgs *p, *n;
1564 struct ktls_session *tls;
1566 STAILQ_HEAD(, mbuf_ext_pgs) local_head;
1568 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
1573 while (STAILQ_EMPTY(&wq->head)) {
1574 wq->running = false;
1575 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
1579 STAILQ_INIT(&local_head);
1580 STAILQ_CONCAT(&local_head, &wq->head);
1581 mtx_unlock(&wq->mtx);
1583 STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) {
1584 if (p->mbuf != NULL) {
1586 counter_u64_add(ktls_cnt_on, -1);
1590 m = __containerof(p, struct mbuf, m_ext_pgs);
1591 uma_zfree(zone_mbuf, m);