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/netisr.h>
59 #include <net/rss_config.h>
61 #if defined(INET) || defined(INET6)
62 #include <netinet/in.h>
63 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_var.h>
66 #include <opencrypto/xform.h>
67 #include <vm/uma_dbg.h>
69 #include <vm/vm_pageout.h>
70 #include <vm/vm_page.h>
74 STAILQ_HEAD(, mbuf_ext_pgs) head;
76 } __aligned(CACHE_LINE_SIZE);
78 static struct ktls_wq *ktls_wq;
79 static struct proc *ktls_proc;
80 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
81 static struct rmlock ktls_backends_lock;
82 static uma_zone_t ktls_session_zone;
83 static uint16_t ktls_cpuid_lookup[MAXCPU];
85 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW, 0,
86 "Kernel TLS offload");
87 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW, 0,
88 "Kernel TLS offload stats");
90 static int ktls_allow_unload;
91 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
92 &ktls_allow_unload, 0, "Allow software crypto modules to unload");
95 static int ktls_bind_threads = 1;
97 static int ktls_bind_threads;
99 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
100 &ktls_bind_threads, 0,
101 "Bind crypto threads to cores or domains at boot");
103 static u_int ktls_maxlen = 16384;
104 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
105 &ktls_maxlen, 0, "Maximum TLS record size");
107 static int ktls_number_threads;
108 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
109 &ktls_number_threads, 0,
110 "Number of TLS threads in thread-pool");
112 static bool ktls_offload_enable;
113 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW,
114 &ktls_offload_enable, 0,
115 "Enable support for kernel TLS offload");
117 static bool ktls_cbc_enable = true;
118 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW,
120 "Enable Support of AES-CBC crypto for kernel TLS");
122 static counter_u64_t ktls_tasks_active;
123 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
124 &ktls_tasks_active, "Number of active tasks");
126 static counter_u64_t ktls_cnt_on;
127 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD,
128 &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto");
130 static counter_u64_t ktls_offload_total;
131 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
132 CTLFLAG_RD, &ktls_offload_total,
133 "Total successful TLS setups (parameters set)");
135 static counter_u64_t ktls_offload_enable_calls;
136 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
137 CTLFLAG_RD, &ktls_offload_enable_calls,
138 "Total number of TLS enable calls made");
140 static counter_u64_t ktls_offload_active;
141 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
142 &ktls_offload_active, "Total Active TLS sessions");
144 static counter_u64_t ktls_offload_failed_crypto;
145 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
146 &ktls_offload_failed_crypto, "Total TLS crypto failures");
148 static counter_u64_t ktls_switch_to_ifnet;
149 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
150 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
152 static counter_u64_t ktls_switch_to_sw;
153 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
154 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
156 static counter_u64_t ktls_switch_failed;
157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
158 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
160 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD, 0,
161 "Software TLS session stats");
162 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD, 0,
163 "Hardware (ifnet) TLS session stats");
165 static counter_u64_t ktls_sw_cbc;
166 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
167 "Active number of software TLS sessions using AES-CBC");
169 static counter_u64_t ktls_sw_gcm;
170 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
171 "Active number of software TLS sessions using AES-GCM");
173 static counter_u64_t ktls_ifnet_cbc;
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
176 "Active number of ifnet TLS sessions using AES-CBC");
178 static counter_u64_t ktls_ifnet_gcm;
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
181 "Active number of ifnet TLS sessions using AES-GCM");
183 static counter_u64_t ktls_ifnet_reset;
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
185 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
187 static counter_u64_t ktls_ifnet_reset_dropped;
188 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
189 &ktls_ifnet_reset_dropped,
190 "TLS sessions dropped after failing to update ifnet send tag");
192 static counter_u64_t ktls_ifnet_reset_failed;
193 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
194 &ktls_ifnet_reset_failed,
195 "TLS sessions that failed to allocate a new ifnet send tag");
197 static int ktls_ifnet_permitted;
198 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
199 &ktls_ifnet_permitted, 1,
200 "Whether to permit hardware (ifnet) TLS sessions");
202 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
204 static void ktls_cleanup(struct ktls_session *tls);
205 #if defined(INET) || defined(INET6)
206 static void ktls_reset_send_tag(void *context, int pending);
208 static void ktls_work_thread(void *ctx);
211 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
213 struct ktls_crypto_backend *curr_be, *tmp;
215 if (be->api_version != KTLS_API_VERSION) {
216 printf("KTLS: API version mismatch (%d vs %d) for %s\n",
217 be->api_version, KTLS_API_VERSION,
222 rm_wlock(&ktls_backends_lock);
223 printf("KTLS: Registering crypto method %s with prio %d\n",
225 if (LIST_EMPTY(&ktls_backends)) {
226 LIST_INSERT_HEAD(&ktls_backends, be, next);
228 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
229 if (curr_be->prio < be->prio) {
230 LIST_INSERT_BEFORE(curr_be, be, next);
233 if (LIST_NEXT(curr_be, next) == NULL) {
234 LIST_INSERT_AFTER(curr_be, be, next);
239 rm_wunlock(&ktls_backends_lock);
244 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
246 struct ktls_crypto_backend *tmp;
249 * Don't error if the backend isn't registered. This permits
250 * MOD_UNLOAD handlers to use this function unconditionally.
252 rm_wlock(&ktls_backends_lock);
253 LIST_FOREACH(tmp, &ktls_backends, next) {
258 rm_wunlock(&ktls_backends_lock);
262 if (!ktls_allow_unload) {
263 rm_wunlock(&ktls_backends_lock);
265 "KTLS: Deregistering crypto method %s is not supported\n",
271 rm_wunlock(&ktls_backends_lock);
275 LIST_REMOVE(be, next);
276 rm_wunlock(&ktls_backends_lock);
280 #if defined(INET) || defined(INET6)
282 ktls_get_cpu(struct socket *so)
289 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
290 if (cpuid != NETISR_CPUID_NONE)
294 * Just use the flowid to shard connections in a repeatable
295 * fashion. Note that some crypto backends rely on the
296 * serialization provided by having the same connection use
299 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
305 ktls_init(void *dummy __unused)
312 ktls_tasks_active = counter_u64_alloc(M_WAITOK);
313 ktls_cnt_on = counter_u64_alloc(M_WAITOK);
314 ktls_offload_total = counter_u64_alloc(M_WAITOK);
315 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
316 ktls_offload_active = counter_u64_alloc(M_WAITOK);
317 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
318 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
319 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
320 ktls_switch_failed = counter_u64_alloc(M_WAITOK);
321 ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
322 ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
323 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
324 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
325 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
326 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
327 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
329 rm_init(&ktls_backends_lock, "ktls backends");
330 LIST_INIT(&ktls_backends);
332 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
335 ktls_session_zone = uma_zcreate("ktls_session",
336 sizeof(struct ktls_session),
338 trash_ctor, trash_dtor, trash_init, trash_fini,
340 NULL, NULL, NULL, NULL,
345 * Initialize the workqueues to run the TLS work. We create a
346 * work queue for each CPU.
349 STAILQ_INIT(&ktls_wq[i].head);
350 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
351 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
352 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
354 panic("Can't add KTLS thread %d error %d", i, error);
357 * Bind threads to cores. If ktls_bind_threads is >
358 * 1, then we bind to the NUMA domain.
360 if (ktls_bind_threads) {
361 if (ktls_bind_threads > 1) {
363 CPU_COPY(&cpuset_domain[pc->pc_domain], &mask);
367 error = cpuset_setthread(td->td_tid, &mask);
370 "Unable to bind KTLS thread for CPU %d error %d",
373 ktls_cpuid_lookup[ktls_number_threads] = i;
374 ktls_number_threads++;
376 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
378 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
380 #if defined(INET) || defined(INET6)
382 ktls_create_session(struct socket *so, struct tls_enable *en,
383 struct ktls_session **tlsp)
385 struct ktls_session *tls;
388 /* Only TLS 1.0 - 1.2 are supported. */
389 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
391 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
392 en->tls_vminor > TLS_MINOR_VER_THREE)
395 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
397 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
399 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
402 /* All supported algorithms require a cipher key. */
403 if (en->cipher_key_len == 0)
406 /* No flags are currently supported. */
410 /* Common checks for supported algorithms. */
411 switch (en->cipher_algorithm) {
412 case CRYPTO_AES_NIST_GCM_16:
414 * auth_algorithm isn't used, but permit GMAC values
417 switch (en->auth_algorithm) {
419 case CRYPTO_AES_128_NIST_GMAC:
420 case CRYPTO_AES_192_NIST_GMAC:
421 case CRYPTO_AES_256_NIST_GMAC:
426 if (en->auth_key_len != 0)
428 if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
429 en->iv_len != TLS_AEAD_GCM_LEN) ||
430 (en->tls_vminor == TLS_MINOR_VER_THREE &&
431 en->iv_len != TLS_1_3_GCM_IV_LEN))
435 switch (en->auth_algorithm) {
436 case CRYPTO_SHA1_HMAC:
438 * TLS 1.0 requires an implicit IV. TLS 1.1+
439 * all use explicit IVs.
441 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
442 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
448 case CRYPTO_SHA2_256_HMAC:
449 case CRYPTO_SHA2_384_HMAC:
450 /* Ignore any supplied IV. */
456 if (en->auth_key_len == 0)
463 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
465 counter_u64_add(ktls_offload_active, 1);
467 refcount_init(&tls->refcount, 1);
468 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
470 tls->wq_index = ktls_get_cpu(so);
472 tls->params.cipher_algorithm = en->cipher_algorithm;
473 tls->params.auth_algorithm = en->auth_algorithm;
474 tls->params.tls_vmajor = en->tls_vmajor;
475 tls->params.tls_vminor = en->tls_vminor;
476 tls->params.flags = en->flags;
477 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
479 /* Set the header and trailer lengths. */
480 tls->params.tls_hlen = sizeof(struct tls_record_layer);
481 switch (en->cipher_algorithm) {
482 case CRYPTO_AES_NIST_GCM_16:
484 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
485 * nonce. TLS 1.3 uses a 12 byte implicit IV.
487 if (en->tls_vminor < TLS_MINOR_VER_THREE)
488 tls->params.tls_hlen += sizeof(uint64_t);
489 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
492 * TLS 1.3 includes optional padding which we
493 * do not support, and also puts the "real" record
494 * type at the end of the encrypted data.
496 if (en->tls_vminor == TLS_MINOR_VER_THREE)
497 tls->params.tls_tlen += sizeof(uint8_t);
499 tls->params.tls_bs = 1;
502 switch (en->auth_algorithm) {
503 case CRYPTO_SHA1_HMAC:
504 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
505 /* Implicit IV, no nonce. */
507 tls->params.tls_hlen += AES_BLOCK_LEN;
509 tls->params.tls_tlen = AES_BLOCK_LEN +
512 case CRYPTO_SHA2_256_HMAC:
513 tls->params.tls_hlen += AES_BLOCK_LEN;
514 tls->params.tls_tlen = AES_BLOCK_LEN +
517 case CRYPTO_SHA2_384_HMAC:
518 tls->params.tls_hlen += AES_BLOCK_LEN;
519 tls->params.tls_tlen = AES_BLOCK_LEN +
523 panic("invalid hmac");
525 tls->params.tls_bs = AES_BLOCK_LEN;
528 panic("invalid cipher");
531 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
532 ("TLS header length too long: %d", tls->params.tls_hlen));
533 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
534 ("TLS trailer length too long: %d", tls->params.tls_tlen));
536 if (en->auth_key_len != 0) {
537 tls->params.auth_key_len = en->auth_key_len;
538 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
540 error = copyin(en->auth_key, tls->params.auth_key,
546 tls->params.cipher_key_len = en->cipher_key_len;
547 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
548 error = copyin(en->cipher_key, tls->params.cipher_key,
554 * This holds the implicit portion of the nonce for GCM and
555 * the initial implicit IV for TLS 1.0. The explicit portions
556 * of the IV are generated in ktls_frame() and ktls_seq().
558 if (en->iv_len != 0) {
559 tls->params.iv_len = en->iv_len;
560 error = copyin(en->iv, tls->params.iv, en->iv_len);
573 static struct ktls_session *
574 ktls_clone_session(struct ktls_session *tls)
576 struct ktls_session *tls_new;
578 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
580 counter_u64_add(ktls_offload_active, 1);
582 refcount_init(&tls_new->refcount, 1);
584 /* Copy fields from existing session. */
585 tls_new->params = tls->params;
586 tls_new->wq_index = tls->wq_index;
588 /* Deep copy keys. */
589 if (tls_new->params.auth_key != NULL) {
590 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
592 memcpy(tls_new->params.auth_key, tls->params.auth_key,
593 tls->params.auth_key_len);
596 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
598 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
599 tls->params.cipher_key_len);
606 ktls_cleanup(struct ktls_session *tls)
609 counter_u64_add(ktls_offload_active, -1);
610 if (tls->free != NULL) {
611 MPASS(tls->be != NULL);
612 switch (tls->params.cipher_algorithm) {
614 counter_u64_add(ktls_sw_cbc, -1);
616 case CRYPTO_AES_NIST_GCM_16:
617 counter_u64_add(ktls_sw_gcm, -1);
621 } else if (tls->snd_tag != NULL) {
622 switch (tls->params.cipher_algorithm) {
624 counter_u64_add(ktls_ifnet_cbc, -1);
626 case CRYPTO_AES_NIST_GCM_16:
627 counter_u64_add(ktls_ifnet_gcm, -1);
630 m_snd_tag_rele(tls->snd_tag);
632 if (tls->params.auth_key != NULL) {
633 explicit_bzero(tls->params.auth_key, tls->params.auth_key_len);
634 free(tls->params.auth_key, M_KTLS);
635 tls->params.auth_key = NULL;
636 tls->params.auth_key_len = 0;
638 if (tls->params.cipher_key != NULL) {
639 explicit_bzero(tls->params.cipher_key,
640 tls->params.cipher_key_len);
641 free(tls->params.cipher_key, M_KTLS);
642 tls->params.cipher_key = NULL;
643 tls->params.cipher_key_len = 0;
645 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
648 #if defined(INET) || defined(INET6)
650 * Common code used when first enabling ifnet TLS on a connection or
651 * when allocating a new ifnet TLS session due to a routing change.
652 * This function allocates a new TLS send tag on whatever interface
653 * the connection is currently routed over.
656 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
657 struct m_snd_tag **mstp)
659 union if_snd_tag_alloc_params params;
666 if (inp->inp_flags2 & INP_FREED) {
670 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
674 if (inp->inp_socket == NULL) {
681 * Check administrative controls on ifnet TLS to determine if
682 * ifnet TLS should be denied.
684 * - Always permit 'force' requests.
685 * - ktls_ifnet_permitted == 0: always deny.
687 if (!force && ktls_ifnet_permitted == 0) {
693 * XXX: Use the cached route in the inpcb to find the
694 * interface. This should perhaps instead use
695 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
696 * enabled after a connection has completed key negotiation in
697 * userland, the cached route will be present in practice.
699 rt = inp->inp_route.ro_rt;
700 if (rt == NULL || rt->rt_ifp == NULL) {
707 params.hdr.type = IF_SND_TAG_TYPE_TLS;
708 params.hdr.flowid = inp->inp_flowid;
709 params.hdr.flowtype = inp->inp_flowtype;
710 params.tls.inp = inp;
711 params.tls.tls = tls;
714 if (ifp->if_snd_tag_alloc == NULL) {
718 if ((ifp->if_capenable & IFCAP_NOMAP) == 0) {
722 if (inp->inp_vflag & INP_IPV6) {
723 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
728 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
733 error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp);
740 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
742 struct m_snd_tag *mst;
745 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
748 switch (tls->params.cipher_algorithm) {
750 counter_u64_add(ktls_ifnet_cbc, 1);
752 case CRYPTO_AES_NIST_GCM_16:
753 counter_u64_add(ktls_ifnet_gcm, 1);
761 ktls_try_sw(struct socket *so, struct ktls_session *tls)
763 struct rm_priotracker prio;
764 struct ktls_crypto_backend *be;
767 * Choose the best software crypto backend. Backends are
768 * stored in sorted priority order (larget value == most
769 * important at the head of the list), so this just stops on
770 * the first backend that claims the session by returning
773 if (ktls_allow_unload)
774 rm_rlock(&ktls_backends_lock, &prio);
775 LIST_FOREACH(be, &ktls_backends, next) {
776 if (be->try(so, tls) == 0)
778 KASSERT(tls->cipher == NULL,
779 ("ktls backend leaked a cipher pointer"));
782 if (ktls_allow_unload)
786 if (ktls_allow_unload)
787 rm_runlock(&ktls_backends_lock, &prio);
790 switch (tls->params.cipher_algorithm) {
792 counter_u64_add(ktls_sw_cbc, 1);
794 case CRYPTO_AES_NIST_GCM_16:
795 counter_u64_add(ktls_sw_gcm, 1);
802 ktls_enable_tx(struct socket *so, struct tls_enable *en)
804 struct ktls_session *tls;
807 if (!ktls_offload_enable)
810 counter_u64_add(ktls_offload_enable_calls, 1);
813 * This should always be true since only the TCP socket option
814 * invokes this function.
816 if (so->so_proto->pr_protocol != IPPROTO_TCP)
820 * XXX: Don't overwrite existing sessions. We should permit
821 * this to support rekeying in the future.
823 if (so->so_snd.sb_tls_info != NULL)
826 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
829 /* TLS requires ext pgs */
830 if (mb_use_ext_pgs == 0)
833 error = ktls_create_session(so, en, &tls);
837 /* Prefer ifnet TLS over software TLS. */
838 error = ktls_try_ifnet(so, tls, false);
840 error = ktls_try_sw(so, tls);
847 error = sblock(&so->so_snd, SBL_WAIT);
853 SOCKBUF_LOCK(&so->so_snd);
854 so->so_snd.sb_tls_info = tls;
855 if (tls->sw_encrypt == NULL)
856 so->so_snd.sb_flags |= SB_TLS_IFNET;
857 SOCKBUF_UNLOCK(&so->so_snd);
858 sbunlock(&so->so_snd);
860 counter_u64_add(ktls_offload_total, 1);
866 ktls_get_tx_mode(struct socket *so)
868 struct ktls_session *tls;
873 INP_WLOCK_ASSERT(inp);
874 SOCKBUF_LOCK(&so->so_snd);
875 tls = so->so_snd.sb_tls_info;
877 mode = TCP_TLS_MODE_NONE;
878 else if (tls->sw_encrypt != NULL)
879 mode = TCP_TLS_MODE_SW;
881 mode = TCP_TLS_MODE_IFNET;
882 SOCKBUF_UNLOCK(&so->so_snd);
887 * Switch between SW and ifnet TLS sessions as requested.
890 ktls_set_tx_mode(struct socket *so, int mode)
892 struct ktls_session *tls, *tls_new;
896 MPASS(mode == TCP_TLS_MODE_SW || mode == TCP_TLS_MODE_IFNET);
899 INP_WLOCK_ASSERT(inp);
900 SOCKBUF_LOCK(&so->so_snd);
901 tls = so->so_snd.sb_tls_info;
903 SOCKBUF_UNLOCK(&so->so_snd);
907 if ((tls->sw_encrypt != NULL && mode == TCP_TLS_MODE_SW) ||
908 (tls->sw_encrypt == NULL && mode == TCP_TLS_MODE_IFNET)) {
909 SOCKBUF_UNLOCK(&so->so_snd);
913 tls = ktls_hold(tls);
914 SOCKBUF_UNLOCK(&so->so_snd);
917 tls_new = ktls_clone_session(tls);
919 if (mode == TCP_TLS_MODE_IFNET)
920 error = ktls_try_ifnet(so, tls_new, true);
922 error = ktls_try_sw(so, tls_new);
924 counter_u64_add(ktls_switch_failed, 1);
931 error = sblock(&so->so_snd, SBL_WAIT);
933 counter_u64_add(ktls_switch_failed, 1);
941 * If we raced with another session change, keep the existing
944 if (tls != so->so_snd.sb_tls_info) {
945 counter_u64_add(ktls_switch_failed, 1);
946 sbunlock(&so->so_snd);
953 SOCKBUF_LOCK(&so->so_snd);
954 so->so_snd.sb_tls_info = tls_new;
955 if (tls_new->sw_encrypt == NULL)
956 so->so_snd.sb_flags |= SB_TLS_IFNET;
957 SOCKBUF_UNLOCK(&so->so_snd);
958 sbunlock(&so->so_snd);
961 * Drop two references on 'tls'. The first is for the
962 * ktls_hold() above. The second drops the reference from the
965 KASSERT(tls->refcount >= 2, ("too few references on old session"));
969 if (mode == TCP_TLS_MODE_IFNET)
970 counter_u64_add(ktls_switch_to_ifnet, 1);
972 counter_u64_add(ktls_switch_to_sw, 1);
979 * Try to allocate a new TLS send tag. This task is scheduled when
980 * ip_output detects a route change while trying to transmit a packet
981 * holding a TLS record. If a new tag is allocated, replace the tag
982 * in the TLS session. Subsequent packets on the connection will use
983 * the new tag. If a new tag cannot be allocated, drop the
987 ktls_reset_send_tag(void *context, int pending)
989 struct epoch_tracker et;
990 struct ktls_session *tls;
991 struct m_snd_tag *old, *new;
1002 * Free the old tag first before allocating a new one.
1003 * ip[6]_output_send() will treat a NULL send tag the same as
1004 * an ifp mismatch and drop packets until a new tag is
1007 * Write-lock the INP when changing tls->snd_tag since
1008 * ip[6]_output_send() holds a read-lock when reading the
1013 tls->snd_tag = NULL;
1016 m_snd_tag_rele(old);
1018 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1023 mtx_pool_lock(mtxpool_sleep, tls);
1024 tls->reset_pending = false;
1025 mtx_pool_unlock(mtxpool_sleep, tls);
1026 if (!in_pcbrele_wlocked(inp))
1029 counter_u64_add(ktls_ifnet_reset, 1);
1032 * XXX: Should we kick tcp_output explicitly now that
1033 * the send tag is fixed or just rely on timers?
1036 INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1038 if (!in_pcbrele_wlocked(inp)) {
1039 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1040 !(inp->inp_flags & INP_DROPPED)) {
1041 tp = intotcpcb(inp);
1042 tp = tcp_drop(tp, ECONNABORTED);
1045 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1049 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1051 counter_u64_add(ktls_ifnet_reset_failed, 1);
1054 * Leave reset_pending true to avoid future tasks while
1055 * the socket goes away.
1063 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1069 INP_LOCK_ASSERT(inp);
1072 * See if we should schedule a task to update the send tag for
1075 mtx_pool_lock(mtxpool_sleep, tls);
1076 if (!tls->reset_pending) {
1077 (void) ktls_hold(tls);
1080 tls->reset_pending = true;
1081 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1083 mtx_pool_unlock(mtxpool_sleep, tls);
1089 ktls_destroy(struct ktls_session *tls)
1091 struct rm_priotracker prio;
1094 if (tls->be != NULL && ktls_allow_unload) {
1095 rm_rlock(&ktls_backends_lock, &prio);
1096 tls->be->use_count--;
1097 rm_runlock(&ktls_backends_lock, &prio);
1099 uma_zfree(ktls_session_zone, tls);
1103 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1105 struct mbuf_ext_pgs *pgs;
1106 struct tls_record_layer *tlshdr;
1109 for (; m != NULL; m = m->m_next) {
1110 KASSERT((m->m_flags & M_NOMAP) != 0,
1111 ("ktls_seq: mapped mbuf %p", m));
1113 pgs = m->m_ext.ext_pgs;
1114 pgs->seqno = sb->sb_tls_seqno;
1117 * Store the sequence number in the TLS header as the
1118 * explicit part of the IV for GCM.
1120 if (pgs->tls->params.cipher_algorithm ==
1121 CRYPTO_AES_NIST_GCM_16) {
1122 tlshdr = (void *)pgs->hdr;
1123 seqno = htobe64(pgs->seqno);
1124 memcpy(tlshdr + 1, &seqno, sizeof(seqno));
1131 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1132 * mbuf in the chain must be an unmapped mbuf. The payload of the
1133 * mbuf must be populated with the payload of each TLS record.
1135 * The record_type argument specifies the TLS record type used when
1136 * populating the TLS header.
1138 * The enq_count argument on return is set to the number of pages of
1139 * payload data for this entire chain that need to be encrypted via SW
1140 * encryption. The returned value should be passed to ktls_enqueue
1141 * when scheduling encryption of this chain of mbufs.
1144 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1145 uint8_t record_type)
1147 struct tls_record_layer *tlshdr;
1149 struct mbuf_ext_pgs *pgs;
1153 maxlen = tls->params.max_frame_len;
1155 for (m = top; m != NULL; m = m->m_next) {
1157 * All mbufs in the chain should be non-empty TLS
1158 * records whose payload does not exceed the maximum
1161 if (m->m_len > maxlen || m->m_len == 0)
1166 * TLS frames require unmapped mbufs to store session
1169 KASSERT((m->m_flags & M_NOMAP) != 0,
1170 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1172 pgs = m->m_ext.ext_pgs;
1174 /* Save a reference to the session. */
1175 pgs->tls = ktls_hold(tls);
1177 pgs->hdr_len = tls->params.tls_hlen;
1178 pgs->trail_len = tls->params.tls_tlen;
1179 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1183 * AES-CBC pads messages to a multiple of the
1184 * block size. Note that the padding is
1185 * applied after the digest and the encryption
1186 * is done on the "plaintext || mac || padding".
1187 * At least one byte of padding is always
1190 * Compute the final trailer length assuming
1191 * at most one block of padding.
1192 * tls->params.sb_tls_tlen is the maximum
1193 * possible trailer length (padding + digest).
1194 * delta holds the number of excess padding
1195 * bytes if the maximum were used. Those
1196 * extra bytes are removed.
1198 bs = tls->params.tls_bs;
1199 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1200 pgs->trail_len -= delta;
1202 m->m_len += pgs->hdr_len + pgs->trail_len;
1204 /* Populate the TLS header. */
1205 tlshdr = (void *)pgs->hdr;
1206 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1209 * TLS 1.3 masquarades as TLS 1.2 with a record type
1210 * of TLS_RLTYPE_APP.
1212 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1213 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1214 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1215 tlshdr->tls_type = TLS_RLTYPE_APP;
1216 /* save the real record type for later */
1217 pgs->record_type = record_type;
1219 tlshdr->tls_vminor = tls->params.tls_vminor;
1220 tlshdr->tls_type = record_type;
1222 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1225 * For GCM, the sequence number is stored in the
1226 * header by ktls_seq(). For CBC, a random nonce is
1227 * inserted for TLS 1.1+.
1229 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1230 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1231 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1234 * When using SW encryption, mark the mbuf not ready.
1235 * It will be marked ready via sbready() after the
1236 * record has been encrypted.
1238 * When using ifnet TLS, unencrypted TLS records are
1239 * sent down the stack to the NIC.
1241 if (tls->sw_encrypt != NULL) {
1242 m->m_flags |= M_NOTREADY;
1243 pgs->nrdy = pgs->npgs;
1244 *enq_cnt += pgs->npgs;
1251 ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs)
1256 /* Mark it for freeing. */
1258 wq = &ktls_wq[pgs->tls->wq_index];
1260 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1261 running = wq->running;
1262 mtx_unlock(&wq->mtx);
1268 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1270 struct mbuf_ext_pgs *pgs;
1274 KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1275 (M_NOMAP | M_NOTREADY)),
1276 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1277 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1279 pgs = m->m_ext.ext_pgs;
1281 KASSERT(pgs->tls->sw_encrypt != NULL, ("ifnet TLS mbuf"));
1283 pgs->enc_cnt = page_count;
1287 * Save a pointer to the socket. The caller is responsible
1288 * for taking an additional reference via soref().
1292 wq = &ktls_wq[pgs->tls->wq_index];
1294 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1295 running = wq->running;
1296 mtx_unlock(&wq->mtx);
1299 counter_u64_add(ktls_cnt_on, 1);
1302 static __noinline void
1303 ktls_encrypt(struct mbuf_ext_pgs *pgs)
1305 struct ktls_session *tls;
1307 struct mbuf *m, *top;
1308 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1309 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1310 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1312 int error, i, len, npages, off, total_pages;
1318 KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs));
1319 KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs));
1324 total_pages = pgs->enc_cnt;
1328 * Encrypt the TLS records in the chain of mbufs starting with
1329 * 'top'. 'total_pages' gives us a total count of pages and is
1330 * used to know when we have finished encrypting the TLS
1331 * records originally queued with 'top'.
1333 * NB: These mbufs are queued in the socket buffer and
1334 * 'm_next' is traversing the mbufs in the socket buffer. The
1335 * socket buffer lock is not held while traversing this chain.
1336 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1337 * pointers should be stable. However, the 'm_next' of the
1338 * last mbuf encrypted is not necessarily NULL. It can point
1339 * to other mbufs appended while 'top' was on the TLS work
1342 * Each mbuf holds an entire TLS record.
1345 for (m = top; npages != total_pages; m = m->m_next) {
1346 pgs = m->m_ext.ext_pgs;
1348 KASSERT(pgs->tls == tls,
1349 ("different TLS sessions in a single mbuf chain: %p vs %p",
1351 KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1352 (M_NOMAP | M_NOTREADY),
1353 ("%p not unready & nomap mbuf (top = %p)\n", m, top));
1354 KASSERT(npages + pgs->npgs <= total_pages,
1355 ("page count mismatch: top %p, total_pages %d, m %p", top,
1359 * Generate source and destination ivoecs to pass to
1360 * the SW encryption backend. For writable mbufs, the
1361 * destination iovec is a copy of the source and
1362 * encryption is done in place. For file-backed mbufs
1363 * (from sendfile), anonymous wired pages are
1364 * allocated and assigned to the destination iovec.
1366 is_anon = (pgs->flags & MBUF_PEXT_FLAG_ANON) != 0;
1368 off = pgs->first_pg_off;
1369 for (i = 0; i < pgs->npgs; i++, off = 0) {
1370 len = mbuf_ext_pg_len(pgs, i, off);
1371 src_iov[i].iov_len = len;
1372 src_iov[i].iov_base =
1373 (char *)(void *)PHYS_TO_DMAP(pgs->pa[i]) + off;
1376 dst_iov[i].iov_base = src_iov[i].iov_base;
1377 dst_iov[i].iov_len = src_iov[i].iov_len;
1381 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
1382 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
1387 parray[i] = VM_PAGE_TO_PHYS(pg);
1388 dst_iov[i].iov_base =
1389 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
1390 dst_iov[i].iov_len = len;
1395 error = (*tls->sw_encrypt)(tls,
1396 (const struct tls_record_layer *)pgs->hdr,
1397 pgs->trail, src_iov, dst_iov, i, pgs->seqno,
1400 counter_u64_add(ktls_offload_failed_crypto, 1);
1405 * For file-backed mbufs, release the file-backed
1406 * pages and replace them in the ext_pgs array with
1407 * the anonymous wired pages allocated above.
1410 /* Free the old pages. */
1411 m->m_ext.ext_free(m);
1413 /* Replace them with the new pages. */
1414 for (i = 0; i < pgs->npgs; i++)
1415 pgs->pa[i] = parray[i];
1417 /* Use the basic free routine. */
1418 m->m_ext.ext_free = mb_free_mext_pgs;
1420 /* Pages are now writable. */
1421 pgs->flags |= MBUF_PEXT_FLAG_ANON;
1425 * Drop a reference to the session now that it is no
1426 * longer needed. Existing code depends on encrypted
1427 * records having no associated session vs
1428 * yet-to-be-encrypted records having an associated
1435 CURVNET_SET(so->so_vnet);
1437 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
1439 so->so_proto->pr_usrreqs->pru_abort(so);
1441 mb_free_notready(top, total_pages);
1450 ktls_work_thread(void *ctx)
1452 struct ktls_wq *wq = ctx;
1453 struct mbuf_ext_pgs *p, *n;
1454 struct ktls_session *tls;
1455 STAILQ_HEAD(, mbuf_ext_pgs) local_head;
1457 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
1462 while (STAILQ_EMPTY(&wq->head)) {
1463 wq->running = false;
1464 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
1468 STAILQ_INIT(&local_head);
1469 STAILQ_CONCAT(&local_head, &wq->head);
1470 mtx_unlock(&wq->mtx);
1472 STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) {
1473 if (p->mbuf != NULL) {
1475 counter_u64_add(ktls_cnt_on, -1);
1479 uma_zfree(zone_extpgs, p);