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>
61 #include <net/rss_config.h>
63 #if defined(INET) || defined(INET6)
64 #include <netinet/in.h>
65 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_var.h>
69 #include <netinet/tcp_offload.h>
71 #include <opencrypto/xform.h>
72 #include <vm/uma_dbg.h>
74 #include <vm/vm_pageout.h>
75 #include <vm/vm_page.h>
79 STAILQ_HEAD(, mbuf_ext_pgs) head;
81 } __aligned(CACHE_LINE_SIZE);
83 static struct ktls_wq *ktls_wq;
84 static struct proc *ktls_proc;
85 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
86 static struct rmlock ktls_backends_lock;
87 static uma_zone_t ktls_session_zone;
88 static uint16_t ktls_cpuid_lookup[MAXCPU];
90 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
91 "Kernel TLS offload");
92 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
93 "Kernel TLS offload stats");
95 static int ktls_allow_unload;
96 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
97 &ktls_allow_unload, 0, "Allow software crypto modules to unload");
100 static int ktls_bind_threads = 1;
102 static int ktls_bind_threads;
104 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
105 &ktls_bind_threads, 0,
106 "Bind crypto threads to cores or domains at boot");
108 static u_int ktls_maxlen = 16384;
109 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
110 &ktls_maxlen, 0, "Maximum TLS record size");
112 static int ktls_number_threads;
113 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
114 &ktls_number_threads, 0,
115 "Number of TLS threads in thread-pool");
117 static bool ktls_offload_enable;
118 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW,
119 &ktls_offload_enable, 0,
120 "Enable support for kernel TLS offload");
122 static bool ktls_cbc_enable = true;
123 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW,
125 "Enable Support of AES-CBC crypto for kernel TLS");
127 static counter_u64_t ktls_tasks_active;
128 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
129 &ktls_tasks_active, "Number of active tasks");
131 static counter_u64_t ktls_cnt_on;
132 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD,
133 &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto");
135 static counter_u64_t ktls_offload_total;
136 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
137 CTLFLAG_RD, &ktls_offload_total,
138 "Total successful TLS setups (parameters set)");
140 static counter_u64_t ktls_offload_enable_calls;
141 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
142 CTLFLAG_RD, &ktls_offload_enable_calls,
143 "Total number of TLS enable calls made");
145 static counter_u64_t ktls_offload_active;
146 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
147 &ktls_offload_active, "Total Active TLS sessions");
149 static counter_u64_t ktls_offload_failed_crypto;
150 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
151 &ktls_offload_failed_crypto, "Total TLS crypto failures");
153 static counter_u64_t ktls_switch_to_ifnet;
154 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
155 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
157 static counter_u64_t ktls_switch_to_sw;
158 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
159 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
161 static counter_u64_t ktls_switch_failed;
162 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
163 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
165 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
166 "Software TLS session stats");
167 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
168 "Hardware (ifnet) TLS session stats");
170 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
171 "TOE TLS session stats");
174 static counter_u64_t ktls_sw_cbc;
175 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
176 "Active number of software TLS sessions using AES-CBC");
178 static counter_u64_t ktls_sw_gcm;
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
180 "Active number of software TLS sessions using AES-GCM");
182 static counter_u64_t ktls_ifnet_cbc;
183 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
185 "Active number of ifnet TLS sessions using AES-CBC");
187 static counter_u64_t ktls_ifnet_gcm;
188 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
190 "Active number of ifnet TLS sessions using AES-GCM");
192 static counter_u64_t ktls_ifnet_reset;
193 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
194 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
196 static counter_u64_t ktls_ifnet_reset_dropped;
197 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
198 &ktls_ifnet_reset_dropped,
199 "TLS sessions dropped after failing to update ifnet send tag");
201 static counter_u64_t ktls_ifnet_reset_failed;
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
203 &ktls_ifnet_reset_failed,
204 "TLS sessions that failed to allocate a new ifnet send tag");
206 static int ktls_ifnet_permitted;
207 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
208 &ktls_ifnet_permitted, 1,
209 "Whether to permit hardware (ifnet) TLS sessions");
212 static counter_u64_t ktls_toe_cbc;
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
215 "Active number of TOE TLS sessions using AES-CBC");
217 static counter_u64_t ktls_toe_gcm;
218 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
220 "Active number of TOE TLS sessions using AES-GCM");
223 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
225 static void ktls_cleanup(struct ktls_session *tls);
226 #if defined(INET) || defined(INET6)
227 static void ktls_reset_send_tag(void *context, int pending);
229 static void ktls_work_thread(void *ctx);
232 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
234 struct ktls_crypto_backend *curr_be, *tmp;
236 if (be->api_version != KTLS_API_VERSION) {
237 printf("KTLS: API version mismatch (%d vs %d) for %s\n",
238 be->api_version, KTLS_API_VERSION,
243 rm_wlock(&ktls_backends_lock);
244 printf("KTLS: Registering crypto method %s with prio %d\n",
246 if (LIST_EMPTY(&ktls_backends)) {
247 LIST_INSERT_HEAD(&ktls_backends, be, next);
249 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
250 if (curr_be->prio < be->prio) {
251 LIST_INSERT_BEFORE(curr_be, be, next);
254 if (LIST_NEXT(curr_be, next) == NULL) {
255 LIST_INSERT_AFTER(curr_be, be, next);
260 rm_wunlock(&ktls_backends_lock);
265 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
267 struct ktls_crypto_backend *tmp;
270 * Don't error if the backend isn't registered. This permits
271 * MOD_UNLOAD handlers to use this function unconditionally.
273 rm_wlock(&ktls_backends_lock);
274 LIST_FOREACH(tmp, &ktls_backends, next) {
279 rm_wunlock(&ktls_backends_lock);
283 if (!ktls_allow_unload) {
284 rm_wunlock(&ktls_backends_lock);
286 "KTLS: Deregistering crypto method %s is not supported\n",
292 rm_wunlock(&ktls_backends_lock);
296 LIST_REMOVE(be, next);
297 rm_wunlock(&ktls_backends_lock);
301 #if defined(INET) || defined(INET6)
303 ktls_get_cpu(struct socket *so)
310 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
311 if (cpuid != NETISR_CPUID_NONE)
315 * Just use the flowid to shard connections in a repeatable
316 * fashion. Note that some crypto backends rely on the
317 * serialization provided by having the same connection use
320 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
326 ktls_init(void *dummy __unused)
333 ktls_tasks_active = counter_u64_alloc(M_WAITOK);
334 ktls_cnt_on = counter_u64_alloc(M_WAITOK);
335 ktls_offload_total = counter_u64_alloc(M_WAITOK);
336 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
337 ktls_offload_active = counter_u64_alloc(M_WAITOK);
338 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
339 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
340 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
341 ktls_switch_failed = counter_u64_alloc(M_WAITOK);
342 ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
343 ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
344 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
345 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
346 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
347 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
348 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
350 ktls_toe_cbc = counter_u64_alloc(M_WAITOK);
351 ktls_toe_gcm = counter_u64_alloc(M_WAITOK);
354 rm_init(&ktls_backends_lock, "ktls backends");
355 LIST_INIT(&ktls_backends);
357 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
360 ktls_session_zone = uma_zcreate("ktls_session",
361 sizeof(struct ktls_session),
362 NULL, NULL, NULL, NULL,
366 * Initialize the workqueues to run the TLS work. We create a
367 * work queue for each CPU.
370 STAILQ_INIT(&ktls_wq[i].head);
371 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
372 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
373 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
375 panic("Can't add KTLS thread %d error %d", i, error);
378 * Bind threads to cores. If ktls_bind_threads is >
379 * 1, then we bind to the NUMA domain.
381 if (ktls_bind_threads) {
382 if (ktls_bind_threads > 1) {
384 CPU_COPY(&cpuset_domain[pc->pc_domain], &mask);
388 error = cpuset_setthread(td->td_tid, &mask);
391 "Unable to bind KTLS thread for CPU %d error %d",
394 ktls_cpuid_lookup[ktls_number_threads] = i;
395 ktls_number_threads++;
397 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
399 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
401 #if defined(INET) || defined(INET6)
403 ktls_create_session(struct socket *so, struct tls_enable *en,
404 struct ktls_session **tlsp)
406 struct ktls_session *tls;
409 /* Only TLS 1.0 - 1.3 are supported. */
410 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
412 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
413 en->tls_vminor > TLS_MINOR_VER_THREE)
416 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
418 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
420 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
423 /* All supported algorithms require a cipher key. */
424 if (en->cipher_key_len == 0)
427 /* No flags are currently supported. */
431 /* Common checks for supported algorithms. */
432 switch (en->cipher_algorithm) {
433 case CRYPTO_AES_NIST_GCM_16:
435 * auth_algorithm isn't used, but permit GMAC values
438 switch (en->auth_algorithm) {
440 case CRYPTO_AES_128_NIST_GMAC:
441 case CRYPTO_AES_192_NIST_GMAC:
442 case CRYPTO_AES_256_NIST_GMAC:
447 if (en->auth_key_len != 0)
449 if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
450 en->iv_len != TLS_AEAD_GCM_LEN) ||
451 (en->tls_vminor == TLS_MINOR_VER_THREE &&
452 en->iv_len != TLS_1_3_GCM_IV_LEN))
456 switch (en->auth_algorithm) {
457 case CRYPTO_SHA1_HMAC:
459 * TLS 1.0 requires an implicit IV. TLS 1.1+
460 * all use explicit IVs.
462 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
463 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
469 case CRYPTO_SHA2_256_HMAC:
470 case CRYPTO_SHA2_384_HMAC:
471 /* Ignore any supplied IV. */
477 if (en->auth_key_len == 0)
484 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
486 counter_u64_add(ktls_offload_active, 1);
488 refcount_init(&tls->refcount, 1);
489 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
491 tls->wq_index = ktls_get_cpu(so);
493 tls->params.cipher_algorithm = en->cipher_algorithm;
494 tls->params.auth_algorithm = en->auth_algorithm;
495 tls->params.tls_vmajor = en->tls_vmajor;
496 tls->params.tls_vminor = en->tls_vminor;
497 tls->params.flags = en->flags;
498 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
500 /* Set the header and trailer lengths. */
501 tls->params.tls_hlen = sizeof(struct tls_record_layer);
502 switch (en->cipher_algorithm) {
503 case CRYPTO_AES_NIST_GCM_16:
505 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
506 * nonce. TLS 1.3 uses a 12 byte implicit IV.
508 if (en->tls_vminor < TLS_MINOR_VER_THREE)
509 tls->params.tls_hlen += sizeof(uint64_t);
510 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
513 * TLS 1.3 includes optional padding which we
514 * do not support, and also puts the "real" record
515 * type at the end of the encrypted data.
517 if (en->tls_vminor == TLS_MINOR_VER_THREE)
518 tls->params.tls_tlen += sizeof(uint8_t);
520 tls->params.tls_bs = 1;
523 switch (en->auth_algorithm) {
524 case CRYPTO_SHA1_HMAC:
525 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
526 /* Implicit IV, no nonce. */
528 tls->params.tls_hlen += AES_BLOCK_LEN;
530 tls->params.tls_tlen = AES_BLOCK_LEN +
533 case CRYPTO_SHA2_256_HMAC:
534 tls->params.tls_hlen += AES_BLOCK_LEN;
535 tls->params.tls_tlen = AES_BLOCK_LEN +
538 case CRYPTO_SHA2_384_HMAC:
539 tls->params.tls_hlen += AES_BLOCK_LEN;
540 tls->params.tls_tlen = AES_BLOCK_LEN +
544 panic("invalid hmac");
546 tls->params.tls_bs = AES_BLOCK_LEN;
549 panic("invalid cipher");
552 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
553 ("TLS header length too long: %d", tls->params.tls_hlen));
554 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
555 ("TLS trailer length too long: %d", tls->params.tls_tlen));
557 if (en->auth_key_len != 0) {
558 tls->params.auth_key_len = en->auth_key_len;
559 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
561 error = copyin(en->auth_key, tls->params.auth_key,
567 tls->params.cipher_key_len = en->cipher_key_len;
568 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
569 error = copyin(en->cipher_key, tls->params.cipher_key,
575 * This holds the implicit portion of the nonce for GCM and
576 * the initial implicit IV for TLS 1.0. The explicit portions
577 * of the IV are generated in ktls_frame().
579 if (en->iv_len != 0) {
580 tls->params.iv_len = en->iv_len;
581 error = copyin(en->iv, tls->params.iv, en->iv_len);
586 * For TLS 1.2, generate an 8-byte nonce as a counter
587 * to generate unique explicit IVs.
589 * Store this counter in the last 8 bytes of the IV
590 * array so that it is 8-byte aligned.
592 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
593 en->tls_vminor == TLS_MINOR_VER_TWO)
594 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
605 static struct ktls_session *
606 ktls_clone_session(struct ktls_session *tls)
608 struct ktls_session *tls_new;
610 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
612 counter_u64_add(ktls_offload_active, 1);
614 refcount_init(&tls_new->refcount, 1);
616 /* Copy fields from existing session. */
617 tls_new->params = tls->params;
618 tls_new->wq_index = tls->wq_index;
620 /* Deep copy keys. */
621 if (tls_new->params.auth_key != NULL) {
622 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
624 memcpy(tls_new->params.auth_key, tls->params.auth_key,
625 tls->params.auth_key_len);
628 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
630 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
631 tls->params.cipher_key_len);
638 ktls_cleanup(struct ktls_session *tls)
641 counter_u64_add(ktls_offload_active, -1);
643 case TCP_TLS_MODE_SW:
644 MPASS(tls->be != NULL);
645 switch (tls->params.cipher_algorithm) {
647 counter_u64_add(ktls_sw_cbc, -1);
649 case CRYPTO_AES_NIST_GCM_16:
650 counter_u64_add(ktls_sw_gcm, -1);
655 case TCP_TLS_MODE_IFNET:
656 switch (tls->params.cipher_algorithm) {
658 counter_u64_add(ktls_ifnet_cbc, -1);
660 case CRYPTO_AES_NIST_GCM_16:
661 counter_u64_add(ktls_ifnet_gcm, -1);
664 m_snd_tag_rele(tls->snd_tag);
667 case TCP_TLS_MODE_TOE:
668 switch (tls->params.cipher_algorithm) {
670 counter_u64_add(ktls_toe_cbc, -1);
672 case CRYPTO_AES_NIST_GCM_16:
673 counter_u64_add(ktls_toe_gcm, -1);
679 if (tls->params.auth_key != NULL) {
680 explicit_bzero(tls->params.auth_key, tls->params.auth_key_len);
681 free(tls->params.auth_key, M_KTLS);
682 tls->params.auth_key = NULL;
683 tls->params.auth_key_len = 0;
685 if (tls->params.cipher_key != NULL) {
686 explicit_bzero(tls->params.cipher_key,
687 tls->params.cipher_key_len);
688 free(tls->params.cipher_key, M_KTLS);
689 tls->params.cipher_key = NULL;
690 tls->params.cipher_key_len = 0;
692 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
695 #if defined(INET) || defined(INET6)
699 ktls_try_toe(struct socket *so, struct ktls_session *tls)
707 if (inp->inp_flags2 & INP_FREED) {
711 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
715 if (inp->inp_socket == NULL) {
720 if (tp->tod == NULL) {
725 error = tcp_offload_alloc_tls_session(tp, tls);
728 tls->mode = TCP_TLS_MODE_TOE;
729 switch (tls->params.cipher_algorithm) {
731 counter_u64_add(ktls_toe_cbc, 1);
733 case CRYPTO_AES_NIST_GCM_16:
734 counter_u64_add(ktls_toe_gcm, 1);
743 * Common code used when first enabling ifnet TLS on a connection or
744 * when allocating a new ifnet TLS session due to a routing change.
745 * This function allocates a new TLS send tag on whatever interface
746 * the connection is currently routed over.
749 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
750 struct m_snd_tag **mstp)
752 union if_snd_tag_alloc_params params;
759 if (inp->inp_flags2 & INP_FREED) {
763 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
767 if (inp->inp_socket == NULL) {
774 * Check administrative controls on ifnet TLS to determine if
775 * ifnet TLS should be denied.
777 * - Always permit 'force' requests.
778 * - ktls_ifnet_permitted == 0: always deny.
780 if (!force && ktls_ifnet_permitted == 0) {
786 * XXX: Use the cached route in the inpcb to find the
787 * interface. This should perhaps instead use
788 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
789 * enabled after a connection has completed key negotiation in
790 * userland, the cached route will be present in practice.
792 rt = inp->inp_route.ro_rt;
793 if (rt == NULL || rt->rt_ifp == NULL) {
800 params.hdr.type = IF_SND_TAG_TYPE_TLS;
801 params.hdr.flowid = inp->inp_flowid;
802 params.hdr.flowtype = inp->inp_flowtype;
803 params.tls.inp = inp;
804 params.tls.tls = tls;
807 if (ifp->if_snd_tag_alloc == NULL) {
811 if ((ifp->if_capenable & IFCAP_NOMAP) == 0) {
815 if (inp->inp_vflag & INP_IPV6) {
816 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
821 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
826 error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp);
833 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
835 struct m_snd_tag *mst;
838 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
840 tls->mode = TCP_TLS_MODE_IFNET;
842 switch (tls->params.cipher_algorithm) {
844 counter_u64_add(ktls_ifnet_cbc, 1);
846 case CRYPTO_AES_NIST_GCM_16:
847 counter_u64_add(ktls_ifnet_gcm, 1);
855 ktls_try_sw(struct socket *so, struct ktls_session *tls)
857 struct rm_priotracker prio;
858 struct ktls_crypto_backend *be;
861 * Choose the best software crypto backend. Backends are
862 * stored in sorted priority order (larget value == most
863 * important at the head of the list), so this just stops on
864 * the first backend that claims the session by returning
867 if (ktls_allow_unload)
868 rm_rlock(&ktls_backends_lock, &prio);
869 LIST_FOREACH(be, &ktls_backends, next) {
870 if (be->try(so, tls) == 0)
872 KASSERT(tls->cipher == NULL,
873 ("ktls backend leaked a cipher pointer"));
876 if (ktls_allow_unload)
880 if (ktls_allow_unload)
881 rm_runlock(&ktls_backends_lock, &prio);
884 tls->mode = TCP_TLS_MODE_SW;
885 switch (tls->params.cipher_algorithm) {
887 counter_u64_add(ktls_sw_cbc, 1);
889 case CRYPTO_AES_NIST_GCM_16:
890 counter_u64_add(ktls_sw_gcm, 1);
897 ktls_enable_tx(struct socket *so, struct tls_enable *en)
899 struct ktls_session *tls;
902 if (!ktls_offload_enable)
905 counter_u64_add(ktls_offload_enable_calls, 1);
908 * This should always be true since only the TCP socket option
909 * invokes this function.
911 if (so->so_proto->pr_protocol != IPPROTO_TCP)
915 * XXX: Don't overwrite existing sessions. We should permit
916 * this to support rekeying in the future.
918 if (so->so_snd.sb_tls_info != NULL)
921 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
924 /* TLS requires ext pgs */
925 if (mb_use_ext_pgs == 0)
928 error = ktls_create_session(so, en, &tls);
932 /* Prefer TOE -> ifnet TLS -> software TLS. */
934 error = ktls_try_toe(so, tls);
937 error = ktls_try_ifnet(so, tls, false);
939 error = ktls_try_sw(so, tls);
946 error = sblock(&so->so_snd, SBL_WAIT);
952 SOCKBUF_LOCK(&so->so_snd);
953 so->so_snd.sb_tls_info = tls;
954 if (tls->mode != TCP_TLS_MODE_SW)
955 so->so_snd.sb_flags |= SB_TLS_IFNET;
956 SOCKBUF_UNLOCK(&so->so_snd);
957 sbunlock(&so->so_snd);
959 counter_u64_add(ktls_offload_total, 1);
965 ktls_get_tx_mode(struct socket *so)
967 struct ktls_session *tls;
972 INP_WLOCK_ASSERT(inp);
973 SOCKBUF_LOCK(&so->so_snd);
974 tls = so->so_snd.sb_tls_info;
976 mode = TCP_TLS_MODE_NONE;
979 SOCKBUF_UNLOCK(&so->so_snd);
984 * Switch between SW and ifnet TLS sessions as requested.
987 ktls_set_tx_mode(struct socket *so, int mode)
989 struct ktls_session *tls, *tls_new;
994 case TCP_TLS_MODE_SW:
995 case TCP_TLS_MODE_IFNET:
1002 INP_WLOCK_ASSERT(inp);
1003 SOCKBUF_LOCK(&so->so_snd);
1004 tls = so->so_snd.sb_tls_info;
1006 SOCKBUF_UNLOCK(&so->so_snd);
1010 if (tls->mode == mode) {
1011 SOCKBUF_UNLOCK(&so->so_snd);
1015 tls = ktls_hold(tls);
1016 SOCKBUF_UNLOCK(&so->so_snd);
1019 tls_new = ktls_clone_session(tls);
1021 if (mode == TCP_TLS_MODE_IFNET)
1022 error = ktls_try_ifnet(so, tls_new, true);
1024 error = ktls_try_sw(so, tls_new);
1026 counter_u64_add(ktls_switch_failed, 1);
1033 error = sblock(&so->so_snd, SBL_WAIT);
1035 counter_u64_add(ktls_switch_failed, 1);
1043 * If we raced with another session change, keep the existing
1046 if (tls != so->so_snd.sb_tls_info) {
1047 counter_u64_add(ktls_switch_failed, 1);
1048 sbunlock(&so->so_snd);
1055 SOCKBUF_LOCK(&so->so_snd);
1056 so->so_snd.sb_tls_info = tls_new;
1057 if (tls_new->mode != TCP_TLS_MODE_SW)
1058 so->so_snd.sb_flags |= SB_TLS_IFNET;
1059 SOCKBUF_UNLOCK(&so->so_snd);
1060 sbunlock(&so->so_snd);
1063 * Drop two references on 'tls'. The first is for the
1064 * ktls_hold() above. The second drops the reference from the
1067 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1071 if (mode == TCP_TLS_MODE_IFNET)
1072 counter_u64_add(ktls_switch_to_ifnet, 1);
1074 counter_u64_add(ktls_switch_to_sw, 1);
1081 * Try to allocate a new TLS send tag. This task is scheduled when
1082 * ip_output detects a route change while trying to transmit a packet
1083 * holding a TLS record. If a new tag is allocated, replace the tag
1084 * in the TLS session. Subsequent packets on the connection will use
1085 * the new tag. If a new tag cannot be allocated, drop the
1089 ktls_reset_send_tag(void *context, int pending)
1091 struct epoch_tracker et;
1092 struct ktls_session *tls;
1093 struct m_snd_tag *old, *new;
1098 MPASS(pending == 1);
1104 * Free the old tag first before allocating a new one.
1105 * ip[6]_output_send() will treat a NULL send tag the same as
1106 * an ifp mismatch and drop packets until a new tag is
1109 * Write-lock the INP when changing tls->snd_tag since
1110 * ip[6]_output_send() holds a read-lock when reading the
1115 tls->snd_tag = NULL;
1118 m_snd_tag_rele(old);
1120 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1125 mtx_pool_lock(mtxpool_sleep, tls);
1126 tls->reset_pending = false;
1127 mtx_pool_unlock(mtxpool_sleep, tls);
1128 if (!in_pcbrele_wlocked(inp))
1131 counter_u64_add(ktls_ifnet_reset, 1);
1134 * XXX: Should we kick tcp_output explicitly now that
1135 * the send tag is fixed or just rely on timers?
1138 NET_EPOCH_ENTER(et);
1140 if (!in_pcbrele_wlocked(inp)) {
1141 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1142 !(inp->inp_flags & INP_DROPPED)) {
1143 tp = intotcpcb(inp);
1144 CURVNET_SET(tp->t_vnet);
1145 tp = tcp_drop(tp, ECONNABORTED);
1149 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1155 counter_u64_add(ktls_ifnet_reset_failed, 1);
1158 * Leave reset_pending true to avoid future tasks while
1159 * the socket goes away.
1167 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1173 INP_LOCK_ASSERT(inp);
1176 * See if we should schedule a task to update the send tag for
1179 mtx_pool_lock(mtxpool_sleep, tls);
1180 if (!tls->reset_pending) {
1181 (void) ktls_hold(tls);
1184 tls->reset_pending = true;
1185 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1187 mtx_pool_unlock(mtxpool_sleep, tls);
1193 ktls_destroy(struct ktls_session *tls)
1195 struct rm_priotracker prio;
1198 if (tls->be != NULL && ktls_allow_unload) {
1199 rm_rlock(&ktls_backends_lock, &prio);
1200 tls->be->use_count--;
1201 rm_runlock(&ktls_backends_lock, &prio);
1203 uma_zfree(ktls_session_zone, tls);
1207 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1209 struct mbuf_ext_pgs *pgs;
1211 for (; m != NULL; m = m->m_next) {
1212 KASSERT((m->m_flags & M_NOMAP) != 0,
1213 ("ktls_seq: mapped mbuf %p", m));
1215 pgs = m->m_ext.ext_pgs;
1216 pgs->seqno = sb->sb_tls_seqno;
1222 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1223 * mbuf in the chain must be an unmapped mbuf. The payload of the
1224 * mbuf must be populated with the payload of each TLS record.
1226 * The record_type argument specifies the TLS record type used when
1227 * populating the TLS header.
1229 * The enq_count argument on return is set to the number of pages of
1230 * payload data for this entire chain that need to be encrypted via SW
1231 * encryption. The returned value should be passed to ktls_enqueue
1232 * when scheduling encryption of this chain of mbufs.
1235 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1236 uint8_t record_type)
1238 struct tls_record_layer *tlshdr;
1240 struct mbuf_ext_pgs *pgs;
1245 maxlen = tls->params.max_frame_len;
1247 for (m = top; m != NULL; m = m->m_next) {
1249 * All mbufs in the chain should be non-empty TLS
1250 * records whose payload does not exceed the maximum
1253 KASSERT(m->m_len <= maxlen && m->m_len > 0,
1254 ("ktls_frame: m %p len %d\n", m, m->m_len));
1256 * TLS frames require unmapped mbufs to store session
1259 KASSERT((m->m_flags & M_NOMAP) != 0,
1260 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1263 pgs = m->m_ext.ext_pgs;
1265 /* Save a reference to the session. */
1266 pgs->tls = ktls_hold(tls);
1268 pgs->hdr_len = tls->params.tls_hlen;
1269 pgs->trail_len = tls->params.tls_tlen;
1270 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1274 * AES-CBC pads messages to a multiple of the
1275 * block size. Note that the padding is
1276 * applied after the digest and the encryption
1277 * is done on the "plaintext || mac || padding".
1278 * At least one byte of padding is always
1281 * Compute the final trailer length assuming
1282 * at most one block of padding.
1283 * tls->params.sb_tls_tlen is the maximum
1284 * possible trailer length (padding + digest).
1285 * delta holds the number of excess padding
1286 * bytes if the maximum were used. Those
1287 * extra bytes are removed.
1289 bs = tls->params.tls_bs;
1290 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1291 pgs->trail_len -= delta;
1293 m->m_len += pgs->hdr_len + pgs->trail_len;
1295 /* Populate the TLS header. */
1296 tlshdr = (void *)pgs->hdr;
1297 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1300 * TLS 1.3 masquarades as TLS 1.2 with a record type
1301 * of TLS_RLTYPE_APP.
1303 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1304 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1305 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1306 tlshdr->tls_type = TLS_RLTYPE_APP;
1307 /* save the real record type for later */
1308 pgs->record_type = record_type;
1310 tlshdr->tls_vminor = tls->params.tls_vminor;
1311 tlshdr->tls_type = record_type;
1313 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1316 * Store nonces / explicit IVs after the end of the
1319 * For GCM with TLS 1.2, an 8 byte nonce is copied
1320 * from the end of the IV. The nonce is then
1321 * incremented for use by the next record.
1323 * For CBC, a random nonce is inserted for TLS 1.1+.
1325 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1326 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1327 noncep = (uint64_t *)(tls->params.iv + 8);
1328 be64enc(tlshdr + 1, *noncep);
1330 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1331 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1332 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1335 * When using SW encryption, mark the mbuf not ready.
1336 * It will be marked ready via sbready() after the
1337 * record has been encrypted.
1339 * When using ifnet TLS, unencrypted TLS records are
1340 * sent down the stack to the NIC.
1342 if (tls->mode == TCP_TLS_MODE_SW) {
1343 m->m_flags |= M_NOTREADY;
1344 pgs->nrdy = pgs->npgs;
1345 *enq_cnt += pgs->npgs;
1351 ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs)
1356 /* Mark it for freeing. */
1358 wq = &ktls_wq[pgs->tls->wq_index];
1360 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1361 running = wq->running;
1362 mtx_unlock(&wq->mtx);
1368 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1370 struct mbuf_ext_pgs *pgs;
1374 KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1375 (M_NOMAP | M_NOTREADY)),
1376 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1377 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1379 pgs = m->m_ext.ext_pgs;
1381 KASSERT(pgs->tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1383 pgs->enc_cnt = page_count;
1387 * Save a pointer to the socket. The caller is responsible
1388 * for taking an additional reference via soref().
1392 wq = &ktls_wq[pgs->tls->wq_index];
1394 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq);
1395 running = wq->running;
1396 mtx_unlock(&wq->mtx);
1399 counter_u64_add(ktls_cnt_on, 1);
1402 static __noinline void
1403 ktls_encrypt(struct mbuf_ext_pgs *pgs)
1405 struct ktls_session *tls;
1407 struct mbuf *m, *top;
1408 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1409 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1410 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1412 int error, i, len, npages, off, total_pages;
1418 KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs));
1419 KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs));
1424 total_pages = pgs->enc_cnt;
1428 * Encrypt the TLS records in the chain of mbufs starting with
1429 * 'top'. 'total_pages' gives us a total count of pages and is
1430 * used to know when we have finished encrypting the TLS
1431 * records originally queued with 'top'.
1433 * NB: These mbufs are queued in the socket buffer and
1434 * 'm_next' is traversing the mbufs in the socket buffer. The
1435 * socket buffer lock is not held while traversing this chain.
1436 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1437 * pointers should be stable. However, the 'm_next' of the
1438 * last mbuf encrypted is not necessarily NULL. It can point
1439 * to other mbufs appended while 'top' was on the TLS work
1442 * Each mbuf holds an entire TLS record.
1445 for (m = top; npages != total_pages; m = m->m_next) {
1446 pgs = m->m_ext.ext_pgs;
1448 KASSERT(pgs->tls == tls,
1449 ("different TLS sessions in a single mbuf chain: %p vs %p",
1451 KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) ==
1452 (M_NOMAP | M_NOTREADY),
1453 ("%p not unready & nomap mbuf (top = %p)\n", m, top));
1454 KASSERT(npages + pgs->npgs <= total_pages,
1455 ("page count mismatch: top %p, total_pages %d, m %p", top,
1459 * Generate source and destination ivoecs to pass to
1460 * the SW encryption backend. For writable mbufs, the
1461 * destination iovec is a copy of the source and
1462 * encryption is done in place. For file-backed mbufs
1463 * (from sendfile), anonymous wired pages are
1464 * allocated and assigned to the destination iovec.
1466 is_anon = (pgs->flags & MBUF_PEXT_FLAG_ANON) != 0;
1468 off = pgs->first_pg_off;
1469 for (i = 0; i < pgs->npgs; i++, off = 0) {
1470 len = mbuf_ext_pg_len(pgs, i, off);
1471 src_iov[i].iov_len = len;
1472 src_iov[i].iov_base =
1473 (char *)(void *)PHYS_TO_DMAP(pgs->pa[i]) + off;
1476 dst_iov[i].iov_base = src_iov[i].iov_base;
1477 dst_iov[i].iov_len = src_iov[i].iov_len;
1481 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
1482 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
1487 parray[i] = VM_PAGE_TO_PHYS(pg);
1488 dst_iov[i].iov_base =
1489 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
1490 dst_iov[i].iov_len = len;
1495 error = (*tls->sw_encrypt)(tls,
1496 (const struct tls_record_layer *)pgs->hdr,
1497 pgs->trail, src_iov, dst_iov, i, pgs->seqno,
1500 counter_u64_add(ktls_offload_failed_crypto, 1);
1505 * For file-backed mbufs, release the file-backed
1506 * pages and replace them in the ext_pgs array with
1507 * the anonymous wired pages allocated above.
1510 /* Free the old pages. */
1511 m->m_ext.ext_free(m);
1513 /* Replace them with the new pages. */
1514 for (i = 0; i < pgs->npgs; i++)
1515 pgs->pa[i] = parray[i];
1517 /* Use the basic free routine. */
1518 m->m_ext.ext_free = mb_free_mext_pgs;
1520 /* Pages are now writable. */
1521 pgs->flags |= MBUF_PEXT_FLAG_ANON;
1525 * Drop a reference to the session now that it is no
1526 * longer needed. Existing code depends on encrypted
1527 * records having no associated session vs
1528 * yet-to-be-encrypted records having an associated
1535 CURVNET_SET(so->so_vnet);
1537 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
1539 so->so_proto->pr_usrreqs->pru_abort(so);
1541 mb_free_notready(top, total_pages);
1550 ktls_work_thread(void *ctx)
1552 struct ktls_wq *wq = ctx;
1553 struct mbuf_ext_pgs *p, *n;
1554 struct ktls_session *tls;
1555 STAILQ_HEAD(, mbuf_ext_pgs) local_head;
1557 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
1562 while (STAILQ_EMPTY(&wq->head)) {
1563 wq->running = false;
1564 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
1568 STAILQ_INIT(&local_head);
1569 STAILQ_CONCAT(&local_head, &wq->head);
1570 mtx_unlock(&wq->mtx);
1572 STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) {
1573 if (p->mbuf != NULL) {
1575 counter_u64_add(ktls_cnt_on, -1);
1579 uma_zfree(zone_extpgs, p);