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>
30 #include "opt_inet6.h"
33 #include <sys/param.h>
34 #include <sys/kernel.h>
35 #include <sys/domainset.h>
39 #include <sys/mutex.h>
40 #include <sys/rmlock.h>
42 #include <sys/protosw.h>
43 #include <sys/refcount.h>
45 #include <sys/socket.h>
46 #include <sys/socketvar.h>
47 #include <sys/sysctl.h>
48 #include <sys/taskqueue.h>
49 #include <sys/kthread.h>
51 #include <sys/vmmeter.h>
52 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
53 #include <machine/pcb.h>
55 #include <machine/vmparam.h>
57 #include <net/if_var.h>
59 #include <net/netisr.h>
60 #include <net/rss_config.h>
62 #include <net/route.h>
63 #include <net/route/nhop.h>
64 #if defined(INET) || defined(INET6)
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_var.h>
70 #include <netinet/tcp_offload.h>
72 #include <opencrypto/xform.h>
73 #include <vm/uma_dbg.h>
75 #include <vm/vm_pageout.h>
76 #include <vm/vm_page.h>
80 STAILQ_HEAD(, mbuf) m_head;
81 STAILQ_HEAD(, socket) so_head;
83 } __aligned(CACHE_LINE_SIZE);
85 struct ktls_domain_info {
90 struct ktls_domain_info ktls_domains[MAXMEMDOM];
91 static struct ktls_wq *ktls_wq;
92 static struct proc *ktls_proc;
93 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
94 static struct rmlock ktls_backends_lock;
95 static uma_zone_t ktls_session_zone;
96 static uint16_t ktls_cpuid_lookup[MAXCPU];
98 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
99 "Kernel TLS offload");
100 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101 "Kernel TLS offload stats");
103 static int ktls_allow_unload;
104 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
105 &ktls_allow_unload, 0, "Allow software crypto modules to unload");
108 static int ktls_bind_threads = 1;
110 static int ktls_bind_threads;
112 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
113 &ktls_bind_threads, 0,
114 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
116 static u_int ktls_maxlen = 16384;
117 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
118 &ktls_maxlen, 0, "Maximum TLS record size");
120 static int ktls_number_threads;
121 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
122 &ktls_number_threads, 0,
123 "Number of TLS threads in thread-pool");
125 static bool ktls_offload_enable;
126 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
127 &ktls_offload_enable, 0,
128 "Enable support for kernel TLS offload");
130 static bool ktls_cbc_enable = true;
131 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
133 "Enable Support of AES-CBC crypto for kernel TLS");
135 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
136 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
137 &ktls_tasks_active, "Number of active tasks");
139 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
140 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
141 &ktls_cnt_tx_pending,
142 "Number of TLS 1.0 records waiting for earlier TLS records");
144 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
145 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
147 "Number of TLS records in queue to tasks for SW encryption");
149 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
150 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
152 "Number of TLS sockets in queue to tasks for SW decryption");
154 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
155 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
156 CTLFLAG_RD, &ktls_offload_total,
157 "Total successful TLS setups (parameters set)");
159 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
160 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
161 CTLFLAG_RD, &ktls_offload_enable_calls,
162 "Total number of TLS enable calls made");
164 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
165 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
166 &ktls_offload_active, "Total Active TLS sessions");
168 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
170 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
172 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
173 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
174 &ktls_offload_failed_crypto, "Total TLS crypto failures");
176 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
177 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
178 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
180 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
182 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
184 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
185 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
186 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
188 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
189 "Software TLS session stats");
190 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
191 "Hardware (ifnet) TLS session stats");
193 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
194 "TOE TLS session stats");
197 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
199 "Active number of software TLS sessions using AES-CBC");
201 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
203 "Active number of software TLS sessions using AES-GCM");
205 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
208 "Active number of software TLS sessions using Chacha20-Poly1305");
210 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
211 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
213 "Active number of ifnet TLS sessions using AES-CBC");
215 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
218 "Active number of ifnet TLS sessions using AES-GCM");
220 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
222 &ktls_ifnet_chacha20,
223 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
225 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
226 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
227 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
229 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
230 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
231 &ktls_ifnet_reset_dropped,
232 "TLS sessions dropped after failing to update ifnet send tag");
234 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
236 &ktls_ifnet_reset_failed,
237 "TLS sessions that failed to allocate a new ifnet send tag");
239 static int ktls_ifnet_permitted;
240 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
241 &ktls_ifnet_permitted, 1,
242 "Whether to permit hardware (ifnet) TLS sessions");
245 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
246 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
248 "Active number of TOE TLS sessions using AES-CBC");
250 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
251 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
253 "Active number of TOE TLS sessions using AES-GCM");
255 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
256 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
258 "Active number of TOE TLS sessions using Chacha20-Poly1305");
261 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
263 static void ktls_cleanup(struct ktls_session *tls);
264 #if defined(INET) || defined(INET6)
265 static void ktls_reset_send_tag(void *context, int pending);
267 static void ktls_work_thread(void *ctx);
270 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
272 struct ktls_crypto_backend *curr_be, *tmp;
274 if (be->api_version != KTLS_API_VERSION) {
275 printf("KTLS: API version mismatch (%d vs %d) for %s\n",
276 be->api_version, KTLS_API_VERSION,
281 rm_wlock(&ktls_backends_lock);
282 printf("KTLS: Registering crypto method %s with prio %d\n",
284 if (LIST_EMPTY(&ktls_backends)) {
285 LIST_INSERT_HEAD(&ktls_backends, be, next);
287 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
288 if (curr_be->prio < be->prio) {
289 LIST_INSERT_BEFORE(curr_be, be, next);
292 if (LIST_NEXT(curr_be, next) == NULL) {
293 LIST_INSERT_AFTER(curr_be, be, next);
298 rm_wunlock(&ktls_backends_lock);
303 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
305 struct ktls_crypto_backend *tmp;
308 * Don't error if the backend isn't registered. This permits
309 * MOD_UNLOAD handlers to use this function unconditionally.
311 rm_wlock(&ktls_backends_lock);
312 LIST_FOREACH(tmp, &ktls_backends, next) {
317 rm_wunlock(&ktls_backends_lock);
321 if (!ktls_allow_unload) {
322 rm_wunlock(&ktls_backends_lock);
324 "KTLS: Deregistering crypto method %s is not supported\n",
330 rm_wunlock(&ktls_backends_lock);
334 LIST_REMOVE(be, next);
335 rm_wunlock(&ktls_backends_lock);
339 #if defined(INET) || defined(INET6)
341 ktls_get_cpu(struct socket *so)
345 struct ktls_domain_info *di;
351 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
352 if (cpuid != NETISR_CPUID_NONE)
356 * Just use the flowid to shard connections in a repeatable
357 * fashion. Note that some crypto backends rely on the
358 * serialization provided by having the same connection use
362 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
363 di = &ktls_domains[inp->inp_numa_domain];
364 cpuid = di->cpu[inp->inp_flowid % di->count];
367 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
373 ktls_init(void *dummy __unused)
378 int count, domain, error, i;
380 rm_init(&ktls_backends_lock, "ktls backends");
381 LIST_INIT(&ktls_backends);
383 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
386 ktls_session_zone = uma_zcreate("ktls_session",
387 sizeof(struct ktls_session),
388 NULL, NULL, NULL, NULL,
392 * Initialize the workqueues to run the TLS work. We create a
393 * work queue for each CPU.
396 STAILQ_INIT(&ktls_wq[i].m_head);
397 STAILQ_INIT(&ktls_wq[i].so_head);
398 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
399 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
400 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
402 panic("Can't add KTLS thread %d error %d", i, error);
405 * Bind threads to cores. If ktls_bind_threads is >
406 * 1, then we bind to the NUMA domain.
408 if (ktls_bind_threads) {
409 if (ktls_bind_threads > 1) {
411 domain = pc->pc_domain;
412 CPU_COPY(&cpuset_domain[domain], &mask);
413 count = ktls_domains[domain].count;
414 ktls_domains[domain].cpu[count] = i;
415 ktls_domains[domain].count++;
419 error = cpuset_setthread(td->td_tid, &mask);
422 "Unable to bind KTLS thread for CPU %d error %d",
425 ktls_cpuid_lookup[ktls_number_threads] = i;
426 ktls_number_threads++;
430 * If we somehow have an empty domain, fall back to choosing
431 * among all KTLS threads.
433 if (ktls_bind_threads > 1) {
434 for (i = 0; i < vm_ndomains; i++) {
435 if (ktls_domains[i].count == 0) {
436 ktls_bind_threads = 1;
443 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
445 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
447 #if defined(INET) || defined(INET6)
449 ktls_create_session(struct socket *so, struct tls_enable *en,
450 struct ktls_session **tlsp)
452 struct ktls_session *tls;
455 /* Only TLS 1.0 - 1.3 are supported. */
456 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
458 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
459 en->tls_vminor > TLS_MINOR_VER_THREE)
462 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
464 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
466 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
469 /* All supported algorithms require a cipher key. */
470 if (en->cipher_key_len == 0)
473 /* No flags are currently supported. */
477 /* Common checks for supported algorithms. */
478 switch (en->cipher_algorithm) {
479 case CRYPTO_AES_NIST_GCM_16:
481 * auth_algorithm isn't used, but permit GMAC values
484 switch (en->auth_algorithm) {
486 #ifdef COMPAT_FREEBSD12
487 /* XXX: Really 13.0-current COMPAT. */
488 case CRYPTO_AES_128_NIST_GMAC:
489 case CRYPTO_AES_192_NIST_GMAC:
490 case CRYPTO_AES_256_NIST_GMAC:
496 if (en->auth_key_len != 0)
498 switch (en->tls_vminor) {
499 case TLS_MINOR_VER_TWO:
500 if (en->iv_len != TLS_AEAD_GCM_LEN)
503 case TLS_MINOR_VER_THREE:
504 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
512 switch (en->auth_algorithm) {
513 case CRYPTO_SHA1_HMAC:
515 case CRYPTO_SHA2_256_HMAC:
516 case CRYPTO_SHA2_384_HMAC:
517 if (en->tls_vminor != TLS_MINOR_VER_TWO)
523 if (en->auth_key_len == 0)
527 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
530 switch (en->tls_vminor) {
531 case TLS_MINOR_VER_ZERO:
532 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
535 case TLS_MINOR_VER_ONE:
536 case TLS_MINOR_VER_TWO:
537 /* Ignore any supplied IV. */
544 case CRYPTO_CHACHA20_POLY1305:
545 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
547 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
548 en->tls_vminor != TLS_MINOR_VER_THREE)
550 if (en->iv_len != TLS_CHACHA20_IV_LEN)
557 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
559 counter_u64_add(ktls_offload_active, 1);
561 refcount_init(&tls->refcount, 1);
562 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
564 tls->wq_index = ktls_get_cpu(so);
566 tls->params.cipher_algorithm = en->cipher_algorithm;
567 tls->params.auth_algorithm = en->auth_algorithm;
568 tls->params.tls_vmajor = en->tls_vmajor;
569 tls->params.tls_vminor = en->tls_vminor;
570 tls->params.flags = en->flags;
571 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
573 /* Set the header and trailer lengths. */
574 tls->params.tls_hlen = sizeof(struct tls_record_layer);
575 switch (en->cipher_algorithm) {
576 case CRYPTO_AES_NIST_GCM_16:
578 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
579 * nonce. TLS 1.3 uses a 12 byte implicit IV.
581 if (en->tls_vminor < TLS_MINOR_VER_THREE)
582 tls->params.tls_hlen += sizeof(uint64_t);
583 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
584 tls->params.tls_bs = 1;
587 switch (en->auth_algorithm) {
588 case CRYPTO_SHA1_HMAC:
589 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
590 /* Implicit IV, no nonce. */
591 tls->sequential_records = true;
592 tls->next_seqno = be64dec(en->rec_seq);
593 STAILQ_INIT(&tls->pending_records);
595 tls->params.tls_hlen += AES_BLOCK_LEN;
597 tls->params.tls_tlen = AES_BLOCK_LEN +
600 case CRYPTO_SHA2_256_HMAC:
601 tls->params.tls_hlen += AES_BLOCK_LEN;
602 tls->params.tls_tlen = AES_BLOCK_LEN +
605 case CRYPTO_SHA2_384_HMAC:
606 tls->params.tls_hlen += AES_BLOCK_LEN;
607 tls->params.tls_tlen = AES_BLOCK_LEN +
611 panic("invalid hmac");
613 tls->params.tls_bs = AES_BLOCK_LEN;
615 case CRYPTO_CHACHA20_POLY1305:
617 * Chacha20 uses a 12 byte implicit IV.
619 tls->params.tls_tlen = POLY1305_HASH_LEN;
620 tls->params.tls_bs = 1;
623 panic("invalid cipher");
627 * TLS 1.3 includes optional padding which we do not support,
628 * and also puts the "real" record type at the end of the
631 if (en->tls_vminor == TLS_MINOR_VER_THREE)
632 tls->params.tls_tlen += sizeof(uint8_t);
634 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
635 ("TLS header length too long: %d", tls->params.tls_hlen));
636 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
637 ("TLS trailer length too long: %d", tls->params.tls_tlen));
639 if (en->auth_key_len != 0) {
640 tls->params.auth_key_len = en->auth_key_len;
641 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
643 error = copyin(en->auth_key, tls->params.auth_key,
649 tls->params.cipher_key_len = en->cipher_key_len;
650 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
651 error = copyin(en->cipher_key, tls->params.cipher_key,
657 * This holds the implicit portion of the nonce for AEAD
658 * ciphers and the initial implicit IV for TLS 1.0. The
659 * explicit portions of the IV are generated in ktls_frame().
661 if (en->iv_len != 0) {
662 tls->params.iv_len = en->iv_len;
663 error = copyin(en->iv, tls->params.iv, en->iv_len);
668 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
669 * counter to generate unique explicit IVs.
671 * Store this counter in the last 8 bytes of the IV
672 * array so that it is 8-byte aligned.
674 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
675 en->tls_vminor == TLS_MINOR_VER_TWO)
676 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
687 static struct ktls_session *
688 ktls_clone_session(struct ktls_session *tls)
690 struct ktls_session *tls_new;
692 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
694 counter_u64_add(ktls_offload_active, 1);
696 refcount_init(&tls_new->refcount, 1);
697 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
699 /* Copy fields from existing session. */
700 tls_new->params = tls->params;
701 tls_new->wq_index = tls->wq_index;
703 /* Deep copy keys. */
704 if (tls_new->params.auth_key != NULL) {
705 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
707 memcpy(tls_new->params.auth_key, tls->params.auth_key,
708 tls->params.auth_key_len);
711 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
713 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
714 tls->params.cipher_key_len);
721 ktls_cleanup(struct ktls_session *tls)
724 counter_u64_add(ktls_offload_active, -1);
726 case TCP_TLS_MODE_SW:
727 MPASS(tls->be != NULL);
728 switch (tls->params.cipher_algorithm) {
730 counter_u64_add(ktls_sw_cbc, -1);
732 case CRYPTO_AES_NIST_GCM_16:
733 counter_u64_add(ktls_sw_gcm, -1);
735 case CRYPTO_CHACHA20_POLY1305:
736 counter_u64_add(ktls_sw_chacha20, -1);
741 case TCP_TLS_MODE_IFNET:
742 switch (tls->params.cipher_algorithm) {
744 counter_u64_add(ktls_ifnet_cbc, -1);
746 case CRYPTO_AES_NIST_GCM_16:
747 counter_u64_add(ktls_ifnet_gcm, -1);
749 case CRYPTO_CHACHA20_POLY1305:
750 counter_u64_add(ktls_ifnet_chacha20, -1);
753 if (tls->snd_tag != NULL)
754 m_snd_tag_rele(tls->snd_tag);
757 case TCP_TLS_MODE_TOE:
758 switch (tls->params.cipher_algorithm) {
760 counter_u64_add(ktls_toe_cbc, -1);
762 case CRYPTO_AES_NIST_GCM_16:
763 counter_u64_add(ktls_toe_gcm, -1);
765 case CRYPTO_CHACHA20_POLY1305:
766 counter_u64_add(ktls_toe_chacha20, -1);
772 if (tls->params.auth_key != NULL) {
773 zfree(tls->params.auth_key, M_KTLS);
774 tls->params.auth_key = NULL;
775 tls->params.auth_key_len = 0;
777 if (tls->params.cipher_key != NULL) {
778 zfree(tls->params.cipher_key, M_KTLS);
779 tls->params.cipher_key = NULL;
780 tls->params.cipher_key_len = 0;
782 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
785 #if defined(INET) || defined(INET6)
789 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
797 if (inp->inp_flags2 & INP_FREED) {
801 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
805 if (inp->inp_socket == NULL) {
810 if (!(tp->t_flags & TF_TOE)) {
815 error = tcp_offload_alloc_tls_session(tp, tls, direction);
818 tls->mode = TCP_TLS_MODE_TOE;
819 switch (tls->params.cipher_algorithm) {
821 counter_u64_add(ktls_toe_cbc, 1);
823 case CRYPTO_AES_NIST_GCM_16:
824 counter_u64_add(ktls_toe_gcm, 1);
826 case CRYPTO_CHACHA20_POLY1305:
827 counter_u64_add(ktls_toe_chacha20, 1);
836 * Common code used when first enabling ifnet TLS on a connection or
837 * when allocating a new ifnet TLS session due to a routing change.
838 * This function allocates a new TLS send tag on whatever interface
839 * the connection is currently routed over.
842 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
843 struct m_snd_tag **mstp)
845 union if_snd_tag_alloc_params params;
847 struct nhop_object *nh;
852 if (inp->inp_flags2 & INP_FREED) {
856 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
860 if (inp->inp_socket == NULL) {
867 * Check administrative controls on ifnet TLS to determine if
868 * ifnet TLS should be denied.
870 * - Always permit 'force' requests.
871 * - ktls_ifnet_permitted == 0: always deny.
873 if (!force && ktls_ifnet_permitted == 0) {
879 * XXX: Use the cached route in the inpcb to find the
880 * interface. This should perhaps instead use
881 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
882 * enabled after a connection has completed key negotiation in
883 * userland, the cached route will be present in practice.
885 nh = inp->inp_route.ro_nh;
894 * Allocate a TLS + ratelimit tag if the connection has an
895 * existing pacing rate.
897 if (tp->t_pacing_rate != -1 &&
898 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
899 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
900 params.tls_rate_limit.inp = inp;
901 params.tls_rate_limit.tls = tls;
902 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
904 params.hdr.type = IF_SND_TAG_TYPE_TLS;
905 params.tls.inp = inp;
906 params.tls.tls = tls;
908 params.hdr.flowid = inp->inp_flowid;
909 params.hdr.flowtype = inp->inp_flowtype;
910 params.hdr.numa_domain = inp->inp_numa_domain;
913 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
917 if (inp->inp_vflag & INP_IPV6) {
918 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
923 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
928 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
935 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
937 struct m_snd_tag *mst;
940 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
942 tls->mode = TCP_TLS_MODE_IFNET;
944 switch (tls->params.cipher_algorithm) {
946 counter_u64_add(ktls_ifnet_cbc, 1);
948 case CRYPTO_AES_NIST_GCM_16:
949 counter_u64_add(ktls_ifnet_gcm, 1);
951 case CRYPTO_CHACHA20_POLY1305:
952 counter_u64_add(ktls_ifnet_chacha20, 1);
960 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
962 struct rm_priotracker prio;
963 struct ktls_crypto_backend *be;
966 * Choose the best software crypto backend. Backends are
967 * stored in sorted priority order (larget value == most
968 * important at the head of the list), so this just stops on
969 * the first backend that claims the session by returning
972 if (ktls_allow_unload)
973 rm_rlock(&ktls_backends_lock, &prio);
974 LIST_FOREACH(be, &ktls_backends, next) {
975 if (be->try(so, tls, direction) == 0)
977 KASSERT(tls->cipher == NULL,
978 ("ktls backend leaked a cipher pointer"));
981 if (ktls_allow_unload)
985 if (ktls_allow_unload)
986 rm_runlock(&ktls_backends_lock, &prio);
989 tls->mode = TCP_TLS_MODE_SW;
990 switch (tls->params.cipher_algorithm) {
992 counter_u64_add(ktls_sw_cbc, 1);
994 case CRYPTO_AES_NIST_GCM_16:
995 counter_u64_add(ktls_sw_gcm, 1);
997 case CRYPTO_CHACHA20_POLY1305:
998 counter_u64_add(ktls_sw_chacha20, 1);
1005 * KTLS RX stores data in the socket buffer as a list of TLS records,
1006 * where each record is stored as a control message containg the TLS
1007 * header followed by data mbufs containing the decrypted data. This
1008 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1009 * both encrypted and decrypted data. TLS records decrypted by a NIC
1010 * should be queued to the socket buffer as records, but encrypted
1011 * data which needs to be decrypted by software arrives as a stream of
1012 * regular mbufs which need to be converted. In addition, there may
1013 * already be pending encrypted data in the socket buffer when KTLS RX
1016 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1019 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1021 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1022 * from the first mbuf. Once all of the data for that TLS record is
1023 * queued, the socket is queued to a worker thread.
1025 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1026 * the TLS chain. Each TLS record is detached from the TLS chain,
1027 * decrypted, and inserted into the regular socket buffer chain as
1028 * record starting with a control message holding the TLS header and
1029 * a chain of mbufs holding the encrypted data.
1033 sb_mark_notready(struct sockbuf *sb)
1040 sb->sb_mbtail = NULL;
1041 sb->sb_lastrecord = NULL;
1042 for (; m != NULL; m = m->m_next) {
1043 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1045 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1047 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1049 m->m_flags |= M_NOTREADY;
1050 sb->sb_acc -= m->m_len;
1051 sb->sb_tlscc += m->m_len;
1052 sb->sb_mtlstail = m;
1054 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1055 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1060 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1062 struct ktls_session *tls;
1065 if (!ktls_offload_enable)
1067 if (SOLISTENING(so))
1070 counter_u64_add(ktls_offload_enable_calls, 1);
1073 * This should always be true since only the TCP socket option
1074 * invokes this function.
1076 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1080 * XXX: Don't overwrite existing sessions. We should permit
1081 * this to support rekeying in the future.
1083 if (so->so_rcv.sb_tls_info != NULL)
1086 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1089 error = ktls_create_session(so, en, &tls);
1094 error = ktls_try_toe(so, tls, KTLS_RX);
1097 error = ktls_try_sw(so, tls, KTLS_RX);
1104 /* Mark the socket as using TLS offload. */
1105 SOCKBUF_LOCK(&so->so_rcv);
1106 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1107 so->so_rcv.sb_tls_info = tls;
1108 so->so_rcv.sb_flags |= SB_TLS_RX;
1110 /* Mark existing data as not ready until it can be decrypted. */
1111 if (tls->mode != TCP_TLS_MODE_TOE) {
1112 sb_mark_notready(&so->so_rcv);
1113 ktls_check_rx(&so->so_rcv);
1115 SOCKBUF_UNLOCK(&so->so_rcv);
1117 counter_u64_add(ktls_offload_total, 1);
1123 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1125 struct ktls_session *tls;
1129 if (!ktls_offload_enable)
1131 if (SOLISTENING(so))
1134 counter_u64_add(ktls_offload_enable_calls, 1);
1137 * This should always be true since only the TCP socket option
1138 * invokes this function.
1140 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1144 * XXX: Don't overwrite existing sessions. We should permit
1145 * this to support rekeying in the future.
1147 if (so->so_snd.sb_tls_info != NULL)
1150 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1153 /* TLS requires ext pgs */
1154 if (mb_use_ext_pgs == 0)
1157 error = ktls_create_session(so, en, &tls);
1161 /* Prefer TOE -> ifnet TLS -> software TLS. */
1163 error = ktls_try_toe(so, tls, KTLS_TX);
1166 error = ktls_try_ifnet(so, tls, false);
1168 error = ktls_try_sw(so, tls, KTLS_TX);
1175 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1182 * Write lock the INP when setting sb_tls_info so that
1183 * routines in tcp_ratelimit.c can read sb_tls_info while
1184 * holding the INP lock.
1188 SOCKBUF_LOCK(&so->so_snd);
1189 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1190 so->so_snd.sb_tls_info = tls;
1191 if (tls->mode != TCP_TLS_MODE_SW)
1192 so->so_snd.sb_flags |= SB_TLS_IFNET;
1193 SOCKBUF_UNLOCK(&so->so_snd);
1195 SOCK_IO_SEND_UNLOCK(so);
1197 counter_u64_add(ktls_offload_total, 1);
1203 ktls_get_rx_mode(struct socket *so)
1205 struct ktls_session *tls;
1209 if (SOLISTENING(so))
1212 INP_WLOCK_ASSERT(inp);
1213 SOCKBUF_LOCK(&so->so_rcv);
1214 tls = so->so_rcv.sb_tls_info;
1216 mode = TCP_TLS_MODE_NONE;
1219 SOCKBUF_UNLOCK(&so->so_rcv);
1224 ktls_get_tx_mode(struct socket *so)
1226 struct ktls_session *tls;
1230 if (SOLISTENING(so))
1233 INP_WLOCK_ASSERT(inp);
1234 SOCKBUF_LOCK(&so->so_snd);
1235 tls = so->so_snd.sb_tls_info;
1237 mode = TCP_TLS_MODE_NONE;
1240 SOCKBUF_UNLOCK(&so->so_snd);
1245 * Switch between SW and ifnet TLS sessions as requested.
1248 ktls_set_tx_mode(struct socket *so, int mode)
1250 struct ktls_session *tls, *tls_new;
1254 if (SOLISTENING(so))
1257 case TCP_TLS_MODE_SW:
1258 case TCP_TLS_MODE_IFNET:
1265 INP_WLOCK_ASSERT(inp);
1266 SOCKBUF_LOCK(&so->so_snd);
1267 tls = so->so_snd.sb_tls_info;
1269 SOCKBUF_UNLOCK(&so->so_snd);
1273 if (tls->mode == mode) {
1274 SOCKBUF_UNLOCK(&so->so_snd);
1278 tls = ktls_hold(tls);
1279 SOCKBUF_UNLOCK(&so->so_snd);
1282 tls_new = ktls_clone_session(tls);
1284 if (mode == TCP_TLS_MODE_IFNET)
1285 error = ktls_try_ifnet(so, tls_new, true);
1287 error = ktls_try_sw(so, tls_new, KTLS_TX);
1289 counter_u64_add(ktls_switch_failed, 1);
1296 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1298 counter_u64_add(ktls_switch_failed, 1);
1306 * If we raced with another session change, keep the existing
1309 if (tls != so->so_snd.sb_tls_info) {
1310 counter_u64_add(ktls_switch_failed, 1);
1311 SOCK_IO_SEND_UNLOCK(so);
1319 SOCKBUF_LOCK(&so->so_snd);
1320 so->so_snd.sb_tls_info = tls_new;
1321 if (tls_new->mode != TCP_TLS_MODE_SW)
1322 so->so_snd.sb_flags |= SB_TLS_IFNET;
1323 SOCKBUF_UNLOCK(&so->so_snd);
1324 SOCK_IO_SEND_UNLOCK(so);
1327 * Drop two references on 'tls'. The first is for the
1328 * ktls_hold() above. The second drops the reference from the
1331 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1335 if (mode == TCP_TLS_MODE_IFNET)
1336 counter_u64_add(ktls_switch_to_ifnet, 1);
1338 counter_u64_add(ktls_switch_to_sw, 1);
1344 * Try to allocate a new TLS send tag. This task is scheduled when
1345 * ip_output detects a route change while trying to transmit a packet
1346 * holding a TLS record. If a new tag is allocated, replace the tag
1347 * in the TLS session. Subsequent packets on the connection will use
1348 * the new tag. If a new tag cannot be allocated, drop the
1352 ktls_reset_send_tag(void *context, int pending)
1354 struct epoch_tracker et;
1355 struct ktls_session *tls;
1356 struct m_snd_tag *old, *new;
1361 MPASS(pending == 1);
1367 * Free the old tag first before allocating a new one.
1368 * ip[6]_output_send() will treat a NULL send tag the same as
1369 * an ifp mismatch and drop packets until a new tag is
1372 * Write-lock the INP when changing tls->snd_tag since
1373 * ip[6]_output_send() holds a read-lock when reading the
1378 tls->snd_tag = NULL;
1381 m_snd_tag_rele(old);
1383 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1388 mtx_pool_lock(mtxpool_sleep, tls);
1389 tls->reset_pending = false;
1390 mtx_pool_unlock(mtxpool_sleep, tls);
1391 if (!in_pcbrele_wlocked(inp))
1394 counter_u64_add(ktls_ifnet_reset, 1);
1397 * XXX: Should we kick tcp_output explicitly now that
1398 * the send tag is fixed or just rely on timers?
1401 NET_EPOCH_ENTER(et);
1403 if (!in_pcbrele_wlocked(inp)) {
1404 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1405 !(inp->inp_flags & INP_DROPPED)) {
1406 tp = intotcpcb(inp);
1407 CURVNET_SET(tp->t_vnet);
1408 tp = tcp_drop(tp, ECONNABORTED);
1412 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1418 counter_u64_add(ktls_ifnet_reset_failed, 1);
1421 * Leave reset_pending true to avoid future tasks while
1422 * the socket goes away.
1430 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1436 INP_LOCK_ASSERT(inp);
1439 * See if we should schedule a task to update the send tag for
1442 mtx_pool_lock(mtxpool_sleep, tls);
1443 if (!tls->reset_pending) {
1444 (void) ktls_hold(tls);
1447 tls->reset_pending = true;
1448 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1450 mtx_pool_unlock(mtxpool_sleep, tls);
1456 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1458 union if_snd_tag_modify_params params = {
1459 .rate_limit.max_rate = max_pacing_rate,
1460 .rate_limit.flags = M_NOWAIT,
1462 struct m_snd_tag *mst;
1466 /* Can't get to the inp, but it should be locked. */
1467 /* INP_LOCK_ASSERT(inp); */
1469 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1471 if (tls->snd_tag == NULL) {
1473 * Resetting send tag, ignore this change. The
1474 * pending reset may or may not see this updated rate
1475 * in the tcpcb. If it doesn't, we will just lose
1484 MPASS(mst->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1487 return (ifp->if_snd_tag_modify(mst, ¶ms));
1493 ktls_destroy(struct ktls_session *tls)
1495 struct rm_priotracker prio;
1497 if (tls->sequential_records) {
1501 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1502 page_count = m->m_epg_enc_cnt;
1503 while (page_count > 0) {
1504 KASSERT(page_count >= m->m_epg_nrdy,
1505 ("%s: too few pages", __func__));
1506 page_count -= m->m_epg_nrdy;
1512 if (tls->be != NULL && ktls_allow_unload) {
1513 rm_rlock(&ktls_backends_lock, &prio);
1514 tls->be->use_count--;
1515 rm_runlock(&ktls_backends_lock, &prio);
1517 uma_zfree(ktls_session_zone, tls);
1521 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1524 for (; m != NULL; m = m->m_next) {
1525 KASSERT((m->m_flags & M_EXTPG) != 0,
1526 ("ktls_seq: mapped mbuf %p", m));
1528 m->m_epg_seqno = sb->sb_tls_seqno;
1534 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1535 * mbuf in the chain must be an unmapped mbuf. The payload of the
1536 * mbuf must be populated with the payload of each TLS record.
1538 * The record_type argument specifies the TLS record type used when
1539 * populating the TLS header.
1541 * The enq_count argument on return is set to the number of pages of
1542 * payload data for this entire chain that need to be encrypted via SW
1543 * encryption. The returned value should be passed to ktls_enqueue
1544 * when scheduling encryption of this chain of mbufs. To handle the
1545 * special case of empty fragments for TLS 1.0 sessions, an empty
1546 * fragment counts as one page.
1549 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1550 uint8_t record_type)
1552 struct tls_record_layer *tlshdr;
1558 maxlen = tls->params.max_frame_len;
1560 for (m = top; m != NULL; m = m->m_next) {
1562 * All mbufs in the chain should be TLS records whose
1563 * payload does not exceed the maximum frame length.
1565 * Empty TLS 1.0 records are permitted when using CBC.
1567 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1568 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1569 ("ktls_frame: m %p len %d", m, m->m_len));
1572 * TLS frames require unmapped mbufs to store session
1575 KASSERT((m->m_flags & M_EXTPG) != 0,
1576 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1580 /* Save a reference to the session. */
1581 m->m_epg_tls = ktls_hold(tls);
1583 m->m_epg_hdrlen = tls->params.tls_hlen;
1584 m->m_epg_trllen = tls->params.tls_tlen;
1585 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1589 * AES-CBC pads messages to a multiple of the
1590 * block size. Note that the padding is
1591 * applied after the digest and the encryption
1592 * is done on the "plaintext || mac || padding".
1593 * At least one byte of padding is always
1596 * Compute the final trailer length assuming
1597 * at most one block of padding.
1598 * tls->params.sb_tls_tlen is the maximum
1599 * possible trailer length (padding + digest).
1600 * delta holds the number of excess padding
1601 * bytes if the maximum were used. Those
1602 * extra bytes are removed.
1604 bs = tls->params.tls_bs;
1605 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1606 m->m_epg_trllen -= delta;
1608 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1610 /* Populate the TLS header. */
1611 tlshdr = (void *)m->m_epg_hdr;
1612 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1615 * TLS 1.3 masquarades as TLS 1.2 with a record type
1616 * of TLS_RLTYPE_APP.
1618 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1619 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1620 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1621 tlshdr->tls_type = TLS_RLTYPE_APP;
1622 /* save the real record type for later */
1623 m->m_epg_record_type = record_type;
1624 m->m_epg_trail[0] = record_type;
1626 tlshdr->tls_vminor = tls->params.tls_vminor;
1627 tlshdr->tls_type = record_type;
1629 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1632 * Store nonces / explicit IVs after the end of the
1635 * For GCM with TLS 1.2, an 8 byte nonce is copied
1636 * from the end of the IV. The nonce is then
1637 * incremented for use by the next record.
1639 * For CBC, a random nonce is inserted for TLS 1.1+.
1641 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1642 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1643 noncep = (uint64_t *)(tls->params.iv + 8);
1644 be64enc(tlshdr + 1, *noncep);
1646 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1647 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1648 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1651 * When using SW encryption, mark the mbuf not ready.
1652 * It will be marked ready via sbready() after the
1653 * record has been encrypted.
1655 * When using ifnet TLS, unencrypted TLS records are
1656 * sent down the stack to the NIC.
1658 if (tls->mode == TCP_TLS_MODE_SW) {
1659 m->m_flags |= M_NOTREADY;
1660 if (__predict_false(tls_len == 0)) {
1661 /* TLS 1.0 empty fragment. */
1664 m->m_epg_nrdy = m->m_epg_npgs;
1665 *enq_cnt += m->m_epg_nrdy;
1671 ktls_permit_empty_frames(struct ktls_session *tls)
1673 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1674 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
1678 ktls_check_rx(struct sockbuf *sb)
1680 struct tls_record_layer hdr;
1685 SOCKBUF_LOCK_ASSERT(sb);
1686 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1688 so = __containerof(sb, struct socket, so_rcv);
1690 if (sb->sb_flags & SB_TLS_RX_RUNNING)
1693 /* Is there enough queued for a TLS header? */
1694 if (sb->sb_tlscc < sizeof(hdr)) {
1695 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1696 so->so_error = EMSGSIZE;
1700 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1702 /* Is the entire record queued? */
1703 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1704 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1705 so->so_error = EMSGSIZE;
1709 sb->sb_flags |= SB_TLS_RX_RUNNING;
1712 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1714 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1715 running = wq->running;
1716 mtx_unlock(&wq->mtx);
1719 counter_u64_add(ktls_cnt_rx_queued, 1);
1722 static struct mbuf *
1723 ktls_detach_record(struct sockbuf *sb, int len)
1725 struct mbuf *m, *n, *top;
1728 SOCKBUF_LOCK_ASSERT(sb);
1729 MPASS(len <= sb->sb_tlscc);
1732 * If TLS chain is the exact size of the record,
1733 * just grab the whole record.
1736 if (sb->sb_tlscc == len) {
1738 sb->sb_mtlstail = NULL;
1743 * While it would be nice to use m_split() here, we need
1744 * to know exactly what m_split() allocates to update the
1745 * accounting, so do it inline instead.
1748 for (m = top; remain > m->m_len; m = m->m_next)
1751 /* Easy case: don't have to split 'm'. */
1752 if (remain == m->m_len) {
1753 sb->sb_mtls = m->m_next;
1754 if (sb->sb_mtls == NULL)
1755 sb->sb_mtlstail = NULL;
1761 * Need to allocate an mbuf to hold the remainder of 'm'. Try
1762 * with M_NOWAIT first.
1764 n = m_get(M_NOWAIT, MT_DATA);
1767 * Use M_WAITOK with socket buffer unlocked. If
1768 * 'sb_mtls' changes while the lock is dropped, return
1769 * NULL to force the caller to retry.
1773 n = m_get(M_WAITOK, MT_DATA);
1776 if (sb->sb_mtls != top) {
1781 n->m_flags |= M_NOTREADY;
1783 /* Store remainder in 'n'. */
1784 n->m_len = m->m_len - remain;
1785 if (m->m_flags & M_EXT) {
1786 n->m_data = m->m_data + remain;
1789 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1792 /* Trim 'm' and update accounting. */
1793 m->m_len -= n->m_len;
1794 sb->sb_tlscc -= n->m_len;
1795 sb->sb_ccc -= n->m_len;
1797 /* Account for 'n'. */
1798 sballoc_ktls_rx(sb, n);
1800 /* Insert 'n' into the TLS chain. */
1802 n->m_next = m->m_next;
1803 if (sb->sb_mtlstail == m)
1804 sb->sb_mtlstail = n;
1806 /* Detach the record from the TLS chain. */
1810 MPASS(m_length(top, NULL) == len);
1811 for (m = top; m != NULL; m = m->m_next)
1812 sbfree_ktls_rx(sb, m);
1813 sb->sb_tlsdcc = len;
1820 * Determine the length of the trailing zero padding and find the real
1821 * record type in the byte before the padding.
1823 * Walking the mbuf chain backwards is clumsy, so another option would
1824 * be to scan forwards remembering the last non-zero byte before the
1825 * trailer. However, it would be expensive to scan the entire record.
1826 * Instead, find the last non-zero byte of each mbuf in the chain
1827 * keeping track of the relative offset of that nonzero byte.
1829 * trail_len is the size of the MAC/tag on input and is set to the
1830 * size of the full trailer including padding and the record type on
1834 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
1835 int *trailer_len, uint8_t *record_typep)
1838 u_int digest_start, last_offset, m_len, offset;
1839 uint8_t record_type;
1841 digest_start = tls_len - *trailer_len;
1844 for (; m != NULL && offset < digest_start;
1845 offset += m->m_len, m = m->m_next) {
1846 /* Don't look for padding in the tag. */
1847 m_len = min(digest_start - offset, m->m_len);
1848 cp = mtod(m, char *);
1850 /* Find last non-zero byte in this mbuf. */
1851 while (m_len > 0 && cp[m_len - 1] == 0)
1854 record_type = cp[m_len - 1];
1855 last_offset = offset + m_len;
1858 if (last_offset < tls->params.tls_hlen)
1861 *record_typep = record_type;
1862 *trailer_len = tls_len - last_offset + 1;
1867 ktls_drop(struct socket *so, int error)
1869 struct epoch_tracker et;
1870 struct inpcb *inp = sotoinpcb(so);
1873 NET_EPOCH_ENTER(et);
1875 if (!(inp->inp_flags & INP_DROPPED)) {
1876 tp = intotcpcb(inp);
1877 CURVNET_SET(inp->inp_vnet);
1878 tp = tcp_drop(tp, error);
1883 so->so_error = error;
1884 SOCKBUF_LOCK(&so->so_rcv);
1885 sorwakeup_locked(so);
1892 ktls_decrypt(struct socket *so)
1894 char tls_header[MBUF_PEXT_HDR_LEN];
1895 struct ktls_session *tls;
1897 struct tls_record_layer *hdr;
1898 struct tls_get_record tgr;
1899 struct mbuf *control, *data, *m;
1901 int error, remain, tls_len, trail_len;
1903 uint8_t vminor, record_type;
1905 hdr = (struct tls_record_layer *)tls_header;
1908 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1909 ("%s: socket %p not running", __func__, so));
1911 tls = sb->sb_tls_info;
1914 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
1916 vminor = TLS_MINOR_VER_TWO;
1918 vminor = tls->params.tls_vminor;
1920 /* Is there enough queued for a TLS header? */
1921 if (sb->sb_tlscc < tls->params.tls_hlen)
1924 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1925 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1927 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1928 hdr->tls_vminor != vminor)
1930 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
1932 else if (tls_len < tls->params.tls_hlen || tls_len >
1933 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1934 tls->params.tls_tlen)
1938 if (__predict_false(error != 0)) {
1940 * We have a corrupted record and are likely
1941 * out of sync. The connection isn't
1942 * recoverable at this point, so abort it.
1945 counter_u64_add(ktls_offload_corrupted_records, 1);
1947 ktls_drop(so, error);
1951 /* Is the entire record queued? */
1952 if (sb->sb_tlscc < tls_len)
1956 * Split out the portion of the mbuf chain containing
1959 data = ktls_detach_record(sb, tls_len);
1962 MPASS(sb->sb_tlsdcc == tls_len);
1964 seqno = sb->sb_tls_seqno;
1969 error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1972 error = tls13_find_record_type(tls, data,
1973 tls_len, &trail_len, &record_type);
1975 record_type = hdr->tls_type;
1978 counter_u64_add(ktls_offload_failed_crypto, 1);
1981 if (sb->sb_tlsdcc == 0) {
1983 * sbcut/drop/flush discarded these
1991 * Drop this TLS record's data, but keep
1992 * decrypting subsequent records.
1994 sb->sb_ccc -= tls_len;
1997 CURVNET_SET(so->so_vnet);
1998 so->so_error = EBADMSG;
1999 sorwakeup_locked(so);
2008 /* Allocate the control mbuf. */
2009 memset(&tgr, 0, sizeof(tgr));
2010 tgr.tls_type = record_type;
2011 tgr.tls_vmajor = hdr->tls_vmajor;
2012 tgr.tls_vminor = hdr->tls_vminor;
2013 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2015 control = sbcreatecontrol_how(&tgr, sizeof(tgr),
2016 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2019 if (sb->sb_tlsdcc == 0) {
2020 /* sbcut/drop/flush discarded these mbufs. */
2021 MPASS(sb->sb_tlscc == 0);
2028 * Clear the 'dcc' accounting in preparation for
2029 * adding the decrypted record.
2031 sb->sb_ccc -= tls_len;
2035 /* If there is no payload, drop all of the data. */
2036 if (tgr.tls_length == htobe16(0)) {
2041 remain = tls->params.tls_hlen;
2042 while (remain > 0) {
2043 if (data->m_len > remain) {
2044 data->m_data += remain;
2045 data->m_len -= remain;
2048 remain -= data->m_len;
2049 data = m_free(data);
2052 /* Trim trailer and clear M_NOTREADY. */
2053 remain = be16toh(tgr.tls_length);
2055 for (m = data; remain > m->m_len; m = m->m_next) {
2056 m->m_flags &= ~M_NOTREADY;
2062 m->m_flags &= ~M_NOTREADY;
2064 /* Set EOR on the final mbuf. */
2065 m->m_flags |= M_EOR;
2068 sbappendcontrol_locked(sb, data, control, 0);
2071 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2073 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2074 so->so_error = EMSGSIZE;
2076 sorwakeup_locked(so);
2079 SOCKBUF_UNLOCK_ASSERT(sb);
2081 CURVNET_SET(so->so_vnet);
2088 ktls_enqueue_to_free(struct mbuf *m)
2093 /* Mark it for freeing. */
2094 m->m_epg_flags |= EPG_FLAG_2FREE;
2095 wq = &ktls_wq[m->m_epg_tls->wq_index];
2097 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2098 running = wq->running;
2099 mtx_unlock(&wq->mtx);
2104 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2106 ktls_batched_records(struct mbuf *m)
2108 int page_count, records;
2111 page_count = m->m_epg_enc_cnt;
2112 while (page_count > 0) {
2114 page_count -= m->m_epg_nrdy;
2117 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2122 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2124 struct ktls_session *tls;
2129 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2130 (M_EXTPG | M_NOTREADY)),
2131 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2132 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2134 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2136 m->m_epg_enc_cnt = page_count;
2139 * Save a pointer to the socket. The caller is responsible
2140 * for taking an additional reference via soref().
2146 wq = &ktls_wq[tls->wq_index];
2148 if (__predict_false(tls->sequential_records)) {
2150 * For TLS 1.0, records must be encrypted
2151 * sequentially. For a given connection, all records
2152 * queued to the associated work queue are processed
2153 * sequentially. However, sendfile(2) might complete
2154 * I/O requests spanning multiple TLS records out of
2155 * order. Here we ensure TLS records are enqueued to
2156 * the work queue in FIFO order.
2158 * tls->next_seqno holds the sequence number of the
2159 * next TLS record that should be enqueued to the work
2160 * queue. If this next record is not tls->next_seqno,
2161 * it must be a future record, so insert it, sorted by
2162 * TLS sequence number, into tls->pending_records and
2165 * If this TLS record matches tls->next_seqno, place
2166 * it in the work queue and then check
2167 * tls->pending_records to see if any
2168 * previously-queued records are now ready for
2171 if (m->m_epg_seqno != tls->next_seqno) {
2175 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2176 if (n->m_epg_seqno > m->m_epg_seqno)
2181 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2184 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2187 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2189 mtx_unlock(&wq->mtx);
2190 counter_u64_add(ktls_cnt_tx_pending, 1);
2194 tls->next_seqno += ktls_batched_records(m);
2195 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2197 while (!STAILQ_EMPTY(&tls->pending_records)) {
2200 n = STAILQ_FIRST(&tls->pending_records);
2201 if (n->m_epg_seqno != tls->next_seqno)
2205 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2206 tls->next_seqno += ktls_batched_records(n);
2207 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2209 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2211 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2213 running = wq->running;
2214 mtx_unlock(&wq->mtx);
2217 counter_u64_add(ktls_cnt_tx_queued, queued);
2220 static __noinline void
2221 ktls_encrypt(struct mbuf *top)
2223 struct ktls_session *tls;
2226 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
2227 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
2228 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
2230 int error, i, len, npages, off, total_pages;
2234 tls = top->m_epg_tls;
2235 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2236 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2238 top->m_epg_so = NULL;
2240 total_pages = top->m_epg_enc_cnt;
2244 * Encrypt the TLS records in the chain of mbufs starting with
2245 * 'top'. 'total_pages' gives us a total count of pages and is
2246 * used to know when we have finished encrypting the TLS
2247 * records originally queued with 'top'.
2249 * NB: These mbufs are queued in the socket buffer and
2250 * 'm_next' is traversing the mbufs in the socket buffer. The
2251 * socket buffer lock is not held while traversing this chain.
2252 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2253 * pointers should be stable. However, the 'm_next' of the
2254 * last mbuf encrypted is not necessarily NULL. It can point
2255 * to other mbufs appended while 'top' was on the TLS work
2258 * Each mbuf holds an entire TLS record.
2261 for (m = top; npages != total_pages; m = m->m_next) {
2262 KASSERT(m->m_epg_tls == tls,
2263 ("different TLS sessions in a single mbuf chain: %p vs %p",
2264 tls, m->m_epg_tls));
2265 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2266 (M_EXTPG | M_NOTREADY),
2267 ("%p not unready & nomap mbuf (top = %p)\n", m, top));
2268 KASSERT(npages + m->m_epg_npgs <= total_pages,
2269 ("page count mismatch: top %p, total_pages %d, m %p", top,
2273 * Generate source and destination ivoecs to pass to
2274 * the SW encryption backend. For writable mbufs, the
2275 * destination iovec is a copy of the source and
2276 * encryption is done in place. For file-backed mbufs
2277 * (from sendfile), anonymous wired pages are
2278 * allocated and assigned to the destination iovec.
2280 is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
2282 off = m->m_epg_1st_off;
2283 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2284 len = m_epg_pagelen(m, i, off);
2285 src_iov[i].iov_len = len;
2286 src_iov[i].iov_base =
2287 (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
2291 dst_iov[i].iov_base = src_iov[i].iov_base;
2292 dst_iov[i].iov_len = src_iov[i].iov_len;
2296 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2302 parray[i] = VM_PAGE_TO_PHYS(pg);
2303 dst_iov[i].iov_base =
2304 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
2305 dst_iov[i].iov_len = len;
2308 npages += m->m_epg_nrdy;
2310 error = (*tls->sw_encrypt)(tls,
2311 (const struct tls_record_layer *)m->m_epg_hdr,
2312 m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
2313 m->m_epg_record_type);
2315 counter_u64_add(ktls_offload_failed_crypto, 1);
2320 * For file-backed mbufs, release the file-backed
2321 * pages and replace them in the ext_pgs array with
2322 * the anonymous wired pages allocated above.
2325 /* Free the old pages. */
2326 m->m_ext.ext_free(m);
2328 /* Replace them with the new pages. */
2329 for (i = 0; i < m->m_epg_npgs; i++)
2330 m->m_epg_pa[i] = parray[i];
2332 /* Use the basic free routine. */
2333 m->m_ext.ext_free = mb_free_mext_pgs;
2335 /* Pages are now writable. */
2336 m->m_epg_flags |= EPG_FLAG_ANON;
2340 * Drop a reference to the session now that it is no
2341 * longer needed. Existing code depends on encrypted
2342 * records having no associated session vs
2343 * yet-to-be-encrypted records having an associated
2346 m->m_epg_tls = NULL;
2350 CURVNET_SET(so->so_vnet);
2352 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2355 mb_free_notready(top, total_pages);
2364 ktls_work_thread(void *ctx)
2366 struct ktls_wq *wq = ctx;
2368 struct socket *so, *son;
2369 STAILQ_HEAD(, mbuf) local_m_head;
2370 STAILQ_HEAD(, socket) local_so_head;
2372 if (ktls_bind_threads > 1) {
2373 curthread->td_domain.dr_policy =
2374 DOMAINSET_PREF(PCPU_GET(domain));
2376 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2381 while (STAILQ_EMPTY(&wq->m_head) &&
2382 STAILQ_EMPTY(&wq->so_head)) {
2383 wq->running = false;
2384 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2388 STAILQ_INIT(&local_m_head);
2389 STAILQ_CONCAT(&local_m_head, &wq->m_head);
2390 STAILQ_INIT(&local_so_head);
2391 STAILQ_CONCAT(&local_so_head, &wq->so_head);
2392 mtx_unlock(&wq->mtx);
2394 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2395 if (m->m_epg_flags & EPG_FLAG_2FREE) {
2396 ktls_free(m->m_epg_tls);
2400 counter_u64_add(ktls_cnt_tx_queued, -1);
2404 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2406 counter_u64_add(ktls_cnt_rx_queued, -1);