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>
37 #include <sys/domainset.h>
41 #include <sys/mutex.h>
42 #include <sys/rmlock.h>
44 #include <sys/protosw.h>
45 #include <sys/refcount.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/taskqueue.h>
51 #include <sys/kthread.h>
53 #include <sys/vmmeter.h>
54 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
55 #include <machine/pcb.h>
57 #include <machine/vmparam.h>
59 #include <net/if_var.h>
61 #include <net/netisr.h>
62 #include <net/rss_config.h>
64 #include <net/route.h>
65 #include <net/route/nhop.h>
66 #if defined(INET) || defined(INET6)
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
70 #include <netinet/tcp_var.h>
72 #include <netinet/tcp_offload.h>
74 #include <opencrypto/xform.h>
75 #include <vm/uma_dbg.h>
77 #include <vm/vm_pageout.h>
78 #include <vm/vm_page.h>
82 STAILQ_HEAD(, mbuf) m_head;
83 STAILQ_HEAD(, socket) so_head;
85 } __aligned(CACHE_LINE_SIZE);
87 struct ktls_domain_info {
92 struct ktls_domain_info ktls_domains[MAXMEMDOM];
93 static struct ktls_wq *ktls_wq;
94 static struct proc *ktls_proc;
95 LIST_HEAD(, ktls_crypto_backend) ktls_backends;
96 static struct rmlock ktls_backends_lock;
97 static uma_zone_t ktls_session_zone;
98 static uint16_t ktls_cpuid_lookup[MAXCPU];
100 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101 "Kernel TLS offload");
102 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
103 "Kernel TLS offload stats");
105 static int ktls_allow_unload;
106 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN,
107 &ktls_allow_unload, 0, "Allow software crypto modules to unload");
110 static int ktls_bind_threads = 1;
112 static int ktls_bind_threads;
114 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
115 &ktls_bind_threads, 0,
116 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
118 static u_int ktls_maxlen = 16384;
119 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN,
120 &ktls_maxlen, 0, "Maximum TLS record size");
122 static int ktls_number_threads;
123 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
124 &ktls_number_threads, 0,
125 "Number of TLS threads in thread-pool");
127 static bool ktls_offload_enable;
128 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
129 &ktls_offload_enable, 0,
130 "Enable support for kernel TLS offload");
132 static bool ktls_cbc_enable = true;
133 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
135 "Enable Support of AES-CBC crypto for kernel TLS");
137 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
138 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
139 &ktls_tasks_active, "Number of active tasks");
141 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
142 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
144 "Number of TLS records in queue to tasks for SW encryption");
146 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
147 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
149 "Number of TLS sockets in queue to tasks for SW decryption");
151 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
152 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
153 CTLFLAG_RD, &ktls_offload_total,
154 "Total successful TLS setups (parameters set)");
156 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
158 CTLFLAG_RD, &ktls_offload_enable_calls,
159 "Total number of TLS enable calls made");
161 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
162 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
163 &ktls_offload_active, "Total Active TLS sessions");
165 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
166 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
167 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
169 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
170 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
171 &ktls_offload_failed_crypto, "Total TLS crypto failures");
173 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
175 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
177 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
178 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
179 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
181 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
182 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
183 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
185 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
186 "Software TLS session stats");
187 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
188 "Hardware (ifnet) TLS session stats");
190 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
191 "TOE TLS session stats");
194 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
195 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
196 "Active number of software TLS sessions using AES-CBC");
198 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
199 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
200 "Active number of software TLS sessions using AES-GCM");
202 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
203 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
205 "Active number of ifnet TLS sessions using AES-CBC");
207 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
208 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
210 "Active number of ifnet TLS sessions using AES-GCM");
212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
214 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
216 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
218 &ktls_ifnet_reset_dropped,
219 "TLS sessions dropped after failing to update ifnet send tag");
221 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
222 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
223 &ktls_ifnet_reset_failed,
224 "TLS sessions that failed to allocate a new ifnet send tag");
226 static int ktls_ifnet_permitted;
227 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
228 &ktls_ifnet_permitted, 1,
229 "Whether to permit hardware (ifnet) TLS sessions");
232 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
233 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
235 "Active number of TOE TLS sessions using AES-CBC");
237 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
238 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
240 "Active number of TOE TLS sessions using AES-GCM");
243 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
245 static void ktls_cleanup(struct ktls_session *tls);
246 #if defined(INET) || defined(INET6)
247 static void ktls_reset_send_tag(void *context, int pending);
249 static void ktls_work_thread(void *ctx);
252 ktls_crypto_backend_register(struct ktls_crypto_backend *be)
254 struct ktls_crypto_backend *curr_be, *tmp;
256 if (be->api_version != KTLS_API_VERSION) {
257 printf("KTLS: API version mismatch (%d vs %d) for %s\n",
258 be->api_version, KTLS_API_VERSION,
263 rm_wlock(&ktls_backends_lock);
264 printf("KTLS: Registering crypto method %s with prio %d\n",
266 if (LIST_EMPTY(&ktls_backends)) {
267 LIST_INSERT_HEAD(&ktls_backends, be, next);
269 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) {
270 if (curr_be->prio < be->prio) {
271 LIST_INSERT_BEFORE(curr_be, be, next);
274 if (LIST_NEXT(curr_be, next) == NULL) {
275 LIST_INSERT_AFTER(curr_be, be, next);
280 rm_wunlock(&ktls_backends_lock);
285 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be)
287 struct ktls_crypto_backend *tmp;
290 * Don't error if the backend isn't registered. This permits
291 * MOD_UNLOAD handlers to use this function unconditionally.
293 rm_wlock(&ktls_backends_lock);
294 LIST_FOREACH(tmp, &ktls_backends, next) {
299 rm_wunlock(&ktls_backends_lock);
303 if (!ktls_allow_unload) {
304 rm_wunlock(&ktls_backends_lock);
306 "KTLS: Deregistering crypto method %s is not supported\n",
312 rm_wunlock(&ktls_backends_lock);
316 LIST_REMOVE(be, next);
317 rm_wunlock(&ktls_backends_lock);
321 #if defined(INET) || defined(INET6)
323 ktls_get_cpu(struct socket *so)
327 struct ktls_domain_info *di;
333 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
334 if (cpuid != NETISR_CPUID_NONE)
338 * Just use the flowid to shard connections in a repeatable
339 * fashion. Note that some crypto backends rely on the
340 * serialization provided by having the same connection use
344 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
345 di = &ktls_domains[inp->inp_numa_domain];
346 cpuid = di->cpu[inp->inp_flowid % di->count];
349 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
355 ktls_init(void *dummy __unused)
360 int count, domain, error, i;
362 rm_init(&ktls_backends_lock, "ktls backends");
363 LIST_INIT(&ktls_backends);
365 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
368 ktls_session_zone = uma_zcreate("ktls_session",
369 sizeof(struct ktls_session),
370 NULL, NULL, NULL, NULL,
374 * Initialize the workqueues to run the TLS work. We create a
375 * work queue for each CPU.
378 STAILQ_INIT(&ktls_wq[i].m_head);
379 STAILQ_INIT(&ktls_wq[i].so_head);
380 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
381 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
382 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
384 panic("Can't add KTLS thread %d error %d", i, error);
387 * Bind threads to cores. If ktls_bind_threads is >
388 * 1, then we bind to the NUMA domain.
390 if (ktls_bind_threads) {
391 if (ktls_bind_threads > 1) {
393 domain = pc->pc_domain;
394 CPU_COPY(&cpuset_domain[domain], &mask);
395 count = ktls_domains[domain].count;
396 ktls_domains[domain].cpu[count] = i;
397 ktls_domains[domain].count++;
401 error = cpuset_setthread(td->td_tid, &mask);
404 "Unable to bind KTLS thread for CPU %d error %d",
407 ktls_cpuid_lookup[ktls_number_threads] = i;
408 ktls_number_threads++;
412 * If we somehow have an empty domain, fall back to choosing
413 * among all KTLS threads.
415 if (ktls_bind_threads > 1) {
416 for (i = 0; i < vm_ndomains; i++) {
417 if (ktls_domains[i].count == 0) {
418 ktls_bind_threads = 1;
425 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
427 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
429 #if defined(INET) || defined(INET6)
431 ktls_create_session(struct socket *so, struct tls_enable *en,
432 struct ktls_session **tlsp)
434 struct ktls_session *tls;
437 /* Only TLS 1.0 - 1.3 are supported. */
438 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
440 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
441 en->tls_vminor > TLS_MINOR_VER_THREE)
444 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
446 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
448 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
451 /* All supported algorithms require a cipher key. */
452 if (en->cipher_key_len == 0)
455 /* No flags are currently supported. */
459 /* Common checks for supported algorithms. */
460 switch (en->cipher_algorithm) {
461 case CRYPTO_AES_NIST_GCM_16:
463 * auth_algorithm isn't used, but permit GMAC values
466 switch (en->auth_algorithm) {
468 #ifdef COMPAT_FREEBSD12
469 /* XXX: Really 13.0-current COMPAT. */
470 case CRYPTO_AES_128_NIST_GMAC:
471 case CRYPTO_AES_192_NIST_GMAC:
472 case CRYPTO_AES_256_NIST_GMAC:
478 if (en->auth_key_len != 0)
480 if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
481 en->iv_len != TLS_AEAD_GCM_LEN) ||
482 (en->tls_vminor == TLS_MINOR_VER_THREE &&
483 en->iv_len != TLS_1_3_GCM_IV_LEN))
487 switch (en->auth_algorithm) {
488 case CRYPTO_SHA1_HMAC:
490 * TLS 1.0 requires an implicit IV. TLS 1.1+
491 * all use explicit IVs.
493 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
494 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
500 case CRYPTO_SHA2_256_HMAC:
501 case CRYPTO_SHA2_384_HMAC:
502 /* Ignore any supplied IV. */
508 if (en->auth_key_len == 0)
515 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
517 counter_u64_add(ktls_offload_active, 1);
519 refcount_init(&tls->refcount, 1);
520 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
522 tls->wq_index = ktls_get_cpu(so);
524 tls->params.cipher_algorithm = en->cipher_algorithm;
525 tls->params.auth_algorithm = en->auth_algorithm;
526 tls->params.tls_vmajor = en->tls_vmajor;
527 tls->params.tls_vminor = en->tls_vminor;
528 tls->params.flags = en->flags;
529 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
531 /* Set the header and trailer lengths. */
532 tls->params.tls_hlen = sizeof(struct tls_record_layer);
533 switch (en->cipher_algorithm) {
534 case CRYPTO_AES_NIST_GCM_16:
536 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
537 * nonce. TLS 1.3 uses a 12 byte implicit IV.
539 if (en->tls_vminor < TLS_MINOR_VER_THREE)
540 tls->params.tls_hlen += sizeof(uint64_t);
541 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
544 * TLS 1.3 includes optional padding which we
545 * do not support, and also puts the "real" record
546 * type at the end of the encrypted data.
548 if (en->tls_vminor == TLS_MINOR_VER_THREE)
549 tls->params.tls_tlen += sizeof(uint8_t);
551 tls->params.tls_bs = 1;
554 switch (en->auth_algorithm) {
555 case CRYPTO_SHA1_HMAC:
556 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
557 /* Implicit IV, no nonce. */
559 tls->params.tls_hlen += AES_BLOCK_LEN;
561 tls->params.tls_tlen = AES_BLOCK_LEN +
564 case CRYPTO_SHA2_256_HMAC:
565 tls->params.tls_hlen += AES_BLOCK_LEN;
566 tls->params.tls_tlen = AES_BLOCK_LEN +
569 case CRYPTO_SHA2_384_HMAC:
570 tls->params.tls_hlen += AES_BLOCK_LEN;
571 tls->params.tls_tlen = AES_BLOCK_LEN +
575 panic("invalid hmac");
577 tls->params.tls_bs = AES_BLOCK_LEN;
580 panic("invalid cipher");
583 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
584 ("TLS header length too long: %d", tls->params.tls_hlen));
585 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
586 ("TLS trailer length too long: %d", tls->params.tls_tlen));
588 if (en->auth_key_len != 0) {
589 tls->params.auth_key_len = en->auth_key_len;
590 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
592 error = copyin(en->auth_key, tls->params.auth_key,
598 tls->params.cipher_key_len = en->cipher_key_len;
599 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
600 error = copyin(en->cipher_key, tls->params.cipher_key,
606 * This holds the implicit portion of the nonce for GCM and
607 * the initial implicit IV for TLS 1.0. The explicit portions
608 * of the IV are generated in ktls_frame().
610 if (en->iv_len != 0) {
611 tls->params.iv_len = en->iv_len;
612 error = copyin(en->iv, tls->params.iv, en->iv_len);
617 * For TLS 1.2, generate an 8-byte nonce as a counter
618 * to generate unique explicit IVs.
620 * Store this counter in the last 8 bytes of the IV
621 * array so that it is 8-byte aligned.
623 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
624 en->tls_vminor == TLS_MINOR_VER_TWO)
625 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
636 static struct ktls_session *
637 ktls_clone_session(struct ktls_session *tls)
639 struct ktls_session *tls_new;
641 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
643 counter_u64_add(ktls_offload_active, 1);
645 refcount_init(&tls_new->refcount, 1);
646 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
648 /* Copy fields from existing session. */
649 tls_new->params = tls->params;
650 tls_new->wq_index = tls->wq_index;
652 /* Deep copy keys. */
653 if (tls_new->params.auth_key != NULL) {
654 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
656 memcpy(tls_new->params.auth_key, tls->params.auth_key,
657 tls->params.auth_key_len);
660 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
662 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
663 tls->params.cipher_key_len);
670 ktls_cleanup(struct ktls_session *tls)
673 counter_u64_add(ktls_offload_active, -1);
675 case TCP_TLS_MODE_SW:
676 MPASS(tls->be != NULL);
677 switch (tls->params.cipher_algorithm) {
679 counter_u64_add(ktls_sw_cbc, -1);
681 case CRYPTO_AES_NIST_GCM_16:
682 counter_u64_add(ktls_sw_gcm, -1);
687 case TCP_TLS_MODE_IFNET:
688 switch (tls->params.cipher_algorithm) {
690 counter_u64_add(ktls_ifnet_cbc, -1);
692 case CRYPTO_AES_NIST_GCM_16:
693 counter_u64_add(ktls_ifnet_gcm, -1);
696 if (tls->snd_tag != NULL)
697 m_snd_tag_rele(tls->snd_tag);
700 case TCP_TLS_MODE_TOE:
701 switch (tls->params.cipher_algorithm) {
703 counter_u64_add(ktls_toe_cbc, -1);
705 case CRYPTO_AES_NIST_GCM_16:
706 counter_u64_add(ktls_toe_gcm, -1);
712 if (tls->params.auth_key != NULL) {
713 zfree(tls->params.auth_key, M_KTLS);
714 tls->params.auth_key = NULL;
715 tls->params.auth_key_len = 0;
717 if (tls->params.cipher_key != NULL) {
718 zfree(tls->params.cipher_key, M_KTLS);
719 tls->params.cipher_key = NULL;
720 tls->params.cipher_key_len = 0;
722 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
725 #if defined(INET) || defined(INET6)
729 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
737 if (inp->inp_flags2 & INP_FREED) {
741 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
745 if (inp->inp_socket == NULL) {
750 if (!(tp->t_flags & TF_TOE)) {
755 error = tcp_offload_alloc_tls_session(tp, tls, direction);
758 tls->mode = TCP_TLS_MODE_TOE;
759 switch (tls->params.cipher_algorithm) {
761 counter_u64_add(ktls_toe_cbc, 1);
763 case CRYPTO_AES_NIST_GCM_16:
764 counter_u64_add(ktls_toe_gcm, 1);
773 * Common code used when first enabling ifnet TLS on a connection or
774 * when allocating a new ifnet TLS session due to a routing change.
775 * This function allocates a new TLS send tag on whatever interface
776 * the connection is currently routed over.
779 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
780 struct m_snd_tag **mstp)
782 union if_snd_tag_alloc_params params;
784 struct nhop_object *nh;
789 if (inp->inp_flags2 & INP_FREED) {
793 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
797 if (inp->inp_socket == NULL) {
804 * Check administrative controls on ifnet TLS to determine if
805 * ifnet TLS should be denied.
807 * - Always permit 'force' requests.
808 * - ktls_ifnet_permitted == 0: always deny.
810 if (!force && ktls_ifnet_permitted == 0) {
816 * XXX: Use the cached route in the inpcb to find the
817 * interface. This should perhaps instead use
818 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
819 * enabled after a connection has completed key negotiation in
820 * userland, the cached route will be present in practice.
822 nh = inp->inp_route.ro_nh;
831 * Allocate a TLS + ratelimit tag if the connection has an
832 * existing pacing rate.
834 if (tp->t_pacing_rate != -1 &&
835 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
836 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
837 params.tls_rate_limit.inp = inp;
838 params.tls_rate_limit.tls = tls;
839 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
841 params.hdr.type = IF_SND_TAG_TYPE_TLS;
842 params.tls.inp = inp;
843 params.tls.tls = tls;
845 params.hdr.flowid = inp->inp_flowid;
846 params.hdr.flowtype = inp->inp_flowtype;
847 params.hdr.numa_domain = inp->inp_numa_domain;
850 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
854 if (inp->inp_vflag & INP_IPV6) {
855 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
860 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
865 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
872 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
874 struct m_snd_tag *mst;
877 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
879 tls->mode = TCP_TLS_MODE_IFNET;
881 switch (tls->params.cipher_algorithm) {
883 counter_u64_add(ktls_ifnet_cbc, 1);
885 case CRYPTO_AES_NIST_GCM_16:
886 counter_u64_add(ktls_ifnet_gcm, 1);
894 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
896 struct rm_priotracker prio;
897 struct ktls_crypto_backend *be;
900 * Choose the best software crypto backend. Backends are
901 * stored in sorted priority order (larget value == most
902 * important at the head of the list), so this just stops on
903 * the first backend that claims the session by returning
906 if (ktls_allow_unload)
907 rm_rlock(&ktls_backends_lock, &prio);
908 LIST_FOREACH(be, &ktls_backends, next) {
909 if (be->try(so, tls, direction) == 0)
911 KASSERT(tls->cipher == NULL,
912 ("ktls backend leaked a cipher pointer"));
915 if (ktls_allow_unload)
919 if (ktls_allow_unload)
920 rm_runlock(&ktls_backends_lock, &prio);
923 tls->mode = TCP_TLS_MODE_SW;
924 switch (tls->params.cipher_algorithm) {
926 counter_u64_add(ktls_sw_cbc, 1);
928 case CRYPTO_AES_NIST_GCM_16:
929 counter_u64_add(ktls_sw_gcm, 1);
936 * KTLS RX stores data in the socket buffer as a list of TLS records,
937 * where each record is stored as a control message containg the TLS
938 * header followed by data mbufs containing the decrypted data. This
939 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
940 * both encrypted and decrypted data. TLS records decrypted by a NIC
941 * should be queued to the socket buffer as records, but encrypted
942 * data which needs to be decrypted by software arrives as a stream of
943 * regular mbufs which need to be converted. In addition, there may
944 * already be pending encrypted data in the socket buffer when KTLS RX
947 * To manage not-yet-decrypted data for KTLS RX, the following scheme
950 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
952 * - ktls_check_rx checks this chain of mbufs reading the TLS header
953 * from the first mbuf. Once all of the data for that TLS record is
954 * queued, the socket is queued to a worker thread.
956 * - The worker thread calls ktls_decrypt to decrypt TLS records in
957 * the TLS chain. Each TLS record is detached from the TLS chain,
958 * decrypted, and inserted into the regular socket buffer chain as
959 * record starting with a control message holding the TLS header and
960 * a chain of mbufs holding the encrypted data.
964 sb_mark_notready(struct sockbuf *sb)
971 sb->sb_mbtail = NULL;
972 sb->sb_lastrecord = NULL;
973 for (; m != NULL; m = m->m_next) {
974 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
976 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
978 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
980 m->m_flags |= M_NOTREADY;
981 sb->sb_acc -= m->m_len;
982 sb->sb_tlscc += m->m_len;
985 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
986 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
991 ktls_enable_rx(struct socket *so, struct tls_enable *en)
993 struct ktls_session *tls;
996 if (!ktls_offload_enable)
1001 counter_u64_add(ktls_offload_enable_calls, 1);
1004 * This should always be true since only the TCP socket option
1005 * invokes this function.
1007 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1011 * XXX: Don't overwrite existing sessions. We should permit
1012 * this to support rekeying in the future.
1014 if (so->so_rcv.sb_tls_info != NULL)
1017 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1020 /* TLS 1.3 is not yet supported. */
1021 if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1022 en->tls_vminor == TLS_MINOR_VER_THREE)
1025 error = ktls_create_session(so, en, &tls);
1030 error = ktls_try_toe(so, tls, KTLS_RX);
1033 error = ktls_try_sw(so, tls, KTLS_RX);
1040 /* Mark the socket as using TLS offload. */
1041 SOCKBUF_LOCK(&so->so_rcv);
1042 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1043 so->so_rcv.sb_tls_info = tls;
1044 so->so_rcv.sb_flags |= SB_TLS_RX;
1046 /* Mark existing data as not ready until it can be decrypted. */
1047 if (tls->mode != TCP_TLS_MODE_TOE) {
1048 sb_mark_notready(&so->so_rcv);
1049 ktls_check_rx(&so->so_rcv);
1051 SOCKBUF_UNLOCK(&so->so_rcv);
1053 counter_u64_add(ktls_offload_total, 1);
1059 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1061 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_snd.sb_tls_info != NULL)
1086 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1089 /* TLS requires ext pgs */
1090 if (mb_use_ext_pgs == 0)
1093 error = ktls_create_session(so, en, &tls);
1097 /* Prefer TOE -> ifnet TLS -> software TLS. */
1099 error = ktls_try_toe(so, tls, KTLS_TX);
1102 error = ktls_try_ifnet(so, tls, false);
1104 error = ktls_try_sw(so, tls, KTLS_TX);
1111 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1118 * Write lock the INP when setting sb_tls_info so that
1119 * routines in tcp_ratelimit.c can read sb_tls_info while
1120 * holding the INP lock.
1124 SOCKBUF_LOCK(&so->so_snd);
1125 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1126 so->so_snd.sb_tls_info = tls;
1127 if (tls->mode != TCP_TLS_MODE_SW)
1128 so->so_snd.sb_flags |= SB_TLS_IFNET;
1129 SOCKBUF_UNLOCK(&so->so_snd);
1131 SOCK_IO_SEND_UNLOCK(so);
1133 counter_u64_add(ktls_offload_total, 1);
1139 ktls_get_rx_mode(struct socket *so)
1141 struct ktls_session *tls;
1145 if (SOLISTENING(so))
1148 INP_WLOCK_ASSERT(inp);
1149 SOCKBUF_LOCK(&so->so_rcv);
1150 tls = so->so_rcv.sb_tls_info;
1152 mode = TCP_TLS_MODE_NONE;
1155 SOCKBUF_UNLOCK(&so->so_rcv);
1160 ktls_get_tx_mode(struct socket *so)
1162 struct ktls_session *tls;
1166 if (SOLISTENING(so))
1169 INP_WLOCK_ASSERT(inp);
1170 SOCKBUF_LOCK(&so->so_snd);
1171 tls = so->so_snd.sb_tls_info;
1173 mode = TCP_TLS_MODE_NONE;
1176 SOCKBUF_UNLOCK(&so->so_snd);
1181 * Switch between SW and ifnet TLS sessions as requested.
1184 ktls_set_tx_mode(struct socket *so, int mode)
1186 struct ktls_session *tls, *tls_new;
1190 if (SOLISTENING(so))
1193 case TCP_TLS_MODE_SW:
1194 case TCP_TLS_MODE_IFNET:
1201 INP_WLOCK_ASSERT(inp);
1202 SOCKBUF_LOCK(&so->so_snd);
1203 tls = so->so_snd.sb_tls_info;
1205 SOCKBUF_UNLOCK(&so->so_snd);
1209 if (tls->mode == mode) {
1210 SOCKBUF_UNLOCK(&so->so_snd);
1214 tls = ktls_hold(tls);
1215 SOCKBUF_UNLOCK(&so->so_snd);
1218 tls_new = ktls_clone_session(tls);
1220 if (mode == TCP_TLS_MODE_IFNET)
1221 error = ktls_try_ifnet(so, tls_new, true);
1223 error = ktls_try_sw(so, tls_new, KTLS_TX);
1225 counter_u64_add(ktls_switch_failed, 1);
1232 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1234 counter_u64_add(ktls_switch_failed, 1);
1242 * If we raced with another session change, keep the existing
1245 if (tls != so->so_snd.sb_tls_info) {
1246 counter_u64_add(ktls_switch_failed, 1);
1247 SOCK_IO_SEND_UNLOCK(so);
1254 SOCKBUF_LOCK(&so->so_snd);
1255 so->so_snd.sb_tls_info = tls_new;
1256 if (tls_new->mode != TCP_TLS_MODE_SW)
1257 so->so_snd.sb_flags |= SB_TLS_IFNET;
1258 SOCKBUF_UNLOCK(&so->so_snd);
1259 SOCK_IO_SEND_UNLOCK(so);
1262 * Drop two references on 'tls'. The first is for the
1263 * ktls_hold() above. The second drops the reference from the
1266 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1270 if (mode == TCP_TLS_MODE_IFNET)
1271 counter_u64_add(ktls_switch_to_ifnet, 1);
1273 counter_u64_add(ktls_switch_to_sw, 1);
1280 * Try to allocate a new TLS send tag. This task is scheduled when
1281 * ip_output detects a route change while trying to transmit a packet
1282 * holding a TLS record. If a new tag is allocated, replace the tag
1283 * in the TLS session. Subsequent packets on the connection will use
1284 * the new tag. If a new tag cannot be allocated, drop the
1288 ktls_reset_send_tag(void *context, int pending)
1290 struct epoch_tracker et;
1291 struct ktls_session *tls;
1292 struct m_snd_tag *old, *new;
1297 MPASS(pending == 1);
1303 * Free the old tag first before allocating a new one.
1304 * ip[6]_output_send() will treat a NULL send tag the same as
1305 * an ifp mismatch and drop packets until a new tag is
1308 * Write-lock the INP when changing tls->snd_tag since
1309 * ip[6]_output_send() holds a read-lock when reading the
1314 tls->snd_tag = NULL;
1317 m_snd_tag_rele(old);
1319 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1324 mtx_pool_lock(mtxpool_sleep, tls);
1325 tls->reset_pending = false;
1326 mtx_pool_unlock(mtxpool_sleep, tls);
1327 if (!in_pcbrele_wlocked(inp))
1330 counter_u64_add(ktls_ifnet_reset, 1);
1333 * XXX: Should we kick tcp_output explicitly now that
1334 * the send tag is fixed or just rely on timers?
1337 NET_EPOCH_ENTER(et);
1339 if (!in_pcbrele_wlocked(inp)) {
1340 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1341 !(inp->inp_flags & INP_DROPPED)) {
1342 tp = intotcpcb(inp);
1343 CURVNET_SET(tp->t_vnet);
1344 tp = tcp_drop(tp, ECONNABORTED);
1348 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1354 counter_u64_add(ktls_ifnet_reset_failed, 1);
1357 * Leave reset_pending true to avoid future tasks while
1358 * the socket goes away.
1366 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1372 INP_LOCK_ASSERT(inp);
1375 * See if we should schedule a task to update the send tag for
1378 mtx_pool_lock(mtxpool_sleep, tls);
1379 if (!tls->reset_pending) {
1380 (void) ktls_hold(tls);
1383 tls->reset_pending = true;
1384 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1386 mtx_pool_unlock(mtxpool_sleep, tls);
1392 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1394 union if_snd_tag_modify_params params = {
1395 .rate_limit.max_rate = max_pacing_rate,
1396 .rate_limit.flags = M_NOWAIT,
1398 struct m_snd_tag *mst;
1402 /* Can't get to the inp, but it should be locked. */
1403 /* INP_LOCK_ASSERT(inp); */
1405 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1407 if (tls->snd_tag == NULL) {
1409 * Resetting send tag, ignore this change. The
1410 * pending reset may or may not see this updated rate
1411 * in the tcpcb. If it doesn't, we will just lose
1417 MPASS(tls->snd_tag != NULL);
1418 MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1422 return (ifp->if_snd_tag_modify(mst, ¶ms));
1428 ktls_destroy(struct ktls_session *tls)
1430 struct rm_priotracker prio;
1433 if (tls->be != NULL && ktls_allow_unload) {
1434 rm_rlock(&ktls_backends_lock, &prio);
1435 tls->be->use_count--;
1436 rm_runlock(&ktls_backends_lock, &prio);
1438 uma_zfree(ktls_session_zone, tls);
1442 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1445 for (; m != NULL; m = m->m_next) {
1446 KASSERT((m->m_flags & M_EXTPG) != 0,
1447 ("ktls_seq: mapped mbuf %p", m));
1449 m->m_epg_seqno = sb->sb_tls_seqno;
1455 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1456 * mbuf in the chain must be an unmapped mbuf. The payload of the
1457 * mbuf must be populated with the payload of each TLS record.
1459 * The record_type argument specifies the TLS record type used when
1460 * populating the TLS header.
1462 * The enq_count argument on return is set to the number of pages of
1463 * payload data for this entire chain that need to be encrypted via SW
1464 * encryption. The returned value should be passed to ktls_enqueue
1465 * when scheduling encryption of this chain of mbufs. To handle the
1466 * special case of empty fragments for TLS 1.0 sessions, an empty
1467 * fragment counts as one page.
1470 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1471 uint8_t record_type)
1473 struct tls_record_layer *tlshdr;
1479 maxlen = tls->params.max_frame_len;
1481 for (m = top; m != NULL; m = m->m_next) {
1483 * All mbufs in the chain should be TLS records whose
1484 * payload does not exceed the maximum frame length.
1486 * Empty TLS records are permitted when using CBC.
1488 KASSERT(m->m_len <= maxlen &&
1489 (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1490 m->m_len >= 0 : m->m_len > 0),
1491 ("ktls_frame: m %p len %d\n", m, m->m_len));
1494 * TLS frames require unmapped mbufs to store session
1497 KASSERT((m->m_flags & M_EXTPG) != 0,
1498 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1502 /* Save a reference to the session. */
1503 m->m_epg_tls = ktls_hold(tls);
1505 m->m_epg_hdrlen = tls->params.tls_hlen;
1506 m->m_epg_trllen = tls->params.tls_tlen;
1507 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1511 * AES-CBC pads messages to a multiple of the
1512 * block size. Note that the padding is
1513 * applied after the digest and the encryption
1514 * is done on the "plaintext || mac || padding".
1515 * At least one byte of padding is always
1518 * Compute the final trailer length assuming
1519 * at most one block of padding.
1520 * tls->params.sb_tls_tlen is the maximum
1521 * possible trailer length (padding + digest).
1522 * delta holds the number of excess padding
1523 * bytes if the maximum were used. Those
1524 * extra bytes are removed.
1526 bs = tls->params.tls_bs;
1527 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1528 m->m_epg_trllen -= delta;
1530 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1532 /* Populate the TLS header. */
1533 tlshdr = (void *)m->m_epg_hdr;
1534 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1537 * TLS 1.3 masquarades as TLS 1.2 with a record type
1538 * of TLS_RLTYPE_APP.
1540 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1541 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1542 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1543 tlshdr->tls_type = TLS_RLTYPE_APP;
1544 /* save the real record type for later */
1545 m->m_epg_record_type = record_type;
1546 m->m_epg_trail[0] = record_type;
1548 tlshdr->tls_vminor = tls->params.tls_vminor;
1549 tlshdr->tls_type = record_type;
1551 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1554 * Store nonces / explicit IVs after the end of the
1557 * For GCM with TLS 1.2, an 8 byte nonce is copied
1558 * from the end of the IV. The nonce is then
1559 * incremented for use by the next record.
1561 * For CBC, a random nonce is inserted for TLS 1.1+.
1563 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1564 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1565 noncep = (uint64_t *)(tls->params.iv + 8);
1566 be64enc(tlshdr + 1, *noncep);
1568 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1569 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1570 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1573 * When using SW encryption, mark the mbuf not ready.
1574 * It will be marked ready via sbready() after the
1575 * record has been encrypted.
1577 * When using ifnet TLS, unencrypted TLS records are
1578 * sent down the stack to the NIC.
1580 if (tls->mode == TCP_TLS_MODE_SW) {
1581 m->m_flags |= M_NOTREADY;
1582 if (__predict_false(tls_len == 0)) {
1583 /* TLS 1.0 empty fragment. */
1586 m->m_epg_nrdy = m->m_epg_npgs;
1587 *enq_cnt += m->m_epg_nrdy;
1593 ktls_check_rx(struct sockbuf *sb)
1595 struct tls_record_layer hdr;
1600 SOCKBUF_LOCK_ASSERT(sb);
1601 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1603 so = __containerof(sb, struct socket, so_rcv);
1605 if (sb->sb_flags & SB_TLS_RX_RUNNING)
1608 /* Is there enough queued for a TLS header? */
1609 if (sb->sb_tlscc < sizeof(hdr)) {
1610 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1611 so->so_error = EMSGSIZE;
1615 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1617 /* Is the entire record queued? */
1618 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1619 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1620 so->so_error = EMSGSIZE;
1624 sb->sb_flags |= SB_TLS_RX_RUNNING;
1627 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1629 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1630 running = wq->running;
1631 mtx_unlock(&wq->mtx);
1634 counter_u64_add(ktls_cnt_rx_queued, 1);
1637 static struct mbuf *
1638 ktls_detach_record(struct sockbuf *sb, int len)
1640 struct mbuf *m, *n, *top;
1643 SOCKBUF_LOCK_ASSERT(sb);
1644 MPASS(len <= sb->sb_tlscc);
1647 * If TLS chain is the exact size of the record,
1648 * just grab the whole record.
1651 if (sb->sb_tlscc == len) {
1653 sb->sb_mtlstail = NULL;
1658 * While it would be nice to use m_split() here, we need
1659 * to know exactly what m_split() allocates to update the
1660 * accounting, so do it inline instead.
1663 for (m = top; remain > m->m_len; m = m->m_next)
1666 /* Easy case: don't have to split 'm'. */
1667 if (remain == m->m_len) {
1668 sb->sb_mtls = m->m_next;
1669 if (sb->sb_mtls == NULL)
1670 sb->sb_mtlstail = NULL;
1676 * Need to allocate an mbuf to hold the remainder of 'm'. Try
1677 * with M_NOWAIT first.
1679 n = m_get(M_NOWAIT, MT_DATA);
1682 * Use M_WAITOK with socket buffer unlocked. If
1683 * 'sb_mtls' changes while the lock is dropped, return
1684 * NULL to force the caller to retry.
1688 n = m_get(M_WAITOK, MT_DATA);
1691 if (sb->sb_mtls != top) {
1696 n->m_flags |= M_NOTREADY;
1698 /* Store remainder in 'n'. */
1699 n->m_len = m->m_len - remain;
1700 if (m->m_flags & M_EXT) {
1701 n->m_data = m->m_data + remain;
1704 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1707 /* Trim 'm' and update accounting. */
1708 m->m_len -= n->m_len;
1709 sb->sb_tlscc -= n->m_len;
1710 sb->sb_ccc -= n->m_len;
1712 /* Account for 'n'. */
1713 sballoc_ktls_rx(sb, n);
1715 /* Insert 'n' into the TLS chain. */
1717 n->m_next = m->m_next;
1718 if (sb->sb_mtlstail == m)
1719 sb->sb_mtlstail = n;
1721 /* Detach the record from the TLS chain. */
1725 MPASS(m_length(top, NULL) == len);
1726 for (m = top; m != NULL; m = m->m_next)
1727 sbfree_ktls_rx(sb, m);
1728 sb->sb_tlsdcc = len;
1735 ktls_decrypt(struct socket *so)
1737 char tls_header[MBUF_PEXT_HDR_LEN];
1738 struct ktls_session *tls;
1740 struct tls_record_layer *hdr;
1741 struct tls_get_record tgr;
1742 struct mbuf *control, *data, *m;
1744 int error, remain, tls_len, trail_len;
1746 hdr = (struct tls_record_layer *)tls_header;
1749 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1750 ("%s: socket %p not running", __func__, so));
1752 tls = sb->sb_tls_info;
1756 /* Is there enough queued for a TLS header? */
1757 if (sb->sb_tlscc < tls->params.tls_hlen)
1760 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1761 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1763 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1764 hdr->tls_vminor != tls->params.tls_vminor)
1766 else if (tls_len < tls->params.tls_hlen || tls_len >
1767 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1768 tls->params.tls_tlen)
1772 if (__predict_false(error != 0)) {
1774 * We have a corrupted record and are likely
1775 * out of sync. The connection isn't
1776 * recoverable at this point, so abort it.
1779 counter_u64_add(ktls_offload_corrupted_records, 1);
1781 CURVNET_SET(so->so_vnet);
1782 so->so_proto->pr_usrreqs->pru_abort(so);
1783 so->so_error = error;
1788 /* Is the entire record queued? */
1789 if (sb->sb_tlscc < tls_len)
1793 * Split out the portion of the mbuf chain containing
1796 data = ktls_detach_record(sb, tls_len);
1799 MPASS(sb->sb_tlsdcc == tls_len);
1801 seqno = sb->sb_tls_seqno;
1806 error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1808 counter_u64_add(ktls_offload_failed_crypto, 1);
1811 if (sb->sb_tlsdcc == 0) {
1813 * sbcut/drop/flush discarded these
1821 * Drop this TLS record's data, but keep
1822 * decrypting subsequent records.
1824 sb->sb_ccc -= tls_len;
1827 CURVNET_SET(so->so_vnet);
1828 so->so_error = EBADMSG;
1829 sorwakeup_locked(so);
1838 /* Allocate the control mbuf. */
1839 tgr.tls_type = hdr->tls_type;
1840 tgr.tls_vmajor = hdr->tls_vmajor;
1841 tgr.tls_vminor = hdr->tls_vminor;
1842 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1844 control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1845 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1848 if (sb->sb_tlsdcc == 0) {
1849 /* sbcut/drop/flush discarded these mbufs. */
1850 MPASS(sb->sb_tlscc == 0);
1857 * Clear the 'dcc' accounting in preparation for
1858 * adding the decrypted record.
1860 sb->sb_ccc -= tls_len;
1864 /* If there is no payload, drop all of the data. */
1865 if (tgr.tls_length == htobe16(0)) {
1870 remain = tls->params.tls_hlen;
1871 while (remain > 0) {
1872 if (data->m_len > remain) {
1873 data->m_data += remain;
1874 data->m_len -= remain;
1877 remain -= data->m_len;
1878 data = m_free(data);
1881 /* Trim trailer and clear M_NOTREADY. */
1882 remain = be16toh(tgr.tls_length);
1884 for (m = data; remain > m->m_len; m = m->m_next) {
1885 m->m_flags &= ~M_NOTREADY;
1891 m->m_flags &= ~M_NOTREADY;
1893 /* Set EOR on the final mbuf. */
1894 m->m_flags |= M_EOR;
1897 sbappendcontrol_locked(sb, data, control, 0);
1900 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1902 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1903 so->so_error = EMSGSIZE;
1905 sorwakeup_locked(so);
1908 SOCKBUF_UNLOCK_ASSERT(sb);
1910 CURVNET_SET(so->so_vnet);
1917 ktls_enqueue_to_free(struct mbuf *m)
1922 /* Mark it for freeing. */
1923 m->m_epg_flags |= EPG_FLAG_2FREE;
1924 wq = &ktls_wq[m->m_epg_tls->wq_index];
1926 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1927 running = wq->running;
1928 mtx_unlock(&wq->mtx);
1934 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1939 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1940 (M_EXTPG | M_NOTREADY)),
1941 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1942 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1944 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1946 m->m_epg_enc_cnt = page_count;
1949 * Save a pointer to the socket. The caller is responsible
1950 * for taking an additional reference via soref().
1954 wq = &ktls_wq[m->m_epg_tls->wq_index];
1956 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1957 running = wq->running;
1958 mtx_unlock(&wq->mtx);
1961 counter_u64_add(ktls_cnt_tx_queued, 1);
1964 static __noinline void
1965 ktls_encrypt(struct mbuf *top)
1967 struct ktls_session *tls;
1970 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1971 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1972 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1974 int error, i, len, npages, off, total_pages;
1978 tls = top->m_epg_tls;
1979 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
1980 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
1982 top->m_epg_so = NULL;
1984 total_pages = top->m_epg_enc_cnt;
1988 * Encrypt the TLS records in the chain of mbufs starting with
1989 * 'top'. 'total_pages' gives us a total count of pages and is
1990 * used to know when we have finished encrypting the TLS
1991 * records originally queued with 'top'.
1993 * NB: These mbufs are queued in the socket buffer and
1994 * 'm_next' is traversing the mbufs in the socket buffer. The
1995 * socket buffer lock is not held while traversing this chain.
1996 * Since the mbufs are all marked M_NOTREADY their 'm_next'
1997 * pointers should be stable. However, the 'm_next' of the
1998 * last mbuf encrypted is not necessarily NULL. It can point
1999 * to other mbufs appended while 'top' was on the TLS work
2002 * Each mbuf holds an entire TLS record.
2005 for (m = top; npages != total_pages; m = m->m_next) {
2006 KASSERT(m->m_epg_tls == tls,
2007 ("different TLS sessions in a single mbuf chain: %p vs %p",
2008 tls, m->m_epg_tls));
2009 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2010 (M_EXTPG | M_NOTREADY),
2011 ("%p not unready & nomap mbuf (top = %p)\n", m, top));
2012 KASSERT(npages + m->m_epg_npgs <= total_pages,
2013 ("page count mismatch: top %p, total_pages %d, m %p", top,
2017 * Generate source and destination ivoecs to pass to
2018 * the SW encryption backend. For writable mbufs, the
2019 * destination iovec is a copy of the source and
2020 * encryption is done in place. For file-backed mbufs
2021 * (from sendfile), anonymous wired pages are
2022 * allocated and assigned to the destination iovec.
2024 is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
2026 off = m->m_epg_1st_off;
2027 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2028 len = m_epg_pagelen(m, i, off);
2029 src_iov[i].iov_len = len;
2030 src_iov[i].iov_base =
2031 (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
2035 dst_iov[i].iov_base = src_iov[i].iov_base;
2036 dst_iov[i].iov_len = src_iov[i].iov_len;
2040 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2046 parray[i] = VM_PAGE_TO_PHYS(pg);
2047 dst_iov[i].iov_base =
2048 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
2049 dst_iov[i].iov_len = len;
2052 npages += m->m_epg_nrdy;
2054 error = (*tls->sw_encrypt)(tls,
2055 (const struct tls_record_layer *)m->m_epg_hdr,
2056 m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
2057 m->m_epg_record_type);
2059 counter_u64_add(ktls_offload_failed_crypto, 1);
2064 * For file-backed mbufs, release the file-backed
2065 * pages and replace them in the ext_pgs array with
2066 * the anonymous wired pages allocated above.
2069 /* Free the old pages. */
2070 m->m_ext.ext_free(m);
2072 /* Replace them with the new pages. */
2073 for (i = 0; i < m->m_epg_npgs; i++)
2074 m->m_epg_pa[i] = parray[i];
2076 /* Use the basic free routine. */
2077 m->m_ext.ext_free = mb_free_mext_pgs;
2079 /* Pages are now writable. */
2080 m->m_epg_flags |= EPG_FLAG_ANON;
2084 * Drop a reference to the session now that it is no
2085 * longer needed. Existing code depends on encrypted
2086 * records having no associated session vs
2087 * yet-to-be-encrypted records having an associated
2090 m->m_epg_tls = NULL;
2094 CURVNET_SET(so->so_vnet);
2096 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2098 so->so_proto->pr_usrreqs->pru_abort(so);
2100 mb_free_notready(top, total_pages);
2109 ktls_work_thread(void *ctx)
2111 struct ktls_wq *wq = ctx;
2113 struct socket *so, *son;
2114 STAILQ_HEAD(, mbuf) local_m_head;
2115 STAILQ_HEAD(, socket) local_so_head;
2117 if (ktls_bind_threads > 1) {
2118 curthread->td_domain.dr_policy =
2119 DOMAINSET_PREF(PCPU_GET(domain));
2121 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2126 while (STAILQ_EMPTY(&wq->m_head) &&
2127 STAILQ_EMPTY(&wq->so_head)) {
2128 wq->running = false;
2129 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2133 STAILQ_INIT(&local_m_head);
2134 STAILQ_CONCAT(&local_m_head, &wq->m_head);
2135 STAILQ_INIT(&local_so_head);
2136 STAILQ_CONCAT(&local_so_head, &wq->so_head);
2137 mtx_unlock(&wq->mtx);
2139 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2140 if (m->m_epg_flags & EPG_FLAG_2FREE) {
2141 ktls_free(m->m_epg_tls);
2145 counter_u64_add(ktls_cnt_tx_queued, -1);
2149 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2151 counter_u64_add(ktls_cnt_rx_queued, -1);