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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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_t 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 ktls_tasks_active = counter_u64_alloc(M_WAITOK);
363 ktls_cnt_tx_queued = counter_u64_alloc(M_WAITOK);
364 ktls_cnt_rx_queued = counter_u64_alloc(M_WAITOK);
365 ktls_offload_total = counter_u64_alloc(M_WAITOK);
366 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK);
367 ktls_offload_active = counter_u64_alloc(M_WAITOK);
368 ktls_offload_corrupted_records = counter_u64_alloc(M_WAITOK);
369 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK);
370 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK);
371 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK);
372 ktls_switch_failed = counter_u64_alloc(M_WAITOK);
373 ktls_sw_cbc = counter_u64_alloc(M_WAITOK);
374 ktls_sw_gcm = counter_u64_alloc(M_WAITOK);
375 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK);
376 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK);
377 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK);
378 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK);
379 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK);
381 ktls_toe_cbc = counter_u64_alloc(M_WAITOK);
382 ktls_toe_gcm = counter_u64_alloc(M_WAITOK);
385 rm_init(&ktls_backends_lock, "ktls backends");
386 LIST_INIT(&ktls_backends);
388 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
391 ktls_session_zone = uma_zcreate("ktls_session",
392 sizeof(struct ktls_session),
393 NULL, NULL, NULL, NULL,
397 * Initialize the workqueues to run the TLS work. We create a
398 * work queue for each CPU.
401 STAILQ_INIT(&ktls_wq[i].m_head);
402 STAILQ_INIT(&ktls_wq[i].so_head);
403 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
404 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
405 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
407 panic("Can't add KTLS thread %d error %d", i, error);
410 * Bind threads to cores. If ktls_bind_threads is >
411 * 1, then we bind to the NUMA domain.
413 if (ktls_bind_threads) {
414 if (ktls_bind_threads > 1) {
416 domain = pc->pc_domain;
417 CPU_COPY(&cpuset_domain[domain], &mask);
418 count = ktls_domains[domain].count;
419 ktls_domains[domain].cpu[count] = i;
420 ktls_domains[domain].count++;
424 error = cpuset_setthread(td->td_tid, &mask);
427 "Unable to bind KTLS thread for CPU %d error %d",
430 ktls_cpuid_lookup[ktls_number_threads] = i;
431 ktls_number_threads++;
435 * If we somehow have an empty domain, fall back to choosing
436 * among all KTLS threads.
438 if (ktls_bind_threads > 1) {
439 for (i = 0; i < vm_ndomains; i++) {
440 if (ktls_domains[i].count == 0) {
441 ktls_bind_threads = 1;
447 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
449 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
451 #if defined(INET) || defined(INET6)
453 ktls_create_session(struct socket *so, struct tls_enable *en,
454 struct ktls_session **tlsp)
456 struct ktls_session *tls;
459 /* Only TLS 1.0 - 1.3 are supported. */
460 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
462 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
463 en->tls_vminor > TLS_MINOR_VER_THREE)
466 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
468 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
470 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
473 /* All supported algorithms require a cipher key. */
474 if (en->cipher_key_len == 0)
477 /* No flags are currently supported. */
481 /* Common checks for supported algorithms. */
482 switch (en->cipher_algorithm) {
483 case CRYPTO_AES_NIST_GCM_16:
485 * auth_algorithm isn't used, but permit GMAC values
488 switch (en->auth_algorithm) {
490 #ifdef COMPAT_FREEBSD12
491 /* XXX: Really 13.0-current COMPAT. */
492 case CRYPTO_AES_128_NIST_GMAC:
493 case CRYPTO_AES_192_NIST_GMAC:
494 case CRYPTO_AES_256_NIST_GMAC:
500 if (en->auth_key_len != 0)
502 if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
503 en->iv_len != TLS_AEAD_GCM_LEN) ||
504 (en->tls_vminor == TLS_MINOR_VER_THREE &&
505 en->iv_len != TLS_1_3_GCM_IV_LEN))
509 switch (en->auth_algorithm) {
510 case CRYPTO_SHA1_HMAC:
512 * TLS 1.0 requires an implicit IV. TLS 1.1+
513 * all use explicit IVs.
515 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
516 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
522 case CRYPTO_SHA2_256_HMAC:
523 case CRYPTO_SHA2_384_HMAC:
524 /* Ignore any supplied IV. */
530 if (en->auth_key_len == 0)
537 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
539 counter_u64_add(ktls_offload_active, 1);
541 refcount_init(&tls->refcount, 1);
542 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
544 tls->wq_index = ktls_get_cpu(so);
546 tls->params.cipher_algorithm = en->cipher_algorithm;
547 tls->params.auth_algorithm = en->auth_algorithm;
548 tls->params.tls_vmajor = en->tls_vmajor;
549 tls->params.tls_vminor = en->tls_vminor;
550 tls->params.flags = en->flags;
551 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
553 /* Set the header and trailer lengths. */
554 tls->params.tls_hlen = sizeof(struct tls_record_layer);
555 switch (en->cipher_algorithm) {
556 case CRYPTO_AES_NIST_GCM_16:
558 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
559 * nonce. TLS 1.3 uses a 12 byte implicit IV.
561 if (en->tls_vminor < TLS_MINOR_VER_THREE)
562 tls->params.tls_hlen += sizeof(uint64_t);
563 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
566 * TLS 1.3 includes optional padding which we
567 * do not support, and also puts the "real" record
568 * type at the end of the encrypted data.
570 if (en->tls_vminor == TLS_MINOR_VER_THREE)
571 tls->params.tls_tlen += sizeof(uint8_t);
573 tls->params.tls_bs = 1;
576 switch (en->auth_algorithm) {
577 case CRYPTO_SHA1_HMAC:
578 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
579 /* Implicit IV, no nonce. */
581 tls->params.tls_hlen += AES_BLOCK_LEN;
583 tls->params.tls_tlen = AES_BLOCK_LEN +
586 case CRYPTO_SHA2_256_HMAC:
587 tls->params.tls_hlen += AES_BLOCK_LEN;
588 tls->params.tls_tlen = AES_BLOCK_LEN +
591 case CRYPTO_SHA2_384_HMAC:
592 tls->params.tls_hlen += AES_BLOCK_LEN;
593 tls->params.tls_tlen = AES_BLOCK_LEN +
597 panic("invalid hmac");
599 tls->params.tls_bs = AES_BLOCK_LEN;
602 panic("invalid cipher");
605 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
606 ("TLS header length too long: %d", tls->params.tls_hlen));
607 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
608 ("TLS trailer length too long: %d", tls->params.tls_tlen));
610 if (en->auth_key_len != 0) {
611 tls->params.auth_key_len = en->auth_key_len;
612 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
614 error = copyin(en->auth_key, tls->params.auth_key,
620 tls->params.cipher_key_len = en->cipher_key_len;
621 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
622 error = copyin(en->cipher_key, tls->params.cipher_key,
628 * This holds the implicit portion of the nonce for GCM and
629 * the initial implicit IV for TLS 1.0. The explicit portions
630 * of the IV are generated in ktls_frame().
632 if (en->iv_len != 0) {
633 tls->params.iv_len = en->iv_len;
634 error = copyin(en->iv, tls->params.iv, en->iv_len);
639 * For TLS 1.2, generate an 8-byte nonce as a counter
640 * to generate unique explicit IVs.
642 * Store this counter in the last 8 bytes of the IV
643 * array so that it is 8-byte aligned.
645 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
646 en->tls_vminor == TLS_MINOR_VER_TWO)
647 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
658 static struct ktls_session *
659 ktls_clone_session(struct ktls_session *tls)
661 struct ktls_session *tls_new;
663 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
665 counter_u64_add(ktls_offload_active, 1);
667 refcount_init(&tls_new->refcount, 1);
669 /* Copy fields from existing session. */
670 tls_new->params = tls->params;
671 tls_new->wq_index = tls->wq_index;
673 /* Deep copy keys. */
674 if (tls_new->params.auth_key != NULL) {
675 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
677 memcpy(tls_new->params.auth_key, tls->params.auth_key,
678 tls->params.auth_key_len);
681 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
683 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
684 tls->params.cipher_key_len);
691 ktls_cleanup(struct ktls_session *tls)
694 counter_u64_add(ktls_offload_active, -1);
696 case TCP_TLS_MODE_SW:
697 MPASS(tls->be != NULL);
698 switch (tls->params.cipher_algorithm) {
700 counter_u64_add(ktls_sw_cbc, -1);
702 case CRYPTO_AES_NIST_GCM_16:
703 counter_u64_add(ktls_sw_gcm, -1);
708 case TCP_TLS_MODE_IFNET:
709 switch (tls->params.cipher_algorithm) {
711 counter_u64_add(ktls_ifnet_cbc, -1);
713 case CRYPTO_AES_NIST_GCM_16:
714 counter_u64_add(ktls_ifnet_gcm, -1);
717 if (tls->snd_tag != NULL)
718 m_snd_tag_rele(tls->snd_tag);
721 case TCP_TLS_MODE_TOE:
722 switch (tls->params.cipher_algorithm) {
724 counter_u64_add(ktls_toe_cbc, -1);
726 case CRYPTO_AES_NIST_GCM_16:
727 counter_u64_add(ktls_toe_gcm, -1);
733 if (tls->params.auth_key != NULL) {
734 zfree(tls->params.auth_key, M_KTLS);
735 tls->params.auth_key = NULL;
736 tls->params.auth_key_len = 0;
738 if (tls->params.cipher_key != NULL) {
739 zfree(tls->params.cipher_key, M_KTLS);
740 tls->params.cipher_key = NULL;
741 tls->params.cipher_key_len = 0;
743 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
746 #if defined(INET) || defined(INET6)
750 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
758 if (inp->inp_flags2 & INP_FREED) {
762 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
766 if (inp->inp_socket == NULL) {
771 if (!(tp->t_flags & TF_TOE)) {
776 error = tcp_offload_alloc_tls_session(tp, tls, direction);
779 tls->mode = TCP_TLS_MODE_TOE;
780 switch (tls->params.cipher_algorithm) {
782 counter_u64_add(ktls_toe_cbc, 1);
784 case CRYPTO_AES_NIST_GCM_16:
785 counter_u64_add(ktls_toe_gcm, 1);
794 * Common code used when first enabling ifnet TLS on a connection or
795 * when allocating a new ifnet TLS session due to a routing change.
796 * This function allocates a new TLS send tag on whatever interface
797 * the connection is currently routed over.
800 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
801 struct m_snd_tag **mstp)
803 union if_snd_tag_alloc_params params;
805 struct nhop_object *nh;
810 if (inp->inp_flags2 & INP_FREED) {
814 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
818 if (inp->inp_socket == NULL) {
825 * Check administrative controls on ifnet TLS to determine if
826 * ifnet TLS should be denied.
828 * - Always permit 'force' requests.
829 * - ktls_ifnet_permitted == 0: always deny.
831 if (!force && ktls_ifnet_permitted == 0) {
837 * XXX: Use the cached route in the inpcb to find the
838 * interface. This should perhaps instead use
839 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
840 * enabled after a connection has completed key negotiation in
841 * userland, the cached route will be present in practice.
843 nh = inp->inp_route.ro_nh;
852 * Allocate a TLS + ratelimit tag if the connection has an
853 * existing pacing rate.
855 if (tp->t_pacing_rate != -1 &&
856 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
857 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
858 params.tls_rate_limit.inp = inp;
859 params.tls_rate_limit.tls = tls;
860 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
862 params.hdr.type = IF_SND_TAG_TYPE_TLS;
863 params.tls.inp = inp;
864 params.tls.tls = tls;
866 params.hdr.flowid = inp->inp_flowid;
867 params.hdr.flowtype = inp->inp_flowtype;
868 params.hdr.numa_domain = inp->inp_numa_domain;
871 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
875 if (inp->inp_vflag & INP_IPV6) {
876 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
881 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
886 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
893 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
895 struct m_snd_tag *mst;
898 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
900 tls->mode = TCP_TLS_MODE_IFNET;
902 switch (tls->params.cipher_algorithm) {
904 counter_u64_add(ktls_ifnet_cbc, 1);
906 case CRYPTO_AES_NIST_GCM_16:
907 counter_u64_add(ktls_ifnet_gcm, 1);
915 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
917 struct rm_priotracker prio;
918 struct ktls_crypto_backend *be;
921 * Choose the best software crypto backend. Backends are
922 * stored in sorted priority order (larget value == most
923 * important at the head of the list), so this just stops on
924 * the first backend that claims the session by returning
927 if (ktls_allow_unload)
928 rm_rlock(&ktls_backends_lock, &prio);
929 LIST_FOREACH(be, &ktls_backends, next) {
930 if (be->try(so, tls, direction) == 0)
932 KASSERT(tls->cipher == NULL,
933 ("ktls backend leaked a cipher pointer"));
936 if (ktls_allow_unload)
940 if (ktls_allow_unload)
941 rm_runlock(&ktls_backends_lock, &prio);
944 tls->mode = TCP_TLS_MODE_SW;
945 switch (tls->params.cipher_algorithm) {
947 counter_u64_add(ktls_sw_cbc, 1);
949 case CRYPTO_AES_NIST_GCM_16:
950 counter_u64_add(ktls_sw_gcm, 1);
957 * KTLS RX stores data in the socket buffer as a list of TLS records,
958 * where each record is stored as a control message containg the TLS
959 * header followed by data mbufs containing the decrypted data. This
960 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
961 * both encrypted and decrypted data. TLS records decrypted by a NIC
962 * should be queued to the socket buffer as records, but encrypted
963 * data which needs to be decrypted by software arrives as a stream of
964 * regular mbufs which need to be converted. In addition, there may
965 * already be pending encrypted data in the socket buffer when KTLS RX
968 * To manage not-yet-decrypted data for KTLS RX, the following scheme
971 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
973 * - ktls_check_rx checks this chain of mbufs reading the TLS header
974 * from the first mbuf. Once all of the data for that TLS record is
975 * queued, the socket is queued to a worker thread.
977 * - The worker thread calls ktls_decrypt to decrypt TLS records in
978 * the TLS chain. Each TLS record is detached from the TLS chain,
979 * decrypted, and inserted into the regular socket buffer chain as
980 * record starting with a control message holding the TLS header and
981 * a chain of mbufs holding the encrypted data.
985 sb_mark_notready(struct sockbuf *sb)
992 sb->sb_mbtail = NULL;
993 sb->sb_lastrecord = NULL;
994 for (; m != NULL; m = m->m_next) {
995 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
997 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
999 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1001 m->m_flags |= M_NOTREADY;
1002 sb->sb_acc -= m->m_len;
1003 sb->sb_tlscc += m->m_len;
1004 sb->sb_mtlstail = m;
1006 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1007 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1012 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1014 struct ktls_session *tls;
1017 if (!ktls_offload_enable)
1019 if (SOLISTENING(so))
1022 counter_u64_add(ktls_offload_enable_calls, 1);
1025 * This should always be true since only the TCP socket option
1026 * invokes this function.
1028 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1032 * XXX: Don't overwrite existing sessions. We should permit
1033 * this to support rekeying in the future.
1035 if (so->so_rcv.sb_tls_info != NULL)
1038 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1041 /* TLS 1.3 is not yet supported. */
1042 if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1043 en->tls_vminor == TLS_MINOR_VER_THREE)
1046 error = ktls_create_session(so, en, &tls);
1051 error = ktls_try_toe(so, tls, KTLS_RX);
1054 error = ktls_try_sw(so, tls, KTLS_RX);
1061 /* Mark the socket as using TLS offload. */
1062 SOCKBUF_LOCK(&so->so_rcv);
1063 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1064 so->so_rcv.sb_tls_info = tls;
1065 so->so_rcv.sb_flags |= SB_TLS_RX;
1067 /* Mark existing data as not ready until it can be decrypted. */
1068 sb_mark_notready(&so->so_rcv);
1069 ktls_check_rx(&so->so_rcv);
1070 SOCKBUF_UNLOCK(&so->so_rcv);
1072 counter_u64_add(ktls_offload_total, 1);
1078 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1080 struct ktls_session *tls;
1084 if (!ktls_offload_enable)
1086 if (SOLISTENING(so))
1089 counter_u64_add(ktls_offload_enable_calls, 1);
1092 * This should always be true since only the TCP socket option
1093 * invokes this function.
1095 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1099 * XXX: Don't overwrite existing sessions. We should permit
1100 * this to support rekeying in the future.
1102 if (so->so_snd.sb_tls_info != NULL)
1105 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1108 /* TLS requires ext pgs */
1109 if (mb_use_ext_pgs == 0)
1112 error = ktls_create_session(so, en, &tls);
1116 /* Prefer TOE -> ifnet TLS -> software TLS. */
1118 error = ktls_try_toe(so, tls, KTLS_TX);
1121 error = ktls_try_ifnet(so, tls, false);
1123 error = ktls_try_sw(so, tls, KTLS_TX);
1130 error = sblock(&so->so_snd, SBL_WAIT);
1137 * Write lock the INP when setting sb_tls_info so that
1138 * routines in tcp_ratelimit.c can read sb_tls_info while
1139 * holding the INP lock.
1143 SOCKBUF_LOCK(&so->so_snd);
1144 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1145 so->so_snd.sb_tls_info = tls;
1146 if (tls->mode != TCP_TLS_MODE_SW)
1147 so->so_snd.sb_flags |= SB_TLS_IFNET;
1148 SOCKBUF_UNLOCK(&so->so_snd);
1150 sbunlock(&so->so_snd);
1152 counter_u64_add(ktls_offload_total, 1);
1158 ktls_get_rx_mode(struct socket *so)
1160 struct ktls_session *tls;
1164 if (SOLISTENING(so))
1167 INP_WLOCK_ASSERT(inp);
1168 SOCKBUF_LOCK(&so->so_rcv);
1169 tls = so->so_rcv.sb_tls_info;
1171 mode = TCP_TLS_MODE_NONE;
1174 SOCKBUF_UNLOCK(&so->so_rcv);
1179 ktls_get_tx_mode(struct socket *so)
1181 struct ktls_session *tls;
1185 if (SOLISTENING(so))
1188 INP_WLOCK_ASSERT(inp);
1189 SOCKBUF_LOCK(&so->so_snd);
1190 tls = so->so_snd.sb_tls_info;
1192 mode = TCP_TLS_MODE_NONE;
1195 SOCKBUF_UNLOCK(&so->so_snd);
1200 * Switch between SW and ifnet TLS sessions as requested.
1203 ktls_set_tx_mode(struct socket *so, int mode)
1205 struct ktls_session *tls, *tls_new;
1209 if (SOLISTENING(so))
1212 case TCP_TLS_MODE_SW:
1213 case TCP_TLS_MODE_IFNET:
1220 INP_WLOCK_ASSERT(inp);
1221 SOCKBUF_LOCK(&so->so_snd);
1222 tls = so->so_snd.sb_tls_info;
1224 SOCKBUF_UNLOCK(&so->so_snd);
1228 if (tls->mode == mode) {
1229 SOCKBUF_UNLOCK(&so->so_snd);
1233 tls = ktls_hold(tls);
1234 SOCKBUF_UNLOCK(&so->so_snd);
1237 tls_new = ktls_clone_session(tls);
1239 if (mode == TCP_TLS_MODE_IFNET)
1240 error = ktls_try_ifnet(so, tls_new, true);
1242 error = ktls_try_sw(so, tls_new, KTLS_TX);
1244 counter_u64_add(ktls_switch_failed, 1);
1251 error = sblock(&so->so_snd, SBL_WAIT);
1253 counter_u64_add(ktls_switch_failed, 1);
1261 * If we raced with another session change, keep the existing
1264 if (tls != so->so_snd.sb_tls_info) {
1265 counter_u64_add(ktls_switch_failed, 1);
1266 sbunlock(&so->so_snd);
1273 SOCKBUF_LOCK(&so->so_snd);
1274 so->so_snd.sb_tls_info = tls_new;
1275 if (tls_new->mode != TCP_TLS_MODE_SW)
1276 so->so_snd.sb_flags |= SB_TLS_IFNET;
1277 SOCKBUF_UNLOCK(&so->so_snd);
1278 sbunlock(&so->so_snd);
1281 * Drop two references on 'tls'. The first is for the
1282 * ktls_hold() above. The second drops the reference from the
1285 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1289 if (mode == TCP_TLS_MODE_IFNET)
1290 counter_u64_add(ktls_switch_to_ifnet, 1);
1292 counter_u64_add(ktls_switch_to_sw, 1);
1299 * Try to allocate a new TLS send tag. This task is scheduled when
1300 * ip_output detects a route change while trying to transmit a packet
1301 * holding a TLS record. If a new tag is allocated, replace the tag
1302 * in the TLS session. Subsequent packets on the connection will use
1303 * the new tag. If a new tag cannot be allocated, drop the
1307 ktls_reset_send_tag(void *context, int pending)
1309 struct epoch_tracker et;
1310 struct ktls_session *tls;
1311 struct m_snd_tag *old, *new;
1316 MPASS(pending == 1);
1322 * Free the old tag first before allocating a new one.
1323 * ip[6]_output_send() will treat a NULL send tag the same as
1324 * an ifp mismatch and drop packets until a new tag is
1327 * Write-lock the INP when changing tls->snd_tag since
1328 * ip[6]_output_send() holds a read-lock when reading the
1333 tls->snd_tag = NULL;
1336 m_snd_tag_rele(old);
1338 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1343 mtx_pool_lock(mtxpool_sleep, tls);
1344 tls->reset_pending = false;
1345 mtx_pool_unlock(mtxpool_sleep, tls);
1346 if (!in_pcbrele_wlocked(inp))
1349 counter_u64_add(ktls_ifnet_reset, 1);
1352 * XXX: Should we kick tcp_output explicitly now that
1353 * the send tag is fixed or just rely on timers?
1356 NET_EPOCH_ENTER(et);
1358 if (!in_pcbrele_wlocked(inp)) {
1359 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1360 !(inp->inp_flags & INP_DROPPED)) {
1361 tp = intotcpcb(inp);
1362 CURVNET_SET(tp->t_vnet);
1363 tp = tcp_drop(tp, ECONNABORTED);
1367 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1373 counter_u64_add(ktls_ifnet_reset_failed, 1);
1376 * Leave reset_pending true to avoid future tasks while
1377 * the socket goes away.
1385 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1391 INP_LOCK_ASSERT(inp);
1394 * See if we should schedule a task to update the send tag for
1397 mtx_pool_lock(mtxpool_sleep, tls);
1398 if (!tls->reset_pending) {
1399 (void) ktls_hold(tls);
1402 tls->reset_pending = true;
1403 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1405 mtx_pool_unlock(mtxpool_sleep, tls);
1411 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1413 union if_snd_tag_modify_params params = {
1414 .rate_limit.max_rate = max_pacing_rate,
1415 .rate_limit.flags = M_NOWAIT,
1417 struct m_snd_tag *mst;
1421 /* Can't get to the inp, but it should be locked. */
1422 /* INP_LOCK_ASSERT(inp); */
1424 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1426 if (tls->snd_tag == NULL) {
1428 * Resetting send tag, ignore this change. The
1429 * pending reset may or may not see this updated rate
1430 * in the tcpcb. If it doesn't, we will just lose
1436 MPASS(tls->snd_tag != NULL);
1437 MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1441 return (ifp->if_snd_tag_modify(mst, ¶ms));
1447 ktls_destroy(struct ktls_session *tls)
1449 struct rm_priotracker prio;
1452 if (tls->be != NULL && ktls_allow_unload) {
1453 rm_rlock(&ktls_backends_lock, &prio);
1454 tls->be->use_count--;
1455 rm_runlock(&ktls_backends_lock, &prio);
1457 uma_zfree(ktls_session_zone, tls);
1461 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1464 for (; m != NULL; m = m->m_next) {
1465 KASSERT((m->m_flags & M_EXTPG) != 0,
1466 ("ktls_seq: mapped mbuf %p", m));
1468 m->m_epg_seqno = sb->sb_tls_seqno;
1474 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1475 * mbuf in the chain must be an unmapped mbuf. The payload of the
1476 * mbuf must be populated with the payload of each TLS record.
1478 * The record_type argument specifies the TLS record type used when
1479 * populating the TLS header.
1481 * The enq_count argument on return is set to the number of pages of
1482 * payload data for this entire chain that need to be encrypted via SW
1483 * encryption. The returned value should be passed to ktls_enqueue
1484 * when scheduling encryption of this chain of mbufs. To handle the
1485 * special case of empty fragments for TLS 1.0 sessions, an empty
1486 * fragment counts as one page.
1489 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1490 uint8_t record_type)
1492 struct tls_record_layer *tlshdr;
1498 maxlen = tls->params.max_frame_len;
1500 for (m = top; m != NULL; m = m->m_next) {
1502 * All mbufs in the chain should be TLS records whose
1503 * payload does not exceed the maximum frame length.
1505 * Empty TLS records are permitted when using CBC.
1507 KASSERT(m->m_len <= maxlen &&
1508 (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1509 m->m_len >= 0 : m->m_len > 0),
1510 ("ktls_frame: m %p len %d\n", m, m->m_len));
1513 * TLS frames require unmapped mbufs to store session
1516 KASSERT((m->m_flags & M_EXTPG) != 0,
1517 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1521 /* Save a reference to the session. */
1522 m->m_epg_tls = ktls_hold(tls);
1524 m->m_epg_hdrlen = tls->params.tls_hlen;
1525 m->m_epg_trllen = tls->params.tls_tlen;
1526 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1530 * AES-CBC pads messages to a multiple of the
1531 * block size. Note that the padding is
1532 * applied after the digest and the encryption
1533 * is done on the "plaintext || mac || padding".
1534 * At least one byte of padding is always
1537 * Compute the final trailer length assuming
1538 * at most one block of padding.
1539 * tls->params.sb_tls_tlen is the maximum
1540 * possible trailer length (padding + digest).
1541 * delta holds the number of excess padding
1542 * bytes if the maximum were used. Those
1543 * extra bytes are removed.
1545 bs = tls->params.tls_bs;
1546 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1547 m->m_epg_trllen -= delta;
1549 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1551 /* Populate the TLS header. */
1552 tlshdr = (void *)m->m_epg_hdr;
1553 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1556 * TLS 1.3 masquarades as TLS 1.2 with a record type
1557 * of TLS_RLTYPE_APP.
1559 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1560 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1561 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1562 tlshdr->tls_type = TLS_RLTYPE_APP;
1563 /* save the real record type for later */
1564 m->m_epg_record_type = record_type;
1565 m->m_epg_trail[0] = record_type;
1567 tlshdr->tls_vminor = tls->params.tls_vminor;
1568 tlshdr->tls_type = record_type;
1570 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1573 * Store nonces / explicit IVs after the end of the
1576 * For GCM with TLS 1.2, an 8 byte nonce is copied
1577 * from the end of the IV. The nonce is then
1578 * incremented for use by the next record.
1580 * For CBC, a random nonce is inserted for TLS 1.1+.
1582 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1583 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1584 noncep = (uint64_t *)(tls->params.iv + 8);
1585 be64enc(tlshdr + 1, *noncep);
1587 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1588 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1589 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1592 * When using SW encryption, mark the mbuf not ready.
1593 * It will be marked ready via sbready() after the
1594 * record has been encrypted.
1596 * When using ifnet TLS, unencrypted TLS records are
1597 * sent down the stack to the NIC.
1599 if (tls->mode == TCP_TLS_MODE_SW) {
1600 m->m_flags |= M_NOTREADY;
1601 m->m_epg_nrdy = m->m_epg_npgs;
1602 if (__predict_false(tls_len == 0)) {
1603 /* TLS 1.0 empty fragment. */
1606 *enq_cnt += m->m_epg_npgs;
1612 ktls_check_rx(struct sockbuf *sb)
1614 struct tls_record_layer hdr;
1619 SOCKBUF_LOCK_ASSERT(sb);
1620 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1622 so = __containerof(sb, struct socket, so_rcv);
1624 if (sb->sb_flags & SB_TLS_RX_RUNNING)
1627 /* Is there enough queued for a TLS header? */
1628 if (sb->sb_tlscc < sizeof(hdr)) {
1629 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1630 so->so_error = EMSGSIZE;
1634 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1636 /* Is the entire record queued? */
1637 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1638 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1639 so->so_error = EMSGSIZE;
1643 sb->sb_flags |= SB_TLS_RX_RUNNING;
1646 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1648 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1649 running = wq->running;
1650 mtx_unlock(&wq->mtx);
1653 counter_u64_add(ktls_cnt_rx_queued, 1);
1656 static struct mbuf *
1657 ktls_detach_record(struct sockbuf *sb, int len)
1659 struct mbuf *m, *n, *top;
1662 SOCKBUF_LOCK_ASSERT(sb);
1663 MPASS(len <= sb->sb_tlscc);
1666 * If TLS chain is the exact size of the record,
1667 * just grab the whole record.
1670 if (sb->sb_tlscc == len) {
1672 sb->sb_mtlstail = NULL;
1677 * While it would be nice to use m_split() here, we need
1678 * to know exactly what m_split() allocates to update the
1679 * accounting, so do it inline instead.
1682 for (m = top; remain > m->m_len; m = m->m_next)
1685 /* Easy case: don't have to split 'm'. */
1686 if (remain == m->m_len) {
1687 sb->sb_mtls = m->m_next;
1688 if (sb->sb_mtls == NULL)
1689 sb->sb_mtlstail = NULL;
1695 * Need to allocate an mbuf to hold the remainder of 'm'. Try
1696 * with M_NOWAIT first.
1698 n = m_get(M_NOWAIT, MT_DATA);
1701 * Use M_WAITOK with socket buffer unlocked. If
1702 * 'sb_mtls' changes while the lock is dropped, return
1703 * NULL to force the caller to retry.
1707 n = m_get(M_WAITOK, MT_DATA);
1710 if (sb->sb_mtls != top) {
1715 n->m_flags |= M_NOTREADY;
1717 /* Store remainder in 'n'. */
1718 n->m_len = m->m_len - remain;
1719 if (m->m_flags & M_EXT) {
1720 n->m_data = m->m_data + remain;
1723 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1726 /* Trim 'm' and update accounting. */
1727 m->m_len -= n->m_len;
1728 sb->sb_tlscc -= n->m_len;
1729 sb->sb_ccc -= n->m_len;
1731 /* Account for 'n'. */
1732 sballoc_ktls_rx(sb, n);
1734 /* Insert 'n' into the TLS chain. */
1736 n->m_next = m->m_next;
1737 if (sb->sb_mtlstail == m)
1738 sb->sb_mtlstail = n;
1740 /* Detach the record from the TLS chain. */
1744 MPASS(m_length(top, NULL) == len);
1745 for (m = top; m != NULL; m = m->m_next)
1746 sbfree_ktls_rx(sb, m);
1747 sb->sb_tlsdcc = len;
1754 ktls_decrypt(struct socket *so)
1756 char tls_header[MBUF_PEXT_HDR_LEN];
1757 struct ktls_session *tls;
1759 struct tls_record_layer *hdr;
1760 struct tls_get_record tgr;
1761 struct mbuf *control, *data, *m;
1763 int error, remain, tls_len, trail_len;
1765 hdr = (struct tls_record_layer *)tls_header;
1768 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1769 ("%s: socket %p not running", __func__, so));
1771 tls = sb->sb_tls_info;
1775 /* Is there enough queued for a TLS header? */
1776 if (sb->sb_tlscc < tls->params.tls_hlen)
1779 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1780 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1782 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1783 hdr->tls_vminor != tls->params.tls_vminor)
1785 else if (tls_len < tls->params.tls_hlen || tls_len >
1786 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1787 tls->params.tls_tlen)
1791 if (__predict_false(error != 0)) {
1793 * We have a corrupted record and are likely
1794 * out of sync. The connection isn't
1795 * recoverable at this point, so abort it.
1798 counter_u64_add(ktls_offload_corrupted_records, 1);
1800 CURVNET_SET(so->so_vnet);
1801 so->so_proto->pr_usrreqs->pru_abort(so);
1802 so->so_error = error;
1807 /* Is the entire record queued? */
1808 if (sb->sb_tlscc < tls_len)
1812 * Split out the portion of the mbuf chain containing
1815 data = ktls_detach_record(sb, tls_len);
1818 MPASS(sb->sb_tlsdcc == tls_len);
1820 seqno = sb->sb_tls_seqno;
1825 error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1827 counter_u64_add(ktls_offload_failed_crypto, 1);
1830 if (sb->sb_tlsdcc == 0) {
1832 * sbcut/drop/flush discarded these
1840 * Drop this TLS record's data, but keep
1841 * decrypting subsequent records.
1843 sb->sb_ccc -= tls_len;
1846 CURVNET_SET(so->so_vnet);
1847 so->so_error = EBADMSG;
1848 sorwakeup_locked(so);
1857 /* Allocate the control mbuf. */
1858 tgr.tls_type = hdr->tls_type;
1859 tgr.tls_vmajor = hdr->tls_vmajor;
1860 tgr.tls_vminor = hdr->tls_vminor;
1861 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1863 control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1864 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1867 if (sb->sb_tlsdcc == 0) {
1868 /* sbcut/drop/flush discarded these mbufs. */
1869 MPASS(sb->sb_tlscc == 0);
1876 * Clear the 'dcc' accounting in preparation for
1877 * adding the decrypted record.
1879 sb->sb_ccc -= tls_len;
1883 /* If there is no payload, drop all of the data. */
1884 if (tgr.tls_length == htobe16(0)) {
1889 remain = tls->params.tls_hlen;
1890 while (remain > 0) {
1891 if (data->m_len > remain) {
1892 data->m_data += remain;
1893 data->m_len -= remain;
1896 remain -= data->m_len;
1897 data = m_free(data);
1900 /* Trim trailer and clear M_NOTREADY. */
1901 remain = be16toh(tgr.tls_length);
1903 for (m = data; remain > m->m_len; m = m->m_next) {
1904 m->m_flags &= ~M_NOTREADY;
1910 m->m_flags &= ~M_NOTREADY;
1912 /* Set EOR on the final mbuf. */
1913 m->m_flags |= M_EOR;
1916 sbappendcontrol_locked(sb, data, control, 0);
1919 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1921 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1922 so->so_error = EMSGSIZE;
1924 sorwakeup_locked(so);
1927 SOCKBUF_UNLOCK_ASSERT(sb);
1929 CURVNET_SET(so->so_vnet);
1936 ktls_enqueue_to_free(struct mbuf *m)
1941 /* Mark it for freeing. */
1942 m->m_epg_flags |= EPG_FLAG_2FREE;
1943 wq = &ktls_wq[m->m_epg_tls->wq_index];
1945 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1946 running = wq->running;
1947 mtx_unlock(&wq->mtx);
1953 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
1958 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
1959 (M_EXTPG | M_NOTREADY)),
1960 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
1961 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
1963 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
1965 m->m_epg_enc_cnt = page_count;
1968 * Save a pointer to the socket. The caller is responsible
1969 * for taking an additional reference via soref().
1973 wq = &ktls_wq[m->m_epg_tls->wq_index];
1975 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1976 running = wq->running;
1977 mtx_unlock(&wq->mtx);
1980 counter_u64_add(ktls_cnt_tx_queued, 1);
1983 static __noinline void
1984 ktls_encrypt(struct mbuf *top)
1986 struct ktls_session *tls;
1989 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1990 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1991 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)];
1993 int error, i, len, npages, off, total_pages;
1997 tls = top->m_epg_tls;
1998 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
1999 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2001 top->m_epg_so = NULL;
2003 total_pages = top->m_epg_enc_cnt;
2007 * Encrypt the TLS records in the chain of mbufs starting with
2008 * 'top'. 'total_pages' gives us a total count of pages and is
2009 * used to know when we have finished encrypting the TLS
2010 * records originally queued with 'top'.
2012 * NB: These mbufs are queued in the socket buffer and
2013 * 'm_next' is traversing the mbufs in the socket buffer. The
2014 * socket buffer lock is not held while traversing this chain.
2015 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2016 * pointers should be stable. However, the 'm_next' of the
2017 * last mbuf encrypted is not necessarily NULL. It can point
2018 * to other mbufs appended while 'top' was on the TLS work
2021 * Each mbuf holds an entire TLS record.
2024 for (m = top; npages != total_pages; m = m->m_next) {
2025 KASSERT(m->m_epg_tls == tls,
2026 ("different TLS sessions in a single mbuf chain: %p vs %p",
2027 tls, m->m_epg_tls));
2028 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2029 (M_EXTPG | M_NOTREADY),
2030 ("%p not unready & nomap mbuf (top = %p)\n", m, top));
2031 KASSERT(npages + m->m_epg_npgs <= total_pages,
2032 ("page count mismatch: top %p, total_pages %d, m %p", top,
2036 * Generate source and destination ivoecs to pass to
2037 * the SW encryption backend. For writable mbufs, the
2038 * destination iovec is a copy of the source and
2039 * encryption is done in place. For file-backed mbufs
2040 * (from sendfile), anonymous wired pages are
2041 * allocated and assigned to the destination iovec.
2043 is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0;
2045 off = m->m_epg_1st_off;
2046 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2047 len = m_epg_pagelen(m, i, off);
2048 src_iov[i].iov_len = len;
2049 src_iov[i].iov_base =
2050 (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) +
2054 dst_iov[i].iov_base = src_iov[i].iov_base;
2055 dst_iov[i].iov_len = src_iov[i].iov_len;
2059 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2060 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
2065 parray[i] = VM_PAGE_TO_PHYS(pg);
2066 dst_iov[i].iov_base =
2067 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off;
2068 dst_iov[i].iov_len = len;
2071 if (__predict_false(m->m_epg_npgs == 0)) {
2072 /* TLS 1.0 empty fragment. */
2077 error = (*tls->sw_encrypt)(tls,
2078 (const struct tls_record_layer *)m->m_epg_hdr,
2079 m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno,
2080 m->m_epg_record_type);
2082 counter_u64_add(ktls_offload_failed_crypto, 1);
2087 * For file-backed mbufs, release the file-backed
2088 * pages and replace them in the ext_pgs array with
2089 * the anonymous wired pages allocated above.
2092 /* Free the old pages. */
2093 m->m_ext.ext_free(m);
2095 /* Replace them with the new pages. */
2096 for (i = 0; i < m->m_epg_npgs; i++)
2097 m->m_epg_pa[i] = parray[i];
2099 /* Use the basic free routine. */
2100 m->m_ext.ext_free = mb_free_mext_pgs;
2102 /* Pages are now writable. */
2103 m->m_epg_flags |= EPG_FLAG_ANON;
2107 * Drop a reference to the session now that it is no
2108 * longer needed. Existing code depends on encrypted
2109 * records having no associated session vs
2110 * yet-to-be-encrypted records having an associated
2113 m->m_epg_tls = NULL;
2117 CURVNET_SET(so->so_vnet);
2119 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2121 so->so_proto->pr_usrreqs->pru_abort(so);
2123 mb_free_notready(top, total_pages);
2132 ktls_work_thread(void *ctx)
2134 struct ktls_wq *wq = ctx;
2136 struct socket *so, *son;
2137 STAILQ_HEAD(, mbuf) local_m_head;
2138 STAILQ_HEAD(, socket) local_so_head;
2140 if (ktls_bind_threads > 1) {
2141 curthread->td_domain.dr_policy =
2142 DOMAINSET_PREF(PCPU_GET(domain));
2144 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2149 while (STAILQ_EMPTY(&wq->m_head) &&
2150 STAILQ_EMPTY(&wq->so_head)) {
2151 wq->running = false;
2152 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2156 STAILQ_INIT(&local_m_head);
2157 STAILQ_CONCAT(&local_m_head, &wq->m_head);
2158 STAILQ_INIT(&local_so_head);
2159 STAILQ_CONCAT(&local_so_head, &wq->so_head);
2160 mtx_unlock(&wq->mtx);
2162 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2163 if (m->m_epg_flags & EPG_FLAG_2FREE) {
2164 ktls_free(m->m_epg_tls);
2165 uma_zfree(zone_mbuf, m);
2168 counter_u64_add(ktls_cnt_tx_queued, -1);
2172 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2174 counter_u64_add(ktls_cnt_rx_queued, -1);