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"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/domainset.h>
40 #include <sys/endian.h>
44 #include <sys/mutex.h>
45 #include <sys/rmlock.h>
47 #include <sys/protosw.h>
48 #include <sys/refcount.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <sys/kthread.h>
56 #include <sys/vmmeter.h>
57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58 #include <machine/pcb.h>
60 #include <machine/vmparam.h>
62 #include <net/if_var.h>
64 #include <net/netisr.h>
65 #include <net/rss_config.h>
67 #include <net/route.h>
68 #include <net/route/nhop.h>
69 #if defined(INET) || defined(INET6)
70 #include <netinet/in.h>
71 #include <netinet/in_pcb.h>
73 #include <netinet/tcp_var.h>
75 #include <netinet/tcp_offload.h>
77 #include <opencrypto/cryptodev.h>
78 #include <opencrypto/ktls.h>
79 #include <vm/uma_dbg.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pagequeue.h>
87 STAILQ_HEAD(, mbuf) m_head;
88 STAILQ_HEAD(, socket) so_head;
91 } __aligned(CACHE_LINE_SIZE);
93 struct ktls_alloc_thread {
100 struct ktls_domain_info {
103 struct ktls_alloc_thread alloc_td;
106 struct ktls_domain_info ktls_domains[MAXMEMDOM];
107 static struct ktls_wq *ktls_wq;
108 static struct proc *ktls_proc;
109 static uma_zone_t ktls_session_zone;
110 static uma_zone_t ktls_buffer_zone;
111 static uint16_t ktls_cpuid_lookup[MAXCPU];
112 static int ktls_init_state;
113 static struct sx ktls_init_lock;
114 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
116 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117 "Kernel TLS offload");
118 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
119 "Kernel TLS offload stats");
122 static int ktls_bind_threads = 1;
124 static int ktls_bind_threads;
126 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
127 &ktls_bind_threads, 0,
128 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
130 static u_int ktls_maxlen = 16384;
131 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
132 &ktls_maxlen, 0, "Maximum TLS record size");
134 static int ktls_number_threads;
135 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
136 &ktls_number_threads, 0,
137 "Number of TLS threads in thread-pool");
139 unsigned int ktls_ifnet_max_rexmit_pct = 2;
140 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
141 &ktls_ifnet_max_rexmit_pct, 2,
142 "Max percent bytes retransmitted before ifnet TLS is disabled");
144 static bool ktls_offload_enable;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
146 &ktls_offload_enable, 0,
147 "Enable support for kernel TLS offload");
149 static bool ktls_cbc_enable = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
152 "Enable support of AES-CBC crypto for kernel TLS");
154 static bool ktls_sw_buffer_cache = true;
155 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
156 &ktls_sw_buffer_cache, 1,
157 "Enable caching of output buffers for SW encryption");
159 static int ktls_max_alloc = 128;
160 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
161 &ktls_max_alloc, 128,
162 "Max number of 16k buffers to allocate in thread context");
164 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
165 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
166 &ktls_tasks_active, "Number of active tasks");
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
170 &ktls_cnt_tx_pending,
171 "Number of TLS 1.0 records waiting for earlier TLS records");
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
176 "Number of TLS records in queue to tasks for SW encryption");
178 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
181 "Number of TLS sockets in queue to tasks for SW decryption");
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
185 CTLFLAG_RD, &ktls_offload_total,
186 "Total successful TLS setups (parameters set)");
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
190 CTLFLAG_RD, &ktls_offload_enable_calls,
191 "Total number of TLS enable calls made");
193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
195 &ktls_offload_active, "Total Active TLS sessions");
197 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
199 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
201 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
203 &ktls_offload_failed_crypto, "Total TLS crypto failures");
205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
207 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
209 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
211 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
213 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
215 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
217 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
218 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
219 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
221 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
222 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
223 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
225 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226 "Software TLS session stats");
227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228 "Hardware (ifnet) TLS session stats");
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231 "TOE TLS session stats");
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236 "Active number of software TLS sessions using AES-CBC");
238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240 "Active number of software TLS sessions using AES-GCM");
242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
245 "Active number of software TLS sessions using Chacha20-Poly1305");
247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
250 "Active number of ifnet TLS sessions using AES-CBC");
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
255 "Active number of ifnet TLS sessions using AES-GCM");
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259 &ktls_ifnet_chacha20,
260 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268 &ktls_ifnet_reset_dropped,
269 "TLS sessions dropped after failing to update ifnet send tag");
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273 &ktls_ifnet_reset_failed,
274 "TLS sessions that failed to allocate a new ifnet send tag");
276 static int ktls_ifnet_permitted;
277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278 &ktls_ifnet_permitted, 1,
279 "Whether to permit hardware (ifnet) TLS sessions");
282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
285 "Active number of TOE TLS sessions using AES-CBC");
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
290 "Active number of TOE TLS sessions using AES-GCM");
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
295 "Active number of TOE TLS sessions using Chacha20-Poly1305");
298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
300 #if defined(INET) || defined(INET6)
301 static void ktls_reset_receive_tag(void *context, int pending);
302 static void ktls_reset_send_tag(void *context, int pending);
304 static void ktls_work_thread(void *ctx);
305 static void ktls_alloc_thread(void *ctx);
307 #if defined(INET) || defined(INET6)
309 ktls_get_cpu(struct socket *so)
313 struct ktls_domain_info *di;
319 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
320 if (cpuid != NETISR_CPUID_NONE)
324 * Just use the flowid to shard connections in a repeatable
325 * fashion. Note that TLS 1.0 sessions rely on the
326 * serialization provided by having the same connection use
330 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
331 di = &ktls_domains[inp->inp_numa_domain];
332 cpuid = di->cpu[inp->inp_flowid % di->count];
335 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
341 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
346 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
347 ("%s: ktls max length %d is not page size-aligned",
348 __func__, ktls_maxlen));
350 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
351 for (i = 0; i < count; i++) {
352 m = vm_page_alloc_noobj_contig_domain(domain, req,
353 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
357 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
363 ktls_buffer_release(void *arg __unused, void **store, int count)
368 for (i = 0; i < count; i++) {
369 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
370 for (j = 0; j < atop(ktls_maxlen); j++) {
371 (void)vm_page_unwire_noq(m + j);
378 ktls_free_mext_contig(struct mbuf *m)
381 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
389 int count, domain, error, i;
391 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
394 ktls_session_zone = uma_zcreate("ktls_session",
395 sizeof(struct ktls_session),
396 NULL, NULL, NULL, NULL,
399 if (ktls_sw_buffer_cache) {
400 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
401 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
402 ktls_buffer_import, ktls_buffer_release, NULL,
403 UMA_ZONE_FIRSTTOUCH);
407 * Initialize the workqueues to run the TLS work. We create a
408 * work queue for each CPU.
411 STAILQ_INIT(&ktls_wq[i].m_head);
412 STAILQ_INIT(&ktls_wq[i].so_head);
413 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
414 if (ktls_bind_threads > 1) {
416 domain = pc->pc_domain;
417 count = ktls_domains[domain].count;
418 ktls_domains[domain].cpu[count] = i;
419 ktls_domains[domain].count++;
421 ktls_cpuid_lookup[ktls_number_threads] = i;
422 ktls_number_threads++;
426 * If we somehow have an empty domain, fall back to choosing
427 * among all KTLS threads.
429 if (ktls_bind_threads > 1) {
430 for (i = 0; i < vm_ndomains; i++) {
431 if (ktls_domains[i].count == 0) {
432 ktls_bind_threads = 1;
438 /* Start kthreads for each workqueue. */
440 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
441 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
443 printf("Can't add KTLS thread %d error %d\n", i, error);
449 * Start an allocation thread per-domain to perform blocking allocations
450 * of 16k physically contiguous TLS crypto destination buffers.
452 if (ktls_sw_buffer_cache) {
453 for (domain = 0; domain < vm_ndomains; domain++) {
454 if (VM_DOMAIN_EMPTY(domain))
456 if (CPU_EMPTY(&cpuset_domain[domain]))
458 error = kproc_kthread_add(ktls_alloc_thread,
459 &ktls_domains[domain], &ktls_proc,
460 &ktls_domains[domain].alloc_td.td,
461 0, 0, "KTLS", "alloc_%d", domain);
463 printf("Can't add KTLS alloc thread %d error %d\n",
471 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
476 ktls_start_kthreads(void)
481 state = atomic_load_acq_int(&ktls_init_state);
482 if (__predict_true(state > 0))
487 sx_xlock(&ktls_init_lock);
488 if (ktls_init_state != 0) {
489 sx_xunlock(&ktls_init_lock);
498 atomic_store_rel_int(&ktls_init_state, state);
499 sx_xunlock(&ktls_init_lock);
503 #if defined(INET) || defined(INET6)
505 ktls_create_session(struct socket *so, struct tls_enable *en,
506 struct ktls_session **tlsp, int direction)
508 struct ktls_session *tls;
511 /* Only TLS 1.0 - 1.3 are supported. */
512 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
514 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
515 en->tls_vminor > TLS_MINOR_VER_THREE)
518 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
520 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
522 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
525 /* All supported algorithms require a cipher key. */
526 if (en->cipher_key_len == 0)
529 /* No flags are currently supported. */
533 /* Common checks for supported algorithms. */
534 switch (en->cipher_algorithm) {
535 case CRYPTO_AES_NIST_GCM_16:
537 * auth_algorithm isn't used, but permit GMAC values
540 switch (en->auth_algorithm) {
542 #ifdef COMPAT_FREEBSD12
543 /* XXX: Really 13.0-current COMPAT. */
544 case CRYPTO_AES_128_NIST_GMAC:
545 case CRYPTO_AES_192_NIST_GMAC:
546 case CRYPTO_AES_256_NIST_GMAC:
552 if (en->auth_key_len != 0)
554 switch (en->tls_vminor) {
555 case TLS_MINOR_VER_TWO:
556 if (en->iv_len != TLS_AEAD_GCM_LEN)
559 case TLS_MINOR_VER_THREE:
560 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
568 switch (en->auth_algorithm) {
569 case CRYPTO_SHA1_HMAC:
571 case CRYPTO_SHA2_256_HMAC:
572 case CRYPTO_SHA2_384_HMAC:
573 if (en->tls_vminor != TLS_MINOR_VER_TWO)
579 if (en->auth_key_len == 0)
583 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
586 switch (en->tls_vminor) {
587 case TLS_MINOR_VER_ZERO:
588 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
591 case TLS_MINOR_VER_ONE:
592 case TLS_MINOR_VER_TWO:
593 /* Ignore any supplied IV. */
600 case CRYPTO_CHACHA20_POLY1305:
601 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
603 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
604 en->tls_vminor != TLS_MINOR_VER_THREE)
606 if (en->iv_len != TLS_CHACHA20_IV_LEN)
613 error = ktls_start_kthreads();
617 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
619 counter_u64_add(ktls_offload_active, 1);
621 refcount_init(&tls->refcount, 1);
622 if (direction == KTLS_RX)
623 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
625 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
627 tls->wq_index = ktls_get_cpu(so);
629 tls->params.cipher_algorithm = en->cipher_algorithm;
630 tls->params.auth_algorithm = en->auth_algorithm;
631 tls->params.tls_vmajor = en->tls_vmajor;
632 tls->params.tls_vminor = en->tls_vminor;
633 tls->params.flags = en->flags;
634 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
636 /* Set the header and trailer lengths. */
637 tls->params.tls_hlen = sizeof(struct tls_record_layer);
638 switch (en->cipher_algorithm) {
639 case CRYPTO_AES_NIST_GCM_16:
641 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
642 * nonce. TLS 1.3 uses a 12 byte implicit IV.
644 if (en->tls_vminor < TLS_MINOR_VER_THREE)
645 tls->params.tls_hlen += sizeof(uint64_t);
646 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
647 tls->params.tls_bs = 1;
650 switch (en->auth_algorithm) {
651 case CRYPTO_SHA1_HMAC:
652 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
653 /* Implicit IV, no nonce. */
654 tls->sequential_records = true;
655 tls->next_seqno = be64dec(en->rec_seq);
656 STAILQ_INIT(&tls->pending_records);
658 tls->params.tls_hlen += AES_BLOCK_LEN;
660 tls->params.tls_tlen = AES_BLOCK_LEN +
663 case CRYPTO_SHA2_256_HMAC:
664 tls->params.tls_hlen += AES_BLOCK_LEN;
665 tls->params.tls_tlen = AES_BLOCK_LEN +
668 case CRYPTO_SHA2_384_HMAC:
669 tls->params.tls_hlen += AES_BLOCK_LEN;
670 tls->params.tls_tlen = AES_BLOCK_LEN +
674 panic("invalid hmac");
676 tls->params.tls_bs = AES_BLOCK_LEN;
678 case CRYPTO_CHACHA20_POLY1305:
680 * Chacha20 uses a 12 byte implicit IV.
682 tls->params.tls_tlen = POLY1305_HASH_LEN;
683 tls->params.tls_bs = 1;
686 panic("invalid cipher");
690 * TLS 1.3 includes optional padding which we do not support,
691 * and also puts the "real" record type at the end of the
694 if (en->tls_vminor == TLS_MINOR_VER_THREE)
695 tls->params.tls_tlen += sizeof(uint8_t);
697 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
698 ("TLS header length too long: %d", tls->params.tls_hlen));
699 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
700 ("TLS trailer length too long: %d", tls->params.tls_tlen));
702 if (en->auth_key_len != 0) {
703 tls->params.auth_key_len = en->auth_key_len;
704 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
706 error = copyin(en->auth_key, tls->params.auth_key,
712 tls->params.cipher_key_len = en->cipher_key_len;
713 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
714 error = copyin(en->cipher_key, tls->params.cipher_key,
720 * This holds the implicit portion of the nonce for AEAD
721 * ciphers and the initial implicit IV for TLS 1.0. The
722 * explicit portions of the IV are generated in ktls_frame().
724 if (en->iv_len != 0) {
725 tls->params.iv_len = en->iv_len;
726 error = copyin(en->iv, tls->params.iv, en->iv_len);
731 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
732 * counter to generate unique explicit IVs.
734 * Store this counter in the last 8 bytes of the IV
735 * array so that it is 8-byte aligned.
737 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
738 en->tls_vminor == TLS_MINOR_VER_TWO)
739 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
750 static struct ktls_session *
751 ktls_clone_session(struct ktls_session *tls, int direction)
753 struct ktls_session *tls_new;
755 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
757 counter_u64_add(ktls_offload_active, 1);
759 refcount_init(&tls_new->refcount, 1);
760 if (direction == KTLS_RX)
761 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
764 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
767 /* Copy fields from existing session. */
768 tls_new->params = tls->params;
769 tls_new->wq_index = tls->wq_index;
771 /* Deep copy keys. */
772 if (tls_new->params.auth_key != NULL) {
773 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
775 memcpy(tls_new->params.auth_key, tls->params.auth_key,
776 tls->params.auth_key_len);
779 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
781 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
782 tls->params.cipher_key_len);
789 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
797 if (inp->inp_flags & INP_DROPPED) {
801 if (inp->inp_socket == NULL) {
806 if (!(tp->t_flags & TF_TOE)) {
811 error = tcp_offload_alloc_tls_session(tp, tls, direction);
814 tls->mode = TCP_TLS_MODE_TOE;
815 switch (tls->params.cipher_algorithm) {
817 counter_u64_add(ktls_toe_cbc, 1);
819 case CRYPTO_AES_NIST_GCM_16:
820 counter_u64_add(ktls_toe_gcm, 1);
822 case CRYPTO_CHACHA20_POLY1305:
823 counter_u64_add(ktls_toe_chacha20, 1);
832 * Common code used when first enabling ifnet TLS on a connection or
833 * when allocating a new ifnet TLS session due to a routing change.
834 * This function allocates a new TLS send tag on whatever interface
835 * the connection is currently routed over.
838 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
839 struct m_snd_tag **mstp)
841 union if_snd_tag_alloc_params params;
843 struct nhop_object *nh;
848 if (inp->inp_flags & INP_DROPPED) {
852 if (inp->inp_socket == NULL) {
859 * Check administrative controls on ifnet TLS to determine if
860 * ifnet TLS should be denied.
862 * - Always permit 'force' requests.
863 * - ktls_ifnet_permitted == 0: always deny.
865 if (!force && ktls_ifnet_permitted == 0) {
871 * XXX: Use the cached route in the inpcb to find the
872 * interface. This should perhaps instead use
873 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
874 * enabled after a connection has completed key negotiation in
875 * userland, the cached route will be present in practice.
877 nh = inp->inp_route.ro_nh;
886 * Allocate a TLS + ratelimit tag if the connection has an
887 * existing pacing rate.
889 if (tp->t_pacing_rate != -1 &&
890 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
891 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
892 params.tls_rate_limit.inp = inp;
893 params.tls_rate_limit.tls = tls;
894 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
896 params.hdr.type = IF_SND_TAG_TYPE_TLS;
897 params.tls.inp = inp;
898 params.tls.tls = tls;
900 params.hdr.flowid = inp->inp_flowid;
901 params.hdr.flowtype = inp->inp_flowtype;
902 params.hdr.numa_domain = inp->inp_numa_domain;
905 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
909 if (inp->inp_vflag & INP_IPV6) {
910 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
915 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
920 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
927 * Allocate an initial TLS receive tag for doing HW decryption of TLS
930 * This function allocates a new TLS receive tag on whatever interface
931 * the connection is currently routed over. If the connection ends up
932 * using a different interface for receive this will get fixed up via
933 * ktls_input_ifp_mismatch as future packets arrive.
936 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
937 struct m_snd_tag **mstp)
939 union if_snd_tag_alloc_params params;
941 struct nhop_object *nh;
944 if (!ktls_ocf_recrypt_supported(tls))
948 if (inp->inp_flags & INP_DROPPED) {
952 if (inp->inp_socket == NULL) {
958 * Check administrative controls on ifnet TLS to determine if
959 * ifnet TLS should be denied.
961 if (ktls_ifnet_permitted == 0) {
967 * XXX: As with ktls_alloc_snd_tag, use the cached route in
968 * the inpcb to find the interface.
970 nh = inp->inp_route.ro_nh;
979 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
980 params.hdr.flowid = inp->inp_flowid;
981 params.hdr.flowtype = inp->inp_flowtype;
982 params.hdr.numa_domain = inp->inp_numa_domain;
983 params.tls_rx.inp = inp;
984 params.tls_rx.tls = tls;
985 params.tls_rx.vlan_id = 0;
989 if (inp->inp_vflag & INP_IPV6) {
990 if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) {
995 if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) {
1000 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1003 * If this connection is over a vlan, vlan_snd_tag_alloc
1004 * rewrites vlan_id with the saved interface. Save the VLAN
1005 * ID for use in ktls_reset_receive_tag which allocates new
1006 * receive tags directly from the leaf interface bypassing
1010 tls->rx_vlan_id = params.tls_rx.vlan_id;
1016 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1019 struct m_snd_tag *mst;
1022 switch (direction) {
1024 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1025 if (__predict_false(error != 0))
1029 KASSERT(!force, ("%s: forced receive tag", __func__));
1030 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1031 if (__predict_false(error != 0))
1035 __assert_unreachable();
1038 tls->mode = TCP_TLS_MODE_IFNET;
1041 switch (tls->params.cipher_algorithm) {
1042 case CRYPTO_AES_CBC:
1043 counter_u64_add(ktls_ifnet_cbc, 1);
1045 case CRYPTO_AES_NIST_GCM_16:
1046 counter_u64_add(ktls_ifnet_gcm, 1);
1048 case CRYPTO_CHACHA20_POLY1305:
1049 counter_u64_add(ktls_ifnet_chacha20, 1);
1059 ktls_use_sw(struct ktls_session *tls)
1061 tls->mode = TCP_TLS_MODE_SW;
1062 switch (tls->params.cipher_algorithm) {
1063 case CRYPTO_AES_CBC:
1064 counter_u64_add(ktls_sw_cbc, 1);
1066 case CRYPTO_AES_NIST_GCM_16:
1067 counter_u64_add(ktls_sw_gcm, 1);
1069 case CRYPTO_CHACHA20_POLY1305:
1070 counter_u64_add(ktls_sw_chacha20, 1);
1076 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1080 error = ktls_ocf_try(so, tls, direction);
1088 * KTLS RX stores data in the socket buffer as a list of TLS records,
1089 * where each record is stored as a control message containg the TLS
1090 * header followed by data mbufs containing the decrypted data. This
1091 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1092 * both encrypted and decrypted data. TLS records decrypted by a NIC
1093 * should be queued to the socket buffer as records, but encrypted
1094 * data which needs to be decrypted by software arrives as a stream of
1095 * regular mbufs which need to be converted. In addition, there may
1096 * already be pending encrypted data in the socket buffer when KTLS RX
1099 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1102 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1104 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1105 * from the first mbuf. Once all of the data for that TLS record is
1106 * queued, the socket is queued to a worker thread.
1108 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1109 * the TLS chain. Each TLS record is detached from the TLS chain,
1110 * decrypted, and inserted into the regular socket buffer chain as
1111 * record starting with a control message holding the TLS header and
1112 * a chain of mbufs holding the encrypted data.
1116 sb_mark_notready(struct sockbuf *sb)
1123 sb->sb_mbtail = NULL;
1124 sb->sb_lastrecord = NULL;
1125 for (; m != NULL; m = m->m_next) {
1126 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1128 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1130 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1132 m->m_flags |= M_NOTREADY;
1133 sb->sb_acc -= m->m_len;
1134 sb->sb_tlscc += m->m_len;
1135 sb->sb_mtlstail = m;
1137 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1138 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1143 * Return information about the pending TLS data in a socket
1144 * buffer. On return, 'seqno' is set to the sequence number
1145 * of the next TLS record to be received, 'resid' is set to
1146 * the amount of bytes still needed for the last pending
1147 * record. The function returns 'false' if the last pending
1148 * record contains a partial TLS header. In that case, 'resid'
1149 * is the number of bytes needed to complete the TLS header.
1152 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1154 struct tls_record_layer hdr;
1158 u_int offset, record_len;
1160 SOCKBUF_LOCK_ASSERT(sb);
1161 MPASS(sb->sb_flags & SB_TLS_RX);
1162 seqno = sb->sb_tls_seqno;
1163 resid = sb->sb_tlscc;
1176 if (resid < sizeof(hdr)) {
1178 *residp = sizeof(hdr) - resid;
1182 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1184 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1185 if (resid <= record_len) {
1187 *residp = record_len - resid;
1190 resid -= record_len;
1192 while (record_len != 0) {
1193 if (m->m_len - offset > record_len) {
1194 offset += record_len;
1198 record_len -= (m->m_len - offset);
1206 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1208 struct ktls_session *tls;
1211 if (!ktls_offload_enable)
1213 if (SOLISTENING(so))
1216 counter_u64_add(ktls_offload_enable_calls, 1);
1219 * This should always be true since only the TCP socket option
1220 * invokes this function.
1222 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1226 * XXX: Don't overwrite existing sessions. We should permit
1227 * this to support rekeying in the future.
1229 if (so->so_rcv.sb_tls_info != NULL)
1232 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1235 error = ktls_create_session(so, en, &tls, KTLS_RX);
1239 error = ktls_ocf_try(so, tls, KTLS_RX);
1245 /* Mark the socket as using TLS offload. */
1246 SOCKBUF_LOCK(&so->so_rcv);
1247 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1248 so->so_rcv.sb_tls_info = tls;
1249 so->so_rcv.sb_flags |= SB_TLS_RX;
1251 /* Mark existing data as not ready until it can be decrypted. */
1252 sb_mark_notready(&so->so_rcv);
1253 ktls_check_rx(&so->so_rcv);
1254 SOCKBUF_UNLOCK(&so->so_rcv);
1256 /* Prefer TOE -> ifnet TLS -> software TLS. */
1258 error = ktls_try_toe(so, tls, KTLS_RX);
1261 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1265 counter_u64_add(ktls_offload_total, 1);
1271 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1273 struct ktls_session *tls;
1277 if (!ktls_offload_enable)
1279 if (SOLISTENING(so))
1282 counter_u64_add(ktls_offload_enable_calls, 1);
1285 * This should always be true since only the TCP socket option
1286 * invokes this function.
1288 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1292 * XXX: Don't overwrite existing sessions. We should permit
1293 * this to support rekeying in the future.
1295 if (so->so_snd.sb_tls_info != NULL)
1298 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1301 /* TLS requires ext pgs */
1302 if (mb_use_ext_pgs == 0)
1305 error = ktls_create_session(so, en, &tls, KTLS_TX);
1309 /* Prefer TOE -> ifnet TLS -> software TLS. */
1311 error = ktls_try_toe(so, tls, KTLS_TX);
1314 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1316 error = ktls_try_sw(so, tls, KTLS_TX);
1323 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1330 * Write lock the INP when setting sb_tls_info so that
1331 * routines in tcp_ratelimit.c can read sb_tls_info while
1332 * holding the INP lock.
1336 SOCKBUF_LOCK(&so->so_snd);
1337 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1338 so->so_snd.sb_tls_info = tls;
1339 if (tls->mode != TCP_TLS_MODE_SW)
1340 so->so_snd.sb_flags |= SB_TLS_IFNET;
1341 SOCKBUF_UNLOCK(&so->so_snd);
1343 SOCK_IO_SEND_UNLOCK(so);
1345 counter_u64_add(ktls_offload_total, 1);
1351 ktls_get_rx_mode(struct socket *so, int *modep)
1353 struct ktls_session *tls;
1354 struct inpcb *inp __diagused;
1356 if (SOLISTENING(so))
1359 INP_WLOCK_ASSERT(inp);
1360 SOCK_RECVBUF_LOCK(so);
1361 tls = so->so_rcv.sb_tls_info;
1363 *modep = TCP_TLS_MODE_NONE;
1366 SOCK_RECVBUF_UNLOCK(so);
1371 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1373 * This function gets information about the next TCP- and TLS-
1374 * sequence number to be processed by the TLS receive worker
1375 * thread. The information is extracted from the given "inpcb"
1376 * structure. The values are stored in host endian format at the two
1377 * given output pointer locations. The TCP sequence number points to
1378 * the beginning of the TLS header.
1380 * This function returns zero on success, else a non-zero error code
1384 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1390 so = inp->inp_socket;
1391 if (__predict_false(so == NULL)) {
1395 if (inp->inp_flags & INP_DROPPED) {
1397 return (ECONNRESET);
1400 tp = intotcpcb(inp);
1403 SOCKBUF_LOCK(&so->so_rcv);
1404 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1405 *tlsseq = so->so_rcv.sb_tls_seqno;
1406 SOCKBUF_UNLOCK(&so->so_rcv);
1414 ktls_get_tx_mode(struct socket *so, int *modep)
1416 struct ktls_session *tls;
1417 struct inpcb *inp __diagused;
1419 if (SOLISTENING(so))
1422 INP_WLOCK_ASSERT(inp);
1423 SOCK_SENDBUF_LOCK(so);
1424 tls = so->so_snd.sb_tls_info;
1426 *modep = TCP_TLS_MODE_NONE;
1429 SOCK_SENDBUF_UNLOCK(so);
1434 * Switch between SW and ifnet TLS sessions as requested.
1437 ktls_set_tx_mode(struct socket *so, int mode)
1439 struct ktls_session *tls, *tls_new;
1443 if (SOLISTENING(so))
1446 case TCP_TLS_MODE_SW:
1447 case TCP_TLS_MODE_IFNET:
1454 INP_WLOCK_ASSERT(inp);
1455 SOCKBUF_LOCK(&so->so_snd);
1456 tls = so->so_snd.sb_tls_info;
1458 SOCKBUF_UNLOCK(&so->so_snd);
1462 if (tls->mode == mode) {
1463 SOCKBUF_UNLOCK(&so->so_snd);
1467 tls = ktls_hold(tls);
1468 SOCKBUF_UNLOCK(&so->so_snd);
1471 tls_new = ktls_clone_session(tls, KTLS_TX);
1473 if (mode == TCP_TLS_MODE_IFNET)
1474 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1476 error = ktls_try_sw(so, tls_new, KTLS_TX);
1478 counter_u64_add(ktls_switch_failed, 1);
1485 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1487 counter_u64_add(ktls_switch_failed, 1);
1495 * If we raced with another session change, keep the existing
1498 if (tls != so->so_snd.sb_tls_info) {
1499 counter_u64_add(ktls_switch_failed, 1);
1500 SOCK_IO_SEND_UNLOCK(so);
1508 SOCKBUF_LOCK(&so->so_snd);
1509 so->so_snd.sb_tls_info = tls_new;
1510 if (tls_new->mode != TCP_TLS_MODE_SW)
1511 so->so_snd.sb_flags |= SB_TLS_IFNET;
1512 SOCKBUF_UNLOCK(&so->so_snd);
1513 SOCK_IO_SEND_UNLOCK(so);
1516 * Drop two references on 'tls'. The first is for the
1517 * ktls_hold() above. The second drops the reference from the
1520 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1524 if (mode == TCP_TLS_MODE_IFNET)
1525 counter_u64_add(ktls_switch_to_ifnet, 1);
1527 counter_u64_add(ktls_switch_to_sw, 1);
1533 * Try to allocate a new TLS receive tag. This task is scheduled when
1534 * sbappend_ktls_rx detects an input path change. If a new tag is
1535 * allocated, replace the tag in the TLS session. If a new tag cannot
1536 * be allocated, let the session fall back to software decryption.
1539 ktls_reset_receive_tag(void *context, int pending)
1541 union if_snd_tag_alloc_params params;
1542 struct ktls_session *tls;
1543 struct m_snd_tag *mst;
1549 MPASS(pending == 1);
1557 if (inp->inp_flags & INP_DROPPED) {
1562 SOCKBUF_LOCK(&so->so_rcv);
1564 tls->snd_tag = NULL;
1566 m_snd_tag_rele(mst);
1570 SOCKBUF_UNLOCK(&so->so_rcv);
1572 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1573 params.hdr.flowid = inp->inp_flowid;
1574 params.hdr.flowtype = inp->inp_flowtype;
1575 params.hdr.numa_domain = inp->inp_numa_domain;
1576 params.tls_rx.inp = inp;
1577 params.tls_rx.tls = tls;
1578 params.tls_rx.vlan_id = tls->rx_vlan_id;
1581 if (inp->inp_vflag & INP_IPV6) {
1582 if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0)
1585 if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0)
1589 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1591 SOCKBUF_LOCK(&so->so_rcv);
1593 SOCKBUF_UNLOCK(&so->so_rcv);
1595 counter_u64_add(ktls_ifnet_reset, 1);
1598 * Just fall back to software decryption if a tag
1599 * cannot be allocated leaving the connection intact.
1600 * If a future input path change switches to another
1601 * interface this connection will resume ifnet TLS.
1603 counter_u64_add(ktls_ifnet_reset_failed, 1);
1607 mtx_pool_lock(mtxpool_sleep, tls);
1608 tls->reset_pending = false;
1609 mtx_pool_unlock(mtxpool_sleep, tls);
1618 * Try to allocate a new TLS send tag. This task is scheduled when
1619 * ip_output detects a route change while trying to transmit a packet
1620 * holding a TLS record. If a new tag is allocated, replace the tag
1621 * in the TLS session. Subsequent packets on the connection will use
1622 * the new tag. If a new tag cannot be allocated, drop the
1626 ktls_reset_send_tag(void *context, int pending)
1628 struct epoch_tracker et;
1629 struct ktls_session *tls;
1630 struct m_snd_tag *old, *new;
1635 MPASS(pending == 1);
1641 * Free the old tag first before allocating a new one.
1642 * ip[6]_output_send() will treat a NULL send tag the same as
1643 * an ifp mismatch and drop packets until a new tag is
1646 * Write-lock the INP when changing tls->snd_tag since
1647 * ip[6]_output_send() holds a read-lock when reading the
1652 tls->snd_tag = NULL;
1655 m_snd_tag_rele(old);
1657 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1662 mtx_pool_lock(mtxpool_sleep, tls);
1663 tls->reset_pending = false;
1664 mtx_pool_unlock(mtxpool_sleep, tls);
1665 if (!in_pcbrele_wlocked(inp))
1668 counter_u64_add(ktls_ifnet_reset, 1);
1671 * XXX: Should we kick tcp_output explicitly now that
1672 * the send tag is fixed or just rely on timers?
1675 NET_EPOCH_ENTER(et);
1677 if (!in_pcbrele_wlocked(inp)) {
1678 if (!(inp->inp_flags & INP_DROPPED)) {
1679 tp = intotcpcb(inp);
1680 CURVNET_SET(inp->inp_vnet);
1681 tp = tcp_drop(tp, ECONNABORTED);
1685 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1691 counter_u64_add(ktls_ifnet_reset_failed, 1);
1694 * Leave reset_pending true to avoid future tasks while
1695 * the socket goes away.
1703 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1705 struct ktls_session *tls;
1708 SOCKBUF_LOCK_ASSERT(sb);
1709 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1711 so = __containerof(sb, struct socket, so_rcv);
1713 tls = sb->sb_tls_info;
1714 if_rele(tls->rx_ifp);
1719 * See if we should schedule a task to update the receive tag for
1722 mtx_pool_lock(mtxpool_sleep, tls);
1723 if (!tls->reset_pending) {
1724 (void) ktls_hold(tls);
1727 tls->reset_pending = true;
1728 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1730 mtx_pool_unlock(mtxpool_sleep, tls);
1734 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1740 INP_LOCK_ASSERT(inp);
1743 * See if we should schedule a task to update the send tag for
1746 mtx_pool_lock(mtxpool_sleep, tls);
1747 if (!tls->reset_pending) {
1748 (void) ktls_hold(tls);
1751 tls->reset_pending = true;
1752 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1754 mtx_pool_unlock(mtxpool_sleep, tls);
1760 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1762 union if_snd_tag_modify_params params = {
1763 .rate_limit.max_rate = max_pacing_rate,
1764 .rate_limit.flags = M_NOWAIT,
1766 struct m_snd_tag *mst;
1768 /* Can't get to the inp, but it should be locked. */
1769 /* INP_LOCK_ASSERT(inp); */
1771 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1773 if (tls->snd_tag == NULL) {
1775 * Resetting send tag, ignore this change. The
1776 * pending reset may or may not see this updated rate
1777 * in the tcpcb. If it doesn't, we will just lose
1786 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1788 return (mst->sw->snd_tag_modify(mst, ¶ms));
1794 ktls_destroy(struct ktls_session *tls)
1796 MPASS(tls->refcount == 0);
1798 if (tls->sequential_records) {
1802 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1803 page_count = m->m_epg_enc_cnt;
1804 while (page_count > 0) {
1805 KASSERT(page_count >= m->m_epg_nrdy,
1806 ("%s: too few pages", __func__));
1807 page_count -= m->m_epg_nrdy;
1813 counter_u64_add(ktls_offload_active, -1);
1814 switch (tls->mode) {
1815 case TCP_TLS_MODE_SW:
1816 switch (tls->params.cipher_algorithm) {
1817 case CRYPTO_AES_CBC:
1818 counter_u64_add(ktls_sw_cbc, -1);
1820 case CRYPTO_AES_NIST_GCM_16:
1821 counter_u64_add(ktls_sw_gcm, -1);
1823 case CRYPTO_CHACHA20_POLY1305:
1824 counter_u64_add(ktls_sw_chacha20, -1);
1828 case TCP_TLS_MODE_IFNET:
1829 switch (tls->params.cipher_algorithm) {
1830 case CRYPTO_AES_CBC:
1831 counter_u64_add(ktls_ifnet_cbc, -1);
1833 case CRYPTO_AES_NIST_GCM_16:
1834 counter_u64_add(ktls_ifnet_gcm, -1);
1836 case CRYPTO_CHACHA20_POLY1305:
1837 counter_u64_add(ktls_ifnet_chacha20, -1);
1840 if (tls->snd_tag != NULL)
1841 m_snd_tag_rele(tls->snd_tag);
1842 if (tls->rx_ifp != NULL)
1843 if_rele(tls->rx_ifp);
1846 case TCP_TLS_MODE_TOE:
1847 switch (tls->params.cipher_algorithm) {
1848 case CRYPTO_AES_CBC:
1849 counter_u64_add(ktls_toe_cbc, -1);
1851 case CRYPTO_AES_NIST_GCM_16:
1852 counter_u64_add(ktls_toe_gcm, -1);
1854 case CRYPTO_CHACHA20_POLY1305:
1855 counter_u64_add(ktls_toe_chacha20, -1);
1861 if (tls->ocf_session != NULL)
1863 if (tls->params.auth_key != NULL) {
1864 zfree(tls->params.auth_key, M_KTLS);
1865 tls->params.auth_key = NULL;
1866 tls->params.auth_key_len = 0;
1868 if (tls->params.cipher_key != NULL) {
1869 zfree(tls->params.cipher_key, M_KTLS);
1870 tls->params.cipher_key = NULL;
1871 tls->params.cipher_key_len = 0;
1873 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1875 uma_zfree(ktls_session_zone, tls);
1879 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1882 for (; m != NULL; m = m->m_next) {
1883 KASSERT((m->m_flags & M_EXTPG) != 0,
1884 ("ktls_seq: mapped mbuf %p", m));
1886 m->m_epg_seqno = sb->sb_tls_seqno;
1892 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1893 * mbuf in the chain must be an unmapped mbuf. The payload of the
1894 * mbuf must be populated with the payload of each TLS record.
1896 * The record_type argument specifies the TLS record type used when
1897 * populating the TLS header.
1899 * The enq_count argument on return is set to the number of pages of
1900 * payload data for this entire chain that need to be encrypted via SW
1901 * encryption. The returned value should be passed to ktls_enqueue
1902 * when scheduling encryption of this chain of mbufs. To handle the
1903 * special case of empty fragments for TLS 1.0 sessions, an empty
1904 * fragment counts as one page.
1907 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1908 uint8_t record_type)
1910 struct tls_record_layer *tlshdr;
1914 int maxlen __diagused;
1916 maxlen = tls->params.max_frame_len;
1918 for (m = top; m != NULL; m = m->m_next) {
1920 * All mbufs in the chain should be TLS records whose
1921 * payload does not exceed the maximum frame length.
1923 * Empty TLS 1.0 records are permitted when using CBC.
1925 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1926 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1927 ("ktls_frame: m %p len %d", m, m->m_len));
1930 * TLS frames require unmapped mbufs to store session
1933 KASSERT((m->m_flags & M_EXTPG) != 0,
1934 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1938 /* Save a reference to the session. */
1939 m->m_epg_tls = ktls_hold(tls);
1941 m->m_epg_hdrlen = tls->params.tls_hlen;
1942 m->m_epg_trllen = tls->params.tls_tlen;
1943 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1947 * AES-CBC pads messages to a multiple of the
1948 * block size. Note that the padding is
1949 * applied after the digest and the encryption
1950 * is done on the "plaintext || mac || padding".
1951 * At least one byte of padding is always
1954 * Compute the final trailer length assuming
1955 * at most one block of padding.
1956 * tls->params.tls_tlen is the maximum
1957 * possible trailer length (padding + digest).
1958 * delta holds the number of excess padding
1959 * bytes if the maximum were used. Those
1960 * extra bytes are removed.
1962 bs = tls->params.tls_bs;
1963 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1964 m->m_epg_trllen -= delta;
1966 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1968 /* Populate the TLS header. */
1969 tlshdr = (void *)m->m_epg_hdr;
1970 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1973 * TLS 1.3 masquarades as TLS 1.2 with a record type
1974 * of TLS_RLTYPE_APP.
1976 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1977 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1978 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1979 tlshdr->tls_type = TLS_RLTYPE_APP;
1980 /* save the real record type for later */
1981 m->m_epg_record_type = record_type;
1982 m->m_epg_trail[0] = record_type;
1984 tlshdr->tls_vminor = tls->params.tls_vminor;
1985 tlshdr->tls_type = record_type;
1987 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1990 * Store nonces / explicit IVs after the end of the
1993 * For GCM with TLS 1.2, an 8 byte nonce is copied
1994 * from the end of the IV. The nonce is then
1995 * incremented for use by the next record.
1997 * For CBC, a random nonce is inserted for TLS 1.1+.
1999 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2000 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2001 noncep = (uint64_t *)(tls->params.iv + 8);
2002 be64enc(tlshdr + 1, *noncep);
2004 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2005 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2006 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2009 * When using SW encryption, mark the mbuf not ready.
2010 * It will be marked ready via sbready() after the
2011 * record has been encrypted.
2013 * When using ifnet TLS, unencrypted TLS records are
2014 * sent down the stack to the NIC.
2016 if (tls->mode == TCP_TLS_MODE_SW) {
2017 m->m_flags |= M_NOTREADY;
2018 if (__predict_false(tls_len == 0)) {
2019 /* TLS 1.0 empty fragment. */
2022 m->m_epg_nrdy = m->m_epg_npgs;
2023 *enq_cnt += m->m_epg_nrdy;
2029 ktls_permit_empty_frames(struct ktls_session *tls)
2031 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2032 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2036 ktls_check_rx(struct sockbuf *sb)
2038 struct tls_record_layer hdr;
2043 SOCKBUF_LOCK_ASSERT(sb);
2044 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2046 so = __containerof(sb, struct socket, so_rcv);
2048 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2051 /* Is there enough queued for a TLS header? */
2052 if (sb->sb_tlscc < sizeof(hdr)) {
2053 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2054 so->so_error = EMSGSIZE;
2058 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2060 /* Is the entire record queued? */
2061 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2062 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2063 so->so_error = EMSGSIZE;
2067 sb->sb_flags |= SB_TLS_RX_RUNNING;
2070 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2072 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2073 running = wq->running;
2074 mtx_unlock(&wq->mtx);
2077 counter_u64_add(ktls_cnt_rx_queued, 1);
2080 static struct mbuf *
2081 ktls_detach_record(struct sockbuf *sb, int len)
2083 struct mbuf *m, *n, *top;
2086 SOCKBUF_LOCK_ASSERT(sb);
2087 MPASS(len <= sb->sb_tlscc);
2090 * If TLS chain is the exact size of the record,
2091 * just grab the whole record.
2094 if (sb->sb_tlscc == len) {
2096 sb->sb_mtlstail = NULL;
2101 * While it would be nice to use m_split() here, we need
2102 * to know exactly what m_split() allocates to update the
2103 * accounting, so do it inline instead.
2106 for (m = top; remain > m->m_len; m = m->m_next)
2109 /* Easy case: don't have to split 'm'. */
2110 if (remain == m->m_len) {
2111 sb->sb_mtls = m->m_next;
2112 if (sb->sb_mtls == NULL)
2113 sb->sb_mtlstail = NULL;
2119 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2120 * with M_NOWAIT first.
2122 n = m_get(M_NOWAIT, MT_DATA);
2125 * Use M_WAITOK with socket buffer unlocked. If
2126 * 'sb_mtls' changes while the lock is dropped, return
2127 * NULL to force the caller to retry.
2131 n = m_get(M_WAITOK, MT_DATA);
2134 if (sb->sb_mtls != top) {
2139 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2141 /* Store remainder in 'n'. */
2142 n->m_len = m->m_len - remain;
2143 if (m->m_flags & M_EXT) {
2144 n->m_data = m->m_data + remain;
2147 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2150 /* Trim 'm' and update accounting. */
2151 m->m_len -= n->m_len;
2152 sb->sb_tlscc -= n->m_len;
2153 sb->sb_ccc -= n->m_len;
2155 /* Account for 'n'. */
2156 sballoc_ktls_rx(sb, n);
2158 /* Insert 'n' into the TLS chain. */
2160 n->m_next = m->m_next;
2161 if (sb->sb_mtlstail == m)
2162 sb->sb_mtlstail = n;
2164 /* Detach the record from the TLS chain. */
2168 MPASS(m_length(top, NULL) == len);
2169 for (m = top; m != NULL; m = m->m_next)
2170 sbfree_ktls_rx(sb, m);
2171 sb->sb_tlsdcc = len;
2178 * Determine the length of the trailing zero padding and find the real
2179 * record type in the byte before the padding.
2181 * Walking the mbuf chain backwards is clumsy, so another option would
2182 * be to scan forwards remembering the last non-zero byte before the
2183 * trailer. However, it would be expensive to scan the entire record.
2184 * Instead, find the last non-zero byte of each mbuf in the chain
2185 * keeping track of the relative offset of that nonzero byte.
2187 * trail_len is the size of the MAC/tag on input and is set to the
2188 * size of the full trailer including padding and the record type on
2192 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2193 int *trailer_len, uint8_t *record_typep)
2196 u_int digest_start, last_offset, m_len, offset;
2197 uint8_t record_type;
2199 digest_start = tls_len - *trailer_len;
2202 for (; m != NULL && offset < digest_start;
2203 offset += m->m_len, m = m->m_next) {
2204 /* Don't look for padding in the tag. */
2205 m_len = min(digest_start - offset, m->m_len);
2206 cp = mtod(m, char *);
2208 /* Find last non-zero byte in this mbuf. */
2209 while (m_len > 0 && cp[m_len - 1] == 0)
2212 record_type = cp[m_len - 1];
2213 last_offset = offset + m_len;
2216 if (last_offset < tls->params.tls_hlen)
2219 *record_typep = record_type;
2220 *trailer_len = tls_len - last_offset + 1;
2225 * Check if a mbuf chain is fully decrypted at the given offset and
2226 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2227 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2228 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2231 ktls_mbuf_crypto_st_t
2232 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2234 int m_flags_ored = 0;
2235 int m_flags_anded = -1;
2237 for (; mb != NULL; mb = mb->m_next) {
2238 if (offset < mb->m_len)
2240 offset -= mb->m_len;
2244 for (; mb != NULL; mb = mb->m_next) {
2245 m_flags_ored |= mb->m_flags;
2246 m_flags_anded &= mb->m_flags;
2248 if (offset <= mb->m_len)
2250 offset -= mb->m_len;
2252 MPASS(mb != NULL || offset == 0);
2254 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2255 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2257 return ((m_flags_ored & M_DECRYPTED) ?
2258 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2259 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2263 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2266 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2268 union if_snd_tag_modify_params params;
2269 struct m_snd_tag *mst;
2273 mst = so->so_rcv.sb_tls_info->snd_tag;
2274 if (__predict_false(mst == NULL))
2277 inp = sotoinpcb(so);
2278 if (__predict_false(inp == NULL))
2282 if (inp->inp_flags & INP_DROPPED) {
2284 return (ECONNRESET);
2287 tp = intotcpcb(inp);
2290 /* Get the TCP sequence number of the next valid TLS header. */
2291 SOCKBUF_LOCK(&so->so_rcv);
2292 params.tls_rx.tls_hdr_tcp_sn =
2293 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2294 params.tls_rx.tls_rec_length = tls_len;
2295 params.tls_rx.tls_seq_number = tls_rcd_num;
2296 SOCKBUF_UNLOCK(&so->so_rcv);
2300 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2301 return (mst->sw->snd_tag_modify(mst, ¶ms));
2305 ktls_drop(struct socket *so, int error)
2307 struct epoch_tracker et;
2308 struct inpcb *inp = sotoinpcb(so);
2311 NET_EPOCH_ENTER(et);
2313 if (!(inp->inp_flags & INP_DROPPED)) {
2314 tp = intotcpcb(inp);
2315 CURVNET_SET(inp->inp_vnet);
2316 tp = tcp_drop(tp, error);
2321 so->so_error = error;
2322 SOCK_RECVBUF_LOCK(so);
2323 sorwakeup_locked(so);
2330 ktls_decrypt(struct socket *so)
2332 char tls_header[MBUF_PEXT_HDR_LEN];
2333 struct ktls_session *tls;
2335 struct tls_record_layer *hdr;
2336 struct tls_get_record tgr;
2337 struct mbuf *control, *data, *m;
2338 ktls_mbuf_crypto_st_t state;
2340 int error, remain, tls_len, trail_len;
2342 uint8_t vminor, record_type;
2344 hdr = (struct tls_record_layer *)tls_header;
2347 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2348 ("%s: socket %p not running", __func__, so));
2350 tls = sb->sb_tls_info;
2353 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2355 vminor = TLS_MINOR_VER_TWO;
2357 vminor = tls->params.tls_vminor;
2359 /* Is there enough queued for a TLS header? */
2360 if (sb->sb_tlscc < tls->params.tls_hlen)
2363 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2364 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2366 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2367 hdr->tls_vminor != vminor)
2369 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2371 else if (tls_len < tls->params.tls_hlen || tls_len >
2372 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2373 tls->params.tls_tlen)
2377 if (__predict_false(error != 0)) {
2379 * We have a corrupted record and are likely
2380 * out of sync. The connection isn't
2381 * recoverable at this point, so abort it.
2384 counter_u64_add(ktls_offload_corrupted_records, 1);
2386 ktls_drop(so, error);
2390 /* Is the entire record queued? */
2391 if (sb->sb_tlscc < tls_len)
2395 * Split out the portion of the mbuf chain containing
2398 data = ktls_detach_record(sb, tls_len);
2401 MPASS(sb->sb_tlsdcc == tls_len);
2403 seqno = sb->sb_tls_seqno;
2408 /* get crypto state for this TLS record */
2409 state = ktls_mbuf_crypto_state(data, 0, tls_len);
2412 case KTLS_MBUF_CRYPTO_ST_MIXED:
2413 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2417 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2418 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2420 if (__predict_true(error == 0)) {
2422 error = tls13_find_record_type(tls, data,
2423 tls_len, &trail_len, &record_type);
2425 record_type = hdr->tls_type;
2429 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2431 * NIC TLS is only supported for AEAD
2432 * ciphersuites which used a fixed sized
2436 trail_len = tls->params.tls_tlen - 1;
2437 error = tls13_find_record_type(tls, data,
2438 tls_len, &trail_len, &record_type);
2440 trail_len = tls->params.tls_tlen;
2442 record_type = hdr->tls_type;
2450 counter_u64_add(ktls_offload_failed_crypto, 1);
2453 if (sb->sb_tlsdcc == 0) {
2455 * sbcut/drop/flush discarded these
2463 * Drop this TLS record's data, but keep
2464 * decrypting subsequent records.
2466 sb->sb_ccc -= tls_len;
2469 if (error != EMSGSIZE)
2471 CURVNET_SET(so->so_vnet);
2472 so->so_error = error;
2473 sorwakeup_locked(so);
2482 /* Allocate the control mbuf. */
2483 memset(&tgr, 0, sizeof(tgr));
2484 tgr.tls_type = record_type;
2485 tgr.tls_vmajor = hdr->tls_vmajor;
2486 tgr.tls_vminor = hdr->tls_vminor;
2487 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2489 control = sbcreatecontrol(&tgr, sizeof(tgr),
2490 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2493 if (sb->sb_tlsdcc == 0) {
2494 /* sbcut/drop/flush discarded these mbufs. */
2495 MPASS(sb->sb_tlscc == 0);
2502 * Clear the 'dcc' accounting in preparation for
2503 * adding the decrypted record.
2505 sb->sb_ccc -= tls_len;
2509 /* If there is no payload, drop all of the data. */
2510 if (tgr.tls_length == htobe16(0)) {
2515 remain = tls->params.tls_hlen;
2516 while (remain > 0) {
2517 if (data->m_len > remain) {
2518 data->m_data += remain;
2519 data->m_len -= remain;
2522 remain -= data->m_len;
2523 data = m_free(data);
2526 /* Trim trailer and clear M_NOTREADY. */
2527 remain = be16toh(tgr.tls_length);
2529 for (m = data; remain > m->m_len; m = m->m_next) {
2530 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2536 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2538 /* Set EOR on the final mbuf. */
2539 m->m_flags |= M_EOR;
2542 sbappendcontrol_locked(sb, data, control, 0);
2544 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2545 sb->sb_flags |= SB_TLS_RX_RESYNC;
2547 ktls_resync_ifnet(so, tls_len, seqno);
2549 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2550 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2552 ktls_resync_ifnet(so, 0, seqno);
2557 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2559 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2560 so->so_error = EMSGSIZE;
2562 sorwakeup_locked(so);
2565 SOCKBUF_UNLOCK_ASSERT(sb);
2567 CURVNET_SET(so->so_vnet);
2573 ktls_enqueue_to_free(struct mbuf *m)
2578 /* Mark it for freeing. */
2579 m->m_epg_flags |= EPG_FLAG_2FREE;
2580 wq = &ktls_wq[m->m_epg_tls->wq_index];
2582 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2583 running = wq->running;
2584 mtx_unlock(&wq->mtx);
2590 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2593 int domain, running;
2595 if (m->m_epg_npgs <= 2)
2597 if (ktls_buffer_zone == NULL)
2599 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2601 * Rate-limit allocation attempts after a failure.
2602 * ktls_buffer_import() will acquire a per-domain mutex to check
2603 * the free page queues and may fail consistently if memory is
2608 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2610 domain = PCPU_GET(domain);
2611 wq->lastallocfail = ticks;
2614 * Note that this check is "racy", but the races are
2615 * harmless, and are either a spurious wakeup if
2616 * multiple threads fail allocations before the alloc
2617 * thread wakes, or waiting an extra second in case we
2618 * see an old value of running == true.
2620 if (!VM_DOMAIN_EMPTY(domain)) {
2621 running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2623 wakeup(&ktls_domains[domain].alloc_td);
2630 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2631 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2634 int error, i, len, off;
2636 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2637 ("%p not unready & nomap mbuf\n", m));
2638 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2639 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2642 /* Anonymous mbufs are encrypted in place. */
2643 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2644 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2647 * For file-backed mbufs (from sendfile), anonymous wired
2648 * pages are allocated and used as the encryption destination.
2650 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2651 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2653 state->dst_iov[0].iov_base = (char *)state->cbuf +
2655 state->dst_iov[0].iov_len = len;
2656 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2659 off = m->m_epg_1st_off;
2660 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2661 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2662 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2663 len = m_epg_pagelen(m, i, off);
2664 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2665 state->dst_iov[i].iov_base =
2666 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2667 state->dst_iov[i].iov_len = len;
2670 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2671 state->dst_iov[i].iov_base = m->m_epg_trail;
2672 state->dst_iov[i].iov_len = m->m_epg_trllen;
2674 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2676 if (__predict_false(error != 0)) {
2677 /* Free the anonymous pages. */
2678 if (state->cbuf != NULL)
2679 uma_zfree(ktls_buffer_zone, state->cbuf);
2681 for (i = 0; i < m->m_epg_npgs; i++) {
2682 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2683 (void)vm_page_unwire_noq(pg);
2691 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2693 ktls_batched_records(struct mbuf *m)
2695 int page_count, records;
2698 page_count = m->m_epg_enc_cnt;
2699 while (page_count > 0) {
2701 page_count -= m->m_epg_nrdy;
2704 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2709 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2711 struct ktls_session *tls;
2716 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2717 (M_EXTPG | M_NOTREADY)),
2718 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2719 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2721 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2723 m->m_epg_enc_cnt = page_count;
2726 * Save a pointer to the socket. The caller is responsible
2727 * for taking an additional reference via soref().
2733 wq = &ktls_wq[tls->wq_index];
2735 if (__predict_false(tls->sequential_records)) {
2737 * For TLS 1.0, records must be encrypted
2738 * sequentially. For a given connection, all records
2739 * queued to the associated work queue are processed
2740 * sequentially. However, sendfile(2) might complete
2741 * I/O requests spanning multiple TLS records out of
2742 * order. Here we ensure TLS records are enqueued to
2743 * the work queue in FIFO order.
2745 * tls->next_seqno holds the sequence number of the
2746 * next TLS record that should be enqueued to the work
2747 * queue. If this next record is not tls->next_seqno,
2748 * it must be a future record, so insert it, sorted by
2749 * TLS sequence number, into tls->pending_records and
2752 * If this TLS record matches tls->next_seqno, place
2753 * it in the work queue and then check
2754 * tls->pending_records to see if any
2755 * previously-queued records are now ready for
2758 if (m->m_epg_seqno != tls->next_seqno) {
2762 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2763 if (n->m_epg_seqno > m->m_epg_seqno)
2768 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2771 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2774 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2776 mtx_unlock(&wq->mtx);
2777 counter_u64_add(ktls_cnt_tx_pending, 1);
2781 tls->next_seqno += ktls_batched_records(m);
2782 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2784 while (!STAILQ_EMPTY(&tls->pending_records)) {
2787 n = STAILQ_FIRST(&tls->pending_records);
2788 if (n->m_epg_seqno != tls->next_seqno)
2792 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2793 tls->next_seqno += ktls_batched_records(n);
2794 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2796 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2798 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2800 running = wq->running;
2801 mtx_unlock(&wq->mtx);
2804 counter_u64_add(ktls_cnt_tx_queued, queued);
2808 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2809 * the pages from the file and replace them with the anonymous pages
2810 * allocated in ktls_encrypt_record().
2813 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2817 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2819 /* Free the old pages. */
2820 m->m_ext.ext_free(m);
2822 /* Replace them with the new pages. */
2823 if (state->cbuf != NULL) {
2824 for (i = 0; i < m->m_epg_npgs; i++)
2825 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2827 /* Contig pages should go back to the cache. */
2828 m->m_ext.ext_free = ktls_free_mext_contig;
2830 for (i = 0; i < m->m_epg_npgs; i++)
2831 m->m_epg_pa[i] = state->parray[i];
2833 /* Use the basic free routine. */
2834 m->m_ext.ext_free = mb_free_mext_pgs;
2837 /* Pages are now writable. */
2838 m->m_epg_flags |= EPG_FLAG_ANON;
2841 static __noinline void
2842 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2844 struct ktls_ocf_encrypt_state state;
2845 struct ktls_session *tls;
2848 int error, npages, total_pages;
2851 tls = top->m_epg_tls;
2852 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2853 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2855 top->m_epg_so = NULL;
2857 total_pages = top->m_epg_enc_cnt;
2861 * Encrypt the TLS records in the chain of mbufs starting with
2862 * 'top'. 'total_pages' gives us a total count of pages and is
2863 * used to know when we have finished encrypting the TLS
2864 * records originally queued with 'top'.
2866 * NB: These mbufs are queued in the socket buffer and
2867 * 'm_next' is traversing the mbufs in the socket buffer. The
2868 * socket buffer lock is not held while traversing this chain.
2869 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2870 * pointers should be stable. However, the 'm_next' of the
2871 * last mbuf encrypted is not necessarily NULL. It can point
2872 * to other mbufs appended while 'top' was on the TLS work
2875 * Each mbuf holds an entire TLS record.
2878 for (m = top; npages != total_pages; m = m->m_next) {
2879 KASSERT(m->m_epg_tls == tls,
2880 ("different TLS sessions in a single mbuf chain: %p vs %p",
2881 tls, m->m_epg_tls));
2882 KASSERT(npages + m->m_epg_npgs <= total_pages,
2883 ("page count mismatch: top %p, total_pages %d, m %p", top,
2886 error = ktls_encrypt_record(wq, m, tls, &state);
2888 counter_u64_add(ktls_offload_failed_crypto, 1);
2892 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2893 ktls_finish_nonanon(m, &state);
2895 npages += m->m_epg_nrdy;
2898 * Drop a reference to the session now that it is no
2899 * longer needed. Existing code depends on encrypted
2900 * records having no associated session vs
2901 * yet-to-be-encrypted records having an associated
2904 m->m_epg_tls = NULL;
2908 CURVNET_SET(so->so_vnet);
2910 (void)so->so_proto->pr_ready(so, top, npages);
2913 mb_free_notready(top, total_pages);
2921 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2923 struct ktls_session *tls;
2930 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2931 ktls_finish_nonanon(m, state);
2934 free(state, M_KTLS);
2937 * Drop a reference to the session now that it is no longer
2938 * needed. Existing code depends on encrypted records having
2939 * no associated session vs yet-to-be-encrypted records having
2940 * an associated session.
2943 m->m_epg_tls = NULL;
2947 counter_u64_add(ktls_offload_failed_crypto, 1);
2949 CURVNET_SET(so->so_vnet);
2950 npages = m->m_epg_nrdy;
2953 (void)so->so_proto->pr_ready(so, m, npages);
2956 mb_free_notready(m, npages);
2964 * Similar to ktls_encrypt, but used with asynchronous OCF backends
2965 * (coprocessors) where encryption does not use host CPU resources and
2966 * it can be beneficial to queue more requests than CPUs.
2968 static __noinline void
2969 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2971 struct ktls_ocf_encrypt_state *state;
2972 struct ktls_session *tls;
2975 int error, mpages, npages, total_pages;
2978 tls = top->m_epg_tls;
2979 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2980 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2982 top->m_epg_so = NULL;
2984 total_pages = top->m_epg_enc_cnt;
2988 for (m = top; npages != total_pages; m = n) {
2989 KASSERT(m->m_epg_tls == tls,
2990 ("different TLS sessions in a single mbuf chain: %p vs %p",
2991 tls, m->m_epg_tls));
2992 KASSERT(npages + m->m_epg_npgs <= total_pages,
2993 ("page count mismatch: top %p, total_pages %d, m %p", top,
2996 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3001 mpages = m->m_epg_nrdy;
3004 error = ktls_encrypt_record(wq, m, tls, state);
3006 counter_u64_add(ktls_offload_failed_crypto, 1);
3007 free(state, M_KTLS);
3008 CURVNET_SET(so->so_vnet);
3017 CURVNET_SET(so->so_vnet);
3020 mb_free_notready(m, total_pages - npages);
3028 ktls_bind_domain(int domain)
3032 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3035 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3040 ktls_alloc_thread(void *ctx)
3042 struct ktls_domain_info *ktls_domain = ctx;
3043 struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3045 struct sysctl_oid *oid;
3047 int domain, error, i, nbufs;
3049 domain = ktls_domain - ktls_domains;
3051 printf("Starting KTLS alloc thread for domain %d\n", domain);
3052 error = ktls_bind_domain(domain);
3054 printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3056 snprintf(name, sizeof(name), "domain%d", domain);
3057 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3058 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3059 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3060 CTLFLAG_RD, &sc->allocs, 0, "buffers allocated");
3061 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3062 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3063 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3064 CTLFLAG_RD, &sc->running, 0, "thread running");
3069 atomic_store_int(&sc->running, 0);
3070 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3071 atomic_store_int(&sc->running, 1);
3073 if (nbufs != ktls_max_alloc) {
3075 nbufs = atomic_load_int(&ktls_max_alloc);
3076 buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3080 * Below we allocate nbufs with different allocation
3081 * flags than we use when allocating normally during
3082 * encryption in the ktls worker thread. We specify
3083 * M_NORECLAIM in the worker thread. However, we omit
3084 * that flag here and add M_WAITOK so that the VM
3085 * system is permitted to perform expensive work to
3086 * defragment memory. We do this here, as it does not
3087 * matter if this thread blocks. If we block a ktls
3088 * worker thread, we risk developing backlogs of
3089 * buffers to be encrypted, leading to surges of
3090 * traffic and potential NIC output drops.
3092 for (i = 0; i < nbufs; i++) {
3093 buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3096 for (i = 0; i < nbufs; i++) {
3097 uma_zfree(ktls_buffer_zone, buf[i]);
3104 ktls_work_thread(void *ctx)
3106 struct ktls_wq *wq = ctx;
3108 struct socket *so, *son;
3109 STAILQ_HEAD(, mbuf) local_m_head;
3110 STAILQ_HEAD(, socket) local_so_head;
3115 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3118 * Bind to a core. If ktls_bind_threads is > 1, then
3119 * we bind to the NUMA domain instead.
3121 if (ktls_bind_threads) {
3124 if (ktls_bind_threads > 1) {
3125 struct pcpu *pc = pcpu_find(cpu);
3127 error = ktls_bind_domain(pc->pc_domain);
3131 CPU_SETOF(cpu, &mask);
3132 error = cpuset_setthread(curthread->td_tid, &mask);
3135 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3138 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3143 while (STAILQ_EMPTY(&wq->m_head) &&
3144 STAILQ_EMPTY(&wq->so_head)) {
3145 wq->running = false;
3146 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3150 STAILQ_INIT(&local_m_head);
3151 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3152 STAILQ_INIT(&local_so_head);
3153 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3154 mtx_unlock(&wq->mtx);
3156 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3157 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3158 ktls_free(m->m_epg_tls);
3161 if (m->m_epg_tls->sync_dispatch)
3162 ktls_encrypt(wq, m);
3164 ktls_encrypt_async(wq, m);
3165 counter_u64_add(ktls_cnt_tx_queued, -1);
3169 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3171 counter_u64_add(ktls_cnt_rx_queued, -1);
3176 #if defined(INET) || defined(INET6)
3178 ktls_disable_ifnet_help(void *context, int pending __unused)
3180 struct ktls_session *tls;
3191 so = inp->inp_socket;
3193 if (inp->inp_flags & INP_DROPPED) {
3197 if (so->so_snd.sb_tls_info != NULL)
3198 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3202 counter_u64_add(ktls_ifnet_disable_ok, 1);
3203 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3204 if ((inp->inp_flags & INP_DROPPED) == 0 &&
3205 (tp = intotcpcb(inp)) != NULL &&
3206 tp->t_fb->tfb_hwtls_change != NULL)
3207 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3209 counter_u64_add(ktls_ifnet_disable_fail, 1);
3213 CURVNET_SET(so->so_vnet);
3216 if (!in_pcbrele_wlocked(inp))
3222 * Called when re-transmits are becoming a substantial portion of the
3223 * sends on this connection. When this happens, we transition the
3224 * connection to software TLS. This is needed because most inline TLS
3225 * NICs keep crypto state only for in-order transmits. This means
3226 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3227 * re-DMA the entire TLS record up to and including the current
3228 * segment. This means that when re-transmitting the last ~1448 byte
3229 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3230 * of magnitude more data than we are sending. This can cause the
3231 * PCIe link to saturate well before the network, which can cause
3232 * output drops, and a general loss of capacity.
3235 ktls_disable_ifnet(void *arg)
3240 struct ktls_session *tls;
3243 inp = tptoinpcb(tp);
3244 INP_WLOCK_ASSERT(inp);
3245 so = inp->inp_socket;
3247 tls = so->so_snd.sb_tls_info;
3248 if (tls->disable_ifnet_pending) {
3254 * note that disable_ifnet_pending is never cleared; disabling
3255 * ifnet can only be done once per session, so we never want
3259 (void)ktls_hold(tls);
3262 tls->disable_ifnet_pending = true;
3265 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3266 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);