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 static void ktls_cleanup(struct ktls_session *tls);
301 #if defined(INET) || defined(INET6)
302 static void ktls_reset_receive_tag(void *context, int pending);
303 static void ktls_reset_send_tag(void *context, int pending);
305 static void ktls_work_thread(void *ctx);
306 static void ktls_alloc_thread(void *ctx);
308 #if defined(INET) || defined(INET6)
310 ktls_get_cpu(struct socket *so)
314 struct ktls_domain_info *di;
320 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
321 if (cpuid != NETISR_CPUID_NONE)
325 * Just use the flowid to shard connections in a repeatable
326 * fashion. Note that TLS 1.0 sessions rely on the
327 * serialization provided by having the same connection use
331 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
332 di = &ktls_domains[inp->inp_numa_domain];
333 cpuid = di->cpu[inp->inp_flowid % di->count];
336 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
342 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
347 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
348 ("%s: ktls max length %d is not page size-aligned",
349 __func__, ktls_maxlen));
351 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
352 for (i = 0; i < count; i++) {
353 m = vm_page_alloc_noobj_contig_domain(domain, req,
354 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
358 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
364 ktls_buffer_release(void *arg __unused, void **store, int count)
369 for (i = 0; i < count; i++) {
370 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
371 for (j = 0; j < atop(ktls_maxlen); j++) {
372 (void)vm_page_unwire_noq(m + j);
379 ktls_free_mext_contig(struct mbuf *m)
382 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
390 int count, domain, error, i;
392 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
395 ktls_session_zone = uma_zcreate("ktls_session",
396 sizeof(struct ktls_session),
397 NULL, NULL, NULL, NULL,
400 if (ktls_sw_buffer_cache) {
401 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
402 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
403 ktls_buffer_import, ktls_buffer_release, NULL,
404 UMA_ZONE_FIRSTTOUCH);
408 * Initialize the workqueues to run the TLS work. We create a
409 * work queue for each CPU.
412 STAILQ_INIT(&ktls_wq[i].m_head);
413 STAILQ_INIT(&ktls_wq[i].so_head);
414 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
415 if (ktls_bind_threads > 1) {
417 domain = pc->pc_domain;
418 count = ktls_domains[domain].count;
419 ktls_domains[domain].cpu[count] = i;
420 ktls_domains[domain].count++;
422 ktls_cpuid_lookup[ktls_number_threads] = i;
423 ktls_number_threads++;
427 * If we somehow have an empty domain, fall back to choosing
428 * among all KTLS threads.
430 if (ktls_bind_threads > 1) {
431 for (i = 0; i < vm_ndomains; i++) {
432 if (ktls_domains[i].count == 0) {
433 ktls_bind_threads = 1;
439 /* Start kthreads for each workqueue. */
441 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
442 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
444 printf("Can't add KTLS thread %d error %d\n", i, error);
450 * Start an allocation thread per-domain to perform blocking allocations
451 * of 16k physically contiguous TLS crypto destination buffers.
453 if (ktls_sw_buffer_cache) {
454 for (domain = 0; domain < vm_ndomains; domain++) {
455 if (VM_DOMAIN_EMPTY(domain))
457 if (CPU_EMPTY(&cpuset_domain[domain]))
459 error = kproc_kthread_add(ktls_alloc_thread,
460 &ktls_domains[domain], &ktls_proc,
461 &ktls_domains[domain].alloc_td.td,
462 0, 0, "KTLS", "alloc_%d", domain);
464 printf("Can't add KTLS alloc thread %d error %d\n",
472 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
477 ktls_start_kthreads(void)
482 state = atomic_load_acq_int(&ktls_init_state);
483 if (__predict_true(state > 0))
488 sx_xlock(&ktls_init_lock);
489 if (ktls_init_state != 0) {
490 sx_xunlock(&ktls_init_lock);
499 atomic_store_rel_int(&ktls_init_state, state);
500 sx_xunlock(&ktls_init_lock);
504 #if defined(INET) || defined(INET6)
506 ktls_create_session(struct socket *so, struct tls_enable *en,
507 struct ktls_session **tlsp, int direction)
509 struct ktls_session *tls;
512 /* Only TLS 1.0 - 1.3 are supported. */
513 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
515 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
516 en->tls_vminor > TLS_MINOR_VER_THREE)
519 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
521 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
523 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
526 /* All supported algorithms require a cipher key. */
527 if (en->cipher_key_len == 0)
530 /* No flags are currently supported. */
534 /* Common checks for supported algorithms. */
535 switch (en->cipher_algorithm) {
536 case CRYPTO_AES_NIST_GCM_16:
538 * auth_algorithm isn't used, but permit GMAC values
541 switch (en->auth_algorithm) {
543 #ifdef COMPAT_FREEBSD12
544 /* XXX: Really 13.0-current COMPAT. */
545 case CRYPTO_AES_128_NIST_GMAC:
546 case CRYPTO_AES_192_NIST_GMAC:
547 case CRYPTO_AES_256_NIST_GMAC:
553 if (en->auth_key_len != 0)
555 switch (en->tls_vminor) {
556 case TLS_MINOR_VER_TWO:
557 if (en->iv_len != TLS_AEAD_GCM_LEN)
560 case TLS_MINOR_VER_THREE:
561 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
569 switch (en->auth_algorithm) {
570 case CRYPTO_SHA1_HMAC:
572 case CRYPTO_SHA2_256_HMAC:
573 case CRYPTO_SHA2_384_HMAC:
574 if (en->tls_vminor != TLS_MINOR_VER_TWO)
580 if (en->auth_key_len == 0)
584 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
587 switch (en->tls_vminor) {
588 case TLS_MINOR_VER_ZERO:
589 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
592 case TLS_MINOR_VER_ONE:
593 case TLS_MINOR_VER_TWO:
594 /* Ignore any supplied IV. */
601 case CRYPTO_CHACHA20_POLY1305:
602 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
604 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
605 en->tls_vminor != TLS_MINOR_VER_THREE)
607 if (en->iv_len != TLS_CHACHA20_IV_LEN)
614 error = ktls_start_kthreads();
618 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
620 counter_u64_add(ktls_offload_active, 1);
622 refcount_init(&tls->refcount, 1);
623 if (direction == KTLS_RX)
624 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
626 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
628 tls->wq_index = ktls_get_cpu(so);
630 tls->params.cipher_algorithm = en->cipher_algorithm;
631 tls->params.auth_algorithm = en->auth_algorithm;
632 tls->params.tls_vmajor = en->tls_vmajor;
633 tls->params.tls_vminor = en->tls_vminor;
634 tls->params.flags = en->flags;
635 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
637 /* Set the header and trailer lengths. */
638 tls->params.tls_hlen = sizeof(struct tls_record_layer);
639 switch (en->cipher_algorithm) {
640 case CRYPTO_AES_NIST_GCM_16:
642 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
643 * nonce. TLS 1.3 uses a 12 byte implicit IV.
645 if (en->tls_vminor < TLS_MINOR_VER_THREE)
646 tls->params.tls_hlen += sizeof(uint64_t);
647 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
648 tls->params.tls_bs = 1;
651 switch (en->auth_algorithm) {
652 case CRYPTO_SHA1_HMAC:
653 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
654 /* Implicit IV, no nonce. */
655 tls->sequential_records = true;
656 tls->next_seqno = be64dec(en->rec_seq);
657 STAILQ_INIT(&tls->pending_records);
659 tls->params.tls_hlen += AES_BLOCK_LEN;
661 tls->params.tls_tlen = AES_BLOCK_LEN +
664 case CRYPTO_SHA2_256_HMAC:
665 tls->params.tls_hlen += AES_BLOCK_LEN;
666 tls->params.tls_tlen = AES_BLOCK_LEN +
669 case CRYPTO_SHA2_384_HMAC:
670 tls->params.tls_hlen += AES_BLOCK_LEN;
671 tls->params.tls_tlen = AES_BLOCK_LEN +
675 panic("invalid hmac");
677 tls->params.tls_bs = AES_BLOCK_LEN;
679 case CRYPTO_CHACHA20_POLY1305:
681 * Chacha20 uses a 12 byte implicit IV.
683 tls->params.tls_tlen = POLY1305_HASH_LEN;
684 tls->params.tls_bs = 1;
687 panic("invalid cipher");
691 * TLS 1.3 includes optional padding which we do not support,
692 * and also puts the "real" record type at the end of the
695 if (en->tls_vminor == TLS_MINOR_VER_THREE)
696 tls->params.tls_tlen += sizeof(uint8_t);
698 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
699 ("TLS header length too long: %d", tls->params.tls_hlen));
700 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
701 ("TLS trailer length too long: %d", tls->params.tls_tlen));
703 if (en->auth_key_len != 0) {
704 tls->params.auth_key_len = en->auth_key_len;
705 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
707 error = copyin(en->auth_key, tls->params.auth_key,
713 tls->params.cipher_key_len = en->cipher_key_len;
714 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
715 error = copyin(en->cipher_key, tls->params.cipher_key,
721 * This holds the implicit portion of the nonce for AEAD
722 * ciphers and the initial implicit IV for TLS 1.0. The
723 * explicit portions of the IV are generated in ktls_frame().
725 if (en->iv_len != 0) {
726 tls->params.iv_len = en->iv_len;
727 error = copyin(en->iv, tls->params.iv, en->iv_len);
732 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
733 * counter to generate unique explicit IVs.
735 * Store this counter in the last 8 bytes of the IV
736 * array so that it is 8-byte aligned.
738 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
739 en->tls_vminor == TLS_MINOR_VER_TWO)
740 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
751 static struct ktls_session *
752 ktls_clone_session(struct ktls_session *tls, int direction)
754 struct ktls_session *tls_new;
756 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
758 counter_u64_add(ktls_offload_active, 1);
760 refcount_init(&tls_new->refcount, 1);
761 if (direction == KTLS_RX)
762 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
765 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
768 /* Copy fields from existing session. */
769 tls_new->params = tls->params;
770 tls_new->wq_index = tls->wq_index;
772 /* Deep copy keys. */
773 if (tls_new->params.auth_key != NULL) {
774 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
776 memcpy(tls_new->params.auth_key, tls->params.auth_key,
777 tls->params.auth_key_len);
780 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
782 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
783 tls->params.cipher_key_len);
790 ktls_cleanup(struct ktls_session *tls)
793 counter_u64_add(ktls_offload_active, -1);
795 case TCP_TLS_MODE_SW:
796 switch (tls->params.cipher_algorithm) {
798 counter_u64_add(ktls_sw_cbc, -1);
800 case CRYPTO_AES_NIST_GCM_16:
801 counter_u64_add(ktls_sw_gcm, -1);
803 case CRYPTO_CHACHA20_POLY1305:
804 counter_u64_add(ktls_sw_chacha20, -1);
808 case TCP_TLS_MODE_IFNET:
809 switch (tls->params.cipher_algorithm) {
811 counter_u64_add(ktls_ifnet_cbc, -1);
813 case CRYPTO_AES_NIST_GCM_16:
814 counter_u64_add(ktls_ifnet_gcm, -1);
816 case CRYPTO_CHACHA20_POLY1305:
817 counter_u64_add(ktls_ifnet_chacha20, -1);
820 if (tls->snd_tag != NULL)
821 m_snd_tag_rele(tls->snd_tag);
822 if (tls->rx_ifp != NULL)
823 if_rele(tls->rx_ifp);
826 case TCP_TLS_MODE_TOE:
827 switch (tls->params.cipher_algorithm) {
829 counter_u64_add(ktls_toe_cbc, -1);
831 case CRYPTO_AES_NIST_GCM_16:
832 counter_u64_add(ktls_toe_gcm, -1);
834 case CRYPTO_CHACHA20_POLY1305:
835 counter_u64_add(ktls_toe_chacha20, -1);
841 if (tls->ocf_session != NULL)
843 if (tls->params.auth_key != NULL) {
844 zfree(tls->params.auth_key, M_KTLS);
845 tls->params.auth_key = NULL;
846 tls->params.auth_key_len = 0;
848 if (tls->params.cipher_key != NULL) {
849 zfree(tls->params.cipher_key, M_KTLS);
850 tls->params.cipher_key = NULL;
851 tls->params.cipher_key_len = 0;
853 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
856 #if defined(INET) || defined(INET6)
860 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
868 if (inp->inp_flags & INP_DROPPED) {
872 if (inp->inp_socket == NULL) {
877 if (!(tp->t_flags & TF_TOE)) {
882 error = tcp_offload_alloc_tls_session(tp, tls, direction);
885 tls->mode = TCP_TLS_MODE_TOE;
886 switch (tls->params.cipher_algorithm) {
888 counter_u64_add(ktls_toe_cbc, 1);
890 case CRYPTO_AES_NIST_GCM_16:
891 counter_u64_add(ktls_toe_gcm, 1);
893 case CRYPTO_CHACHA20_POLY1305:
894 counter_u64_add(ktls_toe_chacha20, 1);
903 * Common code used when first enabling ifnet TLS on a connection or
904 * when allocating a new ifnet TLS session due to a routing change.
905 * This function allocates a new TLS send tag on whatever interface
906 * the connection is currently routed over.
909 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
910 struct m_snd_tag **mstp)
912 union if_snd_tag_alloc_params params;
914 struct nhop_object *nh;
919 if (inp->inp_flags & INP_DROPPED) {
923 if (inp->inp_socket == NULL) {
930 * Check administrative controls on ifnet TLS to determine if
931 * ifnet TLS should be denied.
933 * - Always permit 'force' requests.
934 * - ktls_ifnet_permitted == 0: always deny.
936 if (!force && ktls_ifnet_permitted == 0) {
942 * XXX: Use the cached route in the inpcb to find the
943 * interface. This should perhaps instead use
944 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
945 * enabled after a connection has completed key negotiation in
946 * userland, the cached route will be present in practice.
948 nh = inp->inp_route.ro_nh;
957 * Allocate a TLS + ratelimit tag if the connection has an
958 * existing pacing rate.
960 if (tp->t_pacing_rate != -1 &&
961 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
962 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
963 params.tls_rate_limit.inp = inp;
964 params.tls_rate_limit.tls = tls;
965 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
967 params.hdr.type = IF_SND_TAG_TYPE_TLS;
968 params.tls.inp = inp;
969 params.tls.tls = tls;
971 params.hdr.flowid = inp->inp_flowid;
972 params.hdr.flowtype = inp->inp_flowtype;
973 params.hdr.numa_domain = inp->inp_numa_domain;
976 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
980 if (inp->inp_vflag & INP_IPV6) {
981 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
986 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
991 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
998 * Allocate an initial TLS receive tag for doing HW decryption of TLS
1001 * This function allocates a new TLS receive tag on whatever interface
1002 * the connection is currently routed over. If the connection ends up
1003 * using a different interface for receive this will get fixed up via
1004 * ktls_input_ifp_mismatch as future packets arrive.
1007 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1008 struct m_snd_tag **mstp)
1010 union if_snd_tag_alloc_params params;
1012 struct nhop_object *nh;
1015 if (!ktls_ocf_recrypt_supported(tls))
1019 if (inp->inp_flags & INP_DROPPED) {
1021 return (ECONNRESET);
1023 if (inp->inp_socket == NULL) {
1025 return (ECONNRESET);
1029 * Check administrative controls on ifnet TLS to determine if
1030 * ifnet TLS should be denied.
1032 if (ktls_ifnet_permitted == 0) {
1038 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1039 * the inpcb to find the interface.
1041 nh = inp->inp_route.ro_nh;
1050 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1051 params.hdr.flowid = inp->inp_flowid;
1052 params.hdr.flowtype = inp->inp_flowtype;
1053 params.hdr.numa_domain = inp->inp_numa_domain;
1054 params.tls_rx.inp = inp;
1055 params.tls_rx.tls = tls;
1056 params.tls_rx.vlan_id = 0;
1060 if (inp->inp_vflag & INP_IPV6) {
1061 if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) {
1066 if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) {
1071 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1074 * If this connection is over a vlan, vlan_snd_tag_alloc
1075 * rewrites vlan_id with the saved interface. Save the VLAN
1076 * ID for use in ktls_reset_receive_tag which allocates new
1077 * receive tags directly from the leaf interface bypassing
1081 tls->rx_vlan_id = params.tls_rx.vlan_id;
1087 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1090 struct m_snd_tag *mst;
1093 switch (direction) {
1095 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1096 if (__predict_false(error != 0))
1100 KASSERT(!force, ("%s: forced receive tag", __func__));
1101 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1102 if (__predict_false(error != 0))
1106 __assert_unreachable();
1109 tls->mode = TCP_TLS_MODE_IFNET;
1112 switch (tls->params.cipher_algorithm) {
1113 case CRYPTO_AES_CBC:
1114 counter_u64_add(ktls_ifnet_cbc, 1);
1116 case CRYPTO_AES_NIST_GCM_16:
1117 counter_u64_add(ktls_ifnet_gcm, 1);
1119 case CRYPTO_CHACHA20_POLY1305:
1120 counter_u64_add(ktls_ifnet_chacha20, 1);
1130 ktls_use_sw(struct ktls_session *tls)
1132 tls->mode = TCP_TLS_MODE_SW;
1133 switch (tls->params.cipher_algorithm) {
1134 case CRYPTO_AES_CBC:
1135 counter_u64_add(ktls_sw_cbc, 1);
1137 case CRYPTO_AES_NIST_GCM_16:
1138 counter_u64_add(ktls_sw_gcm, 1);
1140 case CRYPTO_CHACHA20_POLY1305:
1141 counter_u64_add(ktls_sw_chacha20, 1);
1147 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1151 error = ktls_ocf_try(so, tls, direction);
1159 * KTLS RX stores data in the socket buffer as a list of TLS records,
1160 * where each record is stored as a control message containg the TLS
1161 * header followed by data mbufs containing the decrypted data. This
1162 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1163 * both encrypted and decrypted data. TLS records decrypted by a NIC
1164 * should be queued to the socket buffer as records, but encrypted
1165 * data which needs to be decrypted by software arrives as a stream of
1166 * regular mbufs which need to be converted. In addition, there may
1167 * already be pending encrypted data in the socket buffer when KTLS RX
1170 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1173 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1175 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1176 * from the first mbuf. Once all of the data for that TLS record is
1177 * queued, the socket is queued to a worker thread.
1179 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1180 * the TLS chain. Each TLS record is detached from the TLS chain,
1181 * decrypted, and inserted into the regular socket buffer chain as
1182 * record starting with a control message holding the TLS header and
1183 * a chain of mbufs holding the encrypted data.
1187 sb_mark_notready(struct sockbuf *sb)
1194 sb->sb_mbtail = NULL;
1195 sb->sb_lastrecord = NULL;
1196 for (; m != NULL; m = m->m_next) {
1197 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1199 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1201 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1203 m->m_flags |= M_NOTREADY;
1204 sb->sb_acc -= m->m_len;
1205 sb->sb_tlscc += m->m_len;
1206 sb->sb_mtlstail = m;
1208 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1209 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1214 * Return information about the pending TLS data in a socket
1215 * buffer. On return, 'seqno' is set to the sequence number
1216 * of the next TLS record to be received, 'resid' is set to
1217 * the amount of bytes still needed for the last pending
1218 * record. The function returns 'false' if the last pending
1219 * record contains a partial TLS header. In that case, 'resid'
1220 * is the number of bytes needed to complete the TLS header.
1223 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1225 struct tls_record_layer hdr;
1229 u_int offset, record_len;
1231 SOCKBUF_LOCK_ASSERT(sb);
1232 MPASS(sb->sb_flags & SB_TLS_RX);
1233 seqno = sb->sb_tls_seqno;
1234 resid = sb->sb_tlscc;
1247 if (resid < sizeof(hdr)) {
1249 *residp = sizeof(hdr) - resid;
1253 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1255 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1256 if (resid <= record_len) {
1258 *residp = record_len - resid;
1261 resid -= record_len;
1263 while (record_len != 0) {
1264 if (m->m_len - offset > record_len) {
1265 offset += record_len;
1269 record_len -= (m->m_len - offset);
1277 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1279 struct ktls_session *tls;
1282 if (!ktls_offload_enable)
1284 if (SOLISTENING(so))
1287 counter_u64_add(ktls_offload_enable_calls, 1);
1290 * This should always be true since only the TCP socket option
1291 * invokes this function.
1293 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1297 * XXX: Don't overwrite existing sessions. We should permit
1298 * this to support rekeying in the future.
1300 if (so->so_rcv.sb_tls_info != NULL)
1303 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1306 error = ktls_create_session(so, en, &tls, KTLS_RX);
1310 error = ktls_ocf_try(so, tls, KTLS_RX);
1316 /* Mark the socket as using TLS offload. */
1317 SOCKBUF_LOCK(&so->so_rcv);
1318 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1319 so->so_rcv.sb_tls_info = tls;
1320 so->so_rcv.sb_flags |= SB_TLS_RX;
1322 /* Mark existing data as not ready until it can be decrypted. */
1323 sb_mark_notready(&so->so_rcv);
1324 ktls_check_rx(&so->so_rcv);
1325 SOCKBUF_UNLOCK(&so->so_rcv);
1327 /* Prefer TOE -> ifnet TLS -> software TLS. */
1329 error = ktls_try_toe(so, tls, KTLS_RX);
1332 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1336 counter_u64_add(ktls_offload_total, 1);
1342 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1344 struct ktls_session *tls;
1348 if (!ktls_offload_enable)
1350 if (SOLISTENING(so))
1353 counter_u64_add(ktls_offload_enable_calls, 1);
1356 * This should always be true since only the TCP socket option
1357 * invokes this function.
1359 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1363 * XXX: Don't overwrite existing sessions. We should permit
1364 * this to support rekeying in the future.
1366 if (so->so_snd.sb_tls_info != NULL)
1369 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1372 /* TLS requires ext pgs */
1373 if (mb_use_ext_pgs == 0)
1376 error = ktls_create_session(so, en, &tls, KTLS_TX);
1380 /* Prefer TOE -> ifnet TLS -> software TLS. */
1382 error = ktls_try_toe(so, tls, KTLS_TX);
1385 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1387 error = ktls_try_sw(so, tls, KTLS_TX);
1394 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1401 * Write lock the INP when setting sb_tls_info so that
1402 * routines in tcp_ratelimit.c can read sb_tls_info while
1403 * holding the INP lock.
1407 SOCKBUF_LOCK(&so->so_snd);
1408 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1409 so->so_snd.sb_tls_info = tls;
1410 if (tls->mode != TCP_TLS_MODE_SW)
1411 so->so_snd.sb_flags |= SB_TLS_IFNET;
1412 SOCKBUF_UNLOCK(&so->so_snd);
1414 SOCK_IO_SEND_UNLOCK(so);
1416 counter_u64_add(ktls_offload_total, 1);
1422 ktls_get_rx_mode(struct socket *so, int *modep)
1424 struct ktls_session *tls;
1425 struct inpcb *inp __diagused;
1427 if (SOLISTENING(so))
1430 INP_WLOCK_ASSERT(inp);
1431 SOCK_RECVBUF_LOCK(so);
1432 tls = so->so_rcv.sb_tls_info;
1434 *modep = TCP_TLS_MODE_NONE;
1437 SOCK_RECVBUF_UNLOCK(so);
1442 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1444 * This function gets information about the next TCP- and TLS-
1445 * sequence number to be processed by the TLS receive worker
1446 * thread. The information is extracted from the given "inpcb"
1447 * structure. The values are stored in host endian format at the two
1448 * given output pointer locations. The TCP sequence number points to
1449 * the beginning of the TLS header.
1451 * This function returns zero on success, else a non-zero error code
1455 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1461 so = inp->inp_socket;
1462 if (__predict_false(so == NULL)) {
1466 if (inp->inp_flags & INP_DROPPED) {
1468 return (ECONNRESET);
1471 tp = intotcpcb(inp);
1474 SOCKBUF_LOCK(&so->so_rcv);
1475 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1476 *tlsseq = so->so_rcv.sb_tls_seqno;
1477 SOCKBUF_UNLOCK(&so->so_rcv);
1485 ktls_get_tx_mode(struct socket *so, int *modep)
1487 struct ktls_session *tls;
1488 struct inpcb *inp __diagused;
1490 if (SOLISTENING(so))
1493 INP_WLOCK_ASSERT(inp);
1494 SOCK_SENDBUF_LOCK(so);
1495 tls = so->so_snd.sb_tls_info;
1497 *modep = TCP_TLS_MODE_NONE;
1500 SOCK_SENDBUF_UNLOCK(so);
1505 * Switch between SW and ifnet TLS sessions as requested.
1508 ktls_set_tx_mode(struct socket *so, int mode)
1510 struct ktls_session *tls, *tls_new;
1514 if (SOLISTENING(so))
1517 case TCP_TLS_MODE_SW:
1518 case TCP_TLS_MODE_IFNET:
1525 INP_WLOCK_ASSERT(inp);
1526 SOCKBUF_LOCK(&so->so_snd);
1527 tls = so->so_snd.sb_tls_info;
1529 SOCKBUF_UNLOCK(&so->so_snd);
1533 if (tls->mode == mode) {
1534 SOCKBUF_UNLOCK(&so->so_snd);
1538 tls = ktls_hold(tls);
1539 SOCKBUF_UNLOCK(&so->so_snd);
1542 tls_new = ktls_clone_session(tls, KTLS_TX);
1544 if (mode == TCP_TLS_MODE_IFNET)
1545 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1547 error = ktls_try_sw(so, tls_new, KTLS_TX);
1549 counter_u64_add(ktls_switch_failed, 1);
1556 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1558 counter_u64_add(ktls_switch_failed, 1);
1566 * If we raced with another session change, keep the existing
1569 if (tls != so->so_snd.sb_tls_info) {
1570 counter_u64_add(ktls_switch_failed, 1);
1571 SOCK_IO_SEND_UNLOCK(so);
1579 SOCKBUF_LOCK(&so->so_snd);
1580 so->so_snd.sb_tls_info = tls_new;
1581 if (tls_new->mode != TCP_TLS_MODE_SW)
1582 so->so_snd.sb_flags |= SB_TLS_IFNET;
1583 SOCKBUF_UNLOCK(&so->so_snd);
1584 SOCK_IO_SEND_UNLOCK(so);
1587 * Drop two references on 'tls'. The first is for the
1588 * ktls_hold() above. The second drops the reference from the
1591 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1595 if (mode == TCP_TLS_MODE_IFNET)
1596 counter_u64_add(ktls_switch_to_ifnet, 1);
1598 counter_u64_add(ktls_switch_to_sw, 1);
1604 * Try to allocate a new TLS receive tag. This task is scheduled when
1605 * sbappend_ktls_rx detects an input path change. If a new tag is
1606 * allocated, replace the tag in the TLS session. If a new tag cannot
1607 * be allocated, let the session fall back to software decryption.
1610 ktls_reset_receive_tag(void *context, int pending)
1612 union if_snd_tag_alloc_params params;
1613 struct ktls_session *tls;
1614 struct m_snd_tag *mst;
1620 MPASS(pending == 1);
1628 if (inp->inp_flags & INP_DROPPED) {
1633 SOCKBUF_LOCK(&so->so_rcv);
1635 tls->snd_tag = NULL;
1637 m_snd_tag_rele(mst);
1641 SOCKBUF_UNLOCK(&so->so_rcv);
1643 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1644 params.hdr.flowid = inp->inp_flowid;
1645 params.hdr.flowtype = inp->inp_flowtype;
1646 params.hdr.numa_domain = inp->inp_numa_domain;
1647 params.tls_rx.inp = inp;
1648 params.tls_rx.tls = tls;
1649 params.tls_rx.vlan_id = tls->rx_vlan_id;
1652 if (inp->inp_vflag & INP_IPV6) {
1653 if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0)
1656 if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0)
1660 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1662 SOCKBUF_LOCK(&so->so_rcv);
1664 SOCKBUF_UNLOCK(&so->so_rcv);
1666 counter_u64_add(ktls_ifnet_reset, 1);
1669 * Just fall back to software decryption if a tag
1670 * cannot be allocated leaving the connection intact.
1671 * If a future input path change switches to another
1672 * interface this connection will resume ifnet TLS.
1674 counter_u64_add(ktls_ifnet_reset_failed, 1);
1678 mtx_pool_lock(mtxpool_sleep, tls);
1679 tls->reset_pending = false;
1680 mtx_pool_unlock(mtxpool_sleep, tls);
1689 * Try to allocate a new TLS send tag. This task is scheduled when
1690 * ip_output detects a route change while trying to transmit a packet
1691 * holding a TLS record. If a new tag is allocated, replace the tag
1692 * in the TLS session. Subsequent packets on the connection will use
1693 * the new tag. If a new tag cannot be allocated, drop the
1697 ktls_reset_send_tag(void *context, int pending)
1699 struct epoch_tracker et;
1700 struct ktls_session *tls;
1701 struct m_snd_tag *old, *new;
1706 MPASS(pending == 1);
1712 * Free the old tag first before allocating a new one.
1713 * ip[6]_output_send() will treat a NULL send tag the same as
1714 * an ifp mismatch and drop packets until a new tag is
1717 * Write-lock the INP when changing tls->snd_tag since
1718 * ip[6]_output_send() holds a read-lock when reading the
1723 tls->snd_tag = NULL;
1726 m_snd_tag_rele(old);
1728 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1733 mtx_pool_lock(mtxpool_sleep, tls);
1734 tls->reset_pending = false;
1735 mtx_pool_unlock(mtxpool_sleep, tls);
1736 if (!in_pcbrele_wlocked(inp))
1739 counter_u64_add(ktls_ifnet_reset, 1);
1742 * XXX: Should we kick tcp_output explicitly now that
1743 * the send tag is fixed or just rely on timers?
1746 NET_EPOCH_ENTER(et);
1748 if (!in_pcbrele_wlocked(inp)) {
1749 if (!(inp->inp_flags & INP_DROPPED)) {
1750 tp = intotcpcb(inp);
1751 CURVNET_SET(tp->t_vnet);
1752 tp = tcp_drop(tp, ECONNABORTED);
1756 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1762 counter_u64_add(ktls_ifnet_reset_failed, 1);
1765 * Leave reset_pending true to avoid future tasks while
1766 * the socket goes away.
1774 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1776 struct ktls_session *tls;
1779 SOCKBUF_LOCK_ASSERT(sb);
1780 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1782 so = __containerof(sb, struct socket, so_rcv);
1784 tls = sb->sb_tls_info;
1785 if_rele(tls->rx_ifp);
1790 * See if we should schedule a task to update the receive tag for
1793 mtx_pool_lock(mtxpool_sleep, tls);
1794 if (!tls->reset_pending) {
1795 (void) ktls_hold(tls);
1798 tls->reset_pending = true;
1799 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1801 mtx_pool_unlock(mtxpool_sleep, tls);
1805 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1811 INP_LOCK_ASSERT(inp);
1814 * See if we should schedule a task to update the send tag for
1817 mtx_pool_lock(mtxpool_sleep, tls);
1818 if (!tls->reset_pending) {
1819 (void) ktls_hold(tls);
1822 tls->reset_pending = true;
1823 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1825 mtx_pool_unlock(mtxpool_sleep, tls);
1831 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1833 union if_snd_tag_modify_params params = {
1834 .rate_limit.max_rate = max_pacing_rate,
1835 .rate_limit.flags = M_NOWAIT,
1837 struct m_snd_tag *mst;
1839 /* Can't get to the inp, but it should be locked. */
1840 /* INP_LOCK_ASSERT(inp); */
1842 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1844 if (tls->snd_tag == NULL) {
1846 * Resetting send tag, ignore this change. The
1847 * pending reset may or may not see this updated rate
1848 * in the tcpcb. If it doesn't, we will just lose
1857 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1859 return (mst->sw->snd_tag_modify(mst, ¶ms));
1865 ktls_destroy(struct ktls_session *tls)
1868 if (tls->sequential_records) {
1872 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1873 page_count = m->m_epg_enc_cnt;
1874 while (page_count > 0) {
1875 KASSERT(page_count >= m->m_epg_nrdy,
1876 ("%s: too few pages", __func__));
1877 page_count -= m->m_epg_nrdy;
1883 uma_zfree(ktls_session_zone, tls);
1887 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1890 for (; m != NULL; m = m->m_next) {
1891 KASSERT((m->m_flags & M_EXTPG) != 0,
1892 ("ktls_seq: mapped mbuf %p", m));
1894 m->m_epg_seqno = sb->sb_tls_seqno;
1900 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1901 * mbuf in the chain must be an unmapped mbuf. The payload of the
1902 * mbuf must be populated with the payload of each TLS record.
1904 * The record_type argument specifies the TLS record type used when
1905 * populating the TLS header.
1907 * The enq_count argument on return is set to the number of pages of
1908 * payload data for this entire chain that need to be encrypted via SW
1909 * encryption. The returned value should be passed to ktls_enqueue
1910 * when scheduling encryption of this chain of mbufs. To handle the
1911 * special case of empty fragments for TLS 1.0 sessions, an empty
1912 * fragment counts as one page.
1915 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1916 uint8_t record_type)
1918 struct tls_record_layer *tlshdr;
1922 int maxlen __diagused;
1924 maxlen = tls->params.max_frame_len;
1926 for (m = top; m != NULL; m = m->m_next) {
1928 * All mbufs in the chain should be TLS records whose
1929 * payload does not exceed the maximum frame length.
1931 * Empty TLS 1.0 records are permitted when using CBC.
1933 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1934 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1935 ("ktls_frame: m %p len %d", m, m->m_len));
1938 * TLS frames require unmapped mbufs to store session
1941 KASSERT((m->m_flags & M_EXTPG) != 0,
1942 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1946 /* Save a reference to the session. */
1947 m->m_epg_tls = ktls_hold(tls);
1949 m->m_epg_hdrlen = tls->params.tls_hlen;
1950 m->m_epg_trllen = tls->params.tls_tlen;
1951 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1955 * AES-CBC pads messages to a multiple of the
1956 * block size. Note that the padding is
1957 * applied after the digest and the encryption
1958 * is done on the "plaintext || mac || padding".
1959 * At least one byte of padding is always
1962 * Compute the final trailer length assuming
1963 * at most one block of padding.
1964 * tls->params.tls_tlen is the maximum
1965 * possible trailer length (padding + digest).
1966 * delta holds the number of excess padding
1967 * bytes if the maximum were used. Those
1968 * extra bytes are removed.
1970 bs = tls->params.tls_bs;
1971 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1972 m->m_epg_trllen -= delta;
1974 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1976 /* Populate the TLS header. */
1977 tlshdr = (void *)m->m_epg_hdr;
1978 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1981 * TLS 1.3 masquarades as TLS 1.2 with a record type
1982 * of TLS_RLTYPE_APP.
1984 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1985 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1986 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1987 tlshdr->tls_type = TLS_RLTYPE_APP;
1988 /* save the real record type for later */
1989 m->m_epg_record_type = record_type;
1990 m->m_epg_trail[0] = record_type;
1992 tlshdr->tls_vminor = tls->params.tls_vminor;
1993 tlshdr->tls_type = record_type;
1995 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1998 * Store nonces / explicit IVs after the end of the
2001 * For GCM with TLS 1.2, an 8 byte nonce is copied
2002 * from the end of the IV. The nonce is then
2003 * incremented for use by the next record.
2005 * For CBC, a random nonce is inserted for TLS 1.1+.
2007 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2008 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2009 noncep = (uint64_t *)(tls->params.iv + 8);
2010 be64enc(tlshdr + 1, *noncep);
2012 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2013 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2014 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2017 * When using SW encryption, mark the mbuf not ready.
2018 * It will be marked ready via sbready() after the
2019 * record has been encrypted.
2021 * When using ifnet TLS, unencrypted TLS records are
2022 * sent down the stack to the NIC.
2024 if (tls->mode == TCP_TLS_MODE_SW) {
2025 m->m_flags |= M_NOTREADY;
2026 if (__predict_false(tls_len == 0)) {
2027 /* TLS 1.0 empty fragment. */
2030 m->m_epg_nrdy = m->m_epg_npgs;
2031 *enq_cnt += m->m_epg_nrdy;
2037 ktls_permit_empty_frames(struct ktls_session *tls)
2039 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2040 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2044 ktls_check_rx(struct sockbuf *sb)
2046 struct tls_record_layer hdr;
2051 SOCKBUF_LOCK_ASSERT(sb);
2052 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2054 so = __containerof(sb, struct socket, so_rcv);
2056 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2059 /* Is there enough queued for a TLS header? */
2060 if (sb->sb_tlscc < sizeof(hdr)) {
2061 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2062 so->so_error = EMSGSIZE;
2066 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2068 /* Is the entire record queued? */
2069 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2070 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2071 so->so_error = EMSGSIZE;
2075 sb->sb_flags |= SB_TLS_RX_RUNNING;
2078 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2080 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2081 running = wq->running;
2082 mtx_unlock(&wq->mtx);
2085 counter_u64_add(ktls_cnt_rx_queued, 1);
2088 static struct mbuf *
2089 ktls_detach_record(struct sockbuf *sb, int len)
2091 struct mbuf *m, *n, *top;
2094 SOCKBUF_LOCK_ASSERT(sb);
2095 MPASS(len <= sb->sb_tlscc);
2098 * If TLS chain is the exact size of the record,
2099 * just grab the whole record.
2102 if (sb->sb_tlscc == len) {
2104 sb->sb_mtlstail = NULL;
2109 * While it would be nice to use m_split() here, we need
2110 * to know exactly what m_split() allocates to update the
2111 * accounting, so do it inline instead.
2114 for (m = top; remain > m->m_len; m = m->m_next)
2117 /* Easy case: don't have to split 'm'. */
2118 if (remain == m->m_len) {
2119 sb->sb_mtls = m->m_next;
2120 if (sb->sb_mtls == NULL)
2121 sb->sb_mtlstail = NULL;
2127 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2128 * with M_NOWAIT first.
2130 n = m_get(M_NOWAIT, MT_DATA);
2133 * Use M_WAITOK with socket buffer unlocked. If
2134 * 'sb_mtls' changes while the lock is dropped, return
2135 * NULL to force the caller to retry.
2139 n = m_get(M_WAITOK, MT_DATA);
2142 if (sb->sb_mtls != top) {
2147 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2149 /* Store remainder in 'n'. */
2150 n->m_len = m->m_len - remain;
2151 if (m->m_flags & M_EXT) {
2152 n->m_data = m->m_data + remain;
2155 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2158 /* Trim 'm' and update accounting. */
2159 m->m_len -= n->m_len;
2160 sb->sb_tlscc -= n->m_len;
2161 sb->sb_ccc -= n->m_len;
2163 /* Account for 'n'. */
2164 sballoc_ktls_rx(sb, n);
2166 /* Insert 'n' into the TLS chain. */
2168 n->m_next = m->m_next;
2169 if (sb->sb_mtlstail == m)
2170 sb->sb_mtlstail = n;
2172 /* Detach the record from the TLS chain. */
2176 MPASS(m_length(top, NULL) == len);
2177 for (m = top; m != NULL; m = m->m_next)
2178 sbfree_ktls_rx(sb, m);
2179 sb->sb_tlsdcc = len;
2186 * Determine the length of the trailing zero padding and find the real
2187 * record type in the byte before the padding.
2189 * Walking the mbuf chain backwards is clumsy, so another option would
2190 * be to scan forwards remembering the last non-zero byte before the
2191 * trailer. However, it would be expensive to scan the entire record.
2192 * Instead, find the last non-zero byte of each mbuf in the chain
2193 * keeping track of the relative offset of that nonzero byte.
2195 * trail_len is the size of the MAC/tag on input and is set to the
2196 * size of the full trailer including padding and the record type on
2200 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2201 int *trailer_len, uint8_t *record_typep)
2204 u_int digest_start, last_offset, m_len, offset;
2205 uint8_t record_type;
2207 digest_start = tls_len - *trailer_len;
2210 for (; m != NULL && offset < digest_start;
2211 offset += m->m_len, m = m->m_next) {
2212 /* Don't look for padding in the tag. */
2213 m_len = min(digest_start - offset, m->m_len);
2214 cp = mtod(m, char *);
2216 /* Find last non-zero byte in this mbuf. */
2217 while (m_len > 0 && cp[m_len - 1] == 0)
2220 record_type = cp[m_len - 1];
2221 last_offset = offset + m_len;
2224 if (last_offset < tls->params.tls_hlen)
2227 *record_typep = record_type;
2228 *trailer_len = tls_len - last_offset + 1;
2233 * Check if a mbuf chain is fully decrypted at the given offset and
2234 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2235 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2236 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2239 ktls_mbuf_crypto_st_t
2240 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2242 int m_flags_ored = 0;
2243 int m_flags_anded = -1;
2245 for (; mb != NULL; mb = mb->m_next) {
2246 if (offset < mb->m_len)
2248 offset -= mb->m_len;
2252 for (; mb != NULL; mb = mb->m_next) {
2253 m_flags_ored |= mb->m_flags;
2254 m_flags_anded &= mb->m_flags;
2256 if (offset <= mb->m_len)
2258 offset -= mb->m_len;
2260 MPASS(mb != NULL || offset == 0);
2262 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2263 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2265 return ((m_flags_ored & M_DECRYPTED) ?
2266 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2267 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2271 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2274 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2276 union if_snd_tag_modify_params params;
2277 struct m_snd_tag *mst;
2281 mst = so->so_rcv.sb_tls_info->snd_tag;
2282 if (__predict_false(mst == NULL))
2285 inp = sotoinpcb(so);
2286 if (__predict_false(inp == NULL))
2290 if (inp->inp_flags & INP_DROPPED) {
2292 return (ECONNRESET);
2295 tp = intotcpcb(inp);
2298 /* Get the TCP sequence number of the next valid TLS header. */
2299 SOCKBUF_LOCK(&so->so_rcv);
2300 params.tls_rx.tls_hdr_tcp_sn =
2301 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2302 params.tls_rx.tls_rec_length = tls_len;
2303 params.tls_rx.tls_seq_number = tls_rcd_num;
2304 SOCKBUF_UNLOCK(&so->so_rcv);
2308 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2309 return (mst->sw->snd_tag_modify(mst, ¶ms));
2313 ktls_decrypt(struct socket *so)
2315 char tls_header[MBUF_PEXT_HDR_LEN];
2316 struct ktls_session *tls;
2318 struct tls_record_layer *hdr;
2319 struct tls_get_record tgr;
2320 struct mbuf *control, *data, *m;
2321 ktls_mbuf_crypto_st_t state;
2323 int error, remain, tls_len, trail_len;
2325 uint8_t vminor, record_type;
2327 hdr = (struct tls_record_layer *)tls_header;
2330 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2331 ("%s: socket %p not running", __func__, so));
2333 tls = sb->sb_tls_info;
2336 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2338 vminor = TLS_MINOR_VER_TWO;
2340 vminor = tls->params.tls_vminor;
2342 /* Is there enough queued for a TLS header? */
2343 if (sb->sb_tlscc < tls->params.tls_hlen)
2346 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2347 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2349 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2350 hdr->tls_vminor != vminor)
2352 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2354 else if (tls_len < tls->params.tls_hlen || tls_len >
2355 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2356 tls->params.tls_tlen)
2360 if (__predict_false(error != 0)) {
2362 * We have a corrupted record and are likely
2363 * out of sync. The connection isn't
2364 * recoverable at this point, so abort it.
2367 counter_u64_add(ktls_offload_corrupted_records, 1);
2369 CURVNET_SET(so->so_vnet);
2370 so->so_proto->pr_abort(so);
2371 so->so_error = error;
2376 /* Is the entire record queued? */
2377 if (sb->sb_tlscc < tls_len)
2381 * Split out the portion of the mbuf chain containing
2384 data = ktls_detach_record(sb, tls_len);
2387 MPASS(sb->sb_tlsdcc == tls_len);
2389 seqno = sb->sb_tls_seqno;
2394 /* get crypto state for this TLS record */
2395 state = ktls_mbuf_crypto_state(data, 0, tls_len);
2398 case KTLS_MBUF_CRYPTO_ST_MIXED:
2399 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2403 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2404 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2406 if (__predict_true(error == 0)) {
2408 error = tls13_find_record_type(tls, data,
2409 tls_len, &trail_len, &record_type);
2411 record_type = hdr->tls_type;
2415 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2417 * NIC TLS is only supported for AEAD
2418 * ciphersuites which used a fixed sized
2422 trail_len = tls->params.tls_tlen - 1;
2423 error = tls13_find_record_type(tls, data,
2424 tls_len, &trail_len, &record_type);
2426 trail_len = tls->params.tls_tlen;
2428 record_type = hdr->tls_type;
2436 counter_u64_add(ktls_offload_failed_crypto, 1);
2439 if (sb->sb_tlsdcc == 0) {
2441 * sbcut/drop/flush discarded these
2449 * Drop this TLS record's data, but keep
2450 * decrypting subsequent records.
2452 sb->sb_ccc -= tls_len;
2455 CURVNET_SET(so->so_vnet);
2456 so->so_error = EBADMSG;
2457 sorwakeup_locked(so);
2466 /* Allocate the control mbuf. */
2467 memset(&tgr, 0, sizeof(tgr));
2468 tgr.tls_type = record_type;
2469 tgr.tls_vmajor = hdr->tls_vmajor;
2470 tgr.tls_vminor = hdr->tls_vminor;
2471 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2473 control = sbcreatecontrol(&tgr, sizeof(tgr),
2474 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2477 if (sb->sb_tlsdcc == 0) {
2478 /* sbcut/drop/flush discarded these mbufs. */
2479 MPASS(sb->sb_tlscc == 0);
2486 * Clear the 'dcc' accounting in preparation for
2487 * adding the decrypted record.
2489 sb->sb_ccc -= tls_len;
2493 /* If there is no payload, drop all of the data. */
2494 if (tgr.tls_length == htobe16(0)) {
2499 remain = tls->params.tls_hlen;
2500 while (remain > 0) {
2501 if (data->m_len > remain) {
2502 data->m_data += remain;
2503 data->m_len -= remain;
2506 remain -= data->m_len;
2507 data = m_free(data);
2510 /* Trim trailer and clear M_NOTREADY. */
2511 remain = be16toh(tgr.tls_length);
2513 for (m = data; remain > m->m_len; m = m->m_next) {
2514 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2520 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2522 /* Set EOR on the final mbuf. */
2523 m->m_flags |= M_EOR;
2526 sbappendcontrol_locked(sb, data, control, 0);
2528 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2529 sb->sb_flags |= SB_TLS_RX_RESYNC;
2531 ktls_resync_ifnet(so, tls_len, seqno);
2533 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2534 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2536 ktls_resync_ifnet(so, 0, seqno);
2541 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2543 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2544 so->so_error = EMSGSIZE;
2546 sorwakeup_locked(so);
2549 SOCKBUF_UNLOCK_ASSERT(sb);
2551 CURVNET_SET(so->so_vnet);
2557 ktls_enqueue_to_free(struct mbuf *m)
2562 /* Mark it for freeing. */
2563 m->m_epg_flags |= EPG_FLAG_2FREE;
2564 wq = &ktls_wq[m->m_epg_tls->wq_index];
2566 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2567 running = wq->running;
2568 mtx_unlock(&wq->mtx);
2574 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2577 int domain, running;
2579 if (m->m_epg_npgs <= 2)
2581 if (ktls_buffer_zone == NULL)
2583 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2585 * Rate-limit allocation attempts after a failure.
2586 * ktls_buffer_import() will acquire a per-domain mutex to check
2587 * the free page queues and may fail consistently if memory is
2592 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2594 domain = PCPU_GET(domain);
2595 wq->lastallocfail = ticks;
2598 * Note that this check is "racy", but the races are
2599 * harmless, and are either a spurious wakeup if
2600 * multiple threads fail allocations before the alloc
2601 * thread wakes, or waiting an extra second in case we
2602 * see an old value of running == true.
2604 if (!VM_DOMAIN_EMPTY(domain)) {
2605 running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2607 wakeup(&ktls_domains[domain].alloc_td);
2614 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2615 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2618 int error, i, len, off;
2620 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2621 ("%p not unready & nomap mbuf\n", m));
2622 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2623 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2626 /* Anonymous mbufs are encrypted in place. */
2627 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2628 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2631 * For file-backed mbufs (from sendfile), anonymous wired
2632 * pages are allocated and used as the encryption destination.
2634 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2635 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2637 state->dst_iov[0].iov_base = (char *)state->cbuf +
2639 state->dst_iov[0].iov_len = len;
2640 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2643 off = m->m_epg_1st_off;
2644 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2645 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2646 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2647 len = m_epg_pagelen(m, i, off);
2648 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2649 state->dst_iov[i].iov_base =
2650 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2651 state->dst_iov[i].iov_len = len;
2654 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2655 state->dst_iov[i].iov_base = m->m_epg_trail;
2656 state->dst_iov[i].iov_len = m->m_epg_trllen;
2658 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2660 if (__predict_false(error != 0)) {
2661 /* Free the anonymous pages. */
2662 if (state->cbuf != NULL)
2663 uma_zfree(ktls_buffer_zone, state->cbuf);
2665 for (i = 0; i < m->m_epg_npgs; i++) {
2666 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2667 (void)vm_page_unwire_noq(pg);
2675 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2677 ktls_batched_records(struct mbuf *m)
2679 int page_count, records;
2682 page_count = m->m_epg_enc_cnt;
2683 while (page_count > 0) {
2685 page_count -= m->m_epg_nrdy;
2688 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2693 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2695 struct ktls_session *tls;
2700 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2701 (M_EXTPG | M_NOTREADY)),
2702 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2703 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2705 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2707 m->m_epg_enc_cnt = page_count;
2710 * Save a pointer to the socket. The caller is responsible
2711 * for taking an additional reference via soref().
2717 wq = &ktls_wq[tls->wq_index];
2719 if (__predict_false(tls->sequential_records)) {
2721 * For TLS 1.0, records must be encrypted
2722 * sequentially. For a given connection, all records
2723 * queued to the associated work queue are processed
2724 * sequentially. However, sendfile(2) might complete
2725 * I/O requests spanning multiple TLS records out of
2726 * order. Here we ensure TLS records are enqueued to
2727 * the work queue in FIFO order.
2729 * tls->next_seqno holds the sequence number of the
2730 * next TLS record that should be enqueued to the work
2731 * queue. If this next record is not tls->next_seqno,
2732 * it must be a future record, so insert it, sorted by
2733 * TLS sequence number, into tls->pending_records and
2736 * If this TLS record matches tls->next_seqno, place
2737 * it in the work queue and then check
2738 * tls->pending_records to see if any
2739 * previously-queued records are now ready for
2742 if (m->m_epg_seqno != tls->next_seqno) {
2746 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2747 if (n->m_epg_seqno > m->m_epg_seqno)
2752 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2755 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2758 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2760 mtx_unlock(&wq->mtx);
2761 counter_u64_add(ktls_cnt_tx_pending, 1);
2765 tls->next_seqno += ktls_batched_records(m);
2766 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2768 while (!STAILQ_EMPTY(&tls->pending_records)) {
2771 n = STAILQ_FIRST(&tls->pending_records);
2772 if (n->m_epg_seqno != tls->next_seqno)
2776 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2777 tls->next_seqno += ktls_batched_records(n);
2778 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2780 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2782 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2784 running = wq->running;
2785 mtx_unlock(&wq->mtx);
2788 counter_u64_add(ktls_cnt_tx_queued, queued);
2792 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2793 * the pages from the file and replace them with the anonymous pages
2794 * allocated in ktls_encrypt_record().
2797 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2801 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2803 /* Free the old pages. */
2804 m->m_ext.ext_free(m);
2806 /* Replace them with the new pages. */
2807 if (state->cbuf != NULL) {
2808 for (i = 0; i < m->m_epg_npgs; i++)
2809 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2811 /* Contig pages should go back to the cache. */
2812 m->m_ext.ext_free = ktls_free_mext_contig;
2814 for (i = 0; i < m->m_epg_npgs; i++)
2815 m->m_epg_pa[i] = state->parray[i];
2817 /* Use the basic free routine. */
2818 m->m_ext.ext_free = mb_free_mext_pgs;
2821 /* Pages are now writable. */
2822 m->m_epg_flags |= EPG_FLAG_ANON;
2825 static __noinline void
2826 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2828 struct ktls_ocf_encrypt_state state;
2829 struct ktls_session *tls;
2832 int error, npages, total_pages;
2835 tls = top->m_epg_tls;
2836 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2837 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2839 top->m_epg_so = NULL;
2841 total_pages = top->m_epg_enc_cnt;
2845 * Encrypt the TLS records in the chain of mbufs starting with
2846 * 'top'. 'total_pages' gives us a total count of pages and is
2847 * used to know when we have finished encrypting the TLS
2848 * records originally queued with 'top'.
2850 * NB: These mbufs are queued in the socket buffer and
2851 * 'm_next' is traversing the mbufs in the socket buffer. The
2852 * socket buffer lock is not held while traversing this chain.
2853 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2854 * pointers should be stable. However, the 'm_next' of the
2855 * last mbuf encrypted is not necessarily NULL. It can point
2856 * to other mbufs appended while 'top' was on the TLS work
2859 * Each mbuf holds an entire TLS record.
2862 for (m = top; npages != total_pages; m = m->m_next) {
2863 KASSERT(m->m_epg_tls == tls,
2864 ("different TLS sessions in a single mbuf chain: %p vs %p",
2865 tls, m->m_epg_tls));
2866 KASSERT(npages + m->m_epg_npgs <= total_pages,
2867 ("page count mismatch: top %p, total_pages %d, m %p", top,
2870 error = ktls_encrypt_record(wq, m, tls, &state);
2872 counter_u64_add(ktls_offload_failed_crypto, 1);
2876 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2877 ktls_finish_nonanon(m, &state);
2879 npages += m->m_epg_nrdy;
2882 * Drop a reference to the session now that it is no
2883 * longer needed. Existing code depends on encrypted
2884 * records having no associated session vs
2885 * yet-to-be-encrypted records having an associated
2888 m->m_epg_tls = NULL;
2892 CURVNET_SET(so->so_vnet);
2894 (void)so->so_proto->pr_ready(so, top, npages);
2896 so->so_proto->pr_abort(so);
2898 mb_free_notready(top, total_pages);
2906 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2908 struct ktls_session *tls;
2915 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2916 ktls_finish_nonanon(m, state);
2919 free(state, M_KTLS);
2922 * Drop a reference to the session now that it is no longer
2923 * needed. Existing code depends on encrypted records having
2924 * no associated session vs yet-to-be-encrypted records having
2925 * an associated session.
2928 m->m_epg_tls = NULL;
2932 counter_u64_add(ktls_offload_failed_crypto, 1);
2934 CURVNET_SET(so->so_vnet);
2935 npages = m->m_epg_nrdy;
2938 (void)so->so_proto->pr_ready(so, m, npages);
2940 so->so_proto->pr_abort(so);
2942 mb_free_notready(m, npages);
2950 * Similar to ktls_encrypt, but used with asynchronous OCF backends
2951 * (coprocessors) where encryption does not use host CPU resources and
2952 * it can be beneficial to queue more requests than CPUs.
2954 static __noinline void
2955 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2957 struct ktls_ocf_encrypt_state *state;
2958 struct ktls_session *tls;
2961 int error, mpages, npages, total_pages;
2964 tls = top->m_epg_tls;
2965 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2966 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2968 top->m_epg_so = NULL;
2970 total_pages = top->m_epg_enc_cnt;
2974 for (m = top; npages != total_pages; m = n) {
2975 KASSERT(m->m_epg_tls == tls,
2976 ("different TLS sessions in a single mbuf chain: %p vs %p",
2977 tls, m->m_epg_tls));
2978 KASSERT(npages + m->m_epg_npgs <= total_pages,
2979 ("page count mismatch: top %p, total_pages %d, m %p", top,
2982 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2987 mpages = m->m_epg_nrdy;
2990 error = ktls_encrypt_record(wq, m, tls, state);
2992 counter_u64_add(ktls_offload_failed_crypto, 1);
2993 free(state, M_KTLS);
2994 CURVNET_SET(so->so_vnet);
3003 CURVNET_SET(so->so_vnet);
3005 so->so_proto->pr_abort(so);
3007 mb_free_notready(m, total_pages - npages);
3015 ktls_bind_domain(int domain)
3019 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3022 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3027 ktls_alloc_thread(void *ctx)
3029 struct ktls_domain_info *ktls_domain = ctx;
3030 struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3032 struct sysctl_oid *oid;
3034 int domain, error, i, nbufs;
3036 domain = ktls_domain - ktls_domains;
3038 printf("Starting KTLS alloc thread for domain %d\n", domain);
3039 error = ktls_bind_domain(domain);
3041 printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3043 snprintf(name, sizeof(name), "domain%d", domain);
3044 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3045 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3046 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3047 CTLFLAG_RD, &sc->allocs, 0, "buffers allocated");
3048 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3049 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3050 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3051 CTLFLAG_RD, &sc->running, 0, "thread running");
3056 atomic_store_int(&sc->running, 0);
3057 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3058 atomic_store_int(&sc->running, 1);
3060 if (nbufs != ktls_max_alloc) {
3062 nbufs = atomic_load_int(&ktls_max_alloc);
3063 buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3067 * Below we allocate nbufs with different allocation
3068 * flags than we use when allocating normally during
3069 * encryption in the ktls worker thread. We specify
3070 * M_NORECLAIM in the worker thread. However, we omit
3071 * that flag here and add M_WAITOK so that the VM
3072 * system is permitted to perform expensive work to
3073 * defragment memory. We do this here, as it does not
3074 * matter if this thread blocks. If we block a ktls
3075 * worker thread, we risk developing backlogs of
3076 * buffers to be encrypted, leading to surges of
3077 * traffic and potential NIC output drops.
3079 for (i = 0; i < nbufs; i++) {
3080 buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3083 for (i = 0; i < nbufs; i++) {
3084 uma_zfree(ktls_buffer_zone, buf[i]);
3091 ktls_work_thread(void *ctx)
3093 struct ktls_wq *wq = ctx;
3095 struct socket *so, *son;
3096 STAILQ_HEAD(, mbuf) local_m_head;
3097 STAILQ_HEAD(, socket) local_so_head;
3102 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3105 * Bind to a core. If ktls_bind_threads is > 1, then
3106 * we bind to the NUMA domain instead.
3108 if (ktls_bind_threads) {
3111 if (ktls_bind_threads > 1) {
3112 struct pcpu *pc = pcpu_find(cpu);
3114 error = ktls_bind_domain(pc->pc_domain);
3118 CPU_SETOF(cpu, &mask);
3119 error = cpuset_setthread(curthread->td_tid, &mask);
3122 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3125 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3130 while (STAILQ_EMPTY(&wq->m_head) &&
3131 STAILQ_EMPTY(&wq->so_head)) {
3132 wq->running = false;
3133 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3137 STAILQ_INIT(&local_m_head);
3138 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3139 STAILQ_INIT(&local_so_head);
3140 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3141 mtx_unlock(&wq->mtx);
3143 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3144 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3145 ktls_free(m->m_epg_tls);
3148 if (m->m_epg_tls->sync_dispatch)
3149 ktls_encrypt(wq, m);
3151 ktls_encrypt_async(wq, m);
3152 counter_u64_add(ktls_cnt_tx_queued, -1);
3156 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3158 counter_u64_add(ktls_cnt_rx_queued, -1);
3163 #if defined(INET) || defined(INET6)
3165 ktls_disable_ifnet_help(void *context, int pending __unused)
3167 struct ktls_session *tls;
3178 so = inp->inp_socket;
3180 if (inp->inp_flags & INP_DROPPED) {
3184 if (so->so_snd.sb_tls_info != NULL)
3185 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3189 counter_u64_add(ktls_ifnet_disable_ok, 1);
3190 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3191 if ((inp->inp_flags & INP_DROPPED) == 0 &&
3192 (tp = intotcpcb(inp)) != NULL &&
3193 tp->t_fb->tfb_hwtls_change != NULL)
3194 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3196 counter_u64_add(ktls_ifnet_disable_fail, 1);
3201 if (!in_pcbrele_wlocked(inp))
3207 * Called when re-transmits are becoming a substantial portion of the
3208 * sends on this connection. When this happens, we transition the
3209 * connection to software TLS. This is needed because most inline TLS
3210 * NICs keep crypto state only for in-order transmits. This means
3211 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3212 * re-DMA the entire TLS record up to and including the current
3213 * segment. This means that when re-transmitting the last ~1448 byte
3214 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3215 * of magnitude more data than we are sending. This can cause the
3216 * PCIe link to saturate well before the network, which can cause
3217 * output drops, and a general loss of capacity.
3220 ktls_disable_ifnet(void *arg)
3225 struct ktls_session *tls;
3228 inp = tptoinpcb(tp);
3229 INP_WLOCK_ASSERT(inp);
3230 so = inp->inp_socket;
3232 tls = so->so_snd.sb_tls_info;
3233 if (tls->disable_ifnet_pending) {
3239 * note that disable_ifnet_pending is never cleared; disabling
3240 * ifnet can only be done once per session, so we never want
3244 (void)ktls_hold(tls);
3247 tls->disable_ifnet_pending = true;
3250 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3251 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);