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 #include <netinet/in.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/tcp_var.h>
73 #include <netinet/tcp_offload.h>
75 #include <opencrypto/cryptodev.h>
76 #include <opencrypto/ktls.h>
77 #include <vm/uma_dbg.h>
79 #include <vm/vm_pageout.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pagequeue.h>
85 STAILQ_HEAD(, mbuf) m_head;
86 STAILQ_HEAD(, socket) so_head;
89 } __aligned(CACHE_LINE_SIZE);
91 struct ktls_alloc_thread {
98 struct ktls_domain_info {
101 struct ktls_alloc_thread alloc_td;
104 struct ktls_domain_info ktls_domains[MAXMEMDOM];
105 static struct ktls_wq *ktls_wq;
106 static struct proc *ktls_proc;
107 static uma_zone_t ktls_session_zone;
108 static uma_zone_t ktls_buffer_zone;
109 static uint16_t ktls_cpuid_lookup[MAXCPU];
110 static int ktls_init_state;
111 static struct sx ktls_init_lock;
112 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
114 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
115 "Kernel TLS offload");
116 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117 "Kernel TLS offload stats");
120 static int ktls_bind_threads = 1;
122 static int ktls_bind_threads;
124 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
125 &ktls_bind_threads, 0,
126 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
128 static u_int ktls_maxlen = 16384;
129 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
130 &ktls_maxlen, 0, "Maximum TLS record size");
132 static int ktls_number_threads;
133 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
134 &ktls_number_threads, 0,
135 "Number of TLS threads in thread-pool");
137 unsigned int ktls_ifnet_max_rexmit_pct = 2;
138 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
139 &ktls_ifnet_max_rexmit_pct, 2,
140 "Max percent bytes retransmitted before ifnet TLS is disabled");
142 static bool ktls_offload_enable;
143 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
144 &ktls_offload_enable, 0,
145 "Enable support for kernel TLS offload");
147 static bool ktls_cbc_enable = true;
148 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
150 "Enable support of AES-CBC crypto for kernel TLS");
152 static bool ktls_sw_buffer_cache = true;
153 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
154 &ktls_sw_buffer_cache, 1,
155 "Enable caching of output buffers for SW encryption");
157 static int ktls_max_alloc = 128;
158 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
159 &ktls_max_alloc, 128,
160 "Max number of 16k buffers to allocate in thread context");
162 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
163 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
164 &ktls_tasks_active, "Number of active tasks");
166 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
167 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
168 &ktls_cnt_tx_pending,
169 "Number of TLS 1.0 records waiting for earlier TLS records");
171 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
172 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
174 "Number of TLS records in queue to tasks for SW encryption");
176 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
177 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
179 "Number of TLS sockets in queue to tasks for SW decryption");
181 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
182 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
183 CTLFLAG_RD, &ktls_offload_total,
184 "Total successful TLS setups (parameters set)");
186 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
187 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
188 CTLFLAG_RD, &ktls_offload_enable_calls,
189 "Total number of TLS enable calls made");
191 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
192 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
193 &ktls_offload_active, "Total Active TLS sessions");
195 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
197 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
199 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
200 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
201 &ktls_offload_failed_crypto, "Total TLS crypto failures");
203 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
204 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
205 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
207 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
208 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
209 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
211 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
212 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
213 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
215 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
217 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
219 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
220 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
221 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
223 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
224 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
226 "Number of times ktls session was destroyed via taskqueue");
228 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
229 "Software TLS session stats");
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231 "Hardware (ifnet) TLS session stats");
233 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
234 "TOE TLS session stats");
237 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
238 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
239 "Active number of software TLS sessions using AES-CBC");
241 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
242 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
243 "Active number of software TLS sessions using AES-GCM");
245 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
246 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
248 "Active number of software TLS sessions using Chacha20-Poly1305");
250 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
251 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
253 "Active number of ifnet TLS sessions using AES-CBC");
255 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
256 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
258 "Active number of ifnet TLS sessions using AES-GCM");
260 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
261 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
262 &ktls_ifnet_chacha20,
263 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
265 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
266 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
267 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
269 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
270 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
271 &ktls_ifnet_reset_dropped,
272 "TLS sessions dropped after failing to update ifnet send tag");
274 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
275 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
276 &ktls_ifnet_reset_failed,
277 "TLS sessions that failed to allocate a new ifnet send tag");
279 static int ktls_ifnet_permitted;
280 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
281 &ktls_ifnet_permitted, 1,
282 "Whether to permit hardware (ifnet) TLS sessions");
285 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
286 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
288 "Active number of TOE TLS sessions using AES-CBC");
290 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
291 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
293 "Active number of TOE TLS sessions using AES-GCM");
295 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
296 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
298 "Active number of TOE TLS sessions using Chacha20-Poly1305");
301 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
303 static void ktls_reset_receive_tag(void *context, int pending);
304 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);
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];
340 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
345 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
346 ("%s: ktls max length %d is not page size-aligned",
347 __func__, ktls_maxlen));
349 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
350 for (i = 0; i < count; i++) {
351 m = vm_page_alloc_noobj_contig_domain(domain, req,
352 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
356 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
362 ktls_buffer_release(void *arg __unused, void **store, int count)
367 for (i = 0; i < count; i++) {
368 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
369 for (j = 0; j < atop(ktls_maxlen); j++) {
370 (void)vm_page_unwire_noq(m + j);
377 ktls_free_mext_contig(struct mbuf *m)
380 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
388 int count, domain, error, i;
390 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
393 ktls_session_zone = uma_zcreate("ktls_session",
394 sizeof(struct ktls_session),
395 NULL, NULL, NULL, NULL,
398 if (ktls_sw_buffer_cache) {
399 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
400 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
401 ktls_buffer_import, ktls_buffer_release, NULL,
402 UMA_ZONE_FIRSTTOUCH);
406 * Initialize the workqueues to run the TLS work. We create a
407 * work queue for each CPU.
410 STAILQ_INIT(&ktls_wq[i].m_head);
411 STAILQ_INIT(&ktls_wq[i].so_head);
412 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
413 if (ktls_bind_threads > 1) {
415 domain = pc->pc_domain;
416 count = ktls_domains[domain].count;
417 ktls_domains[domain].cpu[count] = i;
418 ktls_domains[domain].count++;
420 ktls_cpuid_lookup[ktls_number_threads] = i;
421 ktls_number_threads++;
425 * If we somehow have an empty domain, fall back to choosing
426 * among all KTLS threads.
428 if (ktls_bind_threads > 1) {
429 for (i = 0; i < vm_ndomains; i++) {
430 if (ktls_domains[i].count == 0) {
431 ktls_bind_threads = 1;
437 /* Start kthreads for each workqueue. */
439 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
440 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
442 printf("Can't add KTLS thread %d error %d\n", i, error);
448 * Start an allocation thread per-domain to perform blocking allocations
449 * of 16k physically contiguous TLS crypto destination buffers.
451 if (ktls_sw_buffer_cache) {
452 for (domain = 0; domain < vm_ndomains; domain++) {
453 if (VM_DOMAIN_EMPTY(domain))
455 if (CPU_EMPTY(&cpuset_domain[domain]))
457 error = kproc_kthread_add(ktls_alloc_thread,
458 &ktls_domains[domain], &ktls_proc,
459 &ktls_domains[domain].alloc_td.td,
460 0, 0, "KTLS", "alloc_%d", domain);
462 printf("Can't add KTLS alloc thread %d error %d\n",
470 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
475 ktls_start_kthreads(void)
480 state = atomic_load_acq_int(&ktls_init_state);
481 if (__predict_true(state > 0))
486 sx_xlock(&ktls_init_lock);
487 if (ktls_init_state != 0) {
488 sx_xunlock(&ktls_init_lock);
497 atomic_store_rel_int(&ktls_init_state, state);
498 sx_xunlock(&ktls_init_lock);
503 ktls_create_session(struct socket *so, struct tls_enable *en,
504 struct ktls_session **tlsp, int direction)
506 struct ktls_session *tls;
509 /* Only TLS 1.0 - 1.3 are supported. */
510 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
512 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
513 en->tls_vminor > TLS_MINOR_VER_THREE)
516 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
518 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
520 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
523 /* All supported algorithms require a cipher key. */
524 if (en->cipher_key_len == 0)
527 /* No flags are currently supported. */
531 /* Common checks for supported algorithms. */
532 switch (en->cipher_algorithm) {
533 case CRYPTO_AES_NIST_GCM_16:
535 * auth_algorithm isn't used, but permit GMAC values
538 switch (en->auth_algorithm) {
540 #ifdef COMPAT_FREEBSD12
541 /* XXX: Really 13.0-current COMPAT. */
542 case CRYPTO_AES_128_NIST_GMAC:
543 case CRYPTO_AES_192_NIST_GMAC:
544 case CRYPTO_AES_256_NIST_GMAC:
550 if (en->auth_key_len != 0)
552 switch (en->tls_vminor) {
553 case TLS_MINOR_VER_TWO:
554 if (en->iv_len != TLS_AEAD_GCM_LEN)
557 case TLS_MINOR_VER_THREE:
558 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
566 switch (en->auth_algorithm) {
567 case CRYPTO_SHA1_HMAC:
569 case CRYPTO_SHA2_256_HMAC:
570 case CRYPTO_SHA2_384_HMAC:
571 if (en->tls_vminor != TLS_MINOR_VER_TWO)
577 if (en->auth_key_len == 0)
581 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
584 switch (en->tls_vminor) {
585 case TLS_MINOR_VER_ZERO:
586 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
589 case TLS_MINOR_VER_ONE:
590 case TLS_MINOR_VER_TWO:
591 /* Ignore any supplied IV. */
598 case CRYPTO_CHACHA20_POLY1305:
599 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
601 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
602 en->tls_vminor != TLS_MINOR_VER_THREE)
604 if (en->iv_len != TLS_CHACHA20_IV_LEN)
611 error = ktls_start_kthreads();
615 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
617 counter_u64_add(ktls_offload_active, 1);
619 refcount_init(&tls->refcount, 1);
620 if (direction == KTLS_RX) {
621 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
623 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
624 tls->inp = so->so_pcb;
629 tls->wq_index = ktls_get_cpu(so);
631 tls->params.cipher_algorithm = en->cipher_algorithm;
632 tls->params.auth_algorithm = en->auth_algorithm;
633 tls->params.tls_vmajor = en->tls_vmajor;
634 tls->params.tls_vminor = en->tls_vminor;
635 tls->params.flags = en->flags;
636 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
638 /* Set the header and trailer lengths. */
639 tls->params.tls_hlen = sizeof(struct tls_record_layer);
640 switch (en->cipher_algorithm) {
641 case CRYPTO_AES_NIST_GCM_16:
643 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
644 * nonce. TLS 1.3 uses a 12 byte implicit IV.
646 if (en->tls_vminor < TLS_MINOR_VER_THREE)
647 tls->params.tls_hlen += sizeof(uint64_t);
648 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
649 tls->params.tls_bs = 1;
652 switch (en->auth_algorithm) {
653 case CRYPTO_SHA1_HMAC:
654 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
655 /* Implicit IV, no nonce. */
656 tls->sequential_records = true;
657 tls->next_seqno = be64dec(en->rec_seq);
658 STAILQ_INIT(&tls->pending_records);
660 tls->params.tls_hlen += AES_BLOCK_LEN;
662 tls->params.tls_tlen = AES_BLOCK_LEN +
665 case CRYPTO_SHA2_256_HMAC:
666 tls->params.tls_hlen += AES_BLOCK_LEN;
667 tls->params.tls_tlen = AES_BLOCK_LEN +
670 case CRYPTO_SHA2_384_HMAC:
671 tls->params.tls_hlen += AES_BLOCK_LEN;
672 tls->params.tls_tlen = AES_BLOCK_LEN +
676 panic("invalid hmac");
678 tls->params.tls_bs = AES_BLOCK_LEN;
680 case CRYPTO_CHACHA20_POLY1305:
682 * Chacha20 uses a 12 byte implicit IV.
684 tls->params.tls_tlen = POLY1305_HASH_LEN;
685 tls->params.tls_bs = 1;
688 panic("invalid cipher");
692 * TLS 1.3 includes optional padding which we do not support,
693 * and also puts the "real" record type at the end of the
696 if (en->tls_vminor == TLS_MINOR_VER_THREE)
697 tls->params.tls_tlen += sizeof(uint8_t);
699 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
700 ("TLS header length too long: %d", tls->params.tls_hlen));
701 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
702 ("TLS trailer length too long: %d", tls->params.tls_tlen));
704 if (en->auth_key_len != 0) {
705 tls->params.auth_key_len = en->auth_key_len;
706 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
708 error = copyin(en->auth_key, tls->params.auth_key,
714 tls->params.cipher_key_len = en->cipher_key_len;
715 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
716 error = copyin(en->cipher_key, tls->params.cipher_key,
722 * This holds the implicit portion of the nonce for AEAD
723 * ciphers and the initial implicit IV for TLS 1.0. The
724 * explicit portions of the IV are generated in ktls_frame().
726 if (en->iv_len != 0) {
727 tls->params.iv_len = en->iv_len;
728 error = copyin(en->iv, tls->params.iv, en->iv_len);
733 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
734 * counter to generate unique explicit IVs.
736 * Store this counter in the last 8 bytes of the IV
737 * array so that it is 8-byte aligned.
739 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
740 en->tls_vminor == TLS_MINOR_VER_TWO)
741 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
752 static struct ktls_session *
753 ktls_clone_session(struct ktls_session *tls, int direction)
755 struct ktls_session *tls_new;
757 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
759 counter_u64_add(ktls_offload_active, 1);
761 refcount_init(&tls_new->refcount, 1);
762 if (direction == KTLS_RX) {
763 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
766 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
768 tls_new->inp = tls->inp;
770 in_pcbref(tls_new->inp);
773 /* Copy fields from existing session. */
774 tls_new->params = tls->params;
775 tls_new->wq_index = tls->wq_index;
777 /* Deep copy keys. */
778 if (tls_new->params.auth_key != NULL) {
779 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
781 memcpy(tls_new->params.auth_key, tls->params.auth_key,
782 tls->params.auth_key_len);
785 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
787 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
788 tls->params.cipher_key_len);
795 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
803 if (inp->inp_flags & INP_DROPPED) {
807 if (inp->inp_socket == NULL) {
812 if (!(tp->t_flags & TF_TOE)) {
817 error = tcp_offload_alloc_tls_session(tp, tls, direction);
820 tls->mode = TCP_TLS_MODE_TOE;
821 switch (tls->params.cipher_algorithm) {
823 counter_u64_add(ktls_toe_cbc, 1);
825 case CRYPTO_AES_NIST_GCM_16:
826 counter_u64_add(ktls_toe_gcm, 1);
828 case CRYPTO_CHACHA20_POLY1305:
829 counter_u64_add(ktls_toe_chacha20, 1);
838 * Common code used when first enabling ifnet TLS on a connection or
839 * when allocating a new ifnet TLS session due to a routing change.
840 * This function allocates a new TLS send tag on whatever interface
841 * the connection is currently routed over.
844 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
845 struct m_snd_tag **mstp)
847 union if_snd_tag_alloc_params params;
849 struct nhop_object *nh;
854 if (inp->inp_flags & INP_DROPPED) {
858 if (inp->inp_socket == NULL) {
865 * Check administrative controls on ifnet TLS to determine if
866 * ifnet TLS should be denied.
868 * - Always permit 'force' requests.
869 * - ktls_ifnet_permitted == 0: always deny.
871 if (!force && ktls_ifnet_permitted == 0) {
877 * XXX: Use the cached route in the inpcb to find the
878 * interface. This should perhaps instead use
879 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
880 * enabled after a connection has completed key negotiation in
881 * userland, the cached route will be present in practice.
883 nh = inp->inp_route.ro_nh;
892 * Allocate a TLS + ratelimit tag if the connection has an
893 * existing pacing rate.
895 if (tp->t_pacing_rate != -1 &&
896 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
897 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
898 params.tls_rate_limit.inp = inp;
899 params.tls_rate_limit.tls = tls;
900 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
902 params.hdr.type = IF_SND_TAG_TYPE_TLS;
903 params.tls.inp = inp;
904 params.tls.tls = tls;
906 params.hdr.flowid = inp->inp_flowid;
907 params.hdr.flowtype = inp->inp_flowtype;
908 params.hdr.numa_domain = inp->inp_numa_domain;
911 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
915 if (inp->inp_vflag & INP_IPV6) {
916 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
921 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
926 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
933 * Allocate an initial TLS receive tag for doing HW decryption of TLS
936 * This function allocates a new TLS receive tag on whatever interface
937 * the connection is currently routed over. If the connection ends up
938 * using a different interface for receive this will get fixed up via
939 * ktls_input_ifp_mismatch as future packets arrive.
942 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
943 struct m_snd_tag **mstp)
945 union if_snd_tag_alloc_params params;
947 struct nhop_object *nh;
950 if (!ktls_ocf_recrypt_supported(tls))
954 if (inp->inp_flags & INP_DROPPED) {
958 if (inp->inp_socket == NULL) {
964 * Check administrative controls on ifnet TLS to determine if
965 * ifnet TLS should be denied.
967 if (ktls_ifnet_permitted == 0) {
973 * XXX: As with ktls_alloc_snd_tag, use the cached route in
974 * the inpcb to find the interface.
976 nh = inp->inp_route.ro_nh;
985 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
986 params.hdr.flowid = inp->inp_flowid;
987 params.hdr.flowtype = inp->inp_flowtype;
988 params.hdr.numa_domain = inp->inp_numa_domain;
989 params.tls_rx.inp = inp;
990 params.tls_rx.tls = tls;
991 params.tls_rx.vlan_id = 0;
995 if (inp->inp_vflag & INP_IPV6) {
996 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0) {
1001 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0) {
1006 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1009 * If this connection is over a vlan, vlan_snd_tag_alloc
1010 * rewrites vlan_id with the saved interface. Save the VLAN
1011 * ID for use in ktls_reset_receive_tag which allocates new
1012 * receive tags directly from the leaf interface bypassing
1016 tls->rx_vlan_id = params.tls_rx.vlan_id;
1022 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1025 struct m_snd_tag *mst;
1028 switch (direction) {
1030 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1031 if (__predict_false(error != 0))
1035 KASSERT(!force, ("%s: forced receive tag", __func__));
1036 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1037 if (__predict_false(error != 0))
1041 __assert_unreachable();
1044 tls->mode = TCP_TLS_MODE_IFNET;
1047 switch (tls->params.cipher_algorithm) {
1048 case CRYPTO_AES_CBC:
1049 counter_u64_add(ktls_ifnet_cbc, 1);
1051 case CRYPTO_AES_NIST_GCM_16:
1052 counter_u64_add(ktls_ifnet_gcm, 1);
1054 case CRYPTO_CHACHA20_POLY1305:
1055 counter_u64_add(ktls_ifnet_chacha20, 1);
1065 ktls_use_sw(struct ktls_session *tls)
1067 tls->mode = TCP_TLS_MODE_SW;
1068 switch (tls->params.cipher_algorithm) {
1069 case CRYPTO_AES_CBC:
1070 counter_u64_add(ktls_sw_cbc, 1);
1072 case CRYPTO_AES_NIST_GCM_16:
1073 counter_u64_add(ktls_sw_gcm, 1);
1075 case CRYPTO_CHACHA20_POLY1305:
1076 counter_u64_add(ktls_sw_chacha20, 1);
1082 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1086 error = ktls_ocf_try(so, tls, direction);
1094 * KTLS RX stores data in the socket buffer as a list of TLS records,
1095 * where each record is stored as a control message containg the TLS
1096 * header followed by data mbufs containing the decrypted data. This
1097 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1098 * both encrypted and decrypted data. TLS records decrypted by a NIC
1099 * should be queued to the socket buffer as records, but encrypted
1100 * data which needs to be decrypted by software arrives as a stream of
1101 * regular mbufs which need to be converted. In addition, there may
1102 * already be pending encrypted data in the socket buffer when KTLS RX
1105 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1108 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1110 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1111 * from the first mbuf. Once all of the data for that TLS record is
1112 * queued, the socket is queued to a worker thread.
1114 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1115 * the TLS chain. Each TLS record is detached from the TLS chain,
1116 * decrypted, and inserted into the regular socket buffer chain as
1117 * record starting with a control message holding the TLS header and
1118 * a chain of mbufs holding the encrypted data.
1122 sb_mark_notready(struct sockbuf *sb)
1129 sb->sb_mbtail = NULL;
1130 sb->sb_lastrecord = NULL;
1131 for (; m != NULL; m = m->m_next) {
1132 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1134 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1136 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1138 m->m_flags |= M_NOTREADY;
1139 sb->sb_acc -= m->m_len;
1140 sb->sb_tlscc += m->m_len;
1141 sb->sb_mtlstail = m;
1143 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1144 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1149 * Return information about the pending TLS data in a socket
1150 * buffer. On return, 'seqno' is set to the sequence number
1151 * of the next TLS record to be received, 'resid' is set to
1152 * the amount of bytes still needed for the last pending
1153 * record. The function returns 'false' if the last pending
1154 * record contains a partial TLS header. In that case, 'resid'
1155 * is the number of bytes needed to complete the TLS header.
1158 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1160 struct tls_record_layer hdr;
1164 u_int offset, record_len;
1166 SOCKBUF_LOCK_ASSERT(sb);
1167 MPASS(sb->sb_flags & SB_TLS_RX);
1168 seqno = sb->sb_tls_seqno;
1169 resid = sb->sb_tlscc;
1182 if (resid < sizeof(hdr)) {
1184 *residp = sizeof(hdr) - resid;
1188 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1190 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1191 if (resid <= record_len) {
1193 *residp = record_len - resid;
1196 resid -= record_len;
1198 while (record_len != 0) {
1199 if (m->m_len - offset > record_len) {
1200 offset += record_len;
1204 record_len -= (m->m_len - offset);
1212 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1214 struct ktls_session *tls;
1217 if (!ktls_offload_enable)
1220 counter_u64_add(ktls_offload_enable_calls, 1);
1223 * This should always be true since only the TCP socket option
1224 * invokes this function.
1226 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1230 * XXX: Don't overwrite existing sessions. We should permit
1231 * this to support rekeying in the future.
1233 if (so->so_rcv.sb_tls_info != NULL)
1236 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1239 error = ktls_create_session(so, en, &tls, KTLS_RX);
1243 error = ktls_ocf_try(so, tls, KTLS_RX);
1249 /* Mark the socket as using TLS offload. */
1250 SOCK_RECVBUF_LOCK(so);
1251 if (SOLISTENING(so)) {
1252 SOCK_RECVBUF_UNLOCK(so);
1256 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1257 so->so_rcv.sb_tls_info = tls;
1258 so->so_rcv.sb_flags |= SB_TLS_RX;
1260 /* Mark existing data as not ready until it can be decrypted. */
1261 sb_mark_notready(&so->so_rcv);
1262 ktls_check_rx(&so->so_rcv);
1263 SOCK_RECVBUF_UNLOCK(so);
1265 /* Prefer TOE -> ifnet TLS -> software TLS. */
1267 error = ktls_try_toe(so, tls, KTLS_RX);
1270 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1274 counter_u64_add(ktls_offload_total, 1);
1280 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1282 struct ktls_session *tls;
1287 if (!ktls_offload_enable)
1290 counter_u64_add(ktls_offload_enable_calls, 1);
1293 * This should always be true since only the TCP socket option
1294 * invokes this function.
1296 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1300 * XXX: Don't overwrite existing sessions. We should permit
1301 * this to support rekeying in the future.
1303 if (so->so_snd.sb_tls_info != NULL)
1306 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1309 /* TLS requires ext pgs */
1310 if (mb_use_ext_pgs == 0)
1313 error = ktls_create_session(so, en, &tls, KTLS_TX);
1317 /* Prefer TOE -> ifnet TLS -> software TLS. */
1319 error = ktls_try_toe(so, tls, KTLS_TX);
1322 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1324 error = ktls_try_sw(so, tls, KTLS_TX);
1332 * Serialize with sosend_generic() and make sure that we're not
1333 * operating on a listening socket.
1335 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1342 * Write lock the INP when setting sb_tls_info so that
1343 * routines in tcp_ratelimit.c can read sb_tls_info while
1344 * holding the INP lock.
1348 SOCK_SENDBUF_LOCK(so);
1349 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1350 so->so_snd.sb_tls_info = tls;
1351 if (tls->mode != TCP_TLS_MODE_SW) {
1352 tp = intotcpcb(inp);
1353 MPASS(tp->t_nic_ktls_xmit == 0);
1354 tp->t_nic_ktls_xmit = 1;
1355 if (tp->t_fb->tfb_hwtls_change != NULL)
1356 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1358 SOCK_SENDBUF_UNLOCK(so);
1360 SOCK_IO_SEND_UNLOCK(so);
1362 counter_u64_add(ktls_offload_total, 1);
1368 ktls_get_rx_mode(struct socket *so, int *modep)
1370 struct ktls_session *tls;
1371 struct inpcb *inp __diagused;
1373 if (SOLISTENING(so))
1376 INP_WLOCK_ASSERT(inp);
1377 SOCK_RECVBUF_LOCK(so);
1378 tls = so->so_rcv.sb_tls_info;
1380 *modep = TCP_TLS_MODE_NONE;
1383 SOCK_RECVBUF_UNLOCK(so);
1388 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1390 * This function gets information about the next TCP- and TLS-
1391 * sequence number to be processed by the TLS receive worker
1392 * thread. The information is extracted from the given "inpcb"
1393 * structure. The values are stored in host endian format at the two
1394 * given output pointer locations. The TCP sequence number points to
1395 * the beginning of the TLS header.
1397 * This function returns zero on success, else a non-zero error code
1401 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1407 so = inp->inp_socket;
1408 if (__predict_false(so == NULL)) {
1412 if (inp->inp_flags & INP_DROPPED) {
1414 return (ECONNRESET);
1417 tp = intotcpcb(inp);
1420 SOCKBUF_LOCK(&so->so_rcv);
1421 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1422 *tlsseq = so->so_rcv.sb_tls_seqno;
1423 SOCKBUF_UNLOCK(&so->so_rcv);
1431 ktls_get_tx_mode(struct socket *so, int *modep)
1433 struct ktls_session *tls;
1434 struct inpcb *inp __diagused;
1436 if (SOLISTENING(so))
1439 INP_WLOCK_ASSERT(inp);
1440 SOCK_SENDBUF_LOCK(so);
1441 tls = so->so_snd.sb_tls_info;
1443 *modep = TCP_TLS_MODE_NONE;
1446 SOCK_SENDBUF_UNLOCK(so);
1451 * Switch between SW and ifnet TLS sessions as requested.
1454 ktls_set_tx_mode(struct socket *so, int mode)
1456 struct ktls_session *tls, *tls_new;
1461 if (SOLISTENING(so))
1464 case TCP_TLS_MODE_SW:
1465 case TCP_TLS_MODE_IFNET:
1472 INP_WLOCK_ASSERT(inp);
1473 tp = intotcpcb(inp);
1475 if (mode == TCP_TLS_MODE_IFNET) {
1476 /* Don't allow enabling ifnet ktls multiple times */
1477 if (tp->t_nic_ktls_xmit)
1481 * Don't enable ifnet ktls if we disabled it due to an
1482 * excessive retransmission rate
1484 if (tp->t_nic_ktls_xmit_dis)
1488 SOCKBUF_LOCK(&so->so_snd);
1489 tls = so->so_snd.sb_tls_info;
1491 SOCKBUF_UNLOCK(&so->so_snd);
1495 if (tls->mode == mode) {
1496 SOCKBUF_UNLOCK(&so->so_snd);
1500 tls = ktls_hold(tls);
1501 SOCKBUF_UNLOCK(&so->so_snd);
1504 tls_new = ktls_clone_session(tls, KTLS_TX);
1506 if (mode == TCP_TLS_MODE_IFNET)
1507 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1509 error = ktls_try_sw(so, tls_new, KTLS_TX);
1511 counter_u64_add(ktls_switch_failed, 1);
1518 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1520 counter_u64_add(ktls_switch_failed, 1);
1528 * If we raced with another session change, keep the existing
1531 if (tls != so->so_snd.sb_tls_info) {
1532 counter_u64_add(ktls_switch_failed, 1);
1533 SOCK_IO_SEND_UNLOCK(so);
1541 SOCKBUF_LOCK(&so->so_snd);
1542 so->so_snd.sb_tls_info = tls_new;
1543 if (tls_new->mode != TCP_TLS_MODE_SW) {
1544 MPASS(tp->t_nic_ktls_xmit == 0);
1545 tp->t_nic_ktls_xmit = 1;
1546 if (tp->t_fb->tfb_hwtls_change != NULL)
1547 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1549 SOCKBUF_UNLOCK(&so->so_snd);
1550 SOCK_IO_SEND_UNLOCK(so);
1553 * Drop two references on 'tls'. The first is for the
1554 * ktls_hold() above. The second drops the reference from the
1557 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1561 if (mode == TCP_TLS_MODE_IFNET)
1562 counter_u64_add(ktls_switch_to_ifnet, 1);
1564 counter_u64_add(ktls_switch_to_sw, 1);
1570 * Try to allocate a new TLS receive tag. This task is scheduled when
1571 * sbappend_ktls_rx detects an input path change. If a new tag is
1572 * allocated, replace the tag in the TLS session. If a new tag cannot
1573 * be allocated, let the session fall back to software decryption.
1576 ktls_reset_receive_tag(void *context, int pending)
1578 union if_snd_tag_alloc_params params;
1579 struct ktls_session *tls;
1580 struct m_snd_tag *mst;
1586 MPASS(pending == 1);
1594 if (inp->inp_flags & INP_DROPPED) {
1599 SOCKBUF_LOCK(&so->so_rcv);
1601 tls->snd_tag = NULL;
1603 m_snd_tag_rele(mst);
1607 SOCKBUF_UNLOCK(&so->so_rcv);
1609 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1610 params.hdr.flowid = inp->inp_flowid;
1611 params.hdr.flowtype = inp->inp_flowtype;
1612 params.hdr.numa_domain = inp->inp_numa_domain;
1613 params.tls_rx.inp = inp;
1614 params.tls_rx.tls = tls;
1615 params.tls_rx.vlan_id = tls->rx_vlan_id;
1618 if (inp->inp_vflag & INP_IPV6) {
1619 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1622 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1626 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1628 SOCKBUF_LOCK(&so->so_rcv);
1630 SOCKBUF_UNLOCK(&so->so_rcv);
1632 counter_u64_add(ktls_ifnet_reset, 1);
1635 * Just fall back to software decryption if a tag
1636 * cannot be allocated leaving the connection intact.
1637 * If a future input path change switches to another
1638 * interface this connection will resume ifnet TLS.
1640 counter_u64_add(ktls_ifnet_reset_failed, 1);
1644 mtx_pool_lock(mtxpool_sleep, tls);
1645 tls->reset_pending = false;
1646 mtx_pool_unlock(mtxpool_sleep, tls);
1655 * Try to allocate a new TLS send tag. This task is scheduled when
1656 * ip_output detects a route change while trying to transmit a packet
1657 * holding a TLS record. If a new tag is allocated, replace the tag
1658 * in the TLS session. Subsequent packets on the connection will use
1659 * the new tag. If a new tag cannot be allocated, drop the
1663 ktls_reset_send_tag(void *context, int pending)
1665 struct epoch_tracker et;
1666 struct ktls_session *tls;
1667 struct m_snd_tag *old, *new;
1672 MPASS(pending == 1);
1678 * Free the old tag first before allocating a new one.
1679 * ip[6]_output_send() will treat a NULL send tag the same as
1680 * an ifp mismatch and drop packets until a new tag is
1683 * Write-lock the INP when changing tls->snd_tag since
1684 * ip[6]_output_send() holds a read-lock when reading the
1689 tls->snd_tag = NULL;
1692 m_snd_tag_rele(old);
1694 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1699 mtx_pool_lock(mtxpool_sleep, tls);
1700 tls->reset_pending = false;
1701 mtx_pool_unlock(mtxpool_sleep, tls);
1704 counter_u64_add(ktls_ifnet_reset, 1);
1707 * XXX: Should we kick tcp_output explicitly now that
1708 * the send tag is fixed or just rely on timers?
1711 NET_EPOCH_ENTER(et);
1713 if (!(inp->inp_flags & INP_DROPPED)) {
1714 tp = intotcpcb(inp);
1715 CURVNET_SET(inp->inp_vnet);
1716 tp = tcp_drop(tp, ECONNABORTED);
1719 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1724 counter_u64_add(ktls_ifnet_reset_failed, 1);
1727 * Leave reset_pending true to avoid future tasks while
1728 * the socket goes away.
1736 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1738 struct ktls_session *tls;
1741 SOCKBUF_LOCK_ASSERT(sb);
1742 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1744 so = __containerof(sb, struct socket, so_rcv);
1746 tls = sb->sb_tls_info;
1747 if_rele(tls->rx_ifp);
1752 * See if we should schedule a task to update the receive tag for
1755 mtx_pool_lock(mtxpool_sleep, tls);
1756 if (!tls->reset_pending) {
1757 (void) ktls_hold(tls);
1760 tls->reset_pending = true;
1761 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1763 mtx_pool_unlock(mtxpool_sleep, tls);
1767 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1773 INP_LOCK_ASSERT(inp);
1776 * See if we should schedule a task to update the send tag for
1779 mtx_pool_lock(mtxpool_sleep, tls);
1780 if (!tls->reset_pending) {
1781 (void) ktls_hold(tls);
1782 tls->reset_pending = true;
1783 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1785 mtx_pool_unlock(mtxpool_sleep, tls);
1791 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1793 union if_snd_tag_modify_params params = {
1794 .rate_limit.max_rate = max_pacing_rate,
1795 .rate_limit.flags = M_NOWAIT,
1797 struct m_snd_tag *mst;
1799 /* Can't get to the inp, but it should be locked. */
1800 /* INP_LOCK_ASSERT(inp); */
1802 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1804 if (tls->snd_tag == NULL) {
1806 * Resetting send tag, ignore this change. The
1807 * pending reset may or may not see this updated rate
1808 * in the tcpcb. If it doesn't, we will just lose
1817 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1819 return (mst->sw->snd_tag_modify(mst, ¶ms));
1824 ktls_destroy_help(void *context, int pending __unused)
1826 ktls_destroy(context);
1830 ktls_destroy(struct ktls_session *tls)
1836 MPASS(tls->refcount == 0);
1840 wlocked = INP_WLOCKED(inp);
1841 if (!wlocked && !INP_TRY_WLOCK(inp)) {
1843 * rwlocks read locks are anonymous, and there
1844 * is no way to know if our current thread
1845 * holds an rlock on the inp. As a rough
1846 * estimate, check to see if the thread holds
1847 * *any* rlocks at all. If it does not, then we
1848 * know that we don't hold the inp rlock, and
1849 * can safely take the wlock
1851 if (curthread->td_rw_rlocks == 0) {
1855 * We might hold the rlock, so let's
1856 * do the destroy in a taskqueue
1857 * context to avoid a potential
1858 * deadlock. This should be very
1861 counter_u64_add(ktls_destroy_task, 1);
1862 TASK_INIT(&tls->destroy_task, 0,
1863 ktls_destroy_help, tls);
1864 (void)taskqueue_enqueue(taskqueue_thread,
1865 &tls->destroy_task);
1871 if (tls->sequential_records) {
1875 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1876 page_count = m->m_epg_enc_cnt;
1877 while (page_count > 0) {
1878 KASSERT(page_count >= m->m_epg_nrdy,
1879 ("%s: too few pages", __func__));
1880 page_count -= m->m_epg_nrdy;
1886 counter_u64_add(ktls_offload_active, -1);
1887 switch (tls->mode) {
1888 case TCP_TLS_MODE_SW:
1889 switch (tls->params.cipher_algorithm) {
1890 case CRYPTO_AES_CBC:
1891 counter_u64_add(ktls_sw_cbc, -1);
1893 case CRYPTO_AES_NIST_GCM_16:
1894 counter_u64_add(ktls_sw_gcm, -1);
1896 case CRYPTO_CHACHA20_POLY1305:
1897 counter_u64_add(ktls_sw_chacha20, -1);
1901 case TCP_TLS_MODE_IFNET:
1902 switch (tls->params.cipher_algorithm) {
1903 case CRYPTO_AES_CBC:
1904 counter_u64_add(ktls_ifnet_cbc, -1);
1906 case CRYPTO_AES_NIST_GCM_16:
1907 counter_u64_add(ktls_ifnet_gcm, -1);
1909 case CRYPTO_CHACHA20_POLY1305:
1910 counter_u64_add(ktls_ifnet_chacha20, -1);
1913 if (tls->snd_tag != NULL)
1914 m_snd_tag_rele(tls->snd_tag);
1915 if (tls->rx_ifp != NULL)
1916 if_rele(tls->rx_ifp);
1918 INP_WLOCK_ASSERT(inp);
1919 tp = intotcpcb(inp);
1920 MPASS(tp->t_nic_ktls_xmit == 1);
1921 tp->t_nic_ktls_xmit = 0;
1925 case TCP_TLS_MODE_TOE:
1926 switch (tls->params.cipher_algorithm) {
1927 case CRYPTO_AES_CBC:
1928 counter_u64_add(ktls_toe_cbc, -1);
1930 case CRYPTO_AES_NIST_GCM_16:
1931 counter_u64_add(ktls_toe_gcm, -1);
1933 case CRYPTO_CHACHA20_POLY1305:
1934 counter_u64_add(ktls_toe_chacha20, -1);
1940 if (tls->ocf_session != NULL)
1942 if (tls->params.auth_key != NULL) {
1943 zfree(tls->params.auth_key, M_KTLS);
1944 tls->params.auth_key = NULL;
1945 tls->params.auth_key_len = 0;
1947 if (tls->params.cipher_key != NULL) {
1948 zfree(tls->params.cipher_key, M_KTLS);
1949 tls->params.cipher_key = NULL;
1950 tls->params.cipher_key_len = 0;
1953 INP_WLOCK_ASSERT(inp);
1954 if (!in_pcbrele_wlocked(inp) && !wlocked)
1957 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1959 uma_zfree(ktls_session_zone, tls);
1963 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1966 for (; m != NULL; m = m->m_next) {
1967 KASSERT((m->m_flags & M_EXTPG) != 0,
1968 ("ktls_seq: mapped mbuf %p", m));
1970 m->m_epg_seqno = sb->sb_tls_seqno;
1976 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1977 * mbuf in the chain must be an unmapped mbuf. The payload of the
1978 * mbuf must be populated with the payload of each TLS record.
1980 * The record_type argument specifies the TLS record type used when
1981 * populating the TLS header.
1983 * The enq_count argument on return is set to the number of pages of
1984 * payload data for this entire chain that need to be encrypted via SW
1985 * encryption. The returned value should be passed to ktls_enqueue
1986 * when scheduling encryption of this chain of mbufs. To handle the
1987 * special case of empty fragments for TLS 1.0 sessions, an empty
1988 * fragment counts as one page.
1991 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1992 uint8_t record_type)
1994 struct tls_record_layer *tlshdr;
1998 int maxlen __diagused;
2000 maxlen = tls->params.max_frame_len;
2002 for (m = top; m != NULL; m = m->m_next) {
2004 * All mbufs in the chain should be TLS records whose
2005 * payload does not exceed the maximum frame length.
2007 * Empty TLS 1.0 records are permitted when using CBC.
2009 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2010 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2011 ("ktls_frame: m %p len %d", m, m->m_len));
2014 * TLS frames require unmapped mbufs to store session
2017 KASSERT((m->m_flags & M_EXTPG) != 0,
2018 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2022 /* Save a reference to the session. */
2023 m->m_epg_tls = ktls_hold(tls);
2025 m->m_epg_hdrlen = tls->params.tls_hlen;
2026 m->m_epg_trllen = tls->params.tls_tlen;
2027 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2031 * AES-CBC pads messages to a multiple of the
2032 * block size. Note that the padding is
2033 * applied after the digest and the encryption
2034 * is done on the "plaintext || mac || padding".
2035 * At least one byte of padding is always
2038 * Compute the final trailer length assuming
2039 * at most one block of padding.
2040 * tls->params.tls_tlen is the maximum
2041 * possible trailer length (padding + digest).
2042 * delta holds the number of excess padding
2043 * bytes if the maximum were used. Those
2044 * extra bytes are removed.
2046 bs = tls->params.tls_bs;
2047 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2048 m->m_epg_trllen -= delta;
2050 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2052 /* Populate the TLS header. */
2053 tlshdr = (void *)m->m_epg_hdr;
2054 tlshdr->tls_vmajor = tls->params.tls_vmajor;
2057 * TLS 1.3 masquarades as TLS 1.2 with a record type
2058 * of TLS_RLTYPE_APP.
2060 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2061 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2062 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2063 tlshdr->tls_type = TLS_RLTYPE_APP;
2064 /* save the real record type for later */
2065 m->m_epg_record_type = record_type;
2066 m->m_epg_trail[0] = record_type;
2068 tlshdr->tls_vminor = tls->params.tls_vminor;
2069 tlshdr->tls_type = record_type;
2071 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2074 * Store nonces / explicit IVs after the end of the
2077 * For GCM with TLS 1.2, an 8 byte nonce is copied
2078 * from the end of the IV. The nonce is then
2079 * incremented for use by the next record.
2081 * For CBC, a random nonce is inserted for TLS 1.1+.
2083 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2084 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2085 noncep = (uint64_t *)(tls->params.iv + 8);
2086 be64enc(tlshdr + 1, *noncep);
2088 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2089 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2090 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2093 * When using SW encryption, mark the mbuf not ready.
2094 * It will be marked ready via sbready() after the
2095 * record has been encrypted.
2097 * When using ifnet TLS, unencrypted TLS records are
2098 * sent down the stack to the NIC.
2100 if (tls->mode == TCP_TLS_MODE_SW) {
2101 m->m_flags |= M_NOTREADY;
2102 if (__predict_false(tls_len == 0)) {
2103 /* TLS 1.0 empty fragment. */
2106 m->m_epg_nrdy = m->m_epg_npgs;
2107 *enq_cnt += m->m_epg_nrdy;
2113 ktls_permit_empty_frames(struct ktls_session *tls)
2115 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2116 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2120 ktls_check_rx(struct sockbuf *sb)
2122 struct tls_record_layer hdr;
2127 SOCKBUF_LOCK_ASSERT(sb);
2128 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2130 so = __containerof(sb, struct socket, so_rcv);
2132 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2135 /* Is there enough queued for a TLS header? */
2136 if (sb->sb_tlscc < sizeof(hdr)) {
2137 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2138 so->so_error = EMSGSIZE;
2142 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2144 /* Is the entire record queued? */
2145 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2146 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2147 so->so_error = EMSGSIZE;
2151 sb->sb_flags |= SB_TLS_RX_RUNNING;
2154 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2156 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2157 running = wq->running;
2158 mtx_unlock(&wq->mtx);
2161 counter_u64_add(ktls_cnt_rx_queued, 1);
2164 static struct mbuf *
2165 ktls_detach_record(struct sockbuf *sb, int len)
2167 struct mbuf *m, *n, *top;
2170 SOCKBUF_LOCK_ASSERT(sb);
2171 MPASS(len <= sb->sb_tlscc);
2174 * If TLS chain is the exact size of the record,
2175 * just grab the whole record.
2178 if (sb->sb_tlscc == len) {
2180 sb->sb_mtlstail = NULL;
2185 * While it would be nice to use m_split() here, we need
2186 * to know exactly what m_split() allocates to update the
2187 * accounting, so do it inline instead.
2190 for (m = top; remain > m->m_len; m = m->m_next)
2193 /* Easy case: don't have to split 'm'. */
2194 if (remain == m->m_len) {
2195 sb->sb_mtls = m->m_next;
2196 if (sb->sb_mtls == NULL)
2197 sb->sb_mtlstail = NULL;
2203 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2204 * with M_NOWAIT first.
2206 n = m_get(M_NOWAIT, MT_DATA);
2209 * Use M_WAITOK with socket buffer unlocked. If
2210 * 'sb_mtls' changes while the lock is dropped, return
2211 * NULL to force the caller to retry.
2215 n = m_get(M_WAITOK, MT_DATA);
2218 if (sb->sb_mtls != top) {
2223 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2225 /* Store remainder in 'n'. */
2226 n->m_len = m->m_len - remain;
2227 if (m->m_flags & M_EXT) {
2228 n->m_data = m->m_data + remain;
2231 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2234 /* Trim 'm' and update accounting. */
2235 m->m_len -= n->m_len;
2236 sb->sb_tlscc -= n->m_len;
2237 sb->sb_ccc -= n->m_len;
2239 /* Account for 'n'. */
2240 sballoc_ktls_rx(sb, n);
2242 /* Insert 'n' into the TLS chain. */
2244 n->m_next = m->m_next;
2245 if (sb->sb_mtlstail == m)
2246 sb->sb_mtlstail = n;
2248 /* Detach the record from the TLS chain. */
2252 MPASS(m_length(top, NULL) == len);
2253 for (m = top; m != NULL; m = m->m_next)
2254 sbfree_ktls_rx(sb, m);
2255 sb->sb_tlsdcc = len;
2262 * Determine the length of the trailing zero padding and find the real
2263 * record type in the byte before the padding.
2265 * Walking the mbuf chain backwards is clumsy, so another option would
2266 * be to scan forwards remembering the last non-zero byte before the
2267 * trailer. However, it would be expensive to scan the entire record.
2268 * Instead, find the last non-zero byte of each mbuf in the chain
2269 * keeping track of the relative offset of that nonzero byte.
2271 * trail_len is the size of the MAC/tag on input and is set to the
2272 * size of the full trailer including padding and the record type on
2276 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2277 int *trailer_len, uint8_t *record_typep)
2280 u_int digest_start, last_offset, m_len, offset;
2281 uint8_t record_type;
2283 digest_start = tls_len - *trailer_len;
2286 for (; m != NULL && offset < digest_start;
2287 offset += m->m_len, m = m->m_next) {
2288 /* Don't look for padding in the tag. */
2289 m_len = min(digest_start - offset, m->m_len);
2290 cp = mtod(m, char *);
2292 /* Find last non-zero byte in this mbuf. */
2293 while (m_len > 0 && cp[m_len - 1] == 0)
2296 record_type = cp[m_len - 1];
2297 last_offset = offset + m_len;
2300 if (last_offset < tls->params.tls_hlen)
2303 *record_typep = record_type;
2304 *trailer_len = tls_len - last_offset + 1;
2309 * Check if a mbuf chain is fully decrypted at the given offset and
2310 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2311 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2312 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2315 ktls_mbuf_crypto_st_t
2316 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2318 int m_flags_ored = 0;
2319 int m_flags_anded = -1;
2321 for (; mb != NULL; mb = mb->m_next) {
2322 if (offset < mb->m_len)
2324 offset -= mb->m_len;
2328 for (; mb != NULL; mb = mb->m_next) {
2329 m_flags_ored |= mb->m_flags;
2330 m_flags_anded &= mb->m_flags;
2332 if (offset <= mb->m_len)
2334 offset -= mb->m_len;
2336 MPASS(mb != NULL || offset == 0);
2338 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2339 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2341 return ((m_flags_ored & M_DECRYPTED) ?
2342 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2343 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2347 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2350 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2352 union if_snd_tag_modify_params params;
2353 struct m_snd_tag *mst;
2357 mst = so->so_rcv.sb_tls_info->snd_tag;
2358 if (__predict_false(mst == NULL))
2361 inp = sotoinpcb(so);
2362 if (__predict_false(inp == NULL))
2366 if (inp->inp_flags & INP_DROPPED) {
2368 return (ECONNRESET);
2371 tp = intotcpcb(inp);
2374 /* Get the TCP sequence number of the next valid TLS header. */
2375 SOCKBUF_LOCK(&so->so_rcv);
2376 params.tls_rx.tls_hdr_tcp_sn =
2377 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2378 params.tls_rx.tls_rec_length = tls_len;
2379 params.tls_rx.tls_seq_number = tls_rcd_num;
2380 SOCKBUF_UNLOCK(&so->so_rcv);
2384 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2385 return (mst->sw->snd_tag_modify(mst, ¶ms));
2389 ktls_drop(struct socket *so, int error)
2391 struct epoch_tracker et;
2392 struct inpcb *inp = sotoinpcb(so);
2395 NET_EPOCH_ENTER(et);
2397 if (!(inp->inp_flags & INP_DROPPED)) {
2398 tp = intotcpcb(inp);
2399 CURVNET_SET(inp->inp_vnet);
2400 tp = tcp_drop(tp, error);
2405 so->so_error = error;
2406 SOCK_RECVBUF_LOCK(so);
2407 sorwakeup_locked(so);
2414 ktls_decrypt(struct socket *so)
2416 char tls_header[MBUF_PEXT_HDR_LEN];
2417 struct ktls_session *tls;
2419 struct tls_record_layer *hdr;
2420 struct tls_get_record tgr;
2421 struct mbuf *control, *data, *m;
2422 ktls_mbuf_crypto_st_t state;
2424 int error, remain, tls_len, trail_len;
2426 uint8_t vminor, record_type;
2428 hdr = (struct tls_record_layer *)tls_header;
2431 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2432 ("%s: socket %p not running", __func__, so));
2434 tls = sb->sb_tls_info;
2437 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2439 vminor = TLS_MINOR_VER_TWO;
2441 vminor = tls->params.tls_vminor;
2443 /* Is there enough queued for a TLS header? */
2444 if (sb->sb_tlscc < tls->params.tls_hlen)
2447 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2448 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2450 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2451 hdr->tls_vminor != vminor)
2453 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2455 else if (tls_len < tls->params.tls_hlen || tls_len >
2456 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2457 tls->params.tls_tlen)
2461 if (__predict_false(error != 0)) {
2463 * We have a corrupted record and are likely
2464 * out of sync. The connection isn't
2465 * recoverable at this point, so abort it.
2468 counter_u64_add(ktls_offload_corrupted_records, 1);
2470 ktls_drop(so, error);
2474 /* Is the entire record queued? */
2475 if (sb->sb_tlscc < tls_len)
2479 * Split out the portion of the mbuf chain containing
2482 data = ktls_detach_record(sb, tls_len);
2485 MPASS(sb->sb_tlsdcc == tls_len);
2487 seqno = sb->sb_tls_seqno;
2492 /* get crypto state for this TLS record */
2493 state = ktls_mbuf_crypto_state(data, 0, tls_len);
2496 case KTLS_MBUF_CRYPTO_ST_MIXED:
2497 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2501 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2502 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2504 if (__predict_true(error == 0)) {
2506 error = tls13_find_record_type(tls, data,
2507 tls_len, &trail_len, &record_type);
2509 record_type = hdr->tls_type;
2513 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2515 * NIC TLS is only supported for AEAD
2516 * ciphersuites which used a fixed sized
2520 trail_len = tls->params.tls_tlen - 1;
2521 error = tls13_find_record_type(tls, data,
2522 tls_len, &trail_len, &record_type);
2524 trail_len = tls->params.tls_tlen;
2526 record_type = hdr->tls_type;
2534 counter_u64_add(ktls_offload_failed_crypto, 1);
2537 if (sb->sb_tlsdcc == 0) {
2539 * sbcut/drop/flush discarded these
2547 * Drop this TLS record's data, but keep
2548 * decrypting subsequent records.
2550 sb->sb_ccc -= tls_len;
2553 if (error != EMSGSIZE)
2555 CURVNET_SET(so->so_vnet);
2556 so->so_error = error;
2557 sorwakeup_locked(so);
2566 /* Allocate the control mbuf. */
2567 memset(&tgr, 0, sizeof(tgr));
2568 tgr.tls_type = record_type;
2569 tgr.tls_vmajor = hdr->tls_vmajor;
2570 tgr.tls_vminor = hdr->tls_vminor;
2571 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2573 control = sbcreatecontrol(&tgr, sizeof(tgr),
2574 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2577 if (sb->sb_tlsdcc == 0) {
2578 /* sbcut/drop/flush discarded these mbufs. */
2579 MPASS(sb->sb_tlscc == 0);
2586 * Clear the 'dcc' accounting in preparation for
2587 * adding the decrypted record.
2589 sb->sb_ccc -= tls_len;
2593 /* If there is no payload, drop all of the data. */
2594 if (tgr.tls_length == htobe16(0)) {
2599 remain = tls->params.tls_hlen;
2600 while (remain > 0) {
2601 if (data->m_len > remain) {
2602 data->m_data += remain;
2603 data->m_len -= remain;
2606 remain -= data->m_len;
2607 data = m_free(data);
2610 /* Trim trailer and clear M_NOTREADY. */
2611 remain = be16toh(tgr.tls_length);
2613 for (m = data; remain > m->m_len; m = m->m_next) {
2614 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2620 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2622 /* Set EOR on the final mbuf. */
2623 m->m_flags |= M_EOR;
2626 sbappendcontrol_locked(sb, data, control, 0);
2628 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2629 sb->sb_flags |= SB_TLS_RX_RESYNC;
2631 ktls_resync_ifnet(so, tls_len, seqno);
2633 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2634 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2636 ktls_resync_ifnet(so, 0, seqno);
2641 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2643 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2644 so->so_error = EMSGSIZE;
2646 sorwakeup_locked(so);
2649 SOCKBUF_UNLOCK_ASSERT(sb);
2651 CURVNET_SET(so->so_vnet);
2657 ktls_enqueue_to_free(struct mbuf *m)
2662 /* Mark it for freeing. */
2663 m->m_epg_flags |= EPG_FLAG_2FREE;
2664 wq = &ktls_wq[m->m_epg_tls->wq_index];
2666 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2667 running = wq->running;
2668 mtx_unlock(&wq->mtx);
2674 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2677 int domain, running;
2679 if (m->m_epg_npgs <= 2)
2681 if (ktls_buffer_zone == NULL)
2683 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2685 * Rate-limit allocation attempts after a failure.
2686 * ktls_buffer_import() will acquire a per-domain mutex to check
2687 * the free page queues and may fail consistently if memory is
2692 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2694 domain = PCPU_GET(domain);
2695 wq->lastallocfail = ticks;
2698 * Note that this check is "racy", but the races are
2699 * harmless, and are either a spurious wakeup if
2700 * multiple threads fail allocations before the alloc
2701 * thread wakes, or waiting an extra second in case we
2702 * see an old value of running == true.
2704 if (!VM_DOMAIN_EMPTY(domain)) {
2705 running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2707 wakeup(&ktls_domains[domain].alloc_td);
2714 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2715 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2718 int error, i, len, off;
2720 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2721 ("%p not unready & nomap mbuf\n", m));
2722 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2723 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2726 /* Anonymous mbufs are encrypted in place. */
2727 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2728 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2731 * For file-backed mbufs (from sendfile), anonymous wired
2732 * pages are allocated and used as the encryption destination.
2734 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2735 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2737 state->dst_iov[0].iov_base = (char *)state->cbuf +
2739 state->dst_iov[0].iov_len = len;
2740 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2743 off = m->m_epg_1st_off;
2744 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2745 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2746 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2747 len = m_epg_pagelen(m, i, off);
2748 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2749 state->dst_iov[i].iov_base =
2750 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2751 state->dst_iov[i].iov_len = len;
2754 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2755 state->dst_iov[i].iov_base = m->m_epg_trail;
2756 state->dst_iov[i].iov_len = m->m_epg_trllen;
2758 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2760 if (__predict_false(error != 0)) {
2761 /* Free the anonymous pages. */
2762 if (state->cbuf != NULL)
2763 uma_zfree(ktls_buffer_zone, state->cbuf);
2765 for (i = 0; i < m->m_epg_npgs; i++) {
2766 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2767 (void)vm_page_unwire_noq(pg);
2775 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2777 ktls_batched_records(struct mbuf *m)
2779 int page_count, records;
2782 page_count = m->m_epg_enc_cnt;
2783 while (page_count > 0) {
2785 page_count -= m->m_epg_nrdy;
2788 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2793 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2795 struct ktls_session *tls;
2800 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2801 (M_EXTPG | M_NOTREADY)),
2802 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2803 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2805 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2807 m->m_epg_enc_cnt = page_count;
2810 * Save a pointer to the socket. The caller is responsible
2811 * for taking an additional reference via soref().
2817 wq = &ktls_wq[tls->wq_index];
2819 if (__predict_false(tls->sequential_records)) {
2821 * For TLS 1.0, records must be encrypted
2822 * sequentially. For a given connection, all records
2823 * queued to the associated work queue are processed
2824 * sequentially. However, sendfile(2) might complete
2825 * I/O requests spanning multiple TLS records out of
2826 * order. Here we ensure TLS records are enqueued to
2827 * the work queue in FIFO order.
2829 * tls->next_seqno holds the sequence number of the
2830 * next TLS record that should be enqueued to the work
2831 * queue. If this next record is not tls->next_seqno,
2832 * it must be a future record, so insert it, sorted by
2833 * TLS sequence number, into tls->pending_records and
2836 * If this TLS record matches tls->next_seqno, place
2837 * it in the work queue and then check
2838 * tls->pending_records to see if any
2839 * previously-queued records are now ready for
2842 if (m->m_epg_seqno != tls->next_seqno) {
2846 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2847 if (n->m_epg_seqno > m->m_epg_seqno)
2852 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2855 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2858 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2860 mtx_unlock(&wq->mtx);
2861 counter_u64_add(ktls_cnt_tx_pending, 1);
2865 tls->next_seqno += ktls_batched_records(m);
2866 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2868 while (!STAILQ_EMPTY(&tls->pending_records)) {
2871 n = STAILQ_FIRST(&tls->pending_records);
2872 if (n->m_epg_seqno != tls->next_seqno)
2876 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2877 tls->next_seqno += ktls_batched_records(n);
2878 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2880 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2882 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2884 running = wq->running;
2885 mtx_unlock(&wq->mtx);
2888 counter_u64_add(ktls_cnt_tx_queued, queued);
2892 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2893 * the pages from the file and replace them with the anonymous pages
2894 * allocated in ktls_encrypt_record().
2897 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2901 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2903 /* Free the old pages. */
2904 m->m_ext.ext_free(m);
2906 /* Replace them with the new pages. */
2907 if (state->cbuf != NULL) {
2908 for (i = 0; i < m->m_epg_npgs; i++)
2909 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2911 /* Contig pages should go back to the cache. */
2912 m->m_ext.ext_free = ktls_free_mext_contig;
2914 for (i = 0; i < m->m_epg_npgs; i++)
2915 m->m_epg_pa[i] = state->parray[i];
2917 /* Use the basic free routine. */
2918 m->m_ext.ext_free = mb_free_mext_pgs;
2921 /* Pages are now writable. */
2922 m->m_epg_flags |= EPG_FLAG_ANON;
2925 static __noinline void
2926 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2928 struct ktls_ocf_encrypt_state state;
2929 struct ktls_session *tls;
2932 int error, npages, total_pages;
2935 tls = top->m_epg_tls;
2936 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2937 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2939 top->m_epg_so = NULL;
2941 total_pages = top->m_epg_enc_cnt;
2945 * Encrypt the TLS records in the chain of mbufs starting with
2946 * 'top'. 'total_pages' gives us a total count of pages and is
2947 * used to know when we have finished encrypting the TLS
2948 * records originally queued with 'top'.
2950 * NB: These mbufs are queued in the socket buffer and
2951 * 'm_next' is traversing the mbufs in the socket buffer. The
2952 * socket buffer lock is not held while traversing this chain.
2953 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2954 * pointers should be stable. However, the 'm_next' of the
2955 * last mbuf encrypted is not necessarily NULL. It can point
2956 * to other mbufs appended while 'top' was on the TLS work
2959 * Each mbuf holds an entire TLS record.
2962 for (m = top; npages != total_pages; m = m->m_next) {
2963 KASSERT(m->m_epg_tls == tls,
2964 ("different TLS sessions in a single mbuf chain: %p vs %p",
2965 tls, m->m_epg_tls));
2966 KASSERT(npages + m->m_epg_npgs <= total_pages,
2967 ("page count mismatch: top %p, total_pages %d, m %p", top,
2970 error = ktls_encrypt_record(wq, m, tls, &state);
2972 counter_u64_add(ktls_offload_failed_crypto, 1);
2976 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2977 ktls_finish_nonanon(m, &state);
2979 npages += m->m_epg_nrdy;
2982 * Drop a reference to the session now that it is no
2983 * longer needed. Existing code depends on encrypted
2984 * records having no associated session vs
2985 * yet-to-be-encrypted records having an associated
2988 m->m_epg_tls = NULL;
2992 CURVNET_SET(so->so_vnet);
2994 (void)so->so_proto->pr_ready(so, top, npages);
2997 mb_free_notready(top, total_pages);
3005 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3007 struct ktls_session *tls;
3014 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3015 ktls_finish_nonanon(m, state);
3018 free(state, M_KTLS);
3021 * Drop a reference to the session now that it is no longer
3022 * needed. Existing code depends on encrypted records having
3023 * no associated session vs yet-to-be-encrypted records having
3024 * an associated session.
3027 m->m_epg_tls = NULL;
3031 counter_u64_add(ktls_offload_failed_crypto, 1);
3033 CURVNET_SET(so->so_vnet);
3034 npages = m->m_epg_nrdy;
3037 (void)so->so_proto->pr_ready(so, m, npages);
3040 mb_free_notready(m, npages);
3048 * Similar to ktls_encrypt, but used with asynchronous OCF backends
3049 * (coprocessors) where encryption does not use host CPU resources and
3050 * it can be beneficial to queue more requests than CPUs.
3052 static __noinline void
3053 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3055 struct ktls_ocf_encrypt_state *state;
3056 struct ktls_session *tls;
3059 int error, mpages, npages, total_pages;
3062 tls = top->m_epg_tls;
3063 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3064 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3066 top->m_epg_so = NULL;
3068 total_pages = top->m_epg_enc_cnt;
3072 for (m = top; npages != total_pages; m = n) {
3073 KASSERT(m->m_epg_tls == tls,
3074 ("different TLS sessions in a single mbuf chain: %p vs %p",
3075 tls, m->m_epg_tls));
3076 KASSERT(npages + m->m_epg_npgs <= total_pages,
3077 ("page count mismatch: top %p, total_pages %d, m %p", top,
3080 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3085 mpages = m->m_epg_nrdy;
3088 error = ktls_encrypt_record(wq, m, tls, state);
3090 counter_u64_add(ktls_offload_failed_crypto, 1);
3091 free(state, M_KTLS);
3092 CURVNET_SET(so->so_vnet);
3101 CURVNET_SET(so->so_vnet);
3104 mb_free_notready(m, total_pages - npages);
3112 ktls_bind_domain(int domain)
3116 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3119 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3124 ktls_alloc_thread(void *ctx)
3126 struct ktls_domain_info *ktls_domain = ctx;
3127 struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
3129 struct sysctl_oid *oid;
3131 int domain, error, i, nbufs;
3133 domain = ktls_domain - ktls_domains;
3135 printf("Starting KTLS alloc thread for domain %d\n", domain);
3136 error = ktls_bind_domain(domain);
3138 printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
3140 snprintf(name, sizeof(name), "domain%d", domain);
3141 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3142 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3143 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
3144 CTLFLAG_RD, &sc->allocs, 0, "buffers allocated");
3145 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3146 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3147 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3148 CTLFLAG_RD, &sc->running, 0, "thread running");
3153 atomic_store_int(&sc->running, 0);
3154 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3155 atomic_store_int(&sc->running, 1);
3157 if (nbufs != ktls_max_alloc) {
3159 nbufs = atomic_load_int(&ktls_max_alloc);
3160 buf = malloc(sizeof(void *) * nbufs, M_KTLS,
3164 * Below we allocate nbufs with different allocation
3165 * flags than we use when allocating normally during
3166 * encryption in the ktls worker thread. We specify
3167 * M_NORECLAIM in the worker thread. However, we omit
3168 * that flag here and add M_WAITOK so that the VM
3169 * system is permitted to perform expensive work to
3170 * defragment memory. We do this here, as it does not
3171 * matter if this thread blocks. If we block a ktls
3172 * worker thread, we risk developing backlogs of
3173 * buffers to be encrypted, leading to surges of
3174 * traffic and potential NIC output drops.
3176 for (i = 0; i < nbufs; i++) {
3177 buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
3180 for (i = 0; i < nbufs; i++) {
3181 uma_zfree(ktls_buffer_zone, buf[i]);
3188 ktls_work_thread(void *ctx)
3190 struct ktls_wq *wq = ctx;
3192 struct socket *so, *son;
3193 STAILQ_HEAD(, mbuf) local_m_head;
3194 STAILQ_HEAD(, socket) local_so_head;
3199 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3202 * Bind to a core. If ktls_bind_threads is > 1, then
3203 * we bind to the NUMA domain instead.
3205 if (ktls_bind_threads) {
3208 if (ktls_bind_threads > 1) {
3209 struct pcpu *pc = pcpu_find(cpu);
3211 error = ktls_bind_domain(pc->pc_domain);
3215 CPU_SETOF(cpu, &mask);
3216 error = cpuset_setthread(curthread->td_tid, &mask);
3219 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3222 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3227 while (STAILQ_EMPTY(&wq->m_head) &&
3228 STAILQ_EMPTY(&wq->so_head)) {
3229 wq->running = false;
3230 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3234 STAILQ_INIT(&local_m_head);
3235 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3236 STAILQ_INIT(&local_so_head);
3237 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3238 mtx_unlock(&wq->mtx);
3240 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3241 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3242 ktls_free(m->m_epg_tls);
3245 if (m->m_epg_tls->sync_dispatch)
3246 ktls_encrypt(wq, m);
3248 ktls_encrypt_async(wq, m);
3249 counter_u64_add(ktls_cnt_tx_queued, -1);
3253 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3255 counter_u64_add(ktls_cnt_rx_queued, -1);
3261 ktls_disable_ifnet_help(void *context, int pending __unused)
3263 struct ktls_session *tls;
3274 so = inp->inp_socket;
3276 if (inp->inp_flags & INP_DROPPED) {
3280 if (so->so_snd.sb_tls_info != NULL)
3281 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3285 counter_u64_add(ktls_ifnet_disable_ok, 1);
3286 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3287 if ((inp->inp_flags & INP_DROPPED) == 0 &&
3288 (tp = intotcpcb(inp)) != NULL &&
3289 tp->t_fb->tfb_hwtls_change != NULL)
3290 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3292 counter_u64_add(ktls_ifnet_disable_fail, 1);
3296 CURVNET_SET(so->so_vnet);
3304 * Called when re-transmits are becoming a substantial portion of the
3305 * sends on this connection. When this happens, we transition the
3306 * connection to software TLS. This is needed because most inline TLS
3307 * NICs keep crypto state only for in-order transmits. This means
3308 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3309 * re-DMA the entire TLS record up to and including the current
3310 * segment. This means that when re-transmitting the last ~1448 byte
3311 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3312 * of magnitude more data than we are sending. This can cause the
3313 * PCIe link to saturate well before the network, which can cause
3314 * output drops, and a general loss of capacity.
3317 ktls_disable_ifnet(void *arg)
3322 struct ktls_session *tls;
3325 inp = tptoinpcb(tp);
3326 INP_WLOCK_ASSERT(inp);
3327 so = inp->inp_socket;
3329 tls = so->so_snd.sb_tls_info;
3330 if (tp->t_nic_ktls_xmit_dis == 1) {
3336 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3337 * ifnet can only be done once per connection, so we never want
3341 (void)ktls_hold(tls);
3343 tp->t_nic_ktls_xmit_dis = 1;
3345 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3346 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);