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];
113 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114 "Kernel TLS offload");
115 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
116 "Kernel TLS offload stats");
119 static int ktls_bind_threads = 1;
121 static int ktls_bind_threads;
123 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
124 &ktls_bind_threads, 0,
125 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
127 static u_int ktls_maxlen = 16384;
128 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
129 &ktls_maxlen, 0, "Maximum TLS record size");
131 static int ktls_number_threads;
132 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
133 &ktls_number_threads, 0,
134 "Number of TLS threads in thread-pool");
136 unsigned int ktls_ifnet_max_rexmit_pct = 2;
137 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
138 &ktls_ifnet_max_rexmit_pct, 2,
139 "Max percent bytes retransmitted before ifnet TLS is disabled");
141 static bool ktls_offload_enable;
142 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
143 &ktls_offload_enable, 0,
144 "Enable support for kernel TLS offload");
146 static bool ktls_cbc_enable = true;
147 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
149 "Enable Support of AES-CBC crypto for kernel TLS");
151 static bool ktls_sw_buffer_cache = true;
152 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
153 &ktls_sw_buffer_cache, 1,
154 "Enable caching of output buffers for SW encryption");
156 static int ktls_max_alloc = 128;
157 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
158 &ktls_max_alloc, 128,
159 "Max number of 16k buffers to allocate in thread context");
161 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
162 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
163 &ktls_tasks_active, "Number of active tasks");
165 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
166 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
168 "Number of TLS records in queue to tasks for SW encryption");
170 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
171 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
173 "Number of TLS sockets in queue to tasks for SW decryption");
175 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
176 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
177 CTLFLAG_RD, &ktls_offload_total,
178 "Total successful TLS setups (parameters set)");
180 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
182 CTLFLAG_RD, &ktls_offload_enable_calls,
183 "Total number of TLS enable calls made");
185 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
186 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
187 &ktls_offload_active, "Total Active TLS sessions");
189 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
190 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
191 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
195 &ktls_offload_failed_crypto, "Total TLS crypto failures");
197 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
199 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
201 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
203 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
207 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
209 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
211 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
213 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
215 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
217 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
218 "Software TLS session stats");
219 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
220 "Hardware (ifnet) TLS session stats");
222 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
223 "TOE TLS session stats");
226 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
227 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
228 "Active number of software TLS sessions using AES-CBC");
230 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
231 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
232 "Active number of software TLS sessions using AES-GCM");
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
237 "Active number of software TLS sessions using Chacha20-Poly1305");
239 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
240 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
242 "Active number of ifnet TLS sessions using AES-CBC");
244 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
245 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
247 "Active number of ifnet TLS sessions using AES-GCM");
249 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
250 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
251 &ktls_ifnet_chacha20,
252 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
254 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
255 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
256 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
258 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
259 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
260 &ktls_ifnet_reset_dropped,
261 "TLS sessions dropped after failing to update ifnet send tag");
263 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
264 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
265 &ktls_ifnet_reset_failed,
266 "TLS sessions that failed to allocate a new ifnet send tag");
268 static int ktls_ifnet_permitted;
269 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
270 &ktls_ifnet_permitted, 1,
271 "Whether to permit hardware (ifnet) TLS sessions");
274 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
275 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
277 "Active number of TOE TLS sessions using AES-CBC");
279 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
280 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
282 "Active number of TOE TLS sessions using AES-GCM");
284 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
285 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
287 "Active number of TOE TLS sessions using Chacha20-Poly1305");
290 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
292 static void ktls_cleanup(struct ktls_session *tls);
293 #if defined(INET) || defined(INET6)
294 static void ktls_reset_send_tag(void *context, int pending);
296 static void ktls_work_thread(void *ctx);
297 static void ktls_alloc_thread(void *ctx);
299 #if defined(INET) || defined(INET6)
301 ktls_get_cpu(struct socket *so)
305 struct ktls_domain_info *di;
311 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
312 if (cpuid != NETISR_CPUID_NONE)
316 * Just use the flowid to shard connections in a repeatable
317 * fashion. Note that TLS 1.0 sessions rely on the
318 * serialization provided by having the same connection use
322 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
323 di = &ktls_domains[inp->inp_numa_domain];
324 cpuid = di->cpu[inp->inp_flowid % di->count];
327 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
333 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
338 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
339 ("%s: ktls max length %d is not page size-aligned",
340 __func__, ktls_maxlen));
342 for (i = 0; i < count; i++) {
343 m = vm_page_alloc_contig_domain(NULL, 0, domain,
344 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
345 VM_ALLOC_NODUMP | malloc2vm_flags(flags),
346 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
350 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
356 ktls_buffer_release(void *arg __unused, void **store, int count)
361 for (i = 0; i < count; i++) {
362 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
363 for (j = 0; j < atop(ktls_maxlen); j++) {
364 (void)vm_page_unwire_noq(m + j);
371 ktls_free_mext_contig(struct mbuf *m)
374 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
378 ktls_init(void *dummy __unused)
383 int count, domain, error, i;
385 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
388 ktls_session_zone = uma_zcreate("ktls_session",
389 sizeof(struct ktls_session),
390 NULL, NULL, NULL, NULL,
393 if (ktls_sw_buffer_cache) {
394 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
395 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
396 ktls_buffer_import, ktls_buffer_release, NULL,
397 UMA_ZONE_FIRSTTOUCH);
401 * Initialize the workqueues to run the TLS work. We create a
402 * work queue for each CPU.
405 STAILQ_INIT(&ktls_wq[i].m_head);
406 STAILQ_INIT(&ktls_wq[i].so_head);
407 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
408 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
409 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
411 panic("Can't add KTLS thread %d error %d", i, error);
414 * Bind threads to cores. If ktls_bind_threads is >
415 * 1, then we bind to the NUMA domain.
417 if (ktls_bind_threads) {
418 if (ktls_bind_threads > 1) {
420 domain = pc->pc_domain;
421 CPU_COPY(&cpuset_domain[domain], &mask);
422 count = ktls_domains[domain].count;
423 ktls_domains[domain].cpu[count] = i;
424 ktls_domains[domain].count++;
428 error = cpuset_setthread(td->td_tid, &mask);
431 "Unable to bind KTLS thread for CPU %d error %d",
434 ktls_cpuid_lookup[ktls_number_threads] = i;
435 ktls_number_threads++;
439 * Start an allocation thread per-domain to perform blocking allocations
440 * of 16k physically contiguous TLS crypto destination buffers.
442 if (ktls_sw_buffer_cache) {
443 for (domain = 0; domain < vm_ndomains; domain++) {
444 if (VM_DOMAIN_EMPTY(domain))
446 if (CPU_EMPTY(&cpuset_domain[domain]))
448 error = kproc_kthread_add(ktls_alloc_thread,
449 &ktls_domains[domain], &ktls_proc,
450 &ktls_domains[domain].alloc_td.td,
451 0, 0, "KTLS", "alloc_%d", domain);
453 panic("Can't add KTLS alloc thread %d error %d",
455 CPU_COPY(&cpuset_domain[domain], &mask);
456 error = cpuset_setthread(ktls_domains[domain].alloc_td.td->td_tid,
459 panic("Unable to bind KTLS alloc %d error %d",
465 * If we somehow have an empty domain, fall back to choosing
466 * among all KTLS threads.
468 if (ktls_bind_threads > 1) {
469 for (i = 0; i < vm_ndomains; i++) {
470 if (ktls_domains[i].count == 0) {
471 ktls_bind_threads = 1;
478 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
480 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL);
482 #if defined(INET) || defined(INET6)
484 ktls_create_session(struct socket *so, struct tls_enable *en,
485 struct ktls_session **tlsp)
487 struct ktls_session *tls;
490 /* Only TLS 1.0 - 1.3 are supported. */
491 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
493 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
494 en->tls_vminor > TLS_MINOR_VER_THREE)
497 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
499 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
501 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
504 /* All supported algorithms require a cipher key. */
505 if (en->cipher_key_len == 0)
508 /* No flags are currently supported. */
512 /* Common checks for supported algorithms. */
513 switch (en->cipher_algorithm) {
514 case CRYPTO_AES_NIST_GCM_16:
516 * auth_algorithm isn't used, but permit GMAC values
519 switch (en->auth_algorithm) {
521 #ifdef COMPAT_FREEBSD12
522 /* XXX: Really 13.0-current COMPAT. */
523 case CRYPTO_AES_128_NIST_GMAC:
524 case CRYPTO_AES_192_NIST_GMAC:
525 case CRYPTO_AES_256_NIST_GMAC:
531 if (en->auth_key_len != 0)
533 if ((en->tls_vminor == TLS_MINOR_VER_TWO &&
534 en->iv_len != TLS_AEAD_GCM_LEN) ||
535 (en->tls_vminor == TLS_MINOR_VER_THREE &&
536 en->iv_len != TLS_1_3_GCM_IV_LEN))
540 switch (en->auth_algorithm) {
541 case CRYPTO_SHA1_HMAC:
543 * TLS 1.0 requires an implicit IV. TLS 1.1+
544 * all use explicit IVs.
546 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
547 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
553 case CRYPTO_SHA2_256_HMAC:
554 case CRYPTO_SHA2_384_HMAC:
555 /* Ignore any supplied IV. */
561 if (en->auth_key_len == 0)
564 case CRYPTO_CHACHA20_POLY1305:
565 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
567 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
568 en->tls_vminor != TLS_MINOR_VER_THREE)
570 if (en->iv_len != TLS_CHACHA20_IV_LEN)
577 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
579 counter_u64_add(ktls_offload_active, 1);
581 refcount_init(&tls->refcount, 1);
582 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
584 tls->wq_index = ktls_get_cpu(so);
586 tls->params.cipher_algorithm = en->cipher_algorithm;
587 tls->params.auth_algorithm = en->auth_algorithm;
588 tls->params.tls_vmajor = en->tls_vmajor;
589 tls->params.tls_vminor = en->tls_vminor;
590 tls->params.flags = en->flags;
591 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
593 /* Set the header and trailer lengths. */
594 tls->params.tls_hlen = sizeof(struct tls_record_layer);
595 switch (en->cipher_algorithm) {
596 case CRYPTO_AES_NIST_GCM_16:
598 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
599 * nonce. TLS 1.3 uses a 12 byte implicit IV.
601 if (en->tls_vminor < TLS_MINOR_VER_THREE)
602 tls->params.tls_hlen += sizeof(uint64_t);
603 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
604 tls->params.tls_bs = 1;
607 switch (en->auth_algorithm) {
608 case CRYPTO_SHA1_HMAC:
609 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
610 /* Implicit IV, no nonce. */
612 tls->params.tls_hlen += AES_BLOCK_LEN;
614 tls->params.tls_tlen = AES_BLOCK_LEN +
617 case CRYPTO_SHA2_256_HMAC:
618 tls->params.tls_hlen += AES_BLOCK_LEN;
619 tls->params.tls_tlen = AES_BLOCK_LEN +
622 case CRYPTO_SHA2_384_HMAC:
623 tls->params.tls_hlen += AES_BLOCK_LEN;
624 tls->params.tls_tlen = AES_BLOCK_LEN +
628 panic("invalid hmac");
630 tls->params.tls_bs = AES_BLOCK_LEN;
632 case CRYPTO_CHACHA20_POLY1305:
634 * Chacha20 uses a 12 byte implicit IV.
636 tls->params.tls_tlen = POLY1305_HASH_LEN;
637 tls->params.tls_bs = 1;
640 panic("invalid cipher");
644 * TLS 1.3 includes optional padding which we do not support,
645 * and also puts the "real" record type at the end of the
648 if (en->tls_vminor == TLS_MINOR_VER_THREE)
649 tls->params.tls_tlen += sizeof(uint8_t);
651 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
652 ("TLS header length too long: %d", tls->params.tls_hlen));
653 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
654 ("TLS trailer length too long: %d", tls->params.tls_tlen));
656 if (en->auth_key_len != 0) {
657 tls->params.auth_key_len = en->auth_key_len;
658 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
660 error = copyin(en->auth_key, tls->params.auth_key,
666 tls->params.cipher_key_len = en->cipher_key_len;
667 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
668 error = copyin(en->cipher_key, tls->params.cipher_key,
674 * This holds the implicit portion of the nonce for AEAD
675 * ciphers and the initial implicit IV for TLS 1.0. The
676 * explicit portions of the IV are generated in ktls_frame().
678 if (en->iv_len != 0) {
679 tls->params.iv_len = en->iv_len;
680 error = copyin(en->iv, tls->params.iv, en->iv_len);
685 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
686 * counter to generate unique explicit IVs.
688 * Store this counter in the last 8 bytes of the IV
689 * array so that it is 8-byte aligned.
691 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
692 en->tls_vminor == TLS_MINOR_VER_TWO)
693 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
704 static struct ktls_session *
705 ktls_clone_session(struct ktls_session *tls)
707 struct ktls_session *tls_new;
709 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
711 counter_u64_add(ktls_offload_active, 1);
713 refcount_init(&tls_new->refcount, 1);
714 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
716 /* Copy fields from existing session. */
717 tls_new->params = tls->params;
718 tls_new->wq_index = tls->wq_index;
720 /* Deep copy keys. */
721 if (tls_new->params.auth_key != NULL) {
722 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
724 memcpy(tls_new->params.auth_key, tls->params.auth_key,
725 tls->params.auth_key_len);
728 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
730 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
731 tls->params.cipher_key_len);
738 ktls_cleanup(struct ktls_session *tls)
741 counter_u64_add(ktls_offload_active, -1);
743 case TCP_TLS_MODE_SW:
744 switch (tls->params.cipher_algorithm) {
746 counter_u64_add(ktls_sw_cbc, -1);
748 case CRYPTO_AES_NIST_GCM_16:
749 counter_u64_add(ktls_sw_gcm, -1);
751 case CRYPTO_CHACHA20_POLY1305:
752 counter_u64_add(ktls_sw_chacha20, -1);
757 case TCP_TLS_MODE_IFNET:
758 switch (tls->params.cipher_algorithm) {
760 counter_u64_add(ktls_ifnet_cbc, -1);
762 case CRYPTO_AES_NIST_GCM_16:
763 counter_u64_add(ktls_ifnet_gcm, -1);
765 case CRYPTO_CHACHA20_POLY1305:
766 counter_u64_add(ktls_ifnet_chacha20, -1);
769 if (tls->snd_tag != NULL)
770 m_snd_tag_rele(tls->snd_tag);
773 case TCP_TLS_MODE_TOE:
774 switch (tls->params.cipher_algorithm) {
776 counter_u64_add(ktls_toe_cbc, -1);
778 case CRYPTO_AES_NIST_GCM_16:
779 counter_u64_add(ktls_toe_gcm, -1);
781 case CRYPTO_CHACHA20_POLY1305:
782 counter_u64_add(ktls_toe_chacha20, -1);
788 if (tls->params.auth_key != NULL) {
789 zfree(tls->params.auth_key, M_KTLS);
790 tls->params.auth_key = NULL;
791 tls->params.auth_key_len = 0;
793 if (tls->params.cipher_key != NULL) {
794 zfree(tls->params.cipher_key, M_KTLS);
795 tls->params.cipher_key = NULL;
796 tls->params.cipher_key_len = 0;
798 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
801 #if defined(INET) || defined(INET6)
805 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
813 if (inp->inp_flags2 & INP_FREED) {
817 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
821 if (inp->inp_socket == NULL) {
826 if (!(tp->t_flags & TF_TOE)) {
831 error = tcp_offload_alloc_tls_session(tp, tls, direction);
834 tls->mode = TCP_TLS_MODE_TOE;
835 switch (tls->params.cipher_algorithm) {
837 counter_u64_add(ktls_toe_cbc, 1);
839 case CRYPTO_AES_NIST_GCM_16:
840 counter_u64_add(ktls_toe_gcm, 1);
842 case CRYPTO_CHACHA20_POLY1305:
843 counter_u64_add(ktls_toe_chacha20, 1);
852 * Common code used when first enabling ifnet TLS on a connection or
853 * when allocating a new ifnet TLS session due to a routing change.
854 * This function allocates a new TLS send tag on whatever interface
855 * the connection is currently routed over.
858 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
859 struct m_snd_tag **mstp)
861 union if_snd_tag_alloc_params params;
863 struct nhop_object *nh;
868 if (inp->inp_flags2 & INP_FREED) {
872 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
876 if (inp->inp_socket == NULL) {
883 * Check administrative controls on ifnet TLS to determine if
884 * ifnet TLS should be denied.
886 * - Always permit 'force' requests.
887 * - ktls_ifnet_permitted == 0: always deny.
889 if (!force && ktls_ifnet_permitted == 0) {
895 * XXX: Use the cached route in the inpcb to find the
896 * interface. This should perhaps instead use
897 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
898 * enabled after a connection has completed key negotiation in
899 * userland, the cached route will be present in practice.
901 nh = inp->inp_route.ro_nh;
910 * Allocate a TLS + ratelimit tag if the connection has an
911 * existing pacing rate.
913 if (tp->t_pacing_rate != -1 &&
914 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
915 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
916 params.tls_rate_limit.inp = inp;
917 params.tls_rate_limit.tls = tls;
918 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
920 params.hdr.type = IF_SND_TAG_TYPE_TLS;
921 params.tls.inp = inp;
922 params.tls.tls = tls;
924 params.hdr.flowid = inp->inp_flowid;
925 params.hdr.flowtype = inp->inp_flowtype;
926 params.hdr.numa_domain = inp->inp_numa_domain;
929 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
933 if (inp->inp_vflag & INP_IPV6) {
934 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
939 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
944 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
951 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
953 struct m_snd_tag *mst;
956 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
958 tls->mode = TCP_TLS_MODE_IFNET;
960 switch (tls->params.cipher_algorithm) {
962 counter_u64_add(ktls_ifnet_cbc, 1);
964 case CRYPTO_AES_NIST_GCM_16:
965 counter_u64_add(ktls_ifnet_gcm, 1);
967 case CRYPTO_CHACHA20_POLY1305:
968 counter_u64_add(ktls_ifnet_chacha20, 1);
976 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
980 error = ktls_ocf_try(so, tls, direction);
983 tls->mode = TCP_TLS_MODE_SW;
984 switch (tls->params.cipher_algorithm) {
986 counter_u64_add(ktls_sw_cbc, 1);
988 case CRYPTO_AES_NIST_GCM_16:
989 counter_u64_add(ktls_sw_gcm, 1);
991 case CRYPTO_CHACHA20_POLY1305:
992 counter_u64_add(ktls_sw_chacha20, 1);
999 * KTLS RX stores data in the socket buffer as a list of TLS records,
1000 * where each record is stored as a control message containg the TLS
1001 * header followed by data mbufs containing the decrypted data. This
1002 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1003 * both encrypted and decrypted data. TLS records decrypted by a NIC
1004 * should be queued to the socket buffer as records, but encrypted
1005 * data which needs to be decrypted by software arrives as a stream of
1006 * regular mbufs which need to be converted. In addition, there may
1007 * already be pending encrypted data in the socket buffer when KTLS RX
1010 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1013 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1015 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1016 * from the first mbuf. Once all of the data for that TLS record is
1017 * queued, the socket is queued to a worker thread.
1019 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1020 * the TLS chain. Each TLS record is detached from the TLS chain,
1021 * decrypted, and inserted into the regular socket buffer chain as
1022 * record starting with a control message holding the TLS header and
1023 * a chain of mbufs holding the encrypted data.
1027 sb_mark_notready(struct sockbuf *sb)
1034 sb->sb_mbtail = NULL;
1035 sb->sb_lastrecord = NULL;
1036 for (; m != NULL; m = m->m_next) {
1037 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1039 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1041 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1043 m->m_flags |= M_NOTREADY;
1044 sb->sb_acc -= m->m_len;
1045 sb->sb_tlscc += m->m_len;
1046 sb->sb_mtlstail = m;
1048 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1049 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1054 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1056 struct ktls_session *tls;
1059 if (!ktls_offload_enable)
1061 if (SOLISTENING(so))
1064 counter_u64_add(ktls_offload_enable_calls, 1);
1067 * This should always be true since only the TCP socket option
1068 * invokes this function.
1070 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1074 * XXX: Don't overwrite existing sessions. We should permit
1075 * this to support rekeying in the future.
1077 if (so->so_rcv.sb_tls_info != NULL)
1080 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1083 /* TLS 1.3 is not yet supported. */
1084 if (en->tls_vmajor == TLS_MAJOR_VER_ONE &&
1085 en->tls_vminor == TLS_MINOR_VER_THREE)
1088 error = ktls_create_session(so, en, &tls);
1093 error = ktls_try_toe(so, tls, KTLS_RX);
1096 error = ktls_try_sw(so, tls, KTLS_RX);
1103 /* Mark the socket as using TLS offload. */
1104 SOCKBUF_LOCK(&so->so_rcv);
1105 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1106 so->so_rcv.sb_tls_info = tls;
1107 so->so_rcv.sb_flags |= SB_TLS_RX;
1109 /* Mark existing data as not ready until it can be decrypted. */
1110 if (tls->mode != TCP_TLS_MODE_TOE) {
1111 sb_mark_notready(&so->so_rcv);
1112 ktls_check_rx(&so->so_rcv);
1114 SOCKBUF_UNLOCK(&so->so_rcv);
1116 counter_u64_add(ktls_offload_total, 1);
1122 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1124 struct ktls_session *tls;
1128 if (!ktls_offload_enable)
1130 if (SOLISTENING(so))
1133 counter_u64_add(ktls_offload_enable_calls, 1);
1136 * This should always be true since only the TCP socket option
1137 * invokes this function.
1139 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1143 * XXX: Don't overwrite existing sessions. We should permit
1144 * this to support rekeying in the future.
1146 if (so->so_snd.sb_tls_info != NULL)
1149 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1152 /* TLS requires ext pgs */
1153 if (mb_use_ext_pgs == 0)
1156 error = ktls_create_session(so, en, &tls);
1160 /* Prefer TOE -> ifnet TLS -> software TLS. */
1162 error = ktls_try_toe(so, tls, KTLS_TX);
1165 error = ktls_try_ifnet(so, tls, false);
1167 error = ktls_try_sw(so, tls, KTLS_TX);
1174 error = sblock(&so->so_snd, SBL_WAIT);
1181 * Write lock the INP when setting sb_tls_info so that
1182 * routines in tcp_ratelimit.c can read sb_tls_info while
1183 * holding the INP lock.
1187 SOCKBUF_LOCK(&so->so_snd);
1188 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1189 so->so_snd.sb_tls_info = tls;
1190 if (tls->mode != TCP_TLS_MODE_SW)
1191 so->so_snd.sb_flags |= SB_TLS_IFNET;
1192 SOCKBUF_UNLOCK(&so->so_snd);
1194 sbunlock(&so->so_snd);
1196 counter_u64_add(ktls_offload_total, 1);
1202 ktls_get_rx_mode(struct socket *so)
1204 struct ktls_session *tls;
1208 if (SOLISTENING(so))
1211 INP_WLOCK_ASSERT(inp);
1212 SOCKBUF_LOCK(&so->so_rcv);
1213 tls = so->so_rcv.sb_tls_info;
1215 mode = TCP_TLS_MODE_NONE;
1218 SOCKBUF_UNLOCK(&so->so_rcv);
1223 ktls_get_tx_mode(struct socket *so)
1225 struct ktls_session *tls;
1229 if (SOLISTENING(so))
1232 INP_WLOCK_ASSERT(inp);
1233 SOCKBUF_LOCK(&so->so_snd);
1234 tls = so->so_snd.sb_tls_info;
1236 mode = TCP_TLS_MODE_NONE;
1239 SOCKBUF_UNLOCK(&so->so_snd);
1244 * Switch between SW and ifnet TLS sessions as requested.
1247 ktls_set_tx_mode(struct socket *so, int mode)
1249 struct ktls_session *tls, *tls_new;
1253 if (SOLISTENING(so))
1256 case TCP_TLS_MODE_SW:
1257 case TCP_TLS_MODE_IFNET:
1264 INP_WLOCK_ASSERT(inp);
1265 SOCKBUF_LOCK(&so->so_snd);
1266 tls = so->so_snd.sb_tls_info;
1268 SOCKBUF_UNLOCK(&so->so_snd);
1272 if (tls->mode == mode) {
1273 SOCKBUF_UNLOCK(&so->so_snd);
1277 tls = ktls_hold(tls);
1278 SOCKBUF_UNLOCK(&so->so_snd);
1281 tls_new = ktls_clone_session(tls);
1283 if (mode == TCP_TLS_MODE_IFNET)
1284 error = ktls_try_ifnet(so, tls_new, true);
1286 error = ktls_try_sw(so, tls_new, KTLS_TX);
1288 counter_u64_add(ktls_switch_failed, 1);
1295 error = sblock(&so->so_snd, SBL_WAIT);
1297 counter_u64_add(ktls_switch_failed, 1);
1305 * If we raced with another session change, keep the existing
1308 if (tls != so->so_snd.sb_tls_info) {
1309 counter_u64_add(ktls_switch_failed, 1);
1310 sbunlock(&so->so_snd);
1317 SOCKBUF_LOCK(&so->so_snd);
1318 so->so_snd.sb_tls_info = tls_new;
1319 if (tls_new->mode != TCP_TLS_MODE_SW)
1320 so->so_snd.sb_flags |= SB_TLS_IFNET;
1321 SOCKBUF_UNLOCK(&so->so_snd);
1322 sbunlock(&so->so_snd);
1325 * Drop two references on 'tls'. The first is for the
1326 * ktls_hold() above. The second drops the reference from the
1329 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1333 if (mode == TCP_TLS_MODE_IFNET)
1334 counter_u64_add(ktls_switch_to_ifnet, 1);
1336 counter_u64_add(ktls_switch_to_sw, 1);
1343 * Try to allocate a new TLS send tag. This task is scheduled when
1344 * ip_output detects a route change while trying to transmit a packet
1345 * holding a TLS record. If a new tag is allocated, replace the tag
1346 * in the TLS session. Subsequent packets on the connection will use
1347 * the new tag. If a new tag cannot be allocated, drop the
1351 ktls_reset_send_tag(void *context, int pending)
1353 struct epoch_tracker et;
1354 struct ktls_session *tls;
1355 struct m_snd_tag *old, *new;
1360 MPASS(pending == 1);
1366 * Free the old tag first before allocating a new one.
1367 * ip[6]_output_send() will treat a NULL send tag the same as
1368 * an ifp mismatch and drop packets until a new tag is
1371 * Write-lock the INP when changing tls->snd_tag since
1372 * ip[6]_output_send() holds a read-lock when reading the
1377 tls->snd_tag = NULL;
1380 m_snd_tag_rele(old);
1382 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1387 mtx_pool_lock(mtxpool_sleep, tls);
1388 tls->reset_pending = false;
1389 mtx_pool_unlock(mtxpool_sleep, tls);
1390 if (!in_pcbrele_wlocked(inp))
1393 counter_u64_add(ktls_ifnet_reset, 1);
1396 * XXX: Should we kick tcp_output explicitly now that
1397 * the send tag is fixed or just rely on timers?
1400 NET_EPOCH_ENTER(et);
1402 if (!in_pcbrele_wlocked(inp)) {
1403 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1404 !(inp->inp_flags & INP_DROPPED)) {
1405 tp = intotcpcb(inp);
1406 CURVNET_SET(tp->t_vnet);
1407 tp = tcp_drop(tp, ECONNABORTED);
1411 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1417 counter_u64_add(ktls_ifnet_reset_failed, 1);
1420 * Leave reset_pending true to avoid future tasks while
1421 * the socket goes away.
1429 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1435 INP_LOCK_ASSERT(inp);
1438 * See if we should schedule a task to update the send tag for
1441 mtx_pool_lock(mtxpool_sleep, tls);
1442 if (!tls->reset_pending) {
1443 (void) ktls_hold(tls);
1446 tls->reset_pending = true;
1447 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1449 mtx_pool_unlock(mtxpool_sleep, tls);
1455 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1457 union if_snd_tag_modify_params params = {
1458 .rate_limit.max_rate = max_pacing_rate,
1459 .rate_limit.flags = M_NOWAIT,
1461 struct m_snd_tag *mst;
1464 /* Can't get to the inp, but it should be locked. */
1465 /* INP_LOCK_ASSERT(inp); */
1467 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1469 if (tls->snd_tag == NULL) {
1471 * Resetting send tag, ignore this change. The
1472 * pending reset may or may not see this updated rate
1473 * in the tcpcb. If it doesn't, we will just lose
1479 MPASS(tls->snd_tag != NULL);
1480 MPASS(tls->snd_tag->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1484 return (ifp->if_snd_tag_modify(mst, ¶ms));
1490 ktls_destroy(struct ktls_session *tls)
1494 uma_zfree(ktls_session_zone, tls);
1498 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1501 for (; m != NULL; m = m->m_next) {
1502 KASSERT((m->m_flags & M_EXTPG) != 0,
1503 ("ktls_seq: mapped mbuf %p", m));
1505 m->m_epg_seqno = sb->sb_tls_seqno;
1511 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1512 * mbuf in the chain must be an unmapped mbuf. The payload of the
1513 * mbuf must be populated with the payload of each TLS record.
1515 * The record_type argument specifies the TLS record type used when
1516 * populating the TLS header.
1518 * The enq_count argument on return is set to the number of pages of
1519 * payload data for this entire chain that need to be encrypted via SW
1520 * encryption. The returned value should be passed to ktls_enqueue
1521 * when scheduling encryption of this chain of mbufs. To handle the
1522 * special case of empty fragments for TLS 1.0 sessions, an empty
1523 * fragment counts as one page.
1526 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1527 uint8_t record_type)
1529 struct tls_record_layer *tlshdr;
1535 maxlen = tls->params.max_frame_len;
1537 for (m = top; m != NULL; m = m->m_next) {
1539 * All mbufs in the chain should be TLS records whose
1540 * payload does not exceed the maximum frame length.
1542 * Empty TLS records are permitted when using CBC.
1544 KASSERT(m->m_len <= maxlen &&
1545 (tls->params.cipher_algorithm == CRYPTO_AES_CBC ?
1546 m->m_len >= 0 : m->m_len > 0),
1547 ("ktls_frame: m %p len %d\n", m, m->m_len));
1550 * TLS frames require unmapped mbufs to store session
1553 KASSERT((m->m_flags & M_EXTPG) != 0,
1554 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top));
1558 /* Save a reference to the session. */
1559 m->m_epg_tls = ktls_hold(tls);
1561 m->m_epg_hdrlen = tls->params.tls_hlen;
1562 m->m_epg_trllen = tls->params.tls_tlen;
1563 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1567 * AES-CBC pads messages to a multiple of the
1568 * block size. Note that the padding is
1569 * applied after the digest and the encryption
1570 * is done on the "plaintext || mac || padding".
1571 * At least one byte of padding is always
1574 * Compute the final trailer length assuming
1575 * at most one block of padding.
1576 * tls->params.tls_tlen is the maximum
1577 * possible trailer length (padding + digest).
1578 * delta holds the number of excess padding
1579 * bytes if the maximum were used. Those
1580 * extra bytes are removed.
1582 bs = tls->params.tls_bs;
1583 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1584 m->m_epg_trllen -= delta;
1586 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1588 /* Populate the TLS header. */
1589 tlshdr = (void *)m->m_epg_hdr;
1590 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1593 * TLS 1.3 masquarades as TLS 1.2 with a record type
1594 * of TLS_RLTYPE_APP.
1596 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1597 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1598 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1599 tlshdr->tls_type = TLS_RLTYPE_APP;
1600 /* save the real record type for later */
1601 m->m_epg_record_type = record_type;
1602 m->m_epg_trail[0] = record_type;
1604 tlshdr->tls_vminor = tls->params.tls_vminor;
1605 tlshdr->tls_type = record_type;
1607 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1610 * Store nonces / explicit IVs after the end of the
1613 * For GCM with TLS 1.2, an 8 byte nonce is copied
1614 * from the end of the IV. The nonce is then
1615 * incremented for use by the next record.
1617 * For CBC, a random nonce is inserted for TLS 1.1+.
1619 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1620 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1621 noncep = (uint64_t *)(tls->params.iv + 8);
1622 be64enc(tlshdr + 1, *noncep);
1624 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1625 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1626 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1629 * When using SW encryption, mark the mbuf not ready.
1630 * It will be marked ready via sbready() after the
1631 * record has been encrypted.
1633 * When using ifnet TLS, unencrypted TLS records are
1634 * sent down the stack to the NIC.
1636 if (tls->mode == TCP_TLS_MODE_SW) {
1637 m->m_flags |= M_NOTREADY;
1638 if (__predict_false(tls_len == 0)) {
1639 /* TLS 1.0 empty fragment. */
1642 m->m_epg_nrdy = m->m_epg_npgs;
1643 *enq_cnt += m->m_epg_nrdy;
1649 ktls_check_rx(struct sockbuf *sb)
1651 struct tls_record_layer hdr;
1656 SOCKBUF_LOCK_ASSERT(sb);
1657 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1659 so = __containerof(sb, struct socket, so_rcv);
1661 if (sb->sb_flags & SB_TLS_RX_RUNNING)
1664 /* Is there enough queued for a TLS header? */
1665 if (sb->sb_tlscc < sizeof(hdr)) {
1666 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1667 so->so_error = EMSGSIZE;
1671 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1673 /* Is the entire record queued? */
1674 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1675 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1676 so->so_error = EMSGSIZE;
1680 sb->sb_flags |= SB_TLS_RX_RUNNING;
1683 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1685 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1686 running = wq->running;
1687 mtx_unlock(&wq->mtx);
1690 counter_u64_add(ktls_cnt_rx_queued, 1);
1693 static struct mbuf *
1694 ktls_detach_record(struct sockbuf *sb, int len)
1696 struct mbuf *m, *n, *top;
1699 SOCKBUF_LOCK_ASSERT(sb);
1700 MPASS(len <= sb->sb_tlscc);
1703 * If TLS chain is the exact size of the record,
1704 * just grab the whole record.
1707 if (sb->sb_tlscc == len) {
1709 sb->sb_mtlstail = NULL;
1714 * While it would be nice to use m_split() here, we need
1715 * to know exactly what m_split() allocates to update the
1716 * accounting, so do it inline instead.
1719 for (m = top; remain > m->m_len; m = m->m_next)
1722 /* Easy case: don't have to split 'm'. */
1723 if (remain == m->m_len) {
1724 sb->sb_mtls = m->m_next;
1725 if (sb->sb_mtls == NULL)
1726 sb->sb_mtlstail = NULL;
1732 * Need to allocate an mbuf to hold the remainder of 'm'. Try
1733 * with M_NOWAIT first.
1735 n = m_get(M_NOWAIT, MT_DATA);
1738 * Use M_WAITOK with socket buffer unlocked. If
1739 * 'sb_mtls' changes while the lock is dropped, return
1740 * NULL to force the caller to retry.
1744 n = m_get(M_WAITOK, MT_DATA);
1747 if (sb->sb_mtls != top) {
1752 n->m_flags |= M_NOTREADY;
1754 /* Store remainder in 'n'. */
1755 n->m_len = m->m_len - remain;
1756 if (m->m_flags & M_EXT) {
1757 n->m_data = m->m_data + remain;
1760 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1763 /* Trim 'm' and update accounting. */
1764 m->m_len -= n->m_len;
1765 sb->sb_tlscc -= n->m_len;
1766 sb->sb_ccc -= n->m_len;
1768 /* Account for 'n'. */
1769 sballoc_ktls_rx(sb, n);
1771 /* Insert 'n' into the TLS chain. */
1773 n->m_next = m->m_next;
1774 if (sb->sb_mtlstail == m)
1775 sb->sb_mtlstail = n;
1777 /* Detach the record from the TLS chain. */
1781 MPASS(m_length(top, NULL) == len);
1782 for (m = top; m != NULL; m = m->m_next)
1783 sbfree_ktls_rx(sb, m);
1784 sb->sb_tlsdcc = len;
1791 ktls_decrypt(struct socket *so)
1793 char tls_header[MBUF_PEXT_HDR_LEN];
1794 struct ktls_session *tls;
1796 struct tls_record_layer *hdr;
1797 struct tls_get_record tgr;
1798 struct mbuf *control, *data, *m;
1800 int error, remain, tls_len, trail_len;
1802 hdr = (struct tls_record_layer *)tls_header;
1805 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
1806 ("%s: socket %p not running", __func__, so));
1808 tls = sb->sb_tls_info;
1812 /* Is there enough queued for a TLS header? */
1813 if (sb->sb_tlscc < tls->params.tls_hlen)
1816 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
1817 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
1819 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
1820 hdr->tls_vminor != tls->params.tls_vminor)
1822 else if (tls_len < tls->params.tls_hlen || tls_len >
1823 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
1824 tls->params.tls_tlen)
1828 if (__predict_false(error != 0)) {
1830 * We have a corrupted record and are likely
1831 * out of sync. The connection isn't
1832 * recoverable at this point, so abort it.
1835 counter_u64_add(ktls_offload_corrupted_records, 1);
1837 CURVNET_SET(so->so_vnet);
1838 so->so_proto->pr_usrreqs->pru_abort(so);
1839 so->so_error = error;
1844 /* Is the entire record queued? */
1845 if (sb->sb_tlscc < tls_len)
1849 * Split out the portion of the mbuf chain containing
1852 data = ktls_detach_record(sb, tls_len);
1855 MPASS(sb->sb_tlsdcc == tls_len);
1857 seqno = sb->sb_tls_seqno;
1862 error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
1864 counter_u64_add(ktls_offload_failed_crypto, 1);
1867 if (sb->sb_tlsdcc == 0) {
1869 * sbcut/drop/flush discarded these
1877 * Drop this TLS record's data, but keep
1878 * decrypting subsequent records.
1880 sb->sb_ccc -= tls_len;
1883 CURVNET_SET(so->so_vnet);
1884 so->so_error = EBADMSG;
1885 sorwakeup_locked(so);
1894 /* Allocate the control mbuf. */
1895 tgr.tls_type = hdr->tls_type;
1896 tgr.tls_vmajor = hdr->tls_vmajor;
1897 tgr.tls_vminor = hdr->tls_vminor;
1898 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
1900 control = sbcreatecontrol_how(&tgr, sizeof(tgr),
1901 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
1904 if (sb->sb_tlsdcc == 0) {
1905 /* sbcut/drop/flush discarded these mbufs. */
1906 MPASS(sb->sb_tlscc == 0);
1913 * Clear the 'dcc' accounting in preparation for
1914 * adding the decrypted record.
1916 sb->sb_ccc -= tls_len;
1920 /* If there is no payload, drop all of the data. */
1921 if (tgr.tls_length == htobe16(0)) {
1926 remain = tls->params.tls_hlen;
1927 while (remain > 0) {
1928 if (data->m_len > remain) {
1929 data->m_data += remain;
1930 data->m_len -= remain;
1933 remain -= data->m_len;
1934 data = m_free(data);
1937 /* Trim trailer and clear M_NOTREADY. */
1938 remain = be16toh(tgr.tls_length);
1940 for (m = data; remain > m->m_len; m = m->m_next) {
1941 m->m_flags &= ~M_NOTREADY;
1947 m->m_flags &= ~M_NOTREADY;
1949 /* Set EOR on the final mbuf. */
1950 m->m_flags |= M_EOR;
1953 sbappendcontrol_locked(sb, data, control, 0);
1956 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
1958 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
1959 so->so_error = EMSGSIZE;
1961 sorwakeup_locked(so);
1964 SOCKBUF_UNLOCK_ASSERT(sb);
1966 CURVNET_SET(so->so_vnet);
1973 ktls_enqueue_to_free(struct mbuf *m)
1978 /* Mark it for freeing. */
1979 m->m_epg_flags |= EPG_FLAG_2FREE;
1980 wq = &ktls_wq[m->m_epg_tls->wq_index];
1982 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
1983 running = wq->running;
1984 mtx_unlock(&wq->mtx);
1990 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
1993 int domain, running;
1995 if (m->m_epg_npgs <= 2)
1997 if (ktls_buffer_zone == NULL)
1999 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2001 * Rate-limit allocation attempts after a failure.
2002 * ktls_buffer_import() will acquire a per-domain mutex to check
2003 * the free page queues and may fail consistently if memory is
2008 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2010 domain = PCPU_GET(domain);
2011 wq->lastallocfail = ticks;
2014 * Note that this check is "racy", but the races are
2015 * harmless, and are either a spurious wakeup if
2016 * multiple threads fail allocations before the alloc
2017 * thread wakes, or waiting an extra second in case we
2018 * see an old value of running == true.
2020 if (!VM_DOMAIN_EMPTY(domain)) {
2021 running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2023 wakeup(&ktls_domains[domain].alloc_td);
2030 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2031 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2034 int error, i, len, off;
2036 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2037 ("%p not unready & nomap mbuf\n", m));
2038 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2039 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2042 /* Anonymous mbufs are encrypted in place. */
2043 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2044 return (tls->sw_encrypt(state, tls, m, NULL, 0));
2047 * For file-backed mbufs (from sendfile), anonymous wired
2048 * pages are allocated and used as the encryption destination.
2050 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2051 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2053 state->dst_iov[0].iov_base = (char *)state->cbuf +
2055 state->dst_iov[0].iov_len = len;
2056 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2059 off = m->m_epg_1st_off;
2060 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2062 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2063 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP |
2064 VM_ALLOC_WIRED | VM_ALLOC_WAITFAIL);
2065 } while (pg == NULL);
2067 len = m_epg_pagelen(m, i, off);
2068 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2069 state->dst_iov[i].iov_base =
2070 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2071 state->dst_iov[i].iov_len = len;
2074 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2075 state->dst_iov[i].iov_base = m->m_epg_trail;
2076 state->dst_iov[i].iov_len = m->m_epg_trllen;
2078 error = tls->sw_encrypt(state, tls, m, state->dst_iov, i + 1);
2080 if (__predict_false(error != 0)) {
2081 /* Free the anonymous pages. */
2082 if (state->cbuf != NULL)
2083 uma_zfree(ktls_buffer_zone, state->cbuf);
2085 for (i = 0; i < m->m_epg_npgs; i++) {
2086 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2087 (void)vm_page_unwire_noq(pg);
2096 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2101 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2102 (M_EXTPG | M_NOTREADY)),
2103 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2104 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2106 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2108 m->m_epg_enc_cnt = page_count;
2111 * Save a pointer to the socket. The caller is responsible
2112 * for taking an additional reference via soref().
2116 wq = &ktls_wq[m->m_epg_tls->wq_index];
2118 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2119 running = wq->running;
2120 mtx_unlock(&wq->mtx);
2123 counter_u64_add(ktls_cnt_tx_queued, 1);
2127 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2128 * the pages from the file and replace them with the anonymous pages
2129 * allocated in ktls_encrypt_record().
2132 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2136 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2138 /* Free the old pages. */
2139 m->m_ext.ext_free(m);
2141 /* Replace them with the new pages. */
2142 if (state->cbuf != NULL) {
2143 for (i = 0; i < m->m_epg_npgs; i++)
2144 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2146 /* Contig pages should go back to the cache. */
2147 m->m_ext.ext_free = ktls_free_mext_contig;
2149 for (i = 0; i < m->m_epg_npgs; i++)
2150 m->m_epg_pa[i] = state->parray[i];
2152 /* Use the basic free routine. */
2153 m->m_ext.ext_free = mb_free_mext_pgs;
2156 /* Pages are now writable. */
2157 m->m_epg_flags |= EPG_FLAG_ANON;
2160 static __noinline void
2161 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2163 struct ktls_ocf_encrypt_state state;
2164 struct ktls_session *tls;
2167 int error, npages, total_pages;
2170 tls = top->m_epg_tls;
2171 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2172 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2174 top->m_epg_so = NULL;
2176 total_pages = top->m_epg_enc_cnt;
2180 * Encrypt the TLS records in the chain of mbufs starting with
2181 * 'top'. 'total_pages' gives us a total count of pages and is
2182 * used to know when we have finished encrypting the TLS
2183 * records originally queued with 'top'.
2185 * NB: These mbufs are queued in the socket buffer and
2186 * 'm_next' is traversing the mbufs in the socket buffer. The
2187 * socket buffer lock is not held while traversing this chain.
2188 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2189 * pointers should be stable. However, the 'm_next' of the
2190 * last mbuf encrypted is not necessarily NULL. It can point
2191 * to other mbufs appended while 'top' was on the TLS work
2194 * Each mbuf holds an entire TLS record.
2197 for (m = top; npages != total_pages; m = m->m_next) {
2198 KASSERT(m->m_epg_tls == tls,
2199 ("different TLS sessions in a single mbuf chain: %p vs %p",
2200 tls, m->m_epg_tls));
2201 KASSERT(npages + m->m_epg_npgs <= total_pages,
2202 ("page count mismatch: top %p, total_pages %d, m %p", top,
2205 error = ktls_encrypt_record(wq, m, tls, &state);
2207 counter_u64_add(ktls_offload_failed_crypto, 1);
2211 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2212 ktls_finish_nonanon(m, &state);
2214 npages += m->m_epg_nrdy;
2217 * Drop a reference to the session now that it is no
2218 * longer needed. Existing code depends on encrypted
2219 * records having no associated session vs
2220 * yet-to-be-encrypted records having an associated
2223 m->m_epg_tls = NULL;
2227 CURVNET_SET(so->so_vnet);
2229 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2231 so->so_proto->pr_usrreqs->pru_abort(so);
2233 mb_free_notready(top, total_pages);
2242 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2244 struct ktls_session *tls;
2251 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2252 ktls_finish_nonanon(m, state);
2255 free(state, M_KTLS);
2258 * Drop a reference to the session now that it is no longer
2259 * needed. Existing code depends on encrypted records having
2260 * no associated session vs yet-to-be-encrypted records having
2261 * an associated session.
2264 m->m_epg_tls = NULL;
2268 counter_u64_add(ktls_offload_failed_crypto, 1);
2270 CURVNET_SET(so->so_vnet);
2271 npages = m->m_epg_nrdy;
2274 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, m, npages);
2276 so->so_proto->pr_usrreqs->pru_abort(so);
2278 mb_free_notready(m, npages);
2287 * Similar to ktls_encrypt, but used with asynchronous OCF backends
2288 * (coprocessors) where encryption does not use host CPU resources and
2289 * it can be beneficial to queue more requests than CPUs.
2291 static __noinline void
2292 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2294 struct ktls_ocf_encrypt_state *state;
2295 struct ktls_session *tls;
2298 int error, mpages, npages, total_pages;
2301 tls = top->m_epg_tls;
2302 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2303 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2305 top->m_epg_so = NULL;
2307 total_pages = top->m_epg_enc_cnt;
2311 for (m = top; npages != total_pages; m = n) {
2312 KASSERT(m->m_epg_tls == tls,
2313 ("different TLS sessions in a single mbuf chain: %p vs %p",
2314 tls, m->m_epg_tls));
2315 KASSERT(npages + m->m_epg_npgs <= total_pages,
2316 ("page count mismatch: top %p, total_pages %d, m %p", top,
2319 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2324 mpages = m->m_epg_nrdy;
2327 error = ktls_encrypt_record(wq, m, tls, state);
2329 counter_u64_add(ktls_offload_failed_crypto, 1);
2330 free(state, M_KTLS);
2331 CURVNET_SET(so->so_vnet);
2341 CURVNET_SET(so->so_vnet);
2343 so->so_proto->pr_usrreqs->pru_abort(so);
2345 mb_free_notready(m, total_pages - npages);
2354 ktls_alloc_thread(void *ctx)
2356 struct ktls_domain_info *ktls_domain = ctx;
2357 struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
2359 struct sysctl_oid *oid;
2363 curthread->td_domain.dr_policy =
2364 DOMAINSET_PREF(PCPU_GET(domain));
2365 snprintf(name, sizeof(name), "domain%d", PCPU_GET(domain));
2367 printf("Starting KTLS alloc thread for domain %d\n",
2369 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
2370 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2371 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
2372 CTLFLAG_RD, &sc->allocs, 0, "buffers allocated");
2373 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
2374 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
2375 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
2376 CTLFLAG_RD, &sc->running, 0, "thread running");
2381 atomic_store_int(&sc->running, 0);
2382 tsleep(sc, PZERO | PNOLOCK, "-", 0);
2383 atomic_store_int(&sc->running, 1);
2385 if (nbufs != ktls_max_alloc) {
2387 nbufs = atomic_load_int(&ktls_max_alloc);
2388 buf = malloc(sizeof(void *) * nbufs, M_KTLS,
2392 * Below we allocate nbufs with different allocation
2393 * flags than we use when allocating normally during
2394 * encryption in the ktls worker thread. We specify
2395 * M_NORECLAIM in the worker thread. However, we omit
2396 * that flag here and add M_WAITOK so that the VM
2397 * system is permitted to perform expensive work to
2398 * defragment memory. We do this here, as it does not
2399 * matter if this thread blocks. If we block a ktls
2400 * worker thread, we risk developing backlogs of
2401 * buffers to be encrypted, leading to surges of
2402 * traffic and potential NIC output drops.
2404 for (i = 0; i < nbufs; i++) {
2405 buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
2408 for (i = 0; i < nbufs; i++) {
2409 uma_zfree(ktls_buffer_zone, buf[i]);
2416 ktls_work_thread(void *ctx)
2418 struct ktls_wq *wq = ctx;
2420 struct socket *so, *son;
2421 STAILQ_HEAD(, mbuf) local_m_head;
2422 STAILQ_HEAD(, socket) local_so_head;
2424 if (ktls_bind_threads > 1) {
2425 curthread->td_domain.dr_policy =
2426 DOMAINSET_PREF(PCPU_GET(domain));
2428 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2433 while (STAILQ_EMPTY(&wq->m_head) &&
2434 STAILQ_EMPTY(&wq->so_head)) {
2435 wq->running = false;
2436 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2440 STAILQ_INIT(&local_m_head);
2441 STAILQ_CONCAT(&local_m_head, &wq->m_head);
2442 STAILQ_INIT(&local_so_head);
2443 STAILQ_CONCAT(&local_so_head, &wq->so_head);
2444 mtx_unlock(&wq->mtx);
2446 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2447 if (m->m_epg_flags & EPG_FLAG_2FREE) {
2448 ktls_free(m->m_epg_tls);
2451 if (m->m_epg_tls->sync_dispatch)
2452 ktls_encrypt(wq, m);
2454 ktls_encrypt_async(wq, m);
2455 counter_u64_add(ktls_cnt_tx_queued, -1);
2459 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2461 counter_u64_add(ktls_cnt_rx_queued, -1);
2466 #if defined(INET) || defined(INET6)
2468 ktls_disable_ifnet_help(void *context, int pending __unused)
2470 struct ktls_session *tls;
2481 so = inp->inp_socket;
2483 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
2484 (inp->inp_flags2 & INP_FREED)) {
2488 if (so->so_snd.sb_tls_info != NULL)
2489 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
2493 counter_u64_add(ktls_ifnet_disable_ok, 1);
2494 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
2495 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 &&
2496 (inp->inp_flags2 & INP_FREED) == 0 &&
2497 (tp = intotcpcb(inp)) != NULL &&
2498 tp->t_fb->tfb_hwtls_change != NULL)
2499 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
2501 counter_u64_add(ktls_ifnet_disable_fail, 1);
2507 if (!in_pcbrele_wlocked(inp))
2513 * Called when re-transmits are becoming a substantial portion of the
2514 * sends on this connection. When this happens, we transition the
2515 * connection to software TLS. This is needed because most inline TLS
2516 * NICs keep crypto state only for in-order transmits. This means
2517 * that to handle a TCP rexmit (which is out-of-order), the NIC must
2518 * re-DMA the entire TLS record up to and including the current
2519 * segment. This means that when re-transmitting the last ~1448 byte
2520 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
2521 * of magnitude more data than we are sending. This can cause the
2522 * PCIe link to saturate well before the network, which can cause
2523 * output drops, and a general loss of capacity.
2526 ktls_disable_ifnet(void *arg)
2531 struct ktls_session *tls;
2535 INP_WLOCK_ASSERT(inp);
2536 so = inp->inp_socket;
2538 tls = so->so_snd.sb_tls_info;
2539 if (tls->disable_ifnet_pending) {
2545 * note that disable_ifnet_pending is never cleared; disabling
2546 * ifnet can only be done once per session, so we never want
2550 (void)ktls_hold(tls);
2553 tls->disable_ifnet_pending = true;
2556 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
2557 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);