FreeBSD kernel kern code
uipc_ktls.c
Go to the documentation of this file.
1/*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2014-2019 Netflix Inc.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
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.
14 *
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
25 * SUCH DAMAGE.
26 */
27
28#include <sys/cdefs.h>
29__FBSDID("$FreeBSD$");
30
31#include "opt_inet.h"
32#include "opt_inet6.h"
33#include "opt_kern_tls.h"
34#include "opt_ratelimit.h"
35#include "opt_rss.h"
36
37#include <sys/param.h>
38#include <sys/kernel.h>
39#include <sys/domainset.h>
40#include <sys/endian.h>
41#include <sys/ktls.h>
42#include <sys/lock.h>
43#include <sys/mbuf.h>
44#include <sys/mutex.h>
45#include <sys/rmlock.h>
46#include <sys/proc.h>
47#include <sys/protosw.h>
48#include <sys/refcount.h>
49#include <sys/smp.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>
55#include <sys/uio.h>
56#include <sys/vmmeter.h>
57#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
58#include <machine/pcb.h>
59#endif
60#include <machine/vmparam.h>
61#include <net/if.h>
62#include <net/if_var.h>
63#ifdef RSS
64#include <net/netisr.h>
65#include <net/rss_config.h>
66#endif
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>
72#endif
73#include <netinet/tcp_var.h>
74#ifdef TCP_OFFLOAD
75#include <netinet/tcp_offload.h>
76#endif
77#include <opencrypto/cryptodev.h>
78#include <opencrypto/ktls.h>
79#include <vm/uma_dbg.h>
80#include <vm/vm.h>
81#include <vm/vm_pageout.h>
82#include <vm/vm_page.h>
83#include <vm/vm_pagequeue.h>
84
85struct ktls_wq {
86 struct mtx mtx;
87 STAILQ_HEAD(, mbuf) m_head;
88 STAILQ_HEAD(, socket) so_head;
89 bool running;
90 int lastallocfail;
91} __aligned(CACHE_LINE_SIZE);
92
93struct ktls_alloc_thread {
94 uint64_t wakeups;
95 uint64_t allocs;
96 struct thread *td;
97 int running;
98};
99
100struct ktls_domain_info {
101 int count;
102 int cpu[MAXCPU];
103 struct ktls_alloc_thread alloc_td;
104};
105
106struct ktls_domain_info ktls_domains[MAXMEMDOM];
107static struct ktls_wq *ktls_wq;
108static struct proc *ktls_proc;
109static uma_zone_t ktls_session_zone;
110static uma_zone_t ktls_buffer_zone;
111static uint16_t ktls_cpuid_lookup[MAXCPU];
112static int ktls_init_state;
113static struct sx ktls_init_lock;
114SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
115
116SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
117 "Kernel TLS offload");
118SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
119 "Kernel TLS offload stats");
120
121#ifdef RSS
122static int ktls_bind_threads = 1;
123#else
124static int ktls_bind_threads;
125#endif
126SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
127 &ktls_bind_threads, 0,
128 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
129
130static u_int ktls_maxlen = 16384;
131SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
132 &ktls_maxlen, 0, "Maximum TLS record size");
133
134static int ktls_number_threads;
135SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
136 &ktls_number_threads, 0,
137 "Number of TLS threads in thread-pool");
138
139unsigned int ktls_ifnet_max_rexmit_pct = 2;
140SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
141 &ktls_ifnet_max_rexmit_pct, 2,
142 "Max percent bytes retransmitted before ifnet TLS is disabled");
143
144static bool ktls_offload_enable;
145SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
146 &ktls_offload_enable, 0,
147 "Enable support for kernel TLS offload");
148
149static bool ktls_cbc_enable = true;
150SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
151 &ktls_cbc_enable, 1,
152 "Enable Support of AES-CBC crypto for kernel TLS");
153
154static bool ktls_sw_buffer_cache = true;
155SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
156 &ktls_sw_buffer_cache, 1,
157 "Enable caching of output buffers for SW encryption");
158
159static int ktls_max_alloc = 128;
160SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN,
161 &ktls_max_alloc, 128,
162 "Max number of 16k buffers to allocate in thread context");
163
164static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
165SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
166 &ktls_tasks_active, "Number of active tasks");
167
168static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
169SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
170 &ktls_cnt_tx_pending,
171 "Number of TLS 1.0 records waiting for earlier TLS records");
172
173static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
174SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
175 &ktls_cnt_tx_queued,
176 "Number of TLS records in queue to tasks for SW encryption");
177
178static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
179SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
180 &ktls_cnt_rx_queued,
181 "Number of TLS sockets in queue to tasks for SW decryption");
182
183static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
184SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
185 CTLFLAG_RD, &ktls_offload_total,
186 "Total successful TLS setups (parameters set)");
187
188static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
189SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
190 CTLFLAG_RD, &ktls_offload_enable_calls,
191 "Total number of TLS enable calls made");
192
193static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
194SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
195 &ktls_offload_active, "Total Active TLS sessions");
196
197static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
198SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
199 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
200
201static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
202SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
203 &ktls_offload_failed_crypto, "Total TLS crypto failures");
204
205static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
206SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
207 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
208
209static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
210SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
211 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
212
213static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
214SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
215 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
216
217static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
218SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
219 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
220
221static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
222SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
223 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
224
225SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226 "Software TLS session stats");
227SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228 "Hardware (ifnet) TLS session stats");
229#ifdef TCP_OFFLOAD
230SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231 "TOE TLS session stats");
232#endif
233
234static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236 "Active number of software TLS sessions using AES-CBC");
237
238static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240 "Active number of software TLS sessions using AES-GCM");
241
242static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
244 &ktls_sw_chacha20,
245 "Active number of software TLS sessions using Chacha20-Poly1305");
246
247static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
249 &ktls_ifnet_cbc,
250 "Active number of ifnet TLS sessions using AES-CBC");
251
252static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
254 &ktls_ifnet_gcm,
255 "Active number of ifnet TLS sessions using AES-GCM");
256
257static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259 &ktls_ifnet_chacha20,
260 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
261
262static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
265
266static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268 &ktls_ifnet_reset_dropped,
269 "TLS sessions dropped after failing to update ifnet send tag");
270
271static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273 &ktls_ifnet_reset_failed,
274 "TLS sessions that failed to allocate a new ifnet send tag");
275
276static int ktls_ifnet_permitted;
277SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278 &ktls_ifnet_permitted, 1,
279 "Whether to permit hardware (ifnet) TLS sessions");
280
281#ifdef TCP_OFFLOAD
282static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
284 &ktls_toe_cbc,
285 "Active number of TOE TLS sessions using AES-CBC");
286
287static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
289 &ktls_toe_gcm,
290 "Active number of TOE TLS sessions using AES-GCM");
291
292static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
294 &ktls_toe_chacha20,
295 "Active number of TOE TLS sessions using Chacha20-Poly1305");
296#endif
297
298static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
299
300static void ktls_cleanup(struct ktls_session *tls);
301#if defined(INET) || defined(INET6)
302static void ktls_reset_send_tag(void *context, int pending);
303#endif
304static void ktls_work_thread(void *ctx);
305static void ktls_alloc_thread(void *ctx);
306
307#if defined(INET) || defined(INET6)
308static u_int
309ktls_get_cpu(struct socket *so)
310{
311 struct inpcb *inp;
312#ifdef NUMA
313 struct ktls_domain_info *di;
314#endif
315 u_int cpuid;
316
317 inp = sotoinpcb(so);
318#ifdef RSS
319 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
320 if (cpuid != NETISR_CPUID_NONE)
321 return (cpuid);
322#endif
323 /*
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
327 * the same queue.
328 */
329#ifdef NUMA
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];
333 } else
334#endif
335 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
336 return (cpuid);
337}
338#endif
339
340static int
341ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
342{
343 vm_page_t m;
344 int i, req;
345
346 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
347 ("%s: ktls max length %d is not page size-aligned",
348 __func__, ktls_maxlen));
349
350 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
351 for (i = 0; i < count; i++) {
352 m = vm_page_alloc_noobj_contig_domain(domain, req,
353 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
354 VM_MEMATTR_DEFAULT);
355 if (m == NULL)
356 break;
357 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
358 }
359 return (i);
360}
361
362static void
363ktls_buffer_release(void *arg __unused, void **store, int count)
364{
365 vm_page_t m;
366 int i, j;
367
368 for (i = 0; i < count; i++) {
369 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
370 for (j = 0; j < atop(ktls_maxlen); j++) {
371 (void)vm_page_unwire_noq(m + j);
372 vm_page_free(m + j);
373 }
374 }
375}
376
377static void
378ktls_free_mext_contig(struct mbuf *m)
379{
380 M_ASSERTEXTPG(m);
381 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
382}
383
384static int
385ktls_init(void)
386{
387 struct thread *td;
388 struct pcpu *pc;
389 int count, domain, error, i;
390
391 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
392 M_WAITOK | M_ZERO);
393
394 ktls_session_zone = uma_zcreate("ktls_session",
395 sizeof(struct ktls_session),
396 NULL, NULL, NULL, NULL,
397 UMA_ALIGN_CACHE, 0);
398
399 if (ktls_sw_buffer_cache) {
400 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
401 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
402 ktls_buffer_import, ktls_buffer_release, NULL,
403 UMA_ZONE_FIRSTTOUCH);
404 }
405
406 /*
407 * Initialize the workqueues to run the TLS work. We create a
408 * work queue for each CPU.
409 */
410 CPU_FOREACH(i) {
411 STAILQ_INIT(&ktls_wq[i].m_head);
412 STAILQ_INIT(&ktls_wq[i].so_head);
413 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
414 if (ktls_bind_threads > 1) {
415 pc = pcpu_find(i);
416 domain = pc->pc_domain;
417 count = ktls_domains[domain].count;
418 ktls_domains[domain].cpu[count] = i;
419 ktls_domains[domain].count++;
420 }
421 ktls_cpuid_lookup[ktls_number_threads] = i;
422 ktls_number_threads++;
423 }
424
425 /*
426 * If we somehow have an empty domain, fall back to choosing
427 * among all KTLS threads.
428 */
429 if (ktls_bind_threads > 1) {
430 for (i = 0; i < vm_ndomains; i++) {
431 if (ktls_domains[i].count == 0) {
432 ktls_bind_threads = 1;
433 break;
434 }
435 }
436 }
437
438 /* Start kthreads for each workqueue. */
439 CPU_FOREACH(i) {
440 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
441 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
442 if (error) {
443 printf("Can't add KTLS thread %d error %d\n", i, error);
444 return (error);
445 }
446 }
447
448 /*
449 * Start an allocation thread per-domain to perform blocking allocations
450 * of 16k physically contiguous TLS crypto destination buffers.
451 */
452 if (ktls_sw_buffer_cache) {
453 for (domain = 0; domain < vm_ndomains; domain++) {
454 if (VM_DOMAIN_EMPTY(domain))
455 continue;
456 if (CPU_EMPTY(&cpuset_domain[domain]))
457 continue;
458 error = kproc_kthread_add(ktls_alloc_thread,
459 &ktls_domains[domain], &ktls_proc,
460 &ktls_domains[domain].alloc_td.td,
461 0, 0, "KTLS", "alloc_%d", domain);
462 if (error) {
463 printf("Can't add KTLS alloc thread %d error %d\n",
464 domain, error);
465 return (error);
466 }
467 }
468 }
469
470 if (bootverbose)
471 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
472 return (0);
473}
474
475static int
476ktls_start_kthreads(void)
477{
478 int error, state;
479
480start:
481 state = atomic_load_acq_int(&ktls_init_state);
482 if (__predict_true(state > 0))
483 return (0);
484 if (state < 0)
485 return (ENXIO);
486
487 sx_xlock(&ktls_init_lock);
488 if (ktls_init_state != 0) {
489 sx_xunlock(&ktls_init_lock);
490 goto start;
491 }
492
493 error = ktls_init();
494 if (error == 0)
495 state = 1;
496 else
497 state = -1;
498 atomic_store_rel_int(&ktls_init_state, state);
499 sx_xunlock(&ktls_init_lock);
500 return (error);
501}
502
503#if defined(INET) || defined(INET6)
504static int
505ktls_create_session(struct socket *so, struct tls_enable *en,
506 struct ktls_session **tlsp)
507{
508 struct ktls_session *tls;
509 int error;
510
511 /* Only TLS 1.0 - 1.3 are supported. */
512 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
513 return (EINVAL);
514 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
515 en->tls_vminor > TLS_MINOR_VER_THREE)
516 return (EINVAL);
517
518 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
519 return (EINVAL);
520 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
521 return (EINVAL);
522 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
523 return (EINVAL);
524
525 /* All supported algorithms require a cipher key. */
526 if (en->cipher_key_len == 0)
527 return (EINVAL);
528
529 /* No flags are currently supported. */
530 if (en->flags != 0)
531 return (EINVAL);
532
533 /* Common checks for supported algorithms. */
534 switch (en->cipher_algorithm) {
535 case CRYPTO_AES_NIST_GCM_16:
536 /*
537 * auth_algorithm isn't used, but permit GMAC values
538 * for compatibility.
539 */
540 switch (en->auth_algorithm) {
541 case 0:
542#ifdef COMPAT_FREEBSD12
543 /* XXX: Really 13.0-current COMPAT. */
544 case CRYPTO_AES_128_NIST_GMAC:
545 case CRYPTO_AES_192_NIST_GMAC:
546 case CRYPTO_AES_256_NIST_GMAC:
547#endif
548 break;
549 default:
550 return (EINVAL);
551 }
552 if (en->auth_key_len != 0)
553 return (EINVAL);
554 switch (en->tls_vminor) {
555 case TLS_MINOR_VER_TWO:
556 if (en->iv_len != TLS_AEAD_GCM_LEN)
557 return (EINVAL);
558 break;
559 case TLS_MINOR_VER_THREE:
560 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
561 return (EINVAL);
562 break;
563 default:
564 return (EINVAL);
565 }
566 break;
567 case CRYPTO_AES_CBC:
568 switch (en->auth_algorithm) {
569 case CRYPTO_SHA1_HMAC:
570 break;
571 case CRYPTO_SHA2_256_HMAC:
572 case CRYPTO_SHA2_384_HMAC:
573 if (en->tls_vminor != TLS_MINOR_VER_TWO)
574 return (EINVAL);
575 break;
576 default:
577 return (EINVAL);
578 }
579 if (en->auth_key_len == 0)
580 return (EINVAL);
581
582 /*
583 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
584 * use explicit IVs.
585 */
586 switch (en->tls_vminor) {
587 case TLS_MINOR_VER_ZERO:
588 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
589 return (EINVAL);
590 break;
591 case TLS_MINOR_VER_ONE:
592 case TLS_MINOR_VER_TWO:
593 /* Ignore any supplied IV. */
594 en->iv_len = 0;
595 break;
596 default:
597 return (EINVAL);
598 }
599 break;
600 case CRYPTO_CHACHA20_POLY1305:
601 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
602 return (EINVAL);
603 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
604 en->tls_vminor != TLS_MINOR_VER_THREE)
605 return (EINVAL);
606 if (en->iv_len != TLS_CHACHA20_IV_LEN)
607 return (EINVAL);
608 break;
609 default:
610 return (EINVAL);
611 }
612
613 error = ktls_start_kthreads();
614 if (error != 0)
615 return (error);
616
617 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
618
619 counter_u64_add(ktls_offload_active, 1);
620
621 refcount_init(&tls->refcount, 1);
622 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
623
624 tls->wq_index = ktls_get_cpu(so);
625
626 tls->params.cipher_algorithm = en->cipher_algorithm;
627 tls->params.auth_algorithm = en->auth_algorithm;
628 tls->params.tls_vmajor = en->tls_vmajor;
629 tls->params.tls_vminor = en->tls_vminor;
630 tls->params.flags = en->flags;
631 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
632
633 /* Set the header and trailer lengths. */
634 tls->params.tls_hlen = sizeof(struct tls_record_layer);
635 switch (en->cipher_algorithm) {
636 case CRYPTO_AES_NIST_GCM_16:
637 /*
638 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
639 * nonce. TLS 1.3 uses a 12 byte implicit IV.
640 */
641 if (en->tls_vminor < TLS_MINOR_VER_THREE)
642 tls->params.tls_hlen += sizeof(uint64_t);
643 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
644 tls->params.tls_bs = 1;
645 break;
646 case CRYPTO_AES_CBC:
647 switch (en->auth_algorithm) {
648 case CRYPTO_SHA1_HMAC:
649 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
650 /* Implicit IV, no nonce. */
651 tls->sequential_records = true;
652 tls->next_seqno = be64dec(en->rec_seq);
653 STAILQ_INIT(&tls->pending_records);
654 } else {
655 tls->params.tls_hlen += AES_BLOCK_LEN;
656 }
657 tls->params.tls_tlen = AES_BLOCK_LEN +
658 SHA1_HASH_LEN;
659 break;
660 case CRYPTO_SHA2_256_HMAC:
661 tls->params.tls_hlen += AES_BLOCK_LEN;
662 tls->params.tls_tlen = AES_BLOCK_LEN +
663 SHA2_256_HASH_LEN;
664 break;
665 case CRYPTO_SHA2_384_HMAC:
666 tls->params.tls_hlen += AES_BLOCK_LEN;
667 tls->params.tls_tlen = AES_BLOCK_LEN +
668 SHA2_384_HASH_LEN;
669 break;
670 default:
671 panic("invalid hmac");
672 }
673 tls->params.tls_bs = AES_BLOCK_LEN;
674 break;
675 case CRYPTO_CHACHA20_POLY1305:
676 /*
677 * Chacha20 uses a 12 byte implicit IV.
678 */
679 tls->params.tls_tlen = POLY1305_HASH_LEN;
680 tls->params.tls_bs = 1;
681 break;
682 default:
683 panic("invalid cipher");
684 }
685
686 /*
687 * TLS 1.3 includes optional padding which we do not support,
688 * and also puts the "real" record type at the end of the
689 * encrypted data.
690 */
691 if (en->tls_vminor == TLS_MINOR_VER_THREE)
692 tls->params.tls_tlen += sizeof(uint8_t);
693
694 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
695 ("TLS header length too long: %d", tls->params.tls_hlen));
696 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
697 ("TLS trailer length too long: %d", tls->params.tls_tlen));
698
699 if (en->auth_key_len != 0) {
700 tls->params.auth_key_len = en->auth_key_len;
701 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
702 M_WAITOK);
703 error = copyin(en->auth_key, tls->params.auth_key,
704 en->auth_key_len);
705 if (error)
706 goto out;
707 }
708
709 tls->params.cipher_key_len = en->cipher_key_len;
710 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
711 error = copyin(en->cipher_key, tls->params.cipher_key,
712 en->cipher_key_len);
713 if (error)
714 goto out;
715
716 /*
717 * This holds the implicit portion of the nonce for AEAD
718 * ciphers and the initial implicit IV for TLS 1.0. The
719 * explicit portions of the IV are generated in ktls_frame().
720 */
721 if (en->iv_len != 0) {
722 tls->params.iv_len = en->iv_len;
723 error = copyin(en->iv, tls->params.iv, en->iv_len);
724 if (error)
725 goto out;
726
727 /*
728 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
729 * counter to generate unique explicit IVs.
730 *
731 * Store this counter in the last 8 bytes of the IV
732 * array so that it is 8-byte aligned.
733 */
734 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
735 en->tls_vminor == TLS_MINOR_VER_TWO)
736 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
737 }
738
739 *tlsp = tls;
740 return (0);
741
742out:
743 ktls_cleanup(tls);
744 return (error);
745}
746
747static struct ktls_session *
748ktls_clone_session(struct ktls_session *tls)
749{
750 struct ktls_session *tls_new;
751
752 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
753
754 counter_u64_add(ktls_offload_active, 1);
755
756 refcount_init(&tls_new->refcount, 1);
757 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new);
758
759 /* Copy fields from existing session. */
760 tls_new->params = tls->params;
761 tls_new->wq_index = tls->wq_index;
762
763 /* Deep copy keys. */
764 if (tls_new->params.auth_key != NULL) {
765 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
766 M_KTLS, M_WAITOK);
767 memcpy(tls_new->params.auth_key, tls->params.auth_key,
768 tls->params.auth_key_len);
769 }
770
771 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
772 M_WAITOK);
773 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
774 tls->params.cipher_key_len);
775
776 return (tls_new);
777}
778#endif
779
780static void
781ktls_cleanup(struct ktls_session *tls)
782{
783
784 counter_u64_add(ktls_offload_active, -1);
785 switch (tls->mode) {
786 case TCP_TLS_MODE_SW:
787 switch (tls->params.cipher_algorithm) {
788 case CRYPTO_AES_CBC:
789 counter_u64_add(ktls_sw_cbc, -1);
790 break;
791 case CRYPTO_AES_NIST_GCM_16:
792 counter_u64_add(ktls_sw_gcm, -1);
793 break;
794 case CRYPTO_CHACHA20_POLY1305:
795 counter_u64_add(ktls_sw_chacha20, -1);
796 break;
797 }
798 break;
799 case TCP_TLS_MODE_IFNET:
800 switch (tls->params.cipher_algorithm) {
801 case CRYPTO_AES_CBC:
802 counter_u64_add(ktls_ifnet_cbc, -1);
803 break;
804 case CRYPTO_AES_NIST_GCM_16:
805 counter_u64_add(ktls_ifnet_gcm, -1);
806 break;
807 case CRYPTO_CHACHA20_POLY1305:
808 counter_u64_add(ktls_ifnet_chacha20, -1);
809 break;
810 }
811 if (tls->snd_tag != NULL)
812 m_snd_tag_rele(tls->snd_tag);
813 break;
814#ifdef TCP_OFFLOAD
815 case TCP_TLS_MODE_TOE:
816 switch (tls->params.cipher_algorithm) {
817 case CRYPTO_AES_CBC:
818 counter_u64_add(ktls_toe_cbc, -1);
819 break;
820 case CRYPTO_AES_NIST_GCM_16:
821 counter_u64_add(ktls_toe_gcm, -1);
822 break;
823 case CRYPTO_CHACHA20_POLY1305:
824 counter_u64_add(ktls_toe_chacha20, -1);
825 break;
826 }
827 break;
828#endif
829 }
830 if (tls->ocf_session != NULL)
831 ktls_ocf_free(tls);
832 if (tls->params.auth_key != NULL) {
833 zfree(tls->params.auth_key, M_KTLS);
834 tls->params.auth_key = NULL;
835 tls->params.auth_key_len = 0;
836 }
837 if (tls->params.cipher_key != NULL) {
838 zfree(tls->params.cipher_key, M_KTLS);
839 tls->params.cipher_key = NULL;
840 tls->params.cipher_key_len = 0;
841 }
842 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
843}
844
845#if defined(INET) || defined(INET6)
846
847#ifdef TCP_OFFLOAD
848static int
849ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
850{
851 struct inpcb *inp;
852 struct tcpcb *tp;
853 int error;
854
855 inp = so->so_pcb;
856 INP_WLOCK(inp);
857 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
858 INP_WUNLOCK(inp);
859 return (ECONNRESET);
860 }
861 if (inp->inp_socket == NULL) {
862 INP_WUNLOCK(inp);
863 return (ECONNRESET);
864 }
865 tp = intotcpcb(inp);
866 if (!(tp->t_flags & TF_TOE)) {
867 INP_WUNLOCK(inp);
868 return (EOPNOTSUPP);
869 }
870
871 error = tcp_offload_alloc_tls_session(tp, tls, direction);
872 INP_WUNLOCK(inp);
873 if (error == 0) {
874 tls->mode = TCP_TLS_MODE_TOE;
875 switch (tls->params.cipher_algorithm) {
876 case CRYPTO_AES_CBC:
877 counter_u64_add(ktls_toe_cbc, 1);
878 break;
879 case CRYPTO_AES_NIST_GCM_16:
880 counter_u64_add(ktls_toe_gcm, 1);
881 break;
882 case CRYPTO_CHACHA20_POLY1305:
883 counter_u64_add(ktls_toe_chacha20, 1);
884 break;
885 }
886 }
887 return (error);
888}
889#endif
890
891/*
892 * Common code used when first enabling ifnet TLS on a connection or
893 * when allocating a new ifnet TLS session due to a routing change.
894 * This function allocates a new TLS send tag on whatever interface
895 * the connection is currently routed over.
896 */
897static int
898ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
899 struct m_snd_tag **mstp)
900{
901 union if_snd_tag_alloc_params params;
902 struct ifnet *ifp;
903 struct nhop_object *nh;
904 struct tcpcb *tp;
905 int error;
906
907 INP_RLOCK(inp);
908 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
909 INP_RUNLOCK(inp);
910 return (ECONNRESET);
911 }
912 if (inp->inp_socket == NULL) {
913 INP_RUNLOCK(inp);
914 return (ECONNRESET);
915 }
916 tp = intotcpcb(inp);
917
918 /*
919 * Check administrative controls on ifnet TLS to determine if
920 * ifnet TLS should be denied.
921 *
922 * - Always permit 'force' requests.
923 * - ktls_ifnet_permitted == 0: always deny.
924 */
925 if (!force && ktls_ifnet_permitted == 0) {
926 INP_RUNLOCK(inp);
927 return (ENXIO);
928 }
929
930 /*
931 * XXX: Use the cached route in the inpcb to find the
932 * interface. This should perhaps instead use
933 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
934 * enabled after a connection has completed key negotiation in
935 * userland, the cached route will be present in practice.
936 */
937 nh = inp->inp_route.ro_nh;
938 if (nh == NULL) {
939 INP_RUNLOCK(inp);
940 return (ENXIO);
941 }
942 ifp = nh->nh_ifp;
943 if_ref(ifp);
944
945 /*
946 * Allocate a TLS + ratelimit tag if the connection has an
947 * existing pacing rate.
948 */
949 if (tp->t_pacing_rate != -1 &&
950 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) {
951 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
952 params.tls_rate_limit.inp = inp;
953 params.tls_rate_limit.tls = tls;
954 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
955 } else {
956 params.hdr.type = IF_SND_TAG_TYPE_TLS;
957 params.tls.inp = inp;
958 params.tls.tls = tls;
959 }
960 params.hdr.flowid = inp->inp_flowid;
961 params.hdr.flowtype = inp->inp_flowtype;
962 params.hdr.numa_domain = inp->inp_numa_domain;
963 INP_RUNLOCK(inp);
964
965 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) {
966 error = EOPNOTSUPP;
967 goto out;
968 }
969 if (inp->inp_vflag & INP_IPV6) {
970 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) {
971 error = EOPNOTSUPP;
972 goto out;
973 }
974 } else {
975 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) {
976 error = EOPNOTSUPP;
977 goto out;
978 }
979 }
980 error = m_snd_tag_alloc(ifp, &params, mstp);
981out:
982 if_rele(ifp);
983 return (error);
984}
985
986static int
987ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force)
988{
989 struct m_snd_tag *mst;
990 int error;
991
992 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
993 if (error == 0) {
994 tls->mode = TCP_TLS_MODE_IFNET;
995 tls->snd_tag = mst;
996 switch (tls->params.cipher_algorithm) {
997 case CRYPTO_AES_CBC:
998 counter_u64_add(ktls_ifnet_cbc, 1);
999 break;
1000 case CRYPTO_AES_NIST_GCM_16:
1001 counter_u64_add(ktls_ifnet_gcm, 1);
1002 break;
1003 case CRYPTO_CHACHA20_POLY1305:
1004 counter_u64_add(ktls_ifnet_chacha20, 1);
1005 break;
1006 }
1007 }
1008 return (error);
1009}
1010
1011static void
1012ktls_use_sw(struct ktls_session *tls)
1013{
1014 tls->mode = TCP_TLS_MODE_SW;
1015 switch (tls->params.cipher_algorithm) {
1016 case CRYPTO_AES_CBC:
1017 counter_u64_add(ktls_sw_cbc, 1);
1018 break;
1019 case CRYPTO_AES_NIST_GCM_16:
1020 counter_u64_add(ktls_sw_gcm, 1);
1021 break;
1022 case CRYPTO_CHACHA20_POLY1305:
1023 counter_u64_add(ktls_sw_chacha20, 1);
1024 break;
1025 }
1026}
1027
1028static int
1029ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction)
1030{
1031 int error;
1032
1033 error = ktls_ocf_try(so, tls, direction);
1034 if (error)
1035 return (error);
1036 ktls_use_sw(tls);
1037 return (0);
1038}
1039
1040/*
1041 * KTLS RX stores data in the socket buffer as a list of TLS records,
1042 * where each record is stored as a control message containg the TLS
1043 * header followed by data mbufs containing the decrypted data. This
1044 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1045 * both encrypted and decrypted data. TLS records decrypted by a NIC
1046 * should be queued to the socket buffer as records, but encrypted
1047 * data which needs to be decrypted by software arrives as a stream of
1048 * regular mbufs which need to be converted. In addition, there may
1049 * already be pending encrypted data in the socket buffer when KTLS RX
1050 * is enabled.
1051 *
1052 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1053 * is used:
1054 *
1055 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1056 *
1057 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1058 * from the first mbuf. Once all of the data for that TLS record is
1059 * queued, the socket is queued to a worker thread.
1060 *
1061 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1062 * the TLS chain. Each TLS record is detached from the TLS chain,
1063 * decrypted, and inserted into the regular socket buffer chain as
1064 * record starting with a control message holding the TLS header and
1065 * a chain of mbufs holding the encrypted data.
1066 */
1067
1068static void
1069sb_mark_notready(struct sockbuf *sb)
1070{
1071 struct mbuf *m;
1072
1073 m = sb->sb_mb;
1074 sb->sb_mtls = m;
1075 sb->sb_mb = NULL;
1076 sb->sb_mbtail = NULL;
1077 sb->sb_lastrecord = NULL;
1078 for (; m != NULL; m = m->m_next) {
1079 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1080 __func__));
1081 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1082 __func__));
1083 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1084 __func__));
1085 m->m_flags |= M_NOTREADY;
1086 sb->sb_acc -= m->m_len;
1087 sb->sb_tlscc += m->m_len;
1088 sb->sb_mtlstail = m;
1089 }
1090 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1091 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1092 sb->sb_ccc));
1093}
1094
1095/*
1096 * Return information about the pending TLS data in a socket
1097 * buffer. On return, 'seqno' is set to the sequence number
1098 * of the next TLS record to be received, 'resid' is set to
1099 * the amount of bytes still needed for the last pending
1100 * record. The function returns 'false' if the last pending
1101 * record contains a partial TLS header. In that case, 'resid'
1102 * is the number of bytes needed to complete the TLS header.
1103 */
1104bool
1105ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1106{
1107 struct tls_record_layer hdr;
1108 struct mbuf *m;
1109 uint64_t seqno;
1110 size_t resid;
1111 u_int offset, record_len;
1112
1113 SOCKBUF_LOCK_ASSERT(sb);
1114 MPASS(sb->sb_flags & SB_TLS_RX);
1115 seqno = sb->sb_tls_seqno;
1116 resid = sb->sb_tlscc;
1117 m = sb->sb_mtls;
1118 offset = 0;
1119
1120 if (resid == 0) {
1121 *seqnop = seqno;
1122 *residp = 0;
1123 return (true);
1124 }
1125
1126 for (;;) {
1127 seqno++;
1128
1129 if (resid < sizeof(hdr)) {
1130 *seqnop = seqno;
1131 *residp = sizeof(hdr) - resid;
1132 return (false);
1133 }
1134
1135 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1136
1137 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1138 if (resid <= record_len) {
1139 *seqnop = seqno;
1140 *residp = record_len - resid;
1141 return (true);
1142 }
1143 resid -= record_len;
1144
1145 while (record_len != 0) {
1146 if (m->m_len - offset > record_len) {
1147 offset += record_len;
1148 break;
1149 }
1150
1151 record_len -= (m->m_len - offset);
1152 offset = 0;
1153 m = m->m_next;
1154 }
1155 }
1156}
1157
1158int
1159ktls_enable_rx(struct socket *so, struct tls_enable *en)
1160{
1161 struct ktls_session *tls;
1162 int error;
1163
1164 if (!ktls_offload_enable)
1165 return (ENOTSUP);
1166 if (SOLISTENING(so))
1167 return (EINVAL);
1168
1169 counter_u64_add(ktls_offload_enable_calls, 1);
1170
1171 /*
1172 * This should always be true since only the TCP socket option
1173 * invokes this function.
1174 */
1175 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1176 return (EINVAL);
1177
1178 /*
1179 * XXX: Don't overwrite existing sessions. We should permit
1180 * this to support rekeying in the future.
1181 */
1182 if (so->so_rcv.sb_tls_info != NULL)
1183 return (EALREADY);
1184
1185 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1186 return (ENOTSUP);
1187
1188 error = ktls_create_session(so, en, &tls);
1189 if (error)
1190 return (error);
1191
1192 error = ktls_ocf_try(so, tls, KTLS_RX);
1193 if (error) {
1194 ktls_cleanup(tls);
1195 return (error);
1196 }
1197
1198 /* Mark the socket as using TLS offload. */
1199 SOCKBUF_LOCK(&so->so_rcv);
1200 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1201 so->so_rcv.sb_tls_info = tls;
1202 so->so_rcv.sb_flags |= SB_TLS_RX;
1203
1204 /* Mark existing data as not ready until it can be decrypted. */
1205 sb_mark_notready(&so->so_rcv);
1206 ktls_check_rx(&so->so_rcv);
1207 SOCKBUF_UNLOCK(&so->so_rcv);
1208
1209#ifdef TCP_OFFLOAD
1210 error = ktls_try_toe(so, tls, KTLS_RX);
1211 if (error)
1212#endif
1213 ktls_use_sw(tls);
1214
1215 counter_u64_add(ktls_offload_total, 1);
1216
1217 return (0);
1218}
1219
1220int
1221ktls_enable_tx(struct socket *so, struct tls_enable *en)
1222{
1223 struct ktls_session *tls;
1224 struct inpcb *inp;
1225 int error;
1226
1227 if (!ktls_offload_enable)
1228 return (ENOTSUP);
1229 if (SOLISTENING(so))
1230 return (EINVAL);
1231
1232 counter_u64_add(ktls_offload_enable_calls, 1);
1233
1234 /*
1235 * This should always be true since only the TCP socket option
1236 * invokes this function.
1237 */
1238 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1239 return (EINVAL);
1240
1241 /*
1242 * XXX: Don't overwrite existing sessions. We should permit
1243 * this to support rekeying in the future.
1244 */
1245 if (so->so_snd.sb_tls_info != NULL)
1246 return (EALREADY);
1247
1248 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1249 return (ENOTSUP);
1250
1251 /* TLS requires ext pgs */
1252 if (mb_use_ext_pgs == 0)
1253 return (ENXIO);
1254
1255 error = ktls_create_session(so, en, &tls);
1256 if (error)
1257 return (error);
1258
1259 /* Prefer TOE -> ifnet TLS -> software TLS. */
1260#ifdef TCP_OFFLOAD
1261 error = ktls_try_toe(so, tls, KTLS_TX);
1262 if (error)
1263#endif
1264 error = ktls_try_ifnet(so, tls, false);
1265 if (error)
1266 error = ktls_try_sw(so, tls, KTLS_TX);
1267
1268 if (error) {
1269 ktls_cleanup(tls);
1270 return (error);
1271 }
1272
1273 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1274 if (error) {
1275 ktls_cleanup(tls);
1276 return (error);
1277 }
1278
1279 /*
1280 * Write lock the INP when setting sb_tls_info so that
1281 * routines in tcp_ratelimit.c can read sb_tls_info while
1282 * holding the INP lock.
1283 */
1284 inp = so->so_pcb;
1285 INP_WLOCK(inp);
1286 SOCKBUF_LOCK(&so->so_snd);
1287 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1288 so->so_snd.sb_tls_info = tls;
1289 if (tls->mode != TCP_TLS_MODE_SW)
1290 so->so_snd.sb_flags |= SB_TLS_IFNET;
1291 SOCKBUF_UNLOCK(&so->so_snd);
1292 INP_WUNLOCK(inp);
1293 SOCK_IO_SEND_UNLOCK(so);
1294
1295 counter_u64_add(ktls_offload_total, 1);
1296
1297 return (0);
1298}
1299
1300int
1301ktls_get_rx_mode(struct socket *so, int *modep)
1302{
1303 struct ktls_session *tls;
1304 struct inpcb *inp __diagused;
1305
1306 if (SOLISTENING(so))
1307 return (EINVAL);
1308 inp = so->so_pcb;
1309 INP_WLOCK_ASSERT(inp);
1310 SOCK_RECVBUF_LOCK(so);
1311 tls = so->so_rcv.sb_tls_info;
1312 if (tls == NULL)
1313 *modep = TCP_TLS_MODE_NONE;
1314 else
1315 *modep = tls->mode;
1316 SOCK_RECVBUF_UNLOCK(so);
1317 return (0);
1318}
1319
1320/*
1321 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1322 *
1323 * This function gets information about the next TCP- and TLS-
1324 * sequence number to be processed by the TLS receive worker
1325 * thread. The information is extracted from the given "inpcb"
1326 * structure. The values are stored in host endian format at the two
1327 * given output pointer locations. The TCP sequence number points to
1328 * the beginning of the TLS header.
1329 *
1330 * This function returns zero on success, else a non-zero error code
1331 * is returned.
1332 */
1333int
1334ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1335{
1336 struct socket *so;
1337 struct tcpcb *tp;
1338
1339 INP_RLOCK(inp);
1340 so = inp->inp_socket;
1341 if (__predict_false(so == NULL)) {
1342 INP_RUNLOCK(inp);
1343 return (EINVAL);
1344 }
1345 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
1346 INP_RUNLOCK(inp);
1347 return (ECONNRESET);
1348 }
1349
1350 tp = intotcpcb(inp);
1351 MPASS(tp != NULL);
1352
1353 SOCKBUF_LOCK(&so->so_rcv);
1354 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1355 *tlsseq = so->so_rcv.sb_tls_seqno;
1356 SOCKBUF_UNLOCK(&so->so_rcv);
1357
1358 INP_RUNLOCK(inp);
1359
1360 return (0);
1361}
1362
1363int
1364ktls_get_tx_mode(struct socket *so, int *modep)
1365{
1366 struct ktls_session *tls;
1367 struct inpcb *inp __diagused;
1368
1369 if (SOLISTENING(so))
1370 return (EINVAL);
1371 inp = so->so_pcb;
1372 INP_WLOCK_ASSERT(inp);
1373 SOCK_SENDBUF_LOCK(so);
1374 tls = so->so_snd.sb_tls_info;
1375 if (tls == NULL)
1376 *modep = TCP_TLS_MODE_NONE;
1377 else
1378 *modep = tls->mode;
1379 SOCK_SENDBUF_UNLOCK(so);
1380 return (0);
1381}
1382
1383/*
1384 * Switch between SW and ifnet TLS sessions as requested.
1385 */
1386int
1387ktls_set_tx_mode(struct socket *so, int mode)
1388{
1389 struct ktls_session *tls, *tls_new;
1390 struct inpcb *inp;
1391 int error;
1392
1393 if (SOLISTENING(so))
1394 return (EINVAL);
1395 switch (mode) {
1396 case TCP_TLS_MODE_SW:
1397 case TCP_TLS_MODE_IFNET:
1398 break;
1399 default:
1400 return (EINVAL);
1401 }
1402
1403 inp = so->so_pcb;
1404 INP_WLOCK_ASSERT(inp);
1405 SOCKBUF_LOCK(&so->so_snd);
1406 tls = so->so_snd.sb_tls_info;
1407 if (tls == NULL) {
1408 SOCKBUF_UNLOCK(&so->so_snd);
1409 return (0);
1410 }
1411
1412 if (tls->mode == mode) {
1413 SOCKBUF_UNLOCK(&so->so_snd);
1414 return (0);
1415 }
1416
1417 tls = ktls_hold(tls);
1418 SOCKBUF_UNLOCK(&so->so_snd);
1419 INP_WUNLOCK(inp);
1420
1421 tls_new = ktls_clone_session(tls);
1422
1423 if (mode == TCP_TLS_MODE_IFNET)
1424 error = ktls_try_ifnet(so, tls_new, true);
1425 else
1426 error = ktls_try_sw(so, tls_new, KTLS_TX);
1427 if (error) {
1428 counter_u64_add(ktls_switch_failed, 1);
1429 ktls_free(tls_new);
1430 ktls_free(tls);
1431 INP_WLOCK(inp);
1432 return (error);
1433 }
1434
1435 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1436 if (error) {
1437 counter_u64_add(ktls_switch_failed, 1);
1438 ktls_free(tls_new);
1439 ktls_free(tls);
1440 INP_WLOCK(inp);
1441 return (error);
1442 }
1443
1444 /*
1445 * If we raced with another session change, keep the existing
1446 * session.
1447 */
1448 if (tls != so->so_snd.sb_tls_info) {
1449 counter_u64_add(ktls_switch_failed, 1);
1450 SOCK_IO_SEND_UNLOCK(so);
1451 ktls_free(tls_new);
1452 ktls_free(tls);
1453 INP_WLOCK(inp);
1454 return (EBUSY);
1455 }
1456
1457 INP_WLOCK(inp);
1458 SOCKBUF_LOCK(&so->so_snd);
1459 so->so_snd.sb_tls_info = tls_new;
1460 if (tls_new->mode != TCP_TLS_MODE_SW)
1461 so->so_snd.sb_flags |= SB_TLS_IFNET;
1462 SOCKBUF_UNLOCK(&so->so_snd);
1463 SOCK_IO_SEND_UNLOCK(so);
1464
1465 /*
1466 * Drop two references on 'tls'. The first is for the
1467 * ktls_hold() above. The second drops the reference from the
1468 * socket buffer.
1469 */
1470 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1471 ktls_free(tls);
1472 ktls_free(tls);
1473
1474 if (mode == TCP_TLS_MODE_IFNET)
1475 counter_u64_add(ktls_switch_to_ifnet, 1);
1476 else
1477 counter_u64_add(ktls_switch_to_sw, 1);
1478
1479 return (0);
1480}
1481
1482/*
1483 * Try to allocate a new TLS send tag. This task is scheduled when
1484 * ip_output detects a route change while trying to transmit a packet
1485 * holding a TLS record. If a new tag is allocated, replace the tag
1486 * in the TLS session. Subsequent packets on the connection will use
1487 * the new tag. If a new tag cannot be allocated, drop the
1488 * connection.
1489 */
1490static void
1491ktls_reset_send_tag(void *context, int pending)
1492{
1493 struct epoch_tracker et;
1494 struct ktls_session *tls;
1495 struct m_snd_tag *old, *new;
1496 struct inpcb *inp;
1497 struct tcpcb *tp;
1498 int error;
1499
1500 MPASS(pending == 1);
1501
1502 tls = context;
1503 inp = tls->inp;
1504
1505 /*
1506 * Free the old tag first before allocating a new one.
1507 * ip[6]_output_send() will treat a NULL send tag the same as
1508 * an ifp mismatch and drop packets until a new tag is
1509 * allocated.
1510 *
1511 * Write-lock the INP when changing tls->snd_tag since
1512 * ip[6]_output_send() holds a read-lock when reading the
1513 * pointer.
1514 */
1515 INP_WLOCK(inp);
1516 old = tls->snd_tag;
1517 tls->snd_tag = NULL;
1518 INP_WUNLOCK(inp);
1519 if (old != NULL)
1520 m_snd_tag_rele(old);
1521
1522 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1523
1524 if (error == 0) {
1525 INP_WLOCK(inp);
1526 tls->snd_tag = new;
1527 mtx_pool_lock(mtxpool_sleep, tls);
1528 tls->reset_pending = false;
1529 mtx_pool_unlock(mtxpool_sleep, tls);
1530 if (!in_pcbrele_wlocked(inp))
1531 INP_WUNLOCK(inp);
1532
1533 counter_u64_add(ktls_ifnet_reset, 1);
1534
1535 /*
1536 * XXX: Should we kick tcp_output explicitly now that
1537 * the send tag is fixed or just rely on timers?
1538 */
1539 } else {
1540 NET_EPOCH_ENTER(et);
1541 INP_WLOCK(inp);
1542 if (!in_pcbrele_wlocked(inp)) {
1543 if (!(inp->inp_flags & INP_TIMEWAIT) &&
1544 !(inp->inp_flags & INP_DROPPED)) {
1545 tp = intotcpcb(inp);
1546 CURVNET_SET(tp->t_vnet);
1547 tp = tcp_drop(tp, ECONNABORTED);
1548 CURVNET_RESTORE();
1549 if (tp != NULL)
1550 INP_WUNLOCK(inp);
1551 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1552 } else
1553 INP_WUNLOCK(inp);
1554 }
1555 NET_EPOCH_EXIT(et);
1556
1557 counter_u64_add(ktls_ifnet_reset_failed, 1);
1558
1559 /*
1560 * Leave reset_pending true to avoid future tasks while
1561 * the socket goes away.
1562 */
1563 }
1564
1565 ktls_free(tls);
1566}
1567
1568int
1569ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1570{
1571
1572 if (inp == NULL)
1573 return (ENOBUFS);
1574
1575 INP_LOCK_ASSERT(inp);
1576
1577 /*
1578 * See if we should schedule a task to update the send tag for
1579 * this session.
1580 */
1581 mtx_pool_lock(mtxpool_sleep, tls);
1582 if (!tls->reset_pending) {
1583 (void) ktls_hold(tls);
1584 in_pcbref(inp);
1585 tls->inp = inp;
1586 tls->reset_pending = true;
1587 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1588 }
1589 mtx_pool_unlock(mtxpool_sleep, tls);
1590 return (ENOBUFS);
1591}
1592
1593#ifdef RATELIMIT
1594int
1595ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1596{
1597 union if_snd_tag_modify_params params = {
1598 .rate_limit.max_rate = max_pacing_rate,
1599 .rate_limit.flags = M_NOWAIT,
1600 };
1601 struct m_snd_tag *mst;
1602
1603 /* Can't get to the inp, but it should be locked. */
1604 /* INP_LOCK_ASSERT(inp); */
1605
1606 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1607
1608 if (tls->snd_tag == NULL) {
1609 /*
1610 * Resetting send tag, ignore this change. The
1611 * pending reset may or may not see this updated rate
1612 * in the tcpcb. If it doesn't, we will just lose
1613 * this rate change.
1614 */
1615 return (0);
1616 }
1617
1618 MPASS(tls->snd_tag != NULL);
1619 MPASS(tls->snd_tag->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1620
1621 mst = tls->snd_tag;
1622 return (mst->sw->snd_tag_modify(mst, &params));
1623}
1624#endif
1625#endif
1626
1627void
1628ktls_destroy(struct ktls_session *tls)
1629{
1630
1631 if (tls->sequential_records) {
1632 struct mbuf *m, *n;
1633 int page_count;
1634
1635 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1636 page_count = m->m_epg_enc_cnt;
1637 while (page_count > 0) {
1638 KASSERT(page_count >= m->m_epg_nrdy,
1639 ("%s: too few pages", __func__));
1640 page_count -= m->m_epg_nrdy;
1641 m = m_free(m);
1642 }
1643 }
1644 }
1645 ktls_cleanup(tls);
1646 uma_zfree(ktls_session_zone, tls);
1647}
1648
1649void
1650ktls_seq(struct sockbuf *sb, struct mbuf *m)
1651{
1652
1653 for (; m != NULL; m = m->m_next) {
1654 KASSERT((m->m_flags & M_EXTPG) != 0,
1655 ("ktls_seq: mapped mbuf %p", m));
1656
1657 m->m_epg_seqno = sb->sb_tls_seqno;
1658 sb->sb_tls_seqno++;
1659 }
1660}
1661
1662/*
1663 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
1664 * mbuf in the chain must be an unmapped mbuf. The payload of the
1665 * mbuf must be populated with the payload of each TLS record.
1666 *
1667 * The record_type argument specifies the TLS record type used when
1668 * populating the TLS header.
1669 *
1670 * The enq_count argument on return is set to the number of pages of
1671 * payload data for this entire chain that need to be encrypted via SW
1672 * encryption. The returned value should be passed to ktls_enqueue
1673 * when scheduling encryption of this chain of mbufs. To handle the
1674 * special case of empty fragments for TLS 1.0 sessions, an empty
1675 * fragment counts as one page.
1676 */
1677void
1678ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
1679 uint8_t record_type)
1680{
1681 struct tls_record_layer *tlshdr;
1682 struct mbuf *m;
1683 uint64_t *noncep;
1684 uint16_t tls_len;
1685 int maxlen __diagused;
1686
1687 maxlen = tls->params.max_frame_len;
1688 *enq_cnt = 0;
1689 for (m = top; m != NULL; m = m->m_next) {
1690 /*
1691 * All mbufs in the chain should be TLS records whose
1692 * payload does not exceed the maximum frame length.
1693 *
1694 * Empty TLS 1.0 records are permitted when using CBC.
1695 */
1696 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
1697 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
1698 ("ktls_frame: m %p len %d", m, m->m_len));
1699
1700 /*
1701 * TLS frames require unmapped mbufs to store session
1702 * info.
1703 */
1704 KASSERT((m->m_flags & M_EXTPG) != 0,
1705 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
1706
1707 tls_len = m->m_len;
1708
1709 /* Save a reference to the session. */
1710 m->m_epg_tls = ktls_hold(tls);
1711
1712 m->m_epg_hdrlen = tls->params.tls_hlen;
1713 m->m_epg_trllen = tls->params.tls_tlen;
1714 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
1715 int bs, delta;
1716
1717 /*
1718 * AES-CBC pads messages to a multiple of the
1719 * block size. Note that the padding is
1720 * applied after the digest and the encryption
1721 * is done on the "plaintext || mac || padding".
1722 * At least one byte of padding is always
1723 * present.
1724 *
1725 * Compute the final trailer length assuming
1726 * at most one block of padding.
1727 * tls->params.tls_tlen is the maximum
1728 * possible trailer length (padding + digest).
1729 * delta holds the number of excess padding
1730 * bytes if the maximum were used. Those
1731 * extra bytes are removed.
1732 */
1733 bs = tls->params.tls_bs;
1734 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
1735 m->m_epg_trllen -= delta;
1736 }
1737 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
1738
1739 /* Populate the TLS header. */
1740 tlshdr = (void *)m->m_epg_hdr;
1741 tlshdr->tls_vmajor = tls->params.tls_vmajor;
1742
1743 /*
1744 * TLS 1.3 masquarades as TLS 1.2 with a record type
1745 * of TLS_RLTYPE_APP.
1746 */
1747 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
1748 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
1749 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
1750 tlshdr->tls_type = TLS_RLTYPE_APP;
1751 /* save the real record type for later */
1752 m->m_epg_record_type = record_type;
1753 m->m_epg_trail[0] = record_type;
1754 } else {
1755 tlshdr->tls_vminor = tls->params.tls_vminor;
1756 tlshdr->tls_type = record_type;
1757 }
1758 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
1759
1760 /*
1761 * Store nonces / explicit IVs after the end of the
1762 * TLS header.
1763 *
1764 * For GCM with TLS 1.2, an 8 byte nonce is copied
1765 * from the end of the IV. The nonce is then
1766 * incremented for use by the next record.
1767 *
1768 * For CBC, a random nonce is inserted for TLS 1.1+.
1769 */
1770 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
1771 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
1772 noncep = (uint64_t *)(tls->params.iv + 8);
1773 be64enc(tlshdr + 1, *noncep);
1774 (*noncep)++;
1775 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1776 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
1777 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
1778
1779 /*
1780 * When using SW encryption, mark the mbuf not ready.
1781 * It will be marked ready via sbready() after the
1782 * record has been encrypted.
1783 *
1784 * When using ifnet TLS, unencrypted TLS records are
1785 * sent down the stack to the NIC.
1786 */
1787 if (tls->mode == TCP_TLS_MODE_SW) {
1788 m->m_flags |= M_NOTREADY;
1789 if (__predict_false(tls_len == 0)) {
1790 /* TLS 1.0 empty fragment. */
1791 m->m_epg_nrdy = 1;
1792 } else
1793 m->m_epg_nrdy = m->m_epg_npgs;
1794 *enq_cnt += m->m_epg_nrdy;
1795 }
1796 }
1797}
1798
1799bool
1800ktls_permit_empty_frames(struct ktls_session *tls)
1801{
1802 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
1803 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
1804}
1805
1806void
1807ktls_check_rx(struct sockbuf *sb)
1808{
1809 struct tls_record_layer hdr;
1810 struct ktls_wq *wq;
1811 struct socket *so;
1812 bool running;
1813
1814 SOCKBUF_LOCK_ASSERT(sb);
1815 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1816 __func__, sb));
1817 so = __containerof(sb, struct socket, so_rcv);
1818
1819 if (sb->sb_flags & SB_TLS_RX_RUNNING)
1820 return;
1821
1822 /* Is there enough queued for a TLS header? */
1823 if (sb->sb_tlscc < sizeof(hdr)) {
1824 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
1825 so->so_error = EMSGSIZE;
1826 return;
1827 }
1828
1829 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
1830
1831 /* Is the entire record queued? */
1832 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
1833 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
1834 so->so_error = EMSGSIZE;
1835 return;
1836 }
1837
1838 sb->sb_flags |= SB_TLS_RX_RUNNING;
1839
1840 soref(so);
1841 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
1842 mtx_lock(&wq->mtx);
1843 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
1844 running = wq->running;
1845 mtx_unlock(&wq->mtx);
1846 if (!running)
1847 wakeup(wq);
1848 counter_u64_add(ktls_cnt_rx_queued, 1);
1849}
1850
1851static struct mbuf *
1852ktls_detach_record(struct sockbuf *sb, int len)
1853{
1854 struct mbuf *m, *n, *top;
1855 int remain;
1856
1857 SOCKBUF_LOCK_ASSERT(sb);
1858 MPASS(len <= sb->sb_tlscc);
1859
1860 /*
1861 * If TLS chain is the exact size of the record,
1862 * just grab the whole record.
1863 */
1864 top = sb->sb_mtls;
1865 if (sb->sb_tlscc == len) {
1866 sb->sb_mtls = NULL;
1867 sb->sb_mtlstail = NULL;
1868 goto out;
1869 }
1870
1871 /*
1872 * While it would be nice to use m_split() here, we need
1873 * to know exactly what m_split() allocates to update the
1874 * accounting, so do it inline instead.
1875 */
1876 remain = len;
1877 for (m = top; remain > m->m_len; m = m->m_next)
1878 remain -= m->m_len;
1879
1880 /* Easy case: don't have to split 'm'. */
1881 if (remain == m->m_len) {
1882 sb->sb_mtls = m->m_next;
1883 if (sb->sb_mtls == NULL)
1884 sb->sb_mtlstail = NULL;
1885 m->m_next = NULL;
1886 goto out;
1887 }
1888
1889 /*
1890 * Need to allocate an mbuf to hold the remainder of 'm'. Try
1891 * with M_NOWAIT first.
1892 */
1893 n = m_get(M_NOWAIT, MT_DATA);
1894 if (n == NULL) {
1895 /*
1896 * Use M_WAITOK with socket buffer unlocked. If
1897 * 'sb_mtls' changes while the lock is dropped, return
1898 * NULL to force the caller to retry.
1899 */
1900 SOCKBUF_UNLOCK(sb);
1901
1902 n = m_get(M_WAITOK, MT_DATA);
1903
1904 SOCKBUF_LOCK(sb);
1905 if (sb->sb_mtls != top) {
1906 m_free(n);
1907 return (NULL);
1908 }
1909 }
1910 n->m_flags |= M_NOTREADY;
1911
1912 /* Store remainder in 'n'. */
1913 n->m_len = m->m_len - remain;
1914 if (m->m_flags & M_EXT) {
1915 n->m_data = m->m_data + remain;
1916 mb_dupcl(n, m);
1917 } else {
1918 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
1919 }
1920
1921 /* Trim 'm' and update accounting. */
1922 m->m_len -= n->m_len;
1923 sb->sb_tlscc -= n->m_len;
1924 sb->sb_ccc -= n->m_len;
1925
1926 /* Account for 'n'. */
1927 sballoc_ktls_rx(sb, n);
1928
1929 /* Insert 'n' into the TLS chain. */
1930 sb->sb_mtls = n;
1931 n->m_next = m->m_next;
1932 if (sb->sb_mtlstail == m)
1933 sb->sb_mtlstail = n;
1934
1935 /* Detach the record from the TLS chain. */
1936 m->m_next = NULL;
1937
1938out:
1939 MPASS(m_length(top, NULL) == len);
1940 for (m = top; m != NULL; m = m->m_next)
1941 sbfree_ktls_rx(sb, m);
1942 sb->sb_tlsdcc = len;
1943 sb->sb_ccc += len;
1944 SBCHECK(sb);
1945 return (top);
1946}
1947
1948/*
1949 * Determine the length of the trailing zero padding and find the real
1950 * record type in the byte before the padding.
1951 *
1952 * Walking the mbuf chain backwards is clumsy, so another option would
1953 * be to scan forwards remembering the last non-zero byte before the
1954 * trailer. However, it would be expensive to scan the entire record.
1955 * Instead, find the last non-zero byte of each mbuf in the chain
1956 * keeping track of the relative offset of that nonzero byte.
1957 *
1958 * trail_len is the size of the MAC/tag on input and is set to the
1959 * size of the full trailer including padding and the record type on
1960 * return.
1961 */
1962static int
1963tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
1964 int *trailer_len, uint8_t *record_typep)
1965{
1966 char *cp;
1967 u_int digest_start, last_offset, m_len, offset;
1968 uint8_t record_type;
1969
1970 digest_start = tls_len - *trailer_len;
1971 last_offset = 0;
1972 offset = 0;
1973 for (; m != NULL && offset < digest_start;
1974 offset += m->m_len, m = m->m_next) {
1975 /* Don't look for padding in the tag. */
1976 m_len = min(digest_start - offset, m->m_len);
1977 cp = mtod(m, char *);
1978
1979 /* Find last non-zero byte in this mbuf. */
1980 while (m_len > 0 && cp[m_len - 1] == 0)
1981 m_len--;
1982 if (m_len > 0) {
1983 record_type = cp[m_len - 1];
1984 last_offset = offset + m_len;
1985 }
1986 }
1987 if (last_offset < tls->params.tls_hlen)
1988 return (EBADMSG);
1989
1990 *record_typep = record_type;
1991 *trailer_len = tls_len - last_offset + 1;
1992 return (0);
1993}
1994
1995static void
1996ktls_decrypt(struct socket *so)
1997{
1998 char tls_header[MBUF_PEXT_HDR_LEN];
1999 struct ktls_session *tls;
2000 struct sockbuf *sb;
2001 struct tls_record_layer *hdr;
2002 struct tls_get_record tgr;
2003 struct mbuf *control, *data, *m;
2004 uint64_t seqno;
2005 int error, remain, tls_len, trail_len;
2006 bool tls13;
2007 uint8_t vminor, record_type;
2008
2009 hdr = (struct tls_record_layer *)tls_header;
2010 sb = &so->so_rcv;
2011 SOCKBUF_LOCK(sb);
2012 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2013 ("%s: socket %p not running", __func__, so));
2014
2015 tls = sb->sb_tls_info;
2016 MPASS(tls != NULL);
2017
2018 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2019 if (tls13)
2020 vminor = TLS_MINOR_VER_TWO;
2021 else
2022 vminor = tls->params.tls_vminor;
2023 for (;;) {
2024 /* Is there enough queued for a TLS header? */
2025 if (sb->sb_tlscc < tls->params.tls_hlen)
2026 break;
2027
2028 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2029 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2030
2031 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2032 hdr->tls_vminor != vminor)
2033 error = EINVAL;
2034 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2035 error = EINVAL;
2036 else if (tls_len < tls->params.tls_hlen || tls_len >
2037 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2038 tls->params.tls_tlen)
2039 error = EMSGSIZE;
2040 else
2041 error = 0;
2042 if (__predict_false(error != 0)) {
2043 /*
2044 * We have a corrupted record and are likely
2045 * out of sync. The connection isn't
2046 * recoverable at this point, so abort it.
2047 */
2048 SOCKBUF_UNLOCK(sb);
2049 counter_u64_add(ktls_offload_corrupted_records, 1);
2050
2051 CURVNET_SET(so->so_vnet);
2052 so->so_proto->pr_usrreqs->pru_abort(so);
2053 so->so_error = error;
2054 CURVNET_RESTORE();
2055 goto deref;
2056 }
2057
2058 /* Is the entire record queued? */
2059 if (sb->sb_tlscc < tls_len)
2060 break;
2061
2062 /*
2063 * Split out the portion of the mbuf chain containing
2064 * this TLS record.
2065 */
2066 data = ktls_detach_record(sb, tls_len);
2067 if (data == NULL)
2068 continue;
2069 MPASS(sb->sb_tlsdcc == tls_len);
2070
2071 seqno = sb->sb_tls_seqno;
2072 sb->sb_tls_seqno++;
2073 SBCHECK(sb);
2074 SOCKBUF_UNLOCK(sb);
2075
2076 error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len);
2077 if (error == 0) {
2078 if (tls13)
2079 error = tls13_find_record_type(tls, data,
2080 tls_len, &trail_len, &record_type);
2081 else
2082 record_type = hdr->tls_type;
2083 }
2084 if (error) {
2085 counter_u64_add(ktls_offload_failed_crypto, 1);
2086
2087 SOCKBUF_LOCK(sb);
2088 if (sb->sb_tlsdcc == 0) {
2089 /*
2090 * sbcut/drop/flush discarded these
2091 * mbufs.
2092 */
2093 m_freem(data);
2094 break;
2095 }
2096
2097 /*
2098 * Drop this TLS record's data, but keep
2099 * decrypting subsequent records.
2100 */
2101 sb->sb_ccc -= tls_len;
2102 sb->sb_tlsdcc = 0;
2103
2104 CURVNET_SET(so->so_vnet);
2105 so->so_error = EBADMSG;
2106 sorwakeup_locked(so);
2107 CURVNET_RESTORE();
2108
2109 m_freem(data);
2110
2111 SOCKBUF_LOCK(sb);
2112 continue;
2113 }
2114
2115 /* Allocate the control mbuf. */
2116 memset(&tgr, 0, sizeof(tgr));
2117 tgr.tls_type = record_type;
2118 tgr.tls_vmajor = hdr->tls_vmajor;
2119 tgr.tls_vminor = hdr->tls_vminor;
2120 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2121 trail_len);
2122 control = sbcreatecontrol_how(&tgr, sizeof(tgr),
2123 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2124
2125 SOCKBUF_LOCK(sb);
2126 if (sb->sb_tlsdcc == 0) {
2127 /* sbcut/drop/flush discarded these mbufs. */
2128 MPASS(sb->sb_tlscc == 0);
2129 m_freem(data);
2130 m_freem(control);
2131 break;
2132 }
2133
2134 /*
2135 * Clear the 'dcc' accounting in preparation for
2136 * adding the decrypted record.
2137 */
2138 sb->sb_ccc -= tls_len;
2139 sb->sb_tlsdcc = 0;
2140 SBCHECK(sb);
2141
2142 /* If there is no payload, drop all of the data. */
2143 if (tgr.tls_length == htobe16(0)) {
2144 m_freem(data);
2145 data = NULL;
2146 } else {
2147 /* Trim header. */
2148 remain = tls->params.tls_hlen;
2149 while (remain > 0) {
2150 if (data->m_len > remain) {
2151 data->m_data += remain;
2152 data->m_len -= remain;
2153 break;
2154 }
2155 remain -= data->m_len;
2156 data = m_free(data);
2157 }
2158
2159 /* Trim trailer and clear M_NOTREADY. */
2160 remain = be16toh(tgr.tls_length);
2161 m = data;
2162 for (m = data; remain > m->m_len; m = m->m_next) {
2163 m->m_flags &= ~M_NOTREADY;
2164 remain -= m->m_len;
2165 }
2166 m->m_len = remain;
2167 m_freem(m->m_next);
2168 m->m_next = NULL;
2169 m->m_flags &= ~M_NOTREADY;
2170
2171 /* Set EOR on the final mbuf. */
2172 m->m_flags |= M_EOR;
2173 }
2174
2175 sbappendcontrol_locked(sb, data, control, 0);
2176 }
2177
2178 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2179
2180 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2181 so->so_error = EMSGSIZE;
2182
2183 sorwakeup_locked(so);
2184
2185deref:
2186 SOCKBUF_UNLOCK_ASSERT(sb);
2187
2188 CURVNET_SET(so->so_vnet);
2189 sorele(so);
2190 CURVNET_RESTORE();
2191}
2192
2193void
2194ktls_enqueue_to_free(struct mbuf *m)
2195{
2196 struct ktls_wq *wq;
2197 bool running;
2198
2199 /* Mark it for freeing. */
2200 m->m_epg_flags |= EPG_FLAG_2FREE;
2201 wq = &ktls_wq[m->m_epg_tls->wq_index];
2202 mtx_lock(&wq->mtx);
2203 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2204 running = wq->running;
2205 mtx_unlock(&wq->mtx);
2206 if (!running)
2207 wakeup(wq);
2208}
2209
2210static void *
2211ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2212{
2213 void *buf;
2214 int domain, running;
2215
2216 if (m->m_epg_npgs <= 2)
2217 return (NULL);
2218 if (ktls_buffer_zone == NULL)
2219 return (NULL);
2220 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2221 /*
2222 * Rate-limit allocation attempts after a failure.
2223 * ktls_buffer_import() will acquire a per-domain mutex to check
2224 * the free page queues and may fail consistently if memory is
2225 * fragmented.
2226 */
2227 return (NULL);
2228 }
2229 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2230 if (buf == NULL) {
2231 domain = PCPU_GET(domain);
2232 wq->lastallocfail = ticks;
2233
2234 /*
2235 * Note that this check is "racy", but the races are
2236 * harmless, and are either a spurious wakeup if
2237 * multiple threads fail allocations before the alloc
2238 * thread wakes, or waiting an extra second in case we
2239 * see an old value of running == true.
2240 */
2241 if (!VM_DOMAIN_EMPTY(domain)) {
2242 running = atomic_load_int(&ktls_domains[domain].alloc_td.running);
2243 if (!running)
2244 wakeup(&ktls_domains[domain].alloc_td);
2245 }
2246 }
2247 return (buf);
2248}
2249
2250static int
2251ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2252 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2253{
2254 vm_page_t pg;
2255 int error, i, len, off;
2256
2257 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2258 ("%p not unready & nomap mbuf\n", m));
2259 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2260 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2261 ktls_maxlen));
2262
2263 /* Anonymous mbufs are encrypted in place. */
2264 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2265 return (tls->sw_encrypt(state, tls, m, NULL, 0));
2266
2267 /*
2268 * For file-backed mbufs (from sendfile), anonymous wired
2269 * pages are allocated and used as the encryption destination.
2270 */
2271 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2272 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2273 m->m_epg_1st_off;
2274 state->dst_iov[0].iov_base = (char *)state->cbuf +
2275 m->m_epg_1st_off;
2276 state->dst_iov[0].iov_len = len;
2277 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2278 i = 1;
2279 } else {
2280 off = m->m_epg_1st_off;
2281 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2282 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2283 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2284 len = m_epg_pagelen(m, i, off);
2285 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2286 state->dst_iov[i].iov_base =
2287 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2288 state->dst_iov[i].iov_len = len;
2289 }
2290 }
2291 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2292 state->dst_iov[i].iov_base = m->m_epg_trail;
2293 state->dst_iov[i].iov_len = m->m_epg_trllen;
2294
2295 error = tls->sw_encrypt(state, tls, m, state->dst_iov, i + 1);
2296
2297 if (__predict_false(error != 0)) {
2298 /* Free the anonymous pages. */
2299 if (state->cbuf != NULL)
2300 uma_zfree(ktls_buffer_zone, state->cbuf);
2301 else {
2302 for (i = 0; i < m->m_epg_npgs; i++) {
2303 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2304 (void)vm_page_unwire_noq(pg);
2305 vm_page_free(pg);
2306 }
2307 }
2308 }
2309 return (error);
2310}
2311
2312/* Number of TLS records in a batch passed to ktls_enqueue(). */
2313static u_int
2314ktls_batched_records(struct mbuf *m)
2315{
2316 int page_count, records;
2317
2318 records = 0;
2319 page_count = m->m_epg_enc_cnt;
2320 while (page_count > 0) {
2321 records++;
2322 page_count -= m->m_epg_nrdy;
2323 m = m->m_next;
2324 }
2325 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2326 return (records);
2327}
2328
2329void
2330ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2331{
2332 struct ktls_session *tls;
2333 struct ktls_wq *wq;
2334 int queued;
2335 bool running;
2336
2337 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2338 (M_EXTPG | M_NOTREADY)),
2339 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2340 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2341
2342 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2343
2344 m->m_epg_enc_cnt = page_count;
2345
2346 /*
2347 * Save a pointer to the socket. The caller is responsible
2348 * for taking an additional reference via soref().
2349 */
2350 m->m_epg_so = so;
2351
2352 queued = 1;
2353 tls = m->m_epg_tls;
2354 wq = &ktls_wq[tls->wq_index];
2355 mtx_lock(&wq->mtx);
2356 if (__predict_false(tls->sequential_records)) {
2357 /*
2358 * For TLS 1.0, records must be encrypted
2359 * sequentially. For a given connection, all records
2360 * queued to the associated work queue are processed
2361 * sequentially. However, sendfile(2) might complete
2362 * I/O requests spanning multiple TLS records out of
2363 * order. Here we ensure TLS records are enqueued to
2364 * the work queue in FIFO order.
2365 *
2366 * tls->next_seqno holds the sequence number of the
2367 * next TLS record that should be enqueued to the work
2368 * queue. If this next record is not tls->next_seqno,
2369 * it must be a future record, so insert it, sorted by
2370 * TLS sequence number, into tls->pending_records and
2371 * return.
2372 *
2373 * If this TLS record matches tls->next_seqno, place
2374 * it in the work queue and then check
2375 * tls->pending_records to see if any
2376 * previously-queued records are now ready for
2377 * encryption.
2378 */
2379 if (m->m_epg_seqno != tls->next_seqno) {
2380 struct mbuf *n, *p;
2381
2382 p = NULL;
2383 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2384 if (n->m_epg_seqno > m->m_epg_seqno)
2385 break;
2386 p = n;
2387 }
2388 if (n == NULL)
2389 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2390 m_epg_stailq);
2391 else if (p == NULL)
2392 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2393 m_epg_stailq);
2394 else
2395 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2396 m_epg_stailq);
2397 mtx_unlock(&wq->mtx);
2398 counter_u64_add(ktls_cnt_tx_pending, 1);
2399 return;
2400 }
2401
2402 tls->next_seqno += ktls_batched_records(m);
2403 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2404
2405 while (!STAILQ_EMPTY(&tls->pending_records)) {
2406 struct mbuf *n;
2407
2408 n = STAILQ_FIRST(&tls->pending_records);
2409 if (n->m_epg_seqno != tls->next_seqno)
2410 break;
2411
2412 queued++;
2413 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2414 tls->next_seqno += ktls_batched_records(n);
2415 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2416 }
2417 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2418 } else
2419 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2420
2421 running = wq->running;
2422 mtx_unlock(&wq->mtx);
2423 if (!running)
2424 wakeup(wq);
2425 counter_u64_add(ktls_cnt_tx_queued, queued);
2426}
2427
2428/*
2429 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2430 * the pages from the file and replace them with the anonymous pages
2431 * allocated in ktls_encrypt_record().
2432 */
2433static void
2434ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2435{
2436 int i;
2437
2438 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2439
2440 /* Free the old pages. */
2441 m->m_ext.ext_free(m);
2442
2443 /* Replace them with the new pages. */
2444 if (state->cbuf != NULL) {
2445 for (i = 0; i < m->m_epg_npgs; i++)
2446 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2447
2448 /* Contig pages should go back to the cache. */
2449 m->m_ext.ext_free = ktls_free_mext_contig;
2450 } else {
2451 for (i = 0; i < m->m_epg_npgs; i++)
2452 m->m_epg_pa[i] = state->parray[i];
2453
2454 /* Use the basic free routine. */
2455 m->m_ext.ext_free = mb_free_mext_pgs;
2456 }
2457
2458 /* Pages are now writable. */
2459 m->m_epg_flags |= EPG_FLAG_ANON;
2460}
2461
2462static __noinline void
2463ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2464{
2465 struct ktls_ocf_encrypt_state state;
2466 struct ktls_session *tls;
2467 struct socket *so;
2468 struct mbuf *m;
2469 int error, npages, total_pages;
2470
2471 so = top->m_epg_so;
2472 tls = top->m_epg_tls;
2473 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2474 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2475#ifdef INVARIANTS
2476 top->m_epg_so = NULL;
2477#endif
2478 total_pages = top->m_epg_enc_cnt;
2479 npages = 0;
2480
2481 /*
2482 * Encrypt the TLS records in the chain of mbufs starting with
2483 * 'top'. 'total_pages' gives us a total count of pages and is
2484 * used to know when we have finished encrypting the TLS
2485 * records originally queued with 'top'.
2486 *
2487 * NB: These mbufs are queued in the socket buffer and
2488 * 'm_next' is traversing the mbufs in the socket buffer. The
2489 * socket buffer lock is not held while traversing this chain.
2490 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2491 * pointers should be stable. However, the 'm_next' of the
2492 * last mbuf encrypted is not necessarily NULL. It can point
2493 * to other mbufs appended while 'top' was on the TLS work
2494 * queue.
2495 *
2496 * Each mbuf holds an entire TLS record.
2497 */
2498 error = 0;
2499 for (m = top; npages != total_pages; m = m->m_next) {
2500 KASSERT(m->m_epg_tls == tls,
2501 ("different TLS sessions in a single mbuf chain: %p vs %p",
2502 tls, m->m_epg_tls));
2503 KASSERT(npages + m->m_epg_npgs <= total_pages,
2504 ("page count mismatch: top %p, total_pages %d, m %p", top,
2505 total_pages, m));
2506
2507 error = ktls_encrypt_record(wq, m, tls, &state);
2508 if (error) {
2509 counter_u64_add(ktls_offload_failed_crypto, 1);
2510 break;
2511 }
2512
2513 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2514 ktls_finish_nonanon(m, &state);
2515
2516 npages += m->m_epg_nrdy;
2517
2518 /*
2519 * Drop a reference to the session now that it is no
2520 * longer needed. Existing code depends on encrypted
2521 * records having no associated session vs
2522 * yet-to-be-encrypted records having an associated
2523 * session.
2524 */
2525 m->m_epg_tls = NULL;
2526 ktls_free(tls);
2527 }
2528
2529 CURVNET_SET(so->so_vnet);
2530 if (error == 0) {
2531 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages);
2532 } else {
2533 so->so_proto->pr_usrreqs->pru_abort(so);
2534 so->so_error = EIO;
2535 mb_free_notready(top, total_pages);
2536 }
2537
2538 sorele(so);
2539 CURVNET_RESTORE();
2540}
2541
2542void
2543ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
2544{
2545 struct ktls_session *tls;
2546 struct socket *so;
2547 struct mbuf *m;
2548 int npages;
2549
2550 m = state->m;
2551
2552 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2553 ktls_finish_nonanon(m, state);
2554
2555 so = state->so;
2556 free(state, M_KTLS);
2557
2558 /*
2559 * Drop a reference to the session now that it is no longer
2560 * needed. Existing code depends on encrypted records having
2561 * no associated session vs yet-to-be-encrypted records having
2562 * an associated session.
2563 */
2564 tls = m->m_epg_tls;
2565 m->m_epg_tls = NULL;
2566 ktls_free(tls);
2567
2568 if (error != 0)
2569 counter_u64_add(ktls_offload_failed_crypto, 1);
2570
2571 CURVNET_SET(so->so_vnet);
2572 npages = m->m_epg_nrdy;
2573
2574 if (error == 0) {
2575 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, m, npages);
2576 } else {
2577 so->so_proto->pr_usrreqs->pru_abort(so);
2578 so->so_error = EIO;
2579 mb_free_notready(m, npages);
2580 }
2581
2582 sorele(so);
2583 CURVNET_RESTORE();
2584}
2585
2586/*
2587 * Similar to ktls_encrypt, but used with asynchronous OCF backends
2588 * (coprocessors) where encryption does not use host CPU resources and
2589 * it can be beneficial to queue more requests than CPUs.
2590 */
2591static __noinline void
2592ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
2593{
2594 struct ktls_ocf_encrypt_state *state;
2595 struct ktls_session *tls;
2596 struct socket *so;
2597 struct mbuf *m, *n;
2598 int error, mpages, npages, total_pages;
2599
2600 so = top->m_epg_so;
2601 tls = top->m_epg_tls;
2602 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2603 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2604#ifdef INVARIANTS
2605 top->m_epg_so = NULL;
2606#endif
2607 total_pages = top->m_epg_enc_cnt;
2608 npages = 0;
2609
2610 error = 0;
2611 for (m = top; npages != total_pages; m = n) {
2612 KASSERT(m->m_epg_tls == tls,
2613 ("different TLS sessions in a single mbuf chain: %p vs %p",
2614 tls, m->m_epg_tls));
2615 KASSERT(npages + m->m_epg_npgs <= total_pages,
2616 ("page count mismatch: top %p, total_pages %d, m %p", top,
2617 total_pages, m));
2618
2619 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
2620 soref(so);
2621 state->so = so;
2622 state->m = m;
2623
2624 mpages = m->m_epg_nrdy;
2625 n = m->m_next;
2626
2627 error = ktls_encrypt_record(wq, m, tls, state);
2628 if (error) {
2629 counter_u64_add(ktls_offload_failed_crypto, 1);
2630 free(state, M_KTLS);
2631 CURVNET_SET(so->so_vnet);
2632 sorele(so);
2633 CURVNET_RESTORE();
2634 break;
2635 }
2636
2637 npages += mpages;
2638 }
2639
2640 CURVNET_SET(so->so_vnet);
2641 if (error != 0) {
2642 so->so_proto->pr_usrreqs->pru_abort(so);
2643 so->so_error = EIO;
2644 mb_free_notready(m, total_pages - npages);
2645 }
2646
2647 sorele(so);
2648 CURVNET_RESTORE();
2649}
2650
2651static int
2652ktls_bind_domain(int domain)
2653{
2654 int error;
2655
2656 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
2657 if (error != 0)
2658 return (error);
2659 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
2660 return (0);
2661}
2662
2663static void
2664ktls_alloc_thread(void *ctx)
2665{
2666 struct ktls_domain_info *ktls_domain = ctx;
2667 struct ktls_alloc_thread *sc = &ktls_domain->alloc_td;
2668 void **buf;
2669 struct sysctl_oid *oid;
2670 char name[80];
2671 int domain, error, i, nbufs;
2672
2673 domain = ktls_domain - ktls_domains;
2674 if (bootverbose)
2675 printf("Starting KTLS alloc thread for domain %d\n", domain);
2676 error = ktls_bind_domain(domain);
2677 if (error)
2678 printf("Unable to bind KTLS alloc thread for domain %d: error %d\n",
2679 domain, error);
2680 snprintf(name, sizeof(name), "domain%d", domain);
2681 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
2682 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2683 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs",
2684 CTLFLAG_RD, &sc->allocs, 0, "buffers allocated");
2685 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
2686 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
2687 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
2688 CTLFLAG_RD, &sc->running, 0, "thread running");
2689
2690 buf = NULL;
2691 nbufs = 0;
2692 for (;;) {
2693 atomic_store_int(&sc->running, 0);
2694 tsleep(sc, PZERO | PNOLOCK, "-", 0);
2695 atomic_store_int(&sc->running, 1);
2696 sc->wakeups++;
2697 if (nbufs != ktls_max_alloc) {
2698 free(buf, M_KTLS);
2699 nbufs = atomic_load_int(&ktls_max_alloc);
2700 buf = malloc(sizeof(void *) * nbufs, M_KTLS,
2701 M_WAITOK | M_ZERO);
2702 }
2703 /*
2704 * Below we allocate nbufs with different allocation
2705 * flags than we use when allocating normally during
2706 * encryption in the ktls worker thread. We specify
2707 * M_NORECLAIM in the worker thread. However, we omit
2708 * that flag here and add M_WAITOK so that the VM
2709 * system is permitted to perform expensive work to
2710 * defragment memory. We do this here, as it does not
2711 * matter if this thread blocks. If we block a ktls
2712 * worker thread, we risk developing backlogs of
2713 * buffers to be encrypted, leading to surges of
2714 * traffic and potential NIC output drops.
2715 */
2716 for (i = 0; i < nbufs; i++) {
2717 buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK);
2718 sc->allocs++;
2719 }
2720 for (i = 0; i < nbufs; i++) {
2721 uma_zfree(ktls_buffer_zone, buf[i]);
2722 buf[i] = NULL;
2723 }
2724 }
2725}
2726
2727static void
2728ktls_work_thread(void *ctx)
2729{
2730 struct ktls_wq *wq = ctx;
2731 struct mbuf *m, *n;
2732 struct socket *so, *son;
2733 STAILQ_HEAD(, mbuf) local_m_head;
2734 STAILQ_HEAD(, socket) local_so_head;
2735 int cpu;
2736
2737 cpu = wq - ktls_wq;
2738 if (bootverbose)
2739 printf("Starting KTLS worker thread for CPU %d\n", cpu);
2740
2741 /*
2742 * Bind to a core. If ktls_bind_threads is > 1, then
2743 * we bind to the NUMA domain instead.
2744 */
2745 if (ktls_bind_threads) {
2746 int error;
2747
2748 if (ktls_bind_threads > 1) {
2749 struct pcpu *pc = pcpu_find(cpu);
2750
2751 error = ktls_bind_domain(pc->pc_domain);
2752 } else {
2753 cpuset_t mask;
2754
2755 CPU_SETOF(cpu, &mask);
2756 error = cpuset_setthread(curthread->td_tid, &mask);
2757 }
2758 if (error)
2759 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
2760 cpu, error);
2761 }
2762#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
2763 fpu_kern_thread(0);
2764#endif
2765 for (;;) {
2766 mtx_lock(&wq->mtx);
2767 while (STAILQ_EMPTY(&wq->m_head) &&
2768 STAILQ_EMPTY(&wq->so_head)) {
2769 wq->running = false;
2770 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
2771 wq->running = true;
2772 }
2773
2774 STAILQ_INIT(&local_m_head);
2775 STAILQ_CONCAT(&local_m_head, &wq->m_head);
2776 STAILQ_INIT(&local_so_head);
2777 STAILQ_CONCAT(&local_so_head, &wq->so_head);
2778 mtx_unlock(&wq->mtx);
2779
2780 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
2781 if (m->m_epg_flags & EPG_FLAG_2FREE) {
2782 ktls_free(m->m_epg_tls);
2783 m_free_raw(m);
2784 } else {
2785 if (m->m_epg_tls->sync_dispatch)
2786 ktls_encrypt(wq, m);
2787 else
2788 ktls_encrypt_async(wq, m);
2789 counter_u64_add(ktls_cnt_tx_queued, -1);
2790 }
2791 }
2792
2793 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
2794 ktls_decrypt(so);
2795 counter_u64_add(ktls_cnt_rx_queued, -1);
2796 }
2797 }
2798}
2799
2800#if defined(INET) || defined(INET6)
2801static void
2802ktls_disable_ifnet_help(void *context, int pending __unused)
2803{
2804 struct ktls_session *tls;
2805 struct inpcb *inp;
2806 struct tcpcb *tp;
2807 struct socket *so;
2808 int err;
2809
2810 tls = context;
2811 inp = tls->inp;
2812 if (inp == NULL)
2813 return;
2814 INP_WLOCK(inp);
2815 so = inp->inp_socket;
2816 MPASS(so != NULL);
2817 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2818 goto out;
2819 }
2820
2821 if (so->so_snd.sb_tls_info != NULL)
2822 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
2823 else
2824 err = ENXIO;
2825 if (err == 0) {
2826 counter_u64_add(ktls_ifnet_disable_ok, 1);
2827 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
2828 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 &&
2829 (tp = intotcpcb(inp)) != NULL &&
2830 tp->t_fb->tfb_hwtls_change != NULL)
2831 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
2832 } else {
2833 counter_u64_add(ktls_ifnet_disable_fail, 1);
2834 }
2835
2836out:
2837 sorele(so);
2838 if (!in_pcbrele_wlocked(inp))
2839 INP_WUNLOCK(inp);
2840 ktls_free(tls);
2841}
2842
2843/*
2844 * Called when re-transmits are becoming a substantial portion of the
2845 * sends on this connection. When this happens, we transition the
2846 * connection to software TLS. This is needed because most inline TLS
2847 * NICs keep crypto state only for in-order transmits. This means
2848 * that to handle a TCP rexmit (which is out-of-order), the NIC must
2849 * re-DMA the entire TLS record up to and including the current
2850 * segment. This means that when re-transmitting the last ~1448 byte
2851 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
2852 * of magnitude more data than we are sending. This can cause the
2853 * PCIe link to saturate well before the network, which can cause
2854 * output drops, and a general loss of capacity.
2855 */
2856void
2857ktls_disable_ifnet(void *arg)
2858{
2859 struct tcpcb *tp;
2860 struct inpcb *inp;
2861 struct socket *so;
2862 struct ktls_session *tls;
2863
2864 tp = arg;
2865 inp = tp->t_inpcb;
2866 INP_WLOCK_ASSERT(inp);
2867 so = inp->inp_socket;
2868 SOCK_LOCK(so);
2869 tls = so->so_snd.sb_tls_info;
2870 if (tls->disable_ifnet_pending) {
2871 SOCK_UNLOCK(so);
2872 return;
2873 }
2874
2875 /*
2876 * note that disable_ifnet_pending is never cleared; disabling
2877 * ifnet can only be done once per session, so we never want
2878 * to do it again
2879 */
2880
2881 (void)ktls_hold(tls);
2882 in_pcbref(inp);
2883 soref(so);
2884 tls->disable_ifnet_pending = true;
2885 tls->inp = inp;
2886 SOCK_UNLOCK(so);
2887 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
2888 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
2889}
2890#endif
count
int * count
Definition: cpufreq_if.m:63
name
const char * name
Definition: kern_fail.c:145
bootverbose
int bootverbose
Definition: init_main.c:131
ticks
volatile int ticks
Definition: kern_clock.c:380
STAILQ_HEAD
static STAILQ_HEAD(cn_device)
Definition: kern_cons.c:88
cpuset_domain
cpuset_t cpuset_domain[MAXMEMDOM]
Definition: kern_cpuset.c:142
cpuset_setthread
int cpuset_setthread(lwpid_t id, cpuset_t *mask)
Definition: kern_cpuset.c:1502
kproc_kthread_add
int kproc_kthread_add(void(*func)(void *), void *arg, struct proc **procptr, struct thread **tdptr, int flags, int pages, const char *procname, const char *fmt,...)
Definition: kern_kthread.c:455
malloc
void *() malloc(size_t size, struct malloc_type *mtp, int flags)
Definition: kern_malloc.c:632
zfree
void zfree(void *addr, struct malloc_type *mtp)
Definition: kern_malloc.c:947
free
void free(void *addr, struct malloc_type *mtp)
Definition: kern_malloc.c:907
m_freem
void m_freem(struct mbuf *mb)
Definition: kern_mbuf.c:1587
m_free_raw
void m_free_raw(struct mbuf *mb)
Definition: kern_mbuf.c:1599
m_snd_tag_alloc
int m_snd_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params, struct m_snd_tag **mstp)
Definition: kern_mbuf.c:1606
mb_free_notready
void mb_free_notready(struct mbuf *m, int count)
Definition: kern_mbuf.c:859
mb_use_ext_pgs
bool mb_use_ext_pgs
Definition: kern_mbuf.c:119
mtxpool_sleep
struct mtx_pool __read_mostly * mtxpool_sleep
Definition: kern_mtxpool.c:84
panic
void panic(const char *fmt,...)
Definition: kern_shutdown.c:884
wakeup
void wakeup(const void *ident)
Definition: kern_synch.c:349
start
void *** start
Definition: linker_if.m:98
domain
struct iommu_domain ** domain
Definition: msi_if.m:96
data
uint32_t * data
Definition: msi_if.m:90
ktls_alloc_thread
Definition: uipc_ktls.c:93
ktls_alloc_thread::td
struct thread * td
Definition: uipc_ktls.c:96
ktls_alloc_thread::allocs
uint64_t allocs
Definition: uipc_ktls.c:95
ktls_alloc_thread::running
int running
Definition: uipc_ktls.c:97
ktls_alloc_thread::wakeups
uint64_t wakeups
Definition: uipc_ktls.c:94
ktls_domain_info
Definition: uipc_ktls.c:100
ktls_domain_info::count
int count
Definition: uipc_ktls.c:101
ktls_domain_info::cpu
int cpu[MAXCPU]
Definition: uipc_ktls.c:102
ktls_domain_info::alloc_td
struct ktls_alloc_thread alloc_td
Definition: uipc_ktls.c:103
ktls_wq
Definition: uipc_ktls.c:85
ktls_wq::mtx
struct mtx mtx
Definition: uipc_ktls.c:86
mask
int mask
Definition: subr_acl_nfs4.c:70
hz
int hz
Definition: subr_param.c:85
pcpu_find
struct pcpu * pcpu_find(u_int cpuid)
Definition: subr_pcpu.c:283
printf
int printf(const char *fmt,...)
Definition: subr_prf.c:397
snprintf
int snprintf(char *str, size_t size, const char *format,...)
Definition: subr_prf.c:550
mp_maxid
u_int mp_maxid
Definition: subr_smp.c:77
flags
uint16_t flags
Definition: subr_stats.c:2
taskqueue_enqueue
int taskqueue_enqueue(struct taskqueue *queue, struct task *task)
Definition: subr_taskqueue.c:277
ktls_cpuid_lookup
static uint16_t ktls_cpuid_lookup[MAXCPU]
Definition: uipc_ktls.c:111
ktls_max_alloc
static int ktls_max_alloc
Definition: uipc_ktls.c:159
COUNTER_U64_DEFINE_EARLY
static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active)
ktls_seq
void ktls_seq(struct sockbuf *sb, struct mbuf *m)
Definition: uipc_ktls.c:1650
ktls_finish_nonanon
static void ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
Definition: uipc_ktls.c:2434
SYSCTL_COUNTER_U64
SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, &ktls_tasks_active, "Number of active tasks")
ktls_enqueue_to_free
void ktls_enqueue_to_free(struct mbuf *m)
Definition: uipc_ktls.c:2194
ktls_maxlen
static u_int ktls_maxlen
Definition: uipc_ktls.c:130
ktls_encrypt_record
static int ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m, struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
Definition: uipc_ktls.c:2251
ktls_encrypt_cb
void ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
Definition: uipc_ktls.c:2543
ktls_start_kthreads
static int ktls_start_kthreads(void)
Definition: uipc_ktls.c:476
ktls_init_state
static int ktls_init_state
Definition: uipc_ktls.c:112
ktls_domains
struct ktls_domain_info ktls_domains[MAXMEMDOM]
Definition: uipc_ktls.c:106
ktls_number_threads
static int ktls_number_threads
Definition: uipc_ktls.c:134
SYSCTL_INT
SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, &ktls_bind_threads, 0, "Bind crypto threads to cores (1) or cores and domains (2) at boot")
ktls_permit_empty_frames
bool ktls_permit_empty_frames(struct ktls_session *tls)
Definition: uipc_ktls.c:1800
ktls_bind_domain
static int ktls_bind_domain(int domain)
Definition: uipc_ktls.c:2652
ktls_check_rx
void ktls_check_rx(struct sockbuf *sb)
Definition: uipc_ktls.c:1807
tls13_find_record_type
static int tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, int *trailer_len, uint8_t *record_typep)
Definition: uipc_ktls.c:1963
ktls_session_zone
static uma_zone_t ktls_session_zone
Definition: uipc_ktls.c:109
ktls_destroy
void ktls_destroy(struct ktls_session *tls)
Definition: uipc_ktls.c:1628
ktls_cbc_enable
static bool ktls_cbc_enable
Definition: uipc_ktls.c:149
ktls_free_mext_contig
static void ktls_free_mext_contig(struct mbuf *m)
Definition: uipc_ktls.c:378
ktls_enqueue
void ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
Definition: uipc_ktls.c:2330
ktls_init_lock
static struct sx ktls_init_lock
Definition: uipc_ktls.c:113
mtx
struct mtx mtx
Definition: uipc_ktls.c:0
ktls_wq
static struct ktls_wq * ktls_wq
Definition: uipc_ktls.c:107
SYSCTL_BOOL
SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, &ktls_offload_enable, 0, "Enable support for kernel TLS offload")
__FBSDID
__FBSDID("$FreeBSD$")
__aligned
struct ktls_alloc_thread __aligned
ktls_cleanup
static void ktls_cleanup(struct ktls_session *tls)
Definition: uipc_ktls.c:781
SYSCTL_UINT
SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN, &ktls_maxlen, 0, "Maximum TLS record size")
ktls_proc
static struct proc * ktls_proc
Definition: uipc_ktls.c:108
ktls_alloc_thread
static void ktls_alloc_thread(void *ctx)
Definition: uipc_ktls.c:2664
ktls_encrypt_async
static __noinline void ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
Definition: uipc_ktls.c:2592
ktls_detach_record
static struct mbuf * ktls_detach_record(struct sockbuf *sb, int len)
Definition: uipc_ktls.c:1852
SYSCTL_NODE
SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW|CTLFLAG_MPSAFE, 0, "Kernel TLS offload")
ktls_buffer_import
static int ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
Definition: uipc_ktls.c:341
ktls_sw_buffer_cache
static bool ktls_sw_buffer_cache
Definition: uipc_ktls.c:154
ktls_offload_enable
static bool ktls_offload_enable
Definition: uipc_ktls.c:144
ktls_init
static int ktls_init(void)
Definition: uipc_ktls.c:385
ktls_encrypt
static __noinline void ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
Definition: uipc_ktls.c:2463
ktls_work_thread
static void ktls_work_thread(void *ctx)
Definition: uipc_ktls.c:2728
ktls_batched_records
static u_int ktls_batched_records(struct mbuf *m)
Definition: uipc_ktls.c:2314
ktls_buffer_zone
static uma_zone_t ktls_buffer_zone
Definition: uipc_ktls.c:110
ktls_frame
void ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, uint8_t record_type)
Definition: uipc_ktls.c:1678
ktls_ifnet_permitted
static int ktls_ifnet_permitted
Definition: uipc_ktls.c:276
MALLOC_DEFINE
static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS")
ktls_decrypt
static void ktls_decrypt(struct socket *so)
Definition: uipc_ktls.c:1996
ktls_ifnet_max_rexmit_pct
unsigned int ktls_ifnet_max_rexmit_pct
Definition: uipc_ktls.c:139
SX_SYSINIT
SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init")
ktls_buffer_release
static void ktls_buffer_release(void *arg __unused, void **store, int count)
Definition: uipc_ktls.c:363
ktls_buffer_alloc
static void * ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
Definition: uipc_ktls.c:2211
ktls_bind_threads
static int ktls_bind_threads
Definition: uipc_ktls.c:124
mb_free_mext_pgs
void mb_free_mext_pgs(struct mbuf *m)
Definition: uipc_mbuf.c:1754
m_copydata
void m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
Definition: uipc_mbuf.c:654
m_length
u_int m_length(struct mbuf *m0, struct mbuf **last)
Definition: uipc_mbuf.c:1445
mb_dupcl
void mb_dupcl(struct mbuf *n, struct mbuf *m)
Definition: uipc_mbuf.c:213
sbappendcontrol_locked
void sbappendcontrol_locked(struct sockbuf *sb, struct mbuf *m0, struct mbuf *control, int flags)
Definition: uipc_sockbuf.c:1242
sbcreatecontrol_how
struct mbuf * sbcreatecontrol_how(void *p, int size, int type, int level, int wait)
Definition: uipc_sockbuf.c:1723
buf
struct stat * buf
Definition: vfs_syscalls.c:2423
mode
mode_t mode
Definition: vfs_syscalls.c:2910