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 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_kern_tls.h"
32 #include "opt_ratelimit.h"
33 #include "opt_rss.h"
34
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/domainset.h>
38 #include <sys/endian.h>
39 #include <sys/ktls.h>
40 #include <sys/lock.h>
41 #include <sys/mbuf.h>
42 #include <sys/mutex.h>
43 #include <sys/rmlock.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/refcount.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/taskqueue.h>
52 #include <sys/kthread.h>
53 #include <sys/uio.h>
54 #include <sys/vmmeter.h>
55 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
56 #include <machine/pcb.h>
57 #endif
58 #include <machine/vmparam.h>
59 #include <net/if.h>
60 #include <net/if_var.h>
61 #ifdef RSS
62 #include <net/netisr.h>
63 #include <net/rss_config.h>
64 #endif
65 #include <net/route.h>
66 #include <net/route/nhop.h>
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_var.h>
70 #ifdef TCP_OFFLOAD
71 #include <netinet/tcp_offload.h>
72 #endif
73 #include <opencrypto/cryptodev.h>
74 #include <opencrypto/ktls.h>
75 #include <vm/vm.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_pagequeue.h>
79
80 struct ktls_wq {
81 struct mtx mtx;
82 STAILQ_HEAD(, mbuf) m_head;
83 STAILQ_HEAD(, socket) so_head;
84 bool running;
85 int lastallocfail;
86 } __aligned(CACHE_LINE_SIZE);
87
88 struct ktls_reclaim_thread {
89 uint64_t wakeups;
90 uint64_t reclaims;
91 struct thread *td;
92 int running;
93 };
94
95 struct ktls_domain_info {
96 int count;
97 int cpu[MAXCPU];
98 struct ktls_reclaim_thread reclaim_td;
99 };
100
101 struct ktls_domain_info ktls_domains[MAXMEMDOM];
102 static struct ktls_wq *ktls_wq;
103 static struct proc *ktls_proc;
104 static uma_zone_t ktls_session_zone;
105 static uma_zone_t ktls_buffer_zone;
106 static uint16_t ktls_cpuid_lookup[MAXCPU];
107 static int ktls_init_state;
108 static struct sx ktls_init_lock;
109 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
110
111 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
112 "Kernel TLS offload");
113 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114 "Kernel TLS offload stats");
115
116 #ifdef RSS
117 static int ktls_bind_threads = 1;
118 #else
119 static int ktls_bind_threads;
120 #endif
121 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
122 &ktls_bind_threads, 0,
123 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
124
125 static u_int ktls_maxlen = 16384;
126 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
127 &ktls_maxlen, 0, "Maximum TLS record size");
128
129 static int ktls_number_threads;
130 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
131 &ktls_number_threads, 0,
132 "Number of TLS threads in thread-pool");
133
134 unsigned int ktls_ifnet_max_rexmit_pct = 2;
135 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
136 &ktls_ifnet_max_rexmit_pct, 2,
137 "Max percent bytes retransmitted before ifnet TLS is disabled");
138
139 static bool ktls_offload_enable = true;
140 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
141 &ktls_offload_enable, 0,
142 "Enable support for kernel TLS offload");
143
144 static bool ktls_cbc_enable = true;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
146 &ktls_cbc_enable, 1,
147 "Enable support of AES-CBC crypto for kernel TLS");
148
149 static bool ktls_sw_buffer_cache = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
151 &ktls_sw_buffer_cache, 1,
152 "Enable caching of output buffers for SW encryption");
153
154 static int ktls_max_reclaim = 1024;
155 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN,
156 &ktls_max_reclaim, 128,
157 "Max number of 16k buffers to reclaim in thread context");
158
159 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
160 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
161 &ktls_tasks_active, "Number of active tasks");
162
163 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
164 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
165 &ktls_cnt_tx_pending,
166 "Number of TLS 1.0 records waiting for earlier TLS records");
167
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
170 &ktls_cnt_tx_queued,
171 "Number of TLS records in queue to tasks for SW encryption");
172
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
175 &ktls_cnt_rx_queued,
176 "Number of TLS sockets in queue to tasks for SW decryption");
177
178 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
180 CTLFLAG_RD, &ktls_offload_total,
181 "Total successful TLS setups (parameters set)");
182
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
185 CTLFLAG_RD, &ktls_offload_enable_calls,
186 "Total number of TLS enable calls made");
187
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
190 &ktls_offload_active, "Total Active TLS sessions");
191
192 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
193 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
194 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
195
196 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
197 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
198 &ktls_offload_failed_crypto, "Total TLS crypto failures");
199
200 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
201 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
202 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
203
204 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
205 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
206 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
207
208 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
209 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
210 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
211
212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
214 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
215
216 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
218 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
219
220 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
222 &ktls_destroy_task,
223 "Number of times ktls session was destroyed via taskqueue");
224
225 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226 "Software TLS session stats");
227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228 "Hardware (ifnet) TLS session stats");
229 #ifdef TCP_OFFLOAD
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231 "TOE TLS session stats");
232 #endif
233
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236 "Active number of software TLS sessions using AES-CBC");
237
238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240 "Active number of software TLS sessions using AES-GCM");
241
242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243 SYSCTL_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
247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248 SYSCTL_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
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253 SYSCTL_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
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259 &ktls_ifnet_chacha20,
260 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
261
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
265
266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268 &ktls_ifnet_reset_dropped,
269 "TLS sessions dropped after failing to update ifnet send tag");
270
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273 &ktls_ifnet_reset_failed,
274 "TLS sessions that failed to allocate a new ifnet send tag");
275
276 static int ktls_ifnet_permitted = 1;
277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278 &ktls_ifnet_permitted, 1,
279 "Whether to permit hardware (ifnet) TLS sessions");
280
281 #ifdef TCP_OFFLOAD
282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283 SYSCTL_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
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288 SYSCTL_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
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293 SYSCTL_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
298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
299
300 static void ktls_reclaim_thread(void *ctx);
301 static void ktls_reset_receive_tag(void *context, int pending);
302 static void ktls_reset_send_tag(void *context, int pending);
303 static void ktls_work_thread(void *ctx);
304
305 int
ktls_copyin_tls_enable(struct sockopt * sopt,struct tls_enable * tls)306 ktls_copyin_tls_enable(struct sockopt *sopt, struct tls_enable *tls)
307 {
308 struct tls_enable_v0 tls_v0;
309 int error;
310 uint8_t *cipher_key = NULL, *iv = NULL, *auth_key = NULL;
311
312 if (sopt->sopt_valsize == sizeof(tls_v0)) {
313 error = sooptcopyin(sopt, &tls_v0, sizeof(tls_v0), sizeof(tls_v0));
314 if (error != 0)
315 goto done;
316 memset(tls, 0, sizeof(*tls));
317 tls->cipher_key = tls_v0.cipher_key;
318 tls->iv = tls_v0.iv;
319 tls->auth_key = tls_v0.auth_key;
320 tls->cipher_algorithm = tls_v0.cipher_algorithm;
321 tls->cipher_key_len = tls_v0.cipher_key_len;
322 tls->iv_len = tls_v0.iv_len;
323 tls->auth_algorithm = tls_v0.auth_algorithm;
324 tls->auth_key_len = tls_v0.auth_key_len;
325 tls->flags = tls_v0.flags;
326 tls->tls_vmajor = tls_v0.tls_vmajor;
327 tls->tls_vminor = tls_v0.tls_vminor;
328 } else
329 error = sooptcopyin(sopt, tls, sizeof(*tls), sizeof(*tls));
330
331 if (error != 0)
332 return (error);
333
334 if (tls->cipher_key_len < 0 || tls->cipher_key_len > TLS_MAX_PARAM_SIZE)
335 return (EINVAL);
336 if (tls->iv_len < 0 || tls->iv_len > sizeof(((struct ktls_session *)NULL)->params.iv))
337 return (EINVAL);
338 if (tls->auth_key_len < 0 || tls->auth_key_len > TLS_MAX_PARAM_SIZE)
339 return (EINVAL);
340
341 /* All supported algorithms require a cipher key. */
342 if (tls->cipher_key_len == 0)
343 return (EINVAL);
344
345 /*
346 * Now do a deep copy of the variable-length arrays in the struct, so that
347 * subsequent consumers of it can reliably assume kernel memory. This
348 * requires doing our own allocations, which we will free in the
349 * error paths so that our caller need only worry about outstanding
350 * allocations existing on successful return.
351 */
352 if (tls->cipher_key_len != 0) {
353 cipher_key = malloc(tls->cipher_key_len, M_KTLS, M_WAITOK);
354 if (sopt->sopt_td != NULL) {
355 error = copyin(tls->cipher_key, cipher_key, tls->cipher_key_len);
356 if (error != 0)
357 goto done;
358 } else {
359 bcopy(tls->cipher_key, cipher_key, tls->cipher_key_len);
360 }
361 }
362 if (tls->iv_len != 0) {
363 iv = malloc(tls->iv_len, M_KTLS, M_WAITOK);
364 if (sopt->sopt_td != NULL) {
365 error = copyin(tls->iv, iv, tls->iv_len);
366 if (error != 0)
367 goto done;
368 } else {
369 bcopy(tls->iv, iv, tls->iv_len);
370 }
371 }
372 if (tls->auth_key_len != 0) {
373 auth_key = malloc(tls->auth_key_len, M_KTLS, M_WAITOK);
374 if (sopt->sopt_td != NULL) {
375 error = copyin(tls->auth_key, auth_key, tls->auth_key_len);
376 if (error != 0)
377 goto done;
378 } else {
379 bcopy(tls->auth_key, auth_key, tls->auth_key_len);
380 }
381 }
382 tls->cipher_key = cipher_key;
383 tls->iv = iv;
384 tls->auth_key = auth_key;
385
386 done:
387 if (error != 0) {
388 zfree(cipher_key, M_KTLS);
389 zfree(iv, M_KTLS);
390 zfree(auth_key, M_KTLS);
391 }
392
393 return (error);
394 }
395
396 void
ktls_cleanup_tls_enable(struct tls_enable * tls)397 ktls_cleanup_tls_enable(struct tls_enable *tls)
398 {
399 zfree(__DECONST(void *, tls->cipher_key), M_KTLS);
400 zfree(__DECONST(void *, tls->iv), M_KTLS);
401 zfree(__DECONST(void *, tls->auth_key), M_KTLS);
402 }
403
404 static u_int
ktls_get_cpu(struct socket * so)405 ktls_get_cpu(struct socket *so)
406 {
407 struct inpcb *inp;
408 #ifdef NUMA
409 struct ktls_domain_info *di;
410 #endif
411 u_int cpuid;
412
413 inp = sotoinpcb(so);
414 #ifdef RSS
415 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
416 if (cpuid != NETISR_CPUID_NONE)
417 return (cpuid);
418 #endif
419 /*
420 * Just use the flowid to shard connections in a repeatable
421 * fashion. Note that TLS 1.0 sessions rely on the
422 * serialization provided by having the same connection use
423 * the same queue.
424 */
425 #ifdef NUMA
426 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
427 di = &ktls_domains[inp->inp_numa_domain];
428 cpuid = di->cpu[inp->inp_flowid % di->count];
429 } else
430 #endif
431 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
432 return (cpuid);
433 }
434
435 static int
ktls_buffer_import(void * arg,void ** store,int count,int domain,int flags)436 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
437 {
438 vm_page_t m;
439 int i, req;
440
441 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
442 ("%s: ktls max length %d is not page size-aligned",
443 __func__, ktls_maxlen));
444
445 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
446 for (i = 0; i < count; i++) {
447 m = vm_page_alloc_noobj_contig_domain(domain, req,
448 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
449 VM_MEMATTR_DEFAULT);
450 if (m == NULL)
451 break;
452 store[i] = VM_PAGE_TO_DMAP(m);
453 }
454 return (i);
455 }
456
457 static void
ktls_buffer_release(void * arg __unused,void ** store,int count)458 ktls_buffer_release(void *arg __unused, void **store, int count)
459 {
460 vm_page_t m;
461 int i, j;
462
463 for (i = 0; i < count; i++) {
464 m = DMAP_TO_VM_PAGE(store[i]);
465 for (j = 0; j < atop(ktls_maxlen); j++) {
466 (void)vm_page_unwire_noq(m + j);
467 vm_page_free(m + j);
468 }
469 }
470 }
471
472 static void
ktls_free_mext_contig(struct mbuf * m)473 ktls_free_mext_contig(struct mbuf *m)
474 {
475 M_ASSERTEXTPG(m);
476 uma_zfree(ktls_buffer_zone, PHYS_TO_DMAP(m->m_epg_pa[0]));
477 }
478
479 static int
ktls_init(void)480 ktls_init(void)
481 {
482 struct thread *td;
483 struct pcpu *pc;
484 int count, domain, error, i;
485
486 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
487 M_WAITOK | M_ZERO);
488
489 ktls_session_zone = uma_zcreate("ktls_session",
490 sizeof(struct ktls_session),
491 NULL, NULL, NULL, NULL,
492 UMA_ALIGN_CACHE, 0);
493
494 if (ktls_sw_buffer_cache) {
495 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
496 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
497 ktls_buffer_import, ktls_buffer_release, NULL,
498 UMA_ZONE_FIRSTTOUCH | UMA_ZONE_NOTRIM);
499 }
500
501 /*
502 * Initialize the workqueues to run the TLS work. We create a
503 * work queue for each CPU.
504 */
505 CPU_FOREACH(i) {
506 STAILQ_INIT(&ktls_wq[i].m_head);
507 STAILQ_INIT(&ktls_wq[i].so_head);
508 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
509 if (ktls_bind_threads > 1) {
510 pc = pcpu_find(i);
511 domain = pc->pc_domain;
512 count = ktls_domains[domain].count;
513 ktls_domains[domain].cpu[count] = i;
514 ktls_domains[domain].count++;
515 }
516 ktls_cpuid_lookup[ktls_number_threads] = i;
517 ktls_number_threads++;
518 }
519
520 /*
521 * If we somehow have an empty domain, fall back to choosing
522 * among all KTLS threads.
523 */
524 if (ktls_bind_threads > 1) {
525 for (i = 0; i < vm_ndomains; i++) {
526 if (ktls_domains[i].count == 0) {
527 ktls_bind_threads = 1;
528 break;
529 }
530 }
531 }
532
533 /* Start kthreads for each workqueue. */
534 CPU_FOREACH(i) {
535 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
536 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
537 if (error) {
538 printf("Can't add KTLS thread %d error %d\n", i, error);
539 return (error);
540 }
541 }
542
543 /*
544 * Start an allocation thread per-domain to perform blocking allocations
545 * of 16k physically contiguous TLS crypto destination buffers.
546 */
547 if (ktls_sw_buffer_cache) {
548 for (domain = 0; domain < vm_ndomains; domain++) {
549 if (VM_DOMAIN_EMPTY(domain))
550 continue;
551 if (CPU_EMPTY(&cpuset_domain[domain]))
552 continue;
553 error = kproc_kthread_add(ktls_reclaim_thread,
554 &ktls_domains[domain], &ktls_proc,
555 &ktls_domains[domain].reclaim_td.td,
556 0, 0, "KTLS", "reclaim_%d", domain);
557 if (error) {
558 printf("Can't add KTLS reclaim thread %d error %d\n",
559 domain, error);
560 return (error);
561 }
562 }
563 }
564
565 if (bootverbose)
566 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
567 return (0);
568 }
569
570 static int
ktls_start_kthreads(void)571 ktls_start_kthreads(void)
572 {
573 int error, state;
574
575 start:
576 state = atomic_load_acq_int(&ktls_init_state);
577 if (__predict_true(state > 0))
578 return (0);
579 if (state < 0)
580 return (ENXIO);
581
582 sx_xlock(&ktls_init_lock);
583 if (ktls_init_state != 0) {
584 sx_xunlock(&ktls_init_lock);
585 goto start;
586 }
587
588 error = ktls_init();
589 if (error == 0)
590 state = 1;
591 else
592 state = -1;
593 atomic_store_rel_int(&ktls_init_state, state);
594 sx_xunlock(&ktls_init_lock);
595 return (error);
596 }
597
598 static int
ktls_create_session(struct socket * so,struct tls_enable * en,struct ktls_session ** tlsp,int direction)599 ktls_create_session(struct socket *so, struct tls_enable *en,
600 struct ktls_session **tlsp, int direction)
601 {
602 struct ktls_session *tls;
603 int error;
604
605 /* Only TLS 1.0 - 1.3 are supported. */
606 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
607 return (EINVAL);
608 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
609 en->tls_vminor > TLS_MINOR_VER_THREE)
610 return (EINVAL);
611
612
613 /* No flags are currently supported. */
614 if (en->flags != 0)
615 return (EINVAL);
616
617 /* Common checks for supported algorithms. */
618 switch (en->cipher_algorithm) {
619 case CRYPTO_AES_NIST_GCM_16:
620 /*
621 * auth_algorithm isn't used, but permit GMAC values
622 * for compatibility.
623 */
624 switch (en->auth_algorithm) {
625 case 0:
626 #ifdef COMPAT_FREEBSD12
627 /* XXX: Really 13.0-current COMPAT. */
628 case CRYPTO_AES_128_NIST_GMAC:
629 case CRYPTO_AES_192_NIST_GMAC:
630 case CRYPTO_AES_256_NIST_GMAC:
631 #endif
632 break;
633 default:
634 return (EINVAL);
635 }
636 if (en->auth_key_len != 0)
637 return (EINVAL);
638 switch (en->tls_vminor) {
639 case TLS_MINOR_VER_TWO:
640 if (en->iv_len != TLS_AEAD_GCM_LEN)
641 return (EINVAL);
642 break;
643 case TLS_MINOR_VER_THREE:
644 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
645 return (EINVAL);
646 break;
647 default:
648 return (EINVAL);
649 }
650 break;
651 case CRYPTO_AES_CBC:
652 switch (en->auth_algorithm) {
653 case CRYPTO_SHA1_HMAC:
654 break;
655 case CRYPTO_SHA2_256_HMAC:
656 case CRYPTO_SHA2_384_HMAC:
657 if (en->tls_vminor != TLS_MINOR_VER_TWO)
658 return (EINVAL);
659 break;
660 default:
661 return (EINVAL);
662 }
663 if (en->auth_key_len == 0)
664 return (EINVAL);
665
666 /*
667 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
668 * use explicit IVs.
669 */
670 switch (en->tls_vminor) {
671 case TLS_MINOR_VER_ZERO:
672 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
673 return (EINVAL);
674 break;
675 case TLS_MINOR_VER_ONE:
676 case TLS_MINOR_VER_TWO:
677 /* Ignore any supplied IV. */
678 en->iv_len = 0;
679 break;
680 default:
681 return (EINVAL);
682 }
683 break;
684 case CRYPTO_CHACHA20_POLY1305:
685 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
686 return (EINVAL);
687 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
688 en->tls_vminor != TLS_MINOR_VER_THREE)
689 return (EINVAL);
690 if (en->iv_len != TLS_CHACHA20_IV_LEN)
691 return (EINVAL);
692 break;
693 default:
694 return (EINVAL);
695 }
696
697 error = ktls_start_kthreads();
698 if (error != 0)
699 return (error);
700
701 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
702
703 counter_u64_add(ktls_offload_active, 1);
704
705 refcount_init(&tls->refcount, 1);
706 if (direction == KTLS_RX) {
707 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
708 } else {
709 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
710 tls->inp = so->so_pcb;
711 in_pcbref(tls->inp);
712 tls->tx = true;
713 }
714
715 tls->wq_index = ktls_get_cpu(so);
716
717 tls->params.cipher_algorithm = en->cipher_algorithm;
718 tls->params.auth_algorithm = en->auth_algorithm;
719 tls->params.tls_vmajor = en->tls_vmajor;
720 tls->params.tls_vminor = en->tls_vminor;
721 tls->params.flags = en->flags;
722 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
723
724 /* Set the header and trailer lengths. */
725 tls->params.tls_hlen = sizeof(struct tls_record_layer);
726 switch (en->cipher_algorithm) {
727 case CRYPTO_AES_NIST_GCM_16:
728 /*
729 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
730 * nonce. TLS 1.3 uses a 12 byte implicit IV.
731 */
732 if (en->tls_vminor < TLS_MINOR_VER_THREE)
733 tls->params.tls_hlen += sizeof(uint64_t);
734 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
735 tls->params.tls_bs = 1;
736 break;
737 case CRYPTO_AES_CBC:
738 switch (en->auth_algorithm) {
739 case CRYPTO_SHA1_HMAC:
740 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
741 /* Implicit IV, no nonce. */
742 tls->sequential_records = true;
743 tls->next_seqno = be64dec(en->rec_seq);
744 STAILQ_INIT(&tls->pending_records);
745 } else {
746 tls->params.tls_hlen += AES_BLOCK_LEN;
747 }
748 tls->params.tls_tlen = AES_BLOCK_LEN +
749 SHA1_HASH_LEN;
750 break;
751 case CRYPTO_SHA2_256_HMAC:
752 tls->params.tls_hlen += AES_BLOCK_LEN;
753 tls->params.tls_tlen = AES_BLOCK_LEN +
754 SHA2_256_HASH_LEN;
755 break;
756 case CRYPTO_SHA2_384_HMAC:
757 tls->params.tls_hlen += AES_BLOCK_LEN;
758 tls->params.tls_tlen = AES_BLOCK_LEN +
759 SHA2_384_HASH_LEN;
760 break;
761 default:
762 panic("invalid hmac");
763 }
764 tls->params.tls_bs = AES_BLOCK_LEN;
765 break;
766 case CRYPTO_CHACHA20_POLY1305:
767 /*
768 * Chacha20 uses a 12 byte implicit IV.
769 */
770 tls->params.tls_tlen = POLY1305_HASH_LEN;
771 tls->params.tls_bs = 1;
772 break;
773 default:
774 panic("invalid cipher");
775 }
776
777 /*
778 * TLS 1.3 includes optional padding which we do not support,
779 * and also puts the "real" record type at the end of the
780 * encrypted data.
781 */
782 if (en->tls_vminor == TLS_MINOR_VER_THREE)
783 tls->params.tls_tlen += sizeof(uint8_t);
784
785 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
786 ("TLS header length too long: %d", tls->params.tls_hlen));
787 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
788 ("TLS trailer length too long: %d", tls->params.tls_tlen));
789
790 if (en->auth_key_len != 0) {
791 tls->params.auth_key_len = en->auth_key_len;
792 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
793 M_WAITOK);
794 bcopy(en->auth_key, tls->params.auth_key, en->auth_key_len);
795 }
796
797 tls->params.cipher_key_len = en->cipher_key_len;
798 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
799 bcopy(en->cipher_key, tls->params.cipher_key, en->cipher_key_len);
800
801 /*
802 * This holds the implicit portion of the nonce for AEAD
803 * ciphers and the initial implicit IV for TLS 1.0. The
804 * explicit portions of the IV are generated in ktls_frame().
805 */
806 if (en->iv_len != 0) {
807 tls->params.iv_len = en->iv_len;
808 bcopy(en->iv, tls->params.iv, en->iv_len);
809
810 /*
811 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
812 * counter to generate unique explicit IVs.
813 *
814 * Store this counter in the last 8 bytes of the IV
815 * array so that it is 8-byte aligned.
816 */
817 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
818 en->tls_vminor == TLS_MINOR_VER_TWO)
819 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
820 }
821
822 tls->gen = 0;
823 *tlsp = tls;
824 return (0);
825 }
826
827 static struct ktls_session *
ktls_clone_session(struct ktls_session * tls,int direction)828 ktls_clone_session(struct ktls_session *tls, int direction)
829 {
830 struct ktls_session *tls_new;
831
832 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
833
834 counter_u64_add(ktls_offload_active, 1);
835
836 refcount_init(&tls_new->refcount, 1);
837 if (direction == KTLS_RX) {
838 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
839 tls_new);
840 } else {
841 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
842 tls_new);
843 tls_new->inp = tls->inp;
844 tls_new->tx = true;
845 in_pcbref(tls_new->inp);
846 }
847
848 /* Copy fields from existing session. */
849 tls_new->params = tls->params;
850 tls_new->wq_index = tls->wq_index;
851
852 /* Deep copy keys. */
853 if (tls_new->params.auth_key != NULL) {
854 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
855 M_KTLS, M_WAITOK);
856 memcpy(tls_new->params.auth_key, tls->params.auth_key,
857 tls->params.auth_key_len);
858 }
859
860 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
861 M_WAITOK);
862 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
863 tls->params.cipher_key_len);
864
865 tls_new->gen = 0;
866 return (tls_new);
867 }
868
869 #ifdef TCP_OFFLOAD
870 static int
ktls_try_toe(struct socket * so,struct ktls_session * tls,int direction)871 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
872 {
873 struct inpcb *inp = sotoinpcb(so);
874 struct tcpcb *tp = intotcpcb(inp);
875 int error;
876
877 INP_WLOCK(inp);
878 if (tp->t_flags & TF_DISCONNECTED) {
879 INP_WUNLOCK(inp);
880 return (ECONNRESET);
881 }
882 if (!(tp->t_flags & TF_TOE)) {
883 INP_WUNLOCK(inp);
884 return (EOPNOTSUPP);
885 }
886
887 error = tcp_offload_alloc_tls_session(tp, tls, direction);
888 INP_WUNLOCK(inp);
889 if (error == 0) {
890 tls->mode = TCP_TLS_MODE_TOE;
891 switch (tls->params.cipher_algorithm) {
892 case CRYPTO_AES_CBC:
893 counter_u64_add(ktls_toe_cbc, 1);
894 break;
895 case CRYPTO_AES_NIST_GCM_16:
896 counter_u64_add(ktls_toe_gcm, 1);
897 break;
898 case CRYPTO_CHACHA20_POLY1305:
899 counter_u64_add(ktls_toe_chacha20, 1);
900 break;
901 }
902 }
903 return (error);
904 }
905 #endif
906
907 /*
908 * Common code used when first enabling ifnet TLS on a connection or
909 * when allocating a new ifnet TLS session due to a routing change.
910 * This function allocates a new TLS send tag on whatever interface
911 * the connection is currently routed over.
912 */
913 static int
ktls_alloc_snd_tag(struct inpcb * inp,struct ktls_session * tls,bool force,struct m_snd_tag ** mstp)914 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
915 struct m_snd_tag **mstp)
916 {
917 union if_snd_tag_alloc_params params;
918 struct ifnet *ifp;
919 struct nhop_object *nh;
920 struct tcpcb *tp = intotcpcb(inp);
921 int error;
922
923 INP_RLOCK(inp);
924 if (tp->t_flags & TF_DISCONNECTED) {
925 INP_RUNLOCK(inp);
926 return (ECONNRESET);
927 }
928
929 /*
930 * Check administrative controls on ifnet TLS to determine if
931 * ifnet TLS should be denied.
932 *
933 * - Always permit 'force' requests.
934 * - ktls_ifnet_permitted == 0: always deny.
935 */
936 if (!force && ktls_ifnet_permitted == 0) {
937 INP_RUNLOCK(inp);
938 return (ENXIO);
939 }
940
941 /*
942 * XXX: Use the cached route in the inpcb to find the
943 * interface. This should perhaps instead use
944 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
945 * enabled after a connection has completed key negotiation in
946 * userland, the cached route will be present in practice.
947 */
948 nh = inp->inp_route.ro_nh;
949 if (nh == NULL) {
950 INP_RUNLOCK(inp);
951 return (ENXIO);
952 }
953 ifp = nh->nh_ifp;
954 if_ref(ifp);
955
956 /*
957 * Allocate a TLS + ratelimit tag if the connection has an
958 * existing pacing rate.
959 */
960 if (tp->t_pacing_rate != -1 &&
961 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
962 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
963 params.tls_rate_limit.inp = inp;
964 params.tls_rate_limit.tls = tls;
965 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
966 } else {
967 params.hdr.type = IF_SND_TAG_TYPE_TLS;
968 params.tls.inp = inp;
969 params.tls.tls = tls;
970 }
971 params.hdr.flowid = inp->inp_flowid;
972 params.hdr.flowtype = inp->inp_flowtype;
973 params.hdr.numa_domain = inp->inp_numa_domain;
974 INP_RUNLOCK(inp);
975
976 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
977 error = EOPNOTSUPP;
978 goto out;
979 }
980 if (inp->inp_vflag & INP_IPV6) {
981 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
982 error = EOPNOTSUPP;
983 goto out;
984 }
985 } else {
986 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
987 error = EOPNOTSUPP;
988 goto out;
989 }
990 }
991 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
992 out:
993 if_rele(ifp);
994 return (error);
995 }
996
997 /*
998 * Allocate an initial TLS receive tag for doing HW decryption of TLS
999 * data.
1000 *
1001 * This function allocates a new TLS receive tag on whatever interface
1002 * the connection is currently routed over. If the connection ends up
1003 * using a different interface for receive this will get fixed up via
1004 * ktls_input_ifp_mismatch as future packets arrive.
1005 */
1006 static int
ktls_alloc_rcv_tag(struct inpcb * inp,struct ktls_session * tls,struct m_snd_tag ** mstp)1007 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1008 struct m_snd_tag **mstp)
1009 {
1010 union if_snd_tag_alloc_params params;
1011 struct ifnet *ifp;
1012 struct nhop_object *nh;
1013 int error;
1014
1015 if (!ktls_ocf_recrypt_supported(tls))
1016 return (ENXIO);
1017
1018 INP_RLOCK(inp);
1019 if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
1020 INP_RUNLOCK(inp);
1021 return (ECONNRESET);
1022 }
1023
1024 /*
1025 * Check administrative controls on ifnet TLS to determine if
1026 * ifnet TLS should be denied.
1027 */
1028 if (ktls_ifnet_permitted == 0) {
1029 INP_RUNLOCK(inp);
1030 return (ENXIO);
1031 }
1032
1033 /*
1034 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1035 * the inpcb to find the interface.
1036 */
1037 nh = inp->inp_route.ro_nh;
1038 if (nh == NULL) {
1039 INP_RUNLOCK(inp);
1040 return (ENXIO);
1041 }
1042 ifp = nh->nh_ifp;
1043 if_ref(ifp);
1044 tls->rx_ifp = ifp;
1045
1046 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1047 params.hdr.flowid = inp->inp_flowid;
1048 params.hdr.flowtype = inp->inp_flowtype;
1049 params.hdr.numa_domain = inp->inp_numa_domain;
1050 params.tls_rx.inp = inp;
1051 params.tls_rx.tls = tls;
1052 params.tls_rx.vlan_id = 0;
1053
1054 INP_RUNLOCK(inp);
1055
1056 if (inp->inp_vflag & INP_IPV6) {
1057 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
1058 error = EOPNOTSUPP;
1059 goto out;
1060 }
1061 } else {
1062 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
1063 error = EOPNOTSUPP;
1064 goto out;
1065 }
1066 }
1067 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1068
1069 /*
1070 * If this connection is over a vlan, vlan_snd_tag_alloc
1071 * rewrites vlan_id with the saved interface. Save the VLAN
1072 * ID for use in ktls_reset_receive_tag which allocates new
1073 * receive tags directly from the leaf interface bypassing
1074 * if_vlan.
1075 */
1076 if (error == 0)
1077 tls->rx_vlan_id = params.tls_rx.vlan_id;
1078 out:
1079 return (error);
1080 }
1081
1082 static int
ktls_try_ifnet(struct socket * so,struct ktls_session * tls,int direction,bool force)1083 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1084 bool force)
1085 {
1086 struct m_snd_tag *mst;
1087 int error;
1088
1089 switch (direction) {
1090 case KTLS_TX:
1091 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1092 if (__predict_false(error != 0))
1093 goto done;
1094 break;
1095 case KTLS_RX:
1096 KASSERT(!force, ("%s: forced receive tag", __func__));
1097 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1098 if (__predict_false(error != 0))
1099 goto done;
1100 break;
1101 default:
1102 __assert_unreachable();
1103 }
1104
1105 tls->mode = TCP_TLS_MODE_IFNET;
1106 tls->snd_tag = mst;
1107
1108 switch (tls->params.cipher_algorithm) {
1109 case CRYPTO_AES_CBC:
1110 counter_u64_add(ktls_ifnet_cbc, 1);
1111 break;
1112 case CRYPTO_AES_NIST_GCM_16:
1113 counter_u64_add(ktls_ifnet_gcm, 1);
1114 break;
1115 case CRYPTO_CHACHA20_POLY1305:
1116 counter_u64_add(ktls_ifnet_chacha20, 1);
1117 break;
1118 default:
1119 break;
1120 }
1121 done:
1122 return (error);
1123 }
1124
1125 static void
ktls_use_sw(struct ktls_session * tls)1126 ktls_use_sw(struct ktls_session *tls)
1127 {
1128 tls->mode = TCP_TLS_MODE_SW;
1129 switch (tls->params.cipher_algorithm) {
1130 case CRYPTO_AES_CBC:
1131 counter_u64_add(ktls_sw_cbc, 1);
1132 break;
1133 case CRYPTO_AES_NIST_GCM_16:
1134 counter_u64_add(ktls_sw_gcm, 1);
1135 break;
1136 case CRYPTO_CHACHA20_POLY1305:
1137 counter_u64_add(ktls_sw_chacha20, 1);
1138 break;
1139 }
1140 }
1141
1142 static int
ktls_try_sw(struct ktls_session * tls,int direction)1143 ktls_try_sw(struct ktls_session *tls, int direction)
1144 {
1145 int error;
1146
1147 error = ktls_ocf_try(tls, direction);
1148 if (error)
1149 return (error);
1150 ktls_use_sw(tls);
1151 return (0);
1152 }
1153
1154 /*
1155 * KTLS RX stores data in the socket buffer as a list of TLS records,
1156 * where each record is stored as a control message containg the TLS
1157 * header followed by data mbufs containing the decrypted data. This
1158 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1159 * both encrypted and decrypted data. TLS records decrypted by a NIC
1160 * should be queued to the socket buffer as records, but encrypted
1161 * data which needs to be decrypted by software arrives as a stream of
1162 * regular mbufs which need to be converted. In addition, there may
1163 * already be pending encrypted data in the socket buffer when KTLS RX
1164 * is enabled.
1165 *
1166 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1167 * is used:
1168 *
1169 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1170 *
1171 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1172 * from the first mbuf. Once all of the data for that TLS record is
1173 * queued, the socket is queued to a worker thread.
1174 *
1175 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1176 * the TLS chain. Each TLS record is detached from the TLS chain,
1177 * decrypted, and inserted into the regular socket buffer chain as
1178 * record starting with a control message holding the TLS header and
1179 * a chain of mbufs holding the encrypted data.
1180 */
1181
1182 static void
sb_mark_notready(struct sockbuf * sb)1183 sb_mark_notready(struct sockbuf *sb)
1184 {
1185 struct mbuf *m;
1186
1187 m = sb->sb_mb;
1188 sb->sb_mtls = m;
1189 sb->sb_mb = NULL;
1190 sb->sb_mbtail = NULL;
1191 sb->sb_lastrecord = NULL;
1192 for (; m != NULL; m = m->m_next) {
1193 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1194 __func__));
1195 KASSERT((m->m_flags & M_NOTREADY) == 0, ("%s: mbuf not ready",
1196 __func__));
1197 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1198 __func__));
1199 m->m_flags |= M_NOTREADY;
1200 sb->sb_acc -= m->m_len;
1201 sb->sb_tlscc += m->m_len;
1202 sb->sb_mtlstail = m;
1203 }
1204 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1205 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1206 sb->sb_ccc));
1207 }
1208
1209 /*
1210 * Return information about the pending TLS data in a socket
1211 * buffer. On return, 'seqno' is set to the sequence number
1212 * of the next TLS record to be received, 'resid' is set to
1213 * the amount of bytes still needed for the last pending
1214 * record. The function returns 'false' if the last pending
1215 * record contains a partial TLS header. In that case, 'resid'
1216 * is the number of bytes needed to complete the TLS header.
1217 */
1218 bool
ktls_pending_rx_info(struct sockbuf * sb,uint64_t * seqnop,size_t * residp)1219 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1220 {
1221 struct tls_record_layer hdr;
1222 struct mbuf *m;
1223 uint64_t seqno;
1224 size_t resid;
1225 u_int offset, record_len;
1226
1227 SOCKBUF_LOCK_ASSERT(sb);
1228 MPASS(sb->sb_flags & SB_TLS_RX);
1229 seqno = sb->sb_tls_seqno;
1230 resid = sb->sb_tlscc;
1231 m = sb->sb_mtls;
1232 offset = 0;
1233
1234 if (resid == 0) {
1235 *seqnop = seqno;
1236 *residp = 0;
1237 return (true);
1238 }
1239
1240 for (;;) {
1241 seqno++;
1242
1243 if (resid < sizeof(hdr)) {
1244 *seqnop = seqno;
1245 *residp = sizeof(hdr) - resid;
1246 return (false);
1247 }
1248
1249 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1250
1251 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1252 if (resid <= record_len) {
1253 *seqnop = seqno;
1254 *residp = record_len - resid;
1255 return (true);
1256 }
1257 resid -= record_len;
1258
1259 while (record_len != 0) {
1260 if (m->m_len - offset > record_len) {
1261 offset += record_len;
1262 break;
1263 }
1264
1265 record_len -= (m->m_len - offset);
1266 offset = 0;
1267 m = m->m_next;
1268 }
1269 }
1270 }
1271
1272 int
ktls_enable_rx(struct socket * so,struct tls_enable * en)1273 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1274 {
1275 struct ktls_session *tls;
1276 int error;
1277
1278 if (!ktls_offload_enable)
1279 return (ENOTSUP);
1280
1281 counter_u64_add(ktls_offload_enable_calls, 1);
1282
1283 /*
1284 * This should always be true since only the TCP socket option
1285 * invokes this function.
1286 */
1287 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1288 return (EINVAL);
1289
1290 /*
1291 * XXX: Don't overwrite existing sessions. We should permit
1292 * this to support rekeying in the future.
1293 */
1294 if (so->so_rcv.sb_tls_info != NULL)
1295 return (EALREADY);
1296
1297 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1298 return (ENOTSUP);
1299
1300 error = ktls_create_session(so, en, &tls, KTLS_RX);
1301 if (error)
1302 return (error);
1303
1304 error = ktls_ocf_try(tls, KTLS_RX);
1305 if (error) {
1306 ktls_free(tls);
1307 return (error);
1308 }
1309
1310 /*
1311 * Serialize with soreceive_generic() and make sure that we're not
1312 * operating on a listening socket.
1313 */
1314 error = SOCK_IO_RECV_LOCK(so, SBL_WAIT);
1315 if (error) {
1316 ktls_free(tls);
1317 return (error);
1318 }
1319
1320 /* Mark the socket as using TLS offload. */
1321 SOCK_RECVBUF_LOCK(so);
1322 if (__predict_false(so->so_rcv.sb_tls_info != NULL))
1323 error = EALREADY;
1324 else if ((so->so_rcv.sb_flags & SB_SPLICED) != 0)
1325 error = EINVAL;
1326 if (error != 0) {
1327 SOCK_RECVBUF_UNLOCK(so);
1328 SOCK_IO_RECV_UNLOCK(so);
1329 ktls_free(tls);
1330 return (EALREADY);
1331 }
1332 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1333 so->so_rcv.sb_tls_info = tls;
1334 so->so_rcv.sb_flags |= SB_TLS_RX;
1335
1336 /* Mark existing data as not ready until it can be decrypted. */
1337 sb_mark_notready(&so->so_rcv);
1338 ktls_check_rx(&so->so_rcv);
1339 SOCK_RECVBUF_UNLOCK(so);
1340 SOCK_IO_RECV_UNLOCK(so);
1341
1342 /* Prefer TOE -> ifnet TLS -> software TLS. */
1343 #ifdef TCP_OFFLOAD
1344 error = ktls_try_toe(so, tls, KTLS_RX);
1345 if (error)
1346 #endif
1347 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1348 if (error)
1349 ktls_use_sw(tls);
1350
1351 counter_u64_add(ktls_offload_total, 1);
1352
1353 return (0);
1354 }
1355
1356 int
ktls_enable_tx(struct socket * so,struct tls_enable * en)1357 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1358 {
1359 struct ktls_session *tls;
1360 struct inpcb *inp;
1361 struct tcpcb *tp;
1362 int error;
1363
1364 if (!ktls_offload_enable)
1365 return (ENOTSUP);
1366
1367 counter_u64_add(ktls_offload_enable_calls, 1);
1368
1369 /*
1370 * This should always be true since only the TCP socket option
1371 * invokes this function.
1372 */
1373 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1374 return (EINVAL);
1375
1376 /*
1377 * XXX: Don't overwrite existing sessions. We should permit
1378 * this to support rekeying in the future.
1379 */
1380 if (so->so_snd.sb_tls_info != NULL)
1381 return (EALREADY);
1382
1383 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1384 return (ENOTSUP);
1385
1386 /* TLS requires ext pgs */
1387 if (mb_use_ext_pgs == 0)
1388 return (ENXIO);
1389
1390 error = ktls_create_session(so, en, &tls, KTLS_TX);
1391 if (error)
1392 return (error);
1393
1394 /* some ktls offload NICs require initial seqno to start offload */
1395 tls->initial_offload_seqno = be64dec(en->rec_seq);
1396
1397 /* Prefer TOE -> ifnet TLS -> software TLS. */
1398 #ifdef TCP_OFFLOAD
1399 error = ktls_try_toe(so, tls, KTLS_TX);
1400 if (error)
1401 #endif
1402 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1403 if (error)
1404 error = ktls_try_sw(tls, KTLS_TX);
1405
1406 if (error) {
1407 ktls_free(tls);
1408 return (error);
1409 }
1410
1411 /*
1412 * Serialize with sosend_generic() and make sure that we're not
1413 * operating on a listening socket.
1414 */
1415 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1416 if (error) {
1417 ktls_free(tls);
1418 return (error);
1419 }
1420
1421 /*
1422 * Write lock the INP when setting sb_tls_info so that
1423 * routines in tcp_ratelimit.c can read sb_tls_info while
1424 * holding the INP lock.
1425 */
1426 inp = so->so_pcb;
1427 INP_WLOCK(inp);
1428 SOCK_SENDBUF_LOCK(so);
1429 if (__predict_false(so->so_snd.sb_tls_info != NULL))
1430 error = EALREADY;
1431 else if ((so->so_snd.sb_flags & SB_SPLICED) != 0)
1432 error = EINVAL;
1433 if (error != 0) {
1434 SOCK_SENDBUF_UNLOCK(so);
1435 INP_WUNLOCK(inp);
1436 SOCK_IO_SEND_UNLOCK(so);
1437 ktls_free(tls);
1438 return (error);
1439 }
1440 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1441 so->so_snd.sb_tls_info = tls;
1442 if (tls->mode != TCP_TLS_MODE_SW) {
1443 tp = intotcpcb(inp);
1444 MPASS(tp->t_nic_ktls_xmit == 0);
1445 tp->t_nic_ktls_xmit = 1;
1446 if (tp->t_fb->tfb_hwtls_change != NULL)
1447 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1448 }
1449 SOCK_SENDBUF_UNLOCK(so);
1450 INP_WUNLOCK(inp);
1451 SOCK_IO_SEND_UNLOCK(so);
1452
1453 counter_u64_add(ktls_offload_total, 1);
1454
1455 return (0);
1456 }
1457
1458 int
ktls_get_rx_mode(struct socket * so,int * modep)1459 ktls_get_rx_mode(struct socket *so, int *modep)
1460 {
1461 struct ktls_session *tls;
1462 struct inpcb *inp __diagused;
1463
1464 if (SOLISTENING(so))
1465 return (EINVAL);
1466 inp = so->so_pcb;
1467 INP_WLOCK_ASSERT(inp);
1468 SOCK_RECVBUF_LOCK(so);
1469 tls = so->so_rcv.sb_tls_info;
1470 if (tls == NULL)
1471 *modep = TCP_TLS_MODE_NONE;
1472 else
1473 *modep = tls->mode;
1474 SOCK_RECVBUF_UNLOCK(so);
1475 return (0);
1476 }
1477
1478 /*
1479 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1480 *
1481 * This function gets information about the next TCP- and TLS-
1482 * sequence number to be processed by the TLS receive worker
1483 * thread. The information is extracted from the given "inpcb"
1484 * structure. The values are stored in host endian format at the two
1485 * given output pointer locations. The TCP sequence number points to
1486 * the beginning of the TLS header.
1487 *
1488 * This function returns zero on success, else a non-zero error code
1489 * is returned.
1490 */
1491 int
ktls_get_rx_sequence(struct inpcb * inp,uint32_t * tcpseq,uint64_t * tlsseq)1492 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1493 {
1494 struct socket *so = inp->inp_socket;
1495 struct tcpcb *tp = intotcpcb(inp);
1496
1497 INP_RLOCK(inp);
1498 if (tp->t_flags & TF_DISCONNECTED) {
1499 INP_RUNLOCK(inp);
1500 return (ECONNRESET);
1501 }
1502
1503 SOCKBUF_LOCK(&so->so_rcv);
1504 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1505 *tlsseq = so->so_rcv.sb_tls_seqno;
1506 SOCKBUF_UNLOCK(&so->so_rcv);
1507
1508 INP_RUNLOCK(inp);
1509
1510 return (0);
1511 }
1512
1513 int
ktls_get_tx_mode(struct socket * so,int * modep)1514 ktls_get_tx_mode(struct socket *so, int *modep)
1515 {
1516 struct ktls_session *tls;
1517 struct inpcb *inp __diagused;
1518
1519 if (SOLISTENING(so))
1520 return (EINVAL);
1521 inp = so->so_pcb;
1522 INP_WLOCK_ASSERT(inp);
1523 SOCK_SENDBUF_LOCK(so);
1524 tls = so->so_snd.sb_tls_info;
1525 if (tls == NULL)
1526 *modep = TCP_TLS_MODE_NONE;
1527 else
1528 *modep = tls->mode;
1529 SOCK_SENDBUF_UNLOCK(so);
1530 return (0);
1531 }
1532
1533 /*
1534 * Switch between SW and ifnet TLS sessions as requested.
1535 */
1536 int
ktls_set_tx_mode(struct socket * so,int mode)1537 ktls_set_tx_mode(struct socket *so, int mode)
1538 {
1539 struct ktls_session *tls, *tls_new;
1540 struct inpcb *inp;
1541 struct tcpcb *tp;
1542 int error;
1543
1544 if (SOLISTENING(so))
1545 return (EINVAL);
1546 switch (mode) {
1547 case TCP_TLS_MODE_SW:
1548 case TCP_TLS_MODE_IFNET:
1549 break;
1550 default:
1551 return (EINVAL);
1552 }
1553
1554 inp = so->so_pcb;
1555 INP_WLOCK_ASSERT(inp);
1556 tp = intotcpcb(inp);
1557
1558 if (mode == TCP_TLS_MODE_IFNET) {
1559 /* Don't allow enabling ifnet ktls multiple times */
1560 if (tp->t_nic_ktls_xmit)
1561 return (EALREADY);
1562
1563 /*
1564 * Don't enable ifnet ktls if we disabled it due to an
1565 * excessive retransmission rate
1566 */
1567 if (tp->t_nic_ktls_xmit_dis)
1568 return (ENXIO);
1569 }
1570
1571 SOCKBUF_LOCK(&so->so_snd);
1572 tls = so->so_snd.sb_tls_info;
1573 if (tls == NULL) {
1574 SOCKBUF_UNLOCK(&so->so_snd);
1575 return (0);
1576 }
1577
1578 if (tls->mode == mode) {
1579 SOCKBUF_UNLOCK(&so->so_snd);
1580 return (0);
1581 }
1582
1583 tls = ktls_hold(tls);
1584 SOCKBUF_UNLOCK(&so->so_snd);
1585 INP_WUNLOCK(inp);
1586
1587 tls_new = ktls_clone_session(tls, KTLS_TX);
1588
1589 if (mode == TCP_TLS_MODE_IFNET)
1590 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1591 else
1592 error = ktls_try_sw(tls_new, KTLS_TX);
1593 if (error) {
1594 counter_u64_add(ktls_switch_failed, 1);
1595 ktls_free(tls_new);
1596 ktls_free(tls);
1597 INP_WLOCK(inp);
1598 return (error);
1599 }
1600
1601 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1602 if (error) {
1603 counter_u64_add(ktls_switch_failed, 1);
1604 ktls_free(tls_new);
1605 ktls_free(tls);
1606 INP_WLOCK(inp);
1607 return (error);
1608 }
1609
1610 /*
1611 * If we raced with another session change, keep the existing
1612 * session.
1613 */
1614 if (tls != so->so_snd.sb_tls_info) {
1615 counter_u64_add(ktls_switch_failed, 1);
1616 SOCK_IO_SEND_UNLOCK(so);
1617 ktls_free(tls_new);
1618 ktls_free(tls);
1619 INP_WLOCK(inp);
1620 return (EBUSY);
1621 }
1622
1623 INP_WLOCK(inp);
1624 SOCKBUF_LOCK(&so->so_snd);
1625 so->so_snd.sb_tls_info = tls_new;
1626 if (tls_new->mode != TCP_TLS_MODE_SW) {
1627 MPASS(tp->t_nic_ktls_xmit == 0);
1628 tp->t_nic_ktls_xmit = 1;
1629 if (tp->t_fb->tfb_hwtls_change != NULL)
1630 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1631 }
1632 SOCKBUF_UNLOCK(&so->so_snd);
1633 SOCK_IO_SEND_UNLOCK(so);
1634
1635 /*
1636 * Drop two references on 'tls'. The first is for the
1637 * ktls_hold() above. The second drops the reference from the
1638 * socket buffer.
1639 */
1640 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1641 ktls_free(tls);
1642 ktls_free(tls);
1643
1644 if (mode == TCP_TLS_MODE_IFNET)
1645 counter_u64_add(ktls_switch_to_ifnet, 1);
1646 else
1647 counter_u64_add(ktls_switch_to_sw, 1);
1648
1649 return (0);
1650 }
1651
1652 /*
1653 * Try to allocate a new TLS receive tag. This task is scheduled when
1654 * sbappend_ktls_rx detects an input path change. If a new tag is
1655 * allocated, replace the tag in the TLS session. If a new tag cannot
1656 * be allocated, let the session fall back to software decryption.
1657 */
1658 static void
ktls_reset_receive_tag(void * context,int pending)1659 ktls_reset_receive_tag(void *context, int pending)
1660 {
1661 union if_snd_tag_alloc_params params;
1662 struct ktls_session *tls;
1663 struct m_snd_tag *mst;
1664 struct inpcb *inp;
1665 struct ifnet *ifp;
1666 struct socket *so;
1667 int error;
1668
1669 MPASS(pending == 1);
1670
1671 tls = context;
1672 so = tls->so;
1673 inp = so->so_pcb;
1674 ifp = NULL;
1675
1676 INP_RLOCK(inp);
1677 if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
1678 INP_RUNLOCK(inp);
1679 goto out;
1680 }
1681
1682 SOCKBUF_LOCK(&so->so_rcv);
1683 mst = tls->snd_tag;
1684 tls->snd_tag = NULL;
1685 if (mst != NULL)
1686 m_snd_tag_rele(mst);
1687
1688 ifp = tls->rx_ifp;
1689 if_ref(ifp);
1690 SOCKBUF_UNLOCK(&so->so_rcv);
1691
1692 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1693 params.hdr.flowid = inp->inp_flowid;
1694 params.hdr.flowtype = inp->inp_flowtype;
1695 params.hdr.numa_domain = inp->inp_numa_domain;
1696 params.tls_rx.inp = inp;
1697 params.tls_rx.tls = tls;
1698 params.tls_rx.vlan_id = tls->rx_vlan_id;
1699 INP_RUNLOCK(inp);
1700
1701 if (inp->inp_vflag & INP_IPV6) {
1702 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1703 goto out;
1704 } else {
1705 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1706 goto out;
1707 }
1708
1709 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1710 if (error == 0) {
1711 SOCKBUF_LOCK(&so->so_rcv);
1712 tls->snd_tag = mst;
1713 SOCKBUF_UNLOCK(&so->so_rcv);
1714
1715 counter_u64_add(ktls_ifnet_reset, 1);
1716 } else {
1717 /*
1718 * Just fall back to software decryption if a tag
1719 * cannot be allocated leaving the connection intact.
1720 * If a future input path change switches to another
1721 * interface this connection will resume ifnet TLS.
1722 */
1723 counter_u64_add(ktls_ifnet_reset_failed, 1);
1724 }
1725
1726 out:
1727 mtx_pool_lock(mtxpool_sleep, tls);
1728 tls->reset_pending = false;
1729 mtx_pool_unlock(mtxpool_sleep, tls);
1730
1731 if (ifp != NULL)
1732 if_rele(ifp);
1733 CURVNET_SET(so->so_vnet);
1734 sorele(so);
1735 CURVNET_RESTORE();
1736 ktls_free(tls);
1737 }
1738
1739 /*
1740 * Try to allocate a new TLS send tag. This task is scheduled when
1741 * ip_output detects a route change while trying to transmit a packet
1742 * holding a TLS record. If a new tag is allocated, replace the tag
1743 * in the TLS session. Subsequent packets on the connection will use
1744 * the new tag. If a new tag cannot be allocated, drop the
1745 * connection.
1746 */
1747 static void
ktls_reset_send_tag(void * context,int pending)1748 ktls_reset_send_tag(void *context, int pending)
1749 {
1750 struct epoch_tracker et;
1751 struct ktls_session *tls;
1752 struct m_snd_tag *old, *new;
1753 struct inpcb *inp;
1754 struct tcpcb *tp;
1755 int error;
1756
1757 MPASS(pending == 1);
1758
1759 tls = context;
1760 inp = tls->inp;
1761
1762 /*
1763 * Free the old tag first before allocating a new one.
1764 * ip[6]_output_send() will treat a NULL send tag the same as
1765 * an ifp mismatch and drop packets until a new tag is
1766 * allocated.
1767 *
1768 * Write-lock the INP when changing tls->snd_tag since
1769 * ip[6]_output_send() holds a read-lock when reading the
1770 * pointer.
1771 */
1772 INP_WLOCK(inp);
1773 old = tls->snd_tag;
1774 tls->snd_tag = NULL;
1775 INP_WUNLOCK(inp);
1776 if (old != NULL)
1777 m_snd_tag_rele(old);
1778
1779 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1780
1781 if (error == 0) {
1782 INP_WLOCK(inp);
1783 tls->snd_tag = new;
1784 mtx_pool_lock(mtxpool_sleep, tls);
1785 tls->reset_pending = false;
1786 mtx_pool_unlock(mtxpool_sleep, tls);
1787 INP_WUNLOCK(inp);
1788
1789 counter_u64_add(ktls_ifnet_reset, 1);
1790
1791 /*
1792 * XXX: Should we kick tcp_output explicitly now that
1793 * the send tag is fixed or just rely on timers?
1794 */
1795 } else {
1796 NET_EPOCH_ENTER(et);
1797 INP_WLOCK(inp);
1798 tp = intotcpcb(inp);
1799 if (!(tp->t_flags & TF_DISCONNECTED)) {
1800 CURVNET_SET(inp->inp_socket->so_vnet);
1801 tp = tcp_drop(tp, ECONNABORTED);
1802 CURVNET_RESTORE();
1803 if (tp != NULL) {
1804 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1805 INP_WUNLOCK(inp);
1806 }
1807 } else
1808 INP_WUNLOCK(inp);
1809 NET_EPOCH_EXIT(et);
1810
1811 counter_u64_add(ktls_ifnet_reset_failed, 1);
1812
1813 /*
1814 * Leave reset_pending true to avoid future tasks while
1815 * the socket goes away.
1816 */
1817 }
1818
1819 ktls_free(tls);
1820 }
1821
1822 void
ktls_input_ifp_mismatch(struct sockbuf * sb,struct ifnet * ifp)1823 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1824 {
1825 struct ktls_session *tls;
1826 struct socket *so;
1827
1828 SOCKBUF_LOCK_ASSERT(sb);
1829 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1830 __func__, sb));
1831 so = __containerof(sb, struct socket, so_rcv);
1832
1833 tls = sb->sb_tls_info;
1834 if_rele(tls->rx_ifp);
1835 if_ref(ifp);
1836 tls->rx_ifp = ifp;
1837
1838 /*
1839 * See if we should schedule a task to update the receive tag for
1840 * this session.
1841 */
1842 mtx_pool_lock(mtxpool_sleep, tls);
1843 if (!tls->reset_pending) {
1844 (void) ktls_hold(tls);
1845 soref(so);
1846 tls->so = so;
1847 tls->reset_pending = true;
1848 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1849 }
1850 mtx_pool_unlock(mtxpool_sleep, tls);
1851 }
1852
1853 int
ktls_output_eagain(struct inpcb * inp,struct ktls_session * tls)1854 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1855 {
1856
1857 if (inp == NULL)
1858 return (ENOBUFS);
1859
1860 INP_LOCK_ASSERT(inp);
1861
1862 /*
1863 * See if we should schedule a task to update the send tag for
1864 * this session.
1865 */
1866 mtx_pool_lock(mtxpool_sleep, tls);
1867 if (!tls->reset_pending) {
1868 (void) ktls_hold(tls);
1869 tls->reset_pending = true;
1870 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1871 }
1872 mtx_pool_unlock(mtxpool_sleep, tls);
1873 return (ENOBUFS);
1874 }
1875
1876 #ifdef RATELIMIT
1877 int
ktls_modify_txrtlmt(struct ktls_session * tls,uint64_t max_pacing_rate)1878 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1879 {
1880 union if_snd_tag_modify_params params = {
1881 .rate_limit.max_rate = max_pacing_rate,
1882 .rate_limit.flags = M_NOWAIT,
1883 };
1884 struct m_snd_tag *mst;
1885
1886 /* Can't get to the inp, but it should be locked. */
1887 /* INP_LOCK_ASSERT(inp); */
1888
1889 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1890
1891 if (tls->snd_tag == NULL) {
1892 /*
1893 * Resetting send tag, ignore this change. The
1894 * pending reset may or may not see this updated rate
1895 * in the tcpcb. If it doesn't, we will just lose
1896 * this rate change.
1897 */
1898 return (0);
1899 }
1900
1901 mst = tls->snd_tag;
1902
1903 MPASS(mst != NULL);
1904 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1905
1906 return (mst->sw->snd_tag_modify(mst, ¶ms));
1907 }
1908 #endif
1909
1910 static void
ktls_destroy_help(void * context,int pending __unused)1911 ktls_destroy_help(void *context, int pending __unused)
1912 {
1913 ktls_destroy(context);
1914 }
1915
1916 void
ktls_destroy(struct ktls_session * tls)1917 ktls_destroy(struct ktls_session *tls)
1918 {
1919 struct inpcb *inp;
1920 struct tcpcb *tp;
1921 bool wlocked;
1922
1923 MPASS(tls->refcount == 0);
1924
1925 inp = tls->inp;
1926 if (tls->tx) {
1927 wlocked = INP_WLOCKED(inp);
1928 if (!wlocked && !INP_TRY_WLOCK(inp)) {
1929 /*
1930 * rwlocks read locks are anonymous, and there
1931 * is no way to know if our current thread
1932 * holds an rlock on the inp. As a rough
1933 * estimate, check to see if the thread holds
1934 * *any* rlocks at all. If it does not, then we
1935 * know that we don't hold the inp rlock, and
1936 * can safely take the wlock
1937 */
1938 if (curthread->td_rw_rlocks == 0) {
1939 INP_WLOCK(inp);
1940 } else {
1941 /*
1942 * We might hold the rlock, so let's
1943 * do the destroy in a taskqueue
1944 * context to avoid a potential
1945 * deadlock. This should be very
1946 * rare.
1947 */
1948 counter_u64_add(ktls_destroy_task, 1);
1949 TASK_INIT(&tls->destroy_task, 0,
1950 ktls_destroy_help, tls);
1951 (void)taskqueue_enqueue(taskqueue_thread,
1952 &tls->destroy_task);
1953 return;
1954 }
1955 }
1956 }
1957
1958 if (tls->sequential_records) {
1959 struct mbuf *m, *n;
1960 int page_count;
1961
1962 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1963 page_count = m->m_epg_enc_cnt;
1964 while (page_count > 0) {
1965 KASSERT(page_count >= m->m_epg_nrdy,
1966 ("%s: too few pages", __func__));
1967 page_count -= m->m_epg_nrdy;
1968 m = m_free(m);
1969 }
1970 }
1971 }
1972
1973 counter_u64_add(ktls_offload_active, -1);
1974 switch (tls->mode) {
1975 case TCP_TLS_MODE_SW:
1976 switch (tls->params.cipher_algorithm) {
1977 case CRYPTO_AES_CBC:
1978 counter_u64_add(ktls_sw_cbc, -1);
1979 break;
1980 case CRYPTO_AES_NIST_GCM_16:
1981 counter_u64_add(ktls_sw_gcm, -1);
1982 break;
1983 case CRYPTO_CHACHA20_POLY1305:
1984 counter_u64_add(ktls_sw_chacha20, -1);
1985 break;
1986 }
1987 break;
1988 case TCP_TLS_MODE_IFNET:
1989 switch (tls->params.cipher_algorithm) {
1990 case CRYPTO_AES_CBC:
1991 counter_u64_add(ktls_ifnet_cbc, -1);
1992 break;
1993 case CRYPTO_AES_NIST_GCM_16:
1994 counter_u64_add(ktls_ifnet_gcm, -1);
1995 break;
1996 case CRYPTO_CHACHA20_POLY1305:
1997 counter_u64_add(ktls_ifnet_chacha20, -1);
1998 break;
1999 }
2000 if (tls->snd_tag != NULL)
2001 m_snd_tag_rele(tls->snd_tag);
2002 if (tls->rx_ifp != NULL)
2003 if_rele(tls->rx_ifp);
2004 if (tls->tx) {
2005 INP_WLOCK_ASSERT(inp);
2006 tp = intotcpcb(inp);
2007 MPASS(tp->t_nic_ktls_xmit == 1);
2008 tp->t_nic_ktls_xmit = 0;
2009 }
2010 break;
2011 #ifdef TCP_OFFLOAD
2012 case TCP_TLS_MODE_TOE:
2013 switch (tls->params.cipher_algorithm) {
2014 case CRYPTO_AES_CBC:
2015 counter_u64_add(ktls_toe_cbc, -1);
2016 break;
2017 case CRYPTO_AES_NIST_GCM_16:
2018 counter_u64_add(ktls_toe_gcm, -1);
2019 break;
2020 case CRYPTO_CHACHA20_POLY1305:
2021 counter_u64_add(ktls_toe_chacha20, -1);
2022 break;
2023 }
2024 break;
2025 #endif
2026 }
2027 if (tls->ocf_session != NULL)
2028 ktls_ocf_free(tls);
2029 if (tls->params.auth_key != NULL) {
2030 zfree(tls->params.auth_key, M_KTLS);
2031 tls->params.auth_key = NULL;
2032 tls->params.auth_key_len = 0;
2033 }
2034 if (tls->params.cipher_key != NULL) {
2035 zfree(tls->params.cipher_key, M_KTLS);
2036 tls->params.cipher_key = NULL;
2037 tls->params.cipher_key_len = 0;
2038 }
2039 if (tls->tx) {
2040 INP_WLOCK_ASSERT(inp);
2041 if (!in_pcbrele_wlocked(inp) && !wlocked)
2042 INP_WUNLOCK(inp);
2043 }
2044 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
2045
2046 uma_zfree(ktls_session_zone, tls);
2047 }
2048
2049 void
ktls_seq(struct sockbuf * sb,struct mbuf * m)2050 ktls_seq(struct sockbuf *sb, struct mbuf *m)
2051 {
2052
2053 for (; m != NULL; m = m->m_next) {
2054 KASSERT((m->m_flags & M_EXTPG) != 0,
2055 ("ktls_seq: mapped mbuf %p", m));
2056
2057 m->m_epg_seqno = sb->sb_tls_seqno;
2058 sb->sb_tls_seqno++;
2059 }
2060 }
2061
2062 /*
2063 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
2064 * mbuf in the chain must be an unmapped mbuf. The payload of the
2065 * mbuf must be populated with the payload of each TLS record.
2066 *
2067 * The record_type argument specifies the TLS record type used when
2068 * populating the TLS header.
2069 *
2070 * The enq_count argument on return is set to the number of pages of
2071 * payload data for this entire chain that need to be encrypted via SW
2072 * encryption. The returned value should be passed to ktls_enqueue
2073 * when scheduling encryption of this chain of mbufs. To handle the
2074 * special case of empty fragments for TLS 1.0 sessions, an empty
2075 * fragment counts as one page.
2076 */
2077 void
ktls_frame(struct mbuf * top,struct ktls_session * tls,int * enq_cnt,uint8_t record_type)2078 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2079 uint8_t record_type)
2080 {
2081 struct tls_record_layer *tlshdr;
2082 struct mbuf *m;
2083 uint64_t *noncep;
2084 uint16_t tls_len;
2085 int maxlen __diagused;
2086
2087 maxlen = tls->params.max_frame_len;
2088 *enq_cnt = 0;
2089 for (m = top; m != NULL; m = m->m_next) {
2090 /*
2091 * All mbufs in the chain should be TLS records whose
2092 * payload does not exceed the maximum frame length.
2093 *
2094 * Empty TLS 1.0 records are permitted when using CBC.
2095 */
2096 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2097 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2098 ("ktls_frame: m %p len %d", m, m->m_len));
2099
2100 /*
2101 * TLS frames require unmapped mbufs to store session
2102 * info.
2103 */
2104 KASSERT((m->m_flags & M_EXTPG) != 0,
2105 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2106
2107 tls_len = m->m_len;
2108
2109 /* Save a reference to the session. */
2110 m->m_epg_tls = ktls_hold(tls);
2111
2112 m->m_epg_hdrlen = tls->params.tls_hlen;
2113 m->m_epg_trllen = tls->params.tls_tlen;
2114 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2115 int bs, delta;
2116
2117 /*
2118 * AES-CBC pads messages to a multiple of the
2119 * block size. Note that the padding is
2120 * applied after the digest and the encryption
2121 * is done on the "plaintext || mac || padding".
2122 * At least one byte of padding is always
2123 * present.
2124 *
2125 * Compute the final trailer length assuming
2126 * at most one block of padding.
2127 * tls->params.tls_tlen is the maximum
2128 * possible trailer length (padding + digest).
2129 * delta holds the number of excess padding
2130 * bytes if the maximum were used. Those
2131 * extra bytes are removed.
2132 */
2133 bs = tls->params.tls_bs;
2134 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2135 m->m_epg_trllen -= delta;
2136 }
2137 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2138
2139 /* Populate the TLS header. */
2140 tlshdr = (void *)m->m_epg_hdr;
2141 tlshdr->tls_vmajor = tls->params.tls_vmajor;
2142
2143 /*
2144 * TLS 1.3 masquarades as TLS 1.2 with a record type
2145 * of TLS_RLTYPE_APP.
2146 */
2147 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2148 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2149 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2150 tlshdr->tls_type = TLS_RLTYPE_APP;
2151 /* save the real record type for later */
2152 m->m_epg_record_type = record_type;
2153 m->m_epg_trail[0] = record_type;
2154 } else {
2155 tlshdr->tls_vminor = tls->params.tls_vminor;
2156 tlshdr->tls_type = record_type;
2157 }
2158 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2159
2160 /*
2161 * Store nonces / explicit IVs after the end of the
2162 * TLS header.
2163 *
2164 * For GCM with TLS 1.2, an 8 byte nonce is copied
2165 * from the end of the IV. The nonce is then
2166 * incremented for use by the next record.
2167 *
2168 * For CBC, a random nonce is inserted for TLS 1.1+.
2169 */
2170 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2171 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2172 noncep = (uint64_t *)(tls->params.iv + 8);
2173 be64enc(tlshdr + 1, *noncep);
2174 (*noncep)++;
2175 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2176 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2177 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2178
2179 /*
2180 * When using SW encryption, mark the mbuf not ready.
2181 * It will be marked ready via sbready() after the
2182 * record has been encrypted.
2183 *
2184 * When using ifnet TLS, unencrypted TLS records are
2185 * sent down the stack to the NIC.
2186 */
2187 if (tls->mode == TCP_TLS_MODE_SW) {
2188 m->m_flags |= M_NOTREADY;
2189 if (__predict_false(tls_len == 0)) {
2190 /* TLS 1.0 empty fragment. */
2191 m->m_epg_nrdy = 1;
2192 } else
2193 m->m_epg_nrdy = m->m_epg_npgs;
2194 *enq_cnt += m->m_epg_nrdy;
2195 }
2196 }
2197 }
2198
2199 bool
ktls_permit_empty_frames(struct ktls_session * tls)2200 ktls_permit_empty_frames(struct ktls_session *tls)
2201 {
2202 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2203 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2204 }
2205
2206 void
ktls_check_rx(struct sockbuf * sb)2207 ktls_check_rx(struct sockbuf *sb)
2208 {
2209 struct tls_record_layer hdr;
2210 struct ktls_wq *wq;
2211 struct socket *so;
2212 bool running;
2213
2214 SOCKBUF_LOCK_ASSERT(sb);
2215 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2216 __func__, sb));
2217 so = __containerof(sb, struct socket, so_rcv);
2218
2219 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2220 return;
2221
2222 /* Is there enough queued for a TLS header? */
2223 if (sb->sb_tlscc < sizeof(hdr)) {
2224 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2225 so->so_error = EMSGSIZE;
2226 return;
2227 }
2228
2229 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2230
2231 /* Is the entire record queued? */
2232 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2233 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2234 so->so_error = EMSGSIZE;
2235 return;
2236 }
2237
2238 sb->sb_flags |= SB_TLS_RX_RUNNING;
2239
2240 soref(so);
2241 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2242 mtx_lock(&wq->mtx);
2243 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2244 running = wq->running;
2245 mtx_unlock(&wq->mtx);
2246 if (!running)
2247 wakeup(wq);
2248 counter_u64_add(ktls_cnt_rx_queued, 1);
2249 }
2250
2251 static struct mbuf *
ktls_detach_record(struct sockbuf * sb,int len)2252 ktls_detach_record(struct sockbuf *sb, int len)
2253 {
2254 struct mbuf *m, *n, *top;
2255 int remain;
2256
2257 SOCKBUF_LOCK_ASSERT(sb);
2258 MPASS(len <= sb->sb_tlscc);
2259
2260 /*
2261 * If TLS chain is the exact size of the record,
2262 * just grab the whole record.
2263 */
2264 top = sb->sb_mtls;
2265 if (sb->sb_tlscc == len) {
2266 sb->sb_mtls = NULL;
2267 sb->sb_mtlstail = NULL;
2268 goto out;
2269 }
2270
2271 /*
2272 * While it would be nice to use m_split() here, we need
2273 * to know exactly what m_split() allocates to update the
2274 * accounting, so do it inline instead.
2275 */
2276 remain = len;
2277 for (m = top; remain > m->m_len; m = m->m_next)
2278 remain -= m->m_len;
2279
2280 /* Easy case: don't have to split 'm'. */
2281 if (remain == m->m_len) {
2282 sb->sb_mtls = m->m_next;
2283 if (sb->sb_mtls == NULL)
2284 sb->sb_mtlstail = NULL;
2285 m->m_next = NULL;
2286 goto out;
2287 }
2288
2289 /*
2290 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2291 * with M_NOWAIT first.
2292 */
2293 n = m_get(M_NOWAIT, MT_DATA);
2294 if (n == NULL) {
2295 /*
2296 * Use M_WAITOK with socket buffer unlocked. If
2297 * 'sb_mtls' changes while the lock is dropped, return
2298 * NULL to force the caller to retry.
2299 */
2300 SOCKBUF_UNLOCK(sb);
2301
2302 n = m_get(M_WAITOK, MT_DATA);
2303
2304 SOCKBUF_LOCK(sb);
2305 if (sb->sb_mtls != top) {
2306 m_free(n);
2307 return (NULL);
2308 }
2309 }
2310 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2311
2312 /* Store remainder in 'n'. */
2313 n->m_len = m->m_len - remain;
2314 if (m->m_flags & M_EXT) {
2315 n->m_data = m->m_data + remain;
2316 mb_dupcl(n, m);
2317 } else {
2318 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2319 }
2320
2321 /* Trim 'm' and update accounting. */
2322 m->m_len -= n->m_len;
2323 sb->sb_tlscc -= n->m_len;
2324 sb->sb_ccc -= n->m_len;
2325
2326 /* Account for 'n'. */
2327 sballoc_ktls_rx(sb, n);
2328
2329 /* Insert 'n' into the TLS chain. */
2330 sb->sb_mtls = n;
2331 n->m_next = m->m_next;
2332 if (sb->sb_mtlstail == m)
2333 sb->sb_mtlstail = n;
2334
2335 /* Detach the record from the TLS chain. */
2336 m->m_next = NULL;
2337
2338 out:
2339 MPASS(m_length(top, NULL) == len);
2340 for (m = top; m != NULL; m = m->m_next)
2341 sbfree_ktls_rx(sb, m);
2342 sb->sb_tlsdcc = len;
2343 sb->sb_ccc += len;
2344 SBCHECK(sb);
2345 return (top);
2346 }
2347
2348 /*
2349 * Determine the length of the trailing zero padding and find the real
2350 * record type in the byte before the padding.
2351 *
2352 * Walking the mbuf chain backwards is clumsy, so another option would
2353 * be to scan forwards remembering the last non-zero byte before the
2354 * trailer. However, it would be expensive to scan the entire record.
2355 * Instead, find the last non-zero byte of each mbuf in the chain
2356 * keeping track of the relative offset of that nonzero byte.
2357 *
2358 * trail_len is the size of the MAC/tag on input and is set to the
2359 * size of the full trailer including padding and the record type on
2360 * return.
2361 */
2362 static int
tls13_find_record_type(struct ktls_session * tls,struct mbuf * m,int tls_len,int * trailer_len,uint8_t * record_typep)2363 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2364 int *trailer_len, uint8_t *record_typep)
2365 {
2366 char *cp;
2367 u_int digest_start, last_offset, m_len, offset;
2368 uint8_t record_type;
2369
2370 digest_start = tls_len - *trailer_len;
2371 last_offset = 0;
2372 offset = 0;
2373 for (; m != NULL && offset < digest_start;
2374 offset += m->m_len, m = m->m_next) {
2375 /* Don't look for padding in the tag. */
2376 m_len = min(digest_start - offset, m->m_len);
2377 cp = mtod(m, char *);
2378
2379 /* Find last non-zero byte in this mbuf. */
2380 while (m_len > 0 && cp[m_len - 1] == 0)
2381 m_len--;
2382 if (m_len > 0) {
2383 record_type = cp[m_len - 1];
2384 last_offset = offset + m_len;
2385 }
2386 }
2387 if (last_offset < tls->params.tls_hlen)
2388 return (EBADMSG);
2389
2390 *record_typep = record_type;
2391 *trailer_len = tls_len - last_offset + 1;
2392 return (0);
2393 }
2394
2395 /*
2396 * Check if a mbuf chain is fully decrypted at the given offset and
2397 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2398 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2399 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2400 * is encrypted.
2401 */
2402 ktls_mbuf_crypto_st_t
ktls_mbuf_crypto_state(struct mbuf * mb,int offset,int len)2403 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2404 {
2405 int m_flags_ored = 0;
2406 int m_flags_anded = -1;
2407
2408 for (; mb != NULL; mb = mb->m_next) {
2409 if (offset < mb->m_len)
2410 break;
2411 offset -= mb->m_len;
2412 }
2413 offset += len;
2414
2415 for (; mb != NULL; mb = mb->m_next) {
2416 m_flags_ored |= mb->m_flags;
2417 m_flags_anded &= mb->m_flags;
2418
2419 if (offset <= mb->m_len)
2420 break;
2421 offset -= mb->m_len;
2422 }
2423 MPASS(mb != NULL || offset == 0);
2424
2425 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2426 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2427 else
2428 return ((m_flags_ored & M_DECRYPTED) ?
2429 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2430 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2431 }
2432
2433 /*
2434 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2435 */
2436 static int
ktls_resync_ifnet(struct socket * so,uint32_t tls_len,uint64_t tls_rcd_num)2437 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2438 {
2439 union if_snd_tag_modify_params params;
2440 struct m_snd_tag *mst;
2441 struct inpcb *inp = sotoinpcb(so);
2442 struct tcpcb *tp = intotcpcb(inp);
2443
2444 mst = so->so_rcv.sb_tls_info->snd_tag;
2445 if (__predict_false(mst == NULL))
2446 return (EINVAL);
2447
2448 INP_RLOCK(inp);
2449 if (tp->t_flags & TF_DISCONNECTED) {
2450 INP_RUNLOCK(inp);
2451 return (ECONNRESET);
2452 }
2453
2454 /* Get the TCP sequence number of the next valid TLS header. */
2455 SOCKBUF_LOCK(&so->so_rcv);
2456 params.tls_rx.tls_hdr_tcp_sn =
2457 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2458 params.tls_rx.tls_rec_length = tls_len;
2459 params.tls_rx.tls_seq_number = tls_rcd_num;
2460 SOCKBUF_UNLOCK(&so->so_rcv);
2461
2462 INP_RUNLOCK(inp);
2463
2464 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2465 return (mst->sw->snd_tag_modify(mst, ¶ms));
2466 }
2467
2468 static void
ktls_drop(struct socket * so,int error)2469 ktls_drop(struct socket *so, int error)
2470 {
2471 struct epoch_tracker et;
2472 struct inpcb *inp = sotoinpcb(so);
2473 struct tcpcb *tp = intotcpcb(inp);
2474
2475 NET_EPOCH_ENTER(et);
2476 INP_WLOCK(inp);
2477 if (!(tp->t_flags & TF_DISCONNECTED)) {
2478 CURVNET_SET(inp->inp_socket->so_vnet);
2479 tp = tcp_drop(tp, error);
2480 CURVNET_RESTORE();
2481 if (tp != NULL)
2482 INP_WUNLOCK(inp);
2483 } else {
2484 so->so_error = error;
2485 SOCK_RECVBUF_LOCK(so);
2486 sorwakeup_locked(so);
2487 INP_WUNLOCK(inp);
2488 }
2489 NET_EPOCH_EXIT(et);
2490 }
2491
2492 static void
ktls_decrypt(struct socket * so)2493 ktls_decrypt(struct socket *so)
2494 {
2495 char tls_header[MBUF_PEXT_HDR_LEN];
2496 struct ktls_session *tls;
2497 struct sockbuf *sb;
2498 struct tls_record_layer *hdr;
2499 struct tls_get_record tgr;
2500 struct mbuf *control, *data, *m;
2501 ktls_mbuf_crypto_st_t state;
2502 uint64_t seqno;
2503 int error, remain, tls_len, trail_len;
2504 bool tls13;
2505 uint8_t vminor, record_type;
2506
2507 hdr = (struct tls_record_layer *)tls_header;
2508 sb = &so->so_rcv;
2509 SOCKBUF_LOCK(sb);
2510 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2511 ("%s: socket %p not running", __func__, so));
2512
2513 tls = sb->sb_tls_info;
2514 MPASS(tls != NULL);
2515
2516 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2517 if (tls13)
2518 vminor = TLS_MINOR_VER_TWO;
2519 else
2520 vminor = tls->params.tls_vminor;
2521 for (;;) {
2522 /* Is there enough queued for a TLS header? */
2523 if (sb->sb_tlscc < tls->params.tls_hlen)
2524 break;
2525
2526 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2527 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2528
2529 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2530 hdr->tls_vminor != vminor)
2531 error = EINVAL;
2532 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2533 error = EINVAL;
2534 else if (tls_len < tls->params.tls_hlen || tls_len >
2535 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2536 tls->params.tls_tlen)
2537 error = EMSGSIZE;
2538 else
2539 error = 0;
2540 if (__predict_false(error != 0)) {
2541 /*
2542 * We have a corrupted record and are likely
2543 * out of sync. The connection isn't
2544 * recoverable at this point, so abort it.
2545 */
2546 SOCKBUF_UNLOCK(sb);
2547 counter_u64_add(ktls_offload_corrupted_records, 1);
2548
2549 ktls_drop(so, error);
2550 goto deref;
2551 }
2552
2553 /* Is the entire record queued? */
2554 if (sb->sb_tlscc < tls_len)
2555 break;
2556
2557 /*
2558 * Split out the portion of the mbuf chain containing
2559 * this TLS record.
2560 */
2561 data = ktls_detach_record(sb, tls_len);
2562 if (data == NULL)
2563 continue;
2564 MPASS(sb->sb_tlsdcc == tls_len);
2565
2566 seqno = sb->sb_tls_seqno;
2567 sb->sb_tls_seqno++;
2568 SBCHECK(sb);
2569 SOCKBUF_UNLOCK(sb);
2570
2571 /* get crypto state for this TLS record */
2572 state = ktls_mbuf_crypto_state(data, 0, tls_len);
2573
2574 switch (state) {
2575 case KTLS_MBUF_CRYPTO_ST_MIXED:
2576 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2577 if (error)
2578 break;
2579 /* FALLTHROUGH */
2580 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2581 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2582 &trail_len);
2583 if (__predict_true(error == 0)) {
2584 if (tls13) {
2585 error = tls13_find_record_type(tls, data,
2586 tls_len, &trail_len, &record_type);
2587 } else {
2588 record_type = hdr->tls_type;
2589 }
2590 }
2591 break;
2592 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2593 /*
2594 * NIC TLS is only supported for AEAD
2595 * ciphersuites which used a fixed sized
2596 * trailer.
2597 */
2598 if (tls13) {
2599 trail_len = tls->params.tls_tlen - 1;
2600 error = tls13_find_record_type(tls, data,
2601 tls_len, &trail_len, &record_type);
2602 } else {
2603 trail_len = tls->params.tls_tlen;
2604 error = 0;
2605 record_type = hdr->tls_type;
2606 }
2607 break;
2608 default:
2609 error = EINVAL;
2610 break;
2611 }
2612 if (error) {
2613 counter_u64_add(ktls_offload_failed_crypto, 1);
2614
2615 SOCKBUF_LOCK(sb);
2616 if (sb->sb_tlsdcc == 0) {
2617 /*
2618 * sbcut/drop/flush discarded these
2619 * mbufs.
2620 */
2621 m_freem(data);
2622 break;
2623 }
2624
2625 /*
2626 * Drop this TLS record's data, but keep
2627 * decrypting subsequent records.
2628 */
2629 sb->sb_ccc -= tls_len;
2630 sb->sb_tlsdcc = 0;
2631
2632 if (error != EMSGSIZE)
2633 error = EBADMSG;
2634 CURVNET_SET(so->so_vnet);
2635 so->so_error = error;
2636 sorwakeup_locked(so);
2637 CURVNET_RESTORE();
2638
2639 m_freem(data);
2640
2641 SOCKBUF_LOCK(sb);
2642 continue;
2643 }
2644
2645 /* Allocate the control mbuf. */
2646 memset(&tgr, 0, sizeof(tgr));
2647 tgr.tls_type = record_type;
2648 tgr.tls_vmajor = hdr->tls_vmajor;
2649 tgr.tls_vminor = hdr->tls_vminor;
2650 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2651 trail_len);
2652 control = sbcreatecontrol(&tgr, sizeof(tgr),
2653 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2654
2655 SOCKBUF_LOCK(sb);
2656 if (sb->sb_tlsdcc == 0) {
2657 /* sbcut/drop/flush discarded these mbufs. */
2658 MPASS(sb->sb_tlscc == 0);
2659 m_freem(data);
2660 m_freem(control);
2661 break;
2662 }
2663
2664 /*
2665 * Clear the 'dcc' accounting in preparation for
2666 * adding the decrypted record.
2667 */
2668 sb->sb_ccc -= tls_len;
2669 sb->sb_tlsdcc = 0;
2670 SBCHECK(sb);
2671
2672 /* If there is no payload, drop all of the data. */
2673 if (tgr.tls_length == htobe16(0)) {
2674 m_freem(data);
2675 data = NULL;
2676 } else {
2677 /* Trim header. */
2678 remain = tls->params.tls_hlen;
2679 while (remain > 0) {
2680 if (data->m_len > remain) {
2681 data->m_data += remain;
2682 data->m_len -= remain;
2683 break;
2684 }
2685 remain -= data->m_len;
2686 data = m_free(data);
2687 }
2688
2689 /* Trim trailer and clear M_NOTREADY. */
2690 remain = be16toh(tgr.tls_length);
2691 m = data;
2692 for (m = data; remain > m->m_len; m = m->m_next) {
2693 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2694 remain -= m->m_len;
2695 }
2696 m->m_len = remain;
2697 m_freem(m->m_next);
2698 m->m_next = NULL;
2699 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2700
2701 /* Set EOR on the final mbuf. */
2702 m->m_flags |= M_EOR;
2703 }
2704
2705 sbappendcontrol_locked(sb, data, control, 0);
2706
2707 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2708 sb->sb_flags |= SB_TLS_RX_RESYNC;
2709 SOCKBUF_UNLOCK(sb);
2710 ktls_resync_ifnet(so, tls_len, seqno);
2711 SOCKBUF_LOCK(sb);
2712 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2713 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2714 SOCKBUF_UNLOCK(sb);
2715 ktls_resync_ifnet(so, 0, seqno);
2716 SOCKBUF_LOCK(sb);
2717 }
2718 }
2719
2720 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2721
2722 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2723 so->so_error = EMSGSIZE;
2724
2725 sorwakeup_locked(so);
2726
2727 deref:
2728 SOCKBUF_UNLOCK_ASSERT(sb);
2729
2730 CURVNET_SET(so->so_vnet);
2731 sorele(so);
2732 CURVNET_RESTORE();
2733 }
2734
2735 void
ktls_enqueue_to_free(struct mbuf * m)2736 ktls_enqueue_to_free(struct mbuf *m)
2737 {
2738 struct ktls_wq *wq;
2739 bool running;
2740
2741 /* Mark it for freeing. */
2742 m->m_epg_flags |= EPG_FLAG_2FREE;
2743 wq = &ktls_wq[m->m_epg_tls->wq_index];
2744 mtx_lock(&wq->mtx);
2745 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2746 running = wq->running;
2747 mtx_unlock(&wq->mtx);
2748 if (!running)
2749 wakeup(wq);
2750 }
2751
2752 static void *
ktls_buffer_alloc(struct ktls_wq * wq,struct mbuf * m)2753 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2754 {
2755 void *buf;
2756 int domain, running;
2757
2758 if (m->m_epg_npgs <= 2)
2759 return (NULL);
2760 if (ktls_buffer_zone == NULL)
2761 return (NULL);
2762 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2763 /*
2764 * Rate-limit allocation attempts after a failure.
2765 * ktls_buffer_import() will acquire a per-domain mutex to check
2766 * the free page queues and may fail consistently if memory is
2767 * fragmented.
2768 */
2769 return (NULL);
2770 }
2771 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2772 if (buf == NULL) {
2773 domain = PCPU_GET(domain);
2774 wq->lastallocfail = ticks;
2775
2776 /*
2777 * Note that this check is "racy", but the races are
2778 * harmless, and are either a spurious wakeup if
2779 * multiple threads fail allocations before the alloc
2780 * thread wakes, or waiting an extra second in case we
2781 * see an old value of running == true.
2782 */
2783 if (!VM_DOMAIN_EMPTY(domain)) {
2784 running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2785 if (!running)
2786 wakeup(&ktls_domains[domain].reclaim_td);
2787 }
2788 }
2789 return (buf);
2790 }
2791
2792 static int
ktls_encrypt_record(struct ktls_wq * wq,struct mbuf * m,struct ktls_session * tls,struct ktls_ocf_encrypt_state * state)2793 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2794 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2795 {
2796 vm_page_t pg;
2797 int error, i, len, off;
2798
2799 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2800 ("%p not unready & nomap mbuf\n", m));
2801 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2802 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2803 ktls_maxlen));
2804
2805 /* Anonymous mbufs are encrypted in place. */
2806 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2807 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2808
2809 /*
2810 * For file-backed mbufs (from sendfile), anonymous wired
2811 * pages are allocated and used as the encryption destination.
2812 */
2813 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2814 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2815 m->m_epg_1st_off;
2816 state->dst_iov[0].iov_base = (char *)state->cbuf +
2817 m->m_epg_1st_off;
2818 state->dst_iov[0].iov_len = len;
2819 state->parray[0] = DMAP_TO_PHYS(state->cbuf);
2820 i = 1;
2821 } else {
2822 off = m->m_epg_1st_off;
2823 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2824 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2825 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2826 len = m_epg_pagelen(m, i, off);
2827 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2828 state->dst_iov[i].iov_base =
2829 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2830 state->dst_iov[i].iov_len = len;
2831 }
2832 }
2833 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2834 state->dst_iov[i].iov_base = m->m_epg_trail;
2835 state->dst_iov[i].iov_len = m->m_epg_trllen;
2836
2837 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2838
2839 if (__predict_false(error != 0)) {
2840 /* Free the anonymous pages. */
2841 if (state->cbuf != NULL)
2842 uma_zfree(ktls_buffer_zone, state->cbuf);
2843 else {
2844 for (i = 0; i < m->m_epg_npgs; i++) {
2845 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2846 (void)vm_page_unwire_noq(pg);
2847 vm_page_free(pg);
2848 }
2849 }
2850 }
2851 return (error);
2852 }
2853
2854 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2855 static u_int
ktls_batched_records(struct mbuf * m)2856 ktls_batched_records(struct mbuf *m)
2857 {
2858 int page_count, records;
2859
2860 records = 0;
2861 page_count = m->m_epg_enc_cnt;
2862 while (page_count > 0) {
2863 records++;
2864 page_count -= m->m_epg_nrdy;
2865 m = m->m_next;
2866 }
2867 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2868 return (records);
2869 }
2870
2871 void
ktls_enqueue(struct mbuf * m,struct socket * so,int page_count)2872 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2873 {
2874 struct ktls_session *tls;
2875 struct ktls_wq *wq;
2876 int queued;
2877 bool running;
2878
2879 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2880 (M_EXTPG | M_NOTREADY)),
2881 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2882 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2883
2884 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2885
2886 m->m_epg_enc_cnt = page_count;
2887
2888 /*
2889 * Save a pointer to the socket. The caller is responsible
2890 * for taking an additional reference via soref().
2891 */
2892 m->m_epg_so = so;
2893
2894 queued = 1;
2895 tls = m->m_epg_tls;
2896 wq = &ktls_wq[tls->wq_index];
2897 mtx_lock(&wq->mtx);
2898 if (__predict_false(tls->sequential_records)) {
2899 /*
2900 * For TLS 1.0, records must be encrypted
2901 * sequentially. For a given connection, all records
2902 * queued to the associated work queue are processed
2903 * sequentially. However, sendfile(2) might complete
2904 * I/O requests spanning multiple TLS records out of
2905 * order. Here we ensure TLS records are enqueued to
2906 * the work queue in FIFO order.
2907 *
2908 * tls->next_seqno holds the sequence number of the
2909 * next TLS record that should be enqueued to the work
2910 * queue. If this next record is not tls->next_seqno,
2911 * it must be a future record, so insert it, sorted by
2912 * TLS sequence number, into tls->pending_records and
2913 * return.
2914 *
2915 * If this TLS record matches tls->next_seqno, place
2916 * it in the work queue and then check
2917 * tls->pending_records to see if any
2918 * previously-queued records are now ready for
2919 * encryption.
2920 */
2921 if (m->m_epg_seqno != tls->next_seqno) {
2922 struct mbuf *n, *p;
2923
2924 p = NULL;
2925 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2926 if (n->m_epg_seqno > m->m_epg_seqno)
2927 break;
2928 p = n;
2929 }
2930 if (n == NULL)
2931 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2932 m_epg_stailq);
2933 else if (p == NULL)
2934 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2935 m_epg_stailq);
2936 else
2937 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2938 m_epg_stailq);
2939 mtx_unlock(&wq->mtx);
2940 counter_u64_add(ktls_cnt_tx_pending, 1);
2941 return;
2942 }
2943
2944 tls->next_seqno += ktls_batched_records(m);
2945 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2946
2947 while (!STAILQ_EMPTY(&tls->pending_records)) {
2948 struct mbuf *n;
2949
2950 n = STAILQ_FIRST(&tls->pending_records);
2951 if (n->m_epg_seqno != tls->next_seqno)
2952 break;
2953
2954 queued++;
2955 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2956 tls->next_seqno += ktls_batched_records(n);
2957 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2958 }
2959 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2960 } else
2961 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2962
2963 running = wq->running;
2964 mtx_unlock(&wq->mtx);
2965 if (!running)
2966 wakeup(wq);
2967 counter_u64_add(ktls_cnt_tx_queued, queued);
2968 }
2969
2970 /*
2971 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2972 * the pages from the file and replace them with the anonymous pages
2973 * allocated in ktls_encrypt_record().
2974 */
2975 static void
ktls_finish_nonanon(struct mbuf * m,struct ktls_ocf_encrypt_state * state)2976 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2977 {
2978 int i;
2979
2980 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2981
2982 /* Free the old pages. */
2983 m->m_ext.ext_free(m);
2984
2985 /* Replace them with the new pages. */
2986 if (state->cbuf != NULL) {
2987 for (i = 0; i < m->m_epg_npgs; i++)
2988 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2989
2990 /* Contig pages should go back to the cache. */
2991 m->m_ext.ext_free = ktls_free_mext_contig;
2992 } else {
2993 for (i = 0; i < m->m_epg_npgs; i++)
2994 m->m_epg_pa[i] = state->parray[i];
2995
2996 /* Use the basic free routine. */
2997 m->m_ext.ext_free = mb_free_mext_pgs;
2998 }
2999
3000 /* Pages are now writable. */
3001 m->m_epg_flags |= EPG_FLAG_ANON;
3002 }
3003
3004 static __noinline void
ktls_encrypt(struct ktls_wq * wq,struct mbuf * top)3005 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
3006 {
3007 struct ktls_ocf_encrypt_state state;
3008 struct ktls_session *tls;
3009 struct socket *so;
3010 struct mbuf *m;
3011 int error, npages, total_pages;
3012
3013 so = top->m_epg_so;
3014 tls = top->m_epg_tls;
3015 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3016 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3017 #ifdef INVARIANTS
3018 top->m_epg_so = NULL;
3019 #endif
3020 total_pages = top->m_epg_enc_cnt;
3021 npages = 0;
3022
3023 /*
3024 * Encrypt the TLS records in the chain of mbufs starting with
3025 * 'top'. 'total_pages' gives us a total count of pages and is
3026 * used to know when we have finished encrypting the TLS
3027 * records originally queued with 'top'.
3028 *
3029 * NB: These mbufs are queued in the socket buffer and
3030 * 'm_next' is traversing the mbufs in the socket buffer. The
3031 * socket buffer lock is not held while traversing this chain.
3032 * Since the mbufs are all marked M_NOTREADY their 'm_next'
3033 * pointers should be stable. However, the 'm_next' of the
3034 * last mbuf encrypted is not necessarily NULL. It can point
3035 * to other mbufs appended while 'top' was on the TLS work
3036 * queue.
3037 *
3038 * Each mbuf holds an entire TLS record.
3039 */
3040 error = 0;
3041 for (m = top; npages != total_pages; m = m->m_next) {
3042 KASSERT(m->m_epg_tls == tls,
3043 ("different TLS sessions in a single mbuf chain: %p vs %p",
3044 tls, m->m_epg_tls));
3045 KASSERT(npages + m->m_epg_npgs <= total_pages,
3046 ("page count mismatch: top %p, total_pages %d, m %p", top,
3047 total_pages, m));
3048
3049 error = ktls_encrypt_record(wq, m, tls, &state);
3050 if (error) {
3051 counter_u64_add(ktls_offload_failed_crypto, 1);
3052 break;
3053 }
3054
3055 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3056 ktls_finish_nonanon(m, &state);
3057 m->m_flags |= M_RDONLY;
3058
3059 npages += m->m_epg_nrdy;
3060
3061 /*
3062 * Drop a reference to the session now that it is no
3063 * longer needed. Existing code depends on encrypted
3064 * records having no associated session vs
3065 * yet-to-be-encrypted records having an associated
3066 * session.
3067 */
3068 m->m_epg_tls = NULL;
3069 ktls_free(tls);
3070 }
3071
3072 CURVNET_SET(so->so_vnet);
3073 if (error == 0) {
3074 (void)so->so_proto->pr_ready(so, top, npages);
3075 } else {
3076 ktls_drop(so, EIO);
3077 mb_free_notready(top, total_pages);
3078 }
3079
3080 sorele(so);
3081 CURVNET_RESTORE();
3082 }
3083
3084 void
ktls_encrypt_cb(struct ktls_ocf_encrypt_state * state,int error)3085 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3086 {
3087 struct ktls_session *tls;
3088 struct socket *so;
3089 struct mbuf *m;
3090 int npages;
3091
3092 m = state->m;
3093
3094 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3095 ktls_finish_nonanon(m, state);
3096 m->m_flags |= M_RDONLY;
3097
3098 so = state->so;
3099 free(state, M_KTLS);
3100
3101 /*
3102 * Drop a reference to the session now that it is no longer
3103 * needed. Existing code depends on encrypted records having
3104 * no associated session vs yet-to-be-encrypted records having
3105 * an associated session.
3106 */
3107 tls = m->m_epg_tls;
3108 m->m_epg_tls = NULL;
3109 ktls_free(tls);
3110
3111 if (error != 0)
3112 counter_u64_add(ktls_offload_failed_crypto, 1);
3113
3114 CURVNET_SET(so->so_vnet);
3115 npages = m->m_epg_nrdy;
3116
3117 if (error == 0) {
3118 (void)so->so_proto->pr_ready(so, m, npages);
3119 } else {
3120 ktls_drop(so, EIO);
3121 mb_free_notready(m, npages);
3122 }
3123
3124 sorele(so);
3125 CURVNET_RESTORE();
3126 }
3127
3128 /*
3129 * Similar to ktls_encrypt, but used with asynchronous OCF backends
3130 * (coprocessors) where encryption does not use host CPU resources and
3131 * it can be beneficial to queue more requests than CPUs.
3132 */
3133 static __noinline void
ktls_encrypt_async(struct ktls_wq * wq,struct mbuf * top)3134 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3135 {
3136 struct ktls_ocf_encrypt_state *state;
3137 struct ktls_session *tls;
3138 struct socket *so;
3139 struct mbuf *m, *n;
3140 int error, mpages, npages, total_pages;
3141
3142 so = top->m_epg_so;
3143 tls = top->m_epg_tls;
3144 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3145 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3146 #ifdef INVARIANTS
3147 top->m_epg_so = NULL;
3148 #endif
3149 total_pages = top->m_epg_enc_cnt;
3150 npages = 0;
3151
3152 error = 0;
3153 for (m = top; npages != total_pages; m = n) {
3154 KASSERT(m->m_epg_tls == tls,
3155 ("different TLS sessions in a single mbuf chain: %p vs %p",
3156 tls, m->m_epg_tls));
3157 KASSERT(npages + m->m_epg_npgs <= total_pages,
3158 ("page count mismatch: top %p, total_pages %d, m %p", top,
3159 total_pages, m));
3160
3161 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3162 soref(so);
3163 state->so = so;
3164 state->m = m;
3165
3166 mpages = m->m_epg_nrdy;
3167 n = m->m_next;
3168
3169 error = ktls_encrypt_record(wq, m, tls, state);
3170 if (error) {
3171 counter_u64_add(ktls_offload_failed_crypto, 1);
3172 free(state, M_KTLS);
3173 CURVNET_SET(so->so_vnet);
3174 sorele(so);
3175 CURVNET_RESTORE();
3176 break;
3177 }
3178
3179 npages += mpages;
3180 }
3181
3182 CURVNET_SET(so->so_vnet);
3183 if (error != 0) {
3184 ktls_drop(so, EIO);
3185 mb_free_notready(m, total_pages - npages);
3186 }
3187
3188 sorele(so);
3189 CURVNET_RESTORE();
3190 }
3191
3192 static int
ktls_bind_domain(int domain)3193 ktls_bind_domain(int domain)
3194 {
3195 int error;
3196
3197 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3198 if (error != 0)
3199 return (error);
3200 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3201 return (0);
3202 }
3203
3204 static void
ktls_reclaim_thread(void * ctx)3205 ktls_reclaim_thread(void *ctx)
3206 {
3207 struct ktls_domain_info *ktls_domain = ctx;
3208 struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3209 struct sysctl_oid *oid;
3210 char name[80];
3211 int error, domain;
3212
3213 domain = ktls_domain - ktls_domains;
3214 if (bootverbose)
3215 printf("Starting KTLS reclaim thread for domain %d\n", domain);
3216 error = ktls_bind_domain(domain);
3217 if (error)
3218 printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3219 domain, error);
3220 snprintf(name, sizeof(name), "domain%d", domain);
3221 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3222 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3223 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3224 CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed");
3225 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3226 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3227 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3228 CTLFLAG_RD, &sc->running, 0, "thread running");
3229
3230 for (;;) {
3231 atomic_store_int(&sc->running, 0);
3232 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3233 atomic_store_int(&sc->running, 1);
3234 sc->wakeups++;
3235 /*
3236 * Below we attempt to reclaim ktls_max_reclaim
3237 * buffers using vm_page_reclaim_contig_domain_ext().
3238 * We do this here, as this function can take several
3239 * seconds to scan all of memory and it does not
3240 * matter if this thread pauses for a while. If we
3241 * block a ktls worker thread, we risk developing
3242 * backlogs of buffers to be encrypted, leading to
3243 * surges of traffic and potential NIC output drops.
3244 */
3245 if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3246 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3247 ktls_max_reclaim) != 0) {
3248 vm_wait_domain(domain);
3249 } else {
3250 sc->reclaims += ktls_max_reclaim;
3251 }
3252 }
3253 }
3254
3255 static void
ktls_work_thread(void * ctx)3256 ktls_work_thread(void *ctx)
3257 {
3258 struct ktls_wq *wq = ctx;
3259 struct mbuf *m, *n;
3260 struct socket *so, *son;
3261 STAILQ_HEAD(, mbuf) local_m_head;
3262 STAILQ_HEAD(, socket) local_so_head;
3263 int cpu;
3264
3265 cpu = wq - ktls_wq;
3266 if (bootverbose)
3267 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3268
3269 /*
3270 * Bind to a core. If ktls_bind_threads is > 1, then
3271 * we bind to the NUMA domain instead.
3272 */
3273 if (ktls_bind_threads) {
3274 int error;
3275
3276 if (ktls_bind_threads > 1) {
3277 struct pcpu *pc = pcpu_find(cpu);
3278
3279 error = ktls_bind_domain(pc->pc_domain);
3280 } else {
3281 cpuset_t mask;
3282
3283 CPU_SETOF(cpu, &mask);
3284 error = cpuset_setthread(curthread->td_tid, &mask);
3285 }
3286 if (error)
3287 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3288 cpu, error);
3289 }
3290 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3291 fpu_kern_thread(0);
3292 #endif
3293 for (;;) {
3294 mtx_lock(&wq->mtx);
3295 while (STAILQ_EMPTY(&wq->m_head) &&
3296 STAILQ_EMPTY(&wq->so_head)) {
3297 wq->running = false;
3298 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3299 wq->running = true;
3300 }
3301
3302 STAILQ_INIT(&local_m_head);
3303 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3304 STAILQ_INIT(&local_so_head);
3305 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3306 mtx_unlock(&wq->mtx);
3307
3308 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3309 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3310 ktls_free(m->m_epg_tls);
3311 m_free_raw(m);
3312 } else {
3313 if (m->m_epg_tls->sync_dispatch)
3314 ktls_encrypt(wq, m);
3315 else
3316 ktls_encrypt_async(wq, m);
3317 counter_u64_add(ktls_cnt_tx_queued, -1);
3318 }
3319 }
3320
3321 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3322 ktls_decrypt(so);
3323 counter_u64_add(ktls_cnt_rx_queued, -1);
3324 }
3325 }
3326 }
3327
3328 static void
ktls_disable_ifnet_help(void * context,int pending __unused)3329 ktls_disable_ifnet_help(void *context, int pending __unused)
3330 {
3331 struct ktls_session *tls;
3332 struct inpcb *inp;
3333 struct tcpcb *tp;
3334 struct socket *so;
3335 int err;
3336
3337 tls = context;
3338 inp = tls->inp;
3339 if (inp == NULL)
3340 return;
3341 INP_WLOCK(inp);
3342 so = inp->inp_socket;
3343 MPASS(so != NULL);
3344 tp = intotcpcb(inp);
3345 if (tp->t_flags & TF_DISCONNECTED) {
3346 goto out;
3347 }
3348
3349 if (so->so_snd.sb_tls_info != NULL)
3350 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3351 else
3352 err = ENXIO;
3353 if (err == 0) {
3354 counter_u64_add(ktls_ifnet_disable_ok, 1);
3355 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3356 if ((tp->t_flags & TF_DISCONNECTED) == 0 &&
3357 tp->t_fb->tfb_hwtls_change != NULL)
3358 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3359 } else {
3360 counter_u64_add(ktls_ifnet_disable_fail, 1);
3361 }
3362
3363 out:
3364 CURVNET_SET(so->so_vnet);
3365 sorele(so);
3366 CURVNET_RESTORE();
3367 INP_WUNLOCK(inp);
3368 ktls_free(tls);
3369 }
3370
3371 /*
3372 * Called when re-transmits are becoming a substantial portion of the
3373 * sends on this connection. When this happens, we transition the
3374 * connection to software TLS. This is needed because most inline TLS
3375 * NICs keep crypto state only for in-order transmits. This means
3376 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3377 * re-DMA the entire TLS record up to and including the current
3378 * segment. This means that when re-transmitting the last ~1448 byte
3379 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3380 * of magnitude more data than we are sending. This can cause the
3381 * PCIe link to saturate well before the network, which can cause
3382 * output drops, and a general loss of capacity.
3383 */
3384 void
ktls_disable_ifnet(void * arg)3385 ktls_disable_ifnet(void *arg)
3386 {
3387 struct tcpcb *tp;
3388 struct inpcb *inp;
3389 struct socket *so;
3390 struct ktls_session *tls;
3391
3392 tp = arg;
3393 inp = tptoinpcb(tp);
3394 INP_WLOCK_ASSERT(inp);
3395 so = inp->inp_socket;
3396 SOCK_LOCK(so);
3397 tls = so->so_snd.sb_tls_info;
3398 if (tp->t_nic_ktls_xmit_dis == 1) {
3399 SOCK_UNLOCK(so);
3400 return;
3401 }
3402
3403 /*
3404 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3405 * ifnet can only be done once per connection, so we never want
3406 * to do it again
3407 */
3408
3409 (void)ktls_hold(tls);
3410 soref(so);
3411 tp->t_nic_ktls_xmit_dis = 1;
3412 SOCK_UNLOCK(so);
3413 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3414 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3415 }
3416
3417 void
ktls_session_to_xktls_onedir(const struct ktls_session * ktls,bool export_keys,struct xktls_session_onedir * xk)3418 ktls_session_to_xktls_onedir(const struct ktls_session *ktls, bool export_keys,
3419 struct xktls_session_onedir *xk)
3420 {
3421 if_t ifp;
3422 struct m_snd_tag *st;
3423
3424 xk->gen = ktls->gen;
3425 #define A(m) xk->m = ktls->params.m
3426 A(cipher_algorithm);
3427 A(auth_algorithm);
3428 A(cipher_key_len);
3429 A(auth_key_len);
3430 A(max_frame_len);
3431 A(tls_vmajor);
3432 A(tls_vminor);
3433 A(tls_hlen);
3434 A(tls_tlen);
3435 A(tls_bs);
3436 A(flags);
3437 if (export_keys) {
3438 memcpy(&xk->iv, &ktls->params.iv, XKTLS_SESSION_IV_BUF_LEN);
3439 A(iv_len);
3440 } else {
3441 memset(&xk->iv, 0, XKTLS_SESSION_IV_BUF_LEN);
3442 xk->iv_len = 0;
3443 }
3444 #undef A
3445 if ((st = ktls->snd_tag) != NULL &&
3446 (ifp = ktls->snd_tag->ifp) != NULL)
3447 strncpy(xk->ifnet, if_name(ifp), sizeof(xk->ifnet));
3448 }
3449
3450 void
ktls_session_copy_keys(const struct ktls_session * ktls,uint8_t * data,size_t * sz)3451 ktls_session_copy_keys(const struct ktls_session *ktls,
3452 uint8_t *data, size_t *sz)
3453 {
3454 size_t t, ta, tc;
3455
3456 if (ktls == NULL) {
3457 *sz = 0;
3458 return;
3459 }
3460 t = *sz;
3461 tc = MIN(t, ktls->params.cipher_key_len);
3462 if (data != NULL)
3463 memcpy(data, ktls->params.cipher_key, tc);
3464 ta = MIN(t - tc, ktls->params.auth_key_len);
3465 if (data != NULL)
3466 memcpy(data + tc, ktls->params.auth_key, ta);
3467 *sz = ta + tc;
3468 }
3469