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] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(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 = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)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, (void *)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;
874 struct tcpcb *tp;
875 int error;
876
877 inp = so->so_pcb;
878 INP_WLOCK(inp);
879 if (inp->inp_flags & INP_DROPPED) {
880 INP_WUNLOCK(inp);
881 return (ECONNRESET);
882 }
883 if (inp->inp_socket == NULL) {
884 INP_WUNLOCK(inp);
885 return (ECONNRESET);
886 }
887 tp = intotcpcb(inp);
888 if (!(tp->t_flags & TF_TOE)) {
889 INP_WUNLOCK(inp);
890 return (EOPNOTSUPP);
891 }
892
893 error = tcp_offload_alloc_tls_session(tp, tls, direction);
894 INP_WUNLOCK(inp);
895 if (error == 0) {
896 tls->mode = TCP_TLS_MODE_TOE;
897 switch (tls->params.cipher_algorithm) {
898 case CRYPTO_AES_CBC:
899 counter_u64_add(ktls_toe_cbc, 1);
900 break;
901 case CRYPTO_AES_NIST_GCM_16:
902 counter_u64_add(ktls_toe_gcm, 1);
903 break;
904 case CRYPTO_CHACHA20_POLY1305:
905 counter_u64_add(ktls_toe_chacha20, 1);
906 break;
907 }
908 }
909 return (error);
910 }
911 #endif
912
913 /*
914 * Common code used when first enabling ifnet TLS on a connection or
915 * when allocating a new ifnet TLS session due to a routing change.
916 * This function allocates a new TLS send tag on whatever interface
917 * the connection is currently routed over.
918 */
919 static int
ktls_alloc_snd_tag(struct inpcb * inp,struct ktls_session * tls,bool force,struct m_snd_tag ** mstp)920 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
921 struct m_snd_tag **mstp)
922 {
923 union if_snd_tag_alloc_params params;
924 struct ifnet *ifp;
925 struct nhop_object *nh;
926 struct tcpcb *tp;
927 int error;
928
929 INP_RLOCK(inp);
930 if (inp->inp_flags & INP_DROPPED) {
931 INP_RUNLOCK(inp);
932 return (ECONNRESET);
933 }
934 if (inp->inp_socket == NULL) {
935 INP_RUNLOCK(inp);
936 return (ECONNRESET);
937 }
938 tp = intotcpcb(inp);
939
940 /*
941 * Check administrative controls on ifnet TLS to determine if
942 * ifnet TLS should be denied.
943 *
944 * - Always permit 'force' requests.
945 * - ktls_ifnet_permitted == 0: always deny.
946 */
947 if (!force && ktls_ifnet_permitted == 0) {
948 INP_RUNLOCK(inp);
949 return (ENXIO);
950 }
951
952 /*
953 * XXX: Use the cached route in the inpcb to find the
954 * interface. This should perhaps instead use
955 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
956 * enabled after a connection has completed key negotiation in
957 * userland, the cached route will be present in practice.
958 */
959 nh = inp->inp_route.ro_nh;
960 if (nh == NULL) {
961 INP_RUNLOCK(inp);
962 return (ENXIO);
963 }
964 ifp = nh->nh_ifp;
965 if_ref(ifp);
966
967 /*
968 * Allocate a TLS + ratelimit tag if the connection has an
969 * existing pacing rate.
970 */
971 if (tp->t_pacing_rate != -1 &&
972 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
973 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
974 params.tls_rate_limit.inp = inp;
975 params.tls_rate_limit.tls = tls;
976 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
977 } else {
978 params.hdr.type = IF_SND_TAG_TYPE_TLS;
979 params.tls.inp = inp;
980 params.tls.tls = tls;
981 }
982 params.hdr.flowid = inp->inp_flowid;
983 params.hdr.flowtype = inp->inp_flowtype;
984 params.hdr.numa_domain = inp->inp_numa_domain;
985 INP_RUNLOCK(inp);
986
987 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
988 error = EOPNOTSUPP;
989 goto out;
990 }
991 if (inp->inp_vflag & INP_IPV6) {
992 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
993 error = EOPNOTSUPP;
994 goto out;
995 }
996 } else {
997 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
998 error = EOPNOTSUPP;
999 goto out;
1000 }
1001 }
1002 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1003 out:
1004 if_rele(ifp);
1005 return (error);
1006 }
1007
1008 /*
1009 * Allocate an initial TLS receive tag for doing HW decryption of TLS
1010 * data.
1011 *
1012 * This function allocates a new TLS receive tag on whatever interface
1013 * the connection is currently routed over. If the connection ends up
1014 * using a different interface for receive this will get fixed up via
1015 * ktls_input_ifp_mismatch as future packets arrive.
1016 */
1017 static int
ktls_alloc_rcv_tag(struct inpcb * inp,struct ktls_session * tls,struct m_snd_tag ** mstp)1018 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1019 struct m_snd_tag **mstp)
1020 {
1021 union if_snd_tag_alloc_params params;
1022 struct ifnet *ifp;
1023 struct nhop_object *nh;
1024 int error;
1025
1026 if (!ktls_ocf_recrypt_supported(tls))
1027 return (ENXIO);
1028
1029 INP_RLOCK(inp);
1030 if (inp->inp_flags & INP_DROPPED) {
1031 INP_RUNLOCK(inp);
1032 return (ECONNRESET);
1033 }
1034 if (inp->inp_socket == NULL) {
1035 INP_RUNLOCK(inp);
1036 return (ECONNRESET);
1037 }
1038
1039 /*
1040 * Check administrative controls on ifnet TLS to determine if
1041 * ifnet TLS should be denied.
1042 */
1043 if (ktls_ifnet_permitted == 0) {
1044 INP_RUNLOCK(inp);
1045 return (ENXIO);
1046 }
1047
1048 /*
1049 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1050 * the inpcb to find the interface.
1051 */
1052 nh = inp->inp_route.ro_nh;
1053 if (nh == NULL) {
1054 INP_RUNLOCK(inp);
1055 return (ENXIO);
1056 }
1057 ifp = nh->nh_ifp;
1058 if_ref(ifp);
1059 tls->rx_ifp = ifp;
1060
1061 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1062 params.hdr.flowid = inp->inp_flowid;
1063 params.hdr.flowtype = inp->inp_flowtype;
1064 params.hdr.numa_domain = inp->inp_numa_domain;
1065 params.tls_rx.inp = inp;
1066 params.tls_rx.tls = tls;
1067 params.tls_rx.vlan_id = 0;
1068
1069 INP_RUNLOCK(inp);
1070
1071 if (inp->inp_vflag & INP_IPV6) {
1072 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
1073 error = EOPNOTSUPP;
1074 goto out;
1075 }
1076 } else {
1077 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
1078 error = EOPNOTSUPP;
1079 goto out;
1080 }
1081 }
1082 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1083
1084 /*
1085 * If this connection is over a vlan, vlan_snd_tag_alloc
1086 * rewrites vlan_id with the saved interface. Save the VLAN
1087 * ID for use in ktls_reset_receive_tag which allocates new
1088 * receive tags directly from the leaf interface bypassing
1089 * if_vlan.
1090 */
1091 if (error == 0)
1092 tls->rx_vlan_id = params.tls_rx.vlan_id;
1093 out:
1094 return (error);
1095 }
1096
1097 static int
ktls_try_ifnet(struct socket * so,struct ktls_session * tls,int direction,bool force)1098 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1099 bool force)
1100 {
1101 struct m_snd_tag *mst;
1102 int error;
1103
1104 switch (direction) {
1105 case KTLS_TX:
1106 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1107 if (__predict_false(error != 0))
1108 goto done;
1109 break;
1110 case KTLS_RX:
1111 KASSERT(!force, ("%s: forced receive tag", __func__));
1112 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1113 if (__predict_false(error != 0))
1114 goto done;
1115 break;
1116 default:
1117 __assert_unreachable();
1118 }
1119
1120 tls->mode = TCP_TLS_MODE_IFNET;
1121 tls->snd_tag = mst;
1122
1123 switch (tls->params.cipher_algorithm) {
1124 case CRYPTO_AES_CBC:
1125 counter_u64_add(ktls_ifnet_cbc, 1);
1126 break;
1127 case CRYPTO_AES_NIST_GCM_16:
1128 counter_u64_add(ktls_ifnet_gcm, 1);
1129 break;
1130 case CRYPTO_CHACHA20_POLY1305:
1131 counter_u64_add(ktls_ifnet_chacha20, 1);
1132 break;
1133 default:
1134 break;
1135 }
1136 done:
1137 return (error);
1138 }
1139
1140 static void
ktls_use_sw(struct ktls_session * tls)1141 ktls_use_sw(struct ktls_session *tls)
1142 {
1143 tls->mode = TCP_TLS_MODE_SW;
1144 switch (tls->params.cipher_algorithm) {
1145 case CRYPTO_AES_CBC:
1146 counter_u64_add(ktls_sw_cbc, 1);
1147 break;
1148 case CRYPTO_AES_NIST_GCM_16:
1149 counter_u64_add(ktls_sw_gcm, 1);
1150 break;
1151 case CRYPTO_CHACHA20_POLY1305:
1152 counter_u64_add(ktls_sw_chacha20, 1);
1153 break;
1154 }
1155 }
1156
1157 static int
ktls_try_sw(struct ktls_session * tls,int direction)1158 ktls_try_sw(struct ktls_session *tls, int direction)
1159 {
1160 int error;
1161
1162 error = ktls_ocf_try(tls, direction);
1163 if (error)
1164 return (error);
1165 ktls_use_sw(tls);
1166 return (0);
1167 }
1168
1169 /*
1170 * KTLS RX stores data in the socket buffer as a list of TLS records,
1171 * where each record is stored as a control message containg the TLS
1172 * header followed by data mbufs containing the decrypted data. This
1173 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1174 * both encrypted and decrypted data. TLS records decrypted by a NIC
1175 * should be queued to the socket buffer as records, but encrypted
1176 * data which needs to be decrypted by software arrives as a stream of
1177 * regular mbufs which need to be converted. In addition, there may
1178 * already be pending encrypted data in the socket buffer when KTLS RX
1179 * is enabled.
1180 *
1181 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1182 * is used:
1183 *
1184 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1185 *
1186 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1187 * from the first mbuf. Once all of the data for that TLS record is
1188 * queued, the socket is queued to a worker thread.
1189 *
1190 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1191 * the TLS chain. Each TLS record is detached from the TLS chain,
1192 * decrypted, and inserted into the regular socket buffer chain as
1193 * record starting with a control message holding the TLS header and
1194 * a chain of mbufs holding the encrypted data.
1195 */
1196
1197 static void
sb_mark_notready(struct sockbuf * sb)1198 sb_mark_notready(struct sockbuf *sb)
1199 {
1200 struct mbuf *m;
1201
1202 m = sb->sb_mb;
1203 sb->sb_mtls = m;
1204 sb->sb_mb = NULL;
1205 sb->sb_mbtail = NULL;
1206 sb->sb_lastrecord = NULL;
1207 for (; m != NULL; m = m->m_next) {
1208 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1209 __func__));
1210 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1211 __func__));
1212 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1213 __func__));
1214 m->m_flags |= M_NOTREADY;
1215 sb->sb_acc -= m->m_len;
1216 sb->sb_tlscc += m->m_len;
1217 sb->sb_mtlstail = m;
1218 }
1219 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1220 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1221 sb->sb_ccc));
1222 }
1223
1224 /*
1225 * Return information about the pending TLS data in a socket
1226 * buffer. On return, 'seqno' is set to the sequence number
1227 * of the next TLS record to be received, 'resid' is set to
1228 * the amount of bytes still needed for the last pending
1229 * record. The function returns 'false' if the last pending
1230 * record contains a partial TLS header. In that case, 'resid'
1231 * is the number of bytes needed to complete the TLS header.
1232 */
1233 bool
ktls_pending_rx_info(struct sockbuf * sb,uint64_t * seqnop,size_t * residp)1234 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1235 {
1236 struct tls_record_layer hdr;
1237 struct mbuf *m;
1238 uint64_t seqno;
1239 size_t resid;
1240 u_int offset, record_len;
1241
1242 SOCKBUF_LOCK_ASSERT(sb);
1243 MPASS(sb->sb_flags & SB_TLS_RX);
1244 seqno = sb->sb_tls_seqno;
1245 resid = sb->sb_tlscc;
1246 m = sb->sb_mtls;
1247 offset = 0;
1248
1249 if (resid == 0) {
1250 *seqnop = seqno;
1251 *residp = 0;
1252 return (true);
1253 }
1254
1255 for (;;) {
1256 seqno++;
1257
1258 if (resid < sizeof(hdr)) {
1259 *seqnop = seqno;
1260 *residp = sizeof(hdr) - resid;
1261 return (false);
1262 }
1263
1264 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1265
1266 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1267 if (resid <= record_len) {
1268 *seqnop = seqno;
1269 *residp = record_len - resid;
1270 return (true);
1271 }
1272 resid -= record_len;
1273
1274 while (record_len != 0) {
1275 if (m->m_len - offset > record_len) {
1276 offset += record_len;
1277 break;
1278 }
1279
1280 record_len -= (m->m_len - offset);
1281 offset = 0;
1282 m = m->m_next;
1283 }
1284 }
1285 }
1286
1287 int
ktls_enable_rx(struct socket * so,struct tls_enable * en)1288 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1289 {
1290 struct ktls_session *tls;
1291 int error;
1292
1293 if (!ktls_offload_enable)
1294 return (ENOTSUP);
1295
1296 counter_u64_add(ktls_offload_enable_calls, 1);
1297
1298 /*
1299 * This should always be true since only the TCP socket option
1300 * invokes this function.
1301 */
1302 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1303 return (EINVAL);
1304
1305 /*
1306 * XXX: Don't overwrite existing sessions. We should permit
1307 * this to support rekeying in the future.
1308 */
1309 if (so->so_rcv.sb_tls_info != NULL)
1310 return (EALREADY);
1311
1312 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1313 return (ENOTSUP);
1314
1315 error = ktls_create_session(so, en, &tls, KTLS_RX);
1316 if (error)
1317 return (error);
1318
1319 error = ktls_ocf_try(tls, KTLS_RX);
1320 if (error) {
1321 ktls_free(tls);
1322 return (error);
1323 }
1324
1325 /*
1326 * Serialize with soreceive_generic() and make sure that we're not
1327 * operating on a listening socket.
1328 */
1329 error = SOCK_IO_RECV_LOCK(so, SBL_WAIT);
1330 if (error) {
1331 ktls_free(tls);
1332 return (error);
1333 }
1334
1335 /* Mark the socket as using TLS offload. */
1336 SOCK_RECVBUF_LOCK(so);
1337 if (__predict_false(so->so_rcv.sb_tls_info != NULL))
1338 error = EALREADY;
1339 else if ((so->so_rcv.sb_flags & SB_SPLICED) != 0)
1340 error = EINVAL;
1341 if (error != 0) {
1342 SOCK_RECVBUF_UNLOCK(so);
1343 SOCK_IO_RECV_UNLOCK(so);
1344 ktls_free(tls);
1345 return (EALREADY);
1346 }
1347 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1348 so->so_rcv.sb_tls_info = tls;
1349 so->so_rcv.sb_flags |= SB_TLS_RX;
1350
1351 /* Mark existing data as not ready until it can be decrypted. */
1352 sb_mark_notready(&so->so_rcv);
1353 ktls_check_rx(&so->so_rcv);
1354 SOCK_RECVBUF_UNLOCK(so);
1355 SOCK_IO_RECV_UNLOCK(so);
1356
1357 /* Prefer TOE -> ifnet TLS -> software TLS. */
1358 #ifdef TCP_OFFLOAD
1359 error = ktls_try_toe(so, tls, KTLS_RX);
1360 if (error)
1361 #endif
1362 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1363 if (error)
1364 ktls_use_sw(tls);
1365
1366 counter_u64_add(ktls_offload_total, 1);
1367
1368 return (0);
1369 }
1370
1371 int
ktls_enable_tx(struct socket * so,struct tls_enable * en)1372 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1373 {
1374 struct ktls_session *tls;
1375 struct inpcb *inp;
1376 struct tcpcb *tp;
1377 int error;
1378
1379 if (!ktls_offload_enable)
1380 return (ENOTSUP);
1381
1382 counter_u64_add(ktls_offload_enable_calls, 1);
1383
1384 /*
1385 * This should always be true since only the TCP socket option
1386 * invokes this function.
1387 */
1388 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1389 return (EINVAL);
1390
1391 /*
1392 * XXX: Don't overwrite existing sessions. We should permit
1393 * this to support rekeying in the future.
1394 */
1395 if (so->so_snd.sb_tls_info != NULL)
1396 return (EALREADY);
1397
1398 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1399 return (ENOTSUP);
1400
1401 /* TLS requires ext pgs */
1402 if (mb_use_ext_pgs == 0)
1403 return (ENXIO);
1404
1405 error = ktls_create_session(so, en, &tls, KTLS_TX);
1406 if (error)
1407 return (error);
1408
1409 /* Prefer TOE -> ifnet TLS -> software TLS. */
1410 #ifdef TCP_OFFLOAD
1411 error = ktls_try_toe(so, tls, KTLS_TX);
1412 if (error)
1413 #endif
1414 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1415 if (error)
1416 error = ktls_try_sw(tls, KTLS_TX);
1417
1418 if (error) {
1419 ktls_free(tls);
1420 return (error);
1421 }
1422
1423 /*
1424 * Serialize with sosend_generic() and make sure that we're not
1425 * operating on a listening socket.
1426 */
1427 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1428 if (error) {
1429 ktls_free(tls);
1430 return (error);
1431 }
1432
1433 /*
1434 * Write lock the INP when setting sb_tls_info so that
1435 * routines in tcp_ratelimit.c can read sb_tls_info while
1436 * holding the INP lock.
1437 */
1438 inp = so->so_pcb;
1439 INP_WLOCK(inp);
1440 SOCK_SENDBUF_LOCK(so);
1441 if (__predict_false(so->so_snd.sb_tls_info != NULL))
1442 error = EALREADY;
1443 else if ((so->so_snd.sb_flags & SB_SPLICED) != 0)
1444 error = EINVAL;
1445 if (error != 0) {
1446 SOCK_SENDBUF_UNLOCK(so);
1447 INP_WUNLOCK(inp);
1448 SOCK_IO_SEND_UNLOCK(so);
1449 ktls_free(tls);
1450 return (error);
1451 }
1452 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1453 so->so_snd.sb_tls_info = tls;
1454 if (tls->mode != TCP_TLS_MODE_SW) {
1455 tp = intotcpcb(inp);
1456 MPASS(tp->t_nic_ktls_xmit == 0);
1457 tp->t_nic_ktls_xmit = 1;
1458 if (tp->t_fb->tfb_hwtls_change != NULL)
1459 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1460 }
1461 SOCK_SENDBUF_UNLOCK(so);
1462 INP_WUNLOCK(inp);
1463 SOCK_IO_SEND_UNLOCK(so);
1464
1465 counter_u64_add(ktls_offload_total, 1);
1466
1467 return (0);
1468 }
1469
1470 int
ktls_get_rx_mode(struct socket * so,int * modep)1471 ktls_get_rx_mode(struct socket *so, int *modep)
1472 {
1473 struct ktls_session *tls;
1474 struct inpcb *inp __diagused;
1475
1476 if (SOLISTENING(so))
1477 return (EINVAL);
1478 inp = so->so_pcb;
1479 INP_WLOCK_ASSERT(inp);
1480 SOCK_RECVBUF_LOCK(so);
1481 tls = so->so_rcv.sb_tls_info;
1482 if (tls == NULL)
1483 *modep = TCP_TLS_MODE_NONE;
1484 else
1485 *modep = tls->mode;
1486 SOCK_RECVBUF_UNLOCK(so);
1487 return (0);
1488 }
1489
1490 /*
1491 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1492 *
1493 * This function gets information about the next TCP- and TLS-
1494 * sequence number to be processed by the TLS receive worker
1495 * thread. The information is extracted from the given "inpcb"
1496 * structure. The values are stored in host endian format at the two
1497 * given output pointer locations. The TCP sequence number points to
1498 * the beginning of the TLS header.
1499 *
1500 * This function returns zero on success, else a non-zero error code
1501 * is returned.
1502 */
1503 int
ktls_get_rx_sequence(struct inpcb * inp,uint32_t * tcpseq,uint64_t * tlsseq)1504 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1505 {
1506 struct socket *so;
1507 struct tcpcb *tp;
1508
1509 INP_RLOCK(inp);
1510 so = inp->inp_socket;
1511 if (__predict_false(so == NULL)) {
1512 INP_RUNLOCK(inp);
1513 return (EINVAL);
1514 }
1515 if (inp->inp_flags & INP_DROPPED) {
1516 INP_RUNLOCK(inp);
1517 return (ECONNRESET);
1518 }
1519
1520 tp = intotcpcb(inp);
1521 MPASS(tp != NULL);
1522
1523 SOCKBUF_LOCK(&so->so_rcv);
1524 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1525 *tlsseq = so->so_rcv.sb_tls_seqno;
1526 SOCKBUF_UNLOCK(&so->so_rcv);
1527
1528 INP_RUNLOCK(inp);
1529
1530 return (0);
1531 }
1532
1533 int
ktls_get_tx_mode(struct socket * so,int * modep)1534 ktls_get_tx_mode(struct socket *so, int *modep)
1535 {
1536 struct ktls_session *tls;
1537 struct inpcb *inp __diagused;
1538
1539 if (SOLISTENING(so))
1540 return (EINVAL);
1541 inp = so->so_pcb;
1542 INP_WLOCK_ASSERT(inp);
1543 SOCK_SENDBUF_LOCK(so);
1544 tls = so->so_snd.sb_tls_info;
1545 if (tls == NULL)
1546 *modep = TCP_TLS_MODE_NONE;
1547 else
1548 *modep = tls->mode;
1549 SOCK_SENDBUF_UNLOCK(so);
1550 return (0);
1551 }
1552
1553 /*
1554 * Switch between SW and ifnet TLS sessions as requested.
1555 */
1556 int
ktls_set_tx_mode(struct socket * so,int mode)1557 ktls_set_tx_mode(struct socket *so, int mode)
1558 {
1559 struct ktls_session *tls, *tls_new;
1560 struct inpcb *inp;
1561 struct tcpcb *tp;
1562 int error;
1563
1564 if (SOLISTENING(so))
1565 return (EINVAL);
1566 switch (mode) {
1567 case TCP_TLS_MODE_SW:
1568 case TCP_TLS_MODE_IFNET:
1569 break;
1570 default:
1571 return (EINVAL);
1572 }
1573
1574 inp = so->so_pcb;
1575 INP_WLOCK_ASSERT(inp);
1576 tp = intotcpcb(inp);
1577
1578 if (mode == TCP_TLS_MODE_IFNET) {
1579 /* Don't allow enabling ifnet ktls multiple times */
1580 if (tp->t_nic_ktls_xmit)
1581 return (EALREADY);
1582
1583 /*
1584 * Don't enable ifnet ktls if we disabled it due to an
1585 * excessive retransmission rate
1586 */
1587 if (tp->t_nic_ktls_xmit_dis)
1588 return (ENXIO);
1589 }
1590
1591 SOCKBUF_LOCK(&so->so_snd);
1592 tls = so->so_snd.sb_tls_info;
1593 if (tls == NULL) {
1594 SOCKBUF_UNLOCK(&so->so_snd);
1595 return (0);
1596 }
1597
1598 if (tls->mode == mode) {
1599 SOCKBUF_UNLOCK(&so->so_snd);
1600 return (0);
1601 }
1602
1603 tls = ktls_hold(tls);
1604 SOCKBUF_UNLOCK(&so->so_snd);
1605 INP_WUNLOCK(inp);
1606
1607 tls_new = ktls_clone_session(tls, KTLS_TX);
1608
1609 if (mode == TCP_TLS_MODE_IFNET)
1610 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1611 else
1612 error = ktls_try_sw(tls_new, KTLS_TX);
1613 if (error) {
1614 counter_u64_add(ktls_switch_failed, 1);
1615 ktls_free(tls_new);
1616 ktls_free(tls);
1617 INP_WLOCK(inp);
1618 return (error);
1619 }
1620
1621 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1622 if (error) {
1623 counter_u64_add(ktls_switch_failed, 1);
1624 ktls_free(tls_new);
1625 ktls_free(tls);
1626 INP_WLOCK(inp);
1627 return (error);
1628 }
1629
1630 /*
1631 * If we raced with another session change, keep the existing
1632 * session.
1633 */
1634 if (tls != so->so_snd.sb_tls_info) {
1635 counter_u64_add(ktls_switch_failed, 1);
1636 SOCK_IO_SEND_UNLOCK(so);
1637 ktls_free(tls_new);
1638 ktls_free(tls);
1639 INP_WLOCK(inp);
1640 return (EBUSY);
1641 }
1642
1643 INP_WLOCK(inp);
1644 SOCKBUF_LOCK(&so->so_snd);
1645 so->so_snd.sb_tls_info = tls_new;
1646 if (tls_new->mode != TCP_TLS_MODE_SW) {
1647 MPASS(tp->t_nic_ktls_xmit == 0);
1648 tp->t_nic_ktls_xmit = 1;
1649 if (tp->t_fb->tfb_hwtls_change != NULL)
1650 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1651 }
1652 SOCKBUF_UNLOCK(&so->so_snd);
1653 SOCK_IO_SEND_UNLOCK(so);
1654
1655 /*
1656 * Drop two references on 'tls'. The first is for the
1657 * ktls_hold() above. The second drops the reference from the
1658 * socket buffer.
1659 */
1660 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1661 ktls_free(tls);
1662 ktls_free(tls);
1663
1664 if (mode == TCP_TLS_MODE_IFNET)
1665 counter_u64_add(ktls_switch_to_ifnet, 1);
1666 else
1667 counter_u64_add(ktls_switch_to_sw, 1);
1668
1669 return (0);
1670 }
1671
1672 /*
1673 * Try to allocate a new TLS receive tag. This task is scheduled when
1674 * sbappend_ktls_rx detects an input path change. If a new tag is
1675 * allocated, replace the tag in the TLS session. If a new tag cannot
1676 * be allocated, let the session fall back to software decryption.
1677 */
1678 static void
ktls_reset_receive_tag(void * context,int pending)1679 ktls_reset_receive_tag(void *context, int pending)
1680 {
1681 union if_snd_tag_alloc_params params;
1682 struct ktls_session *tls;
1683 struct m_snd_tag *mst;
1684 struct inpcb *inp;
1685 struct ifnet *ifp;
1686 struct socket *so;
1687 int error;
1688
1689 MPASS(pending == 1);
1690
1691 tls = context;
1692 so = tls->so;
1693 inp = so->so_pcb;
1694 ifp = NULL;
1695
1696 INP_RLOCK(inp);
1697 if (inp->inp_flags & INP_DROPPED) {
1698 INP_RUNLOCK(inp);
1699 goto out;
1700 }
1701
1702 SOCKBUF_LOCK(&so->so_rcv);
1703 mst = tls->snd_tag;
1704 tls->snd_tag = NULL;
1705 if (mst != NULL)
1706 m_snd_tag_rele(mst);
1707
1708 ifp = tls->rx_ifp;
1709 if_ref(ifp);
1710 SOCKBUF_UNLOCK(&so->so_rcv);
1711
1712 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1713 params.hdr.flowid = inp->inp_flowid;
1714 params.hdr.flowtype = inp->inp_flowtype;
1715 params.hdr.numa_domain = inp->inp_numa_domain;
1716 params.tls_rx.inp = inp;
1717 params.tls_rx.tls = tls;
1718 params.tls_rx.vlan_id = tls->rx_vlan_id;
1719 INP_RUNLOCK(inp);
1720
1721 if (inp->inp_vflag & INP_IPV6) {
1722 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1723 goto out;
1724 } else {
1725 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1726 goto out;
1727 }
1728
1729 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1730 if (error == 0) {
1731 SOCKBUF_LOCK(&so->so_rcv);
1732 tls->snd_tag = mst;
1733 SOCKBUF_UNLOCK(&so->so_rcv);
1734
1735 counter_u64_add(ktls_ifnet_reset, 1);
1736 } else {
1737 /*
1738 * Just fall back to software decryption if a tag
1739 * cannot be allocated leaving the connection intact.
1740 * If a future input path change switches to another
1741 * interface this connection will resume ifnet TLS.
1742 */
1743 counter_u64_add(ktls_ifnet_reset_failed, 1);
1744 }
1745
1746 out:
1747 mtx_pool_lock(mtxpool_sleep, tls);
1748 tls->reset_pending = false;
1749 mtx_pool_unlock(mtxpool_sleep, tls);
1750
1751 if (ifp != NULL)
1752 if_rele(ifp);
1753 CURVNET_SET(so->so_vnet);
1754 sorele(so);
1755 CURVNET_RESTORE();
1756 ktls_free(tls);
1757 }
1758
1759 /*
1760 * Try to allocate a new TLS send tag. This task is scheduled when
1761 * ip_output detects a route change while trying to transmit a packet
1762 * holding a TLS record. If a new tag is allocated, replace the tag
1763 * in the TLS session. Subsequent packets on the connection will use
1764 * the new tag. If a new tag cannot be allocated, drop the
1765 * connection.
1766 */
1767 static void
ktls_reset_send_tag(void * context,int pending)1768 ktls_reset_send_tag(void *context, int pending)
1769 {
1770 struct epoch_tracker et;
1771 struct ktls_session *tls;
1772 struct m_snd_tag *old, *new;
1773 struct inpcb *inp;
1774 struct tcpcb *tp;
1775 int error;
1776
1777 MPASS(pending == 1);
1778
1779 tls = context;
1780 inp = tls->inp;
1781
1782 /*
1783 * Free the old tag first before allocating a new one.
1784 * ip[6]_output_send() will treat a NULL send tag the same as
1785 * an ifp mismatch and drop packets until a new tag is
1786 * allocated.
1787 *
1788 * Write-lock the INP when changing tls->snd_tag since
1789 * ip[6]_output_send() holds a read-lock when reading the
1790 * pointer.
1791 */
1792 INP_WLOCK(inp);
1793 old = tls->snd_tag;
1794 tls->snd_tag = NULL;
1795 INP_WUNLOCK(inp);
1796 if (old != NULL)
1797 m_snd_tag_rele(old);
1798
1799 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1800
1801 if (error == 0) {
1802 INP_WLOCK(inp);
1803 tls->snd_tag = new;
1804 mtx_pool_lock(mtxpool_sleep, tls);
1805 tls->reset_pending = false;
1806 mtx_pool_unlock(mtxpool_sleep, tls);
1807 INP_WUNLOCK(inp);
1808
1809 counter_u64_add(ktls_ifnet_reset, 1);
1810
1811 /*
1812 * XXX: Should we kick tcp_output explicitly now that
1813 * the send tag is fixed or just rely on timers?
1814 */
1815 } else {
1816 NET_EPOCH_ENTER(et);
1817 INP_WLOCK(inp);
1818 if (!(inp->inp_flags & INP_DROPPED)) {
1819 tp = intotcpcb(inp);
1820 CURVNET_SET(inp->inp_vnet);
1821 tp = tcp_drop(tp, ECONNABORTED);
1822 CURVNET_RESTORE();
1823 if (tp != NULL) {
1824 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1825 INP_WUNLOCK(inp);
1826 }
1827 } else
1828 INP_WUNLOCK(inp);
1829 NET_EPOCH_EXIT(et);
1830
1831 counter_u64_add(ktls_ifnet_reset_failed, 1);
1832
1833 /*
1834 * Leave reset_pending true to avoid future tasks while
1835 * the socket goes away.
1836 */
1837 }
1838
1839 ktls_free(tls);
1840 }
1841
1842 void
ktls_input_ifp_mismatch(struct sockbuf * sb,struct ifnet * ifp)1843 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1844 {
1845 struct ktls_session *tls;
1846 struct socket *so;
1847
1848 SOCKBUF_LOCK_ASSERT(sb);
1849 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1850 __func__, sb));
1851 so = __containerof(sb, struct socket, so_rcv);
1852
1853 tls = sb->sb_tls_info;
1854 if_rele(tls->rx_ifp);
1855 if_ref(ifp);
1856 tls->rx_ifp = ifp;
1857
1858 /*
1859 * See if we should schedule a task to update the receive tag for
1860 * this session.
1861 */
1862 mtx_pool_lock(mtxpool_sleep, tls);
1863 if (!tls->reset_pending) {
1864 (void) ktls_hold(tls);
1865 soref(so);
1866 tls->so = so;
1867 tls->reset_pending = true;
1868 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1869 }
1870 mtx_pool_unlock(mtxpool_sleep, tls);
1871 }
1872
1873 int
ktls_output_eagain(struct inpcb * inp,struct ktls_session * tls)1874 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1875 {
1876
1877 if (inp == NULL)
1878 return (ENOBUFS);
1879
1880 INP_LOCK_ASSERT(inp);
1881
1882 /*
1883 * See if we should schedule a task to update the send tag for
1884 * this session.
1885 */
1886 mtx_pool_lock(mtxpool_sleep, tls);
1887 if (!tls->reset_pending) {
1888 (void) ktls_hold(tls);
1889 tls->reset_pending = true;
1890 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1891 }
1892 mtx_pool_unlock(mtxpool_sleep, tls);
1893 return (ENOBUFS);
1894 }
1895
1896 #ifdef RATELIMIT
1897 int
ktls_modify_txrtlmt(struct ktls_session * tls,uint64_t max_pacing_rate)1898 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1899 {
1900 union if_snd_tag_modify_params params = {
1901 .rate_limit.max_rate = max_pacing_rate,
1902 .rate_limit.flags = M_NOWAIT,
1903 };
1904 struct m_snd_tag *mst;
1905
1906 /* Can't get to the inp, but it should be locked. */
1907 /* INP_LOCK_ASSERT(inp); */
1908
1909 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1910
1911 if (tls->snd_tag == NULL) {
1912 /*
1913 * Resetting send tag, ignore this change. The
1914 * pending reset may or may not see this updated rate
1915 * in the tcpcb. If it doesn't, we will just lose
1916 * this rate change.
1917 */
1918 return (0);
1919 }
1920
1921 mst = tls->snd_tag;
1922
1923 MPASS(mst != NULL);
1924 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1925
1926 return (mst->sw->snd_tag_modify(mst, ¶ms));
1927 }
1928 #endif
1929
1930 static void
ktls_destroy_help(void * context,int pending __unused)1931 ktls_destroy_help(void *context, int pending __unused)
1932 {
1933 ktls_destroy(context);
1934 }
1935
1936 void
ktls_destroy(struct ktls_session * tls)1937 ktls_destroy(struct ktls_session *tls)
1938 {
1939 struct inpcb *inp;
1940 struct tcpcb *tp;
1941 bool wlocked;
1942
1943 MPASS(tls->refcount == 0);
1944
1945 inp = tls->inp;
1946 if (tls->tx) {
1947 wlocked = INP_WLOCKED(inp);
1948 if (!wlocked && !INP_TRY_WLOCK(inp)) {
1949 /*
1950 * rwlocks read locks are anonymous, and there
1951 * is no way to know if our current thread
1952 * holds an rlock on the inp. As a rough
1953 * estimate, check to see if the thread holds
1954 * *any* rlocks at all. If it does not, then we
1955 * know that we don't hold the inp rlock, and
1956 * can safely take the wlock
1957 */
1958 if (curthread->td_rw_rlocks == 0) {
1959 INP_WLOCK(inp);
1960 } else {
1961 /*
1962 * We might hold the rlock, so let's
1963 * do the destroy in a taskqueue
1964 * context to avoid a potential
1965 * deadlock. This should be very
1966 * rare.
1967 */
1968 counter_u64_add(ktls_destroy_task, 1);
1969 TASK_INIT(&tls->destroy_task, 0,
1970 ktls_destroy_help, tls);
1971 (void)taskqueue_enqueue(taskqueue_thread,
1972 &tls->destroy_task);
1973 return;
1974 }
1975 }
1976 }
1977
1978 if (tls->sequential_records) {
1979 struct mbuf *m, *n;
1980 int page_count;
1981
1982 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1983 page_count = m->m_epg_enc_cnt;
1984 while (page_count > 0) {
1985 KASSERT(page_count >= m->m_epg_nrdy,
1986 ("%s: too few pages", __func__));
1987 page_count -= m->m_epg_nrdy;
1988 m = m_free(m);
1989 }
1990 }
1991 }
1992
1993 counter_u64_add(ktls_offload_active, -1);
1994 switch (tls->mode) {
1995 case TCP_TLS_MODE_SW:
1996 switch (tls->params.cipher_algorithm) {
1997 case CRYPTO_AES_CBC:
1998 counter_u64_add(ktls_sw_cbc, -1);
1999 break;
2000 case CRYPTO_AES_NIST_GCM_16:
2001 counter_u64_add(ktls_sw_gcm, -1);
2002 break;
2003 case CRYPTO_CHACHA20_POLY1305:
2004 counter_u64_add(ktls_sw_chacha20, -1);
2005 break;
2006 }
2007 break;
2008 case TCP_TLS_MODE_IFNET:
2009 switch (tls->params.cipher_algorithm) {
2010 case CRYPTO_AES_CBC:
2011 counter_u64_add(ktls_ifnet_cbc, -1);
2012 break;
2013 case CRYPTO_AES_NIST_GCM_16:
2014 counter_u64_add(ktls_ifnet_gcm, -1);
2015 break;
2016 case CRYPTO_CHACHA20_POLY1305:
2017 counter_u64_add(ktls_ifnet_chacha20, -1);
2018 break;
2019 }
2020 if (tls->snd_tag != NULL)
2021 m_snd_tag_rele(tls->snd_tag);
2022 if (tls->rx_ifp != NULL)
2023 if_rele(tls->rx_ifp);
2024 if (tls->tx) {
2025 INP_WLOCK_ASSERT(inp);
2026 tp = intotcpcb(inp);
2027 MPASS(tp->t_nic_ktls_xmit == 1);
2028 tp->t_nic_ktls_xmit = 0;
2029 }
2030 break;
2031 #ifdef TCP_OFFLOAD
2032 case TCP_TLS_MODE_TOE:
2033 switch (tls->params.cipher_algorithm) {
2034 case CRYPTO_AES_CBC:
2035 counter_u64_add(ktls_toe_cbc, -1);
2036 break;
2037 case CRYPTO_AES_NIST_GCM_16:
2038 counter_u64_add(ktls_toe_gcm, -1);
2039 break;
2040 case CRYPTO_CHACHA20_POLY1305:
2041 counter_u64_add(ktls_toe_chacha20, -1);
2042 break;
2043 }
2044 break;
2045 #endif
2046 }
2047 if (tls->ocf_session != NULL)
2048 ktls_ocf_free(tls);
2049 if (tls->params.auth_key != NULL) {
2050 zfree(tls->params.auth_key, M_KTLS);
2051 tls->params.auth_key = NULL;
2052 tls->params.auth_key_len = 0;
2053 }
2054 if (tls->params.cipher_key != NULL) {
2055 zfree(tls->params.cipher_key, M_KTLS);
2056 tls->params.cipher_key = NULL;
2057 tls->params.cipher_key_len = 0;
2058 }
2059 if (tls->tx) {
2060 INP_WLOCK_ASSERT(inp);
2061 if (!in_pcbrele_wlocked(inp) && !wlocked)
2062 INP_WUNLOCK(inp);
2063 }
2064 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
2065
2066 uma_zfree(ktls_session_zone, tls);
2067 }
2068
2069 void
ktls_seq(struct sockbuf * sb,struct mbuf * m)2070 ktls_seq(struct sockbuf *sb, struct mbuf *m)
2071 {
2072
2073 for (; m != NULL; m = m->m_next) {
2074 KASSERT((m->m_flags & M_EXTPG) != 0,
2075 ("ktls_seq: mapped mbuf %p", m));
2076
2077 m->m_epg_seqno = sb->sb_tls_seqno;
2078 sb->sb_tls_seqno++;
2079 }
2080 }
2081
2082 /*
2083 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
2084 * mbuf in the chain must be an unmapped mbuf. The payload of the
2085 * mbuf must be populated with the payload of each TLS record.
2086 *
2087 * The record_type argument specifies the TLS record type used when
2088 * populating the TLS header.
2089 *
2090 * The enq_count argument on return is set to the number of pages of
2091 * payload data for this entire chain that need to be encrypted via SW
2092 * encryption. The returned value should be passed to ktls_enqueue
2093 * when scheduling encryption of this chain of mbufs. To handle the
2094 * special case of empty fragments for TLS 1.0 sessions, an empty
2095 * fragment counts as one page.
2096 */
2097 void
ktls_frame(struct mbuf * top,struct ktls_session * tls,int * enq_cnt,uint8_t record_type)2098 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2099 uint8_t record_type)
2100 {
2101 struct tls_record_layer *tlshdr;
2102 struct mbuf *m;
2103 uint64_t *noncep;
2104 uint16_t tls_len;
2105 int maxlen __diagused;
2106
2107 maxlen = tls->params.max_frame_len;
2108 *enq_cnt = 0;
2109 for (m = top; m != NULL; m = m->m_next) {
2110 /*
2111 * All mbufs in the chain should be TLS records whose
2112 * payload does not exceed the maximum frame length.
2113 *
2114 * Empty TLS 1.0 records are permitted when using CBC.
2115 */
2116 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2117 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2118 ("ktls_frame: m %p len %d", m, m->m_len));
2119
2120 /*
2121 * TLS frames require unmapped mbufs to store session
2122 * info.
2123 */
2124 KASSERT((m->m_flags & M_EXTPG) != 0,
2125 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2126
2127 tls_len = m->m_len;
2128
2129 /* Save a reference to the session. */
2130 m->m_epg_tls = ktls_hold(tls);
2131
2132 m->m_epg_hdrlen = tls->params.tls_hlen;
2133 m->m_epg_trllen = tls->params.tls_tlen;
2134 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2135 int bs, delta;
2136
2137 /*
2138 * AES-CBC pads messages to a multiple of the
2139 * block size. Note that the padding is
2140 * applied after the digest and the encryption
2141 * is done on the "plaintext || mac || padding".
2142 * At least one byte of padding is always
2143 * present.
2144 *
2145 * Compute the final trailer length assuming
2146 * at most one block of padding.
2147 * tls->params.tls_tlen is the maximum
2148 * possible trailer length (padding + digest).
2149 * delta holds the number of excess padding
2150 * bytes if the maximum were used. Those
2151 * extra bytes are removed.
2152 */
2153 bs = tls->params.tls_bs;
2154 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2155 m->m_epg_trllen -= delta;
2156 }
2157 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2158
2159 /* Populate the TLS header. */
2160 tlshdr = (void *)m->m_epg_hdr;
2161 tlshdr->tls_vmajor = tls->params.tls_vmajor;
2162
2163 /*
2164 * TLS 1.3 masquarades as TLS 1.2 with a record type
2165 * of TLS_RLTYPE_APP.
2166 */
2167 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2168 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2169 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2170 tlshdr->tls_type = TLS_RLTYPE_APP;
2171 /* save the real record type for later */
2172 m->m_epg_record_type = record_type;
2173 m->m_epg_trail[0] = record_type;
2174 } else {
2175 tlshdr->tls_vminor = tls->params.tls_vminor;
2176 tlshdr->tls_type = record_type;
2177 }
2178 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2179
2180 /*
2181 * Store nonces / explicit IVs after the end of the
2182 * TLS header.
2183 *
2184 * For GCM with TLS 1.2, an 8 byte nonce is copied
2185 * from the end of the IV. The nonce is then
2186 * incremented for use by the next record.
2187 *
2188 * For CBC, a random nonce is inserted for TLS 1.1+.
2189 */
2190 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2191 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2192 noncep = (uint64_t *)(tls->params.iv + 8);
2193 be64enc(tlshdr + 1, *noncep);
2194 (*noncep)++;
2195 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2196 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2197 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2198
2199 /*
2200 * When using SW encryption, mark the mbuf not ready.
2201 * It will be marked ready via sbready() after the
2202 * record has been encrypted.
2203 *
2204 * When using ifnet TLS, unencrypted TLS records are
2205 * sent down the stack to the NIC.
2206 */
2207 if (tls->mode == TCP_TLS_MODE_SW) {
2208 m->m_flags |= M_NOTREADY;
2209 if (__predict_false(tls_len == 0)) {
2210 /* TLS 1.0 empty fragment. */
2211 m->m_epg_nrdy = 1;
2212 } else
2213 m->m_epg_nrdy = m->m_epg_npgs;
2214 *enq_cnt += m->m_epg_nrdy;
2215 }
2216 }
2217 }
2218
2219 bool
ktls_permit_empty_frames(struct ktls_session * tls)2220 ktls_permit_empty_frames(struct ktls_session *tls)
2221 {
2222 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2223 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2224 }
2225
2226 void
ktls_check_rx(struct sockbuf * sb)2227 ktls_check_rx(struct sockbuf *sb)
2228 {
2229 struct tls_record_layer hdr;
2230 struct ktls_wq *wq;
2231 struct socket *so;
2232 bool running;
2233
2234 SOCKBUF_LOCK_ASSERT(sb);
2235 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2236 __func__, sb));
2237 so = __containerof(sb, struct socket, so_rcv);
2238
2239 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2240 return;
2241
2242 /* Is there enough queued for a TLS header? */
2243 if (sb->sb_tlscc < sizeof(hdr)) {
2244 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2245 so->so_error = EMSGSIZE;
2246 return;
2247 }
2248
2249 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2250
2251 /* Is the entire record queued? */
2252 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2253 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2254 so->so_error = EMSGSIZE;
2255 return;
2256 }
2257
2258 sb->sb_flags |= SB_TLS_RX_RUNNING;
2259
2260 soref(so);
2261 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2262 mtx_lock(&wq->mtx);
2263 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2264 running = wq->running;
2265 mtx_unlock(&wq->mtx);
2266 if (!running)
2267 wakeup(wq);
2268 counter_u64_add(ktls_cnt_rx_queued, 1);
2269 }
2270
2271 static struct mbuf *
ktls_detach_record(struct sockbuf * sb,int len)2272 ktls_detach_record(struct sockbuf *sb, int len)
2273 {
2274 struct mbuf *m, *n, *top;
2275 int remain;
2276
2277 SOCKBUF_LOCK_ASSERT(sb);
2278 MPASS(len <= sb->sb_tlscc);
2279
2280 /*
2281 * If TLS chain is the exact size of the record,
2282 * just grab the whole record.
2283 */
2284 top = sb->sb_mtls;
2285 if (sb->sb_tlscc == len) {
2286 sb->sb_mtls = NULL;
2287 sb->sb_mtlstail = NULL;
2288 goto out;
2289 }
2290
2291 /*
2292 * While it would be nice to use m_split() here, we need
2293 * to know exactly what m_split() allocates to update the
2294 * accounting, so do it inline instead.
2295 */
2296 remain = len;
2297 for (m = top; remain > m->m_len; m = m->m_next)
2298 remain -= m->m_len;
2299
2300 /* Easy case: don't have to split 'm'. */
2301 if (remain == m->m_len) {
2302 sb->sb_mtls = m->m_next;
2303 if (sb->sb_mtls == NULL)
2304 sb->sb_mtlstail = NULL;
2305 m->m_next = NULL;
2306 goto out;
2307 }
2308
2309 /*
2310 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2311 * with M_NOWAIT first.
2312 */
2313 n = m_get(M_NOWAIT, MT_DATA);
2314 if (n == NULL) {
2315 /*
2316 * Use M_WAITOK with socket buffer unlocked. If
2317 * 'sb_mtls' changes while the lock is dropped, return
2318 * NULL to force the caller to retry.
2319 */
2320 SOCKBUF_UNLOCK(sb);
2321
2322 n = m_get(M_WAITOK, MT_DATA);
2323
2324 SOCKBUF_LOCK(sb);
2325 if (sb->sb_mtls != top) {
2326 m_free(n);
2327 return (NULL);
2328 }
2329 }
2330 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2331
2332 /* Store remainder in 'n'. */
2333 n->m_len = m->m_len - remain;
2334 if (m->m_flags & M_EXT) {
2335 n->m_data = m->m_data + remain;
2336 mb_dupcl(n, m);
2337 } else {
2338 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2339 }
2340
2341 /* Trim 'm' and update accounting. */
2342 m->m_len -= n->m_len;
2343 sb->sb_tlscc -= n->m_len;
2344 sb->sb_ccc -= n->m_len;
2345
2346 /* Account for 'n'. */
2347 sballoc_ktls_rx(sb, n);
2348
2349 /* Insert 'n' into the TLS chain. */
2350 sb->sb_mtls = n;
2351 n->m_next = m->m_next;
2352 if (sb->sb_mtlstail == m)
2353 sb->sb_mtlstail = n;
2354
2355 /* Detach the record from the TLS chain. */
2356 m->m_next = NULL;
2357
2358 out:
2359 MPASS(m_length(top, NULL) == len);
2360 for (m = top; m != NULL; m = m->m_next)
2361 sbfree_ktls_rx(sb, m);
2362 sb->sb_tlsdcc = len;
2363 sb->sb_ccc += len;
2364 SBCHECK(sb);
2365 return (top);
2366 }
2367
2368 /*
2369 * Determine the length of the trailing zero padding and find the real
2370 * record type in the byte before the padding.
2371 *
2372 * Walking the mbuf chain backwards is clumsy, so another option would
2373 * be to scan forwards remembering the last non-zero byte before the
2374 * trailer. However, it would be expensive to scan the entire record.
2375 * Instead, find the last non-zero byte of each mbuf in the chain
2376 * keeping track of the relative offset of that nonzero byte.
2377 *
2378 * trail_len is the size of the MAC/tag on input and is set to the
2379 * size of the full trailer including padding and the record type on
2380 * return.
2381 */
2382 static int
tls13_find_record_type(struct ktls_session * tls,struct mbuf * m,int tls_len,int * trailer_len,uint8_t * record_typep)2383 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2384 int *trailer_len, uint8_t *record_typep)
2385 {
2386 char *cp;
2387 u_int digest_start, last_offset, m_len, offset;
2388 uint8_t record_type;
2389
2390 digest_start = tls_len - *trailer_len;
2391 last_offset = 0;
2392 offset = 0;
2393 for (; m != NULL && offset < digest_start;
2394 offset += m->m_len, m = m->m_next) {
2395 /* Don't look for padding in the tag. */
2396 m_len = min(digest_start - offset, m->m_len);
2397 cp = mtod(m, char *);
2398
2399 /* Find last non-zero byte in this mbuf. */
2400 while (m_len > 0 && cp[m_len - 1] == 0)
2401 m_len--;
2402 if (m_len > 0) {
2403 record_type = cp[m_len - 1];
2404 last_offset = offset + m_len;
2405 }
2406 }
2407 if (last_offset < tls->params.tls_hlen)
2408 return (EBADMSG);
2409
2410 *record_typep = record_type;
2411 *trailer_len = tls_len - last_offset + 1;
2412 return (0);
2413 }
2414
2415 /*
2416 * Check if a mbuf chain is fully decrypted at the given offset and
2417 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2418 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2419 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2420 * is encrypted.
2421 */
2422 ktls_mbuf_crypto_st_t
ktls_mbuf_crypto_state(struct mbuf * mb,int offset,int len)2423 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2424 {
2425 int m_flags_ored = 0;
2426 int m_flags_anded = -1;
2427
2428 for (; mb != NULL; mb = mb->m_next) {
2429 if (offset < mb->m_len)
2430 break;
2431 offset -= mb->m_len;
2432 }
2433 offset += len;
2434
2435 for (; mb != NULL; mb = mb->m_next) {
2436 m_flags_ored |= mb->m_flags;
2437 m_flags_anded &= mb->m_flags;
2438
2439 if (offset <= mb->m_len)
2440 break;
2441 offset -= mb->m_len;
2442 }
2443 MPASS(mb != NULL || offset == 0);
2444
2445 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2446 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2447 else
2448 return ((m_flags_ored & M_DECRYPTED) ?
2449 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2450 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2451 }
2452
2453 /*
2454 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2455 */
2456 static int
ktls_resync_ifnet(struct socket * so,uint32_t tls_len,uint64_t tls_rcd_num)2457 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2458 {
2459 union if_snd_tag_modify_params params;
2460 struct m_snd_tag *mst;
2461 struct inpcb *inp;
2462 struct tcpcb *tp;
2463
2464 mst = so->so_rcv.sb_tls_info->snd_tag;
2465 if (__predict_false(mst == NULL))
2466 return (EINVAL);
2467
2468 inp = sotoinpcb(so);
2469 if (__predict_false(inp == NULL))
2470 return (EINVAL);
2471
2472 INP_RLOCK(inp);
2473 if (inp->inp_flags & INP_DROPPED) {
2474 INP_RUNLOCK(inp);
2475 return (ECONNRESET);
2476 }
2477
2478 tp = intotcpcb(inp);
2479 MPASS(tp != NULL);
2480
2481 /* Get the TCP sequence number of the next valid TLS header. */
2482 SOCKBUF_LOCK(&so->so_rcv);
2483 params.tls_rx.tls_hdr_tcp_sn =
2484 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2485 params.tls_rx.tls_rec_length = tls_len;
2486 params.tls_rx.tls_seq_number = tls_rcd_num;
2487 SOCKBUF_UNLOCK(&so->so_rcv);
2488
2489 INP_RUNLOCK(inp);
2490
2491 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2492 return (mst->sw->snd_tag_modify(mst, ¶ms));
2493 }
2494
2495 static void
ktls_drop(struct socket * so,int error)2496 ktls_drop(struct socket *so, int error)
2497 {
2498 struct epoch_tracker et;
2499 struct inpcb *inp = sotoinpcb(so);
2500 struct tcpcb *tp;
2501
2502 NET_EPOCH_ENTER(et);
2503 INP_WLOCK(inp);
2504 if (!(inp->inp_flags & INP_DROPPED)) {
2505 tp = intotcpcb(inp);
2506 CURVNET_SET(inp->inp_vnet);
2507 tp = tcp_drop(tp, error);
2508 CURVNET_RESTORE();
2509 if (tp != NULL)
2510 INP_WUNLOCK(inp);
2511 } else {
2512 so->so_error = error;
2513 SOCK_RECVBUF_LOCK(so);
2514 sorwakeup_locked(so);
2515 INP_WUNLOCK(inp);
2516 }
2517 NET_EPOCH_EXIT(et);
2518 }
2519
2520 static void
ktls_decrypt(struct socket * so)2521 ktls_decrypt(struct socket *so)
2522 {
2523 char tls_header[MBUF_PEXT_HDR_LEN];
2524 struct ktls_session *tls;
2525 struct sockbuf *sb;
2526 struct tls_record_layer *hdr;
2527 struct tls_get_record tgr;
2528 struct mbuf *control, *data, *m;
2529 ktls_mbuf_crypto_st_t state;
2530 uint64_t seqno;
2531 int error, remain, tls_len, trail_len;
2532 bool tls13;
2533 uint8_t vminor, record_type;
2534
2535 hdr = (struct tls_record_layer *)tls_header;
2536 sb = &so->so_rcv;
2537 SOCKBUF_LOCK(sb);
2538 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2539 ("%s: socket %p not running", __func__, so));
2540
2541 tls = sb->sb_tls_info;
2542 MPASS(tls != NULL);
2543
2544 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2545 if (tls13)
2546 vminor = TLS_MINOR_VER_TWO;
2547 else
2548 vminor = tls->params.tls_vminor;
2549 for (;;) {
2550 /* Is there enough queued for a TLS header? */
2551 if (sb->sb_tlscc < tls->params.tls_hlen)
2552 break;
2553
2554 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2555 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2556
2557 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2558 hdr->tls_vminor != vminor)
2559 error = EINVAL;
2560 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2561 error = EINVAL;
2562 else if (tls_len < tls->params.tls_hlen || tls_len >
2563 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2564 tls->params.tls_tlen)
2565 error = EMSGSIZE;
2566 else
2567 error = 0;
2568 if (__predict_false(error != 0)) {
2569 /*
2570 * We have a corrupted record and are likely
2571 * out of sync. The connection isn't
2572 * recoverable at this point, so abort it.
2573 */
2574 SOCKBUF_UNLOCK(sb);
2575 counter_u64_add(ktls_offload_corrupted_records, 1);
2576
2577 ktls_drop(so, error);
2578 goto deref;
2579 }
2580
2581 /* Is the entire record queued? */
2582 if (sb->sb_tlscc < tls_len)
2583 break;
2584
2585 /*
2586 * Split out the portion of the mbuf chain containing
2587 * this TLS record.
2588 */
2589 data = ktls_detach_record(sb, tls_len);
2590 if (data == NULL)
2591 continue;
2592 MPASS(sb->sb_tlsdcc == tls_len);
2593
2594 seqno = sb->sb_tls_seqno;
2595 sb->sb_tls_seqno++;
2596 SBCHECK(sb);
2597 SOCKBUF_UNLOCK(sb);
2598
2599 /* get crypto state for this TLS record */
2600 state = ktls_mbuf_crypto_state(data, 0, tls_len);
2601
2602 switch (state) {
2603 case KTLS_MBUF_CRYPTO_ST_MIXED:
2604 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2605 if (error)
2606 break;
2607 /* FALLTHROUGH */
2608 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2609 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2610 &trail_len);
2611 if (__predict_true(error == 0)) {
2612 if (tls13) {
2613 error = tls13_find_record_type(tls, data,
2614 tls_len, &trail_len, &record_type);
2615 } else {
2616 record_type = hdr->tls_type;
2617 }
2618 }
2619 break;
2620 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2621 /*
2622 * NIC TLS is only supported for AEAD
2623 * ciphersuites which used a fixed sized
2624 * trailer.
2625 */
2626 if (tls13) {
2627 trail_len = tls->params.tls_tlen - 1;
2628 error = tls13_find_record_type(tls, data,
2629 tls_len, &trail_len, &record_type);
2630 } else {
2631 trail_len = tls->params.tls_tlen;
2632 error = 0;
2633 record_type = hdr->tls_type;
2634 }
2635 break;
2636 default:
2637 error = EINVAL;
2638 break;
2639 }
2640 if (error) {
2641 counter_u64_add(ktls_offload_failed_crypto, 1);
2642
2643 SOCKBUF_LOCK(sb);
2644 if (sb->sb_tlsdcc == 0) {
2645 /*
2646 * sbcut/drop/flush discarded these
2647 * mbufs.
2648 */
2649 m_freem(data);
2650 break;
2651 }
2652
2653 /*
2654 * Drop this TLS record's data, but keep
2655 * decrypting subsequent records.
2656 */
2657 sb->sb_ccc -= tls_len;
2658 sb->sb_tlsdcc = 0;
2659
2660 if (error != EMSGSIZE)
2661 error = EBADMSG;
2662 CURVNET_SET(so->so_vnet);
2663 so->so_error = error;
2664 sorwakeup_locked(so);
2665 CURVNET_RESTORE();
2666
2667 m_freem(data);
2668
2669 SOCKBUF_LOCK(sb);
2670 continue;
2671 }
2672
2673 /* Allocate the control mbuf. */
2674 memset(&tgr, 0, sizeof(tgr));
2675 tgr.tls_type = record_type;
2676 tgr.tls_vmajor = hdr->tls_vmajor;
2677 tgr.tls_vminor = hdr->tls_vminor;
2678 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2679 trail_len);
2680 control = sbcreatecontrol(&tgr, sizeof(tgr),
2681 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2682
2683 SOCKBUF_LOCK(sb);
2684 if (sb->sb_tlsdcc == 0) {
2685 /* sbcut/drop/flush discarded these mbufs. */
2686 MPASS(sb->sb_tlscc == 0);
2687 m_freem(data);
2688 m_freem(control);
2689 break;
2690 }
2691
2692 /*
2693 * Clear the 'dcc' accounting in preparation for
2694 * adding the decrypted record.
2695 */
2696 sb->sb_ccc -= tls_len;
2697 sb->sb_tlsdcc = 0;
2698 SBCHECK(sb);
2699
2700 /* If there is no payload, drop all of the data. */
2701 if (tgr.tls_length == htobe16(0)) {
2702 m_freem(data);
2703 data = NULL;
2704 } else {
2705 /* Trim header. */
2706 remain = tls->params.tls_hlen;
2707 while (remain > 0) {
2708 if (data->m_len > remain) {
2709 data->m_data += remain;
2710 data->m_len -= remain;
2711 break;
2712 }
2713 remain -= data->m_len;
2714 data = m_free(data);
2715 }
2716
2717 /* Trim trailer and clear M_NOTREADY. */
2718 remain = be16toh(tgr.tls_length);
2719 m = data;
2720 for (m = data; remain > m->m_len; m = m->m_next) {
2721 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2722 remain -= m->m_len;
2723 }
2724 m->m_len = remain;
2725 m_freem(m->m_next);
2726 m->m_next = NULL;
2727 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2728
2729 /* Set EOR on the final mbuf. */
2730 m->m_flags |= M_EOR;
2731 }
2732
2733 sbappendcontrol_locked(sb, data, control, 0);
2734
2735 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2736 sb->sb_flags |= SB_TLS_RX_RESYNC;
2737 SOCKBUF_UNLOCK(sb);
2738 ktls_resync_ifnet(so, tls_len, seqno);
2739 SOCKBUF_LOCK(sb);
2740 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2741 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2742 SOCKBUF_UNLOCK(sb);
2743 ktls_resync_ifnet(so, 0, seqno);
2744 SOCKBUF_LOCK(sb);
2745 }
2746 }
2747
2748 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2749
2750 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2751 so->so_error = EMSGSIZE;
2752
2753 sorwakeup_locked(so);
2754
2755 deref:
2756 SOCKBUF_UNLOCK_ASSERT(sb);
2757
2758 CURVNET_SET(so->so_vnet);
2759 sorele(so);
2760 CURVNET_RESTORE();
2761 }
2762
2763 void
ktls_enqueue_to_free(struct mbuf * m)2764 ktls_enqueue_to_free(struct mbuf *m)
2765 {
2766 struct ktls_wq *wq;
2767 bool running;
2768
2769 /* Mark it for freeing. */
2770 m->m_epg_flags |= EPG_FLAG_2FREE;
2771 wq = &ktls_wq[m->m_epg_tls->wq_index];
2772 mtx_lock(&wq->mtx);
2773 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2774 running = wq->running;
2775 mtx_unlock(&wq->mtx);
2776 if (!running)
2777 wakeup(wq);
2778 }
2779
2780 static void *
ktls_buffer_alloc(struct ktls_wq * wq,struct mbuf * m)2781 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2782 {
2783 void *buf;
2784 int domain, running;
2785
2786 if (m->m_epg_npgs <= 2)
2787 return (NULL);
2788 if (ktls_buffer_zone == NULL)
2789 return (NULL);
2790 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2791 /*
2792 * Rate-limit allocation attempts after a failure.
2793 * ktls_buffer_import() will acquire a per-domain mutex to check
2794 * the free page queues and may fail consistently if memory is
2795 * fragmented.
2796 */
2797 return (NULL);
2798 }
2799 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2800 if (buf == NULL) {
2801 domain = PCPU_GET(domain);
2802 wq->lastallocfail = ticks;
2803
2804 /*
2805 * Note that this check is "racy", but the races are
2806 * harmless, and are either a spurious wakeup if
2807 * multiple threads fail allocations before the alloc
2808 * thread wakes, or waiting an extra second in case we
2809 * see an old value of running == true.
2810 */
2811 if (!VM_DOMAIN_EMPTY(domain)) {
2812 running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2813 if (!running)
2814 wakeup(&ktls_domains[domain].reclaim_td);
2815 }
2816 }
2817 return (buf);
2818 }
2819
2820 static int
ktls_encrypt_record(struct ktls_wq * wq,struct mbuf * m,struct ktls_session * tls,struct ktls_ocf_encrypt_state * state)2821 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2822 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2823 {
2824 vm_page_t pg;
2825 int error, i, len, off;
2826
2827 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2828 ("%p not unready & nomap mbuf\n", m));
2829 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2830 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2831 ktls_maxlen));
2832
2833 /* Anonymous mbufs are encrypted in place. */
2834 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2835 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2836
2837 /*
2838 * For file-backed mbufs (from sendfile), anonymous wired
2839 * pages are allocated and used as the encryption destination.
2840 */
2841 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2842 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2843 m->m_epg_1st_off;
2844 state->dst_iov[0].iov_base = (char *)state->cbuf +
2845 m->m_epg_1st_off;
2846 state->dst_iov[0].iov_len = len;
2847 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2848 i = 1;
2849 } else {
2850 off = m->m_epg_1st_off;
2851 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2852 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2853 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2854 len = m_epg_pagelen(m, i, off);
2855 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2856 state->dst_iov[i].iov_base =
2857 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2858 state->dst_iov[i].iov_len = len;
2859 }
2860 }
2861 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2862 state->dst_iov[i].iov_base = m->m_epg_trail;
2863 state->dst_iov[i].iov_len = m->m_epg_trllen;
2864
2865 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2866
2867 if (__predict_false(error != 0)) {
2868 /* Free the anonymous pages. */
2869 if (state->cbuf != NULL)
2870 uma_zfree(ktls_buffer_zone, state->cbuf);
2871 else {
2872 for (i = 0; i < m->m_epg_npgs; i++) {
2873 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2874 (void)vm_page_unwire_noq(pg);
2875 vm_page_free(pg);
2876 }
2877 }
2878 }
2879 return (error);
2880 }
2881
2882 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2883 static u_int
ktls_batched_records(struct mbuf * m)2884 ktls_batched_records(struct mbuf *m)
2885 {
2886 int page_count, records;
2887
2888 records = 0;
2889 page_count = m->m_epg_enc_cnt;
2890 while (page_count > 0) {
2891 records++;
2892 page_count -= m->m_epg_nrdy;
2893 m = m->m_next;
2894 }
2895 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2896 return (records);
2897 }
2898
2899 void
ktls_enqueue(struct mbuf * m,struct socket * so,int page_count)2900 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2901 {
2902 struct ktls_session *tls;
2903 struct ktls_wq *wq;
2904 int queued;
2905 bool running;
2906
2907 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2908 (M_EXTPG | M_NOTREADY)),
2909 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2910 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2911
2912 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2913
2914 m->m_epg_enc_cnt = page_count;
2915
2916 /*
2917 * Save a pointer to the socket. The caller is responsible
2918 * for taking an additional reference via soref().
2919 */
2920 m->m_epg_so = so;
2921
2922 queued = 1;
2923 tls = m->m_epg_tls;
2924 wq = &ktls_wq[tls->wq_index];
2925 mtx_lock(&wq->mtx);
2926 if (__predict_false(tls->sequential_records)) {
2927 /*
2928 * For TLS 1.0, records must be encrypted
2929 * sequentially. For a given connection, all records
2930 * queued to the associated work queue are processed
2931 * sequentially. However, sendfile(2) might complete
2932 * I/O requests spanning multiple TLS records out of
2933 * order. Here we ensure TLS records are enqueued to
2934 * the work queue in FIFO order.
2935 *
2936 * tls->next_seqno holds the sequence number of the
2937 * next TLS record that should be enqueued to the work
2938 * queue. If this next record is not tls->next_seqno,
2939 * it must be a future record, so insert it, sorted by
2940 * TLS sequence number, into tls->pending_records and
2941 * return.
2942 *
2943 * If this TLS record matches tls->next_seqno, place
2944 * it in the work queue and then check
2945 * tls->pending_records to see if any
2946 * previously-queued records are now ready for
2947 * encryption.
2948 */
2949 if (m->m_epg_seqno != tls->next_seqno) {
2950 struct mbuf *n, *p;
2951
2952 p = NULL;
2953 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2954 if (n->m_epg_seqno > m->m_epg_seqno)
2955 break;
2956 p = n;
2957 }
2958 if (n == NULL)
2959 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2960 m_epg_stailq);
2961 else if (p == NULL)
2962 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2963 m_epg_stailq);
2964 else
2965 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2966 m_epg_stailq);
2967 mtx_unlock(&wq->mtx);
2968 counter_u64_add(ktls_cnt_tx_pending, 1);
2969 return;
2970 }
2971
2972 tls->next_seqno += ktls_batched_records(m);
2973 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2974
2975 while (!STAILQ_EMPTY(&tls->pending_records)) {
2976 struct mbuf *n;
2977
2978 n = STAILQ_FIRST(&tls->pending_records);
2979 if (n->m_epg_seqno != tls->next_seqno)
2980 break;
2981
2982 queued++;
2983 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2984 tls->next_seqno += ktls_batched_records(n);
2985 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2986 }
2987 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2988 } else
2989 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2990
2991 running = wq->running;
2992 mtx_unlock(&wq->mtx);
2993 if (!running)
2994 wakeup(wq);
2995 counter_u64_add(ktls_cnt_tx_queued, queued);
2996 }
2997
2998 /*
2999 * Once a file-backed mbuf (from sendfile) has been encrypted, free
3000 * the pages from the file and replace them with the anonymous pages
3001 * allocated in ktls_encrypt_record().
3002 */
3003 static void
ktls_finish_nonanon(struct mbuf * m,struct ktls_ocf_encrypt_state * state)3004 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
3005 {
3006 int i;
3007
3008 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
3009
3010 /* Free the old pages. */
3011 m->m_ext.ext_free(m);
3012
3013 /* Replace them with the new pages. */
3014 if (state->cbuf != NULL) {
3015 for (i = 0; i < m->m_epg_npgs; i++)
3016 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
3017
3018 /* Contig pages should go back to the cache. */
3019 m->m_ext.ext_free = ktls_free_mext_contig;
3020 } else {
3021 for (i = 0; i < m->m_epg_npgs; i++)
3022 m->m_epg_pa[i] = state->parray[i];
3023
3024 /* Use the basic free routine. */
3025 m->m_ext.ext_free = mb_free_mext_pgs;
3026 }
3027
3028 /* Pages are now writable. */
3029 m->m_epg_flags |= EPG_FLAG_ANON;
3030 }
3031
3032 static __noinline void
ktls_encrypt(struct ktls_wq * wq,struct mbuf * top)3033 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
3034 {
3035 struct ktls_ocf_encrypt_state state;
3036 struct ktls_session *tls;
3037 struct socket *so;
3038 struct mbuf *m;
3039 int error, npages, total_pages;
3040
3041 so = top->m_epg_so;
3042 tls = top->m_epg_tls;
3043 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3044 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3045 #ifdef INVARIANTS
3046 top->m_epg_so = NULL;
3047 #endif
3048 total_pages = top->m_epg_enc_cnt;
3049 npages = 0;
3050
3051 /*
3052 * Encrypt the TLS records in the chain of mbufs starting with
3053 * 'top'. 'total_pages' gives us a total count of pages and is
3054 * used to know when we have finished encrypting the TLS
3055 * records originally queued with 'top'.
3056 *
3057 * NB: These mbufs are queued in the socket buffer and
3058 * 'm_next' is traversing the mbufs in the socket buffer. The
3059 * socket buffer lock is not held while traversing this chain.
3060 * Since the mbufs are all marked M_NOTREADY their 'm_next'
3061 * pointers should be stable. However, the 'm_next' of the
3062 * last mbuf encrypted is not necessarily NULL. It can point
3063 * to other mbufs appended while 'top' was on the TLS work
3064 * queue.
3065 *
3066 * Each mbuf holds an entire TLS record.
3067 */
3068 error = 0;
3069 for (m = top; npages != total_pages; m = m->m_next) {
3070 KASSERT(m->m_epg_tls == tls,
3071 ("different TLS sessions in a single mbuf chain: %p vs %p",
3072 tls, m->m_epg_tls));
3073 KASSERT(npages + m->m_epg_npgs <= total_pages,
3074 ("page count mismatch: top %p, total_pages %d, m %p", top,
3075 total_pages, m));
3076
3077 error = ktls_encrypt_record(wq, m, tls, &state);
3078 if (error) {
3079 counter_u64_add(ktls_offload_failed_crypto, 1);
3080 break;
3081 }
3082
3083 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3084 ktls_finish_nonanon(m, &state);
3085 m->m_flags |= M_RDONLY;
3086
3087 npages += m->m_epg_nrdy;
3088
3089 /*
3090 * Drop a reference to the session now that it is no
3091 * longer needed. Existing code depends on encrypted
3092 * records having no associated session vs
3093 * yet-to-be-encrypted records having an associated
3094 * session.
3095 */
3096 m->m_epg_tls = NULL;
3097 ktls_free(tls);
3098 }
3099
3100 CURVNET_SET(so->so_vnet);
3101 if (error == 0) {
3102 (void)so->so_proto->pr_ready(so, top, npages);
3103 } else {
3104 ktls_drop(so, EIO);
3105 mb_free_notready(top, total_pages);
3106 }
3107
3108 sorele(so);
3109 CURVNET_RESTORE();
3110 }
3111
3112 void
ktls_encrypt_cb(struct ktls_ocf_encrypt_state * state,int error)3113 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3114 {
3115 struct ktls_session *tls;
3116 struct socket *so;
3117 struct mbuf *m;
3118 int npages;
3119
3120 m = state->m;
3121
3122 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3123 ktls_finish_nonanon(m, state);
3124 m->m_flags |= M_RDONLY;
3125
3126 so = state->so;
3127 free(state, M_KTLS);
3128
3129 /*
3130 * Drop a reference to the session now that it is no longer
3131 * needed. Existing code depends on encrypted records having
3132 * no associated session vs yet-to-be-encrypted records having
3133 * an associated session.
3134 */
3135 tls = m->m_epg_tls;
3136 m->m_epg_tls = NULL;
3137 ktls_free(tls);
3138
3139 if (error != 0)
3140 counter_u64_add(ktls_offload_failed_crypto, 1);
3141
3142 CURVNET_SET(so->so_vnet);
3143 npages = m->m_epg_nrdy;
3144
3145 if (error == 0) {
3146 (void)so->so_proto->pr_ready(so, m, npages);
3147 } else {
3148 ktls_drop(so, EIO);
3149 mb_free_notready(m, npages);
3150 }
3151
3152 sorele(so);
3153 CURVNET_RESTORE();
3154 }
3155
3156 /*
3157 * Similar to ktls_encrypt, but used with asynchronous OCF backends
3158 * (coprocessors) where encryption does not use host CPU resources and
3159 * it can be beneficial to queue more requests than CPUs.
3160 */
3161 static __noinline void
ktls_encrypt_async(struct ktls_wq * wq,struct mbuf * top)3162 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3163 {
3164 struct ktls_ocf_encrypt_state *state;
3165 struct ktls_session *tls;
3166 struct socket *so;
3167 struct mbuf *m, *n;
3168 int error, mpages, npages, total_pages;
3169
3170 so = top->m_epg_so;
3171 tls = top->m_epg_tls;
3172 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3173 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3174 #ifdef INVARIANTS
3175 top->m_epg_so = NULL;
3176 #endif
3177 total_pages = top->m_epg_enc_cnt;
3178 npages = 0;
3179
3180 error = 0;
3181 for (m = top; npages != total_pages; m = n) {
3182 KASSERT(m->m_epg_tls == tls,
3183 ("different TLS sessions in a single mbuf chain: %p vs %p",
3184 tls, m->m_epg_tls));
3185 KASSERT(npages + m->m_epg_npgs <= total_pages,
3186 ("page count mismatch: top %p, total_pages %d, m %p", top,
3187 total_pages, m));
3188
3189 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3190 soref(so);
3191 state->so = so;
3192 state->m = m;
3193
3194 mpages = m->m_epg_nrdy;
3195 n = m->m_next;
3196
3197 error = ktls_encrypt_record(wq, m, tls, state);
3198 if (error) {
3199 counter_u64_add(ktls_offload_failed_crypto, 1);
3200 free(state, M_KTLS);
3201 CURVNET_SET(so->so_vnet);
3202 sorele(so);
3203 CURVNET_RESTORE();
3204 break;
3205 }
3206
3207 npages += mpages;
3208 }
3209
3210 CURVNET_SET(so->so_vnet);
3211 if (error != 0) {
3212 ktls_drop(so, EIO);
3213 mb_free_notready(m, total_pages - npages);
3214 }
3215
3216 sorele(so);
3217 CURVNET_RESTORE();
3218 }
3219
3220 static int
ktls_bind_domain(int domain)3221 ktls_bind_domain(int domain)
3222 {
3223 int error;
3224
3225 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3226 if (error != 0)
3227 return (error);
3228 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3229 return (0);
3230 }
3231
3232 static void
ktls_reclaim_thread(void * ctx)3233 ktls_reclaim_thread(void *ctx)
3234 {
3235 struct ktls_domain_info *ktls_domain = ctx;
3236 struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3237 struct sysctl_oid *oid;
3238 char name[80];
3239 int error, domain;
3240
3241 domain = ktls_domain - ktls_domains;
3242 if (bootverbose)
3243 printf("Starting KTLS reclaim thread for domain %d\n", domain);
3244 error = ktls_bind_domain(domain);
3245 if (error)
3246 printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3247 domain, error);
3248 snprintf(name, sizeof(name), "domain%d", domain);
3249 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3250 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3251 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3252 CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed");
3253 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3254 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3255 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3256 CTLFLAG_RD, &sc->running, 0, "thread running");
3257
3258 for (;;) {
3259 atomic_store_int(&sc->running, 0);
3260 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3261 atomic_store_int(&sc->running, 1);
3262 sc->wakeups++;
3263 /*
3264 * Below we attempt to reclaim ktls_max_reclaim
3265 * buffers using vm_page_reclaim_contig_domain_ext().
3266 * We do this here, as this function can take several
3267 * seconds to scan all of memory and it does not
3268 * matter if this thread pauses for a while. If we
3269 * block a ktls worker thread, we risk developing
3270 * backlogs of buffers to be encrypted, leading to
3271 * surges of traffic and potential NIC output drops.
3272 */
3273 if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3274 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3275 ktls_max_reclaim) != 0) {
3276 vm_wait_domain(domain);
3277 } else {
3278 sc->reclaims += ktls_max_reclaim;
3279 }
3280 }
3281 }
3282
3283 static void
ktls_work_thread(void * ctx)3284 ktls_work_thread(void *ctx)
3285 {
3286 struct ktls_wq *wq = ctx;
3287 struct mbuf *m, *n;
3288 struct socket *so, *son;
3289 STAILQ_HEAD(, mbuf) local_m_head;
3290 STAILQ_HEAD(, socket) local_so_head;
3291 int cpu;
3292
3293 cpu = wq - ktls_wq;
3294 if (bootverbose)
3295 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3296
3297 /*
3298 * Bind to a core. If ktls_bind_threads is > 1, then
3299 * we bind to the NUMA domain instead.
3300 */
3301 if (ktls_bind_threads) {
3302 int error;
3303
3304 if (ktls_bind_threads > 1) {
3305 struct pcpu *pc = pcpu_find(cpu);
3306
3307 error = ktls_bind_domain(pc->pc_domain);
3308 } else {
3309 cpuset_t mask;
3310
3311 CPU_SETOF(cpu, &mask);
3312 error = cpuset_setthread(curthread->td_tid, &mask);
3313 }
3314 if (error)
3315 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3316 cpu, error);
3317 }
3318 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3319 fpu_kern_thread(0);
3320 #endif
3321 for (;;) {
3322 mtx_lock(&wq->mtx);
3323 while (STAILQ_EMPTY(&wq->m_head) &&
3324 STAILQ_EMPTY(&wq->so_head)) {
3325 wq->running = false;
3326 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3327 wq->running = true;
3328 }
3329
3330 STAILQ_INIT(&local_m_head);
3331 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3332 STAILQ_INIT(&local_so_head);
3333 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3334 mtx_unlock(&wq->mtx);
3335
3336 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3337 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3338 ktls_free(m->m_epg_tls);
3339 m_free_raw(m);
3340 } else {
3341 if (m->m_epg_tls->sync_dispatch)
3342 ktls_encrypt(wq, m);
3343 else
3344 ktls_encrypt_async(wq, m);
3345 counter_u64_add(ktls_cnt_tx_queued, -1);
3346 }
3347 }
3348
3349 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3350 ktls_decrypt(so);
3351 counter_u64_add(ktls_cnt_rx_queued, -1);
3352 }
3353 }
3354 }
3355
3356 static void
ktls_disable_ifnet_help(void * context,int pending __unused)3357 ktls_disable_ifnet_help(void *context, int pending __unused)
3358 {
3359 struct ktls_session *tls;
3360 struct inpcb *inp;
3361 struct tcpcb *tp;
3362 struct socket *so;
3363 int err;
3364
3365 tls = context;
3366 inp = tls->inp;
3367 if (inp == NULL)
3368 return;
3369 INP_WLOCK(inp);
3370 so = inp->inp_socket;
3371 MPASS(so != NULL);
3372 if (inp->inp_flags & INP_DROPPED) {
3373 goto out;
3374 }
3375
3376 if (so->so_snd.sb_tls_info != NULL)
3377 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3378 else
3379 err = ENXIO;
3380 if (err == 0) {
3381 counter_u64_add(ktls_ifnet_disable_ok, 1);
3382 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3383 if ((inp->inp_flags & INP_DROPPED) == 0 &&
3384 (tp = intotcpcb(inp)) != NULL &&
3385 tp->t_fb->tfb_hwtls_change != NULL)
3386 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3387 } else {
3388 counter_u64_add(ktls_ifnet_disable_fail, 1);
3389 }
3390
3391 out:
3392 CURVNET_SET(so->so_vnet);
3393 sorele(so);
3394 CURVNET_RESTORE();
3395 INP_WUNLOCK(inp);
3396 ktls_free(tls);
3397 }
3398
3399 /*
3400 * Called when re-transmits are becoming a substantial portion of the
3401 * sends on this connection. When this happens, we transition the
3402 * connection to software TLS. This is needed because most inline TLS
3403 * NICs keep crypto state only for in-order transmits. This means
3404 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3405 * re-DMA the entire TLS record up to and including the current
3406 * segment. This means that when re-transmitting the last ~1448 byte
3407 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3408 * of magnitude more data than we are sending. This can cause the
3409 * PCIe link to saturate well before the network, which can cause
3410 * output drops, and a general loss of capacity.
3411 */
3412 void
ktls_disable_ifnet(void * arg)3413 ktls_disable_ifnet(void *arg)
3414 {
3415 struct tcpcb *tp;
3416 struct inpcb *inp;
3417 struct socket *so;
3418 struct ktls_session *tls;
3419
3420 tp = arg;
3421 inp = tptoinpcb(tp);
3422 INP_WLOCK_ASSERT(inp);
3423 so = inp->inp_socket;
3424 SOCK_LOCK(so);
3425 tls = so->so_snd.sb_tls_info;
3426 if (tp->t_nic_ktls_xmit_dis == 1) {
3427 SOCK_UNLOCK(so);
3428 return;
3429 }
3430
3431 /*
3432 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3433 * ifnet can only be done once per connection, so we never want
3434 * to do it again
3435 */
3436
3437 (void)ktls_hold(tls);
3438 soref(so);
3439 tp->t_nic_ktls_xmit_dis = 1;
3440 SOCK_UNLOCK(so);
3441 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3442 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3443 }
3444
3445 void
ktls_session_to_xktls_onedir(const struct ktls_session * ktls,bool export_keys,struct xktls_session_onedir * xk)3446 ktls_session_to_xktls_onedir(const struct ktls_session *ktls, bool export_keys,
3447 struct xktls_session_onedir *xk)
3448 {
3449 if_t ifp;
3450 struct m_snd_tag *st;
3451
3452 xk->gen = ktls->gen;
3453 #define A(m) xk->m = ktls->params.m
3454 A(cipher_algorithm);
3455 A(auth_algorithm);
3456 A(cipher_key_len);
3457 A(auth_key_len);
3458 A(max_frame_len);
3459 A(tls_vmajor);
3460 A(tls_vminor);
3461 A(tls_hlen);
3462 A(tls_tlen);
3463 A(tls_bs);
3464 A(flags);
3465 if (export_keys) {
3466 memcpy(&xk->iv, &ktls->params.iv, XKTLS_SESSION_IV_BUF_LEN);
3467 A(iv_len);
3468 } else {
3469 memset(&xk->iv, 0, XKTLS_SESSION_IV_BUF_LEN);
3470 xk->iv_len = 0;
3471 }
3472 #undef A
3473 if ((st = ktls->snd_tag) != NULL &&
3474 (ifp = ktls->snd_tag->ifp) != NULL)
3475 strncpy(xk->ifnet, if_name(ifp), sizeof(xk->ifnet));
3476 }
3477
3478 void
ktls_session_copy_keys(const struct ktls_session * ktls,uint8_t * data,size_t * sz)3479 ktls_session_copy_keys(const struct ktls_session *ktls,
3480 uint8_t *data, size_t *sz)
3481 {
3482 size_t t, ta, tc;
3483
3484 if (ktls == NULL) {
3485 *sz = 0;
3486 return;
3487 }
3488 t = *sz;
3489 tc = MIN(t, ktls->params.cipher_key_len);
3490 if (data != NULL)
3491 memcpy(data, ktls->params.cipher_key, tc);
3492 ta = MIN(t - tc, ktls->params.auth_key_len);
3493 if (data != NULL)
3494 memcpy(data + tc, ktls->params.auth_key, ta);
3495 *sz = ta + tc;
3496 }
3497