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 typedef enum {
81 KTLS_MBUF_CRYPTO_ST_MIXED = 0,
82 KTLS_MBUF_CRYPTO_ST_ENCRYPTED = 1,
83 KTLS_MBUF_CRYPTO_ST_DECRYPTED = -1,
84 KTLS_MBUF_CRYPTO_ST_SHAREDMBUF = -2,
85 } ktls_mbuf_crypto_st_t;
86
87 struct ktls_wq {
88 struct mtx mtx;
89 STAILQ_HEAD(, mbuf) m_head;
90 STAILQ_HEAD(, socket) so_head;
91 bool running;
92 int lastallocfail;
93 } __aligned(CACHE_LINE_SIZE);
94
95 struct ktls_reclaim_thread {
96 uint64_t wakeups;
97 uint64_t reclaims;
98 struct thread *td;
99 int running;
100 };
101
102 struct ktls_domain_info {
103 int count;
104 int cpu[MAXCPU];
105 struct ktls_reclaim_thread reclaim_td;
106 };
107
108 struct ktls_domain_info ktls_domains[MAXMEMDOM];
109 static struct ktls_wq *ktls_wq;
110 static struct proc *ktls_proc;
111 static uma_zone_t ktls_session_zone;
112 static uma_zone_t ktls_buffer_zone;
113 static uint16_t ktls_cpuid_lookup[MAXCPU];
114 static int ktls_init_state;
115 static struct sx ktls_init_lock;
116 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
117
118 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
119 "Kernel TLS offload");
120 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
121 "Kernel TLS offload stats");
122
123 #ifdef RSS
124 static int ktls_bind_threads = 1;
125 #else
126 static int ktls_bind_threads;
127 #endif
128 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
129 &ktls_bind_threads, 0,
130 "Bind crypto threads to cores (1) or cores and domains (2) at boot");
131
132 static u_int ktls_maxlen = 16384;
133 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
134 &ktls_maxlen, 0, "Maximum TLS record size");
135
136 static int ktls_number_threads;
137 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
138 &ktls_number_threads, 0,
139 "Number of TLS threads in thread-pool");
140
141 unsigned int ktls_ifnet_max_rexmit_pct = 2;
142 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
143 &ktls_ifnet_max_rexmit_pct, 2,
144 "Max percent bytes retransmitted before ifnet TLS is disabled");
145
146 static bool ktls_offload_enable = true;
147 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
148 &ktls_offload_enable, 0,
149 "Enable support for kernel TLS offload");
150
151 static bool ktls_cbc_enable = true;
152 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
153 &ktls_cbc_enable, 1,
154 "Enable support of AES-CBC crypto for kernel TLS");
155
156 static bool ktls_sw_buffer_cache = true;
157 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
158 &ktls_sw_buffer_cache, 1,
159 "Enable caching of output buffers for SW encryption");
160
161 static int ktls_max_reclaim = 1024;
162 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN,
163 &ktls_max_reclaim, 128,
164 "Max number of 16k buffers to reclaim in thread context");
165
166 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
167 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
168 &ktls_tasks_active, "Number of active tasks");
169
170 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
171 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
172 &ktls_cnt_tx_pending,
173 "Number of TLS 1.0 records waiting for earlier TLS records");
174
175 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
176 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
177 &ktls_cnt_tx_queued,
178 "Number of TLS records in queue to tasks for SW encryption");
179
180 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
182 &ktls_cnt_rx_queued,
183 "Number of TLS sockets in queue to tasks for SW decryption");
184
185 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
186 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
187 CTLFLAG_RD, &ktls_offload_total,
188 "Total successful TLS setups (parameters set)");
189
190 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
191 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
192 CTLFLAG_RD, &ktls_offload_enable_calls,
193 "Total number of TLS enable calls made");
194
195 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
197 &ktls_offload_active, "Total Active TLS sessions");
198
199 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
200 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
201 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
202
203 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
204 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
205 &ktls_offload_failed_crypto, "Total TLS crypto failures");
206
207 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
208 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
209 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
210
211 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
212 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
213 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
214
215 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
216 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
217 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
218
219 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
220 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
221 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
222
223 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
224 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
225 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
226
227 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
228 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
229 &ktls_destroy_task,
230 "Number of times ktls session was destroyed via taskqueue");
231
232 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
233 "Software TLS session stats");
234 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
235 "Hardware (ifnet) TLS session stats");
236 #ifdef TCP_OFFLOAD
237 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
238 "TOE TLS session stats");
239 #endif
240
241 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
242 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
243 "Active number of software TLS sessions using AES-CBC");
244
245 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
246 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
247 "Active number of software TLS sessions using AES-GCM");
248
249 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
250 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
251 &ktls_sw_chacha20,
252 "Active number of software TLS sessions using Chacha20-Poly1305");
253
254 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
255 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
256 &ktls_ifnet_cbc,
257 "Active number of ifnet TLS sessions using AES-CBC");
258
259 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
260 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
261 &ktls_ifnet_gcm,
262 "Active number of ifnet TLS sessions using AES-GCM");
263
264 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
265 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
266 &ktls_ifnet_chacha20,
267 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
268
269 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
270 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
271 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
272
273 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
274 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
275 &ktls_ifnet_reset_dropped,
276 "TLS sessions dropped after failing to update ifnet send tag");
277
278 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
279 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
280 &ktls_ifnet_reset_failed,
281 "TLS sessions that failed to allocate a new ifnet send tag");
282
283 static int ktls_ifnet_permitted = 1;
284 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
285 &ktls_ifnet_permitted, 1,
286 "Whether to permit hardware (ifnet) TLS sessions");
287
288 #ifdef TCP_OFFLOAD
289 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
290 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
291 &ktls_toe_cbc,
292 "Active number of TOE TLS sessions using AES-CBC");
293
294 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
295 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
296 &ktls_toe_gcm,
297 "Active number of TOE TLS sessions using AES-GCM");
298
299 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
300 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
301 &ktls_toe_chacha20,
302 "Active number of TOE TLS sessions using Chacha20-Poly1305");
303 #endif
304
305 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
306
307 static void ktls_reclaim_thread(void *ctx);
308 static void ktls_reset_receive_tag(void *context, int pending);
309 static void ktls_reset_send_tag(void *context, int pending);
310 static void ktls_work_thread(void *ctx);
311
312 int
ktls_copyin_tls_enable(struct sockopt * sopt,struct tls_enable * tls)313 ktls_copyin_tls_enable(struct sockopt *sopt, struct tls_enable *tls)
314 {
315 struct tls_enable_v0 tls_v0;
316 int error;
317 uint8_t *cipher_key = NULL, *iv = NULL, *auth_key = NULL;
318
319 if (sopt->sopt_valsize == sizeof(tls_v0)) {
320 error = sooptcopyin(sopt, &tls_v0, sizeof(tls_v0), sizeof(tls_v0));
321 if (error != 0)
322 goto done;
323 memset(tls, 0, sizeof(*tls));
324 tls->cipher_key = tls_v0.cipher_key;
325 tls->iv = tls_v0.iv;
326 tls->auth_key = tls_v0.auth_key;
327 tls->cipher_algorithm = tls_v0.cipher_algorithm;
328 tls->cipher_key_len = tls_v0.cipher_key_len;
329 tls->iv_len = tls_v0.iv_len;
330 tls->auth_algorithm = tls_v0.auth_algorithm;
331 tls->auth_key_len = tls_v0.auth_key_len;
332 tls->flags = tls_v0.flags;
333 tls->tls_vmajor = tls_v0.tls_vmajor;
334 tls->tls_vminor = tls_v0.tls_vminor;
335 } else
336 error = sooptcopyin(sopt, tls, sizeof(*tls), sizeof(*tls));
337
338 if (error != 0)
339 return (error);
340
341 if (tls->cipher_key_len < 0 || tls->cipher_key_len > TLS_MAX_PARAM_SIZE)
342 return (EINVAL);
343 if (tls->iv_len < 0 || tls->iv_len > sizeof(((struct ktls_session *)NULL)->params.iv))
344 return (EINVAL);
345 if (tls->auth_key_len < 0 || tls->auth_key_len > TLS_MAX_PARAM_SIZE)
346 return (EINVAL);
347
348 /* All supported algorithms require a cipher key. */
349 if (tls->cipher_key_len == 0)
350 return (EINVAL);
351
352 /*
353 * Now do a deep copy of the variable-length arrays in the struct, so that
354 * subsequent consumers of it can reliably assume kernel memory. This
355 * requires doing our own allocations, which we will free in the
356 * error paths so that our caller need only worry about outstanding
357 * allocations existing on successful return.
358 */
359 if (tls->cipher_key_len != 0) {
360 cipher_key = malloc(tls->cipher_key_len, M_KTLS, M_WAITOK);
361 if (sopt->sopt_td != NULL) {
362 error = copyin(tls->cipher_key, cipher_key, tls->cipher_key_len);
363 if (error != 0)
364 goto done;
365 } else {
366 bcopy(tls->cipher_key, cipher_key, tls->cipher_key_len);
367 }
368 }
369 if (tls->iv_len != 0) {
370 iv = malloc(tls->iv_len, M_KTLS, M_WAITOK);
371 if (sopt->sopt_td != NULL) {
372 error = copyin(tls->iv, iv, tls->iv_len);
373 if (error != 0)
374 goto done;
375 } else {
376 bcopy(tls->iv, iv, tls->iv_len);
377 }
378 }
379 if (tls->auth_key_len != 0) {
380 auth_key = malloc(tls->auth_key_len, M_KTLS, M_WAITOK);
381 if (sopt->sopt_td != NULL) {
382 error = copyin(tls->auth_key, auth_key, tls->auth_key_len);
383 if (error != 0)
384 goto done;
385 } else {
386 bcopy(tls->auth_key, auth_key, tls->auth_key_len);
387 }
388 }
389 tls->cipher_key = cipher_key;
390 tls->iv = iv;
391 tls->auth_key = auth_key;
392
393 done:
394 if (error != 0) {
395 zfree(cipher_key, M_KTLS);
396 zfree(iv, M_KTLS);
397 zfree(auth_key, M_KTLS);
398 }
399
400 return (error);
401 }
402
403 void
ktls_cleanup_tls_enable(struct tls_enable * tls)404 ktls_cleanup_tls_enable(struct tls_enable *tls)
405 {
406 zfree(__DECONST(void *, tls->cipher_key), M_KTLS);
407 zfree(__DECONST(void *, tls->iv), M_KTLS);
408 zfree(__DECONST(void *, tls->auth_key), M_KTLS);
409 }
410
411 static u_int
ktls_get_cpu(struct socket * so)412 ktls_get_cpu(struct socket *so)
413 {
414 struct inpcb *inp;
415 #ifdef NUMA
416 struct ktls_domain_info *di;
417 #endif
418 u_int cpuid;
419
420 inp = sotoinpcb(so);
421 #ifdef RSS
422 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
423 if (cpuid != NETISR_CPUID_NONE)
424 return (cpuid);
425 #endif
426 /*
427 * Just use the flowid to shard connections in a repeatable
428 * fashion. Note that TLS 1.0 sessions rely on the
429 * serialization provided by having the same connection use
430 * the same queue.
431 */
432 #ifdef NUMA
433 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
434 di = &ktls_domains[inp->inp_numa_domain];
435 cpuid = di->cpu[inp->inp_flowid % di->count];
436 } else
437 #endif
438 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
439 return (cpuid);
440 }
441
442 static int
ktls_buffer_import(void * arg,void ** store,int count,int domain,int flags)443 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
444 {
445 vm_page_t m;
446 int i, req;
447
448 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
449 ("%s: ktls max length %d is not page size-aligned",
450 __func__, ktls_maxlen));
451
452 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
453 for (i = 0; i < count; i++) {
454 m = vm_page_alloc_noobj_contig_domain(domain, req,
455 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
456 VM_MEMATTR_DEFAULT);
457 if (m == NULL)
458 break;
459 store[i] = VM_PAGE_TO_DMAP(m);
460 }
461 return (i);
462 }
463
464 static void
ktls_buffer_release(void * arg __unused,void ** store,int count)465 ktls_buffer_release(void *arg __unused, void **store, int count)
466 {
467 vm_page_t m;
468 int i, j;
469
470 for (i = 0; i < count; i++) {
471 m = DMAP_TO_VM_PAGE(store[i]);
472 for (j = 0; j < atop(ktls_maxlen); j++) {
473 (void)vm_page_unwire_noq(m + j);
474 vm_page_free(m + j);
475 }
476 }
477 }
478
479 static void
ktls_free_mext_contig(struct mbuf * m)480 ktls_free_mext_contig(struct mbuf *m)
481 {
482 M_ASSERTEXTPG(m);
483 uma_zfree(ktls_buffer_zone, PHYS_TO_DMAP(m->m_epg_pa[0]));
484 }
485
486 static int
ktls_init(void)487 ktls_init(void)
488 {
489 struct thread *td;
490 struct pcpu *pc;
491 int count, domain, error, i;
492
493 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
494 M_WAITOK | M_ZERO);
495
496 ktls_session_zone = uma_zcreate("ktls_session",
497 sizeof(struct ktls_session),
498 NULL, NULL, NULL, NULL,
499 UMA_ALIGN_CACHE, 0);
500
501 if (ktls_sw_buffer_cache) {
502 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
503 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
504 ktls_buffer_import, ktls_buffer_release, NULL,
505 UMA_ZONE_FIRSTTOUCH | UMA_ZONE_NOTRIM);
506 }
507
508 /*
509 * Initialize the workqueues to run the TLS work. We create a
510 * work queue for each CPU.
511 */
512 CPU_FOREACH(i) {
513 STAILQ_INIT(&ktls_wq[i].m_head);
514 STAILQ_INIT(&ktls_wq[i].so_head);
515 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
516 if (ktls_bind_threads > 1) {
517 pc = pcpu_find(i);
518 domain = pc->pc_domain;
519 count = ktls_domains[domain].count;
520 ktls_domains[domain].cpu[count] = i;
521 ktls_domains[domain].count++;
522 }
523 ktls_cpuid_lookup[ktls_number_threads] = i;
524 ktls_number_threads++;
525 }
526
527 /*
528 * If we somehow have an empty domain, fall back to choosing
529 * among all KTLS threads.
530 */
531 if (ktls_bind_threads > 1) {
532 for (i = 0; i < vm_ndomains; i++) {
533 if (ktls_domains[i].count == 0) {
534 ktls_bind_threads = 1;
535 break;
536 }
537 }
538 }
539
540 /* Start kthreads for each workqueue. */
541 CPU_FOREACH(i) {
542 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
543 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
544 if (error) {
545 printf("Can't add KTLS thread %d error %d\n", i, error);
546 return (error);
547 }
548 }
549
550 /*
551 * Start an allocation thread per-domain to perform blocking allocations
552 * of 16k physically contiguous TLS crypto destination buffers.
553 */
554 if (ktls_sw_buffer_cache) {
555 for (domain = 0; domain < vm_ndomains; domain++) {
556 if (VM_DOMAIN_EMPTY(domain))
557 continue;
558 if (CPU_EMPTY(&cpuset_domain[domain]))
559 continue;
560 error = kproc_kthread_add(ktls_reclaim_thread,
561 &ktls_domains[domain], &ktls_proc,
562 &ktls_domains[domain].reclaim_td.td,
563 0, 0, "KTLS", "reclaim_%d", domain);
564 if (error) {
565 printf("Can't add KTLS reclaim thread %d error %d\n",
566 domain, error);
567 return (error);
568 }
569 }
570 }
571
572 if (bootverbose)
573 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
574 return (0);
575 }
576
577 static int
ktls_start_kthreads(void)578 ktls_start_kthreads(void)
579 {
580 int error, state;
581
582 start:
583 state = atomic_load_acq_int(&ktls_init_state);
584 if (__predict_true(state > 0))
585 return (0);
586 if (state < 0)
587 return (ENXIO);
588
589 sx_xlock(&ktls_init_lock);
590 if (ktls_init_state != 0) {
591 sx_xunlock(&ktls_init_lock);
592 goto start;
593 }
594
595 error = ktls_init();
596 if (error == 0)
597 state = 1;
598 else
599 state = -1;
600 atomic_store_rel_int(&ktls_init_state, state);
601 sx_xunlock(&ktls_init_lock);
602 return (error);
603 }
604
605 static int
ktls_create_session(struct socket * so,struct tls_enable * en,struct ktls_session ** tlsp,int direction)606 ktls_create_session(struct socket *so, struct tls_enable *en,
607 struct ktls_session **tlsp, int direction)
608 {
609 struct ktls_session *tls;
610 int error;
611
612 /* Only TLS 1.0 - 1.3 are supported. */
613 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
614 return (EINVAL);
615 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
616 en->tls_vminor > TLS_MINOR_VER_THREE)
617 return (EINVAL);
618
619
620 /* No flags are currently supported. */
621 if (en->flags != 0)
622 return (EINVAL);
623
624 /* Common checks for supported algorithms. */
625 switch (en->cipher_algorithm) {
626 case CRYPTO_AES_NIST_GCM_16:
627 /*
628 * auth_algorithm isn't used, but permit GMAC values
629 * for compatibility.
630 */
631 switch (en->auth_algorithm) {
632 case 0:
633 #ifdef COMPAT_FREEBSD12
634 /* XXX: Really 13.0-current COMPAT. */
635 case CRYPTO_AES_128_NIST_GMAC:
636 case CRYPTO_AES_192_NIST_GMAC:
637 case CRYPTO_AES_256_NIST_GMAC:
638 #endif
639 break;
640 default:
641 return (EINVAL);
642 }
643 if (en->auth_key_len != 0)
644 return (EINVAL);
645 switch (en->tls_vminor) {
646 case TLS_MINOR_VER_TWO:
647 if (en->iv_len != TLS_AEAD_GCM_LEN)
648 return (EINVAL);
649 break;
650 case TLS_MINOR_VER_THREE:
651 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
652 return (EINVAL);
653 break;
654 default:
655 return (EINVAL);
656 }
657 break;
658 case CRYPTO_AES_CBC:
659 switch (en->auth_algorithm) {
660 case CRYPTO_SHA1_HMAC:
661 break;
662 case CRYPTO_SHA2_256_HMAC:
663 case CRYPTO_SHA2_384_HMAC:
664 if (en->tls_vminor != TLS_MINOR_VER_TWO)
665 return (EINVAL);
666 break;
667 default:
668 return (EINVAL);
669 }
670 if (en->auth_key_len == 0)
671 return (EINVAL);
672
673 /*
674 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
675 * use explicit IVs.
676 */
677 switch (en->tls_vminor) {
678 case TLS_MINOR_VER_ZERO:
679 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
680 return (EINVAL);
681 break;
682 case TLS_MINOR_VER_ONE:
683 case TLS_MINOR_VER_TWO:
684 /* Ignore any supplied IV. */
685 en->iv_len = 0;
686 break;
687 default:
688 return (EINVAL);
689 }
690 break;
691 case CRYPTO_CHACHA20_POLY1305:
692 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
693 return (EINVAL);
694 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
695 en->tls_vminor != TLS_MINOR_VER_THREE)
696 return (EINVAL);
697 if (en->iv_len != TLS_CHACHA20_IV_LEN)
698 return (EINVAL);
699 break;
700 default:
701 return (EINVAL);
702 }
703
704 error = ktls_start_kthreads();
705 if (error != 0)
706 return (error);
707
708 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
709
710 counter_u64_add(ktls_offload_active, 1);
711
712 refcount_init(&tls->refcount, 1);
713 if (direction == KTLS_RX) {
714 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
715 } else {
716 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
717 tls->inp = so->so_pcb;
718 in_pcbref(tls->inp);
719 tls->tx = true;
720 }
721
722 tls->wq_index = ktls_get_cpu(so);
723
724 tls->params.cipher_algorithm = en->cipher_algorithm;
725 tls->params.auth_algorithm = en->auth_algorithm;
726 tls->params.tls_vmajor = en->tls_vmajor;
727 tls->params.tls_vminor = en->tls_vminor;
728 tls->params.flags = en->flags;
729 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
730
731 /* Set the header and trailer lengths. */
732 tls->params.tls_hlen = sizeof(struct tls_record_layer);
733 switch (en->cipher_algorithm) {
734 case CRYPTO_AES_NIST_GCM_16:
735 /*
736 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
737 * nonce. TLS 1.3 uses a 12 byte implicit IV.
738 */
739 if (en->tls_vminor < TLS_MINOR_VER_THREE)
740 tls->params.tls_hlen += sizeof(uint64_t);
741 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
742 tls->params.tls_bs = 1;
743 break;
744 case CRYPTO_AES_CBC:
745 switch (en->auth_algorithm) {
746 case CRYPTO_SHA1_HMAC:
747 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
748 /* Implicit IV, no nonce. */
749 tls->sequential_records = true;
750 tls->next_seqno = be64dec(en->rec_seq);
751 STAILQ_INIT(&tls->pending_records);
752 } else {
753 tls->params.tls_hlen += AES_BLOCK_LEN;
754 }
755 tls->params.tls_tlen = AES_BLOCK_LEN +
756 SHA1_HASH_LEN;
757 break;
758 case CRYPTO_SHA2_256_HMAC:
759 tls->params.tls_hlen += AES_BLOCK_LEN;
760 tls->params.tls_tlen = AES_BLOCK_LEN +
761 SHA2_256_HASH_LEN;
762 break;
763 case CRYPTO_SHA2_384_HMAC:
764 tls->params.tls_hlen += AES_BLOCK_LEN;
765 tls->params.tls_tlen = AES_BLOCK_LEN +
766 SHA2_384_HASH_LEN;
767 break;
768 default:
769 panic("invalid hmac");
770 }
771 tls->params.tls_bs = AES_BLOCK_LEN;
772 break;
773 case CRYPTO_CHACHA20_POLY1305:
774 /*
775 * Chacha20 uses a 12 byte implicit IV.
776 */
777 tls->params.tls_tlen = POLY1305_HASH_LEN;
778 tls->params.tls_bs = 1;
779 break;
780 default:
781 panic("invalid cipher");
782 }
783
784 /*
785 * TLS 1.3 includes optional padding which we do not support,
786 * and also puts the "real" record type at the end of the
787 * encrypted data.
788 */
789 if (en->tls_vminor == TLS_MINOR_VER_THREE)
790 tls->params.tls_tlen += sizeof(uint8_t);
791
792 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
793 ("TLS header length too long: %d", tls->params.tls_hlen));
794 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
795 ("TLS trailer length too long: %d", tls->params.tls_tlen));
796
797 if (en->auth_key_len != 0) {
798 tls->params.auth_key_len = en->auth_key_len;
799 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
800 M_WAITOK);
801 bcopy(en->auth_key, tls->params.auth_key, en->auth_key_len);
802 }
803
804 tls->params.cipher_key_len = en->cipher_key_len;
805 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
806 bcopy(en->cipher_key, tls->params.cipher_key, en->cipher_key_len);
807
808 /*
809 * This holds the implicit portion of the nonce for AEAD
810 * ciphers and the initial implicit IV for TLS 1.0. The
811 * explicit portions of the IV are generated in ktls_frame().
812 */
813 if (en->iv_len != 0) {
814 tls->params.iv_len = en->iv_len;
815 bcopy(en->iv, tls->params.iv, en->iv_len);
816
817 /*
818 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
819 * counter to generate unique explicit IVs.
820 *
821 * Store this counter in the last 8 bytes of the IV
822 * array so that it is 8-byte aligned.
823 */
824 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
825 en->tls_vminor == TLS_MINOR_VER_TWO)
826 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
827 }
828
829 tls->gen = 0;
830 *tlsp = tls;
831 return (0);
832 }
833
834 static struct ktls_session *
ktls_clone_session(struct ktls_session * tls,int direction)835 ktls_clone_session(struct ktls_session *tls, int direction)
836 {
837 struct ktls_session *tls_new;
838
839 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
840
841 counter_u64_add(ktls_offload_active, 1);
842
843 refcount_init(&tls_new->refcount, 1);
844 if (direction == KTLS_RX) {
845 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
846 tls_new);
847 } else {
848 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
849 tls_new);
850 tls_new->inp = tls->inp;
851 tls_new->tx = true;
852 in_pcbref(tls_new->inp);
853 }
854
855 /* Copy fields from existing session. */
856 tls_new->params = tls->params;
857 tls_new->wq_index = tls->wq_index;
858
859 /* Deep copy keys. */
860 if (tls_new->params.auth_key != NULL) {
861 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
862 M_KTLS, M_WAITOK);
863 memcpy(tls_new->params.auth_key, tls->params.auth_key,
864 tls->params.auth_key_len);
865 }
866
867 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
868 M_WAITOK);
869 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
870 tls->params.cipher_key_len);
871
872 tls_new->gen = 0;
873 return (tls_new);
874 }
875
876 #ifdef TCP_OFFLOAD
877 static int
ktls_try_toe(struct socket * so,struct ktls_session * tls,int direction)878 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
879 {
880 struct inpcb *inp = sotoinpcb(so);
881 struct tcpcb *tp = intotcpcb(inp);
882 int error;
883
884 INP_WLOCK(inp);
885 if (tp->t_flags & TF_DISCONNECTED) {
886 INP_WUNLOCK(inp);
887 return (ECONNRESET);
888 }
889 if (!(tp->t_flags & TF_TOE)) {
890 INP_WUNLOCK(inp);
891 return (EOPNOTSUPP);
892 }
893
894 error = tcp_offload_alloc_tls_session(tp, tls, direction);
895 INP_WUNLOCK(inp);
896 if (error == 0) {
897 tls->mode = TCP_TLS_MODE_TOE;
898 switch (tls->params.cipher_algorithm) {
899 case CRYPTO_AES_CBC:
900 counter_u64_add(ktls_toe_cbc, 1);
901 break;
902 case CRYPTO_AES_NIST_GCM_16:
903 counter_u64_add(ktls_toe_gcm, 1);
904 break;
905 case CRYPTO_CHACHA20_POLY1305:
906 counter_u64_add(ktls_toe_chacha20, 1);
907 break;
908 }
909 }
910 return (error);
911 }
912 #endif
913
914 /*
915 * Common code used when first enabling ifnet TLS on a connection or
916 * when allocating a new ifnet TLS session due to a routing change.
917 * This function allocates a new TLS send tag on whatever interface
918 * the connection is currently routed over.
919 */
920 static int
ktls_alloc_snd_tag(struct inpcb * inp,struct ktls_session * tls,bool force,struct m_snd_tag ** mstp)921 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
922 struct m_snd_tag **mstp)
923 {
924 union if_snd_tag_alloc_params params;
925 struct ifnet *ifp;
926 struct nhop_object *nh;
927 struct tcpcb *tp = intotcpcb(inp);
928 int error;
929
930 INP_RLOCK(inp);
931 if (tp->t_flags & TF_DISCONNECTED) {
932 INP_RUNLOCK(inp);
933 return (ECONNRESET);
934 }
935
936 /*
937 * Check administrative controls on ifnet TLS to determine if
938 * ifnet TLS should be denied.
939 *
940 * - Always permit 'force' requests.
941 * - ktls_ifnet_permitted == 0: always deny.
942 */
943 if (!force && ktls_ifnet_permitted == 0) {
944 INP_RUNLOCK(inp);
945 return (ENXIO);
946 }
947
948 /*
949 * XXX: Use the cached route in the inpcb to find the
950 * interface. This should perhaps instead use
951 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
952 * enabled after a connection has completed key negotiation in
953 * userland, the cached route will be present in practice.
954 */
955 nh = inp->inp_route.ro_nh;
956 if (nh == NULL) {
957 INP_RUNLOCK(inp);
958 return (ENXIO);
959 }
960 ifp = nh->nh_ifp;
961 if_ref(ifp);
962
963 /*
964 * Allocate a TLS + ratelimit tag if the connection has an
965 * existing pacing rate.
966 */
967 if (tp->t_pacing_rate != -1 &&
968 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
969 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
970 params.tls_rate_limit.inp = inp;
971 params.tls_rate_limit.tls = tls;
972 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
973 } else {
974 params.hdr.type = IF_SND_TAG_TYPE_TLS;
975 params.tls.inp = inp;
976 params.tls.tls = tls;
977 }
978 params.hdr.flowid = inp->inp_flowid;
979 params.hdr.flowtype = inp->inp_flowtype;
980 params.hdr.numa_domain = inp->inp_numa_domain;
981 INP_RUNLOCK(inp);
982
983 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
984 error = EOPNOTSUPP;
985 goto out;
986 }
987 if (inp->inp_vflag & INP_IPV6) {
988 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
989 error = EOPNOTSUPP;
990 goto out;
991 }
992 } else {
993 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
994 error = EOPNOTSUPP;
995 goto out;
996 }
997 }
998 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
999 out:
1000 if_rele(ifp);
1001 return (error);
1002 }
1003
1004 /*
1005 * Allocate an initial TLS receive tag for doing HW decryption of TLS
1006 * data.
1007 *
1008 * This function allocates a new TLS receive tag on whatever interface
1009 * the connection is currently routed over. If the connection ends up
1010 * using a different interface for receive this will get fixed up via
1011 * ktls_input_ifp_mismatch as future packets arrive.
1012 */
1013 static int
ktls_alloc_rcv_tag(struct inpcb * inp,struct ktls_session * tls,struct m_snd_tag ** mstp)1014 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1015 struct m_snd_tag **mstp)
1016 {
1017 union if_snd_tag_alloc_params params;
1018 struct ifnet *ifp;
1019 struct nhop_object *nh;
1020 int error;
1021
1022 if (!ktls_ocf_recrypt_supported(tls))
1023 return (ENXIO);
1024
1025 INP_RLOCK(inp);
1026 if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
1027 INP_RUNLOCK(inp);
1028 return (ECONNRESET);
1029 }
1030
1031 /*
1032 * Check administrative controls on ifnet TLS to determine if
1033 * ifnet TLS should be denied.
1034 */
1035 if (ktls_ifnet_permitted == 0) {
1036 INP_RUNLOCK(inp);
1037 return (ENXIO);
1038 }
1039
1040 /*
1041 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1042 * the inpcb to find the interface.
1043 */
1044 nh = inp->inp_route.ro_nh;
1045 if (nh == NULL) {
1046 INP_RUNLOCK(inp);
1047 return (ENXIO);
1048 }
1049 ifp = nh->nh_ifp;
1050 if_ref(ifp);
1051 tls->rx_ifp = ifp;
1052
1053 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1054 params.hdr.flowid = inp->inp_flowid;
1055 params.hdr.flowtype = inp->inp_flowtype;
1056 params.hdr.numa_domain = inp->inp_numa_domain;
1057 params.tls_rx.inp = inp;
1058 params.tls_rx.tls = tls;
1059 params.tls_rx.vlan_id = 0;
1060
1061 INP_RUNLOCK(inp);
1062
1063 if (inp->inp_vflag & INP_IPV6) {
1064 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
1065 error = EOPNOTSUPP;
1066 goto out;
1067 }
1068 } else {
1069 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
1070 error = EOPNOTSUPP;
1071 goto out;
1072 }
1073 }
1074 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1075
1076 /*
1077 * If this connection is over a vlan, vlan_snd_tag_alloc
1078 * rewrites vlan_id with the saved interface. Save the VLAN
1079 * ID for use in ktls_reset_receive_tag which allocates new
1080 * receive tags directly from the leaf interface bypassing
1081 * if_vlan.
1082 */
1083 if (error == 0)
1084 tls->rx_vlan_id = params.tls_rx.vlan_id;
1085 out:
1086 return (error);
1087 }
1088
1089 static int
ktls_try_ifnet(struct socket * so,struct ktls_session * tls,int direction,bool force)1090 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1091 bool force)
1092 {
1093 struct m_snd_tag *mst;
1094 int error;
1095
1096 switch (direction) {
1097 case KTLS_TX:
1098 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1099 if (__predict_false(error != 0))
1100 goto done;
1101 break;
1102 case KTLS_RX:
1103 KASSERT(!force, ("%s: forced receive tag", __func__));
1104 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1105 if (__predict_false(error != 0))
1106 goto done;
1107 break;
1108 default:
1109 __assert_unreachable();
1110 }
1111
1112 tls->mode = TCP_TLS_MODE_IFNET;
1113 tls->snd_tag = mst;
1114
1115 switch (tls->params.cipher_algorithm) {
1116 case CRYPTO_AES_CBC:
1117 counter_u64_add(ktls_ifnet_cbc, 1);
1118 break;
1119 case CRYPTO_AES_NIST_GCM_16:
1120 counter_u64_add(ktls_ifnet_gcm, 1);
1121 break;
1122 case CRYPTO_CHACHA20_POLY1305:
1123 counter_u64_add(ktls_ifnet_chacha20, 1);
1124 break;
1125 default:
1126 break;
1127 }
1128 done:
1129 return (error);
1130 }
1131
1132 static void
ktls_use_sw(struct ktls_session * tls)1133 ktls_use_sw(struct ktls_session *tls)
1134 {
1135 tls->mode = TCP_TLS_MODE_SW;
1136 switch (tls->params.cipher_algorithm) {
1137 case CRYPTO_AES_CBC:
1138 counter_u64_add(ktls_sw_cbc, 1);
1139 break;
1140 case CRYPTO_AES_NIST_GCM_16:
1141 counter_u64_add(ktls_sw_gcm, 1);
1142 break;
1143 case CRYPTO_CHACHA20_POLY1305:
1144 counter_u64_add(ktls_sw_chacha20, 1);
1145 break;
1146 }
1147 }
1148
1149 static int
ktls_try_sw(struct ktls_session * tls,int direction)1150 ktls_try_sw(struct ktls_session *tls, int direction)
1151 {
1152 int error;
1153
1154 error = ktls_ocf_try(tls, direction);
1155 if (error)
1156 return (error);
1157 ktls_use_sw(tls);
1158 return (0);
1159 }
1160
1161 /*
1162 * KTLS RX stores data in the socket buffer as a list of TLS records,
1163 * where each record is stored as a control message containg the TLS
1164 * header followed by data mbufs containing the decrypted data. This
1165 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1166 * both encrypted and decrypted data. TLS records decrypted by a NIC
1167 * should be queued to the socket buffer as records, but encrypted
1168 * data which needs to be decrypted by software arrives as a stream of
1169 * regular mbufs which need to be converted. In addition, there may
1170 * already be pending encrypted data in the socket buffer when KTLS RX
1171 * is enabled.
1172 *
1173 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1174 * is used:
1175 *
1176 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1177 *
1178 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1179 * from the first mbuf. Once all of the data for that TLS record is
1180 * queued, the socket is queued to a worker thread.
1181 *
1182 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1183 * the TLS chain. Each TLS record is detached from the TLS chain,
1184 * decrypted, and inserted into the regular socket buffer chain as
1185 * record starting with a control message holding the TLS header and
1186 * a chain of mbufs holding the encrypted data.
1187 */
1188
1189 static void
sb_mark_notready(struct sockbuf * sb)1190 sb_mark_notready(struct sockbuf *sb)
1191 {
1192 struct mbuf *m;
1193
1194 m = sb->sb_mb;
1195 sb->sb_mtls = m;
1196 sb->sb_mb = NULL;
1197 sb->sb_mbtail = NULL;
1198 sb->sb_lastrecord = NULL;
1199 for (; m != NULL; m = m->m_next) {
1200 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1201 __func__));
1202 KASSERT((m->m_flags & M_NOTREADY) == 0, ("%s: mbuf not ready",
1203 __func__));
1204 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1205 __func__));
1206 m->m_flags |= M_NOTREADY;
1207 sb->sb_acc -= m->m_len;
1208 sb->sb_tlscc += m->m_len;
1209 sb->sb_mtlstail = m;
1210 }
1211 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1212 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1213 sb->sb_ccc));
1214 }
1215
1216 /*
1217 * Return information about the pending TLS data in a socket
1218 * buffer. On return, 'seqno' is set to the sequence number
1219 * of the next TLS record to be received, 'resid' is set to
1220 * the amount of bytes still needed for the last pending
1221 * record. The function returns 'false' if the last pending
1222 * record contains a partial TLS header. In that case, 'resid'
1223 * is the number of bytes needed to complete the TLS header.
1224 */
1225 bool
ktls_pending_rx_info(struct sockbuf * sb,uint64_t * seqnop,size_t * residp)1226 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1227 {
1228 struct tls_record_layer hdr;
1229 struct mbuf *m;
1230 uint64_t seqno;
1231 size_t resid;
1232 u_int offset, record_len;
1233
1234 SOCKBUF_LOCK_ASSERT(sb);
1235 MPASS(sb->sb_flags & SB_TLS_RX);
1236 seqno = sb->sb_tls_seqno;
1237 resid = sb->sb_tlscc;
1238 m = sb->sb_mtls;
1239 offset = 0;
1240
1241 if (resid == 0) {
1242 *seqnop = seqno;
1243 *residp = 0;
1244 return (true);
1245 }
1246
1247 for (;;) {
1248 seqno++;
1249
1250 if (resid < sizeof(hdr)) {
1251 *seqnop = seqno;
1252 *residp = sizeof(hdr) - resid;
1253 return (false);
1254 }
1255
1256 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1257
1258 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1259 if (resid <= record_len) {
1260 *seqnop = seqno;
1261 *residp = record_len - resid;
1262 return (true);
1263 }
1264 resid -= record_len;
1265
1266 while (record_len != 0) {
1267 if (m->m_len - offset > record_len) {
1268 offset += record_len;
1269 break;
1270 }
1271
1272 record_len -= (m->m_len - offset);
1273 offset = 0;
1274 m = m->m_next;
1275 }
1276 }
1277 }
1278
1279 int
ktls_enable_rx(struct socket * so,struct tls_enable * en)1280 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1281 {
1282 struct ktls_session *tls;
1283 int error;
1284
1285 if (!ktls_offload_enable)
1286 return (ENOTSUP);
1287
1288 counter_u64_add(ktls_offload_enable_calls, 1);
1289
1290 /*
1291 * This should always be true since only the TCP socket option
1292 * invokes this function.
1293 */
1294 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1295 return (EINVAL);
1296
1297 /*
1298 * XXX: Don't overwrite existing sessions. We should permit
1299 * this to support rekeying in the future.
1300 */
1301 if (so->so_rcv.sb_tls_info != NULL)
1302 return (EALREADY);
1303
1304 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1305 return (ENOTSUP);
1306
1307 error = ktls_create_session(so, en, &tls, KTLS_RX);
1308 if (error)
1309 return (error);
1310
1311 error = ktls_ocf_try(tls, KTLS_RX);
1312 if (error) {
1313 ktls_free(tls);
1314 return (error);
1315 }
1316
1317 /*
1318 * Serialize with soreceive_generic() and make sure that we're not
1319 * operating on a listening socket.
1320 */
1321 error = SOCK_IO_RECV_LOCK(so, SBL_WAIT);
1322 if (error) {
1323 ktls_free(tls);
1324 return (error);
1325 }
1326
1327 /* Mark the socket as using TLS offload. */
1328 SOCK_RECVBUF_LOCK(so);
1329 if (__predict_false(so->so_rcv.sb_tls_info != NULL))
1330 error = EALREADY;
1331 else if ((so->so_rcv.sb_flags & SB_SPLICED) != 0)
1332 error = EINVAL;
1333 if (error != 0) {
1334 SOCK_RECVBUF_UNLOCK(so);
1335 SOCK_IO_RECV_UNLOCK(so);
1336 ktls_free(tls);
1337 return (EALREADY);
1338 }
1339 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1340 so->so_rcv.sb_tls_info = tls;
1341 so->so_rcv.sb_flags |= SB_TLS_RX;
1342
1343 /* Mark existing data as not ready until it can be decrypted. */
1344 sb_mark_notready(&so->so_rcv);
1345 ktls_check_rx(&so->so_rcv);
1346 SOCK_RECVBUF_UNLOCK(so);
1347 SOCK_IO_RECV_UNLOCK(so);
1348
1349 /* Prefer TOE -> ifnet TLS -> software TLS. */
1350 #ifdef TCP_OFFLOAD
1351 error = ktls_try_toe(so, tls, KTLS_RX);
1352 if (error)
1353 #endif
1354 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1355 if (error)
1356 ktls_use_sw(tls);
1357
1358 counter_u64_add(ktls_offload_total, 1);
1359
1360 return (0);
1361 }
1362
1363 int
ktls_enable_tx(struct socket * so,struct tls_enable * en)1364 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1365 {
1366 struct ktls_session *tls;
1367 struct inpcb *inp;
1368 struct tcpcb *tp;
1369 int error;
1370
1371 if (!ktls_offload_enable)
1372 return (ENOTSUP);
1373
1374 counter_u64_add(ktls_offload_enable_calls, 1);
1375
1376 /*
1377 * This should always be true since only the TCP socket option
1378 * invokes this function.
1379 */
1380 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1381 return (EINVAL);
1382
1383 /*
1384 * XXX: Don't overwrite existing sessions. We should permit
1385 * this to support rekeying in the future.
1386 */
1387 if (so->so_snd.sb_tls_info != NULL)
1388 return (EALREADY);
1389
1390 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1391 return (ENOTSUP);
1392
1393 /* TLS requires ext pgs */
1394 if (mb_use_ext_pgs == 0)
1395 return (ENXIO);
1396
1397 error = ktls_create_session(so, en, &tls, KTLS_TX);
1398 if (error)
1399 return (error);
1400
1401 /* some ktls offload NICs require initial seqno to start offload */
1402 tls->initial_offload_seqno = be64dec(en->rec_seq);
1403
1404 /* Prefer TOE -> ifnet TLS -> software TLS. */
1405 #ifdef TCP_OFFLOAD
1406 error = ktls_try_toe(so, tls, KTLS_TX);
1407 if (error)
1408 #endif
1409 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1410 if (error)
1411 error = ktls_try_sw(tls, KTLS_TX);
1412
1413 if (error) {
1414 ktls_free(tls);
1415 return (error);
1416 }
1417
1418 /*
1419 * Serialize with sosend_generic() and make sure that we're not
1420 * operating on a listening socket.
1421 */
1422 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1423 if (error) {
1424 ktls_free(tls);
1425 return (error);
1426 }
1427
1428 /*
1429 * Write lock the INP when setting sb_tls_info so that
1430 * routines in tcp_ratelimit.c can read sb_tls_info while
1431 * holding the INP lock.
1432 */
1433 inp = so->so_pcb;
1434 INP_WLOCK(inp);
1435 SOCK_SENDBUF_LOCK(so);
1436 if (__predict_false(so->so_snd.sb_tls_info != NULL))
1437 error = EALREADY;
1438 else if ((so->so_snd.sb_flags & SB_SPLICED) != 0)
1439 error = EINVAL;
1440 if (error != 0) {
1441 SOCK_SENDBUF_UNLOCK(so);
1442 INP_WUNLOCK(inp);
1443 SOCK_IO_SEND_UNLOCK(so);
1444 ktls_free(tls);
1445 return (error);
1446 }
1447 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1448 so->so_snd.sb_tls_info = tls;
1449 if (tls->mode != TCP_TLS_MODE_SW) {
1450 tp = intotcpcb(inp);
1451 MPASS(tp->t_nic_ktls_xmit == 0);
1452 tp->t_nic_ktls_xmit = 1;
1453 if (tp->t_fb->tfb_hwtls_change != NULL)
1454 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1455 }
1456 SOCK_SENDBUF_UNLOCK(so);
1457 INP_WUNLOCK(inp);
1458 SOCK_IO_SEND_UNLOCK(so);
1459
1460 counter_u64_add(ktls_offload_total, 1);
1461
1462 return (0);
1463 }
1464
1465 int
ktls_get_rx_mode(struct socket * so,int * modep)1466 ktls_get_rx_mode(struct socket *so, int *modep)
1467 {
1468 struct ktls_session *tls;
1469 struct inpcb *inp __diagused;
1470
1471 if (SOLISTENING(so))
1472 return (EINVAL);
1473 inp = so->so_pcb;
1474 INP_WLOCK_ASSERT(inp);
1475 SOCK_RECVBUF_LOCK(so);
1476 tls = so->so_rcv.sb_tls_info;
1477 if (tls == NULL)
1478 *modep = TCP_TLS_MODE_NONE;
1479 else
1480 *modep = tls->mode;
1481 SOCK_RECVBUF_UNLOCK(so);
1482 return (0);
1483 }
1484
1485 /*
1486 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1487 *
1488 * This function gets information about the next TCP- and TLS-
1489 * sequence number to be processed by the TLS receive worker
1490 * thread. The information is extracted from the given "inpcb"
1491 * structure. The values are stored in host endian format at the two
1492 * given output pointer locations. The TCP sequence number points to
1493 * the beginning of the TLS header.
1494 *
1495 * This function returns zero on success, else a non-zero error code
1496 * is returned.
1497 */
1498 int
ktls_get_rx_sequence(struct inpcb * inp,uint32_t * tcpseq,uint64_t * tlsseq)1499 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1500 {
1501 struct socket *so = inp->inp_socket;
1502 struct tcpcb *tp = intotcpcb(inp);
1503
1504 INP_RLOCK(inp);
1505 if (tp->t_flags & TF_DISCONNECTED) {
1506 INP_RUNLOCK(inp);
1507 return (ECONNRESET);
1508 }
1509
1510 SOCKBUF_LOCK(&so->so_rcv);
1511 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1512 *tlsseq = so->so_rcv.sb_tls_seqno;
1513 SOCKBUF_UNLOCK(&so->so_rcv);
1514
1515 INP_RUNLOCK(inp);
1516
1517 return (0);
1518 }
1519
1520 int
ktls_get_tx_mode(struct socket * so,int * modep)1521 ktls_get_tx_mode(struct socket *so, int *modep)
1522 {
1523 struct ktls_session *tls;
1524 struct inpcb *inp __diagused;
1525
1526 if (SOLISTENING(so))
1527 return (EINVAL);
1528 inp = so->so_pcb;
1529 INP_WLOCK_ASSERT(inp);
1530 SOCK_SENDBUF_LOCK(so);
1531 tls = so->so_snd.sb_tls_info;
1532 if (tls == NULL)
1533 *modep = TCP_TLS_MODE_NONE;
1534 else
1535 *modep = tls->mode;
1536 SOCK_SENDBUF_UNLOCK(so);
1537 return (0);
1538 }
1539
1540 /*
1541 * Switch between SW and ifnet TLS sessions as requested.
1542 */
1543 int
ktls_set_tx_mode(struct socket * so,int mode)1544 ktls_set_tx_mode(struct socket *so, int mode)
1545 {
1546 struct ktls_session *tls, *tls_new;
1547 struct inpcb *inp;
1548 struct tcpcb *tp;
1549 int error;
1550
1551 if (SOLISTENING(so))
1552 return (EINVAL);
1553 switch (mode) {
1554 case TCP_TLS_MODE_SW:
1555 case TCP_TLS_MODE_IFNET:
1556 break;
1557 default:
1558 return (EINVAL);
1559 }
1560
1561 inp = so->so_pcb;
1562 INP_WLOCK_ASSERT(inp);
1563 tp = intotcpcb(inp);
1564
1565 if (mode == TCP_TLS_MODE_IFNET) {
1566 /* Don't allow enabling ifnet ktls multiple times */
1567 if (tp->t_nic_ktls_xmit)
1568 return (EALREADY);
1569
1570 /*
1571 * Don't enable ifnet ktls if we disabled it due to an
1572 * excessive retransmission rate
1573 */
1574 if (tp->t_nic_ktls_xmit_dis)
1575 return (ENXIO);
1576 }
1577
1578 SOCKBUF_LOCK(&so->so_snd);
1579 tls = so->so_snd.sb_tls_info;
1580 if (tls == NULL) {
1581 SOCKBUF_UNLOCK(&so->so_snd);
1582 return (0);
1583 }
1584
1585 if (tls->mode == mode) {
1586 SOCKBUF_UNLOCK(&so->so_snd);
1587 return (0);
1588 }
1589
1590 tls = ktls_hold(tls);
1591 SOCKBUF_UNLOCK(&so->so_snd);
1592 INP_WUNLOCK(inp);
1593
1594 tls_new = ktls_clone_session(tls, KTLS_TX);
1595
1596 if (mode == TCP_TLS_MODE_IFNET)
1597 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1598 else
1599 error = ktls_try_sw(tls_new, KTLS_TX);
1600 if (error) {
1601 counter_u64_add(ktls_switch_failed, 1);
1602 ktls_free(tls_new);
1603 ktls_free(tls);
1604 INP_WLOCK(inp);
1605 return (error);
1606 }
1607
1608 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1609 if (error) {
1610 counter_u64_add(ktls_switch_failed, 1);
1611 ktls_free(tls_new);
1612 ktls_free(tls);
1613 INP_WLOCK(inp);
1614 return (error);
1615 }
1616
1617 /*
1618 * If we raced with another session change, keep the existing
1619 * session.
1620 */
1621 if (tls != so->so_snd.sb_tls_info) {
1622 counter_u64_add(ktls_switch_failed, 1);
1623 SOCK_IO_SEND_UNLOCK(so);
1624 ktls_free(tls_new);
1625 ktls_free(tls);
1626 INP_WLOCK(inp);
1627 return (EBUSY);
1628 }
1629
1630 INP_WLOCK(inp);
1631 SOCKBUF_LOCK(&so->so_snd);
1632 so->so_snd.sb_tls_info = tls_new;
1633 if (tls_new->mode != TCP_TLS_MODE_SW) {
1634 MPASS(tp->t_nic_ktls_xmit == 0);
1635 tp->t_nic_ktls_xmit = 1;
1636 if (tp->t_fb->tfb_hwtls_change != NULL)
1637 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1638 }
1639 SOCKBUF_UNLOCK(&so->so_snd);
1640 SOCK_IO_SEND_UNLOCK(so);
1641
1642 /*
1643 * Drop two references on 'tls'. The first is for the
1644 * ktls_hold() above. The second drops the reference from the
1645 * socket buffer.
1646 */
1647 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1648 ktls_free(tls);
1649 ktls_free(tls);
1650
1651 if (mode == TCP_TLS_MODE_IFNET)
1652 counter_u64_add(ktls_switch_to_ifnet, 1);
1653 else
1654 counter_u64_add(ktls_switch_to_sw, 1);
1655
1656 return (0);
1657 }
1658
1659 /*
1660 * Try to allocate a new TLS receive tag. This task is scheduled when
1661 * sbappend_ktls_rx detects an input path change. If a new tag is
1662 * allocated, replace the tag in the TLS session. If a new tag cannot
1663 * be allocated, let the session fall back to software decryption.
1664 */
1665 static void
ktls_reset_receive_tag(void * context,int pending)1666 ktls_reset_receive_tag(void *context, int pending)
1667 {
1668 union if_snd_tag_alloc_params params;
1669 struct ktls_session *tls;
1670 struct m_snd_tag *mst;
1671 struct inpcb *inp;
1672 struct ifnet *ifp;
1673 struct socket *so;
1674 int error;
1675
1676 MPASS(pending == 1);
1677
1678 tls = context;
1679 so = tls->so;
1680 inp = so->so_pcb;
1681 ifp = NULL;
1682
1683 INP_RLOCK(inp);
1684 if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
1685 INP_RUNLOCK(inp);
1686 goto out;
1687 }
1688
1689 SOCKBUF_LOCK(&so->so_rcv);
1690 mst = tls->snd_tag;
1691 tls->snd_tag = NULL;
1692 if (mst != NULL)
1693 m_snd_tag_rele(mst);
1694
1695 ifp = tls->rx_ifp;
1696 if_ref(ifp);
1697 SOCKBUF_UNLOCK(&so->so_rcv);
1698
1699 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1700 params.hdr.flowid = inp->inp_flowid;
1701 params.hdr.flowtype = inp->inp_flowtype;
1702 params.hdr.numa_domain = inp->inp_numa_domain;
1703 params.tls_rx.inp = inp;
1704 params.tls_rx.tls = tls;
1705 params.tls_rx.vlan_id = tls->rx_vlan_id;
1706 INP_RUNLOCK(inp);
1707
1708 if (inp->inp_vflag & INP_IPV6) {
1709 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1710 goto out;
1711 } else {
1712 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1713 goto out;
1714 }
1715
1716 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1717 if (error == 0) {
1718 SOCKBUF_LOCK(&so->so_rcv);
1719 tls->snd_tag = mst;
1720 SOCKBUF_UNLOCK(&so->so_rcv);
1721
1722 counter_u64_add(ktls_ifnet_reset, 1);
1723 } else {
1724 /*
1725 * Just fall back to software decryption if a tag
1726 * cannot be allocated leaving the connection intact.
1727 * If a future input path change switches to another
1728 * interface this connection will resume ifnet TLS.
1729 */
1730 counter_u64_add(ktls_ifnet_reset_failed, 1);
1731 }
1732
1733 out:
1734 mtx_pool_lock(mtxpool_sleep, tls);
1735 tls->reset_pending = false;
1736 mtx_pool_unlock(mtxpool_sleep, tls);
1737
1738 if (ifp != NULL)
1739 if_rele(ifp);
1740 CURVNET_SET(so->so_vnet);
1741 sorele(so);
1742 CURVNET_RESTORE();
1743 ktls_free(tls);
1744 }
1745
1746 /*
1747 * Try to allocate a new TLS send tag. This task is scheduled when
1748 * ip_output detects a route change while trying to transmit a packet
1749 * holding a TLS record. If a new tag is allocated, replace the tag
1750 * in the TLS session. Subsequent packets on the connection will use
1751 * the new tag. If a new tag cannot be allocated, drop the
1752 * connection.
1753 */
1754 static void
ktls_reset_send_tag(void * context,int pending)1755 ktls_reset_send_tag(void *context, int pending)
1756 {
1757 struct epoch_tracker et;
1758 struct ktls_session *tls;
1759 struct m_snd_tag *old, *new;
1760 struct inpcb *inp;
1761 struct tcpcb *tp;
1762 int error;
1763
1764 MPASS(pending == 1);
1765
1766 tls = context;
1767 inp = tls->inp;
1768
1769 /*
1770 * Free the old tag first before allocating a new one.
1771 * ip[6]_output_send() will treat a NULL send tag the same as
1772 * an ifp mismatch and drop packets until a new tag is
1773 * allocated.
1774 *
1775 * Write-lock the INP when changing tls->snd_tag since
1776 * ip[6]_output_send() holds a read-lock when reading the
1777 * pointer.
1778 */
1779 INP_WLOCK(inp);
1780 old = tls->snd_tag;
1781 tls->snd_tag = NULL;
1782 INP_WUNLOCK(inp);
1783 if (old != NULL)
1784 m_snd_tag_rele(old);
1785
1786 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1787
1788 if (error == 0) {
1789 INP_WLOCK(inp);
1790 tls->snd_tag = new;
1791 mtx_pool_lock(mtxpool_sleep, tls);
1792 tls->reset_pending = false;
1793 mtx_pool_unlock(mtxpool_sleep, tls);
1794 INP_WUNLOCK(inp);
1795
1796 counter_u64_add(ktls_ifnet_reset, 1);
1797
1798 /*
1799 * XXX: Should we kick tcp_output explicitly now that
1800 * the send tag is fixed or just rely on timers?
1801 */
1802 } else {
1803 NET_EPOCH_ENTER(et);
1804 INP_WLOCK(inp);
1805 tp = intotcpcb(inp);
1806 if (!(tp->t_flags & TF_DISCONNECTED)) {
1807 CURVNET_SET(inp->inp_socket->so_vnet);
1808 tp = tcp_drop(tp, ECONNABORTED);
1809 CURVNET_RESTORE();
1810 if (tp != NULL) {
1811 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1812 INP_WUNLOCK(inp);
1813 }
1814 } else
1815 INP_WUNLOCK(inp);
1816 NET_EPOCH_EXIT(et);
1817
1818 counter_u64_add(ktls_ifnet_reset_failed, 1);
1819
1820 /*
1821 * Leave reset_pending true to avoid future tasks while
1822 * the socket goes away.
1823 */
1824 }
1825
1826 ktls_free(tls);
1827 }
1828
1829 void
ktls_input_ifp_mismatch(struct sockbuf * sb,struct ifnet * ifp)1830 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1831 {
1832 struct ktls_session *tls;
1833 struct socket *so;
1834
1835 SOCKBUF_LOCK_ASSERT(sb);
1836 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1837 __func__, sb));
1838 so = __containerof(sb, struct socket, so_rcv);
1839
1840 tls = sb->sb_tls_info;
1841 if_rele(tls->rx_ifp);
1842 if_ref(ifp);
1843 tls->rx_ifp = ifp;
1844
1845 /*
1846 * See if we should schedule a task to update the receive tag for
1847 * this session.
1848 */
1849 mtx_pool_lock(mtxpool_sleep, tls);
1850 if (!tls->reset_pending) {
1851 (void) ktls_hold(tls);
1852 soref(so);
1853 tls->so = so;
1854 tls->reset_pending = true;
1855 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1856 }
1857 mtx_pool_unlock(mtxpool_sleep, tls);
1858 }
1859
1860 int
ktls_output_eagain(struct inpcb * inp,struct ktls_session * tls)1861 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1862 {
1863
1864 if (inp == NULL)
1865 return (ENOBUFS);
1866
1867 INP_LOCK_ASSERT(inp);
1868
1869 /*
1870 * See if we should schedule a task to update the send tag for
1871 * this session.
1872 */
1873 mtx_pool_lock(mtxpool_sleep, tls);
1874 if (!tls->reset_pending) {
1875 (void) ktls_hold(tls);
1876 tls->reset_pending = true;
1877 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1878 }
1879 mtx_pool_unlock(mtxpool_sleep, tls);
1880 return (ENOBUFS);
1881 }
1882
1883 #ifdef RATELIMIT
1884 int
ktls_modify_txrtlmt(struct ktls_session * tls,uint64_t max_pacing_rate)1885 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1886 {
1887 union if_snd_tag_modify_params params = {
1888 .rate_limit.max_rate = max_pacing_rate,
1889 .rate_limit.flags = M_NOWAIT,
1890 };
1891 struct m_snd_tag *mst;
1892
1893 /* Can't get to the inp, but it should be locked. */
1894 /* INP_LOCK_ASSERT(inp); */
1895
1896 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1897
1898 if (tls->snd_tag == NULL) {
1899 /*
1900 * Resetting send tag, ignore this change. The
1901 * pending reset may or may not see this updated rate
1902 * in the tcpcb. If it doesn't, we will just lose
1903 * this rate change.
1904 */
1905 return (0);
1906 }
1907
1908 mst = tls->snd_tag;
1909
1910 MPASS(mst != NULL);
1911 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1912
1913 return (mst->sw->snd_tag_modify(mst, ¶ms));
1914 }
1915 #endif
1916
1917 static void
ktls_destroy_help(void * context,int pending __unused)1918 ktls_destroy_help(void *context, int pending __unused)
1919 {
1920 ktls_destroy(context);
1921 }
1922
1923 void
ktls_destroy(struct ktls_session * tls)1924 ktls_destroy(struct ktls_session *tls)
1925 {
1926 struct inpcb *inp;
1927 struct tcpcb *tp;
1928 bool wlocked;
1929
1930 MPASS(tls->refcount == 0);
1931
1932 inp = tls->inp;
1933 if (tls->tx) {
1934 wlocked = INP_WLOCKED(inp);
1935 if (!wlocked && !INP_TRY_WLOCK(inp)) {
1936 /*
1937 * rwlocks read locks are anonymous, and there
1938 * is no way to know if our current thread
1939 * holds an rlock on the inp. As a rough
1940 * estimate, check to see if the thread holds
1941 * *any* rlocks at all. If it does not, then we
1942 * know that we don't hold the inp rlock, and
1943 * can safely take the wlock
1944 */
1945 if (curthread->td_rw_rlocks == 0) {
1946 INP_WLOCK(inp);
1947 } else {
1948 /*
1949 * We might hold the rlock, so let's
1950 * do the destroy in a taskqueue
1951 * context to avoid a potential
1952 * deadlock. This should be very
1953 * rare.
1954 */
1955 counter_u64_add(ktls_destroy_task, 1);
1956 TASK_INIT(&tls->destroy_task, 0,
1957 ktls_destroy_help, tls);
1958 (void)taskqueue_enqueue(taskqueue_thread,
1959 &tls->destroy_task);
1960 return;
1961 }
1962 }
1963 }
1964
1965 if (tls->sequential_records) {
1966 struct mbuf *m, *n;
1967 int page_count;
1968
1969 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1970 page_count = m->m_epg_enc_cnt;
1971 while (page_count > 0) {
1972 KASSERT(page_count >= m->m_epg_nrdy,
1973 ("%s: too few pages", __func__));
1974 page_count -= m->m_epg_nrdy;
1975 m = m_free(m);
1976 }
1977 }
1978 }
1979
1980 counter_u64_add(ktls_offload_active, -1);
1981 switch (tls->mode) {
1982 case TCP_TLS_MODE_SW:
1983 switch (tls->params.cipher_algorithm) {
1984 case CRYPTO_AES_CBC:
1985 counter_u64_add(ktls_sw_cbc, -1);
1986 break;
1987 case CRYPTO_AES_NIST_GCM_16:
1988 counter_u64_add(ktls_sw_gcm, -1);
1989 break;
1990 case CRYPTO_CHACHA20_POLY1305:
1991 counter_u64_add(ktls_sw_chacha20, -1);
1992 break;
1993 }
1994 break;
1995 case TCP_TLS_MODE_IFNET:
1996 switch (tls->params.cipher_algorithm) {
1997 case CRYPTO_AES_CBC:
1998 counter_u64_add(ktls_ifnet_cbc, -1);
1999 break;
2000 case CRYPTO_AES_NIST_GCM_16:
2001 counter_u64_add(ktls_ifnet_gcm, -1);
2002 break;
2003 case CRYPTO_CHACHA20_POLY1305:
2004 counter_u64_add(ktls_ifnet_chacha20, -1);
2005 break;
2006 }
2007 if (tls->snd_tag != NULL)
2008 m_snd_tag_rele(tls->snd_tag);
2009 if (tls->rx_ifp != NULL)
2010 if_rele(tls->rx_ifp);
2011 if (tls->tx) {
2012 INP_WLOCK_ASSERT(inp);
2013 tp = intotcpcb(inp);
2014 MPASS(tp->t_nic_ktls_xmit == 1);
2015 tp->t_nic_ktls_xmit = 0;
2016 }
2017 break;
2018 #ifdef TCP_OFFLOAD
2019 case TCP_TLS_MODE_TOE:
2020 switch (tls->params.cipher_algorithm) {
2021 case CRYPTO_AES_CBC:
2022 counter_u64_add(ktls_toe_cbc, -1);
2023 break;
2024 case CRYPTO_AES_NIST_GCM_16:
2025 counter_u64_add(ktls_toe_gcm, -1);
2026 break;
2027 case CRYPTO_CHACHA20_POLY1305:
2028 counter_u64_add(ktls_toe_chacha20, -1);
2029 break;
2030 }
2031 break;
2032 #endif
2033 }
2034 if (tls->ocf_session != NULL)
2035 ktls_ocf_free(tls);
2036 if (tls->params.auth_key != NULL) {
2037 zfree(tls->params.auth_key, M_KTLS);
2038 tls->params.auth_key = NULL;
2039 tls->params.auth_key_len = 0;
2040 }
2041 if (tls->params.cipher_key != NULL) {
2042 zfree(tls->params.cipher_key, M_KTLS);
2043 tls->params.cipher_key = NULL;
2044 tls->params.cipher_key_len = 0;
2045 }
2046 if (tls->tx) {
2047 INP_WLOCK_ASSERT(inp);
2048 if (!in_pcbrele_wlocked(inp) && !wlocked)
2049 INP_WUNLOCK(inp);
2050 }
2051 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
2052
2053 uma_zfree(ktls_session_zone, tls);
2054 }
2055
2056 void
ktls_seq(struct sockbuf * sb,struct mbuf * m)2057 ktls_seq(struct sockbuf *sb, struct mbuf *m)
2058 {
2059
2060 for (; m != NULL; m = m->m_next) {
2061 KASSERT((m->m_flags & M_EXTPG) != 0,
2062 ("ktls_seq: mapped mbuf %p", m));
2063
2064 m->m_epg_seqno = sb->sb_tls_seqno;
2065 sb->sb_tls_seqno++;
2066 }
2067 }
2068
2069 /*
2070 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
2071 * mbuf in the chain must be an unmapped mbuf. The payload of the
2072 * mbuf must be populated with the payload of each TLS record.
2073 *
2074 * The record_type argument specifies the TLS record type used when
2075 * populating the TLS header.
2076 *
2077 * The enq_count argument on return is set to the number of pages of
2078 * payload data for this entire chain that need to be encrypted via SW
2079 * encryption. The returned value should be passed to ktls_enqueue
2080 * when scheduling encryption of this chain of mbufs. To handle the
2081 * special case of empty fragments for TLS 1.0 sessions, an empty
2082 * fragment counts as one page.
2083 */
2084 void
ktls_frame(struct mbuf * top,struct ktls_session * tls,int * enq_cnt,uint8_t record_type)2085 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2086 uint8_t record_type)
2087 {
2088 struct tls_record_layer *tlshdr;
2089 struct mbuf *m;
2090 uint64_t *noncep;
2091 uint16_t tls_len;
2092 int maxlen __diagused;
2093
2094 maxlen = tls->params.max_frame_len;
2095 *enq_cnt = 0;
2096 for (m = top; m != NULL; m = m->m_next) {
2097 /*
2098 * All mbufs in the chain should be TLS records whose
2099 * payload does not exceed the maximum frame length.
2100 *
2101 * Empty TLS 1.0 records are permitted when using CBC.
2102 */
2103 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2104 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2105 ("ktls_frame: m %p len %d", m, m->m_len));
2106
2107 /*
2108 * TLS frames require unmapped mbufs to store session
2109 * info.
2110 */
2111 KASSERT((m->m_flags & M_EXTPG) != 0,
2112 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2113
2114 tls_len = m->m_len;
2115
2116 /* Save a reference to the session. */
2117 m->m_epg_tls = ktls_hold(tls);
2118
2119 m->m_epg_hdrlen = tls->params.tls_hlen;
2120 m->m_epg_trllen = tls->params.tls_tlen;
2121 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2122 int bs, delta;
2123
2124 /*
2125 * AES-CBC pads messages to a multiple of the
2126 * block size. Note that the padding is
2127 * applied after the digest and the encryption
2128 * is done on the "plaintext || mac || padding".
2129 * At least one byte of padding is always
2130 * present.
2131 *
2132 * Compute the final trailer length assuming
2133 * at most one block of padding.
2134 * tls->params.tls_tlen is the maximum
2135 * possible trailer length (padding + digest).
2136 * delta holds the number of excess padding
2137 * bytes if the maximum were used. Those
2138 * extra bytes are removed.
2139 */
2140 bs = tls->params.tls_bs;
2141 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2142 m->m_epg_trllen -= delta;
2143 }
2144 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2145
2146 /* Populate the TLS header. */
2147 tlshdr = (void *)m->m_epg_hdr;
2148 tlshdr->tls_vmajor = tls->params.tls_vmajor;
2149
2150 /*
2151 * TLS 1.3 masquarades as TLS 1.2 with a record type
2152 * of TLS_RLTYPE_APP.
2153 */
2154 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2155 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2156 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2157 tlshdr->tls_type = TLS_RLTYPE_APP;
2158 /* save the real record type for later */
2159 m->m_epg_record_type = record_type;
2160 m->m_epg_trail[0] = record_type;
2161 } else {
2162 tlshdr->tls_vminor = tls->params.tls_vminor;
2163 tlshdr->tls_type = record_type;
2164 }
2165 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2166
2167 /*
2168 * Store nonces / explicit IVs after the end of the
2169 * TLS header.
2170 *
2171 * For GCM with TLS 1.2, an 8 byte nonce is copied
2172 * from the end of the IV. The nonce is then
2173 * incremented for use by the next record.
2174 *
2175 * For CBC, a random nonce is inserted for TLS 1.1+.
2176 */
2177 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2178 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2179 noncep = (uint64_t *)(tls->params.iv + 8);
2180 be64enc(tlshdr + 1, *noncep);
2181 (*noncep)++;
2182 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2183 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2184 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2185
2186 /*
2187 * When using SW encryption, mark the mbuf not ready.
2188 * It will be marked ready via sbready() after the
2189 * record has been encrypted.
2190 *
2191 * When using ifnet TLS, unencrypted TLS records are
2192 * sent down the stack to the NIC.
2193 */
2194 if (tls->mode == TCP_TLS_MODE_SW) {
2195 m->m_flags |= M_NOTREADY;
2196 if (__predict_false(tls_len == 0)) {
2197 /* TLS 1.0 empty fragment. */
2198 m->m_epg_nrdy = 1;
2199 } else
2200 m->m_epg_nrdy = m->m_epg_npgs;
2201 *enq_cnt += m->m_epg_nrdy;
2202 }
2203 }
2204 }
2205
2206 bool
ktls_permit_empty_frames(struct ktls_session * tls)2207 ktls_permit_empty_frames(struct ktls_session *tls)
2208 {
2209 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2210 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2211 }
2212
2213 void
ktls_check_rx(struct sockbuf * sb)2214 ktls_check_rx(struct sockbuf *sb)
2215 {
2216 struct tls_record_layer hdr;
2217 struct ktls_wq *wq;
2218 struct socket *so;
2219 bool running;
2220
2221 SOCKBUF_LOCK_ASSERT(sb);
2222 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2223 __func__, sb));
2224 so = __containerof(sb, struct socket, so_rcv);
2225
2226 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2227 return;
2228
2229 /* Is there enough queued for a TLS header? */
2230 if (sb->sb_tlscc < sizeof(hdr)) {
2231 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2232 so->so_error = EMSGSIZE;
2233 return;
2234 }
2235
2236 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2237
2238 /* Is the entire record queued? */
2239 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2240 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2241 so->so_error = EMSGSIZE;
2242 return;
2243 }
2244
2245 sb->sb_flags |= SB_TLS_RX_RUNNING;
2246
2247 soref(so);
2248 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2249 mtx_lock(&wq->mtx);
2250 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2251 running = wq->running;
2252 mtx_unlock(&wq->mtx);
2253 if (!running)
2254 wakeup(wq);
2255 counter_u64_add(ktls_cnt_rx_queued, 1);
2256 }
2257
2258 static struct mbuf *
ktls_detach_record(struct sockbuf * sb,int len)2259 ktls_detach_record(struct sockbuf *sb, int len)
2260 {
2261 struct mbuf *m, *n, *top;
2262 int remain;
2263
2264 SOCKBUF_LOCK_ASSERT(sb);
2265 MPASS(len <= sb->sb_tlscc);
2266
2267 /*
2268 * If TLS chain is the exact size of the record,
2269 * just grab the whole record.
2270 */
2271 top = sb->sb_mtls;
2272 if (sb->sb_tlscc == len) {
2273 sb->sb_mtls = NULL;
2274 sb->sb_mtlstail = NULL;
2275 goto out;
2276 }
2277
2278 /*
2279 * While it would be nice to use m_split() here, we need
2280 * to know exactly what m_split() allocates to update the
2281 * accounting, so do it inline instead.
2282 */
2283 remain = len;
2284 for (m = top; remain > m->m_len; m = m->m_next)
2285 remain -= m->m_len;
2286
2287 /* Easy case: don't have to split 'm'. */
2288 if (remain == m->m_len) {
2289 sb->sb_mtls = m->m_next;
2290 if (sb->sb_mtls == NULL)
2291 sb->sb_mtlstail = NULL;
2292 m->m_next = NULL;
2293 goto out;
2294 }
2295
2296 /*
2297 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2298 * with M_NOWAIT first.
2299 */
2300 n = m_get(M_NOWAIT, MT_DATA);
2301 if (n == NULL) {
2302 /*
2303 * Use M_WAITOK with socket buffer unlocked. If
2304 * 'sb_mtls' changes while the lock is dropped, return
2305 * NULL to force the caller to retry.
2306 */
2307 SOCKBUF_UNLOCK(sb);
2308
2309 n = m_get(M_WAITOK, MT_DATA);
2310
2311 SOCKBUF_LOCK(sb);
2312 if (sb->sb_mtls != top) {
2313 m_free(n);
2314 return (NULL);
2315 }
2316 }
2317 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2318
2319 /* Store remainder in 'n'. */
2320 n->m_len = m->m_len - remain;
2321 if (m->m_flags & M_EXT) {
2322 n->m_data = m->m_data + remain;
2323 mb_dupcl(n, m);
2324 } else {
2325 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2326 }
2327
2328 /* Trim 'm' and update accounting. */
2329 m->m_len -= n->m_len;
2330 sb->sb_tlscc -= n->m_len;
2331 sb->sb_ccc -= n->m_len;
2332
2333 /* Account for 'n'. */
2334 sballoc_ktls_rx(sb, n);
2335
2336 /* Insert 'n' into the TLS chain. */
2337 sb->sb_mtls = n;
2338 n->m_next = m->m_next;
2339 if (sb->sb_mtlstail == m)
2340 sb->sb_mtlstail = n;
2341
2342 /* Detach the record from the TLS chain. */
2343 m->m_next = NULL;
2344
2345 out:
2346 MPASS(m_length(top, NULL) == len);
2347 for (m = top; m != NULL; m = m->m_next)
2348 sbfree_ktls_rx(sb, m);
2349 sb->sb_tlsdcc = len;
2350 sb->sb_ccc += len;
2351 SBCHECK(sb);
2352 return (top);
2353 }
2354
2355 /*
2356 * Determine the length of the trailing zero padding and find the real
2357 * record type in the byte before the padding.
2358 *
2359 * Walking the mbuf chain backwards is clumsy, so another option would
2360 * be to scan forwards remembering the last non-zero byte before the
2361 * trailer. However, it would be expensive to scan the entire record.
2362 * Instead, find the last non-zero byte of each mbuf in the chain
2363 * keeping track of the relative offset of that nonzero byte.
2364 *
2365 * trail_len is the size of the MAC/tag on input and is set to the
2366 * size of the full trailer including padding and the record type on
2367 * return.
2368 */
2369 static int
tls13_find_record_type(struct ktls_session * tls,struct mbuf * m,int tls_len,int * trailer_len,uint8_t * record_typep)2370 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2371 int *trailer_len, uint8_t *record_typep)
2372 {
2373 char *cp;
2374 u_int digest_start, last_offset, m_len, offset;
2375 uint8_t record_type;
2376
2377 digest_start = tls_len - *trailer_len;
2378 last_offset = 0;
2379 offset = 0;
2380 for (; m != NULL && offset < digest_start;
2381 offset += m->m_len, m = m->m_next) {
2382 /* Don't look for padding in the tag. */
2383 m_len = min(digest_start - offset, m->m_len);
2384 cp = mtod(m, char *);
2385
2386 /* Find last non-zero byte in this mbuf. */
2387 while (m_len > 0 && cp[m_len - 1] == 0)
2388 m_len--;
2389 if (m_len > 0) {
2390 record_type = cp[m_len - 1];
2391 last_offset = offset + m_len;
2392 }
2393 }
2394 if (last_offset < tls->params.tls_hlen)
2395 return (EBADMSG);
2396
2397 *record_typep = record_type;
2398 *trailer_len = tls_len - last_offset + 1;
2399 return (0);
2400 }
2401
2402 /*
2403 * Check if a mbuf chain is fully decrypted. Returns
2404 * KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is decrypted.
2405 * KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted and
2406 * decrypted data. KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data is
2407 * encrypted. KTLS_MBUF_CRYPTO_ST_SHAREDMBUF if any mbuf points at
2408 * shared data that must not be modified in place (non-anonymous
2409 * M_EXTPG or sendfile M_EXT buffers).
2410 */
2411 static ktls_mbuf_crypto_st_t
ktls_mbuf_crypto_state(struct mbuf * mb)2412 ktls_mbuf_crypto_state(struct mbuf *mb)
2413 {
2414 bool seen_decrypted, seen_encrypted;
2415
2416 seen_decrypted = false;
2417 seen_encrypted = false;
2418
2419 for (; mb != NULL; mb = mb->m_next) {
2420 if ((mb->m_flags & M_EXTPG) != 0 &&
2421 (mb->m_epg_flags & EPG_FLAG_ANON) == 0)
2422 return (KTLS_MBUF_CRYPTO_ST_SHAREDMBUF);
2423 if ((mb->m_flags & M_EXT) != 0 &&
2424 mb->m_ext.ext_type == EXT_SFBUF)
2425 return (KTLS_MBUF_CRYPTO_ST_SHAREDMBUF);
2426
2427 if (mb->m_flags & M_DECRYPTED)
2428 seen_decrypted = true;
2429 else
2430 seen_encrypted = true;
2431 }
2432
2433 if (seen_decrypted && seen_encrypted)
2434 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2435 else
2436 return (seen_decrypted ?
2437 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2438 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2439 }
2440
2441 /*
2442 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2443 */
2444 static int
ktls_resync_ifnet(struct socket * so,uint32_t tls_len,uint64_t tls_rcd_num)2445 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2446 {
2447 union if_snd_tag_modify_params params;
2448 struct m_snd_tag *mst;
2449 struct inpcb *inp = sotoinpcb(so);
2450 struct tcpcb *tp = intotcpcb(inp);
2451
2452 mst = so->so_rcv.sb_tls_info->snd_tag;
2453 if (__predict_false(mst == NULL))
2454 return (EINVAL);
2455
2456 INP_RLOCK(inp);
2457 if (tp->t_flags & TF_DISCONNECTED) {
2458 INP_RUNLOCK(inp);
2459 return (ECONNRESET);
2460 }
2461
2462 /* Get the TCP sequence number of the next valid TLS header. */
2463 SOCKBUF_LOCK(&so->so_rcv);
2464 params.tls_rx.tls_hdr_tcp_sn =
2465 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2466 params.tls_rx.tls_rec_length = tls_len;
2467 params.tls_rx.tls_seq_number = tls_rcd_num;
2468 SOCKBUF_UNLOCK(&so->so_rcv);
2469
2470 INP_RUNLOCK(inp);
2471
2472 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2473 return (mst->sw->snd_tag_modify(mst, ¶ms));
2474 }
2475
2476 static void
ktls_drop(struct socket * so,int error)2477 ktls_drop(struct socket *so, int error)
2478 {
2479 struct epoch_tracker et;
2480 struct inpcb *inp = sotoinpcb(so);
2481 struct tcpcb *tp = intotcpcb(inp);
2482
2483 NET_EPOCH_ENTER(et);
2484 INP_WLOCK(inp);
2485 if (!(tp->t_flags & TF_DISCONNECTED)) {
2486 CURVNET_SET(inp->inp_socket->so_vnet);
2487 tp = tcp_drop(tp, error);
2488 CURVNET_RESTORE();
2489 if (tp != NULL)
2490 INP_WUNLOCK(inp);
2491 } else {
2492 so->so_error = error;
2493 SOCK_RECVBUF_LOCK(so);
2494 sorwakeup_locked(so);
2495 INP_WUNLOCK(inp);
2496 }
2497 NET_EPOCH_EXIT(et);
2498 }
2499
2500 static void
ktls_decrypt(struct socket * so)2501 ktls_decrypt(struct socket *so)
2502 {
2503 char tls_header[MBUF_PEXT_HDR_LEN];
2504 struct ktls_session *tls;
2505 struct sockbuf *sb;
2506 struct tls_record_layer *hdr;
2507 struct tls_get_record tgr;
2508 struct mbuf *control, *data, *m;
2509 ktls_mbuf_crypto_st_t state;
2510 uint64_t seqno;
2511 int error, remain, tls_len, trail_len;
2512 bool tls13;
2513 uint8_t vminor, record_type;
2514
2515 hdr = (struct tls_record_layer *)tls_header;
2516 sb = &so->so_rcv;
2517 SOCKBUF_LOCK(sb);
2518 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2519 ("%s: socket %p not running", __func__, so));
2520
2521 tls = sb->sb_tls_info;
2522 MPASS(tls != NULL);
2523
2524 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2525 if (tls13)
2526 vminor = TLS_MINOR_VER_TWO;
2527 else
2528 vminor = tls->params.tls_vminor;
2529 for (;;) {
2530 /* Is there enough queued for a TLS header? */
2531 if (sb->sb_tlscc < tls->params.tls_hlen)
2532 break;
2533
2534 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2535 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2536
2537 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2538 hdr->tls_vminor != vminor)
2539 error = EINVAL;
2540 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2541 error = EINVAL;
2542 else if (tls_len < tls->params.tls_hlen || tls_len >
2543 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2544 tls->params.tls_tlen)
2545 error = EMSGSIZE;
2546 else
2547 error = 0;
2548 if (__predict_false(error != 0)) {
2549 /*
2550 * We have a corrupted record and are likely
2551 * out of sync. The connection isn't
2552 * recoverable at this point, so abort it.
2553 */
2554 SOCKBUF_UNLOCK(sb);
2555 counter_u64_add(ktls_offload_corrupted_records, 1);
2556
2557 ktls_drop(so, error);
2558 goto deref;
2559 }
2560
2561 /* Is the entire record queued? */
2562 if (sb->sb_tlscc < tls_len)
2563 break;
2564
2565 /*
2566 * Split out the portion of the mbuf chain containing
2567 * this TLS record.
2568 */
2569 data = ktls_detach_record(sb, tls_len);
2570 if (data == NULL)
2571 continue;
2572 MPASS(sb->sb_tlsdcc == tls_len);
2573
2574 seqno = sb->sb_tls_seqno;
2575 sb->sb_tls_seqno++;
2576 SBCHECK(sb);
2577 SOCKBUF_UNLOCK(sb);
2578
2579 /* get crypto state for this TLS record */
2580 state = ktls_mbuf_crypto_state(data);
2581
2582 switch (state) {
2583 case KTLS_MBUF_CRYPTO_ST_MIXED:
2584 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2585 if (error)
2586 break;
2587 /* FALLTHROUGH */
2588 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2589 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2590 &trail_len);
2591 if (__predict_true(error == 0)) {
2592 if (tls13) {
2593 error = tls13_find_record_type(tls, data,
2594 tls_len, &trail_len, &record_type);
2595 } else {
2596 record_type = hdr->tls_type;
2597 }
2598 }
2599 break;
2600 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2601 /*
2602 * NIC TLS is only supported for AEAD
2603 * ciphersuites which used a fixed sized
2604 * trailer.
2605 */
2606 if (tls13) {
2607 trail_len = tls->params.tls_tlen - 1;
2608 error = tls13_find_record_type(tls, data,
2609 tls_len, &trail_len, &record_type);
2610 } else {
2611 trail_len = tls->params.tls_tlen;
2612 error = 0;
2613 record_type = hdr->tls_type;
2614 }
2615 break;
2616 case KTLS_MBUF_CRYPTO_ST_SHAREDMBUF:
2617 error = EINVAL;
2618 break;
2619 default:
2620 __assert_unreachable();
2621 }
2622 if (error) {
2623 counter_u64_add(ktls_offload_failed_crypto, 1);
2624
2625 SOCKBUF_LOCK(sb);
2626 if (sb->sb_tlsdcc == 0) {
2627 /*
2628 * sbcut/drop/flush discarded these
2629 * mbufs.
2630 */
2631 m_freem(data);
2632 break;
2633 }
2634
2635 /*
2636 * Drop this TLS record's data, but keep
2637 * decrypting subsequent records.
2638 */
2639 sb->sb_ccc -= tls_len;
2640 sb->sb_tlsdcc = 0;
2641
2642 if (error != EMSGSIZE)
2643 error = EBADMSG;
2644 CURVNET_SET(so->so_vnet);
2645 so->so_error = error;
2646 sorwakeup_locked(so);
2647 CURVNET_RESTORE();
2648
2649 m_freem(data);
2650
2651 SOCKBUF_LOCK(sb);
2652 continue;
2653 }
2654
2655 /* Allocate the control mbuf. */
2656 memset(&tgr, 0, sizeof(tgr));
2657 tgr.tls_type = record_type;
2658 tgr.tls_vmajor = hdr->tls_vmajor;
2659 tgr.tls_vminor = hdr->tls_vminor;
2660 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2661 trail_len);
2662 control = sbcreatecontrol(&tgr, sizeof(tgr),
2663 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2664
2665 SOCKBUF_LOCK(sb);
2666 if (sb->sb_tlsdcc == 0) {
2667 /* sbcut/drop/flush discarded these mbufs. */
2668 MPASS(sb->sb_tlscc == 0);
2669 m_freem(data);
2670 m_freem(control);
2671 break;
2672 }
2673
2674 /*
2675 * Clear the 'dcc' accounting in preparation for
2676 * adding the decrypted record.
2677 */
2678 sb->sb_ccc -= tls_len;
2679 sb->sb_tlsdcc = 0;
2680 SBCHECK(sb);
2681
2682 /* If there is no payload, drop all of the data. */
2683 if (tgr.tls_length == htobe16(0)) {
2684 m_freem(data);
2685 data = NULL;
2686 } else {
2687 /* Trim header. */
2688 remain = tls->params.tls_hlen;
2689 while (remain > 0) {
2690 if (data->m_len > remain) {
2691 data->m_data += remain;
2692 data->m_len -= remain;
2693 break;
2694 }
2695 remain -= data->m_len;
2696 data = m_free(data);
2697 }
2698
2699 /* Trim trailer and clear M_NOTREADY. */
2700 remain = be16toh(tgr.tls_length);
2701 m = data;
2702 for (m = data; remain > m->m_len; m = m->m_next) {
2703 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2704 remain -= m->m_len;
2705 }
2706 m->m_len = remain;
2707 m_freem(m->m_next);
2708 m->m_next = NULL;
2709 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2710
2711 /* Set EOR on the final mbuf. */
2712 m->m_flags |= M_EOR;
2713 }
2714
2715 sbappendcontrol_locked(sb, data, control, 0);
2716
2717 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2718 sb->sb_flags |= SB_TLS_RX_RESYNC;
2719 SOCKBUF_UNLOCK(sb);
2720 ktls_resync_ifnet(so, tls_len, seqno);
2721 SOCKBUF_LOCK(sb);
2722 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2723 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2724 SOCKBUF_UNLOCK(sb);
2725 ktls_resync_ifnet(so, 0, seqno);
2726 SOCKBUF_LOCK(sb);
2727 }
2728 }
2729
2730 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2731
2732 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2733 so->so_error = EMSGSIZE;
2734
2735 sorwakeup_locked(so);
2736
2737 deref:
2738 SOCKBUF_UNLOCK_ASSERT(sb);
2739
2740 CURVNET_SET(so->so_vnet);
2741 sorele(so);
2742 CURVNET_RESTORE();
2743 }
2744
2745 void
ktls_enqueue_to_free(struct mbuf * m)2746 ktls_enqueue_to_free(struct mbuf *m)
2747 {
2748 struct ktls_wq *wq;
2749 bool running;
2750
2751 /* Mark it for freeing. */
2752 m->m_epg_flags |= EPG_FLAG_2FREE;
2753 wq = &ktls_wq[m->m_epg_tls->wq_index];
2754 mtx_lock(&wq->mtx);
2755 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2756 running = wq->running;
2757 mtx_unlock(&wq->mtx);
2758 if (!running)
2759 wakeup(wq);
2760 }
2761
2762 static void *
ktls_buffer_alloc(struct ktls_wq * wq,struct mbuf * m)2763 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2764 {
2765 void *buf;
2766 int domain, running;
2767
2768 if (m->m_epg_npgs <= 2)
2769 return (NULL);
2770 if (ktls_buffer_zone == NULL)
2771 return (NULL);
2772 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2773 /*
2774 * Rate-limit allocation attempts after a failure.
2775 * ktls_buffer_import() will acquire a per-domain mutex to check
2776 * the free page queues and may fail consistently if memory is
2777 * fragmented.
2778 */
2779 return (NULL);
2780 }
2781 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2782 if (buf == NULL) {
2783 domain = PCPU_GET(domain);
2784 wq->lastallocfail = ticks;
2785
2786 /*
2787 * Note that this check is "racy", but the races are
2788 * harmless, and are either a spurious wakeup if
2789 * multiple threads fail allocations before the alloc
2790 * thread wakes, or waiting an extra second in case we
2791 * see an old value of running == true.
2792 */
2793 if (!VM_DOMAIN_EMPTY(domain)) {
2794 running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2795 if (!running)
2796 wakeup(&ktls_domains[domain].reclaim_td);
2797 }
2798 }
2799 return (buf);
2800 }
2801
2802 static int
ktls_encrypt_record(struct ktls_wq * wq,struct mbuf * m,struct ktls_session * tls,struct ktls_ocf_encrypt_state * state)2803 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2804 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2805 {
2806 vm_page_t pg;
2807 int error, i, len, off;
2808
2809 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2810 ("%p not unready & nomap mbuf\n", m));
2811 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2812 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2813 ktls_maxlen));
2814
2815 /* Anonymous mbufs are encrypted in place. */
2816 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2817 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2818
2819 /*
2820 * For file-backed mbufs (from sendfile), anonymous wired
2821 * pages are allocated and used as the encryption destination.
2822 */
2823 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2824 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2825 m->m_epg_1st_off;
2826 state->dst_iov[0].iov_base = (char *)state->cbuf +
2827 m->m_epg_1st_off;
2828 state->dst_iov[0].iov_len = len;
2829 state->parray[0] = DMAP_TO_PHYS(state->cbuf);
2830 i = 1;
2831 } else {
2832 off = m->m_epg_1st_off;
2833 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2834 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2835 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2836 len = m_epg_pagelen(m, i, off);
2837 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2838 state->dst_iov[i].iov_base =
2839 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2840 state->dst_iov[i].iov_len = len;
2841 }
2842 }
2843 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2844 state->dst_iov[i].iov_base = m->m_epg_trail;
2845 state->dst_iov[i].iov_len = m->m_epg_trllen;
2846
2847 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2848
2849 if (__predict_false(error != 0)) {
2850 /* Free the anonymous pages. */
2851 if (state->cbuf != NULL)
2852 uma_zfree(ktls_buffer_zone, state->cbuf);
2853 else {
2854 for (i = 0; i < m->m_epg_npgs; i++) {
2855 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2856 (void)vm_page_unwire_noq(pg);
2857 vm_page_free(pg);
2858 }
2859 }
2860 }
2861 return (error);
2862 }
2863
2864 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2865 static u_int
ktls_batched_records(struct mbuf * m)2866 ktls_batched_records(struct mbuf *m)
2867 {
2868 int page_count, records;
2869
2870 records = 0;
2871 page_count = m->m_epg_enc_cnt;
2872 while (page_count > 0) {
2873 records++;
2874 page_count -= m->m_epg_nrdy;
2875 m = m->m_next;
2876 }
2877 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2878 return (records);
2879 }
2880
2881 void
ktls_enqueue(struct mbuf * m,struct socket * so,int page_count)2882 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2883 {
2884 struct ktls_session *tls;
2885 struct ktls_wq *wq;
2886 int queued;
2887 bool running;
2888
2889 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2890 (M_EXTPG | M_NOTREADY)),
2891 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2892 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2893
2894 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2895
2896 m->m_epg_enc_cnt = page_count;
2897
2898 /*
2899 * Save a pointer to the socket. The caller is responsible
2900 * for taking an additional reference via soref().
2901 */
2902 m->m_epg_so = so;
2903
2904 queued = 1;
2905 tls = m->m_epg_tls;
2906 wq = &ktls_wq[tls->wq_index];
2907 mtx_lock(&wq->mtx);
2908 if (__predict_false(tls->sequential_records)) {
2909 /*
2910 * For TLS 1.0, records must be encrypted
2911 * sequentially. For a given connection, all records
2912 * queued to the associated work queue are processed
2913 * sequentially. However, sendfile(2) might complete
2914 * I/O requests spanning multiple TLS records out of
2915 * order. Here we ensure TLS records are enqueued to
2916 * the work queue in FIFO order.
2917 *
2918 * tls->next_seqno holds the sequence number of the
2919 * next TLS record that should be enqueued to the work
2920 * queue. If this next record is not tls->next_seqno,
2921 * it must be a future record, so insert it, sorted by
2922 * TLS sequence number, into tls->pending_records and
2923 * return.
2924 *
2925 * If this TLS record matches tls->next_seqno, place
2926 * it in the work queue and then check
2927 * tls->pending_records to see if any
2928 * previously-queued records are now ready for
2929 * encryption.
2930 */
2931 if (m->m_epg_seqno != tls->next_seqno) {
2932 struct mbuf *n, *p;
2933
2934 p = NULL;
2935 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2936 if (n->m_epg_seqno > m->m_epg_seqno)
2937 break;
2938 p = n;
2939 }
2940 if (n == NULL)
2941 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2942 m_epg_stailq);
2943 else if (p == NULL)
2944 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2945 m_epg_stailq);
2946 else
2947 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2948 m_epg_stailq);
2949 mtx_unlock(&wq->mtx);
2950 counter_u64_add(ktls_cnt_tx_pending, 1);
2951 return;
2952 }
2953
2954 tls->next_seqno += ktls_batched_records(m);
2955 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2956
2957 while (!STAILQ_EMPTY(&tls->pending_records)) {
2958 struct mbuf *n;
2959
2960 n = STAILQ_FIRST(&tls->pending_records);
2961 if (n->m_epg_seqno != tls->next_seqno)
2962 break;
2963
2964 queued++;
2965 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2966 tls->next_seqno += ktls_batched_records(n);
2967 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2968 }
2969 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2970 } else
2971 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2972
2973 running = wq->running;
2974 mtx_unlock(&wq->mtx);
2975 if (!running)
2976 wakeup(wq);
2977 counter_u64_add(ktls_cnt_tx_queued, queued);
2978 }
2979
2980 /*
2981 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2982 * the pages from the file and replace them with the anonymous pages
2983 * allocated in ktls_encrypt_record().
2984 */
2985 static void
ktls_finish_nonanon(struct mbuf * m,struct ktls_ocf_encrypt_state * state)2986 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2987 {
2988 int i;
2989
2990 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2991
2992 /* Free the old pages. */
2993 m->m_ext.ext_free(m);
2994
2995 /* Replace them with the new pages. */
2996 if (state->cbuf != NULL) {
2997 for (i = 0; i < m->m_epg_npgs; i++)
2998 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2999
3000 /* Contig pages should go back to the cache. */
3001 m->m_ext.ext_free = ktls_free_mext_contig;
3002 } else {
3003 for (i = 0; i < m->m_epg_npgs; i++)
3004 m->m_epg_pa[i] = state->parray[i];
3005
3006 /* Use the basic free routine. */
3007 m->m_ext.ext_free = mb_free_mext_pgs;
3008 }
3009
3010 /* Pages are now writable. */
3011 m->m_epg_flags |= EPG_FLAG_ANON;
3012 }
3013
3014 static __noinline void
ktls_encrypt(struct ktls_wq * wq,struct mbuf * top)3015 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
3016 {
3017 struct ktls_ocf_encrypt_state state;
3018 struct ktls_session *tls;
3019 struct socket *so;
3020 struct mbuf *m;
3021 int error, npages, total_pages;
3022
3023 so = top->m_epg_so;
3024 tls = top->m_epg_tls;
3025 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3026 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3027 #ifdef INVARIANTS
3028 top->m_epg_so = NULL;
3029 #endif
3030 total_pages = top->m_epg_enc_cnt;
3031 npages = 0;
3032
3033 /*
3034 * Encrypt the TLS records in the chain of mbufs starting with
3035 * 'top'. 'total_pages' gives us a total count of pages and is
3036 * used to know when we have finished encrypting the TLS
3037 * records originally queued with 'top'.
3038 *
3039 * NB: These mbufs are queued in the socket buffer and
3040 * 'm_next' is traversing the mbufs in the socket buffer. The
3041 * socket buffer lock is not held while traversing this chain.
3042 * Since the mbufs are all marked M_NOTREADY their 'm_next'
3043 * pointers should be stable. However, the 'm_next' of the
3044 * last mbuf encrypted is not necessarily NULL. It can point
3045 * to other mbufs appended while 'top' was on the TLS work
3046 * queue.
3047 *
3048 * Each mbuf holds an entire TLS record.
3049 */
3050 error = 0;
3051 for (m = top; npages != total_pages; m = m->m_next) {
3052 KASSERT(m->m_epg_tls == tls,
3053 ("different TLS sessions in a single mbuf chain: %p vs %p",
3054 tls, m->m_epg_tls));
3055 KASSERT(npages + m->m_epg_npgs <= total_pages,
3056 ("page count mismatch: top %p, total_pages %d, m %p", top,
3057 total_pages, m));
3058
3059 error = ktls_encrypt_record(wq, m, tls, &state);
3060 if (error) {
3061 counter_u64_add(ktls_offload_failed_crypto, 1);
3062 break;
3063 }
3064
3065 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3066 ktls_finish_nonanon(m, &state);
3067 m->m_flags |= M_RDONLY;
3068
3069 npages += m->m_epg_nrdy;
3070
3071 /*
3072 * Drop a reference to the session now that it is no
3073 * longer needed. Existing code depends on encrypted
3074 * records having no associated session vs
3075 * yet-to-be-encrypted records having an associated
3076 * session.
3077 */
3078 m->m_epg_tls = NULL;
3079 ktls_free(tls);
3080 }
3081
3082 CURVNET_SET(so->so_vnet);
3083 if (error == 0) {
3084 (void)so->so_proto->pr_ready(so, top, npages);
3085 } else {
3086 ktls_drop(so, EIO);
3087 mb_free_notready(top, total_pages);
3088 }
3089
3090 sorele(so);
3091 CURVNET_RESTORE();
3092 }
3093
3094 void
ktls_encrypt_cb(struct ktls_ocf_encrypt_state * state,int error)3095 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3096 {
3097 struct ktls_session *tls;
3098 struct socket *so;
3099 struct mbuf *m;
3100 int npages;
3101
3102 m = state->m;
3103
3104 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3105 ktls_finish_nonanon(m, state);
3106 m->m_flags |= M_RDONLY;
3107
3108 so = state->so;
3109 free(state, M_KTLS);
3110
3111 /*
3112 * Drop a reference to the session now that it is no longer
3113 * needed. Existing code depends on encrypted records having
3114 * no associated session vs yet-to-be-encrypted records having
3115 * an associated session.
3116 */
3117 tls = m->m_epg_tls;
3118 m->m_epg_tls = NULL;
3119 ktls_free(tls);
3120
3121 if (error != 0)
3122 counter_u64_add(ktls_offload_failed_crypto, 1);
3123
3124 CURVNET_SET(so->so_vnet);
3125 npages = m->m_epg_nrdy;
3126
3127 if (error == 0) {
3128 (void)so->so_proto->pr_ready(so, m, npages);
3129 } else {
3130 ktls_drop(so, EIO);
3131 mb_free_notready(m, npages);
3132 }
3133
3134 sorele(so);
3135 CURVNET_RESTORE();
3136 }
3137
3138 /*
3139 * Similar to ktls_encrypt, but used with asynchronous OCF backends
3140 * (coprocessors) where encryption does not use host CPU resources and
3141 * it can be beneficial to queue more requests than CPUs.
3142 */
3143 static __noinline void
ktls_encrypt_async(struct ktls_wq * wq,struct mbuf * top)3144 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3145 {
3146 struct ktls_ocf_encrypt_state *state;
3147 struct ktls_session *tls;
3148 struct socket *so;
3149 struct mbuf *m, *n;
3150 int error, mpages, npages, total_pages;
3151
3152 so = top->m_epg_so;
3153 tls = top->m_epg_tls;
3154 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3155 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3156 #ifdef INVARIANTS
3157 top->m_epg_so = NULL;
3158 #endif
3159 total_pages = top->m_epg_enc_cnt;
3160 npages = 0;
3161
3162 error = 0;
3163 for (m = top; npages != total_pages; m = n) {
3164 KASSERT(m->m_epg_tls == tls,
3165 ("different TLS sessions in a single mbuf chain: %p vs %p",
3166 tls, m->m_epg_tls));
3167 KASSERT(npages + m->m_epg_npgs <= total_pages,
3168 ("page count mismatch: top %p, total_pages %d, m %p", top,
3169 total_pages, m));
3170
3171 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3172 soref(so);
3173 state->so = so;
3174 state->m = m;
3175
3176 mpages = m->m_epg_nrdy;
3177 n = m->m_next;
3178
3179 error = ktls_encrypt_record(wq, m, tls, state);
3180 if (error) {
3181 counter_u64_add(ktls_offload_failed_crypto, 1);
3182 free(state, M_KTLS);
3183 CURVNET_SET(so->so_vnet);
3184 sorele(so);
3185 CURVNET_RESTORE();
3186 break;
3187 }
3188
3189 npages += mpages;
3190 }
3191
3192 CURVNET_SET(so->so_vnet);
3193 if (error != 0) {
3194 ktls_drop(so, EIO);
3195 mb_free_notready(m, total_pages - npages);
3196 }
3197
3198 sorele(so);
3199 CURVNET_RESTORE();
3200 }
3201
3202 static int
ktls_bind_domain(int domain)3203 ktls_bind_domain(int domain)
3204 {
3205 int error;
3206
3207 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3208 if (error != 0)
3209 return (error);
3210 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3211 return (0);
3212 }
3213
3214 static void
ktls_reclaim_thread(void * ctx)3215 ktls_reclaim_thread(void *ctx)
3216 {
3217 struct ktls_domain_info *ktls_domain = ctx;
3218 struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3219 struct sysctl_oid *oid;
3220 char name[80];
3221 int error, domain;
3222
3223 domain = ktls_domain - ktls_domains;
3224 if (bootverbose)
3225 printf("Starting KTLS reclaim thread for domain %d\n", domain);
3226 error = ktls_bind_domain(domain);
3227 if (error)
3228 printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3229 domain, error);
3230 snprintf(name, sizeof(name), "domain%d", domain);
3231 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3232 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3233 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3234 CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed");
3235 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3236 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3237 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3238 CTLFLAG_RD, &sc->running, 0, "thread running");
3239
3240 for (;;) {
3241 atomic_store_int(&sc->running, 0);
3242 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3243 atomic_store_int(&sc->running, 1);
3244 sc->wakeups++;
3245 /*
3246 * Below we attempt to reclaim ktls_max_reclaim
3247 * buffers using vm_page_reclaim_contig_domain_ext().
3248 * We do this here, as this function can take several
3249 * seconds to scan all of memory and it does not
3250 * matter if this thread pauses for a while. If we
3251 * block a ktls worker thread, we risk developing
3252 * backlogs of buffers to be encrypted, leading to
3253 * surges of traffic and potential NIC output drops.
3254 */
3255 if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3256 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3257 ktls_max_reclaim) != 0) {
3258 vm_wait_domain(domain);
3259 } else {
3260 sc->reclaims += ktls_max_reclaim;
3261 }
3262 }
3263 }
3264
3265 static void
ktls_work_thread(void * ctx)3266 ktls_work_thread(void *ctx)
3267 {
3268 struct ktls_wq *wq = ctx;
3269 struct mbuf *m, *n;
3270 struct socket *so, *son;
3271 STAILQ_HEAD(, mbuf) local_m_head;
3272 STAILQ_HEAD(, socket) local_so_head;
3273 int cpu;
3274
3275 cpu = wq - ktls_wq;
3276 if (bootverbose)
3277 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3278
3279 /*
3280 * Bind to a core. If ktls_bind_threads is > 1, then
3281 * we bind to the NUMA domain instead.
3282 */
3283 if (ktls_bind_threads) {
3284 int error;
3285
3286 if (ktls_bind_threads > 1) {
3287 struct pcpu *pc = pcpu_find(cpu);
3288
3289 error = ktls_bind_domain(pc->pc_domain);
3290 } else {
3291 cpuset_t mask;
3292
3293 CPU_SETOF(cpu, &mask);
3294 error = cpuset_setthread(curthread->td_tid, &mask);
3295 }
3296 if (error)
3297 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3298 cpu, error);
3299 }
3300 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3301 fpu_kern_thread(0);
3302 #endif
3303 for (;;) {
3304 mtx_lock(&wq->mtx);
3305 while (STAILQ_EMPTY(&wq->m_head) &&
3306 STAILQ_EMPTY(&wq->so_head)) {
3307 wq->running = false;
3308 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3309 wq->running = true;
3310 }
3311
3312 STAILQ_INIT(&local_m_head);
3313 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3314 STAILQ_INIT(&local_so_head);
3315 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3316 mtx_unlock(&wq->mtx);
3317
3318 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3319 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3320 ktls_free(m->m_epg_tls);
3321 m_free_raw(m);
3322 } else {
3323 if (m->m_epg_tls->sync_dispatch)
3324 ktls_encrypt(wq, m);
3325 else
3326 ktls_encrypt_async(wq, m);
3327 counter_u64_add(ktls_cnt_tx_queued, -1);
3328 }
3329 }
3330
3331 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3332 ktls_decrypt(so);
3333 counter_u64_add(ktls_cnt_rx_queued, -1);
3334 }
3335 }
3336 }
3337
3338 static void
ktls_disable_ifnet_help(void * context,int pending __unused)3339 ktls_disable_ifnet_help(void *context, int pending __unused)
3340 {
3341 struct ktls_session *tls;
3342 struct inpcb *inp;
3343 struct tcpcb *tp;
3344 struct socket *so;
3345 int err;
3346
3347 tls = context;
3348 inp = tls->inp;
3349 if (inp == NULL)
3350 return;
3351 INP_WLOCK(inp);
3352 so = inp->inp_socket;
3353 MPASS(so != NULL);
3354 tp = intotcpcb(inp);
3355 if (tp->t_flags & TF_DISCONNECTED) {
3356 goto out;
3357 }
3358
3359 if (so->so_snd.sb_tls_info != NULL)
3360 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3361 else
3362 err = ENXIO;
3363 if (err == 0) {
3364 counter_u64_add(ktls_ifnet_disable_ok, 1);
3365 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3366 if ((tp->t_flags & TF_DISCONNECTED) == 0 &&
3367 tp->t_fb->tfb_hwtls_change != NULL)
3368 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3369 } else {
3370 counter_u64_add(ktls_ifnet_disable_fail, 1);
3371 }
3372
3373 out:
3374 CURVNET_SET(so->so_vnet);
3375 sorele(so);
3376 CURVNET_RESTORE();
3377 INP_WUNLOCK(inp);
3378 ktls_free(tls);
3379 }
3380
3381 /*
3382 * Called when re-transmits are becoming a substantial portion of the
3383 * sends on this connection. When this happens, we transition the
3384 * connection to software TLS. This is needed because most inline TLS
3385 * NICs keep crypto state only for in-order transmits. This means
3386 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3387 * re-DMA the entire TLS record up to and including the current
3388 * segment. This means that when re-transmitting the last ~1448 byte
3389 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3390 * of magnitude more data than we are sending. This can cause the
3391 * PCIe link to saturate well before the network, which can cause
3392 * output drops, and a general loss of capacity.
3393 */
3394 void
ktls_disable_ifnet(void * arg)3395 ktls_disable_ifnet(void *arg)
3396 {
3397 struct tcpcb *tp;
3398 struct inpcb *inp;
3399 struct socket *so;
3400 struct ktls_session *tls;
3401
3402 tp = arg;
3403 inp = tptoinpcb(tp);
3404 INP_WLOCK_ASSERT(inp);
3405 so = inp->inp_socket;
3406 SOCK_LOCK(so);
3407 tls = so->so_snd.sb_tls_info;
3408 if (tp->t_nic_ktls_xmit_dis == 1) {
3409 SOCK_UNLOCK(so);
3410 return;
3411 }
3412
3413 /*
3414 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3415 * ifnet can only be done once per connection, so we never want
3416 * to do it again
3417 */
3418
3419 (void)ktls_hold(tls);
3420 soref(so);
3421 tp->t_nic_ktls_xmit_dis = 1;
3422 SOCK_UNLOCK(so);
3423 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3424 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3425 }
3426
3427 void
ktls_session_to_xktls_onedir(const struct ktls_session * ktls,bool export_keys,struct xktls_session_onedir * xk)3428 ktls_session_to_xktls_onedir(const struct ktls_session *ktls, bool export_keys,
3429 struct xktls_session_onedir *xk)
3430 {
3431 if_t ifp;
3432 struct m_snd_tag *st;
3433
3434 xk->gen = ktls->gen;
3435 #define A(m) xk->m = ktls->params.m
3436 A(cipher_algorithm);
3437 A(auth_algorithm);
3438 A(cipher_key_len);
3439 A(auth_key_len);
3440 A(max_frame_len);
3441 A(tls_vmajor);
3442 A(tls_vminor);
3443 A(tls_hlen);
3444 A(tls_tlen);
3445 A(tls_bs);
3446 A(flags);
3447 if (export_keys) {
3448 memcpy(&xk->iv, &ktls->params.iv, XKTLS_SESSION_IV_BUF_LEN);
3449 A(iv_len);
3450 } else {
3451 memset(&xk->iv, 0, XKTLS_SESSION_IV_BUF_LEN);
3452 xk->iv_len = 0;
3453 }
3454 #undef A
3455 if ((st = ktls->snd_tag) != NULL &&
3456 (ifp = ktls->snd_tag->ifp) != NULL)
3457 strncpy(xk->ifnet, if_name(ifp), sizeof(xk->ifnet));
3458 }
3459
3460 void
ktls_session_copy_keys(const struct ktls_session * ktls,uint8_t * data,size_t * sz)3461 ktls_session_copy_keys(const struct ktls_session *ktls,
3462 uint8_t *data, size_t *sz)
3463 {
3464 size_t t, ta, tc;
3465
3466 if (ktls == NULL) {
3467 *sz = 0;
3468 return;
3469 }
3470 t = *sz;
3471 tc = MIN(t, ktls->params.cipher_key_len);
3472 if (data != NULL)
3473 memcpy(data, ktls->params.cipher_key, tc);
3474 ta = MIN(t - tc, ktls->params.auth_key_len);
3475 if (data != NULL)
3476 memcpy(data + tc, ktls->params.auth_key, ta);
3477 *sz = ta + tc;
3478 }
3479