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_rx_offload_enable = true;
152 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, rx_enable, CTLFLAG_RWTUN,
153 &ktls_rx_offload_enable, 0,
154 "Enable support for kernel TLS receive offload");
155
156 static bool ktls_cbc_enable = true;
157 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
158 &ktls_cbc_enable, 1,
159 "Enable support of AES-CBC crypto for kernel TLS");
160
161 static bool ktls_sw_buffer_cache = true;
162 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
163 &ktls_sw_buffer_cache, 1,
164 "Enable caching of output buffers for SW encryption");
165
166 static int ktls_max_reclaim = 1024;
167 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN,
168 &ktls_max_reclaim, 128,
169 "Max number of 16k buffers to reclaim in thread context");
170
171 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
172 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
173 &ktls_tasks_active, "Number of active tasks");
174
175 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
176 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
177 &ktls_cnt_tx_pending,
178 "Number of TLS 1.0 records waiting for earlier TLS records");
179
180 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
181 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
182 &ktls_cnt_tx_queued,
183 "Number of TLS records in queue to tasks for SW encryption");
184
185 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
186 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
187 &ktls_cnt_rx_queued,
188 "Number of TLS sockets in queue to tasks for SW decryption");
189
190 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
191 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
192 CTLFLAG_RD, &ktls_offload_total,
193 "Total successful TLS setups (parameters set)");
194
195 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
196 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
197 CTLFLAG_RD, &ktls_offload_enable_calls,
198 "Total number of TLS enable calls made");
199
200 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
201 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
202 &ktls_offload_active, "Total Active TLS sessions");
203
204 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
205 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
206 &ktls_offload_corrupted_records, "Total corrupted TLS records received");
207
208 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
209 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
210 &ktls_offload_failed_crypto, "Total TLS crypto failures");
211
212 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
214 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
215
216 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
218 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
219
220 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
222 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
223
224 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
225 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
226 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
227
228 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
229 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
230 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
231
232 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
233 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
234 &ktls_destroy_task,
235 "Number of times ktls session was destroyed via taskqueue");
236
237 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
238 "Software TLS session stats");
239 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
240 "Hardware (ifnet) TLS session stats");
241 #ifdef TCP_OFFLOAD
242 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
243 "TOE TLS session stats");
244 #endif
245
246 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
247 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
248 "Active number of software TLS sessions using AES-CBC");
249
250 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
251 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
252 "Active number of software TLS sessions using AES-GCM");
253
254 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
255 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
256 &ktls_sw_chacha20,
257 "Active number of software TLS sessions using Chacha20-Poly1305");
258
259 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
260 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
261 &ktls_ifnet_cbc,
262 "Active number of ifnet TLS sessions using AES-CBC");
263
264 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
265 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
266 &ktls_ifnet_gcm,
267 "Active number of ifnet TLS sessions using AES-GCM");
268
269 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
270 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
271 &ktls_ifnet_chacha20,
272 "Active number of ifnet TLS sessions using Chacha20-Poly1305");
273
274 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
275 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
276 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
277
278 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
279 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
280 &ktls_ifnet_reset_dropped,
281 "TLS sessions dropped after failing to update ifnet send tag");
282
283 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
284 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
285 &ktls_ifnet_reset_failed,
286 "TLS sessions that failed to allocate a new ifnet send tag");
287
288 static int ktls_ifnet_permitted = 1;
289 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
290 &ktls_ifnet_permitted, 1,
291 "Whether to permit hardware (ifnet) TLS sessions");
292
293 #ifdef TCP_OFFLOAD
294 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
295 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
296 &ktls_toe_cbc,
297 "Active number of TOE TLS sessions using AES-CBC");
298
299 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
300 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
301 &ktls_toe_gcm,
302 "Active number of TOE TLS sessions using AES-GCM");
303
304 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
305 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
306 &ktls_toe_chacha20,
307 "Active number of TOE TLS sessions using Chacha20-Poly1305");
308 #endif
309
310 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
311
312 static void ktls_reclaim_thread(void *ctx);
313 static void ktls_reset_receive_tag(void *context, int pending);
314 static void ktls_reset_send_tag(void *context, int pending);
315 static void ktls_work_thread(void *ctx);
316
317 int
ktls_copyin_tls_enable(struct sockopt * sopt,struct tls_enable * tls)318 ktls_copyin_tls_enable(struct sockopt *sopt, struct tls_enable *tls)
319 {
320 struct tls_enable_v0 tls_v0;
321 int error;
322 uint8_t *cipher_key = NULL, *iv = NULL, *auth_key = NULL;
323
324 if (sopt->sopt_valsize == sizeof(tls_v0)) {
325 error = sooptcopyin(sopt, &tls_v0, sizeof(tls_v0), sizeof(tls_v0));
326 if (error != 0)
327 goto done;
328 memset(tls, 0, sizeof(*tls));
329 tls->cipher_key = tls_v0.cipher_key;
330 tls->iv = tls_v0.iv;
331 tls->auth_key = tls_v0.auth_key;
332 tls->cipher_algorithm = tls_v0.cipher_algorithm;
333 tls->cipher_key_len = tls_v0.cipher_key_len;
334 tls->iv_len = tls_v0.iv_len;
335 tls->auth_algorithm = tls_v0.auth_algorithm;
336 tls->auth_key_len = tls_v0.auth_key_len;
337 tls->flags = tls_v0.flags;
338 tls->tls_vmajor = tls_v0.tls_vmajor;
339 tls->tls_vminor = tls_v0.tls_vminor;
340 } else
341 error = sooptcopyin(sopt, tls, sizeof(*tls), sizeof(*tls));
342
343 if (error != 0)
344 return (error);
345
346 if (tls->cipher_key_len < 0 || tls->cipher_key_len > TLS_MAX_PARAM_SIZE)
347 return (EINVAL);
348 if (tls->iv_len < 0 || tls->iv_len > sizeof(((struct ktls_session *)NULL)->params.iv))
349 return (EINVAL);
350 if (tls->auth_key_len < 0 || tls->auth_key_len > TLS_MAX_PARAM_SIZE)
351 return (EINVAL);
352
353 /* All supported algorithms require a cipher key. */
354 if (tls->cipher_key_len == 0)
355 return (EINVAL);
356
357 /*
358 * Now do a deep copy of the variable-length arrays in the struct, so that
359 * subsequent consumers of it can reliably assume kernel memory. This
360 * requires doing our own allocations, which we will free in the
361 * error paths so that our caller need only worry about outstanding
362 * allocations existing on successful return.
363 */
364 if (tls->cipher_key_len != 0) {
365 cipher_key = malloc(tls->cipher_key_len, M_KTLS, M_WAITOK);
366 if (sopt->sopt_td != NULL) {
367 error = copyin(tls->cipher_key, cipher_key, tls->cipher_key_len);
368 if (error != 0)
369 goto done;
370 } else {
371 bcopy(tls->cipher_key, cipher_key, tls->cipher_key_len);
372 }
373 }
374 if (tls->iv_len != 0) {
375 iv = malloc(tls->iv_len, M_KTLS, M_WAITOK);
376 if (sopt->sopt_td != NULL) {
377 error = copyin(tls->iv, iv, tls->iv_len);
378 if (error != 0)
379 goto done;
380 } else {
381 bcopy(tls->iv, iv, tls->iv_len);
382 }
383 }
384 if (tls->auth_key_len != 0) {
385 auth_key = malloc(tls->auth_key_len, M_KTLS, M_WAITOK);
386 if (sopt->sopt_td != NULL) {
387 error = copyin(tls->auth_key, auth_key, tls->auth_key_len);
388 if (error != 0)
389 goto done;
390 } else {
391 bcopy(tls->auth_key, auth_key, tls->auth_key_len);
392 }
393 }
394 tls->cipher_key = cipher_key;
395 tls->iv = iv;
396 tls->auth_key = auth_key;
397
398 done:
399 if (error != 0) {
400 zfree(cipher_key, M_KTLS);
401 zfree(iv, M_KTLS);
402 zfree(auth_key, M_KTLS);
403 }
404
405 return (error);
406 }
407
408 void
ktls_cleanup_tls_enable(struct tls_enable * tls)409 ktls_cleanup_tls_enable(struct tls_enable *tls)
410 {
411 zfree(__DECONST(void *, tls->cipher_key), M_KTLS);
412 zfree(__DECONST(void *, tls->iv), M_KTLS);
413 zfree(__DECONST(void *, tls->auth_key), M_KTLS);
414 }
415
416 static u_int
ktls_get_cpu(struct socket * so)417 ktls_get_cpu(struct socket *so)
418 {
419 struct inpcb *inp;
420 #ifdef NUMA
421 struct ktls_domain_info *di;
422 #endif
423 u_int cpuid;
424
425 inp = sotoinpcb(so);
426 #ifdef RSS
427 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
428 if (cpuid != NETISR_CPUID_NONE)
429 return (cpuid);
430 #endif
431 /*
432 * Just use the flowid to shard connections in a repeatable
433 * fashion. Note that TLS 1.0 sessions rely on the
434 * serialization provided by having the same connection use
435 * the same queue.
436 */
437 #ifdef NUMA
438 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
439 di = &ktls_domains[inp->inp_numa_domain];
440 cpuid = di->cpu[inp->inp_flowid % di->count];
441 } else
442 #endif
443 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
444 return (cpuid);
445 }
446
447 static int
ktls_buffer_import(void * arg,void ** store,int count,int domain,int flags)448 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
449 {
450 vm_page_t m;
451 int i, req;
452
453 KASSERT((ktls_maxlen & PAGE_MASK) == 0,
454 ("%s: ktls max length %d is not page size-aligned",
455 __func__, ktls_maxlen));
456
457 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
458 for (i = 0; i < count; i++) {
459 m = vm_page_alloc_noobj_contig_domain(domain, req,
460 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
461 VM_MEMATTR_DEFAULT);
462 if (m == NULL)
463 break;
464 store[i] = VM_PAGE_TO_DMAP(m);
465 }
466 return (i);
467 }
468
469 static void
ktls_buffer_release(void * arg __unused,void ** store,int count)470 ktls_buffer_release(void *arg __unused, void **store, int count)
471 {
472 vm_page_t m;
473 int i, j;
474
475 for (i = 0; i < count; i++) {
476 m = DMAP_TO_VM_PAGE(store[i]);
477 for (j = 0; j < atop(ktls_maxlen); j++) {
478 (void)vm_page_unwire_noq(m + j);
479 vm_page_free(m + j);
480 }
481 }
482 }
483
484 static void
ktls_free_mext_contig(struct mbuf * m)485 ktls_free_mext_contig(struct mbuf *m)
486 {
487 M_ASSERTEXTPG(m);
488 uma_zfree(ktls_buffer_zone, PHYS_TO_DMAP(m->m_epg_pa[0]));
489 }
490
491 static int
ktls_init(void)492 ktls_init(void)
493 {
494 struct thread *td;
495 struct pcpu *pc;
496 int count, domain, error, i;
497
498 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
499 M_WAITOK | M_ZERO);
500
501 ktls_session_zone = uma_zcreate("ktls_session",
502 sizeof(struct ktls_session),
503 NULL, NULL, NULL, NULL,
504 UMA_ALIGN_CACHE, 0);
505
506 if (ktls_sw_buffer_cache) {
507 ktls_buffer_zone = uma_zcache_create("ktls_buffers",
508 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
509 ktls_buffer_import, ktls_buffer_release, NULL,
510 UMA_ZONE_FIRSTTOUCH | UMA_ZONE_NOTRIM);
511 }
512
513 /*
514 * Initialize the workqueues to run the TLS work. We create a
515 * work queue for each CPU.
516 */
517 CPU_FOREACH(i) {
518 STAILQ_INIT(&ktls_wq[i].m_head);
519 STAILQ_INIT(&ktls_wq[i].so_head);
520 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
521 if (ktls_bind_threads > 1) {
522 pc = pcpu_find(i);
523 domain = pc->pc_domain;
524 count = ktls_domains[domain].count;
525 ktls_domains[domain].cpu[count] = i;
526 ktls_domains[domain].count++;
527 }
528 ktls_cpuid_lookup[ktls_number_threads] = i;
529 ktls_number_threads++;
530 }
531
532 /*
533 * If we somehow have an empty domain, fall back to choosing
534 * among all KTLS threads.
535 */
536 if (ktls_bind_threads > 1) {
537 for (i = 0; i < vm_ndomains; i++) {
538 if (ktls_domains[i].count == 0) {
539 ktls_bind_threads = 1;
540 break;
541 }
542 }
543 }
544
545 /* Start kthreads for each workqueue. */
546 CPU_FOREACH(i) {
547 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
548 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
549 if (error) {
550 printf("Can't add KTLS thread %d error %d\n", i, error);
551 return (error);
552 }
553 }
554
555 /*
556 * Start an allocation thread per-domain to perform blocking allocations
557 * of 16k physically contiguous TLS crypto destination buffers.
558 */
559 if (ktls_sw_buffer_cache) {
560 for (domain = 0; domain < vm_ndomains; domain++) {
561 if (VM_DOMAIN_EMPTY(domain))
562 continue;
563 if (CPU_EMPTY(&cpuset_domain[domain]))
564 continue;
565 error = kproc_kthread_add(ktls_reclaim_thread,
566 &ktls_domains[domain], &ktls_proc,
567 &ktls_domains[domain].reclaim_td.td,
568 0, 0, "KTLS", "reclaim_%d", domain);
569 if (error) {
570 printf("Can't add KTLS reclaim thread %d error %d\n",
571 domain, error);
572 return (error);
573 }
574 }
575 }
576
577 if (bootverbose)
578 printf("KTLS: Initialized %d threads\n", ktls_number_threads);
579 return (0);
580 }
581
582 static int
ktls_start_kthreads(void)583 ktls_start_kthreads(void)
584 {
585 int error, state;
586
587 start:
588 state = atomic_load_acq_int(&ktls_init_state);
589 if (__predict_true(state > 0))
590 return (0);
591 if (state < 0)
592 return (ENXIO);
593
594 sx_xlock(&ktls_init_lock);
595 if (ktls_init_state != 0) {
596 sx_xunlock(&ktls_init_lock);
597 goto start;
598 }
599
600 error = ktls_init();
601 if (error == 0)
602 state = 1;
603 else
604 state = -1;
605 atomic_store_rel_int(&ktls_init_state, state);
606 sx_xunlock(&ktls_init_lock);
607 return (error);
608 }
609
610 static int
ktls_create_session(struct socket * so,struct tls_enable * en,struct ktls_session ** tlsp,int direction)611 ktls_create_session(struct socket *so, struct tls_enable *en,
612 struct ktls_session **tlsp, int direction)
613 {
614 struct ktls_session *tls;
615 int error;
616
617 /* Only TLS 1.0 - 1.3 are supported. */
618 if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
619 return (EINVAL);
620 if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
621 en->tls_vminor > TLS_MINOR_VER_THREE)
622 return (EINVAL);
623
624
625 /* No flags are currently supported. */
626 if (en->flags != 0)
627 return (EINVAL);
628
629 /* Common checks for supported algorithms. */
630 switch (en->cipher_algorithm) {
631 case CRYPTO_AES_NIST_GCM_16:
632 /*
633 * auth_algorithm isn't used, but permit GMAC values
634 * for compatibility.
635 */
636 switch (en->auth_algorithm) {
637 case 0:
638 #ifdef COMPAT_FREEBSD12
639 /* XXX: Really 13.0-current COMPAT. */
640 case CRYPTO_AES_128_NIST_GMAC:
641 case CRYPTO_AES_192_NIST_GMAC:
642 case CRYPTO_AES_256_NIST_GMAC:
643 #endif
644 break;
645 default:
646 return (EINVAL);
647 }
648 if (en->auth_key_len != 0)
649 return (EINVAL);
650 switch (en->tls_vminor) {
651 case TLS_MINOR_VER_TWO:
652 if (en->iv_len != TLS_AEAD_GCM_LEN)
653 return (EINVAL);
654 break;
655 case TLS_MINOR_VER_THREE:
656 if (en->iv_len != TLS_1_3_GCM_IV_LEN)
657 return (EINVAL);
658 break;
659 default:
660 return (EINVAL);
661 }
662 break;
663 case CRYPTO_AES_CBC:
664 if (!ktls_cbc_enable)
665 return (EOPNOTSUPP);
666
667 switch (en->auth_algorithm) {
668 case CRYPTO_SHA1_HMAC:
669 break;
670 case CRYPTO_SHA2_256_HMAC:
671 case CRYPTO_SHA2_384_HMAC:
672 if (en->tls_vminor != TLS_MINOR_VER_TWO)
673 return (EINVAL);
674 break;
675 default:
676 return (EINVAL);
677 }
678 if (en->auth_key_len == 0)
679 return (EINVAL);
680
681 /*
682 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2
683 * use explicit IVs.
684 */
685 switch (en->tls_vminor) {
686 case TLS_MINOR_VER_ZERO:
687 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
688 return (EINVAL);
689 break;
690 case TLS_MINOR_VER_ONE:
691 case TLS_MINOR_VER_TWO:
692 /* Ignore any supplied IV. */
693 en->iv_len = 0;
694 break;
695 default:
696 return (EINVAL);
697 }
698 break;
699 case CRYPTO_CHACHA20_POLY1305:
700 if (en->auth_algorithm != 0 || en->auth_key_len != 0)
701 return (EINVAL);
702 if (en->tls_vminor != TLS_MINOR_VER_TWO &&
703 en->tls_vminor != TLS_MINOR_VER_THREE)
704 return (EINVAL);
705 if (en->iv_len != TLS_CHACHA20_IV_LEN)
706 return (EINVAL);
707 break;
708 default:
709 return (EINVAL);
710 }
711
712 error = ktls_start_kthreads();
713 if (error != 0)
714 return (error);
715
716 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
717
718 counter_u64_add(ktls_offload_active, 1);
719
720 refcount_init(&tls->refcount, 1);
721 if (direction == KTLS_RX) {
722 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
723 } else {
724 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
725 tls->inp = so->so_pcb;
726 in_pcbref(tls->inp);
727 tls->tx = true;
728 }
729
730 tls->wq_index = ktls_get_cpu(so);
731
732 tls->params.cipher_algorithm = en->cipher_algorithm;
733 tls->params.auth_algorithm = en->auth_algorithm;
734 tls->params.tls_vmajor = en->tls_vmajor;
735 tls->params.tls_vminor = en->tls_vminor;
736 tls->params.flags = en->flags;
737 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
738
739 /* Set the header and trailer lengths. */
740 tls->params.tls_hlen = sizeof(struct tls_record_layer);
741 switch (en->cipher_algorithm) {
742 case CRYPTO_AES_NIST_GCM_16:
743 /*
744 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
745 * nonce. TLS 1.3 uses a 12 byte implicit IV.
746 */
747 if (en->tls_vminor < TLS_MINOR_VER_THREE)
748 tls->params.tls_hlen += sizeof(uint64_t);
749 tls->params.tls_tlen = AES_GMAC_HASH_LEN;
750 tls->params.tls_bs = 1;
751 break;
752 case CRYPTO_AES_CBC:
753 switch (en->auth_algorithm) {
754 case CRYPTO_SHA1_HMAC:
755 if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
756 /* Implicit IV, no nonce. */
757 tls->sequential_records = true;
758 tls->next_seqno = be64dec(en->rec_seq);
759 STAILQ_INIT(&tls->pending_records);
760 } else {
761 tls->params.tls_hlen += AES_BLOCK_LEN;
762 }
763 tls->params.tls_tlen = AES_BLOCK_LEN +
764 SHA1_HASH_LEN;
765 break;
766 case CRYPTO_SHA2_256_HMAC:
767 tls->params.tls_hlen += AES_BLOCK_LEN;
768 tls->params.tls_tlen = AES_BLOCK_LEN +
769 SHA2_256_HASH_LEN;
770 break;
771 case CRYPTO_SHA2_384_HMAC:
772 tls->params.tls_hlen += AES_BLOCK_LEN;
773 tls->params.tls_tlen = AES_BLOCK_LEN +
774 SHA2_384_HASH_LEN;
775 break;
776 default:
777 panic("invalid hmac");
778 }
779 tls->params.tls_bs = AES_BLOCK_LEN;
780 break;
781 case CRYPTO_CHACHA20_POLY1305:
782 /*
783 * Chacha20 uses a 12 byte implicit IV.
784 */
785 tls->params.tls_tlen = POLY1305_HASH_LEN;
786 tls->params.tls_bs = 1;
787 break;
788 default:
789 panic("invalid cipher");
790 }
791
792 /*
793 * TLS 1.3 includes optional padding which we do not support,
794 * and also puts the "real" record type at the end of the
795 * encrypted data.
796 */
797 if (en->tls_vminor == TLS_MINOR_VER_THREE)
798 tls->params.tls_tlen += sizeof(uint8_t);
799
800 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
801 ("TLS header length too long: %d", tls->params.tls_hlen));
802 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
803 ("TLS trailer length too long: %d", tls->params.tls_tlen));
804
805 if (en->auth_key_len != 0) {
806 tls->params.auth_key_len = en->auth_key_len;
807 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
808 M_WAITOK);
809 bcopy(en->auth_key, tls->params.auth_key, en->auth_key_len);
810 }
811
812 tls->params.cipher_key_len = en->cipher_key_len;
813 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
814 bcopy(en->cipher_key, tls->params.cipher_key, en->cipher_key_len);
815
816 /*
817 * This holds the implicit portion of the nonce for AEAD
818 * ciphers and the initial implicit IV for TLS 1.0. The
819 * explicit portions of the IV are generated in ktls_frame().
820 */
821 if (en->iv_len != 0) {
822 tls->params.iv_len = en->iv_len;
823 bcopy(en->iv, tls->params.iv, en->iv_len);
824
825 /*
826 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
827 * counter to generate unique explicit IVs.
828 *
829 * Store this counter in the last 8 bytes of the IV
830 * array so that it is 8-byte aligned.
831 */
832 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
833 en->tls_vminor == TLS_MINOR_VER_TWO)
834 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
835 }
836
837 tls->gen = 0;
838 *tlsp = tls;
839 return (0);
840 }
841
842 static struct ktls_session *
ktls_clone_session(struct ktls_session * tls,int direction)843 ktls_clone_session(struct ktls_session *tls, int direction)
844 {
845 struct ktls_session *tls_new;
846
847 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
848
849 counter_u64_add(ktls_offload_active, 1);
850
851 refcount_init(&tls_new->refcount, 1);
852 if (direction == KTLS_RX) {
853 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
854 tls_new);
855 } else {
856 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
857 tls_new);
858 tls_new->inp = tls->inp;
859 tls_new->tx = true;
860 in_pcbref(tls_new->inp);
861 }
862
863 /* Copy fields from existing session. */
864 tls_new->params = tls->params;
865 tls_new->wq_index = tls->wq_index;
866
867 /* Deep copy keys. */
868 if (tls_new->params.auth_key != NULL) {
869 tls_new->params.auth_key = malloc(tls->params.auth_key_len,
870 M_KTLS, M_WAITOK);
871 memcpy(tls_new->params.auth_key, tls->params.auth_key,
872 tls->params.auth_key_len);
873 }
874
875 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
876 M_WAITOK);
877 memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
878 tls->params.cipher_key_len);
879
880 tls_new->gen = 0;
881 return (tls_new);
882 }
883
884 #ifdef TCP_OFFLOAD
885 static int
ktls_try_toe(struct socket * so,struct ktls_session * tls,int direction)886 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
887 {
888 struct inpcb *inp = sotoinpcb(so);
889 struct tcpcb *tp = intotcpcb(inp);
890 int error;
891
892 INP_WLOCK(inp);
893 if (tp->t_flags & TF_DISCONNECTED) {
894 INP_WUNLOCK(inp);
895 return (ECONNRESET);
896 }
897 if (!(tp->t_flags & TF_TOE)) {
898 INP_WUNLOCK(inp);
899 return (EOPNOTSUPP);
900 }
901
902 error = tcp_offload_alloc_tls_session(tp, tls, direction);
903 INP_WUNLOCK(inp);
904 if (error == 0) {
905 tls->mode = TCP_TLS_MODE_TOE;
906 switch (tls->params.cipher_algorithm) {
907 case CRYPTO_AES_CBC:
908 counter_u64_add(ktls_toe_cbc, 1);
909 break;
910 case CRYPTO_AES_NIST_GCM_16:
911 counter_u64_add(ktls_toe_gcm, 1);
912 break;
913 case CRYPTO_CHACHA20_POLY1305:
914 counter_u64_add(ktls_toe_chacha20, 1);
915 break;
916 }
917 }
918 return (error);
919 }
920 #endif
921
922 /*
923 * Common code used when first enabling ifnet TLS on a connection or
924 * when allocating a new ifnet TLS session due to a routing change.
925 * This function allocates a new TLS send tag on whatever interface
926 * the connection is currently routed over.
927 */
928 static int
ktls_alloc_snd_tag(struct inpcb * inp,struct ktls_session * tls,bool force,struct m_snd_tag ** mstp)929 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
930 struct m_snd_tag **mstp)
931 {
932 union if_snd_tag_alloc_params params;
933 struct ifnet *ifp;
934 struct nhop_object *nh;
935 struct tcpcb *tp = intotcpcb(inp);
936 int error;
937
938 INP_RLOCK(inp);
939 if (tp->t_flags & TF_DISCONNECTED) {
940 INP_RUNLOCK(inp);
941 return (ECONNRESET);
942 }
943
944 /*
945 * Check administrative controls on ifnet TLS to determine if
946 * ifnet TLS should be denied.
947 *
948 * - Always permit 'force' requests.
949 * - ktls_ifnet_permitted == 0: always deny.
950 */
951 if (!force && ktls_ifnet_permitted == 0) {
952 INP_RUNLOCK(inp);
953 return (ENXIO);
954 }
955
956 /*
957 * XXX: Use the cached route in the inpcb to find the
958 * interface. This should perhaps instead use
959 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only
960 * enabled after a connection has completed key negotiation in
961 * userland, the cached route will be present in practice.
962 */
963 nh = inp->inp_route.ro_nh;
964 if (nh == NULL) {
965 INP_RUNLOCK(inp);
966 return (ENXIO);
967 }
968 ifp = nh->nh_ifp;
969 if_ref(ifp);
970
971 /*
972 * Allocate a TLS + ratelimit tag if the connection has an
973 * existing pacing rate.
974 */
975 if (tp->t_pacing_rate != -1 &&
976 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
977 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
978 params.tls_rate_limit.inp = inp;
979 params.tls_rate_limit.tls = tls;
980 params.tls_rate_limit.max_rate = tp->t_pacing_rate;
981 } else {
982 params.hdr.type = IF_SND_TAG_TYPE_TLS;
983 params.tls.inp = inp;
984 params.tls.tls = tls;
985 }
986 params.hdr.flowid = inp->inp_flowid;
987 params.hdr.flowtype = inp->inp_flowtype;
988 params.hdr.numa_domain = inp->inp_numa_domain;
989 INP_RUNLOCK(inp);
990
991 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
992 error = EOPNOTSUPP;
993 goto out;
994 }
995 if (inp->inp_vflag & INP_IPV6) {
996 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
997 error = EOPNOTSUPP;
998 goto out;
999 }
1000 } else {
1001 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
1002 error = EOPNOTSUPP;
1003 goto out;
1004 }
1005 }
1006 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1007 out:
1008 if_rele(ifp);
1009 return (error);
1010 }
1011
1012 /*
1013 * Allocate an initial TLS receive tag for doing HW decryption of TLS
1014 * data.
1015 *
1016 * This function allocates a new TLS receive tag on whatever interface
1017 * the connection is currently routed over. If the connection ends up
1018 * using a different interface for receive this will get fixed up via
1019 * ktls_input_ifp_mismatch as future packets arrive.
1020 */
1021 static int
ktls_alloc_rcv_tag(struct inpcb * inp,struct ktls_session * tls,struct m_snd_tag ** mstp)1022 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
1023 struct m_snd_tag **mstp)
1024 {
1025 union if_snd_tag_alloc_params params;
1026 struct ifnet *ifp;
1027 struct nhop_object *nh;
1028 int error;
1029
1030 if (!ktls_ocf_recrypt_supported(tls))
1031 return (ENXIO);
1032
1033 INP_RLOCK(inp);
1034 if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
1035 INP_RUNLOCK(inp);
1036 return (ECONNRESET);
1037 }
1038
1039 /*
1040 * Check administrative controls on ifnet TLS to determine if
1041 * ifnet TLS should be denied.
1042 */
1043 if (ktls_ifnet_permitted == 0) {
1044 INP_RUNLOCK(inp);
1045 return (ENXIO);
1046 }
1047
1048 /*
1049 * XXX: As with ktls_alloc_snd_tag, use the cached route in
1050 * the inpcb to find the interface.
1051 */
1052 nh = inp->inp_route.ro_nh;
1053 if (nh == NULL) {
1054 INP_RUNLOCK(inp);
1055 return (ENXIO);
1056 }
1057 ifp = nh->nh_ifp;
1058 if_ref(ifp);
1059 tls->rx_ifp = ifp;
1060
1061 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1062 params.hdr.flowid = inp->inp_flowid;
1063 params.hdr.flowtype = inp->inp_flowtype;
1064 params.hdr.numa_domain = inp->inp_numa_domain;
1065 params.tls_rx.inp = inp;
1066 params.tls_rx.tls = tls;
1067 params.tls_rx.vlan_id = 0;
1068
1069 INP_RUNLOCK(inp);
1070
1071 if (inp->inp_vflag & INP_IPV6) {
1072 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
1073 error = EOPNOTSUPP;
1074 goto out;
1075 }
1076 } else {
1077 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
1078 error = EOPNOTSUPP;
1079 goto out;
1080 }
1081 }
1082 error = m_snd_tag_alloc(ifp, ¶ms, mstp);
1083
1084 /*
1085 * If this connection is over a vlan, vlan_snd_tag_alloc
1086 * rewrites vlan_id with the saved interface. Save the VLAN
1087 * ID for use in ktls_reset_receive_tag which allocates new
1088 * receive tags directly from the leaf interface bypassing
1089 * if_vlan.
1090 */
1091 if (error == 0)
1092 tls->rx_vlan_id = params.tls_rx.vlan_id;
1093 out:
1094 return (error);
1095 }
1096
1097 static int
ktls_try_ifnet(struct socket * so,struct ktls_session * tls,int direction,bool force)1098 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1099 bool force)
1100 {
1101 struct m_snd_tag *mst;
1102 int error;
1103
1104 switch (direction) {
1105 case KTLS_TX:
1106 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1107 if (__predict_false(error != 0))
1108 goto done;
1109 break;
1110 case KTLS_RX:
1111 KASSERT(!force, ("%s: forced receive tag", __func__));
1112 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1113 if (__predict_false(error != 0))
1114 goto done;
1115 break;
1116 default:
1117 __assert_unreachable();
1118 }
1119
1120 tls->mode = TCP_TLS_MODE_IFNET;
1121 tls->snd_tag = mst;
1122
1123 switch (tls->params.cipher_algorithm) {
1124 case CRYPTO_AES_CBC:
1125 counter_u64_add(ktls_ifnet_cbc, 1);
1126 break;
1127 case CRYPTO_AES_NIST_GCM_16:
1128 counter_u64_add(ktls_ifnet_gcm, 1);
1129 break;
1130 case CRYPTO_CHACHA20_POLY1305:
1131 counter_u64_add(ktls_ifnet_chacha20, 1);
1132 break;
1133 default:
1134 break;
1135 }
1136 done:
1137 return (error);
1138 }
1139
1140 static void
ktls_use_sw(struct ktls_session * tls)1141 ktls_use_sw(struct ktls_session *tls)
1142 {
1143 tls->mode = TCP_TLS_MODE_SW;
1144 switch (tls->params.cipher_algorithm) {
1145 case CRYPTO_AES_CBC:
1146 counter_u64_add(ktls_sw_cbc, 1);
1147 break;
1148 case CRYPTO_AES_NIST_GCM_16:
1149 counter_u64_add(ktls_sw_gcm, 1);
1150 break;
1151 case CRYPTO_CHACHA20_POLY1305:
1152 counter_u64_add(ktls_sw_chacha20, 1);
1153 break;
1154 }
1155 }
1156
1157 static int
ktls_try_sw(struct ktls_session * tls,int direction)1158 ktls_try_sw(struct ktls_session *tls, int direction)
1159 {
1160 int error;
1161
1162 error = ktls_ocf_try(tls, direction);
1163 if (error)
1164 return (error);
1165 ktls_use_sw(tls);
1166 return (0);
1167 }
1168
1169 /*
1170 * KTLS RX stores data in the socket buffer as a list of TLS records,
1171 * where each record is stored as a control message containg the TLS
1172 * header followed by data mbufs containing the decrypted data. This
1173 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1174 * both encrypted and decrypted data. TLS records decrypted by a NIC
1175 * should be queued to the socket buffer as records, but encrypted
1176 * data which needs to be decrypted by software arrives as a stream of
1177 * regular mbufs which need to be converted. In addition, there may
1178 * already be pending encrypted data in the socket buffer when KTLS RX
1179 * is enabled.
1180 *
1181 * To manage not-yet-decrypted data for KTLS RX, the following scheme
1182 * is used:
1183 *
1184 * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1185 *
1186 * - ktls_check_rx checks this chain of mbufs reading the TLS header
1187 * from the first mbuf. Once all of the data for that TLS record is
1188 * queued, the socket is queued to a worker thread.
1189 *
1190 * - The worker thread calls ktls_decrypt to decrypt TLS records in
1191 * the TLS chain. Each TLS record is detached from the TLS chain,
1192 * decrypted, and inserted into the regular socket buffer chain as
1193 * record starting with a control message holding the TLS header and
1194 * a chain of mbufs holding the encrypted data.
1195 */
1196
1197 static void
sb_mark_notready(struct sockbuf * sb)1198 sb_mark_notready(struct sockbuf *sb)
1199 {
1200 struct mbuf *m;
1201
1202 m = sb->sb_mb;
1203 sb->sb_mtls = m;
1204 sb->sb_mb = NULL;
1205 sb->sb_mbtail = NULL;
1206 sb->sb_lastrecord = NULL;
1207 for (; m != NULL; m = m->m_next) {
1208 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1209 __func__));
1210 KASSERT((m->m_flags & M_NOTREADY) == 0, ("%s: mbuf not ready",
1211 __func__));
1212 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1213 __func__));
1214 m->m_flags |= M_NOTREADY;
1215 sb->sb_acc -= m->m_len;
1216 sb->sb_tlscc += m->m_len;
1217 sb->sb_mtlstail = m;
1218 }
1219 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1220 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1221 sb->sb_ccc));
1222 }
1223
1224 /*
1225 * Return information about the pending TLS data in a socket
1226 * buffer. On return, 'seqno' is set to the sequence number
1227 * of the next TLS record to be received, 'resid' is set to
1228 * the amount of bytes still needed for the last pending
1229 * record. The function returns 'false' if the last pending
1230 * record contains a partial TLS header. In that case, 'resid'
1231 * is the number of bytes needed to complete the TLS header.
1232 */
1233 bool
ktls_pending_rx_info(struct sockbuf * sb,uint64_t * seqnop,size_t * residp)1234 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1235 {
1236 struct tls_record_layer hdr;
1237 struct mbuf *m;
1238 uint64_t seqno;
1239 size_t resid;
1240 u_int offset, record_len;
1241
1242 SOCKBUF_LOCK_ASSERT(sb);
1243 MPASS(sb->sb_flags & SB_TLS_RX);
1244 seqno = sb->sb_tls_seqno;
1245 resid = sb->sb_tlscc;
1246 m = sb->sb_mtls;
1247 offset = 0;
1248
1249 if (resid == 0) {
1250 *seqnop = seqno;
1251 *residp = 0;
1252 return (true);
1253 }
1254
1255 for (;;) {
1256 seqno++;
1257
1258 if (resid < sizeof(hdr)) {
1259 *seqnop = seqno;
1260 *residp = sizeof(hdr) - resid;
1261 return (false);
1262 }
1263
1264 m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1265
1266 record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1267 if (resid <= record_len) {
1268 *seqnop = seqno;
1269 *residp = record_len - resid;
1270 return (true);
1271 }
1272 resid -= record_len;
1273
1274 while (record_len != 0) {
1275 if (m->m_len - offset > record_len) {
1276 offset += record_len;
1277 break;
1278 }
1279
1280 record_len -= (m->m_len - offset);
1281 offset = 0;
1282 m = m->m_next;
1283 }
1284 }
1285 }
1286
1287 int
ktls_enable_rx(struct socket * so,struct tls_enable * en)1288 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1289 {
1290 struct ktls_session *tls;
1291 int error;
1292
1293 if (!ktls_offload_enable || !ktls_rx_offload_enable)
1294 return (ENOTSUP);
1295
1296 counter_u64_add(ktls_offload_enable_calls, 1);
1297
1298 /*
1299 * This should always be true since only the TCP socket option
1300 * invokes this function.
1301 */
1302 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1303 return (EINVAL);
1304
1305 /*
1306 * XXX: Don't overwrite existing sessions. We should permit
1307 * this to support rekeying in the future.
1308 */
1309 if (so->so_rcv.sb_tls_info != NULL)
1310 return (EALREADY);
1311
1312 error = ktls_create_session(so, en, &tls, KTLS_RX);
1313 if (error)
1314 return (error);
1315
1316 error = ktls_ocf_try(tls, KTLS_RX);
1317 if (error) {
1318 ktls_free(tls);
1319 return (error);
1320 }
1321
1322 /*
1323 * Serialize with soreceive_generic() and make sure that we're not
1324 * operating on a listening socket.
1325 */
1326 error = SOCK_IO_RECV_LOCK(so, SBL_WAIT);
1327 if (error) {
1328 ktls_free(tls);
1329 return (error);
1330 }
1331
1332 /* Mark the socket as using TLS offload. */
1333 SOCK_RECVBUF_LOCK(so);
1334 if (__predict_false(so->so_rcv.sb_tls_info != NULL))
1335 error = EALREADY;
1336 else if ((so->so_rcv.sb_flags & SB_SPLICED) != 0)
1337 error = EINVAL;
1338 if (error != 0) {
1339 SOCK_RECVBUF_UNLOCK(so);
1340 SOCK_IO_RECV_UNLOCK(so);
1341 ktls_free(tls);
1342 return (EALREADY);
1343 }
1344 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1345 so->so_rcv.sb_tls_info = tls;
1346 so->so_rcv.sb_flags |= SB_TLS_RX;
1347
1348 /* Mark existing data as not ready until it can be decrypted. */
1349 sb_mark_notready(&so->so_rcv);
1350 ktls_check_rx(&so->so_rcv);
1351 SOCK_RECVBUF_UNLOCK(so);
1352 SOCK_IO_RECV_UNLOCK(so);
1353
1354 /* Prefer TOE -> ifnet TLS -> software TLS. */
1355 #ifdef TCP_OFFLOAD
1356 error = ktls_try_toe(so, tls, KTLS_RX);
1357 if (error)
1358 #endif
1359 error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1360 if (error)
1361 ktls_use_sw(tls);
1362
1363 counter_u64_add(ktls_offload_total, 1);
1364
1365 return (0);
1366 }
1367
1368 int
ktls_enable_tx(struct socket * so,struct tls_enable * en)1369 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1370 {
1371 struct ktls_session *tls;
1372 struct inpcb *inp;
1373 struct tcpcb *tp;
1374 int error;
1375
1376 if (!ktls_offload_enable)
1377 return (ENOTSUP);
1378
1379 counter_u64_add(ktls_offload_enable_calls, 1);
1380
1381 /*
1382 * This should always be true since only the TCP socket option
1383 * invokes this function.
1384 */
1385 if (so->so_proto->pr_protocol != IPPROTO_TCP)
1386 return (EINVAL);
1387
1388 /*
1389 * XXX: Don't overwrite existing sessions. We should permit
1390 * this to support rekeying in the future.
1391 */
1392 if (so->so_snd.sb_tls_info != NULL)
1393 return (EALREADY);
1394
1395 /* TLS requires ext pgs */
1396 if (mb_use_ext_pgs == 0)
1397 return (ENXIO);
1398
1399 error = ktls_create_session(so, en, &tls, KTLS_TX);
1400 if (error)
1401 return (error);
1402
1403 /* some ktls offload NICs require initial seqno to start offload */
1404 tls->initial_offload_seqno = be64dec(en->rec_seq);
1405
1406 /* Prefer TOE -> ifnet TLS -> software TLS. */
1407 #ifdef TCP_OFFLOAD
1408 error = ktls_try_toe(so, tls, KTLS_TX);
1409 if (error)
1410 #endif
1411 error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1412 if (error)
1413 error = ktls_try_sw(tls, KTLS_TX);
1414
1415 if (error) {
1416 ktls_free(tls);
1417 return (error);
1418 }
1419
1420 /*
1421 * Serialize with sosend_generic() and make sure that we're not
1422 * operating on a listening socket.
1423 */
1424 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1425 if (error) {
1426 ktls_free(tls);
1427 return (error);
1428 }
1429
1430 /*
1431 * Write lock the INP when setting sb_tls_info so that
1432 * routines in tcp_ratelimit.c can read sb_tls_info while
1433 * holding the INP lock.
1434 */
1435 inp = so->so_pcb;
1436 INP_WLOCK(inp);
1437 SOCK_SENDBUF_LOCK(so);
1438 if (__predict_false(so->so_snd.sb_tls_info != NULL))
1439 error = EALREADY;
1440 else if ((so->so_snd.sb_flags & SB_SPLICED) != 0)
1441 error = EINVAL;
1442 if (error != 0) {
1443 SOCK_SENDBUF_UNLOCK(so);
1444 INP_WUNLOCK(inp);
1445 SOCK_IO_SEND_UNLOCK(so);
1446 ktls_free(tls);
1447 return (error);
1448 }
1449 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1450 so->so_snd.sb_tls_info = tls;
1451 if (tls->mode != TCP_TLS_MODE_SW) {
1452 tp = intotcpcb(inp);
1453 MPASS(tp->t_nic_ktls_xmit == 0);
1454 tp->t_nic_ktls_xmit = 1;
1455 if (tp->t_fb->tfb_hwtls_change != NULL)
1456 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1457 }
1458 SOCK_SENDBUF_UNLOCK(so);
1459 INP_WUNLOCK(inp);
1460 SOCK_IO_SEND_UNLOCK(so);
1461
1462 counter_u64_add(ktls_offload_total, 1);
1463
1464 return (0);
1465 }
1466
1467 int
ktls_get_rx_mode(struct socket * so,int * modep)1468 ktls_get_rx_mode(struct socket *so, int *modep)
1469 {
1470 struct ktls_session *tls;
1471 struct inpcb *inp __diagused;
1472
1473 if (SOLISTENING(so))
1474 return (EINVAL);
1475 inp = so->so_pcb;
1476 INP_WLOCK_ASSERT(inp);
1477 SOCK_RECVBUF_LOCK(so);
1478 tls = so->so_rcv.sb_tls_info;
1479 if (tls == NULL)
1480 *modep = TCP_TLS_MODE_NONE;
1481 else
1482 *modep = tls->mode;
1483 SOCK_RECVBUF_UNLOCK(so);
1484 return (0);
1485 }
1486
1487 /*
1488 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1489 *
1490 * This function gets information about the next TCP- and TLS-
1491 * sequence number to be processed by the TLS receive worker
1492 * thread. The information is extracted from the given "inpcb"
1493 * structure. The values are stored in host endian format at the two
1494 * given output pointer locations. The TCP sequence number points to
1495 * the beginning of the TLS header.
1496 *
1497 * This function returns zero on success, else a non-zero error code
1498 * is returned.
1499 */
1500 int
ktls_get_rx_sequence(struct inpcb * inp,uint32_t * tcpseq,uint64_t * tlsseq)1501 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1502 {
1503 struct socket *so = inp->inp_socket;
1504 struct tcpcb *tp = intotcpcb(inp);
1505
1506 INP_RLOCK(inp);
1507 if (tp->t_flags & TF_DISCONNECTED) {
1508 INP_RUNLOCK(inp);
1509 return (ECONNRESET);
1510 }
1511
1512 SOCKBUF_LOCK(&so->so_rcv);
1513 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1514 *tlsseq = so->so_rcv.sb_tls_seqno;
1515 SOCKBUF_UNLOCK(&so->so_rcv);
1516
1517 INP_RUNLOCK(inp);
1518
1519 return (0);
1520 }
1521
1522 int
ktls_get_tx_mode(struct socket * so,int * modep)1523 ktls_get_tx_mode(struct socket *so, int *modep)
1524 {
1525 struct ktls_session *tls;
1526 struct inpcb *inp __diagused;
1527
1528 if (SOLISTENING(so))
1529 return (EINVAL);
1530 inp = so->so_pcb;
1531 INP_WLOCK_ASSERT(inp);
1532 SOCK_SENDBUF_LOCK(so);
1533 tls = so->so_snd.sb_tls_info;
1534 if (tls == NULL)
1535 *modep = TCP_TLS_MODE_NONE;
1536 else
1537 *modep = tls->mode;
1538 SOCK_SENDBUF_UNLOCK(so);
1539 return (0);
1540 }
1541
1542 /*
1543 * Switch between SW and ifnet TLS sessions as requested.
1544 */
1545 int
ktls_set_tx_mode(struct socket * so,int mode)1546 ktls_set_tx_mode(struct socket *so, int mode)
1547 {
1548 struct ktls_session *tls, *tls_new;
1549 struct inpcb *inp;
1550 struct tcpcb *tp;
1551 int error;
1552
1553 if (SOLISTENING(so))
1554 return (EINVAL);
1555 switch (mode) {
1556 case TCP_TLS_MODE_SW:
1557 case TCP_TLS_MODE_IFNET:
1558 break;
1559 default:
1560 return (EINVAL);
1561 }
1562
1563 inp = so->so_pcb;
1564 INP_WLOCK_ASSERT(inp);
1565 tp = intotcpcb(inp);
1566
1567 if (mode == TCP_TLS_MODE_IFNET) {
1568 /* Don't allow enabling ifnet ktls multiple times */
1569 if (tp->t_nic_ktls_xmit)
1570 return (EALREADY);
1571
1572 /*
1573 * Don't enable ifnet ktls if we disabled it due to an
1574 * excessive retransmission rate
1575 */
1576 if (tp->t_nic_ktls_xmit_dis)
1577 return (ENXIO);
1578 }
1579
1580 SOCKBUF_LOCK(&so->so_snd);
1581 tls = so->so_snd.sb_tls_info;
1582 if (tls == NULL) {
1583 SOCKBUF_UNLOCK(&so->so_snd);
1584 return (0);
1585 }
1586
1587 if (tls->mode == mode) {
1588 SOCKBUF_UNLOCK(&so->so_snd);
1589 return (0);
1590 }
1591
1592 tls = ktls_hold(tls);
1593 SOCKBUF_UNLOCK(&so->so_snd);
1594 INP_WUNLOCK(inp);
1595
1596 tls_new = ktls_clone_session(tls, KTLS_TX);
1597
1598 if (mode == TCP_TLS_MODE_IFNET)
1599 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1600 else
1601 error = ktls_try_sw(tls_new, KTLS_TX);
1602 if (error) {
1603 counter_u64_add(ktls_switch_failed, 1);
1604 ktls_free(tls_new);
1605 ktls_free(tls);
1606 INP_WLOCK(inp);
1607 return (error);
1608 }
1609
1610 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1611 if (error) {
1612 counter_u64_add(ktls_switch_failed, 1);
1613 ktls_free(tls_new);
1614 ktls_free(tls);
1615 INP_WLOCK(inp);
1616 return (error);
1617 }
1618
1619 /*
1620 * If we raced with another session change, keep the existing
1621 * session.
1622 */
1623 if (tls != so->so_snd.sb_tls_info) {
1624 counter_u64_add(ktls_switch_failed, 1);
1625 SOCK_IO_SEND_UNLOCK(so);
1626 ktls_free(tls_new);
1627 ktls_free(tls);
1628 INP_WLOCK(inp);
1629 return (EBUSY);
1630 }
1631
1632 INP_WLOCK(inp);
1633 SOCKBUF_LOCK(&so->so_snd);
1634 so->so_snd.sb_tls_info = tls_new;
1635 if (tls_new->mode != TCP_TLS_MODE_SW) {
1636 MPASS(tp->t_nic_ktls_xmit == 0);
1637 tp->t_nic_ktls_xmit = 1;
1638 if (tp->t_fb->tfb_hwtls_change != NULL)
1639 (*tp->t_fb->tfb_hwtls_change)(tp, 1);
1640 }
1641 SOCKBUF_UNLOCK(&so->so_snd);
1642 SOCK_IO_SEND_UNLOCK(so);
1643
1644 /*
1645 * Drop two references on 'tls'. The first is for the
1646 * ktls_hold() above. The second drops the reference from the
1647 * socket buffer.
1648 */
1649 KASSERT(tls->refcount >= 2, ("too few references on old session"));
1650 ktls_free(tls);
1651 ktls_free(tls);
1652
1653 if (mode == TCP_TLS_MODE_IFNET)
1654 counter_u64_add(ktls_switch_to_ifnet, 1);
1655 else
1656 counter_u64_add(ktls_switch_to_sw, 1);
1657
1658 return (0);
1659 }
1660
1661 /*
1662 * Try to allocate a new TLS receive tag. This task is scheduled when
1663 * sbappend_ktls_rx detects an input path change. If a new tag is
1664 * allocated, replace the tag in the TLS session. If a new tag cannot
1665 * be allocated, let the session fall back to software decryption.
1666 */
1667 static void
ktls_reset_receive_tag(void * context,int pending)1668 ktls_reset_receive_tag(void *context, int pending)
1669 {
1670 union if_snd_tag_alloc_params params;
1671 struct ktls_session *tls;
1672 struct m_snd_tag *mst;
1673 struct inpcb *inp;
1674 struct ifnet *ifp;
1675 struct socket *so;
1676 int error;
1677
1678 MPASS(pending == 1);
1679
1680 tls = context;
1681 so = tls->so;
1682 inp = so->so_pcb;
1683 ifp = NULL;
1684
1685 INP_RLOCK(inp);
1686 if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
1687 INP_RUNLOCK(inp);
1688 goto out;
1689 }
1690
1691 SOCKBUF_LOCK(&so->so_rcv);
1692 mst = tls->snd_tag;
1693 tls->snd_tag = NULL;
1694 if (mst != NULL)
1695 m_snd_tag_rele(mst);
1696
1697 ifp = tls->rx_ifp;
1698 if_ref(ifp);
1699 SOCKBUF_UNLOCK(&so->so_rcv);
1700
1701 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1702 params.hdr.flowid = inp->inp_flowid;
1703 params.hdr.flowtype = inp->inp_flowtype;
1704 params.hdr.numa_domain = inp->inp_numa_domain;
1705 params.tls_rx.inp = inp;
1706 params.tls_rx.tls = tls;
1707 params.tls_rx.vlan_id = tls->rx_vlan_id;
1708 INP_RUNLOCK(inp);
1709
1710 if (inp->inp_vflag & INP_IPV6) {
1711 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1712 goto out;
1713 } else {
1714 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1715 goto out;
1716 }
1717
1718 error = m_snd_tag_alloc(ifp, ¶ms, &mst);
1719 if (error == 0) {
1720 SOCKBUF_LOCK(&so->so_rcv);
1721 tls->snd_tag = mst;
1722 SOCKBUF_UNLOCK(&so->so_rcv);
1723
1724 counter_u64_add(ktls_ifnet_reset, 1);
1725 } else {
1726 /*
1727 * Just fall back to software decryption if a tag
1728 * cannot be allocated leaving the connection intact.
1729 * If a future input path change switches to another
1730 * interface this connection will resume ifnet TLS.
1731 */
1732 counter_u64_add(ktls_ifnet_reset_failed, 1);
1733 }
1734
1735 out:
1736 mtx_pool_lock(mtxpool_sleep, tls);
1737 tls->reset_pending = false;
1738 mtx_pool_unlock(mtxpool_sleep, tls);
1739
1740 if (ifp != NULL)
1741 if_rele(ifp);
1742 CURVNET_SET(so->so_vnet);
1743 sorele(so);
1744 CURVNET_RESTORE();
1745 ktls_free(tls);
1746 }
1747
1748 /*
1749 * Try to allocate a new TLS send tag. This task is scheduled when
1750 * ip_output detects a route change while trying to transmit a packet
1751 * holding a TLS record. If a new tag is allocated, replace the tag
1752 * in the TLS session. Subsequent packets on the connection will use
1753 * the new tag. If a new tag cannot be allocated, drop the
1754 * connection.
1755 */
1756 static void
ktls_reset_send_tag(void * context,int pending)1757 ktls_reset_send_tag(void *context, int pending)
1758 {
1759 struct epoch_tracker et;
1760 struct ktls_session *tls;
1761 struct m_snd_tag *old, *new;
1762 struct inpcb *inp;
1763 struct tcpcb *tp;
1764 int error;
1765
1766 MPASS(pending == 1);
1767
1768 tls = context;
1769 inp = tls->inp;
1770
1771 /*
1772 * Free the old tag first before allocating a new one.
1773 * ip[6]_output_send() will treat a NULL send tag the same as
1774 * an ifp mismatch and drop packets until a new tag is
1775 * allocated.
1776 *
1777 * Write-lock the INP when changing tls->snd_tag since
1778 * ip[6]_output_send() holds a read-lock when reading the
1779 * pointer.
1780 */
1781 INP_WLOCK(inp);
1782 old = tls->snd_tag;
1783 tls->snd_tag = NULL;
1784 INP_WUNLOCK(inp);
1785 if (old != NULL)
1786 m_snd_tag_rele(old);
1787
1788 error = ktls_alloc_snd_tag(inp, tls, true, &new);
1789
1790 if (error == 0) {
1791 INP_WLOCK(inp);
1792 tls->snd_tag = new;
1793 mtx_pool_lock(mtxpool_sleep, tls);
1794 tls->reset_pending = false;
1795 mtx_pool_unlock(mtxpool_sleep, tls);
1796 INP_WUNLOCK(inp);
1797
1798 counter_u64_add(ktls_ifnet_reset, 1);
1799
1800 /*
1801 * XXX: Should we kick tcp_output explicitly now that
1802 * the send tag is fixed or just rely on timers?
1803 */
1804 } else {
1805 NET_EPOCH_ENTER(et);
1806 INP_WLOCK(inp);
1807 tp = intotcpcb(inp);
1808 if (!(tp->t_flags & TF_DISCONNECTED)) {
1809 CURVNET_SET(inp->inp_socket->so_vnet);
1810 tp = tcp_drop(tp, ECONNABORTED);
1811 CURVNET_RESTORE();
1812 if (tp != NULL) {
1813 counter_u64_add(ktls_ifnet_reset_dropped, 1);
1814 INP_WUNLOCK(inp);
1815 }
1816 } else
1817 INP_WUNLOCK(inp);
1818 NET_EPOCH_EXIT(et);
1819
1820 counter_u64_add(ktls_ifnet_reset_failed, 1);
1821
1822 /*
1823 * Leave reset_pending true to avoid future tasks while
1824 * the socket goes away.
1825 */
1826 }
1827
1828 ktls_free(tls);
1829 }
1830
1831 void
ktls_input_ifp_mismatch(struct sockbuf * sb,struct ifnet * ifp)1832 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1833 {
1834 struct ktls_session *tls;
1835 struct socket *so;
1836
1837 SOCKBUF_LOCK_ASSERT(sb);
1838 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1839 __func__, sb));
1840 so = __containerof(sb, struct socket, so_rcv);
1841
1842 tls = sb->sb_tls_info;
1843 if_rele(tls->rx_ifp);
1844 if_ref(ifp);
1845 tls->rx_ifp = ifp;
1846
1847 /*
1848 * See if we should schedule a task to update the receive tag for
1849 * this session.
1850 */
1851 mtx_pool_lock(mtxpool_sleep, tls);
1852 if (!tls->reset_pending) {
1853 (void) ktls_hold(tls);
1854 soref(so);
1855 tls->so = so;
1856 tls->reset_pending = true;
1857 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1858 }
1859 mtx_pool_unlock(mtxpool_sleep, tls);
1860 }
1861
1862 int
ktls_output_eagain(struct inpcb * inp,struct ktls_session * tls)1863 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1864 {
1865
1866 if (inp == NULL)
1867 return (ENOBUFS);
1868
1869 INP_LOCK_ASSERT(inp);
1870
1871 /*
1872 * See if we should schedule a task to update the send tag for
1873 * this session.
1874 */
1875 mtx_pool_lock(mtxpool_sleep, tls);
1876 if (!tls->reset_pending) {
1877 (void) ktls_hold(tls);
1878 tls->reset_pending = true;
1879 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1880 }
1881 mtx_pool_unlock(mtxpool_sleep, tls);
1882 return (ENOBUFS);
1883 }
1884
1885 #ifdef RATELIMIT
1886 int
ktls_modify_txrtlmt(struct ktls_session * tls,uint64_t max_pacing_rate)1887 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1888 {
1889 union if_snd_tag_modify_params params = {
1890 .rate_limit.max_rate = max_pacing_rate,
1891 .rate_limit.flags = M_NOWAIT,
1892 };
1893 struct m_snd_tag *mst;
1894
1895 /* Can't get to the inp, but it should be locked. */
1896 /* INP_LOCK_ASSERT(inp); */
1897
1898 MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1899
1900 if (tls->snd_tag == NULL) {
1901 /*
1902 * Resetting send tag, ignore this change. The
1903 * pending reset may or may not see this updated rate
1904 * in the tcpcb. If it doesn't, we will just lose
1905 * this rate change.
1906 */
1907 return (0);
1908 }
1909
1910 mst = tls->snd_tag;
1911
1912 MPASS(mst != NULL);
1913 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1914
1915 return (mst->sw->snd_tag_modify(mst, ¶ms));
1916 }
1917 #endif
1918
1919 static void
ktls_destroy_help(void * context,int pending __unused)1920 ktls_destroy_help(void *context, int pending __unused)
1921 {
1922 ktls_destroy(context);
1923 }
1924
1925 void
ktls_destroy(struct ktls_session * tls)1926 ktls_destroy(struct ktls_session *tls)
1927 {
1928 struct inpcb *inp;
1929 struct tcpcb *tp;
1930 bool wlocked;
1931
1932 MPASS(tls->refcount == 0);
1933
1934 inp = tls->inp;
1935 if (tls->tx) {
1936 wlocked = INP_WLOCKED(inp);
1937 if (!wlocked && !INP_TRY_WLOCK(inp)) {
1938 /*
1939 * rwlocks read locks are anonymous, and there
1940 * is no way to know if our current thread
1941 * holds an rlock on the inp. As a rough
1942 * estimate, check to see if the thread holds
1943 * *any* rlocks at all. If it does not, then we
1944 * know that we don't hold the inp rlock, and
1945 * can safely take the wlock
1946 */
1947 if (curthread->td_rw_rlocks == 0) {
1948 INP_WLOCK(inp);
1949 } else {
1950 /*
1951 * We might hold the rlock, so let's
1952 * do the destroy in a taskqueue
1953 * context to avoid a potential
1954 * deadlock. This should be very
1955 * rare.
1956 */
1957 counter_u64_add(ktls_destroy_task, 1);
1958 TASK_INIT(&tls->destroy_task, 0,
1959 ktls_destroy_help, tls);
1960 (void)taskqueue_enqueue(taskqueue_thread,
1961 &tls->destroy_task);
1962 return;
1963 }
1964 }
1965 }
1966
1967 if (tls->sequential_records) {
1968 struct mbuf *m, *n;
1969 int page_count;
1970
1971 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1972 page_count = m->m_epg_enc_cnt;
1973 while (page_count > 0) {
1974 KASSERT(page_count >= m->m_epg_nrdy,
1975 ("%s: too few pages", __func__));
1976 page_count -= m->m_epg_nrdy;
1977 m = m_free(m);
1978 }
1979 }
1980 }
1981
1982 counter_u64_add(ktls_offload_active, -1);
1983 switch (tls->mode) {
1984 case TCP_TLS_MODE_SW:
1985 switch (tls->params.cipher_algorithm) {
1986 case CRYPTO_AES_CBC:
1987 counter_u64_add(ktls_sw_cbc, -1);
1988 break;
1989 case CRYPTO_AES_NIST_GCM_16:
1990 counter_u64_add(ktls_sw_gcm, -1);
1991 break;
1992 case CRYPTO_CHACHA20_POLY1305:
1993 counter_u64_add(ktls_sw_chacha20, -1);
1994 break;
1995 }
1996 break;
1997 case TCP_TLS_MODE_IFNET:
1998 switch (tls->params.cipher_algorithm) {
1999 case CRYPTO_AES_CBC:
2000 counter_u64_add(ktls_ifnet_cbc, -1);
2001 break;
2002 case CRYPTO_AES_NIST_GCM_16:
2003 counter_u64_add(ktls_ifnet_gcm, -1);
2004 break;
2005 case CRYPTO_CHACHA20_POLY1305:
2006 counter_u64_add(ktls_ifnet_chacha20, -1);
2007 break;
2008 }
2009 if (tls->snd_tag != NULL)
2010 m_snd_tag_rele(tls->snd_tag);
2011 if (tls->rx_ifp != NULL)
2012 if_rele(tls->rx_ifp);
2013 if (tls->tx) {
2014 INP_WLOCK_ASSERT(inp);
2015 tp = intotcpcb(inp);
2016 MPASS(tp->t_nic_ktls_xmit == 1);
2017 tp->t_nic_ktls_xmit = 0;
2018 }
2019 break;
2020 #ifdef TCP_OFFLOAD
2021 case TCP_TLS_MODE_TOE:
2022 switch (tls->params.cipher_algorithm) {
2023 case CRYPTO_AES_CBC:
2024 counter_u64_add(ktls_toe_cbc, -1);
2025 break;
2026 case CRYPTO_AES_NIST_GCM_16:
2027 counter_u64_add(ktls_toe_gcm, -1);
2028 break;
2029 case CRYPTO_CHACHA20_POLY1305:
2030 counter_u64_add(ktls_toe_chacha20, -1);
2031 break;
2032 }
2033 break;
2034 #endif
2035 }
2036 if (tls->ocf_session != NULL)
2037 ktls_ocf_free(tls);
2038 if (tls->params.auth_key != NULL) {
2039 zfree(tls->params.auth_key, M_KTLS);
2040 tls->params.auth_key = NULL;
2041 tls->params.auth_key_len = 0;
2042 }
2043 if (tls->params.cipher_key != NULL) {
2044 zfree(tls->params.cipher_key, M_KTLS);
2045 tls->params.cipher_key = NULL;
2046 tls->params.cipher_key_len = 0;
2047 }
2048 if (tls->tx) {
2049 INP_WLOCK_ASSERT(inp);
2050 if (!in_pcbrele_wlocked(inp) && !wlocked)
2051 INP_WUNLOCK(inp);
2052 }
2053 explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
2054
2055 uma_zfree(ktls_session_zone, tls);
2056 }
2057
2058 void
ktls_seq(struct sockbuf * sb,struct mbuf * m)2059 ktls_seq(struct sockbuf *sb, struct mbuf *m)
2060 {
2061
2062 for (; m != NULL; m = m->m_next) {
2063 KASSERT((m->m_flags & M_EXTPG) != 0,
2064 ("ktls_seq: mapped mbuf %p", m));
2065
2066 m->m_epg_seqno = sb->sb_tls_seqno;
2067 sb->sb_tls_seqno++;
2068 }
2069 }
2070
2071 /*
2072 * Add TLS framing (headers and trailers) to a chain of mbufs. Each
2073 * mbuf in the chain must be an unmapped mbuf. The payload of the
2074 * mbuf must be populated with the payload of each TLS record.
2075 *
2076 * The record_type argument specifies the TLS record type used when
2077 * populating the TLS header.
2078 *
2079 * The enq_count argument on return is set to the number of pages of
2080 * payload data for this entire chain that need to be encrypted via SW
2081 * encryption. The returned value should be passed to ktls_enqueue
2082 * when scheduling encryption of this chain of mbufs. To handle the
2083 * special case of empty fragments for TLS 1.0 sessions, an empty
2084 * fragment counts as one page.
2085 */
2086 void
ktls_frame(struct mbuf * top,struct ktls_session * tls,int * enq_cnt,uint8_t record_type)2087 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2088 uint8_t record_type)
2089 {
2090 struct tls_record_layer *tlshdr;
2091 struct mbuf *m;
2092 uint64_t *noncep;
2093 uint16_t tls_len;
2094 int maxlen __diagused;
2095
2096 maxlen = tls->params.max_frame_len;
2097 *enq_cnt = 0;
2098 for (m = top; m != NULL; m = m->m_next) {
2099 /*
2100 * All mbufs in the chain should be TLS records whose
2101 * payload does not exceed the maximum frame length.
2102 *
2103 * Empty TLS 1.0 records are permitted when using CBC.
2104 */
2105 KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2106 (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2107 ("ktls_frame: m %p len %d", m, m->m_len));
2108
2109 /*
2110 * TLS frames require unmapped mbufs to store session
2111 * info.
2112 */
2113 KASSERT((m->m_flags & M_EXTPG) != 0,
2114 ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2115
2116 tls_len = m->m_len;
2117
2118 /* Save a reference to the session. */
2119 m->m_epg_tls = ktls_hold(tls);
2120
2121 m->m_epg_hdrlen = tls->params.tls_hlen;
2122 m->m_epg_trllen = tls->params.tls_tlen;
2123 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2124 int bs, delta;
2125
2126 /*
2127 * AES-CBC pads messages to a multiple of the
2128 * block size. Note that the padding is
2129 * applied after the digest and the encryption
2130 * is done on the "plaintext || mac || padding".
2131 * At least one byte of padding is always
2132 * present.
2133 *
2134 * Compute the final trailer length assuming
2135 * at most one block of padding.
2136 * tls->params.tls_tlen is the maximum
2137 * possible trailer length (padding + digest).
2138 * delta holds the number of excess padding
2139 * bytes if the maximum were used. Those
2140 * extra bytes are removed.
2141 */
2142 bs = tls->params.tls_bs;
2143 delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2144 m->m_epg_trllen -= delta;
2145 }
2146 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2147
2148 /* Populate the TLS header. */
2149 tlshdr = (void *)m->m_epg_hdr;
2150 tlshdr->tls_vmajor = tls->params.tls_vmajor;
2151
2152 /*
2153 * TLS 1.3 masquarades as TLS 1.2 with a record type
2154 * of TLS_RLTYPE_APP.
2155 */
2156 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2157 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2158 tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2159 tlshdr->tls_type = TLS_RLTYPE_APP;
2160 /* save the real record type for later */
2161 m->m_epg_record_type = record_type;
2162 m->m_epg_trail[0] = record_type;
2163 } else {
2164 tlshdr->tls_vminor = tls->params.tls_vminor;
2165 tlshdr->tls_type = record_type;
2166 }
2167 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2168
2169 /*
2170 * Store nonces / explicit IVs after the end of the
2171 * TLS header.
2172 *
2173 * For GCM with TLS 1.2, an 8 byte nonce is copied
2174 * from the end of the IV. The nonce is then
2175 * incremented for use by the next record.
2176 *
2177 * For CBC, a random nonce is inserted for TLS 1.1+.
2178 */
2179 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2180 tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2181 noncep = (uint64_t *)(tls->params.iv + 8);
2182 be64enc(tlshdr + 1, *noncep);
2183 (*noncep)++;
2184 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2185 tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2186 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2187
2188 /*
2189 * When using SW encryption, mark the mbuf not ready.
2190 * It will be marked ready via sbready() after the
2191 * record has been encrypted.
2192 *
2193 * When using ifnet TLS, unencrypted TLS records are
2194 * sent down the stack to the NIC.
2195 */
2196 if (tls->mode == TCP_TLS_MODE_SW) {
2197 m->m_flags |= M_NOTREADY;
2198 if (__predict_false(tls_len == 0)) {
2199 /* TLS 1.0 empty fragment. */
2200 m->m_epg_nrdy = 1;
2201 } else
2202 m->m_epg_nrdy = m->m_epg_npgs;
2203 *enq_cnt += m->m_epg_nrdy;
2204 }
2205 }
2206 }
2207
2208 bool
ktls_permit_empty_frames(struct ktls_session * tls)2209 ktls_permit_empty_frames(struct ktls_session *tls)
2210 {
2211 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2212 tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2213 }
2214
2215 void
ktls_check_rx(struct sockbuf * sb)2216 ktls_check_rx(struct sockbuf *sb)
2217 {
2218 struct tls_record_layer hdr;
2219 struct ktls_wq *wq;
2220 struct socket *so;
2221 bool running;
2222
2223 SOCKBUF_LOCK_ASSERT(sb);
2224 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2225 __func__, sb));
2226 so = __containerof(sb, struct socket, so_rcv);
2227
2228 if (sb->sb_flags & SB_TLS_RX_RUNNING)
2229 return;
2230
2231 /* Is there enough queued for a TLS header? */
2232 if (sb->sb_tlscc < sizeof(hdr)) {
2233 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2234 so->so_error = EMSGSIZE;
2235 return;
2236 }
2237
2238 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2239
2240 /* Is the entire record queued? */
2241 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2242 if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2243 so->so_error = EMSGSIZE;
2244 return;
2245 }
2246
2247 sb->sb_flags |= SB_TLS_RX_RUNNING;
2248
2249 soref(so);
2250 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2251 mtx_lock(&wq->mtx);
2252 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2253 running = wq->running;
2254 mtx_unlock(&wq->mtx);
2255 if (!running)
2256 wakeup(wq);
2257 counter_u64_add(ktls_cnt_rx_queued, 1);
2258 }
2259
2260 static struct mbuf *
ktls_detach_record(struct sockbuf * sb,int len)2261 ktls_detach_record(struct sockbuf *sb, int len)
2262 {
2263 struct mbuf *m, *n, *top;
2264 int remain;
2265
2266 SOCKBUF_LOCK_ASSERT(sb);
2267 MPASS(len <= sb->sb_tlscc);
2268
2269 /*
2270 * If TLS chain is the exact size of the record,
2271 * just grab the whole record.
2272 */
2273 top = sb->sb_mtls;
2274 if (sb->sb_tlscc == len) {
2275 sb->sb_mtls = NULL;
2276 sb->sb_mtlstail = NULL;
2277 goto out;
2278 }
2279
2280 /*
2281 * While it would be nice to use m_split() here, we need
2282 * to know exactly what m_split() allocates to update the
2283 * accounting, so do it inline instead.
2284 */
2285 remain = len;
2286 for (m = top; remain > m->m_len; m = m->m_next)
2287 remain -= m->m_len;
2288
2289 /* Easy case: don't have to split 'm'. */
2290 if (remain == m->m_len) {
2291 sb->sb_mtls = m->m_next;
2292 if (sb->sb_mtls == NULL)
2293 sb->sb_mtlstail = NULL;
2294 m->m_next = NULL;
2295 goto out;
2296 }
2297
2298 /*
2299 * Need to allocate an mbuf to hold the remainder of 'm'. Try
2300 * with M_NOWAIT first.
2301 */
2302 n = m_get(M_NOWAIT, MT_DATA);
2303 if (n == NULL) {
2304 /*
2305 * Use M_WAITOK with socket buffer unlocked. If
2306 * 'sb_mtls' changes while the lock is dropped, return
2307 * NULL to force the caller to retry.
2308 */
2309 SOCKBUF_UNLOCK(sb);
2310
2311 n = m_get(M_WAITOK, MT_DATA);
2312
2313 SOCKBUF_LOCK(sb);
2314 if (sb->sb_mtls != top) {
2315 m_free(n);
2316 return (NULL);
2317 }
2318 }
2319 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2320
2321 /* Store remainder in 'n'. */
2322 n->m_len = m->m_len - remain;
2323 if (m->m_flags & M_EXT) {
2324 n->m_data = m->m_data + remain;
2325 mb_dupcl(n, m);
2326 } else {
2327 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2328 }
2329
2330 /* Trim 'm' and update accounting. */
2331 m->m_len -= n->m_len;
2332 sb->sb_tlscc -= n->m_len;
2333 sb->sb_ccc -= n->m_len;
2334
2335 /* Account for 'n'. */
2336 sballoc_ktls_rx(sb, n);
2337
2338 /* Insert 'n' into the TLS chain. */
2339 sb->sb_mtls = n;
2340 n->m_next = m->m_next;
2341 if (sb->sb_mtlstail == m)
2342 sb->sb_mtlstail = n;
2343
2344 /* Detach the record from the TLS chain. */
2345 m->m_next = NULL;
2346
2347 out:
2348 MPASS(m_length(top, NULL) == len);
2349 for (m = top; m != NULL; m = m->m_next)
2350 sbfree_ktls_rx(sb, m);
2351 sb->sb_tlsdcc = len;
2352 sb->sb_ccc += len;
2353 SBCHECK(sb);
2354 return (top);
2355 }
2356
2357 /*
2358 * Determine the length of the trailing zero padding and find the real
2359 * record type in the byte before the padding.
2360 *
2361 * Walking the mbuf chain backwards is clumsy, so another option would
2362 * be to scan forwards remembering the last non-zero byte before the
2363 * trailer. However, it would be expensive to scan the entire record.
2364 * Instead, find the last non-zero byte of each mbuf in the chain
2365 * keeping track of the relative offset of that nonzero byte.
2366 *
2367 * trail_len is the size of the MAC/tag on input and is set to the
2368 * size of the full trailer including padding and the record type on
2369 * return.
2370 */
2371 static int
tls13_find_record_type(struct ktls_session * tls,struct mbuf * m,int tls_len,int * trailer_len,uint8_t * record_typep)2372 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2373 int *trailer_len, uint8_t *record_typep)
2374 {
2375 char *cp;
2376 u_int digest_start, last_offset, m_len, offset;
2377 uint8_t record_type;
2378
2379 digest_start = tls_len - *trailer_len;
2380 last_offset = 0;
2381 offset = 0;
2382 for (; m != NULL && offset < digest_start;
2383 offset += m->m_len, m = m->m_next) {
2384 /* Don't look for padding in the tag. */
2385 m_len = min(digest_start - offset, m->m_len);
2386 cp = mtod(m, char *);
2387
2388 /* Find last non-zero byte in this mbuf. */
2389 while (m_len > 0 && cp[m_len - 1] == 0)
2390 m_len--;
2391 if (m_len > 0) {
2392 record_type = cp[m_len - 1];
2393 last_offset = offset + m_len - 1;
2394 }
2395 }
2396 if (last_offset < tls->params.tls_hlen)
2397 return (EBADMSG);
2398
2399 *record_typep = record_type;
2400 *trailer_len = tls_len - last_offset;
2401 return (0);
2402 }
2403
2404 /*
2405 * Check if a mbuf chain is fully decrypted. Returns
2406 * KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is decrypted.
2407 * KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted and
2408 * decrypted data. KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data is
2409 * encrypted. KTLS_MBUF_CRYPTO_ST_SHAREDMBUF if any mbuf points at
2410 * shared data that must not be modified in place (non-anonymous
2411 * M_EXTPG or sendfile M_EXT buffers).
2412 */
2413 static ktls_mbuf_crypto_st_t
ktls_mbuf_crypto_state(struct mbuf * mb)2414 ktls_mbuf_crypto_state(struct mbuf *mb)
2415 {
2416 bool seen_decrypted, seen_encrypted;
2417
2418 seen_decrypted = false;
2419 seen_encrypted = false;
2420
2421 for (; mb != NULL; mb = mb->m_next) {
2422 if ((mb->m_flags & M_EXTPG) != 0 &&
2423 (mb->m_epg_flags & EPG_FLAG_ANON) == 0)
2424 return (KTLS_MBUF_CRYPTO_ST_SHAREDMBUF);
2425 if ((mb->m_flags & M_EXT) != 0 &&
2426 mb->m_ext.ext_type == EXT_SFBUF &&
2427 (mb->m_ext.ext_flags & EXT_FLAG_SFBUF_ANON) == 0)
2428 return (KTLS_MBUF_CRYPTO_ST_SHAREDMBUF);
2429
2430 if (mb->m_flags & M_DECRYPTED)
2431 seen_decrypted = true;
2432 else
2433 seen_encrypted = true;
2434 }
2435
2436 if (seen_decrypted && seen_encrypted)
2437 return (KTLS_MBUF_CRYPTO_ST_MIXED);
2438 else
2439 return (seen_decrypted ?
2440 KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2441 KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2442 }
2443
2444 /*
2445 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2446 */
2447 static int
ktls_resync_ifnet(struct socket * so,uint32_t tls_len,uint64_t tls_rcd_num)2448 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2449 {
2450 union if_snd_tag_modify_params params;
2451 struct m_snd_tag *mst;
2452 struct inpcb *inp = sotoinpcb(so);
2453 struct tcpcb *tp = intotcpcb(inp);
2454
2455 mst = so->so_rcv.sb_tls_info->snd_tag;
2456 if (__predict_false(mst == NULL))
2457 return (EINVAL);
2458
2459 INP_RLOCK(inp);
2460 if (tp->t_flags & TF_DISCONNECTED) {
2461 INP_RUNLOCK(inp);
2462 return (ECONNRESET);
2463 }
2464
2465 /* Get the TCP sequence number of the next valid TLS header. */
2466 SOCKBUF_LOCK(&so->so_rcv);
2467 params.tls_rx.tls_hdr_tcp_sn =
2468 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2469 params.tls_rx.tls_rec_length = tls_len;
2470 params.tls_rx.tls_seq_number = tls_rcd_num;
2471 SOCKBUF_UNLOCK(&so->so_rcv);
2472
2473 INP_RUNLOCK(inp);
2474
2475 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2476 return (mst->sw->snd_tag_modify(mst, ¶ms));
2477 }
2478
2479 static void
ktls_drop(struct socket * so,int error)2480 ktls_drop(struct socket *so, int error)
2481 {
2482 struct epoch_tracker et;
2483 struct inpcb *inp = sotoinpcb(so);
2484 struct tcpcb *tp = intotcpcb(inp);
2485
2486 NET_EPOCH_ENTER(et);
2487 INP_WLOCK(inp);
2488 if (!(tp->t_flags & TF_DISCONNECTED)) {
2489 CURVNET_SET(inp->inp_socket->so_vnet);
2490 tp = tcp_drop(tp, error);
2491 CURVNET_RESTORE();
2492 if (tp != NULL)
2493 INP_WUNLOCK(inp);
2494 } else {
2495 so->so_error = error;
2496 SOCK_RECVBUF_LOCK(so);
2497 sorwakeup_locked(so);
2498 INP_WUNLOCK(inp);
2499 }
2500 NET_EPOCH_EXIT(et);
2501 }
2502
2503 static void
ktls_decrypt(struct socket * so)2504 ktls_decrypt(struct socket *so)
2505 {
2506 char tls_header[MBUF_PEXT_HDR_LEN];
2507 struct ktls_session *tls;
2508 struct sockbuf *sb;
2509 struct tls_record_layer *hdr;
2510 struct tls_get_record tgr;
2511 struct mbuf *control, *data, *m;
2512 ktls_mbuf_crypto_st_t state;
2513 uint64_t seqno;
2514 int error, remain, tls_len, trail_len;
2515 bool tls13;
2516 uint8_t vminor, record_type;
2517
2518 hdr = (struct tls_record_layer *)tls_header;
2519 sb = &so->so_rcv;
2520 SOCKBUF_LOCK(sb);
2521 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2522 ("%s: socket %p not running", __func__, so));
2523
2524 tls = sb->sb_tls_info;
2525 MPASS(tls != NULL);
2526
2527 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2528 if (tls13)
2529 vminor = TLS_MINOR_VER_TWO;
2530 else
2531 vminor = tls->params.tls_vminor;
2532 for (;;) {
2533 /* Is there enough queued for a TLS header? */
2534 if (sb->sb_tlscc < tls->params.tls_hlen)
2535 break;
2536
2537 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2538 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2539
2540 if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2541 hdr->tls_vminor != vminor)
2542 error = EINVAL;
2543 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2544 error = EINVAL;
2545 else if (tls_len < tls->params.tls_hlen || tls_len >
2546 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2547 tls->params.tls_tlen)
2548 error = EMSGSIZE;
2549 else
2550 error = 0;
2551 if (__predict_false(error != 0)) {
2552 /*
2553 * We have a corrupted record and are likely
2554 * out of sync. The connection isn't
2555 * recoverable at this point, so abort it.
2556 */
2557 SOCKBUF_UNLOCK(sb);
2558 counter_u64_add(ktls_offload_corrupted_records, 1);
2559
2560 ktls_drop(so, error);
2561 goto deref;
2562 }
2563
2564 /* Is the entire record queued? */
2565 if (sb->sb_tlscc < tls_len)
2566 break;
2567
2568 /*
2569 * Split out the portion of the mbuf chain containing
2570 * this TLS record.
2571 */
2572 data = ktls_detach_record(sb, tls_len);
2573 if (data == NULL)
2574 continue;
2575 MPASS(sb->sb_tlsdcc == tls_len);
2576
2577 seqno = sb->sb_tls_seqno;
2578 sb->sb_tls_seqno++;
2579 SBCHECK(sb);
2580 SOCKBUF_UNLOCK(sb);
2581
2582 /* get crypto state for this TLS record */
2583 state = ktls_mbuf_crypto_state(data);
2584
2585 switch (state) {
2586 case KTLS_MBUF_CRYPTO_ST_MIXED:
2587 error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2588 if (error)
2589 break;
2590 /* FALLTHROUGH */
2591 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2592 error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2593 &trail_len);
2594 if (__predict_true(error == 0)) {
2595 if (tls13) {
2596 error = tls13_find_record_type(tls, data,
2597 tls_len, &trail_len, &record_type);
2598 } else {
2599 record_type = hdr->tls_type;
2600 }
2601 }
2602 break;
2603 case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2604 /*
2605 * NIC TLS is only supported for AEAD
2606 * ciphersuites which used a fixed sized
2607 * trailer.
2608 */
2609 if (tls13) {
2610 trail_len = tls->params.tls_tlen - 1;
2611 error = tls13_find_record_type(tls, data,
2612 tls_len, &trail_len, &record_type);
2613 } else {
2614 trail_len = tls->params.tls_tlen;
2615 error = 0;
2616 record_type = hdr->tls_type;
2617 }
2618 break;
2619 case KTLS_MBUF_CRYPTO_ST_SHAREDMBUF:
2620 error = EINVAL;
2621 break;
2622 default:
2623 __assert_unreachable();
2624 }
2625 if (error) {
2626 counter_u64_add(ktls_offload_failed_crypto, 1);
2627
2628 SOCKBUF_LOCK(sb);
2629 if (sb->sb_tlsdcc == 0) {
2630 /*
2631 * sbcut/drop/flush discarded these
2632 * mbufs.
2633 */
2634 m_freem(data);
2635 break;
2636 }
2637
2638 /*
2639 * Drop this TLS record's data, but keep
2640 * decrypting subsequent records.
2641 */
2642 sb->sb_ccc -= tls_len;
2643 sb->sb_tlsdcc = 0;
2644
2645 if (error != EMSGSIZE)
2646 error = EBADMSG;
2647 CURVNET_SET(so->so_vnet);
2648 so->so_error = error;
2649 sorwakeup_locked(so);
2650 CURVNET_RESTORE();
2651
2652 m_freem(data);
2653
2654 SOCKBUF_LOCK(sb);
2655 continue;
2656 }
2657
2658 /* Allocate the control mbuf. */
2659 memset(&tgr, 0, sizeof(tgr));
2660 tgr.tls_type = record_type;
2661 tgr.tls_vmajor = hdr->tls_vmajor;
2662 tgr.tls_vminor = hdr->tls_vminor;
2663 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2664 trail_len);
2665 control = sbcreatecontrol(&tgr, sizeof(tgr),
2666 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2667
2668 SOCKBUF_LOCK(sb);
2669 if (sb->sb_tlsdcc == 0) {
2670 /* sbcut/drop/flush discarded these mbufs. */
2671 MPASS(sb->sb_tlscc == 0);
2672 m_freem(data);
2673 m_freem(control);
2674 break;
2675 }
2676
2677 /*
2678 * Clear the 'dcc' accounting in preparation for
2679 * adding the decrypted record.
2680 */
2681 sb->sb_ccc -= tls_len;
2682 sb->sb_tlsdcc = 0;
2683 SBCHECK(sb);
2684
2685 /* If there is no payload, drop all of the data. */
2686 if (tgr.tls_length == htobe16(0)) {
2687 m_freem(data);
2688 data = NULL;
2689 } else {
2690 /* Trim header. */
2691 remain = tls->params.tls_hlen;
2692 while (remain > 0) {
2693 if (data->m_len > remain) {
2694 data->m_data += remain;
2695 data->m_len -= remain;
2696 break;
2697 }
2698 remain -= data->m_len;
2699 data = m_free(data);
2700 }
2701
2702 /* Trim trailer and clear M_NOTREADY. */
2703 remain = be16toh(tgr.tls_length);
2704 m = data;
2705 for (m = data; remain > m->m_len; m = m->m_next) {
2706 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2707 remain -= m->m_len;
2708 }
2709 m->m_len = remain;
2710 m_freem(m->m_next);
2711 m->m_next = NULL;
2712 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2713
2714 /* Set EOR on the final mbuf. */
2715 m->m_flags |= M_EOR;
2716 }
2717
2718 sbappendcontrol_locked(sb, data, control, 0);
2719
2720 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2721 sb->sb_flags |= SB_TLS_RX_RESYNC;
2722 SOCKBUF_UNLOCK(sb);
2723 ktls_resync_ifnet(so, tls_len, seqno);
2724 SOCKBUF_LOCK(sb);
2725 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2726 sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2727 SOCKBUF_UNLOCK(sb);
2728 ktls_resync_ifnet(so, 0, seqno);
2729 SOCKBUF_LOCK(sb);
2730 }
2731 }
2732
2733 sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2734
2735 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2736 so->so_error = EMSGSIZE;
2737
2738 sorwakeup_locked(so);
2739
2740 deref:
2741 SOCKBUF_UNLOCK_ASSERT(sb);
2742
2743 CURVNET_SET(so->so_vnet);
2744 sorele(so);
2745 CURVNET_RESTORE();
2746 }
2747
2748 void
ktls_enqueue_to_free(struct mbuf * m)2749 ktls_enqueue_to_free(struct mbuf *m)
2750 {
2751 struct ktls_wq *wq;
2752 bool running;
2753
2754 /* Mark it for freeing. */
2755 m->m_epg_flags |= EPG_FLAG_2FREE;
2756 wq = &ktls_wq[m->m_epg_tls->wq_index];
2757 mtx_lock(&wq->mtx);
2758 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2759 running = wq->running;
2760 mtx_unlock(&wq->mtx);
2761 if (!running)
2762 wakeup(wq);
2763 }
2764
2765 static void *
ktls_buffer_alloc(struct ktls_wq * wq,struct mbuf * m)2766 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2767 {
2768 void *buf;
2769 int domain, running;
2770
2771 if (m->m_epg_npgs <= 2)
2772 return (NULL);
2773 if (ktls_buffer_zone == NULL)
2774 return (NULL);
2775 if ((u_int)(ticks - wq->lastallocfail) < hz) {
2776 /*
2777 * Rate-limit allocation attempts after a failure.
2778 * ktls_buffer_import() will acquire a per-domain mutex to check
2779 * the free page queues and may fail consistently if memory is
2780 * fragmented.
2781 */
2782 return (NULL);
2783 }
2784 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2785 if (buf == NULL) {
2786 domain = PCPU_GET(domain);
2787 wq->lastallocfail = ticks;
2788
2789 /*
2790 * Note that this check is "racy", but the races are
2791 * harmless, and are either a spurious wakeup if
2792 * multiple threads fail allocations before the alloc
2793 * thread wakes, or waiting an extra second in case we
2794 * see an old value of running == true.
2795 */
2796 if (!VM_DOMAIN_EMPTY(domain)) {
2797 running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2798 if (!running)
2799 wakeup(&ktls_domains[domain].reclaim_td);
2800 }
2801 }
2802 return (buf);
2803 }
2804
2805 static int
ktls_encrypt_record(struct ktls_wq * wq,struct mbuf * m,struct ktls_session * tls,struct ktls_ocf_encrypt_state * state)2806 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2807 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2808 {
2809 vm_page_t pg;
2810 int error, i, len, off;
2811
2812 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2813 ("%p not unready & nomap mbuf\n", m));
2814 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2815 ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2816 ktls_maxlen));
2817
2818 /* Anonymous mbufs are encrypted in place. */
2819 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2820 return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2821
2822 /*
2823 * For file-backed mbufs (from sendfile), anonymous wired
2824 * pages are allocated and used as the encryption destination.
2825 */
2826 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2827 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2828 m->m_epg_1st_off;
2829 state->dst_iov[0].iov_base = (char *)state->cbuf +
2830 m->m_epg_1st_off;
2831 state->dst_iov[0].iov_len = len;
2832 state->parray[0] = DMAP_TO_PHYS(state->cbuf);
2833 i = 1;
2834 } else {
2835 off = m->m_epg_1st_off;
2836 for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2837 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2838 VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2839 len = m_epg_pagelen(m, i, off);
2840 state->parray[i] = VM_PAGE_TO_PHYS(pg);
2841 state->dst_iov[i].iov_base =
2842 (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2843 state->dst_iov[i].iov_len = len;
2844 }
2845 }
2846 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2847 state->dst_iov[i].iov_base = m->m_epg_trail;
2848 state->dst_iov[i].iov_len = m->m_epg_trllen;
2849
2850 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2851
2852 if (__predict_false(error != 0)) {
2853 /* Free the anonymous pages. */
2854 if (state->cbuf != NULL)
2855 uma_zfree(ktls_buffer_zone, state->cbuf);
2856 else {
2857 for (i = 0; i < m->m_epg_npgs; i++) {
2858 pg = PHYS_TO_VM_PAGE(state->parray[i]);
2859 (void)vm_page_unwire_noq(pg);
2860 vm_page_free(pg);
2861 }
2862 }
2863 }
2864 return (error);
2865 }
2866
2867 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2868 static u_int
ktls_batched_records(struct mbuf * m)2869 ktls_batched_records(struct mbuf *m)
2870 {
2871 int page_count, records;
2872
2873 records = 0;
2874 page_count = m->m_epg_enc_cnt;
2875 while (page_count > 0) {
2876 records++;
2877 page_count -= m->m_epg_nrdy;
2878 m = m->m_next;
2879 }
2880 KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2881 return (records);
2882 }
2883
2884 void
ktls_enqueue(struct mbuf * m,struct socket * so,int page_count)2885 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2886 {
2887 struct ktls_session *tls;
2888 struct ktls_wq *wq;
2889 int queued;
2890 bool running;
2891
2892 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2893 (M_EXTPG | M_NOTREADY)),
2894 ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2895 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2896
2897 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2898
2899 m->m_epg_enc_cnt = page_count;
2900
2901 /*
2902 * Save a pointer to the socket. The caller is responsible
2903 * for taking an additional reference via soref().
2904 */
2905 m->m_epg_so = so;
2906
2907 queued = 1;
2908 tls = m->m_epg_tls;
2909 wq = &ktls_wq[tls->wq_index];
2910 mtx_lock(&wq->mtx);
2911 if (__predict_false(tls->sequential_records)) {
2912 /*
2913 * For TLS 1.0, records must be encrypted
2914 * sequentially. For a given connection, all records
2915 * queued to the associated work queue are processed
2916 * sequentially. However, sendfile(2) might complete
2917 * I/O requests spanning multiple TLS records out of
2918 * order. Here we ensure TLS records are enqueued to
2919 * the work queue in FIFO order.
2920 *
2921 * tls->next_seqno holds the sequence number of the
2922 * next TLS record that should be enqueued to the work
2923 * queue. If this next record is not tls->next_seqno,
2924 * it must be a future record, so insert it, sorted by
2925 * TLS sequence number, into tls->pending_records and
2926 * return.
2927 *
2928 * If this TLS record matches tls->next_seqno, place
2929 * it in the work queue and then check
2930 * tls->pending_records to see if any
2931 * previously-queued records are now ready for
2932 * encryption.
2933 */
2934 if (m->m_epg_seqno != tls->next_seqno) {
2935 struct mbuf *n, *p;
2936
2937 p = NULL;
2938 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2939 if (n->m_epg_seqno > m->m_epg_seqno)
2940 break;
2941 p = n;
2942 }
2943 if (n == NULL)
2944 STAILQ_INSERT_TAIL(&tls->pending_records, m,
2945 m_epg_stailq);
2946 else if (p == NULL)
2947 STAILQ_INSERT_HEAD(&tls->pending_records, m,
2948 m_epg_stailq);
2949 else
2950 STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2951 m_epg_stailq);
2952 mtx_unlock(&wq->mtx);
2953 counter_u64_add(ktls_cnt_tx_pending, 1);
2954 return;
2955 }
2956
2957 tls->next_seqno += ktls_batched_records(m);
2958 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2959
2960 while (!STAILQ_EMPTY(&tls->pending_records)) {
2961 struct mbuf *n;
2962
2963 n = STAILQ_FIRST(&tls->pending_records);
2964 if (n->m_epg_seqno != tls->next_seqno)
2965 break;
2966
2967 queued++;
2968 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2969 tls->next_seqno += ktls_batched_records(n);
2970 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2971 }
2972 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2973 } else
2974 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2975
2976 running = wq->running;
2977 mtx_unlock(&wq->mtx);
2978 if (!running)
2979 wakeup(wq);
2980 counter_u64_add(ktls_cnt_tx_queued, queued);
2981 }
2982
2983 /*
2984 * Once a file-backed mbuf (from sendfile) has been encrypted, free
2985 * the pages from the file and replace them with the anonymous pages
2986 * allocated in ktls_encrypt_record().
2987 */
2988 static void
ktls_finish_nonanon(struct mbuf * m,struct ktls_ocf_encrypt_state * state)2989 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2990 {
2991 int i;
2992
2993 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2994
2995 /* Free the old pages. */
2996 m->m_ext.ext_free(m);
2997
2998 /* Replace them with the new pages. */
2999 if (state->cbuf != NULL) {
3000 for (i = 0; i < m->m_epg_npgs; i++)
3001 m->m_epg_pa[i] = state->parray[0] + ptoa(i);
3002
3003 /* Contig pages should go back to the cache. */
3004 m->m_ext.ext_free = ktls_free_mext_contig;
3005 } else {
3006 for (i = 0; i < m->m_epg_npgs; i++)
3007 m->m_epg_pa[i] = state->parray[i];
3008
3009 /* Use the basic free routine. */
3010 m->m_ext.ext_free = mb_free_mext_pgs;
3011 }
3012
3013 /* Pages are now writable. */
3014 m->m_epg_flags |= EPG_FLAG_ANON;
3015 }
3016
3017 static __noinline void
ktls_encrypt(struct ktls_wq * wq,struct mbuf * top)3018 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
3019 {
3020 struct ktls_ocf_encrypt_state state;
3021 struct ktls_session *tls;
3022 struct socket *so;
3023 struct mbuf *m;
3024 int error, npages, total_pages;
3025
3026 so = top->m_epg_so;
3027 tls = top->m_epg_tls;
3028 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3029 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3030 #ifdef INVARIANTS
3031 top->m_epg_so = NULL;
3032 #endif
3033 total_pages = top->m_epg_enc_cnt;
3034 npages = 0;
3035
3036 /*
3037 * Encrypt the TLS records in the chain of mbufs starting with
3038 * 'top'. 'total_pages' gives us a total count of pages and is
3039 * used to know when we have finished encrypting the TLS
3040 * records originally queued with 'top'.
3041 *
3042 * NB: These mbufs are queued in the socket buffer and
3043 * 'm_next' is traversing the mbufs in the socket buffer. The
3044 * socket buffer lock is not held while traversing this chain.
3045 * Since the mbufs are all marked M_NOTREADY their 'm_next'
3046 * pointers should be stable. However, the 'm_next' of the
3047 * last mbuf encrypted is not necessarily NULL. It can point
3048 * to other mbufs appended while 'top' was on the TLS work
3049 * queue.
3050 *
3051 * Each mbuf holds an entire TLS record.
3052 */
3053 error = 0;
3054 for (m = top; npages != total_pages; m = m->m_next) {
3055 KASSERT(m->m_epg_tls == tls,
3056 ("different TLS sessions in a single mbuf chain: %p vs %p",
3057 tls, m->m_epg_tls));
3058 KASSERT(npages + m->m_epg_npgs <= total_pages,
3059 ("page count mismatch: top %p, total_pages %d, m %p", top,
3060 total_pages, m));
3061
3062 error = ktls_encrypt_record(wq, m, tls, &state);
3063 if (error) {
3064 counter_u64_add(ktls_offload_failed_crypto, 1);
3065 break;
3066 }
3067
3068 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3069 ktls_finish_nonanon(m, &state);
3070 m->m_flags |= M_RDONLY;
3071
3072 npages += m->m_epg_nrdy;
3073
3074 /*
3075 * Drop a reference to the session now that it is no
3076 * longer needed. Existing code depends on encrypted
3077 * records having no associated session vs
3078 * yet-to-be-encrypted records having an associated
3079 * session.
3080 */
3081 m->m_epg_tls = NULL;
3082 ktls_free(tls);
3083 }
3084
3085 CURVNET_SET(so->so_vnet);
3086 if (error == 0) {
3087 (void)so->so_proto->pr_ready(so, top, npages);
3088 } else {
3089 ktls_drop(so, EIO);
3090 mb_free_notready(top, total_pages);
3091 }
3092
3093 sorele(so);
3094 CURVNET_RESTORE();
3095 }
3096
3097 void
ktls_encrypt_cb(struct ktls_ocf_encrypt_state * state,int error)3098 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3099 {
3100 struct ktls_session *tls;
3101 struct socket *so;
3102 struct mbuf *m;
3103 int npages;
3104
3105 m = state->m;
3106
3107 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3108 ktls_finish_nonanon(m, state);
3109 m->m_flags |= M_RDONLY;
3110
3111 so = state->so;
3112 free(state, M_KTLS);
3113
3114 /*
3115 * Drop a reference to the session now that it is no longer
3116 * needed. Existing code depends on encrypted records having
3117 * no associated session vs yet-to-be-encrypted records having
3118 * an associated session.
3119 */
3120 tls = m->m_epg_tls;
3121 m->m_epg_tls = NULL;
3122 ktls_free(tls);
3123
3124 if (error != 0)
3125 counter_u64_add(ktls_offload_failed_crypto, 1);
3126
3127 CURVNET_SET(so->so_vnet);
3128 npages = m->m_epg_nrdy;
3129
3130 if (error == 0) {
3131 (void)so->so_proto->pr_ready(so, m, npages);
3132 } else {
3133 ktls_drop(so, EIO);
3134 mb_free_notready(m, npages);
3135 }
3136
3137 sorele(so);
3138 CURVNET_RESTORE();
3139 }
3140
3141 /*
3142 * Similar to ktls_encrypt, but used with asynchronous OCF backends
3143 * (coprocessors) where encryption does not use host CPU resources and
3144 * it can be beneficial to queue more requests than CPUs.
3145 */
3146 static __noinline void
ktls_encrypt_async(struct ktls_wq * wq,struct mbuf * top)3147 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3148 {
3149 struct ktls_ocf_encrypt_state *state;
3150 struct ktls_session *tls;
3151 struct socket *so;
3152 struct mbuf *m, *n;
3153 int error, mpages, npages, total_pages;
3154
3155 so = top->m_epg_so;
3156 tls = top->m_epg_tls;
3157 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3158 KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3159 #ifdef INVARIANTS
3160 top->m_epg_so = NULL;
3161 #endif
3162 total_pages = top->m_epg_enc_cnt;
3163 npages = 0;
3164
3165 error = 0;
3166 for (m = top; npages != total_pages; m = n) {
3167 KASSERT(m->m_epg_tls == tls,
3168 ("different TLS sessions in a single mbuf chain: %p vs %p",
3169 tls, m->m_epg_tls));
3170 KASSERT(npages + m->m_epg_npgs <= total_pages,
3171 ("page count mismatch: top %p, total_pages %d, m %p", top,
3172 total_pages, m));
3173
3174 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3175 soref(so);
3176 state->so = so;
3177 state->m = m;
3178
3179 mpages = m->m_epg_nrdy;
3180 n = m->m_next;
3181
3182 error = ktls_encrypt_record(wq, m, tls, state);
3183 if (error) {
3184 counter_u64_add(ktls_offload_failed_crypto, 1);
3185 free(state, M_KTLS);
3186 CURVNET_SET(so->so_vnet);
3187 sorele(so);
3188 CURVNET_RESTORE();
3189 break;
3190 }
3191
3192 npages += mpages;
3193 }
3194
3195 CURVNET_SET(so->so_vnet);
3196 if (error != 0) {
3197 ktls_drop(so, EIO);
3198 mb_free_notready(m, total_pages - npages);
3199 }
3200
3201 sorele(so);
3202 CURVNET_RESTORE();
3203 }
3204
3205 static int
ktls_bind_domain(int domain)3206 ktls_bind_domain(int domain)
3207 {
3208 int error;
3209
3210 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3211 if (error != 0)
3212 return (error);
3213 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3214 return (0);
3215 }
3216
3217 static void
ktls_reclaim_thread(void * ctx)3218 ktls_reclaim_thread(void *ctx)
3219 {
3220 struct ktls_domain_info *ktls_domain = ctx;
3221 struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3222 struct sysctl_oid *oid;
3223 char name[80];
3224 int error, domain;
3225
3226 domain = ktls_domain - ktls_domains;
3227 if (bootverbose)
3228 printf("Starting KTLS reclaim thread for domain %d\n", domain);
3229 error = ktls_bind_domain(domain);
3230 if (error)
3231 printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3232 domain, error);
3233 snprintf(name, sizeof(name), "domain%d", domain);
3234 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3235 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3236 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3237 CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed");
3238 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3239 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups");
3240 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3241 CTLFLAG_RD, &sc->running, 0, "thread running");
3242
3243 for (;;) {
3244 atomic_store_int(&sc->running, 0);
3245 tsleep(sc, PZERO | PNOLOCK, "-", 0);
3246 atomic_store_int(&sc->running, 1);
3247 sc->wakeups++;
3248 /*
3249 * Below we attempt to reclaim ktls_max_reclaim
3250 * buffers using vm_page_reclaim_contig_domain_ext().
3251 * We do this here, as this function can take several
3252 * seconds to scan all of memory and it does not
3253 * matter if this thread pauses for a while. If we
3254 * block a ktls worker thread, we risk developing
3255 * backlogs of buffers to be encrypted, leading to
3256 * surges of traffic and potential NIC output drops.
3257 */
3258 if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3259 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
3260 ktls_max_reclaim) != 0) {
3261 vm_wait_domain(domain);
3262 } else {
3263 sc->reclaims += ktls_max_reclaim;
3264 }
3265 }
3266 }
3267
3268 static void
ktls_work_thread(void * ctx)3269 ktls_work_thread(void *ctx)
3270 {
3271 struct ktls_wq *wq = ctx;
3272 struct mbuf *m, *n;
3273 struct socket *so, *son;
3274 STAILQ_HEAD(, mbuf) local_m_head;
3275 STAILQ_HEAD(, socket) local_so_head;
3276 int cpu;
3277
3278 cpu = wq - ktls_wq;
3279 if (bootverbose)
3280 printf("Starting KTLS worker thread for CPU %d\n", cpu);
3281
3282 /*
3283 * Bind to a core. If ktls_bind_threads is > 1, then
3284 * we bind to the NUMA domain instead.
3285 */
3286 if (ktls_bind_threads) {
3287 int error;
3288
3289 if (ktls_bind_threads > 1) {
3290 struct pcpu *pc = pcpu_find(cpu);
3291
3292 error = ktls_bind_domain(pc->pc_domain);
3293 } else {
3294 cpuset_t mask;
3295
3296 CPU_SETOF(cpu, &mask);
3297 error = cpuset_setthread(curthread->td_tid, &mask);
3298 }
3299 if (error)
3300 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3301 cpu, error);
3302 }
3303 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3304 fpu_kern_thread(0);
3305 #endif
3306 for (;;) {
3307 mtx_lock(&wq->mtx);
3308 while (STAILQ_EMPTY(&wq->m_head) &&
3309 STAILQ_EMPTY(&wq->so_head)) {
3310 wq->running = false;
3311 mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3312 wq->running = true;
3313 }
3314
3315 STAILQ_INIT(&local_m_head);
3316 STAILQ_CONCAT(&local_m_head, &wq->m_head);
3317 STAILQ_INIT(&local_so_head);
3318 STAILQ_CONCAT(&local_so_head, &wq->so_head);
3319 mtx_unlock(&wq->mtx);
3320
3321 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3322 if (m->m_epg_flags & EPG_FLAG_2FREE) {
3323 ktls_free(m->m_epg_tls);
3324 m_free_raw(m);
3325 } else {
3326 if (m->m_epg_tls->sync_dispatch)
3327 ktls_encrypt(wq, m);
3328 else
3329 ktls_encrypt_async(wq, m);
3330 counter_u64_add(ktls_cnt_tx_queued, -1);
3331 }
3332 }
3333
3334 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3335 ktls_decrypt(so);
3336 counter_u64_add(ktls_cnt_rx_queued, -1);
3337 }
3338 }
3339 }
3340
3341 static void
ktls_disable_ifnet_help(void * context,int pending __unused)3342 ktls_disable_ifnet_help(void *context, int pending __unused)
3343 {
3344 struct ktls_session *tls;
3345 struct inpcb *inp;
3346 struct tcpcb *tp;
3347 struct socket *so;
3348 int err;
3349
3350 tls = context;
3351 inp = tls->inp;
3352 if (inp == NULL)
3353 return;
3354 INP_WLOCK(inp);
3355 so = inp->inp_socket;
3356 MPASS(so != NULL);
3357 tp = intotcpcb(inp);
3358 if (tp->t_flags & TF_DISCONNECTED) {
3359 goto out;
3360 }
3361
3362 if (so->so_snd.sb_tls_info != NULL)
3363 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3364 else
3365 err = ENXIO;
3366 if (err == 0) {
3367 counter_u64_add(ktls_ifnet_disable_ok, 1);
3368 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3369 if ((tp->t_flags & TF_DISCONNECTED) == 0 &&
3370 tp->t_fb->tfb_hwtls_change != NULL)
3371 (*tp->t_fb->tfb_hwtls_change)(tp, 0);
3372 } else {
3373 counter_u64_add(ktls_ifnet_disable_fail, 1);
3374 }
3375
3376 out:
3377 CURVNET_SET(so->so_vnet);
3378 sorele(so);
3379 CURVNET_RESTORE();
3380 INP_WUNLOCK(inp);
3381 ktls_free(tls);
3382 }
3383
3384 /*
3385 * Called when re-transmits are becoming a substantial portion of the
3386 * sends on this connection. When this happens, we transition the
3387 * connection to software TLS. This is needed because most inline TLS
3388 * NICs keep crypto state only for in-order transmits. This means
3389 * that to handle a TCP rexmit (which is out-of-order), the NIC must
3390 * re-DMA the entire TLS record up to and including the current
3391 * segment. This means that when re-transmitting the last ~1448 byte
3392 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3393 * of magnitude more data than we are sending. This can cause the
3394 * PCIe link to saturate well before the network, which can cause
3395 * output drops, and a general loss of capacity.
3396 */
3397 void
ktls_disable_ifnet(void * arg)3398 ktls_disable_ifnet(void *arg)
3399 {
3400 struct tcpcb *tp;
3401 struct inpcb *inp;
3402 struct socket *so;
3403 struct ktls_session *tls;
3404
3405 tp = arg;
3406 inp = tptoinpcb(tp);
3407 INP_WLOCK_ASSERT(inp);
3408 so = inp->inp_socket;
3409 SOCK_LOCK(so);
3410 tls = so->so_snd.sb_tls_info;
3411 if (tp->t_nic_ktls_xmit_dis == 1) {
3412 SOCK_UNLOCK(so);
3413 return;
3414 }
3415
3416 /*
3417 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3418 * ifnet can only be done once per connection, so we never want
3419 * to do it again
3420 */
3421
3422 (void)ktls_hold(tls);
3423 soref(so);
3424 tp->t_nic_ktls_xmit_dis = 1;
3425 SOCK_UNLOCK(so);
3426 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3427 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3428 }
3429
3430 void
ktls_session_to_xktls_onedir(const struct ktls_session * ktls,bool export_keys,struct xktls_session_onedir * xk)3431 ktls_session_to_xktls_onedir(const struct ktls_session *ktls, bool export_keys,
3432 struct xktls_session_onedir *xk)
3433 {
3434 if_t ifp;
3435 struct m_snd_tag *st;
3436
3437 xk->gen = ktls->gen;
3438 #define A(m) xk->m = ktls->params.m
3439 A(cipher_algorithm);
3440 A(auth_algorithm);
3441 A(cipher_key_len);
3442 A(auth_key_len);
3443 A(max_frame_len);
3444 A(tls_vmajor);
3445 A(tls_vminor);
3446 A(tls_hlen);
3447 A(tls_tlen);
3448 A(tls_bs);
3449 A(flags);
3450 if (export_keys) {
3451 memcpy(&xk->iv, &ktls->params.iv, XKTLS_SESSION_IV_BUF_LEN);
3452 A(iv_len);
3453 } else {
3454 memset(&xk->iv, 0, XKTLS_SESSION_IV_BUF_LEN);
3455 xk->iv_len = 0;
3456 }
3457 #undef A
3458 if ((st = ktls->snd_tag) != NULL &&
3459 (ifp = ktls->snd_tag->ifp) != NULL)
3460 strncpy(xk->ifnet, if_name(ifp), sizeof(xk->ifnet));
3461 }
3462
3463 void
ktls_session_copy_keys(const struct ktls_session * ktls,uint8_t * data,size_t * sz)3464 ktls_session_copy_keys(const struct ktls_session *ktls,
3465 uint8_t *data, size_t *sz)
3466 {
3467 size_t t, ta, tc;
3468
3469 if (ktls == NULL) {
3470 *sz = 0;
3471 return;
3472 }
3473 t = *sz;
3474 tc = MIN(t, ktls->params.cipher_key_len);
3475 if (data != NULL)
3476 memcpy(data, ktls->params.cipher_key, tc);
3477 ta = MIN(t - tc, ktls->params.auth_key_len);
3478 if (data != NULL)
3479 memcpy(data + tc, ktls->params.auth_key, ta);
3480 *sz = ta + tc;
3481 }
3482