1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2014-2019 Netflix Inc. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 #include "opt_rss.h" 34 35 #include <sys/param.h> 36 #include <sys/kernel.h> 37 #include <sys/ktls.h> 38 #include <sys/lock.h> 39 #include <sys/mbuf.h> 40 #include <sys/mutex.h> 41 #include <sys/rmlock.h> 42 #include <sys/proc.h> 43 #include <sys/protosw.h> 44 #include <sys/refcount.h> 45 #include <sys/smp.h> 46 #include <sys/socket.h> 47 #include <sys/socketvar.h> 48 #include <sys/sysctl.h> 49 #include <sys/taskqueue.h> 50 #include <sys/kthread.h> 51 #include <sys/uio.h> 52 #include <sys/vmmeter.h> 53 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 54 #include <machine/pcb.h> 55 #endif 56 #include <machine/vmparam.h> 57 #include <net/if.h> 58 #include <net/if_var.h> 59 #ifdef RSS 60 #include <net/netisr.h> 61 #include <net/nhop.h> 62 #include <net/rss_config.h> 63 #endif 64 #include <net/route.h> 65 #include <net/route/nhop.h> 66 #if defined(INET) || defined(INET6) 67 #include <netinet/in.h> 68 #include <netinet/in_pcb.h> 69 #endif 70 #include <netinet/tcp_var.h> 71 #ifdef TCP_OFFLOAD 72 #include <netinet/tcp_offload.h> 73 #endif 74 #include <opencrypto/xform.h> 75 #include <vm/uma_dbg.h> 76 #include <vm/vm.h> 77 #include <vm/vm_pageout.h> 78 #include <vm/vm_page.h> 79 80 struct ktls_wq { 81 struct mtx mtx; 82 STAILQ_HEAD(, mbuf_ext_pgs) head; 83 bool running; 84 } __aligned(CACHE_LINE_SIZE); 85 86 static struct ktls_wq *ktls_wq; 87 static struct proc *ktls_proc; 88 LIST_HEAD(, ktls_crypto_backend) ktls_backends; 89 static struct rmlock ktls_backends_lock; 90 static uma_zone_t ktls_session_zone; 91 static uint16_t ktls_cpuid_lookup[MAXCPU]; 92 93 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 94 "Kernel TLS offload"); 95 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 96 "Kernel TLS offload stats"); 97 98 static int ktls_allow_unload; 99 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN, 100 &ktls_allow_unload, 0, "Allow software crypto modules to unload"); 101 102 #ifdef RSS 103 static int ktls_bind_threads = 1; 104 #else 105 static int ktls_bind_threads; 106 #endif 107 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 108 &ktls_bind_threads, 0, 109 "Bind crypto threads to cores or domains at boot"); 110 111 static u_int ktls_maxlen = 16384; 112 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN, 113 &ktls_maxlen, 0, "Maximum TLS record size"); 114 115 static int ktls_number_threads; 116 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 117 &ktls_number_threads, 0, 118 "Number of TLS threads in thread-pool"); 119 120 static bool ktls_offload_enable; 121 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW, 122 &ktls_offload_enable, 0, 123 "Enable support for kernel TLS offload"); 124 125 static bool ktls_cbc_enable = true; 126 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW, 127 &ktls_cbc_enable, 1, 128 "Enable Support of AES-CBC crypto for kernel TLS"); 129 130 static counter_u64_t ktls_tasks_active; 131 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 132 &ktls_tasks_active, "Number of active tasks"); 133 134 static counter_u64_t ktls_cnt_on; 135 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD, 136 &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto"); 137 138 static counter_u64_t ktls_offload_total; 139 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 140 CTLFLAG_RD, &ktls_offload_total, 141 "Total successful TLS setups (parameters set)"); 142 143 static counter_u64_t ktls_offload_enable_calls; 144 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 145 CTLFLAG_RD, &ktls_offload_enable_calls, 146 "Total number of TLS enable calls made"); 147 148 static counter_u64_t ktls_offload_active; 149 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 150 &ktls_offload_active, "Total Active TLS sessions"); 151 152 static counter_u64_t ktls_offload_failed_crypto; 153 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 154 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 155 156 static counter_u64_t ktls_switch_to_ifnet; 157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 158 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 159 160 static counter_u64_t ktls_switch_to_sw; 161 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 162 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 163 164 static counter_u64_t ktls_switch_failed; 165 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 166 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 167 168 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 169 "Software TLS session stats"); 170 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 171 "Hardware (ifnet) TLS session stats"); 172 #ifdef TCP_OFFLOAD 173 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 174 "TOE TLS session stats"); 175 #endif 176 177 static counter_u64_t ktls_sw_cbc; 178 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, 179 "Active number of software TLS sessions using AES-CBC"); 180 181 static counter_u64_t ktls_sw_gcm; 182 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, 183 "Active number of software TLS sessions using AES-GCM"); 184 185 static counter_u64_t ktls_ifnet_cbc; 186 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, 187 &ktls_ifnet_cbc, 188 "Active number of ifnet TLS sessions using AES-CBC"); 189 190 static counter_u64_t ktls_ifnet_gcm; 191 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, 192 &ktls_ifnet_gcm, 193 "Active number of ifnet TLS sessions using AES-GCM"); 194 195 static counter_u64_t ktls_ifnet_reset; 196 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, 197 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); 198 199 static counter_u64_t ktls_ifnet_reset_dropped; 200 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, 201 &ktls_ifnet_reset_dropped, 202 "TLS sessions dropped after failing to update ifnet send tag"); 203 204 static counter_u64_t ktls_ifnet_reset_failed; 205 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, 206 &ktls_ifnet_reset_failed, 207 "TLS sessions that failed to allocate a new ifnet send tag"); 208 209 static int ktls_ifnet_permitted; 210 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, 211 &ktls_ifnet_permitted, 1, 212 "Whether to permit hardware (ifnet) TLS sessions"); 213 214 #ifdef TCP_OFFLOAD 215 static counter_u64_t ktls_toe_cbc; 216 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, 217 &ktls_toe_cbc, 218 "Active number of TOE TLS sessions using AES-CBC"); 219 220 static counter_u64_t ktls_toe_gcm; 221 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, 222 &ktls_toe_gcm, 223 "Active number of TOE TLS sessions using AES-GCM"); 224 #endif 225 226 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); 227 228 static void ktls_cleanup(struct ktls_session *tls); 229 #if defined(INET) || defined(INET6) 230 static void ktls_reset_send_tag(void *context, int pending); 231 #endif 232 static void ktls_work_thread(void *ctx); 233 234 int 235 ktls_crypto_backend_register(struct ktls_crypto_backend *be) 236 { 237 struct ktls_crypto_backend *curr_be, *tmp; 238 239 if (be->api_version != KTLS_API_VERSION) { 240 printf("KTLS: API version mismatch (%d vs %d) for %s\n", 241 be->api_version, KTLS_API_VERSION, 242 be->name); 243 return (EINVAL); 244 } 245 246 rm_wlock(&ktls_backends_lock); 247 printf("KTLS: Registering crypto method %s with prio %d\n", 248 be->name, be->prio); 249 if (LIST_EMPTY(&ktls_backends)) { 250 LIST_INSERT_HEAD(&ktls_backends, be, next); 251 } else { 252 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) { 253 if (curr_be->prio < be->prio) { 254 LIST_INSERT_BEFORE(curr_be, be, next); 255 break; 256 } 257 if (LIST_NEXT(curr_be, next) == NULL) { 258 LIST_INSERT_AFTER(curr_be, be, next); 259 break; 260 } 261 } 262 } 263 rm_wunlock(&ktls_backends_lock); 264 return (0); 265 } 266 267 int 268 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be) 269 { 270 struct ktls_crypto_backend *tmp; 271 272 /* 273 * Don't error if the backend isn't registered. This permits 274 * MOD_UNLOAD handlers to use this function unconditionally. 275 */ 276 rm_wlock(&ktls_backends_lock); 277 LIST_FOREACH(tmp, &ktls_backends, next) { 278 if (tmp == be) 279 break; 280 } 281 if (tmp == NULL) { 282 rm_wunlock(&ktls_backends_lock); 283 return (0); 284 } 285 286 if (!ktls_allow_unload) { 287 rm_wunlock(&ktls_backends_lock); 288 printf( 289 "KTLS: Deregistering crypto method %s is not supported\n", 290 be->name); 291 return (EBUSY); 292 } 293 294 if (be->use_count) { 295 rm_wunlock(&ktls_backends_lock); 296 return (EBUSY); 297 } 298 299 LIST_REMOVE(be, next); 300 rm_wunlock(&ktls_backends_lock); 301 return (0); 302 } 303 304 #if defined(INET) || defined(INET6) 305 static u_int 306 ktls_get_cpu(struct socket *so) 307 { 308 struct inpcb *inp; 309 u_int cpuid; 310 311 inp = sotoinpcb(so); 312 #ifdef RSS 313 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 314 if (cpuid != NETISR_CPUID_NONE) 315 return (cpuid); 316 #endif 317 /* 318 * Just use the flowid to shard connections in a repeatable 319 * fashion. Note that some crypto backends rely on the 320 * serialization provided by having the same connection use 321 * the same queue. 322 */ 323 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 324 return (cpuid); 325 } 326 #endif 327 328 static void 329 ktls_init(void *dummy __unused) 330 { 331 struct thread *td; 332 struct pcpu *pc; 333 cpuset_t mask; 334 int error, i; 335 336 ktls_tasks_active = counter_u64_alloc(M_WAITOK); 337 ktls_cnt_on = counter_u64_alloc(M_WAITOK); 338 ktls_offload_total = counter_u64_alloc(M_WAITOK); 339 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK); 340 ktls_offload_active = counter_u64_alloc(M_WAITOK); 341 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK); 342 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK); 343 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK); 344 ktls_switch_failed = counter_u64_alloc(M_WAITOK); 345 ktls_sw_cbc = counter_u64_alloc(M_WAITOK); 346 ktls_sw_gcm = counter_u64_alloc(M_WAITOK); 347 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK); 348 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK); 349 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK); 350 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK); 351 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK); 352 #ifdef TCP_OFFLOAD 353 ktls_toe_cbc = counter_u64_alloc(M_WAITOK); 354 ktls_toe_gcm = counter_u64_alloc(M_WAITOK); 355 #endif 356 357 rm_init(&ktls_backends_lock, "ktls backends"); 358 LIST_INIT(&ktls_backends); 359 360 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 361 M_WAITOK | M_ZERO); 362 363 ktls_session_zone = uma_zcreate("ktls_session", 364 sizeof(struct ktls_session), 365 NULL, NULL, NULL, NULL, 366 UMA_ALIGN_CACHE, 0); 367 368 /* 369 * Initialize the workqueues to run the TLS work. We create a 370 * work queue for each CPU. 371 */ 372 CPU_FOREACH(i) { 373 STAILQ_INIT(&ktls_wq[i].head); 374 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 375 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 376 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 377 if (error) 378 panic("Can't add KTLS thread %d error %d", i, error); 379 380 /* 381 * Bind threads to cores. If ktls_bind_threads is > 382 * 1, then we bind to the NUMA domain. 383 */ 384 if (ktls_bind_threads) { 385 if (ktls_bind_threads > 1) { 386 pc = pcpu_find(i); 387 CPU_COPY(&cpuset_domain[pc->pc_domain], &mask); 388 } else { 389 CPU_SETOF(i, &mask); 390 } 391 error = cpuset_setthread(td->td_tid, &mask); 392 if (error) 393 panic( 394 "Unable to bind KTLS thread for CPU %d error %d", 395 i, error); 396 } 397 ktls_cpuid_lookup[ktls_number_threads] = i; 398 ktls_number_threads++; 399 } 400 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 401 } 402 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL); 403 404 #if defined(INET) || defined(INET6) 405 static int 406 ktls_create_session(struct socket *so, struct tls_enable *en, 407 struct ktls_session **tlsp) 408 { 409 struct ktls_session *tls; 410 int error; 411 412 /* Only TLS 1.0 - 1.3 are supported. */ 413 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 414 return (EINVAL); 415 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 416 en->tls_vminor > TLS_MINOR_VER_THREE) 417 return (EINVAL); 418 419 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 420 return (EINVAL); 421 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 422 return (EINVAL); 423 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 424 return (EINVAL); 425 426 /* All supported algorithms require a cipher key. */ 427 if (en->cipher_key_len == 0) 428 return (EINVAL); 429 430 /* No flags are currently supported. */ 431 if (en->flags != 0) 432 return (EINVAL); 433 434 /* Common checks for supported algorithms. */ 435 switch (en->cipher_algorithm) { 436 case CRYPTO_AES_NIST_GCM_16: 437 /* 438 * auth_algorithm isn't used, but permit GMAC values 439 * for compatibility. 440 */ 441 switch (en->auth_algorithm) { 442 case 0: 443 #ifdef COMPAT_FREEBSD12 444 /* XXX: Really 13.0-current COMPAT. */ 445 case CRYPTO_AES_128_NIST_GMAC: 446 case CRYPTO_AES_192_NIST_GMAC: 447 case CRYPTO_AES_256_NIST_GMAC: 448 #endif 449 break; 450 default: 451 return (EINVAL); 452 } 453 if (en->auth_key_len != 0) 454 return (EINVAL); 455 if ((en->tls_vminor == TLS_MINOR_VER_TWO && 456 en->iv_len != TLS_AEAD_GCM_LEN) || 457 (en->tls_vminor == TLS_MINOR_VER_THREE && 458 en->iv_len != TLS_1_3_GCM_IV_LEN)) 459 return (EINVAL); 460 break; 461 case CRYPTO_AES_CBC: 462 switch (en->auth_algorithm) { 463 case CRYPTO_SHA1_HMAC: 464 /* 465 * TLS 1.0 requires an implicit IV. TLS 1.1+ 466 * all use explicit IVs. 467 */ 468 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 469 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 470 return (EINVAL); 471 break; 472 } 473 474 /* FALLTHROUGH */ 475 case CRYPTO_SHA2_256_HMAC: 476 case CRYPTO_SHA2_384_HMAC: 477 /* Ignore any supplied IV. */ 478 en->iv_len = 0; 479 break; 480 default: 481 return (EINVAL); 482 } 483 if (en->auth_key_len == 0) 484 return (EINVAL); 485 break; 486 default: 487 return (EINVAL); 488 } 489 490 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 491 492 counter_u64_add(ktls_offload_active, 1); 493 494 refcount_init(&tls->refcount, 1); 495 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 496 497 tls->wq_index = ktls_get_cpu(so); 498 499 tls->params.cipher_algorithm = en->cipher_algorithm; 500 tls->params.auth_algorithm = en->auth_algorithm; 501 tls->params.tls_vmajor = en->tls_vmajor; 502 tls->params.tls_vminor = en->tls_vminor; 503 tls->params.flags = en->flags; 504 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 505 506 /* Set the header and trailer lengths. */ 507 tls->params.tls_hlen = sizeof(struct tls_record_layer); 508 switch (en->cipher_algorithm) { 509 case CRYPTO_AES_NIST_GCM_16: 510 /* 511 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 512 * nonce. TLS 1.3 uses a 12 byte implicit IV. 513 */ 514 if (en->tls_vminor < TLS_MINOR_VER_THREE) 515 tls->params.tls_hlen += sizeof(uint64_t); 516 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 517 518 /* 519 * TLS 1.3 includes optional padding which we 520 * do not support, and also puts the "real" record 521 * type at the end of the encrypted data. 522 */ 523 if (en->tls_vminor == TLS_MINOR_VER_THREE) 524 tls->params.tls_tlen += sizeof(uint8_t); 525 526 tls->params.tls_bs = 1; 527 break; 528 case CRYPTO_AES_CBC: 529 switch (en->auth_algorithm) { 530 case CRYPTO_SHA1_HMAC: 531 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 532 /* Implicit IV, no nonce. */ 533 } else { 534 tls->params.tls_hlen += AES_BLOCK_LEN; 535 } 536 tls->params.tls_tlen = AES_BLOCK_LEN + 537 SHA1_HASH_LEN; 538 break; 539 case CRYPTO_SHA2_256_HMAC: 540 tls->params.tls_hlen += AES_BLOCK_LEN; 541 tls->params.tls_tlen = AES_BLOCK_LEN + 542 SHA2_256_HASH_LEN; 543 break; 544 case CRYPTO_SHA2_384_HMAC: 545 tls->params.tls_hlen += AES_BLOCK_LEN; 546 tls->params.tls_tlen = AES_BLOCK_LEN + 547 SHA2_384_HASH_LEN; 548 break; 549 default: 550 panic("invalid hmac"); 551 } 552 tls->params.tls_bs = AES_BLOCK_LEN; 553 break; 554 default: 555 panic("invalid cipher"); 556 } 557 558 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 559 ("TLS header length too long: %d", tls->params.tls_hlen)); 560 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 561 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 562 563 if (en->auth_key_len != 0) { 564 tls->params.auth_key_len = en->auth_key_len; 565 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 566 M_WAITOK); 567 error = copyin(en->auth_key, tls->params.auth_key, 568 en->auth_key_len); 569 if (error) 570 goto out; 571 } 572 573 tls->params.cipher_key_len = en->cipher_key_len; 574 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 575 error = copyin(en->cipher_key, tls->params.cipher_key, 576 en->cipher_key_len); 577 if (error) 578 goto out; 579 580 /* 581 * This holds the implicit portion of the nonce for GCM and 582 * the initial implicit IV for TLS 1.0. The explicit portions 583 * of the IV are generated in ktls_frame(). 584 */ 585 if (en->iv_len != 0) { 586 tls->params.iv_len = en->iv_len; 587 error = copyin(en->iv, tls->params.iv, en->iv_len); 588 if (error) 589 goto out; 590 591 /* 592 * For TLS 1.2, generate an 8-byte nonce as a counter 593 * to generate unique explicit IVs. 594 * 595 * Store this counter in the last 8 bytes of the IV 596 * array so that it is 8-byte aligned. 597 */ 598 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 599 en->tls_vminor == TLS_MINOR_VER_TWO) 600 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); 601 } 602 603 *tlsp = tls; 604 return (0); 605 606 out: 607 ktls_cleanup(tls); 608 return (error); 609 } 610 611 static struct ktls_session * 612 ktls_clone_session(struct ktls_session *tls) 613 { 614 struct ktls_session *tls_new; 615 616 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 617 618 counter_u64_add(ktls_offload_active, 1); 619 620 refcount_init(&tls_new->refcount, 1); 621 622 /* Copy fields from existing session. */ 623 tls_new->params = tls->params; 624 tls_new->wq_index = tls->wq_index; 625 626 /* Deep copy keys. */ 627 if (tls_new->params.auth_key != NULL) { 628 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 629 M_KTLS, M_WAITOK); 630 memcpy(tls_new->params.auth_key, tls->params.auth_key, 631 tls->params.auth_key_len); 632 } 633 634 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 635 M_WAITOK); 636 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 637 tls->params.cipher_key_len); 638 639 return (tls_new); 640 } 641 #endif 642 643 static void 644 ktls_cleanup(struct ktls_session *tls) 645 { 646 647 counter_u64_add(ktls_offload_active, -1); 648 switch (tls->mode) { 649 case TCP_TLS_MODE_SW: 650 MPASS(tls->be != NULL); 651 switch (tls->params.cipher_algorithm) { 652 case CRYPTO_AES_CBC: 653 counter_u64_add(ktls_sw_cbc, -1); 654 break; 655 case CRYPTO_AES_NIST_GCM_16: 656 counter_u64_add(ktls_sw_gcm, -1); 657 break; 658 } 659 tls->free(tls); 660 break; 661 case TCP_TLS_MODE_IFNET: 662 switch (tls->params.cipher_algorithm) { 663 case CRYPTO_AES_CBC: 664 counter_u64_add(ktls_ifnet_cbc, -1); 665 break; 666 case CRYPTO_AES_NIST_GCM_16: 667 counter_u64_add(ktls_ifnet_gcm, -1); 668 break; 669 } 670 m_snd_tag_rele(tls->snd_tag); 671 break; 672 #ifdef TCP_OFFLOAD 673 case TCP_TLS_MODE_TOE: 674 switch (tls->params.cipher_algorithm) { 675 case CRYPTO_AES_CBC: 676 counter_u64_add(ktls_toe_cbc, -1); 677 break; 678 case CRYPTO_AES_NIST_GCM_16: 679 counter_u64_add(ktls_toe_gcm, -1); 680 break; 681 } 682 break; 683 #endif 684 } 685 if (tls->params.auth_key != NULL) { 686 explicit_bzero(tls->params.auth_key, tls->params.auth_key_len); 687 free(tls->params.auth_key, M_KTLS); 688 tls->params.auth_key = NULL; 689 tls->params.auth_key_len = 0; 690 } 691 if (tls->params.cipher_key != NULL) { 692 explicit_bzero(tls->params.cipher_key, 693 tls->params.cipher_key_len); 694 free(tls->params.cipher_key, M_KTLS); 695 tls->params.cipher_key = NULL; 696 tls->params.cipher_key_len = 0; 697 } 698 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 699 } 700 701 #if defined(INET) || defined(INET6) 702 703 #ifdef TCP_OFFLOAD 704 static int 705 ktls_try_toe(struct socket *so, struct ktls_session *tls) 706 { 707 struct inpcb *inp; 708 struct tcpcb *tp; 709 int error; 710 711 inp = so->so_pcb; 712 INP_WLOCK(inp); 713 if (inp->inp_flags2 & INP_FREED) { 714 INP_WUNLOCK(inp); 715 return (ECONNRESET); 716 } 717 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 718 INP_WUNLOCK(inp); 719 return (ECONNRESET); 720 } 721 if (inp->inp_socket == NULL) { 722 INP_WUNLOCK(inp); 723 return (ECONNRESET); 724 } 725 tp = intotcpcb(inp); 726 if (tp->tod == NULL) { 727 INP_WUNLOCK(inp); 728 return (EOPNOTSUPP); 729 } 730 731 error = tcp_offload_alloc_tls_session(tp, tls); 732 INP_WUNLOCK(inp); 733 if (error == 0) { 734 tls->mode = TCP_TLS_MODE_TOE; 735 switch (tls->params.cipher_algorithm) { 736 case CRYPTO_AES_CBC: 737 counter_u64_add(ktls_toe_cbc, 1); 738 break; 739 case CRYPTO_AES_NIST_GCM_16: 740 counter_u64_add(ktls_toe_gcm, 1); 741 break; 742 } 743 } 744 return (error); 745 } 746 #endif 747 748 /* 749 * Common code used when first enabling ifnet TLS on a connection or 750 * when allocating a new ifnet TLS session due to a routing change. 751 * This function allocates a new TLS send tag on whatever interface 752 * the connection is currently routed over. 753 */ 754 static int 755 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 756 struct m_snd_tag **mstp) 757 { 758 union if_snd_tag_alloc_params params; 759 struct ifnet *ifp; 760 struct nhop_object *nh; 761 struct tcpcb *tp; 762 int error; 763 764 INP_RLOCK(inp); 765 if (inp->inp_flags2 & INP_FREED) { 766 INP_RUNLOCK(inp); 767 return (ECONNRESET); 768 } 769 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 770 INP_RUNLOCK(inp); 771 return (ECONNRESET); 772 } 773 if (inp->inp_socket == NULL) { 774 INP_RUNLOCK(inp); 775 return (ECONNRESET); 776 } 777 tp = intotcpcb(inp); 778 779 /* 780 * Check administrative controls on ifnet TLS to determine if 781 * ifnet TLS should be denied. 782 * 783 * - Always permit 'force' requests. 784 * - ktls_ifnet_permitted == 0: always deny. 785 */ 786 if (!force && ktls_ifnet_permitted == 0) { 787 INP_RUNLOCK(inp); 788 return (ENXIO); 789 } 790 791 /* 792 * XXX: Use the cached route in the inpcb to find the 793 * interface. This should perhaps instead use 794 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 795 * enabled after a connection has completed key negotiation in 796 * userland, the cached route will be present in practice. 797 */ 798 nh = inp->inp_route.ro_nh; 799 if (nh == NULL) { 800 INP_RUNLOCK(inp); 801 return (ENXIO); 802 } 803 ifp = nh->nh_ifp; 804 if_ref(ifp); 805 806 params.hdr.type = IF_SND_TAG_TYPE_TLS; 807 params.hdr.flowid = inp->inp_flowid; 808 params.hdr.flowtype = inp->inp_flowtype; 809 params.hdr.numa_domain = inp->inp_numa_domain; 810 params.tls.inp = inp; 811 params.tls.tls = tls; 812 INP_RUNLOCK(inp); 813 814 if (ifp->if_snd_tag_alloc == NULL) { 815 error = EOPNOTSUPP; 816 goto out; 817 } 818 if ((ifp->if_capenable & IFCAP_NOMAP) == 0) { 819 error = EOPNOTSUPP; 820 goto out; 821 } 822 if (inp->inp_vflag & INP_IPV6) { 823 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { 824 error = EOPNOTSUPP; 825 goto out; 826 } 827 } else { 828 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { 829 error = EOPNOTSUPP; 830 goto out; 831 } 832 } 833 error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp); 834 out: 835 if_rele(ifp); 836 return (error); 837 } 838 839 static int 840 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force) 841 { 842 struct m_snd_tag *mst; 843 int error; 844 845 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 846 if (error == 0) { 847 tls->mode = TCP_TLS_MODE_IFNET; 848 tls->snd_tag = mst; 849 switch (tls->params.cipher_algorithm) { 850 case CRYPTO_AES_CBC: 851 counter_u64_add(ktls_ifnet_cbc, 1); 852 break; 853 case CRYPTO_AES_NIST_GCM_16: 854 counter_u64_add(ktls_ifnet_gcm, 1); 855 break; 856 } 857 } 858 return (error); 859 } 860 861 static int 862 ktls_try_sw(struct socket *so, struct ktls_session *tls) 863 { 864 struct rm_priotracker prio; 865 struct ktls_crypto_backend *be; 866 867 /* 868 * Choose the best software crypto backend. Backends are 869 * stored in sorted priority order (larget value == most 870 * important at the head of the list), so this just stops on 871 * the first backend that claims the session by returning 872 * success. 873 */ 874 if (ktls_allow_unload) 875 rm_rlock(&ktls_backends_lock, &prio); 876 LIST_FOREACH(be, &ktls_backends, next) { 877 if (be->try(so, tls) == 0) 878 break; 879 KASSERT(tls->cipher == NULL, 880 ("ktls backend leaked a cipher pointer")); 881 } 882 if (be != NULL) { 883 if (ktls_allow_unload) 884 be->use_count++; 885 tls->be = be; 886 } 887 if (ktls_allow_unload) 888 rm_runlock(&ktls_backends_lock, &prio); 889 if (be == NULL) 890 return (EOPNOTSUPP); 891 tls->mode = TCP_TLS_MODE_SW; 892 switch (tls->params.cipher_algorithm) { 893 case CRYPTO_AES_CBC: 894 counter_u64_add(ktls_sw_cbc, 1); 895 break; 896 case CRYPTO_AES_NIST_GCM_16: 897 counter_u64_add(ktls_sw_gcm, 1); 898 break; 899 } 900 return (0); 901 } 902 903 int 904 ktls_enable_tx(struct socket *so, struct tls_enable *en) 905 { 906 struct ktls_session *tls; 907 int error; 908 909 if (!ktls_offload_enable) 910 return (ENOTSUP); 911 912 counter_u64_add(ktls_offload_enable_calls, 1); 913 914 /* 915 * This should always be true since only the TCP socket option 916 * invokes this function. 917 */ 918 if (so->so_proto->pr_protocol != IPPROTO_TCP) 919 return (EINVAL); 920 921 /* 922 * XXX: Don't overwrite existing sessions. We should permit 923 * this to support rekeying in the future. 924 */ 925 if (so->so_snd.sb_tls_info != NULL) 926 return (EALREADY); 927 928 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 929 return (ENOTSUP); 930 931 /* TLS requires ext pgs */ 932 if (mb_use_ext_pgs == 0) 933 return (ENXIO); 934 935 error = ktls_create_session(so, en, &tls); 936 if (error) 937 return (error); 938 939 /* Prefer TOE -> ifnet TLS -> software TLS. */ 940 #ifdef TCP_OFFLOAD 941 error = ktls_try_toe(so, tls); 942 if (error) 943 #endif 944 error = ktls_try_ifnet(so, tls, false); 945 if (error) 946 error = ktls_try_sw(so, tls); 947 948 if (error) { 949 ktls_cleanup(tls); 950 return (error); 951 } 952 953 error = sblock(&so->so_snd, SBL_WAIT); 954 if (error) { 955 ktls_cleanup(tls); 956 return (error); 957 } 958 959 SOCKBUF_LOCK(&so->so_snd); 960 so->so_snd.sb_tls_info = tls; 961 if (tls->mode != TCP_TLS_MODE_SW) 962 so->so_snd.sb_flags |= SB_TLS_IFNET; 963 SOCKBUF_UNLOCK(&so->so_snd); 964 sbunlock(&so->so_snd); 965 966 counter_u64_add(ktls_offload_total, 1); 967 968 return (0); 969 } 970 971 int 972 ktls_get_tx_mode(struct socket *so) 973 { 974 struct ktls_session *tls; 975 struct inpcb *inp; 976 int mode; 977 978 inp = so->so_pcb; 979 INP_WLOCK_ASSERT(inp); 980 SOCKBUF_LOCK(&so->so_snd); 981 tls = so->so_snd.sb_tls_info; 982 if (tls == NULL) 983 mode = TCP_TLS_MODE_NONE; 984 else 985 mode = tls->mode; 986 SOCKBUF_UNLOCK(&so->so_snd); 987 return (mode); 988 } 989 990 /* 991 * Switch between SW and ifnet TLS sessions as requested. 992 */ 993 int 994 ktls_set_tx_mode(struct socket *so, int mode) 995 { 996 struct ktls_session *tls, *tls_new; 997 struct inpcb *inp; 998 int error; 999 1000 switch (mode) { 1001 case TCP_TLS_MODE_SW: 1002 case TCP_TLS_MODE_IFNET: 1003 break; 1004 default: 1005 return (EINVAL); 1006 } 1007 1008 inp = so->so_pcb; 1009 INP_WLOCK_ASSERT(inp); 1010 SOCKBUF_LOCK(&so->so_snd); 1011 tls = so->so_snd.sb_tls_info; 1012 if (tls == NULL) { 1013 SOCKBUF_UNLOCK(&so->so_snd); 1014 return (0); 1015 } 1016 1017 if (tls->mode == mode) { 1018 SOCKBUF_UNLOCK(&so->so_snd); 1019 return (0); 1020 } 1021 1022 tls = ktls_hold(tls); 1023 SOCKBUF_UNLOCK(&so->so_snd); 1024 INP_WUNLOCK(inp); 1025 1026 tls_new = ktls_clone_session(tls); 1027 1028 if (mode == TCP_TLS_MODE_IFNET) 1029 error = ktls_try_ifnet(so, tls_new, true); 1030 else 1031 error = ktls_try_sw(so, tls_new); 1032 if (error) { 1033 counter_u64_add(ktls_switch_failed, 1); 1034 ktls_free(tls_new); 1035 ktls_free(tls); 1036 INP_WLOCK(inp); 1037 return (error); 1038 } 1039 1040 error = sblock(&so->so_snd, SBL_WAIT); 1041 if (error) { 1042 counter_u64_add(ktls_switch_failed, 1); 1043 ktls_free(tls_new); 1044 ktls_free(tls); 1045 INP_WLOCK(inp); 1046 return (error); 1047 } 1048 1049 /* 1050 * If we raced with another session change, keep the existing 1051 * session. 1052 */ 1053 if (tls != so->so_snd.sb_tls_info) { 1054 counter_u64_add(ktls_switch_failed, 1); 1055 sbunlock(&so->so_snd); 1056 ktls_free(tls_new); 1057 ktls_free(tls); 1058 INP_WLOCK(inp); 1059 return (EBUSY); 1060 } 1061 1062 SOCKBUF_LOCK(&so->so_snd); 1063 so->so_snd.sb_tls_info = tls_new; 1064 if (tls_new->mode != TCP_TLS_MODE_SW) 1065 so->so_snd.sb_flags |= SB_TLS_IFNET; 1066 SOCKBUF_UNLOCK(&so->so_snd); 1067 sbunlock(&so->so_snd); 1068 1069 /* 1070 * Drop two references on 'tls'. The first is for the 1071 * ktls_hold() above. The second drops the reference from the 1072 * socket buffer. 1073 */ 1074 KASSERT(tls->refcount >= 2, ("too few references on old session")); 1075 ktls_free(tls); 1076 ktls_free(tls); 1077 1078 if (mode == TCP_TLS_MODE_IFNET) 1079 counter_u64_add(ktls_switch_to_ifnet, 1); 1080 else 1081 counter_u64_add(ktls_switch_to_sw, 1); 1082 1083 INP_WLOCK(inp); 1084 return (0); 1085 } 1086 1087 /* 1088 * Try to allocate a new TLS send tag. This task is scheduled when 1089 * ip_output detects a route change while trying to transmit a packet 1090 * holding a TLS record. If a new tag is allocated, replace the tag 1091 * in the TLS session. Subsequent packets on the connection will use 1092 * the new tag. If a new tag cannot be allocated, drop the 1093 * connection. 1094 */ 1095 static void 1096 ktls_reset_send_tag(void *context, int pending) 1097 { 1098 struct epoch_tracker et; 1099 struct ktls_session *tls; 1100 struct m_snd_tag *old, *new; 1101 struct inpcb *inp; 1102 struct tcpcb *tp; 1103 int error; 1104 1105 MPASS(pending == 1); 1106 1107 tls = context; 1108 inp = tls->inp; 1109 1110 /* 1111 * Free the old tag first before allocating a new one. 1112 * ip[6]_output_send() will treat a NULL send tag the same as 1113 * an ifp mismatch and drop packets until a new tag is 1114 * allocated. 1115 * 1116 * Write-lock the INP when changing tls->snd_tag since 1117 * ip[6]_output_send() holds a read-lock when reading the 1118 * pointer. 1119 */ 1120 INP_WLOCK(inp); 1121 old = tls->snd_tag; 1122 tls->snd_tag = NULL; 1123 INP_WUNLOCK(inp); 1124 if (old != NULL) 1125 m_snd_tag_rele(old); 1126 1127 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1128 1129 if (error == 0) { 1130 INP_WLOCK(inp); 1131 tls->snd_tag = new; 1132 mtx_pool_lock(mtxpool_sleep, tls); 1133 tls->reset_pending = false; 1134 mtx_pool_unlock(mtxpool_sleep, tls); 1135 if (!in_pcbrele_wlocked(inp)) 1136 INP_WUNLOCK(inp); 1137 1138 counter_u64_add(ktls_ifnet_reset, 1); 1139 1140 /* 1141 * XXX: Should we kick tcp_output explicitly now that 1142 * the send tag is fixed or just rely on timers? 1143 */ 1144 } else { 1145 NET_EPOCH_ENTER(et); 1146 INP_WLOCK(inp); 1147 if (!in_pcbrele_wlocked(inp)) { 1148 if (!(inp->inp_flags & INP_TIMEWAIT) && 1149 !(inp->inp_flags & INP_DROPPED)) { 1150 tp = intotcpcb(inp); 1151 CURVNET_SET(tp->t_vnet); 1152 tp = tcp_drop(tp, ECONNABORTED); 1153 CURVNET_RESTORE(); 1154 if (tp != NULL) 1155 INP_WUNLOCK(inp); 1156 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1157 } else 1158 INP_WUNLOCK(inp); 1159 } 1160 NET_EPOCH_EXIT(et); 1161 1162 counter_u64_add(ktls_ifnet_reset_failed, 1); 1163 1164 /* 1165 * Leave reset_pending true to avoid future tasks while 1166 * the socket goes away. 1167 */ 1168 } 1169 1170 ktls_free(tls); 1171 } 1172 1173 int 1174 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1175 { 1176 1177 if (inp == NULL) 1178 return (ENOBUFS); 1179 1180 INP_LOCK_ASSERT(inp); 1181 1182 /* 1183 * See if we should schedule a task to update the send tag for 1184 * this session. 1185 */ 1186 mtx_pool_lock(mtxpool_sleep, tls); 1187 if (!tls->reset_pending) { 1188 (void) ktls_hold(tls); 1189 in_pcbref(inp); 1190 tls->inp = inp; 1191 tls->reset_pending = true; 1192 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1193 } 1194 mtx_pool_unlock(mtxpool_sleep, tls); 1195 return (ENOBUFS); 1196 } 1197 #endif 1198 1199 void 1200 ktls_destroy(struct ktls_session *tls) 1201 { 1202 struct rm_priotracker prio; 1203 1204 ktls_cleanup(tls); 1205 if (tls->be != NULL && ktls_allow_unload) { 1206 rm_rlock(&ktls_backends_lock, &prio); 1207 tls->be->use_count--; 1208 rm_runlock(&ktls_backends_lock, &prio); 1209 } 1210 uma_zfree(ktls_session_zone, tls); 1211 } 1212 1213 void 1214 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1215 { 1216 struct mbuf_ext_pgs *pgs; 1217 1218 for (; m != NULL; m = m->m_next) { 1219 KASSERT((m->m_flags & M_NOMAP) != 0, 1220 ("ktls_seq: mapped mbuf %p", m)); 1221 1222 pgs = &m->m_ext_pgs; 1223 pgs->seqno = sb->sb_tls_seqno; 1224 sb->sb_tls_seqno++; 1225 } 1226 } 1227 1228 /* 1229 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1230 * mbuf in the chain must be an unmapped mbuf. The payload of the 1231 * mbuf must be populated with the payload of each TLS record. 1232 * 1233 * The record_type argument specifies the TLS record type used when 1234 * populating the TLS header. 1235 * 1236 * The enq_count argument on return is set to the number of pages of 1237 * payload data for this entire chain that need to be encrypted via SW 1238 * encryption. The returned value should be passed to ktls_enqueue 1239 * when scheduling encryption of this chain of mbufs. 1240 */ 1241 void 1242 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1243 uint8_t record_type) 1244 { 1245 struct tls_record_layer *tlshdr; 1246 struct mbuf *m; 1247 struct mbuf_ext_pgs *pgs; 1248 uint64_t *noncep; 1249 uint16_t tls_len; 1250 int maxlen; 1251 1252 maxlen = tls->params.max_frame_len; 1253 *enq_cnt = 0; 1254 for (m = top; m != NULL; m = m->m_next) { 1255 /* 1256 * All mbufs in the chain should be non-empty TLS 1257 * records whose payload does not exceed the maximum 1258 * frame length. 1259 */ 1260 KASSERT(m->m_len <= maxlen && m->m_len > 0, 1261 ("ktls_frame: m %p len %d\n", m, m->m_len)); 1262 /* 1263 * TLS frames require unmapped mbufs to store session 1264 * info. 1265 */ 1266 KASSERT((m->m_flags & M_NOMAP) != 0, 1267 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top)); 1268 1269 tls_len = m->m_len; 1270 pgs = &m->m_ext_pgs; 1271 1272 /* Save a reference to the session. */ 1273 pgs->tls = ktls_hold(tls); 1274 1275 pgs->hdr_len = tls->params.tls_hlen; 1276 pgs->trail_len = tls->params.tls_tlen; 1277 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 1278 int bs, delta; 1279 1280 /* 1281 * AES-CBC pads messages to a multiple of the 1282 * block size. Note that the padding is 1283 * applied after the digest and the encryption 1284 * is done on the "plaintext || mac || padding". 1285 * At least one byte of padding is always 1286 * present. 1287 * 1288 * Compute the final trailer length assuming 1289 * at most one block of padding. 1290 * tls->params.sb_tls_tlen is the maximum 1291 * possible trailer length (padding + digest). 1292 * delta holds the number of excess padding 1293 * bytes if the maximum were used. Those 1294 * extra bytes are removed. 1295 */ 1296 bs = tls->params.tls_bs; 1297 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 1298 pgs->trail_len -= delta; 1299 } 1300 m->m_len += pgs->hdr_len + pgs->trail_len; 1301 1302 /* Populate the TLS header. */ 1303 tlshdr = (void *)pgs->m_epg_hdr; 1304 tlshdr->tls_vmajor = tls->params.tls_vmajor; 1305 1306 /* 1307 * TLS 1.3 masquarades as TLS 1.2 with a record type 1308 * of TLS_RLTYPE_APP. 1309 */ 1310 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 1311 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 1312 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 1313 tlshdr->tls_type = TLS_RLTYPE_APP; 1314 /* save the real record type for later */ 1315 pgs->record_type = record_type; 1316 pgs->m_epg_trail[0] = record_type; 1317 } else { 1318 tlshdr->tls_vminor = tls->params.tls_vminor; 1319 tlshdr->tls_type = record_type; 1320 } 1321 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 1322 1323 /* 1324 * Store nonces / explicit IVs after the end of the 1325 * TLS header. 1326 * 1327 * For GCM with TLS 1.2, an 8 byte nonce is copied 1328 * from the end of the IV. The nonce is then 1329 * incremented for use by the next record. 1330 * 1331 * For CBC, a random nonce is inserted for TLS 1.1+. 1332 */ 1333 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 1334 tls->params.tls_vminor == TLS_MINOR_VER_TWO) { 1335 noncep = (uint64_t *)(tls->params.iv + 8); 1336 be64enc(tlshdr + 1, *noncep); 1337 (*noncep)++; 1338 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 1339 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 1340 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 1341 1342 /* 1343 * When using SW encryption, mark the mbuf not ready. 1344 * It will be marked ready via sbready() after the 1345 * record has been encrypted. 1346 * 1347 * When using ifnet TLS, unencrypted TLS records are 1348 * sent down the stack to the NIC. 1349 */ 1350 if (tls->mode == TCP_TLS_MODE_SW) { 1351 m->m_flags |= M_NOTREADY; 1352 pgs->nrdy = pgs->npgs; 1353 *enq_cnt += pgs->npgs; 1354 } 1355 } 1356 } 1357 1358 void 1359 ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs) 1360 { 1361 struct ktls_wq *wq; 1362 bool running; 1363 1364 /* Mark it for freeing. */ 1365 pgs->mbuf = NULL; 1366 wq = &ktls_wq[pgs->tls->wq_index]; 1367 mtx_lock(&wq->mtx); 1368 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq); 1369 running = wq->running; 1370 mtx_unlock(&wq->mtx); 1371 if (!running) 1372 wakeup(wq); 1373 } 1374 1375 void 1376 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 1377 { 1378 struct mbuf_ext_pgs *pgs; 1379 struct ktls_wq *wq; 1380 bool running; 1381 1382 KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) == 1383 (M_NOMAP | M_NOTREADY)), 1384 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 1385 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 1386 1387 pgs = &m->m_ext_pgs; 1388 1389 KASSERT(pgs->tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); 1390 1391 pgs->enc_cnt = page_count; 1392 pgs->mbuf = m; 1393 1394 /* 1395 * Save a pointer to the socket. The caller is responsible 1396 * for taking an additional reference via soref(). 1397 */ 1398 pgs->so = so; 1399 1400 wq = &ktls_wq[pgs->tls->wq_index]; 1401 mtx_lock(&wq->mtx); 1402 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq); 1403 running = wq->running; 1404 mtx_unlock(&wq->mtx); 1405 if (!running) 1406 wakeup(wq); 1407 counter_u64_add(ktls_cnt_on, 1); 1408 } 1409 1410 static __noinline void 1411 ktls_encrypt(struct mbuf_ext_pgs *pgs) 1412 { 1413 struct ktls_session *tls; 1414 struct socket *so; 1415 struct mbuf *m, *top; 1416 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1417 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1418 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1419 vm_page_t pg; 1420 int error, i, len, npages, off, total_pages; 1421 bool is_anon; 1422 1423 so = pgs->so; 1424 tls = pgs->tls; 1425 top = pgs->mbuf; 1426 KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs)); 1427 KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs)); 1428 #ifdef INVARIANTS 1429 pgs->so = NULL; 1430 pgs->mbuf = NULL; 1431 #endif 1432 total_pages = pgs->enc_cnt; 1433 npages = 0; 1434 1435 /* 1436 * Encrypt the TLS records in the chain of mbufs starting with 1437 * 'top'. 'total_pages' gives us a total count of pages and is 1438 * used to know when we have finished encrypting the TLS 1439 * records originally queued with 'top'. 1440 * 1441 * NB: These mbufs are queued in the socket buffer and 1442 * 'm_next' is traversing the mbufs in the socket buffer. The 1443 * socket buffer lock is not held while traversing this chain. 1444 * Since the mbufs are all marked M_NOTREADY their 'm_next' 1445 * pointers should be stable. However, the 'm_next' of the 1446 * last mbuf encrypted is not necessarily NULL. It can point 1447 * to other mbufs appended while 'top' was on the TLS work 1448 * queue. 1449 * 1450 * Each mbuf holds an entire TLS record. 1451 */ 1452 error = 0; 1453 for (m = top; npages != total_pages; m = m->m_next) { 1454 pgs = &m->m_ext_pgs; 1455 1456 KASSERT(pgs->tls == tls, 1457 ("different TLS sessions in a single mbuf chain: %p vs %p", 1458 tls, pgs->tls)); 1459 KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) == 1460 (M_NOMAP | M_NOTREADY), 1461 ("%p not unready & nomap mbuf (top = %p)\n", m, top)); 1462 KASSERT(npages + pgs->npgs <= total_pages, 1463 ("page count mismatch: top %p, total_pages %d, m %p", top, 1464 total_pages, m)); 1465 1466 /* 1467 * Generate source and destination ivoecs to pass to 1468 * the SW encryption backend. For writable mbufs, the 1469 * destination iovec is a copy of the source and 1470 * encryption is done in place. For file-backed mbufs 1471 * (from sendfile), anonymous wired pages are 1472 * allocated and assigned to the destination iovec. 1473 */ 1474 is_anon = (pgs->flags & MBUF_PEXT_FLAG_ANON) != 0; 1475 1476 off = pgs->first_pg_off; 1477 for (i = 0; i < pgs->npgs; i++, off = 0) { 1478 len = mbuf_ext_pg_len(pgs, i, off); 1479 src_iov[i].iov_len = len; 1480 src_iov[i].iov_base = 1481 (char *)(void *)PHYS_TO_DMAP(pgs->m_epg_pa[i]) + 1482 off; 1483 1484 if (is_anon) { 1485 dst_iov[i].iov_base = src_iov[i].iov_base; 1486 dst_iov[i].iov_len = src_iov[i].iov_len; 1487 continue; 1488 } 1489 retry_page: 1490 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | 1491 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED); 1492 if (pg == NULL) { 1493 vm_wait(NULL); 1494 goto retry_page; 1495 } 1496 parray[i] = VM_PAGE_TO_PHYS(pg); 1497 dst_iov[i].iov_base = 1498 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off; 1499 dst_iov[i].iov_len = len; 1500 } 1501 1502 npages += i; 1503 1504 error = (*tls->sw_encrypt)(tls, 1505 (const struct tls_record_layer *)pgs->m_epg_hdr, 1506 pgs->m_epg_trail, src_iov, dst_iov, i, pgs->seqno, 1507 pgs->record_type); 1508 if (error) { 1509 counter_u64_add(ktls_offload_failed_crypto, 1); 1510 break; 1511 } 1512 1513 /* 1514 * For file-backed mbufs, release the file-backed 1515 * pages and replace them in the ext_pgs array with 1516 * the anonymous wired pages allocated above. 1517 */ 1518 if (!is_anon) { 1519 /* Free the old pages. */ 1520 m->m_ext.ext_free(m); 1521 1522 /* Replace them with the new pages. */ 1523 for (i = 0; i < pgs->npgs; i++) 1524 pgs->m_epg_pa[i] = parray[i]; 1525 1526 /* Use the basic free routine. */ 1527 m->m_ext.ext_free = mb_free_mext_pgs; 1528 1529 /* Pages are now writable. */ 1530 pgs->flags |= MBUF_PEXT_FLAG_ANON; 1531 } 1532 1533 /* 1534 * Drop a reference to the session now that it is no 1535 * longer needed. Existing code depends on encrypted 1536 * records having no associated session vs 1537 * yet-to-be-encrypted records having an associated 1538 * session. 1539 */ 1540 pgs->tls = NULL; 1541 ktls_free(tls); 1542 } 1543 1544 CURVNET_SET(so->so_vnet); 1545 if (error == 0) { 1546 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages); 1547 } else { 1548 so->so_proto->pr_usrreqs->pru_abort(so); 1549 so->so_error = EIO; 1550 mb_free_notready(top, total_pages); 1551 } 1552 1553 SOCK_LOCK(so); 1554 sorele(so); 1555 CURVNET_RESTORE(); 1556 } 1557 1558 static void 1559 ktls_work_thread(void *ctx) 1560 { 1561 struct ktls_wq *wq = ctx; 1562 struct mbuf_ext_pgs *p, *n; 1563 struct ktls_session *tls; 1564 struct mbuf *m; 1565 STAILQ_HEAD(, mbuf_ext_pgs) local_head; 1566 1567 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 1568 fpu_kern_thread(0); 1569 #endif 1570 for (;;) { 1571 mtx_lock(&wq->mtx); 1572 while (STAILQ_EMPTY(&wq->head)) { 1573 wq->running = false; 1574 mtx_sleep(wq, &wq->mtx, 0, "-", 0); 1575 wq->running = true; 1576 } 1577 1578 STAILQ_INIT(&local_head); 1579 STAILQ_CONCAT(&local_head, &wq->head); 1580 mtx_unlock(&wq->mtx); 1581 1582 STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) { 1583 if (p->mbuf != NULL) { 1584 ktls_encrypt(p); 1585 counter_u64_add(ktls_cnt_on, -1); 1586 } else { 1587 tls = p->tls; 1588 ktls_free(tls); 1589 m = __containerof(p, struct mbuf, m_ext_pgs); 1590 uma_zfree(zone_mbuf, m); 1591 } 1592 } 1593 } 1594 } 1595