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