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 #include <net/route.h> 64 #include <net/route/nhop.h> 65 #if defined(INET) || defined(INET6) 66 #include <netinet/in.h> 67 #include <netinet/in_pcb.h> 68 #endif 69 #include <netinet/tcp_var.h> 70 #ifdef TCP_OFFLOAD 71 #include <netinet/tcp_offload.h> 72 #endif 73 #include <opencrypto/xform.h> 74 #include <vm/uma_dbg.h> 75 #include <vm/vm.h> 76 #include <vm/vm_pageout.h> 77 #include <vm/vm_page.h> 78 79 struct ktls_wq { 80 struct mtx mtx; 81 STAILQ_HEAD(, mbuf) head; 82 bool running; 83 } __aligned(CACHE_LINE_SIZE); 84 85 static struct ktls_wq *ktls_wq; 86 static struct proc *ktls_proc; 87 LIST_HEAD(, ktls_crypto_backend) ktls_backends; 88 static struct rmlock ktls_backends_lock; 89 static uma_zone_t ktls_session_zone; 90 static uint16_t ktls_cpuid_lookup[MAXCPU]; 91 92 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 93 "Kernel TLS offload"); 94 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 95 "Kernel TLS offload stats"); 96 97 static int ktls_allow_unload; 98 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN, 99 &ktls_allow_unload, 0, "Allow software crypto modules to unload"); 100 101 #ifdef RSS 102 static int ktls_bind_threads = 1; 103 #else 104 static int ktls_bind_threads; 105 #endif 106 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 107 &ktls_bind_threads, 0, 108 "Bind crypto threads to cores or domains at boot"); 109 110 static u_int ktls_maxlen = 16384; 111 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN, 112 &ktls_maxlen, 0, "Maximum TLS record size"); 113 114 static int ktls_number_threads; 115 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 116 &ktls_number_threads, 0, 117 "Number of TLS threads in thread-pool"); 118 119 static bool ktls_offload_enable; 120 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW, 121 &ktls_offload_enable, 0, 122 "Enable support for kernel TLS offload"); 123 124 static bool ktls_cbc_enable = true; 125 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW, 126 &ktls_cbc_enable, 1, 127 "Enable Support of AES-CBC crypto for kernel TLS"); 128 129 static counter_u64_t ktls_tasks_active; 130 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 131 &ktls_tasks_active, "Number of active tasks"); 132 133 static counter_u64_t ktls_cnt_on; 134 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD, 135 &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto"); 136 137 static counter_u64_t ktls_offload_total; 138 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 139 CTLFLAG_RD, &ktls_offload_total, 140 "Total successful TLS setups (parameters set)"); 141 142 static counter_u64_t ktls_offload_enable_calls; 143 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 144 CTLFLAG_RD, &ktls_offload_enable_calls, 145 "Total number of TLS enable calls made"); 146 147 static counter_u64_t ktls_offload_active; 148 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 149 &ktls_offload_active, "Total Active TLS sessions"); 150 151 static counter_u64_t ktls_offload_failed_crypto; 152 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 153 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 154 155 static counter_u64_t ktls_switch_to_ifnet; 156 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 157 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 158 159 static counter_u64_t ktls_switch_to_sw; 160 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 161 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 162 163 static counter_u64_t ktls_switch_failed; 164 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 165 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 166 167 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 168 "Software TLS session stats"); 169 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 170 "Hardware (ifnet) TLS session stats"); 171 #ifdef TCP_OFFLOAD 172 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 173 "TOE TLS session stats"); 174 #endif 175 176 static counter_u64_t ktls_sw_cbc; 177 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, 178 "Active number of software TLS sessions using AES-CBC"); 179 180 static counter_u64_t ktls_sw_gcm; 181 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, 182 "Active number of software TLS sessions using AES-GCM"); 183 184 static counter_u64_t ktls_ifnet_cbc; 185 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, 186 &ktls_ifnet_cbc, 187 "Active number of ifnet TLS sessions using AES-CBC"); 188 189 static counter_u64_t ktls_ifnet_gcm; 190 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, 191 &ktls_ifnet_gcm, 192 "Active number of ifnet TLS sessions using AES-GCM"); 193 194 static counter_u64_t ktls_ifnet_reset; 195 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, 196 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); 197 198 static counter_u64_t ktls_ifnet_reset_dropped; 199 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, 200 &ktls_ifnet_reset_dropped, 201 "TLS sessions dropped after failing to update ifnet send tag"); 202 203 static counter_u64_t ktls_ifnet_reset_failed; 204 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, 205 &ktls_ifnet_reset_failed, 206 "TLS sessions that failed to allocate a new ifnet send tag"); 207 208 static int ktls_ifnet_permitted; 209 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, 210 &ktls_ifnet_permitted, 1, 211 "Whether to permit hardware (ifnet) TLS sessions"); 212 213 #ifdef TCP_OFFLOAD 214 static counter_u64_t ktls_toe_cbc; 215 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, 216 &ktls_toe_cbc, 217 "Active number of TOE TLS sessions using AES-CBC"); 218 219 static counter_u64_t ktls_toe_gcm; 220 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, 221 &ktls_toe_gcm, 222 "Active number of TOE TLS sessions using AES-GCM"); 223 #endif 224 225 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); 226 227 static void ktls_cleanup(struct ktls_session *tls); 228 #if defined(INET) || defined(INET6) 229 static void ktls_reset_send_tag(void *context, int pending); 230 #endif 231 static void ktls_work_thread(void *ctx); 232 233 int 234 ktls_crypto_backend_register(struct ktls_crypto_backend *be) 235 { 236 struct ktls_crypto_backend *curr_be, *tmp; 237 238 if (be->api_version != KTLS_API_VERSION) { 239 printf("KTLS: API version mismatch (%d vs %d) for %s\n", 240 be->api_version, KTLS_API_VERSION, 241 be->name); 242 return (EINVAL); 243 } 244 245 rm_wlock(&ktls_backends_lock); 246 printf("KTLS: Registering crypto method %s with prio %d\n", 247 be->name, be->prio); 248 if (LIST_EMPTY(&ktls_backends)) { 249 LIST_INSERT_HEAD(&ktls_backends, be, next); 250 } else { 251 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) { 252 if (curr_be->prio < be->prio) { 253 LIST_INSERT_BEFORE(curr_be, be, next); 254 break; 255 } 256 if (LIST_NEXT(curr_be, next) == NULL) { 257 LIST_INSERT_AFTER(curr_be, be, next); 258 break; 259 } 260 } 261 } 262 rm_wunlock(&ktls_backends_lock); 263 return (0); 264 } 265 266 int 267 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be) 268 { 269 struct ktls_crypto_backend *tmp; 270 271 /* 272 * Don't error if the backend isn't registered. This permits 273 * MOD_UNLOAD handlers to use this function unconditionally. 274 */ 275 rm_wlock(&ktls_backends_lock); 276 LIST_FOREACH(tmp, &ktls_backends, next) { 277 if (tmp == be) 278 break; 279 } 280 if (tmp == NULL) { 281 rm_wunlock(&ktls_backends_lock); 282 return (0); 283 } 284 285 if (!ktls_allow_unload) { 286 rm_wunlock(&ktls_backends_lock); 287 printf( 288 "KTLS: Deregistering crypto method %s is not supported\n", 289 be->name); 290 return (EBUSY); 291 } 292 293 if (be->use_count) { 294 rm_wunlock(&ktls_backends_lock); 295 return (EBUSY); 296 } 297 298 LIST_REMOVE(be, next); 299 rm_wunlock(&ktls_backends_lock); 300 return (0); 301 } 302 303 #if defined(INET) || defined(INET6) 304 static u_int 305 ktls_get_cpu(struct socket *so) 306 { 307 struct inpcb *inp; 308 u_int cpuid; 309 310 inp = sotoinpcb(so); 311 #ifdef RSS 312 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 313 if (cpuid != NETISR_CPUID_NONE) 314 return (cpuid); 315 #endif 316 /* 317 * Just use the flowid to shard connections in a repeatable 318 * fashion. Note that some crypto backends rely on the 319 * serialization provided by having the same connection use 320 * the same queue. 321 */ 322 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 323 return (cpuid); 324 } 325 #endif 326 327 static void 328 ktls_init(void *dummy __unused) 329 { 330 struct thread *td; 331 struct pcpu *pc; 332 cpuset_t mask; 333 int error, i; 334 335 ktls_tasks_active = counter_u64_alloc(M_WAITOK); 336 ktls_cnt_on = counter_u64_alloc(M_WAITOK); 337 ktls_offload_total = counter_u64_alloc(M_WAITOK); 338 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK); 339 ktls_offload_active = counter_u64_alloc(M_WAITOK); 340 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK); 341 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK); 342 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK); 343 ktls_switch_failed = counter_u64_alloc(M_WAITOK); 344 ktls_sw_cbc = counter_u64_alloc(M_WAITOK); 345 ktls_sw_gcm = counter_u64_alloc(M_WAITOK); 346 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK); 347 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK); 348 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK); 349 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK); 350 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK); 351 #ifdef TCP_OFFLOAD 352 ktls_toe_cbc = counter_u64_alloc(M_WAITOK); 353 ktls_toe_gcm = counter_u64_alloc(M_WAITOK); 354 #endif 355 356 rm_init(&ktls_backends_lock, "ktls backends"); 357 LIST_INIT(&ktls_backends); 358 359 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 360 M_WAITOK | M_ZERO); 361 362 ktls_session_zone = uma_zcreate("ktls_session", 363 sizeof(struct ktls_session), 364 NULL, NULL, NULL, NULL, 365 UMA_ALIGN_CACHE, 0); 366 367 /* 368 * Initialize the workqueues to run the TLS work. We create a 369 * work queue for each CPU. 370 */ 371 CPU_FOREACH(i) { 372 STAILQ_INIT(&ktls_wq[i].head); 373 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 374 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 375 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 376 if (error) 377 panic("Can't add KTLS thread %d error %d", i, error); 378 379 /* 380 * Bind threads to cores. If ktls_bind_threads is > 381 * 1, then we bind to the NUMA domain. 382 */ 383 if (ktls_bind_threads) { 384 if (ktls_bind_threads > 1) { 385 pc = pcpu_find(i); 386 CPU_COPY(&cpuset_domain[pc->pc_domain], &mask); 387 } else { 388 CPU_SETOF(i, &mask); 389 } 390 error = cpuset_setthread(td->td_tid, &mask); 391 if (error) 392 panic( 393 "Unable to bind KTLS thread for CPU %d error %d", 394 i, error); 395 } 396 ktls_cpuid_lookup[ktls_number_threads] = i; 397 ktls_number_threads++; 398 } 399 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 400 } 401 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL); 402 403 #if defined(INET) || defined(INET6) 404 static int 405 ktls_create_session(struct socket *so, struct tls_enable *en, 406 struct ktls_session **tlsp) 407 { 408 struct ktls_session *tls; 409 int error; 410 411 /* Only TLS 1.0 - 1.3 are supported. */ 412 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 413 return (EINVAL); 414 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 415 en->tls_vminor > TLS_MINOR_VER_THREE) 416 return (EINVAL); 417 418 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 419 return (EINVAL); 420 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 421 return (EINVAL); 422 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 423 return (EINVAL); 424 425 /* All supported algorithms require a cipher key. */ 426 if (en->cipher_key_len == 0) 427 return (EINVAL); 428 429 /* No flags are currently supported. */ 430 if (en->flags != 0) 431 return (EINVAL); 432 433 /* Common checks for supported algorithms. */ 434 switch (en->cipher_algorithm) { 435 case CRYPTO_AES_NIST_GCM_16: 436 /* 437 * auth_algorithm isn't used, but permit GMAC values 438 * for compatibility. 439 */ 440 switch (en->auth_algorithm) { 441 case 0: 442 #ifdef COMPAT_FREEBSD12 443 /* XXX: Really 13.0-current COMPAT. */ 444 case CRYPTO_AES_128_NIST_GMAC: 445 case CRYPTO_AES_192_NIST_GMAC: 446 case CRYPTO_AES_256_NIST_GMAC: 447 #endif 448 break; 449 default: 450 return (EINVAL); 451 } 452 if (en->auth_key_len != 0) 453 return (EINVAL); 454 if ((en->tls_vminor == TLS_MINOR_VER_TWO && 455 en->iv_len != TLS_AEAD_GCM_LEN) || 456 (en->tls_vminor == TLS_MINOR_VER_THREE && 457 en->iv_len != TLS_1_3_GCM_IV_LEN)) 458 return (EINVAL); 459 break; 460 case CRYPTO_AES_CBC: 461 switch (en->auth_algorithm) { 462 case CRYPTO_SHA1_HMAC: 463 /* 464 * TLS 1.0 requires an implicit IV. TLS 1.1+ 465 * all use explicit IVs. 466 */ 467 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 468 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 469 return (EINVAL); 470 break; 471 } 472 473 /* FALLTHROUGH */ 474 case CRYPTO_SHA2_256_HMAC: 475 case CRYPTO_SHA2_384_HMAC: 476 /* Ignore any supplied IV. */ 477 en->iv_len = 0; 478 break; 479 default: 480 return (EINVAL); 481 } 482 if (en->auth_key_len == 0) 483 return (EINVAL); 484 break; 485 default: 486 return (EINVAL); 487 } 488 489 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 490 491 counter_u64_add(ktls_offload_active, 1); 492 493 refcount_init(&tls->refcount, 1); 494 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 495 496 tls->wq_index = ktls_get_cpu(so); 497 498 tls->params.cipher_algorithm = en->cipher_algorithm; 499 tls->params.auth_algorithm = en->auth_algorithm; 500 tls->params.tls_vmajor = en->tls_vmajor; 501 tls->params.tls_vminor = en->tls_vminor; 502 tls->params.flags = en->flags; 503 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 504 505 /* Set the header and trailer lengths. */ 506 tls->params.tls_hlen = sizeof(struct tls_record_layer); 507 switch (en->cipher_algorithm) { 508 case CRYPTO_AES_NIST_GCM_16: 509 /* 510 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 511 * nonce. TLS 1.3 uses a 12 byte implicit IV. 512 */ 513 if (en->tls_vminor < TLS_MINOR_VER_THREE) 514 tls->params.tls_hlen += sizeof(uint64_t); 515 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 516 517 /* 518 * TLS 1.3 includes optional padding which we 519 * do not support, and also puts the "real" record 520 * type at the end of the encrypted data. 521 */ 522 if (en->tls_vminor == TLS_MINOR_VER_THREE) 523 tls->params.tls_tlen += sizeof(uint8_t); 524 525 tls->params.tls_bs = 1; 526 break; 527 case CRYPTO_AES_CBC: 528 switch (en->auth_algorithm) { 529 case CRYPTO_SHA1_HMAC: 530 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 531 /* Implicit IV, no nonce. */ 532 } else { 533 tls->params.tls_hlen += AES_BLOCK_LEN; 534 } 535 tls->params.tls_tlen = AES_BLOCK_LEN + 536 SHA1_HASH_LEN; 537 break; 538 case CRYPTO_SHA2_256_HMAC: 539 tls->params.tls_hlen += AES_BLOCK_LEN; 540 tls->params.tls_tlen = AES_BLOCK_LEN + 541 SHA2_256_HASH_LEN; 542 break; 543 case CRYPTO_SHA2_384_HMAC: 544 tls->params.tls_hlen += AES_BLOCK_LEN; 545 tls->params.tls_tlen = AES_BLOCK_LEN + 546 SHA2_384_HASH_LEN; 547 break; 548 default: 549 panic("invalid hmac"); 550 } 551 tls->params.tls_bs = AES_BLOCK_LEN; 552 break; 553 default: 554 panic("invalid cipher"); 555 } 556 557 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 558 ("TLS header length too long: %d", tls->params.tls_hlen)); 559 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 560 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 561 562 if (en->auth_key_len != 0) { 563 tls->params.auth_key_len = en->auth_key_len; 564 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 565 M_WAITOK); 566 error = copyin(en->auth_key, tls->params.auth_key, 567 en->auth_key_len); 568 if (error) 569 goto out; 570 } 571 572 tls->params.cipher_key_len = en->cipher_key_len; 573 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 574 error = copyin(en->cipher_key, tls->params.cipher_key, 575 en->cipher_key_len); 576 if (error) 577 goto out; 578 579 /* 580 * This holds the implicit portion of the nonce for GCM and 581 * the initial implicit IV for TLS 1.0. The explicit portions 582 * of the IV are generated in ktls_frame(). 583 */ 584 if (en->iv_len != 0) { 585 tls->params.iv_len = en->iv_len; 586 error = copyin(en->iv, tls->params.iv, en->iv_len); 587 if (error) 588 goto out; 589 590 /* 591 * For TLS 1.2, generate an 8-byte nonce as a counter 592 * to generate unique explicit IVs. 593 * 594 * Store this counter in the last 8 bytes of the IV 595 * array so that it is 8-byte aligned. 596 */ 597 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 598 en->tls_vminor == TLS_MINOR_VER_TWO) 599 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); 600 } 601 602 *tlsp = tls; 603 return (0); 604 605 out: 606 ktls_cleanup(tls); 607 return (error); 608 } 609 610 static struct ktls_session * 611 ktls_clone_session(struct ktls_session *tls) 612 { 613 struct ktls_session *tls_new; 614 615 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 616 617 counter_u64_add(ktls_offload_active, 1); 618 619 refcount_init(&tls_new->refcount, 1); 620 621 /* Copy fields from existing session. */ 622 tls_new->params = tls->params; 623 tls_new->wq_index = tls->wq_index; 624 625 /* Deep copy keys. */ 626 if (tls_new->params.auth_key != NULL) { 627 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 628 M_KTLS, M_WAITOK); 629 memcpy(tls_new->params.auth_key, tls->params.auth_key, 630 tls->params.auth_key_len); 631 } 632 633 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 634 M_WAITOK); 635 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 636 tls->params.cipher_key_len); 637 638 return (tls_new); 639 } 640 #endif 641 642 static void 643 ktls_cleanup(struct ktls_session *tls) 644 { 645 646 counter_u64_add(ktls_offload_active, -1); 647 switch (tls->mode) { 648 case TCP_TLS_MODE_SW: 649 MPASS(tls->be != NULL); 650 switch (tls->params.cipher_algorithm) { 651 case CRYPTO_AES_CBC: 652 counter_u64_add(ktls_sw_cbc, -1); 653 break; 654 case CRYPTO_AES_NIST_GCM_16: 655 counter_u64_add(ktls_sw_gcm, -1); 656 break; 657 } 658 tls->free(tls); 659 break; 660 case TCP_TLS_MODE_IFNET: 661 switch (tls->params.cipher_algorithm) { 662 case CRYPTO_AES_CBC: 663 counter_u64_add(ktls_ifnet_cbc, -1); 664 break; 665 case CRYPTO_AES_NIST_GCM_16: 666 counter_u64_add(ktls_ifnet_gcm, -1); 667 break; 668 } 669 m_snd_tag_rele(tls->snd_tag); 670 break; 671 #ifdef TCP_OFFLOAD 672 case TCP_TLS_MODE_TOE: 673 switch (tls->params.cipher_algorithm) { 674 case CRYPTO_AES_CBC: 675 counter_u64_add(ktls_toe_cbc, -1); 676 break; 677 case CRYPTO_AES_NIST_GCM_16: 678 counter_u64_add(ktls_toe_gcm, -1); 679 break; 680 } 681 break; 682 #endif 683 } 684 if (tls->params.auth_key != NULL) { 685 zfree(tls->params.auth_key, M_KTLS); 686 tls->params.auth_key = NULL; 687 tls->params.auth_key_len = 0; 688 } 689 if (tls->params.cipher_key != NULL) { 690 zfree(tls->params.cipher_key, M_KTLS); 691 tls->params.cipher_key = NULL; 692 tls->params.cipher_key_len = 0; 693 } 694 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 695 } 696 697 #if defined(INET) || defined(INET6) 698 699 #ifdef TCP_OFFLOAD 700 static int 701 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) 702 { 703 struct inpcb *inp; 704 struct tcpcb *tp; 705 int error; 706 707 inp = so->so_pcb; 708 INP_WLOCK(inp); 709 if (inp->inp_flags2 & INP_FREED) { 710 INP_WUNLOCK(inp); 711 return (ECONNRESET); 712 } 713 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 714 INP_WUNLOCK(inp); 715 return (ECONNRESET); 716 } 717 if (inp->inp_socket == NULL) { 718 INP_WUNLOCK(inp); 719 return (ECONNRESET); 720 } 721 tp = intotcpcb(inp); 722 if (tp->tod == NULL) { 723 INP_WUNLOCK(inp); 724 return (EOPNOTSUPP); 725 } 726 727 error = tcp_offload_alloc_tls_session(tp, tls, direction); 728 INP_WUNLOCK(inp); 729 if (error == 0) { 730 tls->mode = TCP_TLS_MODE_TOE; 731 switch (tls->params.cipher_algorithm) { 732 case CRYPTO_AES_CBC: 733 counter_u64_add(ktls_toe_cbc, 1); 734 break; 735 case CRYPTO_AES_NIST_GCM_16: 736 counter_u64_add(ktls_toe_gcm, 1); 737 break; 738 } 739 } 740 return (error); 741 } 742 #endif 743 744 /* 745 * Common code used when first enabling ifnet TLS on a connection or 746 * when allocating a new ifnet TLS session due to a routing change. 747 * This function allocates a new TLS send tag on whatever interface 748 * the connection is currently routed over. 749 */ 750 static int 751 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 752 struct m_snd_tag **mstp) 753 { 754 union if_snd_tag_alloc_params params; 755 struct ifnet *ifp; 756 struct nhop_object *nh; 757 struct tcpcb *tp; 758 int error; 759 760 INP_RLOCK(inp); 761 if (inp->inp_flags2 & INP_FREED) { 762 INP_RUNLOCK(inp); 763 return (ECONNRESET); 764 } 765 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 766 INP_RUNLOCK(inp); 767 return (ECONNRESET); 768 } 769 if (inp->inp_socket == NULL) { 770 INP_RUNLOCK(inp); 771 return (ECONNRESET); 772 } 773 tp = intotcpcb(inp); 774 775 /* 776 * Check administrative controls on ifnet TLS to determine if 777 * ifnet TLS should be denied. 778 * 779 * - Always permit 'force' requests. 780 * - ktls_ifnet_permitted == 0: always deny. 781 */ 782 if (!force && ktls_ifnet_permitted == 0) { 783 INP_RUNLOCK(inp); 784 return (ENXIO); 785 } 786 787 /* 788 * XXX: Use the cached route in the inpcb to find the 789 * interface. This should perhaps instead use 790 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 791 * enabled after a connection has completed key negotiation in 792 * userland, the cached route will be present in practice. 793 */ 794 nh = inp->inp_route.ro_nh; 795 if (nh == NULL) { 796 INP_RUNLOCK(inp); 797 return (ENXIO); 798 } 799 ifp = nh->nh_ifp; 800 if_ref(ifp); 801 802 params.hdr.type = IF_SND_TAG_TYPE_TLS; 803 params.hdr.flowid = inp->inp_flowid; 804 params.hdr.flowtype = inp->inp_flowtype; 805 params.hdr.numa_domain = inp->inp_numa_domain; 806 params.tls.inp = inp; 807 params.tls.tls = tls; 808 INP_RUNLOCK(inp); 809 810 if (ifp->if_snd_tag_alloc == NULL) { 811 error = EOPNOTSUPP; 812 goto out; 813 } 814 if ((ifp->if_capenable & IFCAP_NOMAP) == 0) { 815 error = EOPNOTSUPP; 816 goto out; 817 } 818 if (inp->inp_vflag & INP_IPV6) { 819 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { 820 error = EOPNOTSUPP; 821 goto out; 822 } 823 } else { 824 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { 825 error = EOPNOTSUPP; 826 goto out; 827 } 828 } 829 error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp); 830 out: 831 if_rele(ifp); 832 return (error); 833 } 834 835 static int 836 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force) 837 { 838 struct m_snd_tag *mst; 839 int error; 840 841 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 842 if (error == 0) { 843 tls->mode = TCP_TLS_MODE_IFNET; 844 tls->snd_tag = mst; 845 switch (tls->params.cipher_algorithm) { 846 case CRYPTO_AES_CBC: 847 counter_u64_add(ktls_ifnet_cbc, 1); 848 break; 849 case CRYPTO_AES_NIST_GCM_16: 850 counter_u64_add(ktls_ifnet_gcm, 1); 851 break; 852 } 853 } 854 return (error); 855 } 856 857 static int 858 ktls_try_sw(struct socket *so, struct ktls_session *tls) 859 { 860 struct rm_priotracker prio; 861 struct ktls_crypto_backend *be; 862 863 /* 864 * Choose the best software crypto backend. Backends are 865 * stored in sorted priority order (larget value == most 866 * important at the head of the list), so this just stops on 867 * the first backend that claims the session by returning 868 * success. 869 */ 870 if (ktls_allow_unload) 871 rm_rlock(&ktls_backends_lock, &prio); 872 LIST_FOREACH(be, &ktls_backends, next) { 873 if (be->try(so, tls) == 0) 874 break; 875 KASSERT(tls->cipher == NULL, 876 ("ktls backend leaked a cipher pointer")); 877 } 878 if (be != NULL) { 879 if (ktls_allow_unload) 880 be->use_count++; 881 tls->be = be; 882 } 883 if (ktls_allow_unload) 884 rm_runlock(&ktls_backends_lock, &prio); 885 if (be == NULL) 886 return (EOPNOTSUPP); 887 tls->mode = TCP_TLS_MODE_SW; 888 switch (tls->params.cipher_algorithm) { 889 case CRYPTO_AES_CBC: 890 counter_u64_add(ktls_sw_cbc, 1); 891 break; 892 case CRYPTO_AES_NIST_GCM_16: 893 counter_u64_add(ktls_sw_gcm, 1); 894 break; 895 } 896 return (0); 897 } 898 899 int 900 ktls_enable_rx(struct socket *so, struct tls_enable *en) 901 { 902 struct ktls_session *tls; 903 int error; 904 905 if (!ktls_offload_enable) 906 return (ENOTSUP); 907 908 counter_u64_add(ktls_offload_enable_calls, 1); 909 910 /* 911 * This should always be true since only the TCP socket option 912 * invokes this function. 913 */ 914 if (so->so_proto->pr_protocol != IPPROTO_TCP) 915 return (EINVAL); 916 917 /* 918 * XXX: Don't overwrite existing sessions. We should permit 919 * this to support rekeying in the future. 920 */ 921 if (so->so_rcv.sb_tls_info != NULL) 922 return (EALREADY); 923 924 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 925 return (ENOTSUP); 926 927 error = ktls_create_session(so, en, &tls); 928 if (error) 929 return (error); 930 931 /* TLS RX offload is only supported on TOE currently. */ 932 #ifdef TCP_OFFLOAD 933 error = ktls_try_toe(so, tls, KTLS_RX); 934 #else 935 error = EOPNOTSUPP; 936 #endif 937 938 if (error) { 939 ktls_cleanup(tls); 940 return (error); 941 } 942 943 /* Mark the socket as using TLS offload. */ 944 SOCKBUF_LOCK(&so->so_rcv); 945 so->so_rcv.sb_tls_info = tls; 946 SOCKBUF_UNLOCK(&so->so_rcv); 947 948 counter_u64_add(ktls_offload_total, 1); 949 950 return (0); 951 } 952 953 int 954 ktls_enable_tx(struct socket *so, struct tls_enable *en) 955 { 956 struct ktls_session *tls; 957 int error; 958 959 if (!ktls_offload_enable) 960 return (ENOTSUP); 961 962 counter_u64_add(ktls_offload_enable_calls, 1); 963 964 /* 965 * This should always be true since only the TCP socket option 966 * invokes this function. 967 */ 968 if (so->so_proto->pr_protocol != IPPROTO_TCP) 969 return (EINVAL); 970 971 /* 972 * XXX: Don't overwrite existing sessions. We should permit 973 * this to support rekeying in the future. 974 */ 975 if (so->so_snd.sb_tls_info != NULL) 976 return (EALREADY); 977 978 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 979 return (ENOTSUP); 980 981 /* TLS requires ext pgs */ 982 if (mb_use_ext_pgs == 0) 983 return (ENXIO); 984 985 error = ktls_create_session(so, en, &tls); 986 if (error) 987 return (error); 988 989 /* Prefer TOE -> ifnet TLS -> software TLS. */ 990 #ifdef TCP_OFFLOAD 991 error = ktls_try_toe(so, tls, KTLS_TX); 992 if (error) 993 #endif 994 error = ktls_try_ifnet(so, tls, false); 995 if (error) 996 error = ktls_try_sw(so, tls); 997 998 if (error) { 999 ktls_cleanup(tls); 1000 return (error); 1001 } 1002 1003 error = sblock(&so->so_snd, SBL_WAIT); 1004 if (error) { 1005 ktls_cleanup(tls); 1006 return (error); 1007 } 1008 1009 SOCKBUF_LOCK(&so->so_snd); 1010 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); 1011 so->so_snd.sb_tls_info = tls; 1012 if (tls->mode != TCP_TLS_MODE_SW) 1013 so->so_snd.sb_flags |= SB_TLS_IFNET; 1014 SOCKBUF_UNLOCK(&so->so_snd); 1015 sbunlock(&so->so_snd); 1016 1017 counter_u64_add(ktls_offload_total, 1); 1018 1019 return (0); 1020 } 1021 1022 int 1023 ktls_get_rx_mode(struct socket *so) 1024 { 1025 struct ktls_session *tls; 1026 struct inpcb *inp; 1027 int mode; 1028 1029 inp = so->so_pcb; 1030 INP_WLOCK_ASSERT(inp); 1031 SOCKBUF_LOCK(&so->so_rcv); 1032 tls = so->so_rcv.sb_tls_info; 1033 if (tls == NULL) 1034 mode = TCP_TLS_MODE_NONE; 1035 else 1036 mode = tls->mode; 1037 SOCKBUF_UNLOCK(&so->so_rcv); 1038 return (mode); 1039 } 1040 1041 int 1042 ktls_get_tx_mode(struct socket *so) 1043 { 1044 struct ktls_session *tls; 1045 struct inpcb *inp; 1046 int mode; 1047 1048 inp = so->so_pcb; 1049 INP_WLOCK_ASSERT(inp); 1050 SOCKBUF_LOCK(&so->so_snd); 1051 tls = so->so_snd.sb_tls_info; 1052 if (tls == NULL) 1053 mode = TCP_TLS_MODE_NONE; 1054 else 1055 mode = tls->mode; 1056 SOCKBUF_UNLOCK(&so->so_snd); 1057 return (mode); 1058 } 1059 1060 /* 1061 * Switch between SW and ifnet TLS sessions as requested. 1062 */ 1063 int 1064 ktls_set_tx_mode(struct socket *so, int mode) 1065 { 1066 struct ktls_session *tls, *tls_new; 1067 struct inpcb *inp; 1068 int error; 1069 1070 switch (mode) { 1071 case TCP_TLS_MODE_SW: 1072 case TCP_TLS_MODE_IFNET: 1073 break; 1074 default: 1075 return (EINVAL); 1076 } 1077 1078 inp = so->so_pcb; 1079 INP_WLOCK_ASSERT(inp); 1080 SOCKBUF_LOCK(&so->so_snd); 1081 tls = so->so_snd.sb_tls_info; 1082 if (tls == NULL) { 1083 SOCKBUF_UNLOCK(&so->so_snd); 1084 return (0); 1085 } 1086 1087 if (tls->mode == mode) { 1088 SOCKBUF_UNLOCK(&so->so_snd); 1089 return (0); 1090 } 1091 1092 tls = ktls_hold(tls); 1093 SOCKBUF_UNLOCK(&so->so_snd); 1094 INP_WUNLOCK(inp); 1095 1096 tls_new = ktls_clone_session(tls); 1097 1098 if (mode == TCP_TLS_MODE_IFNET) 1099 error = ktls_try_ifnet(so, tls_new, true); 1100 else 1101 error = ktls_try_sw(so, tls_new); 1102 if (error) { 1103 counter_u64_add(ktls_switch_failed, 1); 1104 ktls_free(tls_new); 1105 ktls_free(tls); 1106 INP_WLOCK(inp); 1107 return (error); 1108 } 1109 1110 error = sblock(&so->so_snd, SBL_WAIT); 1111 if (error) { 1112 counter_u64_add(ktls_switch_failed, 1); 1113 ktls_free(tls_new); 1114 ktls_free(tls); 1115 INP_WLOCK(inp); 1116 return (error); 1117 } 1118 1119 /* 1120 * If we raced with another session change, keep the existing 1121 * session. 1122 */ 1123 if (tls != so->so_snd.sb_tls_info) { 1124 counter_u64_add(ktls_switch_failed, 1); 1125 sbunlock(&so->so_snd); 1126 ktls_free(tls_new); 1127 ktls_free(tls); 1128 INP_WLOCK(inp); 1129 return (EBUSY); 1130 } 1131 1132 SOCKBUF_LOCK(&so->so_snd); 1133 so->so_snd.sb_tls_info = tls_new; 1134 if (tls_new->mode != TCP_TLS_MODE_SW) 1135 so->so_snd.sb_flags |= SB_TLS_IFNET; 1136 SOCKBUF_UNLOCK(&so->so_snd); 1137 sbunlock(&so->so_snd); 1138 1139 /* 1140 * Drop two references on 'tls'. The first is for the 1141 * ktls_hold() above. The second drops the reference from the 1142 * socket buffer. 1143 */ 1144 KASSERT(tls->refcount >= 2, ("too few references on old session")); 1145 ktls_free(tls); 1146 ktls_free(tls); 1147 1148 if (mode == TCP_TLS_MODE_IFNET) 1149 counter_u64_add(ktls_switch_to_ifnet, 1); 1150 else 1151 counter_u64_add(ktls_switch_to_sw, 1); 1152 1153 INP_WLOCK(inp); 1154 return (0); 1155 } 1156 1157 /* 1158 * Try to allocate a new TLS send tag. This task is scheduled when 1159 * ip_output detects a route change while trying to transmit a packet 1160 * holding a TLS record. If a new tag is allocated, replace the tag 1161 * in the TLS session. Subsequent packets on the connection will use 1162 * the new tag. If a new tag cannot be allocated, drop the 1163 * connection. 1164 */ 1165 static void 1166 ktls_reset_send_tag(void *context, int pending) 1167 { 1168 struct epoch_tracker et; 1169 struct ktls_session *tls; 1170 struct m_snd_tag *old, *new; 1171 struct inpcb *inp; 1172 struct tcpcb *tp; 1173 int error; 1174 1175 MPASS(pending == 1); 1176 1177 tls = context; 1178 inp = tls->inp; 1179 1180 /* 1181 * Free the old tag first before allocating a new one. 1182 * ip[6]_output_send() will treat a NULL send tag the same as 1183 * an ifp mismatch and drop packets until a new tag is 1184 * allocated. 1185 * 1186 * Write-lock the INP when changing tls->snd_tag since 1187 * ip[6]_output_send() holds a read-lock when reading the 1188 * pointer. 1189 */ 1190 INP_WLOCK(inp); 1191 old = tls->snd_tag; 1192 tls->snd_tag = NULL; 1193 INP_WUNLOCK(inp); 1194 if (old != NULL) 1195 m_snd_tag_rele(old); 1196 1197 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1198 1199 if (error == 0) { 1200 INP_WLOCK(inp); 1201 tls->snd_tag = new; 1202 mtx_pool_lock(mtxpool_sleep, tls); 1203 tls->reset_pending = false; 1204 mtx_pool_unlock(mtxpool_sleep, tls); 1205 if (!in_pcbrele_wlocked(inp)) 1206 INP_WUNLOCK(inp); 1207 1208 counter_u64_add(ktls_ifnet_reset, 1); 1209 1210 /* 1211 * XXX: Should we kick tcp_output explicitly now that 1212 * the send tag is fixed or just rely on timers? 1213 */ 1214 } else { 1215 NET_EPOCH_ENTER(et); 1216 INP_WLOCK(inp); 1217 if (!in_pcbrele_wlocked(inp)) { 1218 if (!(inp->inp_flags & INP_TIMEWAIT) && 1219 !(inp->inp_flags & INP_DROPPED)) { 1220 tp = intotcpcb(inp); 1221 CURVNET_SET(tp->t_vnet); 1222 tp = tcp_drop(tp, ECONNABORTED); 1223 CURVNET_RESTORE(); 1224 if (tp != NULL) 1225 INP_WUNLOCK(inp); 1226 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1227 } else 1228 INP_WUNLOCK(inp); 1229 } 1230 NET_EPOCH_EXIT(et); 1231 1232 counter_u64_add(ktls_ifnet_reset_failed, 1); 1233 1234 /* 1235 * Leave reset_pending true to avoid future tasks while 1236 * the socket goes away. 1237 */ 1238 } 1239 1240 ktls_free(tls); 1241 } 1242 1243 int 1244 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1245 { 1246 1247 if (inp == NULL) 1248 return (ENOBUFS); 1249 1250 INP_LOCK_ASSERT(inp); 1251 1252 /* 1253 * See if we should schedule a task to update the send tag for 1254 * this session. 1255 */ 1256 mtx_pool_lock(mtxpool_sleep, tls); 1257 if (!tls->reset_pending) { 1258 (void) ktls_hold(tls); 1259 in_pcbref(inp); 1260 tls->inp = inp; 1261 tls->reset_pending = true; 1262 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1263 } 1264 mtx_pool_unlock(mtxpool_sleep, tls); 1265 return (ENOBUFS); 1266 } 1267 #endif 1268 1269 void 1270 ktls_destroy(struct ktls_session *tls) 1271 { 1272 struct rm_priotracker prio; 1273 1274 ktls_cleanup(tls); 1275 if (tls->be != NULL && ktls_allow_unload) { 1276 rm_rlock(&ktls_backends_lock, &prio); 1277 tls->be->use_count--; 1278 rm_runlock(&ktls_backends_lock, &prio); 1279 } 1280 uma_zfree(ktls_session_zone, tls); 1281 } 1282 1283 void 1284 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1285 { 1286 1287 for (; m != NULL; m = m->m_next) { 1288 KASSERT((m->m_flags & M_EXTPG) != 0, 1289 ("ktls_seq: mapped mbuf %p", m)); 1290 1291 m->m_epg_seqno = sb->sb_tls_seqno; 1292 sb->sb_tls_seqno++; 1293 } 1294 } 1295 1296 /* 1297 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1298 * mbuf in the chain must be an unmapped mbuf. The payload of the 1299 * mbuf must be populated with the payload of each TLS record. 1300 * 1301 * The record_type argument specifies the TLS record type used when 1302 * populating the TLS header. 1303 * 1304 * The enq_count argument on return is set to the number of pages of 1305 * payload data for this entire chain that need to be encrypted via SW 1306 * encryption. The returned value should be passed to ktls_enqueue 1307 * when scheduling encryption of this chain of mbufs. 1308 */ 1309 void 1310 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1311 uint8_t record_type) 1312 { 1313 struct tls_record_layer *tlshdr; 1314 struct mbuf *m; 1315 uint64_t *noncep; 1316 uint16_t tls_len; 1317 int maxlen; 1318 1319 maxlen = tls->params.max_frame_len; 1320 *enq_cnt = 0; 1321 for (m = top; m != NULL; m = m->m_next) { 1322 /* 1323 * All mbufs in the chain should be non-empty TLS 1324 * records whose payload does not exceed the maximum 1325 * frame length. 1326 */ 1327 KASSERT(m->m_len <= maxlen && m->m_len > 0, 1328 ("ktls_frame: m %p len %d\n", m, m->m_len)); 1329 /* 1330 * TLS frames require unmapped mbufs to store session 1331 * info. 1332 */ 1333 KASSERT((m->m_flags & M_EXTPG) != 0, 1334 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top)); 1335 1336 tls_len = m->m_len; 1337 1338 /* Save a reference to the session. */ 1339 m->m_epg_tls = ktls_hold(tls); 1340 1341 m->m_epg_hdrlen = tls->params.tls_hlen; 1342 m->m_epg_trllen = tls->params.tls_tlen; 1343 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 1344 int bs, delta; 1345 1346 /* 1347 * AES-CBC pads messages to a multiple of the 1348 * block size. Note that the padding is 1349 * applied after the digest and the encryption 1350 * is done on the "plaintext || mac || padding". 1351 * At least one byte of padding is always 1352 * present. 1353 * 1354 * Compute the final trailer length assuming 1355 * at most one block of padding. 1356 * tls->params.sb_tls_tlen is the maximum 1357 * possible trailer length (padding + digest). 1358 * delta holds the number of excess padding 1359 * bytes if the maximum were used. Those 1360 * extra bytes are removed. 1361 */ 1362 bs = tls->params.tls_bs; 1363 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 1364 m->m_epg_trllen -= delta; 1365 } 1366 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; 1367 1368 /* Populate the TLS header. */ 1369 tlshdr = (void *)m->m_epg_hdr; 1370 tlshdr->tls_vmajor = tls->params.tls_vmajor; 1371 1372 /* 1373 * TLS 1.3 masquarades as TLS 1.2 with a record type 1374 * of TLS_RLTYPE_APP. 1375 */ 1376 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 1377 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 1378 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 1379 tlshdr->tls_type = TLS_RLTYPE_APP; 1380 /* save the real record type for later */ 1381 m->m_epg_record_type = record_type; 1382 m->m_epg_trail[0] = record_type; 1383 } else { 1384 tlshdr->tls_vminor = tls->params.tls_vminor; 1385 tlshdr->tls_type = record_type; 1386 } 1387 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 1388 1389 /* 1390 * Store nonces / explicit IVs after the end of the 1391 * TLS header. 1392 * 1393 * For GCM with TLS 1.2, an 8 byte nonce is copied 1394 * from the end of the IV. The nonce is then 1395 * incremented for use by the next record. 1396 * 1397 * For CBC, a random nonce is inserted for TLS 1.1+. 1398 */ 1399 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 1400 tls->params.tls_vminor == TLS_MINOR_VER_TWO) { 1401 noncep = (uint64_t *)(tls->params.iv + 8); 1402 be64enc(tlshdr + 1, *noncep); 1403 (*noncep)++; 1404 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 1405 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 1406 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 1407 1408 /* 1409 * When using SW encryption, mark the mbuf not ready. 1410 * It will be marked ready via sbready() after the 1411 * record has been encrypted. 1412 * 1413 * When using ifnet TLS, unencrypted TLS records are 1414 * sent down the stack to the NIC. 1415 */ 1416 if (tls->mode == TCP_TLS_MODE_SW) { 1417 m->m_flags |= M_NOTREADY; 1418 m->m_epg_nrdy = m->m_epg_npgs; 1419 *enq_cnt += m->m_epg_npgs; 1420 } 1421 } 1422 } 1423 1424 void 1425 ktls_enqueue_to_free(struct mbuf *m) 1426 { 1427 struct ktls_wq *wq; 1428 bool running; 1429 1430 /* Mark it for freeing. */ 1431 m->m_epg_flags |= EPG_FLAG_2FREE; 1432 wq = &ktls_wq[m->m_epg_tls->wq_index]; 1433 mtx_lock(&wq->mtx); 1434 STAILQ_INSERT_TAIL(&wq->head, m, m_epg_stailq); 1435 running = wq->running; 1436 mtx_unlock(&wq->mtx); 1437 if (!running) 1438 wakeup(wq); 1439 } 1440 1441 void 1442 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 1443 { 1444 struct ktls_wq *wq; 1445 bool running; 1446 1447 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == 1448 (M_EXTPG | M_NOTREADY)), 1449 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 1450 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 1451 1452 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); 1453 1454 m->m_epg_enc_cnt = page_count; 1455 1456 /* 1457 * Save a pointer to the socket. The caller is responsible 1458 * for taking an additional reference via soref(). 1459 */ 1460 m->m_epg_so = so; 1461 1462 wq = &ktls_wq[m->m_epg_tls->wq_index]; 1463 mtx_lock(&wq->mtx); 1464 STAILQ_INSERT_TAIL(&wq->head, m, m_epg_stailq); 1465 running = wq->running; 1466 mtx_unlock(&wq->mtx); 1467 if (!running) 1468 wakeup(wq); 1469 counter_u64_add(ktls_cnt_on, 1); 1470 } 1471 1472 static __noinline void 1473 ktls_encrypt(struct mbuf *top) 1474 { 1475 struct ktls_session *tls; 1476 struct socket *so; 1477 struct mbuf *m; 1478 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1479 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1480 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1481 vm_page_t pg; 1482 int error, i, len, npages, off, total_pages; 1483 bool is_anon; 1484 1485 so = top->m_epg_so; 1486 tls = top->m_epg_tls; 1487 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 1488 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 1489 #ifdef INVARIANTS 1490 top->m_epg_so = NULL; 1491 #endif 1492 total_pages = top->m_epg_enc_cnt; 1493 npages = 0; 1494 1495 /* 1496 * Encrypt the TLS records in the chain of mbufs starting with 1497 * 'top'. 'total_pages' gives us a total count of pages and is 1498 * used to know when we have finished encrypting the TLS 1499 * records originally queued with 'top'. 1500 * 1501 * NB: These mbufs are queued in the socket buffer and 1502 * 'm_next' is traversing the mbufs in the socket buffer. The 1503 * socket buffer lock is not held while traversing this chain. 1504 * Since the mbufs are all marked M_NOTREADY their 'm_next' 1505 * pointers should be stable. However, the 'm_next' of the 1506 * last mbuf encrypted is not necessarily NULL. It can point 1507 * to other mbufs appended while 'top' was on the TLS work 1508 * queue. 1509 * 1510 * Each mbuf holds an entire TLS record. 1511 */ 1512 error = 0; 1513 for (m = top; npages != total_pages; m = m->m_next) { 1514 KASSERT(m->m_epg_tls == tls, 1515 ("different TLS sessions in a single mbuf chain: %p vs %p", 1516 tls, m->m_epg_tls)); 1517 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == 1518 (M_EXTPG | M_NOTREADY), 1519 ("%p not unready & nomap mbuf (top = %p)\n", m, top)); 1520 KASSERT(npages + m->m_epg_npgs <= total_pages, 1521 ("page count mismatch: top %p, total_pages %d, m %p", top, 1522 total_pages, m)); 1523 1524 /* 1525 * Generate source and destination ivoecs to pass to 1526 * the SW encryption backend. For writable mbufs, the 1527 * destination iovec is a copy of the source and 1528 * encryption is done in place. For file-backed mbufs 1529 * (from sendfile), anonymous wired pages are 1530 * allocated and assigned to the destination iovec. 1531 */ 1532 is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0; 1533 1534 off = m->m_epg_1st_off; 1535 for (i = 0; i < m->m_epg_npgs; i++, off = 0) { 1536 len = m_epg_pagelen(m, i, off); 1537 src_iov[i].iov_len = len; 1538 src_iov[i].iov_base = 1539 (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) + 1540 off; 1541 1542 if (is_anon) { 1543 dst_iov[i].iov_base = src_iov[i].iov_base; 1544 dst_iov[i].iov_len = src_iov[i].iov_len; 1545 continue; 1546 } 1547 retry_page: 1548 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | 1549 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED); 1550 if (pg == NULL) { 1551 vm_wait(NULL); 1552 goto retry_page; 1553 } 1554 parray[i] = VM_PAGE_TO_PHYS(pg); 1555 dst_iov[i].iov_base = 1556 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off; 1557 dst_iov[i].iov_len = len; 1558 } 1559 1560 npages += i; 1561 1562 error = (*tls->sw_encrypt)(tls, 1563 (const struct tls_record_layer *)m->m_epg_hdr, 1564 m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno, 1565 m->m_epg_record_type); 1566 if (error) { 1567 counter_u64_add(ktls_offload_failed_crypto, 1); 1568 break; 1569 } 1570 1571 /* 1572 * For file-backed mbufs, release the file-backed 1573 * pages and replace them in the ext_pgs array with 1574 * the anonymous wired pages allocated above. 1575 */ 1576 if (!is_anon) { 1577 /* Free the old pages. */ 1578 m->m_ext.ext_free(m); 1579 1580 /* Replace them with the new pages. */ 1581 for (i = 0; i < m->m_epg_npgs; i++) 1582 m->m_epg_pa[i] = parray[i]; 1583 1584 /* Use the basic free routine. */ 1585 m->m_ext.ext_free = mb_free_mext_pgs; 1586 1587 /* Pages are now writable. */ 1588 m->m_epg_flags |= EPG_FLAG_ANON; 1589 } 1590 1591 /* 1592 * Drop a reference to the session now that it is no 1593 * longer needed. Existing code depends on encrypted 1594 * records having no associated session vs 1595 * yet-to-be-encrypted records having an associated 1596 * session. 1597 */ 1598 m->m_epg_tls = NULL; 1599 ktls_free(tls); 1600 } 1601 1602 CURVNET_SET(so->so_vnet); 1603 if (error == 0) { 1604 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages); 1605 } else { 1606 so->so_proto->pr_usrreqs->pru_abort(so); 1607 so->so_error = EIO; 1608 mb_free_notready(top, total_pages); 1609 } 1610 1611 SOCK_LOCK(so); 1612 sorele(so); 1613 CURVNET_RESTORE(); 1614 } 1615 1616 static void 1617 ktls_work_thread(void *ctx) 1618 { 1619 struct ktls_wq *wq = ctx; 1620 struct mbuf *m, *n; 1621 STAILQ_HEAD(, mbuf) local_head; 1622 1623 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 1624 fpu_kern_thread(0); 1625 #endif 1626 for (;;) { 1627 mtx_lock(&wq->mtx); 1628 while (STAILQ_EMPTY(&wq->head)) { 1629 wq->running = false; 1630 mtx_sleep(wq, &wq->mtx, 0, "-", 0); 1631 wq->running = true; 1632 } 1633 1634 STAILQ_INIT(&local_head); 1635 STAILQ_CONCAT(&local_head, &wq->head); 1636 mtx_unlock(&wq->mtx); 1637 1638 STAILQ_FOREACH_SAFE(m, &local_head, m_epg_stailq, n) { 1639 if (m->m_epg_flags & EPG_FLAG_2FREE) { 1640 ktls_free(m->m_epg_tls); 1641 uma_zfree(zone_mbuf, m); 1642 } else { 1643 ktls_encrypt(m); 1644 counter_u64_add(ktls_cnt_on, -1); 1645 } 1646 } 1647 } 1648 } 1649