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 #ifdef RSS 58 #include <net/netisr.h> 59 #include <net/rss_config.h> 60 #endif 61 #if defined(INET) || defined(INET6) 62 #include <netinet/in.h> 63 #include <netinet/in_pcb.h> 64 #endif 65 #include <netinet/tcp_var.h> 66 #include <opencrypto/xform.h> 67 #include <vm/uma_dbg.h> 68 #include <vm/vm.h> 69 #include <vm/vm_pageout.h> 70 #include <vm/vm_page.h> 71 72 struct ktls_wq { 73 struct mtx mtx; 74 STAILQ_HEAD(, mbuf_ext_pgs) head; 75 bool running; 76 } __aligned(CACHE_LINE_SIZE); 77 78 static struct ktls_wq *ktls_wq; 79 static struct proc *ktls_proc; 80 LIST_HEAD(, ktls_crypto_backend) ktls_backends; 81 static struct rmlock ktls_backends_lock; 82 static uma_zone_t ktls_session_zone; 83 static uint16_t ktls_cpuid_lookup[MAXCPU]; 84 85 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW, 0, 86 "Kernel TLS offload"); 87 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW, 0, 88 "Kernel TLS offload stats"); 89 90 static int ktls_allow_unload; 91 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN, 92 &ktls_allow_unload, 0, "Allow software crypto modules to unload"); 93 94 #ifdef RSS 95 static int ktls_bind_threads = 1; 96 #else 97 static int ktls_bind_threads; 98 #endif 99 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 100 &ktls_bind_threads, 0, 101 "Bind crypto threads to cores or domains at boot"); 102 103 static u_int ktls_maxlen = 16384; 104 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN, 105 &ktls_maxlen, 0, "Maximum TLS record size"); 106 107 static int ktls_number_threads; 108 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 109 &ktls_number_threads, 0, 110 "Number of TLS threads in thread-pool"); 111 112 static bool ktls_offload_enable; 113 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW, 114 &ktls_offload_enable, 0, 115 "Enable support for kernel TLS offload"); 116 117 static bool ktls_cbc_enable = true; 118 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW, 119 &ktls_cbc_enable, 1, 120 "Enable Support of AES-CBC crypto for kernel TLS"); 121 122 static counter_u64_t ktls_tasks_active; 123 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 124 &ktls_tasks_active, "Number of active tasks"); 125 126 static counter_u64_t ktls_cnt_on; 127 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD, 128 &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto"); 129 130 static counter_u64_t ktls_offload_total; 131 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 132 CTLFLAG_RD, &ktls_offload_total, 133 "Total successful TLS setups (parameters set)"); 134 135 static counter_u64_t ktls_offload_enable_calls; 136 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 137 CTLFLAG_RD, &ktls_offload_enable_calls, 138 "Total number of TLS enable calls made"); 139 140 static counter_u64_t ktls_offload_active; 141 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 142 &ktls_offload_active, "Total Active TLS sessions"); 143 144 static counter_u64_t ktls_offload_failed_crypto; 145 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 146 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 147 148 static counter_u64_t ktls_switch_to_ifnet; 149 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 150 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 151 152 static counter_u64_t ktls_switch_to_sw; 153 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 154 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 155 156 static counter_u64_t ktls_switch_failed; 157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 158 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 159 160 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD, 0, 161 "Software TLS session stats"); 162 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD, 0, 163 "Hardware (ifnet) TLS session stats"); 164 165 static counter_u64_t ktls_sw_cbc; 166 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, 167 "Active number of software TLS sessions using AES-CBC"); 168 169 static counter_u64_t ktls_sw_gcm; 170 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, 171 "Active number of software TLS sessions using AES-GCM"); 172 173 static counter_u64_t ktls_ifnet_cbc; 174 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, 175 &ktls_ifnet_cbc, 176 "Active number of ifnet TLS sessions using AES-CBC"); 177 178 static counter_u64_t ktls_ifnet_gcm; 179 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, 180 &ktls_ifnet_gcm, 181 "Active number of ifnet TLS sessions using AES-GCM"); 182 183 static counter_u64_t ktls_ifnet_reset; 184 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, 185 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); 186 187 static counter_u64_t ktls_ifnet_reset_dropped; 188 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, 189 &ktls_ifnet_reset_dropped, 190 "TLS sessions dropped after failing to update ifnet send tag"); 191 192 static counter_u64_t ktls_ifnet_reset_failed; 193 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, 194 &ktls_ifnet_reset_failed, 195 "TLS sessions that failed to allocate a new ifnet send tag"); 196 197 static int ktls_ifnet_permitted; 198 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, 199 &ktls_ifnet_permitted, 1, 200 "Whether to permit hardware (ifnet) TLS sessions"); 201 202 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); 203 204 static void ktls_cleanup(struct ktls_session *tls); 205 #if defined(INET) || defined(INET6) 206 static void ktls_reset_send_tag(void *context, int pending); 207 #endif 208 static void ktls_work_thread(void *ctx); 209 210 int 211 ktls_crypto_backend_register(struct ktls_crypto_backend *be) 212 { 213 struct ktls_crypto_backend *curr_be, *tmp; 214 215 if (be->api_version != KTLS_API_VERSION) { 216 printf("KTLS: API version mismatch (%d vs %d) for %s\n", 217 be->api_version, KTLS_API_VERSION, 218 be->name); 219 return (EINVAL); 220 } 221 222 rm_wlock(&ktls_backends_lock); 223 printf("KTLS: Registering crypto method %s with prio %d\n", 224 be->name, be->prio); 225 if (LIST_EMPTY(&ktls_backends)) { 226 LIST_INSERT_HEAD(&ktls_backends, be, next); 227 } else { 228 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) { 229 if (curr_be->prio < be->prio) { 230 LIST_INSERT_BEFORE(curr_be, be, next); 231 break; 232 } 233 if (LIST_NEXT(curr_be, next) == NULL) { 234 LIST_INSERT_AFTER(curr_be, be, next); 235 break; 236 } 237 } 238 } 239 rm_wunlock(&ktls_backends_lock); 240 return (0); 241 } 242 243 int 244 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be) 245 { 246 struct ktls_crypto_backend *tmp; 247 248 /* 249 * Don't error if the backend isn't registered. This permits 250 * MOD_UNLOAD handlers to use this function unconditionally. 251 */ 252 rm_wlock(&ktls_backends_lock); 253 LIST_FOREACH(tmp, &ktls_backends, next) { 254 if (tmp == be) 255 break; 256 } 257 if (tmp == NULL) { 258 rm_wunlock(&ktls_backends_lock); 259 return (0); 260 } 261 262 if (!ktls_allow_unload) { 263 rm_wunlock(&ktls_backends_lock); 264 printf( 265 "KTLS: Deregistering crypto method %s is not supported\n", 266 be->name); 267 return (EBUSY); 268 } 269 270 if (be->use_count) { 271 rm_wunlock(&ktls_backends_lock); 272 return (EBUSY); 273 } 274 275 LIST_REMOVE(be, next); 276 rm_wunlock(&ktls_backends_lock); 277 return (0); 278 } 279 280 #if defined(INET) || defined(INET6) 281 static uint16_t 282 ktls_get_cpu(struct socket *so) 283 { 284 struct inpcb *inp; 285 uint16_t cpuid; 286 287 inp = sotoinpcb(so); 288 #ifdef RSS 289 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 290 if (cpuid != NETISR_CPUID_NONE) 291 return (cpuid); 292 #endif 293 /* 294 * Just use the flowid to shard connections in a repeatable 295 * fashion. Note that some crypto backends rely on the 296 * serialization provided by having the same connection use 297 * the same queue. 298 */ 299 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 300 return (cpuid); 301 } 302 #endif 303 304 static void 305 ktls_init(void *dummy __unused) 306 { 307 struct thread *td; 308 struct pcpu *pc; 309 cpuset_t mask; 310 int error, i; 311 312 ktls_tasks_active = counter_u64_alloc(M_WAITOK); 313 ktls_cnt_on = counter_u64_alloc(M_WAITOK); 314 ktls_offload_total = counter_u64_alloc(M_WAITOK); 315 ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK); 316 ktls_offload_active = counter_u64_alloc(M_WAITOK); 317 ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK); 318 ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK); 319 ktls_switch_to_sw = counter_u64_alloc(M_WAITOK); 320 ktls_switch_failed = counter_u64_alloc(M_WAITOK); 321 ktls_sw_cbc = counter_u64_alloc(M_WAITOK); 322 ktls_sw_gcm = counter_u64_alloc(M_WAITOK); 323 ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK); 324 ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK); 325 ktls_ifnet_reset = counter_u64_alloc(M_WAITOK); 326 ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK); 327 ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK); 328 329 rm_init(&ktls_backends_lock, "ktls backends"); 330 LIST_INIT(&ktls_backends); 331 332 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 333 M_WAITOK | M_ZERO); 334 335 ktls_session_zone = uma_zcreate("ktls_session", 336 sizeof(struct ktls_session), 337 #ifdef INVARIANTS 338 trash_ctor, trash_dtor, trash_init, trash_fini, 339 #else 340 NULL, NULL, NULL, NULL, 341 #endif 342 UMA_ALIGN_CACHE, 0); 343 344 /* 345 * Initialize the workqueues to run the TLS work. We create a 346 * work queue for each CPU. 347 */ 348 CPU_FOREACH(i) { 349 STAILQ_INIT(&ktls_wq[i].head); 350 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 351 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 352 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 353 if (error) 354 panic("Can't add KTLS thread %d error %d", i, error); 355 356 /* 357 * Bind threads to cores. If ktls_bind_threads is > 358 * 1, then we bind to the NUMA domain. 359 */ 360 if (ktls_bind_threads) { 361 if (ktls_bind_threads > 1) { 362 pc = pcpu_find(i); 363 CPU_COPY(&cpuset_domain[pc->pc_domain], &mask); 364 } else { 365 CPU_SETOF(i, &mask); 366 } 367 error = cpuset_setthread(td->td_tid, &mask); 368 if (error) 369 panic( 370 "Unable to bind KTLS thread for CPU %d error %d", 371 i, error); 372 } 373 ktls_cpuid_lookup[ktls_number_threads] = i; 374 ktls_number_threads++; 375 } 376 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 377 } 378 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL); 379 380 #if defined(INET) || defined(INET6) 381 static int 382 ktls_create_session(struct socket *so, struct tls_enable *en, 383 struct ktls_session **tlsp) 384 { 385 struct ktls_session *tls; 386 int error; 387 388 /* Only TLS 1.0 - 1.2 are supported. */ 389 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 390 return (EINVAL); 391 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 392 en->tls_vminor > TLS_MINOR_VER_THREE) 393 return (EINVAL); 394 395 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 396 return (EINVAL); 397 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 398 return (EINVAL); 399 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 400 return (EINVAL); 401 402 /* All supported algorithms require a cipher key. */ 403 if (en->cipher_key_len == 0) 404 return (EINVAL); 405 406 /* No flags are currently supported. */ 407 if (en->flags != 0) 408 return (EINVAL); 409 410 /* Common checks for supported algorithms. */ 411 switch (en->cipher_algorithm) { 412 case CRYPTO_AES_NIST_GCM_16: 413 /* 414 * auth_algorithm isn't used, but permit GMAC values 415 * for compatibility. 416 */ 417 switch (en->auth_algorithm) { 418 case 0: 419 case CRYPTO_AES_128_NIST_GMAC: 420 case CRYPTO_AES_192_NIST_GMAC: 421 case CRYPTO_AES_256_NIST_GMAC: 422 break; 423 default: 424 return (EINVAL); 425 } 426 if (en->auth_key_len != 0) 427 return (EINVAL); 428 if ((en->tls_vminor == TLS_MINOR_VER_TWO && 429 en->iv_len != TLS_AEAD_GCM_LEN) || 430 (en->tls_vminor == TLS_MINOR_VER_THREE && 431 en->iv_len != TLS_1_3_GCM_IV_LEN)) 432 return (EINVAL); 433 break; 434 case CRYPTO_AES_CBC: 435 switch (en->auth_algorithm) { 436 case CRYPTO_SHA1_HMAC: 437 /* 438 * TLS 1.0 requires an implicit IV. TLS 1.1+ 439 * all use explicit IVs. 440 */ 441 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 442 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 443 return (EINVAL); 444 break; 445 } 446 447 /* FALLTHROUGH */ 448 case CRYPTO_SHA2_256_HMAC: 449 case CRYPTO_SHA2_384_HMAC: 450 /* Ignore any supplied IV. */ 451 en->iv_len = 0; 452 break; 453 default: 454 return (EINVAL); 455 } 456 if (en->auth_key_len == 0) 457 return (EINVAL); 458 break; 459 default: 460 return (EINVAL); 461 } 462 463 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 464 465 counter_u64_add(ktls_offload_active, 1); 466 467 refcount_init(&tls->refcount, 1); 468 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 469 470 tls->wq_index = ktls_get_cpu(so); 471 472 tls->params.cipher_algorithm = en->cipher_algorithm; 473 tls->params.auth_algorithm = en->auth_algorithm; 474 tls->params.tls_vmajor = en->tls_vmajor; 475 tls->params.tls_vminor = en->tls_vminor; 476 tls->params.flags = en->flags; 477 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 478 479 /* Set the header and trailer lengths. */ 480 tls->params.tls_hlen = sizeof(struct tls_record_layer); 481 switch (en->cipher_algorithm) { 482 case CRYPTO_AES_NIST_GCM_16: 483 /* 484 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 485 * nonce. TLS 1.3 uses a 12 byte implicit IV. 486 */ 487 if (en->tls_vminor < TLS_MINOR_VER_THREE) 488 tls->params.tls_hlen += sizeof(uint64_t); 489 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 490 491 /* 492 * TLS 1.3 includes optional padding which we 493 * do not support, and also puts the "real" record 494 * type at the end of the encrypted data. 495 */ 496 if (en->tls_vminor == TLS_MINOR_VER_THREE) 497 tls->params.tls_tlen += sizeof(uint8_t); 498 499 tls->params.tls_bs = 1; 500 break; 501 case CRYPTO_AES_CBC: 502 switch (en->auth_algorithm) { 503 case CRYPTO_SHA1_HMAC: 504 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 505 /* Implicit IV, no nonce. */ 506 } else { 507 tls->params.tls_hlen += AES_BLOCK_LEN; 508 } 509 tls->params.tls_tlen = AES_BLOCK_LEN + 510 SHA1_HASH_LEN; 511 break; 512 case CRYPTO_SHA2_256_HMAC: 513 tls->params.tls_hlen += AES_BLOCK_LEN; 514 tls->params.tls_tlen = AES_BLOCK_LEN + 515 SHA2_256_HASH_LEN; 516 break; 517 case CRYPTO_SHA2_384_HMAC: 518 tls->params.tls_hlen += AES_BLOCK_LEN; 519 tls->params.tls_tlen = AES_BLOCK_LEN + 520 SHA2_384_HASH_LEN; 521 break; 522 default: 523 panic("invalid hmac"); 524 } 525 tls->params.tls_bs = AES_BLOCK_LEN; 526 break; 527 default: 528 panic("invalid cipher"); 529 } 530 531 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 532 ("TLS header length too long: %d", tls->params.tls_hlen)); 533 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 534 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 535 536 if (en->auth_key_len != 0) { 537 tls->params.auth_key_len = en->auth_key_len; 538 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 539 M_WAITOK); 540 error = copyin(en->auth_key, tls->params.auth_key, 541 en->auth_key_len); 542 if (error) 543 goto out; 544 } 545 546 tls->params.cipher_key_len = en->cipher_key_len; 547 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 548 error = copyin(en->cipher_key, tls->params.cipher_key, 549 en->cipher_key_len); 550 if (error) 551 goto out; 552 553 /* 554 * This holds the implicit portion of the nonce for GCM and 555 * the initial implicit IV for TLS 1.0. The explicit portions 556 * of the IV are generated in ktls_frame() and ktls_seq(). 557 */ 558 if (en->iv_len != 0) { 559 tls->params.iv_len = en->iv_len; 560 error = copyin(en->iv, tls->params.iv, en->iv_len); 561 if (error) 562 goto out; 563 } 564 565 *tlsp = tls; 566 return (0); 567 568 out: 569 ktls_cleanup(tls); 570 return (error); 571 } 572 573 static struct ktls_session * 574 ktls_clone_session(struct ktls_session *tls) 575 { 576 struct ktls_session *tls_new; 577 578 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 579 580 counter_u64_add(ktls_offload_active, 1); 581 582 refcount_init(&tls_new->refcount, 1); 583 584 /* Copy fields from existing session. */ 585 tls_new->params = tls->params; 586 tls_new->wq_index = tls->wq_index; 587 588 /* Deep copy keys. */ 589 if (tls_new->params.auth_key != NULL) { 590 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 591 M_KTLS, M_WAITOK); 592 memcpy(tls_new->params.auth_key, tls->params.auth_key, 593 tls->params.auth_key_len); 594 } 595 596 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 597 M_WAITOK); 598 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 599 tls->params.cipher_key_len); 600 601 return (tls_new); 602 } 603 #endif 604 605 static void 606 ktls_cleanup(struct ktls_session *tls) 607 { 608 609 counter_u64_add(ktls_offload_active, -1); 610 if (tls->free != NULL) { 611 MPASS(tls->be != NULL); 612 switch (tls->params.cipher_algorithm) { 613 case CRYPTO_AES_CBC: 614 counter_u64_add(ktls_sw_cbc, -1); 615 break; 616 case CRYPTO_AES_NIST_GCM_16: 617 counter_u64_add(ktls_sw_gcm, -1); 618 break; 619 } 620 tls->free(tls); 621 } else if (tls->snd_tag != NULL) { 622 switch (tls->params.cipher_algorithm) { 623 case CRYPTO_AES_CBC: 624 counter_u64_add(ktls_ifnet_cbc, -1); 625 break; 626 case CRYPTO_AES_NIST_GCM_16: 627 counter_u64_add(ktls_ifnet_gcm, -1); 628 break; 629 } 630 m_snd_tag_rele(tls->snd_tag); 631 } 632 if (tls->params.auth_key != NULL) { 633 explicit_bzero(tls->params.auth_key, tls->params.auth_key_len); 634 free(tls->params.auth_key, M_KTLS); 635 tls->params.auth_key = NULL; 636 tls->params.auth_key_len = 0; 637 } 638 if (tls->params.cipher_key != NULL) { 639 explicit_bzero(tls->params.cipher_key, 640 tls->params.cipher_key_len); 641 free(tls->params.cipher_key, M_KTLS); 642 tls->params.cipher_key = NULL; 643 tls->params.cipher_key_len = 0; 644 } 645 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 646 } 647 648 #if defined(INET) || defined(INET6) 649 /* 650 * Common code used when first enabling ifnet TLS on a connection or 651 * when allocating a new ifnet TLS session due to a routing change. 652 * This function allocates a new TLS send tag on whatever interface 653 * the connection is currently routed over. 654 */ 655 static int 656 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 657 struct m_snd_tag **mstp) 658 { 659 union if_snd_tag_alloc_params params; 660 struct ifnet *ifp; 661 struct rtentry *rt; 662 struct tcpcb *tp; 663 int error; 664 665 INP_RLOCK(inp); 666 if (inp->inp_flags2 & INP_FREED) { 667 INP_RUNLOCK(inp); 668 return (ECONNRESET); 669 } 670 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 671 INP_RUNLOCK(inp); 672 return (ECONNRESET); 673 } 674 if (inp->inp_socket == NULL) { 675 INP_RUNLOCK(inp); 676 return (ECONNRESET); 677 } 678 tp = intotcpcb(inp); 679 680 /* 681 * Check administrative controls on ifnet TLS to determine if 682 * ifnet TLS should be denied. 683 * 684 * - Always permit 'force' requests. 685 * - ktls_ifnet_permitted == 0: always deny. 686 */ 687 if (!force && ktls_ifnet_permitted == 0) { 688 INP_RUNLOCK(inp); 689 return (ENXIO); 690 } 691 692 /* 693 * XXX: Use the cached route in the inpcb to find the 694 * interface. This should perhaps instead use 695 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 696 * enabled after a connection has completed key negotiation in 697 * userland, the cached route will be present in practice. 698 */ 699 rt = inp->inp_route.ro_rt; 700 if (rt == NULL || rt->rt_ifp == NULL) { 701 INP_RUNLOCK(inp); 702 return (ENXIO); 703 } 704 ifp = rt->rt_ifp; 705 if_ref(ifp); 706 707 params.hdr.type = IF_SND_TAG_TYPE_TLS; 708 params.hdr.flowid = inp->inp_flowid; 709 params.hdr.flowtype = inp->inp_flowtype; 710 params.tls.inp = inp; 711 params.tls.tls = tls; 712 INP_RUNLOCK(inp); 713 714 if (ifp->if_snd_tag_alloc == NULL) { 715 error = EOPNOTSUPP; 716 goto out; 717 } 718 if ((ifp->if_capenable & IFCAP_NOMAP) == 0) { 719 error = EOPNOTSUPP; 720 goto out; 721 } 722 if (inp->inp_vflag & INP_IPV6) { 723 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { 724 error = EOPNOTSUPP; 725 goto out; 726 } 727 } else { 728 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { 729 error = EOPNOTSUPP; 730 goto out; 731 } 732 } 733 error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp); 734 out: 735 if_rele(ifp); 736 return (error); 737 } 738 739 static int 740 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force) 741 { 742 struct m_snd_tag *mst; 743 int error; 744 745 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 746 if (error == 0) { 747 tls->snd_tag = mst; 748 switch (tls->params.cipher_algorithm) { 749 case CRYPTO_AES_CBC: 750 counter_u64_add(ktls_ifnet_cbc, 1); 751 break; 752 case CRYPTO_AES_NIST_GCM_16: 753 counter_u64_add(ktls_ifnet_gcm, 1); 754 break; 755 } 756 } 757 return (error); 758 } 759 760 static int 761 ktls_try_sw(struct socket *so, struct ktls_session *tls) 762 { 763 struct rm_priotracker prio; 764 struct ktls_crypto_backend *be; 765 766 /* 767 * Choose the best software crypto backend. Backends are 768 * stored in sorted priority order (larget value == most 769 * important at the head of the list), so this just stops on 770 * the first backend that claims the session by returning 771 * success. 772 */ 773 if (ktls_allow_unload) 774 rm_rlock(&ktls_backends_lock, &prio); 775 LIST_FOREACH(be, &ktls_backends, next) { 776 if (be->try(so, tls) == 0) 777 break; 778 KASSERT(tls->cipher == NULL, 779 ("ktls backend leaked a cipher pointer")); 780 } 781 if (be != NULL) { 782 if (ktls_allow_unload) 783 be->use_count++; 784 tls->be = be; 785 } 786 if (ktls_allow_unload) 787 rm_runlock(&ktls_backends_lock, &prio); 788 if (be == NULL) 789 return (EOPNOTSUPP); 790 switch (tls->params.cipher_algorithm) { 791 case CRYPTO_AES_CBC: 792 counter_u64_add(ktls_sw_cbc, 1); 793 break; 794 case CRYPTO_AES_NIST_GCM_16: 795 counter_u64_add(ktls_sw_gcm, 1); 796 break; 797 } 798 return (0); 799 } 800 801 int 802 ktls_enable_tx(struct socket *so, struct tls_enable *en) 803 { 804 struct ktls_session *tls; 805 int error; 806 807 if (!ktls_offload_enable) 808 return (ENOTSUP); 809 810 counter_u64_add(ktls_offload_enable_calls, 1); 811 812 /* 813 * This should always be true since only the TCP socket option 814 * invokes this function. 815 */ 816 if (so->so_proto->pr_protocol != IPPROTO_TCP) 817 return (EINVAL); 818 819 /* 820 * XXX: Don't overwrite existing sessions. We should permit 821 * this to support rekeying in the future. 822 */ 823 if (so->so_snd.sb_tls_info != NULL) 824 return (EALREADY); 825 826 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 827 return (ENOTSUP); 828 829 /* TLS requires ext pgs */ 830 if (mb_use_ext_pgs == 0) 831 return (ENXIO); 832 833 error = ktls_create_session(so, en, &tls); 834 if (error) 835 return (error); 836 837 /* Prefer ifnet TLS over software TLS. */ 838 error = ktls_try_ifnet(so, tls, false); 839 if (error) 840 error = ktls_try_sw(so, tls); 841 842 if (error) { 843 ktls_cleanup(tls); 844 return (error); 845 } 846 847 error = sblock(&so->so_snd, SBL_WAIT); 848 if (error) { 849 ktls_cleanup(tls); 850 return (error); 851 } 852 853 SOCKBUF_LOCK(&so->so_snd); 854 so->so_snd.sb_tls_info = tls; 855 if (tls->sw_encrypt == NULL) 856 so->so_snd.sb_flags |= SB_TLS_IFNET; 857 SOCKBUF_UNLOCK(&so->so_snd); 858 sbunlock(&so->so_snd); 859 860 counter_u64_add(ktls_offload_total, 1); 861 862 return (0); 863 } 864 865 int 866 ktls_get_tx_mode(struct socket *so) 867 { 868 struct ktls_session *tls; 869 struct inpcb *inp; 870 int mode; 871 872 inp = so->so_pcb; 873 INP_WLOCK_ASSERT(inp); 874 SOCKBUF_LOCK(&so->so_snd); 875 tls = so->so_snd.sb_tls_info; 876 if (tls == NULL) 877 mode = TCP_TLS_MODE_NONE; 878 else if (tls->sw_encrypt != NULL) 879 mode = TCP_TLS_MODE_SW; 880 else 881 mode = TCP_TLS_MODE_IFNET; 882 SOCKBUF_UNLOCK(&so->so_snd); 883 return (mode); 884 } 885 886 /* 887 * Switch between SW and ifnet TLS sessions as requested. 888 */ 889 int 890 ktls_set_tx_mode(struct socket *so, int mode) 891 { 892 struct ktls_session *tls, *tls_new; 893 struct inpcb *inp; 894 int error; 895 896 MPASS(mode == TCP_TLS_MODE_SW || mode == TCP_TLS_MODE_IFNET); 897 898 inp = so->so_pcb; 899 INP_WLOCK_ASSERT(inp); 900 SOCKBUF_LOCK(&so->so_snd); 901 tls = so->so_snd.sb_tls_info; 902 if (tls == NULL) { 903 SOCKBUF_UNLOCK(&so->so_snd); 904 return (0); 905 } 906 907 if ((tls->sw_encrypt != NULL && mode == TCP_TLS_MODE_SW) || 908 (tls->sw_encrypt == NULL && mode == TCP_TLS_MODE_IFNET)) { 909 SOCKBUF_UNLOCK(&so->so_snd); 910 return (0); 911 } 912 913 tls = ktls_hold(tls); 914 SOCKBUF_UNLOCK(&so->so_snd); 915 INP_WUNLOCK(inp); 916 917 tls_new = ktls_clone_session(tls); 918 919 if (mode == TCP_TLS_MODE_IFNET) 920 error = ktls_try_ifnet(so, tls_new, true); 921 else 922 error = ktls_try_sw(so, tls_new); 923 if (error) { 924 counter_u64_add(ktls_switch_failed, 1); 925 ktls_free(tls_new); 926 ktls_free(tls); 927 INP_WLOCK(inp); 928 return (error); 929 } 930 931 error = sblock(&so->so_snd, SBL_WAIT); 932 if (error) { 933 counter_u64_add(ktls_switch_failed, 1); 934 ktls_free(tls_new); 935 ktls_free(tls); 936 INP_WLOCK(inp); 937 return (error); 938 } 939 940 /* 941 * If we raced with another session change, keep the existing 942 * session. 943 */ 944 if (tls != so->so_snd.sb_tls_info) { 945 counter_u64_add(ktls_switch_failed, 1); 946 sbunlock(&so->so_snd); 947 ktls_free(tls_new); 948 ktls_free(tls); 949 INP_WLOCK(inp); 950 return (EBUSY); 951 } 952 953 SOCKBUF_LOCK(&so->so_snd); 954 so->so_snd.sb_tls_info = tls_new; 955 if (tls_new->sw_encrypt == NULL) 956 so->so_snd.sb_flags |= SB_TLS_IFNET; 957 SOCKBUF_UNLOCK(&so->so_snd); 958 sbunlock(&so->so_snd); 959 960 /* 961 * Drop two references on 'tls'. The first is for the 962 * ktls_hold() above. The second drops the reference from the 963 * socket buffer. 964 */ 965 KASSERT(tls->refcount >= 2, ("too few references on old session")); 966 ktls_free(tls); 967 ktls_free(tls); 968 969 if (mode == TCP_TLS_MODE_IFNET) 970 counter_u64_add(ktls_switch_to_ifnet, 1); 971 else 972 counter_u64_add(ktls_switch_to_sw, 1); 973 974 INP_WLOCK(inp); 975 return (0); 976 } 977 978 /* 979 * Try to allocate a new TLS send tag. This task is scheduled when 980 * ip_output detects a route change while trying to transmit a packet 981 * holding a TLS record. If a new tag is allocated, replace the tag 982 * in the TLS session. Subsequent packets on the connection will use 983 * the new tag. If a new tag cannot be allocated, drop the 984 * connection. 985 */ 986 static void 987 ktls_reset_send_tag(void *context, int pending) 988 { 989 struct epoch_tracker et; 990 struct ktls_session *tls; 991 struct m_snd_tag *old, *new; 992 struct inpcb *inp; 993 struct tcpcb *tp; 994 int error; 995 996 MPASS(pending == 1); 997 998 tls = context; 999 inp = tls->inp; 1000 1001 /* 1002 * Free the old tag first before allocating a new one. 1003 * ip[6]_output_send() will treat a NULL send tag the same as 1004 * an ifp mismatch and drop packets until a new tag is 1005 * allocated. 1006 * 1007 * Write-lock the INP when changing tls->snd_tag since 1008 * ip[6]_output_send() holds a read-lock when reading the 1009 * pointer. 1010 */ 1011 INP_WLOCK(inp); 1012 old = tls->snd_tag; 1013 tls->snd_tag = NULL; 1014 INP_WUNLOCK(inp); 1015 if (old != NULL) 1016 m_snd_tag_rele(old); 1017 1018 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1019 1020 if (error == 0) { 1021 INP_WLOCK(inp); 1022 tls->snd_tag = new; 1023 mtx_pool_lock(mtxpool_sleep, tls); 1024 tls->reset_pending = false; 1025 mtx_pool_unlock(mtxpool_sleep, tls); 1026 if (!in_pcbrele_wlocked(inp)) 1027 INP_WUNLOCK(inp); 1028 1029 counter_u64_add(ktls_ifnet_reset, 1); 1030 1031 /* 1032 * XXX: Should we kick tcp_output explicitly now that 1033 * the send tag is fixed or just rely on timers? 1034 */ 1035 } else { 1036 INP_INFO_RLOCK_ET(&V_tcbinfo, et); 1037 INP_WLOCK(inp); 1038 if (!in_pcbrele_wlocked(inp)) { 1039 if (!(inp->inp_flags & INP_TIMEWAIT) && 1040 !(inp->inp_flags & INP_DROPPED)) { 1041 tp = intotcpcb(inp); 1042 tp = tcp_drop(tp, ECONNABORTED); 1043 if (tp != NULL) 1044 INP_WUNLOCK(inp); 1045 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1046 } else 1047 INP_WUNLOCK(inp); 1048 } 1049 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et); 1050 1051 counter_u64_add(ktls_ifnet_reset_failed, 1); 1052 1053 /* 1054 * Leave reset_pending true to avoid future tasks while 1055 * the socket goes away. 1056 */ 1057 } 1058 1059 ktls_free(tls); 1060 } 1061 1062 int 1063 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1064 { 1065 1066 if (inp == NULL) 1067 return (ENOBUFS); 1068 1069 INP_LOCK_ASSERT(inp); 1070 1071 /* 1072 * See if we should schedule a task to update the send tag for 1073 * this session. 1074 */ 1075 mtx_pool_lock(mtxpool_sleep, tls); 1076 if (!tls->reset_pending) { 1077 (void) ktls_hold(tls); 1078 in_pcbref(inp); 1079 tls->inp = inp; 1080 tls->reset_pending = true; 1081 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1082 } 1083 mtx_pool_unlock(mtxpool_sleep, tls); 1084 return (ENOBUFS); 1085 } 1086 #endif 1087 1088 void 1089 ktls_destroy(struct ktls_session *tls) 1090 { 1091 struct rm_priotracker prio; 1092 1093 ktls_cleanup(tls); 1094 if (tls->be != NULL && ktls_allow_unload) { 1095 rm_rlock(&ktls_backends_lock, &prio); 1096 tls->be->use_count--; 1097 rm_runlock(&ktls_backends_lock, &prio); 1098 } 1099 uma_zfree(ktls_session_zone, tls); 1100 } 1101 1102 void 1103 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1104 { 1105 struct mbuf_ext_pgs *pgs; 1106 struct tls_record_layer *tlshdr; 1107 uint64_t seqno; 1108 1109 for (; m != NULL; m = m->m_next) { 1110 KASSERT((m->m_flags & M_NOMAP) != 0, 1111 ("ktls_seq: mapped mbuf %p", m)); 1112 1113 pgs = m->m_ext.ext_pgs; 1114 pgs->seqno = sb->sb_tls_seqno; 1115 1116 /* 1117 * Store the sequence number in the TLS header as the 1118 * explicit part of the IV for GCM. 1119 */ 1120 if (pgs->tls->params.cipher_algorithm == 1121 CRYPTO_AES_NIST_GCM_16) { 1122 tlshdr = (void *)pgs->hdr; 1123 seqno = htobe64(pgs->seqno); 1124 memcpy(tlshdr + 1, &seqno, sizeof(seqno)); 1125 } 1126 sb->sb_tls_seqno++; 1127 } 1128 } 1129 1130 /* 1131 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1132 * mbuf in the chain must be an unmapped mbuf. The payload of the 1133 * mbuf must be populated with the payload of each TLS record. 1134 * 1135 * The record_type argument specifies the TLS record type used when 1136 * populating the TLS header. 1137 * 1138 * The enq_count argument on return is set to the number of pages of 1139 * payload data for this entire chain that need to be encrypted via SW 1140 * encryption. The returned value should be passed to ktls_enqueue 1141 * when scheduling encryption of this chain of mbufs. 1142 */ 1143 int 1144 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1145 uint8_t record_type) 1146 { 1147 struct tls_record_layer *tlshdr; 1148 struct mbuf *m; 1149 struct mbuf_ext_pgs *pgs; 1150 uint16_t tls_len; 1151 int maxlen; 1152 1153 maxlen = tls->params.max_frame_len; 1154 *enq_cnt = 0; 1155 for (m = top; m != NULL; m = m->m_next) { 1156 /* 1157 * All mbufs in the chain should be non-empty TLS 1158 * records whose payload does not exceed the maximum 1159 * frame length. 1160 */ 1161 if (m->m_len > maxlen || m->m_len == 0) 1162 return (EINVAL); 1163 tls_len = m->m_len; 1164 1165 /* 1166 * TLS frames require unmapped mbufs to store session 1167 * info. 1168 */ 1169 KASSERT((m->m_flags & M_NOMAP) != 0, 1170 ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top)); 1171 1172 pgs = m->m_ext.ext_pgs; 1173 1174 /* Save a reference to the session. */ 1175 pgs->tls = ktls_hold(tls); 1176 1177 pgs->hdr_len = tls->params.tls_hlen; 1178 pgs->trail_len = tls->params.tls_tlen; 1179 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 1180 int bs, delta; 1181 1182 /* 1183 * AES-CBC pads messages to a multiple of the 1184 * block size. Note that the padding is 1185 * applied after the digest and the encryption 1186 * is done on the "plaintext || mac || padding". 1187 * At least one byte of padding is always 1188 * present. 1189 * 1190 * Compute the final trailer length assuming 1191 * at most one block of padding. 1192 * tls->params.sb_tls_tlen is the maximum 1193 * possible trailer length (padding + digest). 1194 * delta holds the number of excess padding 1195 * bytes if the maximum were used. Those 1196 * extra bytes are removed. 1197 */ 1198 bs = tls->params.tls_bs; 1199 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 1200 pgs->trail_len -= delta; 1201 } 1202 m->m_len += pgs->hdr_len + pgs->trail_len; 1203 1204 /* Populate the TLS header. */ 1205 tlshdr = (void *)pgs->hdr; 1206 tlshdr->tls_vmajor = tls->params.tls_vmajor; 1207 1208 /* 1209 * TLS 1.3 masquarades as TLS 1.2 with a record type 1210 * of TLS_RLTYPE_APP. 1211 */ 1212 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 1213 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 1214 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 1215 tlshdr->tls_type = TLS_RLTYPE_APP; 1216 /* save the real record type for later */ 1217 pgs->record_type = record_type; 1218 } else { 1219 tlshdr->tls_vminor = tls->params.tls_vminor; 1220 tlshdr->tls_type = record_type; 1221 } 1222 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 1223 1224 /* 1225 * For GCM, the sequence number is stored in the 1226 * header by ktls_seq(). For CBC, a random nonce is 1227 * inserted for TLS 1.1+. 1228 */ 1229 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 1230 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 1231 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 1232 1233 /* 1234 * When using SW encryption, mark the mbuf not ready. 1235 * It will be marked ready via sbready() after the 1236 * record has been encrypted. 1237 * 1238 * When using ifnet TLS, unencrypted TLS records are 1239 * sent down the stack to the NIC. 1240 */ 1241 if (tls->sw_encrypt != NULL) { 1242 m->m_flags |= M_NOTREADY; 1243 pgs->nrdy = pgs->npgs; 1244 *enq_cnt += pgs->npgs; 1245 } 1246 } 1247 return (0); 1248 } 1249 1250 void 1251 ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs) 1252 { 1253 struct ktls_wq *wq; 1254 bool running; 1255 1256 /* Mark it for freeing. */ 1257 pgs->mbuf = NULL; 1258 wq = &ktls_wq[pgs->tls->wq_index]; 1259 mtx_lock(&wq->mtx); 1260 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq); 1261 running = wq->running; 1262 mtx_unlock(&wq->mtx); 1263 if (!running) 1264 wakeup(wq); 1265 } 1266 1267 void 1268 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 1269 { 1270 struct mbuf_ext_pgs *pgs; 1271 struct ktls_wq *wq; 1272 bool running; 1273 1274 KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) == 1275 (M_NOMAP | M_NOTREADY)), 1276 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 1277 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 1278 1279 pgs = m->m_ext.ext_pgs; 1280 1281 KASSERT(pgs->tls->sw_encrypt != NULL, ("ifnet TLS mbuf")); 1282 1283 pgs->enc_cnt = page_count; 1284 pgs->mbuf = m; 1285 1286 /* 1287 * Save a pointer to the socket. The caller is responsible 1288 * for taking an additional reference via soref(). 1289 */ 1290 pgs->so = so; 1291 1292 wq = &ktls_wq[pgs->tls->wq_index]; 1293 mtx_lock(&wq->mtx); 1294 STAILQ_INSERT_TAIL(&wq->head, pgs, stailq); 1295 running = wq->running; 1296 mtx_unlock(&wq->mtx); 1297 if (!running) 1298 wakeup(wq); 1299 counter_u64_add(ktls_cnt_on, 1); 1300 } 1301 1302 static __noinline void 1303 ktls_encrypt(struct mbuf_ext_pgs *pgs) 1304 { 1305 struct ktls_session *tls; 1306 struct socket *so; 1307 struct mbuf *m, *top; 1308 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1309 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1310 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 1311 vm_page_t pg; 1312 int error, i, len, npages, off, total_pages; 1313 bool is_anon; 1314 1315 so = pgs->so; 1316 tls = pgs->tls; 1317 top = pgs->mbuf; 1318 KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs)); 1319 KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs)); 1320 #ifdef INVARIANTS 1321 pgs->so = NULL; 1322 pgs->mbuf = NULL; 1323 #endif 1324 total_pages = pgs->enc_cnt; 1325 npages = 0; 1326 1327 /* 1328 * Encrypt the TLS records in the chain of mbufs starting with 1329 * 'top'. 'total_pages' gives us a total count of pages and is 1330 * used to know when we have finished encrypting the TLS 1331 * records originally queued with 'top'. 1332 * 1333 * NB: These mbufs are queued in the socket buffer and 1334 * 'm_next' is traversing the mbufs in the socket buffer. The 1335 * socket buffer lock is not held while traversing this chain. 1336 * Since the mbufs are all marked M_NOTREADY their 'm_next' 1337 * pointers should be stable. However, the 'm_next' of the 1338 * last mbuf encrypted is not necessarily NULL. It can point 1339 * to other mbufs appended while 'top' was on the TLS work 1340 * queue. 1341 * 1342 * Each mbuf holds an entire TLS record. 1343 */ 1344 error = 0; 1345 for (m = top; npages != total_pages; m = m->m_next) { 1346 pgs = m->m_ext.ext_pgs; 1347 1348 KASSERT(pgs->tls == tls, 1349 ("different TLS sessions in a single mbuf chain: %p vs %p", 1350 tls, pgs->tls)); 1351 KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) == 1352 (M_NOMAP | M_NOTREADY), 1353 ("%p not unready & nomap mbuf (top = %p)\n", m, top)); 1354 KASSERT(npages + pgs->npgs <= total_pages, 1355 ("page count mismatch: top %p, total_pages %d, m %p", top, 1356 total_pages, m)); 1357 1358 /* 1359 * Generate source and destination ivoecs to pass to 1360 * the SW encryption backend. For writable mbufs, the 1361 * destination iovec is a copy of the source and 1362 * encryption is done in place. For file-backed mbufs 1363 * (from sendfile), anonymous wired pages are 1364 * allocated and assigned to the destination iovec. 1365 */ 1366 is_anon = (pgs->flags & MBUF_PEXT_FLAG_ANON) != 0; 1367 1368 off = pgs->first_pg_off; 1369 for (i = 0; i < pgs->npgs; i++, off = 0) { 1370 len = mbuf_ext_pg_len(pgs, i, off); 1371 src_iov[i].iov_len = len; 1372 src_iov[i].iov_base = 1373 (char *)(void *)PHYS_TO_DMAP(pgs->pa[i]) + off; 1374 1375 if (is_anon) { 1376 dst_iov[i].iov_base = src_iov[i].iov_base; 1377 dst_iov[i].iov_len = src_iov[i].iov_len; 1378 continue; 1379 } 1380 retry_page: 1381 pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | 1382 VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED); 1383 if (pg == NULL) { 1384 vm_wait(NULL); 1385 goto retry_page; 1386 } 1387 parray[i] = VM_PAGE_TO_PHYS(pg); 1388 dst_iov[i].iov_base = 1389 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off; 1390 dst_iov[i].iov_len = len; 1391 } 1392 1393 npages += i; 1394 1395 error = (*tls->sw_encrypt)(tls, 1396 (const struct tls_record_layer *)pgs->hdr, 1397 pgs->trail, src_iov, dst_iov, i, pgs->seqno, 1398 pgs->record_type); 1399 if (error) { 1400 counter_u64_add(ktls_offload_failed_crypto, 1); 1401 break; 1402 } 1403 1404 /* 1405 * For file-backed mbufs, release the file-backed 1406 * pages and replace them in the ext_pgs array with 1407 * the anonymous wired pages allocated above. 1408 */ 1409 if (!is_anon) { 1410 /* Free the old pages. */ 1411 m->m_ext.ext_free(m); 1412 1413 /* Replace them with the new pages. */ 1414 for (i = 0; i < pgs->npgs; i++) 1415 pgs->pa[i] = parray[i]; 1416 1417 /* Use the basic free routine. */ 1418 m->m_ext.ext_free = mb_free_mext_pgs; 1419 1420 /* Pages are now writable. */ 1421 pgs->flags |= MBUF_PEXT_FLAG_ANON; 1422 } 1423 1424 /* 1425 * Drop a reference to the session now that it is no 1426 * longer needed. Existing code depends on encrypted 1427 * records having no associated session vs 1428 * yet-to-be-encrypted records having an associated 1429 * session. 1430 */ 1431 pgs->tls = NULL; 1432 ktls_free(tls); 1433 } 1434 1435 CURVNET_SET(so->so_vnet); 1436 if (error == 0) { 1437 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages); 1438 } else { 1439 so->so_proto->pr_usrreqs->pru_abort(so); 1440 so->so_error = EIO; 1441 mb_free_notready(top, total_pages); 1442 } 1443 1444 SOCK_LOCK(so); 1445 sorele(so); 1446 CURVNET_RESTORE(); 1447 } 1448 1449 static void 1450 ktls_work_thread(void *ctx) 1451 { 1452 struct ktls_wq *wq = ctx; 1453 struct mbuf_ext_pgs *p, *n; 1454 struct ktls_session *tls; 1455 STAILQ_HEAD(, mbuf_ext_pgs) local_head; 1456 1457 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 1458 fpu_kern_thread(0); 1459 #endif 1460 for (;;) { 1461 mtx_lock(&wq->mtx); 1462 while (STAILQ_EMPTY(&wq->head)) { 1463 wq->running = false; 1464 mtx_sleep(wq, &wq->mtx, 0, "-", 0); 1465 wq->running = true; 1466 } 1467 1468 STAILQ_INIT(&local_head); 1469 STAILQ_CONCAT(&local_head, &wq->head); 1470 mtx_unlock(&wq->mtx); 1471 1472 STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) { 1473 if (p->mbuf != NULL) { 1474 ktls_encrypt(p); 1475 counter_u64_add(ktls_cnt_on, -1); 1476 } else { 1477 tls = p->tls; 1478 ktls_free(tls); 1479 uma_zfree(zone_extpgs, p); 1480 } 1481 } 1482 } 1483 } 1484