Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/uipc_ktls.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
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/domainset.h> 38 #include <sys/ktls.h> 39 #include <sys/lock.h> 40 #include <sys/mbuf.h> 41 #include <sys/mutex.h> 42 #include <sys/rmlock.h> 43 #include <sys/proc.h> 44 #include <sys/protosw.h> 45 #include <sys/refcount.h> 46 #include <sys/smp.h> 47 #include <sys/socket.h> 48 #include <sys/socketvar.h> 49 #include <sys/sysctl.h> 50 #include <sys/taskqueue.h> 51 #include <sys/kthread.h> 52 #include <sys/uio.h> 53 #include <sys/vmmeter.h> 54 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 55 #include <machine/pcb.h> 56 #endif 57 #include <machine/vmparam.h> 58 #include <net/if.h> 59 #include <net/if_var.h> 60 #ifdef RSS 61 #include <net/netisr.h> 62 #include <net/rss_config.h> 63 #endif 64 #include <net/route.h> 65 #include <net/route/nhop.h> 66 #if defined(INET) || defined(INET6) 67 #include <netinet/in.h> 68 #include <netinet/in_pcb.h> 69 #endif 70 #include <netinet/tcp_var.h> 71 #ifdef TCP_OFFLOAD 72 #include <netinet/tcp_offload.h> 73 #endif 74 #include <opencrypto/xform.h> 75 #include <vm/uma_dbg.h> 76 #include <vm/vm.h> 77 #include <vm/vm_pageout.h> 78 #include <vm/vm_page.h> 79 80 struct ktls_wq { 81 struct mtx mtx; 82 STAILQ_HEAD(, mbuf) m_head; 83 STAILQ_HEAD(, socket) so_head; 84 bool running; 85 } __aligned(CACHE_LINE_SIZE); 86 87 struct ktls_domain_info { 88 int count; 89 int cpu[MAXCPU]; 90 }; 91 92 struct ktls_domain_info ktls_domains[MAXMEMDOM]; 93 static struct ktls_wq *ktls_wq; 94 static struct proc *ktls_proc; 95 LIST_HEAD(, ktls_crypto_backend) ktls_backends; 96 static struct rmlock ktls_backends_lock; 97 static uma_zone_t ktls_session_zone; 98 static uint16_t ktls_cpuid_lookup[MAXCPU]; 99 100 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 101 "Kernel TLS offload"); 102 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 103 "Kernel TLS offload stats"); 104 105 static int ktls_allow_unload; 106 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN, 107 &ktls_allow_unload, 0, "Allow software crypto modules to unload"); 108 109 #ifdef RSS 110 static int ktls_bind_threads = 1; 111 #else 112 static int ktls_bind_threads; 113 #endif 114 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 115 &ktls_bind_threads, 0, 116 "Bind crypto threads to cores (1) or cores and domains (2) at boot"); 117 118 static u_int ktls_maxlen = 16384; 119 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN, 120 &ktls_maxlen, 0, "Maximum TLS record size"); 121 122 static int ktls_number_threads; 123 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 124 &ktls_number_threads, 0, 125 "Number of TLS threads in thread-pool"); 126 127 static bool ktls_offload_enable; 128 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, 129 &ktls_offload_enable, 0, 130 "Enable support for kernel TLS offload"); 131 132 static bool ktls_cbc_enable = true; 133 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN, 134 &ktls_cbc_enable, 1, 135 "Enable Support of AES-CBC crypto for kernel TLS"); 136 137 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active); 138 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 139 &ktls_tasks_active, "Number of active tasks"); 140 141 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending); 142 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD, 143 &ktls_cnt_tx_pending, 144 "Number of TLS 1.0 records waiting for earlier TLS records"); 145 146 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued); 147 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, 148 &ktls_cnt_tx_queued, 149 "Number of TLS records in queue to tasks for SW encryption"); 150 151 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued); 152 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, 153 &ktls_cnt_rx_queued, 154 "Number of TLS sockets in queue to tasks for SW decryption"); 155 156 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total); 157 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 158 CTLFLAG_RD, &ktls_offload_total, 159 "Total successful TLS setups (parameters set)"); 160 161 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls); 162 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 163 CTLFLAG_RD, &ktls_offload_enable_calls, 164 "Total number of TLS enable calls made"); 165 166 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active); 167 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 168 &ktls_offload_active, "Total Active TLS sessions"); 169 170 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records); 171 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, 172 &ktls_offload_corrupted_records, "Total corrupted TLS records received"); 173 174 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto); 175 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 176 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 177 178 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet); 179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 180 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 181 182 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw); 183 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 184 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 185 186 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed); 187 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 188 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 189 190 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 191 "Software TLS session stats"); 192 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 193 "Hardware (ifnet) TLS session stats"); 194 #ifdef TCP_OFFLOAD 195 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 196 "TOE TLS session stats"); 197 #endif 198 199 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc); 200 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, 201 "Active number of software TLS sessions using AES-CBC"); 202 203 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm); 204 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, 205 "Active number of software TLS sessions using AES-GCM"); 206 207 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20); 208 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD, 209 &ktls_sw_chacha20, 210 "Active number of software TLS sessions using Chacha20-Poly1305"); 211 212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc); 213 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, 214 &ktls_ifnet_cbc, 215 "Active number of ifnet TLS sessions using AES-CBC"); 216 217 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm); 218 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, 219 &ktls_ifnet_gcm, 220 "Active number of ifnet TLS sessions using AES-GCM"); 221 222 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20); 223 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD, 224 &ktls_ifnet_chacha20, 225 "Active number of ifnet TLS sessions using Chacha20-Poly1305"); 226 227 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset); 228 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, 229 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); 230 231 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped); 232 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, 233 &ktls_ifnet_reset_dropped, 234 "TLS sessions dropped after failing to update ifnet send tag"); 235 236 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed); 237 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, 238 &ktls_ifnet_reset_failed, 239 "TLS sessions that failed to allocate a new ifnet send tag"); 240 241 static int ktls_ifnet_permitted; 242 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, 243 &ktls_ifnet_permitted, 1, 244 "Whether to permit hardware (ifnet) TLS sessions"); 245 246 #ifdef TCP_OFFLOAD 247 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc); 248 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, 249 &ktls_toe_cbc, 250 "Active number of TOE TLS sessions using AES-CBC"); 251 252 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm); 253 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, 254 &ktls_toe_gcm, 255 "Active number of TOE TLS sessions using AES-GCM"); 256 257 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20); 258 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD, 259 &ktls_toe_chacha20, 260 "Active number of TOE TLS sessions using Chacha20-Poly1305"); 261 #endif 262 263 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); 264 265 static void ktls_cleanup(struct ktls_session *tls); 266 #if defined(INET) || defined(INET6) 267 static void ktls_reset_send_tag(void *context, int pending); 268 #endif 269 static void ktls_work_thread(void *ctx); 270 271 int 272 ktls_crypto_backend_register(struct ktls_crypto_backend *be) 273 { 274 struct ktls_crypto_backend *curr_be, *tmp; 275 276 if (be->api_version != KTLS_API_VERSION) { 277 printf("KTLS: API version mismatch (%d vs %d) for %s\n", 278 be->api_version, KTLS_API_VERSION, 279 be->name); 280 return (EINVAL); 281 } 282 283 rm_wlock(&ktls_backends_lock); 284 printf("KTLS: Registering crypto method %s with prio %d\n", 285 be->name, be->prio); 286 if (LIST_EMPTY(&ktls_backends)) { 287 LIST_INSERT_HEAD(&ktls_backends, be, next); 288 } else { 289 LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) { 290 if (curr_be->prio < be->prio) { 291 LIST_INSERT_BEFORE(curr_be, be, next); 292 break; 293 } 294 if (LIST_NEXT(curr_be, next) == NULL) { 295 LIST_INSERT_AFTER(curr_be, be, next); 296 break; 297 } 298 } 299 } 300 rm_wunlock(&ktls_backends_lock); 301 return (0); 302 } 303 304 int 305 ktls_crypto_backend_deregister(struct ktls_crypto_backend *be) 306 { 307 struct ktls_crypto_backend *tmp; 308 309 /* 310 * Don't error if the backend isn't registered. This permits 311 * MOD_UNLOAD handlers to use this function unconditionally. 312 */ 313 rm_wlock(&ktls_backends_lock); 314 LIST_FOREACH(tmp, &ktls_backends, next) { 315 if (tmp == be) 316 break; 317 } 318 if (tmp == NULL) { 319 rm_wunlock(&ktls_backends_lock); 320 return (0); 321 } 322 323 if (!ktls_allow_unload) { 324 rm_wunlock(&ktls_backends_lock); 325 printf( 326 "KTLS: Deregistering crypto method %s is not supported\n", 327 be->name); 328 return (EBUSY); 329 } 330 331 if (be->use_count) { 332 rm_wunlock(&ktls_backends_lock); 333 return (EBUSY); 334 } 335 336 LIST_REMOVE(be, next); 337 rm_wunlock(&ktls_backends_lock); 338 return (0); 339 } 340 341 #if defined(INET) || defined(INET6) 342 static u_int 343 ktls_get_cpu(struct socket *so) 344 { 345 struct inpcb *inp; 346 #ifdef NUMA 347 struct ktls_domain_info *di; 348 #endif 349 u_int cpuid; 350 351 inp = sotoinpcb(so); 352 #ifdef RSS 353 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 354 if (cpuid != NETISR_CPUID_NONE) 355 return (cpuid); 356 #endif 357 /* 358 * Just use the flowid to shard connections in a repeatable 359 * fashion. Note that some crypto backends rely on the 360 * serialization provided by having the same connection use 361 * the same queue. 362 */ 363 #ifdef NUMA 364 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) { 365 di = &ktls_domains[inp->inp_numa_domain]; 366 cpuid = di->cpu[inp->inp_flowid % di->count]; 367 } else 368 #endif 369 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 370 return (cpuid); 371 } 372 #endif 373 374 static void 375 ktls_init(void *dummy __unused) 376 { 377 struct thread *td; 378 struct pcpu *pc; 379 cpuset_t mask; 380 int count, domain, error, i; 381 382 rm_init(&ktls_backends_lock, "ktls backends"); 383 LIST_INIT(&ktls_backends); 384 385 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 386 M_WAITOK | M_ZERO); 387 388 ktls_session_zone = uma_zcreate("ktls_session", 389 sizeof(struct ktls_session), 390 NULL, NULL, NULL, NULL, 391 UMA_ALIGN_CACHE, 0); 392 393 /* 394 * Initialize the workqueues to run the TLS work. We create a 395 * work queue for each CPU. 396 */ 397 CPU_FOREACH(i) { 398 STAILQ_INIT(&ktls_wq[i].m_head); 399 STAILQ_INIT(&ktls_wq[i].so_head); 400 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 401 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 402 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 403 if (error) 404 panic("Can't add KTLS thread %d error %d", i, error); 405 406 /* 407 * Bind threads to cores. If ktls_bind_threads is > 408 * 1, then we bind to the NUMA domain. 409 */ 410 if (ktls_bind_threads) { 411 if (ktls_bind_threads > 1) { 412 pc = pcpu_find(i); 413 domain = pc->pc_domain; 414 CPU_COPY(&cpuset_domain[domain], &mask); 415 count = ktls_domains[domain].count; 416 ktls_domains[domain].cpu[count] = i; 417 ktls_domains[domain].count++; 418 } else { 419 CPU_SETOF(i, &mask); 420 } 421 error = cpuset_setthread(td->td_tid, &mask); 422 if (error) 423 panic( 424 "Unable to bind KTLS thread for CPU %d error %d", 425 i, error); 426 } 427 ktls_cpuid_lookup[ktls_number_threads] = i; 428 ktls_number_threads++; 429 } 430 431 /* 432 * If we somehow have an empty domain, fall back to choosing 433 * among all KTLS threads. 434 */ 435 if (ktls_bind_threads > 1) { 436 for (i = 0; i < vm_ndomains; i++) { 437 if (ktls_domains[i].count == 0) { 438 ktls_bind_threads = 1; 439 break; 440 } 441 } 442 } 443 444 if (bootverbose) 445 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 446 } 447 SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL); 448 449 #if defined(INET) || defined(INET6) 450 static int 451 ktls_create_session(struct socket *so, struct tls_enable *en, 452 struct ktls_session **tlsp) 453 { 454 struct ktls_session *tls; 455 int error; 456 457 /* Only TLS 1.0 - 1.3 are supported. */ 458 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 459 return (EINVAL); 460 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 461 en->tls_vminor > TLS_MINOR_VER_THREE) 462 return (EINVAL); 463 464 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 465 return (EINVAL); 466 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 467 return (EINVAL); 468 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 469 return (EINVAL); 470 471 /* All supported algorithms require a cipher key. */ 472 if (en->cipher_key_len == 0) 473 return (EINVAL); 474 475 /* No flags are currently supported. */ 476 if (en->flags != 0) 477 return (EINVAL); 478 479 /* Common checks for supported algorithms. */ 480 switch (en->cipher_algorithm) { 481 case CRYPTO_AES_NIST_GCM_16: 482 /* 483 * auth_algorithm isn't used, but permit GMAC values 484 * for compatibility. 485 */ 486 switch (en->auth_algorithm) { 487 case 0: 488 #ifdef COMPAT_FREEBSD12 489 /* XXX: Really 13.0-current COMPAT. */ 490 case CRYPTO_AES_128_NIST_GMAC: 491 case CRYPTO_AES_192_NIST_GMAC: 492 case CRYPTO_AES_256_NIST_GMAC: 493 #endif 494 break; 495 default: 496 return (EINVAL); 497 } 498 if (en->auth_key_len != 0) 499 return (EINVAL); 500 switch (en->tls_vminor) { 501 case TLS_MINOR_VER_TWO: 502 if (en->iv_len != TLS_AEAD_GCM_LEN) 503 return (EINVAL); 504 break; 505 case TLS_MINOR_VER_THREE: 506 if (en->iv_len != TLS_1_3_GCM_IV_LEN) 507 return (EINVAL); 508 break; 509 default: 510 return (EINVAL); 511 } 512 break; 513 case CRYPTO_AES_CBC: 514 switch (en->auth_algorithm) { 515 case CRYPTO_SHA1_HMAC: 516 break; 517 case CRYPTO_SHA2_256_HMAC: 518 case CRYPTO_SHA2_384_HMAC: 519 if (en->tls_vminor != TLS_MINOR_VER_TWO) 520 return (EINVAL); 521 break; 522 default: 523 return (EINVAL); 524 } 525 if (en->auth_key_len == 0) 526 return (EINVAL); 527 528 /* 529 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2 530 * use explicit IVs. 531 */ 532 switch (en->tls_vminor) { 533 case TLS_MINOR_VER_ZERO: 534 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 535 return (EINVAL); 536 break; 537 case TLS_MINOR_VER_ONE: 538 case TLS_MINOR_VER_TWO: 539 /* Ignore any supplied IV. */ 540 en->iv_len = 0; 541 break; 542 default: 543 return (EINVAL); 544 } 545 break; 546 case CRYPTO_CHACHA20_POLY1305: 547 if (en->auth_algorithm != 0 || en->auth_key_len != 0) 548 return (EINVAL); 549 if (en->tls_vminor != TLS_MINOR_VER_TWO && 550 en->tls_vminor != TLS_MINOR_VER_THREE) 551 return (EINVAL); 552 if (en->iv_len != TLS_CHACHA20_IV_LEN) 553 return (EINVAL); 554 break; 555 default: 556 return (EINVAL); 557 } 558 559 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 560 561 counter_u64_add(ktls_offload_active, 1); 562 563 refcount_init(&tls->refcount, 1); 564 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 565 566 tls->wq_index = ktls_get_cpu(so); 567 568 tls->params.cipher_algorithm = en->cipher_algorithm; 569 tls->params.auth_algorithm = en->auth_algorithm; 570 tls->params.tls_vmajor = en->tls_vmajor; 571 tls->params.tls_vminor = en->tls_vminor; 572 tls->params.flags = en->flags; 573 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 574 575 /* Set the header and trailer lengths. */ 576 tls->params.tls_hlen = sizeof(struct tls_record_layer); 577 switch (en->cipher_algorithm) { 578 case CRYPTO_AES_NIST_GCM_16: 579 /* 580 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 581 * nonce. TLS 1.3 uses a 12 byte implicit IV. 582 */ 583 if (en->tls_vminor < TLS_MINOR_VER_THREE) 584 tls->params.tls_hlen += sizeof(uint64_t); 585 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 586 tls->params.tls_bs = 1; 587 break; 588 case CRYPTO_AES_CBC: 589 switch (en->auth_algorithm) { 590 case CRYPTO_SHA1_HMAC: 591 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 592 /* Implicit IV, no nonce. */ 593 tls->sequential_records = true; 594 tls->next_seqno = be64dec(en->rec_seq); 595 STAILQ_INIT(&tls->pending_records); 596 } else { 597 tls->params.tls_hlen += AES_BLOCK_LEN; 598 } 599 tls->params.tls_tlen = AES_BLOCK_LEN + 600 SHA1_HASH_LEN; 601 break; 602 case CRYPTO_SHA2_256_HMAC: 603 tls->params.tls_hlen += AES_BLOCK_LEN; 604 tls->params.tls_tlen = AES_BLOCK_LEN + 605 SHA2_256_HASH_LEN; 606 break; 607 case CRYPTO_SHA2_384_HMAC: 608 tls->params.tls_hlen += AES_BLOCK_LEN; 609 tls->params.tls_tlen = AES_BLOCK_LEN + 610 SHA2_384_HASH_LEN; 611 break; 612 default: 613 panic("invalid hmac"); 614 } 615 tls->params.tls_bs = AES_BLOCK_LEN; 616 break; 617 case CRYPTO_CHACHA20_POLY1305: 618 /* 619 * Chacha20 uses a 12 byte implicit IV. 620 */ 621 tls->params.tls_tlen = POLY1305_HASH_LEN; 622 tls->params.tls_bs = 1; 623 break; 624 default: 625 panic("invalid cipher"); 626 } 627 628 /* 629 * TLS 1.3 includes optional padding which we do not support, 630 * and also puts the "real" record type at the end of the 631 * encrypted data. 632 */ 633 if (en->tls_vminor == TLS_MINOR_VER_THREE) 634 tls->params.tls_tlen += sizeof(uint8_t); 635 636 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 637 ("TLS header length too long: %d", tls->params.tls_hlen)); 638 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 639 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 640 641 if (en->auth_key_len != 0) { 642 tls->params.auth_key_len = en->auth_key_len; 643 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 644 M_WAITOK); 645 error = copyin(en->auth_key, tls->params.auth_key, 646 en->auth_key_len); 647 if (error) 648 goto out; 649 } 650 651 tls->params.cipher_key_len = en->cipher_key_len; 652 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 653 error = copyin(en->cipher_key, tls->params.cipher_key, 654 en->cipher_key_len); 655 if (error) 656 goto out; 657 658 /* 659 * This holds the implicit portion of the nonce for AEAD 660 * ciphers and the initial implicit IV for TLS 1.0. The 661 * explicit portions of the IV are generated in ktls_frame(). 662 */ 663 if (en->iv_len != 0) { 664 tls->params.iv_len = en->iv_len; 665 error = copyin(en->iv, tls->params.iv, en->iv_len); 666 if (error) 667 goto out; 668 669 /* 670 * For TLS 1.2 with GCM, generate an 8-byte nonce as a 671 * counter to generate unique explicit IVs. 672 * 673 * Store this counter in the last 8 bytes of the IV 674 * array so that it is 8-byte aligned. 675 */ 676 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 677 en->tls_vminor == TLS_MINOR_VER_TWO) 678 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); 679 } 680 681 *tlsp = tls; 682 return (0); 683 684 out: 685 ktls_free(tls); 686 return (error); 687 } 688 689 static struct ktls_session * 690 ktls_clone_session(struct ktls_session *tls) 691 { 692 struct ktls_session *tls_new; 693 694 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 695 696 counter_u64_add(ktls_offload_active, 1); 697 698 refcount_init(&tls_new->refcount, 1); 699 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new); 700 701 /* Copy fields from existing session. */ 702 tls_new->params = tls->params; 703 tls_new->wq_index = tls->wq_index; 704 705 /* Deep copy keys. */ 706 if (tls_new->params.auth_key != NULL) { 707 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 708 M_KTLS, M_WAITOK); 709 memcpy(tls_new->params.auth_key, tls->params.auth_key, 710 tls->params.auth_key_len); 711 } 712 713 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 714 M_WAITOK); 715 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 716 tls->params.cipher_key_len); 717 718 return (tls_new); 719 } 720 #endif 721 722 static void 723 ktls_cleanup(struct ktls_session *tls) 724 { 725 726 counter_u64_add(ktls_offload_active, -1); 727 switch (tls->mode) { 728 case TCP_TLS_MODE_SW: 729 MPASS(tls->be != NULL); 730 switch (tls->params.cipher_algorithm) { 731 case CRYPTO_AES_CBC: 732 counter_u64_add(ktls_sw_cbc, -1); 733 break; 734 case CRYPTO_AES_NIST_GCM_16: 735 counter_u64_add(ktls_sw_gcm, -1); 736 break; 737 case CRYPTO_CHACHA20_POLY1305: 738 counter_u64_add(ktls_sw_chacha20, -1); 739 break; 740 } 741 tls->free(tls); 742 break; 743 case TCP_TLS_MODE_IFNET: 744 switch (tls->params.cipher_algorithm) { 745 case CRYPTO_AES_CBC: 746 counter_u64_add(ktls_ifnet_cbc, -1); 747 break; 748 case CRYPTO_AES_NIST_GCM_16: 749 counter_u64_add(ktls_ifnet_gcm, -1); 750 break; 751 case CRYPTO_CHACHA20_POLY1305: 752 counter_u64_add(ktls_ifnet_chacha20, -1); 753 break; 754 } 755 if (tls->snd_tag != NULL) 756 m_snd_tag_rele(tls->snd_tag); 757 break; 758 #ifdef TCP_OFFLOAD 759 case TCP_TLS_MODE_TOE: 760 switch (tls->params.cipher_algorithm) { 761 case CRYPTO_AES_CBC: 762 counter_u64_add(ktls_toe_cbc, -1); 763 break; 764 case CRYPTO_AES_NIST_GCM_16: 765 counter_u64_add(ktls_toe_gcm, -1); 766 break; 767 case CRYPTO_CHACHA20_POLY1305: 768 counter_u64_add(ktls_toe_chacha20, -1); 769 break; 770 } 771 break; 772 #endif 773 } 774 if (tls->params.auth_key != NULL) { 775 zfree(tls->params.auth_key, M_KTLS); 776 tls->params.auth_key = NULL; 777 tls->params.auth_key_len = 0; 778 } 779 if (tls->params.cipher_key != NULL) { 780 zfree(tls->params.cipher_key, M_KTLS); 781 tls->params.cipher_key = NULL; 782 tls->params.cipher_key_len = 0; 783 } 784 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 785 } 786 787 #if defined(INET) || defined(INET6) 788 789 #ifdef TCP_OFFLOAD 790 static int 791 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) 792 { 793 struct inpcb *inp; 794 struct tcpcb *tp; 795 int error; 796 797 inp = so->so_pcb; 798 INP_WLOCK(inp); 799 if (inp->inp_flags2 & INP_FREED) { 800 INP_WUNLOCK(inp); 801 return (ECONNRESET); 802 } 803 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 804 INP_WUNLOCK(inp); 805 return (ECONNRESET); 806 } 807 if (inp->inp_socket == NULL) { 808 INP_WUNLOCK(inp); 809 return (ECONNRESET); 810 } 811 tp = intotcpcb(inp); 812 if (!(tp->t_flags & TF_TOE)) { 813 INP_WUNLOCK(inp); 814 return (EOPNOTSUPP); 815 } 816 817 error = tcp_offload_alloc_tls_session(tp, tls, direction); 818 INP_WUNLOCK(inp); 819 if (error == 0) { 820 tls->mode = TCP_TLS_MODE_TOE; 821 switch (tls->params.cipher_algorithm) { 822 case CRYPTO_AES_CBC: 823 counter_u64_add(ktls_toe_cbc, 1); 824 break; 825 case CRYPTO_AES_NIST_GCM_16: 826 counter_u64_add(ktls_toe_gcm, 1); 827 break; 828 case CRYPTO_CHACHA20_POLY1305: 829 counter_u64_add(ktls_toe_chacha20, 1); 830 break; 831 } 832 } 833 return (error); 834 } 835 #endif 836 837 /* 838 * Common code used when first enabling ifnet TLS on a connection or 839 * when allocating a new ifnet TLS session due to a routing change. 840 * This function allocates a new TLS send tag on whatever interface 841 * the connection is currently routed over. 842 */ 843 static int 844 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 845 struct m_snd_tag **mstp) 846 { 847 union if_snd_tag_alloc_params params; 848 struct ifnet *ifp; 849 struct nhop_object *nh; 850 struct tcpcb *tp; 851 int error; 852 853 INP_RLOCK(inp); 854 if (inp->inp_flags2 & INP_FREED) { 855 INP_RUNLOCK(inp); 856 return (ECONNRESET); 857 } 858 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 859 INP_RUNLOCK(inp); 860 return (ECONNRESET); 861 } 862 if (inp->inp_socket == NULL) { 863 INP_RUNLOCK(inp); 864 return (ECONNRESET); 865 } 866 tp = intotcpcb(inp); 867 868 /* 869 * Check administrative controls on ifnet TLS to determine if 870 * ifnet TLS should be denied. 871 * 872 * - Always permit 'force' requests. 873 * - ktls_ifnet_permitted == 0: always deny. 874 */ 875 if (!force && ktls_ifnet_permitted == 0) { 876 INP_RUNLOCK(inp); 877 return (ENXIO); 878 } 879 880 /* 881 * XXX: Use the cached route in the inpcb to find the 882 * interface. This should perhaps instead use 883 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 884 * enabled after a connection has completed key negotiation in 885 * userland, the cached route will be present in practice. 886 */ 887 nh = inp->inp_route.ro_nh; 888 if (nh == NULL) { 889 INP_RUNLOCK(inp); 890 return (ENXIO); 891 } 892 ifp = nh->nh_ifp; 893 if_ref(ifp); 894 895 /* 896 * Allocate a TLS + ratelimit tag if the connection has an 897 * existing pacing rate. 898 */ 899 if (tp->t_pacing_rate != -1 && 900 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) { 901 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT; 902 params.tls_rate_limit.inp = inp; 903 params.tls_rate_limit.tls = tls; 904 params.tls_rate_limit.max_rate = tp->t_pacing_rate; 905 } else { 906 params.hdr.type = IF_SND_TAG_TYPE_TLS; 907 params.tls.inp = inp; 908 params.tls.tls = tls; 909 } 910 params.hdr.flowid = inp->inp_flowid; 911 params.hdr.flowtype = inp->inp_flowtype; 912 params.hdr.numa_domain = inp->inp_numa_domain; 913 INP_RUNLOCK(inp); 914 915 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) { 916 error = EOPNOTSUPP; 917 goto out; 918 } 919 if (inp->inp_vflag & INP_IPV6) { 920 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { 921 error = EOPNOTSUPP; 922 goto out; 923 } 924 } else { 925 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { 926 error = EOPNOTSUPP; 927 goto out; 928 } 929 } 930 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 931 out: 932 if_rele(ifp); 933 return (error); 934 } 935 936 static int 937 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force) 938 { 939 struct m_snd_tag *mst; 940 int error; 941 942 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 943 if (error == 0) { 944 tls->mode = TCP_TLS_MODE_IFNET; 945 tls->snd_tag = mst; 946 switch (tls->params.cipher_algorithm) { 947 case CRYPTO_AES_CBC: 948 counter_u64_add(ktls_ifnet_cbc, 1); 949 break; 950 case CRYPTO_AES_NIST_GCM_16: 951 counter_u64_add(ktls_ifnet_gcm, 1); 952 break; 953 case CRYPTO_CHACHA20_POLY1305: 954 counter_u64_add(ktls_ifnet_chacha20, 1); 955 break; 956 } 957 } 958 return (error); 959 } 960 961 static int 962 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction) 963 { 964 struct rm_priotracker prio; 965 struct ktls_crypto_backend *be; 966 967 /* 968 * Choose the best software crypto backend. Backends are 969 * stored in sorted priority order (larget value == most 970 * important at the head of the list), so this just stops on 971 * the first backend that claims the session by returning 972 * success. 973 */ 974 if (ktls_allow_unload) 975 rm_rlock(&ktls_backends_lock, &prio); 976 LIST_FOREACH(be, &ktls_backends, next) { 977 if (be->try(so, tls, direction) == 0) 978 break; 979 KASSERT(tls->cipher == NULL, 980 ("ktls backend leaked a cipher pointer")); 981 } 982 if (be != NULL) { 983 if (ktls_allow_unload) 984 be->use_count++; 985 tls->be = be; 986 } 987 if (ktls_allow_unload) 988 rm_runlock(&ktls_backends_lock, &prio); 989 if (be == NULL) 990 return (EOPNOTSUPP); 991 tls->mode = TCP_TLS_MODE_SW; 992 switch (tls->params.cipher_algorithm) { 993 case CRYPTO_AES_CBC: 994 counter_u64_add(ktls_sw_cbc, 1); 995 break; 996 case CRYPTO_AES_NIST_GCM_16: 997 counter_u64_add(ktls_sw_gcm, 1); 998 break; 999 case CRYPTO_CHACHA20_POLY1305: 1000 counter_u64_add(ktls_sw_chacha20, 1); 1001 break; 1002 } 1003 return (0); 1004 } 1005 1006 /* 1007 * KTLS RX stores data in the socket buffer as a list of TLS records, 1008 * where each record is stored as a control message containg the TLS 1009 * header followed by data mbufs containing the decrypted data. This 1010 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for 1011 * both encrypted and decrypted data. TLS records decrypted by a NIC 1012 * should be queued to the socket buffer as records, but encrypted 1013 * data which needs to be decrypted by software arrives as a stream of 1014 * regular mbufs which need to be converted. In addition, there may 1015 * already be pending encrypted data in the socket buffer when KTLS RX 1016 * is enabled. 1017 * 1018 * To manage not-yet-decrypted data for KTLS RX, the following scheme 1019 * is used: 1020 * 1021 * - A single chain of NOTREADY mbufs is hung off of sb_mtls. 1022 * 1023 * - ktls_check_rx checks this chain of mbufs reading the TLS header 1024 * from the first mbuf. Once all of the data for that TLS record is 1025 * queued, the socket is queued to a worker thread. 1026 * 1027 * - The worker thread calls ktls_decrypt to decrypt TLS records in 1028 * the TLS chain. Each TLS record is detached from the TLS chain, 1029 * decrypted, and inserted into the regular socket buffer chain as 1030 * record starting with a control message holding the TLS header and 1031 * a chain of mbufs holding the encrypted data. 1032 */ 1033 1034 static void 1035 sb_mark_notready(struct sockbuf *sb) 1036 { 1037 struct mbuf *m; 1038 1039 m = sb->sb_mb; 1040 sb->sb_mtls = m; 1041 sb->sb_mb = NULL; 1042 sb->sb_mbtail = NULL; 1043 sb->sb_lastrecord = NULL; 1044 for (; m != NULL; m = m->m_next) { 1045 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", 1046 __func__)); 1047 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", 1048 __func__)); 1049 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", 1050 __func__)); 1051 m->m_flags |= M_NOTREADY; 1052 sb->sb_acc -= m->m_len; 1053 sb->sb_tlscc += m->m_len; 1054 sb->sb_mtlstail = m; 1055 } 1056 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, 1057 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, 1058 sb->sb_ccc)); 1059 } 1060 1061 int 1062 ktls_enable_rx(struct socket *so, struct tls_enable *en) 1063 { 1064 struct ktls_session *tls; 1065 int error; 1066 1067 if (!ktls_offload_enable) 1068 return (ENOTSUP); 1069 if (SOLISTENING(so)) 1070 return (EINVAL); 1071 1072 counter_u64_add(ktls_offload_enable_calls, 1); 1073 1074 /* 1075 * This should always be true since only the TCP socket option 1076 * invokes this function. 1077 */ 1078 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1079 return (EINVAL); 1080 1081 /* 1082 * XXX: Don't overwrite existing sessions. We should permit 1083 * this to support rekeying in the future. 1084 */ 1085 if (so->so_rcv.sb_tls_info != NULL) 1086 return (EALREADY); 1087 1088 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1089 return (ENOTSUP); 1090 1091 error = ktls_create_session(so, en, &tls); 1092 if (error) 1093 return (error); 1094 1095 #ifdef TCP_OFFLOAD 1096 error = ktls_try_toe(so, tls, KTLS_RX); 1097 if (error) 1098 #endif 1099 error = ktls_try_sw(so, tls, KTLS_RX); 1100 1101 if (error) { 1102 ktls_free(tls); 1103 return (error); 1104 } 1105 1106 /* Mark the socket as using TLS offload. */ 1107 SOCKBUF_LOCK(&so->so_rcv); 1108 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); 1109 so->so_rcv.sb_tls_info = tls; 1110 so->so_rcv.sb_flags |= SB_TLS_RX; 1111 1112 /* Mark existing data as not ready until it can be decrypted. */ 1113 if (tls->mode != TCP_TLS_MODE_TOE) { 1114 sb_mark_notready(&so->so_rcv); 1115 ktls_check_rx(&so->so_rcv); 1116 } 1117 SOCKBUF_UNLOCK(&so->so_rcv); 1118 1119 counter_u64_add(ktls_offload_total, 1); 1120 1121 return (0); 1122 } 1123 1124 int 1125 ktls_enable_tx(struct socket *so, struct tls_enable *en) 1126 { 1127 struct ktls_session *tls; 1128 struct inpcb *inp; 1129 int error; 1130 1131 if (!ktls_offload_enable) 1132 return (ENOTSUP); 1133 if (SOLISTENING(so)) 1134 return (EINVAL); 1135 1136 counter_u64_add(ktls_offload_enable_calls, 1); 1137 1138 /* 1139 * This should always be true since only the TCP socket option 1140 * invokes this function. 1141 */ 1142 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1143 return (EINVAL); 1144 1145 /* 1146 * XXX: Don't overwrite existing sessions. We should permit 1147 * this to support rekeying in the future. 1148 */ 1149 if (so->so_snd.sb_tls_info != NULL) 1150 return (EALREADY); 1151 1152 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1153 return (ENOTSUP); 1154 1155 /* TLS requires ext pgs */ 1156 if (mb_use_ext_pgs == 0) 1157 return (ENXIO); 1158 1159 error = ktls_create_session(so, en, &tls); 1160 if (error) 1161 return (error); 1162 1163 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1164 #ifdef TCP_OFFLOAD 1165 error = ktls_try_toe(so, tls, KTLS_TX); 1166 if (error) 1167 #endif 1168 error = ktls_try_ifnet(so, tls, false); 1169 if (error) 1170 error = ktls_try_sw(so, tls, KTLS_TX); 1171 1172 if (error) { 1173 ktls_free(tls); 1174 return (error); 1175 } 1176 1177 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1178 if (error) { 1179 ktls_free(tls); 1180 return (error); 1181 } 1182 1183 /* 1184 * Write lock the INP when setting sb_tls_info so that 1185 * routines in tcp_ratelimit.c can read sb_tls_info while 1186 * holding the INP lock. 1187 */ 1188 inp = so->so_pcb; 1189 INP_WLOCK(inp); 1190 SOCKBUF_LOCK(&so->so_snd); 1191 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); 1192 so->so_snd.sb_tls_info = tls; 1193 if (tls->mode != TCP_TLS_MODE_SW) 1194 so->so_snd.sb_flags |= SB_TLS_IFNET; 1195 SOCKBUF_UNLOCK(&so->so_snd); 1196 INP_WUNLOCK(inp); 1197 SOCK_IO_SEND_UNLOCK(so); 1198 1199 counter_u64_add(ktls_offload_total, 1); 1200 1201 return (0); 1202 } 1203 1204 int 1205 ktls_get_rx_mode(struct socket *so) 1206 { 1207 struct ktls_session *tls; 1208 struct inpcb *inp; 1209 int mode; 1210 1211 if (SOLISTENING(so)) 1212 return (EINVAL); 1213 inp = so->so_pcb; 1214 INP_WLOCK_ASSERT(inp); 1215 SOCKBUF_LOCK(&so->so_rcv); 1216 tls = so->so_rcv.sb_tls_info; 1217 if (tls == NULL) 1218 mode = TCP_TLS_MODE_NONE; 1219 else 1220 mode = tls->mode; 1221 SOCKBUF_UNLOCK(&so->so_rcv); 1222 return (mode); 1223 } 1224 1225 int 1226 ktls_get_tx_mode(struct socket *so) 1227 { 1228 struct ktls_session *tls; 1229 struct inpcb *inp; 1230 int mode; 1231 1232 if (SOLISTENING(so)) 1233 return (EINVAL); 1234 inp = so->so_pcb; 1235 INP_WLOCK_ASSERT(inp); 1236 SOCKBUF_LOCK(&so->so_snd); 1237 tls = so->so_snd.sb_tls_info; 1238 if (tls == NULL) 1239 mode = TCP_TLS_MODE_NONE; 1240 else 1241 mode = tls->mode; 1242 SOCKBUF_UNLOCK(&so->so_snd); 1243 return (mode); 1244 } 1245 1246 /* 1247 * Switch between SW and ifnet TLS sessions as requested. 1248 */ 1249 int 1250 ktls_set_tx_mode(struct socket *so, int mode) 1251 { 1252 struct ktls_session *tls, *tls_new; 1253 struct inpcb *inp; 1254 int error; 1255 1256 if (SOLISTENING(so)) 1257 return (EINVAL); 1258 switch (mode) { 1259 case TCP_TLS_MODE_SW: 1260 case TCP_TLS_MODE_IFNET: 1261 break; 1262 default: 1263 return (EINVAL); 1264 } 1265 1266 inp = so->so_pcb; 1267 INP_WLOCK_ASSERT(inp); 1268 SOCKBUF_LOCK(&so->so_snd); 1269 tls = so->so_snd.sb_tls_info; 1270 if (tls == NULL) { 1271 SOCKBUF_UNLOCK(&so->so_snd); 1272 return (0); 1273 } 1274 1275 if (tls->mode == mode) { 1276 SOCKBUF_UNLOCK(&so->so_snd); 1277 return (0); 1278 } 1279 1280 tls = ktls_hold(tls); 1281 SOCKBUF_UNLOCK(&so->so_snd); 1282 INP_WUNLOCK(inp); 1283 1284 tls_new = ktls_clone_session(tls); 1285 1286 if (mode == TCP_TLS_MODE_IFNET) 1287 error = ktls_try_ifnet(so, tls_new, true); 1288 else 1289 error = ktls_try_sw(so, tls_new, KTLS_TX); 1290 if (error) { 1291 counter_u64_add(ktls_switch_failed, 1); 1292 ktls_free(tls_new); 1293 ktls_free(tls); 1294 INP_WLOCK(inp); 1295 return (error); 1296 } 1297 1298 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1299 if (error) { 1300 counter_u64_add(ktls_switch_failed, 1); 1301 ktls_free(tls_new); 1302 ktls_free(tls); 1303 INP_WLOCK(inp); 1304 return (error); 1305 } 1306 1307 /* 1308 * If we raced with another session change, keep the existing 1309 * session. 1310 */ 1311 if (tls != so->so_snd.sb_tls_info) { 1312 counter_u64_add(ktls_switch_failed, 1); 1313 SOCK_IO_SEND_UNLOCK(so); 1314 ktls_free(tls_new); 1315 ktls_free(tls); 1316 INP_WLOCK(inp); 1317 return (EBUSY); 1318 } 1319 1320 INP_WLOCK(inp); 1321 SOCKBUF_LOCK(&so->so_snd); 1322 so->so_snd.sb_tls_info = tls_new; 1323 if (tls_new->mode != TCP_TLS_MODE_SW) 1324 so->so_snd.sb_flags |= SB_TLS_IFNET; 1325 SOCKBUF_UNLOCK(&so->so_snd); 1326 SOCK_IO_SEND_UNLOCK(so); 1327 1328 /* 1329 * Drop two references on 'tls'. The first is for the 1330 * ktls_hold() above. The second drops the reference from the 1331 * socket buffer. 1332 */ 1333 KASSERT(tls->refcount >= 2, ("too few references on old session")); 1334 ktls_free(tls); 1335 ktls_free(tls); 1336 1337 if (mode == TCP_TLS_MODE_IFNET) 1338 counter_u64_add(ktls_switch_to_ifnet, 1); 1339 else 1340 counter_u64_add(ktls_switch_to_sw, 1); 1341 1342 return (0); 1343 } 1344 1345 /* 1346 * Try to allocate a new TLS send tag. This task is scheduled when 1347 * ip_output detects a route change while trying to transmit a packet 1348 * holding a TLS record. If a new tag is allocated, replace the tag 1349 * in the TLS session. Subsequent packets on the connection will use 1350 * the new tag. If a new tag cannot be allocated, drop the 1351 * connection. 1352 */ 1353 static void 1354 ktls_reset_send_tag(void *context, int pending) 1355 { 1356 struct epoch_tracker et; 1357 struct ktls_session *tls; 1358 struct m_snd_tag *old, *new; 1359 struct inpcb *inp; 1360 struct tcpcb *tp; 1361 int error; 1362 1363 MPASS(pending == 1); 1364 1365 tls = context; 1366 inp = tls->inp; 1367 1368 /* 1369 * Free the old tag first before allocating a new one. 1370 * ip[6]_output_send() will treat a NULL send tag the same as 1371 * an ifp mismatch and drop packets until a new tag is 1372 * allocated. 1373 * 1374 * Write-lock the INP when changing tls->snd_tag since 1375 * ip[6]_output_send() holds a read-lock when reading the 1376 * pointer. 1377 */ 1378 INP_WLOCK(inp); 1379 old = tls->snd_tag; 1380 tls->snd_tag = NULL; 1381 INP_WUNLOCK(inp); 1382 if (old != NULL) 1383 m_snd_tag_rele(old); 1384 1385 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1386 1387 if (error == 0) { 1388 INP_WLOCK(inp); 1389 tls->snd_tag = new; 1390 mtx_pool_lock(mtxpool_sleep, tls); 1391 tls->reset_pending = false; 1392 mtx_pool_unlock(mtxpool_sleep, tls); 1393 if (!in_pcbrele_wlocked(inp)) 1394 INP_WUNLOCK(inp); 1395 1396 counter_u64_add(ktls_ifnet_reset, 1); 1397 1398 /* 1399 * XXX: Should we kick tcp_output explicitly now that 1400 * the send tag is fixed or just rely on timers? 1401 */ 1402 } else { 1403 NET_EPOCH_ENTER(et); 1404 INP_WLOCK(inp); 1405 if (!in_pcbrele_wlocked(inp)) { 1406 if (!(inp->inp_flags & INP_TIMEWAIT) && 1407 !(inp->inp_flags & INP_DROPPED)) { 1408 tp = intotcpcb(inp); 1409 CURVNET_SET(tp->t_vnet); 1410 tp = tcp_drop(tp, ECONNABORTED); 1411 CURVNET_RESTORE(); 1412 if (tp != NULL) 1413 INP_WUNLOCK(inp); 1414 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1415 } else 1416 INP_WUNLOCK(inp); 1417 } 1418 NET_EPOCH_EXIT(et); 1419 1420 counter_u64_add(ktls_ifnet_reset_failed, 1); 1421 1422 /* 1423 * Leave reset_pending true to avoid future tasks while 1424 * the socket goes away. 1425 */ 1426 } 1427 1428 ktls_free(tls); 1429 } 1430 1431 int 1432 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1433 { 1434 1435 if (inp == NULL) 1436 return (ENOBUFS); 1437 1438 INP_LOCK_ASSERT(inp); 1439 1440 /* 1441 * See if we should schedule a task to update the send tag for 1442 * this session. 1443 */ 1444 mtx_pool_lock(mtxpool_sleep, tls); 1445 if (!tls->reset_pending) { 1446 (void) ktls_hold(tls); 1447 in_pcbref(inp); 1448 tls->inp = inp; 1449 tls->reset_pending = true; 1450 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1451 } 1452 mtx_pool_unlock(mtxpool_sleep, tls); 1453 return (ENOBUFS); 1454 } 1455 1456 #ifdef RATELIMIT 1457 int 1458 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate) 1459 { 1460 union if_snd_tag_modify_params params = { 1461 .rate_limit.max_rate = max_pacing_rate, 1462 .rate_limit.flags = M_NOWAIT, 1463 }; 1464 struct m_snd_tag *mst; 1465 struct ifnet *ifp; 1466 int error; 1467 1468 /* Can't get to the inp, but it should be locked. */ 1469 /* INP_LOCK_ASSERT(inp); */ 1470 1471 MPASS(tls->mode == TCP_TLS_MODE_IFNET); 1472 1473 if (tls->snd_tag == NULL) { 1474 /* 1475 * Resetting send tag, ignore this change. The 1476 * pending reset may or may not see this updated rate 1477 * in the tcpcb. If it doesn't, we will just lose 1478 * this rate change. 1479 */ 1480 return (0); 1481 } 1482 1483 mst = tls->snd_tag; 1484 1485 MPASS(mst != NULL); 1486 MPASS(mst->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT); 1487 1488 ifp = mst->ifp; 1489 return (ifp->if_snd_tag_modify(mst, ¶ms)); 1490 } 1491 #endif 1492 #endif 1493 1494 void 1495 ktls_destroy(struct ktls_session *tls) 1496 { 1497 struct rm_priotracker prio; 1498 1499 if (tls->sequential_records) { 1500 struct mbuf *m, *n; 1501 int page_count; 1502 1503 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) { 1504 page_count = m->m_epg_enc_cnt; 1505 while (page_count > 0) { 1506 KASSERT(page_count >= m->m_epg_nrdy, 1507 ("%s: too few pages", __func__)); 1508 page_count -= m->m_epg_nrdy; 1509 m = m_free(m); 1510 } 1511 } 1512 } 1513 ktls_cleanup(tls); 1514 if (tls->be != NULL && ktls_allow_unload) { 1515 rm_rlock(&ktls_backends_lock, &prio); 1516 tls->be->use_count--; 1517 rm_runlock(&ktls_backends_lock, &prio); 1518 } 1519 uma_zfree(ktls_session_zone, tls); 1520 } 1521 1522 void 1523 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1524 { 1525 1526 for (; m != NULL; m = m->m_next) { 1527 KASSERT((m->m_flags & M_EXTPG) != 0, 1528 ("ktls_seq: mapped mbuf %p", m)); 1529 1530 m->m_epg_seqno = sb->sb_tls_seqno; 1531 sb->sb_tls_seqno++; 1532 } 1533 } 1534 1535 /* 1536 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1537 * mbuf in the chain must be an unmapped mbuf. The payload of the 1538 * mbuf must be populated with the payload of each TLS record. 1539 * 1540 * The record_type argument specifies the TLS record type used when 1541 * populating the TLS header. 1542 * 1543 * The enq_count argument on return is set to the number of pages of 1544 * payload data for this entire chain that need to be encrypted via SW 1545 * encryption. The returned value should be passed to ktls_enqueue 1546 * when scheduling encryption of this chain of mbufs. To handle the 1547 * special case of empty fragments for TLS 1.0 sessions, an empty 1548 * fragment counts as one page. 1549 */ 1550 void 1551 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1552 uint8_t record_type) 1553 { 1554 struct tls_record_layer *tlshdr; 1555 struct mbuf *m; 1556 uint64_t *noncep; 1557 uint16_t tls_len; 1558 int maxlen; 1559 1560 maxlen = tls->params.max_frame_len; 1561 *enq_cnt = 0; 1562 for (m = top; m != NULL; m = m->m_next) { 1563 /* 1564 * All mbufs in the chain should be TLS records whose 1565 * payload does not exceed the maximum frame length. 1566 * 1567 * Empty TLS 1.0 records are permitted when using CBC. 1568 */ 1569 KASSERT(m->m_len <= maxlen && m->m_len >= 0 && 1570 (m->m_len > 0 || ktls_permit_empty_frames(tls)), 1571 ("ktls_frame: m %p len %d", m, m->m_len)); 1572 1573 /* 1574 * TLS frames require unmapped mbufs to store session 1575 * info. 1576 */ 1577 KASSERT((m->m_flags & M_EXTPG) != 0, 1578 ("ktls_frame: mapped mbuf %p (top = %p)", m, top)); 1579 1580 tls_len = m->m_len; 1581 1582 /* Save a reference to the session. */ 1583 m->m_epg_tls = ktls_hold(tls); 1584 1585 m->m_epg_hdrlen = tls->params.tls_hlen; 1586 m->m_epg_trllen = tls->params.tls_tlen; 1587 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 1588 int bs, delta; 1589 1590 /* 1591 * AES-CBC pads messages to a multiple of the 1592 * block size. Note that the padding is 1593 * applied after the digest and the encryption 1594 * is done on the "plaintext || mac || padding". 1595 * At least one byte of padding is always 1596 * present. 1597 * 1598 * Compute the final trailer length assuming 1599 * at most one block of padding. 1600 * tls->params.sb_tls_tlen is the maximum 1601 * possible trailer length (padding + digest). 1602 * delta holds the number of excess padding 1603 * bytes if the maximum were used. Those 1604 * extra bytes are removed. 1605 */ 1606 bs = tls->params.tls_bs; 1607 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 1608 m->m_epg_trllen -= delta; 1609 } 1610 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; 1611 1612 /* Populate the TLS header. */ 1613 tlshdr = (void *)m->m_epg_hdr; 1614 tlshdr->tls_vmajor = tls->params.tls_vmajor; 1615 1616 /* 1617 * TLS 1.3 masquarades as TLS 1.2 with a record type 1618 * of TLS_RLTYPE_APP. 1619 */ 1620 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 1621 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 1622 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 1623 tlshdr->tls_type = TLS_RLTYPE_APP; 1624 /* save the real record type for later */ 1625 m->m_epg_record_type = record_type; 1626 m->m_epg_trail[0] = record_type; 1627 } else { 1628 tlshdr->tls_vminor = tls->params.tls_vminor; 1629 tlshdr->tls_type = record_type; 1630 } 1631 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 1632 1633 /* 1634 * Store nonces / explicit IVs after the end of the 1635 * TLS header. 1636 * 1637 * For GCM with TLS 1.2, an 8 byte nonce is copied 1638 * from the end of the IV. The nonce is then 1639 * incremented for use by the next record. 1640 * 1641 * For CBC, a random nonce is inserted for TLS 1.1+. 1642 */ 1643 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 1644 tls->params.tls_vminor == TLS_MINOR_VER_TWO) { 1645 noncep = (uint64_t *)(tls->params.iv + 8); 1646 be64enc(tlshdr + 1, *noncep); 1647 (*noncep)++; 1648 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 1649 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 1650 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 1651 1652 /* 1653 * When using SW encryption, mark the mbuf not ready. 1654 * It will be marked ready via sbready() after the 1655 * record has been encrypted. 1656 * 1657 * When using ifnet TLS, unencrypted TLS records are 1658 * sent down the stack to the NIC. 1659 */ 1660 if (tls->mode == TCP_TLS_MODE_SW) { 1661 m->m_flags |= M_NOTREADY; 1662 if (__predict_false(tls_len == 0)) { 1663 /* TLS 1.0 empty fragment. */ 1664 m->m_epg_nrdy = 1; 1665 } else 1666 m->m_epg_nrdy = m->m_epg_npgs; 1667 *enq_cnt += m->m_epg_nrdy; 1668 } 1669 } 1670 } 1671 1672 bool 1673 ktls_permit_empty_frames(struct ktls_session *tls) 1674 { 1675 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 1676 tls->params.tls_vminor == TLS_MINOR_VER_ZERO); 1677 } 1678 1679 void 1680 ktls_check_rx(struct sockbuf *sb) 1681 { 1682 struct tls_record_layer hdr; 1683 struct ktls_wq *wq; 1684 struct socket *so; 1685 bool running; 1686 1687 SOCKBUF_LOCK_ASSERT(sb); 1688 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 1689 __func__, sb)); 1690 so = __containerof(sb, struct socket, so_rcv); 1691 1692 if (sb->sb_flags & SB_TLS_RX_RUNNING) 1693 return; 1694 1695 /* Is there enough queued for a TLS header? */ 1696 if (sb->sb_tlscc < sizeof(hdr)) { 1697 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) 1698 so->so_error = EMSGSIZE; 1699 return; 1700 } 1701 1702 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); 1703 1704 /* Is the entire record queued? */ 1705 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { 1706 if ((sb->sb_state & SBS_CANTRCVMORE) != 0) 1707 so->so_error = EMSGSIZE; 1708 return; 1709 } 1710 1711 sb->sb_flags |= SB_TLS_RX_RUNNING; 1712 1713 soref(so); 1714 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; 1715 mtx_lock(&wq->mtx); 1716 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); 1717 running = wq->running; 1718 mtx_unlock(&wq->mtx); 1719 if (!running) 1720 wakeup(wq); 1721 counter_u64_add(ktls_cnt_rx_queued, 1); 1722 } 1723 1724 static struct mbuf * 1725 ktls_detach_record(struct sockbuf *sb, int len) 1726 { 1727 struct mbuf *m, *n, *top; 1728 int remain; 1729 1730 SOCKBUF_LOCK_ASSERT(sb); 1731 MPASS(len <= sb->sb_tlscc); 1732 1733 /* 1734 * If TLS chain is the exact size of the record, 1735 * just grab the whole record. 1736 */ 1737 top = sb->sb_mtls; 1738 if (sb->sb_tlscc == len) { 1739 sb->sb_mtls = NULL; 1740 sb->sb_mtlstail = NULL; 1741 goto out; 1742 } 1743 1744 /* 1745 * While it would be nice to use m_split() here, we need 1746 * to know exactly what m_split() allocates to update the 1747 * accounting, so do it inline instead. 1748 */ 1749 remain = len; 1750 for (m = top; remain > m->m_len; m = m->m_next) 1751 remain -= m->m_len; 1752 1753 /* Easy case: don't have to split 'm'. */ 1754 if (remain == m->m_len) { 1755 sb->sb_mtls = m->m_next; 1756 if (sb->sb_mtls == NULL) 1757 sb->sb_mtlstail = NULL; 1758 m->m_next = NULL; 1759 goto out; 1760 } 1761 1762 /* 1763 * Need to allocate an mbuf to hold the remainder of 'm'. Try 1764 * with M_NOWAIT first. 1765 */ 1766 n = m_get(M_NOWAIT, MT_DATA); 1767 if (n == NULL) { 1768 /* 1769 * Use M_WAITOK with socket buffer unlocked. If 1770 * 'sb_mtls' changes while the lock is dropped, return 1771 * NULL to force the caller to retry. 1772 */ 1773 SOCKBUF_UNLOCK(sb); 1774 1775 n = m_get(M_WAITOK, MT_DATA); 1776 1777 SOCKBUF_LOCK(sb); 1778 if (sb->sb_mtls != top) { 1779 m_free(n); 1780 return (NULL); 1781 } 1782 } 1783 n->m_flags |= M_NOTREADY; 1784 1785 /* Store remainder in 'n'. */ 1786 n->m_len = m->m_len - remain; 1787 if (m->m_flags & M_EXT) { 1788 n->m_data = m->m_data + remain; 1789 mb_dupcl(n, m); 1790 } else { 1791 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); 1792 } 1793 1794 /* Trim 'm' and update accounting. */ 1795 m->m_len -= n->m_len; 1796 sb->sb_tlscc -= n->m_len; 1797 sb->sb_ccc -= n->m_len; 1798 1799 /* Account for 'n'. */ 1800 sballoc_ktls_rx(sb, n); 1801 1802 /* Insert 'n' into the TLS chain. */ 1803 sb->sb_mtls = n; 1804 n->m_next = m->m_next; 1805 if (sb->sb_mtlstail == m) 1806 sb->sb_mtlstail = n; 1807 1808 /* Detach the record from the TLS chain. */ 1809 m->m_next = NULL; 1810 1811 out: 1812 MPASS(m_length(top, NULL) == len); 1813 for (m = top; m != NULL; m = m->m_next) 1814 sbfree_ktls_rx(sb, m); 1815 sb->sb_tlsdcc = len; 1816 sb->sb_ccc += len; 1817 SBCHECK(sb); 1818 return (top); 1819 } 1820 1821 /* 1822 * Determine the length of the trailing zero padding and find the real 1823 * record type in the byte before the padding. 1824 * 1825 * Walking the mbuf chain backwards is clumsy, so another option would 1826 * be to scan forwards remembering the last non-zero byte before the 1827 * trailer. However, it would be expensive to scan the entire record. 1828 * Instead, find the last non-zero byte of each mbuf in the chain 1829 * keeping track of the relative offset of that nonzero byte. 1830 * 1831 * trail_len is the size of the MAC/tag on input and is set to the 1832 * size of the full trailer including padding and the record type on 1833 * return. 1834 */ 1835 static int 1836 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, 1837 int *trailer_len, uint8_t *record_typep) 1838 { 1839 char *cp; 1840 u_int digest_start, last_offset, m_len, offset; 1841 uint8_t record_type; 1842 1843 digest_start = tls_len - *trailer_len; 1844 last_offset = 0; 1845 offset = 0; 1846 for (; m != NULL && offset < digest_start; 1847 offset += m->m_len, m = m->m_next) { 1848 /* Don't look for padding in the tag. */ 1849 m_len = min(digest_start - offset, m->m_len); 1850 cp = mtod(m, char *); 1851 1852 /* Find last non-zero byte in this mbuf. */ 1853 while (m_len > 0 && cp[m_len - 1] == 0) 1854 m_len--; 1855 if (m_len > 0) { 1856 record_type = cp[m_len - 1]; 1857 last_offset = offset + m_len; 1858 } 1859 } 1860 if (last_offset < tls->params.tls_hlen) 1861 return (EBADMSG); 1862 1863 *record_typep = record_type; 1864 *trailer_len = tls_len - last_offset + 1; 1865 return (0); 1866 } 1867 1868 static void 1869 ktls_drop(struct socket *so, int error) 1870 { 1871 struct epoch_tracker et; 1872 struct inpcb *inp = sotoinpcb(so); 1873 struct tcpcb *tp; 1874 1875 NET_EPOCH_ENTER(et); 1876 INP_WLOCK(inp); 1877 if (!(inp->inp_flags & INP_DROPPED)) { 1878 tp = intotcpcb(inp); 1879 CURVNET_SET(inp->inp_vnet); 1880 tp = tcp_drop(tp, error); 1881 CURVNET_RESTORE(); 1882 if (tp != NULL) 1883 INP_WUNLOCK(inp); 1884 } else { 1885 so->so_error = error; 1886 SOCKBUF_LOCK(&so->so_rcv); 1887 sorwakeup_locked(so); 1888 INP_WUNLOCK(inp); 1889 } 1890 NET_EPOCH_EXIT(et); 1891 } 1892 1893 static void 1894 ktls_decrypt(struct socket *so) 1895 { 1896 char tls_header[MBUF_PEXT_HDR_LEN]; 1897 struct ktls_session *tls; 1898 struct sockbuf *sb; 1899 struct tls_record_layer *hdr; 1900 struct tls_get_record tgr; 1901 struct mbuf *control, *data, *m; 1902 uint64_t seqno; 1903 int error, remain, tls_len, trail_len; 1904 bool tls13; 1905 uint8_t vminor, record_type; 1906 1907 hdr = (struct tls_record_layer *)tls_header; 1908 sb = &so->so_rcv; 1909 SOCKBUF_LOCK(sb); 1910 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, 1911 ("%s: socket %p not running", __func__, so)); 1912 1913 tls = sb->sb_tls_info; 1914 MPASS(tls != NULL); 1915 1916 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE); 1917 if (tls13) 1918 vminor = TLS_MINOR_VER_TWO; 1919 else 1920 vminor = tls->params.tls_vminor; 1921 for (;;) { 1922 /* Is there enough queued for a TLS header? */ 1923 if (sb->sb_tlscc < tls->params.tls_hlen) 1924 break; 1925 1926 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); 1927 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); 1928 1929 if (hdr->tls_vmajor != tls->params.tls_vmajor || 1930 hdr->tls_vminor != vminor) 1931 error = EINVAL; 1932 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP) 1933 error = EINVAL; 1934 else if (tls_len < tls->params.tls_hlen || tls_len > 1935 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + 1936 tls->params.tls_tlen) 1937 error = EMSGSIZE; 1938 else 1939 error = 0; 1940 if (__predict_false(error != 0)) { 1941 /* 1942 * We have a corrupted record and are likely 1943 * out of sync. The connection isn't 1944 * recoverable at this point, so abort it. 1945 */ 1946 SOCKBUF_UNLOCK(sb); 1947 counter_u64_add(ktls_offload_corrupted_records, 1); 1948 1949 ktls_drop(so, error); 1950 goto deref; 1951 } 1952 1953 /* Is the entire record queued? */ 1954 if (sb->sb_tlscc < tls_len) 1955 break; 1956 1957 /* 1958 * Split out the portion of the mbuf chain containing 1959 * this TLS record. 1960 */ 1961 data = ktls_detach_record(sb, tls_len); 1962 if (data == NULL) 1963 continue; 1964 MPASS(sb->sb_tlsdcc == tls_len); 1965 1966 seqno = sb->sb_tls_seqno; 1967 sb->sb_tls_seqno++; 1968 SBCHECK(sb); 1969 SOCKBUF_UNLOCK(sb); 1970 1971 error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len); 1972 if (error == 0) { 1973 if (tls13) 1974 error = tls13_find_record_type(tls, data, 1975 tls_len, &trail_len, &record_type); 1976 else 1977 record_type = hdr->tls_type; 1978 } 1979 if (error) { 1980 counter_u64_add(ktls_offload_failed_crypto, 1); 1981 1982 SOCKBUF_LOCK(sb); 1983 if (sb->sb_tlsdcc == 0) { 1984 /* 1985 * sbcut/drop/flush discarded these 1986 * mbufs. 1987 */ 1988 m_freem(data); 1989 break; 1990 } 1991 1992 /* 1993 * Drop this TLS record's data, but keep 1994 * decrypting subsequent records. 1995 */ 1996 sb->sb_ccc -= tls_len; 1997 sb->sb_tlsdcc = 0; 1998 1999 CURVNET_SET(so->so_vnet); 2000 so->so_error = EBADMSG; 2001 sorwakeup_locked(so); 2002 CURVNET_RESTORE(); 2003 2004 m_freem(data); 2005 2006 SOCKBUF_LOCK(sb); 2007 continue; 2008 } 2009 2010 /* Allocate the control mbuf. */ 2011 memset(&tgr, 0, sizeof(tgr)); 2012 tgr.tls_type = record_type; 2013 tgr.tls_vmajor = hdr->tls_vmajor; 2014 tgr.tls_vminor = hdr->tls_vminor; 2015 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - 2016 trail_len); 2017 control = sbcreatecontrol_how(&tgr, sizeof(tgr), 2018 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); 2019 2020 SOCKBUF_LOCK(sb); 2021 if (sb->sb_tlsdcc == 0) { 2022 /* sbcut/drop/flush discarded these mbufs. */ 2023 MPASS(sb->sb_tlscc == 0); 2024 m_freem(data); 2025 m_freem(control); 2026 break; 2027 } 2028 2029 /* 2030 * Clear the 'dcc' accounting in preparation for 2031 * adding the decrypted record. 2032 */ 2033 sb->sb_ccc -= tls_len; 2034 sb->sb_tlsdcc = 0; 2035 SBCHECK(sb); 2036 2037 /* If there is no payload, drop all of the data. */ 2038 if (tgr.tls_length == htobe16(0)) { 2039 m_freem(data); 2040 data = NULL; 2041 } else { 2042 /* Trim header. */ 2043 remain = tls->params.tls_hlen; 2044 while (remain > 0) { 2045 if (data->m_len > remain) { 2046 data->m_data += remain; 2047 data->m_len -= remain; 2048 break; 2049 } 2050 remain -= data->m_len; 2051 data = m_free(data); 2052 } 2053 2054 /* Trim trailer and clear M_NOTREADY. */ 2055 remain = be16toh(tgr.tls_length); 2056 m = data; 2057 for (m = data; remain > m->m_len; m = m->m_next) { 2058 m->m_flags &= ~M_NOTREADY; 2059 remain -= m->m_len; 2060 } 2061 m->m_len = remain; 2062 m_freem(m->m_next); 2063 m->m_next = NULL; 2064 m->m_flags &= ~M_NOTREADY; 2065 2066 /* Set EOR on the final mbuf. */ 2067 m->m_flags |= M_EOR; 2068 } 2069 2070 sbappendcontrol_locked(sb, data, control, 0); 2071 } 2072 2073 sb->sb_flags &= ~SB_TLS_RX_RUNNING; 2074 2075 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) 2076 so->so_error = EMSGSIZE; 2077 2078 sorwakeup_locked(so); 2079 2080 deref: 2081 SOCKBUF_UNLOCK_ASSERT(sb); 2082 2083 CURVNET_SET(so->so_vnet); 2084 SOCK_LOCK(so); 2085 sorele(so); 2086 CURVNET_RESTORE(); 2087 } 2088 2089 void 2090 ktls_enqueue_to_free(struct mbuf *m) 2091 { 2092 struct ktls_wq *wq; 2093 bool running; 2094 2095 /* Mark it for freeing. */ 2096 m->m_epg_flags |= EPG_FLAG_2FREE; 2097 wq = &ktls_wq[m->m_epg_tls->wq_index]; 2098 mtx_lock(&wq->mtx); 2099 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2100 running = wq->running; 2101 mtx_unlock(&wq->mtx); 2102 if (!running) 2103 wakeup(wq); 2104 } 2105 2106 /* Number of TLS records in a batch passed to ktls_enqueue(). */ 2107 static u_int 2108 ktls_batched_records(struct mbuf *m) 2109 { 2110 int page_count, records; 2111 2112 records = 0; 2113 page_count = m->m_epg_enc_cnt; 2114 while (page_count > 0) { 2115 records++; 2116 page_count -= m->m_epg_nrdy; 2117 m = m->m_next; 2118 } 2119 KASSERT(page_count == 0, ("%s: mismatched page count", __func__)); 2120 return (records); 2121 } 2122 2123 void 2124 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 2125 { 2126 struct ktls_session *tls; 2127 struct ktls_wq *wq; 2128 int queued; 2129 bool running; 2130 2131 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == 2132 (M_EXTPG | M_NOTREADY)), 2133 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 2134 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 2135 2136 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); 2137 2138 m->m_epg_enc_cnt = page_count; 2139 2140 /* 2141 * Save a pointer to the socket. The caller is responsible 2142 * for taking an additional reference via soref(). 2143 */ 2144 m->m_epg_so = so; 2145 2146 queued = 1; 2147 tls = m->m_epg_tls; 2148 wq = &ktls_wq[tls->wq_index]; 2149 mtx_lock(&wq->mtx); 2150 if (__predict_false(tls->sequential_records)) { 2151 /* 2152 * For TLS 1.0, records must be encrypted 2153 * sequentially. For a given connection, all records 2154 * queued to the associated work queue are processed 2155 * sequentially. However, sendfile(2) might complete 2156 * I/O requests spanning multiple TLS records out of 2157 * order. Here we ensure TLS records are enqueued to 2158 * the work queue in FIFO order. 2159 * 2160 * tls->next_seqno holds the sequence number of the 2161 * next TLS record that should be enqueued to the work 2162 * queue. If this next record is not tls->next_seqno, 2163 * it must be a future record, so insert it, sorted by 2164 * TLS sequence number, into tls->pending_records and 2165 * return. 2166 * 2167 * If this TLS record matches tls->next_seqno, place 2168 * it in the work queue and then check 2169 * tls->pending_records to see if any 2170 * previously-queued records are now ready for 2171 * encryption. 2172 */ 2173 if (m->m_epg_seqno != tls->next_seqno) { 2174 struct mbuf *n, *p; 2175 2176 p = NULL; 2177 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) { 2178 if (n->m_epg_seqno > m->m_epg_seqno) 2179 break; 2180 p = n; 2181 } 2182 if (n == NULL) 2183 STAILQ_INSERT_TAIL(&tls->pending_records, m, 2184 m_epg_stailq); 2185 else if (p == NULL) 2186 STAILQ_INSERT_HEAD(&tls->pending_records, m, 2187 m_epg_stailq); 2188 else 2189 STAILQ_INSERT_AFTER(&tls->pending_records, p, m, 2190 m_epg_stailq); 2191 mtx_unlock(&wq->mtx); 2192 counter_u64_add(ktls_cnt_tx_pending, 1); 2193 return; 2194 } 2195 2196 tls->next_seqno += ktls_batched_records(m); 2197 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2198 2199 while (!STAILQ_EMPTY(&tls->pending_records)) { 2200 struct mbuf *n; 2201 2202 n = STAILQ_FIRST(&tls->pending_records); 2203 if (n->m_epg_seqno != tls->next_seqno) 2204 break; 2205 2206 queued++; 2207 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq); 2208 tls->next_seqno += ktls_batched_records(n); 2209 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq); 2210 } 2211 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1)); 2212 } else 2213 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2214 2215 running = wq->running; 2216 mtx_unlock(&wq->mtx); 2217 if (!running) 2218 wakeup(wq); 2219 counter_u64_add(ktls_cnt_tx_queued, queued); 2220 } 2221 2222 static __noinline void 2223 ktls_encrypt(struct mbuf *top) 2224 { 2225 struct ktls_session *tls; 2226 struct socket *so; 2227 struct mbuf *m; 2228 vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 2229 struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 2230 struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; 2231 vm_page_t pg; 2232 int error, i, len, npages, off, total_pages; 2233 bool is_anon; 2234 2235 so = top->m_epg_so; 2236 tls = top->m_epg_tls; 2237 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 2238 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 2239 #ifdef INVARIANTS 2240 top->m_epg_so = NULL; 2241 #endif 2242 total_pages = top->m_epg_enc_cnt; 2243 npages = 0; 2244 2245 /* 2246 * Encrypt the TLS records in the chain of mbufs starting with 2247 * 'top'. 'total_pages' gives us a total count of pages and is 2248 * used to know when we have finished encrypting the TLS 2249 * records originally queued with 'top'. 2250 * 2251 * NB: These mbufs are queued in the socket buffer and 2252 * 'm_next' is traversing the mbufs in the socket buffer. The 2253 * socket buffer lock is not held while traversing this chain. 2254 * Since the mbufs are all marked M_NOTREADY their 'm_next' 2255 * pointers should be stable. However, the 'm_next' of the 2256 * last mbuf encrypted is not necessarily NULL. It can point 2257 * to other mbufs appended while 'top' was on the TLS work 2258 * queue. 2259 * 2260 * Each mbuf holds an entire TLS record. 2261 */ 2262 error = 0; 2263 for (m = top; npages != total_pages; m = m->m_next) { 2264 KASSERT(m->m_epg_tls == tls, 2265 ("different TLS sessions in a single mbuf chain: %p vs %p", 2266 tls, m->m_epg_tls)); 2267 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == 2268 (M_EXTPG | M_NOTREADY), 2269 ("%p not unready & nomap mbuf (top = %p)\n", m, top)); 2270 KASSERT(npages + m->m_epg_npgs <= total_pages, 2271 ("page count mismatch: top %p, total_pages %d, m %p", top, 2272 total_pages, m)); 2273 2274 /* 2275 * Generate source and destination ivoecs to pass to 2276 * the SW encryption backend. For writable mbufs, the 2277 * destination iovec is a copy of the source and 2278 * encryption is done in place. For file-backed mbufs 2279 * (from sendfile), anonymous wired pages are 2280 * allocated and assigned to the destination iovec. 2281 */ 2282 is_anon = (m->m_epg_flags & EPG_FLAG_ANON) != 0; 2283 2284 off = m->m_epg_1st_off; 2285 for (i = 0; i < m->m_epg_npgs; i++, off = 0) { 2286 len = m_epg_pagelen(m, i, off); 2287 src_iov[i].iov_len = len; 2288 src_iov[i].iov_base = 2289 (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]) + 2290 off; 2291 2292 if (is_anon) { 2293 dst_iov[i].iov_base = src_iov[i].iov_base; 2294 dst_iov[i].iov_len = src_iov[i].iov_len; 2295 continue; 2296 } 2297 retry_page: 2298 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP | 2299 VM_ALLOC_WIRED); 2300 if (pg == NULL) { 2301 vm_wait(NULL); 2302 goto retry_page; 2303 } 2304 parray[i] = VM_PAGE_TO_PHYS(pg); 2305 dst_iov[i].iov_base = 2306 (char *)(void *)PHYS_TO_DMAP(parray[i]) + off; 2307 dst_iov[i].iov_len = len; 2308 } 2309 2310 npages += m->m_epg_nrdy; 2311 2312 error = (*tls->sw_encrypt)(tls, 2313 (const struct tls_record_layer *)m->m_epg_hdr, 2314 m->m_epg_trail, src_iov, dst_iov, i, m->m_epg_seqno, 2315 m->m_epg_record_type); 2316 if (error) { 2317 counter_u64_add(ktls_offload_failed_crypto, 1); 2318 break; 2319 } 2320 2321 /* 2322 * For file-backed mbufs, release the file-backed 2323 * pages and replace them in the ext_pgs array with 2324 * the anonymous wired pages allocated above. 2325 */ 2326 if (!is_anon) { 2327 /* Free the old pages. */ 2328 m->m_ext.ext_free(m); 2329 2330 /* Replace them with the new pages. */ 2331 for (i = 0; i < m->m_epg_npgs; i++) 2332 m->m_epg_pa[i] = parray[i]; 2333 2334 /* Use the basic free routine. */ 2335 m->m_ext.ext_free = mb_free_mext_pgs; 2336 2337 /* Pages are now writable. */ 2338 m->m_epg_flags |= EPG_FLAG_ANON; 2339 } 2340 2341 /* 2342 * Drop a reference to the session now that it is no 2343 * longer needed. Existing code depends on encrypted 2344 * records having no associated session vs 2345 * yet-to-be-encrypted records having an associated 2346 * session. 2347 */ 2348 m->m_epg_tls = NULL; 2349 ktls_free(tls); 2350 } 2351 2352 CURVNET_SET(so->so_vnet); 2353 if (error == 0) { 2354 (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages); 2355 } else { 2356 ktls_drop(so, EIO); 2357 mb_free_notready(top, total_pages); 2358 } 2359 2360 SOCK_LOCK(so); 2361 sorele(so); 2362 CURVNET_RESTORE(); 2363 } 2364 2365 static void 2366 ktls_work_thread(void *ctx) 2367 { 2368 struct ktls_wq *wq = ctx; 2369 struct mbuf *m, *n; 2370 struct socket *so, *son; 2371 STAILQ_HEAD(, mbuf) local_m_head; 2372 STAILQ_HEAD(, socket) local_so_head; 2373 2374 if (ktls_bind_threads > 1) { 2375 curthread->td_domain.dr_policy = 2376 DOMAINSET_PREF(PCPU_GET(domain)); 2377 } 2378 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 2379 fpu_kern_thread(0); 2380 #endif 2381 for (;;) { 2382 mtx_lock(&wq->mtx); 2383 while (STAILQ_EMPTY(&wq->m_head) && 2384 STAILQ_EMPTY(&wq->so_head)) { 2385 wq->running = false; 2386 mtx_sleep(wq, &wq->mtx, 0, "-", 0); 2387 wq->running = true; 2388 } 2389 2390 STAILQ_INIT(&local_m_head); 2391 STAILQ_CONCAT(&local_m_head, &wq->m_head); 2392 STAILQ_INIT(&local_so_head); 2393 STAILQ_CONCAT(&local_so_head, &wq->so_head); 2394 mtx_unlock(&wq->mtx); 2395 2396 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) { 2397 if (m->m_epg_flags & EPG_FLAG_2FREE) { 2398 ktls_free(m->m_epg_tls); 2399 m_free_raw(m); 2400 } else { 2401 ktls_encrypt(m); 2402 counter_u64_add(ktls_cnt_tx_queued, -1); 2403 } 2404 } 2405 2406 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) { 2407 ktls_decrypt(so); 2408 counter_u64_add(ktls_cnt_rx_queued, -1); 2409 } 2410 } 2411 }
Cache object: 70c3f012460a08b9a836d744658d8431
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]