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_kern_tls.h" 34 #include "opt_ratelimit.h" 35 #include "opt_rss.h" 36 37 #include <sys/param.h> 38 #include <sys/kernel.h> 39 #include <sys/domainset.h> 40 #include <sys/endian.h> 41 #include <sys/ktls.h> 42 #include <sys/lock.h> 43 #include <sys/mbuf.h> 44 #include <sys/mutex.h> 45 #include <sys/rmlock.h> 46 #include <sys/proc.h> 47 #include <sys/protosw.h> 48 #include <sys/refcount.h> 49 #include <sys/smp.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/sysctl.h> 53 #include <sys/taskqueue.h> 54 #include <sys/kthread.h> 55 #include <sys/uio.h> 56 #include <sys/vmmeter.h> 57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 58 #include <machine/pcb.h> 59 #endif 60 #include <machine/vmparam.h> 61 #include <net/if.h> 62 #include <net/if_var.h> 63 #ifdef RSS 64 #include <net/netisr.h> 65 #include <net/rss_config.h> 66 #endif 67 #include <net/route.h> 68 #include <net/route/nhop.h> 69 #if defined(INET) || defined(INET6) 70 #include <netinet/in.h> 71 #include <netinet/in_pcb.h> 72 #endif 73 #include <netinet/tcp_var.h> 74 #ifdef TCP_OFFLOAD 75 #include <netinet/tcp_offload.h> 76 #endif 77 #include <opencrypto/cryptodev.h> 78 #include <opencrypto/ktls.h> 79 #include <vm/uma_dbg.h> 80 #include <vm/vm.h> 81 #include <vm/vm_pageout.h> 82 #include <vm/vm_page.h> 83 #include <vm/vm_pagequeue.h> 84 85 struct ktls_wq { 86 struct mtx mtx; 87 STAILQ_HEAD(, mbuf) m_head; 88 STAILQ_HEAD(, socket) so_head; 89 bool running; 90 int lastallocfail; 91 } __aligned(CACHE_LINE_SIZE); 92 93 struct ktls_alloc_thread { 94 uint64_t wakeups; 95 uint64_t allocs; 96 struct thread *td; 97 int running; 98 }; 99 100 struct ktls_domain_info { 101 int count; 102 int cpu[MAXCPU]; 103 struct ktls_alloc_thread alloc_td; 104 }; 105 106 struct ktls_domain_info ktls_domains[MAXMEMDOM]; 107 static struct ktls_wq *ktls_wq; 108 static struct proc *ktls_proc; 109 static uma_zone_t ktls_session_zone; 110 static uma_zone_t ktls_buffer_zone; 111 static uint16_t ktls_cpuid_lookup[MAXCPU]; 112 static int ktls_init_state; 113 static struct sx ktls_init_lock; 114 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init"); 115 116 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 117 "Kernel TLS offload"); 118 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 119 "Kernel TLS offload stats"); 120 121 #ifdef RSS 122 static int ktls_bind_threads = 1; 123 #else 124 static int ktls_bind_threads; 125 #endif 126 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 127 &ktls_bind_threads, 0, 128 "Bind crypto threads to cores (1) or cores and domains (2) at boot"); 129 130 static u_int ktls_maxlen = 16384; 131 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN, 132 &ktls_maxlen, 0, "Maximum TLS record size"); 133 134 static int ktls_number_threads; 135 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 136 &ktls_number_threads, 0, 137 "Number of TLS threads in thread-pool"); 138 139 unsigned int ktls_ifnet_max_rexmit_pct = 2; 140 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN, 141 &ktls_ifnet_max_rexmit_pct, 2, 142 "Max percent bytes retransmitted before ifnet TLS is disabled"); 143 144 static bool ktls_offload_enable; 145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, 146 &ktls_offload_enable, 0, 147 "Enable support for kernel TLS offload"); 148 149 static bool ktls_cbc_enable = true; 150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN, 151 &ktls_cbc_enable, 1, 152 "Enable Support of AES-CBC crypto for kernel TLS"); 153 154 static bool ktls_sw_buffer_cache = true; 155 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN, 156 &ktls_sw_buffer_cache, 1, 157 "Enable caching of output buffers for SW encryption"); 158 159 static int ktls_max_alloc = 128; 160 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN, 161 &ktls_max_alloc, 128, 162 "Max number of 16k buffers to allocate in thread context"); 163 164 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active); 165 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 166 &ktls_tasks_active, "Number of active tasks"); 167 168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending); 169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD, 170 &ktls_cnt_tx_pending, 171 "Number of TLS 1.0 records waiting for earlier TLS records"); 172 173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued); 174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, 175 &ktls_cnt_tx_queued, 176 "Number of TLS records in queue to tasks for SW encryption"); 177 178 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued); 179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, 180 &ktls_cnt_rx_queued, 181 "Number of TLS sockets in queue to tasks for SW decryption"); 182 183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total); 184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 185 CTLFLAG_RD, &ktls_offload_total, 186 "Total successful TLS setups (parameters set)"); 187 188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls); 189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 190 CTLFLAG_RD, &ktls_offload_enable_calls, 191 "Total number of TLS enable calls made"); 192 193 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active); 194 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 195 &ktls_offload_active, "Total Active TLS sessions"); 196 197 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records); 198 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, 199 &ktls_offload_corrupted_records, "Total corrupted TLS records received"); 200 201 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto); 202 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 203 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 204 205 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet); 206 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 207 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 208 209 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw); 210 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 211 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 212 213 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed); 214 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 215 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 216 217 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail); 218 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD, 219 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet"); 220 221 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok); 222 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD, 223 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet"); 224 225 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 226 "Software TLS session stats"); 227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 228 "Hardware (ifnet) TLS session stats"); 229 #ifdef TCP_OFFLOAD 230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 231 "TOE TLS session stats"); 232 #endif 233 234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc); 235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, 236 "Active number of software TLS sessions using AES-CBC"); 237 238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm); 239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, 240 "Active number of software TLS sessions using AES-GCM"); 241 242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20); 243 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD, 244 &ktls_sw_chacha20, 245 "Active number of software TLS sessions using Chacha20-Poly1305"); 246 247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc); 248 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, 249 &ktls_ifnet_cbc, 250 "Active number of ifnet TLS sessions using AES-CBC"); 251 252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm); 253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, 254 &ktls_ifnet_gcm, 255 "Active number of ifnet TLS sessions using AES-GCM"); 256 257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20); 258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD, 259 &ktls_ifnet_chacha20, 260 "Active number of ifnet TLS sessions using Chacha20-Poly1305"); 261 262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset); 263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, 264 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); 265 266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped); 267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, 268 &ktls_ifnet_reset_dropped, 269 "TLS sessions dropped after failing to update ifnet send tag"); 270 271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed); 272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, 273 &ktls_ifnet_reset_failed, 274 "TLS sessions that failed to allocate a new ifnet send tag"); 275 276 static int ktls_ifnet_permitted; 277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, 278 &ktls_ifnet_permitted, 1, 279 "Whether to permit hardware (ifnet) TLS sessions"); 280 281 #ifdef TCP_OFFLOAD 282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc); 283 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, 284 &ktls_toe_cbc, 285 "Active number of TOE TLS sessions using AES-CBC"); 286 287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm); 288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, 289 &ktls_toe_gcm, 290 "Active number of TOE TLS sessions using AES-GCM"); 291 292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20); 293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD, 294 &ktls_toe_chacha20, 295 "Active number of TOE TLS sessions using Chacha20-Poly1305"); 296 #endif 297 298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); 299 300 static void ktls_cleanup(struct ktls_session *tls); 301 #if defined(INET) || defined(INET6) 302 static void ktls_reset_receive_tag(void *context, int pending); 303 static void ktls_reset_send_tag(void *context, int pending); 304 #endif 305 static void ktls_work_thread(void *ctx); 306 static void ktls_alloc_thread(void *ctx); 307 308 #if defined(INET) || defined(INET6) 309 static u_int 310 ktls_get_cpu(struct socket *so) 311 { 312 struct inpcb *inp; 313 #ifdef NUMA 314 struct ktls_domain_info *di; 315 #endif 316 u_int cpuid; 317 318 inp = sotoinpcb(so); 319 #ifdef RSS 320 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 321 if (cpuid != NETISR_CPUID_NONE) 322 return (cpuid); 323 #endif 324 /* 325 * Just use the flowid to shard connections in a repeatable 326 * fashion. Note that TLS 1.0 sessions rely on the 327 * serialization provided by having the same connection use 328 * the same queue. 329 */ 330 #ifdef NUMA 331 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) { 332 di = &ktls_domains[inp->inp_numa_domain]; 333 cpuid = di->cpu[inp->inp_flowid % di->count]; 334 } else 335 #endif 336 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 337 return (cpuid); 338 } 339 #endif 340 341 static int 342 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags) 343 { 344 vm_page_t m; 345 int i, req; 346 347 KASSERT((ktls_maxlen & PAGE_MASK) == 0, 348 ("%s: ktls max length %d is not page size-aligned", 349 __func__, ktls_maxlen)); 350 351 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags); 352 for (i = 0; i < count; i++) { 353 m = vm_page_alloc_noobj_contig_domain(domain, req, 354 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, 355 VM_MEMATTR_DEFAULT); 356 if (m == NULL) 357 break; 358 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 359 } 360 return (i); 361 } 362 363 static void 364 ktls_buffer_release(void *arg __unused, void **store, int count) 365 { 366 vm_page_t m; 367 int i, j; 368 369 for (i = 0; i < count; i++) { 370 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i])); 371 for (j = 0; j < atop(ktls_maxlen); j++) { 372 (void)vm_page_unwire_noq(m + j); 373 vm_page_free(m + j); 374 } 375 } 376 } 377 378 static void 379 ktls_free_mext_contig(struct mbuf *m) 380 { 381 M_ASSERTEXTPG(m); 382 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0])); 383 } 384 385 static int 386 ktls_init(void) 387 { 388 struct thread *td; 389 struct pcpu *pc; 390 int count, domain, error, i; 391 392 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 393 M_WAITOK | M_ZERO); 394 395 ktls_session_zone = uma_zcreate("ktls_session", 396 sizeof(struct ktls_session), 397 NULL, NULL, NULL, NULL, 398 UMA_ALIGN_CACHE, 0); 399 400 if (ktls_sw_buffer_cache) { 401 ktls_buffer_zone = uma_zcache_create("ktls_buffers", 402 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL, 403 ktls_buffer_import, ktls_buffer_release, NULL, 404 UMA_ZONE_FIRSTTOUCH); 405 } 406 407 /* 408 * Initialize the workqueues to run the TLS work. We create a 409 * work queue for each CPU. 410 */ 411 CPU_FOREACH(i) { 412 STAILQ_INIT(&ktls_wq[i].m_head); 413 STAILQ_INIT(&ktls_wq[i].so_head); 414 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 415 if (ktls_bind_threads > 1) { 416 pc = pcpu_find(i); 417 domain = pc->pc_domain; 418 count = ktls_domains[domain].count; 419 ktls_domains[domain].cpu[count] = i; 420 ktls_domains[domain].count++; 421 } 422 ktls_cpuid_lookup[ktls_number_threads] = i; 423 ktls_number_threads++; 424 } 425 426 /* 427 * If we somehow have an empty domain, fall back to choosing 428 * among all KTLS threads. 429 */ 430 if (ktls_bind_threads > 1) { 431 for (i = 0; i < vm_ndomains; i++) { 432 if (ktls_domains[i].count == 0) { 433 ktls_bind_threads = 1; 434 break; 435 } 436 } 437 } 438 439 /* Start kthreads for each workqueue. */ 440 CPU_FOREACH(i) { 441 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 442 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 443 if (error) { 444 printf("Can't add KTLS thread %d error %d\n", i, error); 445 return (error); 446 } 447 } 448 449 /* 450 * Start an allocation thread per-domain to perform blocking allocations 451 * of 16k physically contiguous TLS crypto destination buffers. 452 */ 453 if (ktls_sw_buffer_cache) { 454 for (domain = 0; domain < vm_ndomains; domain++) { 455 if (VM_DOMAIN_EMPTY(domain)) 456 continue; 457 if (CPU_EMPTY(&cpuset_domain[domain])) 458 continue; 459 error = kproc_kthread_add(ktls_alloc_thread, 460 &ktls_domains[domain], &ktls_proc, 461 &ktls_domains[domain].alloc_td.td, 462 0, 0, "KTLS", "alloc_%d", domain); 463 if (error) { 464 printf("Can't add KTLS alloc thread %d error %d\n", 465 domain, error); 466 return (error); 467 } 468 } 469 } 470 471 if (bootverbose) 472 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 473 return (0); 474 } 475 476 static int 477 ktls_start_kthreads(void) 478 { 479 int error, state; 480 481 start: 482 state = atomic_load_acq_int(&ktls_init_state); 483 if (__predict_true(state > 0)) 484 return (0); 485 if (state < 0) 486 return (ENXIO); 487 488 sx_xlock(&ktls_init_lock); 489 if (ktls_init_state != 0) { 490 sx_xunlock(&ktls_init_lock); 491 goto start; 492 } 493 494 error = ktls_init(); 495 if (error == 0) 496 state = 1; 497 else 498 state = -1; 499 atomic_store_rel_int(&ktls_init_state, state); 500 sx_xunlock(&ktls_init_lock); 501 return (error); 502 } 503 504 #if defined(INET) || defined(INET6) 505 static int 506 ktls_create_session(struct socket *so, struct tls_enable *en, 507 struct ktls_session **tlsp, int direction) 508 { 509 struct ktls_session *tls; 510 int error; 511 512 /* Only TLS 1.0 - 1.3 are supported. */ 513 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 514 return (EINVAL); 515 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 516 en->tls_vminor > TLS_MINOR_VER_THREE) 517 return (EINVAL); 518 519 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 520 return (EINVAL); 521 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 522 return (EINVAL); 523 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 524 return (EINVAL); 525 526 /* All supported algorithms require a cipher key. */ 527 if (en->cipher_key_len == 0) 528 return (EINVAL); 529 530 /* No flags are currently supported. */ 531 if (en->flags != 0) 532 return (EINVAL); 533 534 /* Common checks for supported algorithms. */ 535 switch (en->cipher_algorithm) { 536 case CRYPTO_AES_NIST_GCM_16: 537 /* 538 * auth_algorithm isn't used, but permit GMAC values 539 * for compatibility. 540 */ 541 switch (en->auth_algorithm) { 542 case 0: 543 #ifdef COMPAT_FREEBSD12 544 /* XXX: Really 13.0-current COMPAT. */ 545 case CRYPTO_AES_128_NIST_GMAC: 546 case CRYPTO_AES_192_NIST_GMAC: 547 case CRYPTO_AES_256_NIST_GMAC: 548 #endif 549 break; 550 default: 551 return (EINVAL); 552 } 553 if (en->auth_key_len != 0) 554 return (EINVAL); 555 switch (en->tls_vminor) { 556 case TLS_MINOR_VER_TWO: 557 if (en->iv_len != TLS_AEAD_GCM_LEN) 558 return (EINVAL); 559 break; 560 case TLS_MINOR_VER_THREE: 561 if (en->iv_len != TLS_1_3_GCM_IV_LEN) 562 return (EINVAL); 563 break; 564 default: 565 return (EINVAL); 566 } 567 break; 568 case CRYPTO_AES_CBC: 569 switch (en->auth_algorithm) { 570 case CRYPTO_SHA1_HMAC: 571 break; 572 case CRYPTO_SHA2_256_HMAC: 573 case CRYPTO_SHA2_384_HMAC: 574 if (en->tls_vminor != TLS_MINOR_VER_TWO) 575 return (EINVAL); 576 break; 577 default: 578 return (EINVAL); 579 } 580 if (en->auth_key_len == 0) 581 return (EINVAL); 582 583 /* 584 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2 585 * use explicit IVs. 586 */ 587 switch (en->tls_vminor) { 588 case TLS_MINOR_VER_ZERO: 589 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 590 return (EINVAL); 591 break; 592 case TLS_MINOR_VER_ONE: 593 case TLS_MINOR_VER_TWO: 594 /* Ignore any supplied IV. */ 595 en->iv_len = 0; 596 break; 597 default: 598 return (EINVAL); 599 } 600 break; 601 case CRYPTO_CHACHA20_POLY1305: 602 if (en->auth_algorithm != 0 || en->auth_key_len != 0) 603 return (EINVAL); 604 if (en->tls_vminor != TLS_MINOR_VER_TWO && 605 en->tls_vminor != TLS_MINOR_VER_THREE) 606 return (EINVAL); 607 if (en->iv_len != TLS_CHACHA20_IV_LEN) 608 return (EINVAL); 609 break; 610 default: 611 return (EINVAL); 612 } 613 614 error = ktls_start_kthreads(); 615 if (error != 0) 616 return (error); 617 618 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 619 620 counter_u64_add(ktls_offload_active, 1); 621 622 refcount_init(&tls->refcount, 1); 623 if (direction == KTLS_RX) 624 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls); 625 else 626 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 627 628 tls->wq_index = ktls_get_cpu(so); 629 630 tls->params.cipher_algorithm = en->cipher_algorithm; 631 tls->params.auth_algorithm = en->auth_algorithm; 632 tls->params.tls_vmajor = en->tls_vmajor; 633 tls->params.tls_vminor = en->tls_vminor; 634 tls->params.flags = en->flags; 635 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 636 637 /* Set the header and trailer lengths. */ 638 tls->params.tls_hlen = sizeof(struct tls_record_layer); 639 switch (en->cipher_algorithm) { 640 case CRYPTO_AES_NIST_GCM_16: 641 /* 642 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 643 * nonce. TLS 1.3 uses a 12 byte implicit IV. 644 */ 645 if (en->tls_vminor < TLS_MINOR_VER_THREE) 646 tls->params.tls_hlen += sizeof(uint64_t); 647 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 648 tls->params.tls_bs = 1; 649 break; 650 case CRYPTO_AES_CBC: 651 switch (en->auth_algorithm) { 652 case CRYPTO_SHA1_HMAC: 653 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 654 /* Implicit IV, no nonce. */ 655 tls->sequential_records = true; 656 tls->next_seqno = be64dec(en->rec_seq); 657 STAILQ_INIT(&tls->pending_records); 658 } else { 659 tls->params.tls_hlen += AES_BLOCK_LEN; 660 } 661 tls->params.tls_tlen = AES_BLOCK_LEN + 662 SHA1_HASH_LEN; 663 break; 664 case CRYPTO_SHA2_256_HMAC: 665 tls->params.tls_hlen += AES_BLOCK_LEN; 666 tls->params.tls_tlen = AES_BLOCK_LEN + 667 SHA2_256_HASH_LEN; 668 break; 669 case CRYPTO_SHA2_384_HMAC: 670 tls->params.tls_hlen += AES_BLOCK_LEN; 671 tls->params.tls_tlen = AES_BLOCK_LEN + 672 SHA2_384_HASH_LEN; 673 break; 674 default: 675 panic("invalid hmac"); 676 } 677 tls->params.tls_bs = AES_BLOCK_LEN; 678 break; 679 case CRYPTO_CHACHA20_POLY1305: 680 /* 681 * Chacha20 uses a 12 byte implicit IV. 682 */ 683 tls->params.tls_tlen = POLY1305_HASH_LEN; 684 tls->params.tls_bs = 1; 685 break; 686 default: 687 panic("invalid cipher"); 688 } 689 690 /* 691 * TLS 1.3 includes optional padding which we do not support, 692 * and also puts the "real" record type at the end of the 693 * encrypted data. 694 */ 695 if (en->tls_vminor == TLS_MINOR_VER_THREE) 696 tls->params.tls_tlen += sizeof(uint8_t); 697 698 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 699 ("TLS header length too long: %d", tls->params.tls_hlen)); 700 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 701 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 702 703 if (en->auth_key_len != 0) { 704 tls->params.auth_key_len = en->auth_key_len; 705 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 706 M_WAITOK); 707 error = copyin(en->auth_key, tls->params.auth_key, 708 en->auth_key_len); 709 if (error) 710 goto out; 711 } 712 713 tls->params.cipher_key_len = en->cipher_key_len; 714 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 715 error = copyin(en->cipher_key, tls->params.cipher_key, 716 en->cipher_key_len); 717 if (error) 718 goto out; 719 720 /* 721 * This holds the implicit portion of the nonce for AEAD 722 * ciphers and the initial implicit IV for TLS 1.0. The 723 * explicit portions of the IV are generated in ktls_frame(). 724 */ 725 if (en->iv_len != 0) { 726 tls->params.iv_len = en->iv_len; 727 error = copyin(en->iv, tls->params.iv, en->iv_len); 728 if (error) 729 goto out; 730 731 /* 732 * For TLS 1.2 with GCM, generate an 8-byte nonce as a 733 * counter to generate unique explicit IVs. 734 * 735 * Store this counter in the last 8 bytes of the IV 736 * array so that it is 8-byte aligned. 737 */ 738 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 739 en->tls_vminor == TLS_MINOR_VER_TWO) 740 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); 741 } 742 743 *tlsp = tls; 744 return (0); 745 746 out: 747 ktls_cleanup(tls); 748 return (error); 749 } 750 751 static struct ktls_session * 752 ktls_clone_session(struct ktls_session *tls, int direction) 753 { 754 struct ktls_session *tls_new; 755 756 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 757 758 counter_u64_add(ktls_offload_active, 1); 759 760 refcount_init(&tls_new->refcount, 1); 761 if (direction == KTLS_RX) 762 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag, 763 tls_new); 764 else 765 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, 766 tls_new); 767 768 /* Copy fields from existing session. */ 769 tls_new->params = tls->params; 770 tls_new->wq_index = tls->wq_index; 771 772 /* Deep copy keys. */ 773 if (tls_new->params.auth_key != NULL) { 774 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 775 M_KTLS, M_WAITOK); 776 memcpy(tls_new->params.auth_key, tls->params.auth_key, 777 tls->params.auth_key_len); 778 } 779 780 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 781 M_WAITOK); 782 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 783 tls->params.cipher_key_len); 784 785 return (tls_new); 786 } 787 #endif 788 789 static void 790 ktls_cleanup(struct ktls_session *tls) 791 { 792 793 counter_u64_add(ktls_offload_active, -1); 794 switch (tls->mode) { 795 case TCP_TLS_MODE_SW: 796 switch (tls->params.cipher_algorithm) { 797 case CRYPTO_AES_CBC: 798 counter_u64_add(ktls_sw_cbc, -1); 799 break; 800 case CRYPTO_AES_NIST_GCM_16: 801 counter_u64_add(ktls_sw_gcm, -1); 802 break; 803 case CRYPTO_CHACHA20_POLY1305: 804 counter_u64_add(ktls_sw_chacha20, -1); 805 break; 806 } 807 break; 808 case TCP_TLS_MODE_IFNET: 809 switch (tls->params.cipher_algorithm) { 810 case CRYPTO_AES_CBC: 811 counter_u64_add(ktls_ifnet_cbc, -1); 812 break; 813 case CRYPTO_AES_NIST_GCM_16: 814 counter_u64_add(ktls_ifnet_gcm, -1); 815 break; 816 case CRYPTO_CHACHA20_POLY1305: 817 counter_u64_add(ktls_ifnet_chacha20, -1); 818 break; 819 } 820 if (tls->snd_tag != NULL) 821 m_snd_tag_rele(tls->snd_tag); 822 if (tls->rx_ifp != NULL) 823 if_rele(tls->rx_ifp); 824 break; 825 #ifdef TCP_OFFLOAD 826 case TCP_TLS_MODE_TOE: 827 switch (tls->params.cipher_algorithm) { 828 case CRYPTO_AES_CBC: 829 counter_u64_add(ktls_toe_cbc, -1); 830 break; 831 case CRYPTO_AES_NIST_GCM_16: 832 counter_u64_add(ktls_toe_gcm, -1); 833 break; 834 case CRYPTO_CHACHA20_POLY1305: 835 counter_u64_add(ktls_toe_chacha20, -1); 836 break; 837 } 838 break; 839 #endif 840 } 841 if (tls->ocf_session != NULL) 842 ktls_ocf_free(tls); 843 if (tls->params.auth_key != NULL) { 844 zfree(tls->params.auth_key, M_KTLS); 845 tls->params.auth_key = NULL; 846 tls->params.auth_key_len = 0; 847 } 848 if (tls->params.cipher_key != NULL) { 849 zfree(tls->params.cipher_key, M_KTLS); 850 tls->params.cipher_key = NULL; 851 tls->params.cipher_key_len = 0; 852 } 853 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 854 } 855 856 #if defined(INET) || defined(INET6) 857 858 #ifdef TCP_OFFLOAD 859 static int 860 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) 861 { 862 struct inpcb *inp; 863 struct tcpcb *tp; 864 int error; 865 866 inp = so->so_pcb; 867 INP_WLOCK(inp); 868 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 869 INP_WUNLOCK(inp); 870 return (ECONNRESET); 871 } 872 if (inp->inp_socket == NULL) { 873 INP_WUNLOCK(inp); 874 return (ECONNRESET); 875 } 876 tp = intotcpcb(inp); 877 if (!(tp->t_flags & TF_TOE)) { 878 INP_WUNLOCK(inp); 879 return (EOPNOTSUPP); 880 } 881 882 error = tcp_offload_alloc_tls_session(tp, tls, direction); 883 INP_WUNLOCK(inp); 884 if (error == 0) { 885 tls->mode = TCP_TLS_MODE_TOE; 886 switch (tls->params.cipher_algorithm) { 887 case CRYPTO_AES_CBC: 888 counter_u64_add(ktls_toe_cbc, 1); 889 break; 890 case CRYPTO_AES_NIST_GCM_16: 891 counter_u64_add(ktls_toe_gcm, 1); 892 break; 893 case CRYPTO_CHACHA20_POLY1305: 894 counter_u64_add(ktls_toe_chacha20, 1); 895 break; 896 } 897 } 898 return (error); 899 } 900 #endif 901 902 /* 903 * Common code used when first enabling ifnet TLS on a connection or 904 * when allocating a new ifnet TLS session due to a routing change. 905 * This function allocates a new TLS send tag on whatever interface 906 * the connection is currently routed over. 907 */ 908 static int 909 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 910 struct m_snd_tag **mstp) 911 { 912 union if_snd_tag_alloc_params params; 913 struct ifnet *ifp; 914 struct nhop_object *nh; 915 struct tcpcb *tp; 916 int error; 917 918 INP_RLOCK(inp); 919 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 920 INP_RUNLOCK(inp); 921 return (ECONNRESET); 922 } 923 if (inp->inp_socket == NULL) { 924 INP_RUNLOCK(inp); 925 return (ECONNRESET); 926 } 927 tp = intotcpcb(inp); 928 929 /* 930 * Check administrative controls on ifnet TLS to determine if 931 * ifnet TLS should be denied. 932 * 933 * - Always permit 'force' requests. 934 * - ktls_ifnet_permitted == 0: always deny. 935 */ 936 if (!force && ktls_ifnet_permitted == 0) { 937 INP_RUNLOCK(inp); 938 return (ENXIO); 939 } 940 941 /* 942 * XXX: Use the cached route in the inpcb to find the 943 * interface. This should perhaps instead use 944 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 945 * enabled after a connection has completed key negotiation in 946 * userland, the cached route will be present in practice. 947 */ 948 nh = inp->inp_route.ro_nh; 949 if (nh == NULL) { 950 INP_RUNLOCK(inp); 951 return (ENXIO); 952 } 953 ifp = nh->nh_ifp; 954 if_ref(ifp); 955 956 /* 957 * Allocate a TLS + ratelimit tag if the connection has an 958 * existing pacing rate. 959 */ 960 if (tp->t_pacing_rate != -1 && 961 (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) { 962 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT; 963 params.tls_rate_limit.inp = inp; 964 params.tls_rate_limit.tls = tls; 965 params.tls_rate_limit.max_rate = tp->t_pacing_rate; 966 } else { 967 params.hdr.type = IF_SND_TAG_TYPE_TLS; 968 params.tls.inp = inp; 969 params.tls.tls = tls; 970 } 971 params.hdr.flowid = inp->inp_flowid; 972 params.hdr.flowtype = inp->inp_flowtype; 973 params.hdr.numa_domain = inp->inp_numa_domain; 974 INP_RUNLOCK(inp); 975 976 if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) { 977 error = EOPNOTSUPP; 978 goto out; 979 } 980 if (inp->inp_vflag & INP_IPV6) { 981 if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { 982 error = EOPNOTSUPP; 983 goto out; 984 } 985 } else { 986 if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { 987 error = EOPNOTSUPP; 988 goto out; 989 } 990 } 991 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 992 out: 993 if_rele(ifp); 994 return (error); 995 } 996 997 /* 998 * Allocate an initial TLS receive tag for doing HW decryption of TLS 999 * data. 1000 * 1001 * This function allocates a new TLS receive tag on whatever interface 1002 * the connection is currently routed over. If the connection ends up 1003 * using a different interface for receive this will get fixed up via 1004 * ktls_input_ifp_mismatch as future packets arrive. 1005 */ 1006 static int 1007 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls, 1008 struct m_snd_tag **mstp) 1009 { 1010 union if_snd_tag_alloc_params params; 1011 struct ifnet *ifp; 1012 struct nhop_object *nh; 1013 int error; 1014 1015 if (!ktls_ocf_recrypt_supported(tls)) 1016 return (ENXIO); 1017 1018 INP_RLOCK(inp); 1019 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 1020 INP_RUNLOCK(inp); 1021 return (ECONNRESET); 1022 } 1023 if (inp->inp_socket == NULL) { 1024 INP_RUNLOCK(inp); 1025 return (ECONNRESET); 1026 } 1027 1028 /* 1029 * Check administrative controls on ifnet TLS to determine if 1030 * ifnet TLS should be denied. 1031 */ 1032 if (ktls_ifnet_permitted == 0) { 1033 INP_RUNLOCK(inp); 1034 return (ENXIO); 1035 } 1036 1037 /* 1038 * XXX: As with ktls_alloc_snd_tag, use the cached route in 1039 * the inpcb to find the interface. 1040 */ 1041 nh = inp->inp_route.ro_nh; 1042 if (nh == NULL) { 1043 INP_RUNLOCK(inp); 1044 return (ENXIO); 1045 } 1046 ifp = nh->nh_ifp; 1047 if_ref(ifp); 1048 tls->rx_ifp = ifp; 1049 1050 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; 1051 params.hdr.flowid = inp->inp_flowid; 1052 params.hdr.flowtype = inp->inp_flowtype; 1053 params.hdr.numa_domain = inp->inp_numa_domain; 1054 params.tls_rx.inp = inp; 1055 params.tls_rx.tls = tls; 1056 params.tls_rx.vlan_id = 0; 1057 1058 INP_RUNLOCK(inp); 1059 1060 if (inp->inp_vflag & INP_IPV6) { 1061 if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) { 1062 error = EOPNOTSUPP; 1063 goto out; 1064 } 1065 } else { 1066 if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) { 1067 error = EOPNOTSUPP; 1068 goto out; 1069 } 1070 } 1071 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 1072 1073 /* 1074 * If this connection is over a vlan, vlan_snd_tag_alloc 1075 * rewrites vlan_id with the saved interface. Save the VLAN 1076 * ID for use in ktls_reset_receive_tag which allocates new 1077 * receive tags directly from the leaf interface bypassing 1078 * if_vlan. 1079 */ 1080 if (error == 0) 1081 tls->rx_vlan_id = params.tls_rx.vlan_id; 1082 out: 1083 return (error); 1084 } 1085 1086 static int 1087 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction, 1088 bool force) 1089 { 1090 struct m_snd_tag *mst; 1091 int error; 1092 1093 switch (direction) { 1094 case KTLS_TX: 1095 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 1096 if (__predict_false(error != 0)) 1097 goto done; 1098 break; 1099 case KTLS_RX: 1100 KASSERT(!force, ("%s: forced receive tag", __func__)); 1101 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst); 1102 if (__predict_false(error != 0)) 1103 goto done; 1104 break; 1105 default: 1106 __assert_unreachable(); 1107 } 1108 1109 tls->mode = TCP_TLS_MODE_IFNET; 1110 tls->snd_tag = mst; 1111 1112 switch (tls->params.cipher_algorithm) { 1113 case CRYPTO_AES_CBC: 1114 counter_u64_add(ktls_ifnet_cbc, 1); 1115 break; 1116 case CRYPTO_AES_NIST_GCM_16: 1117 counter_u64_add(ktls_ifnet_gcm, 1); 1118 break; 1119 case CRYPTO_CHACHA20_POLY1305: 1120 counter_u64_add(ktls_ifnet_chacha20, 1); 1121 break; 1122 default: 1123 break; 1124 } 1125 done: 1126 return (error); 1127 } 1128 1129 static void 1130 ktls_use_sw(struct ktls_session *tls) 1131 { 1132 tls->mode = TCP_TLS_MODE_SW; 1133 switch (tls->params.cipher_algorithm) { 1134 case CRYPTO_AES_CBC: 1135 counter_u64_add(ktls_sw_cbc, 1); 1136 break; 1137 case CRYPTO_AES_NIST_GCM_16: 1138 counter_u64_add(ktls_sw_gcm, 1); 1139 break; 1140 case CRYPTO_CHACHA20_POLY1305: 1141 counter_u64_add(ktls_sw_chacha20, 1); 1142 break; 1143 } 1144 } 1145 1146 static int 1147 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction) 1148 { 1149 int error; 1150 1151 error = ktls_ocf_try(so, tls, direction); 1152 if (error) 1153 return (error); 1154 ktls_use_sw(tls); 1155 return (0); 1156 } 1157 1158 /* 1159 * KTLS RX stores data in the socket buffer as a list of TLS records, 1160 * where each record is stored as a control message containg the TLS 1161 * header followed by data mbufs containing the decrypted data. This 1162 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for 1163 * both encrypted and decrypted data. TLS records decrypted by a NIC 1164 * should be queued to the socket buffer as records, but encrypted 1165 * data which needs to be decrypted by software arrives as a stream of 1166 * regular mbufs which need to be converted. In addition, there may 1167 * already be pending encrypted data in the socket buffer when KTLS RX 1168 * is enabled. 1169 * 1170 * To manage not-yet-decrypted data for KTLS RX, the following scheme 1171 * is used: 1172 * 1173 * - A single chain of NOTREADY mbufs is hung off of sb_mtls. 1174 * 1175 * - ktls_check_rx checks this chain of mbufs reading the TLS header 1176 * from the first mbuf. Once all of the data for that TLS record is 1177 * queued, the socket is queued to a worker thread. 1178 * 1179 * - The worker thread calls ktls_decrypt to decrypt TLS records in 1180 * the TLS chain. Each TLS record is detached from the TLS chain, 1181 * decrypted, and inserted into the regular socket buffer chain as 1182 * record starting with a control message holding the TLS header and 1183 * a chain of mbufs holding the encrypted data. 1184 */ 1185 1186 static void 1187 sb_mark_notready(struct sockbuf *sb) 1188 { 1189 struct mbuf *m; 1190 1191 m = sb->sb_mb; 1192 sb->sb_mtls = m; 1193 sb->sb_mb = NULL; 1194 sb->sb_mbtail = NULL; 1195 sb->sb_lastrecord = NULL; 1196 for (; m != NULL; m = m->m_next) { 1197 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", 1198 __func__)); 1199 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", 1200 __func__)); 1201 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", 1202 __func__)); 1203 m->m_flags |= M_NOTREADY; 1204 sb->sb_acc -= m->m_len; 1205 sb->sb_tlscc += m->m_len; 1206 sb->sb_mtlstail = m; 1207 } 1208 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, 1209 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, 1210 sb->sb_ccc)); 1211 } 1212 1213 /* 1214 * Return information about the pending TLS data in a socket 1215 * buffer. On return, 'seqno' is set to the sequence number 1216 * of the next TLS record to be received, 'resid' is set to 1217 * the amount of bytes still needed for the last pending 1218 * record. The function returns 'false' if the last pending 1219 * record contains a partial TLS header. In that case, 'resid' 1220 * is the number of bytes needed to complete the TLS header. 1221 */ 1222 bool 1223 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp) 1224 { 1225 struct tls_record_layer hdr; 1226 struct mbuf *m; 1227 uint64_t seqno; 1228 size_t resid; 1229 u_int offset, record_len; 1230 1231 SOCKBUF_LOCK_ASSERT(sb); 1232 MPASS(sb->sb_flags & SB_TLS_RX); 1233 seqno = sb->sb_tls_seqno; 1234 resid = sb->sb_tlscc; 1235 m = sb->sb_mtls; 1236 offset = 0; 1237 1238 if (resid == 0) { 1239 *seqnop = seqno; 1240 *residp = 0; 1241 return (true); 1242 } 1243 1244 for (;;) { 1245 seqno++; 1246 1247 if (resid < sizeof(hdr)) { 1248 *seqnop = seqno; 1249 *residp = sizeof(hdr) - resid; 1250 return (false); 1251 } 1252 1253 m_copydata(m, offset, sizeof(hdr), (void *)&hdr); 1254 1255 record_len = sizeof(hdr) + ntohs(hdr.tls_length); 1256 if (resid <= record_len) { 1257 *seqnop = seqno; 1258 *residp = record_len - resid; 1259 return (true); 1260 } 1261 resid -= record_len; 1262 1263 while (record_len != 0) { 1264 if (m->m_len - offset > record_len) { 1265 offset += record_len; 1266 break; 1267 } 1268 1269 record_len -= (m->m_len - offset); 1270 offset = 0; 1271 m = m->m_next; 1272 } 1273 } 1274 } 1275 1276 int 1277 ktls_enable_rx(struct socket *so, struct tls_enable *en) 1278 { 1279 struct ktls_session *tls; 1280 int error; 1281 1282 if (!ktls_offload_enable) 1283 return (ENOTSUP); 1284 if (SOLISTENING(so)) 1285 return (EINVAL); 1286 1287 counter_u64_add(ktls_offload_enable_calls, 1); 1288 1289 /* 1290 * This should always be true since only the TCP socket option 1291 * invokes this function. 1292 */ 1293 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1294 return (EINVAL); 1295 1296 /* 1297 * XXX: Don't overwrite existing sessions. We should permit 1298 * this to support rekeying in the future. 1299 */ 1300 if (so->so_rcv.sb_tls_info != NULL) 1301 return (EALREADY); 1302 1303 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1304 return (ENOTSUP); 1305 1306 error = ktls_create_session(so, en, &tls, KTLS_RX); 1307 if (error) 1308 return (error); 1309 1310 error = ktls_ocf_try(so, tls, KTLS_RX); 1311 if (error) { 1312 ktls_cleanup(tls); 1313 return (error); 1314 } 1315 1316 /* Mark the socket as using TLS offload. */ 1317 SOCKBUF_LOCK(&so->so_rcv); 1318 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); 1319 so->so_rcv.sb_tls_info = tls; 1320 so->so_rcv.sb_flags |= SB_TLS_RX; 1321 1322 /* Mark existing data as not ready until it can be decrypted. */ 1323 sb_mark_notready(&so->so_rcv); 1324 ktls_check_rx(&so->so_rcv); 1325 SOCKBUF_UNLOCK(&so->so_rcv); 1326 1327 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1328 #ifdef TCP_OFFLOAD 1329 error = ktls_try_toe(so, tls, KTLS_RX); 1330 if (error) 1331 #endif 1332 error = ktls_try_ifnet(so, tls, KTLS_RX, false); 1333 if (error) 1334 ktls_use_sw(tls); 1335 1336 counter_u64_add(ktls_offload_total, 1); 1337 1338 return (0); 1339 } 1340 1341 int 1342 ktls_enable_tx(struct socket *so, struct tls_enable *en) 1343 { 1344 struct ktls_session *tls; 1345 struct inpcb *inp; 1346 int error; 1347 1348 if (!ktls_offload_enable) 1349 return (ENOTSUP); 1350 if (SOLISTENING(so)) 1351 return (EINVAL); 1352 1353 counter_u64_add(ktls_offload_enable_calls, 1); 1354 1355 /* 1356 * This should always be true since only the TCP socket option 1357 * invokes this function. 1358 */ 1359 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1360 return (EINVAL); 1361 1362 /* 1363 * XXX: Don't overwrite existing sessions. We should permit 1364 * this to support rekeying in the future. 1365 */ 1366 if (so->so_snd.sb_tls_info != NULL) 1367 return (EALREADY); 1368 1369 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1370 return (ENOTSUP); 1371 1372 /* TLS requires ext pgs */ 1373 if (mb_use_ext_pgs == 0) 1374 return (ENXIO); 1375 1376 error = ktls_create_session(so, en, &tls, KTLS_TX); 1377 if (error) 1378 return (error); 1379 1380 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1381 #ifdef TCP_OFFLOAD 1382 error = ktls_try_toe(so, tls, KTLS_TX); 1383 if (error) 1384 #endif 1385 error = ktls_try_ifnet(so, tls, KTLS_TX, false); 1386 if (error) 1387 error = ktls_try_sw(so, tls, KTLS_TX); 1388 1389 if (error) { 1390 ktls_cleanup(tls); 1391 return (error); 1392 } 1393 1394 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1395 if (error) { 1396 ktls_cleanup(tls); 1397 return (error); 1398 } 1399 1400 /* 1401 * Write lock the INP when setting sb_tls_info so that 1402 * routines in tcp_ratelimit.c can read sb_tls_info while 1403 * holding the INP lock. 1404 */ 1405 inp = so->so_pcb; 1406 INP_WLOCK(inp); 1407 SOCKBUF_LOCK(&so->so_snd); 1408 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); 1409 so->so_snd.sb_tls_info = tls; 1410 if (tls->mode != TCP_TLS_MODE_SW) 1411 so->so_snd.sb_flags |= SB_TLS_IFNET; 1412 SOCKBUF_UNLOCK(&so->so_snd); 1413 INP_WUNLOCK(inp); 1414 SOCK_IO_SEND_UNLOCK(so); 1415 1416 counter_u64_add(ktls_offload_total, 1); 1417 1418 return (0); 1419 } 1420 1421 int 1422 ktls_get_rx_mode(struct socket *so, int *modep) 1423 { 1424 struct ktls_session *tls; 1425 struct inpcb *inp __diagused; 1426 1427 if (SOLISTENING(so)) 1428 return (EINVAL); 1429 inp = so->so_pcb; 1430 INP_WLOCK_ASSERT(inp); 1431 SOCK_RECVBUF_LOCK(so); 1432 tls = so->so_rcv.sb_tls_info; 1433 if (tls == NULL) 1434 *modep = TCP_TLS_MODE_NONE; 1435 else 1436 *modep = tls->mode; 1437 SOCK_RECVBUF_UNLOCK(so); 1438 return (0); 1439 } 1440 1441 /* 1442 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number. 1443 * 1444 * This function gets information about the next TCP- and TLS- 1445 * sequence number to be processed by the TLS receive worker 1446 * thread. The information is extracted from the given "inpcb" 1447 * structure. The values are stored in host endian format at the two 1448 * given output pointer locations. The TCP sequence number points to 1449 * the beginning of the TLS header. 1450 * 1451 * This function returns zero on success, else a non-zero error code 1452 * is returned. 1453 */ 1454 int 1455 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq) 1456 { 1457 struct socket *so; 1458 struct tcpcb *tp; 1459 1460 INP_RLOCK(inp); 1461 so = inp->inp_socket; 1462 if (__predict_false(so == NULL)) { 1463 INP_RUNLOCK(inp); 1464 return (EINVAL); 1465 } 1466 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 1467 INP_RUNLOCK(inp); 1468 return (ECONNRESET); 1469 } 1470 1471 tp = intotcpcb(inp); 1472 MPASS(tp != NULL); 1473 1474 SOCKBUF_LOCK(&so->so_rcv); 1475 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc; 1476 *tlsseq = so->so_rcv.sb_tls_seqno; 1477 SOCKBUF_UNLOCK(&so->so_rcv); 1478 1479 INP_RUNLOCK(inp); 1480 1481 return (0); 1482 } 1483 1484 int 1485 ktls_get_tx_mode(struct socket *so, int *modep) 1486 { 1487 struct ktls_session *tls; 1488 struct inpcb *inp __diagused; 1489 1490 if (SOLISTENING(so)) 1491 return (EINVAL); 1492 inp = so->so_pcb; 1493 INP_WLOCK_ASSERT(inp); 1494 SOCK_SENDBUF_LOCK(so); 1495 tls = so->so_snd.sb_tls_info; 1496 if (tls == NULL) 1497 *modep = TCP_TLS_MODE_NONE; 1498 else 1499 *modep = tls->mode; 1500 SOCK_SENDBUF_UNLOCK(so); 1501 return (0); 1502 } 1503 1504 /* 1505 * Switch between SW and ifnet TLS sessions as requested. 1506 */ 1507 int 1508 ktls_set_tx_mode(struct socket *so, int mode) 1509 { 1510 struct ktls_session *tls, *tls_new; 1511 struct inpcb *inp; 1512 int error; 1513 1514 if (SOLISTENING(so)) 1515 return (EINVAL); 1516 switch (mode) { 1517 case TCP_TLS_MODE_SW: 1518 case TCP_TLS_MODE_IFNET: 1519 break; 1520 default: 1521 return (EINVAL); 1522 } 1523 1524 inp = so->so_pcb; 1525 INP_WLOCK_ASSERT(inp); 1526 SOCKBUF_LOCK(&so->so_snd); 1527 tls = so->so_snd.sb_tls_info; 1528 if (tls == NULL) { 1529 SOCKBUF_UNLOCK(&so->so_snd); 1530 return (0); 1531 } 1532 1533 if (tls->mode == mode) { 1534 SOCKBUF_UNLOCK(&so->so_snd); 1535 return (0); 1536 } 1537 1538 tls = ktls_hold(tls); 1539 SOCKBUF_UNLOCK(&so->so_snd); 1540 INP_WUNLOCK(inp); 1541 1542 tls_new = ktls_clone_session(tls, KTLS_TX); 1543 1544 if (mode == TCP_TLS_MODE_IFNET) 1545 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true); 1546 else 1547 error = ktls_try_sw(so, tls_new, KTLS_TX); 1548 if (error) { 1549 counter_u64_add(ktls_switch_failed, 1); 1550 ktls_free(tls_new); 1551 ktls_free(tls); 1552 INP_WLOCK(inp); 1553 return (error); 1554 } 1555 1556 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1557 if (error) { 1558 counter_u64_add(ktls_switch_failed, 1); 1559 ktls_free(tls_new); 1560 ktls_free(tls); 1561 INP_WLOCK(inp); 1562 return (error); 1563 } 1564 1565 /* 1566 * If we raced with another session change, keep the existing 1567 * session. 1568 */ 1569 if (tls != so->so_snd.sb_tls_info) { 1570 counter_u64_add(ktls_switch_failed, 1); 1571 SOCK_IO_SEND_UNLOCK(so); 1572 ktls_free(tls_new); 1573 ktls_free(tls); 1574 INP_WLOCK(inp); 1575 return (EBUSY); 1576 } 1577 1578 INP_WLOCK(inp); 1579 SOCKBUF_LOCK(&so->so_snd); 1580 so->so_snd.sb_tls_info = tls_new; 1581 if (tls_new->mode != TCP_TLS_MODE_SW) 1582 so->so_snd.sb_flags |= SB_TLS_IFNET; 1583 SOCKBUF_UNLOCK(&so->so_snd); 1584 SOCK_IO_SEND_UNLOCK(so); 1585 1586 /* 1587 * Drop two references on 'tls'. The first is for the 1588 * ktls_hold() above. The second drops the reference from the 1589 * socket buffer. 1590 */ 1591 KASSERT(tls->refcount >= 2, ("too few references on old session")); 1592 ktls_free(tls); 1593 ktls_free(tls); 1594 1595 if (mode == TCP_TLS_MODE_IFNET) 1596 counter_u64_add(ktls_switch_to_ifnet, 1); 1597 else 1598 counter_u64_add(ktls_switch_to_sw, 1); 1599 1600 return (0); 1601 } 1602 1603 /* 1604 * Try to allocate a new TLS receive tag. This task is scheduled when 1605 * sbappend_ktls_rx detects an input path change. If a new tag is 1606 * allocated, replace the tag in the TLS session. If a new tag cannot 1607 * be allocated, let the session fall back to software decryption. 1608 */ 1609 static void 1610 ktls_reset_receive_tag(void *context, int pending) 1611 { 1612 union if_snd_tag_alloc_params params; 1613 struct ktls_session *tls; 1614 struct m_snd_tag *mst; 1615 struct inpcb *inp; 1616 struct ifnet *ifp; 1617 struct socket *so; 1618 int error; 1619 1620 MPASS(pending == 1); 1621 1622 tls = context; 1623 so = tls->so; 1624 inp = so->so_pcb; 1625 ifp = NULL; 1626 1627 INP_RLOCK(inp); 1628 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 1629 INP_RUNLOCK(inp); 1630 goto out; 1631 } 1632 1633 SOCKBUF_LOCK(&so->so_rcv); 1634 mst = tls->snd_tag; 1635 tls->snd_tag = NULL; 1636 if (mst != NULL) 1637 m_snd_tag_rele(mst); 1638 1639 ifp = tls->rx_ifp; 1640 if_ref(ifp); 1641 SOCKBUF_UNLOCK(&so->so_rcv); 1642 1643 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; 1644 params.hdr.flowid = inp->inp_flowid; 1645 params.hdr.flowtype = inp->inp_flowtype; 1646 params.hdr.numa_domain = inp->inp_numa_domain; 1647 params.tls_rx.inp = inp; 1648 params.tls_rx.tls = tls; 1649 params.tls_rx.vlan_id = tls->rx_vlan_id; 1650 INP_RUNLOCK(inp); 1651 1652 if (inp->inp_vflag & INP_IPV6) { 1653 if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) 1654 goto out; 1655 } else { 1656 if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) 1657 goto out; 1658 } 1659 1660 error = m_snd_tag_alloc(ifp, ¶ms, &mst); 1661 if (error == 0) { 1662 SOCKBUF_LOCK(&so->so_rcv); 1663 tls->snd_tag = mst; 1664 SOCKBUF_UNLOCK(&so->so_rcv); 1665 1666 counter_u64_add(ktls_ifnet_reset, 1); 1667 } else { 1668 /* 1669 * Just fall back to software decryption if a tag 1670 * cannot be allocated leaving the connection intact. 1671 * If a future input path change switches to another 1672 * interface this connection will resume ifnet TLS. 1673 */ 1674 counter_u64_add(ktls_ifnet_reset_failed, 1); 1675 } 1676 1677 out: 1678 mtx_pool_lock(mtxpool_sleep, tls); 1679 tls->reset_pending = false; 1680 mtx_pool_unlock(mtxpool_sleep, tls); 1681 1682 if (ifp != NULL) 1683 if_rele(ifp); 1684 sorele(so); 1685 ktls_free(tls); 1686 } 1687 1688 /* 1689 * Try to allocate a new TLS send tag. This task is scheduled when 1690 * ip_output detects a route change while trying to transmit a packet 1691 * holding a TLS record. If a new tag is allocated, replace the tag 1692 * in the TLS session. Subsequent packets on the connection will use 1693 * the new tag. If a new tag cannot be allocated, drop the 1694 * connection. 1695 */ 1696 static void 1697 ktls_reset_send_tag(void *context, int pending) 1698 { 1699 struct epoch_tracker et; 1700 struct ktls_session *tls; 1701 struct m_snd_tag *old, *new; 1702 struct inpcb *inp; 1703 struct tcpcb *tp; 1704 int error; 1705 1706 MPASS(pending == 1); 1707 1708 tls = context; 1709 inp = tls->inp; 1710 1711 /* 1712 * Free the old tag first before allocating a new one. 1713 * ip[6]_output_send() will treat a NULL send tag the same as 1714 * an ifp mismatch and drop packets until a new tag is 1715 * allocated. 1716 * 1717 * Write-lock the INP when changing tls->snd_tag since 1718 * ip[6]_output_send() holds a read-lock when reading the 1719 * pointer. 1720 */ 1721 INP_WLOCK(inp); 1722 old = tls->snd_tag; 1723 tls->snd_tag = NULL; 1724 INP_WUNLOCK(inp); 1725 if (old != NULL) 1726 m_snd_tag_rele(old); 1727 1728 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1729 1730 if (error == 0) { 1731 INP_WLOCK(inp); 1732 tls->snd_tag = new; 1733 mtx_pool_lock(mtxpool_sleep, tls); 1734 tls->reset_pending = false; 1735 mtx_pool_unlock(mtxpool_sleep, tls); 1736 if (!in_pcbrele_wlocked(inp)) 1737 INP_WUNLOCK(inp); 1738 1739 counter_u64_add(ktls_ifnet_reset, 1); 1740 1741 /* 1742 * XXX: Should we kick tcp_output explicitly now that 1743 * the send tag is fixed or just rely on timers? 1744 */ 1745 } else { 1746 NET_EPOCH_ENTER(et); 1747 INP_WLOCK(inp); 1748 if (!in_pcbrele_wlocked(inp)) { 1749 if (!(inp->inp_flags & INP_TIMEWAIT) && 1750 !(inp->inp_flags & INP_DROPPED)) { 1751 tp = intotcpcb(inp); 1752 CURVNET_SET(tp->t_vnet); 1753 tp = tcp_drop(tp, ECONNABORTED); 1754 CURVNET_RESTORE(); 1755 if (tp != NULL) 1756 INP_WUNLOCK(inp); 1757 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1758 } else 1759 INP_WUNLOCK(inp); 1760 } 1761 NET_EPOCH_EXIT(et); 1762 1763 counter_u64_add(ktls_ifnet_reset_failed, 1); 1764 1765 /* 1766 * Leave reset_pending true to avoid future tasks while 1767 * the socket goes away. 1768 */ 1769 } 1770 1771 ktls_free(tls); 1772 } 1773 1774 void 1775 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp) 1776 { 1777 struct ktls_session *tls; 1778 struct socket *so; 1779 1780 SOCKBUF_LOCK_ASSERT(sb); 1781 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 1782 __func__, sb)); 1783 so = __containerof(sb, struct socket, so_rcv); 1784 1785 tls = sb->sb_tls_info; 1786 if_rele(tls->rx_ifp); 1787 if_ref(ifp); 1788 tls->rx_ifp = ifp; 1789 1790 /* 1791 * See if we should schedule a task to update the receive tag for 1792 * this session. 1793 */ 1794 mtx_pool_lock(mtxpool_sleep, tls); 1795 if (!tls->reset_pending) { 1796 (void) ktls_hold(tls); 1797 soref(so); 1798 tls->so = so; 1799 tls->reset_pending = true; 1800 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1801 } 1802 mtx_pool_unlock(mtxpool_sleep, tls); 1803 } 1804 1805 int 1806 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1807 { 1808 1809 if (inp == NULL) 1810 return (ENOBUFS); 1811 1812 INP_LOCK_ASSERT(inp); 1813 1814 /* 1815 * See if we should schedule a task to update the send tag for 1816 * this session. 1817 */ 1818 mtx_pool_lock(mtxpool_sleep, tls); 1819 if (!tls->reset_pending) { 1820 (void) ktls_hold(tls); 1821 in_pcbref(inp); 1822 tls->inp = inp; 1823 tls->reset_pending = true; 1824 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1825 } 1826 mtx_pool_unlock(mtxpool_sleep, tls); 1827 return (ENOBUFS); 1828 } 1829 1830 #ifdef RATELIMIT 1831 int 1832 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate) 1833 { 1834 union if_snd_tag_modify_params params = { 1835 .rate_limit.max_rate = max_pacing_rate, 1836 .rate_limit.flags = M_NOWAIT, 1837 }; 1838 struct m_snd_tag *mst; 1839 1840 /* Can't get to the inp, but it should be locked. */ 1841 /* INP_LOCK_ASSERT(inp); */ 1842 1843 MPASS(tls->mode == TCP_TLS_MODE_IFNET); 1844 1845 if (tls->snd_tag == NULL) { 1846 /* 1847 * Resetting send tag, ignore this change. The 1848 * pending reset may or may not see this updated rate 1849 * in the tcpcb. If it doesn't, we will just lose 1850 * this rate change. 1851 */ 1852 return (0); 1853 } 1854 1855 mst = tls->snd_tag; 1856 1857 MPASS(mst != NULL); 1858 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT); 1859 1860 return (mst->sw->snd_tag_modify(mst, ¶ms)); 1861 } 1862 #endif 1863 #endif 1864 1865 void 1866 ktls_destroy(struct ktls_session *tls) 1867 { 1868 1869 if (tls->sequential_records) { 1870 struct mbuf *m, *n; 1871 int page_count; 1872 1873 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) { 1874 page_count = m->m_epg_enc_cnt; 1875 while (page_count > 0) { 1876 KASSERT(page_count >= m->m_epg_nrdy, 1877 ("%s: too few pages", __func__)); 1878 page_count -= m->m_epg_nrdy; 1879 m = m_free(m); 1880 } 1881 } 1882 } 1883 ktls_cleanup(tls); 1884 uma_zfree(ktls_session_zone, tls); 1885 } 1886 1887 void 1888 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1889 { 1890 1891 for (; m != NULL; m = m->m_next) { 1892 KASSERT((m->m_flags & M_EXTPG) != 0, 1893 ("ktls_seq: mapped mbuf %p", m)); 1894 1895 m->m_epg_seqno = sb->sb_tls_seqno; 1896 sb->sb_tls_seqno++; 1897 } 1898 } 1899 1900 /* 1901 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1902 * mbuf in the chain must be an unmapped mbuf. The payload of the 1903 * mbuf must be populated with the payload of each TLS record. 1904 * 1905 * The record_type argument specifies the TLS record type used when 1906 * populating the TLS header. 1907 * 1908 * The enq_count argument on return is set to the number of pages of 1909 * payload data for this entire chain that need to be encrypted via SW 1910 * encryption. The returned value should be passed to ktls_enqueue 1911 * when scheduling encryption of this chain of mbufs. To handle the 1912 * special case of empty fragments for TLS 1.0 sessions, an empty 1913 * fragment counts as one page. 1914 */ 1915 void 1916 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1917 uint8_t record_type) 1918 { 1919 struct tls_record_layer *tlshdr; 1920 struct mbuf *m; 1921 uint64_t *noncep; 1922 uint16_t tls_len; 1923 int maxlen __diagused; 1924 1925 maxlen = tls->params.max_frame_len; 1926 *enq_cnt = 0; 1927 for (m = top; m != NULL; m = m->m_next) { 1928 /* 1929 * All mbufs in the chain should be TLS records whose 1930 * payload does not exceed the maximum frame length. 1931 * 1932 * Empty TLS 1.0 records are permitted when using CBC. 1933 */ 1934 KASSERT(m->m_len <= maxlen && m->m_len >= 0 && 1935 (m->m_len > 0 || ktls_permit_empty_frames(tls)), 1936 ("ktls_frame: m %p len %d", m, m->m_len)); 1937 1938 /* 1939 * TLS frames require unmapped mbufs to store session 1940 * info. 1941 */ 1942 KASSERT((m->m_flags & M_EXTPG) != 0, 1943 ("ktls_frame: mapped mbuf %p (top = %p)", m, top)); 1944 1945 tls_len = m->m_len; 1946 1947 /* Save a reference to the session. */ 1948 m->m_epg_tls = ktls_hold(tls); 1949 1950 m->m_epg_hdrlen = tls->params.tls_hlen; 1951 m->m_epg_trllen = tls->params.tls_tlen; 1952 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 1953 int bs, delta; 1954 1955 /* 1956 * AES-CBC pads messages to a multiple of the 1957 * block size. Note that the padding is 1958 * applied after the digest and the encryption 1959 * is done on the "plaintext || mac || padding". 1960 * At least one byte of padding is always 1961 * present. 1962 * 1963 * Compute the final trailer length assuming 1964 * at most one block of padding. 1965 * tls->params.tls_tlen is the maximum 1966 * possible trailer length (padding + digest). 1967 * delta holds the number of excess padding 1968 * bytes if the maximum were used. Those 1969 * extra bytes are removed. 1970 */ 1971 bs = tls->params.tls_bs; 1972 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 1973 m->m_epg_trllen -= delta; 1974 } 1975 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; 1976 1977 /* Populate the TLS header. */ 1978 tlshdr = (void *)m->m_epg_hdr; 1979 tlshdr->tls_vmajor = tls->params.tls_vmajor; 1980 1981 /* 1982 * TLS 1.3 masquarades as TLS 1.2 with a record type 1983 * of TLS_RLTYPE_APP. 1984 */ 1985 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 1986 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 1987 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 1988 tlshdr->tls_type = TLS_RLTYPE_APP; 1989 /* save the real record type for later */ 1990 m->m_epg_record_type = record_type; 1991 m->m_epg_trail[0] = record_type; 1992 } else { 1993 tlshdr->tls_vminor = tls->params.tls_vminor; 1994 tlshdr->tls_type = record_type; 1995 } 1996 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 1997 1998 /* 1999 * Store nonces / explicit IVs after the end of the 2000 * TLS header. 2001 * 2002 * For GCM with TLS 1.2, an 8 byte nonce is copied 2003 * from the end of the IV. The nonce is then 2004 * incremented for use by the next record. 2005 * 2006 * For CBC, a random nonce is inserted for TLS 1.1+. 2007 */ 2008 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 2009 tls->params.tls_vminor == TLS_MINOR_VER_TWO) { 2010 noncep = (uint64_t *)(tls->params.iv + 8); 2011 be64enc(tlshdr + 1, *noncep); 2012 (*noncep)++; 2013 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 2014 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 2015 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 2016 2017 /* 2018 * When using SW encryption, mark the mbuf not ready. 2019 * It will be marked ready via sbready() after the 2020 * record has been encrypted. 2021 * 2022 * When using ifnet TLS, unencrypted TLS records are 2023 * sent down the stack to the NIC. 2024 */ 2025 if (tls->mode == TCP_TLS_MODE_SW) { 2026 m->m_flags |= M_NOTREADY; 2027 if (__predict_false(tls_len == 0)) { 2028 /* TLS 1.0 empty fragment. */ 2029 m->m_epg_nrdy = 1; 2030 } else 2031 m->m_epg_nrdy = m->m_epg_npgs; 2032 *enq_cnt += m->m_epg_nrdy; 2033 } 2034 } 2035 } 2036 2037 bool 2038 ktls_permit_empty_frames(struct ktls_session *tls) 2039 { 2040 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 2041 tls->params.tls_vminor == TLS_MINOR_VER_ZERO); 2042 } 2043 2044 void 2045 ktls_check_rx(struct sockbuf *sb) 2046 { 2047 struct tls_record_layer hdr; 2048 struct ktls_wq *wq; 2049 struct socket *so; 2050 bool running; 2051 2052 SOCKBUF_LOCK_ASSERT(sb); 2053 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 2054 __func__, sb)); 2055 so = __containerof(sb, struct socket, so_rcv); 2056 2057 if (sb->sb_flags & SB_TLS_RX_RUNNING) 2058 return; 2059 2060 /* Is there enough queued for a TLS header? */ 2061 if (sb->sb_tlscc < sizeof(hdr)) { 2062 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) 2063 so->so_error = EMSGSIZE; 2064 return; 2065 } 2066 2067 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); 2068 2069 /* Is the entire record queued? */ 2070 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { 2071 if ((sb->sb_state & SBS_CANTRCVMORE) != 0) 2072 so->so_error = EMSGSIZE; 2073 return; 2074 } 2075 2076 sb->sb_flags |= SB_TLS_RX_RUNNING; 2077 2078 soref(so); 2079 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; 2080 mtx_lock(&wq->mtx); 2081 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); 2082 running = wq->running; 2083 mtx_unlock(&wq->mtx); 2084 if (!running) 2085 wakeup(wq); 2086 counter_u64_add(ktls_cnt_rx_queued, 1); 2087 } 2088 2089 static struct mbuf * 2090 ktls_detach_record(struct sockbuf *sb, int len) 2091 { 2092 struct mbuf *m, *n, *top; 2093 int remain; 2094 2095 SOCKBUF_LOCK_ASSERT(sb); 2096 MPASS(len <= sb->sb_tlscc); 2097 2098 /* 2099 * If TLS chain is the exact size of the record, 2100 * just grab the whole record. 2101 */ 2102 top = sb->sb_mtls; 2103 if (sb->sb_tlscc == len) { 2104 sb->sb_mtls = NULL; 2105 sb->sb_mtlstail = NULL; 2106 goto out; 2107 } 2108 2109 /* 2110 * While it would be nice to use m_split() here, we need 2111 * to know exactly what m_split() allocates to update the 2112 * accounting, so do it inline instead. 2113 */ 2114 remain = len; 2115 for (m = top; remain > m->m_len; m = m->m_next) 2116 remain -= m->m_len; 2117 2118 /* Easy case: don't have to split 'm'. */ 2119 if (remain == m->m_len) { 2120 sb->sb_mtls = m->m_next; 2121 if (sb->sb_mtls == NULL) 2122 sb->sb_mtlstail = NULL; 2123 m->m_next = NULL; 2124 goto out; 2125 } 2126 2127 /* 2128 * Need to allocate an mbuf to hold the remainder of 'm'. Try 2129 * with M_NOWAIT first. 2130 */ 2131 n = m_get(M_NOWAIT, MT_DATA); 2132 if (n == NULL) { 2133 /* 2134 * Use M_WAITOK with socket buffer unlocked. If 2135 * 'sb_mtls' changes while the lock is dropped, return 2136 * NULL to force the caller to retry. 2137 */ 2138 SOCKBUF_UNLOCK(sb); 2139 2140 n = m_get(M_WAITOK, MT_DATA); 2141 2142 SOCKBUF_LOCK(sb); 2143 if (sb->sb_mtls != top) { 2144 m_free(n); 2145 return (NULL); 2146 } 2147 } 2148 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED)); 2149 2150 /* Store remainder in 'n'. */ 2151 n->m_len = m->m_len - remain; 2152 if (m->m_flags & M_EXT) { 2153 n->m_data = m->m_data + remain; 2154 mb_dupcl(n, m); 2155 } else { 2156 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); 2157 } 2158 2159 /* Trim 'm' and update accounting. */ 2160 m->m_len -= n->m_len; 2161 sb->sb_tlscc -= n->m_len; 2162 sb->sb_ccc -= n->m_len; 2163 2164 /* Account for 'n'. */ 2165 sballoc_ktls_rx(sb, n); 2166 2167 /* Insert 'n' into the TLS chain. */ 2168 sb->sb_mtls = n; 2169 n->m_next = m->m_next; 2170 if (sb->sb_mtlstail == m) 2171 sb->sb_mtlstail = n; 2172 2173 /* Detach the record from the TLS chain. */ 2174 m->m_next = NULL; 2175 2176 out: 2177 MPASS(m_length(top, NULL) == len); 2178 for (m = top; m != NULL; m = m->m_next) 2179 sbfree_ktls_rx(sb, m); 2180 sb->sb_tlsdcc = len; 2181 sb->sb_ccc += len; 2182 SBCHECK(sb); 2183 return (top); 2184 } 2185 2186 /* 2187 * Determine the length of the trailing zero padding and find the real 2188 * record type in the byte before the padding. 2189 * 2190 * Walking the mbuf chain backwards is clumsy, so another option would 2191 * be to scan forwards remembering the last non-zero byte before the 2192 * trailer. However, it would be expensive to scan the entire record. 2193 * Instead, find the last non-zero byte of each mbuf in the chain 2194 * keeping track of the relative offset of that nonzero byte. 2195 * 2196 * trail_len is the size of the MAC/tag on input and is set to the 2197 * size of the full trailer including padding and the record type on 2198 * return. 2199 */ 2200 static int 2201 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, 2202 int *trailer_len, uint8_t *record_typep) 2203 { 2204 char *cp; 2205 u_int digest_start, last_offset, m_len, offset; 2206 uint8_t record_type; 2207 2208 digest_start = tls_len - *trailer_len; 2209 last_offset = 0; 2210 offset = 0; 2211 for (; m != NULL && offset < digest_start; 2212 offset += m->m_len, m = m->m_next) { 2213 /* Don't look for padding in the tag. */ 2214 m_len = min(digest_start - offset, m->m_len); 2215 cp = mtod(m, char *); 2216 2217 /* Find last non-zero byte in this mbuf. */ 2218 while (m_len > 0 && cp[m_len - 1] == 0) 2219 m_len--; 2220 if (m_len > 0) { 2221 record_type = cp[m_len - 1]; 2222 last_offset = offset + m_len; 2223 } 2224 } 2225 if (last_offset < tls->params.tls_hlen) 2226 return (EBADMSG); 2227 2228 *record_typep = record_type; 2229 *trailer_len = tls_len - last_offset + 1; 2230 return (0); 2231 } 2232 2233 /* 2234 * Check if a mbuf chain is fully decrypted at the given offset and 2235 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is 2236 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted 2237 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data 2238 * is encrypted. 2239 */ 2240 ktls_mbuf_crypto_st_t 2241 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len) 2242 { 2243 int m_flags_ored = 0; 2244 int m_flags_anded = -1; 2245 2246 for (; mb != NULL; mb = mb->m_next) { 2247 if (offset < mb->m_len) 2248 break; 2249 offset -= mb->m_len; 2250 } 2251 offset += len; 2252 2253 for (; mb != NULL; mb = mb->m_next) { 2254 m_flags_ored |= mb->m_flags; 2255 m_flags_anded &= mb->m_flags; 2256 2257 if (offset <= mb->m_len) 2258 break; 2259 offset -= mb->m_len; 2260 } 2261 MPASS(mb != NULL || offset == 0); 2262 2263 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED) 2264 return (KTLS_MBUF_CRYPTO_ST_MIXED); 2265 else 2266 return ((m_flags_ored & M_DECRYPTED) ? 2267 KTLS_MBUF_CRYPTO_ST_DECRYPTED : 2268 KTLS_MBUF_CRYPTO_ST_ENCRYPTED); 2269 } 2270 2271 /* 2272 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss 2273 */ 2274 static int 2275 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num) 2276 { 2277 union if_snd_tag_modify_params params; 2278 struct m_snd_tag *mst; 2279 struct inpcb *inp; 2280 struct tcpcb *tp; 2281 2282 mst = so->so_rcv.sb_tls_info->snd_tag; 2283 if (__predict_false(mst == NULL)) 2284 return (EINVAL); 2285 2286 inp = sotoinpcb(so); 2287 if (__predict_false(inp == NULL)) 2288 return (EINVAL); 2289 2290 INP_RLOCK(inp); 2291 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 2292 INP_RUNLOCK(inp); 2293 return (ECONNRESET); 2294 } 2295 2296 tp = intotcpcb(inp); 2297 MPASS(tp != NULL); 2298 2299 /* Get the TCP sequence number of the next valid TLS header. */ 2300 SOCKBUF_LOCK(&so->so_rcv); 2301 params.tls_rx.tls_hdr_tcp_sn = 2302 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len; 2303 params.tls_rx.tls_rec_length = tls_len; 2304 params.tls_rx.tls_seq_number = tls_rcd_num; 2305 SOCKBUF_UNLOCK(&so->so_rcv); 2306 2307 INP_RUNLOCK(inp); 2308 2309 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX); 2310 return (mst->sw->snd_tag_modify(mst, ¶ms)); 2311 } 2312 2313 static void 2314 ktls_decrypt(struct socket *so) 2315 { 2316 char tls_header[MBUF_PEXT_HDR_LEN]; 2317 struct ktls_session *tls; 2318 struct sockbuf *sb; 2319 struct tls_record_layer *hdr; 2320 struct tls_get_record tgr; 2321 struct mbuf *control, *data, *m; 2322 ktls_mbuf_crypto_st_t state; 2323 uint64_t seqno; 2324 int error, remain, tls_len, trail_len; 2325 bool tls13; 2326 uint8_t vminor, record_type; 2327 2328 hdr = (struct tls_record_layer *)tls_header; 2329 sb = &so->so_rcv; 2330 SOCKBUF_LOCK(sb); 2331 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, 2332 ("%s: socket %p not running", __func__, so)); 2333 2334 tls = sb->sb_tls_info; 2335 MPASS(tls != NULL); 2336 2337 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE); 2338 if (tls13) 2339 vminor = TLS_MINOR_VER_TWO; 2340 else 2341 vminor = tls->params.tls_vminor; 2342 for (;;) { 2343 /* Is there enough queued for a TLS header? */ 2344 if (sb->sb_tlscc < tls->params.tls_hlen) 2345 break; 2346 2347 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); 2348 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); 2349 2350 if (hdr->tls_vmajor != tls->params.tls_vmajor || 2351 hdr->tls_vminor != vminor) 2352 error = EINVAL; 2353 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP) 2354 error = EINVAL; 2355 else if (tls_len < tls->params.tls_hlen || tls_len > 2356 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + 2357 tls->params.tls_tlen) 2358 error = EMSGSIZE; 2359 else 2360 error = 0; 2361 if (__predict_false(error != 0)) { 2362 /* 2363 * We have a corrupted record and are likely 2364 * out of sync. The connection isn't 2365 * recoverable at this point, so abort it. 2366 */ 2367 SOCKBUF_UNLOCK(sb); 2368 counter_u64_add(ktls_offload_corrupted_records, 1); 2369 2370 CURVNET_SET(so->so_vnet); 2371 so->so_proto->pr_abort(so); 2372 so->so_error = error; 2373 CURVNET_RESTORE(); 2374 goto deref; 2375 } 2376 2377 /* Is the entire record queued? */ 2378 if (sb->sb_tlscc < tls_len) 2379 break; 2380 2381 /* 2382 * Split out the portion of the mbuf chain containing 2383 * this TLS record. 2384 */ 2385 data = ktls_detach_record(sb, tls_len); 2386 if (data == NULL) 2387 continue; 2388 MPASS(sb->sb_tlsdcc == tls_len); 2389 2390 seqno = sb->sb_tls_seqno; 2391 sb->sb_tls_seqno++; 2392 SBCHECK(sb); 2393 SOCKBUF_UNLOCK(sb); 2394 2395 /* get crypto state for this TLS record */ 2396 state = ktls_mbuf_crypto_state(data, 0, tls_len); 2397 2398 switch (state) { 2399 case KTLS_MBUF_CRYPTO_ST_MIXED: 2400 error = ktls_ocf_recrypt(tls, hdr, data, seqno); 2401 if (error) 2402 break; 2403 /* FALLTHROUGH */ 2404 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED: 2405 error = ktls_ocf_decrypt(tls, hdr, data, seqno, 2406 &trail_len); 2407 if (__predict_true(error == 0)) { 2408 if (tls13) { 2409 error = tls13_find_record_type(tls, data, 2410 tls_len, &trail_len, &record_type); 2411 } else { 2412 record_type = hdr->tls_type; 2413 } 2414 } 2415 break; 2416 case KTLS_MBUF_CRYPTO_ST_DECRYPTED: 2417 /* 2418 * NIC TLS is only supported for AEAD 2419 * ciphersuites which used a fixed sized 2420 * trailer. 2421 */ 2422 if (tls13) { 2423 trail_len = tls->params.tls_tlen - 1; 2424 error = tls13_find_record_type(tls, data, 2425 tls_len, &trail_len, &record_type); 2426 } else { 2427 trail_len = tls->params.tls_tlen; 2428 error = 0; 2429 record_type = hdr->tls_type; 2430 } 2431 break; 2432 default: 2433 error = EINVAL; 2434 break; 2435 } 2436 if (error) { 2437 counter_u64_add(ktls_offload_failed_crypto, 1); 2438 2439 SOCKBUF_LOCK(sb); 2440 if (sb->sb_tlsdcc == 0) { 2441 /* 2442 * sbcut/drop/flush discarded these 2443 * mbufs. 2444 */ 2445 m_freem(data); 2446 break; 2447 } 2448 2449 /* 2450 * Drop this TLS record's data, but keep 2451 * decrypting subsequent records. 2452 */ 2453 sb->sb_ccc -= tls_len; 2454 sb->sb_tlsdcc = 0; 2455 2456 CURVNET_SET(so->so_vnet); 2457 so->so_error = EBADMSG; 2458 sorwakeup_locked(so); 2459 CURVNET_RESTORE(); 2460 2461 m_freem(data); 2462 2463 SOCKBUF_LOCK(sb); 2464 continue; 2465 } 2466 2467 /* Allocate the control mbuf. */ 2468 memset(&tgr, 0, sizeof(tgr)); 2469 tgr.tls_type = record_type; 2470 tgr.tls_vmajor = hdr->tls_vmajor; 2471 tgr.tls_vminor = hdr->tls_vminor; 2472 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - 2473 trail_len); 2474 control = sbcreatecontrol(&tgr, sizeof(tgr), 2475 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); 2476 2477 SOCKBUF_LOCK(sb); 2478 if (sb->sb_tlsdcc == 0) { 2479 /* sbcut/drop/flush discarded these mbufs. */ 2480 MPASS(sb->sb_tlscc == 0); 2481 m_freem(data); 2482 m_freem(control); 2483 break; 2484 } 2485 2486 /* 2487 * Clear the 'dcc' accounting in preparation for 2488 * adding the decrypted record. 2489 */ 2490 sb->sb_ccc -= tls_len; 2491 sb->sb_tlsdcc = 0; 2492 SBCHECK(sb); 2493 2494 /* If there is no payload, drop all of the data. */ 2495 if (tgr.tls_length == htobe16(0)) { 2496 m_freem(data); 2497 data = NULL; 2498 } else { 2499 /* Trim header. */ 2500 remain = tls->params.tls_hlen; 2501 while (remain > 0) { 2502 if (data->m_len > remain) { 2503 data->m_data += remain; 2504 data->m_len -= remain; 2505 break; 2506 } 2507 remain -= data->m_len; 2508 data = m_free(data); 2509 } 2510 2511 /* Trim trailer and clear M_NOTREADY. */ 2512 remain = be16toh(tgr.tls_length); 2513 m = data; 2514 for (m = data; remain > m->m_len; m = m->m_next) { 2515 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); 2516 remain -= m->m_len; 2517 } 2518 m->m_len = remain; 2519 m_freem(m->m_next); 2520 m->m_next = NULL; 2521 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); 2522 2523 /* Set EOR on the final mbuf. */ 2524 m->m_flags |= M_EOR; 2525 } 2526 2527 sbappendcontrol_locked(sb, data, control, 0); 2528 2529 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) { 2530 sb->sb_flags |= SB_TLS_RX_RESYNC; 2531 SOCKBUF_UNLOCK(sb); 2532 ktls_resync_ifnet(so, tls_len, seqno); 2533 SOCKBUF_LOCK(sb); 2534 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) { 2535 sb->sb_flags &= ~SB_TLS_RX_RESYNC; 2536 SOCKBUF_UNLOCK(sb); 2537 ktls_resync_ifnet(so, 0, seqno); 2538 SOCKBUF_LOCK(sb); 2539 } 2540 } 2541 2542 sb->sb_flags &= ~SB_TLS_RX_RUNNING; 2543 2544 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) 2545 so->so_error = EMSGSIZE; 2546 2547 sorwakeup_locked(so); 2548 2549 deref: 2550 SOCKBUF_UNLOCK_ASSERT(sb); 2551 2552 CURVNET_SET(so->so_vnet); 2553 sorele(so); 2554 CURVNET_RESTORE(); 2555 } 2556 2557 void 2558 ktls_enqueue_to_free(struct mbuf *m) 2559 { 2560 struct ktls_wq *wq; 2561 bool running; 2562 2563 /* Mark it for freeing. */ 2564 m->m_epg_flags |= EPG_FLAG_2FREE; 2565 wq = &ktls_wq[m->m_epg_tls->wq_index]; 2566 mtx_lock(&wq->mtx); 2567 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2568 running = wq->running; 2569 mtx_unlock(&wq->mtx); 2570 if (!running) 2571 wakeup(wq); 2572 } 2573 2574 static void * 2575 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m) 2576 { 2577 void *buf; 2578 int domain, running; 2579 2580 if (m->m_epg_npgs <= 2) 2581 return (NULL); 2582 if (ktls_buffer_zone == NULL) 2583 return (NULL); 2584 if ((u_int)(ticks - wq->lastallocfail) < hz) { 2585 /* 2586 * Rate-limit allocation attempts after a failure. 2587 * ktls_buffer_import() will acquire a per-domain mutex to check 2588 * the free page queues and may fail consistently if memory is 2589 * fragmented. 2590 */ 2591 return (NULL); 2592 } 2593 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM); 2594 if (buf == NULL) { 2595 domain = PCPU_GET(domain); 2596 wq->lastallocfail = ticks; 2597 2598 /* 2599 * Note that this check is "racy", but the races are 2600 * harmless, and are either a spurious wakeup if 2601 * multiple threads fail allocations before the alloc 2602 * thread wakes, or waiting an extra second in case we 2603 * see an old value of running == true. 2604 */ 2605 if (!VM_DOMAIN_EMPTY(domain)) { 2606 running = atomic_load_int(&ktls_domains[domain].alloc_td.running); 2607 if (!running) 2608 wakeup(&ktls_domains[domain].alloc_td); 2609 } 2610 } 2611 return (buf); 2612 } 2613 2614 static int 2615 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m, 2616 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state) 2617 { 2618 vm_page_t pg; 2619 int error, i, len, off; 2620 2621 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY), 2622 ("%p not unready & nomap mbuf\n", m)); 2623 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen, 2624 ("page count %d larger than maximum frame length %d", m->m_epg_npgs, 2625 ktls_maxlen)); 2626 2627 /* Anonymous mbufs are encrypted in place. */ 2628 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0) 2629 return (ktls_ocf_encrypt(state, tls, m, NULL, 0)); 2630 2631 /* 2632 * For file-backed mbufs (from sendfile), anonymous wired 2633 * pages are allocated and used as the encryption destination. 2634 */ 2635 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) { 2636 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len - 2637 m->m_epg_1st_off; 2638 state->dst_iov[0].iov_base = (char *)state->cbuf + 2639 m->m_epg_1st_off; 2640 state->dst_iov[0].iov_len = len; 2641 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf); 2642 i = 1; 2643 } else { 2644 off = m->m_epg_1st_off; 2645 for (i = 0; i < m->m_epg_npgs; i++, off = 0) { 2646 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP | 2647 VM_ALLOC_WIRED | VM_ALLOC_WAITOK); 2648 len = m_epg_pagelen(m, i, off); 2649 state->parray[i] = VM_PAGE_TO_PHYS(pg); 2650 state->dst_iov[i].iov_base = 2651 (char *)PHYS_TO_DMAP(state->parray[i]) + off; 2652 state->dst_iov[i].iov_len = len; 2653 } 2654 } 2655 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small")); 2656 state->dst_iov[i].iov_base = m->m_epg_trail; 2657 state->dst_iov[i].iov_len = m->m_epg_trllen; 2658 2659 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1); 2660 2661 if (__predict_false(error != 0)) { 2662 /* Free the anonymous pages. */ 2663 if (state->cbuf != NULL) 2664 uma_zfree(ktls_buffer_zone, state->cbuf); 2665 else { 2666 for (i = 0; i < m->m_epg_npgs; i++) { 2667 pg = PHYS_TO_VM_PAGE(state->parray[i]); 2668 (void)vm_page_unwire_noq(pg); 2669 vm_page_free(pg); 2670 } 2671 } 2672 } 2673 return (error); 2674 } 2675 2676 /* Number of TLS records in a batch passed to ktls_enqueue(). */ 2677 static u_int 2678 ktls_batched_records(struct mbuf *m) 2679 { 2680 int page_count, records; 2681 2682 records = 0; 2683 page_count = m->m_epg_enc_cnt; 2684 while (page_count > 0) { 2685 records++; 2686 page_count -= m->m_epg_nrdy; 2687 m = m->m_next; 2688 } 2689 KASSERT(page_count == 0, ("%s: mismatched page count", __func__)); 2690 return (records); 2691 } 2692 2693 void 2694 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 2695 { 2696 struct ktls_session *tls; 2697 struct ktls_wq *wq; 2698 int queued; 2699 bool running; 2700 2701 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == 2702 (M_EXTPG | M_NOTREADY)), 2703 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 2704 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 2705 2706 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); 2707 2708 m->m_epg_enc_cnt = page_count; 2709 2710 /* 2711 * Save a pointer to the socket. The caller is responsible 2712 * for taking an additional reference via soref(). 2713 */ 2714 m->m_epg_so = so; 2715 2716 queued = 1; 2717 tls = m->m_epg_tls; 2718 wq = &ktls_wq[tls->wq_index]; 2719 mtx_lock(&wq->mtx); 2720 if (__predict_false(tls->sequential_records)) { 2721 /* 2722 * For TLS 1.0, records must be encrypted 2723 * sequentially. For a given connection, all records 2724 * queued to the associated work queue are processed 2725 * sequentially. However, sendfile(2) might complete 2726 * I/O requests spanning multiple TLS records out of 2727 * order. Here we ensure TLS records are enqueued to 2728 * the work queue in FIFO order. 2729 * 2730 * tls->next_seqno holds the sequence number of the 2731 * next TLS record that should be enqueued to the work 2732 * queue. If this next record is not tls->next_seqno, 2733 * it must be a future record, so insert it, sorted by 2734 * TLS sequence number, into tls->pending_records and 2735 * return. 2736 * 2737 * If this TLS record matches tls->next_seqno, place 2738 * it in the work queue and then check 2739 * tls->pending_records to see if any 2740 * previously-queued records are now ready for 2741 * encryption. 2742 */ 2743 if (m->m_epg_seqno != tls->next_seqno) { 2744 struct mbuf *n, *p; 2745 2746 p = NULL; 2747 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) { 2748 if (n->m_epg_seqno > m->m_epg_seqno) 2749 break; 2750 p = n; 2751 } 2752 if (n == NULL) 2753 STAILQ_INSERT_TAIL(&tls->pending_records, m, 2754 m_epg_stailq); 2755 else if (p == NULL) 2756 STAILQ_INSERT_HEAD(&tls->pending_records, m, 2757 m_epg_stailq); 2758 else 2759 STAILQ_INSERT_AFTER(&tls->pending_records, p, m, 2760 m_epg_stailq); 2761 mtx_unlock(&wq->mtx); 2762 counter_u64_add(ktls_cnt_tx_pending, 1); 2763 return; 2764 } 2765 2766 tls->next_seqno += ktls_batched_records(m); 2767 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2768 2769 while (!STAILQ_EMPTY(&tls->pending_records)) { 2770 struct mbuf *n; 2771 2772 n = STAILQ_FIRST(&tls->pending_records); 2773 if (n->m_epg_seqno != tls->next_seqno) 2774 break; 2775 2776 queued++; 2777 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq); 2778 tls->next_seqno += ktls_batched_records(n); 2779 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq); 2780 } 2781 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1)); 2782 } else 2783 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2784 2785 running = wq->running; 2786 mtx_unlock(&wq->mtx); 2787 if (!running) 2788 wakeup(wq); 2789 counter_u64_add(ktls_cnt_tx_queued, queued); 2790 } 2791 2792 /* 2793 * Once a file-backed mbuf (from sendfile) has been encrypted, free 2794 * the pages from the file and replace them with the anonymous pages 2795 * allocated in ktls_encrypt_record(). 2796 */ 2797 static void 2798 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state) 2799 { 2800 int i; 2801 2802 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0); 2803 2804 /* Free the old pages. */ 2805 m->m_ext.ext_free(m); 2806 2807 /* Replace them with the new pages. */ 2808 if (state->cbuf != NULL) { 2809 for (i = 0; i < m->m_epg_npgs; i++) 2810 m->m_epg_pa[i] = state->parray[0] + ptoa(i); 2811 2812 /* Contig pages should go back to the cache. */ 2813 m->m_ext.ext_free = ktls_free_mext_contig; 2814 } else { 2815 for (i = 0; i < m->m_epg_npgs; i++) 2816 m->m_epg_pa[i] = state->parray[i]; 2817 2818 /* Use the basic free routine. */ 2819 m->m_ext.ext_free = mb_free_mext_pgs; 2820 } 2821 2822 /* Pages are now writable. */ 2823 m->m_epg_flags |= EPG_FLAG_ANON; 2824 } 2825 2826 static __noinline void 2827 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top) 2828 { 2829 struct ktls_ocf_encrypt_state state; 2830 struct ktls_session *tls; 2831 struct socket *so; 2832 struct mbuf *m; 2833 int error, npages, total_pages; 2834 2835 so = top->m_epg_so; 2836 tls = top->m_epg_tls; 2837 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 2838 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 2839 #ifdef INVARIANTS 2840 top->m_epg_so = NULL; 2841 #endif 2842 total_pages = top->m_epg_enc_cnt; 2843 npages = 0; 2844 2845 /* 2846 * Encrypt the TLS records in the chain of mbufs starting with 2847 * 'top'. 'total_pages' gives us a total count of pages and is 2848 * used to know when we have finished encrypting the TLS 2849 * records originally queued with 'top'. 2850 * 2851 * NB: These mbufs are queued in the socket buffer and 2852 * 'm_next' is traversing the mbufs in the socket buffer. The 2853 * socket buffer lock is not held while traversing this chain. 2854 * Since the mbufs are all marked M_NOTREADY their 'm_next' 2855 * pointers should be stable. However, the 'm_next' of the 2856 * last mbuf encrypted is not necessarily NULL. It can point 2857 * to other mbufs appended while 'top' was on the TLS work 2858 * queue. 2859 * 2860 * Each mbuf holds an entire TLS record. 2861 */ 2862 error = 0; 2863 for (m = top; npages != total_pages; m = m->m_next) { 2864 KASSERT(m->m_epg_tls == tls, 2865 ("different TLS sessions in a single mbuf chain: %p vs %p", 2866 tls, m->m_epg_tls)); 2867 KASSERT(npages + m->m_epg_npgs <= total_pages, 2868 ("page count mismatch: top %p, total_pages %d, m %p", top, 2869 total_pages, m)); 2870 2871 error = ktls_encrypt_record(wq, m, tls, &state); 2872 if (error) { 2873 counter_u64_add(ktls_offload_failed_crypto, 1); 2874 break; 2875 } 2876 2877 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) 2878 ktls_finish_nonanon(m, &state); 2879 2880 npages += m->m_epg_nrdy; 2881 2882 /* 2883 * Drop a reference to the session now that it is no 2884 * longer needed. Existing code depends on encrypted 2885 * records having no associated session vs 2886 * yet-to-be-encrypted records having an associated 2887 * session. 2888 */ 2889 m->m_epg_tls = NULL; 2890 ktls_free(tls); 2891 } 2892 2893 CURVNET_SET(so->so_vnet); 2894 if (error == 0) { 2895 (void)so->so_proto->pr_ready(so, top, npages); 2896 } else { 2897 so->so_proto->pr_abort(so); 2898 so->so_error = EIO; 2899 mb_free_notready(top, total_pages); 2900 } 2901 2902 sorele(so); 2903 CURVNET_RESTORE(); 2904 } 2905 2906 void 2907 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error) 2908 { 2909 struct ktls_session *tls; 2910 struct socket *so; 2911 struct mbuf *m; 2912 int npages; 2913 2914 m = state->m; 2915 2916 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) 2917 ktls_finish_nonanon(m, state); 2918 2919 so = state->so; 2920 free(state, M_KTLS); 2921 2922 /* 2923 * Drop a reference to the session now that it is no longer 2924 * needed. Existing code depends on encrypted records having 2925 * no associated session vs yet-to-be-encrypted records having 2926 * an associated session. 2927 */ 2928 tls = m->m_epg_tls; 2929 m->m_epg_tls = NULL; 2930 ktls_free(tls); 2931 2932 if (error != 0) 2933 counter_u64_add(ktls_offload_failed_crypto, 1); 2934 2935 CURVNET_SET(so->so_vnet); 2936 npages = m->m_epg_nrdy; 2937 2938 if (error == 0) { 2939 (void)so->so_proto->pr_ready(so, m, npages); 2940 } else { 2941 so->so_proto->pr_abort(so); 2942 so->so_error = EIO; 2943 mb_free_notready(m, npages); 2944 } 2945 2946 sorele(so); 2947 CURVNET_RESTORE(); 2948 } 2949 2950 /* 2951 * Similar to ktls_encrypt, but used with asynchronous OCF backends 2952 * (coprocessors) where encryption does not use host CPU resources and 2953 * it can be beneficial to queue more requests than CPUs. 2954 */ 2955 static __noinline void 2956 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top) 2957 { 2958 struct ktls_ocf_encrypt_state *state; 2959 struct ktls_session *tls; 2960 struct socket *so; 2961 struct mbuf *m, *n; 2962 int error, mpages, npages, total_pages; 2963 2964 so = top->m_epg_so; 2965 tls = top->m_epg_tls; 2966 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 2967 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 2968 #ifdef INVARIANTS 2969 top->m_epg_so = NULL; 2970 #endif 2971 total_pages = top->m_epg_enc_cnt; 2972 npages = 0; 2973 2974 error = 0; 2975 for (m = top; npages != total_pages; m = n) { 2976 KASSERT(m->m_epg_tls == tls, 2977 ("different TLS sessions in a single mbuf chain: %p vs %p", 2978 tls, m->m_epg_tls)); 2979 KASSERT(npages + m->m_epg_npgs <= total_pages, 2980 ("page count mismatch: top %p, total_pages %d, m %p", top, 2981 total_pages, m)); 2982 2983 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO); 2984 soref(so); 2985 state->so = so; 2986 state->m = m; 2987 2988 mpages = m->m_epg_nrdy; 2989 n = m->m_next; 2990 2991 error = ktls_encrypt_record(wq, m, tls, state); 2992 if (error) { 2993 counter_u64_add(ktls_offload_failed_crypto, 1); 2994 free(state, M_KTLS); 2995 CURVNET_SET(so->so_vnet); 2996 sorele(so); 2997 CURVNET_RESTORE(); 2998 break; 2999 } 3000 3001 npages += mpages; 3002 } 3003 3004 CURVNET_SET(so->so_vnet); 3005 if (error != 0) { 3006 so->so_proto->pr_abort(so); 3007 so->so_error = EIO; 3008 mb_free_notready(m, total_pages - npages); 3009 } 3010 3011 sorele(so); 3012 CURVNET_RESTORE(); 3013 } 3014 3015 static int 3016 ktls_bind_domain(int domain) 3017 { 3018 int error; 3019 3020 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]); 3021 if (error != 0) 3022 return (error); 3023 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain); 3024 return (0); 3025 } 3026 3027 static void 3028 ktls_alloc_thread(void *ctx) 3029 { 3030 struct ktls_domain_info *ktls_domain = ctx; 3031 struct ktls_alloc_thread *sc = &ktls_domain->alloc_td; 3032 void **buf; 3033 struct sysctl_oid *oid; 3034 char name[80]; 3035 int domain, error, i, nbufs; 3036 3037 domain = ktls_domain - ktls_domains; 3038 if (bootverbose) 3039 printf("Starting KTLS alloc thread for domain %d\n", domain); 3040 error = ktls_bind_domain(domain); 3041 if (error) 3042 printf("Unable to bind KTLS alloc thread for domain %d: error %d\n", 3043 domain, error); 3044 snprintf(name, sizeof(name), "domain%d", domain); 3045 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO, 3046 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 3047 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs", 3048 CTLFLAG_RD, &sc->allocs, 0, "buffers allocated"); 3049 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups", 3050 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups"); 3051 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running", 3052 CTLFLAG_RD, &sc->running, 0, "thread running"); 3053 3054 buf = NULL; 3055 nbufs = 0; 3056 for (;;) { 3057 atomic_store_int(&sc->running, 0); 3058 tsleep(sc, PZERO | PNOLOCK, "-", 0); 3059 atomic_store_int(&sc->running, 1); 3060 sc->wakeups++; 3061 if (nbufs != ktls_max_alloc) { 3062 free(buf, M_KTLS); 3063 nbufs = atomic_load_int(&ktls_max_alloc); 3064 buf = malloc(sizeof(void *) * nbufs, M_KTLS, 3065 M_WAITOK | M_ZERO); 3066 } 3067 /* 3068 * Below we allocate nbufs with different allocation 3069 * flags than we use when allocating normally during 3070 * encryption in the ktls worker thread. We specify 3071 * M_NORECLAIM in the worker thread. However, we omit 3072 * that flag here and add M_WAITOK so that the VM 3073 * system is permitted to perform expensive work to 3074 * defragment memory. We do this here, as it does not 3075 * matter if this thread blocks. If we block a ktls 3076 * worker thread, we risk developing backlogs of 3077 * buffers to be encrypted, leading to surges of 3078 * traffic and potential NIC output drops. 3079 */ 3080 for (i = 0; i < nbufs; i++) { 3081 buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK); 3082 sc->allocs++; 3083 } 3084 for (i = 0; i < nbufs; i++) { 3085 uma_zfree(ktls_buffer_zone, buf[i]); 3086 buf[i] = NULL; 3087 } 3088 } 3089 } 3090 3091 static void 3092 ktls_work_thread(void *ctx) 3093 { 3094 struct ktls_wq *wq = ctx; 3095 struct mbuf *m, *n; 3096 struct socket *so, *son; 3097 STAILQ_HEAD(, mbuf) local_m_head; 3098 STAILQ_HEAD(, socket) local_so_head; 3099 int cpu; 3100 3101 cpu = wq - ktls_wq; 3102 if (bootverbose) 3103 printf("Starting KTLS worker thread for CPU %d\n", cpu); 3104 3105 /* 3106 * Bind to a core. If ktls_bind_threads is > 1, then 3107 * we bind to the NUMA domain instead. 3108 */ 3109 if (ktls_bind_threads) { 3110 int error; 3111 3112 if (ktls_bind_threads > 1) { 3113 struct pcpu *pc = pcpu_find(cpu); 3114 3115 error = ktls_bind_domain(pc->pc_domain); 3116 } else { 3117 cpuset_t mask; 3118 3119 CPU_SETOF(cpu, &mask); 3120 error = cpuset_setthread(curthread->td_tid, &mask); 3121 } 3122 if (error) 3123 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n", 3124 cpu, error); 3125 } 3126 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 3127 fpu_kern_thread(0); 3128 #endif 3129 for (;;) { 3130 mtx_lock(&wq->mtx); 3131 while (STAILQ_EMPTY(&wq->m_head) && 3132 STAILQ_EMPTY(&wq->so_head)) { 3133 wq->running = false; 3134 mtx_sleep(wq, &wq->mtx, 0, "-", 0); 3135 wq->running = true; 3136 } 3137 3138 STAILQ_INIT(&local_m_head); 3139 STAILQ_CONCAT(&local_m_head, &wq->m_head); 3140 STAILQ_INIT(&local_so_head); 3141 STAILQ_CONCAT(&local_so_head, &wq->so_head); 3142 mtx_unlock(&wq->mtx); 3143 3144 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) { 3145 if (m->m_epg_flags & EPG_FLAG_2FREE) { 3146 ktls_free(m->m_epg_tls); 3147 m_free_raw(m); 3148 } else { 3149 if (m->m_epg_tls->sync_dispatch) 3150 ktls_encrypt(wq, m); 3151 else 3152 ktls_encrypt_async(wq, m); 3153 counter_u64_add(ktls_cnt_tx_queued, -1); 3154 } 3155 } 3156 3157 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) { 3158 ktls_decrypt(so); 3159 counter_u64_add(ktls_cnt_rx_queued, -1); 3160 } 3161 } 3162 } 3163 3164 #if defined(INET) || defined(INET6) 3165 static void 3166 ktls_disable_ifnet_help(void *context, int pending __unused) 3167 { 3168 struct ktls_session *tls; 3169 struct inpcb *inp; 3170 struct tcpcb *tp; 3171 struct socket *so; 3172 int err; 3173 3174 tls = context; 3175 inp = tls->inp; 3176 if (inp == NULL) 3177 return; 3178 INP_WLOCK(inp); 3179 so = inp->inp_socket; 3180 MPASS(so != NULL); 3181 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 3182 goto out; 3183 } 3184 3185 if (so->so_snd.sb_tls_info != NULL) 3186 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW); 3187 else 3188 err = ENXIO; 3189 if (err == 0) { 3190 counter_u64_add(ktls_ifnet_disable_ok, 1); 3191 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */ 3192 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 && 3193 (tp = intotcpcb(inp)) != NULL && 3194 tp->t_fb->tfb_hwtls_change != NULL) 3195 (*tp->t_fb->tfb_hwtls_change)(tp, 0); 3196 } else { 3197 counter_u64_add(ktls_ifnet_disable_fail, 1); 3198 } 3199 3200 out: 3201 sorele(so); 3202 if (!in_pcbrele_wlocked(inp)) 3203 INP_WUNLOCK(inp); 3204 ktls_free(tls); 3205 } 3206 3207 /* 3208 * Called when re-transmits are becoming a substantial portion of the 3209 * sends on this connection. When this happens, we transition the 3210 * connection to software TLS. This is needed because most inline TLS 3211 * NICs keep crypto state only for in-order transmits. This means 3212 * that to handle a TCP rexmit (which is out-of-order), the NIC must 3213 * re-DMA the entire TLS record up to and including the current 3214 * segment. This means that when re-transmitting the last ~1448 byte 3215 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order 3216 * of magnitude more data than we are sending. This can cause the 3217 * PCIe link to saturate well before the network, which can cause 3218 * output drops, and a general loss of capacity. 3219 */ 3220 void 3221 ktls_disable_ifnet(void *arg) 3222 { 3223 struct tcpcb *tp; 3224 struct inpcb *inp; 3225 struct socket *so; 3226 struct ktls_session *tls; 3227 3228 tp = arg; 3229 inp = tp->t_inpcb; 3230 INP_WLOCK_ASSERT(inp); 3231 so = inp->inp_socket; 3232 SOCK_LOCK(so); 3233 tls = so->so_snd.sb_tls_info; 3234 if (tls->disable_ifnet_pending) { 3235 SOCK_UNLOCK(so); 3236 return; 3237 } 3238 3239 /* 3240 * note that disable_ifnet_pending is never cleared; disabling 3241 * ifnet can only be done once per session, so we never want 3242 * to do it again 3243 */ 3244 3245 (void)ktls_hold(tls); 3246 in_pcbref(inp); 3247 soref(so); 3248 tls->disable_ifnet_pending = true; 3249 tls->inp = inp; 3250 SOCK_UNLOCK(so); 3251 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls); 3252 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task); 3253 } 3254 #endif 3255