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 #include "opt_inet.h" 30 #include "opt_inet6.h" 31 #include "opt_kern_tls.h" 32 #include "opt_ratelimit.h" 33 #include "opt_rss.h" 34 35 #include <sys/param.h> 36 #include <sys/kernel.h> 37 #include <sys/domainset.h> 38 #include <sys/endian.h> 39 #include <sys/ktls.h> 40 #include <sys/lock.h> 41 #include <sys/mbuf.h> 42 #include <sys/mutex.h> 43 #include <sys/rmlock.h> 44 #include <sys/proc.h> 45 #include <sys/protosw.h> 46 #include <sys/refcount.h> 47 #include <sys/smp.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/sysctl.h> 51 #include <sys/taskqueue.h> 52 #include <sys/kthread.h> 53 #include <sys/uio.h> 54 #include <sys/vmmeter.h> 55 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 56 #include <machine/pcb.h> 57 #endif 58 #include <machine/vmparam.h> 59 #include <net/if.h> 60 #include <net/if_var.h> 61 #ifdef RSS 62 #include <net/netisr.h> 63 #include <net/rss_config.h> 64 #endif 65 #include <net/route.h> 66 #include <net/route/nhop.h> 67 #include <netinet/in.h> 68 #include <netinet/in_pcb.h> 69 #include <netinet/tcp_var.h> 70 #ifdef TCP_OFFLOAD 71 #include <netinet/tcp_offload.h> 72 #endif 73 #include <opencrypto/cryptodev.h> 74 #include <opencrypto/ktls.h> 75 #include <vm/vm.h> 76 #include <vm/vm_pageout.h> 77 #include <vm/vm_page.h> 78 #include <vm/vm_pagequeue.h> 79 80 struct ktls_wq { 81 struct mtx mtx; 82 STAILQ_HEAD(, mbuf) m_head; 83 STAILQ_HEAD(, socket) so_head; 84 bool running; 85 int lastallocfail; 86 } __aligned(CACHE_LINE_SIZE); 87 88 struct ktls_reclaim_thread { 89 uint64_t wakeups; 90 uint64_t reclaims; 91 struct thread *td; 92 int running; 93 }; 94 95 struct ktls_domain_info { 96 int count; 97 int cpu[MAXCPU]; 98 struct ktls_reclaim_thread reclaim_td; 99 }; 100 101 struct ktls_domain_info ktls_domains[MAXMEMDOM]; 102 static struct ktls_wq *ktls_wq; 103 static struct proc *ktls_proc; 104 static uma_zone_t ktls_session_zone; 105 static uma_zone_t ktls_buffer_zone; 106 static uint16_t ktls_cpuid_lookup[MAXCPU]; 107 static int ktls_init_state; 108 static struct sx ktls_init_lock; 109 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init"); 110 111 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 112 "Kernel TLS offload"); 113 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 114 "Kernel TLS offload stats"); 115 116 #ifdef RSS 117 static int ktls_bind_threads = 1; 118 #else 119 static int ktls_bind_threads; 120 #endif 121 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 122 &ktls_bind_threads, 0, 123 "Bind crypto threads to cores (1) or cores and domains (2) at boot"); 124 125 static u_int ktls_maxlen = 16384; 126 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN, 127 &ktls_maxlen, 0, "Maximum TLS record size"); 128 129 static int ktls_number_threads; 130 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 131 &ktls_number_threads, 0, 132 "Number of TLS threads in thread-pool"); 133 134 unsigned int ktls_ifnet_max_rexmit_pct = 2; 135 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN, 136 &ktls_ifnet_max_rexmit_pct, 2, 137 "Max percent bytes retransmitted before ifnet TLS is disabled"); 138 139 static bool ktls_offload_enable; 140 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, 141 &ktls_offload_enable, 0, 142 "Enable support for kernel TLS offload"); 143 144 static bool ktls_cbc_enable = true; 145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN, 146 &ktls_cbc_enable, 1, 147 "Enable support of AES-CBC crypto for kernel TLS"); 148 149 static bool ktls_sw_buffer_cache = true; 150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN, 151 &ktls_sw_buffer_cache, 1, 152 "Enable caching of output buffers for SW encryption"); 153 154 static int ktls_max_reclaim = 1024; 155 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN, 156 &ktls_max_reclaim, 128, 157 "Max number of 16k buffers to reclaim in thread context"); 158 159 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active); 160 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 161 &ktls_tasks_active, "Number of active tasks"); 162 163 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending); 164 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD, 165 &ktls_cnt_tx_pending, 166 "Number of TLS 1.0 records waiting for earlier TLS records"); 167 168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued); 169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, 170 &ktls_cnt_tx_queued, 171 "Number of TLS records in queue to tasks for SW encryption"); 172 173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued); 174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, 175 &ktls_cnt_rx_queued, 176 "Number of TLS sockets in queue to tasks for SW decryption"); 177 178 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total); 179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 180 CTLFLAG_RD, &ktls_offload_total, 181 "Total successful TLS setups (parameters set)"); 182 183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls); 184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 185 CTLFLAG_RD, &ktls_offload_enable_calls, 186 "Total number of TLS enable calls made"); 187 188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active); 189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 190 &ktls_offload_active, "Total Active TLS sessions"); 191 192 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records); 193 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, 194 &ktls_offload_corrupted_records, "Total corrupted TLS records received"); 195 196 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto); 197 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 198 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 199 200 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet); 201 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 202 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 203 204 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw); 205 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 206 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 207 208 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed); 209 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 210 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 211 212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail); 213 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD, 214 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet"); 215 216 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok); 217 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD, 218 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet"); 219 220 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task); 221 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD, 222 &ktls_destroy_task, 223 "Number of times ktls session was destroyed via taskqueue"); 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_reset_receive_tag(void *context, int pending); 301 static void ktls_reset_send_tag(void *context, int pending); 302 static void ktls_work_thread(void *ctx); 303 static void ktls_reclaim_thread(void *ctx); 304 305 static u_int 306 ktls_get_cpu(struct socket *so) 307 { 308 struct inpcb *inp; 309 #ifdef NUMA 310 struct ktls_domain_info *di; 311 #endif 312 u_int cpuid; 313 314 inp = sotoinpcb(so); 315 #ifdef RSS 316 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 317 if (cpuid != NETISR_CPUID_NONE) 318 return (cpuid); 319 #endif 320 /* 321 * Just use the flowid to shard connections in a repeatable 322 * fashion. Note that TLS 1.0 sessions rely on the 323 * serialization provided by having the same connection use 324 * the same queue. 325 */ 326 #ifdef NUMA 327 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) { 328 di = &ktls_domains[inp->inp_numa_domain]; 329 cpuid = di->cpu[inp->inp_flowid % di->count]; 330 } else 331 #endif 332 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 333 return (cpuid); 334 } 335 336 static int 337 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags) 338 { 339 vm_page_t m; 340 int i, req; 341 342 KASSERT((ktls_maxlen & PAGE_MASK) == 0, 343 ("%s: ktls max length %d is not page size-aligned", 344 __func__, ktls_maxlen)); 345 346 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags); 347 for (i = 0; i < count; i++) { 348 m = vm_page_alloc_noobj_contig_domain(domain, req, 349 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, 350 VM_MEMATTR_DEFAULT); 351 if (m == NULL) 352 break; 353 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 354 } 355 return (i); 356 } 357 358 static void 359 ktls_buffer_release(void *arg __unused, void **store, int count) 360 { 361 vm_page_t m; 362 int i, j; 363 364 for (i = 0; i < count; i++) { 365 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i])); 366 for (j = 0; j < atop(ktls_maxlen); j++) { 367 (void)vm_page_unwire_noq(m + j); 368 vm_page_free(m + j); 369 } 370 } 371 } 372 373 static void 374 ktls_free_mext_contig(struct mbuf *m) 375 { 376 M_ASSERTEXTPG(m); 377 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0])); 378 } 379 380 static int 381 ktls_init(void) 382 { 383 struct thread *td; 384 struct pcpu *pc; 385 int count, domain, error, i; 386 387 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 388 M_WAITOK | M_ZERO); 389 390 ktls_session_zone = uma_zcreate("ktls_session", 391 sizeof(struct ktls_session), 392 NULL, NULL, NULL, NULL, 393 UMA_ALIGN_CACHE, 0); 394 395 if (ktls_sw_buffer_cache) { 396 ktls_buffer_zone = uma_zcache_create("ktls_buffers", 397 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL, 398 ktls_buffer_import, ktls_buffer_release, NULL, 399 UMA_ZONE_FIRSTTOUCH); 400 } 401 402 /* 403 * Initialize the workqueues to run the TLS work. We create a 404 * work queue for each CPU. 405 */ 406 CPU_FOREACH(i) { 407 STAILQ_INIT(&ktls_wq[i].m_head); 408 STAILQ_INIT(&ktls_wq[i].so_head); 409 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 410 if (ktls_bind_threads > 1) { 411 pc = pcpu_find(i); 412 domain = pc->pc_domain; 413 count = ktls_domains[domain].count; 414 ktls_domains[domain].cpu[count] = i; 415 ktls_domains[domain].count++; 416 } 417 ktls_cpuid_lookup[ktls_number_threads] = i; 418 ktls_number_threads++; 419 } 420 421 /* 422 * If we somehow have an empty domain, fall back to choosing 423 * among all KTLS threads. 424 */ 425 if (ktls_bind_threads > 1) { 426 for (i = 0; i < vm_ndomains; i++) { 427 if (ktls_domains[i].count == 0) { 428 ktls_bind_threads = 1; 429 break; 430 } 431 } 432 } 433 434 /* Start kthreads for each workqueue. */ 435 CPU_FOREACH(i) { 436 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 437 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 438 if (error) { 439 printf("Can't add KTLS thread %d error %d\n", i, error); 440 return (error); 441 } 442 } 443 444 /* 445 * Start an allocation thread per-domain to perform blocking allocations 446 * of 16k physically contiguous TLS crypto destination buffers. 447 */ 448 if (ktls_sw_buffer_cache) { 449 for (domain = 0; domain < vm_ndomains; domain++) { 450 if (VM_DOMAIN_EMPTY(domain)) 451 continue; 452 if (CPU_EMPTY(&cpuset_domain[domain])) 453 continue; 454 error = kproc_kthread_add(ktls_reclaim_thread, 455 &ktls_domains[domain], &ktls_proc, 456 &ktls_domains[domain].reclaim_td.td, 457 0, 0, "KTLS", "reclaim_%d", domain); 458 if (error) { 459 printf("Can't add KTLS reclaim thread %d error %d\n", 460 domain, error); 461 return (error); 462 } 463 } 464 } 465 466 if (bootverbose) 467 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 468 return (0); 469 } 470 471 static int 472 ktls_start_kthreads(void) 473 { 474 int error, state; 475 476 start: 477 state = atomic_load_acq_int(&ktls_init_state); 478 if (__predict_true(state > 0)) 479 return (0); 480 if (state < 0) 481 return (ENXIO); 482 483 sx_xlock(&ktls_init_lock); 484 if (ktls_init_state != 0) { 485 sx_xunlock(&ktls_init_lock); 486 goto start; 487 } 488 489 error = ktls_init(); 490 if (error == 0) 491 state = 1; 492 else 493 state = -1; 494 atomic_store_rel_int(&ktls_init_state, state); 495 sx_xunlock(&ktls_init_lock); 496 return (error); 497 } 498 499 static int 500 ktls_create_session(struct socket *so, struct tls_enable *en, 501 struct ktls_session **tlsp, int direction) 502 { 503 struct ktls_session *tls; 504 int error; 505 506 /* Only TLS 1.0 - 1.3 are supported. */ 507 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 508 return (EINVAL); 509 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 510 en->tls_vminor > TLS_MINOR_VER_THREE) 511 return (EINVAL); 512 513 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 514 return (EINVAL); 515 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 516 return (EINVAL); 517 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 518 return (EINVAL); 519 520 /* All supported algorithms require a cipher key. */ 521 if (en->cipher_key_len == 0) 522 return (EINVAL); 523 524 /* No flags are currently supported. */ 525 if (en->flags != 0) 526 return (EINVAL); 527 528 /* Common checks for supported algorithms. */ 529 switch (en->cipher_algorithm) { 530 case CRYPTO_AES_NIST_GCM_16: 531 /* 532 * auth_algorithm isn't used, but permit GMAC values 533 * for compatibility. 534 */ 535 switch (en->auth_algorithm) { 536 case 0: 537 #ifdef COMPAT_FREEBSD12 538 /* XXX: Really 13.0-current COMPAT. */ 539 case CRYPTO_AES_128_NIST_GMAC: 540 case CRYPTO_AES_192_NIST_GMAC: 541 case CRYPTO_AES_256_NIST_GMAC: 542 #endif 543 break; 544 default: 545 return (EINVAL); 546 } 547 if (en->auth_key_len != 0) 548 return (EINVAL); 549 switch (en->tls_vminor) { 550 case TLS_MINOR_VER_TWO: 551 if (en->iv_len != TLS_AEAD_GCM_LEN) 552 return (EINVAL); 553 break; 554 case TLS_MINOR_VER_THREE: 555 if (en->iv_len != TLS_1_3_GCM_IV_LEN) 556 return (EINVAL); 557 break; 558 default: 559 return (EINVAL); 560 } 561 break; 562 case CRYPTO_AES_CBC: 563 switch (en->auth_algorithm) { 564 case CRYPTO_SHA1_HMAC: 565 break; 566 case CRYPTO_SHA2_256_HMAC: 567 case CRYPTO_SHA2_384_HMAC: 568 if (en->tls_vminor != TLS_MINOR_VER_TWO) 569 return (EINVAL); 570 break; 571 default: 572 return (EINVAL); 573 } 574 if (en->auth_key_len == 0) 575 return (EINVAL); 576 577 /* 578 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2 579 * use explicit IVs. 580 */ 581 switch (en->tls_vminor) { 582 case TLS_MINOR_VER_ZERO: 583 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 584 return (EINVAL); 585 break; 586 case TLS_MINOR_VER_ONE: 587 case TLS_MINOR_VER_TWO: 588 /* Ignore any supplied IV. */ 589 en->iv_len = 0; 590 break; 591 default: 592 return (EINVAL); 593 } 594 break; 595 case CRYPTO_CHACHA20_POLY1305: 596 if (en->auth_algorithm != 0 || en->auth_key_len != 0) 597 return (EINVAL); 598 if (en->tls_vminor != TLS_MINOR_VER_TWO && 599 en->tls_vminor != TLS_MINOR_VER_THREE) 600 return (EINVAL); 601 if (en->iv_len != TLS_CHACHA20_IV_LEN) 602 return (EINVAL); 603 break; 604 default: 605 return (EINVAL); 606 } 607 608 error = ktls_start_kthreads(); 609 if (error != 0) 610 return (error); 611 612 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 613 614 counter_u64_add(ktls_offload_active, 1); 615 616 refcount_init(&tls->refcount, 1); 617 if (direction == KTLS_RX) { 618 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls); 619 } else { 620 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 621 tls->inp = so->so_pcb; 622 in_pcbref(tls->inp); 623 tls->tx = true; 624 } 625 626 tls->wq_index = ktls_get_cpu(so); 627 628 tls->params.cipher_algorithm = en->cipher_algorithm; 629 tls->params.auth_algorithm = en->auth_algorithm; 630 tls->params.tls_vmajor = en->tls_vmajor; 631 tls->params.tls_vminor = en->tls_vminor; 632 tls->params.flags = en->flags; 633 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 634 635 /* Set the header and trailer lengths. */ 636 tls->params.tls_hlen = sizeof(struct tls_record_layer); 637 switch (en->cipher_algorithm) { 638 case CRYPTO_AES_NIST_GCM_16: 639 /* 640 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 641 * nonce. TLS 1.3 uses a 12 byte implicit IV. 642 */ 643 if (en->tls_vminor < TLS_MINOR_VER_THREE) 644 tls->params.tls_hlen += sizeof(uint64_t); 645 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 646 tls->params.tls_bs = 1; 647 break; 648 case CRYPTO_AES_CBC: 649 switch (en->auth_algorithm) { 650 case CRYPTO_SHA1_HMAC: 651 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 652 /* Implicit IV, no nonce. */ 653 tls->sequential_records = true; 654 tls->next_seqno = be64dec(en->rec_seq); 655 STAILQ_INIT(&tls->pending_records); 656 } else { 657 tls->params.tls_hlen += AES_BLOCK_LEN; 658 } 659 tls->params.tls_tlen = AES_BLOCK_LEN + 660 SHA1_HASH_LEN; 661 break; 662 case CRYPTO_SHA2_256_HMAC: 663 tls->params.tls_hlen += AES_BLOCK_LEN; 664 tls->params.tls_tlen = AES_BLOCK_LEN + 665 SHA2_256_HASH_LEN; 666 break; 667 case CRYPTO_SHA2_384_HMAC: 668 tls->params.tls_hlen += AES_BLOCK_LEN; 669 tls->params.tls_tlen = AES_BLOCK_LEN + 670 SHA2_384_HASH_LEN; 671 break; 672 default: 673 panic("invalid hmac"); 674 } 675 tls->params.tls_bs = AES_BLOCK_LEN; 676 break; 677 case CRYPTO_CHACHA20_POLY1305: 678 /* 679 * Chacha20 uses a 12 byte implicit IV. 680 */ 681 tls->params.tls_tlen = POLY1305_HASH_LEN; 682 tls->params.tls_bs = 1; 683 break; 684 default: 685 panic("invalid cipher"); 686 } 687 688 /* 689 * TLS 1.3 includes optional padding which we do not support, 690 * and also puts the "real" record type at the end of the 691 * encrypted data. 692 */ 693 if (en->tls_vminor == TLS_MINOR_VER_THREE) 694 tls->params.tls_tlen += sizeof(uint8_t); 695 696 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 697 ("TLS header length too long: %d", tls->params.tls_hlen)); 698 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 699 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 700 701 if (en->auth_key_len != 0) { 702 tls->params.auth_key_len = en->auth_key_len; 703 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 704 M_WAITOK); 705 error = copyin(en->auth_key, tls->params.auth_key, 706 en->auth_key_len); 707 if (error) 708 goto out; 709 } 710 711 tls->params.cipher_key_len = en->cipher_key_len; 712 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 713 error = copyin(en->cipher_key, tls->params.cipher_key, 714 en->cipher_key_len); 715 if (error) 716 goto out; 717 718 /* 719 * This holds the implicit portion of the nonce for AEAD 720 * ciphers and the initial implicit IV for TLS 1.0. The 721 * explicit portions of the IV are generated in ktls_frame(). 722 */ 723 if (en->iv_len != 0) { 724 tls->params.iv_len = en->iv_len; 725 error = copyin(en->iv, tls->params.iv, en->iv_len); 726 if (error) 727 goto out; 728 729 /* 730 * For TLS 1.2 with GCM, generate an 8-byte nonce as a 731 * counter to generate unique explicit IVs. 732 * 733 * Store this counter in the last 8 bytes of the IV 734 * array so that it is 8-byte aligned. 735 */ 736 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 737 en->tls_vminor == TLS_MINOR_VER_TWO) 738 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); 739 } 740 741 *tlsp = tls; 742 return (0); 743 744 out: 745 ktls_free(tls); 746 return (error); 747 } 748 749 static struct ktls_session * 750 ktls_clone_session(struct ktls_session *tls, int direction) 751 { 752 struct ktls_session *tls_new; 753 754 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 755 756 counter_u64_add(ktls_offload_active, 1); 757 758 refcount_init(&tls_new->refcount, 1); 759 if (direction == KTLS_RX) { 760 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag, 761 tls_new); 762 } else { 763 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, 764 tls_new); 765 tls_new->inp = tls->inp; 766 tls_new->tx = true; 767 in_pcbref(tls_new->inp); 768 } 769 770 /* Copy fields from existing session. */ 771 tls_new->params = tls->params; 772 tls_new->wq_index = tls->wq_index; 773 774 /* Deep copy keys. */ 775 if (tls_new->params.auth_key != NULL) { 776 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 777 M_KTLS, M_WAITOK); 778 memcpy(tls_new->params.auth_key, tls->params.auth_key, 779 tls->params.auth_key_len); 780 } 781 782 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 783 M_WAITOK); 784 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 785 tls->params.cipher_key_len); 786 787 return (tls_new); 788 } 789 790 #ifdef TCP_OFFLOAD 791 static int 792 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) 793 { 794 struct inpcb *inp; 795 struct tcpcb *tp; 796 int error; 797 798 inp = so->so_pcb; 799 INP_WLOCK(inp); 800 if (inp->inp_flags & INP_DROPPED) { 801 INP_WUNLOCK(inp); 802 return (ECONNRESET); 803 } 804 if (inp->inp_socket == NULL) { 805 INP_WUNLOCK(inp); 806 return (ECONNRESET); 807 } 808 tp = intotcpcb(inp); 809 if (!(tp->t_flags & TF_TOE)) { 810 INP_WUNLOCK(inp); 811 return (EOPNOTSUPP); 812 } 813 814 error = tcp_offload_alloc_tls_session(tp, tls, direction); 815 INP_WUNLOCK(inp); 816 if (error == 0) { 817 tls->mode = TCP_TLS_MODE_TOE; 818 switch (tls->params.cipher_algorithm) { 819 case CRYPTO_AES_CBC: 820 counter_u64_add(ktls_toe_cbc, 1); 821 break; 822 case CRYPTO_AES_NIST_GCM_16: 823 counter_u64_add(ktls_toe_gcm, 1); 824 break; 825 case CRYPTO_CHACHA20_POLY1305: 826 counter_u64_add(ktls_toe_chacha20, 1); 827 break; 828 } 829 } 830 return (error); 831 } 832 #endif 833 834 /* 835 * Common code used when first enabling ifnet TLS on a connection or 836 * when allocating a new ifnet TLS session due to a routing change. 837 * This function allocates a new TLS send tag on whatever interface 838 * the connection is currently routed over. 839 */ 840 static int 841 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 842 struct m_snd_tag **mstp) 843 { 844 union if_snd_tag_alloc_params params; 845 struct ifnet *ifp; 846 struct nhop_object *nh; 847 struct tcpcb *tp; 848 int error; 849 850 INP_RLOCK(inp); 851 if (inp->inp_flags & INP_DROPPED) { 852 INP_RUNLOCK(inp); 853 return (ECONNRESET); 854 } 855 if (inp->inp_socket == NULL) { 856 INP_RUNLOCK(inp); 857 return (ECONNRESET); 858 } 859 tp = intotcpcb(inp); 860 861 /* 862 * Check administrative controls on ifnet TLS to determine if 863 * ifnet TLS should be denied. 864 * 865 * - Always permit 'force' requests. 866 * - ktls_ifnet_permitted == 0: always deny. 867 */ 868 if (!force && ktls_ifnet_permitted == 0) { 869 INP_RUNLOCK(inp); 870 return (ENXIO); 871 } 872 873 /* 874 * XXX: Use the cached route in the inpcb to find the 875 * interface. This should perhaps instead use 876 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 877 * enabled after a connection has completed key negotiation in 878 * userland, the cached route will be present in practice. 879 */ 880 nh = inp->inp_route.ro_nh; 881 if (nh == NULL) { 882 INP_RUNLOCK(inp); 883 return (ENXIO); 884 } 885 ifp = nh->nh_ifp; 886 if_ref(ifp); 887 888 /* 889 * Allocate a TLS + ratelimit tag if the connection has an 890 * existing pacing rate. 891 */ 892 if (tp->t_pacing_rate != -1 && 893 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) { 894 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT; 895 params.tls_rate_limit.inp = inp; 896 params.tls_rate_limit.tls = tls; 897 params.tls_rate_limit.max_rate = tp->t_pacing_rate; 898 } else { 899 params.hdr.type = IF_SND_TAG_TYPE_TLS; 900 params.tls.inp = inp; 901 params.tls.tls = tls; 902 } 903 params.hdr.flowid = inp->inp_flowid; 904 params.hdr.flowtype = inp->inp_flowtype; 905 params.hdr.numa_domain = inp->inp_numa_domain; 906 INP_RUNLOCK(inp); 907 908 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) { 909 error = EOPNOTSUPP; 910 goto out; 911 } 912 if (inp->inp_vflag & INP_IPV6) { 913 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) { 914 error = EOPNOTSUPP; 915 goto out; 916 } 917 } else { 918 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) { 919 error = EOPNOTSUPP; 920 goto out; 921 } 922 } 923 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 924 out: 925 if_rele(ifp); 926 return (error); 927 } 928 929 /* 930 * Allocate an initial TLS receive tag for doing HW decryption of TLS 931 * data. 932 * 933 * This function allocates a new TLS receive tag on whatever interface 934 * the connection is currently routed over. If the connection ends up 935 * using a different interface for receive this will get fixed up via 936 * ktls_input_ifp_mismatch as future packets arrive. 937 */ 938 static int 939 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls, 940 struct m_snd_tag **mstp) 941 { 942 union if_snd_tag_alloc_params params; 943 struct ifnet *ifp; 944 struct nhop_object *nh; 945 int error; 946 947 if (!ktls_ocf_recrypt_supported(tls)) 948 return (ENXIO); 949 950 INP_RLOCK(inp); 951 if (inp->inp_flags & INP_DROPPED) { 952 INP_RUNLOCK(inp); 953 return (ECONNRESET); 954 } 955 if (inp->inp_socket == NULL) { 956 INP_RUNLOCK(inp); 957 return (ECONNRESET); 958 } 959 960 /* 961 * Check administrative controls on ifnet TLS to determine if 962 * ifnet TLS should be denied. 963 */ 964 if (ktls_ifnet_permitted == 0) { 965 INP_RUNLOCK(inp); 966 return (ENXIO); 967 } 968 969 /* 970 * XXX: As with ktls_alloc_snd_tag, use the cached route in 971 * the inpcb to find the interface. 972 */ 973 nh = inp->inp_route.ro_nh; 974 if (nh == NULL) { 975 INP_RUNLOCK(inp); 976 return (ENXIO); 977 } 978 ifp = nh->nh_ifp; 979 if_ref(ifp); 980 tls->rx_ifp = ifp; 981 982 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; 983 params.hdr.flowid = inp->inp_flowid; 984 params.hdr.flowtype = inp->inp_flowtype; 985 params.hdr.numa_domain = inp->inp_numa_domain; 986 params.tls_rx.inp = inp; 987 params.tls_rx.tls = tls; 988 params.tls_rx.vlan_id = 0; 989 990 INP_RUNLOCK(inp); 991 992 if (inp->inp_vflag & INP_IPV6) { 993 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) { 994 error = EOPNOTSUPP; 995 goto out; 996 } 997 } else { 998 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) { 999 error = EOPNOTSUPP; 1000 goto out; 1001 } 1002 } 1003 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 1004 1005 /* 1006 * If this connection is over a vlan, vlan_snd_tag_alloc 1007 * rewrites vlan_id with the saved interface. Save the VLAN 1008 * ID for use in ktls_reset_receive_tag which allocates new 1009 * receive tags directly from the leaf interface bypassing 1010 * if_vlan. 1011 */ 1012 if (error == 0) 1013 tls->rx_vlan_id = params.tls_rx.vlan_id; 1014 out: 1015 return (error); 1016 } 1017 1018 static int 1019 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction, 1020 bool force) 1021 { 1022 struct m_snd_tag *mst; 1023 int error; 1024 1025 switch (direction) { 1026 case KTLS_TX: 1027 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 1028 if (__predict_false(error != 0)) 1029 goto done; 1030 break; 1031 case KTLS_RX: 1032 KASSERT(!force, ("%s: forced receive tag", __func__)); 1033 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst); 1034 if (__predict_false(error != 0)) 1035 goto done; 1036 break; 1037 default: 1038 __assert_unreachable(); 1039 } 1040 1041 tls->mode = TCP_TLS_MODE_IFNET; 1042 tls->snd_tag = mst; 1043 1044 switch (tls->params.cipher_algorithm) { 1045 case CRYPTO_AES_CBC: 1046 counter_u64_add(ktls_ifnet_cbc, 1); 1047 break; 1048 case CRYPTO_AES_NIST_GCM_16: 1049 counter_u64_add(ktls_ifnet_gcm, 1); 1050 break; 1051 case CRYPTO_CHACHA20_POLY1305: 1052 counter_u64_add(ktls_ifnet_chacha20, 1); 1053 break; 1054 default: 1055 break; 1056 } 1057 done: 1058 return (error); 1059 } 1060 1061 static void 1062 ktls_use_sw(struct ktls_session *tls) 1063 { 1064 tls->mode = TCP_TLS_MODE_SW; 1065 switch (tls->params.cipher_algorithm) { 1066 case CRYPTO_AES_CBC: 1067 counter_u64_add(ktls_sw_cbc, 1); 1068 break; 1069 case CRYPTO_AES_NIST_GCM_16: 1070 counter_u64_add(ktls_sw_gcm, 1); 1071 break; 1072 case CRYPTO_CHACHA20_POLY1305: 1073 counter_u64_add(ktls_sw_chacha20, 1); 1074 break; 1075 } 1076 } 1077 1078 static int 1079 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction) 1080 { 1081 int error; 1082 1083 error = ktls_ocf_try(so, tls, direction); 1084 if (error) 1085 return (error); 1086 ktls_use_sw(tls); 1087 return (0); 1088 } 1089 1090 /* 1091 * KTLS RX stores data in the socket buffer as a list of TLS records, 1092 * where each record is stored as a control message containg the TLS 1093 * header followed by data mbufs containing the decrypted data. This 1094 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for 1095 * both encrypted and decrypted data. TLS records decrypted by a NIC 1096 * should be queued to the socket buffer as records, but encrypted 1097 * data which needs to be decrypted by software arrives as a stream of 1098 * regular mbufs which need to be converted. In addition, there may 1099 * already be pending encrypted data in the socket buffer when KTLS RX 1100 * is enabled. 1101 * 1102 * To manage not-yet-decrypted data for KTLS RX, the following scheme 1103 * is used: 1104 * 1105 * - A single chain of NOTREADY mbufs is hung off of sb_mtls. 1106 * 1107 * - ktls_check_rx checks this chain of mbufs reading the TLS header 1108 * from the first mbuf. Once all of the data for that TLS record is 1109 * queued, the socket is queued to a worker thread. 1110 * 1111 * - The worker thread calls ktls_decrypt to decrypt TLS records in 1112 * the TLS chain. Each TLS record is detached from the TLS chain, 1113 * decrypted, and inserted into the regular socket buffer chain as 1114 * record starting with a control message holding the TLS header and 1115 * a chain of mbufs holding the encrypted data. 1116 */ 1117 1118 static void 1119 sb_mark_notready(struct sockbuf *sb) 1120 { 1121 struct mbuf *m; 1122 1123 m = sb->sb_mb; 1124 sb->sb_mtls = m; 1125 sb->sb_mb = NULL; 1126 sb->sb_mbtail = NULL; 1127 sb->sb_lastrecord = NULL; 1128 for (; m != NULL; m = m->m_next) { 1129 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", 1130 __func__)); 1131 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", 1132 __func__)); 1133 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", 1134 __func__)); 1135 m->m_flags |= M_NOTREADY; 1136 sb->sb_acc -= m->m_len; 1137 sb->sb_tlscc += m->m_len; 1138 sb->sb_mtlstail = m; 1139 } 1140 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, 1141 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, 1142 sb->sb_ccc)); 1143 } 1144 1145 /* 1146 * Return information about the pending TLS data in a socket 1147 * buffer. On return, 'seqno' is set to the sequence number 1148 * of the next TLS record to be received, 'resid' is set to 1149 * the amount of bytes still needed for the last pending 1150 * record. The function returns 'false' if the last pending 1151 * record contains a partial TLS header. In that case, 'resid' 1152 * is the number of bytes needed to complete the TLS header. 1153 */ 1154 bool 1155 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp) 1156 { 1157 struct tls_record_layer hdr; 1158 struct mbuf *m; 1159 uint64_t seqno; 1160 size_t resid; 1161 u_int offset, record_len; 1162 1163 SOCKBUF_LOCK_ASSERT(sb); 1164 MPASS(sb->sb_flags & SB_TLS_RX); 1165 seqno = sb->sb_tls_seqno; 1166 resid = sb->sb_tlscc; 1167 m = sb->sb_mtls; 1168 offset = 0; 1169 1170 if (resid == 0) { 1171 *seqnop = seqno; 1172 *residp = 0; 1173 return (true); 1174 } 1175 1176 for (;;) { 1177 seqno++; 1178 1179 if (resid < sizeof(hdr)) { 1180 *seqnop = seqno; 1181 *residp = sizeof(hdr) - resid; 1182 return (false); 1183 } 1184 1185 m_copydata(m, offset, sizeof(hdr), (void *)&hdr); 1186 1187 record_len = sizeof(hdr) + ntohs(hdr.tls_length); 1188 if (resid <= record_len) { 1189 *seqnop = seqno; 1190 *residp = record_len - resid; 1191 return (true); 1192 } 1193 resid -= record_len; 1194 1195 while (record_len != 0) { 1196 if (m->m_len - offset > record_len) { 1197 offset += record_len; 1198 break; 1199 } 1200 1201 record_len -= (m->m_len - offset); 1202 offset = 0; 1203 m = m->m_next; 1204 } 1205 } 1206 } 1207 1208 int 1209 ktls_enable_rx(struct socket *so, struct tls_enable *en) 1210 { 1211 struct ktls_session *tls; 1212 int error; 1213 1214 if (!ktls_offload_enable) 1215 return (ENOTSUP); 1216 1217 counter_u64_add(ktls_offload_enable_calls, 1); 1218 1219 /* 1220 * This should always be true since only the TCP socket option 1221 * invokes this function. 1222 */ 1223 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1224 return (EINVAL); 1225 1226 /* 1227 * XXX: Don't overwrite existing sessions. We should permit 1228 * this to support rekeying in the future. 1229 */ 1230 if (so->so_rcv.sb_tls_info != NULL) 1231 return (EALREADY); 1232 1233 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1234 return (ENOTSUP); 1235 1236 error = ktls_create_session(so, en, &tls, KTLS_RX); 1237 if (error) 1238 return (error); 1239 1240 error = ktls_ocf_try(so, tls, KTLS_RX); 1241 if (error) { 1242 ktls_free(tls); 1243 return (error); 1244 } 1245 1246 /* Mark the socket as using TLS offload. */ 1247 SOCK_RECVBUF_LOCK(so); 1248 if (SOLISTENING(so)) { 1249 SOCK_RECVBUF_UNLOCK(so); 1250 ktls_free(tls); 1251 return (EINVAL); 1252 } 1253 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); 1254 so->so_rcv.sb_tls_info = tls; 1255 so->so_rcv.sb_flags |= SB_TLS_RX; 1256 1257 /* Mark existing data as not ready until it can be decrypted. */ 1258 sb_mark_notready(&so->so_rcv); 1259 ktls_check_rx(&so->so_rcv); 1260 SOCK_RECVBUF_UNLOCK(so); 1261 1262 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1263 #ifdef TCP_OFFLOAD 1264 error = ktls_try_toe(so, tls, KTLS_RX); 1265 if (error) 1266 #endif 1267 error = ktls_try_ifnet(so, tls, KTLS_RX, false); 1268 if (error) 1269 ktls_use_sw(tls); 1270 1271 counter_u64_add(ktls_offload_total, 1); 1272 1273 return (0); 1274 } 1275 1276 int 1277 ktls_enable_tx(struct socket *so, struct tls_enable *en) 1278 { 1279 struct ktls_session *tls; 1280 struct inpcb *inp; 1281 struct tcpcb *tp; 1282 int error; 1283 1284 if (!ktls_offload_enable) 1285 return (ENOTSUP); 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_snd.sb_tls_info != NULL) 1301 return (EALREADY); 1302 1303 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1304 return (ENOTSUP); 1305 1306 /* TLS requires ext pgs */ 1307 if (mb_use_ext_pgs == 0) 1308 return (ENXIO); 1309 1310 error = ktls_create_session(so, en, &tls, KTLS_TX); 1311 if (error) 1312 return (error); 1313 1314 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1315 #ifdef TCP_OFFLOAD 1316 error = ktls_try_toe(so, tls, KTLS_TX); 1317 if (error) 1318 #endif 1319 error = ktls_try_ifnet(so, tls, KTLS_TX, false); 1320 if (error) 1321 error = ktls_try_sw(so, tls, KTLS_TX); 1322 1323 if (error) { 1324 ktls_free(tls); 1325 return (error); 1326 } 1327 1328 /* 1329 * Serialize with sosend_generic() and make sure that we're not 1330 * operating on a listening socket. 1331 */ 1332 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1333 if (error) { 1334 ktls_free(tls); 1335 return (error); 1336 } 1337 1338 /* 1339 * Write lock the INP when setting sb_tls_info so that 1340 * routines in tcp_ratelimit.c can read sb_tls_info while 1341 * holding the INP lock. 1342 */ 1343 inp = so->so_pcb; 1344 INP_WLOCK(inp); 1345 SOCK_SENDBUF_LOCK(so); 1346 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); 1347 so->so_snd.sb_tls_info = tls; 1348 if (tls->mode != TCP_TLS_MODE_SW) { 1349 tp = intotcpcb(inp); 1350 MPASS(tp->t_nic_ktls_xmit == 0); 1351 tp->t_nic_ktls_xmit = 1; 1352 if (tp->t_fb->tfb_hwtls_change != NULL) 1353 (*tp->t_fb->tfb_hwtls_change)(tp, 1); 1354 } 1355 SOCK_SENDBUF_UNLOCK(so); 1356 INP_WUNLOCK(inp); 1357 SOCK_IO_SEND_UNLOCK(so); 1358 1359 counter_u64_add(ktls_offload_total, 1); 1360 1361 return (0); 1362 } 1363 1364 int 1365 ktls_get_rx_mode(struct socket *so, int *modep) 1366 { 1367 struct ktls_session *tls; 1368 struct inpcb *inp __diagused; 1369 1370 if (SOLISTENING(so)) 1371 return (EINVAL); 1372 inp = so->so_pcb; 1373 INP_WLOCK_ASSERT(inp); 1374 SOCK_RECVBUF_LOCK(so); 1375 tls = so->so_rcv.sb_tls_info; 1376 if (tls == NULL) 1377 *modep = TCP_TLS_MODE_NONE; 1378 else 1379 *modep = tls->mode; 1380 SOCK_RECVBUF_UNLOCK(so); 1381 return (0); 1382 } 1383 1384 /* 1385 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number. 1386 * 1387 * This function gets information about the next TCP- and TLS- 1388 * sequence number to be processed by the TLS receive worker 1389 * thread. The information is extracted from the given "inpcb" 1390 * structure. The values are stored in host endian format at the two 1391 * given output pointer locations. The TCP sequence number points to 1392 * the beginning of the TLS header. 1393 * 1394 * This function returns zero on success, else a non-zero error code 1395 * is returned. 1396 */ 1397 int 1398 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq) 1399 { 1400 struct socket *so; 1401 struct tcpcb *tp; 1402 1403 INP_RLOCK(inp); 1404 so = inp->inp_socket; 1405 if (__predict_false(so == NULL)) { 1406 INP_RUNLOCK(inp); 1407 return (EINVAL); 1408 } 1409 if (inp->inp_flags & INP_DROPPED) { 1410 INP_RUNLOCK(inp); 1411 return (ECONNRESET); 1412 } 1413 1414 tp = intotcpcb(inp); 1415 MPASS(tp != NULL); 1416 1417 SOCKBUF_LOCK(&so->so_rcv); 1418 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc; 1419 *tlsseq = so->so_rcv.sb_tls_seqno; 1420 SOCKBUF_UNLOCK(&so->so_rcv); 1421 1422 INP_RUNLOCK(inp); 1423 1424 return (0); 1425 } 1426 1427 int 1428 ktls_get_tx_mode(struct socket *so, int *modep) 1429 { 1430 struct ktls_session *tls; 1431 struct inpcb *inp __diagused; 1432 1433 if (SOLISTENING(so)) 1434 return (EINVAL); 1435 inp = so->so_pcb; 1436 INP_WLOCK_ASSERT(inp); 1437 SOCK_SENDBUF_LOCK(so); 1438 tls = so->so_snd.sb_tls_info; 1439 if (tls == NULL) 1440 *modep = TCP_TLS_MODE_NONE; 1441 else 1442 *modep = tls->mode; 1443 SOCK_SENDBUF_UNLOCK(so); 1444 return (0); 1445 } 1446 1447 /* 1448 * Switch between SW and ifnet TLS sessions as requested. 1449 */ 1450 int 1451 ktls_set_tx_mode(struct socket *so, int mode) 1452 { 1453 struct ktls_session *tls, *tls_new; 1454 struct inpcb *inp; 1455 struct tcpcb *tp; 1456 int error; 1457 1458 if (SOLISTENING(so)) 1459 return (EINVAL); 1460 switch (mode) { 1461 case TCP_TLS_MODE_SW: 1462 case TCP_TLS_MODE_IFNET: 1463 break; 1464 default: 1465 return (EINVAL); 1466 } 1467 1468 inp = so->so_pcb; 1469 INP_WLOCK_ASSERT(inp); 1470 tp = intotcpcb(inp); 1471 1472 if (mode == TCP_TLS_MODE_IFNET) { 1473 /* Don't allow enabling ifnet ktls multiple times */ 1474 if (tp->t_nic_ktls_xmit) 1475 return (EALREADY); 1476 1477 /* 1478 * Don't enable ifnet ktls if we disabled it due to an 1479 * excessive retransmission rate 1480 */ 1481 if (tp->t_nic_ktls_xmit_dis) 1482 return (ENXIO); 1483 } 1484 1485 SOCKBUF_LOCK(&so->so_snd); 1486 tls = so->so_snd.sb_tls_info; 1487 if (tls == NULL) { 1488 SOCKBUF_UNLOCK(&so->so_snd); 1489 return (0); 1490 } 1491 1492 if (tls->mode == mode) { 1493 SOCKBUF_UNLOCK(&so->so_snd); 1494 return (0); 1495 } 1496 1497 tls = ktls_hold(tls); 1498 SOCKBUF_UNLOCK(&so->so_snd); 1499 INP_WUNLOCK(inp); 1500 1501 tls_new = ktls_clone_session(tls, KTLS_TX); 1502 1503 if (mode == TCP_TLS_MODE_IFNET) 1504 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true); 1505 else 1506 error = ktls_try_sw(so, tls_new, KTLS_TX); 1507 if (error) { 1508 counter_u64_add(ktls_switch_failed, 1); 1509 ktls_free(tls_new); 1510 ktls_free(tls); 1511 INP_WLOCK(inp); 1512 return (error); 1513 } 1514 1515 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1516 if (error) { 1517 counter_u64_add(ktls_switch_failed, 1); 1518 ktls_free(tls_new); 1519 ktls_free(tls); 1520 INP_WLOCK(inp); 1521 return (error); 1522 } 1523 1524 /* 1525 * If we raced with another session change, keep the existing 1526 * session. 1527 */ 1528 if (tls != so->so_snd.sb_tls_info) { 1529 counter_u64_add(ktls_switch_failed, 1); 1530 SOCK_IO_SEND_UNLOCK(so); 1531 ktls_free(tls_new); 1532 ktls_free(tls); 1533 INP_WLOCK(inp); 1534 return (EBUSY); 1535 } 1536 1537 INP_WLOCK(inp); 1538 SOCKBUF_LOCK(&so->so_snd); 1539 so->so_snd.sb_tls_info = tls_new; 1540 if (tls_new->mode != TCP_TLS_MODE_SW) { 1541 MPASS(tp->t_nic_ktls_xmit == 0); 1542 tp->t_nic_ktls_xmit = 1; 1543 if (tp->t_fb->tfb_hwtls_change != NULL) 1544 (*tp->t_fb->tfb_hwtls_change)(tp, 1); 1545 } 1546 SOCKBUF_UNLOCK(&so->so_snd); 1547 SOCK_IO_SEND_UNLOCK(so); 1548 1549 /* 1550 * Drop two references on 'tls'. The first is for the 1551 * ktls_hold() above. The second drops the reference from the 1552 * socket buffer. 1553 */ 1554 KASSERT(tls->refcount >= 2, ("too few references on old session")); 1555 ktls_free(tls); 1556 ktls_free(tls); 1557 1558 if (mode == TCP_TLS_MODE_IFNET) 1559 counter_u64_add(ktls_switch_to_ifnet, 1); 1560 else 1561 counter_u64_add(ktls_switch_to_sw, 1); 1562 1563 return (0); 1564 } 1565 1566 /* 1567 * Try to allocate a new TLS receive tag. This task is scheduled when 1568 * sbappend_ktls_rx detects an input path change. If a new tag is 1569 * allocated, replace the tag in the TLS session. If a new tag cannot 1570 * be allocated, let the session fall back to software decryption. 1571 */ 1572 static void 1573 ktls_reset_receive_tag(void *context, int pending) 1574 { 1575 union if_snd_tag_alloc_params params; 1576 struct ktls_session *tls; 1577 struct m_snd_tag *mst; 1578 struct inpcb *inp; 1579 struct ifnet *ifp; 1580 struct socket *so; 1581 int error; 1582 1583 MPASS(pending == 1); 1584 1585 tls = context; 1586 so = tls->so; 1587 inp = so->so_pcb; 1588 ifp = NULL; 1589 1590 INP_RLOCK(inp); 1591 if (inp->inp_flags & INP_DROPPED) { 1592 INP_RUNLOCK(inp); 1593 goto out; 1594 } 1595 1596 SOCKBUF_LOCK(&so->so_rcv); 1597 mst = tls->snd_tag; 1598 tls->snd_tag = NULL; 1599 if (mst != NULL) 1600 m_snd_tag_rele(mst); 1601 1602 ifp = tls->rx_ifp; 1603 if_ref(ifp); 1604 SOCKBUF_UNLOCK(&so->so_rcv); 1605 1606 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; 1607 params.hdr.flowid = inp->inp_flowid; 1608 params.hdr.flowtype = inp->inp_flowtype; 1609 params.hdr.numa_domain = inp->inp_numa_domain; 1610 params.tls_rx.inp = inp; 1611 params.tls_rx.tls = tls; 1612 params.tls_rx.vlan_id = tls->rx_vlan_id; 1613 INP_RUNLOCK(inp); 1614 1615 if (inp->inp_vflag & INP_IPV6) { 1616 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0) 1617 goto out; 1618 } else { 1619 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0) 1620 goto out; 1621 } 1622 1623 error = m_snd_tag_alloc(ifp, ¶ms, &mst); 1624 if (error == 0) { 1625 SOCKBUF_LOCK(&so->so_rcv); 1626 tls->snd_tag = mst; 1627 SOCKBUF_UNLOCK(&so->so_rcv); 1628 1629 counter_u64_add(ktls_ifnet_reset, 1); 1630 } else { 1631 /* 1632 * Just fall back to software decryption if a tag 1633 * cannot be allocated leaving the connection intact. 1634 * If a future input path change switches to another 1635 * interface this connection will resume ifnet TLS. 1636 */ 1637 counter_u64_add(ktls_ifnet_reset_failed, 1); 1638 } 1639 1640 out: 1641 mtx_pool_lock(mtxpool_sleep, tls); 1642 tls->reset_pending = false; 1643 mtx_pool_unlock(mtxpool_sleep, tls); 1644 1645 if (ifp != NULL) 1646 if_rele(ifp); 1647 CURVNET_SET(so->so_vnet); 1648 sorele(so); 1649 CURVNET_RESTORE(); 1650 ktls_free(tls); 1651 } 1652 1653 /* 1654 * Try to allocate a new TLS send tag. This task is scheduled when 1655 * ip_output detects a route change while trying to transmit a packet 1656 * holding a TLS record. If a new tag is allocated, replace the tag 1657 * in the TLS session. Subsequent packets on the connection will use 1658 * the new tag. If a new tag cannot be allocated, drop the 1659 * connection. 1660 */ 1661 static void 1662 ktls_reset_send_tag(void *context, int pending) 1663 { 1664 struct epoch_tracker et; 1665 struct ktls_session *tls; 1666 struct m_snd_tag *old, *new; 1667 struct inpcb *inp; 1668 struct tcpcb *tp; 1669 int error; 1670 1671 MPASS(pending == 1); 1672 1673 tls = context; 1674 inp = tls->inp; 1675 1676 /* 1677 * Free the old tag first before allocating a new one. 1678 * ip[6]_output_send() will treat a NULL send tag the same as 1679 * an ifp mismatch and drop packets until a new tag is 1680 * allocated. 1681 * 1682 * Write-lock the INP when changing tls->snd_tag since 1683 * ip[6]_output_send() holds a read-lock when reading the 1684 * pointer. 1685 */ 1686 INP_WLOCK(inp); 1687 old = tls->snd_tag; 1688 tls->snd_tag = NULL; 1689 INP_WUNLOCK(inp); 1690 if (old != NULL) 1691 m_snd_tag_rele(old); 1692 1693 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1694 1695 if (error == 0) { 1696 INP_WLOCK(inp); 1697 tls->snd_tag = new; 1698 mtx_pool_lock(mtxpool_sleep, tls); 1699 tls->reset_pending = false; 1700 mtx_pool_unlock(mtxpool_sleep, tls); 1701 INP_WUNLOCK(inp); 1702 1703 counter_u64_add(ktls_ifnet_reset, 1); 1704 1705 /* 1706 * XXX: Should we kick tcp_output explicitly now that 1707 * the send tag is fixed or just rely on timers? 1708 */ 1709 } else { 1710 NET_EPOCH_ENTER(et); 1711 INP_WLOCK(inp); 1712 if (!(inp->inp_flags & INP_DROPPED)) { 1713 tp = intotcpcb(inp); 1714 CURVNET_SET(inp->inp_vnet); 1715 tp = tcp_drop(tp, ECONNABORTED); 1716 CURVNET_RESTORE(); 1717 if (tp != NULL) { 1718 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1719 INP_WUNLOCK(inp); 1720 } 1721 } else 1722 INP_WUNLOCK(inp); 1723 NET_EPOCH_EXIT(et); 1724 1725 counter_u64_add(ktls_ifnet_reset_failed, 1); 1726 1727 /* 1728 * Leave reset_pending true to avoid future tasks while 1729 * the socket goes away. 1730 */ 1731 } 1732 1733 ktls_free(tls); 1734 } 1735 1736 void 1737 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp) 1738 { 1739 struct ktls_session *tls; 1740 struct socket *so; 1741 1742 SOCKBUF_LOCK_ASSERT(sb); 1743 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 1744 __func__, sb)); 1745 so = __containerof(sb, struct socket, so_rcv); 1746 1747 tls = sb->sb_tls_info; 1748 if_rele(tls->rx_ifp); 1749 if_ref(ifp); 1750 tls->rx_ifp = ifp; 1751 1752 /* 1753 * See if we should schedule a task to update the receive tag for 1754 * this session. 1755 */ 1756 mtx_pool_lock(mtxpool_sleep, tls); 1757 if (!tls->reset_pending) { 1758 (void) ktls_hold(tls); 1759 soref(so); 1760 tls->so = so; 1761 tls->reset_pending = true; 1762 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1763 } 1764 mtx_pool_unlock(mtxpool_sleep, tls); 1765 } 1766 1767 int 1768 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1769 { 1770 1771 if (inp == NULL) 1772 return (ENOBUFS); 1773 1774 INP_LOCK_ASSERT(inp); 1775 1776 /* 1777 * See if we should schedule a task to update the send tag for 1778 * this session. 1779 */ 1780 mtx_pool_lock(mtxpool_sleep, tls); 1781 if (!tls->reset_pending) { 1782 (void) ktls_hold(tls); 1783 tls->reset_pending = true; 1784 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1785 } 1786 mtx_pool_unlock(mtxpool_sleep, tls); 1787 return (ENOBUFS); 1788 } 1789 1790 #ifdef RATELIMIT 1791 int 1792 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate) 1793 { 1794 union if_snd_tag_modify_params params = { 1795 .rate_limit.max_rate = max_pacing_rate, 1796 .rate_limit.flags = M_NOWAIT, 1797 }; 1798 struct m_snd_tag *mst; 1799 1800 /* Can't get to the inp, but it should be locked. */ 1801 /* INP_LOCK_ASSERT(inp); */ 1802 1803 MPASS(tls->mode == TCP_TLS_MODE_IFNET); 1804 1805 if (tls->snd_tag == NULL) { 1806 /* 1807 * Resetting send tag, ignore this change. The 1808 * pending reset may or may not see this updated rate 1809 * in the tcpcb. If it doesn't, we will just lose 1810 * this rate change. 1811 */ 1812 return (0); 1813 } 1814 1815 mst = tls->snd_tag; 1816 1817 MPASS(mst != NULL); 1818 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT); 1819 1820 return (mst->sw->snd_tag_modify(mst, ¶ms)); 1821 } 1822 #endif 1823 1824 static void 1825 ktls_destroy_help(void *context, int pending __unused) 1826 { 1827 ktls_destroy(context); 1828 } 1829 1830 void 1831 ktls_destroy(struct ktls_session *tls) 1832 { 1833 struct inpcb *inp; 1834 struct tcpcb *tp; 1835 bool wlocked; 1836 1837 MPASS(tls->refcount == 0); 1838 1839 inp = tls->inp; 1840 if (tls->tx) { 1841 wlocked = INP_WLOCKED(inp); 1842 if (!wlocked && !INP_TRY_WLOCK(inp)) { 1843 /* 1844 * rwlocks read locks are anonymous, and there 1845 * is no way to know if our current thread 1846 * holds an rlock on the inp. As a rough 1847 * estimate, check to see if the thread holds 1848 * *any* rlocks at all. If it does not, then we 1849 * know that we don't hold the inp rlock, and 1850 * can safely take the wlock 1851 */ 1852 if (curthread->td_rw_rlocks == 0) { 1853 INP_WLOCK(inp); 1854 } else { 1855 /* 1856 * We might hold the rlock, so let's 1857 * do the destroy in a taskqueue 1858 * context to avoid a potential 1859 * deadlock. This should be very 1860 * rare. 1861 */ 1862 counter_u64_add(ktls_destroy_task, 1); 1863 TASK_INIT(&tls->destroy_task, 0, 1864 ktls_destroy_help, tls); 1865 (void)taskqueue_enqueue(taskqueue_thread, 1866 &tls->destroy_task); 1867 return; 1868 } 1869 } 1870 } 1871 1872 if (tls->sequential_records) { 1873 struct mbuf *m, *n; 1874 int page_count; 1875 1876 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) { 1877 page_count = m->m_epg_enc_cnt; 1878 while (page_count > 0) { 1879 KASSERT(page_count >= m->m_epg_nrdy, 1880 ("%s: too few pages", __func__)); 1881 page_count -= m->m_epg_nrdy; 1882 m = m_free(m); 1883 } 1884 } 1885 } 1886 1887 counter_u64_add(ktls_offload_active, -1); 1888 switch (tls->mode) { 1889 case TCP_TLS_MODE_SW: 1890 switch (tls->params.cipher_algorithm) { 1891 case CRYPTO_AES_CBC: 1892 counter_u64_add(ktls_sw_cbc, -1); 1893 break; 1894 case CRYPTO_AES_NIST_GCM_16: 1895 counter_u64_add(ktls_sw_gcm, -1); 1896 break; 1897 case CRYPTO_CHACHA20_POLY1305: 1898 counter_u64_add(ktls_sw_chacha20, -1); 1899 break; 1900 } 1901 break; 1902 case TCP_TLS_MODE_IFNET: 1903 switch (tls->params.cipher_algorithm) { 1904 case CRYPTO_AES_CBC: 1905 counter_u64_add(ktls_ifnet_cbc, -1); 1906 break; 1907 case CRYPTO_AES_NIST_GCM_16: 1908 counter_u64_add(ktls_ifnet_gcm, -1); 1909 break; 1910 case CRYPTO_CHACHA20_POLY1305: 1911 counter_u64_add(ktls_ifnet_chacha20, -1); 1912 break; 1913 } 1914 if (tls->snd_tag != NULL) 1915 m_snd_tag_rele(tls->snd_tag); 1916 if (tls->rx_ifp != NULL) 1917 if_rele(tls->rx_ifp); 1918 if (tls->tx) { 1919 INP_WLOCK_ASSERT(inp); 1920 tp = intotcpcb(inp); 1921 MPASS(tp->t_nic_ktls_xmit == 1); 1922 tp->t_nic_ktls_xmit = 0; 1923 } 1924 break; 1925 #ifdef TCP_OFFLOAD 1926 case TCP_TLS_MODE_TOE: 1927 switch (tls->params.cipher_algorithm) { 1928 case CRYPTO_AES_CBC: 1929 counter_u64_add(ktls_toe_cbc, -1); 1930 break; 1931 case CRYPTO_AES_NIST_GCM_16: 1932 counter_u64_add(ktls_toe_gcm, -1); 1933 break; 1934 case CRYPTO_CHACHA20_POLY1305: 1935 counter_u64_add(ktls_toe_chacha20, -1); 1936 break; 1937 } 1938 break; 1939 #endif 1940 } 1941 if (tls->ocf_session != NULL) 1942 ktls_ocf_free(tls); 1943 if (tls->params.auth_key != NULL) { 1944 zfree(tls->params.auth_key, M_KTLS); 1945 tls->params.auth_key = NULL; 1946 tls->params.auth_key_len = 0; 1947 } 1948 if (tls->params.cipher_key != NULL) { 1949 zfree(tls->params.cipher_key, M_KTLS); 1950 tls->params.cipher_key = NULL; 1951 tls->params.cipher_key_len = 0; 1952 } 1953 if (tls->tx) { 1954 INP_WLOCK_ASSERT(inp); 1955 if (!in_pcbrele_wlocked(inp) && !wlocked) 1956 INP_WUNLOCK(inp); 1957 } 1958 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 1959 1960 uma_zfree(ktls_session_zone, tls); 1961 } 1962 1963 void 1964 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1965 { 1966 1967 for (; m != NULL; m = m->m_next) { 1968 KASSERT((m->m_flags & M_EXTPG) != 0, 1969 ("ktls_seq: mapped mbuf %p", m)); 1970 1971 m->m_epg_seqno = sb->sb_tls_seqno; 1972 sb->sb_tls_seqno++; 1973 } 1974 } 1975 1976 /* 1977 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1978 * mbuf in the chain must be an unmapped mbuf. The payload of the 1979 * mbuf must be populated with the payload of each TLS record. 1980 * 1981 * The record_type argument specifies the TLS record type used when 1982 * populating the TLS header. 1983 * 1984 * The enq_count argument on return is set to the number of pages of 1985 * payload data for this entire chain that need to be encrypted via SW 1986 * encryption. The returned value should be passed to ktls_enqueue 1987 * when scheduling encryption of this chain of mbufs. To handle the 1988 * special case of empty fragments for TLS 1.0 sessions, an empty 1989 * fragment counts as one page. 1990 */ 1991 void 1992 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1993 uint8_t record_type) 1994 { 1995 struct tls_record_layer *tlshdr; 1996 struct mbuf *m; 1997 uint64_t *noncep; 1998 uint16_t tls_len; 1999 int maxlen __diagused; 2000 2001 maxlen = tls->params.max_frame_len; 2002 *enq_cnt = 0; 2003 for (m = top; m != NULL; m = m->m_next) { 2004 /* 2005 * All mbufs in the chain should be TLS records whose 2006 * payload does not exceed the maximum frame length. 2007 * 2008 * Empty TLS 1.0 records are permitted when using CBC. 2009 */ 2010 KASSERT(m->m_len <= maxlen && m->m_len >= 0 && 2011 (m->m_len > 0 || ktls_permit_empty_frames(tls)), 2012 ("ktls_frame: m %p len %d", m, m->m_len)); 2013 2014 /* 2015 * TLS frames require unmapped mbufs to store session 2016 * info. 2017 */ 2018 KASSERT((m->m_flags & M_EXTPG) != 0, 2019 ("ktls_frame: mapped mbuf %p (top = %p)", m, top)); 2020 2021 tls_len = m->m_len; 2022 2023 /* Save a reference to the session. */ 2024 m->m_epg_tls = ktls_hold(tls); 2025 2026 m->m_epg_hdrlen = tls->params.tls_hlen; 2027 m->m_epg_trllen = tls->params.tls_tlen; 2028 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 2029 int bs, delta; 2030 2031 /* 2032 * AES-CBC pads messages to a multiple of the 2033 * block size. Note that the padding is 2034 * applied after the digest and the encryption 2035 * is done on the "plaintext || mac || padding". 2036 * At least one byte of padding is always 2037 * present. 2038 * 2039 * Compute the final trailer length assuming 2040 * at most one block of padding. 2041 * tls->params.tls_tlen is the maximum 2042 * possible trailer length (padding + digest). 2043 * delta holds the number of excess padding 2044 * bytes if the maximum were used. Those 2045 * extra bytes are removed. 2046 */ 2047 bs = tls->params.tls_bs; 2048 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 2049 m->m_epg_trllen -= delta; 2050 } 2051 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; 2052 2053 /* Populate the TLS header. */ 2054 tlshdr = (void *)m->m_epg_hdr; 2055 tlshdr->tls_vmajor = tls->params.tls_vmajor; 2056 2057 /* 2058 * TLS 1.3 masquarades as TLS 1.2 with a record type 2059 * of TLS_RLTYPE_APP. 2060 */ 2061 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 2062 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 2063 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 2064 tlshdr->tls_type = TLS_RLTYPE_APP; 2065 /* save the real record type for later */ 2066 m->m_epg_record_type = record_type; 2067 m->m_epg_trail[0] = record_type; 2068 } else { 2069 tlshdr->tls_vminor = tls->params.tls_vminor; 2070 tlshdr->tls_type = record_type; 2071 } 2072 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 2073 2074 /* 2075 * Store nonces / explicit IVs after the end of the 2076 * TLS header. 2077 * 2078 * For GCM with TLS 1.2, an 8 byte nonce is copied 2079 * from the end of the IV. The nonce is then 2080 * incremented for use by the next record. 2081 * 2082 * For CBC, a random nonce is inserted for TLS 1.1+. 2083 */ 2084 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 2085 tls->params.tls_vminor == TLS_MINOR_VER_TWO) { 2086 noncep = (uint64_t *)(tls->params.iv + 8); 2087 be64enc(tlshdr + 1, *noncep); 2088 (*noncep)++; 2089 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 2090 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 2091 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 2092 2093 /* 2094 * When using SW encryption, mark the mbuf not ready. 2095 * It will be marked ready via sbready() after the 2096 * record has been encrypted. 2097 * 2098 * When using ifnet TLS, unencrypted TLS records are 2099 * sent down the stack to the NIC. 2100 */ 2101 if (tls->mode == TCP_TLS_MODE_SW) { 2102 m->m_flags |= M_NOTREADY; 2103 if (__predict_false(tls_len == 0)) { 2104 /* TLS 1.0 empty fragment. */ 2105 m->m_epg_nrdy = 1; 2106 } else 2107 m->m_epg_nrdy = m->m_epg_npgs; 2108 *enq_cnt += m->m_epg_nrdy; 2109 } 2110 } 2111 } 2112 2113 bool 2114 ktls_permit_empty_frames(struct ktls_session *tls) 2115 { 2116 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 2117 tls->params.tls_vminor == TLS_MINOR_VER_ZERO); 2118 } 2119 2120 void 2121 ktls_check_rx(struct sockbuf *sb) 2122 { 2123 struct tls_record_layer hdr; 2124 struct ktls_wq *wq; 2125 struct socket *so; 2126 bool running; 2127 2128 SOCKBUF_LOCK_ASSERT(sb); 2129 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 2130 __func__, sb)); 2131 so = __containerof(sb, struct socket, so_rcv); 2132 2133 if (sb->sb_flags & SB_TLS_RX_RUNNING) 2134 return; 2135 2136 /* Is there enough queued for a TLS header? */ 2137 if (sb->sb_tlscc < sizeof(hdr)) { 2138 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) 2139 so->so_error = EMSGSIZE; 2140 return; 2141 } 2142 2143 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); 2144 2145 /* Is the entire record queued? */ 2146 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { 2147 if ((sb->sb_state & SBS_CANTRCVMORE) != 0) 2148 so->so_error = EMSGSIZE; 2149 return; 2150 } 2151 2152 sb->sb_flags |= SB_TLS_RX_RUNNING; 2153 2154 soref(so); 2155 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; 2156 mtx_lock(&wq->mtx); 2157 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); 2158 running = wq->running; 2159 mtx_unlock(&wq->mtx); 2160 if (!running) 2161 wakeup(wq); 2162 counter_u64_add(ktls_cnt_rx_queued, 1); 2163 } 2164 2165 static struct mbuf * 2166 ktls_detach_record(struct sockbuf *sb, int len) 2167 { 2168 struct mbuf *m, *n, *top; 2169 int remain; 2170 2171 SOCKBUF_LOCK_ASSERT(sb); 2172 MPASS(len <= sb->sb_tlscc); 2173 2174 /* 2175 * If TLS chain is the exact size of the record, 2176 * just grab the whole record. 2177 */ 2178 top = sb->sb_mtls; 2179 if (sb->sb_tlscc == len) { 2180 sb->sb_mtls = NULL; 2181 sb->sb_mtlstail = NULL; 2182 goto out; 2183 } 2184 2185 /* 2186 * While it would be nice to use m_split() here, we need 2187 * to know exactly what m_split() allocates to update the 2188 * accounting, so do it inline instead. 2189 */ 2190 remain = len; 2191 for (m = top; remain > m->m_len; m = m->m_next) 2192 remain -= m->m_len; 2193 2194 /* Easy case: don't have to split 'm'. */ 2195 if (remain == m->m_len) { 2196 sb->sb_mtls = m->m_next; 2197 if (sb->sb_mtls == NULL) 2198 sb->sb_mtlstail = NULL; 2199 m->m_next = NULL; 2200 goto out; 2201 } 2202 2203 /* 2204 * Need to allocate an mbuf to hold the remainder of 'm'. Try 2205 * with M_NOWAIT first. 2206 */ 2207 n = m_get(M_NOWAIT, MT_DATA); 2208 if (n == NULL) { 2209 /* 2210 * Use M_WAITOK with socket buffer unlocked. If 2211 * 'sb_mtls' changes while the lock is dropped, return 2212 * NULL to force the caller to retry. 2213 */ 2214 SOCKBUF_UNLOCK(sb); 2215 2216 n = m_get(M_WAITOK, MT_DATA); 2217 2218 SOCKBUF_LOCK(sb); 2219 if (sb->sb_mtls != top) { 2220 m_free(n); 2221 return (NULL); 2222 } 2223 } 2224 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED)); 2225 2226 /* Store remainder in 'n'. */ 2227 n->m_len = m->m_len - remain; 2228 if (m->m_flags & M_EXT) { 2229 n->m_data = m->m_data + remain; 2230 mb_dupcl(n, m); 2231 } else { 2232 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); 2233 } 2234 2235 /* Trim 'm' and update accounting. */ 2236 m->m_len -= n->m_len; 2237 sb->sb_tlscc -= n->m_len; 2238 sb->sb_ccc -= n->m_len; 2239 2240 /* Account for 'n'. */ 2241 sballoc_ktls_rx(sb, n); 2242 2243 /* Insert 'n' into the TLS chain. */ 2244 sb->sb_mtls = n; 2245 n->m_next = m->m_next; 2246 if (sb->sb_mtlstail == m) 2247 sb->sb_mtlstail = n; 2248 2249 /* Detach the record from the TLS chain. */ 2250 m->m_next = NULL; 2251 2252 out: 2253 MPASS(m_length(top, NULL) == len); 2254 for (m = top; m != NULL; m = m->m_next) 2255 sbfree_ktls_rx(sb, m); 2256 sb->sb_tlsdcc = len; 2257 sb->sb_ccc += len; 2258 SBCHECK(sb); 2259 return (top); 2260 } 2261 2262 /* 2263 * Determine the length of the trailing zero padding and find the real 2264 * record type in the byte before the padding. 2265 * 2266 * Walking the mbuf chain backwards is clumsy, so another option would 2267 * be to scan forwards remembering the last non-zero byte before the 2268 * trailer. However, it would be expensive to scan the entire record. 2269 * Instead, find the last non-zero byte of each mbuf in the chain 2270 * keeping track of the relative offset of that nonzero byte. 2271 * 2272 * trail_len is the size of the MAC/tag on input and is set to the 2273 * size of the full trailer including padding and the record type on 2274 * return. 2275 */ 2276 static int 2277 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, 2278 int *trailer_len, uint8_t *record_typep) 2279 { 2280 char *cp; 2281 u_int digest_start, last_offset, m_len, offset; 2282 uint8_t record_type; 2283 2284 digest_start = tls_len - *trailer_len; 2285 last_offset = 0; 2286 offset = 0; 2287 for (; m != NULL && offset < digest_start; 2288 offset += m->m_len, m = m->m_next) { 2289 /* Don't look for padding in the tag. */ 2290 m_len = min(digest_start - offset, m->m_len); 2291 cp = mtod(m, char *); 2292 2293 /* Find last non-zero byte in this mbuf. */ 2294 while (m_len > 0 && cp[m_len - 1] == 0) 2295 m_len--; 2296 if (m_len > 0) { 2297 record_type = cp[m_len - 1]; 2298 last_offset = offset + m_len; 2299 } 2300 } 2301 if (last_offset < tls->params.tls_hlen) 2302 return (EBADMSG); 2303 2304 *record_typep = record_type; 2305 *trailer_len = tls_len - last_offset + 1; 2306 return (0); 2307 } 2308 2309 /* 2310 * Check if a mbuf chain is fully decrypted at the given offset and 2311 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is 2312 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted 2313 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data 2314 * is encrypted. 2315 */ 2316 ktls_mbuf_crypto_st_t 2317 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len) 2318 { 2319 int m_flags_ored = 0; 2320 int m_flags_anded = -1; 2321 2322 for (; mb != NULL; mb = mb->m_next) { 2323 if (offset < mb->m_len) 2324 break; 2325 offset -= mb->m_len; 2326 } 2327 offset += len; 2328 2329 for (; mb != NULL; mb = mb->m_next) { 2330 m_flags_ored |= mb->m_flags; 2331 m_flags_anded &= mb->m_flags; 2332 2333 if (offset <= mb->m_len) 2334 break; 2335 offset -= mb->m_len; 2336 } 2337 MPASS(mb != NULL || offset == 0); 2338 2339 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED) 2340 return (KTLS_MBUF_CRYPTO_ST_MIXED); 2341 else 2342 return ((m_flags_ored & M_DECRYPTED) ? 2343 KTLS_MBUF_CRYPTO_ST_DECRYPTED : 2344 KTLS_MBUF_CRYPTO_ST_ENCRYPTED); 2345 } 2346 2347 /* 2348 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss 2349 */ 2350 static int 2351 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num) 2352 { 2353 union if_snd_tag_modify_params params; 2354 struct m_snd_tag *mst; 2355 struct inpcb *inp; 2356 struct tcpcb *tp; 2357 2358 mst = so->so_rcv.sb_tls_info->snd_tag; 2359 if (__predict_false(mst == NULL)) 2360 return (EINVAL); 2361 2362 inp = sotoinpcb(so); 2363 if (__predict_false(inp == NULL)) 2364 return (EINVAL); 2365 2366 INP_RLOCK(inp); 2367 if (inp->inp_flags & INP_DROPPED) { 2368 INP_RUNLOCK(inp); 2369 return (ECONNRESET); 2370 } 2371 2372 tp = intotcpcb(inp); 2373 MPASS(tp != NULL); 2374 2375 /* Get the TCP sequence number of the next valid TLS header. */ 2376 SOCKBUF_LOCK(&so->so_rcv); 2377 params.tls_rx.tls_hdr_tcp_sn = 2378 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len; 2379 params.tls_rx.tls_rec_length = tls_len; 2380 params.tls_rx.tls_seq_number = tls_rcd_num; 2381 SOCKBUF_UNLOCK(&so->so_rcv); 2382 2383 INP_RUNLOCK(inp); 2384 2385 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX); 2386 return (mst->sw->snd_tag_modify(mst, ¶ms)); 2387 } 2388 2389 static void 2390 ktls_drop(struct socket *so, int error) 2391 { 2392 struct epoch_tracker et; 2393 struct inpcb *inp = sotoinpcb(so); 2394 struct tcpcb *tp; 2395 2396 NET_EPOCH_ENTER(et); 2397 INP_WLOCK(inp); 2398 if (!(inp->inp_flags & INP_DROPPED)) { 2399 tp = intotcpcb(inp); 2400 CURVNET_SET(inp->inp_vnet); 2401 tp = tcp_drop(tp, error); 2402 CURVNET_RESTORE(); 2403 if (tp != NULL) 2404 INP_WUNLOCK(inp); 2405 } else { 2406 so->so_error = error; 2407 SOCK_RECVBUF_LOCK(so); 2408 sorwakeup_locked(so); 2409 INP_WUNLOCK(inp); 2410 } 2411 NET_EPOCH_EXIT(et); 2412 } 2413 2414 static void 2415 ktls_decrypt(struct socket *so) 2416 { 2417 char tls_header[MBUF_PEXT_HDR_LEN]; 2418 struct ktls_session *tls; 2419 struct sockbuf *sb; 2420 struct tls_record_layer *hdr; 2421 struct tls_get_record tgr; 2422 struct mbuf *control, *data, *m; 2423 ktls_mbuf_crypto_st_t state; 2424 uint64_t seqno; 2425 int error, remain, tls_len, trail_len; 2426 bool tls13; 2427 uint8_t vminor, record_type; 2428 2429 hdr = (struct tls_record_layer *)tls_header; 2430 sb = &so->so_rcv; 2431 SOCKBUF_LOCK(sb); 2432 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, 2433 ("%s: socket %p not running", __func__, so)); 2434 2435 tls = sb->sb_tls_info; 2436 MPASS(tls != NULL); 2437 2438 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE); 2439 if (tls13) 2440 vminor = TLS_MINOR_VER_TWO; 2441 else 2442 vminor = tls->params.tls_vminor; 2443 for (;;) { 2444 /* Is there enough queued for a TLS header? */ 2445 if (sb->sb_tlscc < tls->params.tls_hlen) 2446 break; 2447 2448 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); 2449 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); 2450 2451 if (hdr->tls_vmajor != tls->params.tls_vmajor || 2452 hdr->tls_vminor != vminor) 2453 error = EINVAL; 2454 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP) 2455 error = EINVAL; 2456 else if (tls_len < tls->params.tls_hlen || tls_len > 2457 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + 2458 tls->params.tls_tlen) 2459 error = EMSGSIZE; 2460 else 2461 error = 0; 2462 if (__predict_false(error != 0)) { 2463 /* 2464 * We have a corrupted record and are likely 2465 * out of sync. The connection isn't 2466 * recoverable at this point, so abort it. 2467 */ 2468 SOCKBUF_UNLOCK(sb); 2469 counter_u64_add(ktls_offload_corrupted_records, 1); 2470 2471 ktls_drop(so, error); 2472 goto deref; 2473 } 2474 2475 /* Is the entire record queued? */ 2476 if (sb->sb_tlscc < tls_len) 2477 break; 2478 2479 /* 2480 * Split out the portion of the mbuf chain containing 2481 * this TLS record. 2482 */ 2483 data = ktls_detach_record(sb, tls_len); 2484 if (data == NULL) 2485 continue; 2486 MPASS(sb->sb_tlsdcc == tls_len); 2487 2488 seqno = sb->sb_tls_seqno; 2489 sb->sb_tls_seqno++; 2490 SBCHECK(sb); 2491 SOCKBUF_UNLOCK(sb); 2492 2493 /* get crypto state for this TLS record */ 2494 state = ktls_mbuf_crypto_state(data, 0, tls_len); 2495 2496 switch (state) { 2497 case KTLS_MBUF_CRYPTO_ST_MIXED: 2498 error = ktls_ocf_recrypt(tls, hdr, data, seqno); 2499 if (error) 2500 break; 2501 /* FALLTHROUGH */ 2502 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED: 2503 error = ktls_ocf_decrypt(tls, hdr, data, seqno, 2504 &trail_len); 2505 if (__predict_true(error == 0)) { 2506 if (tls13) { 2507 error = tls13_find_record_type(tls, data, 2508 tls_len, &trail_len, &record_type); 2509 } else { 2510 record_type = hdr->tls_type; 2511 } 2512 } 2513 break; 2514 case KTLS_MBUF_CRYPTO_ST_DECRYPTED: 2515 /* 2516 * NIC TLS is only supported for AEAD 2517 * ciphersuites which used a fixed sized 2518 * trailer. 2519 */ 2520 if (tls13) { 2521 trail_len = tls->params.tls_tlen - 1; 2522 error = tls13_find_record_type(tls, data, 2523 tls_len, &trail_len, &record_type); 2524 } else { 2525 trail_len = tls->params.tls_tlen; 2526 error = 0; 2527 record_type = hdr->tls_type; 2528 } 2529 break; 2530 default: 2531 error = EINVAL; 2532 break; 2533 } 2534 if (error) { 2535 counter_u64_add(ktls_offload_failed_crypto, 1); 2536 2537 SOCKBUF_LOCK(sb); 2538 if (sb->sb_tlsdcc == 0) { 2539 /* 2540 * sbcut/drop/flush discarded these 2541 * mbufs. 2542 */ 2543 m_freem(data); 2544 break; 2545 } 2546 2547 /* 2548 * Drop this TLS record's data, but keep 2549 * decrypting subsequent records. 2550 */ 2551 sb->sb_ccc -= tls_len; 2552 sb->sb_tlsdcc = 0; 2553 2554 if (error != EMSGSIZE) 2555 error = EBADMSG; 2556 CURVNET_SET(so->so_vnet); 2557 so->so_error = error; 2558 sorwakeup_locked(so); 2559 CURVNET_RESTORE(); 2560 2561 m_freem(data); 2562 2563 SOCKBUF_LOCK(sb); 2564 continue; 2565 } 2566 2567 /* Allocate the control mbuf. */ 2568 memset(&tgr, 0, sizeof(tgr)); 2569 tgr.tls_type = record_type; 2570 tgr.tls_vmajor = hdr->tls_vmajor; 2571 tgr.tls_vminor = hdr->tls_vminor; 2572 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - 2573 trail_len); 2574 control = sbcreatecontrol(&tgr, sizeof(tgr), 2575 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); 2576 2577 SOCKBUF_LOCK(sb); 2578 if (sb->sb_tlsdcc == 0) { 2579 /* sbcut/drop/flush discarded these mbufs. */ 2580 MPASS(sb->sb_tlscc == 0); 2581 m_freem(data); 2582 m_freem(control); 2583 break; 2584 } 2585 2586 /* 2587 * Clear the 'dcc' accounting in preparation for 2588 * adding the decrypted record. 2589 */ 2590 sb->sb_ccc -= tls_len; 2591 sb->sb_tlsdcc = 0; 2592 SBCHECK(sb); 2593 2594 /* If there is no payload, drop all of the data. */ 2595 if (tgr.tls_length == htobe16(0)) { 2596 m_freem(data); 2597 data = NULL; 2598 } else { 2599 /* Trim header. */ 2600 remain = tls->params.tls_hlen; 2601 while (remain > 0) { 2602 if (data->m_len > remain) { 2603 data->m_data += remain; 2604 data->m_len -= remain; 2605 break; 2606 } 2607 remain -= data->m_len; 2608 data = m_free(data); 2609 } 2610 2611 /* Trim trailer and clear M_NOTREADY. */ 2612 remain = be16toh(tgr.tls_length); 2613 m = data; 2614 for (m = data; remain > m->m_len; m = m->m_next) { 2615 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); 2616 remain -= m->m_len; 2617 } 2618 m->m_len = remain; 2619 m_freem(m->m_next); 2620 m->m_next = NULL; 2621 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); 2622 2623 /* Set EOR on the final mbuf. */ 2624 m->m_flags |= M_EOR; 2625 } 2626 2627 sbappendcontrol_locked(sb, data, control, 0); 2628 2629 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) { 2630 sb->sb_flags |= SB_TLS_RX_RESYNC; 2631 SOCKBUF_UNLOCK(sb); 2632 ktls_resync_ifnet(so, tls_len, seqno); 2633 SOCKBUF_LOCK(sb); 2634 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) { 2635 sb->sb_flags &= ~SB_TLS_RX_RESYNC; 2636 SOCKBUF_UNLOCK(sb); 2637 ktls_resync_ifnet(so, 0, seqno); 2638 SOCKBUF_LOCK(sb); 2639 } 2640 } 2641 2642 sb->sb_flags &= ~SB_TLS_RX_RUNNING; 2643 2644 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) 2645 so->so_error = EMSGSIZE; 2646 2647 sorwakeup_locked(so); 2648 2649 deref: 2650 SOCKBUF_UNLOCK_ASSERT(sb); 2651 2652 CURVNET_SET(so->so_vnet); 2653 sorele(so); 2654 CURVNET_RESTORE(); 2655 } 2656 2657 void 2658 ktls_enqueue_to_free(struct mbuf *m) 2659 { 2660 struct ktls_wq *wq; 2661 bool running; 2662 2663 /* Mark it for freeing. */ 2664 m->m_epg_flags |= EPG_FLAG_2FREE; 2665 wq = &ktls_wq[m->m_epg_tls->wq_index]; 2666 mtx_lock(&wq->mtx); 2667 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2668 running = wq->running; 2669 mtx_unlock(&wq->mtx); 2670 if (!running) 2671 wakeup(wq); 2672 } 2673 2674 static void * 2675 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m) 2676 { 2677 void *buf; 2678 int domain, running; 2679 2680 if (m->m_epg_npgs <= 2) 2681 return (NULL); 2682 if (ktls_buffer_zone == NULL) 2683 return (NULL); 2684 if ((u_int)(ticks - wq->lastallocfail) < hz) { 2685 /* 2686 * Rate-limit allocation attempts after a failure. 2687 * ktls_buffer_import() will acquire a per-domain mutex to check 2688 * the free page queues and may fail consistently if memory is 2689 * fragmented. 2690 */ 2691 return (NULL); 2692 } 2693 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM); 2694 if (buf == NULL) { 2695 domain = PCPU_GET(domain); 2696 wq->lastallocfail = ticks; 2697 2698 /* 2699 * Note that this check is "racy", but the races are 2700 * harmless, and are either a spurious wakeup if 2701 * multiple threads fail allocations before the alloc 2702 * thread wakes, or waiting an extra second in case we 2703 * see an old value of running == true. 2704 */ 2705 if (!VM_DOMAIN_EMPTY(domain)) { 2706 running = atomic_load_int(&ktls_domains[domain].reclaim_td.running); 2707 if (!running) 2708 wakeup(&ktls_domains[domain].reclaim_td); 2709 } 2710 } 2711 return (buf); 2712 } 2713 2714 static int 2715 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m, 2716 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state) 2717 { 2718 vm_page_t pg; 2719 int error, i, len, off; 2720 2721 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY), 2722 ("%p not unready & nomap mbuf\n", m)); 2723 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen, 2724 ("page count %d larger than maximum frame length %d", m->m_epg_npgs, 2725 ktls_maxlen)); 2726 2727 /* Anonymous mbufs are encrypted in place. */ 2728 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0) 2729 return (ktls_ocf_encrypt(state, tls, m, NULL, 0)); 2730 2731 /* 2732 * For file-backed mbufs (from sendfile), anonymous wired 2733 * pages are allocated and used as the encryption destination. 2734 */ 2735 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) { 2736 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len - 2737 m->m_epg_1st_off; 2738 state->dst_iov[0].iov_base = (char *)state->cbuf + 2739 m->m_epg_1st_off; 2740 state->dst_iov[0].iov_len = len; 2741 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf); 2742 i = 1; 2743 } else { 2744 off = m->m_epg_1st_off; 2745 for (i = 0; i < m->m_epg_npgs; i++, off = 0) { 2746 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP | 2747 VM_ALLOC_WIRED | VM_ALLOC_WAITOK); 2748 len = m_epg_pagelen(m, i, off); 2749 state->parray[i] = VM_PAGE_TO_PHYS(pg); 2750 state->dst_iov[i].iov_base = 2751 (char *)PHYS_TO_DMAP(state->parray[i]) + off; 2752 state->dst_iov[i].iov_len = len; 2753 } 2754 } 2755 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small")); 2756 state->dst_iov[i].iov_base = m->m_epg_trail; 2757 state->dst_iov[i].iov_len = m->m_epg_trllen; 2758 2759 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1); 2760 2761 if (__predict_false(error != 0)) { 2762 /* Free the anonymous pages. */ 2763 if (state->cbuf != NULL) 2764 uma_zfree(ktls_buffer_zone, state->cbuf); 2765 else { 2766 for (i = 0; i < m->m_epg_npgs; i++) { 2767 pg = PHYS_TO_VM_PAGE(state->parray[i]); 2768 (void)vm_page_unwire_noq(pg); 2769 vm_page_free(pg); 2770 } 2771 } 2772 } 2773 return (error); 2774 } 2775 2776 /* Number of TLS records in a batch passed to ktls_enqueue(). */ 2777 static u_int 2778 ktls_batched_records(struct mbuf *m) 2779 { 2780 int page_count, records; 2781 2782 records = 0; 2783 page_count = m->m_epg_enc_cnt; 2784 while (page_count > 0) { 2785 records++; 2786 page_count -= m->m_epg_nrdy; 2787 m = m->m_next; 2788 } 2789 KASSERT(page_count == 0, ("%s: mismatched page count", __func__)); 2790 return (records); 2791 } 2792 2793 void 2794 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 2795 { 2796 struct ktls_session *tls; 2797 struct ktls_wq *wq; 2798 int queued; 2799 bool running; 2800 2801 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == 2802 (M_EXTPG | M_NOTREADY)), 2803 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 2804 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 2805 2806 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); 2807 2808 m->m_epg_enc_cnt = page_count; 2809 2810 /* 2811 * Save a pointer to the socket. The caller is responsible 2812 * for taking an additional reference via soref(). 2813 */ 2814 m->m_epg_so = so; 2815 2816 queued = 1; 2817 tls = m->m_epg_tls; 2818 wq = &ktls_wq[tls->wq_index]; 2819 mtx_lock(&wq->mtx); 2820 if (__predict_false(tls->sequential_records)) { 2821 /* 2822 * For TLS 1.0, records must be encrypted 2823 * sequentially. For a given connection, all records 2824 * queued to the associated work queue are processed 2825 * sequentially. However, sendfile(2) might complete 2826 * I/O requests spanning multiple TLS records out of 2827 * order. Here we ensure TLS records are enqueued to 2828 * the work queue in FIFO order. 2829 * 2830 * tls->next_seqno holds the sequence number of the 2831 * next TLS record that should be enqueued to the work 2832 * queue. If this next record is not tls->next_seqno, 2833 * it must be a future record, so insert it, sorted by 2834 * TLS sequence number, into tls->pending_records and 2835 * return. 2836 * 2837 * If this TLS record matches tls->next_seqno, place 2838 * it in the work queue and then check 2839 * tls->pending_records to see if any 2840 * previously-queued records are now ready for 2841 * encryption. 2842 */ 2843 if (m->m_epg_seqno != tls->next_seqno) { 2844 struct mbuf *n, *p; 2845 2846 p = NULL; 2847 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) { 2848 if (n->m_epg_seqno > m->m_epg_seqno) 2849 break; 2850 p = n; 2851 } 2852 if (n == NULL) 2853 STAILQ_INSERT_TAIL(&tls->pending_records, m, 2854 m_epg_stailq); 2855 else if (p == NULL) 2856 STAILQ_INSERT_HEAD(&tls->pending_records, m, 2857 m_epg_stailq); 2858 else 2859 STAILQ_INSERT_AFTER(&tls->pending_records, p, m, 2860 m_epg_stailq); 2861 mtx_unlock(&wq->mtx); 2862 counter_u64_add(ktls_cnt_tx_pending, 1); 2863 return; 2864 } 2865 2866 tls->next_seqno += ktls_batched_records(m); 2867 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2868 2869 while (!STAILQ_EMPTY(&tls->pending_records)) { 2870 struct mbuf *n; 2871 2872 n = STAILQ_FIRST(&tls->pending_records); 2873 if (n->m_epg_seqno != tls->next_seqno) 2874 break; 2875 2876 queued++; 2877 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq); 2878 tls->next_seqno += ktls_batched_records(n); 2879 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq); 2880 } 2881 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1)); 2882 } else 2883 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2884 2885 running = wq->running; 2886 mtx_unlock(&wq->mtx); 2887 if (!running) 2888 wakeup(wq); 2889 counter_u64_add(ktls_cnt_tx_queued, queued); 2890 } 2891 2892 /* 2893 * Once a file-backed mbuf (from sendfile) has been encrypted, free 2894 * the pages from the file and replace them with the anonymous pages 2895 * allocated in ktls_encrypt_record(). 2896 */ 2897 static void 2898 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state) 2899 { 2900 int i; 2901 2902 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0); 2903 2904 /* Free the old pages. */ 2905 m->m_ext.ext_free(m); 2906 2907 /* Replace them with the new pages. */ 2908 if (state->cbuf != NULL) { 2909 for (i = 0; i < m->m_epg_npgs; i++) 2910 m->m_epg_pa[i] = state->parray[0] + ptoa(i); 2911 2912 /* Contig pages should go back to the cache. */ 2913 m->m_ext.ext_free = ktls_free_mext_contig; 2914 } else { 2915 for (i = 0; i < m->m_epg_npgs; i++) 2916 m->m_epg_pa[i] = state->parray[i]; 2917 2918 /* Use the basic free routine. */ 2919 m->m_ext.ext_free = mb_free_mext_pgs; 2920 } 2921 2922 /* Pages are now writable. */ 2923 m->m_epg_flags |= EPG_FLAG_ANON; 2924 } 2925 2926 static __noinline void 2927 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top) 2928 { 2929 struct ktls_ocf_encrypt_state state; 2930 struct ktls_session *tls; 2931 struct socket *so; 2932 struct mbuf *m; 2933 int error, npages, total_pages; 2934 2935 so = top->m_epg_so; 2936 tls = top->m_epg_tls; 2937 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 2938 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 2939 #ifdef INVARIANTS 2940 top->m_epg_so = NULL; 2941 #endif 2942 total_pages = top->m_epg_enc_cnt; 2943 npages = 0; 2944 2945 /* 2946 * Encrypt the TLS records in the chain of mbufs starting with 2947 * 'top'. 'total_pages' gives us a total count of pages and is 2948 * used to know when we have finished encrypting the TLS 2949 * records originally queued with 'top'. 2950 * 2951 * NB: These mbufs are queued in the socket buffer and 2952 * 'm_next' is traversing the mbufs in the socket buffer. The 2953 * socket buffer lock is not held while traversing this chain. 2954 * Since the mbufs are all marked M_NOTREADY their 'm_next' 2955 * pointers should be stable. However, the 'm_next' of the 2956 * last mbuf encrypted is not necessarily NULL. It can point 2957 * to other mbufs appended while 'top' was on the TLS work 2958 * queue. 2959 * 2960 * Each mbuf holds an entire TLS record. 2961 */ 2962 error = 0; 2963 for (m = top; npages != total_pages; m = m->m_next) { 2964 KASSERT(m->m_epg_tls == tls, 2965 ("different TLS sessions in a single mbuf chain: %p vs %p", 2966 tls, m->m_epg_tls)); 2967 KASSERT(npages + m->m_epg_npgs <= total_pages, 2968 ("page count mismatch: top %p, total_pages %d, m %p", top, 2969 total_pages, m)); 2970 2971 error = ktls_encrypt_record(wq, m, tls, &state); 2972 if (error) { 2973 counter_u64_add(ktls_offload_failed_crypto, 1); 2974 break; 2975 } 2976 2977 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) 2978 ktls_finish_nonanon(m, &state); 2979 2980 npages += m->m_epg_nrdy; 2981 2982 /* 2983 * Drop a reference to the session now that it is no 2984 * longer needed. Existing code depends on encrypted 2985 * records having no associated session vs 2986 * yet-to-be-encrypted records having an associated 2987 * session. 2988 */ 2989 m->m_epg_tls = NULL; 2990 ktls_free(tls); 2991 } 2992 2993 CURVNET_SET(so->so_vnet); 2994 if (error == 0) { 2995 (void)so->so_proto->pr_ready(so, top, npages); 2996 } else { 2997 ktls_drop(so, EIO); 2998 mb_free_notready(top, total_pages); 2999 } 3000 3001 sorele(so); 3002 CURVNET_RESTORE(); 3003 } 3004 3005 void 3006 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error) 3007 { 3008 struct ktls_session *tls; 3009 struct socket *so; 3010 struct mbuf *m; 3011 int npages; 3012 3013 m = state->m; 3014 3015 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) 3016 ktls_finish_nonanon(m, state); 3017 3018 so = state->so; 3019 free(state, M_KTLS); 3020 3021 /* 3022 * Drop a reference to the session now that it is no longer 3023 * needed. Existing code depends on encrypted records having 3024 * no associated session vs yet-to-be-encrypted records having 3025 * an associated session. 3026 */ 3027 tls = m->m_epg_tls; 3028 m->m_epg_tls = NULL; 3029 ktls_free(tls); 3030 3031 if (error != 0) 3032 counter_u64_add(ktls_offload_failed_crypto, 1); 3033 3034 CURVNET_SET(so->so_vnet); 3035 npages = m->m_epg_nrdy; 3036 3037 if (error == 0) { 3038 (void)so->so_proto->pr_ready(so, m, npages); 3039 } else { 3040 ktls_drop(so, EIO); 3041 mb_free_notready(m, npages); 3042 } 3043 3044 sorele(so); 3045 CURVNET_RESTORE(); 3046 } 3047 3048 /* 3049 * Similar to ktls_encrypt, but used with asynchronous OCF backends 3050 * (coprocessors) where encryption does not use host CPU resources and 3051 * it can be beneficial to queue more requests than CPUs. 3052 */ 3053 static __noinline void 3054 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top) 3055 { 3056 struct ktls_ocf_encrypt_state *state; 3057 struct ktls_session *tls; 3058 struct socket *so; 3059 struct mbuf *m, *n; 3060 int error, mpages, npages, total_pages; 3061 3062 so = top->m_epg_so; 3063 tls = top->m_epg_tls; 3064 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 3065 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 3066 #ifdef INVARIANTS 3067 top->m_epg_so = NULL; 3068 #endif 3069 total_pages = top->m_epg_enc_cnt; 3070 npages = 0; 3071 3072 error = 0; 3073 for (m = top; npages != total_pages; m = n) { 3074 KASSERT(m->m_epg_tls == tls, 3075 ("different TLS sessions in a single mbuf chain: %p vs %p", 3076 tls, m->m_epg_tls)); 3077 KASSERT(npages + m->m_epg_npgs <= total_pages, 3078 ("page count mismatch: top %p, total_pages %d, m %p", top, 3079 total_pages, m)); 3080 3081 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO); 3082 soref(so); 3083 state->so = so; 3084 state->m = m; 3085 3086 mpages = m->m_epg_nrdy; 3087 n = m->m_next; 3088 3089 error = ktls_encrypt_record(wq, m, tls, state); 3090 if (error) { 3091 counter_u64_add(ktls_offload_failed_crypto, 1); 3092 free(state, M_KTLS); 3093 CURVNET_SET(so->so_vnet); 3094 sorele(so); 3095 CURVNET_RESTORE(); 3096 break; 3097 } 3098 3099 npages += mpages; 3100 } 3101 3102 CURVNET_SET(so->so_vnet); 3103 if (error != 0) { 3104 ktls_drop(so, EIO); 3105 mb_free_notready(m, total_pages - npages); 3106 } 3107 3108 sorele(so); 3109 CURVNET_RESTORE(); 3110 } 3111 3112 static int 3113 ktls_bind_domain(int domain) 3114 { 3115 int error; 3116 3117 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]); 3118 if (error != 0) 3119 return (error); 3120 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain); 3121 return (0); 3122 } 3123 3124 static void 3125 ktls_reclaim_thread(void *ctx) 3126 { 3127 struct ktls_domain_info *ktls_domain = ctx; 3128 struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td; 3129 struct sysctl_oid *oid; 3130 char name[80]; 3131 int error, domain; 3132 3133 domain = ktls_domain - ktls_domains; 3134 if (bootverbose) 3135 printf("Starting KTLS reclaim thread for domain %d\n", domain); 3136 error = ktls_bind_domain(domain); 3137 if (error) 3138 printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n", 3139 domain, error); 3140 snprintf(name, sizeof(name), "domain%d", domain); 3141 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO, 3142 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 3143 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims", 3144 CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed"); 3145 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups", 3146 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups"); 3147 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running", 3148 CTLFLAG_RD, &sc->running, 0, "thread running"); 3149 3150 for (;;) { 3151 atomic_store_int(&sc->running, 0); 3152 tsleep(sc, PZERO | PNOLOCK, "-", 0); 3153 atomic_store_int(&sc->running, 1); 3154 sc->wakeups++; 3155 /* 3156 * Below we attempt to reclaim ktls_max_reclaim 3157 * buffers using vm_page_reclaim_contig_domain_ext(). 3158 * We do this here, as this function can take several 3159 * seconds to scan all of memory and it does not 3160 * matter if this thread pauses for a while. If we 3161 * block a ktls worker thread, we risk developing 3162 * backlogs of buffers to be encrypted, leading to 3163 * surges of traffic and potential NIC output drops. 3164 */ 3165 if (vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL, 3166 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, 3167 ktls_max_reclaim) != 0) { 3168 vm_wait_domain(domain); 3169 } else { 3170 sc->reclaims += ktls_max_reclaim; 3171 } 3172 } 3173 } 3174 3175 static void 3176 ktls_work_thread(void *ctx) 3177 { 3178 struct ktls_wq *wq = ctx; 3179 struct mbuf *m, *n; 3180 struct socket *so, *son; 3181 STAILQ_HEAD(, mbuf) local_m_head; 3182 STAILQ_HEAD(, socket) local_so_head; 3183 int cpu; 3184 3185 cpu = wq - ktls_wq; 3186 if (bootverbose) 3187 printf("Starting KTLS worker thread for CPU %d\n", cpu); 3188 3189 /* 3190 * Bind to a core. If ktls_bind_threads is > 1, then 3191 * we bind to the NUMA domain instead. 3192 */ 3193 if (ktls_bind_threads) { 3194 int error; 3195 3196 if (ktls_bind_threads > 1) { 3197 struct pcpu *pc = pcpu_find(cpu); 3198 3199 error = ktls_bind_domain(pc->pc_domain); 3200 } else { 3201 cpuset_t mask; 3202 3203 CPU_SETOF(cpu, &mask); 3204 error = cpuset_setthread(curthread->td_tid, &mask); 3205 } 3206 if (error) 3207 printf("Unable to bind KTLS worker thread for CPU %d: error %d\n", 3208 cpu, error); 3209 } 3210 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 3211 fpu_kern_thread(0); 3212 #endif 3213 for (;;) { 3214 mtx_lock(&wq->mtx); 3215 while (STAILQ_EMPTY(&wq->m_head) && 3216 STAILQ_EMPTY(&wq->so_head)) { 3217 wq->running = false; 3218 mtx_sleep(wq, &wq->mtx, 0, "-", 0); 3219 wq->running = true; 3220 } 3221 3222 STAILQ_INIT(&local_m_head); 3223 STAILQ_CONCAT(&local_m_head, &wq->m_head); 3224 STAILQ_INIT(&local_so_head); 3225 STAILQ_CONCAT(&local_so_head, &wq->so_head); 3226 mtx_unlock(&wq->mtx); 3227 3228 STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) { 3229 if (m->m_epg_flags & EPG_FLAG_2FREE) { 3230 ktls_free(m->m_epg_tls); 3231 m_free_raw(m); 3232 } else { 3233 if (m->m_epg_tls->sync_dispatch) 3234 ktls_encrypt(wq, m); 3235 else 3236 ktls_encrypt_async(wq, m); 3237 counter_u64_add(ktls_cnt_tx_queued, -1); 3238 } 3239 } 3240 3241 STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) { 3242 ktls_decrypt(so); 3243 counter_u64_add(ktls_cnt_rx_queued, -1); 3244 } 3245 } 3246 } 3247 3248 static void 3249 ktls_disable_ifnet_help(void *context, int pending __unused) 3250 { 3251 struct ktls_session *tls; 3252 struct inpcb *inp; 3253 struct tcpcb *tp; 3254 struct socket *so; 3255 int err; 3256 3257 tls = context; 3258 inp = tls->inp; 3259 if (inp == NULL) 3260 return; 3261 INP_WLOCK(inp); 3262 so = inp->inp_socket; 3263 MPASS(so != NULL); 3264 if (inp->inp_flags & INP_DROPPED) { 3265 goto out; 3266 } 3267 3268 if (so->so_snd.sb_tls_info != NULL) 3269 err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW); 3270 else 3271 err = ENXIO; 3272 if (err == 0) { 3273 counter_u64_add(ktls_ifnet_disable_ok, 1); 3274 /* ktls_set_tx_mode() drops inp wlock, so recheck flags */ 3275 if ((inp->inp_flags & INP_DROPPED) == 0 && 3276 (tp = intotcpcb(inp)) != NULL && 3277 tp->t_fb->tfb_hwtls_change != NULL) 3278 (*tp->t_fb->tfb_hwtls_change)(tp, 0); 3279 } else { 3280 counter_u64_add(ktls_ifnet_disable_fail, 1); 3281 } 3282 3283 out: 3284 CURVNET_SET(so->so_vnet); 3285 sorele(so); 3286 CURVNET_RESTORE(); 3287 INP_WUNLOCK(inp); 3288 ktls_free(tls); 3289 } 3290 3291 /* 3292 * Called when re-transmits are becoming a substantial portion of the 3293 * sends on this connection. When this happens, we transition the 3294 * connection to software TLS. This is needed because most inline TLS 3295 * NICs keep crypto state only for in-order transmits. This means 3296 * that to handle a TCP rexmit (which is out-of-order), the NIC must 3297 * re-DMA the entire TLS record up to and including the current 3298 * segment. This means that when re-transmitting the last ~1448 byte 3299 * segment of a 16KB TLS record, we could wind up re-DMA'ing an order 3300 * of magnitude more data than we are sending. This can cause the 3301 * PCIe link to saturate well before the network, which can cause 3302 * output drops, and a general loss of capacity. 3303 */ 3304 void 3305 ktls_disable_ifnet(void *arg) 3306 { 3307 struct tcpcb *tp; 3308 struct inpcb *inp; 3309 struct socket *so; 3310 struct ktls_session *tls; 3311 3312 tp = arg; 3313 inp = tptoinpcb(tp); 3314 INP_WLOCK_ASSERT(inp); 3315 so = inp->inp_socket; 3316 SOCK_LOCK(so); 3317 tls = so->so_snd.sb_tls_info; 3318 if (tp->t_nic_ktls_xmit_dis == 1) { 3319 SOCK_UNLOCK(so); 3320 return; 3321 } 3322 3323 /* 3324 * note that t_nic_ktls_xmit_dis is never cleared; disabling 3325 * ifnet can only be done once per connection, so we never want 3326 * to do it again 3327 */ 3328 3329 (void)ktls_hold(tls); 3330 soref(so); 3331 tp->t_nic_ktls_xmit_dis = 1; 3332 SOCK_UNLOCK(so); 3333 TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls); 3334 (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task); 3335 } 3336