1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2014-2019 Netflix Inc. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 #include "opt_kern_tls.h" 34 #include "opt_ratelimit.h" 35 #include "opt_rss.h" 36 37 #include <sys/param.h> 38 #include <sys/kernel.h> 39 #include <sys/domainset.h> 40 #include <sys/endian.h> 41 #include <sys/ktls.h> 42 #include <sys/lock.h> 43 #include <sys/mbuf.h> 44 #include <sys/mutex.h> 45 #include <sys/rmlock.h> 46 #include <sys/proc.h> 47 #include <sys/protosw.h> 48 #include <sys/refcount.h> 49 #include <sys/smp.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/sysctl.h> 53 #include <sys/taskqueue.h> 54 #include <sys/kthread.h> 55 #include <sys/uio.h> 56 #include <sys/vmmeter.h> 57 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) 58 #include <machine/pcb.h> 59 #endif 60 #include <machine/vmparam.h> 61 #include <net/if.h> 62 #include <net/if_var.h> 63 #ifdef RSS 64 #include <net/netisr.h> 65 #include <net/rss_config.h> 66 #endif 67 #include <net/route.h> 68 #include <net/route/nhop.h> 69 #include <netinet/in.h> 70 #include <netinet/in_pcb.h> 71 #include <netinet/tcp_var.h> 72 #ifdef TCP_OFFLOAD 73 #include <netinet/tcp_offload.h> 74 #endif 75 #include <opencrypto/cryptodev.h> 76 #include <opencrypto/ktls.h> 77 #include <vm/uma_dbg.h> 78 #include <vm/vm.h> 79 #include <vm/vm_pageout.h> 80 #include <vm/vm_page.h> 81 #include <vm/vm_pagequeue.h> 82 83 struct ktls_wq { 84 struct mtx mtx; 85 STAILQ_HEAD(, mbuf) m_head; 86 STAILQ_HEAD(, socket) so_head; 87 bool running; 88 int lastallocfail; 89 } __aligned(CACHE_LINE_SIZE); 90 91 struct ktls_reclaim_thread { 92 uint64_t wakeups; 93 uint64_t reclaims; 94 struct thread *td; 95 int running; 96 }; 97 98 struct ktls_domain_info { 99 int count; 100 int cpu[MAXCPU]; 101 struct ktls_reclaim_thread reclaim_td; 102 }; 103 104 struct ktls_domain_info ktls_domains[MAXMEMDOM]; 105 static struct ktls_wq *ktls_wq; 106 static struct proc *ktls_proc; 107 static uma_zone_t ktls_session_zone; 108 static uma_zone_t ktls_buffer_zone; 109 static uint16_t ktls_cpuid_lookup[MAXCPU]; 110 static int ktls_init_state; 111 static struct sx ktls_init_lock; 112 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init"); 113 114 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 115 "Kernel TLS offload"); 116 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 117 "Kernel TLS offload stats"); 118 119 #ifdef RSS 120 static int ktls_bind_threads = 1; 121 #else 122 static int ktls_bind_threads; 123 #endif 124 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, 125 &ktls_bind_threads, 0, 126 "Bind crypto threads to cores (1) or cores and domains (2) at boot"); 127 128 static u_int ktls_maxlen = 16384; 129 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN, 130 &ktls_maxlen, 0, "Maximum TLS record size"); 131 132 static int ktls_number_threads; 133 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, 134 &ktls_number_threads, 0, 135 "Number of TLS threads in thread-pool"); 136 137 unsigned int ktls_ifnet_max_rexmit_pct = 2; 138 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN, 139 &ktls_ifnet_max_rexmit_pct, 2, 140 "Max percent bytes retransmitted before ifnet TLS is disabled"); 141 142 static bool ktls_offload_enable; 143 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, 144 &ktls_offload_enable, 0, 145 "Enable support for kernel TLS offload"); 146 147 static bool ktls_cbc_enable = true; 148 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN, 149 &ktls_cbc_enable, 1, 150 "Enable support of AES-CBC crypto for kernel TLS"); 151 152 static bool ktls_sw_buffer_cache = true; 153 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN, 154 &ktls_sw_buffer_cache, 1, 155 "Enable caching of output buffers for SW encryption"); 156 157 static int ktls_max_reclaim = 1024; 158 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN, 159 &ktls_max_reclaim, 128, 160 "Max number of 16k buffers to reclaim in thread context"); 161 162 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active); 163 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, 164 &ktls_tasks_active, "Number of active tasks"); 165 166 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending); 167 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD, 168 &ktls_cnt_tx_pending, 169 "Number of TLS 1.0 records waiting for earlier TLS records"); 170 171 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued); 172 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, 173 &ktls_cnt_tx_queued, 174 "Number of TLS records in queue to tasks for SW encryption"); 175 176 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued); 177 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, 178 &ktls_cnt_rx_queued, 179 "Number of TLS sockets in queue to tasks for SW decryption"); 180 181 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total); 182 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, 183 CTLFLAG_RD, &ktls_offload_total, 184 "Total successful TLS setups (parameters set)"); 185 186 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls); 187 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, 188 CTLFLAG_RD, &ktls_offload_enable_calls, 189 "Total number of TLS enable calls made"); 190 191 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active); 192 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, 193 &ktls_offload_active, "Total Active TLS sessions"); 194 195 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records); 196 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, 197 &ktls_offload_corrupted_records, "Total corrupted TLS records received"); 198 199 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto); 200 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, 201 &ktls_offload_failed_crypto, "Total TLS crypto failures"); 202 203 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet); 204 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, 205 &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); 206 207 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw); 208 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, 209 &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); 210 211 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed); 212 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, 213 &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); 214 215 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail); 216 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD, 217 &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet"); 218 219 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok); 220 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD, 221 &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet"); 222 223 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task); 224 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD, 225 &ktls_destroy_task, 226 "Number of times ktls session was destroyed via taskqueue"); 227 228 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 229 "Software TLS session stats"); 230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 231 "Hardware (ifnet) TLS session stats"); 232 #ifdef TCP_OFFLOAD 233 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 234 "TOE TLS session stats"); 235 #endif 236 237 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc); 238 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, 239 "Active number of software TLS sessions using AES-CBC"); 240 241 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm); 242 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, 243 "Active number of software TLS sessions using AES-GCM"); 244 245 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20); 246 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD, 247 &ktls_sw_chacha20, 248 "Active number of software TLS sessions using Chacha20-Poly1305"); 249 250 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc); 251 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, 252 &ktls_ifnet_cbc, 253 "Active number of ifnet TLS sessions using AES-CBC"); 254 255 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm); 256 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, 257 &ktls_ifnet_gcm, 258 "Active number of ifnet TLS sessions using AES-GCM"); 259 260 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20); 261 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD, 262 &ktls_ifnet_chacha20, 263 "Active number of ifnet TLS sessions using Chacha20-Poly1305"); 264 265 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset); 266 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, 267 &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); 268 269 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped); 270 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, 271 &ktls_ifnet_reset_dropped, 272 "TLS sessions dropped after failing to update ifnet send tag"); 273 274 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed); 275 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, 276 &ktls_ifnet_reset_failed, 277 "TLS sessions that failed to allocate a new ifnet send tag"); 278 279 static int ktls_ifnet_permitted; 280 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, 281 &ktls_ifnet_permitted, 1, 282 "Whether to permit hardware (ifnet) TLS sessions"); 283 284 #ifdef TCP_OFFLOAD 285 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc); 286 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, 287 &ktls_toe_cbc, 288 "Active number of TOE TLS sessions using AES-CBC"); 289 290 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm); 291 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, 292 &ktls_toe_gcm, 293 "Active number of TOE TLS sessions using AES-GCM"); 294 295 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20); 296 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD, 297 &ktls_toe_chacha20, 298 "Active number of TOE TLS sessions using Chacha20-Poly1305"); 299 #endif 300 301 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); 302 303 static void ktls_reset_receive_tag(void *context, int pending); 304 static void ktls_reset_send_tag(void *context, int pending); 305 static void ktls_work_thread(void *ctx); 306 static void ktls_reclaim_thread(void *ctx); 307 308 static u_int 309 ktls_get_cpu(struct socket *so) 310 { 311 struct inpcb *inp; 312 #ifdef NUMA 313 struct ktls_domain_info *di; 314 #endif 315 u_int cpuid; 316 317 inp = sotoinpcb(so); 318 #ifdef RSS 319 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 320 if (cpuid != NETISR_CPUID_NONE) 321 return (cpuid); 322 #endif 323 /* 324 * Just use the flowid to shard connections in a repeatable 325 * fashion. Note that TLS 1.0 sessions rely on the 326 * serialization provided by having the same connection use 327 * the same queue. 328 */ 329 #ifdef NUMA 330 if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) { 331 di = &ktls_domains[inp->inp_numa_domain]; 332 cpuid = di->cpu[inp->inp_flowid % di->count]; 333 } else 334 #endif 335 cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; 336 return (cpuid); 337 } 338 339 static int 340 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags) 341 { 342 vm_page_t m; 343 int i, req; 344 345 KASSERT((ktls_maxlen & PAGE_MASK) == 0, 346 ("%s: ktls max length %d is not page size-aligned", 347 __func__, ktls_maxlen)); 348 349 req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags); 350 for (i = 0; i < count; i++) { 351 m = vm_page_alloc_noobj_contig_domain(domain, req, 352 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, 353 VM_MEMATTR_DEFAULT); 354 if (m == NULL) 355 break; 356 store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 357 } 358 return (i); 359 } 360 361 static void 362 ktls_buffer_release(void *arg __unused, void **store, int count) 363 { 364 vm_page_t m; 365 int i, j; 366 367 for (i = 0; i < count; i++) { 368 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i])); 369 for (j = 0; j < atop(ktls_maxlen); j++) { 370 (void)vm_page_unwire_noq(m + j); 371 vm_page_free(m + j); 372 } 373 } 374 } 375 376 static void 377 ktls_free_mext_contig(struct mbuf *m) 378 { 379 M_ASSERTEXTPG(m); 380 uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0])); 381 } 382 383 static int 384 ktls_init(void) 385 { 386 struct thread *td; 387 struct pcpu *pc; 388 int count, domain, error, i; 389 390 ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, 391 M_WAITOK | M_ZERO); 392 393 ktls_session_zone = uma_zcreate("ktls_session", 394 sizeof(struct ktls_session), 395 NULL, NULL, NULL, NULL, 396 UMA_ALIGN_CACHE, 0); 397 398 if (ktls_sw_buffer_cache) { 399 ktls_buffer_zone = uma_zcache_create("ktls_buffers", 400 roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL, 401 ktls_buffer_import, ktls_buffer_release, NULL, 402 UMA_ZONE_FIRSTTOUCH); 403 } 404 405 /* 406 * Initialize the workqueues to run the TLS work. We create a 407 * work queue for each CPU. 408 */ 409 CPU_FOREACH(i) { 410 STAILQ_INIT(&ktls_wq[i].m_head); 411 STAILQ_INIT(&ktls_wq[i].so_head); 412 mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); 413 if (ktls_bind_threads > 1) { 414 pc = pcpu_find(i); 415 domain = pc->pc_domain; 416 count = ktls_domains[domain].count; 417 ktls_domains[domain].cpu[count] = i; 418 ktls_domains[domain].count++; 419 } 420 ktls_cpuid_lookup[ktls_number_threads] = i; 421 ktls_number_threads++; 422 } 423 424 /* 425 * If we somehow have an empty domain, fall back to choosing 426 * among all KTLS threads. 427 */ 428 if (ktls_bind_threads > 1) { 429 for (i = 0; i < vm_ndomains; i++) { 430 if (ktls_domains[i].count == 0) { 431 ktls_bind_threads = 1; 432 break; 433 } 434 } 435 } 436 437 /* Start kthreads for each workqueue. */ 438 CPU_FOREACH(i) { 439 error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], 440 &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); 441 if (error) { 442 printf("Can't add KTLS thread %d error %d\n", i, error); 443 return (error); 444 } 445 } 446 447 /* 448 * Start an allocation thread per-domain to perform blocking allocations 449 * of 16k physically contiguous TLS crypto destination buffers. 450 */ 451 if (ktls_sw_buffer_cache) { 452 for (domain = 0; domain < vm_ndomains; domain++) { 453 if (VM_DOMAIN_EMPTY(domain)) 454 continue; 455 if (CPU_EMPTY(&cpuset_domain[domain])) 456 continue; 457 error = kproc_kthread_add(ktls_reclaim_thread, 458 &ktls_domains[domain], &ktls_proc, 459 &ktls_domains[domain].reclaim_td.td, 460 0, 0, "KTLS", "reclaim_%d", domain); 461 if (error) { 462 printf("Can't add KTLS reclaim thread %d error %d\n", 463 domain, error); 464 return (error); 465 } 466 } 467 } 468 469 if (bootverbose) 470 printf("KTLS: Initialized %d threads\n", ktls_number_threads); 471 return (0); 472 } 473 474 static int 475 ktls_start_kthreads(void) 476 { 477 int error, state; 478 479 start: 480 state = atomic_load_acq_int(&ktls_init_state); 481 if (__predict_true(state > 0)) 482 return (0); 483 if (state < 0) 484 return (ENXIO); 485 486 sx_xlock(&ktls_init_lock); 487 if (ktls_init_state != 0) { 488 sx_xunlock(&ktls_init_lock); 489 goto start; 490 } 491 492 error = ktls_init(); 493 if (error == 0) 494 state = 1; 495 else 496 state = -1; 497 atomic_store_rel_int(&ktls_init_state, state); 498 sx_xunlock(&ktls_init_lock); 499 return (error); 500 } 501 502 static int 503 ktls_create_session(struct socket *so, struct tls_enable *en, 504 struct ktls_session **tlsp, int direction) 505 { 506 struct ktls_session *tls; 507 int error; 508 509 /* Only TLS 1.0 - 1.3 are supported. */ 510 if (en->tls_vmajor != TLS_MAJOR_VER_ONE) 511 return (EINVAL); 512 if (en->tls_vminor < TLS_MINOR_VER_ZERO || 513 en->tls_vminor > TLS_MINOR_VER_THREE) 514 return (EINVAL); 515 516 if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) 517 return (EINVAL); 518 if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) 519 return (EINVAL); 520 if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) 521 return (EINVAL); 522 523 /* All supported algorithms require a cipher key. */ 524 if (en->cipher_key_len == 0) 525 return (EINVAL); 526 527 /* No flags are currently supported. */ 528 if (en->flags != 0) 529 return (EINVAL); 530 531 /* Common checks for supported algorithms. */ 532 switch (en->cipher_algorithm) { 533 case CRYPTO_AES_NIST_GCM_16: 534 /* 535 * auth_algorithm isn't used, but permit GMAC values 536 * for compatibility. 537 */ 538 switch (en->auth_algorithm) { 539 case 0: 540 #ifdef COMPAT_FREEBSD12 541 /* XXX: Really 13.0-current COMPAT. */ 542 case CRYPTO_AES_128_NIST_GMAC: 543 case CRYPTO_AES_192_NIST_GMAC: 544 case CRYPTO_AES_256_NIST_GMAC: 545 #endif 546 break; 547 default: 548 return (EINVAL); 549 } 550 if (en->auth_key_len != 0) 551 return (EINVAL); 552 switch (en->tls_vminor) { 553 case TLS_MINOR_VER_TWO: 554 if (en->iv_len != TLS_AEAD_GCM_LEN) 555 return (EINVAL); 556 break; 557 case TLS_MINOR_VER_THREE: 558 if (en->iv_len != TLS_1_3_GCM_IV_LEN) 559 return (EINVAL); 560 break; 561 default: 562 return (EINVAL); 563 } 564 break; 565 case CRYPTO_AES_CBC: 566 switch (en->auth_algorithm) { 567 case CRYPTO_SHA1_HMAC: 568 break; 569 case CRYPTO_SHA2_256_HMAC: 570 case CRYPTO_SHA2_384_HMAC: 571 if (en->tls_vminor != TLS_MINOR_VER_TWO) 572 return (EINVAL); 573 break; 574 default: 575 return (EINVAL); 576 } 577 if (en->auth_key_len == 0) 578 return (EINVAL); 579 580 /* 581 * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2 582 * use explicit IVs. 583 */ 584 switch (en->tls_vminor) { 585 case TLS_MINOR_VER_ZERO: 586 if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) 587 return (EINVAL); 588 break; 589 case TLS_MINOR_VER_ONE: 590 case TLS_MINOR_VER_TWO: 591 /* Ignore any supplied IV. */ 592 en->iv_len = 0; 593 break; 594 default: 595 return (EINVAL); 596 } 597 break; 598 case CRYPTO_CHACHA20_POLY1305: 599 if (en->auth_algorithm != 0 || en->auth_key_len != 0) 600 return (EINVAL); 601 if (en->tls_vminor != TLS_MINOR_VER_TWO && 602 en->tls_vminor != TLS_MINOR_VER_THREE) 603 return (EINVAL); 604 if (en->iv_len != TLS_CHACHA20_IV_LEN) 605 return (EINVAL); 606 break; 607 default: 608 return (EINVAL); 609 } 610 611 error = ktls_start_kthreads(); 612 if (error != 0) 613 return (error); 614 615 tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 616 617 counter_u64_add(ktls_offload_active, 1); 618 619 refcount_init(&tls->refcount, 1); 620 if (direction == KTLS_RX) { 621 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls); 622 } else { 623 TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); 624 tls->inp = so->so_pcb; 625 in_pcbref(tls->inp); 626 tls->tx = true; 627 } 628 629 tls->wq_index = ktls_get_cpu(so); 630 631 tls->params.cipher_algorithm = en->cipher_algorithm; 632 tls->params.auth_algorithm = en->auth_algorithm; 633 tls->params.tls_vmajor = en->tls_vmajor; 634 tls->params.tls_vminor = en->tls_vminor; 635 tls->params.flags = en->flags; 636 tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); 637 638 /* Set the header and trailer lengths. */ 639 tls->params.tls_hlen = sizeof(struct tls_record_layer); 640 switch (en->cipher_algorithm) { 641 case CRYPTO_AES_NIST_GCM_16: 642 /* 643 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte 644 * nonce. TLS 1.3 uses a 12 byte implicit IV. 645 */ 646 if (en->tls_vminor < TLS_MINOR_VER_THREE) 647 tls->params.tls_hlen += sizeof(uint64_t); 648 tls->params.tls_tlen = AES_GMAC_HASH_LEN; 649 tls->params.tls_bs = 1; 650 break; 651 case CRYPTO_AES_CBC: 652 switch (en->auth_algorithm) { 653 case CRYPTO_SHA1_HMAC: 654 if (en->tls_vminor == TLS_MINOR_VER_ZERO) { 655 /* Implicit IV, no nonce. */ 656 tls->sequential_records = true; 657 tls->next_seqno = be64dec(en->rec_seq); 658 STAILQ_INIT(&tls->pending_records); 659 } else { 660 tls->params.tls_hlen += AES_BLOCK_LEN; 661 } 662 tls->params.tls_tlen = AES_BLOCK_LEN + 663 SHA1_HASH_LEN; 664 break; 665 case CRYPTO_SHA2_256_HMAC: 666 tls->params.tls_hlen += AES_BLOCK_LEN; 667 tls->params.tls_tlen = AES_BLOCK_LEN + 668 SHA2_256_HASH_LEN; 669 break; 670 case CRYPTO_SHA2_384_HMAC: 671 tls->params.tls_hlen += AES_BLOCK_LEN; 672 tls->params.tls_tlen = AES_BLOCK_LEN + 673 SHA2_384_HASH_LEN; 674 break; 675 default: 676 panic("invalid hmac"); 677 } 678 tls->params.tls_bs = AES_BLOCK_LEN; 679 break; 680 case CRYPTO_CHACHA20_POLY1305: 681 /* 682 * Chacha20 uses a 12 byte implicit IV. 683 */ 684 tls->params.tls_tlen = POLY1305_HASH_LEN; 685 tls->params.tls_bs = 1; 686 break; 687 default: 688 panic("invalid cipher"); 689 } 690 691 /* 692 * TLS 1.3 includes optional padding which we do not support, 693 * and also puts the "real" record type at the end of the 694 * encrypted data. 695 */ 696 if (en->tls_vminor == TLS_MINOR_VER_THREE) 697 tls->params.tls_tlen += sizeof(uint8_t); 698 699 KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, 700 ("TLS header length too long: %d", tls->params.tls_hlen)); 701 KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, 702 ("TLS trailer length too long: %d", tls->params.tls_tlen)); 703 704 if (en->auth_key_len != 0) { 705 tls->params.auth_key_len = en->auth_key_len; 706 tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, 707 M_WAITOK); 708 error = copyin(en->auth_key, tls->params.auth_key, 709 en->auth_key_len); 710 if (error) 711 goto out; 712 } 713 714 tls->params.cipher_key_len = en->cipher_key_len; 715 tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); 716 error = copyin(en->cipher_key, tls->params.cipher_key, 717 en->cipher_key_len); 718 if (error) 719 goto out; 720 721 /* 722 * This holds the implicit portion of the nonce for AEAD 723 * ciphers and the initial implicit IV for TLS 1.0. The 724 * explicit portions of the IV are generated in ktls_frame(). 725 */ 726 if (en->iv_len != 0) { 727 tls->params.iv_len = en->iv_len; 728 error = copyin(en->iv, tls->params.iv, en->iv_len); 729 if (error) 730 goto out; 731 732 /* 733 * For TLS 1.2 with GCM, generate an 8-byte nonce as a 734 * counter to generate unique explicit IVs. 735 * 736 * Store this counter in the last 8 bytes of the IV 737 * array so that it is 8-byte aligned. 738 */ 739 if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 740 en->tls_vminor == TLS_MINOR_VER_TWO) 741 arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); 742 } 743 744 *tlsp = tls; 745 return (0); 746 747 out: 748 ktls_free(tls); 749 return (error); 750 } 751 752 static struct ktls_session * 753 ktls_clone_session(struct ktls_session *tls, int direction) 754 { 755 struct ktls_session *tls_new; 756 757 tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); 758 759 counter_u64_add(ktls_offload_active, 1); 760 761 refcount_init(&tls_new->refcount, 1); 762 if (direction == KTLS_RX) { 763 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag, 764 tls_new); 765 } else { 766 TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, 767 tls_new); 768 tls_new->inp = tls->inp; 769 tls_new->tx = true; 770 in_pcbref(tls_new->inp); 771 } 772 773 /* Copy fields from existing session. */ 774 tls_new->params = tls->params; 775 tls_new->wq_index = tls->wq_index; 776 777 /* Deep copy keys. */ 778 if (tls_new->params.auth_key != NULL) { 779 tls_new->params.auth_key = malloc(tls->params.auth_key_len, 780 M_KTLS, M_WAITOK); 781 memcpy(tls_new->params.auth_key, tls->params.auth_key, 782 tls->params.auth_key_len); 783 } 784 785 tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, 786 M_WAITOK); 787 memcpy(tls_new->params.cipher_key, tls->params.cipher_key, 788 tls->params.cipher_key_len); 789 790 return (tls_new); 791 } 792 793 #ifdef TCP_OFFLOAD 794 static int 795 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) 796 { 797 struct inpcb *inp; 798 struct tcpcb *tp; 799 int error; 800 801 inp = so->so_pcb; 802 INP_WLOCK(inp); 803 if (inp->inp_flags & INP_DROPPED) { 804 INP_WUNLOCK(inp); 805 return (ECONNRESET); 806 } 807 if (inp->inp_socket == NULL) { 808 INP_WUNLOCK(inp); 809 return (ECONNRESET); 810 } 811 tp = intotcpcb(inp); 812 if (!(tp->t_flags & TF_TOE)) { 813 INP_WUNLOCK(inp); 814 return (EOPNOTSUPP); 815 } 816 817 error = tcp_offload_alloc_tls_session(tp, tls, direction); 818 INP_WUNLOCK(inp); 819 if (error == 0) { 820 tls->mode = TCP_TLS_MODE_TOE; 821 switch (tls->params.cipher_algorithm) { 822 case CRYPTO_AES_CBC: 823 counter_u64_add(ktls_toe_cbc, 1); 824 break; 825 case CRYPTO_AES_NIST_GCM_16: 826 counter_u64_add(ktls_toe_gcm, 1); 827 break; 828 case CRYPTO_CHACHA20_POLY1305: 829 counter_u64_add(ktls_toe_chacha20, 1); 830 break; 831 } 832 } 833 return (error); 834 } 835 #endif 836 837 /* 838 * Common code used when first enabling ifnet TLS on a connection or 839 * when allocating a new ifnet TLS session due to a routing change. 840 * This function allocates a new TLS send tag on whatever interface 841 * the connection is currently routed over. 842 */ 843 static int 844 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, 845 struct m_snd_tag **mstp) 846 { 847 union if_snd_tag_alloc_params params; 848 struct ifnet *ifp; 849 struct nhop_object *nh; 850 struct tcpcb *tp; 851 int error; 852 853 INP_RLOCK(inp); 854 if (inp->inp_flags & INP_DROPPED) { 855 INP_RUNLOCK(inp); 856 return (ECONNRESET); 857 } 858 if (inp->inp_socket == NULL) { 859 INP_RUNLOCK(inp); 860 return (ECONNRESET); 861 } 862 tp = intotcpcb(inp); 863 864 /* 865 * Check administrative controls on ifnet TLS to determine if 866 * ifnet TLS should be denied. 867 * 868 * - Always permit 'force' requests. 869 * - ktls_ifnet_permitted == 0: always deny. 870 */ 871 if (!force && ktls_ifnet_permitted == 0) { 872 INP_RUNLOCK(inp); 873 return (ENXIO); 874 } 875 876 /* 877 * XXX: Use the cached route in the inpcb to find the 878 * interface. This should perhaps instead use 879 * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only 880 * enabled after a connection has completed key negotiation in 881 * userland, the cached route will be present in practice. 882 */ 883 nh = inp->inp_route.ro_nh; 884 if (nh == NULL) { 885 INP_RUNLOCK(inp); 886 return (ENXIO); 887 } 888 ifp = nh->nh_ifp; 889 if_ref(ifp); 890 891 /* 892 * Allocate a TLS + ratelimit tag if the connection has an 893 * existing pacing rate. 894 */ 895 if (tp->t_pacing_rate != -1 && 896 (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) { 897 params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT; 898 params.tls_rate_limit.inp = inp; 899 params.tls_rate_limit.tls = tls; 900 params.tls_rate_limit.max_rate = tp->t_pacing_rate; 901 } else { 902 params.hdr.type = IF_SND_TAG_TYPE_TLS; 903 params.tls.inp = inp; 904 params.tls.tls = tls; 905 } 906 params.hdr.flowid = inp->inp_flowid; 907 params.hdr.flowtype = inp->inp_flowtype; 908 params.hdr.numa_domain = inp->inp_numa_domain; 909 INP_RUNLOCK(inp); 910 911 if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) { 912 error = EOPNOTSUPP; 913 goto out; 914 } 915 if (inp->inp_vflag & INP_IPV6) { 916 if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) { 917 error = EOPNOTSUPP; 918 goto out; 919 } 920 } else { 921 if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) { 922 error = EOPNOTSUPP; 923 goto out; 924 } 925 } 926 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 927 out: 928 if_rele(ifp); 929 return (error); 930 } 931 932 /* 933 * Allocate an initial TLS receive tag for doing HW decryption of TLS 934 * data. 935 * 936 * This function allocates a new TLS receive tag on whatever interface 937 * the connection is currently routed over. If the connection ends up 938 * using a different interface for receive this will get fixed up via 939 * ktls_input_ifp_mismatch as future packets arrive. 940 */ 941 static int 942 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls, 943 struct m_snd_tag **mstp) 944 { 945 union if_snd_tag_alloc_params params; 946 struct ifnet *ifp; 947 struct nhop_object *nh; 948 int error; 949 950 if (!ktls_ocf_recrypt_supported(tls)) 951 return (ENXIO); 952 953 INP_RLOCK(inp); 954 if (inp->inp_flags & INP_DROPPED) { 955 INP_RUNLOCK(inp); 956 return (ECONNRESET); 957 } 958 if (inp->inp_socket == NULL) { 959 INP_RUNLOCK(inp); 960 return (ECONNRESET); 961 } 962 963 /* 964 * Check administrative controls on ifnet TLS to determine if 965 * ifnet TLS should be denied. 966 */ 967 if (ktls_ifnet_permitted == 0) { 968 INP_RUNLOCK(inp); 969 return (ENXIO); 970 } 971 972 /* 973 * XXX: As with ktls_alloc_snd_tag, use the cached route in 974 * the inpcb to find the interface. 975 */ 976 nh = inp->inp_route.ro_nh; 977 if (nh == NULL) { 978 INP_RUNLOCK(inp); 979 return (ENXIO); 980 } 981 ifp = nh->nh_ifp; 982 if_ref(ifp); 983 tls->rx_ifp = ifp; 984 985 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; 986 params.hdr.flowid = inp->inp_flowid; 987 params.hdr.flowtype = inp->inp_flowtype; 988 params.hdr.numa_domain = inp->inp_numa_domain; 989 params.tls_rx.inp = inp; 990 params.tls_rx.tls = tls; 991 params.tls_rx.vlan_id = 0; 992 993 INP_RUNLOCK(inp); 994 995 if (inp->inp_vflag & INP_IPV6) { 996 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) { 997 error = EOPNOTSUPP; 998 goto out; 999 } 1000 } else { 1001 if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) { 1002 error = EOPNOTSUPP; 1003 goto out; 1004 } 1005 } 1006 error = m_snd_tag_alloc(ifp, ¶ms, mstp); 1007 1008 /* 1009 * If this connection is over a vlan, vlan_snd_tag_alloc 1010 * rewrites vlan_id with the saved interface. Save the VLAN 1011 * ID for use in ktls_reset_receive_tag which allocates new 1012 * receive tags directly from the leaf interface bypassing 1013 * if_vlan. 1014 */ 1015 if (error == 0) 1016 tls->rx_vlan_id = params.tls_rx.vlan_id; 1017 out: 1018 return (error); 1019 } 1020 1021 static int 1022 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction, 1023 bool force) 1024 { 1025 struct m_snd_tag *mst; 1026 int error; 1027 1028 switch (direction) { 1029 case KTLS_TX: 1030 error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); 1031 if (__predict_false(error != 0)) 1032 goto done; 1033 break; 1034 case KTLS_RX: 1035 KASSERT(!force, ("%s: forced receive tag", __func__)); 1036 error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst); 1037 if (__predict_false(error != 0)) 1038 goto done; 1039 break; 1040 default: 1041 __assert_unreachable(); 1042 } 1043 1044 tls->mode = TCP_TLS_MODE_IFNET; 1045 tls->snd_tag = mst; 1046 1047 switch (tls->params.cipher_algorithm) { 1048 case CRYPTO_AES_CBC: 1049 counter_u64_add(ktls_ifnet_cbc, 1); 1050 break; 1051 case CRYPTO_AES_NIST_GCM_16: 1052 counter_u64_add(ktls_ifnet_gcm, 1); 1053 break; 1054 case CRYPTO_CHACHA20_POLY1305: 1055 counter_u64_add(ktls_ifnet_chacha20, 1); 1056 break; 1057 default: 1058 break; 1059 } 1060 done: 1061 return (error); 1062 } 1063 1064 static void 1065 ktls_use_sw(struct ktls_session *tls) 1066 { 1067 tls->mode = TCP_TLS_MODE_SW; 1068 switch (tls->params.cipher_algorithm) { 1069 case CRYPTO_AES_CBC: 1070 counter_u64_add(ktls_sw_cbc, 1); 1071 break; 1072 case CRYPTO_AES_NIST_GCM_16: 1073 counter_u64_add(ktls_sw_gcm, 1); 1074 break; 1075 case CRYPTO_CHACHA20_POLY1305: 1076 counter_u64_add(ktls_sw_chacha20, 1); 1077 break; 1078 } 1079 } 1080 1081 static int 1082 ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction) 1083 { 1084 int error; 1085 1086 error = ktls_ocf_try(so, tls, direction); 1087 if (error) 1088 return (error); 1089 ktls_use_sw(tls); 1090 return (0); 1091 } 1092 1093 /* 1094 * KTLS RX stores data in the socket buffer as a list of TLS records, 1095 * where each record is stored as a control message containg the TLS 1096 * header followed by data mbufs containing the decrypted data. This 1097 * is different from KTLS TX which always uses an mb_ext_pgs mbuf for 1098 * both encrypted and decrypted data. TLS records decrypted by a NIC 1099 * should be queued to the socket buffer as records, but encrypted 1100 * data which needs to be decrypted by software arrives as a stream of 1101 * regular mbufs which need to be converted. In addition, there may 1102 * already be pending encrypted data in the socket buffer when KTLS RX 1103 * is enabled. 1104 * 1105 * To manage not-yet-decrypted data for KTLS RX, the following scheme 1106 * is used: 1107 * 1108 * - A single chain of NOTREADY mbufs is hung off of sb_mtls. 1109 * 1110 * - ktls_check_rx checks this chain of mbufs reading the TLS header 1111 * from the first mbuf. Once all of the data for that TLS record is 1112 * queued, the socket is queued to a worker thread. 1113 * 1114 * - The worker thread calls ktls_decrypt to decrypt TLS records in 1115 * the TLS chain. Each TLS record is detached from the TLS chain, 1116 * decrypted, and inserted into the regular socket buffer chain as 1117 * record starting with a control message holding the TLS header and 1118 * a chain of mbufs holding the encrypted data. 1119 */ 1120 1121 static void 1122 sb_mark_notready(struct sockbuf *sb) 1123 { 1124 struct mbuf *m; 1125 1126 m = sb->sb_mb; 1127 sb->sb_mtls = m; 1128 sb->sb_mb = NULL; 1129 sb->sb_mbtail = NULL; 1130 sb->sb_lastrecord = NULL; 1131 for (; m != NULL; m = m->m_next) { 1132 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", 1133 __func__)); 1134 KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", 1135 __func__)); 1136 KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", 1137 __func__)); 1138 m->m_flags |= M_NOTREADY; 1139 sb->sb_acc -= m->m_len; 1140 sb->sb_tlscc += m->m_len; 1141 sb->sb_mtlstail = m; 1142 } 1143 KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, 1144 ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, 1145 sb->sb_ccc)); 1146 } 1147 1148 /* 1149 * Return information about the pending TLS data in a socket 1150 * buffer. On return, 'seqno' is set to the sequence number 1151 * of the next TLS record to be received, 'resid' is set to 1152 * the amount of bytes still needed for the last pending 1153 * record. The function returns 'false' if the last pending 1154 * record contains a partial TLS header. In that case, 'resid' 1155 * is the number of bytes needed to complete the TLS header. 1156 */ 1157 bool 1158 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp) 1159 { 1160 struct tls_record_layer hdr; 1161 struct mbuf *m; 1162 uint64_t seqno; 1163 size_t resid; 1164 u_int offset, record_len; 1165 1166 SOCKBUF_LOCK_ASSERT(sb); 1167 MPASS(sb->sb_flags & SB_TLS_RX); 1168 seqno = sb->sb_tls_seqno; 1169 resid = sb->sb_tlscc; 1170 m = sb->sb_mtls; 1171 offset = 0; 1172 1173 if (resid == 0) { 1174 *seqnop = seqno; 1175 *residp = 0; 1176 return (true); 1177 } 1178 1179 for (;;) { 1180 seqno++; 1181 1182 if (resid < sizeof(hdr)) { 1183 *seqnop = seqno; 1184 *residp = sizeof(hdr) - resid; 1185 return (false); 1186 } 1187 1188 m_copydata(m, offset, sizeof(hdr), (void *)&hdr); 1189 1190 record_len = sizeof(hdr) + ntohs(hdr.tls_length); 1191 if (resid <= record_len) { 1192 *seqnop = seqno; 1193 *residp = record_len - resid; 1194 return (true); 1195 } 1196 resid -= record_len; 1197 1198 while (record_len != 0) { 1199 if (m->m_len - offset > record_len) { 1200 offset += record_len; 1201 break; 1202 } 1203 1204 record_len -= (m->m_len - offset); 1205 offset = 0; 1206 m = m->m_next; 1207 } 1208 } 1209 } 1210 1211 int 1212 ktls_enable_rx(struct socket *so, struct tls_enable *en) 1213 { 1214 struct ktls_session *tls; 1215 int error; 1216 1217 if (!ktls_offload_enable) 1218 return (ENOTSUP); 1219 1220 counter_u64_add(ktls_offload_enable_calls, 1); 1221 1222 /* 1223 * This should always be true since only the TCP socket option 1224 * invokes this function. 1225 */ 1226 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1227 return (EINVAL); 1228 1229 /* 1230 * XXX: Don't overwrite existing sessions. We should permit 1231 * this to support rekeying in the future. 1232 */ 1233 if (so->so_rcv.sb_tls_info != NULL) 1234 return (EALREADY); 1235 1236 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1237 return (ENOTSUP); 1238 1239 error = ktls_create_session(so, en, &tls, KTLS_RX); 1240 if (error) 1241 return (error); 1242 1243 error = ktls_ocf_try(so, tls, KTLS_RX); 1244 if (error) { 1245 ktls_free(tls); 1246 return (error); 1247 } 1248 1249 /* Mark the socket as using TLS offload. */ 1250 SOCK_RECVBUF_LOCK(so); 1251 if (SOLISTENING(so)) { 1252 SOCK_RECVBUF_UNLOCK(so); 1253 ktls_free(tls); 1254 return (EINVAL); 1255 } 1256 so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); 1257 so->so_rcv.sb_tls_info = tls; 1258 so->so_rcv.sb_flags |= SB_TLS_RX; 1259 1260 /* Mark existing data as not ready until it can be decrypted. */ 1261 sb_mark_notready(&so->so_rcv); 1262 ktls_check_rx(&so->so_rcv); 1263 SOCK_RECVBUF_UNLOCK(so); 1264 1265 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1266 #ifdef TCP_OFFLOAD 1267 error = ktls_try_toe(so, tls, KTLS_RX); 1268 if (error) 1269 #endif 1270 error = ktls_try_ifnet(so, tls, KTLS_RX, false); 1271 if (error) 1272 ktls_use_sw(tls); 1273 1274 counter_u64_add(ktls_offload_total, 1); 1275 1276 return (0); 1277 } 1278 1279 int 1280 ktls_enable_tx(struct socket *so, struct tls_enable *en) 1281 { 1282 struct ktls_session *tls; 1283 struct inpcb *inp; 1284 struct tcpcb *tp; 1285 int error; 1286 1287 if (!ktls_offload_enable) 1288 return (ENOTSUP); 1289 1290 counter_u64_add(ktls_offload_enable_calls, 1); 1291 1292 /* 1293 * This should always be true since only the TCP socket option 1294 * invokes this function. 1295 */ 1296 if (so->so_proto->pr_protocol != IPPROTO_TCP) 1297 return (EINVAL); 1298 1299 /* 1300 * XXX: Don't overwrite existing sessions. We should permit 1301 * this to support rekeying in the future. 1302 */ 1303 if (so->so_snd.sb_tls_info != NULL) 1304 return (EALREADY); 1305 1306 if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) 1307 return (ENOTSUP); 1308 1309 /* TLS requires ext pgs */ 1310 if (mb_use_ext_pgs == 0) 1311 return (ENXIO); 1312 1313 error = ktls_create_session(so, en, &tls, KTLS_TX); 1314 if (error) 1315 return (error); 1316 1317 /* Prefer TOE -> ifnet TLS -> software TLS. */ 1318 #ifdef TCP_OFFLOAD 1319 error = ktls_try_toe(so, tls, KTLS_TX); 1320 if (error) 1321 #endif 1322 error = ktls_try_ifnet(so, tls, KTLS_TX, false); 1323 if (error) 1324 error = ktls_try_sw(so, tls, KTLS_TX); 1325 1326 if (error) { 1327 ktls_free(tls); 1328 return (error); 1329 } 1330 1331 /* 1332 * Serialize with sosend_generic() and make sure that we're not 1333 * operating on a listening socket. 1334 */ 1335 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1336 if (error) { 1337 ktls_free(tls); 1338 return (error); 1339 } 1340 1341 /* 1342 * Write lock the INP when setting sb_tls_info so that 1343 * routines in tcp_ratelimit.c can read sb_tls_info while 1344 * holding the INP lock. 1345 */ 1346 inp = so->so_pcb; 1347 INP_WLOCK(inp); 1348 SOCK_SENDBUF_LOCK(so); 1349 so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); 1350 so->so_snd.sb_tls_info = tls; 1351 if (tls->mode != TCP_TLS_MODE_SW) { 1352 tp = intotcpcb(inp); 1353 MPASS(tp->t_nic_ktls_xmit == 0); 1354 tp->t_nic_ktls_xmit = 1; 1355 if (tp->t_fb->tfb_hwtls_change != NULL) 1356 (*tp->t_fb->tfb_hwtls_change)(tp, 1); 1357 } 1358 SOCK_SENDBUF_UNLOCK(so); 1359 INP_WUNLOCK(inp); 1360 SOCK_IO_SEND_UNLOCK(so); 1361 1362 counter_u64_add(ktls_offload_total, 1); 1363 1364 return (0); 1365 } 1366 1367 int 1368 ktls_get_rx_mode(struct socket *so, int *modep) 1369 { 1370 struct ktls_session *tls; 1371 struct inpcb *inp __diagused; 1372 1373 if (SOLISTENING(so)) 1374 return (EINVAL); 1375 inp = so->so_pcb; 1376 INP_WLOCK_ASSERT(inp); 1377 SOCK_RECVBUF_LOCK(so); 1378 tls = so->so_rcv.sb_tls_info; 1379 if (tls == NULL) 1380 *modep = TCP_TLS_MODE_NONE; 1381 else 1382 *modep = tls->mode; 1383 SOCK_RECVBUF_UNLOCK(so); 1384 return (0); 1385 } 1386 1387 /* 1388 * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number. 1389 * 1390 * This function gets information about the next TCP- and TLS- 1391 * sequence number to be processed by the TLS receive worker 1392 * thread. The information is extracted from the given "inpcb" 1393 * structure. The values are stored in host endian format at the two 1394 * given output pointer locations. The TCP sequence number points to 1395 * the beginning of the TLS header. 1396 * 1397 * This function returns zero on success, else a non-zero error code 1398 * is returned. 1399 */ 1400 int 1401 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq) 1402 { 1403 struct socket *so; 1404 struct tcpcb *tp; 1405 1406 INP_RLOCK(inp); 1407 so = inp->inp_socket; 1408 if (__predict_false(so == NULL)) { 1409 INP_RUNLOCK(inp); 1410 return (EINVAL); 1411 } 1412 if (inp->inp_flags & INP_DROPPED) { 1413 INP_RUNLOCK(inp); 1414 return (ECONNRESET); 1415 } 1416 1417 tp = intotcpcb(inp); 1418 MPASS(tp != NULL); 1419 1420 SOCKBUF_LOCK(&so->so_rcv); 1421 *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc; 1422 *tlsseq = so->so_rcv.sb_tls_seqno; 1423 SOCKBUF_UNLOCK(&so->so_rcv); 1424 1425 INP_RUNLOCK(inp); 1426 1427 return (0); 1428 } 1429 1430 int 1431 ktls_get_tx_mode(struct socket *so, int *modep) 1432 { 1433 struct ktls_session *tls; 1434 struct inpcb *inp __diagused; 1435 1436 if (SOLISTENING(so)) 1437 return (EINVAL); 1438 inp = so->so_pcb; 1439 INP_WLOCK_ASSERT(inp); 1440 SOCK_SENDBUF_LOCK(so); 1441 tls = so->so_snd.sb_tls_info; 1442 if (tls == NULL) 1443 *modep = TCP_TLS_MODE_NONE; 1444 else 1445 *modep = tls->mode; 1446 SOCK_SENDBUF_UNLOCK(so); 1447 return (0); 1448 } 1449 1450 /* 1451 * Switch between SW and ifnet TLS sessions as requested. 1452 */ 1453 int 1454 ktls_set_tx_mode(struct socket *so, int mode) 1455 { 1456 struct ktls_session *tls, *tls_new; 1457 struct inpcb *inp; 1458 struct tcpcb *tp; 1459 int error; 1460 1461 if (SOLISTENING(so)) 1462 return (EINVAL); 1463 switch (mode) { 1464 case TCP_TLS_MODE_SW: 1465 case TCP_TLS_MODE_IFNET: 1466 break; 1467 default: 1468 return (EINVAL); 1469 } 1470 1471 inp = so->so_pcb; 1472 INP_WLOCK_ASSERT(inp); 1473 tp = intotcpcb(inp); 1474 1475 if (mode == TCP_TLS_MODE_IFNET) { 1476 /* Don't allow enabling ifnet ktls multiple times */ 1477 if (tp->t_nic_ktls_xmit) 1478 return (EALREADY); 1479 1480 /* 1481 * Don't enable ifnet ktls if we disabled it due to an 1482 * excessive retransmission rate 1483 */ 1484 if (tp->t_nic_ktls_xmit_dis) 1485 return (ENXIO); 1486 } 1487 1488 SOCKBUF_LOCK(&so->so_snd); 1489 tls = so->so_snd.sb_tls_info; 1490 if (tls == NULL) { 1491 SOCKBUF_UNLOCK(&so->so_snd); 1492 return (0); 1493 } 1494 1495 if (tls->mode == mode) { 1496 SOCKBUF_UNLOCK(&so->so_snd); 1497 return (0); 1498 } 1499 1500 tls = ktls_hold(tls); 1501 SOCKBUF_UNLOCK(&so->so_snd); 1502 INP_WUNLOCK(inp); 1503 1504 tls_new = ktls_clone_session(tls, KTLS_TX); 1505 1506 if (mode == TCP_TLS_MODE_IFNET) 1507 error = ktls_try_ifnet(so, tls_new, KTLS_TX, true); 1508 else 1509 error = ktls_try_sw(so, tls_new, KTLS_TX); 1510 if (error) { 1511 counter_u64_add(ktls_switch_failed, 1); 1512 ktls_free(tls_new); 1513 ktls_free(tls); 1514 INP_WLOCK(inp); 1515 return (error); 1516 } 1517 1518 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); 1519 if (error) { 1520 counter_u64_add(ktls_switch_failed, 1); 1521 ktls_free(tls_new); 1522 ktls_free(tls); 1523 INP_WLOCK(inp); 1524 return (error); 1525 } 1526 1527 /* 1528 * If we raced with another session change, keep the existing 1529 * session. 1530 */ 1531 if (tls != so->so_snd.sb_tls_info) { 1532 counter_u64_add(ktls_switch_failed, 1); 1533 SOCK_IO_SEND_UNLOCK(so); 1534 ktls_free(tls_new); 1535 ktls_free(tls); 1536 INP_WLOCK(inp); 1537 return (EBUSY); 1538 } 1539 1540 INP_WLOCK(inp); 1541 SOCKBUF_LOCK(&so->so_snd); 1542 so->so_snd.sb_tls_info = tls_new; 1543 if (tls_new->mode != TCP_TLS_MODE_SW) { 1544 MPASS(tp->t_nic_ktls_xmit == 0); 1545 tp->t_nic_ktls_xmit = 1; 1546 if (tp->t_fb->tfb_hwtls_change != NULL) 1547 (*tp->t_fb->tfb_hwtls_change)(tp, 1); 1548 } 1549 SOCKBUF_UNLOCK(&so->so_snd); 1550 SOCK_IO_SEND_UNLOCK(so); 1551 1552 /* 1553 * Drop two references on 'tls'. The first is for the 1554 * ktls_hold() above. The second drops the reference from the 1555 * socket buffer. 1556 */ 1557 KASSERT(tls->refcount >= 2, ("too few references on old session")); 1558 ktls_free(tls); 1559 ktls_free(tls); 1560 1561 if (mode == TCP_TLS_MODE_IFNET) 1562 counter_u64_add(ktls_switch_to_ifnet, 1); 1563 else 1564 counter_u64_add(ktls_switch_to_sw, 1); 1565 1566 return (0); 1567 } 1568 1569 /* 1570 * Try to allocate a new TLS receive tag. This task is scheduled when 1571 * sbappend_ktls_rx detects an input path change. If a new tag is 1572 * allocated, replace the tag in the TLS session. If a new tag cannot 1573 * be allocated, let the session fall back to software decryption. 1574 */ 1575 static void 1576 ktls_reset_receive_tag(void *context, int pending) 1577 { 1578 union if_snd_tag_alloc_params params; 1579 struct ktls_session *tls; 1580 struct m_snd_tag *mst; 1581 struct inpcb *inp; 1582 struct ifnet *ifp; 1583 struct socket *so; 1584 int error; 1585 1586 MPASS(pending == 1); 1587 1588 tls = context; 1589 so = tls->so; 1590 inp = so->so_pcb; 1591 ifp = NULL; 1592 1593 INP_RLOCK(inp); 1594 if (inp->inp_flags & INP_DROPPED) { 1595 INP_RUNLOCK(inp); 1596 goto out; 1597 } 1598 1599 SOCKBUF_LOCK(&so->so_rcv); 1600 mst = tls->snd_tag; 1601 tls->snd_tag = NULL; 1602 if (mst != NULL) 1603 m_snd_tag_rele(mst); 1604 1605 ifp = tls->rx_ifp; 1606 if_ref(ifp); 1607 SOCKBUF_UNLOCK(&so->so_rcv); 1608 1609 params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; 1610 params.hdr.flowid = inp->inp_flowid; 1611 params.hdr.flowtype = inp->inp_flowtype; 1612 params.hdr.numa_domain = inp->inp_numa_domain; 1613 params.tls_rx.inp = inp; 1614 params.tls_rx.tls = tls; 1615 params.tls_rx.vlan_id = tls->rx_vlan_id; 1616 INP_RUNLOCK(inp); 1617 1618 if (inp->inp_vflag & INP_IPV6) { 1619 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0) 1620 goto out; 1621 } else { 1622 if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0) 1623 goto out; 1624 } 1625 1626 error = m_snd_tag_alloc(ifp, ¶ms, &mst); 1627 if (error == 0) { 1628 SOCKBUF_LOCK(&so->so_rcv); 1629 tls->snd_tag = mst; 1630 SOCKBUF_UNLOCK(&so->so_rcv); 1631 1632 counter_u64_add(ktls_ifnet_reset, 1); 1633 } else { 1634 /* 1635 * Just fall back to software decryption if a tag 1636 * cannot be allocated leaving the connection intact. 1637 * If a future input path change switches to another 1638 * interface this connection will resume ifnet TLS. 1639 */ 1640 counter_u64_add(ktls_ifnet_reset_failed, 1); 1641 } 1642 1643 out: 1644 mtx_pool_lock(mtxpool_sleep, tls); 1645 tls->reset_pending = false; 1646 mtx_pool_unlock(mtxpool_sleep, tls); 1647 1648 if (ifp != NULL) 1649 if_rele(ifp); 1650 sorele(so); 1651 ktls_free(tls); 1652 } 1653 1654 /* 1655 * Try to allocate a new TLS send tag. This task is scheduled when 1656 * ip_output detects a route change while trying to transmit a packet 1657 * holding a TLS record. If a new tag is allocated, replace the tag 1658 * in the TLS session. Subsequent packets on the connection will use 1659 * the new tag. If a new tag cannot be allocated, drop the 1660 * connection. 1661 */ 1662 static void 1663 ktls_reset_send_tag(void *context, int pending) 1664 { 1665 struct epoch_tracker et; 1666 struct ktls_session *tls; 1667 struct m_snd_tag *old, *new; 1668 struct inpcb *inp; 1669 struct tcpcb *tp; 1670 int error; 1671 1672 MPASS(pending == 1); 1673 1674 tls = context; 1675 inp = tls->inp; 1676 1677 /* 1678 * Free the old tag first before allocating a new one. 1679 * ip[6]_output_send() will treat a NULL send tag the same as 1680 * an ifp mismatch and drop packets until a new tag is 1681 * allocated. 1682 * 1683 * Write-lock the INP when changing tls->snd_tag since 1684 * ip[6]_output_send() holds a read-lock when reading the 1685 * pointer. 1686 */ 1687 INP_WLOCK(inp); 1688 old = tls->snd_tag; 1689 tls->snd_tag = NULL; 1690 INP_WUNLOCK(inp); 1691 if (old != NULL) 1692 m_snd_tag_rele(old); 1693 1694 error = ktls_alloc_snd_tag(inp, tls, true, &new); 1695 1696 if (error == 0) { 1697 INP_WLOCK(inp); 1698 tls->snd_tag = new; 1699 mtx_pool_lock(mtxpool_sleep, tls); 1700 tls->reset_pending = false; 1701 mtx_pool_unlock(mtxpool_sleep, tls); 1702 INP_WUNLOCK(inp); 1703 1704 counter_u64_add(ktls_ifnet_reset, 1); 1705 1706 /* 1707 * XXX: Should we kick tcp_output explicitly now that 1708 * the send tag is fixed or just rely on timers? 1709 */ 1710 } else { 1711 NET_EPOCH_ENTER(et); 1712 INP_WLOCK(inp); 1713 if (!(inp->inp_flags & INP_DROPPED)) { 1714 tp = intotcpcb(inp); 1715 CURVNET_SET(inp->inp_vnet); 1716 tp = tcp_drop(tp, ECONNABORTED); 1717 CURVNET_RESTORE(); 1718 if (tp != NULL) { 1719 counter_u64_add(ktls_ifnet_reset_dropped, 1); 1720 INP_WUNLOCK(inp); 1721 } 1722 } else 1723 INP_WUNLOCK(inp); 1724 NET_EPOCH_EXIT(et); 1725 1726 counter_u64_add(ktls_ifnet_reset_failed, 1); 1727 1728 /* 1729 * Leave reset_pending true to avoid future tasks while 1730 * the socket goes away. 1731 */ 1732 } 1733 1734 ktls_free(tls); 1735 } 1736 1737 void 1738 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp) 1739 { 1740 struct ktls_session *tls; 1741 struct socket *so; 1742 1743 SOCKBUF_LOCK_ASSERT(sb); 1744 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 1745 __func__, sb)); 1746 so = __containerof(sb, struct socket, so_rcv); 1747 1748 tls = sb->sb_tls_info; 1749 if_rele(tls->rx_ifp); 1750 if_ref(ifp); 1751 tls->rx_ifp = ifp; 1752 1753 /* 1754 * See if we should schedule a task to update the receive tag for 1755 * this session. 1756 */ 1757 mtx_pool_lock(mtxpool_sleep, tls); 1758 if (!tls->reset_pending) { 1759 (void) ktls_hold(tls); 1760 soref(so); 1761 tls->so = so; 1762 tls->reset_pending = true; 1763 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1764 } 1765 mtx_pool_unlock(mtxpool_sleep, tls); 1766 } 1767 1768 int 1769 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) 1770 { 1771 1772 if (inp == NULL) 1773 return (ENOBUFS); 1774 1775 INP_LOCK_ASSERT(inp); 1776 1777 /* 1778 * See if we should schedule a task to update the send tag for 1779 * this session. 1780 */ 1781 mtx_pool_lock(mtxpool_sleep, tls); 1782 if (!tls->reset_pending) { 1783 (void) ktls_hold(tls); 1784 tls->reset_pending = true; 1785 taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); 1786 } 1787 mtx_pool_unlock(mtxpool_sleep, tls); 1788 return (ENOBUFS); 1789 } 1790 1791 #ifdef RATELIMIT 1792 int 1793 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate) 1794 { 1795 union if_snd_tag_modify_params params = { 1796 .rate_limit.max_rate = max_pacing_rate, 1797 .rate_limit.flags = M_NOWAIT, 1798 }; 1799 struct m_snd_tag *mst; 1800 1801 /* Can't get to the inp, but it should be locked. */ 1802 /* INP_LOCK_ASSERT(inp); */ 1803 1804 MPASS(tls->mode == TCP_TLS_MODE_IFNET); 1805 1806 if (tls->snd_tag == NULL) { 1807 /* 1808 * Resetting send tag, ignore this change. The 1809 * pending reset may or may not see this updated rate 1810 * in the tcpcb. If it doesn't, we will just lose 1811 * this rate change. 1812 */ 1813 return (0); 1814 } 1815 1816 mst = tls->snd_tag; 1817 1818 MPASS(mst != NULL); 1819 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT); 1820 1821 return (mst->sw->snd_tag_modify(mst, ¶ms)); 1822 } 1823 #endif 1824 1825 static void 1826 ktls_destroy_help(void *context, int pending __unused) 1827 { 1828 ktls_destroy(context); 1829 } 1830 1831 void 1832 ktls_destroy(struct ktls_session *tls) 1833 { 1834 struct inpcb *inp; 1835 struct tcpcb *tp; 1836 bool wlocked; 1837 1838 MPASS(tls->refcount == 0); 1839 1840 inp = tls->inp; 1841 if (tls->tx) { 1842 wlocked = INP_WLOCKED(inp); 1843 if (!wlocked && !INP_TRY_WLOCK(inp)) { 1844 /* 1845 * rwlocks read locks are anonymous, and there 1846 * is no way to know if our current thread 1847 * holds an rlock on the inp. As a rough 1848 * estimate, check to see if the thread holds 1849 * *any* rlocks at all. If it does not, then we 1850 * know that we don't hold the inp rlock, and 1851 * can safely take the wlock 1852 */ 1853 if (curthread->td_rw_rlocks == 0) { 1854 INP_WLOCK(inp); 1855 } else { 1856 /* 1857 * We might hold the rlock, so let's 1858 * do the destroy in a taskqueue 1859 * context to avoid a potential 1860 * deadlock. This should be very 1861 * rare. 1862 */ 1863 counter_u64_add(ktls_destroy_task, 1); 1864 TASK_INIT(&tls->destroy_task, 0, 1865 ktls_destroy_help, tls); 1866 (void)taskqueue_enqueue(taskqueue_thread, 1867 &tls->destroy_task); 1868 return; 1869 } 1870 } 1871 } 1872 1873 if (tls->sequential_records) { 1874 struct mbuf *m, *n; 1875 int page_count; 1876 1877 STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) { 1878 page_count = m->m_epg_enc_cnt; 1879 while (page_count > 0) { 1880 KASSERT(page_count >= m->m_epg_nrdy, 1881 ("%s: too few pages", __func__)); 1882 page_count -= m->m_epg_nrdy; 1883 m = m_free(m); 1884 } 1885 } 1886 } 1887 1888 counter_u64_add(ktls_offload_active, -1); 1889 switch (tls->mode) { 1890 case TCP_TLS_MODE_SW: 1891 switch (tls->params.cipher_algorithm) { 1892 case CRYPTO_AES_CBC: 1893 counter_u64_add(ktls_sw_cbc, -1); 1894 break; 1895 case CRYPTO_AES_NIST_GCM_16: 1896 counter_u64_add(ktls_sw_gcm, -1); 1897 break; 1898 case CRYPTO_CHACHA20_POLY1305: 1899 counter_u64_add(ktls_sw_chacha20, -1); 1900 break; 1901 } 1902 break; 1903 case TCP_TLS_MODE_IFNET: 1904 switch (tls->params.cipher_algorithm) { 1905 case CRYPTO_AES_CBC: 1906 counter_u64_add(ktls_ifnet_cbc, -1); 1907 break; 1908 case CRYPTO_AES_NIST_GCM_16: 1909 counter_u64_add(ktls_ifnet_gcm, -1); 1910 break; 1911 case CRYPTO_CHACHA20_POLY1305: 1912 counter_u64_add(ktls_ifnet_chacha20, -1); 1913 break; 1914 } 1915 if (tls->snd_tag != NULL) 1916 m_snd_tag_rele(tls->snd_tag); 1917 if (tls->rx_ifp != NULL) 1918 if_rele(tls->rx_ifp); 1919 if (tls->tx) { 1920 INP_WLOCK_ASSERT(inp); 1921 tp = intotcpcb(inp); 1922 MPASS(tp->t_nic_ktls_xmit == 1); 1923 tp->t_nic_ktls_xmit = 0; 1924 } 1925 break; 1926 #ifdef TCP_OFFLOAD 1927 case TCP_TLS_MODE_TOE: 1928 switch (tls->params.cipher_algorithm) { 1929 case CRYPTO_AES_CBC: 1930 counter_u64_add(ktls_toe_cbc, -1); 1931 break; 1932 case CRYPTO_AES_NIST_GCM_16: 1933 counter_u64_add(ktls_toe_gcm, -1); 1934 break; 1935 case CRYPTO_CHACHA20_POLY1305: 1936 counter_u64_add(ktls_toe_chacha20, -1); 1937 break; 1938 } 1939 break; 1940 #endif 1941 } 1942 if (tls->ocf_session != NULL) 1943 ktls_ocf_free(tls); 1944 if (tls->params.auth_key != NULL) { 1945 zfree(tls->params.auth_key, M_KTLS); 1946 tls->params.auth_key = NULL; 1947 tls->params.auth_key_len = 0; 1948 } 1949 if (tls->params.cipher_key != NULL) { 1950 zfree(tls->params.cipher_key, M_KTLS); 1951 tls->params.cipher_key = NULL; 1952 tls->params.cipher_key_len = 0; 1953 } 1954 if (tls->tx) { 1955 INP_WLOCK_ASSERT(inp); 1956 if (!in_pcbrele_wlocked(inp) && !wlocked) 1957 INP_WUNLOCK(inp); 1958 } 1959 explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); 1960 1961 uma_zfree(ktls_session_zone, tls); 1962 } 1963 1964 void 1965 ktls_seq(struct sockbuf *sb, struct mbuf *m) 1966 { 1967 1968 for (; m != NULL; m = m->m_next) { 1969 KASSERT((m->m_flags & M_EXTPG) != 0, 1970 ("ktls_seq: mapped mbuf %p", m)); 1971 1972 m->m_epg_seqno = sb->sb_tls_seqno; 1973 sb->sb_tls_seqno++; 1974 } 1975 } 1976 1977 /* 1978 * Add TLS framing (headers and trailers) to a chain of mbufs. Each 1979 * mbuf in the chain must be an unmapped mbuf. The payload of the 1980 * mbuf must be populated with the payload of each TLS record. 1981 * 1982 * The record_type argument specifies the TLS record type used when 1983 * populating the TLS header. 1984 * 1985 * The enq_count argument on return is set to the number of pages of 1986 * payload data for this entire chain that need to be encrypted via SW 1987 * encryption. The returned value should be passed to ktls_enqueue 1988 * when scheduling encryption of this chain of mbufs. To handle the 1989 * special case of empty fragments for TLS 1.0 sessions, an empty 1990 * fragment counts as one page. 1991 */ 1992 void 1993 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, 1994 uint8_t record_type) 1995 { 1996 struct tls_record_layer *tlshdr; 1997 struct mbuf *m; 1998 uint64_t *noncep; 1999 uint16_t tls_len; 2000 int maxlen __diagused; 2001 2002 maxlen = tls->params.max_frame_len; 2003 *enq_cnt = 0; 2004 for (m = top; m != NULL; m = m->m_next) { 2005 /* 2006 * All mbufs in the chain should be TLS records whose 2007 * payload does not exceed the maximum frame length. 2008 * 2009 * Empty TLS 1.0 records are permitted when using CBC. 2010 */ 2011 KASSERT(m->m_len <= maxlen && m->m_len >= 0 && 2012 (m->m_len > 0 || ktls_permit_empty_frames(tls)), 2013 ("ktls_frame: m %p len %d", m, m->m_len)); 2014 2015 /* 2016 * TLS frames require unmapped mbufs to store session 2017 * info. 2018 */ 2019 KASSERT((m->m_flags & M_EXTPG) != 0, 2020 ("ktls_frame: mapped mbuf %p (top = %p)", m, top)); 2021 2022 tls_len = m->m_len; 2023 2024 /* Save a reference to the session. */ 2025 m->m_epg_tls = ktls_hold(tls); 2026 2027 m->m_epg_hdrlen = tls->params.tls_hlen; 2028 m->m_epg_trllen = tls->params.tls_tlen; 2029 if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { 2030 int bs, delta; 2031 2032 /* 2033 * AES-CBC pads messages to a multiple of the 2034 * block size. Note that the padding is 2035 * applied after the digest and the encryption 2036 * is done on the "plaintext || mac || padding". 2037 * At least one byte of padding is always 2038 * present. 2039 * 2040 * Compute the final trailer length assuming 2041 * at most one block of padding. 2042 * tls->params.tls_tlen is the maximum 2043 * possible trailer length (padding + digest). 2044 * delta holds the number of excess padding 2045 * bytes if the maximum were used. Those 2046 * extra bytes are removed. 2047 */ 2048 bs = tls->params.tls_bs; 2049 delta = (tls_len + tls->params.tls_tlen) & (bs - 1); 2050 m->m_epg_trllen -= delta; 2051 } 2052 m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; 2053 2054 /* Populate the TLS header. */ 2055 tlshdr = (void *)m->m_epg_hdr; 2056 tlshdr->tls_vmajor = tls->params.tls_vmajor; 2057 2058 /* 2059 * TLS 1.3 masquarades as TLS 1.2 with a record type 2060 * of TLS_RLTYPE_APP. 2061 */ 2062 if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && 2063 tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { 2064 tlshdr->tls_vminor = TLS_MINOR_VER_TWO; 2065 tlshdr->tls_type = TLS_RLTYPE_APP; 2066 /* save the real record type for later */ 2067 m->m_epg_record_type = record_type; 2068 m->m_epg_trail[0] = record_type; 2069 } else { 2070 tlshdr->tls_vminor = tls->params.tls_vminor; 2071 tlshdr->tls_type = record_type; 2072 } 2073 tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); 2074 2075 /* 2076 * Store nonces / explicit IVs after the end of the 2077 * TLS header. 2078 * 2079 * For GCM with TLS 1.2, an 8 byte nonce is copied 2080 * from the end of the IV. The nonce is then 2081 * incremented for use by the next record. 2082 * 2083 * For CBC, a random nonce is inserted for TLS 1.1+. 2084 */ 2085 if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && 2086 tls->params.tls_vminor == TLS_MINOR_VER_TWO) { 2087 noncep = (uint64_t *)(tls->params.iv + 8); 2088 be64enc(tlshdr + 1, *noncep); 2089 (*noncep)++; 2090 } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 2091 tls->params.tls_vminor >= TLS_MINOR_VER_ONE) 2092 arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); 2093 2094 /* 2095 * When using SW encryption, mark the mbuf not ready. 2096 * It will be marked ready via sbready() after the 2097 * record has been encrypted. 2098 * 2099 * When using ifnet TLS, unencrypted TLS records are 2100 * sent down the stack to the NIC. 2101 */ 2102 if (tls->mode == TCP_TLS_MODE_SW) { 2103 m->m_flags |= M_NOTREADY; 2104 if (__predict_false(tls_len == 0)) { 2105 /* TLS 1.0 empty fragment. */ 2106 m->m_epg_nrdy = 1; 2107 } else 2108 m->m_epg_nrdy = m->m_epg_npgs; 2109 *enq_cnt += m->m_epg_nrdy; 2110 } 2111 } 2112 } 2113 2114 bool 2115 ktls_permit_empty_frames(struct ktls_session *tls) 2116 { 2117 return (tls->params.cipher_algorithm == CRYPTO_AES_CBC && 2118 tls->params.tls_vminor == TLS_MINOR_VER_ZERO); 2119 } 2120 2121 void 2122 ktls_check_rx(struct sockbuf *sb) 2123 { 2124 struct tls_record_layer hdr; 2125 struct ktls_wq *wq; 2126 struct socket *so; 2127 bool running; 2128 2129 SOCKBUF_LOCK_ASSERT(sb); 2130 KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", 2131 __func__, sb)); 2132 so = __containerof(sb, struct socket, so_rcv); 2133 2134 if (sb->sb_flags & SB_TLS_RX_RUNNING) 2135 return; 2136 2137 /* Is there enough queued for a TLS header? */ 2138 if (sb->sb_tlscc < sizeof(hdr)) { 2139 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) 2140 so->so_error = EMSGSIZE; 2141 return; 2142 } 2143 2144 m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); 2145 2146 /* Is the entire record queued? */ 2147 if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { 2148 if ((sb->sb_state & SBS_CANTRCVMORE) != 0) 2149 so->so_error = EMSGSIZE; 2150 return; 2151 } 2152 2153 sb->sb_flags |= SB_TLS_RX_RUNNING; 2154 2155 soref(so); 2156 wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; 2157 mtx_lock(&wq->mtx); 2158 STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); 2159 running = wq->running; 2160 mtx_unlock(&wq->mtx); 2161 if (!running) 2162 wakeup(wq); 2163 counter_u64_add(ktls_cnt_rx_queued, 1); 2164 } 2165 2166 static struct mbuf * 2167 ktls_detach_record(struct sockbuf *sb, int len) 2168 { 2169 struct mbuf *m, *n, *top; 2170 int remain; 2171 2172 SOCKBUF_LOCK_ASSERT(sb); 2173 MPASS(len <= sb->sb_tlscc); 2174 2175 /* 2176 * If TLS chain is the exact size of the record, 2177 * just grab the whole record. 2178 */ 2179 top = sb->sb_mtls; 2180 if (sb->sb_tlscc == len) { 2181 sb->sb_mtls = NULL; 2182 sb->sb_mtlstail = NULL; 2183 goto out; 2184 } 2185 2186 /* 2187 * While it would be nice to use m_split() here, we need 2188 * to know exactly what m_split() allocates to update the 2189 * accounting, so do it inline instead. 2190 */ 2191 remain = len; 2192 for (m = top; remain > m->m_len; m = m->m_next) 2193 remain -= m->m_len; 2194 2195 /* Easy case: don't have to split 'm'. */ 2196 if (remain == m->m_len) { 2197 sb->sb_mtls = m->m_next; 2198 if (sb->sb_mtls == NULL) 2199 sb->sb_mtlstail = NULL; 2200 m->m_next = NULL; 2201 goto out; 2202 } 2203 2204 /* 2205 * Need to allocate an mbuf to hold the remainder of 'm'. Try 2206 * with M_NOWAIT first. 2207 */ 2208 n = m_get(M_NOWAIT, MT_DATA); 2209 if (n == NULL) { 2210 /* 2211 * Use M_WAITOK with socket buffer unlocked. If 2212 * 'sb_mtls' changes while the lock is dropped, return 2213 * NULL to force the caller to retry. 2214 */ 2215 SOCKBUF_UNLOCK(sb); 2216 2217 n = m_get(M_WAITOK, MT_DATA); 2218 2219 SOCKBUF_LOCK(sb); 2220 if (sb->sb_mtls != top) { 2221 m_free(n); 2222 return (NULL); 2223 } 2224 } 2225 n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED)); 2226 2227 /* Store remainder in 'n'. */ 2228 n->m_len = m->m_len - remain; 2229 if (m->m_flags & M_EXT) { 2230 n->m_data = m->m_data + remain; 2231 mb_dupcl(n, m); 2232 } else { 2233 bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); 2234 } 2235 2236 /* Trim 'm' and update accounting. */ 2237 m->m_len -= n->m_len; 2238 sb->sb_tlscc -= n->m_len; 2239 sb->sb_ccc -= n->m_len; 2240 2241 /* Account for 'n'. */ 2242 sballoc_ktls_rx(sb, n); 2243 2244 /* Insert 'n' into the TLS chain. */ 2245 sb->sb_mtls = n; 2246 n->m_next = m->m_next; 2247 if (sb->sb_mtlstail == m) 2248 sb->sb_mtlstail = n; 2249 2250 /* Detach the record from the TLS chain. */ 2251 m->m_next = NULL; 2252 2253 out: 2254 MPASS(m_length(top, NULL) == len); 2255 for (m = top; m != NULL; m = m->m_next) 2256 sbfree_ktls_rx(sb, m); 2257 sb->sb_tlsdcc = len; 2258 sb->sb_ccc += len; 2259 SBCHECK(sb); 2260 return (top); 2261 } 2262 2263 /* 2264 * Determine the length of the trailing zero padding and find the real 2265 * record type in the byte before the padding. 2266 * 2267 * Walking the mbuf chain backwards is clumsy, so another option would 2268 * be to scan forwards remembering the last non-zero byte before the 2269 * trailer. However, it would be expensive to scan the entire record. 2270 * Instead, find the last non-zero byte of each mbuf in the chain 2271 * keeping track of the relative offset of that nonzero byte. 2272 * 2273 * trail_len is the size of the MAC/tag on input and is set to the 2274 * size of the full trailer including padding and the record type on 2275 * return. 2276 */ 2277 static int 2278 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, 2279 int *trailer_len, uint8_t *record_typep) 2280 { 2281 char *cp; 2282 u_int digest_start, last_offset, m_len, offset; 2283 uint8_t record_type; 2284 2285 digest_start = tls_len - *trailer_len; 2286 last_offset = 0; 2287 offset = 0; 2288 for (; m != NULL && offset < digest_start; 2289 offset += m->m_len, m = m->m_next) { 2290 /* Don't look for padding in the tag. */ 2291 m_len = min(digest_start - offset, m->m_len); 2292 cp = mtod(m, char *); 2293 2294 /* Find last non-zero byte in this mbuf. */ 2295 while (m_len > 0 && cp[m_len - 1] == 0) 2296 m_len--; 2297 if (m_len > 0) { 2298 record_type = cp[m_len - 1]; 2299 last_offset = offset + m_len; 2300 } 2301 } 2302 if (last_offset < tls->params.tls_hlen) 2303 return (EBADMSG); 2304 2305 *record_typep = record_type; 2306 *trailer_len = tls_len - last_offset + 1; 2307 return (0); 2308 } 2309 2310 /* 2311 * Check if a mbuf chain is fully decrypted at the given offset and 2312 * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is 2313 * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted 2314 * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data 2315 * is encrypted. 2316 */ 2317 ktls_mbuf_crypto_st_t 2318 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len) 2319 { 2320 int m_flags_ored = 0; 2321 int m_flags_anded = -1; 2322 2323 for (; mb != NULL; mb = mb->m_next) { 2324 if (offset < mb->m_len) 2325 break; 2326 offset -= mb->m_len; 2327 } 2328 offset += len; 2329 2330 for (; mb != NULL; mb = mb->m_next) { 2331 m_flags_ored |= mb->m_flags; 2332 m_flags_anded &= mb->m_flags; 2333 2334 if (offset <= mb->m_len) 2335 break; 2336 offset -= mb->m_len; 2337 } 2338 MPASS(mb != NULL || offset == 0); 2339 2340 if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED) 2341 return (KTLS_MBUF_CRYPTO_ST_MIXED); 2342 else 2343 return ((m_flags_ored & M_DECRYPTED) ? 2344 KTLS_MBUF_CRYPTO_ST_DECRYPTED : 2345 KTLS_MBUF_CRYPTO_ST_ENCRYPTED); 2346 } 2347 2348 /* 2349 * ktls_resync_ifnet - get HW TLS RX back on track after packet loss 2350 */ 2351 static int 2352 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num) 2353 { 2354 union if_snd_tag_modify_params params; 2355 struct m_snd_tag *mst; 2356 struct inpcb *inp; 2357 struct tcpcb *tp; 2358 2359 mst = so->so_rcv.sb_tls_info->snd_tag; 2360 if (__predict_false(mst == NULL)) 2361 return (EINVAL); 2362 2363 inp = sotoinpcb(so); 2364 if (__predict_false(inp == NULL)) 2365 return (EINVAL); 2366 2367 INP_RLOCK(inp); 2368 if (inp->inp_flags & INP_DROPPED) { 2369 INP_RUNLOCK(inp); 2370 return (ECONNRESET); 2371 } 2372 2373 tp = intotcpcb(inp); 2374 MPASS(tp != NULL); 2375 2376 /* Get the TCP sequence number of the next valid TLS header. */ 2377 SOCKBUF_LOCK(&so->so_rcv); 2378 params.tls_rx.tls_hdr_tcp_sn = 2379 tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len; 2380 params.tls_rx.tls_rec_length = tls_len; 2381 params.tls_rx.tls_seq_number = tls_rcd_num; 2382 SOCKBUF_UNLOCK(&so->so_rcv); 2383 2384 INP_RUNLOCK(inp); 2385 2386 MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX); 2387 return (mst->sw->snd_tag_modify(mst, ¶ms)); 2388 } 2389 2390 static void 2391 ktls_drop(struct socket *so, int error) 2392 { 2393 struct epoch_tracker et; 2394 struct inpcb *inp = sotoinpcb(so); 2395 struct tcpcb *tp; 2396 2397 NET_EPOCH_ENTER(et); 2398 INP_WLOCK(inp); 2399 if (!(inp->inp_flags & INP_DROPPED)) { 2400 tp = intotcpcb(inp); 2401 CURVNET_SET(inp->inp_vnet); 2402 tp = tcp_drop(tp, error); 2403 CURVNET_RESTORE(); 2404 if (tp != NULL) 2405 INP_WUNLOCK(inp); 2406 } else { 2407 so->so_error = error; 2408 SOCK_RECVBUF_LOCK(so); 2409 sorwakeup_locked(so); 2410 INP_WUNLOCK(inp); 2411 } 2412 NET_EPOCH_EXIT(et); 2413 } 2414 2415 static void 2416 ktls_decrypt(struct socket *so) 2417 { 2418 char tls_header[MBUF_PEXT_HDR_LEN]; 2419 struct ktls_session *tls; 2420 struct sockbuf *sb; 2421 struct tls_record_layer *hdr; 2422 struct tls_get_record tgr; 2423 struct mbuf *control, *data, *m; 2424 ktls_mbuf_crypto_st_t state; 2425 uint64_t seqno; 2426 int error, remain, tls_len, trail_len; 2427 bool tls13; 2428 uint8_t vminor, record_type; 2429 2430 hdr = (struct tls_record_layer *)tls_header; 2431 sb = &so->so_rcv; 2432 SOCKBUF_LOCK(sb); 2433 KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, 2434 ("%s: socket %p not running", __func__, so)); 2435 2436 tls = sb->sb_tls_info; 2437 MPASS(tls != NULL); 2438 2439 tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE); 2440 if (tls13) 2441 vminor = TLS_MINOR_VER_TWO; 2442 else 2443 vminor = tls->params.tls_vminor; 2444 for (;;) { 2445 /* Is there enough queued for a TLS header? */ 2446 if (sb->sb_tlscc < tls->params.tls_hlen) 2447 break; 2448 2449 m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); 2450 tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); 2451 2452 if (hdr->tls_vmajor != tls->params.tls_vmajor || 2453 hdr->tls_vminor != vminor) 2454 error = EINVAL; 2455 else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP) 2456 error = EINVAL; 2457 else if (tls_len < tls->params.tls_hlen || tls_len > 2458 tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + 2459 tls->params.tls_tlen) 2460 error = EMSGSIZE; 2461 else 2462 error = 0; 2463 if (__predict_false(error != 0)) { 2464 /* 2465 * We have a corrupted record and are likely 2466 * out of sync. The connection isn't 2467 * recoverable at this point, so abort it. 2468 */ 2469 SOCKBUF_UNLOCK(sb); 2470 counter_u64_add(ktls_offload_corrupted_records, 1); 2471 2472 ktls_drop(so, error); 2473 goto deref; 2474 } 2475 2476 /* Is the entire record queued? */ 2477 if (sb->sb_tlscc < tls_len) 2478 break; 2479 2480 /* 2481 * Split out the portion of the mbuf chain containing 2482 * this TLS record. 2483 */ 2484 data = ktls_detach_record(sb, tls_len); 2485 if (data == NULL) 2486 continue; 2487 MPASS(sb->sb_tlsdcc == tls_len); 2488 2489 seqno = sb->sb_tls_seqno; 2490 sb->sb_tls_seqno++; 2491 SBCHECK(sb); 2492 SOCKBUF_UNLOCK(sb); 2493 2494 /* get crypto state for this TLS record */ 2495 state = ktls_mbuf_crypto_state(data, 0, tls_len); 2496 2497 switch (state) { 2498 case KTLS_MBUF_CRYPTO_ST_MIXED: 2499 error = ktls_ocf_recrypt(tls, hdr, data, seqno); 2500 if (error) 2501 break; 2502 /* FALLTHROUGH */ 2503 case KTLS_MBUF_CRYPTO_ST_ENCRYPTED: 2504 error = ktls_ocf_decrypt(tls, hdr, data, seqno, 2505 &trail_len); 2506 if (__predict_true(error == 0)) { 2507 if (tls13) { 2508 error = tls13_find_record_type(tls, data, 2509 tls_len, &trail_len, &record_type); 2510 } else { 2511 record_type = hdr->tls_type; 2512 } 2513 } 2514 break; 2515 case KTLS_MBUF_CRYPTO_ST_DECRYPTED: 2516 /* 2517 * NIC TLS is only supported for AEAD 2518 * ciphersuites which used a fixed sized 2519 * trailer. 2520 */ 2521 if (tls13) { 2522 trail_len = tls->params.tls_tlen - 1; 2523 error = tls13_find_record_type(tls, data, 2524 tls_len, &trail_len, &record_type); 2525 } else { 2526 trail_len = tls->params.tls_tlen; 2527 error = 0; 2528 record_type = hdr->tls_type; 2529 } 2530 break; 2531 default: 2532 error = EINVAL; 2533 break; 2534 } 2535 if (error) { 2536 counter_u64_add(ktls_offload_failed_crypto, 1); 2537 2538 SOCKBUF_LOCK(sb); 2539 if (sb->sb_tlsdcc == 0) { 2540 /* 2541 * sbcut/drop/flush discarded these 2542 * mbufs. 2543 */ 2544 m_freem(data); 2545 break; 2546 } 2547 2548 /* 2549 * Drop this TLS record's data, but keep 2550 * decrypting subsequent records. 2551 */ 2552 sb->sb_ccc -= tls_len; 2553 sb->sb_tlsdcc = 0; 2554 2555 if (error != EMSGSIZE) 2556 error = EBADMSG; 2557 CURVNET_SET(so->so_vnet); 2558 so->so_error = error; 2559 sorwakeup_locked(so); 2560 CURVNET_RESTORE(); 2561 2562 m_freem(data); 2563 2564 SOCKBUF_LOCK(sb); 2565 continue; 2566 } 2567 2568 /* Allocate the control mbuf. */ 2569 memset(&tgr, 0, sizeof(tgr)); 2570 tgr.tls_type = record_type; 2571 tgr.tls_vmajor = hdr->tls_vmajor; 2572 tgr.tls_vminor = hdr->tls_vminor; 2573 tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - 2574 trail_len); 2575 control = sbcreatecontrol(&tgr, sizeof(tgr), 2576 TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); 2577 2578 SOCKBUF_LOCK(sb); 2579 if (sb->sb_tlsdcc == 0) { 2580 /* sbcut/drop/flush discarded these mbufs. */ 2581 MPASS(sb->sb_tlscc == 0); 2582 m_freem(data); 2583 m_freem(control); 2584 break; 2585 } 2586 2587 /* 2588 * Clear the 'dcc' accounting in preparation for 2589 * adding the decrypted record. 2590 */ 2591 sb->sb_ccc -= tls_len; 2592 sb->sb_tlsdcc = 0; 2593 SBCHECK(sb); 2594 2595 /* If there is no payload, drop all of the data. */ 2596 if (tgr.tls_length == htobe16(0)) { 2597 m_freem(data); 2598 data = NULL; 2599 } else { 2600 /* Trim header. */ 2601 remain = tls->params.tls_hlen; 2602 while (remain > 0) { 2603 if (data->m_len > remain) { 2604 data->m_data += remain; 2605 data->m_len -= remain; 2606 break; 2607 } 2608 remain -= data->m_len; 2609 data = m_free(data); 2610 } 2611 2612 /* Trim trailer and clear M_NOTREADY. */ 2613 remain = be16toh(tgr.tls_length); 2614 m = data; 2615 for (m = data; remain > m->m_len; m = m->m_next) { 2616 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); 2617 remain -= m->m_len; 2618 } 2619 m->m_len = remain; 2620 m_freem(m->m_next); 2621 m->m_next = NULL; 2622 m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); 2623 2624 /* Set EOR on the final mbuf. */ 2625 m->m_flags |= M_EOR; 2626 } 2627 2628 sbappendcontrol_locked(sb, data, control, 0); 2629 2630 if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) { 2631 sb->sb_flags |= SB_TLS_RX_RESYNC; 2632 SOCKBUF_UNLOCK(sb); 2633 ktls_resync_ifnet(so, tls_len, seqno); 2634 SOCKBUF_LOCK(sb); 2635 } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) { 2636 sb->sb_flags &= ~SB_TLS_RX_RESYNC; 2637 SOCKBUF_UNLOCK(sb); 2638 ktls_resync_ifnet(so, 0, seqno); 2639 SOCKBUF_LOCK(sb); 2640 } 2641 } 2642 2643 sb->sb_flags &= ~SB_TLS_RX_RUNNING; 2644 2645 if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) 2646 so->so_error = EMSGSIZE; 2647 2648 sorwakeup_locked(so); 2649 2650 deref: 2651 SOCKBUF_UNLOCK_ASSERT(sb); 2652 2653 CURVNET_SET(so->so_vnet); 2654 sorele(so); 2655 CURVNET_RESTORE(); 2656 } 2657 2658 void 2659 ktls_enqueue_to_free(struct mbuf *m) 2660 { 2661 struct ktls_wq *wq; 2662 bool running; 2663 2664 /* Mark it for freeing. */ 2665 m->m_epg_flags |= EPG_FLAG_2FREE; 2666 wq = &ktls_wq[m->m_epg_tls->wq_index]; 2667 mtx_lock(&wq->mtx); 2668 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2669 running = wq->running; 2670 mtx_unlock(&wq->mtx); 2671 if (!running) 2672 wakeup(wq); 2673 } 2674 2675 static void * 2676 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m) 2677 { 2678 void *buf; 2679 int domain, running; 2680 2681 if (m->m_epg_npgs <= 2) 2682 return (NULL); 2683 if (ktls_buffer_zone == NULL) 2684 return (NULL); 2685 if ((u_int)(ticks - wq->lastallocfail) < hz) { 2686 /* 2687 * Rate-limit allocation attempts after a failure. 2688 * ktls_buffer_import() will acquire a per-domain mutex to check 2689 * the free page queues and may fail consistently if memory is 2690 * fragmented. 2691 */ 2692 return (NULL); 2693 } 2694 buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM); 2695 if (buf == NULL) { 2696 domain = PCPU_GET(domain); 2697 wq->lastallocfail = ticks; 2698 2699 /* 2700 * Note that this check is "racy", but the races are 2701 * harmless, and are either a spurious wakeup if 2702 * multiple threads fail allocations before the alloc 2703 * thread wakes, or waiting an extra second in case we 2704 * see an old value of running == true. 2705 */ 2706 if (!VM_DOMAIN_EMPTY(domain)) { 2707 running = atomic_load_int(&ktls_domains[domain].reclaim_td.running); 2708 if (!running) 2709 wakeup(&ktls_domains[domain].reclaim_td); 2710 } 2711 } 2712 return (buf); 2713 } 2714 2715 static int 2716 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m, 2717 struct ktls_session *tls, struct ktls_ocf_encrypt_state *state) 2718 { 2719 vm_page_t pg; 2720 int error, i, len, off; 2721 2722 KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY), 2723 ("%p not unready & nomap mbuf\n", m)); 2724 KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen, 2725 ("page count %d larger than maximum frame length %d", m->m_epg_npgs, 2726 ktls_maxlen)); 2727 2728 /* Anonymous mbufs are encrypted in place. */ 2729 if ((m->m_epg_flags & EPG_FLAG_ANON) != 0) 2730 return (ktls_ocf_encrypt(state, tls, m, NULL, 0)); 2731 2732 /* 2733 * For file-backed mbufs (from sendfile), anonymous wired 2734 * pages are allocated and used as the encryption destination. 2735 */ 2736 if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) { 2737 len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len - 2738 m->m_epg_1st_off; 2739 state->dst_iov[0].iov_base = (char *)state->cbuf + 2740 m->m_epg_1st_off; 2741 state->dst_iov[0].iov_len = len; 2742 state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf); 2743 i = 1; 2744 } else { 2745 off = m->m_epg_1st_off; 2746 for (i = 0; i < m->m_epg_npgs; i++, off = 0) { 2747 pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP | 2748 VM_ALLOC_WIRED | VM_ALLOC_WAITOK); 2749 len = m_epg_pagelen(m, i, off); 2750 state->parray[i] = VM_PAGE_TO_PHYS(pg); 2751 state->dst_iov[i].iov_base = 2752 (char *)PHYS_TO_DMAP(state->parray[i]) + off; 2753 state->dst_iov[i].iov_len = len; 2754 } 2755 } 2756 KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small")); 2757 state->dst_iov[i].iov_base = m->m_epg_trail; 2758 state->dst_iov[i].iov_len = m->m_epg_trllen; 2759 2760 error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1); 2761 2762 if (__predict_false(error != 0)) { 2763 /* Free the anonymous pages. */ 2764 if (state->cbuf != NULL) 2765 uma_zfree(ktls_buffer_zone, state->cbuf); 2766 else { 2767 for (i = 0; i < m->m_epg_npgs; i++) { 2768 pg = PHYS_TO_VM_PAGE(state->parray[i]); 2769 (void)vm_page_unwire_noq(pg); 2770 vm_page_free(pg); 2771 } 2772 } 2773 } 2774 return (error); 2775 } 2776 2777 /* Number of TLS records in a batch passed to ktls_enqueue(). */ 2778 static u_int 2779 ktls_batched_records(struct mbuf *m) 2780 { 2781 int page_count, records; 2782 2783 records = 0; 2784 page_count = m->m_epg_enc_cnt; 2785 while (page_count > 0) { 2786 records++; 2787 page_count -= m->m_epg_nrdy; 2788 m = m->m_next; 2789 } 2790 KASSERT(page_count == 0, ("%s: mismatched page count", __func__)); 2791 return (records); 2792 } 2793 2794 void 2795 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) 2796 { 2797 struct ktls_session *tls; 2798 struct ktls_wq *wq; 2799 int queued; 2800 bool running; 2801 2802 KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == 2803 (M_EXTPG | M_NOTREADY)), 2804 ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); 2805 KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); 2806 2807 KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); 2808 2809 m->m_epg_enc_cnt = page_count; 2810 2811 /* 2812 * Save a pointer to the socket. The caller is responsible 2813 * for taking an additional reference via soref(). 2814 */ 2815 m->m_epg_so = so; 2816 2817 queued = 1; 2818 tls = m->m_epg_tls; 2819 wq = &ktls_wq[tls->wq_index]; 2820 mtx_lock(&wq->mtx); 2821 if (__predict_false(tls->sequential_records)) { 2822 /* 2823 * For TLS 1.0, records must be encrypted 2824 * sequentially. For a given connection, all records 2825 * queued to the associated work queue are processed 2826 * sequentially. However, sendfile(2) might complete 2827 * I/O requests spanning multiple TLS records out of 2828 * order. Here we ensure TLS records are enqueued to 2829 * the work queue in FIFO order. 2830 * 2831 * tls->next_seqno holds the sequence number of the 2832 * next TLS record that should be enqueued to the work 2833 * queue. If this next record is not tls->next_seqno, 2834 * it must be a future record, so insert it, sorted by 2835 * TLS sequence number, into tls->pending_records and 2836 * return. 2837 * 2838 * If this TLS record matches tls->next_seqno, place 2839 * it in the work queue and then check 2840 * tls->pending_records to see if any 2841 * previously-queued records are now ready for 2842 * encryption. 2843 */ 2844 if (m->m_epg_seqno != tls->next_seqno) { 2845 struct mbuf *n, *p; 2846 2847 p = NULL; 2848 STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) { 2849 if (n->m_epg_seqno > m->m_epg_seqno) 2850 break; 2851 p = n; 2852 } 2853 if (n == NULL) 2854 STAILQ_INSERT_TAIL(&tls->pending_records, m, 2855 m_epg_stailq); 2856 else if (p == NULL) 2857 STAILQ_INSERT_HEAD(&tls->pending_records, m, 2858 m_epg_stailq); 2859 else 2860 STAILQ_INSERT_AFTER(&tls->pending_records, p, m, 2861 m_epg_stailq); 2862 mtx_unlock(&wq->mtx); 2863 counter_u64_add(ktls_cnt_tx_pending, 1); 2864 return; 2865 } 2866 2867 tls->next_seqno += ktls_batched_records(m); 2868 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2869 2870 while (!STAILQ_EMPTY(&tls->pending_records)) { 2871 struct mbuf *n; 2872 2873 n = STAILQ_FIRST(&tls->pending_records); 2874 if (n->m_epg_seqno != tls->next_seqno) 2875 break; 2876 2877 queued++; 2878 STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq); 2879 tls->next_seqno += ktls_batched_records(n); 2880 STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq); 2881 } 2882 counter_u64_add(ktls_cnt_tx_pending, -(queued - 1)); 2883 } else 2884 STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); 2885 2886 running = wq->running; 2887 mtx_unlock(&wq->mtx); 2888 if (!running) 2889 wakeup(wq); 2890 counter_u64_add(ktls_cnt_tx_queued, queued); 2891 } 2892 2893 /* 2894 * Once a file-backed mbuf (from sendfile) has been encrypted, free 2895 * the pages from the file and replace them with the anonymous pages 2896 * allocated in ktls_encrypt_record(). 2897 */ 2898 static void 2899 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state) 2900 { 2901 int i; 2902 2903 MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0); 2904 2905 /* Free the old pages. */ 2906 m->m_ext.ext_free(m); 2907 2908 /* Replace them with the new pages. */ 2909 if (state->cbuf != NULL) { 2910 for (i = 0; i < m->m_epg_npgs; i++) 2911 m->m_epg_pa[i] = state->parray[0] + ptoa(i); 2912 2913 /* Contig pages should go back to the cache. */ 2914 m->m_ext.ext_free = ktls_free_mext_contig; 2915 } else { 2916 for (i = 0; i < m->m_epg_npgs; i++) 2917 m->m_epg_pa[i] = state->parray[i]; 2918 2919 /* Use the basic free routine. */ 2920 m->m_ext.ext_free = mb_free_mext_pgs; 2921 } 2922 2923 /* Pages are now writable. */ 2924 m->m_epg_flags |= EPG_FLAG_ANON; 2925 } 2926 2927 static __noinline void 2928 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top) 2929 { 2930 struct ktls_ocf_encrypt_state state; 2931 struct ktls_session *tls; 2932 struct socket *so; 2933 struct mbuf *m; 2934 int error, npages, total_pages; 2935 2936 so = top->m_epg_so; 2937 tls = top->m_epg_tls; 2938 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 2939 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 2940 #ifdef INVARIANTS 2941 top->m_epg_so = NULL; 2942 #endif 2943 total_pages = top->m_epg_enc_cnt; 2944 npages = 0; 2945 2946 /* 2947 * Encrypt the TLS records in the chain of mbufs starting with 2948 * 'top'. 'total_pages' gives us a total count of pages and is 2949 * used to know when we have finished encrypting the TLS 2950 * records originally queued with 'top'. 2951 * 2952 * NB: These mbufs are queued in the socket buffer and 2953 * 'm_next' is traversing the mbufs in the socket buffer. The 2954 * socket buffer lock is not held while traversing this chain. 2955 * Since the mbufs are all marked M_NOTREADY their 'm_next' 2956 * pointers should be stable. However, the 'm_next' of the 2957 * last mbuf encrypted is not necessarily NULL. It can point 2958 * to other mbufs appended while 'top' was on the TLS work 2959 * queue. 2960 * 2961 * Each mbuf holds an entire TLS record. 2962 */ 2963 error = 0; 2964 for (m = top; npages != total_pages; m = m->m_next) { 2965 KASSERT(m->m_epg_tls == tls, 2966 ("different TLS sessions in a single mbuf chain: %p vs %p", 2967 tls, m->m_epg_tls)); 2968 KASSERT(npages + m->m_epg_npgs <= total_pages, 2969 ("page count mismatch: top %p, total_pages %d, m %p", top, 2970 total_pages, m)); 2971 2972 error = ktls_encrypt_record(wq, m, tls, &state); 2973 if (error) { 2974 counter_u64_add(ktls_offload_failed_crypto, 1); 2975 break; 2976 } 2977 2978 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) 2979 ktls_finish_nonanon(m, &state); 2980 2981 npages += m->m_epg_nrdy; 2982 2983 /* 2984 * Drop a reference to the session now that it is no 2985 * longer needed. Existing code depends on encrypted 2986 * records having no associated session vs 2987 * yet-to-be-encrypted records having an associated 2988 * session. 2989 */ 2990 m->m_epg_tls = NULL; 2991 ktls_free(tls); 2992 } 2993 2994 CURVNET_SET(so->so_vnet); 2995 if (error == 0) { 2996 (void)so->so_proto->pr_ready(so, top, npages); 2997 } else { 2998 ktls_drop(so, EIO); 2999 mb_free_notready(top, total_pages); 3000 } 3001 3002 sorele(so); 3003 CURVNET_RESTORE(); 3004 } 3005 3006 void 3007 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error) 3008 { 3009 struct ktls_session *tls; 3010 struct socket *so; 3011 struct mbuf *m; 3012 int npages; 3013 3014 m = state->m; 3015 3016 if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) 3017 ktls_finish_nonanon(m, state); 3018 3019 so = state->so; 3020 free(state, M_KTLS); 3021 3022 /* 3023 * Drop a reference to the session now that it is no longer 3024 * needed. Existing code depends on encrypted records having 3025 * no associated session vs yet-to-be-encrypted records having 3026 * an associated session. 3027 */ 3028 tls = m->m_epg_tls; 3029 m->m_epg_tls = NULL; 3030 ktls_free(tls); 3031 3032 if (error != 0) 3033 counter_u64_add(ktls_offload_failed_crypto, 1); 3034 3035 CURVNET_SET(so->so_vnet); 3036 npages = m->m_epg_nrdy; 3037 3038 if (error == 0) { 3039 (void)so->so_proto->pr_ready(so, m, npages); 3040 } else { 3041 ktls_drop(so, EIO); 3042 mb_free_notready(m, npages); 3043 } 3044 3045 sorele(so); 3046 CURVNET_RESTORE(); 3047 } 3048 3049 /* 3050 * Similar to ktls_encrypt, but used with asynchronous OCF backends 3051 * (coprocessors) where encryption does not use host CPU resources and 3052 * it can be beneficial to queue more requests than CPUs. 3053 */ 3054 static __noinline void 3055 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top) 3056 { 3057 struct ktls_ocf_encrypt_state *state; 3058 struct ktls_session *tls; 3059 struct socket *so; 3060 struct mbuf *m, *n; 3061 int error, mpages, npages, total_pages; 3062 3063 so = top->m_epg_so; 3064 tls = top->m_epg_tls; 3065 KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); 3066 KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); 3067 #ifdef INVARIANTS 3068 top->m_epg_so = NULL; 3069 #endif 3070 total_pages = top->m_epg_enc_cnt; 3071 npages = 0; 3072 3073 error = 0; 3074 for (m = top; npages != total_pages; m = n) { 3075 KASSERT(m->m_epg_tls == tls, 3076 ("different TLS sessions in a single mbuf chain: %p vs %p", 3077 tls, m->m_epg_tls)); 3078 KASSERT(npages + m->m_epg_npgs <= total_pages, 3079 ("page count mismatch: top %p, total_pages %d, m %p", top, 3080 total_pages, m)); 3081 3082 state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO); 3083 soref(so); 3084 state->so = so; 3085 state->m = m; 3086 3087 mpages = m->m_epg_nrdy; 3088 n = m->m_next; 3089 3090 error = ktls_encrypt_record(wq, m, tls, state); 3091 if (error) { 3092 counter_u64_add(ktls_offload_failed_crypto, 1); 3093 free(state, M_KTLS); 3094 CURVNET_SET(so->so_vnet); 3095 sorele(so); 3096 CURVNET_RESTORE(); 3097 break; 3098 } 3099 3100 npages += mpages; 3101 } 3102 3103 CURVNET_SET(so->so_vnet); 3104 if (error != 0) { 3105 ktls_drop(so, EIO); 3106 mb_free_notready(m, total_pages - npages); 3107 } 3108 3109 sorele(so); 3110 CURVNET_RESTORE(); 3111 } 3112 3113 static int 3114 ktls_bind_domain(int domain) 3115 { 3116 int error; 3117 3118 error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]); 3119 if (error != 0) 3120 return (error); 3121 curthread->td_domain.dr_policy = DOMAINSET_PREF(domain); 3122 return (0); 3123 } 3124 3125 static void 3126 ktls_reclaim_thread(void *ctx) 3127 { 3128 struct ktls_domain_info *ktls_domain = ctx; 3129 struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td; 3130 struct sysctl_oid *oid; 3131 char name[80]; 3132 int error, domain; 3133 3134 domain = ktls_domain - ktls_domains; 3135 if (bootverbose) 3136 printf("Starting KTLS reclaim thread for domain %d\n", domain); 3137 error = ktls_bind_domain(domain); 3138 if (error) 3139 printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n", 3140 domain, error); 3141 snprintf(name, sizeof(name), "domain%d", domain); 3142 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO, 3143 name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 3144 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims", 3145 CTLFLAG_RD, &sc->reclaims, 0, "buffers reclaimed"); 3146 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups", 3147 CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups"); 3148 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running", 3149 CTLFLAG_RD, &sc->running, 0, "thread running"); 3150 3151 for (;;) { 3152 atomic_store_int(&sc->running, 0); 3153 tsleep(sc, PZERO | PNOLOCK, "-", 0); 3154 atomic_store_int(&sc->running, 1); 3155 sc->wakeups++; 3156 /* 3157 * Below we attempt to reclaim ktls_max_reclaim 3158 * buffers using vm_page_reclaim_contig_domain_ext(). 3159 * We do this here, as this function can take several 3160 * seconds to scan all of memory and it does not 3161 * matter if this thread pauses for a while. If we 3162 * block a ktls worker thread, we risk developing 3163 * backlogs of buffers to be encrypted, leading to 3164 * surges of traffic and potential NIC output drops. 3165 */ 3166 if (!vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL, 3167 atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, ktls_max_reclaim)) { 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