xref: /freebsd-14.2/sys/kern/uipc_ktls.c (revision b14a4912)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2014-2019 Netflix Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_kern_tls.h"
32 #include "opt_ratelimit.h"
33 #include "opt_rss.h"
34 
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/domainset.h>
38 #include <sys/endian.h>
39 #include <sys/ktls.h>
40 #include <sys/lock.h>
41 #include <sys/mbuf.h>
42 #include <sys/mutex.h>
43 #include <sys/rmlock.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/refcount.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/taskqueue.h>
52 #include <sys/kthread.h>
53 #include <sys/uio.h>
54 #include <sys/vmmeter.h>
55 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
56 #include <machine/pcb.h>
57 #endif
58 #include <machine/vmparam.h>
59 #include <net/if.h>
60 #include <net/if_var.h>
61 #ifdef RSS
62 #include <net/netisr.h>
63 #include <net/rss_config.h>
64 #endif
65 #include <net/route.h>
66 #include <net/route/nhop.h>
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_var.h>
70 #ifdef TCP_OFFLOAD
71 #include <netinet/tcp_offload.h>
72 #endif
73 #include <opencrypto/cryptodev.h>
74 #include <opencrypto/ktls.h>
75 #include <vm/vm.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_pagequeue.h>
79 
80 struct ktls_wq {
81 	struct mtx	mtx;
82 	STAILQ_HEAD(, mbuf) m_head;
83 	STAILQ_HEAD(, socket) so_head;
84 	bool		running;
85 	int		lastallocfail;
86 } __aligned(CACHE_LINE_SIZE);
87 
88 struct ktls_reclaim_thread {
89 	uint64_t wakeups;
90 	uint64_t reclaims;
91 	struct thread *td;
92 	int running;
93 };
94 
95 struct ktls_domain_info {
96 	int count;
97 	int cpu[MAXCPU];
98 	struct ktls_reclaim_thread reclaim_td;
99 };
100 
101 struct ktls_domain_info ktls_domains[MAXMEMDOM];
102 static struct ktls_wq *ktls_wq;
103 static struct proc *ktls_proc;
104 static uma_zone_t ktls_session_zone;
105 static uma_zone_t ktls_buffer_zone;
106 static uint16_t ktls_cpuid_lookup[MAXCPU];
107 static int ktls_init_state;
108 static struct sx ktls_init_lock;
109 SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init");
110 
111 SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
112     "Kernel TLS offload");
113 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114     "Kernel TLS offload stats");
115 
116 #ifdef RSS
117 static int ktls_bind_threads = 1;
118 #else
119 static int ktls_bind_threads;
120 #endif
121 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN,
122     &ktls_bind_threads, 0,
123     "Bind crypto threads to cores (1) or cores and domains (2) at boot");
124 
125 static u_int ktls_maxlen = 16384;
126 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN,
127     &ktls_maxlen, 0, "Maximum TLS record size");
128 
129 static int ktls_number_threads;
130 SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD,
131     &ktls_number_threads, 0,
132     "Number of TLS threads in thread-pool");
133 
134 unsigned int ktls_ifnet_max_rexmit_pct = 2;
135 SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN,
136     &ktls_ifnet_max_rexmit_pct, 2,
137     "Max percent bytes retransmitted before ifnet TLS is disabled");
138 
139 static bool ktls_offload_enable;
140 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN,
141     &ktls_offload_enable, 0,
142     "Enable support for kernel TLS offload");
143 
144 static bool ktls_cbc_enable = true;
145 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN,
146     &ktls_cbc_enable, 1,
147     "Enable support of AES-CBC crypto for kernel TLS");
148 
149 static bool ktls_sw_buffer_cache = true;
150 SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN,
151     &ktls_sw_buffer_cache, 1,
152     "Enable caching of output buffers for SW encryption");
153 
154 static int ktls_max_reclaim = 1024;
155 SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_reclaim, CTLFLAG_RWTUN,
156     &ktls_max_reclaim, 128,
157     "Max number of 16k buffers to reclaim in thread context");
158 
159 static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active);
160 SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD,
161     &ktls_tasks_active, "Number of active tasks");
162 
163 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending);
164 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD,
165     &ktls_cnt_tx_pending,
166     "Number of TLS 1.0 records waiting for earlier TLS records");
167 
168 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued);
169 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD,
170     &ktls_cnt_tx_queued,
171     "Number of TLS records in queue to tasks for SW encryption");
172 
173 static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued);
174 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD,
175     &ktls_cnt_rx_queued,
176     "Number of TLS sockets in queue to tasks for SW decryption");
177 
178 static COUNTER_U64_DEFINE_EARLY(ktls_offload_total);
179 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total,
180     CTLFLAG_RD, &ktls_offload_total,
181     "Total successful TLS setups (parameters set)");
182 
183 static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls);
184 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls,
185     CTLFLAG_RD, &ktls_offload_enable_calls,
186     "Total number of TLS enable calls made");
187 
188 static COUNTER_U64_DEFINE_EARLY(ktls_offload_active);
189 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD,
190     &ktls_offload_active, "Total Active TLS sessions");
191 
192 static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records);
193 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD,
194     &ktls_offload_corrupted_records, "Total corrupted TLS records received");
195 
196 static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto);
197 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD,
198     &ktls_offload_failed_crypto, "Total TLS crypto failures");
199 
200 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet);
201 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD,
202     &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet");
203 
204 static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw);
205 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD,
206     &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW");
207 
208 static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed);
209 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD,
210     &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet");
211 
212 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail);
213 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD,
214     &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet");
215 
216 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok);
217 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD,
218     &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet");
219 
220 static COUNTER_U64_DEFINE_EARLY(ktls_destroy_task);
221 SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, destroy_task, CTLFLAG_RD,
222     &ktls_destroy_task,
223     "Number of times ktls session was destroyed via taskqueue");
224 
225 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226     "Software TLS session stats");
227 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
228     "Hardware (ifnet) TLS session stats");
229 #ifdef TCP_OFFLOAD
230 SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
231     "TOE TLS session stats");
232 #endif
233 
234 static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc);
235 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc,
236     "Active number of software TLS sessions using AES-CBC");
237 
238 static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm);
239 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm,
240     "Active number of software TLS sessions using AES-GCM");
241 
242 static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20);
243 SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD,
244     &ktls_sw_chacha20,
245     "Active number of software TLS sessions using Chacha20-Poly1305");
246 
247 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc);
248 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD,
249     &ktls_ifnet_cbc,
250     "Active number of ifnet TLS sessions using AES-CBC");
251 
252 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm);
253 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD,
254     &ktls_ifnet_gcm,
255     "Active number of ifnet TLS sessions using AES-GCM");
256 
257 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20);
258 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD,
259     &ktls_ifnet_chacha20,
260     "Active number of ifnet TLS sessions using Chacha20-Poly1305");
261 
262 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset);
263 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD,
264     &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag");
265 
266 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped);
267 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD,
268     &ktls_ifnet_reset_dropped,
269     "TLS sessions dropped after failing to update ifnet send tag");
270 
271 static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed);
272 SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD,
273     &ktls_ifnet_reset_failed,
274     "TLS sessions that failed to allocate a new ifnet send tag");
275 
276 static int ktls_ifnet_permitted;
277 SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN,
278     &ktls_ifnet_permitted, 1,
279     "Whether to permit hardware (ifnet) TLS sessions");
280 
281 #ifdef TCP_OFFLOAD
282 static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc);
283 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD,
284     &ktls_toe_cbc,
285     "Active number of TOE TLS sessions using AES-CBC");
286 
287 static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm);
288 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD,
289     &ktls_toe_gcm,
290     "Active number of TOE TLS sessions using AES-GCM");
291 
292 static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20);
293 SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD,
294     &ktls_toe_chacha20,
295     "Active number of TOE TLS sessions using Chacha20-Poly1305");
296 #endif
297 
298 static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS");
299 
300 static void ktls_reset_receive_tag(void *context, int pending);
301 static void ktls_reset_send_tag(void *context, int pending);
302 static void ktls_work_thread(void *ctx);
303 static void ktls_reclaim_thread(void *ctx);
304 
305 static u_int
ktls_get_cpu(struct socket * so)306 ktls_get_cpu(struct socket *so)
307 {
308 	struct inpcb *inp;
309 #ifdef NUMA
310 	struct ktls_domain_info *di;
311 #endif
312 	u_int cpuid;
313 
314 	inp = sotoinpcb(so);
315 #ifdef RSS
316 	cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
317 	if (cpuid != NETISR_CPUID_NONE)
318 		return (cpuid);
319 #endif
320 	/*
321 	 * Just use the flowid to shard connections in a repeatable
322 	 * fashion.  Note that TLS 1.0 sessions rely on the
323 	 * serialization provided by having the same connection use
324 	 * the same queue.
325 	 */
326 #ifdef NUMA
327 	if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) {
328 		di = &ktls_domains[inp->inp_numa_domain];
329 		cpuid = di->cpu[inp->inp_flowid % di->count];
330 	} else
331 #endif
332 		cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads];
333 	return (cpuid);
334 }
335 
336 static int
ktls_buffer_import(void * arg,void ** store,int count,int domain,int flags)337 ktls_buffer_import(void *arg, void **store, int count, int domain, int flags)
338 {
339 	vm_page_t m;
340 	int i, req;
341 
342 	KASSERT((ktls_maxlen & PAGE_MASK) == 0,
343 	    ("%s: ktls max length %d is not page size-aligned",
344 	    __func__, ktls_maxlen));
345 
346 	req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags);
347 	for (i = 0; i < count; i++) {
348 		m = vm_page_alloc_noobj_contig_domain(domain, req,
349 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0,
350 		    VM_MEMATTR_DEFAULT);
351 		if (m == NULL)
352 			break;
353 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
354 	}
355 	return (i);
356 }
357 
358 static void
ktls_buffer_release(void * arg __unused,void ** store,int count)359 ktls_buffer_release(void *arg __unused, void **store, int count)
360 {
361 	vm_page_t m;
362 	int i, j;
363 
364 	for (i = 0; i < count; i++) {
365 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
366 		for (j = 0; j < atop(ktls_maxlen); j++) {
367 			(void)vm_page_unwire_noq(m + j);
368 			vm_page_free(m + j);
369 		}
370 	}
371 }
372 
373 static void
ktls_free_mext_contig(struct mbuf * m)374 ktls_free_mext_contig(struct mbuf *m)
375 {
376 	M_ASSERTEXTPG(m);
377 	uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0]));
378 }
379 
380 static int
ktls_init(void)381 ktls_init(void)
382 {
383 	struct thread *td;
384 	struct pcpu *pc;
385 	int count, domain, error, i;
386 
387 	ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS,
388 	    M_WAITOK | M_ZERO);
389 
390 	ktls_session_zone = uma_zcreate("ktls_session",
391 	    sizeof(struct ktls_session),
392 	    NULL, NULL, NULL, NULL,
393 	    UMA_ALIGN_CACHE, 0);
394 
395 	if (ktls_sw_buffer_cache) {
396 		ktls_buffer_zone = uma_zcache_create("ktls_buffers",
397 		    roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL,
398 		    ktls_buffer_import, ktls_buffer_release, NULL,
399 		    UMA_ZONE_FIRSTTOUCH);
400 	}
401 
402 	/*
403 	 * Initialize the workqueues to run the TLS work.  We create a
404 	 * work queue for each CPU.
405 	 */
406 	CPU_FOREACH(i) {
407 		STAILQ_INIT(&ktls_wq[i].m_head);
408 		STAILQ_INIT(&ktls_wq[i].so_head);
409 		mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF);
410 		if (ktls_bind_threads > 1) {
411 			pc = pcpu_find(i);
412 			domain = pc->pc_domain;
413 			count = ktls_domains[domain].count;
414 			ktls_domains[domain].cpu[count] = i;
415 			ktls_domains[domain].count++;
416 		}
417 		ktls_cpuid_lookup[ktls_number_threads] = i;
418 		ktls_number_threads++;
419 	}
420 
421 	/*
422 	 * If we somehow have an empty domain, fall back to choosing
423 	 * among all KTLS threads.
424 	 */
425 	if (ktls_bind_threads > 1) {
426 		for (i = 0; i < vm_ndomains; i++) {
427 			if (ktls_domains[i].count == 0) {
428 				ktls_bind_threads = 1;
429 				break;
430 			}
431 		}
432 	}
433 
434 	/* Start kthreads for each workqueue. */
435 	CPU_FOREACH(i) {
436 		error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i],
437 		    &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i);
438 		if (error) {
439 			printf("Can't add KTLS thread %d error %d\n", i, error);
440 			return (error);
441 		}
442 	}
443 
444 	/*
445 	 * Start an allocation thread per-domain to perform blocking allocations
446 	 * of 16k physically contiguous TLS crypto destination buffers.
447 	 */
448 	if (ktls_sw_buffer_cache) {
449 		for (domain = 0; domain < vm_ndomains; domain++) {
450 			if (VM_DOMAIN_EMPTY(domain))
451 				continue;
452 			if (CPU_EMPTY(&cpuset_domain[domain]))
453 				continue;
454 			error = kproc_kthread_add(ktls_reclaim_thread,
455 			    &ktls_domains[domain], &ktls_proc,
456 			    &ktls_domains[domain].reclaim_td.td,
457 			    0, 0, "KTLS", "reclaim_%d", domain);
458 			if (error) {
459 				printf("Can't add KTLS reclaim thread %d error %d\n",
460 				    domain, error);
461 				return (error);
462 			}
463 		}
464 	}
465 
466 	if (bootverbose)
467 		printf("KTLS: Initialized %d threads\n", ktls_number_threads);
468 	return (0);
469 }
470 
471 static int
ktls_start_kthreads(void)472 ktls_start_kthreads(void)
473 {
474 	int error, state;
475 
476 start:
477 	state = atomic_load_acq_int(&ktls_init_state);
478 	if (__predict_true(state > 0))
479 		return (0);
480 	if (state < 0)
481 		return (ENXIO);
482 
483 	sx_xlock(&ktls_init_lock);
484 	if (ktls_init_state != 0) {
485 		sx_xunlock(&ktls_init_lock);
486 		goto start;
487 	}
488 
489 	error = ktls_init();
490 	if (error == 0)
491 		state = 1;
492 	else
493 		state = -1;
494 	atomic_store_rel_int(&ktls_init_state, state);
495 	sx_xunlock(&ktls_init_lock);
496 	return (error);
497 }
498 
499 static int
ktls_create_session(struct socket * so,struct tls_enable * en,struct ktls_session ** tlsp,int direction)500 ktls_create_session(struct socket *so, struct tls_enable *en,
501     struct ktls_session **tlsp, int direction)
502 {
503 	struct ktls_session *tls;
504 	int error;
505 
506 	/* Only TLS 1.0 - 1.3 are supported. */
507 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE)
508 		return (EINVAL);
509 	if (en->tls_vminor < TLS_MINOR_VER_ZERO ||
510 	    en->tls_vminor > TLS_MINOR_VER_THREE)
511 		return (EINVAL);
512 
513 	if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE)
514 		return (EINVAL);
515 	if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE)
516 		return (EINVAL);
517 	if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv))
518 		return (EINVAL);
519 
520 	/* All supported algorithms require a cipher key. */
521 	if (en->cipher_key_len == 0)
522 		return (EINVAL);
523 
524 	/* No flags are currently supported. */
525 	if (en->flags != 0)
526 		return (EINVAL);
527 
528 	/* Common checks for supported algorithms. */
529 	switch (en->cipher_algorithm) {
530 	case CRYPTO_AES_NIST_GCM_16:
531 		/*
532 		 * auth_algorithm isn't used, but permit GMAC values
533 		 * for compatibility.
534 		 */
535 		switch (en->auth_algorithm) {
536 		case 0:
537 #ifdef COMPAT_FREEBSD12
538 		/* XXX: Really 13.0-current COMPAT. */
539 		case CRYPTO_AES_128_NIST_GMAC:
540 		case CRYPTO_AES_192_NIST_GMAC:
541 		case CRYPTO_AES_256_NIST_GMAC:
542 #endif
543 			break;
544 		default:
545 			return (EINVAL);
546 		}
547 		if (en->auth_key_len != 0)
548 			return (EINVAL);
549 		switch (en->tls_vminor) {
550 		case TLS_MINOR_VER_TWO:
551 			if (en->iv_len != TLS_AEAD_GCM_LEN)
552 				return (EINVAL);
553 			break;
554 		case TLS_MINOR_VER_THREE:
555 			if (en->iv_len != TLS_1_3_GCM_IV_LEN)
556 				return (EINVAL);
557 			break;
558 		default:
559 			return (EINVAL);
560 		}
561 		break;
562 	case CRYPTO_AES_CBC:
563 		switch (en->auth_algorithm) {
564 		case CRYPTO_SHA1_HMAC:
565 			break;
566 		case CRYPTO_SHA2_256_HMAC:
567 		case CRYPTO_SHA2_384_HMAC:
568 			if (en->tls_vminor != TLS_MINOR_VER_TWO)
569 				return (EINVAL);
570 			break;
571 		default:
572 			return (EINVAL);
573 		}
574 		if (en->auth_key_len == 0)
575 			return (EINVAL);
576 
577 		/*
578 		 * TLS 1.0 requires an implicit IV.  TLS 1.1 and 1.2
579 		 * use explicit IVs.
580 		 */
581 		switch (en->tls_vminor) {
582 		case TLS_MINOR_VER_ZERO:
583 			if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN)
584 				return (EINVAL);
585 			break;
586 		case TLS_MINOR_VER_ONE:
587 		case TLS_MINOR_VER_TWO:
588 			/* Ignore any supplied IV. */
589 			en->iv_len = 0;
590 			break;
591 		default:
592 			return (EINVAL);
593 		}
594 		break;
595 	case CRYPTO_CHACHA20_POLY1305:
596 		if (en->auth_algorithm != 0 || en->auth_key_len != 0)
597 			return (EINVAL);
598 		if (en->tls_vminor != TLS_MINOR_VER_TWO &&
599 		    en->tls_vminor != TLS_MINOR_VER_THREE)
600 			return (EINVAL);
601 		if (en->iv_len != TLS_CHACHA20_IV_LEN)
602 			return (EINVAL);
603 		break;
604 	default:
605 		return (EINVAL);
606 	}
607 
608 	error = ktls_start_kthreads();
609 	if (error != 0)
610 		return (error);
611 
612 	tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
613 
614 	counter_u64_add(ktls_offload_active, 1);
615 
616 	refcount_init(&tls->refcount, 1);
617 	if (direction == KTLS_RX) {
618 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls);
619 	} else {
620 		TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls);
621 		tls->inp = so->so_pcb;
622 		in_pcbref(tls->inp);
623 		tls->tx = true;
624 	}
625 
626 	tls->wq_index = ktls_get_cpu(so);
627 
628 	tls->params.cipher_algorithm = en->cipher_algorithm;
629 	tls->params.auth_algorithm = en->auth_algorithm;
630 	tls->params.tls_vmajor = en->tls_vmajor;
631 	tls->params.tls_vminor = en->tls_vminor;
632 	tls->params.flags = en->flags;
633 	tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen);
634 
635 	/* Set the header and trailer lengths. */
636 	tls->params.tls_hlen = sizeof(struct tls_record_layer);
637 	switch (en->cipher_algorithm) {
638 	case CRYPTO_AES_NIST_GCM_16:
639 		/*
640 		 * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte
641 		 * nonce.  TLS 1.3 uses a 12 byte implicit IV.
642 		 */
643 		if (en->tls_vminor < TLS_MINOR_VER_THREE)
644 			tls->params.tls_hlen += sizeof(uint64_t);
645 		tls->params.tls_tlen = AES_GMAC_HASH_LEN;
646 		tls->params.tls_bs = 1;
647 		break;
648 	case CRYPTO_AES_CBC:
649 		switch (en->auth_algorithm) {
650 		case CRYPTO_SHA1_HMAC:
651 			if (en->tls_vminor == TLS_MINOR_VER_ZERO) {
652 				/* Implicit IV, no nonce. */
653 				tls->sequential_records = true;
654 				tls->next_seqno = be64dec(en->rec_seq);
655 				STAILQ_INIT(&tls->pending_records);
656 			} else {
657 				tls->params.tls_hlen += AES_BLOCK_LEN;
658 			}
659 			tls->params.tls_tlen = AES_BLOCK_LEN +
660 			    SHA1_HASH_LEN;
661 			break;
662 		case CRYPTO_SHA2_256_HMAC:
663 			tls->params.tls_hlen += AES_BLOCK_LEN;
664 			tls->params.tls_tlen = AES_BLOCK_LEN +
665 			    SHA2_256_HASH_LEN;
666 			break;
667 		case CRYPTO_SHA2_384_HMAC:
668 			tls->params.tls_hlen += AES_BLOCK_LEN;
669 			tls->params.tls_tlen = AES_BLOCK_LEN +
670 			    SHA2_384_HASH_LEN;
671 			break;
672 		default:
673 			panic("invalid hmac");
674 		}
675 		tls->params.tls_bs = AES_BLOCK_LEN;
676 		break;
677 	case CRYPTO_CHACHA20_POLY1305:
678 		/*
679 		 * Chacha20 uses a 12 byte implicit IV.
680 		 */
681 		tls->params.tls_tlen = POLY1305_HASH_LEN;
682 		tls->params.tls_bs = 1;
683 		break;
684 	default:
685 		panic("invalid cipher");
686 	}
687 
688 	/*
689 	 * TLS 1.3 includes optional padding which we do not support,
690 	 * and also puts the "real" record type at the end of the
691 	 * encrypted data.
692 	 */
693 	if (en->tls_vminor == TLS_MINOR_VER_THREE)
694 		tls->params.tls_tlen += sizeof(uint8_t);
695 
696 	KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN,
697 	    ("TLS header length too long: %d", tls->params.tls_hlen));
698 	KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN,
699 	    ("TLS trailer length too long: %d", tls->params.tls_tlen));
700 
701 	if (en->auth_key_len != 0) {
702 		tls->params.auth_key_len = en->auth_key_len;
703 		tls->params.auth_key = malloc(en->auth_key_len, M_KTLS,
704 		    M_WAITOK);
705 		error = copyin(en->auth_key, tls->params.auth_key,
706 		    en->auth_key_len);
707 		if (error)
708 			goto out;
709 	}
710 
711 	tls->params.cipher_key_len = en->cipher_key_len;
712 	tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK);
713 	error = copyin(en->cipher_key, tls->params.cipher_key,
714 	    en->cipher_key_len);
715 	if (error)
716 		goto out;
717 
718 	/*
719 	 * This holds the implicit portion of the nonce for AEAD
720 	 * ciphers and the initial implicit IV for TLS 1.0.  The
721 	 * explicit portions of the IV are generated in ktls_frame().
722 	 */
723 	if (en->iv_len != 0) {
724 		tls->params.iv_len = en->iv_len;
725 		error = copyin(en->iv, tls->params.iv, en->iv_len);
726 		if (error)
727 			goto out;
728 
729 		/*
730 		 * For TLS 1.2 with GCM, generate an 8-byte nonce as a
731 		 * counter to generate unique explicit IVs.
732 		 *
733 		 * Store this counter in the last 8 bytes of the IV
734 		 * array so that it is 8-byte aligned.
735 		 */
736 		if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
737 		    en->tls_vminor == TLS_MINOR_VER_TWO)
738 			arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0);
739 	}
740 
741 	*tlsp = tls;
742 	return (0);
743 
744 out:
745 	ktls_free(tls);
746 	return (error);
747 }
748 
749 static struct ktls_session *
ktls_clone_session(struct ktls_session * tls,int direction)750 ktls_clone_session(struct ktls_session *tls, int direction)
751 {
752 	struct ktls_session *tls_new;
753 
754 	tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO);
755 
756 	counter_u64_add(ktls_offload_active, 1);
757 
758 	refcount_init(&tls_new->refcount, 1);
759 	if (direction == KTLS_RX) {
760 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag,
761 		    tls_new);
762 	} else {
763 		TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag,
764 		    tls_new);
765 		tls_new->inp = tls->inp;
766 		tls_new->tx = true;
767 		in_pcbref(tls_new->inp);
768 	}
769 
770 	/* Copy fields from existing session. */
771 	tls_new->params = tls->params;
772 	tls_new->wq_index = tls->wq_index;
773 
774 	/* Deep copy keys. */
775 	if (tls_new->params.auth_key != NULL) {
776 		tls_new->params.auth_key = malloc(tls->params.auth_key_len,
777 		    M_KTLS, M_WAITOK);
778 		memcpy(tls_new->params.auth_key, tls->params.auth_key,
779 		    tls->params.auth_key_len);
780 	}
781 
782 	tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS,
783 	    M_WAITOK);
784 	memcpy(tls_new->params.cipher_key, tls->params.cipher_key,
785 	    tls->params.cipher_key_len);
786 
787 	return (tls_new);
788 }
789 
790 #ifdef TCP_OFFLOAD
791 static int
ktls_try_toe(struct socket * so,struct ktls_session * tls,int direction)792 ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction)
793 {
794 	struct inpcb *inp;
795 	struct tcpcb *tp;
796 	int error;
797 
798 	inp = so->so_pcb;
799 	INP_WLOCK(inp);
800 	if (inp->inp_flags & INP_DROPPED) {
801 		INP_WUNLOCK(inp);
802 		return (ECONNRESET);
803 	}
804 	if (inp->inp_socket == NULL) {
805 		INP_WUNLOCK(inp);
806 		return (ECONNRESET);
807 	}
808 	tp = intotcpcb(inp);
809 	if (!(tp->t_flags & TF_TOE)) {
810 		INP_WUNLOCK(inp);
811 		return (EOPNOTSUPP);
812 	}
813 
814 	error = tcp_offload_alloc_tls_session(tp, tls, direction);
815 	INP_WUNLOCK(inp);
816 	if (error == 0) {
817 		tls->mode = TCP_TLS_MODE_TOE;
818 		switch (tls->params.cipher_algorithm) {
819 		case CRYPTO_AES_CBC:
820 			counter_u64_add(ktls_toe_cbc, 1);
821 			break;
822 		case CRYPTO_AES_NIST_GCM_16:
823 			counter_u64_add(ktls_toe_gcm, 1);
824 			break;
825 		case CRYPTO_CHACHA20_POLY1305:
826 			counter_u64_add(ktls_toe_chacha20, 1);
827 			break;
828 		}
829 	}
830 	return (error);
831 }
832 #endif
833 
834 /*
835  * Common code used when first enabling ifnet TLS on a connection or
836  * when allocating a new ifnet TLS session due to a routing change.
837  * This function allocates a new TLS send tag on whatever interface
838  * the connection is currently routed over.
839  */
840 static int
ktls_alloc_snd_tag(struct inpcb * inp,struct ktls_session * tls,bool force,struct m_snd_tag ** mstp)841 ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force,
842     struct m_snd_tag **mstp)
843 {
844 	union if_snd_tag_alloc_params params;
845 	struct ifnet *ifp;
846 	struct nhop_object *nh;
847 	struct tcpcb *tp;
848 	int error;
849 
850 	INP_RLOCK(inp);
851 	if (inp->inp_flags & INP_DROPPED) {
852 		INP_RUNLOCK(inp);
853 		return (ECONNRESET);
854 	}
855 	if (inp->inp_socket == NULL) {
856 		INP_RUNLOCK(inp);
857 		return (ECONNRESET);
858 	}
859 	tp = intotcpcb(inp);
860 
861 	/*
862 	 * Check administrative controls on ifnet TLS to determine if
863 	 * ifnet TLS should be denied.
864 	 *
865 	 * - Always permit 'force' requests.
866 	 * - ktls_ifnet_permitted == 0: always deny.
867 	 */
868 	if (!force && ktls_ifnet_permitted == 0) {
869 		INP_RUNLOCK(inp);
870 		return (ENXIO);
871 	}
872 
873 	/*
874 	 * XXX: Use the cached route in the inpcb to find the
875 	 * interface.  This should perhaps instead use
876 	 * rtalloc1_fib(dst, 0, 0, fibnum).  Since KTLS is only
877 	 * enabled after a connection has completed key negotiation in
878 	 * userland, the cached route will be present in practice.
879 	 */
880 	nh = inp->inp_route.ro_nh;
881 	if (nh == NULL) {
882 		INP_RUNLOCK(inp);
883 		return (ENXIO);
884 	}
885 	ifp = nh->nh_ifp;
886 	if_ref(ifp);
887 
888 	/*
889 	 * Allocate a TLS + ratelimit tag if the connection has an
890 	 * existing pacing rate.
891 	 */
892 	if (tp->t_pacing_rate != -1 &&
893 	    (if_getcapenable(ifp) & IFCAP_TXTLS_RTLMT) != 0) {
894 		params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT;
895 		params.tls_rate_limit.inp = inp;
896 		params.tls_rate_limit.tls = tls;
897 		params.tls_rate_limit.max_rate = tp->t_pacing_rate;
898 	} else {
899 		params.hdr.type = IF_SND_TAG_TYPE_TLS;
900 		params.tls.inp = inp;
901 		params.tls.tls = tls;
902 	}
903 	params.hdr.flowid = inp->inp_flowid;
904 	params.hdr.flowtype = inp->inp_flowtype;
905 	params.hdr.numa_domain = inp->inp_numa_domain;
906 	INP_RUNLOCK(inp);
907 
908 	if ((if_getcapenable(ifp) & IFCAP_MEXTPG) == 0) {
909 		error = EOPNOTSUPP;
910 		goto out;
911 	}
912 	if (inp->inp_vflag & INP_IPV6) {
913 		if ((if_getcapenable(ifp) & IFCAP_TXTLS6) == 0) {
914 			error = EOPNOTSUPP;
915 			goto out;
916 		}
917 	} else {
918 		if ((if_getcapenable(ifp) & IFCAP_TXTLS4) == 0) {
919 			error = EOPNOTSUPP;
920 			goto out;
921 		}
922 	}
923 	error = m_snd_tag_alloc(ifp, &params, mstp);
924 out:
925 	if_rele(ifp);
926 	return (error);
927 }
928 
929 /*
930  * Allocate an initial TLS receive tag for doing HW decryption of TLS
931  * data.
932  *
933  * This function allocates a new TLS receive tag on whatever interface
934  * the connection is currently routed over.  If the connection ends up
935  * using a different interface for receive this will get fixed up via
936  * ktls_input_ifp_mismatch as future packets arrive.
937  */
938 static int
ktls_alloc_rcv_tag(struct inpcb * inp,struct ktls_session * tls,struct m_snd_tag ** mstp)939 ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls,
940     struct m_snd_tag **mstp)
941 {
942 	union if_snd_tag_alloc_params params;
943 	struct ifnet *ifp;
944 	struct nhop_object *nh;
945 	int error;
946 
947 	if (!ktls_ocf_recrypt_supported(tls))
948 		return (ENXIO);
949 
950 	INP_RLOCK(inp);
951 	if (inp->inp_flags & INP_DROPPED) {
952 		INP_RUNLOCK(inp);
953 		return (ECONNRESET);
954 	}
955 	if (inp->inp_socket == NULL) {
956 		INP_RUNLOCK(inp);
957 		return (ECONNRESET);
958 	}
959 
960 	/*
961 	 * Check administrative controls on ifnet TLS to determine if
962 	 * ifnet TLS should be denied.
963 	 */
964 	if (ktls_ifnet_permitted == 0) {
965 		INP_RUNLOCK(inp);
966 		return (ENXIO);
967 	}
968 
969 	/*
970 	 * XXX: As with ktls_alloc_snd_tag, use the cached route in
971 	 * the inpcb to find the interface.
972 	 */
973 	nh = inp->inp_route.ro_nh;
974 	if (nh == NULL) {
975 		INP_RUNLOCK(inp);
976 		return (ENXIO);
977 	}
978 	ifp = nh->nh_ifp;
979 	if_ref(ifp);
980 	tls->rx_ifp = ifp;
981 
982 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
983 	params.hdr.flowid = inp->inp_flowid;
984 	params.hdr.flowtype = inp->inp_flowtype;
985 	params.hdr.numa_domain = inp->inp_numa_domain;
986 	params.tls_rx.inp = inp;
987 	params.tls_rx.tls = tls;
988 	params.tls_rx.vlan_id = 0;
989 
990 	INP_RUNLOCK(inp);
991 
992 	if (inp->inp_vflag & INP_IPV6) {
993 		if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS6)) == 0) {
994 			error = EOPNOTSUPP;
995 			goto out;
996 		}
997 	} else {
998 		if ((if_getcapenable2(ifp) & IFCAP2_BIT(IFCAP2_RXTLS4)) == 0) {
999 			error = EOPNOTSUPP;
1000 			goto out;
1001 		}
1002 	}
1003 	error = m_snd_tag_alloc(ifp, &params, mstp);
1004 
1005 	/*
1006 	 * If this connection is over a vlan, vlan_snd_tag_alloc
1007 	 * rewrites vlan_id with the saved interface.  Save the VLAN
1008 	 * ID for use in ktls_reset_receive_tag which allocates new
1009 	 * receive tags directly from the leaf interface bypassing
1010 	 * if_vlan.
1011 	 */
1012 	if (error == 0)
1013 		tls->rx_vlan_id = params.tls_rx.vlan_id;
1014 out:
1015 	return (error);
1016 }
1017 
1018 static int
ktls_try_ifnet(struct socket * so,struct ktls_session * tls,int direction,bool force)1019 ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction,
1020     bool force)
1021 {
1022 	struct m_snd_tag *mst;
1023 	int error;
1024 
1025 	switch (direction) {
1026 	case KTLS_TX:
1027 		error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst);
1028 		if (__predict_false(error != 0))
1029 			goto done;
1030 		break;
1031 	case KTLS_RX:
1032 		KASSERT(!force, ("%s: forced receive tag", __func__));
1033 		error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst);
1034 		if (__predict_false(error != 0))
1035 			goto done;
1036 		break;
1037 	default:
1038 		__assert_unreachable();
1039 	}
1040 
1041 	tls->mode = TCP_TLS_MODE_IFNET;
1042 	tls->snd_tag = mst;
1043 
1044 	switch (tls->params.cipher_algorithm) {
1045 	case CRYPTO_AES_CBC:
1046 		counter_u64_add(ktls_ifnet_cbc, 1);
1047 		break;
1048 	case CRYPTO_AES_NIST_GCM_16:
1049 		counter_u64_add(ktls_ifnet_gcm, 1);
1050 		break;
1051 	case CRYPTO_CHACHA20_POLY1305:
1052 		counter_u64_add(ktls_ifnet_chacha20, 1);
1053 		break;
1054 	default:
1055 		break;
1056 	}
1057 done:
1058 	return (error);
1059 }
1060 
1061 static void
ktls_use_sw(struct ktls_session * tls)1062 ktls_use_sw(struct ktls_session *tls)
1063 {
1064 	tls->mode = TCP_TLS_MODE_SW;
1065 	switch (tls->params.cipher_algorithm) {
1066 	case CRYPTO_AES_CBC:
1067 		counter_u64_add(ktls_sw_cbc, 1);
1068 		break;
1069 	case CRYPTO_AES_NIST_GCM_16:
1070 		counter_u64_add(ktls_sw_gcm, 1);
1071 		break;
1072 	case CRYPTO_CHACHA20_POLY1305:
1073 		counter_u64_add(ktls_sw_chacha20, 1);
1074 		break;
1075 	}
1076 }
1077 
1078 static int
ktls_try_sw(struct ktls_session * tls,int direction)1079 ktls_try_sw(struct ktls_session *tls, int direction)
1080 {
1081 	int error;
1082 
1083 	error = ktls_ocf_try(tls, direction);
1084 	if (error)
1085 		return (error);
1086 	ktls_use_sw(tls);
1087 	return (0);
1088 }
1089 
1090 /*
1091  * KTLS RX stores data in the socket buffer as a list of TLS records,
1092  * where each record is stored as a control message containg the TLS
1093  * header followed by data mbufs containing the decrypted data.  This
1094  * is different from KTLS TX which always uses an mb_ext_pgs mbuf for
1095  * both encrypted and decrypted data.  TLS records decrypted by a NIC
1096  * should be queued to the socket buffer as records, but encrypted
1097  * data which needs to be decrypted by software arrives as a stream of
1098  * regular mbufs which need to be converted.  In addition, there may
1099  * already be pending encrypted data in the socket buffer when KTLS RX
1100  * is enabled.
1101  *
1102  * To manage not-yet-decrypted data for KTLS RX, the following scheme
1103  * is used:
1104  *
1105  * - A single chain of NOTREADY mbufs is hung off of sb_mtls.
1106  *
1107  * - ktls_check_rx checks this chain of mbufs reading the TLS header
1108  *   from the first mbuf.  Once all of the data for that TLS record is
1109  *   queued, the socket is queued to a worker thread.
1110  *
1111  * - The worker thread calls ktls_decrypt to decrypt TLS records in
1112  *   the TLS chain.  Each TLS record is detached from the TLS chain,
1113  *   decrypted, and inserted into the regular socket buffer chain as
1114  *   record starting with a control message holding the TLS header and
1115  *   a chain of mbufs holding the encrypted data.
1116  */
1117 
1118 static void
sb_mark_notready(struct sockbuf * sb)1119 sb_mark_notready(struct sockbuf *sb)
1120 {
1121 	struct mbuf *m;
1122 
1123 	m = sb->sb_mb;
1124 	sb->sb_mtls = m;
1125 	sb->sb_mb = NULL;
1126 	sb->sb_mbtail = NULL;
1127 	sb->sb_lastrecord = NULL;
1128 	for (; m != NULL; m = m->m_next) {
1129 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL",
1130 		    __func__));
1131 		KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail",
1132 		    __func__));
1133 		KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len",
1134 		    __func__));
1135 		m->m_flags |= M_NOTREADY;
1136 		sb->sb_acc -= m->m_len;
1137 		sb->sb_tlscc += m->m_len;
1138 		sb->sb_mtlstail = m;
1139 	}
1140 	KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc,
1141 	    ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc,
1142 	    sb->sb_ccc));
1143 }
1144 
1145 /*
1146  * Return information about the pending TLS data in a socket
1147  * buffer.  On return, 'seqno' is set to the sequence number
1148  * of the next TLS record to be received, 'resid' is set to
1149  * the amount of bytes still needed for the last pending
1150  * record.  The function returns 'false' if the last pending
1151  * record contains a partial TLS header.  In that case, 'resid'
1152  * is the number of bytes needed to complete the TLS header.
1153  */
1154 bool
ktls_pending_rx_info(struct sockbuf * sb,uint64_t * seqnop,size_t * residp)1155 ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp)
1156 {
1157 	struct tls_record_layer hdr;
1158 	struct mbuf *m;
1159 	uint64_t seqno;
1160 	size_t resid;
1161 	u_int offset, record_len;
1162 
1163 	SOCKBUF_LOCK_ASSERT(sb);
1164 	MPASS(sb->sb_flags & SB_TLS_RX);
1165 	seqno = sb->sb_tls_seqno;
1166 	resid = sb->sb_tlscc;
1167 	m = sb->sb_mtls;
1168 	offset = 0;
1169 
1170 	if (resid == 0) {
1171 		*seqnop = seqno;
1172 		*residp = 0;
1173 		return (true);
1174 	}
1175 
1176 	for (;;) {
1177 		seqno++;
1178 
1179 		if (resid < sizeof(hdr)) {
1180 			*seqnop = seqno;
1181 			*residp = sizeof(hdr) - resid;
1182 			return (false);
1183 		}
1184 
1185 		m_copydata(m, offset, sizeof(hdr), (void *)&hdr);
1186 
1187 		record_len = sizeof(hdr) + ntohs(hdr.tls_length);
1188 		if (resid <= record_len) {
1189 			*seqnop = seqno;
1190 			*residp = record_len - resid;
1191 			return (true);
1192 		}
1193 		resid -= record_len;
1194 
1195 		while (record_len != 0) {
1196 			if (m->m_len - offset > record_len) {
1197 				offset += record_len;
1198 				break;
1199 			}
1200 
1201 			record_len -= (m->m_len - offset);
1202 			offset = 0;
1203 			m = m->m_next;
1204 		}
1205 	}
1206 }
1207 
1208 int
ktls_enable_rx(struct socket * so,struct tls_enable * en)1209 ktls_enable_rx(struct socket *so, struct tls_enable *en)
1210 {
1211 	struct ktls_session *tls;
1212 	int error;
1213 
1214 	if (!ktls_offload_enable)
1215 		return (ENOTSUP);
1216 
1217 	counter_u64_add(ktls_offload_enable_calls, 1);
1218 
1219 	/*
1220 	 * This should always be true since only the TCP socket option
1221 	 * invokes this function.
1222 	 */
1223 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1224 		return (EINVAL);
1225 
1226 	/*
1227 	 * XXX: Don't overwrite existing sessions.  We should permit
1228 	 * this to support rekeying in the future.
1229 	 */
1230 	if (so->so_rcv.sb_tls_info != NULL)
1231 		return (EALREADY);
1232 
1233 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1234 		return (ENOTSUP);
1235 
1236 	error = ktls_create_session(so, en, &tls, KTLS_RX);
1237 	if (error)
1238 		return (error);
1239 
1240 	error = ktls_ocf_try(tls, KTLS_RX);
1241 	if (error) {
1242 		ktls_free(tls);
1243 		return (error);
1244 	}
1245 
1246 	/*
1247 	 * Serialize with soreceive_generic() and make sure that we're not
1248 	 * operating on a listening socket.
1249 	 */
1250 	error = SOCK_IO_RECV_LOCK(so, SBL_WAIT);
1251 	if (error) {
1252 		ktls_free(tls);
1253 		return (error);
1254 	}
1255 
1256 	/* Mark the socket as using TLS offload. */
1257 	SOCK_RECVBUF_LOCK(so);
1258 	if (__predict_false(so->so_rcv.sb_tls_info != NULL)) {
1259 		SOCK_RECVBUF_UNLOCK(so);
1260 		SOCK_IO_RECV_UNLOCK(so);
1261 		ktls_free(tls);
1262 		return (EALREADY);
1263 	}
1264 	so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq);
1265 	so->so_rcv.sb_tls_info = tls;
1266 	so->so_rcv.sb_flags |= SB_TLS_RX;
1267 
1268 	/* Mark existing data as not ready until it can be decrypted. */
1269 	sb_mark_notready(&so->so_rcv);
1270 	ktls_check_rx(&so->so_rcv);
1271 	SOCK_RECVBUF_UNLOCK(so);
1272 	SOCK_IO_RECV_UNLOCK(so);
1273 
1274 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1275 #ifdef TCP_OFFLOAD
1276 	error = ktls_try_toe(so, tls, KTLS_RX);
1277 	if (error)
1278 #endif
1279 		error = ktls_try_ifnet(so, tls, KTLS_RX, false);
1280 	if (error)
1281 		ktls_use_sw(tls);
1282 
1283 	counter_u64_add(ktls_offload_total, 1);
1284 
1285 	return (0);
1286 }
1287 
1288 int
ktls_enable_tx(struct socket * so,struct tls_enable * en)1289 ktls_enable_tx(struct socket *so, struct tls_enable *en)
1290 {
1291 	struct ktls_session *tls;
1292 	struct inpcb *inp;
1293 	struct tcpcb *tp;
1294 	int error;
1295 
1296 	if (!ktls_offload_enable)
1297 		return (ENOTSUP);
1298 
1299 	counter_u64_add(ktls_offload_enable_calls, 1);
1300 
1301 	/*
1302 	 * This should always be true since only the TCP socket option
1303 	 * invokes this function.
1304 	 */
1305 	if (so->so_proto->pr_protocol != IPPROTO_TCP)
1306 		return (EINVAL);
1307 
1308 	/*
1309 	 * XXX: Don't overwrite existing sessions.  We should permit
1310 	 * this to support rekeying in the future.
1311 	 */
1312 	if (so->so_snd.sb_tls_info != NULL)
1313 		return (EALREADY);
1314 
1315 	if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable)
1316 		return (ENOTSUP);
1317 
1318 	/* TLS requires ext pgs */
1319 	if (mb_use_ext_pgs == 0)
1320 		return (ENXIO);
1321 
1322 	error = ktls_create_session(so, en, &tls, KTLS_TX);
1323 	if (error)
1324 		return (error);
1325 
1326 	/* Prefer TOE -> ifnet TLS -> software TLS. */
1327 #ifdef TCP_OFFLOAD
1328 	error = ktls_try_toe(so, tls, KTLS_TX);
1329 	if (error)
1330 #endif
1331 		error = ktls_try_ifnet(so, tls, KTLS_TX, false);
1332 	if (error)
1333 		error = ktls_try_sw(tls, KTLS_TX);
1334 
1335 	if (error) {
1336 		ktls_free(tls);
1337 		return (error);
1338 	}
1339 
1340 	/*
1341 	 * Serialize with sosend_generic() and make sure that we're not
1342 	 * operating on a listening socket.
1343 	 */
1344 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1345 	if (error) {
1346 		ktls_free(tls);
1347 		return (error);
1348 	}
1349 
1350 	/*
1351 	 * Write lock the INP when setting sb_tls_info so that
1352 	 * routines in tcp_ratelimit.c can read sb_tls_info while
1353 	 * holding the INP lock.
1354 	 */
1355 	inp = so->so_pcb;
1356 	INP_WLOCK(inp);
1357 	SOCK_SENDBUF_LOCK(so);
1358 	if (__predict_false(so->so_snd.sb_tls_info != NULL)) {
1359 		SOCK_SENDBUF_UNLOCK(so);
1360 		INP_WUNLOCK(inp);
1361 		SOCK_IO_SEND_UNLOCK(so);
1362 		ktls_free(tls);
1363 		return (EALREADY);
1364 	}
1365 	so->so_snd.sb_tls_seqno = be64dec(en->rec_seq);
1366 	so->so_snd.sb_tls_info = tls;
1367 	if (tls->mode != TCP_TLS_MODE_SW) {
1368 		tp = intotcpcb(inp);
1369 		MPASS(tp->t_nic_ktls_xmit == 0);
1370 		tp->t_nic_ktls_xmit = 1;
1371 		if (tp->t_fb->tfb_hwtls_change != NULL)
1372 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1373 	}
1374 	SOCK_SENDBUF_UNLOCK(so);
1375 	INP_WUNLOCK(inp);
1376 	SOCK_IO_SEND_UNLOCK(so);
1377 
1378 	counter_u64_add(ktls_offload_total, 1);
1379 
1380 	return (0);
1381 }
1382 
1383 int
ktls_get_rx_mode(struct socket * so,int * modep)1384 ktls_get_rx_mode(struct socket *so, int *modep)
1385 {
1386 	struct ktls_session *tls;
1387 	struct inpcb *inp __diagused;
1388 
1389 	if (SOLISTENING(so))
1390 		return (EINVAL);
1391 	inp = so->so_pcb;
1392 	INP_WLOCK_ASSERT(inp);
1393 	SOCK_RECVBUF_LOCK(so);
1394 	tls = so->so_rcv.sb_tls_info;
1395 	if (tls == NULL)
1396 		*modep = TCP_TLS_MODE_NONE;
1397 	else
1398 		*modep = tls->mode;
1399 	SOCK_RECVBUF_UNLOCK(so);
1400 	return (0);
1401 }
1402 
1403 /*
1404  * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number.
1405  *
1406  * This function gets information about the next TCP- and TLS-
1407  * sequence number to be processed by the TLS receive worker
1408  * thread. The information is extracted from the given "inpcb"
1409  * structure. The values are stored in host endian format at the two
1410  * given output pointer locations. The TCP sequence number points to
1411  * the beginning of the TLS header.
1412  *
1413  * This function returns zero on success, else a non-zero error code
1414  * is returned.
1415  */
1416 int
ktls_get_rx_sequence(struct inpcb * inp,uint32_t * tcpseq,uint64_t * tlsseq)1417 ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq)
1418 {
1419 	struct socket *so;
1420 	struct tcpcb *tp;
1421 
1422 	INP_RLOCK(inp);
1423 	so = inp->inp_socket;
1424 	if (__predict_false(so == NULL)) {
1425 		INP_RUNLOCK(inp);
1426 		return (EINVAL);
1427 	}
1428 	if (inp->inp_flags & INP_DROPPED) {
1429 		INP_RUNLOCK(inp);
1430 		return (ECONNRESET);
1431 	}
1432 
1433 	tp = intotcpcb(inp);
1434 	MPASS(tp != NULL);
1435 
1436 	SOCKBUF_LOCK(&so->so_rcv);
1437 	*tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc;
1438 	*tlsseq = so->so_rcv.sb_tls_seqno;
1439 	SOCKBUF_UNLOCK(&so->so_rcv);
1440 
1441 	INP_RUNLOCK(inp);
1442 
1443 	return (0);
1444 }
1445 
1446 int
ktls_get_tx_mode(struct socket * so,int * modep)1447 ktls_get_tx_mode(struct socket *so, int *modep)
1448 {
1449 	struct ktls_session *tls;
1450 	struct inpcb *inp __diagused;
1451 
1452 	if (SOLISTENING(so))
1453 		return (EINVAL);
1454 	inp = so->so_pcb;
1455 	INP_WLOCK_ASSERT(inp);
1456 	SOCK_SENDBUF_LOCK(so);
1457 	tls = so->so_snd.sb_tls_info;
1458 	if (tls == NULL)
1459 		*modep = TCP_TLS_MODE_NONE;
1460 	else
1461 		*modep = tls->mode;
1462 	SOCK_SENDBUF_UNLOCK(so);
1463 	return (0);
1464 }
1465 
1466 /*
1467  * Switch between SW and ifnet TLS sessions as requested.
1468  */
1469 int
ktls_set_tx_mode(struct socket * so,int mode)1470 ktls_set_tx_mode(struct socket *so, int mode)
1471 {
1472 	struct ktls_session *tls, *tls_new;
1473 	struct inpcb *inp;
1474 	struct tcpcb *tp;
1475 	int error;
1476 
1477 	if (SOLISTENING(so))
1478 		return (EINVAL);
1479 	switch (mode) {
1480 	case TCP_TLS_MODE_SW:
1481 	case TCP_TLS_MODE_IFNET:
1482 		break;
1483 	default:
1484 		return (EINVAL);
1485 	}
1486 
1487 	inp = so->so_pcb;
1488 	INP_WLOCK_ASSERT(inp);
1489 	tp = intotcpcb(inp);
1490 
1491 	if (mode == TCP_TLS_MODE_IFNET) {
1492 		/* Don't allow enabling ifnet ktls multiple times */
1493 		if (tp->t_nic_ktls_xmit)
1494 			return (EALREADY);
1495 
1496 		/*
1497 		 * Don't enable ifnet ktls if we disabled it due to an
1498 		 * excessive retransmission rate
1499 		 */
1500 		if (tp->t_nic_ktls_xmit_dis)
1501 			return (ENXIO);
1502 	}
1503 
1504 	SOCKBUF_LOCK(&so->so_snd);
1505 	tls = so->so_snd.sb_tls_info;
1506 	if (tls == NULL) {
1507 		SOCKBUF_UNLOCK(&so->so_snd);
1508 		return (0);
1509 	}
1510 
1511 	if (tls->mode == mode) {
1512 		SOCKBUF_UNLOCK(&so->so_snd);
1513 		return (0);
1514 	}
1515 
1516 	tls = ktls_hold(tls);
1517 	SOCKBUF_UNLOCK(&so->so_snd);
1518 	INP_WUNLOCK(inp);
1519 
1520 	tls_new = ktls_clone_session(tls, KTLS_TX);
1521 
1522 	if (mode == TCP_TLS_MODE_IFNET)
1523 		error = ktls_try_ifnet(so, tls_new, KTLS_TX, true);
1524 	else
1525 		error = ktls_try_sw(tls_new, KTLS_TX);
1526 	if (error) {
1527 		counter_u64_add(ktls_switch_failed, 1);
1528 		ktls_free(tls_new);
1529 		ktls_free(tls);
1530 		INP_WLOCK(inp);
1531 		return (error);
1532 	}
1533 
1534 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
1535 	if (error) {
1536 		counter_u64_add(ktls_switch_failed, 1);
1537 		ktls_free(tls_new);
1538 		ktls_free(tls);
1539 		INP_WLOCK(inp);
1540 		return (error);
1541 	}
1542 
1543 	/*
1544 	 * If we raced with another session change, keep the existing
1545 	 * session.
1546 	 */
1547 	if (tls != so->so_snd.sb_tls_info) {
1548 		counter_u64_add(ktls_switch_failed, 1);
1549 		SOCK_IO_SEND_UNLOCK(so);
1550 		ktls_free(tls_new);
1551 		ktls_free(tls);
1552 		INP_WLOCK(inp);
1553 		return (EBUSY);
1554 	}
1555 
1556 	INP_WLOCK(inp);
1557 	SOCKBUF_LOCK(&so->so_snd);
1558 	so->so_snd.sb_tls_info = tls_new;
1559 	if (tls_new->mode != TCP_TLS_MODE_SW) {
1560 		MPASS(tp->t_nic_ktls_xmit == 0);
1561 		tp->t_nic_ktls_xmit = 1;
1562 		if (tp->t_fb->tfb_hwtls_change != NULL)
1563 			(*tp->t_fb->tfb_hwtls_change)(tp, 1);
1564 	}
1565 	SOCKBUF_UNLOCK(&so->so_snd);
1566 	SOCK_IO_SEND_UNLOCK(so);
1567 
1568 	/*
1569 	 * Drop two references on 'tls'.  The first is for the
1570 	 * ktls_hold() above.  The second drops the reference from the
1571 	 * socket buffer.
1572 	 */
1573 	KASSERT(tls->refcount >= 2, ("too few references on old session"));
1574 	ktls_free(tls);
1575 	ktls_free(tls);
1576 
1577 	if (mode == TCP_TLS_MODE_IFNET)
1578 		counter_u64_add(ktls_switch_to_ifnet, 1);
1579 	else
1580 		counter_u64_add(ktls_switch_to_sw, 1);
1581 
1582 	return (0);
1583 }
1584 
1585 /*
1586  * Try to allocate a new TLS receive tag.  This task is scheduled when
1587  * sbappend_ktls_rx detects an input path change.  If a new tag is
1588  * allocated, replace the tag in the TLS session.  If a new tag cannot
1589  * be allocated, let the session fall back to software decryption.
1590  */
1591 static void
ktls_reset_receive_tag(void * context,int pending)1592 ktls_reset_receive_tag(void *context, int pending)
1593 {
1594 	union if_snd_tag_alloc_params params;
1595 	struct ktls_session *tls;
1596 	struct m_snd_tag *mst;
1597 	struct inpcb *inp;
1598 	struct ifnet *ifp;
1599 	struct socket *so;
1600 	int error;
1601 
1602 	MPASS(pending == 1);
1603 
1604 	tls = context;
1605 	so = tls->so;
1606 	inp = so->so_pcb;
1607 	ifp = NULL;
1608 
1609 	INP_RLOCK(inp);
1610 	if (inp->inp_flags & INP_DROPPED) {
1611 		INP_RUNLOCK(inp);
1612 		goto out;
1613 	}
1614 
1615 	SOCKBUF_LOCK(&so->so_rcv);
1616 	mst = tls->snd_tag;
1617 	tls->snd_tag = NULL;
1618 	if (mst != NULL)
1619 		m_snd_tag_rele(mst);
1620 
1621 	ifp = tls->rx_ifp;
1622 	if_ref(ifp);
1623 	SOCKBUF_UNLOCK(&so->so_rcv);
1624 
1625 	params.hdr.type = IF_SND_TAG_TYPE_TLS_RX;
1626 	params.hdr.flowid = inp->inp_flowid;
1627 	params.hdr.flowtype = inp->inp_flowtype;
1628 	params.hdr.numa_domain = inp->inp_numa_domain;
1629 	params.tls_rx.inp = inp;
1630 	params.tls_rx.tls = tls;
1631 	params.tls_rx.vlan_id = tls->rx_vlan_id;
1632 	INP_RUNLOCK(inp);
1633 
1634 	if (inp->inp_vflag & INP_IPV6) {
1635 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS6) == 0)
1636 			goto out;
1637 	} else {
1638 		if ((if_getcapenable2(ifp) & IFCAP2_RXTLS4) == 0)
1639 			goto out;
1640 	}
1641 
1642 	error = m_snd_tag_alloc(ifp, &params, &mst);
1643 	if (error == 0) {
1644 		SOCKBUF_LOCK(&so->so_rcv);
1645 		tls->snd_tag = mst;
1646 		SOCKBUF_UNLOCK(&so->so_rcv);
1647 
1648 		counter_u64_add(ktls_ifnet_reset, 1);
1649 	} else {
1650 		/*
1651 		 * Just fall back to software decryption if a tag
1652 		 * cannot be allocated leaving the connection intact.
1653 		 * If a future input path change switches to another
1654 		 * interface this connection will resume ifnet TLS.
1655 		 */
1656 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1657 	}
1658 
1659 out:
1660 	mtx_pool_lock(mtxpool_sleep, tls);
1661 	tls->reset_pending = false;
1662 	mtx_pool_unlock(mtxpool_sleep, tls);
1663 
1664 	if (ifp != NULL)
1665 		if_rele(ifp);
1666 	CURVNET_SET(so->so_vnet);
1667 	sorele(so);
1668 	CURVNET_RESTORE();
1669 	ktls_free(tls);
1670 }
1671 
1672 /*
1673  * Try to allocate a new TLS send tag.  This task is scheduled when
1674  * ip_output detects a route change while trying to transmit a packet
1675  * holding a TLS record.  If a new tag is allocated, replace the tag
1676  * in the TLS session.  Subsequent packets on the connection will use
1677  * the new tag.  If a new tag cannot be allocated, drop the
1678  * connection.
1679  */
1680 static void
ktls_reset_send_tag(void * context,int pending)1681 ktls_reset_send_tag(void *context, int pending)
1682 {
1683 	struct epoch_tracker et;
1684 	struct ktls_session *tls;
1685 	struct m_snd_tag *old, *new;
1686 	struct inpcb *inp;
1687 	struct tcpcb *tp;
1688 	int error;
1689 
1690 	MPASS(pending == 1);
1691 
1692 	tls = context;
1693 	inp = tls->inp;
1694 
1695 	/*
1696 	 * Free the old tag first before allocating a new one.
1697 	 * ip[6]_output_send() will treat a NULL send tag the same as
1698 	 * an ifp mismatch and drop packets until a new tag is
1699 	 * allocated.
1700 	 *
1701 	 * Write-lock the INP when changing tls->snd_tag since
1702 	 * ip[6]_output_send() holds a read-lock when reading the
1703 	 * pointer.
1704 	 */
1705 	INP_WLOCK(inp);
1706 	old = tls->snd_tag;
1707 	tls->snd_tag = NULL;
1708 	INP_WUNLOCK(inp);
1709 	if (old != NULL)
1710 		m_snd_tag_rele(old);
1711 
1712 	error = ktls_alloc_snd_tag(inp, tls, true, &new);
1713 
1714 	if (error == 0) {
1715 		INP_WLOCK(inp);
1716 		tls->snd_tag = new;
1717 		mtx_pool_lock(mtxpool_sleep, tls);
1718 		tls->reset_pending = false;
1719 		mtx_pool_unlock(mtxpool_sleep, tls);
1720 		INP_WUNLOCK(inp);
1721 
1722 		counter_u64_add(ktls_ifnet_reset, 1);
1723 
1724 		/*
1725 		 * XXX: Should we kick tcp_output explicitly now that
1726 		 * the send tag is fixed or just rely on timers?
1727 		 */
1728 	} else {
1729 		NET_EPOCH_ENTER(et);
1730 		INP_WLOCK(inp);
1731 		if (!(inp->inp_flags & INP_DROPPED)) {
1732 			tp = intotcpcb(inp);
1733 			CURVNET_SET(inp->inp_vnet);
1734 			tp = tcp_drop(tp, ECONNABORTED);
1735 			CURVNET_RESTORE();
1736 			if (tp != NULL) {
1737 				counter_u64_add(ktls_ifnet_reset_dropped, 1);
1738 				INP_WUNLOCK(inp);
1739 			}
1740 		} else
1741 			INP_WUNLOCK(inp);
1742 		NET_EPOCH_EXIT(et);
1743 
1744 		counter_u64_add(ktls_ifnet_reset_failed, 1);
1745 
1746 		/*
1747 		 * Leave reset_pending true to avoid future tasks while
1748 		 * the socket goes away.
1749 		 */
1750 	}
1751 
1752 	ktls_free(tls);
1753 }
1754 
1755 void
ktls_input_ifp_mismatch(struct sockbuf * sb,struct ifnet * ifp)1756 ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp)
1757 {
1758 	struct ktls_session *tls;
1759 	struct socket *so;
1760 
1761 	SOCKBUF_LOCK_ASSERT(sb);
1762 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
1763 	    __func__, sb));
1764 	so = __containerof(sb, struct socket, so_rcv);
1765 
1766 	tls = sb->sb_tls_info;
1767 	if_rele(tls->rx_ifp);
1768 	if_ref(ifp);
1769 	tls->rx_ifp = ifp;
1770 
1771 	/*
1772 	 * See if we should schedule a task to update the receive tag for
1773 	 * this session.
1774 	 */
1775 	mtx_pool_lock(mtxpool_sleep, tls);
1776 	if (!tls->reset_pending) {
1777 		(void) ktls_hold(tls);
1778 		soref(so);
1779 		tls->so = so;
1780 		tls->reset_pending = true;
1781 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1782 	}
1783 	mtx_pool_unlock(mtxpool_sleep, tls);
1784 }
1785 
1786 int
ktls_output_eagain(struct inpcb * inp,struct ktls_session * tls)1787 ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls)
1788 {
1789 
1790 	if (inp == NULL)
1791 		return (ENOBUFS);
1792 
1793 	INP_LOCK_ASSERT(inp);
1794 
1795 	/*
1796 	 * See if we should schedule a task to update the send tag for
1797 	 * this session.
1798 	 */
1799 	mtx_pool_lock(mtxpool_sleep, tls);
1800 	if (!tls->reset_pending) {
1801 		(void) ktls_hold(tls);
1802 		tls->reset_pending = true;
1803 		taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task);
1804 	}
1805 	mtx_pool_unlock(mtxpool_sleep, tls);
1806 	return (ENOBUFS);
1807 }
1808 
1809 #ifdef RATELIMIT
1810 int
ktls_modify_txrtlmt(struct ktls_session * tls,uint64_t max_pacing_rate)1811 ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate)
1812 {
1813 	union if_snd_tag_modify_params params = {
1814 		.rate_limit.max_rate = max_pacing_rate,
1815 		.rate_limit.flags = M_NOWAIT,
1816 	};
1817 	struct m_snd_tag *mst;
1818 
1819 	/* Can't get to the inp, but it should be locked. */
1820 	/* INP_LOCK_ASSERT(inp); */
1821 
1822 	MPASS(tls->mode == TCP_TLS_MODE_IFNET);
1823 
1824 	if (tls->snd_tag == NULL) {
1825 		/*
1826 		 * Resetting send tag, ignore this change.  The
1827 		 * pending reset may or may not see this updated rate
1828 		 * in the tcpcb.  If it doesn't, we will just lose
1829 		 * this rate change.
1830 		 */
1831 		return (0);
1832 	}
1833 
1834 	mst = tls->snd_tag;
1835 
1836 	MPASS(mst != NULL);
1837 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT);
1838 
1839 	return (mst->sw->snd_tag_modify(mst, &params));
1840 }
1841 #endif
1842 
1843 static void
ktls_destroy_help(void * context,int pending __unused)1844 ktls_destroy_help(void *context, int pending __unused)
1845 {
1846 	ktls_destroy(context);
1847 }
1848 
1849 void
ktls_destroy(struct ktls_session * tls)1850 ktls_destroy(struct ktls_session *tls)
1851 {
1852 	struct inpcb *inp;
1853 	struct tcpcb *tp;
1854 	bool wlocked;
1855 
1856 	MPASS(tls->refcount == 0);
1857 
1858 	inp = tls->inp;
1859 	if (tls->tx) {
1860 		wlocked = INP_WLOCKED(inp);
1861 		if (!wlocked && !INP_TRY_WLOCK(inp)) {
1862 			/*
1863 			 * rwlocks read locks are anonymous, and there
1864 			 * is no way to know if our current thread
1865 			 * holds an rlock on the inp.  As a rough
1866 			 * estimate, check to see if the thread holds
1867 			 * *any* rlocks at all.  If it does not, then we
1868 			 * know that we don't hold the inp rlock, and
1869 			 * can safely take the wlock
1870 			 */
1871 			if (curthread->td_rw_rlocks == 0) {
1872 				INP_WLOCK(inp);
1873 			} else {
1874 				/*
1875 				 * We might hold the rlock, so let's
1876 				 * do the destroy in a taskqueue
1877 				 * context to avoid a potential
1878 				 * deadlock.  This should be very
1879 				 * rare.
1880 				 */
1881 				counter_u64_add(ktls_destroy_task, 1);
1882 				TASK_INIT(&tls->destroy_task, 0,
1883 				    ktls_destroy_help, tls);
1884 				(void)taskqueue_enqueue(taskqueue_thread,
1885 				    &tls->destroy_task);
1886 				return;
1887 			}
1888 		}
1889 	}
1890 
1891 	if (tls->sequential_records) {
1892 		struct mbuf *m, *n;
1893 		int page_count;
1894 
1895 		STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) {
1896 			page_count = m->m_epg_enc_cnt;
1897 			while (page_count > 0) {
1898 				KASSERT(page_count >= m->m_epg_nrdy,
1899 				    ("%s: too few pages", __func__));
1900 				page_count -= m->m_epg_nrdy;
1901 				m = m_free(m);
1902 			}
1903 		}
1904 	}
1905 
1906 	counter_u64_add(ktls_offload_active, -1);
1907 	switch (tls->mode) {
1908 	case TCP_TLS_MODE_SW:
1909 		switch (tls->params.cipher_algorithm) {
1910 		case CRYPTO_AES_CBC:
1911 			counter_u64_add(ktls_sw_cbc, -1);
1912 			break;
1913 		case CRYPTO_AES_NIST_GCM_16:
1914 			counter_u64_add(ktls_sw_gcm, -1);
1915 			break;
1916 		case CRYPTO_CHACHA20_POLY1305:
1917 			counter_u64_add(ktls_sw_chacha20, -1);
1918 			break;
1919 		}
1920 		break;
1921 	case TCP_TLS_MODE_IFNET:
1922 		switch (tls->params.cipher_algorithm) {
1923 		case CRYPTO_AES_CBC:
1924 			counter_u64_add(ktls_ifnet_cbc, -1);
1925 			break;
1926 		case CRYPTO_AES_NIST_GCM_16:
1927 			counter_u64_add(ktls_ifnet_gcm, -1);
1928 			break;
1929 		case CRYPTO_CHACHA20_POLY1305:
1930 			counter_u64_add(ktls_ifnet_chacha20, -1);
1931 			break;
1932 		}
1933 		if (tls->snd_tag != NULL)
1934 			m_snd_tag_rele(tls->snd_tag);
1935 		if (tls->rx_ifp != NULL)
1936 			if_rele(tls->rx_ifp);
1937 		if (tls->tx) {
1938 			INP_WLOCK_ASSERT(inp);
1939 			tp = intotcpcb(inp);
1940 			MPASS(tp->t_nic_ktls_xmit == 1);
1941 			tp->t_nic_ktls_xmit = 0;
1942 		}
1943 		break;
1944 #ifdef TCP_OFFLOAD
1945 	case TCP_TLS_MODE_TOE:
1946 		switch (tls->params.cipher_algorithm) {
1947 		case CRYPTO_AES_CBC:
1948 			counter_u64_add(ktls_toe_cbc, -1);
1949 			break;
1950 		case CRYPTO_AES_NIST_GCM_16:
1951 			counter_u64_add(ktls_toe_gcm, -1);
1952 			break;
1953 		case CRYPTO_CHACHA20_POLY1305:
1954 			counter_u64_add(ktls_toe_chacha20, -1);
1955 			break;
1956 		}
1957 		break;
1958 #endif
1959 	}
1960 	if (tls->ocf_session != NULL)
1961 		ktls_ocf_free(tls);
1962 	if (tls->params.auth_key != NULL) {
1963 		zfree(tls->params.auth_key, M_KTLS);
1964 		tls->params.auth_key = NULL;
1965 		tls->params.auth_key_len = 0;
1966 	}
1967 	if (tls->params.cipher_key != NULL) {
1968 		zfree(tls->params.cipher_key, M_KTLS);
1969 		tls->params.cipher_key = NULL;
1970 		tls->params.cipher_key_len = 0;
1971 	}
1972 	if (tls->tx) {
1973 		INP_WLOCK_ASSERT(inp);
1974 		if (!in_pcbrele_wlocked(inp) && !wlocked)
1975 			INP_WUNLOCK(inp);
1976 	}
1977 	explicit_bzero(tls->params.iv, sizeof(tls->params.iv));
1978 
1979 	uma_zfree(ktls_session_zone, tls);
1980 }
1981 
1982 void
ktls_seq(struct sockbuf * sb,struct mbuf * m)1983 ktls_seq(struct sockbuf *sb, struct mbuf *m)
1984 {
1985 
1986 	for (; m != NULL; m = m->m_next) {
1987 		KASSERT((m->m_flags & M_EXTPG) != 0,
1988 		    ("ktls_seq: mapped mbuf %p", m));
1989 
1990 		m->m_epg_seqno = sb->sb_tls_seqno;
1991 		sb->sb_tls_seqno++;
1992 	}
1993 }
1994 
1995 /*
1996  * Add TLS framing (headers and trailers) to a chain of mbufs.  Each
1997  * mbuf in the chain must be an unmapped mbuf.  The payload of the
1998  * mbuf must be populated with the payload of each TLS record.
1999  *
2000  * The record_type argument specifies the TLS record type used when
2001  * populating the TLS header.
2002  *
2003  * The enq_count argument on return is set to the number of pages of
2004  * payload data for this entire chain that need to be encrypted via SW
2005  * encryption.  The returned value should be passed to ktls_enqueue
2006  * when scheduling encryption of this chain of mbufs.  To handle the
2007  * special case of empty fragments for TLS 1.0 sessions, an empty
2008  * fragment counts as one page.
2009  */
2010 void
ktls_frame(struct mbuf * top,struct ktls_session * tls,int * enq_cnt,uint8_t record_type)2011 ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt,
2012     uint8_t record_type)
2013 {
2014 	struct tls_record_layer *tlshdr;
2015 	struct mbuf *m;
2016 	uint64_t *noncep;
2017 	uint16_t tls_len;
2018 	int maxlen __diagused;
2019 
2020 	maxlen = tls->params.max_frame_len;
2021 	*enq_cnt = 0;
2022 	for (m = top; m != NULL; m = m->m_next) {
2023 		/*
2024 		 * All mbufs in the chain should be TLS records whose
2025 		 * payload does not exceed the maximum frame length.
2026 		 *
2027 		 * Empty TLS 1.0 records are permitted when using CBC.
2028 		 */
2029 		KASSERT(m->m_len <= maxlen && m->m_len >= 0 &&
2030 		    (m->m_len > 0 || ktls_permit_empty_frames(tls)),
2031 		    ("ktls_frame: m %p len %d", m, m->m_len));
2032 
2033 		/*
2034 		 * TLS frames require unmapped mbufs to store session
2035 		 * info.
2036 		 */
2037 		KASSERT((m->m_flags & M_EXTPG) != 0,
2038 		    ("ktls_frame: mapped mbuf %p (top = %p)", m, top));
2039 
2040 		tls_len = m->m_len;
2041 
2042 		/* Save a reference to the session. */
2043 		m->m_epg_tls = ktls_hold(tls);
2044 
2045 		m->m_epg_hdrlen = tls->params.tls_hlen;
2046 		m->m_epg_trllen = tls->params.tls_tlen;
2047 		if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) {
2048 			int bs, delta;
2049 
2050 			/*
2051 			 * AES-CBC pads messages to a multiple of the
2052 			 * block size.  Note that the padding is
2053 			 * applied after the digest and the encryption
2054 			 * is done on the "plaintext || mac || padding".
2055 			 * At least one byte of padding is always
2056 			 * present.
2057 			 *
2058 			 * Compute the final trailer length assuming
2059 			 * at most one block of padding.
2060 			 * tls->params.tls_tlen is the maximum
2061 			 * possible trailer length (padding + digest).
2062 			 * delta holds the number of excess padding
2063 			 * bytes if the maximum were used.  Those
2064 			 * extra bytes are removed.
2065 			 */
2066 			bs = tls->params.tls_bs;
2067 			delta = (tls_len + tls->params.tls_tlen) & (bs - 1);
2068 			m->m_epg_trllen -= delta;
2069 		}
2070 		m->m_len += m->m_epg_hdrlen + m->m_epg_trllen;
2071 
2072 		/* Populate the TLS header. */
2073 		tlshdr = (void *)m->m_epg_hdr;
2074 		tlshdr->tls_vmajor = tls->params.tls_vmajor;
2075 
2076 		/*
2077 		 * TLS 1.3 masquarades as TLS 1.2 with a record type
2078 		 * of TLS_RLTYPE_APP.
2079 		 */
2080 		if (tls->params.tls_vminor == TLS_MINOR_VER_THREE &&
2081 		    tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) {
2082 			tlshdr->tls_vminor = TLS_MINOR_VER_TWO;
2083 			tlshdr->tls_type = TLS_RLTYPE_APP;
2084 			/* save the real record type for later */
2085 			m->m_epg_record_type = record_type;
2086 			m->m_epg_trail[0] = record_type;
2087 		} else {
2088 			tlshdr->tls_vminor = tls->params.tls_vminor;
2089 			tlshdr->tls_type = record_type;
2090 		}
2091 		tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr));
2092 
2093 		/*
2094 		 * Store nonces / explicit IVs after the end of the
2095 		 * TLS header.
2096 		 *
2097 		 * For GCM with TLS 1.2, an 8 byte nonce is copied
2098 		 * from the end of the IV.  The nonce is then
2099 		 * incremented for use by the next record.
2100 		 *
2101 		 * For CBC, a random nonce is inserted for TLS 1.1+.
2102 		 */
2103 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 &&
2104 		    tls->params.tls_vminor == TLS_MINOR_VER_TWO) {
2105 			noncep = (uint64_t *)(tls->params.iv + 8);
2106 			be64enc(tlshdr + 1, *noncep);
2107 			(*noncep)++;
2108 		} else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2109 		    tls->params.tls_vminor >= TLS_MINOR_VER_ONE)
2110 			arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0);
2111 
2112 		/*
2113 		 * When using SW encryption, mark the mbuf not ready.
2114 		 * It will be marked ready via sbready() after the
2115 		 * record has been encrypted.
2116 		 *
2117 		 * When using ifnet TLS, unencrypted TLS records are
2118 		 * sent down the stack to the NIC.
2119 		 */
2120 		if (tls->mode == TCP_TLS_MODE_SW) {
2121 			m->m_flags |= M_NOTREADY;
2122 			if (__predict_false(tls_len == 0)) {
2123 				/* TLS 1.0 empty fragment. */
2124 				m->m_epg_nrdy = 1;
2125 			} else
2126 				m->m_epg_nrdy = m->m_epg_npgs;
2127 			*enq_cnt += m->m_epg_nrdy;
2128 		}
2129 	}
2130 }
2131 
2132 bool
ktls_permit_empty_frames(struct ktls_session * tls)2133 ktls_permit_empty_frames(struct ktls_session *tls)
2134 {
2135 	return (tls->params.cipher_algorithm == CRYPTO_AES_CBC &&
2136 	    tls->params.tls_vminor == TLS_MINOR_VER_ZERO);
2137 }
2138 
2139 void
ktls_check_rx(struct sockbuf * sb)2140 ktls_check_rx(struct sockbuf *sb)
2141 {
2142 	struct tls_record_layer hdr;
2143 	struct ktls_wq *wq;
2144 	struct socket *so;
2145 	bool running;
2146 
2147 	SOCKBUF_LOCK_ASSERT(sb);
2148 	KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX",
2149 	    __func__, sb));
2150 	so = __containerof(sb, struct socket, so_rcv);
2151 
2152 	if (sb->sb_flags & SB_TLS_RX_RUNNING)
2153 		return;
2154 
2155 	/* Is there enough queued for a TLS header? */
2156 	if (sb->sb_tlscc < sizeof(hdr)) {
2157 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0)
2158 			so->so_error = EMSGSIZE;
2159 		return;
2160 	}
2161 
2162 	m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr);
2163 
2164 	/* Is the entire record queued? */
2165 	if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) {
2166 		if ((sb->sb_state & SBS_CANTRCVMORE) != 0)
2167 			so->so_error = EMSGSIZE;
2168 		return;
2169 	}
2170 
2171 	sb->sb_flags |= SB_TLS_RX_RUNNING;
2172 
2173 	soref(so);
2174 	wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index];
2175 	mtx_lock(&wq->mtx);
2176 	STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list);
2177 	running = wq->running;
2178 	mtx_unlock(&wq->mtx);
2179 	if (!running)
2180 		wakeup(wq);
2181 	counter_u64_add(ktls_cnt_rx_queued, 1);
2182 }
2183 
2184 static struct mbuf *
ktls_detach_record(struct sockbuf * sb,int len)2185 ktls_detach_record(struct sockbuf *sb, int len)
2186 {
2187 	struct mbuf *m, *n, *top;
2188 	int remain;
2189 
2190 	SOCKBUF_LOCK_ASSERT(sb);
2191 	MPASS(len <= sb->sb_tlscc);
2192 
2193 	/*
2194 	 * If TLS chain is the exact size of the record,
2195 	 * just grab the whole record.
2196 	 */
2197 	top = sb->sb_mtls;
2198 	if (sb->sb_tlscc == len) {
2199 		sb->sb_mtls = NULL;
2200 		sb->sb_mtlstail = NULL;
2201 		goto out;
2202 	}
2203 
2204 	/*
2205 	 * While it would be nice to use m_split() here, we need
2206 	 * to know exactly what m_split() allocates to update the
2207 	 * accounting, so do it inline instead.
2208 	 */
2209 	remain = len;
2210 	for (m = top; remain > m->m_len; m = m->m_next)
2211 		remain -= m->m_len;
2212 
2213 	/* Easy case: don't have to split 'm'. */
2214 	if (remain == m->m_len) {
2215 		sb->sb_mtls = m->m_next;
2216 		if (sb->sb_mtls == NULL)
2217 			sb->sb_mtlstail = NULL;
2218 		m->m_next = NULL;
2219 		goto out;
2220 	}
2221 
2222 	/*
2223 	 * Need to allocate an mbuf to hold the remainder of 'm'.  Try
2224 	 * with M_NOWAIT first.
2225 	 */
2226 	n = m_get(M_NOWAIT, MT_DATA);
2227 	if (n == NULL) {
2228 		/*
2229 		 * Use M_WAITOK with socket buffer unlocked.  If
2230 		 * 'sb_mtls' changes while the lock is dropped, return
2231 		 * NULL to force the caller to retry.
2232 		 */
2233 		SOCKBUF_UNLOCK(sb);
2234 
2235 		n = m_get(M_WAITOK, MT_DATA);
2236 
2237 		SOCKBUF_LOCK(sb);
2238 		if (sb->sb_mtls != top) {
2239 			m_free(n);
2240 			return (NULL);
2241 		}
2242 	}
2243 	n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED));
2244 
2245 	/* Store remainder in 'n'. */
2246 	n->m_len = m->m_len - remain;
2247 	if (m->m_flags & M_EXT) {
2248 		n->m_data = m->m_data + remain;
2249 		mb_dupcl(n, m);
2250 	} else {
2251 		bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len);
2252 	}
2253 
2254 	/* Trim 'm' and update accounting. */
2255 	m->m_len -= n->m_len;
2256 	sb->sb_tlscc -= n->m_len;
2257 	sb->sb_ccc -= n->m_len;
2258 
2259 	/* Account for 'n'. */
2260 	sballoc_ktls_rx(sb, n);
2261 
2262 	/* Insert 'n' into the TLS chain. */
2263 	sb->sb_mtls = n;
2264 	n->m_next = m->m_next;
2265 	if (sb->sb_mtlstail == m)
2266 		sb->sb_mtlstail = n;
2267 
2268 	/* Detach the record from the TLS chain. */
2269 	m->m_next = NULL;
2270 
2271 out:
2272 	MPASS(m_length(top, NULL) == len);
2273 	for (m = top; m != NULL; m = m->m_next)
2274 		sbfree_ktls_rx(sb, m);
2275 	sb->sb_tlsdcc = len;
2276 	sb->sb_ccc += len;
2277 	SBCHECK(sb);
2278 	return (top);
2279 }
2280 
2281 /*
2282  * Determine the length of the trailing zero padding and find the real
2283  * record type in the byte before the padding.
2284  *
2285  * Walking the mbuf chain backwards is clumsy, so another option would
2286  * be to scan forwards remembering the last non-zero byte before the
2287  * trailer.  However, it would be expensive to scan the entire record.
2288  * Instead, find the last non-zero byte of each mbuf in the chain
2289  * keeping track of the relative offset of that nonzero byte.
2290  *
2291  * trail_len is the size of the MAC/tag on input and is set to the
2292  * size of the full trailer including padding and the record type on
2293  * return.
2294  */
2295 static int
tls13_find_record_type(struct ktls_session * tls,struct mbuf * m,int tls_len,int * trailer_len,uint8_t * record_typep)2296 tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len,
2297     int *trailer_len, uint8_t *record_typep)
2298 {
2299 	char *cp;
2300 	u_int digest_start, last_offset, m_len, offset;
2301 	uint8_t record_type;
2302 
2303 	digest_start = tls_len - *trailer_len;
2304 	last_offset = 0;
2305 	offset = 0;
2306 	for (; m != NULL && offset < digest_start;
2307 	     offset += m->m_len, m = m->m_next) {
2308 		/* Don't look for padding in the tag. */
2309 		m_len = min(digest_start - offset, m->m_len);
2310 		cp = mtod(m, char *);
2311 
2312 		/* Find last non-zero byte in this mbuf. */
2313 		while (m_len > 0 && cp[m_len - 1] == 0)
2314 			m_len--;
2315 		if (m_len > 0) {
2316 			record_type = cp[m_len - 1];
2317 			last_offset = offset + m_len;
2318 		}
2319 	}
2320 	if (last_offset < tls->params.tls_hlen)
2321 		return (EBADMSG);
2322 
2323 	*record_typep = record_type;
2324 	*trailer_len = tls_len - last_offset + 1;
2325 	return (0);
2326 }
2327 
2328 /*
2329  * Check if a mbuf chain is fully decrypted at the given offset and
2330  * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is
2331  * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted
2332  * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data
2333  * is encrypted.
2334  */
2335 ktls_mbuf_crypto_st_t
ktls_mbuf_crypto_state(struct mbuf * mb,int offset,int len)2336 ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len)
2337 {
2338 	int m_flags_ored = 0;
2339 	int m_flags_anded = -1;
2340 
2341 	for (; mb != NULL; mb = mb->m_next) {
2342 		if (offset < mb->m_len)
2343 			break;
2344 		offset -= mb->m_len;
2345 	}
2346 	offset += len;
2347 
2348 	for (; mb != NULL; mb = mb->m_next) {
2349 		m_flags_ored |= mb->m_flags;
2350 		m_flags_anded &= mb->m_flags;
2351 
2352 		if (offset <= mb->m_len)
2353 			break;
2354 		offset -= mb->m_len;
2355 	}
2356 	MPASS(mb != NULL || offset == 0);
2357 
2358 	if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED)
2359 		return (KTLS_MBUF_CRYPTO_ST_MIXED);
2360 	else
2361 		return ((m_flags_ored & M_DECRYPTED) ?
2362 		    KTLS_MBUF_CRYPTO_ST_DECRYPTED :
2363 		    KTLS_MBUF_CRYPTO_ST_ENCRYPTED);
2364 }
2365 
2366 /*
2367  * ktls_resync_ifnet - get HW TLS RX back on track after packet loss
2368  */
2369 static int
ktls_resync_ifnet(struct socket * so,uint32_t tls_len,uint64_t tls_rcd_num)2370 ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num)
2371 {
2372 	union if_snd_tag_modify_params params;
2373 	struct m_snd_tag *mst;
2374 	struct inpcb *inp;
2375 	struct tcpcb *tp;
2376 
2377 	mst = so->so_rcv.sb_tls_info->snd_tag;
2378 	if (__predict_false(mst == NULL))
2379 		return (EINVAL);
2380 
2381 	inp = sotoinpcb(so);
2382 	if (__predict_false(inp == NULL))
2383 		return (EINVAL);
2384 
2385 	INP_RLOCK(inp);
2386 	if (inp->inp_flags & INP_DROPPED) {
2387 		INP_RUNLOCK(inp);
2388 		return (ECONNRESET);
2389 	}
2390 
2391 	tp = intotcpcb(inp);
2392 	MPASS(tp != NULL);
2393 
2394 	/* Get the TCP sequence number of the next valid TLS header. */
2395 	SOCKBUF_LOCK(&so->so_rcv);
2396 	params.tls_rx.tls_hdr_tcp_sn =
2397 	    tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len;
2398 	params.tls_rx.tls_rec_length = tls_len;
2399 	params.tls_rx.tls_seq_number = tls_rcd_num;
2400 	SOCKBUF_UNLOCK(&so->so_rcv);
2401 
2402 	INP_RUNLOCK(inp);
2403 
2404 	MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX);
2405 	return (mst->sw->snd_tag_modify(mst, &params));
2406 }
2407 
2408 static void
ktls_drop(struct socket * so,int error)2409 ktls_drop(struct socket *so, int error)
2410 {
2411 	struct epoch_tracker et;
2412 	struct inpcb *inp = sotoinpcb(so);
2413 	struct tcpcb *tp;
2414 
2415 	NET_EPOCH_ENTER(et);
2416 	INP_WLOCK(inp);
2417 	if (!(inp->inp_flags & INP_DROPPED)) {
2418 		tp = intotcpcb(inp);
2419 		CURVNET_SET(inp->inp_vnet);
2420 		tp = tcp_drop(tp, error);
2421 		CURVNET_RESTORE();
2422 		if (tp != NULL)
2423 			INP_WUNLOCK(inp);
2424 	} else {
2425 		so->so_error = error;
2426 		SOCK_RECVBUF_LOCK(so);
2427 		sorwakeup_locked(so);
2428 		INP_WUNLOCK(inp);
2429 	}
2430 	NET_EPOCH_EXIT(et);
2431 }
2432 
2433 static void
ktls_decrypt(struct socket * so)2434 ktls_decrypt(struct socket *so)
2435 {
2436 	char tls_header[MBUF_PEXT_HDR_LEN];
2437 	struct ktls_session *tls;
2438 	struct sockbuf *sb;
2439 	struct tls_record_layer *hdr;
2440 	struct tls_get_record tgr;
2441 	struct mbuf *control, *data, *m;
2442 	ktls_mbuf_crypto_st_t state;
2443 	uint64_t seqno;
2444 	int error, remain, tls_len, trail_len;
2445 	bool tls13;
2446 	uint8_t vminor, record_type;
2447 
2448 	hdr = (struct tls_record_layer *)tls_header;
2449 	sb = &so->so_rcv;
2450 	SOCKBUF_LOCK(sb);
2451 	KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING,
2452 	    ("%s: socket %p not running", __func__, so));
2453 
2454 	tls = sb->sb_tls_info;
2455 	MPASS(tls != NULL);
2456 
2457 	tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE);
2458 	if (tls13)
2459 		vminor = TLS_MINOR_VER_TWO;
2460 	else
2461 		vminor = tls->params.tls_vminor;
2462 	for (;;) {
2463 		/* Is there enough queued for a TLS header? */
2464 		if (sb->sb_tlscc < tls->params.tls_hlen)
2465 			break;
2466 
2467 		m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header);
2468 		tls_len = sizeof(*hdr) + ntohs(hdr->tls_length);
2469 
2470 		if (hdr->tls_vmajor != tls->params.tls_vmajor ||
2471 		    hdr->tls_vminor != vminor)
2472 			error = EINVAL;
2473 		else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP)
2474 			error = EINVAL;
2475 		else if (tls_len < tls->params.tls_hlen || tls_len >
2476 		    tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 +
2477 		    tls->params.tls_tlen)
2478 			error = EMSGSIZE;
2479 		else
2480 			error = 0;
2481 		if (__predict_false(error != 0)) {
2482 			/*
2483 			 * We have a corrupted record and are likely
2484 			 * out of sync.  The connection isn't
2485 			 * recoverable at this point, so abort it.
2486 			 */
2487 			SOCKBUF_UNLOCK(sb);
2488 			counter_u64_add(ktls_offload_corrupted_records, 1);
2489 
2490 			ktls_drop(so, error);
2491 			goto deref;
2492 		}
2493 
2494 		/* Is the entire record queued? */
2495 		if (sb->sb_tlscc < tls_len)
2496 			break;
2497 
2498 		/*
2499 		 * Split out the portion of the mbuf chain containing
2500 		 * this TLS record.
2501 		 */
2502 		data = ktls_detach_record(sb, tls_len);
2503 		if (data == NULL)
2504 			continue;
2505 		MPASS(sb->sb_tlsdcc == tls_len);
2506 
2507 		seqno = sb->sb_tls_seqno;
2508 		sb->sb_tls_seqno++;
2509 		SBCHECK(sb);
2510 		SOCKBUF_UNLOCK(sb);
2511 
2512 		/* get crypto state for this TLS record */
2513 		state = ktls_mbuf_crypto_state(data, 0, tls_len);
2514 
2515 		switch (state) {
2516 		case KTLS_MBUF_CRYPTO_ST_MIXED:
2517 			error = ktls_ocf_recrypt(tls, hdr, data, seqno);
2518 			if (error)
2519 				break;
2520 			/* FALLTHROUGH */
2521 		case KTLS_MBUF_CRYPTO_ST_ENCRYPTED:
2522 			error = ktls_ocf_decrypt(tls, hdr, data, seqno,
2523 			    &trail_len);
2524 			if (__predict_true(error == 0)) {
2525 				if (tls13) {
2526 					error = tls13_find_record_type(tls, data,
2527 					    tls_len, &trail_len, &record_type);
2528 				} else {
2529 					record_type = hdr->tls_type;
2530 				}
2531 			}
2532 			break;
2533 		case KTLS_MBUF_CRYPTO_ST_DECRYPTED:
2534 			/*
2535 			 * NIC TLS is only supported for AEAD
2536 			 * ciphersuites which used a fixed sized
2537 			 * trailer.
2538 			 */
2539 			if (tls13) {
2540 				trail_len = tls->params.tls_tlen - 1;
2541 				error = tls13_find_record_type(tls, data,
2542 				    tls_len, &trail_len, &record_type);
2543 			} else {
2544 				trail_len = tls->params.tls_tlen;
2545 				error = 0;
2546 				record_type = hdr->tls_type;
2547 			}
2548 			break;
2549 		default:
2550 			error = EINVAL;
2551 			break;
2552 		}
2553 		if (error) {
2554 			counter_u64_add(ktls_offload_failed_crypto, 1);
2555 
2556 			SOCKBUF_LOCK(sb);
2557 			if (sb->sb_tlsdcc == 0) {
2558 				/*
2559 				 * sbcut/drop/flush discarded these
2560 				 * mbufs.
2561 				 */
2562 				m_freem(data);
2563 				break;
2564 			}
2565 
2566 			/*
2567 			 * Drop this TLS record's data, but keep
2568 			 * decrypting subsequent records.
2569 			 */
2570 			sb->sb_ccc -= tls_len;
2571 			sb->sb_tlsdcc = 0;
2572 
2573 			if (error != EMSGSIZE)
2574 				error = EBADMSG;
2575 			CURVNET_SET(so->so_vnet);
2576 			so->so_error = error;
2577 			sorwakeup_locked(so);
2578 			CURVNET_RESTORE();
2579 
2580 			m_freem(data);
2581 
2582 			SOCKBUF_LOCK(sb);
2583 			continue;
2584 		}
2585 
2586 		/* Allocate the control mbuf. */
2587 		memset(&tgr, 0, sizeof(tgr));
2588 		tgr.tls_type = record_type;
2589 		tgr.tls_vmajor = hdr->tls_vmajor;
2590 		tgr.tls_vminor = hdr->tls_vminor;
2591 		tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen -
2592 		    trail_len);
2593 		control = sbcreatecontrol(&tgr, sizeof(tgr),
2594 		    TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK);
2595 
2596 		SOCKBUF_LOCK(sb);
2597 		if (sb->sb_tlsdcc == 0) {
2598 			/* sbcut/drop/flush discarded these mbufs. */
2599 			MPASS(sb->sb_tlscc == 0);
2600 			m_freem(data);
2601 			m_freem(control);
2602 			break;
2603 		}
2604 
2605 		/*
2606 		 * Clear the 'dcc' accounting in preparation for
2607 		 * adding the decrypted record.
2608 		 */
2609 		sb->sb_ccc -= tls_len;
2610 		sb->sb_tlsdcc = 0;
2611 		SBCHECK(sb);
2612 
2613 		/* If there is no payload, drop all of the data. */
2614 		if (tgr.tls_length == htobe16(0)) {
2615 			m_freem(data);
2616 			data = NULL;
2617 		} else {
2618 			/* Trim header. */
2619 			remain = tls->params.tls_hlen;
2620 			while (remain > 0) {
2621 				if (data->m_len > remain) {
2622 					data->m_data += remain;
2623 					data->m_len -= remain;
2624 					break;
2625 				}
2626 				remain -= data->m_len;
2627 				data = m_free(data);
2628 			}
2629 
2630 			/* Trim trailer and clear M_NOTREADY. */
2631 			remain = be16toh(tgr.tls_length);
2632 			m = data;
2633 			for (m = data; remain > m->m_len; m = m->m_next) {
2634 				m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2635 				remain -= m->m_len;
2636 			}
2637 			m->m_len = remain;
2638 			m_freem(m->m_next);
2639 			m->m_next = NULL;
2640 			m->m_flags &= ~(M_NOTREADY | M_DECRYPTED);
2641 
2642 			/* Set EOR on the final mbuf. */
2643 			m->m_flags |= M_EOR;
2644 		}
2645 
2646 		sbappendcontrol_locked(sb, data, control, 0);
2647 
2648 		if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) {
2649 			sb->sb_flags |= SB_TLS_RX_RESYNC;
2650 			SOCKBUF_UNLOCK(sb);
2651 			ktls_resync_ifnet(so, tls_len, seqno);
2652 			SOCKBUF_LOCK(sb);
2653 		} else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) {
2654 			sb->sb_flags &= ~SB_TLS_RX_RESYNC;
2655 			SOCKBUF_UNLOCK(sb);
2656 			ktls_resync_ifnet(so, 0, seqno);
2657 			SOCKBUF_LOCK(sb);
2658 		}
2659 	}
2660 
2661 	sb->sb_flags &= ~SB_TLS_RX_RUNNING;
2662 
2663 	if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0)
2664 		so->so_error = EMSGSIZE;
2665 
2666 	sorwakeup_locked(so);
2667 
2668 deref:
2669 	SOCKBUF_UNLOCK_ASSERT(sb);
2670 
2671 	CURVNET_SET(so->so_vnet);
2672 	sorele(so);
2673 	CURVNET_RESTORE();
2674 }
2675 
2676 void
ktls_enqueue_to_free(struct mbuf * m)2677 ktls_enqueue_to_free(struct mbuf *m)
2678 {
2679 	struct ktls_wq *wq;
2680 	bool running;
2681 
2682 	/* Mark it for freeing. */
2683 	m->m_epg_flags |= EPG_FLAG_2FREE;
2684 	wq = &ktls_wq[m->m_epg_tls->wq_index];
2685 	mtx_lock(&wq->mtx);
2686 	STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2687 	running = wq->running;
2688 	mtx_unlock(&wq->mtx);
2689 	if (!running)
2690 		wakeup(wq);
2691 }
2692 
2693 static void *
ktls_buffer_alloc(struct ktls_wq * wq,struct mbuf * m)2694 ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m)
2695 {
2696 	void *buf;
2697 	int domain, running;
2698 
2699 	if (m->m_epg_npgs <= 2)
2700 		return (NULL);
2701 	if (ktls_buffer_zone == NULL)
2702 		return (NULL);
2703 	if ((u_int)(ticks - wq->lastallocfail) < hz) {
2704 		/*
2705 		 * Rate-limit allocation attempts after a failure.
2706 		 * ktls_buffer_import() will acquire a per-domain mutex to check
2707 		 * the free page queues and may fail consistently if memory is
2708 		 * fragmented.
2709 		 */
2710 		return (NULL);
2711 	}
2712 	buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM);
2713 	if (buf == NULL) {
2714 		domain = PCPU_GET(domain);
2715 		wq->lastallocfail = ticks;
2716 
2717 		/*
2718 		 * Note that this check is "racy", but the races are
2719 		 * harmless, and are either a spurious wakeup if
2720 		 * multiple threads fail allocations before the alloc
2721 		 * thread wakes, or waiting an extra second in case we
2722 		 * see an old value of running == true.
2723 		 */
2724 		if (!VM_DOMAIN_EMPTY(domain)) {
2725 			running = atomic_load_int(&ktls_domains[domain].reclaim_td.running);
2726 			if (!running)
2727 				wakeup(&ktls_domains[domain].reclaim_td);
2728 		}
2729 	}
2730 	return (buf);
2731 }
2732 
2733 static int
ktls_encrypt_record(struct ktls_wq * wq,struct mbuf * m,struct ktls_session * tls,struct ktls_ocf_encrypt_state * state)2734 ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m,
2735     struct ktls_session *tls, struct ktls_ocf_encrypt_state *state)
2736 {
2737 	vm_page_t pg;
2738 	int error, i, len, off;
2739 
2740 	KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY),
2741 	    ("%p not unready & nomap mbuf\n", m));
2742 	KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen,
2743 	    ("page count %d larger than maximum frame length %d", m->m_epg_npgs,
2744 	    ktls_maxlen));
2745 
2746 	/* Anonymous mbufs are encrypted in place. */
2747 	if ((m->m_epg_flags & EPG_FLAG_ANON) != 0)
2748 		return (ktls_ocf_encrypt(state, tls, m, NULL, 0));
2749 
2750 	/*
2751 	 * For file-backed mbufs (from sendfile), anonymous wired
2752 	 * pages are allocated and used as the encryption destination.
2753 	 */
2754 	if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) {
2755 		len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len -
2756 		    m->m_epg_1st_off;
2757 		state->dst_iov[0].iov_base = (char *)state->cbuf +
2758 		    m->m_epg_1st_off;
2759 		state->dst_iov[0].iov_len = len;
2760 		state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf);
2761 		i = 1;
2762 	} else {
2763 		off = m->m_epg_1st_off;
2764 		for (i = 0; i < m->m_epg_npgs; i++, off = 0) {
2765 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
2766 			    VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2767 			len = m_epg_pagelen(m, i, off);
2768 			state->parray[i] = VM_PAGE_TO_PHYS(pg);
2769 			state->dst_iov[i].iov_base =
2770 			    (char *)PHYS_TO_DMAP(state->parray[i]) + off;
2771 			state->dst_iov[i].iov_len = len;
2772 		}
2773 	}
2774 	KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small"));
2775 	state->dst_iov[i].iov_base = m->m_epg_trail;
2776 	state->dst_iov[i].iov_len = m->m_epg_trllen;
2777 
2778 	error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1);
2779 
2780 	if (__predict_false(error != 0)) {
2781 		/* Free the anonymous pages. */
2782 		if (state->cbuf != NULL)
2783 			uma_zfree(ktls_buffer_zone, state->cbuf);
2784 		else {
2785 			for (i = 0; i < m->m_epg_npgs; i++) {
2786 				pg = PHYS_TO_VM_PAGE(state->parray[i]);
2787 				(void)vm_page_unwire_noq(pg);
2788 				vm_page_free(pg);
2789 			}
2790 		}
2791 	}
2792 	return (error);
2793 }
2794 
2795 /* Number of TLS records in a batch passed to ktls_enqueue(). */
2796 static u_int
ktls_batched_records(struct mbuf * m)2797 ktls_batched_records(struct mbuf *m)
2798 {
2799 	int page_count, records;
2800 
2801 	records = 0;
2802 	page_count = m->m_epg_enc_cnt;
2803 	while (page_count > 0) {
2804 		records++;
2805 		page_count -= m->m_epg_nrdy;
2806 		m = m->m_next;
2807 	}
2808 	KASSERT(page_count == 0, ("%s: mismatched page count", __func__));
2809 	return (records);
2810 }
2811 
2812 void
ktls_enqueue(struct mbuf * m,struct socket * so,int page_count)2813 ktls_enqueue(struct mbuf *m, struct socket *so, int page_count)
2814 {
2815 	struct ktls_session *tls;
2816 	struct ktls_wq *wq;
2817 	int queued;
2818 	bool running;
2819 
2820 	KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) ==
2821 	    (M_EXTPG | M_NOTREADY)),
2822 	    ("ktls_enqueue: %p not unready & nomap mbuf\n", m));
2823 	KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count"));
2824 
2825 	KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf"));
2826 
2827 	m->m_epg_enc_cnt = page_count;
2828 
2829 	/*
2830 	 * Save a pointer to the socket.  The caller is responsible
2831 	 * for taking an additional reference via soref().
2832 	 */
2833 	m->m_epg_so = so;
2834 
2835 	queued = 1;
2836 	tls = m->m_epg_tls;
2837 	wq = &ktls_wq[tls->wq_index];
2838 	mtx_lock(&wq->mtx);
2839 	if (__predict_false(tls->sequential_records)) {
2840 		/*
2841 		 * For TLS 1.0, records must be encrypted
2842 		 * sequentially.  For a given connection, all records
2843 		 * queued to the associated work queue are processed
2844 		 * sequentially.  However, sendfile(2) might complete
2845 		 * I/O requests spanning multiple TLS records out of
2846 		 * order.  Here we ensure TLS records are enqueued to
2847 		 * the work queue in FIFO order.
2848 		 *
2849 		 * tls->next_seqno holds the sequence number of the
2850 		 * next TLS record that should be enqueued to the work
2851 		 * queue.  If this next record is not tls->next_seqno,
2852 		 * it must be a future record, so insert it, sorted by
2853 		 * TLS sequence number, into tls->pending_records and
2854 		 * return.
2855 		 *
2856 		 * If this TLS record matches tls->next_seqno, place
2857 		 * it in the work queue and then check
2858 		 * tls->pending_records to see if any
2859 		 * previously-queued records are now ready for
2860 		 * encryption.
2861 		 */
2862 		if (m->m_epg_seqno != tls->next_seqno) {
2863 			struct mbuf *n, *p;
2864 
2865 			p = NULL;
2866 			STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) {
2867 				if (n->m_epg_seqno > m->m_epg_seqno)
2868 					break;
2869 				p = n;
2870 			}
2871 			if (n == NULL)
2872 				STAILQ_INSERT_TAIL(&tls->pending_records, m,
2873 				    m_epg_stailq);
2874 			else if (p == NULL)
2875 				STAILQ_INSERT_HEAD(&tls->pending_records, m,
2876 				    m_epg_stailq);
2877 			else
2878 				STAILQ_INSERT_AFTER(&tls->pending_records, p, m,
2879 				    m_epg_stailq);
2880 			mtx_unlock(&wq->mtx);
2881 			counter_u64_add(ktls_cnt_tx_pending, 1);
2882 			return;
2883 		}
2884 
2885 		tls->next_seqno += ktls_batched_records(m);
2886 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2887 
2888 		while (!STAILQ_EMPTY(&tls->pending_records)) {
2889 			struct mbuf *n;
2890 
2891 			n = STAILQ_FIRST(&tls->pending_records);
2892 			if (n->m_epg_seqno != tls->next_seqno)
2893 				break;
2894 
2895 			queued++;
2896 			STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq);
2897 			tls->next_seqno += ktls_batched_records(n);
2898 			STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq);
2899 		}
2900 		counter_u64_add(ktls_cnt_tx_pending, -(queued - 1));
2901 	} else
2902 		STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq);
2903 
2904 	running = wq->running;
2905 	mtx_unlock(&wq->mtx);
2906 	if (!running)
2907 		wakeup(wq);
2908 	counter_u64_add(ktls_cnt_tx_queued, queued);
2909 }
2910 
2911 /*
2912  * Once a file-backed mbuf (from sendfile) has been encrypted, free
2913  * the pages from the file and replace them with the anonymous pages
2914  * allocated in ktls_encrypt_record().
2915  */
2916 static void
ktls_finish_nonanon(struct mbuf * m,struct ktls_ocf_encrypt_state * state)2917 ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state)
2918 {
2919 	int i;
2920 
2921 	MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0);
2922 
2923 	/* Free the old pages. */
2924 	m->m_ext.ext_free(m);
2925 
2926 	/* Replace them with the new pages. */
2927 	if (state->cbuf != NULL) {
2928 		for (i = 0; i < m->m_epg_npgs; i++)
2929 			m->m_epg_pa[i] = state->parray[0] + ptoa(i);
2930 
2931 		/* Contig pages should go back to the cache. */
2932 		m->m_ext.ext_free = ktls_free_mext_contig;
2933 	} else {
2934 		for (i = 0; i < m->m_epg_npgs; i++)
2935 			m->m_epg_pa[i] = state->parray[i];
2936 
2937 		/* Use the basic free routine. */
2938 		m->m_ext.ext_free = mb_free_mext_pgs;
2939 	}
2940 
2941 	/* Pages are now writable. */
2942 	m->m_epg_flags |= EPG_FLAG_ANON;
2943 }
2944 
2945 static __noinline void
ktls_encrypt(struct ktls_wq * wq,struct mbuf * top)2946 ktls_encrypt(struct ktls_wq *wq, struct mbuf *top)
2947 {
2948 	struct ktls_ocf_encrypt_state state;
2949 	struct ktls_session *tls;
2950 	struct socket *so;
2951 	struct mbuf *m;
2952 	int error, npages, total_pages;
2953 
2954 	so = top->m_epg_so;
2955 	tls = top->m_epg_tls;
2956 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
2957 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
2958 #ifdef INVARIANTS
2959 	top->m_epg_so = NULL;
2960 #endif
2961 	total_pages = top->m_epg_enc_cnt;
2962 	npages = 0;
2963 
2964 	/*
2965 	 * Encrypt the TLS records in the chain of mbufs starting with
2966 	 * 'top'.  'total_pages' gives us a total count of pages and is
2967 	 * used to know when we have finished encrypting the TLS
2968 	 * records originally queued with 'top'.
2969 	 *
2970 	 * NB: These mbufs are queued in the socket buffer and
2971 	 * 'm_next' is traversing the mbufs in the socket buffer.  The
2972 	 * socket buffer lock is not held while traversing this chain.
2973 	 * Since the mbufs are all marked M_NOTREADY their 'm_next'
2974 	 * pointers should be stable.  However, the 'm_next' of the
2975 	 * last mbuf encrypted is not necessarily NULL.  It can point
2976 	 * to other mbufs appended while 'top' was on the TLS work
2977 	 * queue.
2978 	 *
2979 	 * Each mbuf holds an entire TLS record.
2980 	 */
2981 	error = 0;
2982 	for (m = top; npages != total_pages; m = m->m_next) {
2983 		KASSERT(m->m_epg_tls == tls,
2984 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
2985 		    tls, m->m_epg_tls));
2986 		KASSERT(npages + m->m_epg_npgs <= total_pages,
2987 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
2988 		    total_pages, m));
2989 
2990 		error = ktls_encrypt_record(wq, m, tls, &state);
2991 		if (error) {
2992 			counter_u64_add(ktls_offload_failed_crypto, 1);
2993 			break;
2994 		}
2995 
2996 		if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
2997 			ktls_finish_nonanon(m, &state);
2998 
2999 		npages += m->m_epg_nrdy;
3000 
3001 		/*
3002 		 * Drop a reference to the session now that it is no
3003 		 * longer needed.  Existing code depends on encrypted
3004 		 * records having no associated session vs
3005 		 * yet-to-be-encrypted records having an associated
3006 		 * session.
3007 		 */
3008 		m->m_epg_tls = NULL;
3009 		ktls_free(tls);
3010 	}
3011 
3012 	CURVNET_SET(so->so_vnet);
3013 	if (error == 0) {
3014 		(void)so->so_proto->pr_ready(so, top, npages);
3015 	} else {
3016 		ktls_drop(so, EIO);
3017 		mb_free_notready(top, total_pages);
3018 	}
3019 
3020 	sorele(so);
3021 	CURVNET_RESTORE();
3022 }
3023 
3024 void
ktls_encrypt_cb(struct ktls_ocf_encrypt_state * state,int error)3025 ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error)
3026 {
3027 	struct ktls_session *tls;
3028 	struct socket *so;
3029 	struct mbuf *m;
3030 	int npages;
3031 
3032 	m = state->m;
3033 
3034 	if ((m->m_epg_flags & EPG_FLAG_ANON) == 0)
3035 		ktls_finish_nonanon(m, state);
3036 
3037 	so = state->so;
3038 	free(state, M_KTLS);
3039 
3040 	/*
3041 	 * Drop a reference to the session now that it is no longer
3042 	 * needed.  Existing code depends on encrypted records having
3043 	 * no associated session vs yet-to-be-encrypted records having
3044 	 * an associated session.
3045 	 */
3046 	tls = m->m_epg_tls;
3047 	m->m_epg_tls = NULL;
3048 	ktls_free(tls);
3049 
3050 	if (error != 0)
3051 		counter_u64_add(ktls_offload_failed_crypto, 1);
3052 
3053 	CURVNET_SET(so->so_vnet);
3054 	npages = m->m_epg_nrdy;
3055 
3056 	if (error == 0) {
3057 		(void)so->so_proto->pr_ready(so, m, npages);
3058 	} else {
3059 		ktls_drop(so, EIO);
3060 		mb_free_notready(m, npages);
3061 	}
3062 
3063 	sorele(so);
3064 	CURVNET_RESTORE();
3065 }
3066 
3067 /*
3068  * Similar to ktls_encrypt, but used with asynchronous OCF backends
3069  * (coprocessors) where encryption does not use host CPU resources and
3070  * it can be beneficial to queue more requests than CPUs.
3071  */
3072 static __noinline void
ktls_encrypt_async(struct ktls_wq * wq,struct mbuf * top)3073 ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top)
3074 {
3075 	struct ktls_ocf_encrypt_state *state;
3076 	struct ktls_session *tls;
3077 	struct socket *so;
3078 	struct mbuf *m, *n;
3079 	int error, mpages, npages, total_pages;
3080 
3081 	so = top->m_epg_so;
3082 	tls = top->m_epg_tls;
3083 	KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top));
3084 	KASSERT(so != NULL, ("so = NULL, top = %p\n", top));
3085 #ifdef INVARIANTS
3086 	top->m_epg_so = NULL;
3087 #endif
3088 	total_pages = top->m_epg_enc_cnt;
3089 	npages = 0;
3090 
3091 	error = 0;
3092 	for (m = top; npages != total_pages; m = n) {
3093 		KASSERT(m->m_epg_tls == tls,
3094 		    ("different TLS sessions in a single mbuf chain: %p vs %p",
3095 		    tls, m->m_epg_tls));
3096 		KASSERT(npages + m->m_epg_npgs <= total_pages,
3097 		    ("page count mismatch: top %p, total_pages %d, m %p", top,
3098 		    total_pages, m));
3099 
3100 		state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO);
3101 		soref(so);
3102 		state->so = so;
3103 		state->m = m;
3104 
3105 		mpages = m->m_epg_nrdy;
3106 		n = m->m_next;
3107 
3108 		error = ktls_encrypt_record(wq, m, tls, state);
3109 		if (error) {
3110 			counter_u64_add(ktls_offload_failed_crypto, 1);
3111 			free(state, M_KTLS);
3112 			CURVNET_SET(so->so_vnet);
3113 			sorele(so);
3114 			CURVNET_RESTORE();
3115 			break;
3116 		}
3117 
3118 		npages += mpages;
3119 	}
3120 
3121 	CURVNET_SET(so->so_vnet);
3122 	if (error != 0) {
3123 		ktls_drop(so, EIO);
3124 		mb_free_notready(m, total_pages - npages);
3125 	}
3126 
3127 	sorele(so);
3128 	CURVNET_RESTORE();
3129 }
3130 
3131 static int
ktls_bind_domain(int domain)3132 ktls_bind_domain(int domain)
3133 {
3134 	int error;
3135 
3136 	error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]);
3137 	if (error != 0)
3138 		return (error);
3139 	curthread->td_domain.dr_policy = DOMAINSET_PREF(domain);
3140 	return (0);
3141 }
3142 
3143 static void
ktls_reclaim_thread(void * ctx)3144 ktls_reclaim_thread(void *ctx)
3145 {
3146 	struct ktls_domain_info *ktls_domain = ctx;
3147 	struct ktls_reclaim_thread *sc = &ktls_domain->reclaim_td;
3148 	struct sysctl_oid *oid;
3149 	char name[80];
3150 	int error, domain;
3151 
3152 	domain = ktls_domain - ktls_domains;
3153 	if (bootverbose)
3154 		printf("Starting KTLS reclaim thread for domain %d\n", domain);
3155 	error = ktls_bind_domain(domain);
3156 	if (error)
3157 		printf("Unable to bind KTLS reclaim thread for domain %d: error %d\n",
3158 		    domain, error);
3159 	snprintf(name, sizeof(name), "domain%d", domain);
3160 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO,
3161 	    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
3162 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "reclaims",
3163 	    CTLFLAG_RD,  &sc->reclaims, 0, "buffers reclaimed");
3164 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups",
3165 	    CTLFLAG_RD,  &sc->wakeups, 0, "thread wakeups");
3166 	SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running",
3167 	    CTLFLAG_RD,  &sc->running, 0, "thread running");
3168 
3169 	for (;;) {
3170 		atomic_store_int(&sc->running, 0);
3171 		tsleep(sc, PZERO | PNOLOCK, "-",  0);
3172 		atomic_store_int(&sc->running, 1);
3173 		sc->wakeups++;
3174 		/*
3175 		 * Below we attempt to reclaim ktls_max_reclaim
3176 		 * buffers using vm_page_reclaim_contig_domain_ext().
3177 		 * We do this here, as this function can take several
3178 		 * seconds to scan all of memory and it does not
3179 		 * matter if this thread pauses for a while.  If we
3180 		 * block a ktls worker thread, we risk developing
3181 		 * backlogs of buffers to be encrypted, leading to
3182 		 * surges of traffic and potential NIC output drops.
3183 		 */
3184 		if (!vm_page_reclaim_contig_domain_ext(domain, VM_ALLOC_NORMAL,
3185 		    atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, ktls_max_reclaim)) {
3186 			vm_wait_domain(domain);
3187 		} else {
3188 			sc->reclaims += ktls_max_reclaim;
3189 		}
3190 	}
3191 }
3192 
3193 static void
ktls_work_thread(void * ctx)3194 ktls_work_thread(void *ctx)
3195 {
3196 	struct ktls_wq *wq = ctx;
3197 	struct mbuf *m, *n;
3198 	struct socket *so, *son;
3199 	STAILQ_HEAD(, mbuf) local_m_head;
3200 	STAILQ_HEAD(, socket) local_so_head;
3201 	int cpu;
3202 
3203 	cpu = wq - ktls_wq;
3204 	if (bootverbose)
3205 		printf("Starting KTLS worker thread for CPU %d\n", cpu);
3206 
3207 	/*
3208 	 * Bind to a core.  If ktls_bind_threads is > 1, then
3209 	 * we bind to the NUMA domain instead.
3210 	 */
3211 	if (ktls_bind_threads) {
3212 		int error;
3213 
3214 		if (ktls_bind_threads > 1) {
3215 			struct pcpu *pc = pcpu_find(cpu);
3216 
3217 			error = ktls_bind_domain(pc->pc_domain);
3218 		} else {
3219 			cpuset_t mask;
3220 
3221 			CPU_SETOF(cpu, &mask);
3222 			error = cpuset_setthread(curthread->td_tid, &mask);
3223 		}
3224 		if (error)
3225 			printf("Unable to bind KTLS worker thread for CPU %d: error %d\n",
3226 				cpu, error);
3227 	}
3228 #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
3229 	fpu_kern_thread(0);
3230 #endif
3231 	for (;;) {
3232 		mtx_lock(&wq->mtx);
3233 		while (STAILQ_EMPTY(&wq->m_head) &&
3234 		    STAILQ_EMPTY(&wq->so_head)) {
3235 			wq->running = false;
3236 			mtx_sleep(wq, &wq->mtx, 0, "-", 0);
3237 			wq->running = true;
3238 		}
3239 
3240 		STAILQ_INIT(&local_m_head);
3241 		STAILQ_CONCAT(&local_m_head, &wq->m_head);
3242 		STAILQ_INIT(&local_so_head);
3243 		STAILQ_CONCAT(&local_so_head, &wq->so_head);
3244 		mtx_unlock(&wq->mtx);
3245 
3246 		STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) {
3247 			if (m->m_epg_flags & EPG_FLAG_2FREE) {
3248 				ktls_free(m->m_epg_tls);
3249 				m_free_raw(m);
3250 			} else {
3251 				if (m->m_epg_tls->sync_dispatch)
3252 					ktls_encrypt(wq, m);
3253 				else
3254 					ktls_encrypt_async(wq, m);
3255 				counter_u64_add(ktls_cnt_tx_queued, -1);
3256 			}
3257 		}
3258 
3259 		STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) {
3260 			ktls_decrypt(so);
3261 			counter_u64_add(ktls_cnt_rx_queued, -1);
3262 		}
3263 	}
3264 }
3265 
3266 static void
ktls_disable_ifnet_help(void * context,int pending __unused)3267 ktls_disable_ifnet_help(void *context, int pending __unused)
3268 {
3269 	struct ktls_session *tls;
3270 	struct inpcb *inp;
3271 	struct tcpcb *tp;
3272 	struct socket *so;
3273 	int err;
3274 
3275 	tls = context;
3276 	inp = tls->inp;
3277 	if (inp == NULL)
3278 		return;
3279 	INP_WLOCK(inp);
3280 	so = inp->inp_socket;
3281 	MPASS(so != NULL);
3282 	if (inp->inp_flags & INP_DROPPED) {
3283 		goto out;
3284 	}
3285 
3286 	if (so->so_snd.sb_tls_info != NULL)
3287 		err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW);
3288 	else
3289 		err = ENXIO;
3290 	if (err == 0) {
3291 		counter_u64_add(ktls_ifnet_disable_ok, 1);
3292 		/* ktls_set_tx_mode() drops inp wlock, so recheck flags */
3293 		if ((inp->inp_flags & INP_DROPPED) == 0 &&
3294 		    (tp = intotcpcb(inp)) != NULL &&
3295 		    tp->t_fb->tfb_hwtls_change != NULL)
3296 			(*tp->t_fb->tfb_hwtls_change)(tp, 0);
3297 	} else {
3298 		counter_u64_add(ktls_ifnet_disable_fail, 1);
3299 	}
3300 
3301 out:
3302 	CURVNET_SET(so->so_vnet);
3303 	sorele(so);
3304 	CURVNET_RESTORE();
3305 	INP_WUNLOCK(inp);
3306 	ktls_free(tls);
3307 }
3308 
3309 /*
3310  * Called when re-transmits are becoming a substantial portion of the
3311  * sends on this connection.  When this happens, we transition the
3312  * connection to software TLS.  This is needed because most inline TLS
3313  * NICs keep crypto state only for in-order transmits.  This means
3314  * that to handle a TCP rexmit (which is out-of-order), the NIC must
3315  * re-DMA the entire TLS record up to and including the current
3316  * segment.  This means that when re-transmitting the last ~1448 byte
3317  * segment of a 16KB TLS record, we could wind up re-DMA'ing an order
3318  * of magnitude more data than we are sending.  This can cause the
3319  * PCIe link to saturate well before the network, which can cause
3320  * output drops, and a general loss of capacity.
3321  */
3322 void
ktls_disable_ifnet(void * arg)3323 ktls_disable_ifnet(void *arg)
3324 {
3325 	struct tcpcb *tp;
3326 	struct inpcb *inp;
3327 	struct socket *so;
3328 	struct ktls_session *tls;
3329 
3330 	tp = arg;
3331 	inp = tptoinpcb(tp);
3332 	INP_WLOCK_ASSERT(inp);
3333 	so = inp->inp_socket;
3334 	SOCK_LOCK(so);
3335 	tls = so->so_snd.sb_tls_info;
3336 	if (tp->t_nic_ktls_xmit_dis == 1) {
3337 		SOCK_UNLOCK(so);
3338 		return;
3339 	}
3340 
3341 	/*
3342 	 * note that t_nic_ktls_xmit_dis is never cleared; disabling
3343 	 * ifnet can only be done once per connection, so we never want
3344 	 * to do it again
3345 	 */
3346 
3347 	(void)ktls_hold(tls);
3348 	soref(so);
3349 	tp->t_nic_ktls_xmit_dis = 1;
3350 	SOCK_UNLOCK(so);
3351 	TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls);
3352 	(void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task);
3353 }
3354