xref: /openbmc/linux/net/tls/tls_device_fallback.c (revision 801b27e8)
1 /* Copyright (c) 2018, Mellanox Technologies All rights reserved.
2  *
3  * This software is available to you under a choice of one of two
4  * licenses.  You may choose to be licensed under the terms of the GNU
5  * General Public License (GPL) Version 2, available from the file
6  * COPYING in the main directory of this source tree, or the
7  * OpenIB.org BSD license below:
8  *
9  *     Redistribution and use in source and binary forms, with or
10  *     without modification, are permitted provided that the following
11  *     conditions are met:
12  *
13  *      - Redistributions of source code must retain the above
14  *        copyright notice, this list of conditions and the following
15  *        disclaimer.
16  *
17  *      - Redistributions in binary form must reproduce the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer in the documentation and/or other materials
20  *        provided with the distribution.
21  *
22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
23  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
24  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
25  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
26  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
27  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
28  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
29  * SOFTWARE.
30  */
31 
32 #include <net/tls.h>
33 #include <crypto/aead.h>
34 #include <crypto/scatterwalk.h>
35 #include <net/ip6_checksum.h>
36 
37 #include "tls.h"
38 
39 static void chain_to_walk(struct scatterlist *sg, struct scatter_walk *walk)
40 {
41 	struct scatterlist *src = walk->sg;
42 	int diff = walk->offset - src->offset;
43 
44 	sg_set_page(sg, sg_page(src),
45 		    src->length - diff, walk->offset);
46 
47 	scatterwalk_crypto_chain(sg, sg_next(src), 2);
48 }
49 
50 static int tls_enc_record(struct aead_request *aead_req,
51 			  struct crypto_aead *aead, char *aad,
52 			  char *iv, __be64 rcd_sn,
53 			  struct scatter_walk *in,
54 			  struct scatter_walk *out, int *in_len,
55 			  struct tls_prot_info *prot)
56 {
57 	unsigned char buf[TLS_HEADER_SIZE + MAX_IV_SIZE];
58 	const struct tls_cipher_size_desc *cipher_sz;
59 	struct scatterlist sg_in[3];
60 	struct scatterlist sg_out[3];
61 	unsigned int buf_size;
62 	u16 len;
63 	int rc;
64 
65 	switch (prot->cipher_type) {
66 	case TLS_CIPHER_AES_GCM_128:
67 	case TLS_CIPHER_AES_GCM_256:
68 		break;
69 	default:
70 		return -EINVAL;
71 	}
72 	cipher_sz = &tls_cipher_size_desc[prot->cipher_type];
73 
74 	buf_size = TLS_HEADER_SIZE + cipher_sz->iv;
75 	len = min_t(int, *in_len, buf_size);
76 
77 	scatterwalk_copychunks(buf, in, len, 0);
78 	scatterwalk_copychunks(buf, out, len, 1);
79 
80 	*in_len -= len;
81 	if (!*in_len)
82 		return 0;
83 
84 	scatterwalk_pagedone(in, 0, 1);
85 	scatterwalk_pagedone(out, 1, 1);
86 
87 	len = buf[4] | (buf[3] << 8);
88 	len -= cipher_sz->iv;
89 
90 	tls_make_aad(aad, len - cipher_sz->tag, (char *)&rcd_sn, buf[0], prot);
91 
92 	memcpy(iv + cipher_sz->salt, buf + TLS_HEADER_SIZE, cipher_sz->iv);
93 
94 	sg_init_table(sg_in, ARRAY_SIZE(sg_in));
95 	sg_init_table(sg_out, ARRAY_SIZE(sg_out));
96 	sg_set_buf(sg_in, aad, TLS_AAD_SPACE_SIZE);
97 	sg_set_buf(sg_out, aad, TLS_AAD_SPACE_SIZE);
98 	chain_to_walk(sg_in + 1, in);
99 	chain_to_walk(sg_out + 1, out);
100 
101 	*in_len -= len;
102 	if (*in_len < 0) {
103 		*in_len += cipher_sz->tag;
104 		/* the input buffer doesn't contain the entire record.
105 		 * trim len accordingly. The resulting authentication tag
106 		 * will contain garbage, but we don't care, so we won't
107 		 * include any of it in the output skb
108 		 * Note that we assume the output buffer length
109 		 * is larger then input buffer length + tag size
110 		 */
111 		if (*in_len < 0)
112 			len += *in_len;
113 
114 		*in_len = 0;
115 	}
116 
117 	if (*in_len) {
118 		scatterwalk_copychunks(NULL, in, len, 2);
119 		scatterwalk_pagedone(in, 0, 1);
120 		scatterwalk_copychunks(NULL, out, len, 2);
121 		scatterwalk_pagedone(out, 1, 1);
122 	}
123 
124 	len -= cipher_sz->tag;
125 	aead_request_set_crypt(aead_req, sg_in, sg_out, len, iv);
126 
127 	rc = crypto_aead_encrypt(aead_req);
128 
129 	return rc;
130 }
131 
132 static void tls_init_aead_request(struct aead_request *aead_req,
133 				  struct crypto_aead *aead)
134 {
135 	aead_request_set_tfm(aead_req, aead);
136 	aead_request_set_ad(aead_req, TLS_AAD_SPACE_SIZE);
137 }
138 
139 static struct aead_request *tls_alloc_aead_request(struct crypto_aead *aead,
140 						   gfp_t flags)
141 {
142 	unsigned int req_size = sizeof(struct aead_request) +
143 		crypto_aead_reqsize(aead);
144 	struct aead_request *aead_req;
145 
146 	aead_req = kzalloc(req_size, flags);
147 	if (aead_req)
148 		tls_init_aead_request(aead_req, aead);
149 	return aead_req;
150 }
151 
152 static int tls_enc_records(struct aead_request *aead_req,
153 			   struct crypto_aead *aead, struct scatterlist *sg_in,
154 			   struct scatterlist *sg_out, char *aad, char *iv,
155 			   u64 rcd_sn, int len, struct tls_prot_info *prot)
156 {
157 	struct scatter_walk out, in;
158 	int rc;
159 
160 	scatterwalk_start(&in, sg_in);
161 	scatterwalk_start(&out, sg_out);
162 
163 	do {
164 		rc = tls_enc_record(aead_req, aead, aad, iv,
165 				    cpu_to_be64(rcd_sn), &in, &out, &len, prot);
166 		rcd_sn++;
167 
168 	} while (rc == 0 && len);
169 
170 	scatterwalk_done(&in, 0, 0);
171 	scatterwalk_done(&out, 1, 0);
172 
173 	return rc;
174 }
175 
176 /* Can't use icsk->icsk_af_ops->send_check here because the ip addresses
177  * might have been changed by NAT.
178  */
179 static void update_chksum(struct sk_buff *skb, int headln)
180 {
181 	struct tcphdr *th = tcp_hdr(skb);
182 	int datalen = skb->len - headln;
183 	const struct ipv6hdr *ipv6h;
184 	const struct iphdr *iph;
185 
186 	/* We only changed the payload so if we are using partial we don't
187 	 * need to update anything.
188 	 */
189 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL))
190 		return;
191 
192 	skb->ip_summed = CHECKSUM_PARTIAL;
193 	skb->csum_start = skb_transport_header(skb) - skb->head;
194 	skb->csum_offset = offsetof(struct tcphdr, check);
195 
196 	if (skb->sk->sk_family == AF_INET6) {
197 		ipv6h = ipv6_hdr(skb);
198 		th->check = ~csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
199 					     datalen, IPPROTO_TCP, 0);
200 	} else {
201 		iph = ip_hdr(skb);
202 		th->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr, datalen,
203 					       IPPROTO_TCP, 0);
204 	}
205 }
206 
207 static void complete_skb(struct sk_buff *nskb, struct sk_buff *skb, int headln)
208 {
209 	struct sock *sk = skb->sk;
210 	int delta;
211 
212 	skb_copy_header(nskb, skb);
213 
214 	skb_put(nskb, skb->len);
215 	memcpy(nskb->data, skb->data, headln);
216 
217 	nskb->destructor = skb->destructor;
218 	nskb->sk = sk;
219 	skb->destructor = NULL;
220 	skb->sk = NULL;
221 
222 	update_chksum(nskb, headln);
223 
224 	/* sock_efree means skb must gone through skb_orphan_partial() */
225 	if (nskb->destructor == sock_efree)
226 		return;
227 
228 	delta = nskb->truesize - skb->truesize;
229 	if (likely(delta < 0))
230 		WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
231 	else if (delta)
232 		refcount_add(delta, &sk->sk_wmem_alloc);
233 }
234 
235 /* This function may be called after the user socket is already
236  * closed so make sure we don't use anything freed during
237  * tls_sk_proto_close here
238  */
239 
240 static int fill_sg_in(struct scatterlist *sg_in,
241 		      struct sk_buff *skb,
242 		      struct tls_offload_context_tx *ctx,
243 		      u64 *rcd_sn,
244 		      s32 *sync_size,
245 		      int *resync_sgs)
246 {
247 	int tcp_payload_offset = skb_tcp_all_headers(skb);
248 	int payload_len = skb->len - tcp_payload_offset;
249 	u32 tcp_seq = ntohl(tcp_hdr(skb)->seq);
250 	struct tls_record_info *record;
251 	unsigned long flags;
252 	int remaining;
253 	int i;
254 
255 	spin_lock_irqsave(&ctx->lock, flags);
256 	record = tls_get_record(ctx, tcp_seq, rcd_sn);
257 	if (!record) {
258 		spin_unlock_irqrestore(&ctx->lock, flags);
259 		return -EINVAL;
260 	}
261 
262 	*sync_size = tcp_seq - tls_record_start_seq(record);
263 	if (*sync_size < 0) {
264 		int is_start_marker = tls_record_is_start_marker(record);
265 
266 		spin_unlock_irqrestore(&ctx->lock, flags);
267 		/* This should only occur if the relevant record was
268 		 * already acked. In that case it should be ok
269 		 * to drop the packet and avoid retransmission.
270 		 *
271 		 * There is a corner case where the packet contains
272 		 * both an acked and a non-acked record.
273 		 * We currently don't handle that case and rely
274 		 * on TCP to retransmit a packet that doesn't contain
275 		 * already acked payload.
276 		 */
277 		if (!is_start_marker)
278 			*sync_size = 0;
279 		return -EINVAL;
280 	}
281 
282 	remaining = *sync_size;
283 	for (i = 0; remaining > 0; i++) {
284 		skb_frag_t *frag = &record->frags[i];
285 
286 		__skb_frag_ref(frag);
287 		sg_set_page(sg_in + i, skb_frag_page(frag),
288 			    skb_frag_size(frag), skb_frag_off(frag));
289 
290 		remaining -= skb_frag_size(frag);
291 
292 		if (remaining < 0)
293 			sg_in[i].length += remaining;
294 	}
295 	*resync_sgs = i;
296 
297 	spin_unlock_irqrestore(&ctx->lock, flags);
298 	if (skb_to_sgvec(skb, &sg_in[i], tcp_payload_offset, payload_len) < 0)
299 		return -EINVAL;
300 
301 	return 0;
302 }
303 
304 static void fill_sg_out(struct scatterlist sg_out[3], void *buf,
305 			struct tls_context *tls_ctx,
306 			struct sk_buff *nskb,
307 			int tcp_payload_offset,
308 			int payload_len,
309 			int sync_size,
310 			void *dummy_buf)
311 {
312 	const struct tls_cipher_size_desc *cipher_sz =
313 		&tls_cipher_size_desc[tls_ctx->crypto_send.info.cipher_type];
314 
315 	sg_set_buf(&sg_out[0], dummy_buf, sync_size);
316 	sg_set_buf(&sg_out[1], nskb->data + tcp_payload_offset, payload_len);
317 	/* Add room for authentication tag produced by crypto */
318 	dummy_buf += sync_size;
319 	sg_set_buf(&sg_out[2], dummy_buf, cipher_sz->tag);
320 }
321 
322 static struct sk_buff *tls_enc_skb(struct tls_context *tls_ctx,
323 				   struct scatterlist sg_out[3],
324 				   struct scatterlist *sg_in,
325 				   struct sk_buff *skb,
326 				   s32 sync_size, u64 rcd_sn)
327 {
328 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
329 	int tcp_payload_offset = skb_tcp_all_headers(skb);
330 	int payload_len = skb->len - tcp_payload_offset;
331 	const struct tls_cipher_size_desc *cipher_sz;
332 	void *buf, *iv, *aad, *dummy_buf, *salt;
333 	struct aead_request *aead_req;
334 	struct sk_buff *nskb = NULL;
335 	int buf_len;
336 
337 	aead_req = tls_alloc_aead_request(ctx->aead_send, GFP_ATOMIC);
338 	if (!aead_req)
339 		return NULL;
340 
341 	switch (tls_ctx->crypto_send.info.cipher_type) {
342 	case TLS_CIPHER_AES_GCM_128:
343 		salt = tls_ctx->crypto_send.aes_gcm_128.salt;
344 		break;
345 	case TLS_CIPHER_AES_GCM_256:
346 		salt = tls_ctx->crypto_send.aes_gcm_256.salt;
347 		break;
348 	default:
349 		goto free_req;
350 	}
351 	cipher_sz = &tls_cipher_size_desc[tls_ctx->crypto_send.info.cipher_type];
352 	buf_len = cipher_sz->salt + cipher_sz->iv + TLS_AAD_SPACE_SIZE +
353 		  sync_size + cipher_sz->tag;
354 	buf = kmalloc(buf_len, GFP_ATOMIC);
355 	if (!buf)
356 		goto free_req;
357 
358 	iv = buf;
359 	memcpy(iv, salt, cipher_sz->salt);
360 	aad = buf + cipher_sz->salt + cipher_sz->iv;
361 	dummy_buf = aad + TLS_AAD_SPACE_SIZE;
362 
363 	nskb = alloc_skb(skb_headroom(skb) + skb->len, GFP_ATOMIC);
364 	if (!nskb)
365 		goto free_buf;
366 
367 	skb_reserve(nskb, skb_headroom(skb));
368 
369 	fill_sg_out(sg_out, buf, tls_ctx, nskb, tcp_payload_offset,
370 		    payload_len, sync_size, dummy_buf);
371 
372 	if (tls_enc_records(aead_req, ctx->aead_send, sg_in, sg_out, aad, iv,
373 			    rcd_sn, sync_size + payload_len,
374 			    &tls_ctx->prot_info) < 0)
375 		goto free_nskb;
376 
377 	complete_skb(nskb, skb, tcp_payload_offset);
378 
379 	/* validate_xmit_skb_list assumes that if the skb wasn't segmented
380 	 * nskb->prev will point to the skb itself
381 	 */
382 	nskb->prev = nskb;
383 
384 free_buf:
385 	kfree(buf);
386 free_req:
387 	kfree(aead_req);
388 	return nskb;
389 free_nskb:
390 	kfree_skb(nskb);
391 	nskb = NULL;
392 	goto free_buf;
393 }
394 
395 static struct sk_buff *tls_sw_fallback(struct sock *sk, struct sk_buff *skb)
396 {
397 	int tcp_payload_offset = skb_tcp_all_headers(skb);
398 	struct tls_context *tls_ctx = tls_get_ctx(sk);
399 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
400 	int payload_len = skb->len - tcp_payload_offset;
401 	struct scatterlist *sg_in, sg_out[3];
402 	struct sk_buff *nskb = NULL;
403 	int sg_in_max_elements;
404 	int resync_sgs = 0;
405 	s32 sync_size = 0;
406 	u64 rcd_sn;
407 
408 	/* worst case is:
409 	 * MAX_SKB_FRAGS in tls_record_info
410 	 * MAX_SKB_FRAGS + 1 in SKB head and frags.
411 	 */
412 	sg_in_max_elements = 2 * MAX_SKB_FRAGS + 1;
413 
414 	if (!payload_len)
415 		return skb;
416 
417 	sg_in = kmalloc_array(sg_in_max_elements, sizeof(*sg_in), GFP_ATOMIC);
418 	if (!sg_in)
419 		goto free_orig;
420 
421 	sg_init_table(sg_in, sg_in_max_elements);
422 	sg_init_table(sg_out, ARRAY_SIZE(sg_out));
423 
424 	if (fill_sg_in(sg_in, skb, ctx, &rcd_sn, &sync_size, &resync_sgs)) {
425 		/* bypass packets before kernel TLS socket option was set */
426 		if (sync_size < 0 && payload_len <= -sync_size)
427 			nskb = skb_get(skb);
428 		goto put_sg;
429 	}
430 
431 	nskb = tls_enc_skb(tls_ctx, sg_out, sg_in, skb, sync_size, rcd_sn);
432 
433 put_sg:
434 	while (resync_sgs)
435 		put_page(sg_page(&sg_in[--resync_sgs]));
436 	kfree(sg_in);
437 free_orig:
438 	if (nskb)
439 		consume_skb(skb);
440 	else
441 		kfree_skb(skb);
442 	return nskb;
443 }
444 
445 struct sk_buff *tls_validate_xmit_skb(struct sock *sk,
446 				      struct net_device *dev,
447 				      struct sk_buff *skb)
448 {
449 	if (dev == rcu_dereference_bh(tls_get_ctx(sk)->netdev) ||
450 	    netif_is_bond_master(dev))
451 		return skb;
452 
453 	return tls_sw_fallback(sk, skb);
454 }
455 EXPORT_SYMBOL_GPL(tls_validate_xmit_skb);
456 
457 struct sk_buff *tls_validate_xmit_skb_sw(struct sock *sk,
458 					 struct net_device *dev,
459 					 struct sk_buff *skb)
460 {
461 	return tls_sw_fallback(sk, skb);
462 }
463 
464 struct sk_buff *tls_encrypt_skb(struct sk_buff *skb)
465 {
466 	return tls_sw_fallback(skb->sk, skb);
467 }
468 EXPORT_SYMBOL_GPL(tls_encrypt_skb);
469 
470 int tls_sw_fallback_init(struct sock *sk,
471 			 struct tls_offload_context_tx *offload_ctx,
472 			 struct tls_crypto_info *crypto_info)
473 {
474 	const struct tls_cipher_size_desc *cipher_sz;
475 	const u8 *key;
476 	int rc;
477 
478 	offload_ctx->aead_send =
479 	    crypto_alloc_aead("gcm(aes)", 0, CRYPTO_ALG_ASYNC);
480 	if (IS_ERR(offload_ctx->aead_send)) {
481 		rc = PTR_ERR(offload_ctx->aead_send);
482 		pr_err_ratelimited("crypto_alloc_aead failed rc=%d\n", rc);
483 		offload_ctx->aead_send = NULL;
484 		goto err_out;
485 	}
486 
487 	switch (crypto_info->cipher_type) {
488 	case TLS_CIPHER_AES_GCM_128:
489 		key = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->key;
490 		break;
491 	case TLS_CIPHER_AES_GCM_256:
492 		key = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->key;
493 		break;
494 	default:
495 		rc = -EINVAL;
496 		goto free_aead;
497 	}
498 	cipher_sz = &tls_cipher_size_desc[crypto_info->cipher_type];
499 
500 	rc = crypto_aead_setkey(offload_ctx->aead_send, key, cipher_sz->key);
501 	if (rc)
502 		goto free_aead;
503 
504 	rc = crypto_aead_setauthsize(offload_ctx->aead_send, cipher_sz->tag);
505 	if (rc)
506 		goto free_aead;
507 
508 	return 0;
509 free_aead:
510 	crypto_free_aead(offload_ctx->aead_send);
511 err_out:
512 	return rc;
513 }
514