xref: /freebsd/sys/dev/cxgbe/tom/t4_tls.c (revision f2d48b5e2c3b45850585e4d7aee324fe148afbf2)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2017-2018 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: John Baldwin <jhb@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include "opt_inet.h"
31 #include "opt_kern_tls.h"
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/ktr.h>
38 #ifdef KERN_TLS
39 #include <sys/ktls.h>
40 #endif
41 #include <sys/sglist.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/systm.h>
45 #include <netinet/in.h>
46 #include <netinet/in_pcb.h>
47 #include <netinet/tcp_var.h>
48 #include <netinet/toecore.h>
49 #ifdef KERN_TLS
50 #include <opencrypto/cryptodev.h>
51 #include <opencrypto/xform.h>
52 #endif
53 
54 #ifdef TCP_OFFLOAD
55 #include "common/common.h"
56 #include "common/t4_tcb.h"
57 #include "crypto/t4_crypto.h"
58 #include "tom/t4_tom_l2t.h"
59 #include "tom/t4_tom.h"
60 
61 /*
62  * The TCP sequence number of a CPL_TLS_DATA mbuf is saved here while
63  * the mbuf is in the ulp_pdu_reclaimq.
64  */
65 #define	tls_tcp_seq	PH_loc.thirtytwo[0]
66 
67 static void
68 t4_set_tls_tcb_field(struct toepcb *toep, uint16_t word, uint64_t mask,
69     uint64_t val)
70 {
71 	struct adapter *sc = td_adapter(toep->td);
72 
73 	t4_set_tcb_field(sc, toep->ofld_txq, toep, word, mask, val, 0, 0);
74 }
75 
76 /* TLS and DTLS common routines */
77 bool
78 can_tls_offload(struct adapter *sc)
79 {
80 
81 	return (sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS);
82 }
83 
84 int
85 tls_tx_key(struct toepcb *toep)
86 {
87 	struct tls_ofld_info *tls_ofld = &toep->tls;
88 
89 	return (tls_ofld->tx_key_addr >= 0);
90 }
91 
92 int
93 tls_rx_key(struct toepcb *toep)
94 {
95 	struct tls_ofld_info *tls_ofld = &toep->tls;
96 
97 	return (tls_ofld->rx_key_addr >= 0);
98 }
99 
100 static int
101 key_size(struct toepcb *toep)
102 {
103 	struct tls_ofld_info *tls_ofld = &toep->tls;
104 
105 	return ((tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE) ?
106 		tls_ofld->k_ctx.tx_key_info_size : KEY_IN_DDR_SIZE);
107 }
108 
109 /* Set TLS Key-Id in TCB */
110 static void
111 t4_set_tls_keyid(struct toepcb *toep, unsigned int key_id)
112 {
113 
114 	t4_set_tls_tcb_field(toep, W_TCB_RX_TLS_KEY_TAG,
115 			 V_TCB_RX_TLS_KEY_TAG(M_TCB_RX_TLS_BUF_TAG),
116 			 V_TCB_RX_TLS_KEY_TAG(key_id));
117 }
118 
119 /* Clear TF_RX_QUIESCE to re-enable receive. */
120 static void
121 t4_clear_rx_quiesce(struct toepcb *toep)
122 {
123 
124 	t4_set_tls_tcb_field(toep, W_TCB_T_FLAGS, V_TF_RX_QUIESCE(1), 0);
125 }
126 
127 static void
128 tls_clr_ofld_mode(struct toepcb *toep)
129 {
130 
131 	tls_stop_handshake_timer(toep);
132 
133 	KASSERT(toep->tls.rx_key_addr == -1,
134 	    ("%s: tid %d has RX key", __func__, toep->tid));
135 
136 	/* Switch to plain TOE mode. */
137 	t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
138 	    V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)),
139 	    V_TCB_ULP_RAW(V_TF_TLS_ENABLE(0)));
140 	t4_set_tls_tcb_field(toep, W_TCB_ULP_TYPE,
141 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE), V_TCB_ULP_TYPE(ULP_MODE_NONE));
142 	t4_clear_rx_quiesce(toep);
143 
144 	toep->flags &= ~(TPF_FORCE_CREDITS | TPF_TLS_ESTABLISHED);
145 	toep->params.ulp_mode = ULP_MODE_NONE;
146 }
147 
148 static void
149 tls_clr_quiesce(struct toepcb *toep)
150 {
151 
152 	tls_stop_handshake_timer(toep);
153 	t4_clear_rx_quiesce(toep);
154 }
155 
156 /*
157  * Calculate the TLS data expansion size
158  */
159 static int
160 tls_expansion_size(struct toepcb *toep, int data_len, int full_pdus_only,
161     unsigned short *pdus_per_ulp)
162 {
163 	struct tls_ofld_info *tls_ofld = &toep->tls;
164 	struct tls_scmd *scmd = &tls_ofld->scmd0;
165 	int expn_size = 0, frag_count = 0, pad_per_pdu = 0,
166 	    pad_last_pdu = 0, last_frag_size = 0, max_frag_size = 0;
167 	int exp_per_pdu = 0;
168 	int hdr_len = TLS_HEADER_LENGTH;
169 
170 	do {
171 		max_frag_size = tls_ofld->k_ctx.frag_size;
172 		if (G_SCMD_CIPH_MODE(scmd->seqno_numivs) ==
173 		   SCMD_CIPH_MODE_AES_GCM) {
174 			frag_count = (data_len / max_frag_size);
175 			exp_per_pdu = GCM_TAG_SIZE + AEAD_EXPLICIT_DATA_SIZE +
176 				hdr_len;
177 			expn_size =  frag_count * exp_per_pdu;
178 			if (full_pdus_only) {
179 				*pdus_per_ulp = data_len / (exp_per_pdu +
180 					max_frag_size);
181 				if (*pdus_per_ulp > 32)
182 					*pdus_per_ulp = 32;
183 				else if(!*pdus_per_ulp)
184 					*pdus_per_ulp = 1;
185 				expn_size = (*pdus_per_ulp) * exp_per_pdu;
186 				break;
187 			}
188 			if ((last_frag_size = data_len % max_frag_size) > 0) {
189 				frag_count += 1;
190 				expn_size += exp_per_pdu;
191 			}
192 			break;
193 		} else if (G_SCMD_CIPH_MODE(scmd->seqno_numivs) !=
194 			   SCMD_CIPH_MODE_NOP) {
195 			/* Calculate the number of fragments we can make */
196 			frag_count  = (data_len / max_frag_size);
197 			if (frag_count > 0) {
198 				pad_per_pdu = (((howmany((max_frag_size +
199 						       tls_ofld->mac_length),
200 						      CIPHER_BLOCK_SIZE)) *
201 						CIPHER_BLOCK_SIZE) -
202 					       (max_frag_size +
203 						tls_ofld->mac_length));
204 				if (!pad_per_pdu)
205 					pad_per_pdu = CIPHER_BLOCK_SIZE;
206 				exp_per_pdu = pad_per_pdu +
207 				       	tls_ofld->mac_length +
208 					hdr_len + CIPHER_BLOCK_SIZE;
209 				expn_size = frag_count * exp_per_pdu;
210 			}
211 			if (full_pdus_only) {
212 				*pdus_per_ulp = data_len / (exp_per_pdu +
213 					max_frag_size);
214 				if (*pdus_per_ulp > 32)
215 					*pdus_per_ulp = 32;
216 				else if (!*pdus_per_ulp)
217 					*pdus_per_ulp = 1;
218 				expn_size = (*pdus_per_ulp) * exp_per_pdu;
219 				break;
220 			}
221 			/* Consider the last fragment */
222 			if ((last_frag_size = data_len % max_frag_size) > 0) {
223 				pad_last_pdu = (((howmany((last_frag_size +
224 							tls_ofld->mac_length),
225 						       CIPHER_BLOCK_SIZE)) *
226 						 CIPHER_BLOCK_SIZE) -
227 						(last_frag_size +
228 						 tls_ofld->mac_length));
229 				if (!pad_last_pdu)
230 					pad_last_pdu = CIPHER_BLOCK_SIZE;
231 				expn_size += (pad_last_pdu +
232 					      tls_ofld->mac_length + hdr_len +
233 					      CIPHER_BLOCK_SIZE);
234 			}
235 		}
236 	} while (0);
237 
238 	return (expn_size);
239 }
240 
241 /* Copy Key to WR */
242 static void
243 tls_copy_tx_key(struct toepcb *toep, void *dst)
244 {
245 	struct tls_ofld_info *tls_ofld = &toep->tls;
246 	struct ulptx_sc_memrd *sc_memrd;
247 	struct ulptx_idata *sc;
248 
249 	if (tls_ofld->k_ctx.tx_key_info_size <= 0)
250 		return;
251 
252 	if (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_DDR) {
253 		sc = dst;
254 		sc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
255 		sc->len = htobe32(0);
256 		sc_memrd = (struct ulptx_sc_memrd *)(sc + 1);
257 		sc_memrd->cmd_to_len = htobe32(V_ULPTX_CMD(ULP_TX_SC_MEMRD) |
258 		    V_ULP_TX_SC_MORE(1) |
259 		    V_ULPTX_LEN16(tls_ofld->k_ctx.tx_key_info_size >> 4));
260 		sc_memrd->addr = htobe32(tls_ofld->tx_key_addr >> 5);
261 	} else if (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE) {
262 		memcpy(dst, &tls_ofld->k_ctx.tx,
263 		    tls_ofld->k_ctx.tx_key_info_size);
264 	}
265 }
266 
267 /* TLS/DTLS content type  for CPL SFO */
268 static inline unsigned char
269 tls_content_type(unsigned char content_type)
270 {
271 	/*
272 	 * XXX: Shouldn't this map CONTENT_TYPE_APP_DATA to DATA and
273 	 * default to "CUSTOM" for all other types including
274 	 * heartbeat?
275 	 */
276 	switch (content_type) {
277 	case CONTENT_TYPE_CCS:
278 		return CPL_TX_TLS_SFO_TYPE_CCS;
279 	case CONTENT_TYPE_ALERT:
280 		return CPL_TX_TLS_SFO_TYPE_ALERT;
281 	case CONTENT_TYPE_HANDSHAKE:
282 		return CPL_TX_TLS_SFO_TYPE_HANDSHAKE;
283 	case CONTENT_TYPE_HEARTBEAT:
284 		return CPL_TX_TLS_SFO_TYPE_HEARTBEAT;
285 	}
286 	return CPL_TX_TLS_SFO_TYPE_DATA;
287 }
288 
289 static unsigned char
290 get_cipher_key_size(unsigned int ck_size)
291 {
292 	switch (ck_size) {
293 	case AES_NOP: /* NOP */
294 		return 15;
295 	case AES_128: /* AES128 */
296 		return CH_CK_SIZE_128;
297 	case AES_192: /* AES192 */
298 		return CH_CK_SIZE_192;
299 	case AES_256: /* AES256 */
300 		return CH_CK_SIZE_256;
301 	default:
302 		return CH_CK_SIZE_256;
303 	}
304 }
305 
306 static unsigned char
307 get_mac_key_size(unsigned int mk_size)
308 {
309 	switch (mk_size) {
310 	case SHA_NOP: /* NOP */
311 		return CH_MK_SIZE_128;
312 	case SHA_GHASH: /* GHASH */
313 	case SHA_512: /* SHA512 */
314 		return CH_MK_SIZE_512;
315 	case SHA_224: /* SHA2-224 */
316 		return CH_MK_SIZE_192;
317 	case SHA_256: /* SHA2-256*/
318 		return CH_MK_SIZE_256;
319 	case SHA_384: /* SHA384 */
320 		return CH_MK_SIZE_512;
321 	case SHA1: /* SHA1 */
322 	default:
323 		return CH_MK_SIZE_160;
324 	}
325 }
326 
327 static unsigned int
328 get_proto_ver(int proto_ver)
329 {
330 	switch (proto_ver) {
331 	case TLS1_2_VERSION:
332 		return TLS_1_2_VERSION;
333 	case TLS1_1_VERSION:
334 		return TLS_1_1_VERSION;
335 	case DTLS1_2_VERSION:
336 		return DTLS_1_2_VERSION;
337 	default:
338 		return TLS_VERSION_MAX;
339 	}
340 }
341 
342 static void
343 tls_rxkey_flit1(struct tls_keyctx *kwr, struct tls_key_context *kctx)
344 {
345 
346 	if (kctx->state.enc_mode == CH_EVP_CIPH_GCM_MODE) {
347 		kwr->u.rxhdr.ivinsert_to_authinsrt =
348 		    htobe64(V_TLS_KEYCTX_TX_WR_IVINSERT(6ULL) |
349 			V_TLS_KEYCTX_TX_WR_AADSTRTOFST(1ULL) |
350 			V_TLS_KEYCTX_TX_WR_AADSTOPOFST(5ULL) |
351 			V_TLS_KEYCTX_TX_WR_AUTHSRTOFST(14ULL) |
352 			V_TLS_KEYCTX_TX_WR_AUTHSTOPOFST(16ULL) |
353 			V_TLS_KEYCTX_TX_WR_CIPHERSRTOFST(14ULL) |
354 			V_TLS_KEYCTX_TX_WR_CIPHERSTOPOFST(0ULL) |
355 			V_TLS_KEYCTX_TX_WR_AUTHINSRT(16ULL));
356 		kwr->u.rxhdr.ivpresent_to_rxmk_size &=
357 			~(V_TLS_KEYCTX_TX_WR_RXOPAD_PRESENT(1));
358 		kwr->u.rxhdr.authmode_to_rxvalid &=
359 			~(V_TLS_KEYCTX_TX_WR_CIPHAUTHSEQCTRL(1));
360 	} else {
361 		kwr->u.rxhdr.ivinsert_to_authinsrt =
362 		    htobe64(V_TLS_KEYCTX_TX_WR_IVINSERT(6ULL) |
363 			V_TLS_KEYCTX_TX_WR_AADSTRTOFST(1ULL) |
364 			V_TLS_KEYCTX_TX_WR_AADSTOPOFST(5ULL) |
365 			V_TLS_KEYCTX_TX_WR_AUTHSRTOFST(22ULL) |
366 			V_TLS_KEYCTX_TX_WR_AUTHSTOPOFST(0ULL) |
367 			V_TLS_KEYCTX_TX_WR_CIPHERSRTOFST(22ULL) |
368 			V_TLS_KEYCTX_TX_WR_CIPHERSTOPOFST(0ULL) |
369 			V_TLS_KEYCTX_TX_WR_AUTHINSRT(0ULL));
370 	}
371 }
372 
373 /* Rx key */
374 static void
375 prepare_rxkey_wr(struct tls_keyctx *kwr, struct tls_key_context *kctx)
376 {
377 	unsigned int ck_size = kctx->cipher_secret_size;
378 	unsigned int mk_size = kctx->mac_secret_size;
379 	int proto_ver = kctx->proto_ver;
380 
381 	kwr->u.rxhdr.flitcnt_hmacctrl =
382 		((kctx->rx_key_info_size >> 4) << 3) | kctx->hmac_ctrl;
383 
384 	kwr->u.rxhdr.protover_ciphmode =
385 		V_TLS_KEYCTX_TX_WR_PROTOVER(get_proto_ver(proto_ver)) |
386 		V_TLS_KEYCTX_TX_WR_CIPHMODE(kctx->state.enc_mode);
387 
388 	kwr->u.rxhdr.authmode_to_rxvalid =
389 		V_TLS_KEYCTX_TX_WR_AUTHMODE(kctx->state.auth_mode) |
390 		V_TLS_KEYCTX_TX_WR_CIPHAUTHSEQCTRL(1) |
391 		V_TLS_KEYCTX_TX_WR_SEQNUMCTRL(3) |
392 		V_TLS_KEYCTX_TX_WR_RXVALID(1);
393 
394 	kwr->u.rxhdr.ivpresent_to_rxmk_size =
395 		V_TLS_KEYCTX_TX_WR_IVPRESENT(0) |
396 		V_TLS_KEYCTX_TX_WR_RXOPAD_PRESENT(1) |
397 		V_TLS_KEYCTX_TX_WR_RXCK_SIZE(get_cipher_key_size(ck_size)) |
398 		V_TLS_KEYCTX_TX_WR_RXMK_SIZE(get_mac_key_size(mk_size));
399 
400 	tls_rxkey_flit1(kwr, kctx);
401 
402 	/* No key reversal for GCM */
403 	if (kctx->state.enc_mode != CH_EVP_CIPH_GCM_MODE) {
404 		t4_aes_getdeckey(kwr->keys.edkey, kctx->rx.key,
405 				 (kctx->cipher_secret_size << 3));
406 		memcpy(kwr->keys.edkey + kctx->cipher_secret_size,
407 		       kctx->rx.key + kctx->cipher_secret_size,
408 		       (IPAD_SIZE + OPAD_SIZE));
409 	} else {
410 		memcpy(kwr->keys.edkey, kctx->rx.key,
411 		       (kctx->rx_key_info_size - SALT_SIZE));
412 		memcpy(kwr->u.rxhdr.rxsalt, kctx->rx.salt, SALT_SIZE);
413 	}
414 }
415 
416 /* Tx key */
417 static void
418 prepare_txkey_wr(struct tls_keyctx *kwr, struct tls_key_context *kctx)
419 {
420 	unsigned int ck_size = kctx->cipher_secret_size;
421 	unsigned int mk_size = kctx->mac_secret_size;
422 
423 	kwr->u.txhdr.ctxlen =
424 		(kctx->tx_key_info_size >> 4);
425 	kwr->u.txhdr.dualck_to_txvalid =
426 		V_TLS_KEYCTX_TX_WR_TXOPAD_PRESENT(1) |
427 		V_TLS_KEYCTX_TX_WR_SALT_PRESENT(1) |
428 		V_TLS_KEYCTX_TX_WR_TXCK_SIZE(get_cipher_key_size(ck_size)) |
429 		V_TLS_KEYCTX_TX_WR_TXMK_SIZE(get_mac_key_size(mk_size)) |
430 		V_TLS_KEYCTX_TX_WR_TXVALID(1);
431 
432 	memcpy(kwr->keys.edkey, kctx->tx.key, HDR_KCTX_SIZE);
433 	if (kctx->state.enc_mode == CH_EVP_CIPH_GCM_MODE) {
434 		memcpy(kwr->u.txhdr.txsalt, kctx->tx.salt, SALT_SIZE);
435 		kwr->u.txhdr.dualck_to_txvalid &=
436 			~(V_TLS_KEYCTX_TX_WR_TXOPAD_PRESENT(1));
437 	}
438 	kwr->u.txhdr.dualck_to_txvalid = htons(kwr->u.txhdr.dualck_to_txvalid);
439 }
440 
441 /* TLS Key memory management */
442 static int
443 get_new_keyid(struct toepcb *toep)
444 {
445 	struct adapter *sc = td_adapter(toep->td);
446 	vmem_addr_t addr;
447 
448 	if (vmem_alloc(sc->key_map, TLS_KEY_CONTEXT_SZ, M_NOWAIT | M_FIRSTFIT,
449 	    &addr) != 0)
450 		return (-1);
451 
452 	return (addr);
453 }
454 
455 static void
456 free_keyid(struct toepcb *toep, int keyid)
457 {
458 	struct adapter *sc = td_adapter(toep->td);
459 
460 	vmem_free(sc->key_map, keyid, TLS_KEY_CONTEXT_SZ);
461 }
462 
463 static void
464 clear_tls_keyid(struct toepcb *toep)
465 {
466 	struct tls_ofld_info *tls_ofld = &toep->tls;
467 
468 	if (tls_ofld->rx_key_addr >= 0) {
469 		free_keyid(toep, tls_ofld->rx_key_addr);
470 		tls_ofld->rx_key_addr = -1;
471 	}
472 	if (tls_ofld->tx_key_addr >= 0) {
473 		free_keyid(toep, tls_ofld->tx_key_addr);
474 		tls_ofld->tx_key_addr = -1;
475 	}
476 }
477 
478 static int
479 get_keyid(struct tls_ofld_info *tls_ofld, unsigned int ops)
480 {
481 	return (ops & KEY_WRITE_RX ? tls_ofld->rx_key_addr :
482 		((ops & KEY_WRITE_TX) ? tls_ofld->tx_key_addr : -1));
483 }
484 
485 static int
486 get_tp_plen_max(struct tls_ofld_info *tls_ofld)
487 {
488 	int plen = ((min(3*4096, TP_TX_PG_SZ))/1448) * 1448;
489 
490 	return (tls_ofld->k_ctx.frag_size <= 8192 ? plen : FC_TP_PLEN_MAX);
491 }
492 
493 /* Send request to get the key-id */
494 static int
495 tls_program_key_id(struct toepcb *toep, struct tls_key_context *k_ctx)
496 {
497 	struct tls_ofld_info *tls_ofld = &toep->tls;
498 	struct adapter *sc = td_adapter(toep->td);
499 	struct ofld_tx_sdesc *txsd;
500 	int kwrlen, kctxlen, keyid, len;
501 	struct wrqe *wr;
502 	struct tls_key_req *kwr;
503 	struct tls_keyctx *kctx;
504 
505 	kwrlen = sizeof(*kwr);
506 	kctxlen = roundup2(sizeof(*kctx), 32);
507 	len = roundup2(kwrlen + kctxlen, 16);
508 
509 	if (toep->txsd_avail == 0)
510 		return (EAGAIN);
511 
512 	/* Dont initialize key for re-neg */
513 	if (!G_KEY_CLR_LOC(k_ctx->l_p_key)) {
514 		if ((keyid = get_new_keyid(toep)) < 0) {
515 			return (ENOSPC);
516 		}
517 	} else {
518 		keyid = get_keyid(tls_ofld, k_ctx->l_p_key);
519 	}
520 
521 	wr = alloc_wrqe(len, toep->ofld_txq);
522 	if (wr == NULL) {
523 		free_keyid(toep, keyid);
524 		return (ENOMEM);
525 	}
526 	kwr = wrtod(wr);
527 	memset(kwr, 0, kwrlen);
528 
529 	kwr->wr_hi = htobe32(V_FW_WR_OP(FW_ULPTX_WR) | F_FW_WR_COMPL |
530 	    F_FW_WR_ATOMIC);
531 	kwr->wr_mid = htobe32(V_FW_WR_LEN16(DIV_ROUND_UP(len, 16)) |
532 	    V_FW_WR_FLOWID(toep->tid));
533 	kwr->protocol = get_proto_ver(k_ctx->proto_ver);
534 	kwr->mfs = htons(k_ctx->frag_size);
535 	kwr->reneg_to_write_rx = k_ctx->l_p_key;
536 
537 	/* master command */
538 	kwr->cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE) |
539 	    V_T5_ULP_MEMIO_ORDER(1) | V_T5_ULP_MEMIO_IMM(1));
540 	kwr->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(kctxlen >> 5));
541 	kwr->len16 = htobe32((toep->tid << 8) |
542 	    DIV_ROUND_UP(len - sizeof(struct work_request_hdr), 16));
543 	kwr->kaddr = htobe32(V_ULP_MEMIO_ADDR(keyid >> 5));
544 
545 	/* sub command */
546 	kwr->sc_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
547 	kwr->sc_len = htobe32(kctxlen);
548 
549 	kctx = (struct tls_keyctx *)(kwr + 1);
550 	memset(kctx, 0, kctxlen);
551 
552 	if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_TX) {
553 		tls_ofld->tx_key_addr = keyid;
554 		prepare_txkey_wr(kctx, k_ctx);
555 	} else if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
556 		tls_ofld->rx_key_addr = keyid;
557 		prepare_rxkey_wr(kctx, k_ctx);
558 	}
559 
560 	txsd = &toep->txsd[toep->txsd_pidx];
561 	txsd->tx_credits = DIV_ROUND_UP(len, 16);
562 	txsd->plen = 0;
563 	toep->tx_credits -= txsd->tx_credits;
564 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
565 		toep->txsd_pidx = 0;
566 	toep->txsd_avail--;
567 
568 	t4_wrq_tx(sc, wr);
569 
570 	return (0);
571 }
572 
573 /* Store a key received from SSL in DDR. */
574 static int
575 program_key_context(struct tcpcb *tp, struct toepcb *toep,
576     struct tls_key_context *uk_ctx)
577 {
578 	struct adapter *sc = td_adapter(toep->td);
579 	struct tls_ofld_info *tls_ofld = &toep->tls;
580 	struct tls_key_context *k_ctx;
581 	int error, key_offset;
582 
583 	if (tp->t_state != TCPS_ESTABLISHED) {
584 		/*
585 		 * XXX: Matches Linux driver, but not sure this is a
586 		 * very appropriate error.
587 		 */
588 		return (ENOENT);
589 	}
590 
591 	/* Stop timer on handshake completion */
592 	tls_stop_handshake_timer(toep);
593 
594 	toep->flags &= ~TPF_FORCE_CREDITS;
595 
596 	CTR4(KTR_CXGBE, "%s: tid %d %s proto_ver %#x", __func__, toep->tid,
597 	    G_KEY_GET_LOC(uk_ctx->l_p_key) == KEY_WRITE_RX ? "KEY_WRITE_RX" :
598 	    "KEY_WRITE_TX", uk_ctx->proto_ver);
599 
600 	if (G_KEY_GET_LOC(uk_ctx->l_p_key) == KEY_WRITE_RX &&
601 	    ulp_mode(toep) != ULP_MODE_TLS)
602 		return (EOPNOTSUPP);
603 
604 	/* Don't copy the 'tx' and 'rx' fields. */
605 	k_ctx = &tls_ofld->k_ctx;
606 	memcpy(&k_ctx->l_p_key, &uk_ctx->l_p_key,
607 	    sizeof(*k_ctx) - offsetof(struct tls_key_context, l_p_key));
608 
609 	/* TLS version != 1.1 and !1.2 OR DTLS != 1.2 */
610 	if (get_proto_ver(k_ctx->proto_ver) > DTLS_1_2_VERSION) {
611 		if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
612 			tls_ofld->rx_key_addr = -1;
613 			t4_clear_rx_quiesce(toep);
614 		} else {
615 			tls_ofld->tx_key_addr = -1;
616 		}
617 		return (0);
618 	}
619 
620 	if (k_ctx->state.enc_mode == CH_EVP_CIPH_GCM_MODE) {
621 		k_ctx->iv_size = 4;
622 		k_ctx->mac_first = 0;
623 		k_ctx->hmac_ctrl = 0;
624 	} else {
625 		k_ctx->iv_size = 8; /* for CBC, iv is 16B, unit of 2B */
626 		k_ctx->mac_first = 1;
627 	}
628 
629 	tls_ofld->scmd0.seqno_numivs =
630 		(V_SCMD_SEQ_NO_CTRL(3) |
631 		 V_SCMD_PROTO_VERSION(get_proto_ver(k_ctx->proto_ver)) |
632 		 V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
633 		 V_SCMD_CIPH_AUTH_SEQ_CTRL((k_ctx->mac_first == 0)) |
634 		 V_SCMD_CIPH_MODE(k_ctx->state.enc_mode) |
635 		 V_SCMD_AUTH_MODE(k_ctx->state.auth_mode) |
636 		 V_SCMD_HMAC_CTRL(k_ctx->hmac_ctrl) |
637 		 V_SCMD_IV_SIZE(k_ctx->iv_size));
638 
639 	tls_ofld->scmd0.ivgen_hdrlen =
640 		(V_SCMD_IV_GEN_CTRL(k_ctx->iv_ctrl) |
641 		 V_SCMD_KEY_CTX_INLINE(0) |
642 		 V_SCMD_TLS_FRAG_ENABLE(1));
643 
644 	tls_ofld->mac_length = k_ctx->mac_secret_size;
645 
646 	if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
647 		k_ctx->rx = uk_ctx->rx;
648 		/* Dont initialize key for re-neg */
649 		if (!G_KEY_CLR_LOC(k_ctx->l_p_key))
650 			tls_ofld->rx_key_addr = -1;
651 	} else {
652 		k_ctx->tx = uk_ctx->tx;
653 		/* Dont initialize key for re-neg */
654 		if (!G_KEY_CLR_LOC(k_ctx->l_p_key))
655 			tls_ofld->tx_key_addr = -1;
656 	}
657 
658 	/* Flush pending data before new Tx key becomes active */
659 	if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_TX) {
660 		struct sockbuf *sb;
661 
662 		/* XXX: This might not drain everything. */
663 		t4_push_frames(sc, toep, 0);
664 		sb = &toep->inp->inp_socket->so_snd;
665 		SOCKBUF_LOCK(sb);
666 
667 		/* XXX: This asserts that everything has been pushed. */
668 		MPASS(sb->sb_sndptr == NULL || sb->sb_sndptr->m_next == NULL);
669 		sb->sb_sndptr = NULL;
670 		tls_ofld->sb_off = sbavail(sb);
671 		SOCKBUF_UNLOCK(sb);
672 		tls_ofld->tx_seq_no = 0;
673 	}
674 
675 	if ((G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) ||
676 	    (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_DDR)) {
677 
678 		/*
679 		 * XXX: The userland library sets tx_key_info_size, not
680 		 * rx_key_info_size.
681 		 */
682 		k_ctx->rx_key_info_size = k_ctx->tx_key_info_size;
683 
684 		error = tls_program_key_id(toep, k_ctx);
685 		if (error) {
686 			/* XXX: Only clear quiesce for KEY_WRITE_RX? */
687 			t4_clear_rx_quiesce(toep);
688 			return (error);
689 		}
690 	}
691 
692 	if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
693 		/*
694 		 * RX key tags are an index into the key portion of MA
695 		 * memory stored as an offset from the base address in
696 		 * units of 64 bytes.
697 		 */
698 		key_offset = tls_ofld->rx_key_addr - sc->vres.key.start;
699 		t4_set_tls_keyid(toep, key_offset / 64);
700 		t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
701 				 V_TCB_ULP_RAW(M_TCB_ULP_RAW),
702 				 V_TCB_ULP_RAW((V_TF_TLS_KEY_SIZE(3) |
703 						V_TF_TLS_CONTROL(1) |
704 						V_TF_TLS_ACTIVE(1) |
705 						V_TF_TLS_ENABLE(1))));
706 		t4_set_tls_tcb_field(toep, W_TCB_TLS_SEQ,
707 				 V_TCB_TLS_SEQ(M_TCB_TLS_SEQ),
708 				 V_TCB_TLS_SEQ(0));
709 		t4_clear_rx_quiesce(toep);
710 
711 		toep->flags |= TPF_TLS_RECEIVE;
712 	} else {
713 		unsigned short pdus_per_ulp;
714 
715 		if (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE)
716 			tls_ofld->tx_key_addr = 1;
717 
718 		tls_ofld->fcplenmax = get_tp_plen_max(tls_ofld);
719 		tls_ofld->expn_per_ulp = tls_expansion_size(toep,
720 				tls_ofld->fcplenmax, 1, &pdus_per_ulp);
721 		tls_ofld->pdus_per_ulp = pdus_per_ulp;
722 		tls_ofld->adjusted_plen = tls_ofld->pdus_per_ulp *
723 			((tls_ofld->expn_per_ulp/tls_ofld->pdus_per_ulp) +
724 			 tls_ofld->k_ctx.frag_size);
725 	}
726 
727 	return (0);
728 }
729 
730 /*
731  * In some cases a client connection can hang without sending the
732  * ServerHelloDone message from the NIC to the host.  Send a dummy
733  * RX_DATA_ACK with RX_MODULATE to unstick the connection.
734  */
735 static void
736 tls_send_handshake_ack(void *arg)
737 {
738 	struct toepcb *toep = arg;
739 	struct tls_ofld_info *tls_ofld = &toep->tls;
740 	struct adapter *sc = td_adapter(toep->td);
741 
742 	/*
743 	 * If this connection has timed out without receiving more
744 	 * data, downgrade to plain TOE mode and don't re-arm the
745 	 * timer.
746 	 */
747 	if (sc->tt.tls_rx_timeout != 0) {
748 		struct inpcb *inp;
749 		struct tcpcb *tp;
750 
751 		inp = toep->inp;
752 		tp = intotcpcb(inp);
753 		if ((ticks - tp->t_rcvtime) >= sc->tt.tls_rx_timeout) {
754 			CTR2(KTR_CXGBE, "%s: tid %d clr_ofld_mode", __func__,
755 			    toep->tid);
756 			tls_clr_ofld_mode(toep);
757 			return;
758 		}
759 	}
760 
761 	/*
762 	 * XXX: Does not have the t4_get_tcb() checks to refine the
763 	 * workaround.
764 	 */
765 	callout_schedule(&tls_ofld->handshake_timer, TLS_SRV_HELLO_RD_TM * hz);
766 
767 	CTR2(KTR_CXGBE, "%s: tid %d sending RX_DATA_ACK", __func__, toep->tid);
768 	send_rx_modulate(sc, toep);
769 }
770 
771 static void
772 tls_start_handshake_timer(struct toepcb *toep)
773 {
774 	struct tls_ofld_info *tls_ofld = &toep->tls;
775 
776 	INP_WLOCK_ASSERT(toep->inp);
777 	callout_reset(&tls_ofld->handshake_timer, TLS_SRV_HELLO_BKOFF_TM * hz,
778 	    tls_send_handshake_ack, toep);
779 }
780 
781 void
782 tls_stop_handshake_timer(struct toepcb *toep)
783 {
784 	struct tls_ofld_info *tls_ofld = &toep->tls;
785 
786 	INP_WLOCK_ASSERT(toep->inp);
787 	callout_stop(&tls_ofld->handshake_timer);
788 }
789 
790 int
791 t4_ctloutput_tls(struct socket *so, struct sockopt *sopt)
792 {
793 	struct tls_key_context uk_ctx;
794 	struct inpcb *inp;
795 	struct tcpcb *tp;
796 	struct toepcb *toep;
797 	int error, optval;
798 
799 	error = 0;
800 	if (sopt->sopt_dir == SOPT_SET &&
801 	    sopt->sopt_name == TCP_TLSOM_SET_TLS_CONTEXT) {
802 		error = sooptcopyin(sopt, &uk_ctx, sizeof(uk_ctx),
803 		    sizeof(uk_ctx));
804 		if (error)
805 			return (error);
806 	}
807 
808 	inp = sotoinpcb(so);
809 	KASSERT(inp != NULL, ("tcp_ctloutput: inp == NULL"));
810 	INP_WLOCK(inp);
811 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
812 		INP_WUNLOCK(inp);
813 		return (ECONNRESET);
814 	}
815 	tp = intotcpcb(inp);
816 	toep = tp->t_toe;
817 	switch (sopt->sopt_dir) {
818 	case SOPT_SET:
819 		switch (sopt->sopt_name) {
820 		case TCP_TLSOM_SET_TLS_CONTEXT:
821 			if (toep->tls.mode == TLS_MODE_KTLS)
822 				error = EINVAL;
823 			else {
824 				error = program_key_context(tp, toep, &uk_ctx);
825 				if (error == 0)
826 					toep->tls.mode = TLS_MODE_TLSOM;
827 			}
828 			INP_WUNLOCK(inp);
829 			break;
830 		case TCP_TLSOM_CLR_TLS_TOM:
831 			if (toep->tls.mode == TLS_MODE_KTLS)
832 				error = EINVAL;
833 			else if (ulp_mode(toep) == ULP_MODE_TLS) {
834 				CTR2(KTR_CXGBE, "%s: tid %d CLR_TLS_TOM",
835 				    __func__, toep->tid);
836 				tls_clr_ofld_mode(toep);
837 			} else
838 				error = EOPNOTSUPP;
839 			INP_WUNLOCK(inp);
840 			break;
841 		case TCP_TLSOM_CLR_QUIES:
842 			if (toep->tls.mode == TLS_MODE_KTLS)
843 				error = EINVAL;
844 			else if (ulp_mode(toep) == ULP_MODE_TLS) {
845 				CTR2(KTR_CXGBE, "%s: tid %d CLR_QUIES",
846 				    __func__, toep->tid);
847 				tls_clr_quiesce(toep);
848 			} else
849 				error = EOPNOTSUPP;
850 			INP_WUNLOCK(inp);
851 			break;
852 		default:
853 			INP_WUNLOCK(inp);
854 			error = EOPNOTSUPP;
855 			break;
856 		}
857 		break;
858 	case SOPT_GET:
859 		switch (sopt->sopt_name) {
860 		case TCP_TLSOM_GET_TLS_TOM:
861 			/*
862 			 * TLS TX is permitted on any TOE socket, but
863 			 * TLS RX requires a TLS ULP mode.
864 			 */
865 			optval = TLS_TOM_NONE;
866 			if (can_tls_offload(td_adapter(toep->td)) &&
867 			    toep->tls.mode != TLS_MODE_KTLS) {
868 				switch (ulp_mode(toep)) {
869 				case ULP_MODE_NONE:
870 				case ULP_MODE_TCPDDP:
871 					optval = TLS_TOM_TXONLY;
872 					break;
873 				case ULP_MODE_TLS:
874 					optval = TLS_TOM_BOTH;
875 					break;
876 				}
877 			}
878 			CTR3(KTR_CXGBE, "%s: tid %d GET_TLS_TOM = %d",
879 			    __func__, toep->tid, optval);
880 			INP_WUNLOCK(inp);
881 			error = sooptcopyout(sopt, &optval, sizeof(optval));
882 			break;
883 		default:
884 			INP_WUNLOCK(inp);
885 			error = EOPNOTSUPP;
886 			break;
887 		}
888 		break;
889 	}
890 	return (error);
891 }
892 
893 #ifdef KERN_TLS
894 static void
895 init_ktls_key_context(struct ktls_session *tls, struct tls_key_context *k_ctx,
896     int direction)
897 {
898 	struct auth_hash *axf;
899 	u_int key_info_size, mac_key_size;
900 	char *hash, *key;
901 
902 	k_ctx->l_p_key = V_KEY_GET_LOC(direction == KTLS_TX ? KEY_WRITE_TX :
903 	    KEY_WRITE_RX);
904 	k_ctx->proto_ver = tls->params.tls_vmajor << 8 | tls->params.tls_vminor;
905 	k_ctx->cipher_secret_size = tls->params.cipher_key_len;
906 	key_info_size = sizeof(struct tx_keyctx_hdr) +
907 	    k_ctx->cipher_secret_size;
908 	if (direction == KTLS_TX)
909 		key = k_ctx->tx.key;
910 	else
911 		key = k_ctx->rx.key;
912 	memcpy(key, tls->params.cipher_key, tls->params.cipher_key_len);
913 	hash = key + tls->params.cipher_key_len;
914 	if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16) {
915 		k_ctx->state.auth_mode = SCMD_AUTH_MODE_GHASH;
916 		k_ctx->state.enc_mode = SCMD_CIPH_MODE_AES_GCM;
917 		k_ctx->iv_size = 4;
918 		k_ctx->mac_first = 0;
919 		k_ctx->hmac_ctrl = SCMD_HMAC_CTRL_NOP;
920 		key_info_size += GMAC_BLOCK_LEN;
921 		k_ctx->mac_secret_size = 0;
922 		if (direction == KTLS_TX)
923 			memcpy(k_ctx->tx.salt, tls->params.iv, SALT_SIZE);
924 		else
925 			memcpy(k_ctx->rx.salt, tls->params.iv, SALT_SIZE);
926 		t4_init_gmac_hash(tls->params.cipher_key,
927 		    tls->params.cipher_key_len, hash);
928 	} else {
929 		switch (tls->params.auth_algorithm) {
930 		case CRYPTO_SHA1_HMAC:
931 			axf = &auth_hash_hmac_sha1;
932 			mac_key_size = SHA1_HASH_LEN;
933 			k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA1;
934 			break;
935 		case CRYPTO_SHA2_256_HMAC:
936 			axf = &auth_hash_hmac_sha2_256;
937 			mac_key_size = SHA2_256_HASH_LEN;
938 			k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA256;
939 			break;
940 		case CRYPTO_SHA2_384_HMAC:
941 			axf = &auth_hash_hmac_sha2_384;
942 			mac_key_size = SHA2_512_HASH_LEN;
943 			k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA512_384;
944 			break;
945 		default:
946 			panic("bad auth mode");
947 		}
948 		k_ctx->state.enc_mode = SCMD_CIPH_MODE_AES_CBC;
949 		k_ctx->iv_size = 8; /* for CBC, iv is 16B, unit of 2B */
950 		k_ctx->mac_first = 1;
951 		k_ctx->hmac_ctrl = SCMD_HMAC_CTRL_NO_TRUNC;
952 		key_info_size += roundup2(mac_key_size, 16) * 2;
953 		k_ctx->mac_secret_size = mac_key_size;
954 		t4_init_hmac_digest(axf, mac_key_size, tls->params.auth_key,
955 		    tls->params.auth_key_len, hash);
956 	}
957 
958 	if (direction == KTLS_TX)
959 		k_ctx->tx_key_info_size = key_info_size;
960 	else
961 		k_ctx->rx_key_info_size = key_info_size;
962 	k_ctx->frag_size = tls->params.max_frame_len;
963 	k_ctx->iv_ctrl = 1;
964 }
965 
966 int
967 tls_alloc_ktls(struct toepcb *toep, struct ktls_session *tls, int direction)
968 {
969 	struct adapter *sc = td_adapter(toep->td);
970 	struct tls_key_context *k_ctx;
971 	int error, key_offset;
972 
973 	if (toep->tls.mode == TLS_MODE_TLSOM)
974 		return (EINVAL);
975 	if (!can_tls_offload(td_adapter(toep->td)))
976 		return (EINVAL);
977 	switch (ulp_mode(toep)) {
978 	case ULP_MODE_TLS:
979 		break;
980 	case ULP_MODE_NONE:
981 	case ULP_MODE_TCPDDP:
982 		if (direction != KTLS_TX)
983 			return (EINVAL);
984 		break;
985 	default:
986 		return (EINVAL);
987 	}
988 
989 	switch (tls->params.cipher_algorithm) {
990 	case CRYPTO_AES_CBC:
991 		/* XXX: Explicitly ignore any provided IV. */
992 		switch (tls->params.cipher_key_len) {
993 		case 128 / 8:
994 		case 192 / 8:
995 		case 256 / 8:
996 			break;
997 		default:
998 			error = EINVAL;
999 			goto clr_ofld;
1000 		}
1001 		switch (tls->params.auth_algorithm) {
1002 		case CRYPTO_SHA1_HMAC:
1003 		case CRYPTO_SHA2_256_HMAC:
1004 		case CRYPTO_SHA2_384_HMAC:
1005 			break;
1006 		default:
1007 			error = EPROTONOSUPPORT;
1008 			goto clr_ofld;
1009 		}
1010 		break;
1011 	case CRYPTO_AES_NIST_GCM_16:
1012 		if (tls->params.iv_len != SALT_SIZE) {
1013 			error = EINVAL;
1014 			goto clr_ofld;
1015 		}
1016 		switch (tls->params.cipher_key_len) {
1017 		case 128 / 8:
1018 		case 192 / 8:
1019 		case 256 / 8:
1020 			break;
1021 		default:
1022 			error = EINVAL;
1023 			goto clr_ofld;
1024 		}
1025 		break;
1026 	default:
1027 		error = EPROTONOSUPPORT;
1028 		goto clr_ofld;
1029 	}
1030 
1031 	/* Only TLS 1.1 and TLS 1.2 are currently supported. */
1032 	if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE ||
1033 	    tls->params.tls_vminor < TLS_MINOR_VER_ONE ||
1034 	    tls->params.tls_vminor > TLS_MINOR_VER_TWO) {
1035 		error = EPROTONOSUPPORT;
1036 		goto clr_ofld;
1037 	}
1038 
1039 	/* Bail if we already have a key. */
1040 	if (direction == KTLS_TX) {
1041 		if (toep->tls.tx_key_addr != -1)
1042 			return (EOPNOTSUPP);
1043 	} else {
1044 		if (toep->tls.rx_key_addr != -1)
1045 			return (EOPNOTSUPP);
1046 	}
1047 
1048 	/*
1049 	 * XXX: This assumes no key renegotation.  If KTLS ever supports
1050 	 * that we will want to allocate TLS sessions dynamically rather
1051 	 * than as a static member of toep.
1052 	 */
1053 	k_ctx = &toep->tls.k_ctx;
1054 	init_ktls_key_context(tls, k_ctx, direction);
1055 
1056 	error = tls_program_key_id(toep, k_ctx);
1057 	if (error) {
1058 		if (direction == KTLS_RX)
1059 			goto clr_ofld;
1060 		return (error);
1061 	}
1062 
1063 	if (direction == KTLS_TX) {
1064 		toep->tls.scmd0.seqno_numivs =
1065 			(V_SCMD_SEQ_NO_CTRL(3) |
1066 			 V_SCMD_PROTO_VERSION(get_proto_ver(k_ctx->proto_ver)) |
1067 			 V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
1068 			 V_SCMD_CIPH_AUTH_SEQ_CTRL((k_ctx->mac_first == 0)) |
1069 			 V_SCMD_CIPH_MODE(k_ctx->state.enc_mode) |
1070 			 V_SCMD_AUTH_MODE(k_ctx->state.auth_mode) |
1071 			 V_SCMD_HMAC_CTRL(k_ctx->hmac_ctrl) |
1072 			 V_SCMD_IV_SIZE(k_ctx->iv_size));
1073 
1074 		toep->tls.scmd0.ivgen_hdrlen =
1075 			(V_SCMD_IV_GEN_CTRL(k_ctx->iv_ctrl) |
1076 			 V_SCMD_KEY_CTX_INLINE(0) |
1077 			 V_SCMD_TLS_FRAG_ENABLE(1));
1078 
1079 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16)
1080 			toep->tls.iv_len = 8;
1081 		else
1082 			toep->tls.iv_len = AES_BLOCK_LEN;
1083 
1084 		toep->tls.mac_length = k_ctx->mac_secret_size;
1085 
1086 		toep->tls.fcplenmax = get_tp_plen_max(&toep->tls);
1087 		toep->tls.expn_per_ulp = tls->params.tls_hlen +
1088 		    tls->params.tls_tlen;
1089 		toep->tls.pdus_per_ulp = 1;
1090 		toep->tls.adjusted_plen = toep->tls.expn_per_ulp +
1091 		    toep->tls.k_ctx.frag_size;
1092 	} else {
1093 		/* Stop timer on handshake completion */
1094 		tls_stop_handshake_timer(toep);
1095 
1096 		toep->flags &= ~TPF_FORCE_CREDITS;
1097 		toep->flags |= TPF_TLS_RECEIVE;
1098 
1099 		/*
1100 		 * RX key tags are an index into the key portion of MA
1101 		 * memory stored as an offset from the base address in
1102 		 * units of 64 bytes.
1103 		 */
1104 		key_offset = toep->tls.rx_key_addr - sc->vres.key.start;
1105 		t4_set_tls_keyid(toep, key_offset / 64);
1106 		t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
1107 				 V_TCB_ULP_RAW(M_TCB_ULP_RAW),
1108 				 V_TCB_ULP_RAW((V_TF_TLS_KEY_SIZE(3) |
1109 						V_TF_TLS_CONTROL(1) |
1110 						V_TF_TLS_ACTIVE(1) |
1111 						V_TF_TLS_ENABLE(1))));
1112 		t4_set_tls_tcb_field(toep, W_TCB_TLS_SEQ,
1113 				 V_TCB_TLS_SEQ(M_TCB_TLS_SEQ),
1114 				 V_TCB_TLS_SEQ(0));
1115 		t4_clear_rx_quiesce(toep);
1116 	}
1117 
1118 	toep->tls.mode = TLS_MODE_KTLS;
1119 
1120 	return (0);
1121 
1122 clr_ofld:
1123 	if (ulp_mode(toep) == ULP_MODE_TLS) {
1124 		CTR2(KTR_CXGBE, "%s: tid %d clr_ofld_mode", __func__,
1125 		    toep->tid);
1126 		tls_clr_ofld_mode(toep);
1127 	}
1128 	return (error);
1129 }
1130 #endif
1131 
1132 void
1133 tls_init_toep(struct toepcb *toep)
1134 {
1135 	struct tls_ofld_info *tls_ofld = &toep->tls;
1136 
1137 	tls_ofld->mode = TLS_MODE_OFF;
1138 	tls_ofld->key_location = TLS_SFO_WR_CONTEXTLOC_DDR;
1139 	tls_ofld->rx_key_addr = -1;
1140 	tls_ofld->tx_key_addr = -1;
1141 }
1142 
1143 void
1144 tls_establish(struct toepcb *toep)
1145 {
1146 
1147 	/*
1148 	 * Enable PDU extraction.
1149 	 *
1150 	 * XXX: Supposedly this should be done by the firmware when
1151 	 * the ULP_MODE FLOWC parameter is set in send_flowc_wr(), but
1152 	 * in practice this seems to be required.
1153 	 */
1154 	CTR2(KTR_CXGBE, "%s: tid %d setting TLS_ENABLE", __func__, toep->tid);
1155 	t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW, V_TCB_ULP_RAW(M_TCB_ULP_RAW),
1156 	    V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)));
1157 
1158 	toep->flags |= TPF_FORCE_CREDITS | TPF_TLS_ESTABLISHED;
1159 
1160 	callout_init_rw(&toep->tls.handshake_timer, &toep->inp->inp_lock, 0);
1161 	tls_start_handshake_timer(toep);
1162 }
1163 
1164 void
1165 tls_detach(struct toepcb *toep)
1166 {
1167 
1168 	if (toep->flags & TPF_TLS_ESTABLISHED) {
1169 		tls_stop_handshake_timer(toep);
1170 		toep->flags &= ~TPF_TLS_ESTABLISHED;
1171 	}
1172 }
1173 
1174 void
1175 tls_uninit_toep(struct toepcb *toep)
1176 {
1177 
1178 	MPASS((toep->flags & TPF_TLS_ESTABLISHED) == 0);
1179 	clear_tls_keyid(toep);
1180 }
1181 
1182 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
1183 #define	MIN_OFLD_TLSTX_CREDITS(toep)					\
1184 	(howmany(sizeof(struct fw_tlstx_data_wr) +			\
1185 	    sizeof(struct cpl_tx_tls_sfo) + key_size((toep)) +		\
1186 	    CIPHER_BLOCK_SIZE + 1, 16))
1187 
1188 static inline u_int
1189 max_imm_tls_space(int tx_credits)
1190 {
1191 	const int n = 2;	/* Use only up to 2 desc for imm. data WR */
1192 	int space;
1193 
1194 	KASSERT(tx_credits >= 0 &&
1195 		tx_credits <= MAX_OFLD_TX_CREDITS,
1196 		("%s: %d credits", __func__, tx_credits));
1197 
1198 	if (tx_credits >= (n * EQ_ESIZE) / 16)
1199 		space = (n * EQ_ESIZE);
1200 	else
1201 		space = tx_credits * 16;
1202 	return (space);
1203 }
1204 
1205 static int
1206 count_mbuf_segs(struct mbuf *m, int skip, int len, int *max_nsegs_1mbufp)
1207 {
1208 	int max_nsegs_1mbuf, n, nsegs;
1209 
1210 	while (skip >= m->m_len) {
1211 		skip -= m->m_len;
1212 		m = m->m_next;
1213 	}
1214 
1215 	nsegs = 0;
1216 	max_nsegs_1mbuf = 0;
1217 	while (len > 0) {
1218 		n = sglist_count(mtod(m, char *) + skip, m->m_len - skip);
1219 		if (n > max_nsegs_1mbuf)
1220 			max_nsegs_1mbuf = n;
1221 		nsegs += n;
1222 		len -= m->m_len - skip;
1223 		skip = 0;
1224 		m = m->m_next;
1225 	}
1226 	*max_nsegs_1mbufp = max_nsegs_1mbuf;
1227 	return (nsegs);
1228 }
1229 
1230 static void
1231 write_tlstx_wr(struct fw_tlstx_data_wr *txwr, struct toepcb *toep,
1232     unsigned int immdlen, unsigned int plen, unsigned int expn,
1233     unsigned int pdus, uint8_t credits, int shove, int imm_ivs)
1234 {
1235 	struct tls_ofld_info *tls_ofld = &toep->tls;
1236 	unsigned int len = plen + expn;
1237 
1238 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_TLSTX_DATA_WR) |
1239 	    V_FW_TLSTX_DATA_WR_COMPL(1) |
1240 	    V_FW_TLSTX_DATA_WR_IMMDLEN(immdlen));
1241 	txwr->flowid_len16 = htobe32(V_FW_TLSTX_DATA_WR_FLOWID(toep->tid) |
1242 	    V_FW_TLSTX_DATA_WR_LEN16(credits));
1243 	txwr->plen = htobe32(len);
1244 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ULP_MODE_TLS) |
1245 	    V_TX_URG(0) | /* F_T6_TX_FORCE | */ V_TX_SHOVE(shove));
1246 	txwr->ctxloc_to_exp = htobe32(V_FW_TLSTX_DATA_WR_NUMIVS(pdus) |
1247 	    V_FW_TLSTX_DATA_WR_EXP(expn) |
1248 	    V_FW_TLSTX_DATA_WR_CTXLOC(tls_ofld->key_location) |
1249 	    V_FW_TLSTX_DATA_WR_IVDSGL(!imm_ivs) |
1250 	    V_FW_TLSTX_DATA_WR_KEYSIZE(tls_ofld->k_ctx.tx_key_info_size >> 4));
1251 	txwr->mfs = htobe16(tls_ofld->k_ctx.frag_size);
1252 	txwr->adjustedplen_pkd = htobe16(
1253 	    V_FW_TLSTX_DATA_WR_ADJUSTEDPLEN(tls_ofld->adjusted_plen));
1254 	txwr->expinplenmax_pkd = htobe16(
1255 	    V_FW_TLSTX_DATA_WR_EXPINPLENMAX(tls_ofld->expn_per_ulp));
1256 	txwr->pdusinplenmax_pkd =
1257 	    V_FW_TLSTX_DATA_WR_PDUSINPLENMAX(tls_ofld->pdus_per_ulp);
1258 }
1259 
1260 static void
1261 write_tlstx_cpl(struct cpl_tx_tls_sfo *cpl, struct toepcb *toep,
1262     struct tls_hdr *tls_hdr, unsigned int plen, unsigned int pdus)
1263 {
1264 	struct tls_ofld_info *tls_ofld = &toep->tls;
1265 	int data_type, seglen;
1266 
1267 	if (plen < tls_ofld->k_ctx.frag_size)
1268 		seglen = plen;
1269 	else
1270 		seglen = tls_ofld->k_ctx.frag_size;
1271 	data_type = tls_content_type(tls_hdr->type);
1272 	cpl->op_to_seg_len = htobe32(V_CPL_TX_TLS_SFO_OPCODE(CPL_TX_TLS_SFO) |
1273 	    V_CPL_TX_TLS_SFO_DATA_TYPE(data_type) |
1274 	    V_CPL_TX_TLS_SFO_CPL_LEN(2) | V_CPL_TX_TLS_SFO_SEG_LEN(seglen));
1275 	cpl->pld_len = htobe32(plen);
1276 	if (data_type == CPL_TX_TLS_SFO_TYPE_HEARTBEAT)
1277 		cpl->type_protover = htobe32(
1278 		    V_CPL_TX_TLS_SFO_TYPE(tls_hdr->type));
1279 	cpl->seqno_numivs = htobe32(tls_ofld->scmd0.seqno_numivs |
1280 	    V_SCMD_NUM_IVS(pdus));
1281 	cpl->ivgen_hdrlen = htobe32(tls_ofld->scmd0.ivgen_hdrlen);
1282 	cpl->scmd1 = htobe64(tls_ofld->tx_seq_no);
1283 	tls_ofld->tx_seq_no += pdus;
1284 }
1285 
1286 /*
1287  * Similar to write_tx_sgl() except that it accepts an optional
1288  * trailer buffer for IVs.
1289  */
1290 static void
1291 write_tlstx_sgl(void *dst, struct mbuf *start, int skip, int plen,
1292     void *iv_buffer, int iv_len, int nsegs, int n)
1293 {
1294 	struct mbuf *m;
1295 	struct ulptx_sgl *usgl = dst;
1296 	int i, j, rc;
1297 	struct sglist sg;
1298 	struct sglist_seg segs[n];
1299 
1300 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
1301 
1302 	sglist_init(&sg, n, segs);
1303 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1304 	    V_ULPTX_NSGE(nsegs));
1305 
1306 	for (m = start; skip >= m->m_len; m = m->m_next)
1307 		skip -= m->m_len;
1308 
1309 	i = -1;
1310 	for (m = start; plen > 0; m = m->m_next) {
1311 		rc = sglist_append(&sg, mtod(m, char *) + skip,
1312 		    m->m_len - skip);
1313 		if (__predict_false(rc != 0))
1314 			panic("%s: sglist_append %d", __func__, rc);
1315 		plen -= m->m_len - skip;
1316 		skip = 0;
1317 
1318 		for (j = 0; j < sg.sg_nseg; i++, j++) {
1319 			if (i < 0) {
1320 				usgl->len0 = htobe32(segs[j].ss_len);
1321 				usgl->addr0 = htobe64(segs[j].ss_paddr);
1322 			} else {
1323 				usgl->sge[i / 2].len[i & 1] =
1324 				    htobe32(segs[j].ss_len);
1325 				usgl->sge[i / 2].addr[i & 1] =
1326 				    htobe64(segs[j].ss_paddr);
1327 			}
1328 #ifdef INVARIANTS
1329 			nsegs--;
1330 #endif
1331 		}
1332 		sglist_reset(&sg);
1333 	}
1334 	if (iv_buffer != NULL) {
1335 		rc = sglist_append(&sg, iv_buffer, iv_len);
1336 		if (__predict_false(rc != 0))
1337 			panic("%s: sglist_append %d", __func__, rc);
1338 
1339 		for (j = 0; j < sg.sg_nseg; i++, j++) {
1340 			if (i < 0) {
1341 				usgl->len0 = htobe32(segs[j].ss_len);
1342 				usgl->addr0 = htobe64(segs[j].ss_paddr);
1343 			} else {
1344 				usgl->sge[i / 2].len[i & 1] =
1345 				    htobe32(segs[j].ss_len);
1346 				usgl->sge[i / 2].addr[i & 1] =
1347 				    htobe64(segs[j].ss_paddr);
1348 			}
1349 #ifdef INVARIANTS
1350 			nsegs--;
1351 #endif
1352 		}
1353 	}
1354 	if (i & 1)
1355 		usgl->sge[i / 2].len[1] = htobe32(0);
1356 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, iv_buffer %p",
1357 	    __func__, nsegs, start, iv_buffer));
1358 }
1359 
1360 /*
1361  * Similar to t4_push_frames() but handles TLS sockets when TLS offload
1362  * is enabled.  Rather than transmitting bulk data, the socket buffer
1363  * contains TLS records.  The work request requires a full TLS record,
1364  * so batch mbufs up until a full TLS record is seen.  This requires
1365  * reading the TLS header out of the start of each record to determine
1366  * its length.
1367  */
1368 void
1369 t4_push_tls_records(struct adapter *sc, struct toepcb *toep, int drop)
1370 {
1371 	struct tls_hdr thdr;
1372 	struct mbuf *sndptr;
1373 	struct fw_tlstx_data_wr *txwr;
1374 	struct cpl_tx_tls_sfo *cpl;
1375 	struct wrqe *wr;
1376 	u_int plen, nsegs, credits, space, max_nsegs_1mbuf, wr_len;
1377 	u_int expn_size, iv_len, pdus, sndptroff;
1378 	struct tls_ofld_info *tls_ofld = &toep->tls;
1379 	struct inpcb *inp = toep->inp;
1380 	struct tcpcb *tp = intotcpcb(inp);
1381 	struct socket *so = inp->inp_socket;
1382 	struct sockbuf *sb = &so->so_snd;
1383 	int tls_size, tx_credits, shove, /* compl,*/ sowwakeup;
1384 	struct ofld_tx_sdesc *txsd;
1385 	bool imm_ivs, imm_payload;
1386 	void *iv_buffer, *iv_dst, *buf;
1387 
1388 	INP_WLOCK_ASSERT(inp);
1389 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1390 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1391 
1392 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
1393 	    ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS,
1394 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1395 	KASSERT(tls_tx_key(toep),
1396 	    ("%s: TX key not set for toep %p", __func__, toep));
1397 
1398 #ifdef VERBOSE_TRACES
1399 	CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
1400 	    __func__, toep->tid, toep->flags, tp->t_flags);
1401 #endif
1402 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1403 		return;
1404 
1405 #ifdef RATELIMIT
1406 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
1407 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
1408 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
1409 	}
1410 #endif
1411 
1412 	/*
1413 	 * This function doesn't resume by itself.  Someone else must clear the
1414 	 * flag and call this function.
1415 	 */
1416 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1417 		KASSERT(drop == 0,
1418 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1419 		return;
1420 	}
1421 
1422 	txsd = &toep->txsd[toep->txsd_pidx];
1423 	for (;;) {
1424 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1425 		space = max_imm_tls_space(tx_credits);
1426 		wr_len = sizeof(struct fw_tlstx_data_wr) +
1427 		    sizeof(struct cpl_tx_tls_sfo) + key_size(toep);
1428 		if (wr_len + CIPHER_BLOCK_SIZE + 1 > space) {
1429 #ifdef VERBOSE_TRACES
1430 			CTR5(KTR_CXGBE,
1431 			    "%s: tid %d tx_credits %d min_wr %d space %d",
1432 			    __func__, toep->tid, tx_credits, wr_len +
1433 			    CIPHER_BLOCK_SIZE + 1, space);
1434 #endif
1435 			return;
1436 		}
1437 
1438 		SOCKBUF_LOCK(sb);
1439 		sowwakeup = drop;
1440 		if (drop) {
1441 			sbdrop_locked(sb, drop);
1442 			MPASS(tls_ofld->sb_off >= drop);
1443 			tls_ofld->sb_off -= drop;
1444 			drop = 0;
1445 		}
1446 
1447 		/*
1448 		 * Send a FIN if requested, but only if there's no
1449 		 * more data to send.
1450 		 */
1451 		if (sbavail(sb) == tls_ofld->sb_off &&
1452 		    toep->flags & TPF_SEND_FIN) {
1453 			if (sowwakeup)
1454 				sowwakeup_locked(so);
1455 			else
1456 				SOCKBUF_UNLOCK(sb);
1457 			SOCKBUF_UNLOCK_ASSERT(sb);
1458 			t4_close_conn(sc, toep);
1459 			return;
1460 		}
1461 
1462 		if (sbavail(sb) < tls_ofld->sb_off + TLS_HEADER_LENGTH) {
1463 			/*
1464 			 * A full TLS header is not yet queued, stop
1465 			 * for now until more data is added to the
1466 			 * socket buffer.  However, if the connection
1467 			 * has been closed, we will never get the rest
1468 			 * of the header so just discard the partial
1469 			 * header and close the connection.
1470 			 */
1471 #ifdef VERBOSE_TRACES
1472 			CTR5(KTR_CXGBE, "%s: tid %d sbavail %d sb_off %d%s",
1473 			    __func__, toep->tid, sbavail(sb), tls_ofld->sb_off,
1474 			    toep->flags & TPF_SEND_FIN ? "" : " SEND_FIN");
1475 #endif
1476 			if (sowwakeup)
1477 				sowwakeup_locked(so);
1478 			else
1479 				SOCKBUF_UNLOCK(sb);
1480 			SOCKBUF_UNLOCK_ASSERT(sb);
1481 			if (toep->flags & TPF_SEND_FIN)
1482 				t4_close_conn(sc, toep);
1483 			return;
1484 		}
1485 
1486 		/* Read the header of the next TLS record. */
1487 		sndptr = sbsndmbuf(sb, tls_ofld->sb_off, &sndptroff);
1488 		m_copydata(sndptr, sndptroff, sizeof(thdr), (caddr_t)&thdr);
1489 		tls_size = htons(thdr.length);
1490 		plen = TLS_HEADER_LENGTH + tls_size;
1491 		pdus = howmany(tls_size, tls_ofld->k_ctx.frag_size);
1492 		iv_len = pdus * CIPHER_BLOCK_SIZE;
1493 
1494 		if (sbavail(sb) < tls_ofld->sb_off + plen) {
1495 			/*
1496 			 * The full TLS record is not yet queued, stop
1497 			 * for now until more data is added to the
1498 			 * socket buffer.  However, if the connection
1499 			 * has been closed, we will never get the rest
1500 			 * of the record so just discard the partial
1501 			 * record and close the connection.
1502 			 */
1503 #ifdef VERBOSE_TRACES
1504 			CTR6(KTR_CXGBE,
1505 			    "%s: tid %d sbavail %d sb_off %d plen %d%s",
1506 			    __func__, toep->tid, sbavail(sb), tls_ofld->sb_off,
1507 			    plen, toep->flags & TPF_SEND_FIN ? "" :
1508 			    " SEND_FIN");
1509 #endif
1510 			if (sowwakeup)
1511 				sowwakeup_locked(so);
1512 			else
1513 				SOCKBUF_UNLOCK(sb);
1514 			SOCKBUF_UNLOCK_ASSERT(sb);
1515 			if (toep->flags & TPF_SEND_FIN)
1516 				t4_close_conn(sc, toep);
1517 			return;
1518 		}
1519 
1520 		/* Shove if there is no additional data pending. */
1521 		shove = (sbavail(sb) == tls_ofld->sb_off + plen) &&
1522 		    !(tp->t_flags & TF_MORETOCOME);
1523 
1524 		if (sb->sb_flags & SB_AUTOSIZE &&
1525 		    V_tcp_do_autosndbuf &&
1526 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
1527 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
1528 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
1529 			    V_tcp_autosndbuf_max);
1530 
1531 			if (!sbreserve_locked(sb, newsize, so, NULL))
1532 				sb->sb_flags &= ~SB_AUTOSIZE;
1533 			else
1534 				sowwakeup = 1;	/* room available */
1535 		}
1536 		if (sowwakeup)
1537 			sowwakeup_locked(so);
1538 		else
1539 			SOCKBUF_UNLOCK(sb);
1540 		SOCKBUF_UNLOCK_ASSERT(sb);
1541 
1542 		if (__predict_false(toep->flags & TPF_FIN_SENT))
1543 			panic("%s: excess tx.", __func__);
1544 
1545 		/* Determine whether to use immediate vs SGL. */
1546 		imm_payload = false;
1547 		imm_ivs = false;
1548 		if (wr_len + iv_len <= space) {
1549 			imm_ivs = true;
1550 			wr_len += iv_len;
1551 			if (wr_len + tls_size <= space) {
1552 				wr_len += tls_size;
1553 				imm_payload = true;
1554 			}
1555 		}
1556 
1557 		/* Allocate space for IVs if needed. */
1558 		if (!imm_ivs) {
1559 			iv_buffer = malloc(iv_len, M_CXGBE, M_NOWAIT);
1560 			if (iv_buffer == NULL) {
1561 				/*
1562 				 * XXX: How to restart this?
1563 				 */
1564 				if (sowwakeup)
1565 					sowwakeup_locked(so);
1566 				else
1567 					SOCKBUF_UNLOCK(sb);
1568 				SOCKBUF_UNLOCK_ASSERT(sb);
1569 				CTR3(KTR_CXGBE,
1570 			    "%s: tid %d failed to alloc IV space len %d",
1571 				    __func__, toep->tid, iv_len);
1572 				return;
1573 			}
1574 		} else
1575 			iv_buffer = NULL;
1576 
1577 		/* Determine size of SGL. */
1578 		nsegs = 0;
1579 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1580 		if (!imm_payload) {
1581 			nsegs = count_mbuf_segs(sndptr, sndptroff +
1582 			    TLS_HEADER_LENGTH, tls_size, &max_nsegs_1mbuf);
1583 			if (!imm_ivs) {
1584 				int n = sglist_count(iv_buffer, iv_len);
1585 				nsegs += n;
1586 				if (n > max_nsegs_1mbuf)
1587 					max_nsegs_1mbuf = n;
1588 			}
1589 
1590 			/* Account for SGL in work request length. */
1591 			wr_len += sizeof(struct ulptx_sgl) +
1592 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1593 		}
1594 
1595 		wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1596 		if (wr == NULL) {
1597 			/* XXX: how will we recover from this? */
1598 			toep->flags |= TPF_TX_SUSPENDED;
1599 			return;
1600 		}
1601 
1602 #ifdef VERBOSE_TRACES
1603 		CTR5(KTR_CXGBE, "%s: tid %d TLS record %d len %#x pdus %d",
1604 		    __func__, toep->tid, thdr.type, tls_size, pdus);
1605 #endif
1606 		txwr = wrtod(wr);
1607 		cpl = (struct cpl_tx_tls_sfo *)(txwr + 1);
1608 		memset(txwr, 0, roundup2(wr_len, 16));
1609 		credits = howmany(wr_len, 16);
1610 		expn_size = tls_expansion_size(toep, tls_size, 0, NULL);
1611 		write_tlstx_wr(txwr, toep, imm_payload ? tls_size : 0,
1612 		    tls_size, expn_size, pdus, credits, shove, imm_ivs ? 1 : 0);
1613 		write_tlstx_cpl(cpl, toep, &thdr, tls_size, pdus);
1614 		tls_copy_tx_key(toep, cpl + 1);
1615 
1616 		/* Generate random IVs */
1617 		buf = (char *)(cpl + 1) + key_size(toep);
1618 		if (imm_ivs) {
1619 			MPASS(iv_buffer == NULL);
1620 			iv_dst = buf;
1621 			buf = (char *)iv_dst + iv_len;
1622 		} else
1623 			iv_dst = iv_buffer;
1624 		arc4rand(iv_dst, iv_len, 0);
1625 
1626 		if (imm_payload) {
1627 			m_copydata(sndptr, sndptroff + TLS_HEADER_LENGTH,
1628 			    tls_size, buf);
1629 		} else {
1630 			write_tlstx_sgl(buf, sndptr,
1631 			    sndptroff + TLS_HEADER_LENGTH, tls_size, iv_buffer,
1632 			    iv_len, nsegs, max_nsegs_1mbuf);
1633 		}
1634 
1635 		KASSERT(toep->tx_credits >= credits,
1636 			("%s: not enough credits", __func__));
1637 
1638 		toep->tx_credits -= credits;
1639 
1640 		tp->snd_nxt += plen;
1641 		tp->snd_max += plen;
1642 
1643 		SOCKBUF_LOCK(sb);
1644 		sbsndptr_adv(sb, sb->sb_sndptr, plen);
1645 		tls_ofld->sb_off += plen;
1646 		SOCKBUF_UNLOCK(sb);
1647 
1648 		toep->flags |= TPF_TX_DATA_SENT;
1649 		if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep))
1650 			toep->flags |= TPF_TX_SUSPENDED;
1651 
1652 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1653 		txsd->plen = plen;
1654 		txsd->tx_credits = credits;
1655 		txsd->iv_buffer = iv_buffer;
1656 		txsd++;
1657 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1658 			toep->txsd_pidx = 0;
1659 			txsd = &toep->txsd[0];
1660 		}
1661 		toep->txsd_avail--;
1662 
1663 		atomic_add_long(&toep->vi->pi->tx_toe_tls_records, 1);
1664 		atomic_add_long(&toep->vi->pi->tx_toe_tls_octets, plen);
1665 
1666 		t4_l2t_send(sc, wr, toep->l2te);
1667 	}
1668 }
1669 
1670 #ifdef KERN_TLS
1671 static int
1672 count_ext_pgs_segs(struct mbuf *m)
1673 {
1674 	vm_paddr_t nextpa;
1675 	u_int i, nsegs;
1676 
1677 	MPASS(m->m_epg_npgs > 0);
1678 	nsegs = 1;
1679 	nextpa = m->m_epg_pa[0] + PAGE_SIZE;
1680 	for (i = 1; i < m->m_epg_npgs; i++) {
1681 		if (nextpa != m->m_epg_pa[i])
1682 			nsegs++;
1683 		nextpa = m->m_epg_pa[i] + PAGE_SIZE;
1684 	}
1685 	return (nsegs);
1686 }
1687 
1688 static void
1689 write_ktlstx_sgl(void *dst, struct mbuf *m, int nsegs)
1690 {
1691 	struct ulptx_sgl *usgl = dst;
1692 	vm_paddr_t pa;
1693 	uint32_t len;
1694 	int i, j;
1695 
1696 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
1697 
1698 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1699 	    V_ULPTX_NSGE(nsegs));
1700 
1701 	/* Figure out the first S/G length. */
1702 	pa = m->m_epg_pa[0] + m->m_epg_1st_off;
1703 	usgl->addr0 = htobe64(pa);
1704 	len = m_epg_pagelen(m, 0, m->m_epg_1st_off);
1705 	pa += len;
1706 	for (i = 1; i < m->m_epg_npgs; i++) {
1707 		if (m->m_epg_pa[i] != pa)
1708 			break;
1709 		len += m_epg_pagelen(m, i, 0);
1710 		pa += m_epg_pagelen(m, i, 0);
1711 	}
1712 	usgl->len0 = htobe32(len);
1713 #ifdef INVARIANTS
1714 	nsegs--;
1715 #endif
1716 
1717 	j = -1;
1718 	for (; i < m->m_epg_npgs; i++) {
1719 		if (j == -1 || m->m_epg_pa[i] != pa) {
1720 			if (j >= 0)
1721 				usgl->sge[j / 2].len[j & 1] = htobe32(len);
1722 			j++;
1723 #ifdef INVARIANTS
1724 			nsegs--;
1725 #endif
1726 			pa = m->m_epg_pa[i];
1727 			usgl->sge[j / 2].addr[j & 1] = htobe64(pa);
1728 			len = m_epg_pagelen(m, i, 0);
1729 			pa += len;
1730 		} else {
1731 			len += m_epg_pagelen(m, i, 0);
1732 			pa += m_epg_pagelen(m, i, 0);
1733 		}
1734 	}
1735 	if (j >= 0) {
1736 		usgl->sge[j / 2].len[j & 1] = htobe32(len);
1737 
1738 		if ((j & 1) == 0)
1739 			usgl->sge[j / 2].len[1] = htobe32(0);
1740 	}
1741 	KASSERT(nsegs == 0, ("%s: nsegs %d, m %p", __func__, nsegs, m));
1742 }
1743 
1744 /*
1745  * Similar to t4_push_frames() but handles sockets that contain TLS
1746  * record mbufs.  Unlike TLSOM, each mbuf is a complete TLS record and
1747  * corresponds to a single work request.
1748  */
1749 void
1750 t4_push_ktls(struct adapter *sc, struct toepcb *toep, int drop)
1751 {
1752 	struct tls_hdr *thdr;
1753 	struct fw_tlstx_data_wr *txwr;
1754 	struct cpl_tx_tls_sfo *cpl;
1755 	struct wrqe *wr;
1756 	struct mbuf *m;
1757 	u_int nsegs, credits, wr_len;
1758 	u_int expn_size;
1759 	struct inpcb *inp = toep->inp;
1760 	struct tcpcb *tp = intotcpcb(inp);
1761 	struct socket *so = inp->inp_socket;
1762 	struct sockbuf *sb = &so->so_snd;
1763 	int tls_size, tx_credits, shove, sowwakeup;
1764 	struct ofld_tx_sdesc *txsd;
1765 	char *buf;
1766 
1767 	INP_WLOCK_ASSERT(inp);
1768 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1769 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1770 
1771 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
1772 	    ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS,
1773 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1774 	KASSERT(tls_tx_key(toep),
1775 	    ("%s: TX key not set for toep %p", __func__, toep));
1776 
1777 #ifdef VERBOSE_TRACES
1778 	CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
1779 	    __func__, toep->tid, toep->flags, tp->t_flags);
1780 #endif
1781 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1782 		return;
1783 
1784 #ifdef RATELIMIT
1785 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
1786 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
1787 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
1788 	}
1789 #endif
1790 
1791 	/*
1792 	 * This function doesn't resume by itself.  Someone else must clear the
1793 	 * flag and call this function.
1794 	 */
1795 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1796 		KASSERT(drop == 0,
1797 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1798 		return;
1799 	}
1800 
1801 	txsd = &toep->txsd[toep->txsd_pidx];
1802 	for (;;) {
1803 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1804 
1805 		SOCKBUF_LOCK(sb);
1806 		sowwakeup = drop;
1807 		if (drop) {
1808 			sbdrop_locked(sb, drop);
1809 			drop = 0;
1810 		}
1811 
1812 		m = sb->sb_sndptr != NULL ? sb->sb_sndptr->m_next : sb->sb_mb;
1813 
1814 		/*
1815 		 * Send a FIN if requested, but only if there's no
1816 		 * more data to send.
1817 		 */
1818 		if (m == NULL && toep->flags & TPF_SEND_FIN) {
1819 			if (sowwakeup)
1820 				sowwakeup_locked(so);
1821 			else
1822 				SOCKBUF_UNLOCK(sb);
1823 			SOCKBUF_UNLOCK_ASSERT(sb);
1824 			t4_close_conn(sc, toep);
1825 			return;
1826 		}
1827 
1828 		/*
1829 		 * If there is no ready data to send, wait until more
1830 		 * data arrives.
1831 		 */
1832 		if (m == NULL || (m->m_flags & M_NOTAVAIL) != 0) {
1833 			if (sowwakeup)
1834 				sowwakeup_locked(so);
1835 			else
1836 				SOCKBUF_UNLOCK(sb);
1837 			SOCKBUF_UNLOCK_ASSERT(sb);
1838 #ifdef VERBOSE_TRACES
1839 			CTR2(KTR_CXGBE, "%s: tid %d no ready data to send",
1840 			    __func__, toep->tid);
1841 #endif
1842 			return;
1843 		}
1844 
1845 		KASSERT(m->m_flags & M_EXTPG, ("%s: mbuf %p is not NOMAP",
1846 		    __func__, m));
1847 		KASSERT(m->m_epg_tls != NULL,
1848 		    ("%s: mbuf %p doesn't have TLS session", __func__, m));
1849 
1850 		/* Calculate WR length. */
1851 		wr_len = sizeof(struct fw_tlstx_data_wr) +
1852 		    sizeof(struct cpl_tx_tls_sfo) + key_size(toep);
1853 
1854 		/* Explicit IVs for AES-CBC and AES-GCM are <= 16. */
1855 		MPASS(toep->tls.iv_len <= AES_BLOCK_LEN);
1856 		wr_len += AES_BLOCK_LEN;
1857 
1858 		/* Account for SGL in work request length. */
1859 		nsegs = count_ext_pgs_segs(m);
1860 		wr_len += sizeof(struct ulptx_sgl) +
1861 		    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1862 
1863 		/* Not enough credits for this work request. */
1864 		if (howmany(wr_len, 16) > tx_credits) {
1865 			if (sowwakeup)
1866 				sowwakeup_locked(so);
1867 			else
1868 				SOCKBUF_UNLOCK(sb);
1869 			SOCKBUF_UNLOCK_ASSERT(sb);
1870 #ifdef VERBOSE_TRACES
1871 			CTR5(KTR_CXGBE,
1872 	    "%s: tid %d mbuf %p requires %d credits, but only %d available",
1873 			    __func__, toep->tid, m, howmany(wr_len, 16),
1874 			    tx_credits);
1875 #endif
1876 			toep->flags |= TPF_TX_SUSPENDED;
1877 			return;
1878 		}
1879 
1880 		/* Shove if there is no additional data pending. */
1881 		shove = ((m->m_next == NULL ||
1882 		    (m->m_next->m_flags & M_NOTAVAIL) != 0)) &&
1883 		    (tp->t_flags & TF_MORETOCOME) == 0;
1884 
1885 		if (sb->sb_flags & SB_AUTOSIZE &&
1886 		    V_tcp_do_autosndbuf &&
1887 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
1888 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
1889 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
1890 			    V_tcp_autosndbuf_max);
1891 
1892 			if (!sbreserve_locked(sb, newsize, so, NULL))
1893 				sb->sb_flags &= ~SB_AUTOSIZE;
1894 			else
1895 				sowwakeup = 1;	/* room available */
1896 		}
1897 		if (sowwakeup)
1898 			sowwakeup_locked(so);
1899 		else
1900 			SOCKBUF_UNLOCK(sb);
1901 		SOCKBUF_UNLOCK_ASSERT(sb);
1902 
1903 		if (__predict_false(toep->flags & TPF_FIN_SENT))
1904 			panic("%s: excess tx.", __func__);
1905 
1906 		wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1907 		if (wr == NULL) {
1908 			/* XXX: how will we recover from this? */
1909 			toep->flags |= TPF_TX_SUSPENDED;
1910 			return;
1911 		}
1912 
1913 		thdr = (struct tls_hdr *)&m->m_epg_hdr;
1914 #ifdef VERBOSE_TRACES
1915 		CTR5(KTR_CXGBE, "%s: tid %d TLS record %ju type %d len %#x",
1916 		    __func__, toep->tid, m->m_epg_seqno, thdr->type,
1917 		    m->m_len);
1918 #endif
1919 		txwr = wrtod(wr);
1920 		cpl = (struct cpl_tx_tls_sfo *)(txwr + 1);
1921 		memset(txwr, 0, roundup2(wr_len, 16));
1922 		credits = howmany(wr_len, 16);
1923 		expn_size = m->m_epg_hdrlen +
1924 		    m->m_epg_trllen;
1925 		tls_size = m->m_len - expn_size;
1926 		write_tlstx_wr(txwr, toep, 0,
1927 		    tls_size, expn_size, 1, credits, shove, 1);
1928 		toep->tls.tx_seq_no = m->m_epg_seqno;
1929 		write_tlstx_cpl(cpl, toep, thdr, tls_size, 1);
1930 		tls_copy_tx_key(toep, cpl + 1);
1931 
1932 		/* Copy IV. */
1933 		buf = (char *)(cpl + 1) + key_size(toep);
1934 		memcpy(buf, thdr + 1, toep->tls.iv_len);
1935 		buf += AES_BLOCK_LEN;
1936 
1937 		write_ktlstx_sgl(buf, m, nsegs);
1938 
1939 		KASSERT(toep->tx_credits >= credits,
1940 			("%s: not enough credits", __func__));
1941 
1942 		toep->tx_credits -= credits;
1943 
1944 		tp->snd_nxt += m->m_len;
1945 		tp->snd_max += m->m_len;
1946 
1947 		SOCKBUF_LOCK(sb);
1948 		sb->sb_sndptr = m;
1949 		SOCKBUF_UNLOCK(sb);
1950 
1951 		toep->flags |= TPF_TX_DATA_SENT;
1952 		if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep))
1953 			toep->flags |= TPF_TX_SUSPENDED;
1954 
1955 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1956 		txsd->plen = m->m_len;
1957 		txsd->tx_credits = credits;
1958 		txsd++;
1959 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1960 			toep->txsd_pidx = 0;
1961 			txsd = &toep->txsd[0];
1962 		}
1963 		toep->txsd_avail--;
1964 
1965 		atomic_add_long(&toep->vi->pi->tx_toe_tls_records, 1);
1966 		atomic_add_long(&toep->vi->pi->tx_toe_tls_octets, m->m_len);
1967 
1968 		t4_l2t_send(sc, wr, toep->l2te);
1969 	}
1970 }
1971 #endif
1972 
1973 /*
1974  * For TLS data we place received mbufs received via CPL_TLS_DATA into
1975  * an mbufq in the TLS offload state.  When CPL_RX_TLS_CMP is
1976  * received, the completed PDUs are placed into the socket receive
1977  * buffer.
1978  *
1979  * The TLS code reuses the ulp_pdu_reclaimq to hold the pending mbufs.
1980  */
1981 static int
1982 do_tls_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1983 {
1984 	struct adapter *sc = iq->adapter;
1985 	const struct cpl_tls_data *cpl = mtod(m, const void *);
1986 	unsigned int tid = GET_TID(cpl);
1987 	struct toepcb *toep = lookup_tid(sc, tid);
1988 	struct inpcb *inp = toep->inp;
1989 	struct tcpcb *tp;
1990 	int len;
1991 
1992 	/* XXX: Should this match do_rx_data instead? */
1993 	KASSERT(!(toep->flags & TPF_SYNQE),
1994 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
1995 
1996 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
1997 
1998 	/* strip off CPL header */
1999 	m_adj(m, sizeof(*cpl));
2000 	len = m->m_pkthdr.len;
2001 
2002 	atomic_add_long(&toep->vi->pi->rx_toe_tls_octets, len);
2003 
2004 	KASSERT(len == G_CPL_TLS_DATA_LENGTH(be32toh(cpl->length_pkd)),
2005 	    ("%s: payload length mismatch", __func__));
2006 
2007 	INP_WLOCK(inp);
2008 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
2009 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
2010 		    __func__, tid, len, inp->inp_flags);
2011 		INP_WUNLOCK(inp);
2012 		m_freem(m);
2013 		return (0);
2014 	}
2015 
2016 	/* Save TCP sequence number. */
2017 	m->m_pkthdr.tls_tcp_seq = be32toh(cpl->seq);
2018 
2019 	if (mbufq_enqueue(&toep->ulp_pdu_reclaimq, m)) {
2020 #ifdef INVARIANTS
2021 		panic("Failed to queue TLS data packet");
2022 #else
2023 		printf("%s: Failed to queue TLS data packet\n", __func__);
2024 		INP_WUNLOCK(inp);
2025 		m_freem(m);
2026 		return (0);
2027 #endif
2028 	}
2029 
2030 	tp = intotcpcb(inp);
2031 	tp->t_rcvtime = ticks;
2032 
2033 #ifdef VERBOSE_TRACES
2034 	CTR4(KTR_CXGBE, "%s: tid %u len %d seq %u", __func__, tid, len,
2035 	    be32toh(cpl->seq));
2036 #endif
2037 
2038 	INP_WUNLOCK(inp);
2039 	return (0);
2040 }
2041 
2042 static int
2043 do_rx_tls_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
2044 {
2045 	struct adapter *sc = iq->adapter;
2046 	const struct cpl_rx_tls_cmp *cpl = mtod(m, const void *);
2047 	struct tlsrx_hdr_pkt *tls_hdr_pkt;
2048 	unsigned int tid = GET_TID(cpl);
2049 	struct toepcb *toep = lookup_tid(sc, tid);
2050 	struct inpcb *inp = toep->inp;
2051 	struct tcpcb *tp;
2052 	struct socket *so;
2053 	struct sockbuf *sb;
2054 	struct mbuf *tls_data;
2055 #ifdef KERN_TLS
2056 	struct tls_get_record *tgr;
2057 	struct mbuf *control;
2058 #endif
2059 	int len, pdu_length, rx_credits;
2060 
2061 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
2062 	KASSERT(!(toep->flags & TPF_SYNQE),
2063 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
2064 
2065 	/* strip off CPL header */
2066 	m_adj(m, sizeof(*cpl));
2067 	len = m->m_pkthdr.len;
2068 
2069 	atomic_add_long(&toep->vi->pi->rx_toe_tls_records, 1);
2070 
2071 	KASSERT(len == G_CPL_RX_TLS_CMP_LENGTH(be32toh(cpl->pdulength_length)),
2072 	    ("%s: payload length mismatch", __func__));
2073 
2074 	INP_WLOCK(inp);
2075 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
2076 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
2077 		    __func__, tid, len, inp->inp_flags);
2078 		INP_WUNLOCK(inp);
2079 		m_freem(m);
2080 		return (0);
2081 	}
2082 
2083 	pdu_length = G_CPL_RX_TLS_CMP_PDULENGTH(be32toh(cpl->pdulength_length));
2084 
2085 	so = inp_inpcbtosocket(inp);
2086 	tp = intotcpcb(inp);
2087 
2088 #ifdef VERBOSE_TRACES
2089 	CTR6(KTR_CXGBE, "%s: tid %u PDU len %d len %d seq %u, rcv_nxt %u",
2090 	    __func__, tid, pdu_length, len, be32toh(cpl->seq), tp->rcv_nxt);
2091 #endif
2092 
2093 	tp->rcv_nxt += pdu_length;
2094 	KASSERT(tp->rcv_wnd >= pdu_length,
2095 	    ("%s: negative window size", __func__));
2096 	tp->rcv_wnd -= pdu_length;
2097 
2098 	/* XXX: Not sure what to do about urgent data. */
2099 
2100 	/*
2101 	 * The payload of this CPL is the TLS header followed by
2102 	 * additional fields.
2103 	 */
2104 	KASSERT(m->m_len >= sizeof(*tls_hdr_pkt),
2105 	    ("%s: payload too small", __func__));
2106 	tls_hdr_pkt = mtod(m, void *);
2107 
2108 	tls_data = mbufq_dequeue(&toep->ulp_pdu_reclaimq);
2109 	if (tls_data != NULL) {
2110 		KASSERT(be32toh(cpl->seq) == tls_data->m_pkthdr.tls_tcp_seq,
2111 		    ("%s: sequence mismatch", __func__));
2112 	}
2113 
2114 #ifdef KERN_TLS
2115 	if (toep->tls.mode == TLS_MODE_KTLS) {
2116 		/* Report decryption errors as EBADMSG. */
2117 		if ((tls_hdr_pkt->res_to_mac_error & M_TLSRX_HDR_PKT_ERROR) !=
2118 		    0) {
2119 			m_freem(m);
2120 			m_freem(tls_data);
2121 
2122 			CURVNET_SET(toep->vnet);
2123 			so->so_error = EBADMSG;
2124 			sorwakeup(so);
2125 
2126 			INP_WUNLOCK(inp);
2127 			CURVNET_RESTORE();
2128 
2129 			return (0);
2130 		}
2131 
2132 		/* Allocate the control message mbuf. */
2133 		control = sbcreatecontrol(NULL, sizeof(*tgr), TLS_GET_RECORD,
2134 		    IPPROTO_TCP);
2135 		if (control == NULL) {
2136 			m_freem(m);
2137 			m_freem(tls_data);
2138 
2139 			CURVNET_SET(toep->vnet);
2140 			so->so_error = ENOBUFS;
2141 			sorwakeup(so);
2142 
2143 			INP_WUNLOCK(inp);
2144 			CURVNET_RESTORE();
2145 
2146 			return (0);
2147 		}
2148 
2149 		tgr = (struct tls_get_record *)
2150 		    CMSG_DATA(mtod(control, struct cmsghdr *));
2151 		tgr->tls_type = tls_hdr_pkt->type;
2152 		tgr->tls_vmajor = be16toh(tls_hdr_pkt->version) >> 8;
2153 		tgr->tls_vminor = be16toh(tls_hdr_pkt->version) & 0xff;
2154 
2155 		m_freem(m);
2156 
2157 		if (tls_data != NULL) {
2158 			m_last(tls_data)->m_flags |= M_EOR;
2159 			tgr->tls_length = htobe16(tls_data->m_pkthdr.len);
2160 		} else
2161 			tgr->tls_length = 0;
2162 		m = tls_data;
2163 	} else
2164 #endif
2165 	{
2166 		/*
2167 		 * Only the TLS header is sent to OpenSSL, so report
2168 		 * errors by altering the record type.
2169 		 */
2170 		if ((tls_hdr_pkt->res_to_mac_error & M_TLSRX_HDR_PKT_ERROR) !=
2171 		    0)
2172 			tls_hdr_pkt->type = CONTENT_TYPE_ERROR;
2173 
2174 		/* Trim this CPL's mbuf to only include the TLS header. */
2175 		KASSERT(m->m_len == len && m->m_next == NULL,
2176 		    ("%s: CPL spans multiple mbufs", __func__));
2177 		m->m_len = TLS_HEADER_LENGTH;
2178 		m->m_pkthdr.len = TLS_HEADER_LENGTH;
2179 
2180 		if (tls_data != NULL) {
2181 			/*
2182 			 * Update the TLS header length to be the length of
2183 			 * the payload data.
2184 			 */
2185 			tls_hdr_pkt->length = htobe16(tls_data->m_pkthdr.len);
2186 
2187 			m->m_next = tls_data;
2188 			m->m_pkthdr.len += tls_data->m_len;
2189 		}
2190 
2191 #ifdef KERN_TLS
2192 		control = NULL;
2193 #endif
2194 	}
2195 
2196 	sb = &so->so_rcv;
2197 	SOCKBUF_LOCK(sb);
2198 
2199 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
2200 		struct epoch_tracker et;
2201 
2202 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
2203 		    __func__, tid, pdu_length);
2204 		m_freem(m);
2205 #ifdef KERN_TLS
2206 		m_freem(control);
2207 #endif
2208 		SOCKBUF_UNLOCK(sb);
2209 		INP_WUNLOCK(inp);
2210 
2211 		CURVNET_SET(toep->vnet);
2212 		NET_EPOCH_ENTER(et);
2213 		INP_WLOCK(inp);
2214 		tp = tcp_drop(tp, ECONNRESET);
2215 		if (tp)
2216 			INP_WUNLOCK(inp);
2217 		NET_EPOCH_EXIT(et);
2218 		CURVNET_RESTORE();
2219 
2220 		return (0);
2221 	}
2222 
2223 	/*
2224 	 * Not all of the bytes on the wire are included in the socket buffer
2225 	 * (e.g. the MAC of the TLS record).  However, those bytes are included
2226 	 * in the TCP sequence space.
2227 	 */
2228 
2229 	/* receive buffer autosize */
2230 	MPASS(toep->vnet == so->so_vnet);
2231 	CURVNET_SET(toep->vnet);
2232 	if (sb->sb_flags & SB_AUTOSIZE &&
2233 	    V_tcp_do_autorcvbuf &&
2234 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
2235 	    m->m_pkthdr.len > (sbspace(sb) / 8 * 7)) {
2236 		unsigned int hiwat = sb->sb_hiwat;
2237 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
2238 		    V_tcp_autorcvbuf_max);
2239 
2240 		if (!sbreserve_locked(sb, newsize, so, NULL))
2241 			sb->sb_flags &= ~SB_AUTOSIZE;
2242 	}
2243 
2244 #ifdef KERN_TLS
2245 	if (control != NULL)
2246 		sbappendcontrol_locked(sb, m, control, 0);
2247 	else
2248 #endif
2249 		sbappendstream_locked(sb, m, 0);
2250 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
2251 #ifdef VERBOSE_TRACES
2252 	CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u",
2253 	    __func__, tid, rx_credits, tp->rcv_wnd);
2254 #endif
2255 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
2256 		rx_credits = send_rx_credits(sc, toep, rx_credits);
2257 		tp->rcv_wnd += rx_credits;
2258 		tp->rcv_adv += rx_credits;
2259 	}
2260 
2261 	sorwakeup_locked(so);
2262 	SOCKBUF_UNLOCK_ASSERT(sb);
2263 
2264 	INP_WUNLOCK(inp);
2265 	CURVNET_RESTORE();
2266 	return (0);
2267 }
2268 
2269 void
2270 do_rx_data_tls(const struct cpl_rx_data *cpl, struct toepcb *toep,
2271     struct mbuf *m)
2272 {
2273 	struct inpcb *inp = toep->inp;
2274 	struct tls_ofld_info *tls_ofld = &toep->tls;
2275 	struct tls_hdr *hdr;
2276 	struct tcpcb *tp;
2277 	struct socket *so;
2278 	struct sockbuf *sb;
2279 	int error, len, rx_credits;
2280 
2281 	len = m->m_pkthdr.len;
2282 
2283 	INP_WLOCK_ASSERT(inp);
2284 
2285 	so = inp_inpcbtosocket(inp);
2286 	tp = intotcpcb(inp);
2287 	sb = &so->so_rcv;
2288 	SOCKBUF_LOCK(sb);
2289 	CURVNET_SET(toep->vnet);
2290 
2291 	tp->rcv_nxt += len;
2292 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
2293 	tp->rcv_wnd -= len;
2294 
2295 	/* Do we have a full TLS header? */
2296 	if (len < sizeof(*hdr)) {
2297 		CTR3(KTR_CXGBE, "%s: tid %u len %d: too short for a TLS header",
2298 		    __func__, toep->tid, len);
2299 		so->so_error = EMSGSIZE;
2300 		goto out;
2301 	}
2302 	hdr = mtod(m, struct tls_hdr *);
2303 
2304 	/* Is the header valid? */
2305 	if (be16toh(hdr->version) != tls_ofld->k_ctx.proto_ver) {
2306 		CTR3(KTR_CXGBE, "%s: tid %u invalid version %04x",
2307 		    __func__, toep->tid, be16toh(hdr->version));
2308 		error = EINVAL;
2309 		goto report_error;
2310 	}
2311 	if (be16toh(hdr->length) < sizeof(*hdr)) {
2312 		CTR3(KTR_CXGBE, "%s: tid %u invalid length %u",
2313 		    __func__, toep->tid, be16toh(hdr->length));
2314 		error = EBADMSG;
2315 		goto report_error;
2316 	}
2317 
2318 	/* Did we get a truncated record? */
2319 	if (len < be16toh(hdr->length)) {
2320 		CTR4(KTR_CXGBE, "%s: tid %u truncated TLS record (%d vs %u)",
2321 		    __func__, toep->tid, len, be16toh(hdr->length));
2322 
2323 		error = EMSGSIZE;
2324 		goto report_error;
2325 	}
2326 
2327 	/* Is the header type unknown? */
2328 	switch (hdr->type) {
2329 	case CONTENT_TYPE_CCS:
2330 	case CONTENT_TYPE_ALERT:
2331 	case CONTENT_TYPE_APP_DATA:
2332 	case CONTENT_TYPE_HANDSHAKE:
2333 		break;
2334 	default:
2335 		CTR3(KTR_CXGBE, "%s: tid %u invalid TLS record type %u",
2336 		    __func__, toep->tid, hdr->type);
2337 		error = EBADMSG;
2338 		goto report_error;
2339 	}
2340 
2341 	/*
2342 	 * Just punt.  Although this could fall back to software
2343 	 * decryption, this case should never really happen.
2344 	 */
2345 	CTR4(KTR_CXGBE, "%s: tid %u dropping TLS record type %u, length %u",
2346 	    __func__, toep->tid, hdr->type, be16toh(hdr->length));
2347 	error = EBADMSG;
2348 
2349 report_error:
2350 #ifdef KERN_TLS
2351 	if (toep->tls.mode == TLS_MODE_KTLS)
2352 		so->so_error = error;
2353 	else
2354 #endif
2355 	{
2356 		/*
2357 		 * Report errors by sending an empty TLS record
2358 		 * with an error record type.
2359 		 */
2360 		hdr->type = CONTENT_TYPE_ERROR;
2361 
2362 		/* Trim this CPL's mbuf to only include the TLS header. */
2363 		KASSERT(m->m_len == len && m->m_next == NULL,
2364 		    ("%s: CPL spans multiple mbufs", __func__));
2365 		m->m_len = TLS_HEADER_LENGTH;
2366 		m->m_pkthdr.len = TLS_HEADER_LENGTH;
2367 
2368 		sbappendstream_locked(sb, m, 0);
2369 		m = NULL;
2370 	}
2371 
2372 out:
2373 	/*
2374 	 * This connection is going to die anyway, so probably don't
2375 	 * need to bother with returning credits.
2376 	 */
2377 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
2378 #ifdef VERBOSE_TRACES
2379 	CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u",
2380 	    __func__, toep->tid, rx_credits, tp->rcv_wnd);
2381 #endif
2382 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
2383 		rx_credits = send_rx_credits(toep->vi->adapter, toep,
2384 		    rx_credits);
2385 		tp->rcv_wnd += rx_credits;
2386 		tp->rcv_adv += rx_credits;
2387 	}
2388 
2389 	sorwakeup_locked(so);
2390 	SOCKBUF_UNLOCK_ASSERT(sb);
2391 
2392 	INP_WUNLOCK(inp);
2393 	CURVNET_RESTORE();
2394 
2395 	m_freem(m);
2396 }
2397 
2398 void
2399 t4_tls_mod_load(void)
2400 {
2401 
2402 	t4_register_cpl_handler(CPL_TLS_DATA, do_tls_data);
2403 	t4_register_cpl_handler(CPL_RX_TLS_CMP, do_rx_tls_cmp);
2404 }
2405 
2406 void
2407 t4_tls_mod_unload(void)
2408 {
2409 
2410 	t4_register_cpl_handler(CPL_TLS_DATA, NULL);
2411 	t4_register_cpl_handler(CPL_RX_TLS_CMP, NULL);
2412 }
2413 #endif	/* TCP_OFFLOAD */
2414