xref: /freebsd/sys/dev/cxgbe/tom/t4_tls.c (revision 8a82be504499c444e8fe9cb9eb41bee10d358b4d)
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 /*
68  * Handshake lock used for the handshake timer.  Having a global lock
69  * is perhaps not ideal, but it avoids having to use callout_drain()
70  * in tls_uninit_toep() which can't block.  Also, the timer shouldn't
71  * actually fire for most connections.
72  */
73 static struct mtx tls_handshake_lock;
74 
75 static void
76 t4_set_tls_tcb_field(struct toepcb *toep, uint16_t word, uint64_t mask,
77     uint64_t val)
78 {
79 	struct adapter *sc = td_adapter(toep->td);
80 
81 	t4_set_tcb_field(sc, toep->ofld_txq, toep, word, mask, val, 0, 0);
82 }
83 
84 /* TLS and DTLS common routines */
85 bool
86 can_tls_offload(struct adapter *sc)
87 {
88 
89 	return (sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS);
90 }
91 
92 int
93 tls_tx_key(struct toepcb *toep)
94 {
95 	struct tls_ofld_info *tls_ofld = &toep->tls;
96 
97 	return (tls_ofld->tx_key_addr >= 0);
98 }
99 
100 int
101 tls_rx_key(struct toepcb *toep)
102 {
103 	struct tls_ofld_info *tls_ofld = &toep->tls;
104 
105 	return (tls_ofld->rx_key_addr >= 0);
106 }
107 
108 static int
109 key_size(struct toepcb *toep)
110 {
111 	struct tls_ofld_info *tls_ofld = &toep->tls;
112 
113 	return ((tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE) ?
114 		tls_ofld->k_ctx.tx_key_info_size : KEY_IN_DDR_SIZE);
115 }
116 
117 /* Set TLS Key-Id in TCB */
118 static void
119 t4_set_tls_keyid(struct toepcb *toep, unsigned int key_id)
120 {
121 
122 	t4_set_tls_tcb_field(toep, W_TCB_RX_TLS_KEY_TAG,
123 			 V_TCB_RX_TLS_KEY_TAG(M_TCB_RX_TLS_BUF_TAG),
124 			 V_TCB_RX_TLS_KEY_TAG(key_id));
125 }
126 
127 /* Clear TF_RX_QUIESCE to re-enable receive. */
128 static void
129 t4_clear_rx_quiesce(struct toepcb *toep)
130 {
131 
132 	t4_set_tls_tcb_field(toep, W_TCB_T_FLAGS, V_TF_RX_QUIESCE(1), 0);
133 }
134 
135 static void
136 tls_clr_ofld_mode(struct toepcb *toep)
137 {
138 
139 	tls_stop_handshake_timer(toep);
140 
141 	/* Operate in PDU extraction mode only. */
142 	t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
143 	    V_TCB_ULP_RAW(M_TCB_ULP_RAW),
144 	    V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)));
145 	t4_clear_rx_quiesce(toep);
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 	 * XXX: Does not have the t4_get_tcb() checks to refine the
744 	 * workaround.
745 	 */
746 	callout_schedule(&tls_ofld->handshake_timer, TLS_SRV_HELLO_RD_TM * hz);
747 
748 	CTR2(KTR_CXGBE, "%s: tid %d sending RX_DATA_ACK", __func__, toep->tid);
749 	send_rx_modulate(sc, toep);
750 }
751 
752 static void
753 tls_start_handshake_timer(struct toepcb *toep)
754 {
755 	struct tls_ofld_info *tls_ofld = &toep->tls;
756 
757 	mtx_lock(&tls_handshake_lock);
758 	callout_reset(&tls_ofld->handshake_timer, TLS_SRV_HELLO_BKOFF_TM * hz,
759 	    tls_send_handshake_ack, toep);
760 	mtx_unlock(&tls_handshake_lock);
761 }
762 
763 void
764 tls_stop_handshake_timer(struct toepcb *toep)
765 {
766 	struct tls_ofld_info *tls_ofld = &toep->tls;
767 
768 	mtx_lock(&tls_handshake_lock);
769 	callout_stop(&tls_ofld->handshake_timer);
770 	mtx_unlock(&tls_handshake_lock);
771 }
772 
773 int
774 t4_ctloutput_tls(struct socket *so, struct sockopt *sopt)
775 {
776 	struct tls_key_context uk_ctx;
777 	struct inpcb *inp;
778 	struct tcpcb *tp;
779 	struct toepcb *toep;
780 	int error, optval;
781 
782 	error = 0;
783 	if (sopt->sopt_dir == SOPT_SET &&
784 	    sopt->sopt_name == TCP_TLSOM_SET_TLS_CONTEXT) {
785 		error = sooptcopyin(sopt, &uk_ctx, sizeof(uk_ctx),
786 		    sizeof(uk_ctx));
787 		if (error)
788 			return (error);
789 	}
790 
791 	inp = sotoinpcb(so);
792 	KASSERT(inp != NULL, ("tcp_ctloutput: inp == NULL"));
793 	INP_WLOCK(inp);
794 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
795 		INP_WUNLOCK(inp);
796 		return (ECONNRESET);
797 	}
798 	tp = intotcpcb(inp);
799 	toep = tp->t_toe;
800 	switch (sopt->sopt_dir) {
801 	case SOPT_SET:
802 		switch (sopt->sopt_name) {
803 		case TCP_TLSOM_SET_TLS_CONTEXT:
804 			if (toep->tls.mode == TLS_MODE_KTLS)
805 				error = EINVAL;
806 			else {
807 				error = program_key_context(tp, toep, &uk_ctx);
808 				if (error == 0)
809 					toep->tls.mode = TLS_MODE_TLSOM;
810 			}
811 			INP_WUNLOCK(inp);
812 			break;
813 		case TCP_TLSOM_CLR_TLS_TOM:
814 			if (toep->tls.mode == TLS_MODE_KTLS)
815 				error = EINVAL;
816 			else if (ulp_mode(toep) == ULP_MODE_TLS) {
817 				CTR2(KTR_CXGBE, "%s: tid %d CLR_TLS_TOM",
818 				    __func__, toep->tid);
819 				tls_clr_ofld_mode(toep);
820 			} else
821 				error = EOPNOTSUPP;
822 			INP_WUNLOCK(inp);
823 			break;
824 		case TCP_TLSOM_CLR_QUIES:
825 			if (toep->tls.mode == TLS_MODE_KTLS)
826 				error = EINVAL;
827 			else if (ulp_mode(toep) == ULP_MODE_TLS) {
828 				CTR2(KTR_CXGBE, "%s: tid %d CLR_QUIES",
829 				    __func__, toep->tid);
830 				tls_clr_quiesce(toep);
831 			} else
832 				error = EOPNOTSUPP;
833 			INP_WUNLOCK(inp);
834 			break;
835 		default:
836 			INP_WUNLOCK(inp);
837 			error = EOPNOTSUPP;
838 			break;
839 		}
840 		break;
841 	case SOPT_GET:
842 		switch (sopt->sopt_name) {
843 		case TCP_TLSOM_GET_TLS_TOM:
844 			/*
845 			 * TLS TX is permitted on any TOE socket, but
846 			 * TLS RX requires a TLS ULP mode.
847 			 */
848 			optval = TLS_TOM_NONE;
849 			if (can_tls_offload(td_adapter(toep->td)) &&
850 			    toep->tls.mode != TLS_MODE_KTLS) {
851 				switch (ulp_mode(toep)) {
852 				case ULP_MODE_NONE:
853 				case ULP_MODE_TCPDDP:
854 					optval = TLS_TOM_TXONLY;
855 					break;
856 				case ULP_MODE_TLS:
857 					optval = TLS_TOM_BOTH;
858 					break;
859 				}
860 			}
861 			CTR3(KTR_CXGBE, "%s: tid %d GET_TLS_TOM = %d",
862 			    __func__, toep->tid, optval);
863 			INP_WUNLOCK(inp);
864 			error = sooptcopyout(sopt, &optval, sizeof(optval));
865 			break;
866 		default:
867 			INP_WUNLOCK(inp);
868 			error = EOPNOTSUPP;
869 			break;
870 		}
871 		break;
872 	}
873 	return (error);
874 }
875 
876 #ifdef KERN_TLS
877 static void
878 init_ktls_key_context(struct ktls_session *tls, struct tls_key_context *k_ctx,
879     int direction)
880 {
881 	struct auth_hash *axf;
882 	u_int key_info_size, mac_key_size;
883 	char *hash, *key;
884 
885 	k_ctx->l_p_key = V_KEY_GET_LOC(direction == KTLS_TX ? KEY_WRITE_TX :
886 	    KEY_WRITE_RX);
887 	k_ctx->proto_ver = tls->params.tls_vmajor << 8 | tls->params.tls_vminor;
888 	k_ctx->cipher_secret_size = tls->params.cipher_key_len;
889 	key_info_size = sizeof(struct tx_keyctx_hdr) +
890 	    k_ctx->cipher_secret_size;
891 	if (direction == KTLS_TX)
892 		key = k_ctx->tx.key;
893 	else
894 		key = k_ctx->rx.key;
895 	memcpy(key, tls->params.cipher_key, tls->params.cipher_key_len);
896 	hash = key + tls->params.cipher_key_len;
897 	if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16) {
898 		k_ctx->state.auth_mode = SCMD_AUTH_MODE_GHASH;
899 		k_ctx->state.enc_mode = SCMD_CIPH_MODE_AES_GCM;
900 		k_ctx->iv_size = 4;
901 		k_ctx->mac_first = 0;
902 		k_ctx->hmac_ctrl = SCMD_HMAC_CTRL_NOP;
903 		key_info_size += GMAC_BLOCK_LEN;
904 		k_ctx->mac_secret_size = 0;
905 		if (direction == KTLS_TX)
906 			memcpy(k_ctx->tx.salt, tls->params.iv, SALT_SIZE);
907 		else
908 			memcpy(k_ctx->rx.salt, tls->params.iv, SALT_SIZE);
909 		t4_init_gmac_hash(tls->params.cipher_key,
910 		    tls->params.cipher_key_len, hash);
911 	} else {
912 		switch (tls->params.auth_algorithm) {
913 		case CRYPTO_SHA1_HMAC:
914 			axf = &auth_hash_hmac_sha1;
915 			mac_key_size = SHA1_HASH_LEN;
916 			k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA1;
917 			break;
918 		case CRYPTO_SHA2_256_HMAC:
919 			axf = &auth_hash_hmac_sha2_256;
920 			mac_key_size = SHA2_256_HASH_LEN;
921 			k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA256;
922 			break;
923 		case CRYPTO_SHA2_384_HMAC:
924 			axf = &auth_hash_hmac_sha2_384;
925 			mac_key_size = SHA2_512_HASH_LEN;
926 			k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA512_384;
927 			break;
928 		default:
929 			panic("bad auth mode");
930 		}
931 		k_ctx->state.enc_mode = SCMD_CIPH_MODE_AES_CBC;
932 		k_ctx->iv_size = 8; /* for CBC, iv is 16B, unit of 2B */
933 		k_ctx->mac_first = 1;
934 		k_ctx->hmac_ctrl = SCMD_HMAC_CTRL_NO_TRUNC;
935 		key_info_size += roundup2(mac_key_size, 16) * 2;
936 		k_ctx->mac_secret_size = mac_key_size;
937 		t4_init_hmac_digest(axf, mac_key_size, tls->params.auth_key,
938 		    tls->params.auth_key_len, hash);
939 	}
940 
941 	if (direction == KTLS_TX)
942 		k_ctx->tx_key_info_size = key_info_size;
943 	else
944 		k_ctx->rx_key_info_size = key_info_size;
945 	k_ctx->frag_size = tls->params.max_frame_len;
946 	k_ctx->iv_ctrl = 1;
947 }
948 
949 int
950 tls_alloc_ktls(struct toepcb *toep, struct ktls_session *tls, int direction)
951 {
952 	struct adapter *sc = td_adapter(toep->td);
953 	struct tls_key_context *k_ctx;
954 	int error, key_offset;
955 
956 	if (toep->tls.mode == TLS_MODE_TLSOM)
957 		return (EINVAL);
958 	if (!can_tls_offload(td_adapter(toep->td)))
959 		return (EINVAL);
960 	switch (ulp_mode(toep)) {
961 	case ULP_MODE_TLS:
962 		break;
963 	case ULP_MODE_NONE:
964 	case ULP_MODE_TCPDDP:
965 		if (direction != KTLS_TX)
966 			return (EINVAL);
967 		break;
968 	default:
969 		return (EINVAL);
970 	}
971 
972 	switch (tls->params.cipher_algorithm) {
973 	case CRYPTO_AES_CBC:
974 		/* XXX: Explicitly ignore any provided IV. */
975 		switch (tls->params.cipher_key_len) {
976 		case 128 / 8:
977 		case 192 / 8:
978 		case 256 / 8:
979 			break;
980 		default:
981 			return (EINVAL);
982 		}
983 		switch (tls->params.auth_algorithm) {
984 		case CRYPTO_SHA1_HMAC:
985 		case CRYPTO_SHA2_256_HMAC:
986 		case CRYPTO_SHA2_384_HMAC:
987 			break;
988 		default:
989 			return (EPROTONOSUPPORT);
990 		}
991 		break;
992 	case CRYPTO_AES_NIST_GCM_16:
993 		if (tls->params.iv_len != SALT_SIZE)
994 			return (EINVAL);
995 		switch (tls->params.cipher_key_len) {
996 		case 128 / 8:
997 		case 192 / 8:
998 		case 256 / 8:
999 			break;
1000 		default:
1001 			return (EINVAL);
1002 		}
1003 		break;
1004 	default:
1005 		return (EPROTONOSUPPORT);
1006 	}
1007 
1008 	/* Only TLS 1.1 and TLS 1.2 are currently supported. */
1009 	if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE ||
1010 	    tls->params.tls_vminor < TLS_MINOR_VER_ONE ||
1011 	    tls->params.tls_vminor > TLS_MINOR_VER_TWO)
1012 		return (EPROTONOSUPPORT);
1013 
1014 	/* Bail if we already have a key. */
1015 	if (direction == KTLS_TX) {
1016 		if (toep->tls.tx_key_addr != -1)
1017 			return (EOPNOTSUPP);
1018 	} else {
1019 		if (toep->tls.rx_key_addr != -1)
1020 			return (EOPNOTSUPP);
1021 	}
1022 
1023 	/*
1024 	 * XXX: This assumes no key renegotation.  If KTLS ever supports
1025 	 * that we will want to allocate TLS sessions dynamically rather
1026 	 * than as a static member of toep.
1027 	 */
1028 	k_ctx = &toep->tls.k_ctx;
1029 	init_ktls_key_context(tls, k_ctx, direction);
1030 
1031 	error = tls_program_key_id(toep, k_ctx);
1032 	if (error)
1033 		return (error);
1034 
1035 	if (direction == KTLS_TX) {
1036 		toep->tls.scmd0.seqno_numivs =
1037 			(V_SCMD_SEQ_NO_CTRL(3) |
1038 			 V_SCMD_PROTO_VERSION(get_proto_ver(k_ctx->proto_ver)) |
1039 			 V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
1040 			 V_SCMD_CIPH_AUTH_SEQ_CTRL((k_ctx->mac_first == 0)) |
1041 			 V_SCMD_CIPH_MODE(k_ctx->state.enc_mode) |
1042 			 V_SCMD_AUTH_MODE(k_ctx->state.auth_mode) |
1043 			 V_SCMD_HMAC_CTRL(k_ctx->hmac_ctrl) |
1044 			 V_SCMD_IV_SIZE(k_ctx->iv_size));
1045 
1046 		toep->tls.scmd0.ivgen_hdrlen =
1047 			(V_SCMD_IV_GEN_CTRL(k_ctx->iv_ctrl) |
1048 			 V_SCMD_KEY_CTX_INLINE(0) |
1049 			 V_SCMD_TLS_FRAG_ENABLE(1));
1050 
1051 		if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16)
1052 			toep->tls.iv_len = 8;
1053 		else
1054 			toep->tls.iv_len = AES_BLOCK_LEN;
1055 
1056 		toep->tls.mac_length = k_ctx->mac_secret_size;
1057 
1058 		toep->tls.fcplenmax = get_tp_plen_max(&toep->tls);
1059 		toep->tls.expn_per_ulp = tls->params.tls_hlen +
1060 		    tls->params.tls_tlen;
1061 		toep->tls.pdus_per_ulp = 1;
1062 		toep->tls.adjusted_plen = toep->tls.expn_per_ulp +
1063 		    toep->tls.k_ctx.frag_size;
1064 	} else {
1065 		/* Stop timer on handshake completion */
1066 		tls_stop_handshake_timer(toep);
1067 
1068 		toep->flags &= ~TPF_FORCE_CREDITS;
1069 		toep->flags |= TPF_TLS_RECEIVE;
1070 
1071 		/*
1072 		 * RX key tags are an index into the key portion of MA
1073 		 * memory stored as an offset from the base address in
1074 		 * units of 64 bytes.
1075 		 */
1076 		key_offset = toep->tls.rx_key_addr - sc->vres.key.start;
1077 		t4_set_tls_keyid(toep, key_offset / 64);
1078 		t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
1079 				 V_TCB_ULP_RAW(M_TCB_ULP_RAW),
1080 				 V_TCB_ULP_RAW((V_TF_TLS_KEY_SIZE(3) |
1081 						V_TF_TLS_CONTROL(1) |
1082 						V_TF_TLS_ACTIVE(1) |
1083 						V_TF_TLS_ENABLE(1))));
1084 		t4_set_tls_tcb_field(toep, W_TCB_TLS_SEQ,
1085 				 V_TCB_TLS_SEQ(M_TCB_TLS_SEQ),
1086 				 V_TCB_TLS_SEQ(0));
1087 		t4_clear_rx_quiesce(toep);
1088 	}
1089 
1090 	toep->tls.mode = TLS_MODE_KTLS;
1091 
1092 	return (0);
1093 }
1094 #endif
1095 
1096 void
1097 tls_init_toep(struct toepcb *toep)
1098 {
1099 	struct tls_ofld_info *tls_ofld = &toep->tls;
1100 
1101 	tls_ofld->mode = TLS_MODE_OFF;
1102 	tls_ofld->key_location = TLS_SFO_WR_CONTEXTLOC_DDR;
1103 	tls_ofld->rx_key_addr = -1;
1104 	tls_ofld->tx_key_addr = -1;
1105 	if (ulp_mode(toep) == ULP_MODE_TLS)
1106 		callout_init_mtx(&tls_ofld->handshake_timer,
1107 		    &tls_handshake_lock, 0);
1108 }
1109 
1110 void
1111 tls_establish(struct toepcb *toep)
1112 {
1113 
1114 	/*
1115 	 * Enable PDU extraction.
1116 	 *
1117 	 * XXX: Supposedly this should be done by the firmware when
1118 	 * the ULP_MODE FLOWC parameter is set in send_flowc_wr(), but
1119 	 * in practice this seems to be required.
1120 	 */
1121 	CTR2(KTR_CXGBE, "%s: tid %d setting TLS_ENABLE", __func__, toep->tid);
1122 	t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW, V_TCB_ULP_RAW(M_TCB_ULP_RAW),
1123 	    V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)));
1124 
1125 	toep->flags |= TPF_FORCE_CREDITS;
1126 
1127 	tls_start_handshake_timer(toep);
1128 }
1129 
1130 void
1131 tls_uninit_toep(struct toepcb *toep)
1132 {
1133 
1134 	if (ulp_mode(toep) == ULP_MODE_TLS)
1135 		tls_stop_handshake_timer(toep);
1136 	clear_tls_keyid(toep);
1137 }
1138 
1139 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
1140 #define	MIN_OFLD_TLSTX_CREDITS(toep)					\
1141 	(howmany(sizeof(struct fw_tlstx_data_wr) +			\
1142 	    sizeof(struct cpl_tx_tls_sfo) + key_size((toep)) +		\
1143 	    CIPHER_BLOCK_SIZE + 1, 16))
1144 
1145 static inline u_int
1146 max_imm_tls_space(int tx_credits)
1147 {
1148 	const int n = 2;	/* Use only up to 2 desc for imm. data WR */
1149 	int space;
1150 
1151 	KASSERT(tx_credits >= 0 &&
1152 		tx_credits <= MAX_OFLD_TX_CREDITS,
1153 		("%s: %d credits", __func__, tx_credits));
1154 
1155 	if (tx_credits >= (n * EQ_ESIZE) / 16)
1156 		space = (n * EQ_ESIZE);
1157 	else
1158 		space = tx_credits * 16;
1159 	return (space);
1160 }
1161 
1162 static int
1163 count_mbuf_segs(struct mbuf *m, int skip, int len, int *max_nsegs_1mbufp)
1164 {
1165 	int max_nsegs_1mbuf, n, nsegs;
1166 
1167 	while (skip >= m->m_len) {
1168 		skip -= m->m_len;
1169 		m = m->m_next;
1170 	}
1171 
1172 	nsegs = 0;
1173 	max_nsegs_1mbuf = 0;
1174 	while (len > 0) {
1175 		n = sglist_count(mtod(m, char *) + skip, m->m_len - skip);
1176 		if (n > max_nsegs_1mbuf)
1177 			max_nsegs_1mbuf = n;
1178 		nsegs += n;
1179 		len -= m->m_len - skip;
1180 		skip = 0;
1181 		m = m->m_next;
1182 	}
1183 	*max_nsegs_1mbufp = max_nsegs_1mbuf;
1184 	return (nsegs);
1185 }
1186 
1187 static void
1188 write_tlstx_wr(struct fw_tlstx_data_wr *txwr, struct toepcb *toep,
1189     unsigned int immdlen, unsigned int plen, unsigned int expn,
1190     unsigned int pdus, uint8_t credits, int shove, int imm_ivs)
1191 {
1192 	struct tls_ofld_info *tls_ofld = &toep->tls;
1193 	unsigned int len = plen + expn;
1194 
1195 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_TLSTX_DATA_WR) |
1196 	    V_FW_TLSTX_DATA_WR_COMPL(1) |
1197 	    V_FW_TLSTX_DATA_WR_IMMDLEN(immdlen));
1198 	txwr->flowid_len16 = htobe32(V_FW_TLSTX_DATA_WR_FLOWID(toep->tid) |
1199 	    V_FW_TLSTX_DATA_WR_LEN16(credits));
1200 	txwr->plen = htobe32(len);
1201 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ULP_MODE_TLS) |
1202 	    V_TX_URG(0) | /* F_T6_TX_FORCE | */ V_TX_SHOVE(shove));
1203 	txwr->ctxloc_to_exp = htobe32(V_FW_TLSTX_DATA_WR_NUMIVS(pdus) |
1204 	    V_FW_TLSTX_DATA_WR_EXP(expn) |
1205 	    V_FW_TLSTX_DATA_WR_CTXLOC(tls_ofld->key_location) |
1206 	    V_FW_TLSTX_DATA_WR_IVDSGL(!imm_ivs) |
1207 	    V_FW_TLSTX_DATA_WR_KEYSIZE(tls_ofld->k_ctx.tx_key_info_size >> 4));
1208 	txwr->mfs = htobe16(tls_ofld->k_ctx.frag_size);
1209 	txwr->adjustedplen_pkd = htobe16(
1210 	    V_FW_TLSTX_DATA_WR_ADJUSTEDPLEN(tls_ofld->adjusted_plen));
1211 	txwr->expinplenmax_pkd = htobe16(
1212 	    V_FW_TLSTX_DATA_WR_EXPINPLENMAX(tls_ofld->expn_per_ulp));
1213 	txwr->pdusinplenmax_pkd =
1214 	    V_FW_TLSTX_DATA_WR_PDUSINPLENMAX(tls_ofld->pdus_per_ulp);
1215 }
1216 
1217 static void
1218 write_tlstx_cpl(struct cpl_tx_tls_sfo *cpl, struct toepcb *toep,
1219     struct tls_hdr *tls_hdr, unsigned int plen, unsigned int pdus)
1220 {
1221 	struct tls_ofld_info *tls_ofld = &toep->tls;
1222 	int data_type, seglen;
1223 
1224 	if (plen < tls_ofld->k_ctx.frag_size)
1225 		seglen = plen;
1226 	else
1227 		seglen = tls_ofld->k_ctx.frag_size;
1228 	data_type = tls_content_type(tls_hdr->type);
1229 	cpl->op_to_seg_len = htobe32(V_CPL_TX_TLS_SFO_OPCODE(CPL_TX_TLS_SFO) |
1230 	    V_CPL_TX_TLS_SFO_DATA_TYPE(data_type) |
1231 	    V_CPL_TX_TLS_SFO_CPL_LEN(2) | V_CPL_TX_TLS_SFO_SEG_LEN(seglen));
1232 	cpl->pld_len = htobe32(plen);
1233 	if (data_type == CPL_TX_TLS_SFO_TYPE_HEARTBEAT)
1234 		cpl->type_protover = htobe32(
1235 		    V_CPL_TX_TLS_SFO_TYPE(tls_hdr->type));
1236 	cpl->seqno_numivs = htobe32(tls_ofld->scmd0.seqno_numivs |
1237 	    V_SCMD_NUM_IVS(pdus));
1238 	cpl->ivgen_hdrlen = htobe32(tls_ofld->scmd0.ivgen_hdrlen);
1239 	cpl->scmd1 = htobe64(tls_ofld->tx_seq_no);
1240 	tls_ofld->tx_seq_no += pdus;
1241 }
1242 
1243 /*
1244  * Similar to write_tx_sgl() except that it accepts an optional
1245  * trailer buffer for IVs.
1246  */
1247 static void
1248 write_tlstx_sgl(void *dst, struct mbuf *start, int skip, int plen,
1249     void *iv_buffer, int iv_len, int nsegs, int n)
1250 {
1251 	struct mbuf *m;
1252 	struct ulptx_sgl *usgl = dst;
1253 	int i, j, rc;
1254 	struct sglist sg;
1255 	struct sglist_seg segs[n];
1256 
1257 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
1258 
1259 	sglist_init(&sg, n, segs);
1260 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1261 	    V_ULPTX_NSGE(nsegs));
1262 
1263 	for (m = start; skip >= m->m_len; m = m->m_next)
1264 		skip -= m->m_len;
1265 
1266 	i = -1;
1267 	for (m = start; plen > 0; m = m->m_next) {
1268 		rc = sglist_append(&sg, mtod(m, char *) + skip,
1269 		    m->m_len - skip);
1270 		if (__predict_false(rc != 0))
1271 			panic("%s: sglist_append %d", __func__, rc);
1272 		plen -= m->m_len - skip;
1273 		skip = 0;
1274 
1275 		for (j = 0; j < sg.sg_nseg; i++, j++) {
1276 			if (i < 0) {
1277 				usgl->len0 = htobe32(segs[j].ss_len);
1278 				usgl->addr0 = htobe64(segs[j].ss_paddr);
1279 			} else {
1280 				usgl->sge[i / 2].len[i & 1] =
1281 				    htobe32(segs[j].ss_len);
1282 				usgl->sge[i / 2].addr[i & 1] =
1283 				    htobe64(segs[j].ss_paddr);
1284 			}
1285 #ifdef INVARIANTS
1286 			nsegs--;
1287 #endif
1288 		}
1289 		sglist_reset(&sg);
1290 	}
1291 	if (iv_buffer != NULL) {
1292 		rc = sglist_append(&sg, iv_buffer, iv_len);
1293 		if (__predict_false(rc != 0))
1294 			panic("%s: sglist_append %d", __func__, rc);
1295 
1296 		for (j = 0; j < sg.sg_nseg; i++, j++) {
1297 			if (i < 0) {
1298 				usgl->len0 = htobe32(segs[j].ss_len);
1299 				usgl->addr0 = htobe64(segs[j].ss_paddr);
1300 			} else {
1301 				usgl->sge[i / 2].len[i & 1] =
1302 				    htobe32(segs[j].ss_len);
1303 				usgl->sge[i / 2].addr[i & 1] =
1304 				    htobe64(segs[j].ss_paddr);
1305 			}
1306 #ifdef INVARIANTS
1307 			nsegs--;
1308 #endif
1309 		}
1310 	}
1311 	if (i & 1)
1312 		usgl->sge[i / 2].len[1] = htobe32(0);
1313 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, iv_buffer %p",
1314 	    __func__, nsegs, start, iv_buffer));
1315 }
1316 
1317 /*
1318  * Similar to t4_push_frames() but handles TLS sockets when TLS offload
1319  * is enabled.  Rather than transmitting bulk data, the socket buffer
1320  * contains TLS records.  The work request requires a full TLS record,
1321  * so batch mbufs up until a full TLS record is seen.  This requires
1322  * reading the TLS header out of the start of each record to determine
1323  * its length.
1324  */
1325 void
1326 t4_push_tls_records(struct adapter *sc, struct toepcb *toep, int drop)
1327 {
1328 	struct tls_hdr thdr;
1329 	struct mbuf *sndptr;
1330 	struct fw_tlstx_data_wr *txwr;
1331 	struct cpl_tx_tls_sfo *cpl;
1332 	struct wrqe *wr;
1333 	u_int plen, nsegs, credits, space, max_nsegs_1mbuf, wr_len;
1334 	u_int expn_size, iv_len, pdus, sndptroff;
1335 	struct tls_ofld_info *tls_ofld = &toep->tls;
1336 	struct inpcb *inp = toep->inp;
1337 	struct tcpcb *tp = intotcpcb(inp);
1338 	struct socket *so = inp->inp_socket;
1339 	struct sockbuf *sb = &so->so_snd;
1340 	int tls_size, tx_credits, shove, /* compl,*/ sowwakeup;
1341 	struct ofld_tx_sdesc *txsd;
1342 	bool imm_ivs, imm_payload;
1343 	void *iv_buffer, *iv_dst, *buf;
1344 
1345 	INP_WLOCK_ASSERT(inp);
1346 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1347 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1348 
1349 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
1350 	    ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS,
1351 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1352 	KASSERT(tls_tx_key(toep),
1353 	    ("%s: TX key not set for toep %p", __func__, toep));
1354 
1355 #ifdef VERBOSE_TRACES
1356 	CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
1357 	    __func__, toep->tid, toep->flags, tp->t_flags);
1358 #endif
1359 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1360 		return;
1361 
1362 #ifdef RATELIMIT
1363 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
1364 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
1365 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
1366 	}
1367 #endif
1368 
1369 	/*
1370 	 * This function doesn't resume by itself.  Someone else must clear the
1371 	 * flag and call this function.
1372 	 */
1373 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1374 		KASSERT(drop == 0,
1375 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1376 		return;
1377 	}
1378 
1379 	txsd = &toep->txsd[toep->txsd_pidx];
1380 	for (;;) {
1381 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1382 		space = max_imm_tls_space(tx_credits);
1383 		wr_len = sizeof(struct fw_tlstx_data_wr) +
1384 		    sizeof(struct cpl_tx_tls_sfo) + key_size(toep);
1385 		if (wr_len + CIPHER_BLOCK_SIZE + 1 > space) {
1386 #ifdef VERBOSE_TRACES
1387 			CTR5(KTR_CXGBE,
1388 			    "%s: tid %d tx_credits %d min_wr %d space %d",
1389 			    __func__, toep->tid, tx_credits, wr_len +
1390 			    CIPHER_BLOCK_SIZE + 1, space);
1391 #endif
1392 			return;
1393 		}
1394 
1395 		SOCKBUF_LOCK(sb);
1396 		sowwakeup = drop;
1397 		if (drop) {
1398 			sbdrop_locked(sb, drop);
1399 			MPASS(tls_ofld->sb_off >= drop);
1400 			tls_ofld->sb_off -= drop;
1401 			drop = 0;
1402 		}
1403 
1404 		/*
1405 		 * Send a FIN if requested, but only if there's no
1406 		 * more data to send.
1407 		 */
1408 		if (sbavail(sb) == tls_ofld->sb_off &&
1409 		    toep->flags & TPF_SEND_FIN) {
1410 			if (sowwakeup)
1411 				sowwakeup_locked(so);
1412 			else
1413 				SOCKBUF_UNLOCK(sb);
1414 			SOCKBUF_UNLOCK_ASSERT(sb);
1415 			t4_close_conn(sc, toep);
1416 			return;
1417 		}
1418 
1419 		if (sbavail(sb) < tls_ofld->sb_off + TLS_HEADER_LENGTH) {
1420 			/*
1421 			 * A full TLS header is not yet queued, stop
1422 			 * for now until more data is added to the
1423 			 * socket buffer.  However, if the connection
1424 			 * has been closed, we will never get the rest
1425 			 * of the header so just discard the partial
1426 			 * header and close the connection.
1427 			 */
1428 #ifdef VERBOSE_TRACES
1429 			CTR5(KTR_CXGBE, "%s: tid %d sbavail %d sb_off %d%s",
1430 			    __func__, toep->tid, sbavail(sb), tls_ofld->sb_off,
1431 			    toep->flags & TPF_SEND_FIN ? "" : " SEND_FIN");
1432 #endif
1433 			if (sowwakeup)
1434 				sowwakeup_locked(so);
1435 			else
1436 				SOCKBUF_UNLOCK(sb);
1437 			SOCKBUF_UNLOCK_ASSERT(sb);
1438 			if (toep->flags & TPF_SEND_FIN)
1439 				t4_close_conn(sc, toep);
1440 			return;
1441 		}
1442 
1443 		/* Read the header of the next TLS record. */
1444 		sndptr = sbsndmbuf(sb, tls_ofld->sb_off, &sndptroff);
1445 		m_copydata(sndptr, sndptroff, sizeof(thdr), (caddr_t)&thdr);
1446 		tls_size = htons(thdr.length);
1447 		plen = TLS_HEADER_LENGTH + tls_size;
1448 		pdus = howmany(tls_size, tls_ofld->k_ctx.frag_size);
1449 		iv_len = pdus * CIPHER_BLOCK_SIZE;
1450 
1451 		if (sbavail(sb) < tls_ofld->sb_off + plen) {
1452 			/*
1453 			 * The full TLS record is not yet queued, stop
1454 			 * for now until more data is added to the
1455 			 * socket buffer.  However, if the connection
1456 			 * has been closed, we will never get the rest
1457 			 * of the record so just discard the partial
1458 			 * record and close the connection.
1459 			 */
1460 #ifdef VERBOSE_TRACES
1461 			CTR6(KTR_CXGBE,
1462 			    "%s: tid %d sbavail %d sb_off %d plen %d%s",
1463 			    __func__, toep->tid, sbavail(sb), tls_ofld->sb_off,
1464 			    plen, toep->flags & TPF_SEND_FIN ? "" :
1465 			    " SEND_FIN");
1466 #endif
1467 			if (sowwakeup)
1468 				sowwakeup_locked(so);
1469 			else
1470 				SOCKBUF_UNLOCK(sb);
1471 			SOCKBUF_UNLOCK_ASSERT(sb);
1472 			if (toep->flags & TPF_SEND_FIN)
1473 				t4_close_conn(sc, toep);
1474 			return;
1475 		}
1476 
1477 		/* Shove if there is no additional data pending. */
1478 		shove = (sbavail(sb) == tls_ofld->sb_off + plen) &&
1479 		    !(tp->t_flags & TF_MORETOCOME);
1480 
1481 		if (sb->sb_flags & SB_AUTOSIZE &&
1482 		    V_tcp_do_autosndbuf &&
1483 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
1484 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
1485 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
1486 			    V_tcp_autosndbuf_max);
1487 
1488 			if (!sbreserve_locked(sb, newsize, so, NULL))
1489 				sb->sb_flags &= ~SB_AUTOSIZE;
1490 			else
1491 				sowwakeup = 1;	/* room available */
1492 		}
1493 		if (sowwakeup)
1494 			sowwakeup_locked(so);
1495 		else
1496 			SOCKBUF_UNLOCK(sb);
1497 		SOCKBUF_UNLOCK_ASSERT(sb);
1498 
1499 		if (__predict_false(toep->flags & TPF_FIN_SENT))
1500 			panic("%s: excess tx.", __func__);
1501 
1502 		/* Determine whether to use immediate vs SGL. */
1503 		imm_payload = false;
1504 		imm_ivs = false;
1505 		if (wr_len + iv_len <= space) {
1506 			imm_ivs = true;
1507 			wr_len += iv_len;
1508 			if (wr_len + tls_size <= space) {
1509 				wr_len += tls_size;
1510 				imm_payload = true;
1511 			}
1512 		}
1513 
1514 		/* Allocate space for IVs if needed. */
1515 		if (!imm_ivs) {
1516 			iv_buffer = malloc(iv_len, M_CXGBE, M_NOWAIT);
1517 			if (iv_buffer == NULL) {
1518 				/*
1519 				 * XXX: How to restart this?
1520 				 */
1521 				if (sowwakeup)
1522 					sowwakeup_locked(so);
1523 				else
1524 					SOCKBUF_UNLOCK(sb);
1525 				SOCKBUF_UNLOCK_ASSERT(sb);
1526 				CTR3(KTR_CXGBE,
1527 			    "%s: tid %d failed to alloc IV space len %d",
1528 				    __func__, toep->tid, iv_len);
1529 				return;
1530 			}
1531 		} else
1532 			iv_buffer = NULL;
1533 
1534 		/* Determine size of SGL. */
1535 		nsegs = 0;
1536 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1537 		if (!imm_payload) {
1538 			nsegs = count_mbuf_segs(sndptr, sndptroff +
1539 			    TLS_HEADER_LENGTH, tls_size, &max_nsegs_1mbuf);
1540 			if (!imm_ivs) {
1541 				int n = sglist_count(iv_buffer, iv_len);
1542 				nsegs += n;
1543 				if (n > max_nsegs_1mbuf)
1544 					max_nsegs_1mbuf = n;
1545 			}
1546 
1547 			/* Account for SGL in work request length. */
1548 			wr_len += sizeof(struct ulptx_sgl) +
1549 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1550 		}
1551 
1552 		wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1553 		if (wr == NULL) {
1554 			/* XXX: how will we recover from this? */
1555 			toep->flags |= TPF_TX_SUSPENDED;
1556 			return;
1557 		}
1558 
1559 #ifdef VERBOSE_TRACES
1560 		CTR5(KTR_CXGBE, "%s: tid %d TLS record %d len %#x pdus %d",
1561 		    __func__, toep->tid, thdr.type, tls_size, pdus);
1562 #endif
1563 		txwr = wrtod(wr);
1564 		cpl = (struct cpl_tx_tls_sfo *)(txwr + 1);
1565 		memset(txwr, 0, roundup2(wr_len, 16));
1566 		credits = howmany(wr_len, 16);
1567 		expn_size = tls_expansion_size(toep, tls_size, 0, NULL);
1568 		write_tlstx_wr(txwr, toep, imm_payload ? tls_size : 0,
1569 		    tls_size, expn_size, pdus, credits, shove, imm_ivs ? 1 : 0);
1570 		write_tlstx_cpl(cpl, toep, &thdr, tls_size, pdus);
1571 		tls_copy_tx_key(toep, cpl + 1);
1572 
1573 		/* Generate random IVs */
1574 		buf = (char *)(cpl + 1) + key_size(toep);
1575 		if (imm_ivs) {
1576 			MPASS(iv_buffer == NULL);
1577 			iv_dst = buf;
1578 			buf = (char *)iv_dst + iv_len;
1579 		} else
1580 			iv_dst = iv_buffer;
1581 		arc4rand(iv_dst, iv_len, 0);
1582 
1583 		if (imm_payload) {
1584 			m_copydata(sndptr, sndptroff + TLS_HEADER_LENGTH,
1585 			    tls_size, buf);
1586 		} else {
1587 			write_tlstx_sgl(buf, sndptr,
1588 			    sndptroff + TLS_HEADER_LENGTH, tls_size, iv_buffer,
1589 			    iv_len, nsegs, max_nsegs_1mbuf);
1590 		}
1591 
1592 		KASSERT(toep->tx_credits >= credits,
1593 			("%s: not enough credits", __func__));
1594 
1595 		toep->tx_credits -= credits;
1596 
1597 		tp->snd_nxt += plen;
1598 		tp->snd_max += plen;
1599 
1600 		SOCKBUF_LOCK(sb);
1601 		sbsndptr_adv(sb, sb->sb_sndptr, plen);
1602 		tls_ofld->sb_off += plen;
1603 		SOCKBUF_UNLOCK(sb);
1604 
1605 		toep->flags |= TPF_TX_DATA_SENT;
1606 		if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep))
1607 			toep->flags |= TPF_TX_SUSPENDED;
1608 
1609 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1610 		txsd->plen = plen;
1611 		txsd->tx_credits = credits;
1612 		txsd->iv_buffer = iv_buffer;
1613 		txsd++;
1614 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1615 			toep->txsd_pidx = 0;
1616 			txsd = &toep->txsd[0];
1617 		}
1618 		toep->txsd_avail--;
1619 
1620 		atomic_add_long(&toep->vi->pi->tx_toe_tls_records, 1);
1621 		atomic_add_long(&toep->vi->pi->tx_toe_tls_octets, plen);
1622 
1623 		t4_l2t_send(sc, wr, toep->l2te);
1624 	}
1625 }
1626 
1627 #ifdef KERN_TLS
1628 static int
1629 count_ext_pgs_segs(struct mbuf *m)
1630 {
1631 	vm_paddr_t nextpa;
1632 	u_int i, nsegs;
1633 
1634 	MPASS(m->m_epg_npgs > 0);
1635 	nsegs = 1;
1636 	nextpa = m->m_epg_pa[0] + PAGE_SIZE;
1637 	for (i = 1; i < m->m_epg_npgs; i++) {
1638 		if (nextpa != m->m_epg_pa[i])
1639 			nsegs++;
1640 		nextpa = m->m_epg_pa[i] + PAGE_SIZE;
1641 	}
1642 	return (nsegs);
1643 }
1644 
1645 static void
1646 write_ktlstx_sgl(void *dst, struct mbuf *m, int nsegs)
1647 {
1648 	struct ulptx_sgl *usgl = dst;
1649 	vm_paddr_t pa;
1650 	uint32_t len;
1651 	int i, j;
1652 
1653 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
1654 
1655 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1656 	    V_ULPTX_NSGE(nsegs));
1657 
1658 	/* Figure out the first S/G length. */
1659 	pa = m->m_epg_pa[0] + m->m_epg_1st_off;
1660 	usgl->addr0 = htobe64(pa);
1661 	len = m_epg_pagelen(m, 0, m->m_epg_1st_off);
1662 	pa += len;
1663 	for (i = 1; i < m->m_epg_npgs; i++) {
1664 		if (m->m_epg_pa[i] != pa)
1665 			break;
1666 		len += m_epg_pagelen(m, i, 0);
1667 		pa += m_epg_pagelen(m, i, 0);
1668 	}
1669 	usgl->len0 = htobe32(len);
1670 #ifdef INVARIANTS
1671 	nsegs--;
1672 #endif
1673 
1674 	j = -1;
1675 	for (; i < m->m_epg_npgs; i++) {
1676 		if (j == -1 || m->m_epg_pa[i] != pa) {
1677 			if (j >= 0)
1678 				usgl->sge[j / 2].len[j & 1] = htobe32(len);
1679 			j++;
1680 #ifdef INVARIANTS
1681 			nsegs--;
1682 #endif
1683 			pa = m->m_epg_pa[i];
1684 			usgl->sge[j / 2].addr[j & 1] = htobe64(pa);
1685 			len = m_epg_pagelen(m, i, 0);
1686 			pa += len;
1687 		} else {
1688 			len += m_epg_pagelen(m, i, 0);
1689 			pa += m_epg_pagelen(m, i, 0);
1690 		}
1691 	}
1692 	if (j >= 0) {
1693 		usgl->sge[j / 2].len[j & 1] = htobe32(len);
1694 
1695 		if ((j & 1) == 0)
1696 			usgl->sge[j / 2].len[1] = htobe32(0);
1697 	}
1698 	KASSERT(nsegs == 0, ("%s: nsegs %d, m %p", __func__, nsegs, m));
1699 }
1700 
1701 /*
1702  * Similar to t4_push_frames() but handles sockets that contain TLS
1703  * record mbufs.  Unlike TLSOM, each mbuf is a complete TLS record and
1704  * corresponds to a single work request.
1705  */
1706 void
1707 t4_push_ktls(struct adapter *sc, struct toepcb *toep, int drop)
1708 {
1709 	struct tls_hdr *thdr;
1710 	struct fw_tlstx_data_wr *txwr;
1711 	struct cpl_tx_tls_sfo *cpl;
1712 	struct wrqe *wr;
1713 	struct mbuf *m;
1714 	u_int nsegs, credits, wr_len;
1715 	u_int expn_size;
1716 	struct inpcb *inp = toep->inp;
1717 	struct tcpcb *tp = intotcpcb(inp);
1718 	struct socket *so = inp->inp_socket;
1719 	struct sockbuf *sb = &so->so_snd;
1720 	int tls_size, tx_credits, shove, sowwakeup;
1721 	struct ofld_tx_sdesc *txsd;
1722 	char *buf;
1723 
1724 	INP_WLOCK_ASSERT(inp);
1725 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1726 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1727 
1728 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
1729 	    ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS,
1730 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1731 	KASSERT(tls_tx_key(toep),
1732 	    ("%s: TX key not set for toep %p", __func__, toep));
1733 
1734 #ifdef VERBOSE_TRACES
1735 	CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
1736 	    __func__, toep->tid, toep->flags, tp->t_flags);
1737 #endif
1738 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1739 		return;
1740 
1741 #ifdef RATELIMIT
1742 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
1743 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
1744 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
1745 	}
1746 #endif
1747 
1748 	/*
1749 	 * This function doesn't resume by itself.  Someone else must clear the
1750 	 * flag and call this function.
1751 	 */
1752 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1753 		KASSERT(drop == 0,
1754 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1755 		return;
1756 	}
1757 
1758 	txsd = &toep->txsd[toep->txsd_pidx];
1759 	for (;;) {
1760 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1761 
1762 		SOCKBUF_LOCK(sb);
1763 		sowwakeup = drop;
1764 		if (drop) {
1765 			sbdrop_locked(sb, drop);
1766 			drop = 0;
1767 		}
1768 
1769 		m = sb->sb_sndptr != NULL ? sb->sb_sndptr->m_next : sb->sb_mb;
1770 
1771 		/*
1772 		 * Send a FIN if requested, but only if there's no
1773 		 * more data to send.
1774 		 */
1775 		if (m == NULL && toep->flags & TPF_SEND_FIN) {
1776 			if (sowwakeup)
1777 				sowwakeup_locked(so);
1778 			else
1779 				SOCKBUF_UNLOCK(sb);
1780 			SOCKBUF_UNLOCK_ASSERT(sb);
1781 			t4_close_conn(sc, toep);
1782 			return;
1783 		}
1784 
1785 		/*
1786 		 * If there is no ready data to send, wait until more
1787 		 * data arrives.
1788 		 */
1789 		if (m == NULL || (m->m_flags & M_NOTAVAIL) != 0) {
1790 			if (sowwakeup)
1791 				sowwakeup_locked(so);
1792 			else
1793 				SOCKBUF_UNLOCK(sb);
1794 			SOCKBUF_UNLOCK_ASSERT(sb);
1795 #ifdef VERBOSE_TRACES
1796 			CTR2(KTR_CXGBE, "%s: tid %d no ready data to send",
1797 			    __func__, toep->tid);
1798 #endif
1799 			return;
1800 		}
1801 
1802 		KASSERT(m->m_flags & M_EXTPG, ("%s: mbuf %p is not NOMAP",
1803 		    __func__, m));
1804 		KASSERT(m->m_epg_tls != NULL,
1805 		    ("%s: mbuf %p doesn't have TLS session", __func__, m));
1806 
1807 		/* Calculate WR length. */
1808 		wr_len = sizeof(struct fw_tlstx_data_wr) +
1809 		    sizeof(struct cpl_tx_tls_sfo) + key_size(toep);
1810 
1811 		/* Explicit IVs for AES-CBC and AES-GCM are <= 16. */
1812 		MPASS(toep->tls.iv_len <= AES_BLOCK_LEN);
1813 		wr_len += AES_BLOCK_LEN;
1814 
1815 		/* Account for SGL in work request length. */
1816 		nsegs = count_ext_pgs_segs(m);
1817 		wr_len += sizeof(struct ulptx_sgl) +
1818 		    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1819 
1820 		/* Not enough credits for this work request. */
1821 		if (howmany(wr_len, 16) > tx_credits) {
1822 			if (sowwakeup)
1823 				sowwakeup_locked(so);
1824 			else
1825 				SOCKBUF_UNLOCK(sb);
1826 			SOCKBUF_UNLOCK_ASSERT(sb);
1827 #ifdef VERBOSE_TRACES
1828 			CTR5(KTR_CXGBE,
1829 	    "%s: tid %d mbuf %p requires %d credits, but only %d available",
1830 			    __func__, toep->tid, m, howmany(wr_len, 16),
1831 			    tx_credits);
1832 #endif
1833 			toep->flags |= TPF_TX_SUSPENDED;
1834 			return;
1835 		}
1836 
1837 		/* Shove if there is no additional data pending. */
1838 		shove = ((m->m_next == NULL ||
1839 		    (m->m_next->m_flags & M_NOTAVAIL) != 0)) &&
1840 		    (tp->t_flags & TF_MORETOCOME) == 0;
1841 
1842 		if (sb->sb_flags & SB_AUTOSIZE &&
1843 		    V_tcp_do_autosndbuf &&
1844 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
1845 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
1846 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
1847 			    V_tcp_autosndbuf_max);
1848 
1849 			if (!sbreserve_locked(sb, newsize, so, NULL))
1850 				sb->sb_flags &= ~SB_AUTOSIZE;
1851 			else
1852 				sowwakeup = 1;	/* room available */
1853 		}
1854 		if (sowwakeup)
1855 			sowwakeup_locked(so);
1856 		else
1857 			SOCKBUF_UNLOCK(sb);
1858 		SOCKBUF_UNLOCK_ASSERT(sb);
1859 
1860 		if (__predict_false(toep->flags & TPF_FIN_SENT))
1861 			panic("%s: excess tx.", __func__);
1862 
1863 		wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1864 		if (wr == NULL) {
1865 			/* XXX: how will we recover from this? */
1866 			toep->flags |= TPF_TX_SUSPENDED;
1867 			return;
1868 		}
1869 
1870 		thdr = (struct tls_hdr *)&m->m_epg_hdr;
1871 #ifdef VERBOSE_TRACES
1872 		CTR5(KTR_CXGBE, "%s: tid %d TLS record %ju type %d len %#x",
1873 		    __func__, toep->tid, m->m_epg_seqno, thdr->type,
1874 		    m->m_len);
1875 #endif
1876 		txwr = wrtod(wr);
1877 		cpl = (struct cpl_tx_tls_sfo *)(txwr + 1);
1878 		memset(txwr, 0, roundup2(wr_len, 16));
1879 		credits = howmany(wr_len, 16);
1880 		expn_size = m->m_epg_hdrlen +
1881 		    m->m_epg_trllen;
1882 		tls_size = m->m_len - expn_size;
1883 		write_tlstx_wr(txwr, toep, 0,
1884 		    tls_size, expn_size, 1, credits, shove, 1);
1885 		toep->tls.tx_seq_no = m->m_epg_seqno;
1886 		write_tlstx_cpl(cpl, toep, thdr, tls_size, 1);
1887 		tls_copy_tx_key(toep, cpl + 1);
1888 
1889 		/* Copy IV. */
1890 		buf = (char *)(cpl + 1) + key_size(toep);
1891 		memcpy(buf, thdr + 1, toep->tls.iv_len);
1892 		buf += AES_BLOCK_LEN;
1893 
1894 		write_ktlstx_sgl(buf, m, nsegs);
1895 
1896 		KASSERT(toep->tx_credits >= credits,
1897 			("%s: not enough credits", __func__));
1898 
1899 		toep->tx_credits -= credits;
1900 
1901 		tp->snd_nxt += m->m_len;
1902 		tp->snd_max += m->m_len;
1903 
1904 		SOCKBUF_LOCK(sb);
1905 		sb->sb_sndptr = m;
1906 		SOCKBUF_UNLOCK(sb);
1907 
1908 		toep->flags |= TPF_TX_DATA_SENT;
1909 		if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep))
1910 			toep->flags |= TPF_TX_SUSPENDED;
1911 
1912 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1913 		txsd->plen = m->m_len;
1914 		txsd->tx_credits = credits;
1915 		txsd++;
1916 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1917 			toep->txsd_pidx = 0;
1918 			txsd = &toep->txsd[0];
1919 		}
1920 		toep->txsd_avail--;
1921 
1922 		atomic_add_long(&toep->vi->pi->tx_toe_tls_records, 1);
1923 		atomic_add_long(&toep->vi->pi->tx_toe_tls_octets, m->m_len);
1924 
1925 		t4_l2t_send(sc, wr, toep->l2te);
1926 	}
1927 }
1928 #endif
1929 
1930 /*
1931  * For TLS data we place received mbufs received via CPL_TLS_DATA into
1932  * an mbufq in the TLS offload state.  When CPL_RX_TLS_CMP is
1933  * received, the completed PDUs are placed into the socket receive
1934  * buffer.
1935  *
1936  * The TLS code reuses the ulp_pdu_reclaimq to hold the pending mbufs.
1937  */
1938 static int
1939 do_tls_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1940 {
1941 	struct adapter *sc = iq->adapter;
1942 	const struct cpl_tls_data *cpl = mtod(m, const void *);
1943 	unsigned int tid = GET_TID(cpl);
1944 	struct toepcb *toep = lookup_tid(sc, tid);
1945 	struct inpcb *inp = toep->inp;
1946 	struct tcpcb *tp;
1947 	int len;
1948 
1949 	/* XXX: Should this match do_rx_data instead? */
1950 	KASSERT(!(toep->flags & TPF_SYNQE),
1951 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
1952 
1953 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
1954 
1955 	/* strip off CPL header */
1956 	m_adj(m, sizeof(*cpl));
1957 	len = m->m_pkthdr.len;
1958 
1959 	atomic_add_long(&toep->vi->pi->rx_toe_tls_octets, len);
1960 
1961 	KASSERT(len == G_CPL_TLS_DATA_LENGTH(be32toh(cpl->length_pkd)),
1962 	    ("%s: payload length mismatch", __func__));
1963 
1964 	INP_WLOCK(inp);
1965 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1966 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1967 		    __func__, tid, len, inp->inp_flags);
1968 		INP_WUNLOCK(inp);
1969 		m_freem(m);
1970 		return (0);
1971 	}
1972 
1973 	/* Save TCP sequence number. */
1974 	m->m_pkthdr.tls_tcp_seq = be32toh(cpl->seq);
1975 
1976 	if (mbufq_enqueue(&toep->ulp_pdu_reclaimq, m)) {
1977 #ifdef INVARIANTS
1978 		panic("Failed to queue TLS data packet");
1979 #else
1980 		printf("%s: Failed to queue TLS data packet\n", __func__);
1981 		INP_WUNLOCK(inp);
1982 		m_freem(m);
1983 		return (0);
1984 #endif
1985 	}
1986 
1987 	tp = intotcpcb(inp);
1988 	tp->t_rcvtime = ticks;
1989 
1990 #ifdef VERBOSE_TRACES
1991 	CTR4(KTR_CXGBE, "%s: tid %u len %d seq %u", __func__, tid, len,
1992 	    be32toh(cpl->seq));
1993 #endif
1994 
1995 	INP_WUNLOCK(inp);
1996 	return (0);
1997 }
1998 
1999 static int
2000 do_rx_tls_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
2001 {
2002 	struct adapter *sc = iq->adapter;
2003 	const struct cpl_rx_tls_cmp *cpl = mtod(m, const void *);
2004 	struct tlsrx_hdr_pkt *tls_hdr_pkt;
2005 	unsigned int tid = GET_TID(cpl);
2006 	struct toepcb *toep = lookup_tid(sc, tid);
2007 	struct inpcb *inp = toep->inp;
2008 	struct tcpcb *tp;
2009 	struct socket *so;
2010 	struct sockbuf *sb;
2011 	struct mbuf *tls_data;
2012 #ifdef KERN_TLS
2013 	struct tls_get_record *tgr;
2014 	struct mbuf *control;
2015 #endif
2016 	int len, pdu_length, rx_credits;
2017 
2018 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
2019 	KASSERT(!(toep->flags & TPF_SYNQE),
2020 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
2021 
2022 	/* strip off CPL header */
2023 	m_adj(m, sizeof(*cpl));
2024 	len = m->m_pkthdr.len;
2025 
2026 	atomic_add_long(&toep->vi->pi->rx_toe_tls_records, 1);
2027 
2028 	KASSERT(len == G_CPL_RX_TLS_CMP_LENGTH(be32toh(cpl->pdulength_length)),
2029 	    ("%s: payload length mismatch", __func__));
2030 
2031 	INP_WLOCK(inp);
2032 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
2033 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
2034 		    __func__, tid, len, inp->inp_flags);
2035 		INP_WUNLOCK(inp);
2036 		m_freem(m);
2037 		return (0);
2038 	}
2039 
2040 	pdu_length = G_CPL_RX_TLS_CMP_PDULENGTH(be32toh(cpl->pdulength_length));
2041 
2042 	so = inp_inpcbtosocket(inp);
2043 	tp = intotcpcb(inp);
2044 
2045 #ifdef VERBOSE_TRACES
2046 	CTR6(KTR_CXGBE, "%s: tid %u PDU len %d len %d seq %u, rcv_nxt %u",
2047 	    __func__, tid, pdu_length, len, be32toh(cpl->seq), tp->rcv_nxt);
2048 #endif
2049 
2050 	tp->rcv_nxt += pdu_length;
2051 	KASSERT(tp->rcv_wnd >= pdu_length,
2052 	    ("%s: negative window size", __func__));
2053 	tp->rcv_wnd -= pdu_length;
2054 
2055 	/* XXX: Not sure what to do about urgent data. */
2056 
2057 	/*
2058 	 * The payload of this CPL is the TLS header followed by
2059 	 * additional fields.
2060 	 */
2061 	KASSERT(m->m_len >= sizeof(*tls_hdr_pkt),
2062 	    ("%s: payload too small", __func__));
2063 	tls_hdr_pkt = mtod(m, void *);
2064 
2065 	tls_data = mbufq_dequeue(&toep->ulp_pdu_reclaimq);
2066 	if (tls_data != NULL) {
2067 		KASSERT(be32toh(cpl->seq) == tls_data->m_pkthdr.tls_tcp_seq,
2068 		    ("%s: sequence mismatch", __func__));
2069 	}
2070 
2071 #ifdef KERN_TLS
2072 	if (toep->tls.mode == TLS_MODE_KTLS) {
2073 		/* Report decryption errors as EBADMSG. */
2074 		if ((tls_hdr_pkt->res_to_mac_error & M_TLSRX_HDR_PKT_ERROR) !=
2075 		    0) {
2076 			m_freem(m);
2077 			m_freem(tls_data);
2078 
2079 			CURVNET_SET(toep->vnet);
2080 			so->so_error = EBADMSG;
2081 			sorwakeup(so);
2082 
2083 			INP_WUNLOCK(inp);
2084 			CURVNET_RESTORE();
2085 
2086 			return (0);
2087 		}
2088 
2089 		/* Allocate the control message mbuf. */
2090 		control = sbcreatecontrol(NULL, sizeof(*tgr), TLS_GET_RECORD,
2091 		    IPPROTO_TCP);
2092 		if (control == NULL) {
2093 			m_freem(m);
2094 			m_freem(tls_data);
2095 
2096 			CURVNET_SET(toep->vnet);
2097 			so->so_error = ENOBUFS;
2098 			sorwakeup(so);
2099 
2100 			INP_WUNLOCK(inp);
2101 			CURVNET_RESTORE();
2102 
2103 			return (0);
2104 		}
2105 
2106 		tgr = (struct tls_get_record *)
2107 		    CMSG_DATA(mtod(control, struct cmsghdr *));
2108 		tgr->tls_type = tls_hdr_pkt->type;
2109 		tgr->tls_vmajor = be16toh(tls_hdr_pkt->version) >> 8;
2110 		tgr->tls_vminor = be16toh(tls_hdr_pkt->version) & 0xff;
2111 
2112 		m_freem(m);
2113 
2114 		if (tls_data != NULL) {
2115 			m_last(tls_data)->m_flags |= M_EOR;
2116 			tgr->tls_length = htobe16(tls_data->m_pkthdr.len);
2117 		} else
2118 			tgr->tls_length = 0;
2119 		m = tls_data;
2120 	} else
2121 #endif
2122 	{
2123 		/*
2124 		 * Only the TLS header is sent to OpenSSL, so report
2125 		 * errors by altering the record type.
2126 		 */
2127 		if ((tls_hdr_pkt->res_to_mac_error & M_TLSRX_HDR_PKT_ERROR) !=
2128 		    0)
2129 			tls_hdr_pkt->type = CONTENT_TYPE_ERROR;
2130 
2131 		/* Trim this CPL's mbuf to only include the TLS header. */
2132 		KASSERT(m->m_len == len && m->m_next == NULL,
2133 		    ("%s: CPL spans multiple mbufs", __func__));
2134 		m->m_len = TLS_HEADER_LENGTH;
2135 		m->m_pkthdr.len = TLS_HEADER_LENGTH;
2136 
2137 		if (tls_data != NULL) {
2138 			/*
2139 			 * Update the TLS header length to be the length of
2140 			 * the payload data.
2141 			 */
2142 			tls_hdr_pkt->length = htobe16(tls_data->m_pkthdr.len);
2143 
2144 			m->m_next = tls_data;
2145 			m->m_pkthdr.len += tls_data->m_len;
2146 		}
2147 
2148 #ifdef KERN_TLS
2149 		control = NULL;
2150 #endif
2151 	}
2152 
2153 	sb = &so->so_rcv;
2154 	SOCKBUF_LOCK(sb);
2155 
2156 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
2157 		struct epoch_tracker et;
2158 
2159 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
2160 		    __func__, tid, pdu_length);
2161 		m_freem(m);
2162 #ifdef KERN_TLS
2163 		m_freem(control);
2164 #endif
2165 		SOCKBUF_UNLOCK(sb);
2166 		INP_WUNLOCK(inp);
2167 
2168 		CURVNET_SET(toep->vnet);
2169 		NET_EPOCH_ENTER(et);
2170 		INP_WLOCK(inp);
2171 		tp = tcp_drop(tp, ECONNRESET);
2172 		if (tp)
2173 			INP_WUNLOCK(inp);
2174 		NET_EPOCH_EXIT(et);
2175 		CURVNET_RESTORE();
2176 
2177 		return (0);
2178 	}
2179 
2180 	/*
2181 	 * Not all of the bytes on the wire are included in the socket buffer
2182 	 * (e.g. the MAC of the TLS record).  However, those bytes are included
2183 	 * in the TCP sequence space.
2184 	 */
2185 
2186 	/* receive buffer autosize */
2187 	MPASS(toep->vnet == so->so_vnet);
2188 	CURVNET_SET(toep->vnet);
2189 	if (sb->sb_flags & SB_AUTOSIZE &&
2190 	    V_tcp_do_autorcvbuf &&
2191 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
2192 	    m->m_pkthdr.len > (sbspace(sb) / 8 * 7)) {
2193 		unsigned int hiwat = sb->sb_hiwat;
2194 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
2195 		    V_tcp_autorcvbuf_max);
2196 
2197 		if (!sbreserve_locked(sb, newsize, so, NULL))
2198 			sb->sb_flags &= ~SB_AUTOSIZE;
2199 	}
2200 
2201 #ifdef KERN_TLS
2202 	if (control != NULL)
2203 		sbappendcontrol_locked(sb, m, control, 0);
2204 	else
2205 #endif
2206 		sbappendstream_locked(sb, m, 0);
2207 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
2208 #ifdef VERBOSE_TRACES
2209 	CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u",
2210 	    __func__, tid, rx_credits, tp->rcv_wnd);
2211 #endif
2212 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
2213 		rx_credits = send_rx_credits(sc, toep, rx_credits);
2214 		tp->rcv_wnd += rx_credits;
2215 		tp->rcv_adv += rx_credits;
2216 	}
2217 
2218 	sorwakeup_locked(so);
2219 	SOCKBUF_UNLOCK_ASSERT(sb);
2220 
2221 	INP_WUNLOCK(inp);
2222 	CURVNET_RESTORE();
2223 	return (0);
2224 }
2225 
2226 void
2227 do_rx_data_tls(const struct cpl_rx_data *cpl, struct toepcb *toep,
2228     struct mbuf *m)
2229 {
2230 	struct inpcb *inp = toep->inp;
2231 	struct tls_ofld_info *tls_ofld = &toep->tls;
2232 	struct tls_hdr *hdr;
2233 	struct tcpcb *tp;
2234 	struct socket *so;
2235 	struct sockbuf *sb;
2236 	int error, len, rx_credits;
2237 
2238 	len = m->m_pkthdr.len;
2239 
2240 	INP_WLOCK_ASSERT(inp);
2241 
2242 	so = inp_inpcbtosocket(inp);
2243 	tp = intotcpcb(inp);
2244 	sb = &so->so_rcv;
2245 	SOCKBUF_LOCK(sb);
2246 	CURVNET_SET(toep->vnet);
2247 
2248 	tp->rcv_nxt += len;
2249 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
2250 	tp->rcv_wnd -= len;
2251 
2252 	/* Do we have a full TLS header? */
2253 	if (len < sizeof(*hdr)) {
2254 		CTR3(KTR_CXGBE, "%s: tid %u len %d: too short for a TLS header",
2255 		    __func__, toep->tid, len);
2256 		so->so_error = EMSGSIZE;
2257 		goto out;
2258 	}
2259 	hdr = mtod(m, struct tls_hdr *);
2260 
2261 	/* Is the header valid? */
2262 	if (be16toh(hdr->version) != tls_ofld->k_ctx.proto_ver) {
2263 		CTR3(KTR_CXGBE, "%s: tid %u invalid version %04x",
2264 		    __func__, toep->tid, be16toh(hdr->version));
2265 		error = EINVAL;
2266 		goto report_error;
2267 	}
2268 	if (be16toh(hdr->length) < sizeof(*hdr)) {
2269 		CTR3(KTR_CXGBE, "%s: tid %u invalid length %u",
2270 		    __func__, toep->tid, be16toh(hdr->length));
2271 		error = EBADMSG;
2272 		goto report_error;
2273 	}
2274 
2275 	/* Did we get a truncated record? */
2276 	if (len < be16toh(hdr->length)) {
2277 		CTR4(KTR_CXGBE, "%s: tid %u truncated TLS record (%d vs %u)",
2278 		    __func__, toep->tid, len, be16toh(hdr->length));
2279 
2280 		error = EMSGSIZE;
2281 		goto report_error;
2282 	}
2283 
2284 	/* Is the header type unknown? */
2285 	switch (hdr->type) {
2286 	case CONTENT_TYPE_CCS:
2287 	case CONTENT_TYPE_ALERT:
2288 	case CONTENT_TYPE_APP_DATA:
2289 	case CONTENT_TYPE_HANDSHAKE:
2290 		break;
2291 	default:
2292 		CTR3(KTR_CXGBE, "%s: tid %u invalid TLS record type %u",
2293 		    __func__, toep->tid, hdr->type);
2294 		error = EBADMSG;
2295 		goto report_error;
2296 	}
2297 
2298 	/*
2299 	 * Just punt.  Although this could fall back to software
2300 	 * decryption, this case should never really happen.
2301 	 */
2302 	CTR4(KTR_CXGBE, "%s: tid %u dropping TLS record type %u, length %u",
2303 	    __func__, toep->tid, hdr->type, be16toh(hdr->length));
2304 	error = EBADMSG;
2305 
2306 report_error:
2307 #ifdef KERN_TLS
2308 	if (toep->tls.mode == TLS_MODE_KTLS)
2309 		so->so_error = error;
2310 	else
2311 #endif
2312 	{
2313 		/*
2314 		 * Report errors by sending an empty TLS record
2315 		 * with an error record type.
2316 		 */
2317 		hdr->type = CONTENT_TYPE_ERROR;
2318 
2319 		/* Trim this CPL's mbuf to only include the TLS header. */
2320 		KASSERT(m->m_len == len && m->m_next == NULL,
2321 		    ("%s: CPL spans multiple mbufs", __func__));
2322 		m->m_len = TLS_HEADER_LENGTH;
2323 		m->m_pkthdr.len = TLS_HEADER_LENGTH;
2324 
2325 		sbappendstream_locked(sb, m, 0);
2326 		m = NULL;
2327 	}
2328 
2329 out:
2330 	/*
2331 	 * This connection is going to die anyway, so probably don't
2332 	 * need to bother with returning credits.
2333 	 */
2334 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
2335 #ifdef VERBOSE_TRACES
2336 	CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u",
2337 	    __func__, toep->tid, rx_credits, tp->rcv_wnd);
2338 #endif
2339 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
2340 		rx_credits = send_rx_credits(toep->vi->adapter, toep,
2341 		    rx_credits);
2342 		tp->rcv_wnd += rx_credits;
2343 		tp->rcv_adv += rx_credits;
2344 	}
2345 
2346 	sorwakeup_locked(so);
2347 	SOCKBUF_UNLOCK_ASSERT(sb);
2348 
2349 	INP_WUNLOCK(inp);
2350 	CURVNET_RESTORE();
2351 
2352 	m_freem(m);
2353 }
2354 
2355 void
2356 t4_tls_mod_load(void)
2357 {
2358 
2359 	mtx_init(&tls_handshake_lock, "t4tls handshake", NULL, MTX_DEF);
2360 	t4_register_cpl_handler(CPL_TLS_DATA, do_tls_data);
2361 	t4_register_cpl_handler(CPL_RX_TLS_CMP, do_rx_tls_cmp);
2362 }
2363 
2364 void
2365 t4_tls_mod_unload(void)
2366 {
2367 
2368 	t4_register_cpl_handler(CPL_TLS_DATA, NULL);
2369 	t4_register_cpl_handler(CPL_RX_TLS_CMP, NULL);
2370 	mtx_destroy(&tls_handshake_lock);
2371 }
2372 #endif	/* TCP_OFFLOAD */
2373