xref: /freebsd/sys/dev/cxgbe/crypto/t7_kern_tls.c (revision ed5fc8066f98e639f624de9997b33727ed883c32)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2025 Chelsio Communications
5  * Written by: John Baldwin <jhb@FreeBSD.org>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_kern_tls.h"
32 
33 #include <sys/param.h>
34 #include <sys/ktr.h>
35 #include <sys/ktls.h>
36 #include <sys/sglist.h>
37 #include <sys/socket.h>
38 #include <sys/socketvar.h>
39 #include <sys/sockbuf.h>
40 #include <netinet/in.h>
41 #include <netinet/in_pcb.h>
42 #include <netinet/ip.h>
43 #include <netinet/ip6.h>
44 #include <netinet/tcp_var.h>
45 #include <opencrypto/cryptodev.h>
46 #include <opencrypto/xform.h>
47 #include <vm/vm.h>
48 #include <vm/pmap.h>
49 
50 #include "common/common.h"
51 #include "common/t4_regs.h"
52 #include "common/t4_regs_values.h"
53 #include "common/t4_tcb.h"
54 #include "t4_l2t.h"
55 #include "t4_clip.h"
56 #include "t4_mp_ring.h"
57 #include "crypto/t4_crypto.h"
58 
59 #if defined(INET) || defined(INET6)
60 
61 #define TLS_HEADER_LENGTH		5
62 
63 struct tls_scmd {
64 	__be32 seqno_numivs;
65 	__be32 ivgen_hdrlen;
66 };
67 
68 struct tlspcb {
69 	struct m_snd_tag com;
70 	struct vi_info *vi;	/* virtual interface */
71 	struct adapter *sc;
72 	struct sge_txq *txq;
73 
74 	int tx_key_addr;
75 	bool inline_key;
76 	bool tls13;
77 	unsigned char enc_mode;
78 
79 	struct tls_scmd scmd0;
80 	struct tls_scmd scmd0_partial;
81 	struct tls_scmd scmd0_short;
82 
83 	unsigned int tx_key_info_size;
84 
85 	uint16_t prev_mss;
86 
87 	/* Fields used for GCM records using GHASH state. */
88 	uint16_t ghash_offset;
89 	uint64_t ghash_tls_seqno;
90 	char ghash[AES_GMAC_HASH_LEN];
91 	bool ghash_valid;
92 	bool ghash_pending;
93 	bool ghash_lcb;
94 	bool queue_mbufs;
95 	uint8_t rx_chid;
96 	uint16_t rx_qid;
97 	struct mbufq pending_mbufs;
98 
99 	/*
100 	 * Only used outside of setup and teardown when using inline
101 	 * keys or for partial GCM mode.
102 	 */
103 	struct tls_keyctx keyctx;
104 };
105 
106 static void t7_tls_tag_free(struct m_snd_tag *mst);
107 static int ktls_setup_keys(struct tlspcb *tlsp,
108     const struct ktls_session *tls, struct sge_txq *txq);
109 
110 static void *zero_buffer;
111 static vm_paddr_t zero_buffer_pa;
112 
113 static const struct if_snd_tag_sw t7_tls_tag_sw = {
114 	.snd_tag_free = t7_tls_tag_free,
115 	.type = IF_SND_TAG_TYPE_TLS
116 };
117 
118 static inline struct tlspcb *
mst_to_tls(struct m_snd_tag * t)119 mst_to_tls(struct m_snd_tag *t)
120 {
121 	return (__containerof(t, struct tlspcb, com));
122 }
123 
124 static struct tlspcb *
alloc_tlspcb(struct ifnet * ifp,struct vi_info * vi,int flags)125 alloc_tlspcb(struct ifnet *ifp, struct vi_info *vi, int flags)
126 {
127 	struct port_info *pi = vi->pi;
128 	struct adapter *sc = pi->adapter;
129 	struct tlspcb *tlsp;
130 
131 	tlsp = malloc(sizeof(*tlsp), M_CXGBE, M_ZERO | flags);
132 	if (tlsp == NULL)
133 		return (NULL);
134 
135 	m_snd_tag_init(&tlsp->com, ifp, &t7_tls_tag_sw);
136 	tlsp->vi = vi;
137 	tlsp->sc = sc;
138 	tlsp->tx_key_addr = -1;
139 	tlsp->ghash_offset = -1;
140 	tlsp->rx_chid = pi->rx_chan;
141 	tlsp->rx_qid = -1;
142 	tlsp->txq = NULL;
143 	mbufq_init(&tlsp->pending_mbufs, INT_MAX);
144 
145 	return (tlsp);
146 }
147 
148 int
t7_tls_tag_alloc(struct ifnet * ifp,union if_snd_tag_alloc_params * params,struct m_snd_tag ** pt)149 t7_tls_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params,
150     struct m_snd_tag **pt)
151 {
152 	const struct ktls_session *tls;
153 	struct tlspcb *tlsp;
154 	struct adapter *sc;
155 	struct vi_info *vi;
156 	struct inpcb *inp;
157 	struct sge_txq *txq;
158 	int error, iv_size, keyid, mac_first, qidx;
159 	uint32_t flowid;
160 
161 	tls = params->tls.tls;
162 
163 	/* TLS 1.1 through TLS 1.3 are currently supported. */
164 	if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE ||
165 	    tls->params.tls_vminor < TLS_MINOR_VER_ONE ||
166 	    tls->params.tls_vminor > TLS_MINOR_VER_THREE)
167 		return (EPROTONOSUPPORT);
168 
169 	/* Sanity check values in *tls. */
170 	switch (tls->params.cipher_algorithm) {
171 	case CRYPTO_AES_CBC:
172 		/* XXX: Explicitly ignore any provided IV. */
173 		switch (tls->params.cipher_key_len) {
174 		case 128 / 8:
175 		case 192 / 8:
176 		case 256 / 8:
177 			break;
178 		default:
179 			return (EINVAL);
180 		}
181 		switch (tls->params.auth_algorithm) {
182 		case CRYPTO_SHA1_HMAC:
183 		case CRYPTO_SHA2_256_HMAC:
184 		case CRYPTO_SHA2_384_HMAC:
185 			break;
186 		default:
187 			return (EPROTONOSUPPORT);
188 		}
189 		iv_size = AES_BLOCK_LEN;
190 		mac_first = 1;
191 		break;
192 	case CRYPTO_AES_NIST_GCM_16:
193 		switch (tls->params.cipher_key_len) {
194 		case 128 / 8:
195 		case 192 / 8:
196 		case 256 / 8:
197 			break;
198 		default:
199 			return (EINVAL);
200 		}
201 
202 		/*
203 		 * The IV size for TLS 1.2 is the explicit IV in the
204 		 * record header.  For TLS 1.3 it is the size of the
205 		 * sequence number.
206 		 */
207 		iv_size = 8;
208 		mac_first = 0;
209 		break;
210 	default:
211 		return (EPROTONOSUPPORT);
212 	}
213 
214 	vi = if_getsoftc(ifp);
215 	sc = vi->adapter;
216 
217 	tlsp = alloc_tlspcb(ifp, vi, M_WAITOK);
218 
219 	/*
220 	 * Pointers with the low bit set in the pointer can't
221 	 * be stored as the cookie in the CPL_FW6_PLD reply.
222 	 */
223 	if (((uintptr_t)tlsp & CPL_FW6_COOKIE_MASK) != 0) {
224 		error = EINVAL;
225 		goto failed;
226 	}
227 
228 	tlsp->tls13 = tls->params.tls_vminor == TLS_MINOR_VER_THREE;
229 
230 	if (sc->tlst.inline_keys)
231 		keyid = -1;
232 	else
233 		keyid = t4_alloc_tls_keyid(sc);
234 	if (keyid < 0) {
235 		CTR(KTR_CXGBE, "%s: %p using immediate key ctx", __func__,
236 		    tlsp);
237 		tlsp->inline_key = true;
238 	} else {
239 		tlsp->tx_key_addr = keyid;
240 		CTR(KTR_CXGBE, "%s: %p allocated TX key addr %#x", __func__,
241 		    tlsp, tlsp->tx_key_addr);
242 	}
243 
244 	inp = params->tls.inp;
245 	INP_RLOCK(inp);
246 	if (intotcpcb(inp)->t_flags & TF_DISCONNECTED) {
247 		INP_RUNLOCK(inp);
248 		error = ECONNRESET;
249 		goto failed;
250 	}
251 
252 	if (inp->inp_flowtype != M_HASHTYPE_NONE)
253 		flowid = inp->inp_flowid;
254 	else
255 		flowid = arc4random();
256 	qidx = flowid % vi->nrxq + vi->first_rxq;
257 	tlsp->rx_qid = sc->sge.rxq[qidx].iq.abs_id;
258 	qidx = (flowid % (vi->ntxq - vi->rsrv_noflowq)) + vi->rsrv_noflowq +
259 	    vi->first_txq;
260 	tlsp->txq = txq = &sc->sge.txq[qidx];
261 	INP_RUNLOCK(inp);
262 
263 	error = ktls_setup_keys(tlsp, tls, txq);
264 	if (error)
265 		goto failed;
266 
267 	tlsp->enc_mode = t4_tls_cipher_mode(tls);
268 	tlsp->tx_key_info_size = t4_tls_key_info_size(tls);
269 
270 	/* The SCMD fields used when encrypting a full TLS record. */
271 	if (tlsp->tls13)
272 		tlsp->scmd0.seqno_numivs = V_SCMD_SEQ_NO_CTRL(0);
273 	else
274 		tlsp->scmd0.seqno_numivs = V_SCMD_SEQ_NO_CTRL(3);
275 	tlsp->scmd0.seqno_numivs |=
276 	    V_SCMD_PROTO_VERSION(t4_tls_proto_ver(tls)) |
277 	    V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
278 	    V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) |
279 	    V_SCMD_CIPH_MODE(tlsp->enc_mode) |
280 	    V_SCMD_AUTH_MODE(t4_tls_auth_mode(tls)) |
281 	    V_SCMD_HMAC_CTRL(t4_tls_hmac_ctrl(tls)) |
282 	    V_SCMD_IV_SIZE(iv_size / 2) | V_SCMD_NUM_IVS(1);
283 	tlsp->scmd0.seqno_numivs = htobe32(tlsp->scmd0.seqno_numivs);
284 
285 	tlsp->scmd0.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) |
286 	    V_SCMD_TLS_FRAG_ENABLE(0);
287 	if (tlsp->inline_key)
288 		tlsp->scmd0.ivgen_hdrlen |= V_SCMD_KEY_CTX_INLINE(1);
289 
290 	/*
291 	 * The SCMD fields used when encrypting a short TLS record
292 	 * (no trailer and possibly a truncated payload).
293 	 */
294 	tlsp->scmd0_short.seqno_numivs = V_SCMD_SEQ_NO_CTRL(0) |
295 	    V_SCMD_PROTO_VERSION(SCMD_PROTO_VERSION_GENERIC) |
296 	    V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
297 	    V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) |
298 	    V_SCMD_AUTH_MODE(SCMD_AUTH_MODE_NOP) |
299 	    V_SCMD_HMAC_CTRL(SCMD_HMAC_CTRL_NOP) |
300 	    V_SCMD_IV_SIZE(AES_BLOCK_LEN / 2) | V_SCMD_NUM_IVS(0);
301 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM)
302 		tlsp->scmd0_short.seqno_numivs |=
303 		    V_SCMD_CIPH_MODE(SCMD_CIPH_MODE_AES_CTR);
304 	else
305 		tlsp->scmd0_short.seqno_numivs |=
306 		    V_SCMD_CIPH_MODE(tlsp->enc_mode);
307 	tlsp->scmd0_short.seqno_numivs =
308 	    htobe32(tlsp->scmd0_short.seqno_numivs);
309 
310 	tlsp->scmd0_short.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) |
311 	    V_SCMD_TLS_FRAG_ENABLE(0) | V_SCMD_AADIVDROP(1);
312 	if (tlsp->inline_key)
313 		tlsp->scmd0_short.ivgen_hdrlen |= V_SCMD_KEY_CTX_INLINE(1);
314 
315 	/*
316 	 * The SCMD fields used when encrypting a short TLS record
317 	 * using a partial GHASH.
318 	 */
319 	tlsp->scmd0_partial.seqno_numivs = V_SCMD_SEQ_NO_CTRL(0) |
320 	    V_SCMD_PROTO_VERSION(SCMD_PROTO_VERSION_GENERIC) |
321 	    V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
322 	    V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) |
323 	    V_SCMD_CIPH_MODE(tlsp->enc_mode) |
324 	    V_SCMD_AUTH_MODE(t4_tls_auth_mode(tls)) |
325 	    V_SCMD_HMAC_CTRL(t4_tls_hmac_ctrl(tls)) |
326 	    V_SCMD_IV_SIZE(AES_BLOCK_LEN / 2) | V_SCMD_NUM_IVS(1);
327 	tlsp->scmd0_partial.seqno_numivs =
328 	    htobe32(tlsp->scmd0_partial.seqno_numivs);
329 
330 	tlsp->scmd0_partial.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) |
331 	    V_SCMD_TLS_FRAG_ENABLE(0) | V_SCMD_AADIVDROP(1) |
332 	    V_SCMD_KEY_CTX_INLINE(1);
333 
334 	TXQ_LOCK(txq);
335 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM)
336 		txq->kern_tls_gcm++;
337 	else
338 		txq->kern_tls_cbc++;
339 	TXQ_UNLOCK(txq);
340 	*pt = &tlsp->com;
341 	return (0);
342 
343 failed:
344 	m_snd_tag_rele(&tlsp->com);
345 	return (error);
346 }
347 
348 static int
ktls_setup_keys(struct tlspcb * tlsp,const struct ktls_session * tls,struct sge_txq * txq)349 ktls_setup_keys(struct tlspcb *tlsp, const struct ktls_session *tls,
350     struct sge_txq *txq)
351 {
352 	struct tls_key_req *kwr;
353 	struct tls_keyctx *kctx;
354 	void *items[1];
355 	struct mbuf *m;
356 	int error;
357 
358 	/*
359 	 * Store the salt and keys in the key context.  For
360 	 * connections with an inline key, this key context is passed
361 	 * as immediate data in each work request.  For connections
362 	 * storing the key in DDR, a work request is used to store a
363 	 * copy of the key context in DDR.
364 	 */
365 	t4_tls_key_ctx(tls, KTLS_TX, &tlsp->keyctx);
366 	if (tlsp->inline_key)
367 		return (0);
368 
369 	/* Populate key work request. */
370         m = alloc_wr_mbuf(TLS_KEY_WR_SZ, M_NOWAIT);
371 	if (m == NULL) {
372 		CTR(KTR_CXGBE, "%s: %p failed to alloc WR mbuf", __func__,
373 		    tlsp);
374 		return (ENOMEM);
375 	}
376 	m->m_pkthdr.snd_tag = m_snd_tag_ref(&tlsp->com);
377 	m->m_pkthdr.csum_flags |= CSUM_SND_TAG;
378 	kwr = mtod(m, void *);
379 	memset(kwr, 0, TLS_KEY_WR_SZ);
380 
381 	t4_write_tlskey_wr(tls, KTLS_TX, 0, 0, tlsp->tx_key_addr, kwr);
382 	kctx = (struct tls_keyctx *)(kwr + 1);
383 	memcpy(kctx, &tlsp->keyctx, sizeof(*kctx));
384 
385 	/*
386 	 * Place the key work request in the transmit queue.  It
387 	 * should be sent to the NIC before any TLS packets using this
388 	 * session.
389 	 */
390 	items[0] = m;
391 	error = mp_ring_enqueue(txq->r, items, 1, 1);
392 	if (error)
393 		m_free(m);
394 	else
395 		CTR(KTR_CXGBE, "%s: %p sent key WR", __func__, tlsp);
396 	return (error);
397 }
398 
399 static u_int
ktls_base_wr_size(struct tlspcb * tlsp,bool inline_key)400 ktls_base_wr_size(struct tlspcb *tlsp, bool inline_key)
401 {
402 	u_int wr_len;
403 
404 	wr_len = sizeof(struct fw_ulptx_wr);	// 16
405 	wr_len += sizeof(struct ulp_txpkt);	// 8
406 	wr_len += sizeof(struct ulptx_idata);	// 8
407 	wr_len += sizeof(struct cpl_tx_sec_pdu);// 32
408 	if (inline_key)
409 		wr_len += tlsp->tx_key_info_size;
410 	else {
411 		wr_len += sizeof(struct ulptx_sc_memrd);// 8
412 		wr_len += sizeof(struct ulptx_idata);	// 8
413 	}
414 	/* SplitMode CPL_RX_PHYS_DSGL here if needed. */
415 	/* CPL_TX_*_LSO here if needed. */
416 	wr_len += sizeof(struct cpl_tx_pkt_core);// 16
417 	return (wr_len);
418 }
419 
420 static u_int
ktls_sgl_size(u_int nsegs)421 ktls_sgl_size(u_int nsegs)
422 {
423 	u_int wr_len;
424 
425 	/* First segment is part of ulptx_sgl. */
426 	nsegs--;
427 
428 	wr_len = sizeof(struct ulptx_sgl);
429 	wr_len += 8 * ((3 * nsegs) / 2 + (nsegs & 1));
430 	return (wr_len);
431 }
432 
433 /*
434  * A request that doesn't need to generate the TLS trailer is a short
435  * record.  For these requests, part of the TLS record payload is
436  * encrypted without invoking the MAC.
437  *
438  * Returns true if this record should be sent as a short record.  In
439  * either case, the remaining outputs describe the how much of the
440  * TLS record to send as input to the crypto block and the amount of
441  * crypto output to trim via SplitMode:
442  *
443  * *header_len - Number of bytes of TLS header to pass as immediate
444  *               data
445  *
446  * *offset - Start offset of TLS record payload to pass as DSGL data
447  *
448  * *plen - Length of TLS record payload to pass as DSGL data
449  *
450  * *leading_waste - amount of non-packet-header bytes to drop at the
451  *                  start of the crypto output
452  *
453  * *trailing_waste - amount of crypto output to drop from the end
454  */
455 static bool
ktls_is_short_record(struct tlspcb * tlsp,struct mbuf * m_tls,u_int tlen,u_int rlen,u_int * header_len,u_int * offset,u_int * plen,u_int * leading_waste,u_int * trailing_waste,bool send_partial_ghash,bool request_ghash)456 ktls_is_short_record(struct tlspcb *tlsp, struct mbuf *m_tls, u_int tlen,
457     u_int rlen, u_int *header_len, u_int *offset, u_int *plen,
458     u_int *leading_waste, u_int *trailing_waste, bool send_partial_ghash,
459     bool request_ghash)
460 {
461 	u_int new_tlen, trailer_len;
462 
463 	MPASS(tlen > m_tls->m_epg_hdrlen);
464 
465 	/*
466 	 * For TLS 1.3 treat the inner record type stored as the first
467 	 * byte of the trailer as part of the payload rather than part
468 	 * of the trailer.
469 	 */
470 	trailer_len = m_tls->m_epg_trllen;
471 	if (tlsp->tls13)
472 		trailer_len--;
473 
474 	/*
475 	 * Default to sending the full record as input to the crypto
476 	 * engine and relying on SplitMode to drop any waste.
477 	 */
478 	*header_len = m_tls->m_epg_hdrlen;
479 	*offset = 0;
480 	*plen = rlen - (m_tls->m_epg_hdrlen + trailer_len);
481 	*leading_waste = mtod(m_tls, vm_offset_t);
482 	*trailing_waste = rlen - tlen;
483 	if (!tlsp->sc->tlst.short_records)
484 		return (false);
485 
486 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_CBC) {
487 		/*
488 		 * For AES-CBC we have to send input from the start of
489 		 * the TLS record payload that is a multiple of the
490 		 * block size.  new_tlen rounds up tlen to the end of
491 		 * the containing AES block.  If this last block
492 		 * overlaps with the trailer, send the full record to
493 		 * generate the MAC.
494 		 */
495 		new_tlen = TLS_HEADER_LENGTH +
496 		    roundup2(tlen - TLS_HEADER_LENGTH, AES_BLOCK_LEN);
497 		if (rlen - new_tlen < trailer_len)
498 			return (false);
499 
500 		*trailing_waste = new_tlen - tlen;
501 		*plen = new_tlen - m_tls->m_epg_hdrlen;
502 	} else {
503 		if (rlen - tlen < trailer_len ||
504 		    (rlen - tlen == trailer_len && request_ghash)) {
505 			/*
506 			 * For AES-GCM we have to send the full record
507 			 * if the end overlaps with the trailer and a
508 			 * partial GHASH isn't being sent.
509 			 */
510 			if (!send_partial_ghash)
511 				return (false);
512 
513 			/*
514 			 * Will need to treat any excess trailer bytes as
515 			 * trailing waste.  *trailing_waste is already
516 			 * correct.
517 			 */
518 		} else {
519 			/*
520 			 * We can use AES-CTR or AES-GCM in partial GHASH
521 			 * mode to encrypt a partial PDU.
522 			 *
523 			 * The last block can be partially encrypted
524 			 * without any trailing waste.
525 			 */
526 			*trailing_waste = 0;
527 			*plen = tlen - m_tls->m_epg_hdrlen;
528 		}
529 
530 		/*
531 		 * If this request starts at the first byte of the
532 		 * payload (so the previous request sent the full TLS
533 		 * header as a tunnel packet) and a partial GHASH is
534 		 * being requested, the full TLS header must be sent
535 		 * as input for the GHASH.
536 		 */
537 		if (mtod(m_tls, vm_offset_t) == m_tls->m_epg_hdrlen &&
538 		    request_ghash)
539 			return (true);
540 
541 		/*
542 		 * In addition, we can minimize leading waste by
543 		 * starting encryption at the start of the closest AES
544 		 * block.
545 		 */
546 		if (mtod(m_tls, vm_offset_t) >= m_tls->m_epg_hdrlen) {
547 			*header_len = 0;
548 			*offset = mtod(m_tls, vm_offset_t) -
549 			    m_tls->m_epg_hdrlen;
550 			if (*offset >= *plen)
551 				*offset = *plen;
552 			else
553 				*offset = rounddown2(*offset, AES_BLOCK_LEN);
554 
555 			/*
556 			 * If the request is just bytes from the trailer,
557 			 * trim the offset to the end of the payload.
558 			 */
559 			*offset = min(*offset, *plen);
560 			*plen -= *offset;
561 			*leading_waste -= (m_tls->m_epg_hdrlen + *offset);
562 		}
563 	}
564 	return (true);
565 }
566 
567 /* Size of the AES-GCM TLS AAD for a given connection. */
568 static int
ktls_gcm_aad_len(struct tlspcb * tlsp)569 ktls_gcm_aad_len(struct tlspcb *tlsp)
570 {
571 	return (tlsp->tls13 ? sizeof(struct tls_aead_data_13) :
572 	    sizeof(struct tls_aead_data));
573 }
574 
575 static int
ktls_wr_len(struct tlspcb * tlsp,struct mbuf * m,struct mbuf * m_tls,int * nsegsp)576 ktls_wr_len(struct tlspcb *tlsp, struct mbuf *m, struct mbuf *m_tls,
577     int *nsegsp)
578 {
579 	const struct tls_record_layer *hdr;
580 	u_int header_len, imm_len, offset, plen, rlen, tlen, wr_len;
581 	u_int leading_waste, trailing_waste;
582 	bool inline_key, last_ghash_frag, request_ghash, send_partial_ghash;
583 	bool short_record;
584 
585 	M_ASSERTEXTPG(m_tls);
586 
587 	/*
588 	 * The relative offset of the last byte to send from the TLS
589 	 * record.
590 	 */
591 	tlen = mtod(m_tls, vm_offset_t) + m_tls->m_len;
592 	if (tlen <= m_tls->m_epg_hdrlen) {
593 		/*
594 		 * For requests that only want to send the TLS header,
595 		 * send a tunnelled packet as immediate data.
596 		 */
597 		wr_len = sizeof(struct fw_eth_tx_pkt_wr) +
598 		    sizeof(struct cpl_tx_pkt_core) +
599 		    roundup2(m->m_len + m_tls->m_len, 16);
600 		if (wr_len > SGE_MAX_WR_LEN) {
601 			CTR(KTR_CXGBE,
602 		    "%s: %p TLS header-only packet too long (len %d)",
603 			    __func__, tlsp, m->m_len + m_tls->m_len);
604 		}
605 
606 		/* This should always be the last TLS record in a chain. */
607 		MPASS(m_tls->m_next == NULL);
608 		*nsegsp = 0;
609 		return (wr_len);
610 	}
611 
612 	hdr = (void *)m_tls->m_epg_hdr;
613 	rlen = TLS_HEADER_LENGTH + ntohs(hdr->tls_length);
614 
615 	/*
616 	 * See if this request might make use of GHASH state.  This
617 	 * errs on the side of over-budgeting the WR size.
618 	 */
619 	last_ghash_frag = false;
620 	request_ghash = false;
621 	send_partial_ghash = false;
622 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM &&
623 	    tlsp->sc->tlst.partial_ghash && tlsp->sc->tlst.short_records) {
624 		u_int trailer_len;
625 
626 		trailer_len = m_tls->m_epg_trllen;
627 		if (tlsp->tls13)
628 			trailer_len--;
629 		KASSERT(trailer_len == AES_GMAC_HASH_LEN,
630 		    ("invalid trailer length for AES-GCM"));
631 
632 		/* Is this the start of a TLS record? */
633 		if (mtod(m_tls, vm_offset_t) <= m_tls->m_epg_hdrlen) {
634 			/*
635 			 * Might use partial GHASH if this doesn't
636 			 * send the full record.
637 			 */
638 			if (tlen < rlen) {
639 				if (tlen < (rlen - trailer_len))
640 					send_partial_ghash = true;
641 				request_ghash = true;
642 			}
643 		} else {
644 			send_partial_ghash = true;
645 			if (tlen < rlen)
646 				request_ghash = true;
647 			if (tlen >= (rlen - trailer_len))
648 				last_ghash_frag = true;
649 		}
650 	}
651 
652 	/*
653 	 * Assume not sending partial GHASH for this call to get the
654 	 * larger size.
655 	 */
656 	short_record = ktls_is_short_record(tlsp, m_tls, tlen, rlen,
657 	    &header_len, &offset, &plen, &leading_waste, &trailing_waste,
658 	    false, request_ghash);
659 
660 	inline_key = send_partial_ghash || tlsp->inline_key;
661 
662 	/* Calculate the size of the work request. */
663 	wr_len = ktls_base_wr_size(tlsp, inline_key);
664 
665 	if (send_partial_ghash)
666 		wr_len += AES_GMAC_HASH_LEN;
667 
668 	if (leading_waste != 0 || trailing_waste != 0) {
669 		/*
670 		 * Partial records might require a SplitMode
671 		 * CPL_RX_PHYS_DSGL.
672 		 */
673 		wr_len += sizeof(struct cpl_t7_rx_phys_dsgl);
674 	}
675 
676 	/* Budget for an LSO header even if we don't use it. */
677 	wr_len += sizeof(struct cpl_tx_pkt_lso_core);
678 
679 	/*
680 	 * Headers (including the TLS header) are always sent as
681 	 * immediate data.  Short records include a raw AES IV as
682 	 * immediate data.  TLS 1.3 non-short records include a
683 	 * placeholder for the sequence number as immediate data.
684 	 * Short records using a partial hash may also need to send
685 	 * TLS AAD.  If a partial hash might be sent, assume a short
686 	 * record to get the larger size.
687 	 */
688 	imm_len = m->m_len + header_len;
689 	if (short_record || send_partial_ghash) {
690 		imm_len += AES_BLOCK_LEN;
691 		if (send_partial_ghash && header_len != 0)
692 			imm_len += ktls_gcm_aad_len(tlsp);
693 	} else if (tlsp->tls13)
694 		imm_len += sizeof(uint64_t);
695 	wr_len += roundup2(imm_len, 16);
696 
697 	/*
698 	 * TLS record payload via DSGL.  For partial GCM mode we
699 	 * might need an extra SG entry for a placeholder.
700 	 */
701 	*nsegsp = sglist_count_mbuf_epg(m_tls, m_tls->m_epg_hdrlen + offset,
702 	    plen);
703 	wr_len += ktls_sgl_size(*nsegsp + (last_ghash_frag ? 1 : 0));
704 
705 	if (request_ghash) {
706 		/* AES-GCM records might return a partial hash. */
707 		wr_len += sizeof(struct ulp_txpkt);
708 		wr_len += sizeof(struct ulptx_idata);
709 		wr_len += sizeof(struct cpl_tx_tls_ack);
710 		wr_len += sizeof(struct rss_header) +
711 		    sizeof(struct cpl_fw6_pld);
712 		wr_len += AES_GMAC_HASH_LEN;
713 	}
714 
715 	wr_len = roundup2(wr_len, 16);
716 	return (wr_len);
717 }
718 
719 /* Queue the next pending packet. */
720 static void
ktls_queue_next_packet(struct tlspcb * tlsp,bool enqueue_only)721 ktls_queue_next_packet(struct tlspcb *tlsp, bool enqueue_only)
722 {
723 #ifdef KTR
724 	struct ether_header *eh;
725 	struct tcphdr *tcp;
726 	tcp_seq tcp_seqno;
727 #endif
728 	struct mbuf *m;
729 	void *items[1];
730 	int rc;
731 
732 	TXQ_LOCK_ASSERT_OWNED(tlsp->txq);
733 	KASSERT(tlsp->queue_mbufs, ("%s: mbufs not being queued for %p",
734 	    __func__, tlsp));
735 	for (;;) {
736 		m = mbufq_dequeue(&tlsp->pending_mbufs);
737 		if (m == NULL) {
738 			tlsp->queue_mbufs = false;
739 			return;
740 		}
741 
742 #ifdef KTR
743 		eh = mtod(m, struct ether_header *);
744 		tcp = (struct tcphdr *)((char *)eh + m->m_pkthdr.l2hlen +
745 		    m->m_pkthdr.l3hlen);
746 		tcp_seqno = ntohl(tcp->th_seq);
747 #ifdef VERBOSE_TRACES
748 		CTR(KTR_CXGBE, "%s: pkt len %d TCP seq %u", __func__,
749 		    m->m_pkthdr.len, tcp_seqno);
750 #endif
751 #endif
752 
753 		items[0] = m;
754 		if (enqueue_only)
755 			rc = mp_ring_enqueue_only(tlsp->txq->r, items, 1);
756 		else {
757 			TXQ_UNLOCK(tlsp->txq);
758 			rc = mp_ring_enqueue(tlsp->txq->r, items, 1, 256);
759 			TXQ_LOCK(tlsp->txq);
760 		}
761 		if (__predict_true(rc == 0))
762 			return;
763 
764 		CTR(KTR_CXGBE, "%s: pkt len %d TCP seq %u dropped", __func__,
765 		    m->m_pkthdr.len, tcp_seqno);
766 		m_freem(m);
767 	}
768 }
769 
770 int
t7_ktls_parse_pkt(struct mbuf * m)771 t7_ktls_parse_pkt(struct mbuf *m)
772 {
773 	struct tlspcb *tlsp;
774 	struct ether_header *eh;
775 	struct ip *ip;
776 	struct ip6_hdr *ip6;
777 	struct tcphdr *tcp;
778 	struct mbuf *m_tls;
779 	void *items[1];
780 	int error, nsegs;
781 	u_int wr_len, tot_len;
782 	uint16_t eh_type;
783 
784 	/*
785 	 * Locate headers in initial mbuf.
786 	 *
787 	 * XXX: This assumes all of the headers are in the initial mbuf.
788 	 * Could perhaps use m_advance() like parse_pkt() if that turns
789 	 * out to not be true.
790 	 */
791 	M_ASSERTPKTHDR(m);
792 	MPASS(m->m_pkthdr.snd_tag != NULL);
793 	tlsp = mst_to_tls(m->m_pkthdr.snd_tag);
794 
795 	if (m->m_len <= sizeof(*eh) + sizeof(*ip)) {
796 		CTR(KTR_CXGBE, "%s: %p header mbuf too short", __func__, tlsp);
797 		return (EINVAL);
798 	}
799 	eh = mtod(m, struct ether_header *);
800 	eh_type = ntohs(eh->ether_type);
801 	if (eh_type == ETHERTYPE_VLAN) {
802 		struct ether_vlan_header *evh = (void *)eh;
803 
804 		eh_type = ntohs(evh->evl_proto);
805 		m->m_pkthdr.l2hlen = sizeof(*evh);
806 	} else
807 		m->m_pkthdr.l2hlen = sizeof(*eh);
808 
809 	switch (eh_type) {
810 	case ETHERTYPE_IP:
811 		ip = (struct ip *)(eh + 1);
812 		if (ip->ip_p != IPPROTO_TCP) {
813 			CTR(KTR_CXGBE, "%s: %p mbuf not IPPROTO_TCP", __func__,
814 			    tlsp);
815 			return (EINVAL);
816 		}
817 		m->m_pkthdr.l3hlen = ip->ip_hl * 4;
818 		break;
819 	case ETHERTYPE_IPV6:
820 		ip6 = (struct ip6_hdr *)(eh + 1);
821 		if (ip6->ip6_nxt != IPPROTO_TCP) {
822 			CTR(KTR_CXGBE, "%s: %p, mbuf not IPPROTO_TCP (%u)",
823 			    __func__, tlsp, ip6->ip6_nxt);
824 			return (EINVAL);
825 		}
826 		m->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
827 		break;
828 	default:
829 		CTR(KTR_CXGBE, "%s: %p mbuf not ETHERTYPE_IP{,V6}", __func__,
830 		    tlsp);
831 		return (EINVAL);
832 	}
833 	if (m->m_len < m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
834 	    sizeof(*tcp)) {
835 		CTR(KTR_CXGBE, "%s: %p header mbuf too short (2)", __func__,
836 		    tlsp);
837 		return (EINVAL);
838 	}
839 	tcp = (struct tcphdr *)((char *)(eh + 1) + m->m_pkthdr.l3hlen);
840 	m->m_pkthdr.l4hlen = tcp->th_off * 4;
841 
842 	/* Bail if there is TCP payload before the TLS record. */
843 	if (m->m_len != m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
844 	    m->m_pkthdr.l4hlen) {
845 		CTR(KTR_CXGBE,
846 		    "%s: %p header mbuf bad length (%d + %d + %d != %d)",
847 		    __func__, tlsp, m->m_pkthdr.l2hlen, m->m_pkthdr.l3hlen,
848 		    m->m_pkthdr.l4hlen, m->m_len);
849 		return (EINVAL);
850 	}
851 
852 	/* Assume all headers are in 'm' for now. */
853 	MPASS(m->m_next != NULL);
854 	MPASS(m->m_next->m_flags & M_EXTPG);
855 
856 	tot_len = 0;
857 
858 	/*
859 	 * Each of the remaining mbufs in the chain should reference a
860 	 * TLS record.
861 	 */
862 	for (m_tls = m->m_next; m_tls != NULL; m_tls = m_tls->m_next) {
863 		MPASS(m_tls->m_flags & M_EXTPG);
864 
865 		wr_len = ktls_wr_len(tlsp, m, m_tls, &nsegs);
866 #ifdef VERBOSE_TRACES
867 		CTR(KTR_CXGBE, "%s: %p wr_len %d nsegs %d", __func__, tlsp,
868 		    wr_len, nsegs);
869 #endif
870 		if (wr_len > SGE_MAX_WR_LEN || nsegs > TX_SGL_SEGS)
871 			return (EFBIG);
872 		tot_len += roundup2(wr_len, EQ_ESIZE);
873 
874 		/*
875 		 * Store 'nsegs' for the first TLS record in the
876 		 * header mbuf's metadata.
877 		 */
878 		if (m_tls == m->m_next)
879 			set_mbuf_nsegs(m, nsegs);
880 	}
881 
882 	MPASS(tot_len != 0);
883 	set_mbuf_len16(m, tot_len / 16);
884 
885 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
886 		/* Defer packets beyond what has been sent so far. */
887 		TXQ_LOCK(tlsp->txq);
888 		if (tlsp->queue_mbufs) {
889 			error = mbufq_enqueue(&tlsp->pending_mbufs, m);
890 			if (error == 0) {
891 #ifdef VERBOSE_TRACES
892 				CTR(KTR_CXGBE,
893 				    "%s: %p len16 %d nsegs %d TCP seq %u deferred",
894 				    __func__, tlsp, mbuf_len16(m),
895 				    mbuf_nsegs(m), ntohl(tcp->th_seq));
896 #endif
897 			}
898 			TXQ_UNLOCK(tlsp->txq);
899 			return (error);
900 		}
901 		tlsp->queue_mbufs = true;
902 		TXQ_UNLOCK(tlsp->txq);
903 	}
904 
905 #ifdef VERBOSE_TRACES
906 	CTR(KTR_CXGBE, "%s: %p len16 %d nsegs %d", __func__, tlsp,
907 	    mbuf_len16(m), mbuf_nsegs(m));
908 #endif
909 	items[0] = m;
910 	error = mp_ring_enqueue(tlsp->txq->r, items, 1, 256);
911 	if (__predict_false(error != 0)) {
912 		if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
913 			TXQ_LOCK(tlsp->txq);
914 			ktls_queue_next_packet(tlsp, false);
915 			TXQ_UNLOCK(tlsp->txq);
916 		}
917 	}
918 	return (error);
919 }
920 
921 static inline bool
needs_vlan_insertion(struct mbuf * m)922 needs_vlan_insertion(struct mbuf *m)
923 {
924 
925 	M_ASSERTPKTHDR(m);
926 
927 	return (m->m_flags & M_VLANTAG);
928 }
929 
930 static inline uint64_t
pkt_ctrl1(struct sge_txq * txq,struct mbuf * m,uint16_t eh_type)931 pkt_ctrl1(struct sge_txq *txq, struct mbuf *m, uint16_t eh_type)
932 {
933 	uint64_t ctrl1;
934 
935 	/* Checksums are always offloaded */
936 	if (eh_type == ETHERTYPE_IP) {
937 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) |
938 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
939 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
940 	} else {
941 		MPASS(m->m_pkthdr.l3hlen == sizeof(struct ip6_hdr));
942 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) |
943 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
944 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
945 	}
946 	txq->txcsum++;
947 
948 	/* VLAN tag insertion */
949 	if (needs_vlan_insertion(m)) {
950 		ctrl1 |= F_TXPKT_VLAN_VLD |
951 		    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
952 		txq->vlan_insertion++;
953 	}
954 
955 	return (ctrl1);
956 }
957 
958 static inline void *
write_lso_cpl(void * cpl,struct mbuf * m0,uint16_t mss,uint16_t eh_type,int total_len)959 write_lso_cpl(void *cpl, struct mbuf *m0, uint16_t mss, uint16_t eh_type,
960     int total_len)
961 {
962 	struct cpl_tx_pkt_lso_core *lso;
963 	uint32_t ctrl;
964 
965 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
966 	    m0->m_pkthdr.l4hlen > 0,
967 	    ("%s: mbuf %p needs TSO but missing header lengths",
968 		__func__, m0));
969 
970 	ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
971 	    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
972 	    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
973 	    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
974 	    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
975 	if (eh_type == ETHERTYPE_IPV6)
976 		ctrl |= F_LSO_IPV6;
977 
978 	lso = cpl;
979 	lso->lso_ctrl = htobe32(ctrl);
980 	lso->ipid_ofst = htobe16(0);
981 	lso->mss = htobe16(mss);
982 	lso->seqno_offset = htobe32(0);
983 	lso->len = htobe32(total_len);
984 
985 	return (lso + 1);
986 }
987 
988 static inline void *
write_tx_tls_ack(void * dst,u_int rx_chid,u_int hash_len,bool ghash_lcb)989 write_tx_tls_ack(void *dst, u_int rx_chid, u_int hash_len, bool ghash_lcb)
990 {
991 	struct cpl_tx_tls_ack *cpl;
992 	uint32_t flags;
993 
994 	flags = ghash_lcb ? F_CPL_TX_TLS_ACK_LCB : F_CPL_TX_TLS_ACK_PHASH;
995 	cpl = dst;
996 	cpl->op_to_Rsvd2 = htobe32(V_CPL_TX_TLS_ACK_OPCODE(CPL_TX_TLS_ACK) |
997 	    V_T7_CPL_TX_TLS_ACK_RXCHID(rx_chid) | F_CPL_TX_TLS_ACK_ULPTXLPBK |
998 	    flags);
999 
1000 	/* 32 == AckEncCpl, 16 == LCB */
1001 	cpl->PldLen = htobe32(V_CPL_TX_TLS_ACK_PLDLEN(32 + 16 + hash_len));
1002 	cpl->Rsvd3 = 0;
1003 
1004 	return (cpl + 1);
1005 }
1006 
1007 static inline void *
write_fw6_pld(void * dst,u_int rx_chid,u_int rx_qid,u_int hash_len,uint64_t cookie)1008 write_fw6_pld(void *dst, u_int rx_chid, u_int rx_qid, u_int hash_len,
1009     uint64_t cookie)
1010 {
1011 	struct rss_header *rss;
1012 	struct cpl_fw6_pld *cpl;
1013 
1014 	rss = dst;
1015 	memset(rss, 0, sizeof(*rss));
1016 	rss->opcode = CPL_FW6_PLD;
1017 	rss->qid = htobe16(rx_qid);
1018 	rss->channel = rx_chid;
1019 
1020 	cpl = (void *)(rss + 1);
1021 	memset(cpl, 0, sizeof(*cpl));
1022 	cpl->opcode = CPL_FW6_PLD;
1023 	cpl->len = htobe16(hash_len);
1024 	cpl->data[1] = htobe64(cookie);
1025 
1026 	return (cpl + 1);
1027 }
1028 
1029 static inline void *
write_split_mode_rx_phys(void * dst,struct mbuf * m,struct mbuf * m_tls,u_int crypto_hdr_len,u_int leading_waste,u_int trailing_waste)1030 write_split_mode_rx_phys(void *dst, struct mbuf *m, struct mbuf *m_tls,
1031     u_int crypto_hdr_len, u_int leading_waste, u_int trailing_waste)
1032 {
1033 	struct cpl_t7_rx_phys_dsgl *cpl;
1034 	uint16_t *len;
1035 	uint8_t numsge;
1036 
1037 	/* Forward first (3) and third (1) segments. */
1038 	numsge = 0xa;
1039 
1040 	cpl = dst;
1041 	cpl->ot.opcode = CPL_RX_PHYS_DSGL;
1042 	cpl->PhysAddrFields_lo_to_NumSGE =
1043 	    htobe32(F_CPL_T7_RX_PHYS_DSGL_SPLITMODE |
1044 	    V_CPL_T7_RX_PHYS_DSGL_NUMSGE(numsge));
1045 
1046 	len = (uint16_t *)(cpl->RSSCopy);
1047 
1048 	/*
1049 	 * First segment always contains packet headers as well as
1050 	 * transmit-related CPLs.
1051 	 */
1052 	len[0] = htobe16(crypto_hdr_len);
1053 
1054 	/*
1055 	 * Second segment is "gap" of data to drop at the front of the
1056 	 * TLS record.
1057 	 */
1058 	len[1] = htobe16(leading_waste);
1059 
1060 	/* Third segment is how much of the TLS record to send. */
1061 	len[2] = htobe16(m_tls->m_len);
1062 
1063 	/* Fourth segment is how much data to drop at the end. */
1064 	len[3] = htobe16(trailing_waste);
1065 
1066 #ifdef VERBOSE_TRACES
1067 	CTR(KTR_CXGBE, "%s: forward %u skip %u forward %u skip %u",
1068 	    __func__, be16toh(len[0]), be16toh(len[1]), be16toh(len[2]),
1069 	    be16toh(len[3]));
1070 #endif
1071 	return (cpl + 1);
1072 }
1073 
1074 /*
1075  * If the SGL ends on an address that is not 16 byte aligned, this function will
1076  * add a 0 filled flit at the end.
1077  */
1078 static void *
write_gl_to_buf(struct sglist * gl,caddr_t to)1079 write_gl_to_buf(struct sglist *gl, caddr_t to)
1080 {
1081 	struct sglist_seg *seg;
1082 	__be64 *flitp;
1083 	struct ulptx_sgl *usgl;
1084 	int i, nflits, nsegs;
1085 
1086 	KASSERT(((uintptr_t)to & 0xf) == 0,
1087 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, to));
1088 
1089 	nsegs = gl->sg_nseg;
1090 	MPASS(nsegs > 0);
1091 
1092 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
1093 	flitp = (__be64 *)to;
1094 	seg = &gl->sg_segs[0];
1095 	usgl = (void *)flitp;
1096 
1097 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1098 	    V_ULPTX_NSGE(nsegs));
1099 	usgl->len0 = htobe32(seg->ss_len);
1100 	usgl->addr0 = htobe64(seg->ss_paddr);
1101 	seg++;
1102 
1103 	for (i = 0; i < nsegs - 1; i++, seg++) {
1104 		usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
1105 		usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
1106 	}
1107 	if (i & 1)
1108 		usgl->sge[i / 2].len[1] = htobe32(0);
1109 	flitp += nflits;
1110 
1111 	if (nflits & 1) {
1112 		MPASS(((uintptr_t)flitp) & 0xf);
1113 		*flitp++ = 0;
1114 	}
1115 
1116 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
1117 	return (flitp);
1118 }
1119 
1120 static inline void
copy_to_txd(struct sge_eq * eq,const char * from,caddr_t * to,int len)1121 copy_to_txd(struct sge_eq *eq, const char *from, caddr_t *to, int len)
1122 {
1123 
1124 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
1125 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
1126 
1127 	if (__predict_true((uintptr_t)(*to) + len <=
1128 	    (uintptr_t)&eq->desc[eq->sidx])) {
1129 		bcopy(from, *to, len);
1130 		(*to) += len;
1131 		if ((uintptr_t)(*to) == (uintptr_t)&eq->desc[eq->sidx])
1132 			(*to) = (caddr_t)eq->desc;
1133 	} else {
1134 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
1135 
1136 		bcopy(from, *to, portion);
1137 		from += portion;
1138 		portion = len - portion;	/* remaining */
1139 		bcopy(from, (void *)eq->desc, portion);
1140 		(*to) = (caddr_t)eq->desc + portion;
1141 	}
1142 }
1143 
1144 static int
ktls_write_tunnel_packet(struct sge_txq * txq,void * dst,struct mbuf * m,const void * src,u_int len,u_int available,tcp_seq tcp_seqno,u_int pidx,uint16_t eh_type,bool last_wr)1145 ktls_write_tunnel_packet(struct sge_txq *txq, void *dst, struct mbuf *m,
1146     const void *src, u_int len, u_int available, tcp_seq tcp_seqno, u_int pidx,
1147     uint16_t eh_type, bool last_wr)
1148 {
1149 	struct tx_sdesc *txsd;
1150 	struct fw_eth_tx_pkt_wr *wr;
1151 	struct cpl_tx_pkt_core *cpl;
1152 	uint32_t ctrl;
1153 	int len16, ndesc, pktlen;
1154 	struct ether_header *eh;
1155 	struct ip *ip, newip;
1156 	struct ip6_hdr *ip6, newip6;
1157 	struct tcphdr *tcp, newtcp;
1158 	caddr_t out;
1159 
1160 	TXQ_LOCK_ASSERT_OWNED(txq);
1161 	M_ASSERTPKTHDR(m);
1162 
1163 	wr = dst;
1164 	pktlen = m->m_len + len;
1165 	ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen;
1166 	len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16);
1167 	ndesc = tx_len16_to_desc(len16);
1168 	MPASS(ndesc <= available);
1169 
1170 	/* Firmware work request header */
1171 	/* TODO: Handle VF work request. */
1172 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
1173 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
1174 
1175 	ctrl = V_FW_WR_LEN16(len16);
1176 	wr->equiq_to_len16 = htobe32(ctrl);
1177 	wr->r3 = 0;
1178 
1179 	cpl = (void *)(wr + 1);
1180 
1181 	/* CPL header */
1182 	cpl->ctrl0 = txq->cpl_ctrl0;
1183 	cpl->pack = 0;
1184 	cpl->len = htobe16(pktlen);
1185 
1186 	out = (void *)(cpl + 1);
1187 
1188 	/* Copy over Ethernet header. */
1189 	eh = mtod(m, struct ether_header *);
1190 	copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen);
1191 
1192 	/* Fixup length in IP header and copy out. */
1193 	if (eh_type == ETHERTYPE_IP) {
1194 		ip = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1195 		newip = *ip;
1196 		newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen);
1197 		copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip));
1198 		if (m->m_pkthdr.l3hlen > sizeof(*ip))
1199 			copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out,
1200 			    m->m_pkthdr.l3hlen - sizeof(*ip));
1201 	} else {
1202 		ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1203 		newip6 = *ip6;
1204 		newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen -
1205 		    sizeof(*ip6));
1206 		copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6));
1207 		MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6));
1208 	}
1209 	cpl->ctrl1 = htobe64(pkt_ctrl1(txq, m, eh_type));
1210 
1211 	/* Set sequence number in TCP header. */
1212 	tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen);
1213 	newtcp = *tcp;
1214 	newtcp.th_seq = htonl(tcp_seqno);
1215 	copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp));
1216 
1217 	/* Copy rest of TCP header. */
1218 	copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, m->m_len -
1219 	    (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp)));
1220 
1221 	/* Copy the payload data. */
1222 	copy_to_txd(&txq->eq, src, &out, len);
1223 	txq->imm_wrs++;
1224 
1225 	txq->txpkt_wrs++;
1226 
1227 	txsd = &txq->sdesc[pidx];
1228 	if (last_wr)
1229 		txsd->m = m;
1230 	else
1231 		txsd->m = NULL;
1232 	txsd->desc_used = ndesc;
1233 
1234 	return (ndesc);
1235 }
1236 
1237 static int
ktls_write_tls_wr(struct tlspcb * tlsp,struct sge_txq * txq,void * dst,struct mbuf * m,struct tcphdr * tcp,struct mbuf * m_tls,u_int available,tcp_seq tcp_seqno,u_int pidx,uint16_t eh_type,uint16_t mss)1238 ktls_write_tls_wr(struct tlspcb *tlsp, struct sge_txq *txq,
1239     void *dst, struct mbuf *m, struct tcphdr *tcp, struct mbuf *m_tls,
1240     u_int available, tcp_seq tcp_seqno, u_int pidx, uint16_t eh_type,
1241     uint16_t mss)
1242 {
1243 	struct sge_eq *eq = &txq->eq;
1244 	struct tx_sdesc *txsd;
1245 	struct fw_ulptx_wr *wr;
1246 	struct ulp_txpkt *txpkt;
1247 	struct ulptx_sc_memrd *memrd;
1248 	struct ulptx_idata *idata;
1249 	struct cpl_tx_sec_pdu *sec_pdu;
1250 	struct cpl_tx_pkt_core *tx_pkt;
1251 	const struct tls_record_layer *hdr;
1252 	struct ip *ip;
1253 	struct ip6_hdr *ip6;
1254 	struct tcphdr *newtcp;
1255 	char *iv, *out;
1256 	u_int aad_start, aad_stop;
1257 	u_int auth_start, auth_stop, auth_insert;
1258 	u_int cipher_start, cipher_stop, iv_offset;
1259 	u_int header_len, offset, plen, rlen, tlen;
1260 	u_int imm_len, ndesc, nsegs, txpkt_lens[2], wr_len;
1261 	u_int cpl_len, crypto_hdr_len, post_key_context_len;
1262 	u_int leading_waste, trailing_waste;
1263 	u_short ip_len;
1264 	bool inline_key, ghash_lcb, last_ghash_frag, last_wr, need_lso;
1265 	bool request_ghash, send_partial_ghash, short_record, split_mode;
1266 	bool using_scratch;
1267 
1268 	MPASS(tlsp->txq == txq);
1269 	M_ASSERTEXTPG(m_tls);
1270 
1271 	/* Final work request for this mbuf chain? */
1272 	last_wr = (m_tls->m_next == NULL);
1273 
1274 	/*
1275 	 * The relative offset of the last byte to send from the TLS
1276 	 * record.
1277 	 */
1278 	tlen = mtod(m_tls, vm_offset_t) + m_tls->m_len;
1279 	if (tlen <= m_tls->m_epg_hdrlen) {
1280 		/*
1281 		 * For requests that only want to send the TLS header,
1282 		 * send a tunnelled packet as immediate data.
1283 		 */
1284 #ifdef VERBOSE_TRACES
1285 		CTR(KTR_CXGBE, "%s: %p header-only TLS record %u", __func__,
1286 		    tlsp, (u_int)m_tls->m_epg_seqno);
1287 #endif
1288 		/* This should always be the last TLS record in a chain. */
1289 		MPASS(last_wr);
1290 
1291 		txq->kern_tls_header++;
1292 
1293 		return (ktls_write_tunnel_packet(txq, dst, m,
1294 		    (char *)m_tls->m_epg_hdr + mtod(m_tls, vm_offset_t),
1295 		    m_tls->m_len, available, tcp_seqno, pidx, eh_type,
1296 		    last_wr));
1297 	}
1298 
1299 	/* Locate the TLS header. */
1300 	hdr = (void *)m_tls->m_epg_hdr;
1301 	rlen = TLS_HEADER_LENGTH + ntohs(hdr->tls_length);
1302 
1303 #ifdef VERBOSE_TRACES
1304 	CTR(KTR_CXGBE, "%s: offset %lu len %u TCP seq %u TLS record %u",
1305 	    __func__, mtod(m_tls, vm_offset_t), m_tls->m_len, tcp_seqno,
1306 	    (u_int)m_tls->m_epg_seqno);
1307 #endif
1308 
1309 	/* Should this request make use of GHASH state? */
1310 	ghash_lcb = false;
1311 	last_ghash_frag = false;
1312 	request_ghash = false;
1313 	send_partial_ghash = false;
1314 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM &&
1315 	    tlsp->sc->tlst.partial_ghash && tlsp->sc->tlst.short_records) {
1316 		u_int trailer_len;
1317 
1318 		trailer_len = m_tls->m_epg_trllen;
1319 		if (tlsp->tls13)
1320 			trailer_len--;
1321 		KASSERT(trailer_len == AES_GMAC_HASH_LEN,
1322 		    ("invalid trailer length for AES-GCM"));
1323 
1324 		/* Is this the start of a TLS record? */
1325 		if (mtod(m_tls, vm_offset_t) <= m_tls->m_epg_hdrlen) {
1326 			/*
1327 			 * If this is the very first TLS record or
1328 			 * if this is a newer TLS record, request a partial
1329 			 * hash, but not if we are going to send the whole
1330 			 * thing.
1331 			 */
1332 			if ((tlsp->ghash_tls_seqno == 0 ||
1333 			    tlsp->ghash_tls_seqno < m_tls->m_epg_seqno) &&
1334 			    tlen < rlen) {
1335 				/*
1336 				 * If we are only missing part or all
1337 				 * of the trailer, send a normal full
1338 				 * record but request the hash.
1339 				 * Otherwise, use partial GHASH mode.
1340 				 */
1341 				if (tlen >= (rlen - trailer_len))
1342 					ghash_lcb = true;
1343 				else
1344 					send_partial_ghash = true;
1345 				request_ghash = true;
1346 				tlsp->ghash_tls_seqno = m_tls->m_epg_seqno;
1347 			}
1348 		} else if (tlsp->ghash_tls_seqno == m_tls->m_epg_seqno &&
1349 		    tlsp->ghash_valid) {
1350 			/*
1351 			 * Compute the offset of the first AES block as
1352 			 * is done in ktls_is_short_record.
1353 			 */
1354 			if (rlen - tlen < trailer_len)
1355 				plen = rlen - (m_tls->m_epg_hdrlen +
1356 				    trailer_len);
1357 			else
1358 				plen = tlen - m_tls->m_epg_hdrlen;
1359 			offset = mtod(m_tls, vm_offset_t) - m_tls->m_epg_hdrlen;
1360 			if (offset >= plen)
1361 				offset = plen;
1362 			else
1363 				offset = rounddown2(offset, AES_BLOCK_LEN);
1364 			if (tlsp->ghash_offset == offset) {
1365 				if (offset == plen) {
1366 					/*
1367 					 * Send a partial trailer as a
1368 					 * tunnelled packet as
1369 					 * immediate data.
1370 					 */
1371 #ifdef VERBOSE_TRACES
1372 					CTR(KTR_CXGBE,
1373 					    "%s: %p trailer-only TLS record %u",
1374 					    __func__, tlsp,
1375 					    (u_int)m_tls->m_epg_seqno);
1376 #endif
1377 
1378 					txq->kern_tls_trailer++;
1379 
1380 					offset = mtod(m_tls, vm_offset_t) -
1381 					    (m_tls->m_epg_hdrlen + plen);
1382 					KASSERT(offset <= AES_GMAC_HASH_LEN,
1383 					    ("offset outside of trailer"));
1384 					return (ktls_write_tunnel_packet(txq,
1385 					    dst, m, tlsp->ghash + offset,
1386 					    m_tls->m_len, available, tcp_seqno,
1387 					    pidx, eh_type, last_wr));
1388 				}
1389 
1390 				/*
1391 				 * If this request sends the end of
1392 				 * the payload, it is the last
1393 				 * fragment.
1394 				 */
1395 				if (tlen >= (rlen - trailer_len)) {
1396 					last_ghash_frag = true;
1397 					ghash_lcb = true;
1398 				}
1399 
1400 				/*
1401 				 * Only use partial GCM mode (rather
1402 				 * than an AES-CTR short record) if
1403 				 * there is input auth data to pass to
1404 				 * the GHASH.  That is true so long as
1405 				 * there is at least one full block of
1406 				 * payload data, or if the remaining
1407 				 * payload data is the final partial
1408 				 * block.
1409 				 */
1410 				if (plen - offset >= GMAC_BLOCK_LEN ||
1411 				    last_ghash_frag) {
1412 					send_partial_ghash = true;
1413 
1414 					/*
1415 					 * If not sending the complete
1416 					 * end of the record, this is
1417 					 * a middle request so needs
1418 					 * to request an updated
1419 					 * partial hash.
1420 					 */
1421 					if (tlen < rlen)
1422 						request_ghash = true;
1423 				}
1424 			}
1425 		}
1426 	}
1427 
1428 	short_record = ktls_is_short_record(tlsp, m_tls, tlen, rlen,
1429 	    &header_len, &offset, &plen, &leading_waste, &trailing_waste,
1430 	    send_partial_ghash, request_ghash);
1431 
1432 	if (short_record) {
1433 #ifdef VERBOSE_TRACES
1434 		CTR(KTR_CXGBE,
1435 		    "%s: %p short TLS record %u hdr %u offs %u plen %u",
1436 		    __func__, tlsp, (u_int)m_tls->m_epg_seqno, header_len,
1437 		    offset, plen);
1438 		if (send_partial_ghash) {
1439 			if (header_len != 0)
1440 				CTR(KTR_CXGBE, "%s: %p sending initial GHASH",
1441 				    __func__, tlsp);
1442 			else
1443 				CTR(KTR_CXGBE, "%s: %p sending partial GHASH for offset %u%s",
1444 				    __func__, tlsp, tlsp->ghash_offset,
1445 				    last_ghash_frag ? ", last_frag" : "");
1446 		}
1447 #endif
1448 		KASSERT(send_partial_ghash || !request_ghash,
1449 		    ("requesting but not sending partial hash for short record"));
1450 	} else {
1451 		KASSERT(!send_partial_ghash,
1452 		    ("sending partial hash with full record"));
1453 	}
1454 
1455 	if (tlen < rlen && m_tls->m_next == NULL &&
1456 	    (tcp->th_flags & TH_FIN) != 0) {
1457 		txq->kern_tls_fin_short++;
1458 #ifdef INVARIANTS
1459 		panic("%s: FIN on short TLS record", __func__);
1460 #endif
1461 	}
1462 
1463 	/*
1464 	 * Use cached value for first record in chain if not using
1465 	 * partial GCM mode. ktls_parse_pkt() calculates nsegs based
1466 	 * on send_partial_ghash being false.
1467 	 */
1468 	if (m->m_next == m_tls && !send_partial_ghash)
1469 		nsegs = mbuf_nsegs(m);
1470 	else
1471 		nsegs = sglist_count_mbuf_epg(m_tls,
1472 		    m_tls->m_epg_hdrlen + offset, plen);
1473 
1474 	/* Determine if we need an LSO header. */
1475 	need_lso = (m_tls->m_len > mss);
1476 
1477 	/* Calculate the size of the TLS work request. */
1478 	inline_key = send_partial_ghash || tlsp->inline_key;
1479 	wr_len = ktls_base_wr_size(tlsp, inline_key);
1480 
1481 	if (send_partial_ghash) {
1482 		/* Inline key context includes partial hash in OPAD. */
1483 		wr_len += AES_GMAC_HASH_LEN;
1484 	}
1485 
1486 	/*
1487 	 * SplitMode is required if there is any thing we need to trim
1488 	 * from the crypto output, either at the front or end of the
1489 	 * record.  Note that short records might not need trimming.
1490 	 */
1491 	split_mode = leading_waste != 0 || trailing_waste != 0;
1492 	if (split_mode) {
1493 		/*
1494 		 * Partial records require a SplitMode
1495 		 * CPL_RX_PHYS_DSGL.
1496 		 */
1497 		wr_len += sizeof(struct cpl_t7_rx_phys_dsgl);
1498 	}
1499 
1500 	if (need_lso)
1501 		wr_len += sizeof(struct cpl_tx_pkt_lso_core);
1502 
1503 	imm_len = m->m_len + header_len;
1504 	if (short_record) {
1505 		imm_len += AES_BLOCK_LEN;
1506 		if (send_partial_ghash && header_len != 0)
1507 			imm_len += ktls_gcm_aad_len(tlsp);
1508 	} else if (tlsp->tls13)
1509 		imm_len += sizeof(uint64_t);
1510 	wr_len += roundup2(imm_len, 16);
1511 	wr_len += ktls_sgl_size(nsegs + (last_ghash_frag ? 1 : 0));
1512 	wr_len = roundup2(wr_len, 16);
1513 	txpkt_lens[0] = wr_len - sizeof(*wr);
1514 
1515 	if (request_ghash) {
1516 		/*
1517 		 * Requesting the hash entails a second ULP_TX_PKT
1518 		 * containing CPL_TX_TLS_ACK, CPL_FW6_PLD, and space
1519 		 * for the hash.
1520 		 */
1521 		txpkt_lens[1] = sizeof(struct ulp_txpkt);
1522 		txpkt_lens[1] += sizeof(struct ulptx_idata);
1523 		txpkt_lens[1] += sizeof(struct cpl_tx_tls_ack);
1524 		txpkt_lens[1] += sizeof(struct rss_header) +
1525 		    sizeof(struct cpl_fw6_pld);
1526 		txpkt_lens[1] += AES_GMAC_HASH_LEN;
1527 		wr_len += txpkt_lens[1];
1528 	} else
1529 		txpkt_lens[1] = 0;
1530 
1531 	ndesc = howmany(wr_len, EQ_ESIZE);
1532 	MPASS(ndesc <= available);
1533 
1534 	/*
1535 	 * Use the per-txq scratch pad if near the end of the ring to
1536 	 * simplify handling of wrap-around.
1537 	 */
1538 	using_scratch = (eq->sidx - pidx < ndesc);
1539 	if (using_scratch)
1540 		wr = (void *)txq->ss;
1541 	else
1542 		wr = dst;
1543 
1544 	/* FW_ULPTX_WR */
1545 	wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR));
1546 	wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA |
1547 	    V_FW_WR_LEN16(wr_len / 16));
1548 	wr->cookie = 0;
1549 
1550 	/* ULP_TXPKT */
1551 	txpkt = (void *)(wr + 1);
1552 	txpkt->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
1553 	    V_ULP_TXPKT_DATAMODIFY(0) |
1554 	    V_T7_ULP_TXPKT_CHANNELID(tlsp->vi->pi->port_id) |
1555 	    V_ULP_TXPKT_DEST(0) |
1556 	    V_ULP_TXPKT_CMDMORE(request_ghash ? 1 : 0) |
1557 	    V_ULP_TXPKT_FID(txq->eq.iqid) | V_ULP_TXPKT_RO(1));
1558 	txpkt->len = htobe32(howmany(txpkt_lens[0], 16));
1559 
1560 	/* ULPTX_IDATA sub-command */
1561 	idata = (void *)(txpkt + 1);
1562 	idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
1563 	    V_ULP_TX_SC_MORE(1));
1564 	idata->len = sizeof(struct cpl_tx_sec_pdu);
1565 
1566 	/*
1567 	 * After the key context comes CPL_RX_PHYS_DSGL, CPL_TX_*, and
1568 	 * immediate data containing headers.  When using an inline
1569 	 * key, these are counted as part of this ULPTX_IDATA.  When
1570 	 * reading the key from memory, these are part of a separate
1571 	 * ULPTX_IDATA.
1572 	 */
1573 	cpl_len = sizeof(struct cpl_tx_pkt_core);
1574 	if (need_lso)
1575 		cpl_len += sizeof(struct cpl_tx_pkt_lso_core);
1576 	if (split_mode)
1577 		cpl_len += sizeof(struct cpl_t7_rx_phys_dsgl);
1578 	post_key_context_len = cpl_len + imm_len;
1579 
1580 	if (inline_key) {
1581 		idata->len += tlsp->tx_key_info_size + post_key_context_len;
1582 		if (send_partial_ghash) {
1583 			/* Partial GHASH in key context. */
1584 			idata->len += AES_GMAC_HASH_LEN;
1585 		}
1586 	}
1587 	idata->len = htobe32(idata->len);
1588 
1589 	/* CPL_TX_SEC_PDU */
1590 	sec_pdu = (void *)(idata + 1);
1591 
1592 	/*
1593 	 * Packet headers are passed through unchanged by the crypto
1594 	 * engine by marking them as header data in SCMD0.
1595 	 */
1596 	crypto_hdr_len = m->m_len;
1597 
1598 	if (send_partial_ghash) {
1599 		/*
1600 		 * For short records using a partial hash, the TLS
1601 		 * header is counted as header data in SCMD0.  TLS AAD
1602 		 * is next (if AAD is present) followed by the AES-CTR
1603 		 * IV.  Last is the cipher region for the payload.
1604 		 */
1605 		if (header_len != 0) {
1606 			aad_start = 1;
1607 			aad_stop = ktls_gcm_aad_len(tlsp);
1608 		} else {
1609 			aad_start = 0;
1610 			aad_stop = 0;
1611 		}
1612 		iv_offset = aad_stop + 1;
1613 		cipher_start = iv_offset + AES_BLOCK_LEN;
1614 		cipher_stop = 0;
1615 		if (last_ghash_frag) {
1616 			auth_start = cipher_start;
1617 			auth_stop = AES_GMAC_HASH_LEN;
1618 			auth_insert = auth_stop;
1619 		} else if (plen < GMAC_BLOCK_LEN) {
1620 			/*
1621 			 * A request that sends part of the first AES
1622 			 * block will only have AAD.
1623 			 */
1624 			KASSERT(header_len != 0,
1625 			    ("%s: partial GHASH with no auth", __func__));
1626 			auth_start = 0;
1627 			auth_stop = 0;
1628 			auth_insert = 0;
1629 		} else {
1630 			auth_start = cipher_start;
1631 			auth_stop = plen % GMAC_BLOCK_LEN;
1632 			auth_insert = 0;
1633 		}
1634 
1635 		sec_pdu->pldlen = htobe32(aad_stop + AES_BLOCK_LEN + plen +
1636 		    (last_ghash_frag ? AES_GMAC_HASH_LEN : 0));
1637 
1638 		/*
1639 		 * For short records, the TLS header is treated as
1640 		 * header data.
1641 		 */
1642 		crypto_hdr_len += header_len;
1643 
1644 		/* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */
1645 		sec_pdu->seqno_numivs = tlsp->scmd0_partial.seqno_numivs;
1646 		sec_pdu->ivgen_hdrlen = tlsp->scmd0_partial.ivgen_hdrlen;
1647 		if (last_ghash_frag)
1648 			sec_pdu->ivgen_hdrlen |= V_SCMD_LAST_FRAG(1);
1649 		else
1650 			sec_pdu->ivgen_hdrlen |= V_SCMD_MORE_FRAGS(1);
1651 		sec_pdu->ivgen_hdrlen = htobe32(sec_pdu->ivgen_hdrlen |
1652 		    V_SCMD_HDR_LEN(crypto_hdr_len));
1653 
1654 		txq->kern_tls_partial_ghash++;
1655 	} else if (short_record) {
1656 		/*
1657 		 * For short records without a partial hash, the TLS
1658 		 * header is counted as header data in SCMD0 and the
1659 		 * IV is next, followed by a cipher region for the
1660 		 * payload.
1661 		 */
1662 		aad_start = 0;
1663 		aad_stop = 0;
1664 		iv_offset = 1;
1665 		auth_start = 0;
1666 		auth_stop = 0;
1667 		auth_insert = 0;
1668 		cipher_start = AES_BLOCK_LEN + 1;
1669 		cipher_stop = 0;
1670 
1671 		sec_pdu->pldlen = htobe32(AES_BLOCK_LEN + plen);
1672 
1673 		/*
1674 		 * For short records, the TLS header is treated as
1675 		 * header data.
1676 		 */
1677 		crypto_hdr_len += header_len;
1678 
1679 		/* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */
1680 		sec_pdu->seqno_numivs = tlsp->scmd0_short.seqno_numivs;
1681 		sec_pdu->ivgen_hdrlen = htobe32(
1682 		    tlsp->scmd0_short.ivgen_hdrlen |
1683 		    V_SCMD_HDR_LEN(crypto_hdr_len));
1684 
1685 		txq->kern_tls_short++;
1686 	} else {
1687 		/*
1688 		 * AAD is TLS header.  IV is after AAD for TLS < 1.3.
1689 		 * For TLS 1.3, a placeholder for the TLS sequence
1690 		 * number is provided as an IV before the AAD.  The
1691 		 * cipher region starts after the AAD and IV.  See
1692 		 * comments in ccr_authenc() and ccr_gmac() in
1693 		 * t4_crypto.c regarding cipher and auth start/stop
1694 		 * values.
1695 		 */
1696 		if (tlsp->tls13) {
1697 			iv_offset = 1;
1698 			aad_start = 1 + sizeof(uint64_t);
1699 			aad_stop = sizeof(uint64_t) + TLS_HEADER_LENGTH;
1700 			cipher_start = aad_stop + 1;
1701 		} else {
1702 			aad_start = 1;
1703 			aad_stop = TLS_HEADER_LENGTH;
1704 			iv_offset = TLS_HEADER_LENGTH + 1;
1705 			cipher_start = m_tls->m_epg_hdrlen + 1;
1706 		}
1707 		if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
1708 			cipher_stop = 0;
1709 			auth_start = cipher_start;
1710 			auth_stop = 0;
1711 			auth_insert = 0;
1712 		} else {
1713 			cipher_stop = 0;
1714 			auth_start = cipher_start;
1715 			auth_stop = 0;
1716 			auth_insert = 0;
1717 		}
1718 
1719 		sec_pdu->pldlen = htobe32((tlsp->tls13 ? sizeof(uint64_t) : 0) +
1720 		    m_tls->m_epg_hdrlen + plen);
1721 
1722 		/* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */
1723 		sec_pdu->seqno_numivs = tlsp->scmd0.seqno_numivs;
1724 		sec_pdu->ivgen_hdrlen = htobe32(tlsp->scmd0.ivgen_hdrlen |
1725 		    V_SCMD_HDR_LEN(crypto_hdr_len));
1726 
1727 		if (split_mode)
1728 			txq->kern_tls_partial++;
1729 		else
1730 			txq->kern_tls_full++;
1731 	}
1732 	sec_pdu->op_ivinsrtofst = htobe32(
1733 	    V_CPL_TX_SEC_PDU_OPCODE(CPL_TX_SEC_PDU) |
1734 	    V_CPL_TX_SEC_PDU_CPLLEN(cpl_len / 8) |
1735 	    V_CPL_TX_SEC_PDU_PLACEHOLDER(send_partial_ghash ? 1 : 0) |
1736 	    V_CPL_TX_SEC_PDU_IVINSRTOFST(iv_offset));
1737 	sec_pdu->aadstart_cipherstop_hi = htobe32(
1738 	    V_CPL_TX_SEC_PDU_AADSTART(aad_start) |
1739 	    V_CPL_TX_SEC_PDU_AADSTOP(aad_stop) |
1740 	    V_CPL_TX_SEC_PDU_CIPHERSTART(cipher_start) |
1741 	    V_CPL_TX_SEC_PDU_CIPHERSTOP_HI(cipher_stop >> 4));
1742 	sec_pdu->cipherstop_lo_authinsert = htobe32(
1743 	    V_CPL_TX_SEC_PDU_CIPHERSTOP_LO(cipher_stop & 0xf) |
1744 	    V_CPL_TX_SEC_PDU_AUTHSTART(auth_start) |
1745 	    V_CPL_TX_SEC_PDU_AUTHSTOP(auth_stop) |
1746 	    V_CPL_TX_SEC_PDU_AUTHINSERT(auth_insert));
1747 
1748 	if (send_partial_ghash && last_ghash_frag) {
1749 		uint64_t aad_len, cipher_len;
1750 
1751 		aad_len = ktls_gcm_aad_len(tlsp);
1752 		cipher_len = rlen - (m_tls->m_epg_hdrlen + AES_GMAC_HASH_LEN);
1753 		sec_pdu->scmd1 = htobe64(aad_len << 44 | cipher_len);
1754 	} else
1755 		sec_pdu->scmd1 = htobe64(m_tls->m_epg_seqno);
1756 
1757 	/* Key context */
1758 	out = (void *)(sec_pdu + 1);
1759 	if (inline_key) {
1760 		memcpy(out, &tlsp->keyctx, tlsp->tx_key_info_size);
1761 		if (send_partial_ghash) {
1762 			struct tls_keyctx *keyctx = (void *)out;
1763 
1764 			keyctx->u.txhdr.ctxlen++;
1765 			keyctx->u.txhdr.dualck_to_txvalid &= ~htobe16(
1766 			    V_KEY_CONTEXT_MK_SIZE(M_KEY_CONTEXT_MK_SIZE));
1767 			keyctx->u.txhdr.dualck_to_txvalid |= htobe16(
1768 			    F_KEY_CONTEXT_OPAD_PRESENT |
1769 			    V_KEY_CONTEXT_MK_SIZE(0));
1770 		}
1771 		out += tlsp->tx_key_info_size;
1772 		if (send_partial_ghash) {
1773 			if (header_len != 0)
1774 				memset(out, 0, AES_GMAC_HASH_LEN);
1775 			else
1776 				memcpy(out, tlsp->ghash, AES_GMAC_HASH_LEN);
1777 			out += AES_GMAC_HASH_LEN;
1778 		}
1779 	} else {
1780 		/* ULPTX_SC_MEMRD to read key context. */
1781 		memrd = (void *)out;
1782 		memrd->cmd_to_len = htobe32(V_ULPTX_CMD(ULP_TX_SC_MEMRD) |
1783 		    V_ULP_TX_SC_MORE(1) |
1784 		    V_ULPTX_LEN16(tlsp->tx_key_info_size >> 4));
1785 		memrd->addr = htobe32(tlsp->tx_key_addr >> 5);
1786 
1787 		/* ULPTX_IDATA for CPL_TX_* and headers. */
1788 		idata = (void *)(memrd + 1);
1789 		idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
1790 		    V_ULP_TX_SC_MORE(1));
1791 		idata->len = htobe32(post_key_context_len);
1792 
1793 		out = (void *)(idata + 1);
1794 	}
1795 
1796 	/* CPL_RX_PHYS_DSGL */
1797 	if (split_mode) {
1798 		crypto_hdr_len = sizeof(struct cpl_tx_pkt_core);
1799 		if (need_lso)
1800 			crypto_hdr_len += sizeof(struct cpl_tx_pkt_lso_core);
1801 		crypto_hdr_len += m->m_len;
1802 		out = write_split_mode_rx_phys(out, m, m_tls, crypto_hdr_len,
1803 		    leading_waste, trailing_waste);
1804 	}
1805 
1806 	/* CPL_TX_PKT_LSO */
1807 	if (need_lso) {
1808 		out = write_lso_cpl(out, m, mss, eh_type, m->m_len +
1809 		    m_tls->m_len);
1810 		txq->tso_wrs++;
1811 	}
1812 
1813 	/* CPL_TX_PKT_XT */
1814 	tx_pkt = (void *)out;
1815 	tx_pkt->ctrl0 = txq->cpl_ctrl0;
1816 	tx_pkt->ctrl1 = htobe64(pkt_ctrl1(txq, m, eh_type));
1817 	tx_pkt->pack = 0;
1818 	tx_pkt->len = htobe16(m->m_len + m_tls->m_len);
1819 
1820 	/* Copy the packet headers. */
1821 	out = (void *)(tx_pkt + 1);
1822 	memcpy(out, mtod(m, char *), m->m_len);
1823 
1824 	/* Modify the packet length in the IP header. */
1825 	ip_len = m->m_len + m_tls->m_len - m->m_pkthdr.l2hlen;
1826 	if (eh_type == ETHERTYPE_IP) {
1827 		ip = (void *)(out + m->m_pkthdr.l2hlen);
1828 		be16enc(&ip->ip_len, ip_len);
1829 	} else {
1830 		ip6 = (void *)(out + m->m_pkthdr.l2hlen);
1831 		be16enc(&ip6->ip6_plen, ip_len - sizeof(*ip6));
1832 	}
1833 
1834 	/* Modify sequence number and flags in TCP header. */
1835 	newtcp = (void *)(out + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen);
1836 	be32enc(&newtcp->th_seq, tcp_seqno);
1837 	if (!last_wr)
1838 		newtcp->th_flags = tcp->th_flags & ~(TH_PUSH | TH_FIN);
1839 	out += m->m_len;
1840 
1841 	/*
1842 	 * Insert placeholder for sequence number as IV for TLS 1.3
1843 	 * non-short records.
1844 	 */
1845 	if (tlsp->tls13 && !short_record) {
1846 		memset(out, 0, sizeof(uint64_t));
1847 		out += sizeof(uint64_t);
1848 	}
1849 
1850 	/* Populate the TLS header */
1851 	memcpy(out, m_tls->m_epg_hdr, header_len);
1852 	out += header_len;
1853 
1854 	/* TLS AAD for short records using a partial hash. */
1855 	if (send_partial_ghash && header_len != 0) {
1856 		if (tlsp->tls13) {
1857 			struct tls_aead_data_13 ad;
1858 
1859 			ad.type = hdr->tls_type;
1860 			ad.tls_vmajor = hdr->tls_vmajor;
1861 			ad.tls_vminor = hdr->tls_vminor;
1862 			ad.tls_length = hdr->tls_length;
1863 			memcpy(out, &ad, sizeof(ad));
1864 			out += sizeof(ad);
1865 		} else {
1866 			struct tls_aead_data ad;
1867 			uint16_t cipher_len;
1868 
1869 			cipher_len = rlen -
1870 			    (m_tls->m_epg_hdrlen + AES_GMAC_HASH_LEN);
1871 			ad.seq = htobe64(m_tls->m_epg_seqno);
1872 			ad.type = hdr->tls_type;
1873 			ad.tls_vmajor = hdr->tls_vmajor;
1874 			ad.tls_vminor = hdr->tls_vminor;
1875 			ad.tls_length = htons(cipher_len);
1876 			memcpy(out, &ad, sizeof(ad));
1877 			out += sizeof(ad);
1878 		}
1879 	}
1880 
1881 	/* AES IV for a short record. */
1882 	if (short_record) {
1883 		iv = out;
1884 		if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
1885 			memcpy(iv, tlsp->keyctx.u.txhdr.txsalt, SALT_SIZE);
1886 			if (tlsp->tls13) {
1887 				uint64_t value;
1888 
1889 				value = be64dec(tlsp->keyctx.u.txhdr.txsalt +
1890 				    4);
1891 				value ^= m_tls->m_epg_seqno;
1892 				be64enc(iv + 4, value);
1893 			} else
1894 				memcpy(iv + 4, hdr + 1, 8);
1895 			if (send_partial_ghash)
1896 				be32enc(iv + 12, 1 + offset / AES_BLOCK_LEN);
1897 			else
1898 				be32enc(iv + 12, 2 + offset / AES_BLOCK_LEN);
1899 		} else
1900 			memcpy(iv, hdr + 1, AES_BLOCK_LEN);
1901 		out += AES_BLOCK_LEN;
1902 	}
1903 
1904 	if (imm_len % 16 != 0) {
1905 		if (imm_len % 8 != 0) {
1906 			/* Zero pad to an 8-byte boundary. */
1907 			memset(out, 0, 8 - (imm_len % 8));
1908 			out += 8 - (imm_len % 8);
1909 		}
1910 
1911 		/*
1912 		 * Insert a ULP_TX_SC_NOOP if needed so the SGL is
1913 		 * 16-byte aligned.
1914 		 */
1915 		if (imm_len % 16 <= 8) {
1916 			idata = (void *)out;
1917 			idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP) |
1918 			    V_ULP_TX_SC_MORE(1));
1919 			idata->len = htobe32(0);
1920 			out = (void *)(idata + 1);
1921 		}
1922 	}
1923 
1924 	/* SGL for record payload */
1925 	sglist_reset(txq->gl);
1926 	if (sglist_append_mbuf_epg(txq->gl, m_tls, m_tls->m_epg_hdrlen + offset,
1927 	    plen) != 0) {
1928 #ifdef INVARIANTS
1929 		panic("%s: failed to append sglist", __func__);
1930 #endif
1931 	}
1932 	if (last_ghash_frag) {
1933 		if (sglist_append_phys(txq->gl, zero_buffer_pa,
1934 		    AES_GMAC_HASH_LEN) != 0) {
1935 #ifdef INVARIANTS
1936 			panic("%s: failed to append sglist (2)", __func__);
1937 #endif
1938 		}
1939 	}
1940 	out = write_gl_to_buf(txq->gl, out);
1941 
1942 	if (request_ghash) {
1943 		/* ULP_TXPKT */
1944 		txpkt = (void *)out;
1945 		txpkt->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
1946 		    V_ULP_TXPKT_DATAMODIFY(0) |
1947 		    V_T7_ULP_TXPKT_CHANNELID(tlsp->vi->pi->port_id) |
1948 		    V_ULP_TXPKT_DEST(0) |
1949 		    V_ULP_TXPKT_FID(txq->eq.iqid) | V_ULP_TXPKT_RO(1));
1950 		txpkt->len = htobe32(howmany(txpkt_lens[1], 16));
1951 
1952 		/* ULPTX_IDATA sub-command */
1953 		idata = (void *)(txpkt + 1);
1954 		idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
1955 		    V_ULP_TX_SC_MORE(0));
1956 		idata->len = sizeof(struct cpl_tx_tls_ack);
1957 		idata->len += sizeof(struct rss_header) +
1958 		    sizeof(struct cpl_fw6_pld);
1959 		idata->len += AES_GMAC_HASH_LEN;
1960 		idata->len = htobe32(idata->len);
1961 		out = (void *)(idata + 1);
1962 
1963 		/* CPL_TX_TLS_ACK */
1964 		out = write_tx_tls_ack(out, tlsp->rx_chid, AES_GMAC_HASH_LEN,
1965 		    ghash_lcb);
1966 
1967 		/* CPL_FW6_PLD */
1968 		out = write_fw6_pld(out, tlsp->rx_chid, tlsp->rx_qid,
1969 		    AES_GMAC_HASH_LEN, (uintptr_t)tlsp | CPL_FW6_COOKIE_KTLS);
1970 
1971 		/* Space for partial hash. */
1972 		memset(out, 0, AES_GMAC_HASH_LEN);
1973 		out += AES_GMAC_HASH_LEN;
1974 
1975 		tlsp->ghash_pending = true;
1976 		tlsp->ghash_valid = false;
1977 		tlsp->ghash_lcb = ghash_lcb;
1978 		if (ghash_lcb)
1979 			tlsp->ghash_offset = offset + plen;
1980 		else
1981 			tlsp->ghash_offset = rounddown2(offset + plen,
1982 			    GMAC_BLOCK_LEN);
1983 #ifdef VERBOSE_TRACES
1984 		CTR(KTR_CXGBE, "%s: %p requesting GHASH for offset %u",
1985 		    __func__, tlsp, tlsp->ghash_offset);
1986 #endif
1987 		m_snd_tag_ref(&tlsp->com);
1988 
1989 		txq->kern_tls_ghash_requested++;
1990 	}
1991 
1992 	if (using_scratch) {
1993 		out = dst;
1994 		copy_to_txd(eq, txq->ss, &out, wr_len);
1995 	}
1996 
1997 	txq->kern_tls_records++;
1998 	txq->kern_tls_octets += m_tls->m_len;
1999 	if (split_mode) {
2000 		txq->kern_tls_splitmode++;
2001 		txq->kern_tls_waste += leading_waste + trailing_waste;
2002 	}
2003 	if (need_lso)
2004 		txq->kern_tls_lso++;
2005 
2006 	txsd = &txq->sdesc[pidx];
2007 	if (last_wr)
2008 		txsd->m = m;
2009 	else
2010 		txsd->m = NULL;
2011 	txsd->desc_used = ndesc;
2012 
2013 	return (ndesc);
2014 }
2015 
2016 int
t7_ktls_write_wr(struct sge_txq * txq,void * dst,struct mbuf * m,u_int available)2017 t7_ktls_write_wr(struct sge_txq *txq, void *dst, struct mbuf *m,
2018     u_int available)
2019 {
2020 	struct sge_eq *eq = &txq->eq;
2021 	struct tlspcb *tlsp;
2022 	struct tcphdr *tcp;
2023 	struct mbuf *m_tls;
2024 	struct ether_header *eh;
2025 	tcp_seq tcp_seqno;
2026 	u_int ndesc, pidx, totdesc;
2027 	uint16_t eh_type, mss;
2028 
2029 	TXQ_LOCK_ASSERT_OWNED(txq);
2030 	M_ASSERTPKTHDR(m);
2031 	MPASS(m->m_pkthdr.snd_tag != NULL);
2032 	tlsp = mst_to_tls(m->m_pkthdr.snd_tag);
2033 
2034 	totdesc = 0;
2035 	eh = mtod(m, struct ether_header *);
2036 	eh_type = ntohs(eh->ether_type);
2037 	if (eh_type == ETHERTYPE_VLAN) {
2038 		struct ether_vlan_header *evh = (void *)eh;
2039 
2040 		eh_type = ntohs(evh->evl_proto);
2041 	}
2042 
2043 	tcp = (struct tcphdr *)((char *)eh + m->m_pkthdr.l2hlen +
2044 	    m->m_pkthdr.l3hlen);
2045 	pidx = eq->pidx;
2046 
2047 	/* Determine MSS. */
2048 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2049 		mss = m->m_pkthdr.tso_segsz;
2050 		tlsp->prev_mss = mss;
2051 	} else if (tlsp->prev_mss != 0)
2052 		mss = tlsp->prev_mss;
2053 	else
2054 		mss = if_getmtu(tlsp->vi->ifp) -
2055 		    (m->m_pkthdr.l3hlen + m->m_pkthdr.l4hlen);
2056 
2057 	/* Fetch the starting TCP sequence number for this chain. */
2058 	tcp_seqno = ntohl(tcp->th_seq);
2059 #ifdef VERBOSE_TRACES
2060 	CTR(KTR_CXGBE, "%s: pkt len %d TCP seq %u", __func__, m->m_pkthdr.len,
2061 	    tcp_seqno);
2062 #endif
2063 	KASSERT(!tlsp->ghash_pending, ("%s: GHASH pending for send", __func__));
2064 
2065 	/*
2066 	 * Iterate over each TLS record constructing a work request
2067 	 * for that record.
2068 	 */
2069 	for (m_tls = m->m_next; m_tls != NULL; m_tls = m_tls->m_next) {
2070 		MPASS(m_tls->m_flags & M_EXTPG);
2071 
2072 		ndesc = ktls_write_tls_wr(tlsp, txq, dst, m, tcp, m_tls,
2073 		    available - totdesc, tcp_seqno, pidx, eh_type, mss);
2074 		totdesc += ndesc;
2075 		IDXINCR(pidx, ndesc, eq->sidx);
2076 		dst = &eq->desc[pidx];
2077 
2078 		tcp_seqno += m_tls->m_len;
2079 	}
2080 
2081 	/*
2082 	 * Queue another packet if this was a GCM request that didn't
2083 	 * request a GHASH response.
2084 	 */
2085 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM && !tlsp->ghash_pending)
2086 		ktls_queue_next_packet(tlsp, true);
2087 
2088 	MPASS(totdesc <= available);
2089 	return (totdesc);
2090 }
2091 
2092 static void
t7_tls_tag_free(struct m_snd_tag * mst)2093 t7_tls_tag_free(struct m_snd_tag *mst)
2094 {
2095 	struct adapter *sc;
2096 	struct tlspcb *tlsp;
2097 
2098 	tlsp = mst_to_tls(mst);
2099 	sc = tlsp->sc;
2100 
2101 	CTR2(KTR_CXGBE, "%s: %p", __func__, tlsp);
2102 
2103 	if (tlsp->tx_key_addr >= 0)
2104 		t4_free_tls_keyid(sc, tlsp->tx_key_addr);
2105 
2106 	KASSERT(mbufq_len(&tlsp->pending_mbufs) == 0,
2107 	    ("%s: pending mbufs", __func__));
2108 
2109 	zfree(tlsp, M_CXGBE);
2110 }
2111 
2112 static int
ktls_fw6_pld(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)2113 ktls_fw6_pld(struct sge_iq *iq, const struct rss_header *rss,
2114     struct mbuf *m)
2115 {
2116 	const struct cpl_fw6_pld *cpl;
2117 	struct tlspcb *tlsp;
2118 	const void *ghash;
2119 
2120 	if (m != NULL)
2121 		cpl = mtod(m, const void *);
2122 	else
2123 		cpl = (const void *)(rss + 1);
2124 
2125 	tlsp = (struct tlspcb *)(uintptr_t)CPL_FW6_PLD_COOKIE(cpl);
2126 	KASSERT(cpl->data[0] == 0, ("%s: error status returned", __func__));
2127 
2128 	TXQ_LOCK(tlsp->txq);
2129 #ifdef VERBOSE_TRACES
2130 	CTR(KTR_CXGBE, "%s: %p received GHASH for offset %u%s", __func__, tlsp,
2131 	    tlsp->ghash_offset, tlsp->ghash_lcb ? " in LCB" : "");
2132 #endif
2133 	if (tlsp->ghash_lcb)
2134 		ghash = &cpl->data[2];
2135 	else
2136 		ghash = cpl + 1;
2137 	memcpy(tlsp->ghash, ghash, AES_GMAC_HASH_LEN);
2138 	tlsp->ghash_valid = true;
2139 	tlsp->ghash_pending = false;
2140 	tlsp->txq->kern_tls_ghash_received++;
2141 
2142 	ktls_queue_next_packet(tlsp, false);
2143 	TXQ_UNLOCK(tlsp->txq);
2144 
2145 	m_snd_tag_rele(&tlsp->com);
2146 	m_freem(m);
2147 	return (0);
2148 }
2149 
2150 void
t7_ktls_modload(void)2151 t7_ktls_modload(void)
2152 {
2153 	zero_buffer = malloc_aligned(AES_GMAC_HASH_LEN, AES_GMAC_HASH_LEN,
2154 	    M_CXGBE, M_ZERO | M_WAITOK);
2155 	zero_buffer_pa = vtophys(zero_buffer);
2156 	t4_register_shared_cpl_handler(CPL_FW6_PLD, ktls_fw6_pld,
2157 	    CPL_FW6_COOKIE_KTLS);
2158 }
2159 
2160 void
t7_ktls_modunload(void)2161 t7_ktls_modunload(void)
2162 {
2163 	free(zero_buffer, M_CXGBE);
2164 	t4_register_shared_cpl_handler(CPL_FW6_PLD, NULL, CPL_FW6_COOKIE_KTLS);
2165 }
2166 
2167 #else
2168 
2169 int
t7_tls_tag_alloc(struct ifnet * ifp,union if_snd_tag_alloc_params * params,struct m_snd_tag ** pt)2170 t7_tls_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params,
2171     struct m_snd_tag **pt)
2172 {
2173 	return (ENXIO);
2174 }
2175 
2176 int
t7_ktls_parse_pkt(struct mbuf * m)2177 t7_ktls_parse_pkt(struct mbuf *m)
2178 {
2179 	return (EINVAL);
2180 }
2181 
2182 int
t7_ktls_write_wr(struct sge_txq * txq,void * dst,struct mbuf * m,u_int available)2183 t7_ktls_write_wr(struct sge_txq *txq, void *dst, struct mbuf *m,
2184     u_int available)
2185 {
2186 	panic("can't happen");
2187 }
2188 
2189 void
t7_ktls_modload(void)2190 t7_ktls_modload(void)
2191 {
2192 }
2193 
2194 void
t7_ktls_modunload(void)2195 t7_ktls_modunload(void)
2196 {
2197 }
2198 
2199 #endif
2200