xref: /freebsd/sys/dev/cxgbe/crypto/t6_kern_tls.c (revision f76effed)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2018-2019 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_inet6.h"
32 #include "opt_kern_tls.h"
33 
34 #include <sys/param.h>
35 #include <sys/ktr.h>
36 #include <sys/ktls.h>
37 #include <sys/sglist.h>
38 #include <sys/socket.h>
39 #include <sys/socketvar.h>
40 #include <sys/sockbuf.h>
41 #include <netinet/in.h>
42 #include <netinet/in_pcb.h>
43 #include <netinet/ip.h>
44 #include <netinet/ip6.h>
45 #include <netinet/tcp_var.h>
46 #include <opencrypto/cryptodev.h>
47 #include <opencrypto/xform.h>
48 
49 #include "common/common.h"
50 #include "common/t4_regs.h"
51 #include "common/t4_regs_values.h"
52 #include "common/t4_tcb.h"
53 #include "t4_l2t.h"
54 #include "t4_clip.h"
55 #include "t4_mp_ring.h"
56 #include "crypto/t4_crypto.h"
57 
58 #if defined(INET) || defined(INET6)
59 
60 #define TLS_HEADER_LENGTH		5
61 
62 struct tls_scmd {
63 	__be32 seqno_numivs;
64 	__be32 ivgen_hdrlen;
65 };
66 
67 struct tlspcb {
68 	struct m_snd_tag com;
69 	struct vi_info *vi;	/* virtual interface */
70 	struct adapter *sc;
71 	struct l2t_entry *l2te;	/* L2 table entry used by this connection */
72 	struct sge_txq *txq;
73 	int tid;		/* Connection identifier */
74 
75 	int tx_key_addr;
76 	bool inline_key;
77 	bool using_timestamps;
78 	unsigned char enc_mode;
79 
80 	struct tls_scmd scmd0;
81 	struct tls_scmd scmd0_short;
82 
83 	unsigned int tx_key_info_size;
84 
85 	uint32_t prev_seq;
86 	uint32_t prev_ack;
87 	uint32_t prev_tsecr;
88 	uint16_t prev_win;
89 	uint16_t prev_mss;
90 
91 	/* Only used outside of setup and teardown when using inline keys. */
92 	struct tls_keyctx keyctx;
93 
94 	/* Fields only used during setup and teardown. */
95 	struct inpcb *inp;	/* backpointer to host stack's PCB */
96 	struct sge_wrq *ctrlq;
97 	struct clip_entry *ce;	/* CLIP table entry used by this tid */
98 
99 	bool open_pending;
100 };
101 
102 static void t6_tls_tag_free(struct m_snd_tag *mst);
103 static int ktls_setup_keys(struct tlspcb *tlsp,
104     const struct ktls_session *tls, struct sge_txq *txq);
105 
106 static const struct if_snd_tag_sw t6_tls_tag_sw = {
107 	.snd_tag_free = t6_tls_tag_free,
108 	.type = IF_SND_TAG_TYPE_TLS
109 };
110 
111 static inline struct tlspcb *
mst_to_tls(struct m_snd_tag * t)112 mst_to_tls(struct m_snd_tag *t)
113 {
114 	return (__containerof(t, struct tlspcb, com));
115 }
116 
117 static struct tlspcb *
alloc_tlspcb(if_t ifp,struct vi_info * vi,int flags)118 alloc_tlspcb(if_t ifp, struct vi_info *vi, int flags)
119 {
120 	struct port_info *pi = vi->pi;
121 	struct adapter *sc = pi->adapter;
122 	struct tlspcb *tlsp;
123 
124 	tlsp = malloc(sizeof(*tlsp), M_CXGBE, M_ZERO | flags);
125 	if (tlsp == NULL)
126 		return (NULL);
127 
128 	m_snd_tag_init(&tlsp->com, ifp, &t6_tls_tag_sw);
129 	tlsp->vi = vi;
130 	tlsp->sc = sc;
131 	tlsp->ctrlq = &sc->sge.ctrlq[pi->port_id];
132 	tlsp->tid = -1;
133 	tlsp->tx_key_addr = -1;
134 
135 	return (tlsp);
136 }
137 
138 static int
ktls_act_open_cpl_size(bool isipv6)139 ktls_act_open_cpl_size(bool isipv6)
140 {
141 
142 	if (isipv6)
143 		return (sizeof(struct cpl_t6_act_open_req6));
144 	else
145 		return (sizeof(struct cpl_t6_act_open_req));
146 }
147 
148 static void
mk_ktls_act_open_req(struct adapter * sc,struct vi_info * vi,struct inpcb * inp,struct tlspcb * tlsp,int atid,void * dst)149 mk_ktls_act_open_req(struct adapter *sc, struct vi_info *vi, struct inpcb *inp,
150     struct tlspcb *tlsp, int atid, void *dst)
151 {
152 	struct tcpcb *tp = intotcpcb(inp);
153 	struct cpl_t6_act_open_req *cpl6;
154 	struct cpl_act_open_req *cpl;
155 	uint64_t options;
156 	int qid_atid;
157 
158 	cpl6 = dst;
159 	cpl = (struct cpl_act_open_req *)cpl6;
160 	INIT_TP_WR(cpl6, 0);
161 	qid_atid = V_TID_QID(sc->sge.fwq.abs_id) | V_TID_TID(atid) |
162 	    V_TID_COOKIE(CPL_COOKIE_KERN_TLS);
163 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ,
164 		qid_atid));
165 	inp_4tuple_get(inp, &cpl->local_ip, &cpl->local_port,
166 	    &cpl->peer_ip, &cpl->peer_port);
167 
168 	options = F_TCAM_BYPASS | V_ULP_MODE(ULP_MODE_NONE);
169 	options |= V_SMAC_SEL(vi->smt_idx) | V_TX_CHAN(vi->pi->tx_chan);
170 	options |= F_NON_OFFLOAD;
171 	cpl->opt0 = htobe64(options);
172 
173 	options = V_TX_QUEUE(TX_MODQ(vi->pi->tx_chan));
174 	if (tp->t_flags & TF_REQ_TSTMP)
175 		options |= F_TSTAMPS_EN;
176 	cpl->opt2 = htobe32(options);
177 }
178 
179 static void
mk_ktls_act_open_req6(struct adapter * sc,struct vi_info * vi,struct inpcb * inp,struct tlspcb * tlsp,int atid,void * dst)180 mk_ktls_act_open_req6(struct adapter *sc, struct vi_info *vi,
181     struct inpcb *inp, struct tlspcb *tlsp, int atid, void *dst)
182 {
183 	struct tcpcb *tp = intotcpcb(inp);
184 	struct cpl_t6_act_open_req6 *cpl6;
185 	struct cpl_act_open_req6 *cpl;
186 	uint64_t options;
187 	int qid_atid;
188 
189 	cpl6 = dst;
190 	cpl = (struct cpl_act_open_req6 *)cpl6;
191 	INIT_TP_WR(cpl6, 0);
192 	qid_atid = V_TID_QID(sc->sge.fwq.abs_id) | V_TID_TID(atid) |
193 	    V_TID_COOKIE(CPL_COOKIE_KERN_TLS);
194 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ6,
195 		qid_atid));
196 	cpl->local_port = inp->inp_lport;
197 	cpl->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0];
198 	cpl->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8];
199 	cpl->peer_port = inp->inp_fport;
200 	cpl->peer_ip_hi = *(uint64_t *)&inp->in6p_faddr.s6_addr[0];
201 	cpl->peer_ip_lo = *(uint64_t *)&inp->in6p_faddr.s6_addr[8];
202 
203 	options = F_TCAM_BYPASS | V_ULP_MODE(ULP_MODE_NONE);
204 	options |= V_SMAC_SEL(vi->smt_idx) | V_TX_CHAN(vi->pi->tx_chan);
205 	options |= F_NON_OFFLOAD;
206 	cpl->opt0 = htobe64(options);
207 
208 	options = V_TX_QUEUE(TX_MODQ(vi->pi->tx_chan));
209 	if (tp->t_flags & TF_REQ_TSTMP)
210 		options |= F_TSTAMPS_EN;
211 	cpl->opt2 = htobe32(options);
212 }
213 
214 static int
send_ktls_act_open_req(struct adapter * sc,struct vi_info * vi,struct inpcb * inp,struct tlspcb * tlsp,int atid)215 send_ktls_act_open_req(struct adapter *sc, struct vi_info *vi,
216     struct inpcb *inp, struct tlspcb *tlsp, int atid)
217 {
218 	struct wrqe *wr;
219 	bool isipv6;
220 
221 	isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
222 	if (isipv6) {
223 		tlsp->ce = t4_get_clip_entry(sc, &inp->in6p_laddr, true);
224 		if (tlsp->ce == NULL)
225 			return (ENOENT);
226 	}
227 
228 	wr = alloc_wrqe(ktls_act_open_cpl_size(isipv6), tlsp->ctrlq);
229 	if (wr == NULL) {
230 		CTR2(KTR_CXGBE, "%s: atid %d failed to alloc WR", __func__,
231 		    atid);
232 		return (ENOMEM);
233 	}
234 
235 	if (isipv6)
236 		mk_ktls_act_open_req6(sc, vi, inp, tlsp, atid, wrtod(wr));
237 	else
238 		mk_ktls_act_open_req(sc, vi, inp, tlsp, atid, wrtod(wr));
239 
240 	tlsp->open_pending = true;
241 	t4_wrq_tx(sc, wr);
242 	return (0);
243 }
244 
245 static int
ktls_act_open_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)246 ktls_act_open_rpl(struct sge_iq *iq, const struct rss_header *rss,
247     struct mbuf *m)
248 {
249 	struct adapter *sc = iq->adapter;
250 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
251 	u_int atid = G_TID_TID(G_AOPEN_ATID(be32toh(cpl->atid_status)));
252 	u_int status = G_AOPEN_STATUS(be32toh(cpl->atid_status));
253 	struct tlspcb *tlsp = lookup_atid(sc, atid);
254 	struct inpcb *inp = tlsp->inp;
255 
256 	CTR3(KTR_CXGBE, "%s: atid %d status %d", __func__, atid, status);
257 	free_atid(sc, atid);
258 	if (status == 0)
259 		tlsp->tid = GET_TID(cpl);
260 
261 	INP_WLOCK(inp);
262 	tlsp->open_pending = false;
263 	wakeup(tlsp);
264 	INP_WUNLOCK(inp);
265 	return (0);
266 }
267 
268 /* SET_TCB_FIELD sent as a ULP command looks like this */
269 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
270     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
271 
272 _Static_assert((LEN__SET_TCB_FIELD_ULP + sizeof(struct ulptx_idata)) % 16 == 0,
273     "CPL_SET_TCB_FIELD ULP command not 16-byte aligned");
274 
275 static void
write_set_tcb_field_ulp(struct tlspcb * tlsp,void * dst,struct sge_txq * txq,uint16_t word,uint64_t mask,uint64_t val)276 write_set_tcb_field_ulp(struct tlspcb *tlsp, void *dst, struct sge_txq *txq,
277     uint16_t word, uint64_t mask, uint64_t val)
278 {
279 	struct ulp_txpkt *txpkt;
280 	struct ulptx_idata *idata;
281 	struct cpl_set_tcb_field_core *cpl;
282 
283 	/* ULP_TXPKT */
284 	txpkt = dst;
285 	txpkt->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
286 	    V_ULP_TXPKT_DATAMODIFY(0) |
287 	    V_ULP_TXPKT_CHANNELID(tlsp->vi->pi->port_id) | V_ULP_TXPKT_DEST(0) |
288 	    V_ULP_TXPKT_FID(txq->eq.cntxt_id) | V_ULP_TXPKT_RO(1));
289 	txpkt->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
290 
291 	/* ULPTX_IDATA sub-command */
292 	idata = (struct ulptx_idata *)(txpkt + 1);
293 	idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
294 	idata->len = htobe32(sizeof(*cpl));
295 
296 	/* CPL_SET_TCB_FIELD */
297 	cpl = (struct cpl_set_tcb_field_core *)(idata + 1);
298 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tlsp->tid));
299 	cpl->reply_ctrl = htobe16(F_NO_REPLY);
300 	cpl->word_cookie = htobe16(V_WORD(word));
301 	cpl->mask = htobe64(mask);
302 	cpl->val = htobe64(val);
303 
304 	/* ULPTX_NOOP */
305 	idata = (struct ulptx_idata *)(cpl + 1);
306 	idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
307 	idata->len = htobe32(0);
308 }
309 
310 static int
ktls_set_tcb_fields(struct tlspcb * tlsp,struct tcpcb * tp,struct sge_txq * txq)311 ktls_set_tcb_fields(struct tlspcb *tlsp, struct tcpcb *tp, struct sge_txq *txq)
312 {
313 	struct fw_ulptx_wr *wr;
314 	struct mbuf *m;
315 	char *dst;
316 	void *items[1];
317 	int error, len;
318 
319 	len = sizeof(*wr) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
320 	if (tp->t_flags & TF_REQ_TSTMP)
321 		len += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
322 	m = alloc_wr_mbuf(len, M_NOWAIT);
323 	if (m == NULL) {
324 		CTR2(KTR_CXGBE, "%s: tid %d failed to alloc WR mbuf", __func__,
325 		    tlsp->tid);
326 		return (ENOMEM);
327 	}
328 	m->m_pkthdr.snd_tag = m_snd_tag_ref(&tlsp->com);
329 	m->m_pkthdr.csum_flags |= CSUM_SND_TAG;
330 
331 	/* FW_ULPTX_WR */
332 	wr = mtod(m, void *);
333 	wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR));
334 	wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA |
335 	    V_FW_WR_LEN16(len / 16));
336 	wr->cookie = 0;
337 	dst = (char *)(wr + 1);
338 
339         /* Clear TF_NON_OFFLOAD and set TF_CORE_BYPASS */
340 	write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_T_FLAGS,
341 	    V_TCB_T_FLAGS(V_TF_CORE_BYPASS(1) | V_TF_NON_OFFLOAD(1)),
342 	    V_TCB_T_FLAGS(V_TF_CORE_BYPASS(1)));
343 	dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
344 
345 	/* Clear the SND_UNA_RAW, SND_NXT_RAW, and SND_MAX_RAW offsets. */
346 	write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_SND_UNA_RAW,
347 	    V_TCB_SND_NXT_RAW(M_TCB_SND_NXT_RAW) |
348 	    V_TCB_SND_UNA_RAW(M_TCB_SND_UNA_RAW),
349 	    V_TCB_SND_NXT_RAW(0) | V_TCB_SND_UNA_RAW(0));
350 	dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
351 
352 	write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_SND_MAX_RAW,
353 	    V_TCB_SND_MAX_RAW(M_TCB_SND_MAX_RAW), V_TCB_SND_MAX_RAW(0));
354 	dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
355 
356 	if (tp->t_flags & TF_REQ_TSTMP) {
357 		write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_TIMESTAMP_OFFSET,
358 		    V_TCB_TIMESTAMP_OFFSET(M_TCB_TIMESTAMP_OFFSET),
359 		    V_TCB_TIMESTAMP_OFFSET(tp->ts_offset >> 28));
360 		dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
361 	}
362 
363 	KASSERT(dst - (char *)wr == len, ("%s: length mismatch", __func__));
364 
365 	items[0] = m;
366 	error = mp_ring_enqueue(txq->r, items, 1, 1);
367 	if (error)
368 		m_free(m);
369 	return (error);
370 }
371 
372 int
t6_tls_tag_alloc(if_t ifp,union if_snd_tag_alloc_params * params,struct m_snd_tag ** pt)373 t6_tls_tag_alloc(if_t ifp, union if_snd_tag_alloc_params *params,
374     struct m_snd_tag **pt)
375 {
376 	const struct ktls_session *tls;
377 	struct tlspcb *tlsp;
378 	struct adapter *sc;
379 	struct vi_info *vi;
380 	struct inpcb *inp;
381 	struct tcpcb *tp;
382 	struct sge_txq *txq;
383 	int atid, error, explicit_iv_size, keyid, mac_first;
384 
385 	tls = params->tls.tls;
386 
387 	/* Only TLS 1.1 and TLS 1.2 are currently supported. */
388 	if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE ||
389 	    tls->params.tls_vminor < TLS_MINOR_VER_ONE ||
390 	    tls->params.tls_vminor > TLS_MINOR_VER_TWO)
391 		return (EPROTONOSUPPORT);
392 
393 	/* Sanity check values in *tls. */
394 	switch (tls->params.cipher_algorithm) {
395 	case CRYPTO_AES_CBC:
396 		/* XXX: Explicitly ignore any provided IV. */
397 		switch (tls->params.cipher_key_len) {
398 		case 128 / 8:
399 		case 192 / 8:
400 		case 256 / 8:
401 			break;
402 		default:
403 			return (EINVAL);
404 		}
405 		switch (tls->params.auth_algorithm) {
406 		case CRYPTO_SHA1_HMAC:
407 		case CRYPTO_SHA2_256_HMAC:
408 		case CRYPTO_SHA2_384_HMAC:
409 			break;
410 		default:
411 			return (EPROTONOSUPPORT);
412 		}
413 		explicit_iv_size = AES_BLOCK_LEN;
414 		mac_first = 1;
415 		break;
416 	case CRYPTO_AES_NIST_GCM_16:
417 		if (tls->params.iv_len != SALT_SIZE)
418 			return (EINVAL);
419 		switch (tls->params.cipher_key_len) {
420 		case 128 / 8:
421 		case 192 / 8:
422 		case 256 / 8:
423 			break;
424 		default:
425 			return (EINVAL);
426 		}
427 		explicit_iv_size = 8;
428 		mac_first = 0;
429 		break;
430 	default:
431 		return (EPROTONOSUPPORT);
432 	}
433 
434 	vi = if_getsoftc(ifp);
435 	sc = vi->adapter;
436 
437 	tlsp = alloc_tlspcb(ifp, vi, M_WAITOK);
438 
439 	atid = alloc_atid(sc, tlsp);
440 	if (atid < 0) {
441 		error = ENOMEM;
442 		goto failed;
443 	}
444 
445 	if (sc->tlst.inline_keys)
446 		keyid = -1;
447 	else
448 		keyid = t4_alloc_tls_keyid(sc);
449 	if (keyid < 0) {
450 		CTR2(KTR_CXGBE, "%s: atid %d using immediate key ctx", __func__,
451 		    atid);
452 		tlsp->inline_key = true;
453 	} else {
454 		tlsp->tx_key_addr = keyid;
455 		CTR3(KTR_CXGBE, "%s: atid %d allocated TX key addr %#x",
456 		    __func__,
457 		    atid, tlsp->tx_key_addr);
458 	}
459 
460 	inp = params->tls.inp;
461 	INP_RLOCK(inp);
462 	if (inp->inp_flags & INP_DROPPED) {
463 		INP_RUNLOCK(inp);
464 		error = ECONNRESET;
465 		goto failed;
466 	}
467 	tlsp->inp = inp;
468 
469 	tp = intotcpcb(inp);
470 	if (tp->t_flags & TF_REQ_TSTMP) {
471 		tlsp->using_timestamps = true;
472 		if ((tp->ts_offset & 0xfffffff) != 0) {
473 			INP_RUNLOCK(inp);
474 			error = EINVAL;
475 			goto failed;
476 		}
477 	} else
478 		tlsp->using_timestamps = false;
479 
480 	error = send_ktls_act_open_req(sc, vi, inp, tlsp, atid);
481 	if (error) {
482 		INP_RUNLOCK(inp);
483 		goto failed;
484 	}
485 
486 	/* Wait for reply to active open. */
487 	CTR2(KTR_CXGBE, "%s: atid %d sent CPL_ACT_OPEN_REQ", __func__,
488 	    atid);
489 	while (tlsp->open_pending) {
490 		/*
491 		 * XXX: PCATCH?  We would then have to discard the PCB
492 		 * when the completion CPL arrived.
493 		 */
494 		error = rw_sleep(tlsp, &inp->inp_lock, 0, "t6tlsop", 0);
495 	}
496 
497 	atid = -1;
498 	if (tlsp->tid < 0) {
499 		INP_RUNLOCK(inp);
500 		error = ENOMEM;
501 		goto failed;
502 	}
503 
504 	if (inp->inp_flags & INP_DROPPED) {
505 		INP_RUNLOCK(inp);
506 		error = ECONNRESET;
507 		goto failed;
508 	}
509 
510 	txq = &sc->sge.txq[vi->first_txq];
511 	if (inp->inp_flowtype != M_HASHTYPE_NONE)
512 		txq += ((inp->inp_flowid % (vi->ntxq - vi->rsrv_noflowq)) +
513 		    vi->rsrv_noflowq);
514 	tlsp->txq = txq;
515 
516 	error = ktls_set_tcb_fields(tlsp, tp, txq);
517 	INP_RUNLOCK(inp);
518 	if (error)
519 		goto failed;
520 
521 	error = ktls_setup_keys(tlsp, tls, txq);
522 	if (error)
523 		goto failed;
524 
525 	tlsp->enc_mode = t4_tls_cipher_mode(tls);
526 	tlsp->tx_key_info_size = t4_tls_key_info_size(tls);
527 
528 	/* The SCMD fields used when encrypting a full TLS record. */
529 	tlsp->scmd0.seqno_numivs = htobe32(V_SCMD_SEQ_NO_CTRL(3) |
530 	    V_SCMD_PROTO_VERSION(t4_tls_proto_ver(tls)) |
531 	    V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
532 	    V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) |
533 	    V_SCMD_CIPH_MODE(tlsp->enc_mode) |
534 	    V_SCMD_AUTH_MODE(t4_tls_auth_mode(tls)) |
535 	    V_SCMD_HMAC_CTRL(t4_tls_hmac_ctrl(tls)) |
536 	    V_SCMD_IV_SIZE(explicit_iv_size / 2) | V_SCMD_NUM_IVS(1));
537 
538 	tlsp->scmd0.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) |
539 	    V_SCMD_TLS_FRAG_ENABLE(0);
540 	if (tlsp->inline_key)
541 		tlsp->scmd0.ivgen_hdrlen |= V_SCMD_KEY_CTX_INLINE(1);
542 	tlsp->scmd0.ivgen_hdrlen = htobe32(tlsp->scmd0.ivgen_hdrlen);
543 
544 	/*
545 	 * The SCMD fields used when encrypting a partial TLS record
546 	 * (no trailer and possibly a truncated payload).
547 	 */
548 	tlsp->scmd0_short.seqno_numivs = V_SCMD_SEQ_NO_CTRL(0) |
549 	    V_SCMD_PROTO_VERSION(SCMD_PROTO_VERSION_GENERIC) |
550 	    V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
551 	    V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) |
552 	    V_SCMD_AUTH_MODE(SCMD_AUTH_MODE_NOP) |
553 	    V_SCMD_HMAC_CTRL(SCMD_HMAC_CTRL_NOP) |
554 	    V_SCMD_IV_SIZE(AES_BLOCK_LEN / 2) | V_SCMD_NUM_IVS(0);
555 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM)
556 		tlsp->scmd0_short.seqno_numivs |=
557 		    V_SCMD_CIPH_MODE(SCMD_CIPH_MODE_AES_CTR);
558 	else
559 		tlsp->scmd0_short.seqno_numivs |=
560 		    V_SCMD_CIPH_MODE(tlsp->enc_mode);
561 	tlsp->scmd0_short.seqno_numivs =
562 	    htobe32(tlsp->scmd0_short.seqno_numivs);
563 
564 	tlsp->scmd0_short.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) |
565 	    V_SCMD_TLS_FRAG_ENABLE(0) |
566 	    V_SCMD_AADIVDROP(1);
567 	if (tlsp->inline_key)
568 		tlsp->scmd0_short.ivgen_hdrlen |= V_SCMD_KEY_CTX_INLINE(1);
569 
570 	TXQ_LOCK(txq);
571 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM)
572 		txq->kern_tls_gcm++;
573 	else
574 		txq->kern_tls_cbc++;
575 	TXQ_UNLOCK(txq);
576 	*pt = &tlsp->com;
577 	return (0);
578 
579 failed:
580 	if (atid >= 0)
581 		free_atid(sc, atid);
582 	m_snd_tag_rele(&tlsp->com);
583 	return (error);
584 }
585 
586 static int
ktls_setup_keys(struct tlspcb * tlsp,const struct ktls_session * tls,struct sge_txq * txq)587 ktls_setup_keys(struct tlspcb *tlsp, const struct ktls_session *tls,
588     struct sge_txq *txq)
589 {
590 	struct tls_key_req *kwr;
591 	struct tls_keyctx *kctx;
592 	void *items[1];
593 	struct mbuf *m;
594 	int error;
595 
596 	/*
597 	 * Store the salt and keys in the key context.  For
598 	 * connections with an inline key, this key context is passed
599 	 * as immediate data in each work request.  For connections
600 	 * storing the key in DDR, a work request is used to store a
601 	 * copy of the key context in DDR.
602 	 */
603 	t4_tls_key_ctx(tls, KTLS_TX, &tlsp->keyctx);
604 	if (tlsp->inline_key)
605 		return (0);
606 
607 	/* Populate key work request. */
608         m = alloc_wr_mbuf(TLS_KEY_WR_SZ, M_NOWAIT);
609 	if (m == NULL) {
610 		CTR2(KTR_CXGBE, "%s: tid %d failed to alloc WR mbuf", __func__,
611 		    tlsp->tid);
612 		return (ENOMEM);
613 	}
614 	m->m_pkthdr.snd_tag = m_snd_tag_ref(&tlsp->com);
615 	m->m_pkthdr.csum_flags |= CSUM_SND_TAG;
616 	kwr = mtod(m, void *);
617 	memset(kwr, 0, TLS_KEY_WR_SZ);
618 
619 	t4_write_tlskey_wr(tls, KTLS_TX, tlsp->tid, 0, tlsp->tx_key_addr, kwr);
620 	kctx = (struct tls_keyctx *)(kwr + 1);
621 	memcpy(kctx, &tlsp->keyctx, sizeof(*kctx));
622 
623 	/*
624 	 * Place the key work request in the transmit queue.  It
625 	 * should be sent to the NIC before any TLS packets using this
626 	 * session.
627 	 */
628 	items[0] = m;
629 	error = mp_ring_enqueue(txq->r, items, 1, 1);
630 	if (error)
631 		m_free(m);
632 	else
633 		CTR2(KTR_CXGBE, "%s: tid %d sent key WR", __func__, tlsp->tid);
634 	return (error);
635 }
636 
637 static u_int
ktls_base_wr_size(struct tlspcb * tlsp)638 ktls_base_wr_size(struct tlspcb *tlsp)
639 {
640 	u_int wr_len;
641 
642 	wr_len = sizeof(struct fw_ulptx_wr);	// 16
643 	wr_len += sizeof(struct ulp_txpkt);	// 8
644 	wr_len += sizeof(struct ulptx_idata);	// 8
645 	wr_len += sizeof(struct cpl_tx_sec_pdu);// 32
646 	if (tlsp->inline_key)
647 		wr_len += tlsp->tx_key_info_size;
648 	else {
649 		wr_len += sizeof(struct ulptx_sc_memrd);// 8
650 		wr_len += sizeof(struct ulptx_idata);	// 8
651 	}
652 	wr_len += sizeof(struct cpl_tx_data);	// 16
653 	return (wr_len);
654 }
655 
656 /* How many bytes of TCP payload to send for a given TLS record. */
657 static u_int
ktls_tcp_payload_length(struct tlspcb * tlsp,struct mbuf * m_tls)658 ktls_tcp_payload_length(struct tlspcb *tlsp, struct mbuf *m_tls)
659 {
660 	struct tls_record_layer *hdr;
661 	u_int plen, mlen;
662 
663 	M_ASSERTEXTPG(m_tls);
664 	hdr = (void *)m_tls->m_epg_hdr;
665 	plen = ntohs(hdr->tls_length);
666 
667 	/*
668 	 * What range of the TLS record is the mbuf requesting to be
669 	 * sent.
670 	 */
671 	mlen = mtod(m_tls, vm_offset_t) + m_tls->m_len;
672 
673 	/* Always send complete records. */
674 	if (mlen == TLS_HEADER_LENGTH + plen)
675 		return (mlen);
676 
677 	/*
678 	 * If the host stack has asked to send part of the trailer,
679 	 * trim the length to avoid sending any of the trailer.  There
680 	 * is no way to send a partial trailer currently.
681 	 */
682 	if (mlen > TLS_HEADER_LENGTH + plen - m_tls->m_epg_trllen)
683 		mlen = TLS_HEADER_LENGTH + plen - m_tls->m_epg_trllen;
684 
685 
686 	/*
687 	 * For AES-CBC adjust the ciphertext length for the block
688 	 * size.
689 	 */
690 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_CBC &&
691 	    mlen > TLS_HEADER_LENGTH) {
692 		mlen = TLS_HEADER_LENGTH + rounddown(mlen - TLS_HEADER_LENGTH,
693 		    AES_BLOCK_LEN);
694 	}
695 
696 #ifdef VERBOSE_TRACES
697 	CTR4(KTR_CXGBE, "%s: tid %d short TLS record (%u vs %u)",
698 	    __func__, tlsp->tid, mlen, TLS_HEADER_LENGTH + plen);
699 #endif
700 	return (mlen);
701 }
702 
703 /*
704  * For a "short" TLS record, determine the offset into the TLS record
705  * payload to send.  This offset does not include the TLS header, but
706  * a non-zero offset implies that a header will not be sent.
707  */
708 static u_int
ktls_payload_offset(struct tlspcb * tlsp,struct mbuf * m_tls)709 ktls_payload_offset(struct tlspcb *tlsp, struct mbuf *m_tls)
710 {
711 	struct tls_record_layer *hdr;
712 	u_int offset, plen;
713 #ifdef INVARIANTS
714 	u_int mlen;
715 #endif
716 
717 	M_ASSERTEXTPG(m_tls);
718 	hdr = (void *)m_tls->m_epg_hdr;
719 	plen = ntohs(hdr->tls_length);
720 #ifdef INVARIANTS
721 	mlen = mtod(m_tls, vm_offset_t) + m_tls->m_len;
722 	MPASS(mlen < TLS_HEADER_LENGTH + plen);
723 #endif
724 	if (mtod(m_tls, vm_offset_t) <= m_tls->m_epg_hdrlen)
725 		return (0);
726 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
727 		/*
728 		 * Always send something.  This function is only called
729 		 * if we aren't sending the tag at all, but if the
730 		 * request starts in the tag then we are in an odd
731 		 * state where would effectively send nothing.  Cap
732 		 * the offset at the last byte of the record payload
733 		 * to send the last cipher block.
734 		 */
735 		offset = min(mtod(m_tls, vm_offset_t) - m_tls->m_epg_hdrlen,
736 		    (plen - TLS_HEADER_LENGTH - m_tls->m_epg_trllen) - 1);
737 		return (rounddown(offset, AES_BLOCK_LEN));
738 	}
739 	return (0);
740 }
741 
742 static u_int
ktls_sgl_size(u_int nsegs)743 ktls_sgl_size(u_int nsegs)
744 {
745 	u_int wr_len;
746 
747 	/* First segment is part of ulptx_sgl. */
748 	nsegs--;
749 
750 	wr_len = sizeof(struct ulptx_sgl);
751 	wr_len += 8 * ((3 * nsegs) / 2 + (nsegs & 1));
752 	return (wr_len);
753 }
754 
755 static int
ktls_wr_len(struct tlspcb * tlsp,struct mbuf * m,struct mbuf * m_tls,int * nsegsp)756 ktls_wr_len(struct tlspcb *tlsp, struct mbuf *m, struct mbuf *m_tls,
757     int *nsegsp)
758 {
759 	struct tls_record_layer *hdr;
760 	u_int imm_len, offset, plen, wr_len, tlen;
761 
762 	M_ASSERTEXTPG(m_tls);
763 
764 	/*
765 	 * Determine the size of the TLS record payload to send
766 	 * excluding header and trailer.
767 	 */
768 	tlen = ktls_tcp_payload_length(tlsp, m_tls);
769 	if (tlen <= m_tls->m_epg_hdrlen) {
770 		/*
771 		 * For requests that only want to send the TLS header,
772 		 * send a tunnelled packet as immediate data.
773 		 */
774 		wr_len = sizeof(struct fw_eth_tx_pkt_wr) +
775 		    sizeof(struct cpl_tx_pkt_core) +
776 		    roundup2(m->m_len + m_tls->m_len, 16);
777 		if (wr_len > SGE_MAX_WR_LEN) {
778 			CTR3(KTR_CXGBE,
779 		    "%s: tid %d TLS header-only packet too long (len %d)",
780 			    __func__, tlsp->tid, m->m_len + m_tls->m_len);
781 		}
782 
783 		/* This should always be the last TLS record in a chain. */
784 		MPASS(m_tls->m_next == NULL);
785 		*nsegsp = 0;
786 		return (wr_len);
787 	}
788 
789 	hdr = (void *)m_tls->m_epg_hdr;
790 	plen = TLS_HEADER_LENGTH + ntohs(hdr->tls_length) - m_tls->m_epg_trllen;
791 	if (tlen < plen) {
792 		plen = tlen;
793 		offset = ktls_payload_offset(tlsp, m_tls);
794 	} else
795 		offset = 0;
796 
797 	/* Calculate the size of the work request. */
798 	wr_len = ktls_base_wr_size(tlsp);
799 
800 	/*
801 	 * Full records and short records with an offset of 0 include
802 	 * the TLS header as immediate data.  Short records include a
803 	 * raw AES IV as immediate data.
804 	 */
805 	imm_len = 0;
806 	if (offset == 0)
807 		imm_len += m_tls->m_epg_hdrlen;
808 	if (plen == tlen)
809 		imm_len += AES_BLOCK_LEN;
810 	wr_len += roundup2(imm_len, 16);
811 
812 	/* TLS record payload via DSGL. */
813 	*nsegsp = sglist_count_mbuf_epg(m_tls, m_tls->m_epg_hdrlen + offset,
814 	    plen - (m_tls->m_epg_hdrlen + offset));
815 	wr_len += ktls_sgl_size(*nsegsp);
816 
817 	wr_len = roundup2(wr_len, 16);
818 	return (wr_len);
819 }
820 
821 /*
822  * See if we have any TCP options requiring a dedicated options-only
823  * packet.
824  */
825 static int
ktls_has_tcp_options(struct tcphdr * tcp)826 ktls_has_tcp_options(struct tcphdr *tcp)
827 {
828 	u_char *cp;
829 	int cnt, opt, optlen;
830 
831 	cp = (u_char *)(tcp + 1);
832 	cnt = tcp->th_off * 4 - sizeof(struct tcphdr);
833 	for (; cnt > 0; cnt -= optlen, cp += optlen) {
834 		opt = cp[0];
835 		if (opt == TCPOPT_EOL)
836 			break;
837 		if (opt == TCPOPT_NOP)
838 			optlen = 1;
839 		else {
840 			if (cnt < 2)
841 				break;
842 			optlen = cp[1];
843 			if (optlen < 2 || optlen > cnt)
844 				break;
845 		}
846 		switch (opt) {
847 		case TCPOPT_NOP:
848 		case TCPOPT_TIMESTAMP:
849 			break;
850 		default:
851 			return (1);
852 		}
853 	}
854 	return (0);
855 }
856 
857 /*
858  * Find the TCP timestamp option.
859  */
860 static void *
ktls_find_tcp_timestamps(struct tcphdr * tcp)861 ktls_find_tcp_timestamps(struct tcphdr *tcp)
862 {
863 	u_char *cp;
864 	int cnt, opt, optlen;
865 
866 	cp = (u_char *)(tcp + 1);
867 	cnt = tcp->th_off * 4 - sizeof(struct tcphdr);
868 	for (; cnt > 0; cnt -= optlen, cp += optlen) {
869 		opt = cp[0];
870 		if (opt == TCPOPT_EOL)
871 			break;
872 		if (opt == TCPOPT_NOP)
873 			optlen = 1;
874 		else {
875 			if (cnt < 2)
876 				break;
877 			optlen = cp[1];
878 			if (optlen < 2 || optlen > cnt)
879 				break;
880 		}
881 		if (opt == TCPOPT_TIMESTAMP && optlen == TCPOLEN_TIMESTAMP)
882 			return (cp + 2);
883 	}
884 	return (NULL);
885 }
886 
887 int
t6_ktls_parse_pkt(struct mbuf * m)888 t6_ktls_parse_pkt(struct mbuf *m)
889 {
890 	struct tlspcb *tlsp;
891 	struct ether_header *eh;
892 	struct ip *ip;
893 	struct ip6_hdr *ip6;
894 	struct tcphdr *tcp;
895 	struct mbuf *m_tls;
896 	void *items[1];
897 	int nsegs;
898 	u_int wr_len, tot_len;
899 
900 	/*
901 	 * Locate headers in initial mbuf.
902 	 *
903 	 * XXX: This assumes all of the headers are in the initial mbuf.
904 	 * Could perhaps use m_advance() like parse_pkt() if that turns
905 	 * out to not be true.
906 	 */
907 	M_ASSERTPKTHDR(m);
908 	MPASS(m->m_pkthdr.snd_tag != NULL);
909 	tlsp = mst_to_tls(m->m_pkthdr.snd_tag);
910 
911 	if (m->m_len <= sizeof(*eh) + sizeof(*ip)) {
912 		CTR2(KTR_CXGBE, "%s: tid %d header mbuf too short", __func__,
913 		    tlsp->tid);
914 		return (EINVAL);
915 	}
916 	eh = mtod(m, struct ether_header *);
917 	if (ntohs(eh->ether_type) != ETHERTYPE_IP &&
918 	    ntohs(eh->ether_type) != ETHERTYPE_IPV6) {
919 		CTR2(KTR_CXGBE, "%s: tid %d mbuf not ETHERTYPE_IP{,V6}",
920 		    __func__, tlsp->tid);
921 		return (EINVAL);
922 	}
923 	m->m_pkthdr.l2hlen = sizeof(*eh);
924 
925 	/* XXX: Reject unsupported IP options? */
926 	if (ntohs(eh->ether_type) == ETHERTYPE_IP) {
927 		ip = (struct ip *)(eh + 1);
928 		if (ip->ip_p != IPPROTO_TCP) {
929 			CTR2(KTR_CXGBE, "%s: tid %d mbuf not IPPROTO_TCP",
930 			    __func__, tlsp->tid);
931 			return (EINVAL);
932 		}
933 		m->m_pkthdr.l3hlen = ip->ip_hl * 4;
934 	} else {
935 		ip6 = (struct ip6_hdr *)(eh + 1);
936 		if (ip6->ip6_nxt != IPPROTO_TCP) {
937 			CTR3(KTR_CXGBE, "%s: tid %d mbuf not IPPROTO_TCP (%u)",
938 			    __func__, tlsp->tid, ip6->ip6_nxt);
939 			return (EINVAL);
940 		}
941 		m->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
942 	}
943 	if (m->m_len < m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
944 	    sizeof(*tcp)) {
945 		CTR2(KTR_CXGBE, "%s: tid %d header mbuf too short (2)",
946 		    __func__, tlsp->tid);
947 		return (EINVAL);
948 	}
949 	tcp = (struct tcphdr *)((char *)(eh + 1) + m->m_pkthdr.l3hlen);
950 	m->m_pkthdr.l4hlen = tcp->th_off * 4;
951 
952 	/* Bail if there is TCP payload before the TLS record. */
953 	if (m->m_len != m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
954 	    m->m_pkthdr.l4hlen) {
955 		CTR6(KTR_CXGBE,
956 		    "%s: tid %d header mbuf bad length (%d + %d + %d != %d)",
957 		    __func__, tlsp->tid, m->m_pkthdr.l2hlen,
958 		    m->m_pkthdr.l3hlen, m->m_pkthdr.l4hlen, m->m_len);
959 		return (EINVAL);
960 	}
961 
962 	/* Assume all headers are in 'm' for now. */
963 	MPASS(m->m_next != NULL);
964 	MPASS(m->m_next->m_flags & M_EXTPG);
965 
966 	tot_len = 0;
967 
968 	/*
969 	 * Each of the remaining mbufs in the chain should reference a
970 	 * TLS record.
971 	 */
972 	for (m_tls = m->m_next; m_tls != NULL; m_tls = m_tls->m_next) {
973 		MPASS(m_tls->m_flags & M_EXTPG);
974 
975 		wr_len = ktls_wr_len(tlsp, m, m_tls, &nsegs);
976 #ifdef VERBOSE_TRACES
977 		CTR4(KTR_CXGBE, "%s: tid %d wr_len %d nsegs %d", __func__,
978 		    tlsp->tid, wr_len, nsegs);
979 #endif
980 		if (wr_len > SGE_MAX_WR_LEN || nsegs > TX_SGL_SEGS)
981 			return (EFBIG);
982 		tot_len += roundup2(wr_len, EQ_ESIZE);
983 
984 		/*
985 		 * Store 'nsegs' for the first TLS record in the
986 		 * header mbuf's metadata.
987 		 */
988 		if (m_tls == m->m_next)
989 			set_mbuf_nsegs(m, nsegs);
990 	}
991 
992 	MPASS(tot_len != 0);
993 
994 	/*
995 	 * See if we have any TCP options or a FIN requiring a
996 	 * dedicated packet.
997 	 */
998 	if ((tcp->th_flags & TH_FIN) != 0 || ktls_has_tcp_options(tcp)) {
999 		wr_len = sizeof(struct fw_eth_tx_pkt_wr) +
1000 		    sizeof(struct cpl_tx_pkt_core) + roundup2(m->m_len, 16);
1001 		if (wr_len > SGE_MAX_WR_LEN) {
1002 			CTR3(KTR_CXGBE,
1003 			    "%s: tid %d options-only packet too long (len %d)",
1004 			    __func__, tlsp->tid, m->m_len);
1005 			return (EINVAL);
1006 		}
1007 		tot_len += roundup2(wr_len, EQ_ESIZE);
1008 	}
1009 
1010 	/* Include room for a TP work request to program an L2T entry. */
1011 	tot_len += EQ_ESIZE;
1012 
1013 	/*
1014 	 * Include room for a ULPTX work request including up to 5
1015 	 * CPL_SET_TCB_FIELD commands before the first TLS work
1016 	 * request.
1017 	 */
1018 	wr_len = sizeof(struct fw_ulptx_wr) +
1019 	    5 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1020 
1021 	/*
1022 	 * If timestamps are present, reserve 1 more command for
1023 	 * setting the echoed timestamp.
1024 	 */
1025 	if (tlsp->using_timestamps)
1026 		wr_len += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1027 
1028 	tot_len += roundup2(wr_len, EQ_ESIZE);
1029 
1030 	set_mbuf_len16(m, tot_len / 16);
1031 #ifdef VERBOSE_TRACES
1032 	CTR4(KTR_CXGBE, "%s: tid %d len16 %d nsegs %d", __func__,
1033 	    tlsp->tid, mbuf_len16(m), mbuf_nsegs(m));
1034 #endif
1035 	items[0] = m;
1036 	return (mp_ring_enqueue(tlsp->txq->r, items, 1, 256));
1037 }
1038 
1039 /*
1040  * If the SGL ends on an address that is not 16 byte aligned, this function will
1041  * add a 0 filled flit at the end.
1042  */
1043 static void
write_gl_to_buf(struct sglist * gl,caddr_t to)1044 write_gl_to_buf(struct sglist *gl, caddr_t to)
1045 {
1046 	struct sglist_seg *seg;
1047 	__be64 *flitp;
1048 	struct ulptx_sgl *usgl;
1049 	int i, nflits, nsegs;
1050 
1051 	KASSERT(((uintptr_t)to & 0xf) == 0,
1052 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, to));
1053 
1054 	nsegs = gl->sg_nseg;
1055 	MPASS(nsegs > 0);
1056 
1057 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
1058 	flitp = (__be64 *)to;
1059 	seg = &gl->sg_segs[0];
1060 	usgl = (void *)flitp;
1061 
1062 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1063 	    V_ULPTX_NSGE(nsegs));
1064 	usgl->len0 = htobe32(seg->ss_len);
1065 	usgl->addr0 = htobe64(seg->ss_paddr);
1066 	seg++;
1067 
1068 	for (i = 0; i < nsegs - 1; i++, seg++) {
1069 		usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
1070 		usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
1071 	}
1072 	if (i & 1)
1073 		usgl->sge[i / 2].len[1] = htobe32(0);
1074 	flitp += nflits;
1075 
1076 	if (nflits & 1) {
1077 		MPASS(((uintptr_t)flitp) & 0xf);
1078 		*flitp++ = 0;
1079 	}
1080 
1081 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
1082 }
1083 
1084 static inline void
copy_to_txd(struct sge_eq * eq,caddr_t from,caddr_t * to,int len)1085 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
1086 {
1087 
1088 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
1089 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
1090 
1091 	if (__predict_true((uintptr_t)(*to) + len <=
1092 	    (uintptr_t)&eq->desc[eq->sidx])) {
1093 		bcopy(from, *to, len);
1094 		(*to) += len;
1095 		if ((uintptr_t)(*to) == (uintptr_t)&eq->desc[eq->sidx])
1096 			(*to) = (caddr_t)eq->desc;
1097 	} else {
1098 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
1099 
1100 		bcopy(from, *to, portion);
1101 		from += portion;
1102 		portion = len - portion;	/* remaining */
1103 		bcopy(from, (void *)eq->desc, portion);
1104 		(*to) = (caddr_t)eq->desc + portion;
1105 	}
1106 }
1107 
1108 static int
ktls_write_tcp_options(struct sge_txq * txq,void * dst,struct mbuf * m,u_int available,u_int pidx)1109 ktls_write_tcp_options(struct sge_txq *txq, void *dst, struct mbuf *m,
1110     u_int available, u_int pidx)
1111 {
1112 	struct tx_sdesc *txsd;
1113 	struct fw_eth_tx_pkt_wr *wr;
1114 	struct cpl_tx_pkt_core *cpl;
1115 	uint32_t ctrl;
1116 	uint64_t ctrl1;
1117 	int len16, ndesc, pktlen;
1118 	struct ether_header *eh;
1119 	struct ip *ip, newip;
1120 	struct ip6_hdr *ip6, newip6;
1121 	struct tcphdr *tcp, newtcp;
1122 	caddr_t out;
1123 
1124 	TXQ_LOCK_ASSERT_OWNED(txq);
1125 	M_ASSERTPKTHDR(m);
1126 
1127 	wr = dst;
1128 	pktlen = m->m_len;
1129 	ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen;
1130 	len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16);
1131 	ndesc = tx_len16_to_desc(len16);
1132 	MPASS(ndesc <= available);
1133 
1134 	/* Firmware work request header */
1135 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
1136 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
1137 
1138 	ctrl = V_FW_WR_LEN16(len16);
1139 	wr->equiq_to_len16 = htobe32(ctrl);
1140 	wr->r3 = 0;
1141 
1142 	cpl = (void *)(wr + 1);
1143 
1144 	/* CPL header */
1145 	cpl->ctrl0 = txq->cpl_ctrl0;
1146 	cpl->pack = 0;
1147 	cpl->len = htobe16(pktlen);
1148 
1149 	out = (void *)(cpl + 1);
1150 
1151 	/* Copy over Ethernet header. */
1152 	eh = mtod(m, struct ether_header *);
1153 	copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen);
1154 
1155 	/* Fixup length in IP header and copy out. */
1156 	if (ntohs(eh->ether_type) == ETHERTYPE_IP) {
1157 		ip = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1158 		newip = *ip;
1159 		newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen);
1160 		copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip));
1161 		if (m->m_pkthdr.l3hlen > sizeof(*ip))
1162 			copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out,
1163 			    m->m_pkthdr.l3hlen - sizeof(*ip));
1164 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) |
1165 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
1166 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
1167 	} else {
1168 		ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1169 		newip6 = *ip6;
1170 		newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen);
1171 		copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6));
1172 		MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6));
1173 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) |
1174 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
1175 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
1176 	}
1177 	cpl->ctrl1 = htobe64(ctrl1);
1178 	txq->txcsum++;
1179 
1180 	/* Clear PUSH and FIN in the TCP header if present. */
1181 	tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen);
1182 	newtcp = *tcp;
1183 	newtcp.th_flags &= ~(TH_PUSH | TH_FIN);
1184 	copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp));
1185 
1186 	/* Copy rest of packet. */
1187 	copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, pktlen -
1188 	    (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp)));
1189 	txq->imm_wrs++;
1190 
1191 	txq->txpkt_wrs++;
1192 
1193 	txq->kern_tls_options++;
1194 
1195 	txsd = &txq->sdesc[pidx];
1196 	txsd->m = NULL;
1197 	txsd->desc_used = ndesc;
1198 
1199 	return (ndesc);
1200 }
1201 
1202 static int
ktls_write_tunnel_packet(struct sge_txq * txq,void * dst,struct mbuf * m,struct mbuf * m_tls,u_int available,tcp_seq tcp_seqno,u_int pidx)1203 ktls_write_tunnel_packet(struct sge_txq *txq, void *dst, struct mbuf *m,
1204     struct mbuf *m_tls, u_int available, tcp_seq tcp_seqno, u_int pidx)
1205 {
1206 	struct tx_sdesc *txsd;
1207 	struct fw_eth_tx_pkt_wr *wr;
1208 	struct cpl_tx_pkt_core *cpl;
1209 	uint32_t ctrl;
1210 	uint64_t ctrl1;
1211 	int len16, ndesc, pktlen;
1212 	struct ether_header *eh;
1213 	struct ip *ip, newip;
1214 	struct ip6_hdr *ip6, newip6;
1215 	struct tcphdr *tcp, newtcp;
1216 	caddr_t out;
1217 
1218 	TXQ_LOCK_ASSERT_OWNED(txq);
1219 	M_ASSERTPKTHDR(m);
1220 
1221 	/* Locate the template TLS header. */
1222 	M_ASSERTEXTPG(m_tls);
1223 
1224 	/* This should always be the last TLS record in a chain. */
1225 	MPASS(m_tls->m_next == NULL);
1226 
1227 	wr = dst;
1228 	pktlen = m->m_len + m_tls->m_len;
1229 	ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen;
1230 	len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16);
1231 	ndesc = tx_len16_to_desc(len16);
1232 	MPASS(ndesc <= available);
1233 
1234 	/* Firmware work request header */
1235 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
1236 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
1237 
1238 	ctrl = V_FW_WR_LEN16(len16);
1239 	wr->equiq_to_len16 = htobe32(ctrl);
1240 	wr->r3 = 0;
1241 
1242 	cpl = (void *)(wr + 1);
1243 
1244 	/* CPL header */
1245 	cpl->ctrl0 = txq->cpl_ctrl0;
1246 	cpl->pack = 0;
1247 	cpl->len = htobe16(pktlen);
1248 
1249 	out = (void *)(cpl + 1);
1250 
1251 	/* Copy over Ethernet header. */
1252 	eh = mtod(m, struct ether_header *);
1253 	copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen);
1254 
1255 	/* Fixup length in IP header and copy out. */
1256 	if (ntohs(eh->ether_type) == ETHERTYPE_IP) {
1257 		ip = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1258 		newip = *ip;
1259 		newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen);
1260 		copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip));
1261 		if (m->m_pkthdr.l3hlen > sizeof(*ip))
1262 			copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out,
1263 			    m->m_pkthdr.l3hlen - sizeof(*ip));
1264 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) |
1265 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
1266 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
1267 	} else {
1268 		ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1269 		newip6 = *ip6;
1270 		newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen);
1271 		copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6));
1272 		MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6));
1273 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) |
1274 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
1275 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
1276 	}
1277 	cpl->ctrl1 = htobe64(ctrl1);
1278 	txq->txcsum++;
1279 
1280 	/* Set sequence number in TCP header. */
1281 	tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen);
1282 	newtcp = *tcp;
1283 	newtcp.th_seq = htonl(tcp_seqno + mtod(m_tls, vm_offset_t));
1284 	copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp));
1285 
1286 	/* Copy rest of TCP header. */
1287 	copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, m->m_len -
1288 	    (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp)));
1289 
1290 	/* Copy the subset of the TLS header requested. */
1291 	copy_to_txd(&txq->eq, (char *)m_tls->m_epg_hdr +
1292 	    mtod(m_tls, vm_offset_t), &out, m_tls->m_len);
1293 	txq->imm_wrs++;
1294 
1295 	txq->txpkt_wrs++;
1296 
1297 	txq->kern_tls_header++;
1298 
1299 	txsd = &txq->sdesc[pidx];
1300 	txsd->m = m;
1301 	txsd->desc_used = ndesc;
1302 
1303 	return (ndesc);
1304 }
1305 
1306 _Static_assert(sizeof(struct cpl_set_tcb_field) <= EQ_ESIZE,
1307     "CPL_SET_TCB_FIELD must be smaller than a single TX descriptor");
1308 _Static_assert(W_TCB_SND_UNA_RAW == W_TCB_SND_NXT_RAW,
1309     "SND_NXT_RAW and SND_UNA_RAW are in different words");
1310 
1311 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,uint32_t * tsopt,u_int pidx,bool set_l2t_idx)1312 ktls_write_tls_wr(struct tlspcb *tlsp, struct sge_txq *txq, void *dst,
1313     struct mbuf *m, struct tcphdr *tcp, struct mbuf *m_tls, u_int available,
1314     tcp_seq tcp_seqno, uint32_t *tsopt, u_int pidx, bool set_l2t_idx)
1315 {
1316 	struct sge_eq *eq = &txq->eq;
1317 	struct tx_sdesc *txsd;
1318 	struct fw_ulptx_wr *wr;
1319 	struct ulp_txpkt *txpkt;
1320 	struct ulptx_sc_memrd *memrd;
1321 	struct ulptx_idata *idata;
1322 	struct cpl_tx_sec_pdu *sec_pdu;
1323 	struct cpl_tx_data *tx_data;
1324 	struct tls_record_layer *hdr;
1325 	char *iv, *out;
1326 	u_int aad_start, aad_stop;
1327 	u_int auth_start, auth_stop, auth_insert;
1328 	u_int cipher_start, cipher_stop, iv_offset;
1329 	u_int imm_len, mss, ndesc, offset, plen, tlen, twr_len, wr_len;
1330 	u_int fields, nsegs, tx_max_offset, tx_max;
1331 	bool first_wr, last_wr, using_scratch;
1332 
1333 	ndesc = 0;
1334 	MPASS(tlsp->txq == txq);
1335 
1336 	first_wr = (tlsp->prev_seq == 0 && tlsp->prev_ack == 0 &&
1337 	    tlsp->prev_win == 0);
1338 
1339 	/*
1340 	 * Use the per-txq scratch pad if near the end of the ring to
1341 	 * simplify handling of wrap-around.  This uses a simple but
1342 	 * not quite perfect test of using the scratch buffer if we
1343 	 * can't fit a maximal work request in without wrapping.
1344 	 */
1345 	using_scratch = (eq->sidx - pidx < SGE_MAX_WR_LEN / EQ_ESIZE);
1346 
1347 	/* Locate the TLS header. */
1348 	M_ASSERTEXTPG(m_tls);
1349 	hdr = (void *)m_tls->m_epg_hdr;
1350 	plen = TLS_HEADER_LENGTH + ntohs(hdr->tls_length) - m_tls->m_epg_trllen;
1351 
1352 	/* Determine how much of the TLS record to send. */
1353 	tlen = ktls_tcp_payload_length(tlsp, m_tls);
1354 	if (tlen <= m_tls->m_epg_hdrlen) {
1355 		/*
1356 		 * For requests that only want to send the TLS header,
1357 		 * send a tunnelled packet as immediate data.
1358 		 */
1359 #ifdef VERBOSE_TRACES
1360 		CTR3(KTR_CXGBE, "%s: tid %d header-only TLS record %u",
1361 		    __func__, tlsp->tid, (u_int)m_tls->m_epg_seqno);
1362 #endif
1363 		return (ktls_write_tunnel_packet(txq, dst, m, m_tls, available,
1364 		    tcp_seqno, pidx));
1365 	}
1366 	if (tlen < plen) {
1367 		plen = tlen;
1368 		offset = ktls_payload_offset(tlsp, m_tls);
1369 #ifdef VERBOSE_TRACES
1370 		CTR4(KTR_CXGBE, "%s: tid %d short TLS record %u with offset %u",
1371 		    __func__, tlsp->tid, (u_int)m_tls->m_epg_seqno, offset);
1372 #endif
1373 		if (m_tls->m_next == NULL && (tcp->th_flags & TH_FIN) != 0) {
1374 			txq->kern_tls_fin_short++;
1375 #ifdef INVARIANTS
1376 			panic("%s: FIN on short TLS record", __func__);
1377 #endif
1378 		}
1379 	} else
1380 		offset = 0;
1381 
1382 	/*
1383 	 * This is the last work request for a given TLS mbuf chain if
1384 	 * it is the last mbuf in the chain and FIN is not set.  If
1385 	 * FIN is set, then ktls_write_tcp_fin() will write out the
1386 	 * last work request.
1387 	 */
1388 	last_wr = m_tls->m_next == NULL && (tcp->th_flags & TH_FIN) == 0;
1389 
1390 	/*
1391 	 * The host stack may ask us to not send part of the start of
1392 	 * a TLS record.  (For example, the stack might have
1393 	 * previously sent a "short" TLS record and might later send
1394 	 * down an mbuf that requests to send the remainder of the TLS
1395 	 * record.)  The crypto engine must process a TLS record from
1396 	 * the beginning if computing a GCM tag or HMAC, so we always
1397 	 * send the TLS record from the beginning as input to the
1398 	 * crypto engine and via CPL_TX_DATA to TP.  However, TP will
1399 	 * drop individual packets after they have been chopped up
1400 	 * into MSS-sized chunks if the entire sequence range of those
1401 	 * packets is less than SND_UNA.  SND_UNA is computed as
1402 	 * TX_MAX - SND_UNA_RAW.  Thus, use the offset stored in
1403 	 * m_data to set TX_MAX to the first byte in the TCP sequence
1404 	 * space the host actually wants us to send and set
1405 	 * SND_UNA_RAW to 0.
1406 	 *
1407 	 * If the host sends us back to back requests that span the
1408 	 * trailer of a single TLS record (first request ends "in" the
1409 	 * trailer and second request starts at the next byte but
1410 	 * still "in" the trailer), the initial bytes of the trailer
1411 	 * that the first request drops will not be retransmitted.  If
1412 	 * the host uses the same requests when retransmitting the
1413 	 * connection will hang.  To handle this, always transmit the
1414 	 * full trailer for a request that begins "in" the trailer
1415 	 * (the second request in the example above).  This should
1416 	 * also help to avoid retransmits for the common case.
1417 	 *
1418 	 * A similar condition exists when using CBC for back to back
1419 	 * requests that span a single AES block.  The first request
1420 	 * will be truncated to end at the end of the previous AES
1421 	 * block.  To handle this, always begin transmission at the
1422 	 * start of the current AES block.
1423 	 */
1424 	tx_max_offset = mtod(m_tls, vm_offset_t);
1425 	if (tx_max_offset > TLS_HEADER_LENGTH + ntohs(hdr->tls_length) -
1426 	    m_tls->m_epg_trllen) {
1427 		/* Always send the full trailer. */
1428 		tx_max_offset = TLS_HEADER_LENGTH + ntohs(hdr->tls_length) -
1429 		    m_tls->m_epg_trllen;
1430 	}
1431 	if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_CBC &&
1432 	    tx_max_offset > TLS_HEADER_LENGTH) {
1433 		/* Always send all of the first AES block. */
1434 		tx_max_offset = TLS_HEADER_LENGTH +
1435 		    rounddown(tx_max_offset - TLS_HEADER_LENGTH,
1436 		    AES_BLOCK_LEN);
1437 	}
1438 	tx_max = tcp_seqno + tx_max_offset;
1439 
1440 	/*
1441 	 * Update TCB fields.  Reserve space for the FW_ULPTX_WR header
1442 	 * but don't populate it until we know how many field updates
1443 	 * are required.
1444 	 */
1445 	if (using_scratch)
1446 		wr = (void *)txq->ss;
1447 	else
1448 		wr = dst;
1449 	out = (void *)(wr + 1);
1450 	fields = 0;
1451 	if (set_l2t_idx) {
1452 		KASSERT(m->m_next == m_tls,
1453 		    ("trying to set L2T_IX for subsequent TLS WR"));
1454 #ifdef VERBOSE_TRACES
1455 		CTR3(KTR_CXGBE, "%s: tid %d set L2T_IX to %d", __func__,
1456 		    tlsp->tid, tlsp->l2te->idx);
1457 #endif
1458 		write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_L2T_IX,
1459 		    V_TCB_L2T_IX(M_TCB_L2T_IX), V_TCB_L2T_IX(tlsp->l2te->idx));
1460 		out += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1461 		fields++;
1462 	}
1463 	if (tsopt != NULL && tlsp->prev_tsecr != ntohl(tsopt[1])) {
1464 		KASSERT(m->m_next == m_tls,
1465 		    ("trying to set T_RTSEQ_RECENT for subsequent TLS WR"));
1466 #ifdef VERBOSE_TRACES
1467 		CTR2(KTR_CXGBE, "%s: tid %d wrote updated T_RTSEQ_RECENT",
1468 		    __func__, tlsp->tid);
1469 #endif
1470 		write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_T_RTSEQ_RECENT,
1471 		    V_TCB_T_RTSEQ_RECENT(M_TCB_T_RTSEQ_RECENT),
1472 		    V_TCB_T_RTSEQ_RECENT(ntohl(tsopt[1])));
1473 		out += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1474 		fields++;
1475 
1476 		tlsp->prev_tsecr = ntohl(tsopt[1]);
1477 	}
1478 
1479 	if (first_wr || tlsp->prev_seq != tx_max) {
1480 		KASSERT(m->m_next == m_tls,
1481 		    ("trying to set TX_MAX for subsequent TLS WR"));
1482 #ifdef VERBOSE_TRACES
1483 		CTR4(KTR_CXGBE,
1484 		    "%s: tid %d setting TX_MAX to %u (tcp_seqno %u)",
1485 		    __func__, tlsp->tid, tx_max, tcp_seqno);
1486 #endif
1487 		write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_TX_MAX,
1488 		    V_TCB_TX_MAX(M_TCB_TX_MAX), V_TCB_TX_MAX(tx_max));
1489 		out += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1490 		fields++;
1491 	}
1492 
1493 	/*
1494 	 * If there is data to drop at the beginning of this TLS
1495 	 * record or if this is a retransmit,
1496 	 * reset SND_UNA_RAW to 0 so that SND_UNA == TX_MAX.
1497 	 */
1498 	if (tlsp->prev_seq != tx_max || mtod(m_tls, vm_offset_t) != 0) {
1499 		KASSERT(m->m_next == m_tls,
1500 		    ("trying to clear SND_UNA_RAW for subsequent TLS WR"));
1501 #ifdef VERBOSE_TRACES
1502 		CTR2(KTR_CXGBE, "%s: tid %d clearing SND_UNA_RAW", __func__,
1503 		    tlsp->tid);
1504 #endif
1505 		write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_SND_UNA_RAW,
1506 		    V_TCB_SND_UNA_RAW(M_TCB_SND_UNA_RAW),
1507 		    V_TCB_SND_UNA_RAW(0));
1508 		out += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1509 		fields++;
1510 	}
1511 
1512 	/*
1513 	 * Store the expected sequence number of the next byte after
1514 	 * this record.
1515 	 */
1516 	tlsp->prev_seq = tcp_seqno + tlen;
1517 
1518 	if (first_wr || tlsp->prev_ack != ntohl(tcp->th_ack)) {
1519 		KASSERT(m->m_next == m_tls,
1520 		    ("trying to set RCV_NXT for subsequent TLS WR"));
1521 		write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_RCV_NXT,
1522 		    V_TCB_RCV_NXT(M_TCB_RCV_NXT),
1523 		    V_TCB_RCV_NXT(ntohl(tcp->th_ack)));
1524 		out += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1525 		fields++;
1526 
1527 		tlsp->prev_ack = ntohl(tcp->th_ack);
1528 	}
1529 
1530 	if (first_wr || tlsp->prev_win != ntohs(tcp->th_win)) {
1531 		KASSERT(m->m_next == m_tls,
1532 		    ("trying to set RCV_WND for subsequent TLS WR"));
1533 		write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_RCV_WND,
1534 		    V_TCB_RCV_WND(M_TCB_RCV_WND),
1535 		    V_TCB_RCV_WND(ntohs(tcp->th_win)));
1536 		out += roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1537 		fields++;
1538 
1539 		tlsp->prev_win = ntohs(tcp->th_win);
1540 	}
1541 
1542 	/* Use cached value for first record in chain. */
1543 	if (m->m_next == m_tls)
1544 		nsegs = mbuf_nsegs(m);
1545 	else
1546 		nsegs = sglist_count_mbuf_epg(m_tls, m_tls->m_epg_hdrlen +
1547 		    offset, plen - (m_tls->m_epg_hdrlen + offset));
1548 
1549 	/* Calculate the size of the TLS work request. */
1550 	twr_len = ktls_base_wr_size(tlsp);
1551 
1552 	imm_len = 0;
1553 	if (offset == 0)
1554 		imm_len += m_tls->m_epg_hdrlen;
1555 	if (plen == tlen)
1556 		imm_len += AES_BLOCK_LEN;
1557 	twr_len += roundup2(imm_len, 16);
1558 	twr_len += ktls_sgl_size(nsegs);
1559 
1560 	/*
1561 	 * If any field updates were required, determine if they can
1562 	 * be included in the TLS work request.  If not, use the
1563 	 * FW_ULPTX_WR work request header at 'wr' as a dedicated work
1564 	 * request for the field updates and start a new work request
1565 	 * for the TLS work request afterward.
1566 	 */
1567 	if (fields != 0) {
1568 		wr_len = fields * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
1569 		if (twr_len + wr_len <= SGE_MAX_WR_LEN &&
1570 		    tlsp->sc->tlst.combo_wrs) {
1571 			wr_len += twr_len;
1572 			txpkt = (void *)out;
1573 		} else {
1574 			wr_len += sizeof(*wr);
1575 			wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR));
1576 			wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA |
1577 			    V_FW_WR_LEN16(wr_len / 16));
1578 			wr->cookie = 0;
1579 
1580 			/*
1581 			 * If we were using scratch space, copy the
1582 			 * field updates work request to the ring.
1583 			 */
1584 			if (using_scratch) {
1585 				out = dst;
1586 				copy_to_txd(eq, txq->ss, &out, wr_len);
1587 			}
1588 
1589 			ndesc = howmany(wr_len, EQ_ESIZE);
1590 			MPASS(ndesc <= available);
1591 
1592 			txq->raw_wrs++;
1593 			txsd = &txq->sdesc[pidx];
1594 			txsd->m = NULL;
1595 			txsd->desc_used = ndesc;
1596 			IDXINCR(pidx, ndesc, eq->sidx);
1597 			dst = &eq->desc[pidx];
1598 
1599 			/*
1600 			 * Determine if we should use scratch space
1601 			 * for the TLS work request based on the
1602 			 * available space after advancing pidx for
1603 			 * the field updates work request.
1604 			 */
1605 			wr_len = twr_len;
1606 			using_scratch = (eq->sidx - pidx <
1607 			    howmany(wr_len, EQ_ESIZE));
1608 			if (using_scratch)
1609 				wr = (void *)txq->ss;
1610 			else
1611 				wr = dst;
1612 			txpkt = (void *)(wr + 1);
1613 		}
1614 	} else {
1615 		wr_len = twr_len;
1616 		txpkt = (void *)out;
1617 	}
1618 
1619 	wr_len = roundup2(wr_len, 16);
1620 	MPASS(ndesc + howmany(wr_len, EQ_ESIZE) <= available);
1621 
1622 	/* FW_ULPTX_WR */
1623 	wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR));
1624 	wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA |
1625 	    V_FW_WR_LEN16(wr_len / 16));
1626 	wr->cookie = 0;
1627 
1628 	/* ULP_TXPKT */
1629 	txpkt->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
1630 	    V_ULP_TXPKT_DATAMODIFY(0) |
1631 	    V_ULP_TXPKT_CHANNELID(tlsp->vi->pi->port_id) | V_ULP_TXPKT_DEST(0) |
1632 	    V_ULP_TXPKT_FID(txq->eq.cntxt_id) | V_ULP_TXPKT_RO(1));
1633 	txpkt->len = htobe32(howmany(twr_len - sizeof(*wr), 16));
1634 
1635 	/* ULPTX_IDATA sub-command */
1636 	idata = (void *)(txpkt + 1);
1637 	idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
1638 	    V_ULP_TX_SC_MORE(1));
1639 	idata->len = sizeof(struct cpl_tx_sec_pdu);
1640 
1641 	/*
1642 	 * The key context, CPL_TX_DATA, and immediate data are part
1643 	 * of this ULPTX_IDATA when using an inline key.  When reading
1644 	 * the key from memory, the CPL_TX_DATA and immediate data are
1645 	 * part of a separate ULPTX_IDATA.
1646 	 */
1647 	if (tlsp->inline_key)
1648 		idata->len += tlsp->tx_key_info_size +
1649 		    sizeof(struct cpl_tx_data) + imm_len;
1650 	idata->len = htobe32(idata->len);
1651 
1652 	/* CPL_TX_SEC_PDU */
1653 	sec_pdu = (void *)(idata + 1);
1654 
1655 	/*
1656 	 * For short records, AAD is counted as header data in SCMD0,
1657 	 * the IV is next followed by a cipher region for the payload.
1658 	 */
1659 	if (plen == tlen) {
1660 		aad_start = 0;
1661 		aad_stop = 0;
1662 		iv_offset = 1;
1663 		auth_start = 0;
1664 		auth_stop = 0;
1665 		auth_insert = 0;
1666 		cipher_start = AES_BLOCK_LEN + 1;
1667 		cipher_stop = 0;
1668 
1669 		sec_pdu->pldlen = htobe32(16 + plen -
1670 		    (m_tls->m_epg_hdrlen + offset));
1671 
1672 		/* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */
1673 		sec_pdu->seqno_numivs = tlsp->scmd0_short.seqno_numivs;
1674 		sec_pdu->ivgen_hdrlen = htobe32(
1675 		    tlsp->scmd0_short.ivgen_hdrlen |
1676 		    V_SCMD_HDR_LEN(offset == 0 ? m_tls->m_epg_hdrlen : 0));
1677 
1678 		txq->kern_tls_short++;
1679 	} else {
1680 		/*
1681 		 * AAD is TLS header.  IV is after AAD.  The cipher region
1682 		 * starts after the IV.  See comments in ccr_authenc() and
1683 		 * ccr_gmac() in t4_crypto.c regarding cipher and auth
1684 		 * start/stop values.
1685 		 */
1686 		aad_start = 1;
1687 		aad_stop = TLS_HEADER_LENGTH;
1688 		iv_offset = TLS_HEADER_LENGTH + 1;
1689 		cipher_start = m_tls->m_epg_hdrlen + 1;
1690 		if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
1691 			cipher_stop = 0;
1692 			auth_start = cipher_start;
1693 			auth_stop = 0;
1694 			auth_insert = 0;
1695 		} else {
1696 			cipher_stop = 0;
1697 			auth_start = cipher_start;
1698 			auth_stop = 0;
1699 			auth_insert = 0;
1700 		}
1701 
1702 		sec_pdu->pldlen = htobe32(plen);
1703 
1704 		/* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */
1705 		sec_pdu->seqno_numivs = tlsp->scmd0.seqno_numivs;
1706 		sec_pdu->ivgen_hdrlen = tlsp->scmd0.ivgen_hdrlen;
1707 
1708 		if (mtod(m_tls, vm_offset_t) == 0)
1709 			txq->kern_tls_full++;
1710 		else
1711 			txq->kern_tls_partial++;
1712 	}
1713 	sec_pdu->op_ivinsrtofst = htobe32(
1714 	    V_CPL_TX_SEC_PDU_OPCODE(CPL_TX_SEC_PDU) |
1715 	    V_CPL_TX_SEC_PDU_CPLLEN(2) | V_CPL_TX_SEC_PDU_PLACEHOLDER(0) |
1716 	    V_CPL_TX_SEC_PDU_IVINSRTOFST(iv_offset));
1717 	sec_pdu->aadstart_cipherstop_hi = htobe32(
1718 	    V_CPL_TX_SEC_PDU_AADSTART(aad_start) |
1719 	    V_CPL_TX_SEC_PDU_AADSTOP(aad_stop) |
1720 	    V_CPL_TX_SEC_PDU_CIPHERSTART(cipher_start) |
1721 	    V_CPL_TX_SEC_PDU_CIPHERSTOP_HI(cipher_stop >> 4));
1722 	sec_pdu->cipherstop_lo_authinsert = htobe32(
1723 	    V_CPL_TX_SEC_PDU_CIPHERSTOP_LO(cipher_stop & 0xf) |
1724 	    V_CPL_TX_SEC_PDU_AUTHSTART(auth_start) |
1725 	    V_CPL_TX_SEC_PDU_AUTHSTOP(auth_stop) |
1726 	    V_CPL_TX_SEC_PDU_AUTHINSERT(auth_insert));
1727 
1728 	sec_pdu->scmd1 = htobe64(m_tls->m_epg_seqno);
1729 
1730 	/* Key context */
1731 	out = (void *)(sec_pdu + 1);
1732 	if (tlsp->inline_key) {
1733 		memcpy(out, &tlsp->keyctx, tlsp->tx_key_info_size);
1734 		out += tlsp->tx_key_info_size;
1735 	} else {
1736 		/* ULPTX_SC_MEMRD to read key context. */
1737 		memrd = (void *)out;
1738 		memrd->cmd_to_len = htobe32(V_ULPTX_CMD(ULP_TX_SC_MEMRD) |
1739 		    V_ULP_TX_SC_MORE(1) |
1740 		    V_ULPTX_LEN16(tlsp->tx_key_info_size >> 4));
1741 		memrd->addr = htobe32(tlsp->tx_key_addr >> 5);
1742 
1743 		/* ULPTX_IDATA for CPL_TX_DATA and TLS header. */
1744 		idata = (void *)(memrd + 1);
1745 		idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
1746 		    V_ULP_TX_SC_MORE(1));
1747 		idata->len = htobe32(sizeof(struct cpl_tx_data) + imm_len);
1748 
1749 		out = (void *)(idata + 1);
1750 	}
1751 
1752 	/* CPL_TX_DATA */
1753 	tx_data = (void *)out;
1754 	OPCODE_TID(tx_data) = htonl(MK_OPCODE_TID(CPL_TX_DATA, tlsp->tid));
1755 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1756 		mss = m->m_pkthdr.tso_segsz;
1757 		tlsp->prev_mss = mss;
1758 	} else if (tlsp->prev_mss != 0)
1759 		mss = tlsp->prev_mss;
1760 	else
1761 		mss = if_getmtu(tlsp->vi->ifp) -
1762 		    (m->m_pkthdr.l3hlen + m->m_pkthdr.l4hlen);
1763 	if (offset == 0) {
1764 		tx_data->len = htobe32(V_TX_DATA_MSS(mss) | V_TX_LENGTH(tlen));
1765 		tx_data->rsvd = htobe32(tcp_seqno);
1766 	} else {
1767 		tx_data->len = htobe32(V_TX_DATA_MSS(mss) |
1768 		    V_TX_LENGTH(tlen - (m_tls->m_epg_hdrlen + offset)));
1769 		tx_data->rsvd = htobe32(tcp_seqno + m_tls->m_epg_hdrlen + offset);
1770 	}
1771 	tx_data->flags = htobe32(F_TX_BYPASS);
1772 	if (last_wr && tcp->th_flags & TH_PUSH)
1773 		tx_data->flags |= htobe32(F_TX_PUSH | F_TX_SHOVE);
1774 
1775 	/* Populate the TLS header */
1776 	out = (void *)(tx_data + 1);
1777 	if (offset == 0) {
1778 		memcpy(out, m_tls->m_epg_hdr, m_tls->m_epg_hdrlen);
1779 		out += m_tls->m_epg_hdrlen;
1780 	}
1781 
1782 	/* AES IV for a short record. */
1783 	if (plen == tlen) {
1784 		iv = out;
1785 		if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) {
1786 			memcpy(iv, tlsp->keyctx.u.txhdr.txsalt, SALT_SIZE);
1787 			memcpy(iv + 4, hdr + 1, 8);
1788 			*(uint32_t *)(iv + 12) = htobe32(2 +
1789 			    offset / AES_BLOCK_LEN);
1790 		} else
1791 			memcpy(iv, hdr + 1, AES_BLOCK_LEN);
1792 		out += AES_BLOCK_LEN;
1793 	}
1794 
1795 	if (imm_len % 16 != 0) {
1796 		/* Zero pad to an 8-byte boundary. */
1797 		memset(out, 0, 8 - (imm_len % 8));
1798 		out += 8 - (imm_len % 8);
1799 
1800 		/*
1801 		 * Insert a ULP_TX_SC_NOOP if needed so the SGL is
1802 		 * 16-byte aligned.
1803 		 */
1804 		if (imm_len % 16 <= 8) {
1805 			idata = (void *)out;
1806 			idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
1807 			idata->len = htobe32(0);
1808 			out = (void *)(idata + 1);
1809 		}
1810 	}
1811 
1812 	/* SGL for record payload */
1813 	sglist_reset(txq->gl);
1814 	if (sglist_append_mbuf_epg(txq->gl, m_tls, m_tls->m_epg_hdrlen + offset,
1815 	    plen - (m_tls->m_epg_hdrlen + offset)) != 0) {
1816 #ifdef INVARIANTS
1817 		panic("%s: failed to append sglist", __func__);
1818 #endif
1819 	}
1820 	write_gl_to_buf(txq->gl, out);
1821 
1822 	if (using_scratch) {
1823 		out = dst;
1824 		copy_to_txd(eq, txq->ss, &out, wr_len);
1825 	}
1826 
1827 	ndesc += howmany(wr_len, EQ_ESIZE);
1828 	MPASS(ndesc <= available);
1829 
1830 	txq->kern_tls_records++;
1831 	txq->kern_tls_octets += tlen - mtod(m_tls, vm_offset_t);
1832 	if (mtod(m_tls, vm_offset_t) != 0) {
1833 		if (offset == 0)
1834 			txq->kern_tls_waste += mtod(m_tls, vm_offset_t);
1835 		else
1836 			txq->kern_tls_waste += mtod(m_tls, vm_offset_t) -
1837 			    (m_tls->m_epg_hdrlen + offset);
1838 	}
1839 
1840 	txsd = &txq->sdesc[pidx];
1841 	if (last_wr)
1842 		txsd->m = m;
1843 	else
1844 		txsd->m = NULL;
1845 	txsd->desc_used = howmany(wr_len, EQ_ESIZE);
1846 
1847 	return (ndesc);
1848 }
1849 
1850 static int
ktls_write_tcp_fin(struct sge_txq * txq,void * dst,struct mbuf * m,u_int available,tcp_seq tcp_seqno,u_int pidx)1851 ktls_write_tcp_fin(struct sge_txq *txq, void *dst, struct mbuf *m,
1852     u_int available, tcp_seq tcp_seqno, u_int pidx)
1853 {
1854 	struct tx_sdesc *txsd;
1855 	struct fw_eth_tx_pkt_wr *wr;
1856 	struct cpl_tx_pkt_core *cpl;
1857 	uint32_t ctrl;
1858 	uint64_t ctrl1;
1859 	int len16, ndesc, pktlen;
1860 	struct ether_header *eh;
1861 	struct ip *ip, newip;
1862 	struct ip6_hdr *ip6, newip6;
1863 	struct tcphdr *tcp, newtcp;
1864 	caddr_t out;
1865 
1866 	TXQ_LOCK_ASSERT_OWNED(txq);
1867 	M_ASSERTPKTHDR(m);
1868 
1869 	wr = dst;
1870 	pktlen = m->m_len;
1871 	ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen;
1872 	len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16);
1873 	ndesc = tx_len16_to_desc(len16);
1874 	MPASS(ndesc <= available);
1875 
1876 	/* Firmware work request header */
1877 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
1878 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
1879 
1880 	ctrl = V_FW_WR_LEN16(len16);
1881 	wr->equiq_to_len16 = htobe32(ctrl);
1882 	wr->r3 = 0;
1883 
1884 	cpl = (void *)(wr + 1);
1885 
1886 	/* CPL header */
1887 	cpl->ctrl0 = txq->cpl_ctrl0;
1888 	cpl->pack = 0;
1889 	cpl->len = htobe16(pktlen);
1890 
1891 	out = (void *)(cpl + 1);
1892 
1893 	/* Copy over Ethernet header. */
1894 	eh = mtod(m, struct ether_header *);
1895 	copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen);
1896 
1897 	/* Fixup length in IP header and copy out. */
1898 	if (ntohs(eh->ether_type) == ETHERTYPE_IP) {
1899 		ip = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1900 		newip = *ip;
1901 		newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen);
1902 		copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip));
1903 		if (m->m_pkthdr.l3hlen > sizeof(*ip))
1904 			copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out,
1905 			    m->m_pkthdr.l3hlen - sizeof(*ip));
1906 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) |
1907 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
1908 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
1909 	} else {
1910 		ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen);
1911 		newip6 = *ip6;
1912 		newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen);
1913 		copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6));
1914 		MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6));
1915 		ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) |
1916 		    V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) |
1917 		    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
1918 	}
1919 	cpl->ctrl1 = htobe64(ctrl1);
1920 	txq->txcsum++;
1921 
1922 	/* Set sequence number in TCP header. */
1923 	tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen);
1924 	newtcp = *tcp;
1925 	newtcp.th_seq = htonl(tcp_seqno);
1926 	copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp));
1927 
1928 	/* Copy rest of packet. */
1929 	copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, m->m_len -
1930 	    (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp)));
1931 	txq->imm_wrs++;
1932 
1933 	txq->txpkt_wrs++;
1934 
1935 	txq->kern_tls_fin++;
1936 
1937 	txsd = &txq->sdesc[pidx];
1938 	txsd->m = m;
1939 	txsd->desc_used = ndesc;
1940 
1941 	return (ndesc);
1942 }
1943 
1944 int
t6_ktls_write_wr(struct sge_txq * txq,void * dst,struct mbuf * m,u_int available)1945 t6_ktls_write_wr(struct sge_txq *txq, void *dst, struct mbuf *m,
1946     u_int available)
1947 {
1948 	struct sge_eq *eq = &txq->eq;
1949 	struct tx_sdesc *txsd;
1950 	struct tlspcb *tlsp;
1951 	struct tcphdr *tcp;
1952 	struct mbuf *m_tls;
1953 	struct ether_header *eh;
1954 	tcp_seq tcp_seqno;
1955 	u_int ndesc, pidx, totdesc;
1956 	uint16_t vlan_tag;
1957 	bool has_fin, set_l2t_idx;
1958 	void *tsopt;
1959 
1960 	M_ASSERTPKTHDR(m);
1961 	MPASS(m->m_pkthdr.snd_tag != NULL);
1962 	tlsp = mst_to_tls(m->m_pkthdr.snd_tag);
1963 
1964 	totdesc = 0;
1965 	eh = mtod(m, struct ether_header *);
1966 	tcp = (struct tcphdr *)((char *)eh + m->m_pkthdr.l2hlen +
1967 	    m->m_pkthdr.l3hlen);
1968 	pidx = eq->pidx;
1969 	has_fin = (tcp->th_flags & TH_FIN) != 0;
1970 
1971 	/*
1972 	 * If this TLS record has a FIN, then we will send any
1973 	 * requested options as part of the FIN packet.
1974 	 */
1975 	if (!has_fin && ktls_has_tcp_options(tcp)) {
1976 		ndesc = ktls_write_tcp_options(txq, dst, m, available, pidx);
1977 		totdesc += ndesc;
1978 		IDXINCR(pidx, ndesc, eq->sidx);
1979 		dst = &eq->desc[pidx];
1980 #ifdef VERBOSE_TRACES
1981 		CTR2(KTR_CXGBE, "%s: tid %d wrote TCP options packet", __func__,
1982 		    tlsp->tid);
1983 #endif
1984 	}
1985 
1986 	/*
1987 	 * Allocate a new L2T entry if necessary.  This may write out
1988 	 * a work request to the txq.
1989 	 */
1990 	if (m->m_flags & M_VLANTAG)
1991 		vlan_tag = m->m_pkthdr.ether_vtag;
1992 	else
1993 		vlan_tag = 0xfff;
1994 	set_l2t_idx = false;
1995 	if (tlsp->l2te == NULL || tlsp->l2te->vlan != vlan_tag ||
1996 	    memcmp(tlsp->l2te->dmac, eh->ether_dhost, ETHER_ADDR_LEN) != 0) {
1997 		set_l2t_idx = true;
1998 		if (tlsp->l2te)
1999 			t4_l2t_release(tlsp->l2te);
2000 		tlsp->l2te = t4_l2t_alloc_tls(tlsp->sc, txq, dst, &ndesc,
2001 		    vlan_tag, tlsp->vi->pi->lport, eh->ether_dhost);
2002 		if (tlsp->l2te == NULL)
2003 			CXGBE_UNIMPLEMENTED("failed to allocate TLS L2TE");
2004 		if (ndesc != 0) {
2005 			MPASS(ndesc <= available - totdesc);
2006 
2007 			txq->raw_wrs++;
2008 			txsd = &txq->sdesc[pidx];
2009 			txsd->m = NULL;
2010 			txsd->desc_used = ndesc;
2011 			totdesc += ndesc;
2012 			IDXINCR(pidx, ndesc, eq->sidx);
2013 			dst = &eq->desc[pidx];
2014 		}
2015 	}
2016 
2017 	/*
2018 	 * Iterate over each TLS record constructing a work request
2019 	 * for that record.
2020 	 */
2021 	for (m_tls = m->m_next; m_tls != NULL; m_tls = m_tls->m_next) {
2022 		MPASS(m_tls->m_flags & M_EXTPG);
2023 
2024 		/*
2025 		 * Determine the initial TCP sequence number for this
2026 		 * record.
2027 		 */
2028 		tsopt = NULL;
2029 		if (m_tls == m->m_next) {
2030 			tcp_seqno = ntohl(tcp->th_seq) -
2031 			    mtod(m_tls, vm_offset_t);
2032 			if (tlsp->using_timestamps)
2033 				tsopt = ktls_find_tcp_timestamps(tcp);
2034 		} else {
2035 			MPASS(mtod(m_tls, vm_offset_t) == 0);
2036 			tcp_seqno = tlsp->prev_seq;
2037 		}
2038 
2039 		ndesc = ktls_write_tls_wr(tlsp, txq, dst, m, tcp, m_tls,
2040 		    available - totdesc, tcp_seqno, tsopt, pidx, set_l2t_idx);
2041 		totdesc += ndesc;
2042 		IDXINCR(pidx, ndesc, eq->sidx);
2043 		dst = &eq->desc[pidx];
2044 
2045 		/* Only need to set the L2T index once. */
2046 		set_l2t_idx = false;
2047 	}
2048 
2049 	if (has_fin) {
2050 		/*
2051 		 * If the TCP header for this chain has FIN sent, then
2052 		 * explicitly send a packet that has FIN set.  This
2053 		 * will also have PUSH set if requested.  This assumes
2054 		 * we sent at least one TLS record work request and
2055 		 * uses the TCP sequence number after that reqeust as
2056 		 * the sequence number for the FIN packet.
2057 		 */
2058 		ndesc = ktls_write_tcp_fin(txq, dst, m, available,
2059 		    tlsp->prev_seq, pidx);
2060 		totdesc += ndesc;
2061 	}
2062 
2063 	MPASS(totdesc <= available);
2064 	return (totdesc);
2065 }
2066 
2067 static void
t6_tls_tag_free(struct m_snd_tag * mst)2068 t6_tls_tag_free(struct m_snd_tag *mst)
2069 {
2070 	struct adapter *sc;
2071 	struct tlspcb *tlsp;
2072 
2073 	tlsp = mst_to_tls(mst);
2074 	sc = tlsp->sc;
2075 
2076 	CTR2(KTR_CXGBE, "%s: tid %d", __func__, tlsp->tid);
2077 
2078 	if (tlsp->l2te)
2079 		t4_l2t_release(tlsp->l2te);
2080 	if (tlsp->tid >= 0)
2081 		release_tid(sc, tlsp->tid, tlsp->ctrlq);
2082 	if (tlsp->ce)
2083 		t4_release_clip_entry(sc, tlsp->ce);
2084 	if (tlsp->tx_key_addr >= 0)
2085 		t4_free_tls_keyid(sc, tlsp->tx_key_addr);
2086 
2087 	zfree(tlsp, M_CXGBE);
2088 }
2089 
2090 void
t6_ktls_modload(void)2091 t6_ktls_modload(void)
2092 {
2093 
2094 	t4_register_shared_cpl_handler(CPL_ACT_OPEN_RPL, ktls_act_open_rpl,
2095 	    CPL_COOKIE_KERN_TLS);
2096 }
2097 
2098 void
t6_ktls_modunload(void)2099 t6_ktls_modunload(void)
2100 {
2101 
2102 	t4_register_shared_cpl_handler(CPL_ACT_OPEN_RPL, NULL,
2103 	    CPL_COOKIE_KERN_TLS);
2104 }
2105 
2106 #else
2107 
2108 int
t6_tls_tag_alloc(if_t ifp,union if_snd_tag_alloc_params * params,struct m_snd_tag ** pt)2109 t6_tls_tag_alloc(if_t ifp, union if_snd_tag_alloc_params *params,
2110     struct m_snd_tag **pt)
2111 {
2112 	return (ENXIO);
2113 }
2114 
2115 int
t6_ktls_parse_pkt(struct mbuf * m)2116 t6_ktls_parse_pkt(struct mbuf *m)
2117 {
2118 	return (EINVAL);
2119 }
2120 
2121 int
t6_ktls_write_wr(struct sge_txq * txq,void * dst,struct mbuf * m,u_int available)2122 t6_ktls_write_wr(struct sge_txq *txq, void *dst, struct mbuf *m,
2123     u_int available)
2124 {
2125 	panic("can't happen");
2126 }
2127 
2128 void
t6_ktls_modload(void)2129 t6_ktls_modload(void)
2130 {
2131 }
2132 
2133 void
t6_ktls_modunload(void)2134 t6_ktls_modunload(void)
2135 {
2136 }
2137 
2138 #endif
2139