xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_hw_tls.c (revision 1323ec57)
1 /*-
2  * Copyright (c) 2019-2021 Mellanox Technologies. All rights reserved.
3  * Copyright (c) 2022 NVIDIA corporation & affiliates.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28 
29 #include "opt_kern_tls.h"
30 #include "opt_rss.h"
31 #include "opt_ratelimit.h"
32 
33 #include <dev/mlx5/mlx5_en/en.h>
34 
35 #include <dev/mlx5/tls.h>
36 
37 #include <linux/delay.h>
38 #include <sys/ktls.h>
39 #include <opencrypto/cryptodev.h>
40 
41 #ifdef KERN_TLS
42 
43 #ifdef RATELIMIT
44 static if_snd_tag_modify_t mlx5e_tls_rl_snd_tag_modify;
45 #endif
46 static if_snd_tag_query_t mlx5e_tls_snd_tag_query;
47 static if_snd_tag_free_t mlx5e_tls_snd_tag_free;
48 
49 static const struct if_snd_tag_sw mlx5e_tls_snd_tag_sw = {
50 	.snd_tag_query = mlx5e_tls_snd_tag_query,
51 	.snd_tag_free = mlx5e_tls_snd_tag_free,
52 	.type = IF_SND_TAG_TYPE_TLS
53 };
54 
55 #ifdef RATELIMIT
56 static const struct if_snd_tag_sw mlx5e_tls_rl_snd_tag_sw = {
57 	.snd_tag_modify = mlx5e_tls_rl_snd_tag_modify,
58 	.snd_tag_query = mlx5e_tls_snd_tag_query,
59 	.snd_tag_free = mlx5e_tls_snd_tag_free,
60 	.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT
61 };
62 #endif
63 
64 MALLOC_DEFINE(M_MLX5E_TLS, "MLX5E_TLS", "MLX5 ethernet HW TLS");
65 
66 /* software TLS context */
67 struct mlx5_ifc_sw_tls_cntx_bits {
68 	struct mlx5_ifc_tls_static_params_bits param;
69 	struct mlx5_ifc_tls_progress_params_bits progress;
70 	struct {
71 		uint8_t key_data[8][0x20];
72 		uint8_t key_len[0x20];
73 	} key;
74 };
75 
76 CTASSERT(MLX5_ST_SZ_BYTES(sw_tls_cntx) <= sizeof(((struct mlx5e_tls_tag *)0)->crypto_params));
77 CTASSERT(MLX5_ST_SZ_BYTES(mkc) == sizeof(((struct mlx5e_tx_umr_wqe *)0)->mkc));
78 
79 static const char *mlx5e_tls_stats_desc[] = {
80 	MLX5E_TLS_STATS(MLX5E_STATS_DESC)
81 };
82 
83 static void mlx5e_tls_work(struct work_struct *);
84 
85 static int
86 mlx5e_tls_tag_import(void *arg, void **store, int cnt, int domain, int flags)
87 {
88 	struct mlx5e_tls_tag *ptag;
89 	int i;
90 
91 	for (i = 0; i != cnt; i++) {
92 		ptag = malloc_domainset(sizeof(*ptag), M_MLX5E_TLS,
93 		    mlx5_dev_domainset(arg), flags | M_ZERO);
94 		mtx_init(&ptag->mtx, "mlx5-tls-tag-mtx", NULL, MTX_DEF);
95 		INIT_WORK(&ptag->work, mlx5e_tls_work);
96 		store[i] = ptag;
97 	}
98 	return (i);
99 }
100 
101 static void
102 mlx5e_tls_tag_release(void *arg, void **store, int cnt)
103 {
104 	struct mlx5e_tls_tag *ptag;
105 	struct mlx5e_priv *priv;
106 	struct mlx5e_tls *ptls;
107 	int i;
108 
109 	for (i = 0; i != cnt; i++) {
110 		ptag = store[i];
111 		ptls = ptag->tls;
112 		priv = container_of(ptls, struct mlx5e_priv, tls);
113 
114 		flush_work(&ptag->work);
115 
116 		if (ptag->tisn != 0) {
117 			mlx5_tls_close_tis(priv->mdev, ptag->tisn);
118 			atomic_add_32(&ptls->num_resources, -1U);
119 		}
120 
121 		mtx_destroy(&ptag->mtx);
122 
123 		free(ptag, M_MLX5E_TLS);
124 	}
125 }
126 
127 static void
128 mlx5e_tls_tag_zfree(struct mlx5e_tls_tag *ptag)
129 {
130 	/* make sure any unhandled taskqueue events are ignored */
131 	ptag->state = MLX5E_TLS_ST_FREED;
132 
133 	/* reset some variables */
134 	ptag->dek_index = 0;
135 	ptag->dek_index_ok = 0;
136 
137 	/* avoid leaking keys */
138 	memset(ptag->crypto_params, 0, sizeof(ptag->crypto_params));
139 
140 	/* update number of TIS contexts */
141 	if (ptag->tisn == 0)
142 		atomic_add_32(&ptag->tls->num_resources, -1U);
143 
144 	/* return tag to UMA */
145 	uma_zfree(ptag->tls->zone, ptag);
146 }
147 
148 int
149 mlx5e_tls_init(struct mlx5e_priv *priv)
150 {
151 	struct mlx5e_tls *ptls = &priv->tls;
152 	struct sysctl_oid *node;
153 	uint32_t x;
154 
155 	if (MLX5_CAP_GEN(priv->mdev, tls_tx) == 0 ||
156 	    MLX5_CAP_GEN(priv->mdev, log_max_dek) == 0)
157 		return (0);
158 
159 	ptls->wq = create_singlethread_workqueue("mlx5-tls-wq");
160 	if (ptls->wq == NULL)
161 		return (ENOMEM);
162 
163 	sysctl_ctx_init(&ptls->ctx);
164 
165 	snprintf(ptls->zname, sizeof(ptls->zname),
166 	    "mlx5_%u_tls", device_get_unit(priv->mdev->pdev->dev.bsddev));
167 
168 	ptls->zone = uma_zcache_create(ptls->zname,
169 	     sizeof(struct mlx5e_tls_tag), NULL, NULL, NULL, NULL,
170 	     mlx5e_tls_tag_import, mlx5e_tls_tag_release, priv->mdev,
171 	     UMA_ZONE_UNMANAGED);
172 
173 	/* shared between RX and TX TLS */
174 	ptls->max_resources = 1U << (MLX5_CAP_GEN(priv->mdev, log_max_dek) - 1);
175 
176 	for (x = 0; x != MLX5E_TLS_STATS_NUM; x++)
177 		ptls->stats.arg[x] = counter_u64_alloc(M_WAITOK);
178 
179 	ptls->init = 1;
180 
181 	node = SYSCTL_ADD_NODE(&priv->sysctl_ctx,
182 	    SYSCTL_CHILDREN(priv->sysctl_ifnet), OID_AUTO,
183 	    "tls", CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "Hardware TLS offload");
184 	if (node == NULL)
185 		return (0);
186 
187 	mlx5e_create_counter_stats(&ptls->ctx,
188 	    SYSCTL_CHILDREN(node), "stats",
189 	    mlx5e_tls_stats_desc, MLX5E_TLS_STATS_NUM,
190 	    ptls->stats.arg);
191 
192 	return (0);
193 }
194 
195 void
196 mlx5e_tls_cleanup(struct mlx5e_priv *priv)
197 {
198 	struct mlx5e_tls *ptls = &priv->tls;
199 	uint32_t x;
200 
201 	if (ptls->init == 0)
202 		return;
203 
204 	ptls->init = 0;
205 	flush_workqueue(ptls->wq);
206 	sysctl_ctx_free(&ptls->ctx);
207 	uma_zdestroy(ptls->zone);
208 	destroy_workqueue(ptls->wq);
209 
210 	/* check if all resources are freed */
211 	MPASS(priv->tls.num_resources == 0);
212 
213 	for (x = 0; x != MLX5E_TLS_STATS_NUM; x++)
214 		counter_u64_free(ptls->stats.arg[x]);
215 }
216 
217 static void
218 mlx5e_tls_work(struct work_struct *work)
219 {
220 	struct mlx5e_tls_tag *ptag;
221 	struct mlx5e_priv *priv;
222 	int err;
223 
224 	ptag = container_of(work, struct mlx5e_tls_tag, work);
225 	priv = container_of(ptag->tls, struct mlx5e_priv, tls);
226 
227 	switch (ptag->state) {
228 	case MLX5E_TLS_ST_INIT:
229 		/* try to open TIS, if not present */
230 		if (ptag->tisn == 0) {
231 			err = mlx5_tls_open_tis(priv->mdev, 0, priv->tdn,
232 			    priv->pdn, &ptag->tisn);
233 			if (err) {
234 				MLX5E_TLS_STAT_INC(ptag, tx_error, 1);
235 				break;
236 			}
237 		}
238 		MLX5_SET(sw_tls_cntx, ptag->crypto_params, progress.pd, ptag->tisn);
239 
240 		/* try to allocate a DEK context ID */
241 		err = mlx5_encryption_key_create(priv->mdev, priv->pdn,
242 		    MLX5_ADDR_OF(sw_tls_cntx, ptag->crypto_params, key.key_data),
243 		    MLX5_GET(sw_tls_cntx, ptag->crypto_params, key.key_len),
244 		    &ptag->dek_index);
245 		if (err) {
246 			MLX5E_TLS_STAT_INC(ptag, tx_error, 1);
247 			break;
248 		}
249 
250 		MLX5_SET(sw_tls_cntx, ptag->crypto_params, param.dek_index, ptag->dek_index);
251 
252 		ptag->dek_index_ok = 1;
253 
254 		MLX5E_TLS_TAG_LOCK(ptag);
255 		if (ptag->state == MLX5E_TLS_ST_INIT)
256 			ptag->state = MLX5E_TLS_ST_SETUP;
257 		MLX5E_TLS_TAG_UNLOCK(ptag);
258 		break;
259 
260 	case MLX5E_TLS_ST_RELEASE:
261 		/* try to destroy DEK context by ID */
262 		if (ptag->dek_index_ok)
263 			err = mlx5_encryption_key_destroy(priv->mdev, ptag->dek_index);
264 
265 		/* free tag */
266 		mlx5e_tls_tag_zfree(ptag);
267 		break;
268 
269 	default:
270 		break;
271 	}
272 }
273 
274 static int
275 mlx5e_tls_set_params(void *ctx, const struct tls_session_params *en)
276 {
277 
278 	MLX5_SET(sw_tls_cntx, ctx, param.const_2, 2);
279 	if (en->tls_vminor == TLS_MINOR_VER_TWO)
280 		MLX5_SET(sw_tls_cntx, ctx, param.tls_version, 2); /* v1.2 */
281 	else
282 		MLX5_SET(sw_tls_cntx, ctx, param.tls_version, 3); /* v1.3 */
283 	MLX5_SET(sw_tls_cntx, ctx, param.const_1, 1);
284 	MLX5_SET(sw_tls_cntx, ctx, param.encryption_standard, 1); /* TLS */
285 
286 	/* copy the initial vector in place */
287 	switch (en->iv_len) {
288 	case MLX5_FLD_SZ_BYTES(sw_tls_cntx, param.gcm_iv):
289 	case MLX5_FLD_SZ_BYTES(sw_tls_cntx, param.gcm_iv) +
290 	     MLX5_FLD_SZ_BYTES(sw_tls_cntx, param.implicit_iv):
291 		memcpy(MLX5_ADDR_OF(sw_tls_cntx, ctx, param.gcm_iv),
292 		    en->iv, en->iv_len);
293 		break;
294 	default:
295 		return (EINVAL);
296 	}
297 
298 	if (en->cipher_key_len <= MLX5_FLD_SZ_BYTES(sw_tls_cntx, key.key_data)) {
299 		memcpy(MLX5_ADDR_OF(sw_tls_cntx, ctx, key.key_data),
300 		    en->cipher_key, en->cipher_key_len);
301 		MLX5_SET(sw_tls_cntx, ctx, key.key_len, en->cipher_key_len);
302 	} else {
303 		return (EINVAL);
304 	}
305 	return (0);
306 }
307 
308 /* Verify zero default */
309 CTASSERT(MLX5E_TLS_ST_INIT == 0);
310 
311 int
312 mlx5e_tls_snd_tag_alloc(struct ifnet *ifp,
313     union if_snd_tag_alloc_params *params,
314     struct m_snd_tag **ppmt)
315 {
316 	union if_snd_tag_alloc_params rl_params;
317 	const struct if_snd_tag_sw *snd_tag_sw;
318 	struct mlx5e_priv *priv;
319 	struct mlx5e_tls_tag *ptag;
320 	const struct tls_session_params *en;
321 	int error;
322 
323 	priv = ifp->if_softc;
324 
325 	if (priv->gone != 0 || priv->tls.init == 0)
326 		return (EOPNOTSUPP);
327 
328 	/* allocate new tag from zone, if any */
329 	ptag = uma_zalloc(priv->tls.zone, M_NOWAIT);
330 	if (ptag == NULL)
331 		return (ENOMEM);
332 
333 	/* sanity check default values */
334 	MPASS(ptag->dek_index == 0);
335 	MPASS(ptag->dek_index_ok == 0);
336 
337 	/* setup TLS tag */
338 	ptag->tls = &priv->tls;
339 
340 	/* check if there is no TIS context */
341 	if (ptag->tisn == 0) {
342 		uint32_t value;
343 
344 		value = atomic_fetchadd_32(&priv->tls.num_resources, 1U);
345 
346 		/* check resource limits */
347 		if (value >= priv->tls.max_resources) {
348 			error = ENOMEM;
349 			goto failure;
350 		}
351 	}
352 
353 	en = &params->tls.tls->params;
354 
355 	/* only TLS v1.2 and v1.3 is currently supported */
356 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE ||
357 	    (en->tls_vminor != TLS_MINOR_VER_TWO
358 #ifdef TLS_MINOR_VER_THREE
359 	     && en->tls_vminor != TLS_MINOR_VER_THREE
360 #endif
361 	     )) {
362 		error = EPROTONOSUPPORT;
363 		goto failure;
364 	}
365 
366 	switch (en->cipher_algorithm) {
367 	case CRYPTO_AES_NIST_GCM_16:
368 		switch (en->cipher_key_len) {
369 		case 128 / 8:
370 			if (en->tls_vminor == TLS_MINOR_VER_TWO) {
371 				if (MLX5_CAP_TLS(priv->mdev, tls_1_2_aes_gcm_128) == 0) {
372 					error = EPROTONOSUPPORT;
373 					goto failure;
374 				}
375 			} else {
376 				if (MLX5_CAP_TLS(priv->mdev, tls_1_3_aes_gcm_128) == 0) {
377 					error = EPROTONOSUPPORT;
378 					goto failure;
379 				}
380 			}
381 			error = mlx5e_tls_set_params(ptag->crypto_params, en);
382 			if (error)
383 				goto failure;
384 			break;
385 
386 		case 256 / 8:
387 			if (en->tls_vminor == TLS_MINOR_VER_TWO) {
388 				if (MLX5_CAP_TLS(priv->mdev, tls_1_2_aes_gcm_256) == 0) {
389 					error = EPROTONOSUPPORT;
390 					goto failure;
391 				}
392 			} else {
393 				if (MLX5_CAP_TLS(priv->mdev, tls_1_3_aes_gcm_256) == 0) {
394 					error = EPROTONOSUPPORT;
395 					goto failure;
396 				}
397 			}
398 			error = mlx5e_tls_set_params(ptag->crypto_params, en);
399 			if (error)
400 				goto failure;
401 			break;
402 
403 		default:
404 			error = EINVAL;
405 			goto failure;
406 		}
407 		break;
408 	default:
409 		error = EPROTONOSUPPORT;
410 		goto failure;
411 	}
412 
413 	memset(&rl_params, 0, sizeof(rl_params));
414 	rl_params.hdr = params->hdr;
415 	switch (params->hdr.type) {
416 #ifdef RATELIMIT
417 	case IF_SND_TAG_TYPE_TLS_RATE_LIMIT:
418 		rl_params.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT;
419 		rl_params.rate_limit.max_rate = params->tls_rate_limit.max_rate;
420 		snd_tag_sw = &mlx5e_tls_rl_snd_tag_sw;
421 		break;
422 #endif
423 	case IF_SND_TAG_TYPE_TLS:
424 		rl_params.hdr.type = IF_SND_TAG_TYPE_UNLIMITED;
425 		snd_tag_sw = &mlx5e_tls_snd_tag_sw;
426 		break;
427 	default:
428 		error = EOPNOTSUPP;
429 		goto failure;
430 	}
431 
432 	error = m_snd_tag_alloc(ifp, &rl_params, &ptag->rl_tag);
433 	if (error)
434 		goto failure;
435 
436 	/* store pointer to mbuf tag */
437 	MPASS(ptag->tag.refcount == 0);
438 	m_snd_tag_init(&ptag->tag, ifp, snd_tag_sw);
439 	*ppmt = &ptag->tag;
440 
441 	/* reset state */
442 	ptag->state = MLX5E_TLS_ST_INIT;
443 
444 	queue_work(priv->tls.wq, &ptag->work);
445 	flush_work(&ptag->work);
446 
447 	return (0);
448 
449 failure:
450 	mlx5e_tls_tag_zfree(ptag);
451 	return (error);
452 }
453 
454 #ifdef RATELIMIT
455 static int
456 mlx5e_tls_rl_snd_tag_modify(struct m_snd_tag *pmt, union if_snd_tag_modify_params *params)
457 {
458 	union if_snd_tag_modify_params rl_params;
459 	struct mlx5e_tls_tag *ptag =
460 	    container_of(pmt, struct mlx5e_tls_tag, tag);
461 	int error;
462 
463 	memset(&rl_params, 0, sizeof(rl_params));
464 	rl_params.rate_limit.max_rate = params->tls_rate_limit.max_rate;
465 	error = ptag->rl_tag->sw->snd_tag_modify(ptag->rl_tag, &rl_params);
466 	return (error);
467 }
468 #endif
469 
470 static int
471 mlx5e_tls_snd_tag_query(struct m_snd_tag *pmt, union if_snd_tag_query_params *params)
472 {
473 	struct mlx5e_tls_tag *ptag =
474 	    container_of(pmt, struct mlx5e_tls_tag, tag);
475 
476 	return (ptag->rl_tag->sw->snd_tag_query(ptag->rl_tag, params));
477 }
478 
479 static void
480 mlx5e_tls_snd_tag_free(struct m_snd_tag *pmt)
481 {
482 	struct mlx5e_tls_tag *ptag =
483 	    container_of(pmt, struct mlx5e_tls_tag, tag);
484 	struct mlx5e_priv *priv;
485 
486 	m_snd_tag_rele(ptag->rl_tag);
487 
488 	MLX5E_TLS_TAG_LOCK(ptag);
489 	ptag->state = MLX5E_TLS_ST_RELEASE;
490 	MLX5E_TLS_TAG_UNLOCK(ptag);
491 
492 	priv = ptag->tag.ifp->if_softc;
493 	queue_work(priv->tls.wq, &ptag->work);
494 }
495 
496 CTASSERT((MLX5_FLD_SZ_BYTES(sw_tls_cntx, param) % 16) == 0);
497 
498 static void
499 mlx5e_tls_send_static_parameters(struct mlx5e_sq *sq, struct mlx5e_tls_tag *ptag)
500 {
501 	const u32 ds_cnt = DIV_ROUND_UP(sizeof(struct mlx5e_tx_umr_wqe) +
502 	    MLX5_FLD_SZ_BYTES(sw_tls_cntx, param), MLX5_SEND_WQE_DS);
503 	struct mlx5e_tx_umr_wqe *wqe;
504 	u16 pi;
505 
506 	pi = sq->pc & sq->wq.sz_m1;
507 	wqe = mlx5_wq_cyc_get_wqe(&sq->wq, pi);
508 
509 	memset(wqe, 0, sizeof(*wqe));
510 
511 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((sq->pc << 8) |
512 	    MLX5_OPCODE_UMR | (MLX5_OPCODE_MOD_UMR_TLS_TIS_STATIC_PARAMS << 24));
513 	wqe->ctrl.qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
514 	wqe->ctrl.imm = cpu_to_be32(ptag->tisn << 8);
515 
516 	if (mlx5e_do_send_cqe(sq))
517 		wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE | MLX5_FENCE_MODE_INITIATOR_SMALL;
518 	else
519 		wqe->ctrl.fm_ce_se = MLX5_FENCE_MODE_INITIATOR_SMALL;
520 
521 	/* fill out UMR control segment */
522 	wqe->umr.flags = 0x80;	/* inline data */
523 	wqe->umr.bsf_octowords = cpu_to_be16(MLX5_FLD_SZ_BYTES(sw_tls_cntx, param) / 16);
524 
525 	/* copy in the static crypto parameters */
526 	memcpy(wqe + 1, MLX5_ADDR_OF(sw_tls_cntx, ptag->crypto_params, param),
527 	    MLX5_FLD_SZ_BYTES(sw_tls_cntx, param));
528 
529 	/* copy data for doorbell */
530 	memcpy(sq->doorbell.d32, &wqe->ctrl, sizeof(sq->doorbell.d32));
531 
532 	sq->mbuf[pi].mbuf = NULL;
533 	sq->mbuf[pi].num_bytes = 0;
534 	sq->mbuf[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
535 	sq->mbuf[pi].mst = m_snd_tag_ref(&ptag->tag);
536 
537 	sq->pc += sq->mbuf[pi].num_wqebbs;
538 }
539 
540 CTASSERT(MLX5_FLD_SZ_BYTES(sw_tls_cntx, progress) ==
541     sizeof(((struct mlx5e_tx_psv_wqe *)0)->psv));
542 
543 static void
544 mlx5e_tls_send_progress_parameters(struct mlx5e_sq *sq, struct mlx5e_tls_tag *ptag)
545 {
546 	const u32 ds_cnt = DIV_ROUND_UP(sizeof(struct mlx5e_tx_psv_wqe),
547 	    MLX5_SEND_WQE_DS);
548 	struct mlx5e_tx_psv_wqe *wqe;
549 	u16 pi;
550 
551 	pi = sq->pc & sq->wq.sz_m1;
552 	wqe = mlx5_wq_cyc_get_wqe(&sq->wq, pi);
553 
554 	memset(wqe, 0, sizeof(*wqe));
555 
556 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((sq->pc << 8) |
557 	    MLX5_OPCODE_SET_PSV | (MLX5_OPCODE_MOD_PSV_TLS_TIS_PROGRESS_PARAMS << 24));
558 	wqe->ctrl.qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
559 
560 	if (mlx5e_do_send_cqe(sq))
561 		wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
562 
563 	/* copy in the PSV control segment */
564 	memcpy(&wqe->psv, MLX5_ADDR_OF(sw_tls_cntx, ptag->crypto_params, progress),
565 	    sizeof(wqe->psv));
566 
567 	/* copy data for doorbell */
568 	memcpy(sq->doorbell.d32, &wqe->ctrl, sizeof(sq->doorbell.d32));
569 
570 	sq->mbuf[pi].mbuf = NULL;
571 	sq->mbuf[pi].num_bytes = 0;
572 	sq->mbuf[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
573 	sq->mbuf[pi].mst = m_snd_tag_ref(&ptag->tag);
574 
575 	sq->pc += sq->mbuf[pi].num_wqebbs;
576 }
577 
578 static void
579 mlx5e_tls_send_nop(struct mlx5e_sq *sq, struct mlx5e_tls_tag *ptag)
580 {
581 	const u32 ds_cnt = MLX5_SEND_WQEBB_NUM_DS;
582 	struct mlx5e_tx_wqe *wqe;
583 	u16 pi;
584 
585 	pi = sq->pc & sq->wq.sz_m1;
586 	wqe = mlx5_wq_cyc_get_wqe(&sq->wq, pi);
587 
588 	memset(&wqe->ctrl, 0, sizeof(wqe->ctrl));
589 
590 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_NOP);
591 	wqe->ctrl.qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
592 	if (mlx5e_do_send_cqe(sq))
593 		wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE | MLX5_FENCE_MODE_INITIATOR_SMALL;
594 	else
595 		wqe->ctrl.fm_ce_se = MLX5_FENCE_MODE_INITIATOR_SMALL;
596 
597 	/* Copy data for doorbell */
598 	memcpy(sq->doorbell.d32, &wqe->ctrl, sizeof(sq->doorbell.d32));
599 
600 	sq->mbuf[pi].mbuf = NULL;
601 	sq->mbuf[pi].num_bytes = 0;
602 	sq->mbuf[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
603 	sq->mbuf[pi].mst = m_snd_tag_ref(&ptag->tag);
604 
605 	sq->pc += sq->mbuf[pi].num_wqebbs;
606 }
607 
608 #define	SBTLS_MBUF_NO_DATA ((struct mbuf *)1)
609 
610 static struct mbuf *
611 sbtls_recover_record(struct mbuf *mb, int wait, uint32_t tcp_old, uint32_t *ptcp_seq, bool *pis_start)
612 {
613 	struct mbuf *mr, *top;
614 	uint32_t offset;
615 	uint32_t delta;
616 
617 	/* check format of incoming mbuf */
618 	if (mb->m_next == NULL ||
619 	    (mb->m_next->m_flags & (M_EXTPG | M_EXT)) != (M_EXTPG | M_EXT)) {
620 		top = NULL;
621 		goto done;
622 	}
623 
624 	/* get unmapped data offset */
625 	offset = mtod(mb->m_next, uintptr_t);
626 
627 	/* check if we don't need to re-transmit anything */
628 	if (offset == 0) {
629 		top = SBTLS_MBUF_NO_DATA;
630 		*pis_start = true;
631 		goto done;
632 	}
633 
634 	/* try to get a new  packet header */
635 	top = m_gethdr(wait, MT_DATA);
636 	if (top == NULL)
637 		goto done;
638 
639 	mr = m_get(wait, MT_DATA);
640 	if (mr == NULL) {
641 		m_free(top);
642 		top = NULL;
643 		goto done;
644 	}
645 
646 	top->m_next = mr;
647 
648 	mb_dupcl(mr, mb->m_next);
649 
650 	/* the beginning of the TLS record */
651 	mr->m_data = NULL;
652 
653 	/* setup packet header length */
654 	top->m_pkthdr.len = mr->m_len = offset;
655 	top->m_len = 0;
656 
657 	/* check for partial re-transmit */
658 	delta = *ptcp_seq - tcp_old;
659 
660 	if (delta < offset) {
661 		m_adj(top, offset - delta);
662 		offset = delta;
663 
664 		/* continue where we left off */
665 		*pis_start = false;
666 	} else {
667 		*pis_start = true;
668 	}
669 
670 	/*
671 	 * Rewind the TCP sequence number by the amount of data
672 	 * retransmitted:
673 	 */
674 	*ptcp_seq -= offset;
675 done:
676 	return (top);
677 }
678 
679 static int
680 mlx5e_sq_tls_populate(struct mbuf *mb, uint64_t *pseq)
681 {
682 
683 	for (; mb != NULL; mb = mb->m_next) {
684 		if (!(mb->m_flags & M_EXTPG))
685 			continue;
686 		*pseq = mb->m_epg_seqno;
687 		return (1);
688 	}
689 	return (0);
690 }
691 
692 int
693 mlx5e_sq_tls_xmit(struct mlx5e_sq *sq, struct mlx5e_xmit_args *parg, struct mbuf **ppmb)
694 {
695 	struct mlx5e_tls_tag *ptls_tag;
696 	struct m_snd_tag *ptag;
697 	const struct tcphdr *th;
698 	struct mbuf *mb = *ppmb;
699 	u64 rcd_sn;
700 	u32 header_size;
701 	u32 mb_seq;
702 
703 	if ((mb->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0)
704 		return (MLX5E_TLS_CONTINUE);
705 
706 	ptag = mb->m_pkthdr.snd_tag;
707 
708 	if (
709 #ifdef RATELIMIT
710 	    ptag->sw->type != IF_SND_TAG_TYPE_TLS_RATE_LIMIT &&
711 #endif
712 	    ptag->sw->type != IF_SND_TAG_TYPE_TLS)
713 		return (MLX5E_TLS_CONTINUE);
714 
715 	ptls_tag = container_of(ptag, struct mlx5e_tls_tag, tag);
716 
717 	header_size = mlx5e_get_full_header_size(mb, &th);
718 	if (unlikely(header_size == 0 || th == NULL))
719 		return (MLX5E_TLS_FAILURE);
720 
721 	/*
722 	 * Send non-TLS TCP packets AS-IS:
723 	 */
724 	if (header_size == mb->m_pkthdr.len ||
725 	    mlx5e_sq_tls_populate(mb, &rcd_sn) == 0) {
726 		parg->tisn = 0;
727 		parg->ihs = header_size;
728 		return (MLX5E_TLS_CONTINUE);
729 	}
730 
731 	mb_seq = ntohl(th->th_seq);
732 
733 	MLX5E_TLS_TAG_LOCK(ptls_tag);
734 	switch (ptls_tag->state) {
735 	case MLX5E_TLS_ST_INIT:
736 		MLX5E_TLS_TAG_UNLOCK(ptls_tag);
737 		return (MLX5E_TLS_FAILURE);
738 	case MLX5E_TLS_ST_SETUP:
739 		ptls_tag->state = MLX5E_TLS_ST_TXRDY;
740 		ptls_tag->expected_seq = ~mb_seq;	/* force setup */
741 	default:
742 		MLX5E_TLS_TAG_UNLOCK(ptls_tag);
743 		break;
744 	}
745 
746 	if (unlikely(ptls_tag->expected_seq != mb_seq)) {
747 		bool is_start;
748 		struct mbuf *r_mb;
749 		uint32_t tcp_seq = mb_seq;
750 
751 		r_mb = sbtls_recover_record(mb, M_NOWAIT, ptls_tag->expected_seq, &tcp_seq, &is_start);
752 		if (r_mb == NULL) {
753 			MLX5E_TLS_STAT_INC(ptls_tag, tx_error, 1);
754 			return (MLX5E_TLS_FAILURE);
755 		}
756 
757 		MLX5E_TLS_STAT_INC(ptls_tag, tx_packets_ooo, 1);
758 
759 		/* check if this is the first fragment of a TLS record */
760 		if (is_start) {
761 			/* setup TLS static parameters */
762 			MLX5_SET64(sw_tls_cntx, ptls_tag->crypto_params,
763 			    param.initial_record_number, rcd_sn);
764 
765 			/*
766 			 * NOTE: The sendqueue should have enough room to
767 			 * carry both the static and the progress parameters
768 			 * when we get here!
769 			 */
770 			mlx5e_tls_send_static_parameters(sq, ptls_tag);
771 			mlx5e_tls_send_progress_parameters(sq, ptls_tag);
772 
773 			if (r_mb == SBTLS_MBUF_NO_DATA) {
774 				mlx5e_tls_send_nop(sq, ptls_tag);
775 				ptls_tag->expected_seq = mb_seq;
776 				return (MLX5E_TLS_LOOP);
777 			}
778 		}
779 
780 		MLX5E_TLS_STAT_INC(ptls_tag, tx_bytes_ooo, r_mb->m_pkthdr.len);
781 
782 		/* setup transmit arguments */
783 		parg->tisn = ptls_tag->tisn;
784 		parg->mst = &ptls_tag->tag;
785 
786 		/* try to send DUMP data */
787 		if (mlx5e_sq_dump_xmit(sq, parg, &r_mb) != 0) {
788 			m_freem(r_mb);
789 			ptls_tag->expected_seq = tcp_seq;
790 			return (MLX5E_TLS_FAILURE);
791 		} else {
792 			ptls_tag->expected_seq = mb_seq;
793 			return (MLX5E_TLS_LOOP);
794 		}
795 	} else {
796 		MLX5E_TLS_STAT_INC(ptls_tag, tx_packets, 1);
797 		MLX5E_TLS_STAT_INC(ptls_tag, tx_bytes, mb->m_pkthdr.len);
798 	}
799 	ptls_tag->expected_seq += mb->m_pkthdr.len - header_size;
800 
801 	parg->tisn = ptls_tag->tisn;
802 	parg->ihs = header_size;
803 	parg->mst = &ptls_tag->tag;
804 	return (MLX5E_TLS_CONTINUE);
805 }
806 
807 #else
808 
809 int
810 mlx5e_tls_init(struct mlx5e_priv *priv)
811 {
812 
813 	return (0);
814 }
815 
816 void
817 mlx5e_tls_cleanup(struct mlx5e_priv *priv)
818 {
819 	/* NOP */
820 }
821 
822 #endif		/* KERN_TLS */
823