1 /*-
2  * Copyright (c) 2021-2022 NVIDIA corporation & affiliates.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include "opt_kern_tls.h"
29 #include "opt_rss.h"
30 #include "opt_ratelimit.h"
31 
32 #include <dev/mlx5/mlx5_en/en.h>
33 
34 #include <dev/mlx5/tls.h>
35 
36 #include <dev/mlx5/fs.h>
37 #include <dev/mlx5/mlx5_core/fs_tcp.h>
38 
39 #include <sys/ktls.h>
40 #include <opencrypto/cryptodev.h>
41 
42 #ifdef KERN_TLS
43 
44 static if_snd_tag_free_t mlx5e_tls_rx_snd_tag_free;
45 static if_snd_tag_modify_t mlx5e_tls_rx_snd_tag_modify;
46 
47 static const struct if_snd_tag_sw mlx5e_tls_rx_snd_tag_sw = {
48 	.snd_tag_modify = mlx5e_tls_rx_snd_tag_modify,
49 	.snd_tag_free = mlx5e_tls_rx_snd_tag_free,
50 	.type = IF_SND_TAG_TYPE_TLS_RX
51 };
52 
53 MALLOC_DEFINE(M_MLX5E_TLS_RX, "MLX5E_TLS_RX", "MLX5 ethernet HW TLS RX");
54 
55 /* software TLS RX context */
56 struct mlx5_ifc_sw_tls_rx_cntx_bits {
57 	struct mlx5_ifc_tls_static_params_bits param;
58 	struct mlx5_ifc_tls_progress_params_bits progress;
59 	struct {
60 		uint8_t key_data[8][0x20];
61 		uint8_t key_len[0x20];
62 	} key;
63 };
64 
65 CTASSERT(MLX5_ST_SZ_BYTES(sw_tls_rx_cntx) <= sizeof(((struct mlx5e_tls_rx_tag *)NULL)->crypto_params));
66 CTASSERT(MLX5_ST_SZ_BYTES(mkc) == sizeof(((struct mlx5e_tx_umr_wqe *)NULL)->mkc));
67 
68 static const char *mlx5e_tls_rx_stats_desc[] = {
69 	MLX5E_TLS_RX_STATS(MLX5E_STATS_DESC)
70 };
71 
72 static void mlx5e_tls_rx_work(struct work_struct *);
73 static bool mlx5e_tls_rx_snd_tag_find_tcp_sn_and_tls_rcd(struct mlx5e_tls_rx_tag *,
74     uint32_t, uint32_t *, uint64_t *);
75 
76 CTASSERT((MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param) % 16) == 0);
77 
78 static uint32_t
79 mlx5e_tls_rx_get_ch(struct mlx5e_priv *priv, uint32_t flowid, uint32_t flowtype)
80 {
81 	u32 ch;
82 #ifdef RSS
83 	u32 temp;
84 #endif
85 
86 	/* keep this code synced with mlx5e_select_queue() */
87 	ch = priv->params.num_channels;
88 #ifdef RSS
89 	if (rss_hash2bucket(flowid, flowtype, &temp) == 0)
90 		ch = temp % ch;
91 	else
92 #endif
93 		ch = (flowid % 128) % ch;
94 	return (ch);
95 }
96 
97 /*
98  * This function gets a pointer to an internal queue, IQ, based on the
99  * provided "flowid" and "flowtype". The IQ returned may in some rare
100  * cases not be activated or running, but this is all handled by the
101  * "mlx5e_iq_get_producer_index()" function.
102  *
103  * The idea behind this function is to spread the IQ traffic as much
104  * as possible and to avoid congestion on the same IQ when processing
105  * RX traffic.
106  */
107 static struct mlx5e_iq *
108 mlx5e_tls_rx_get_iq(struct mlx5e_priv *priv, uint32_t flowid, uint32_t flowtype)
109 {
110 	/*
111 	 * NOTE: The channels array is only freed at detach
112 	 * and it safe to return a pointer to the send tag
113 	 * inside the channels structure as long as we
114 	 * reference the priv.
115 	 */
116 	return (&priv->channel[mlx5e_tls_rx_get_ch(priv, flowid, flowtype)].iq);
117 }
118 
119 static void
120 mlx5e_tls_rx_send_static_parameters_cb(void *arg)
121 {
122 	struct mlx5e_tls_rx_tag *ptag;
123 
124 	ptag = (struct mlx5e_tls_rx_tag *)arg;
125 
126 	m_snd_tag_rele(&ptag->tag);
127 }
128 
129 /*
130  * This function sends the so-called TLS RX static parameters to the
131  * hardware. These parameters are temporarily stored in the
132  * "crypto_params" field of the TLS RX tag.  Most importantly this
133  * function sets the TCP sequence number (32-bit) and TLS record
134  * number (64-bit) where the decryption can resume.
135  *
136  * Zero is returned upon success. Else some error happend.
137  */
138 static int
139 mlx5e_tls_rx_send_static_parameters(struct mlx5e_iq *iq, struct mlx5e_tls_rx_tag *ptag)
140 {
141 	const u32 ds_cnt = DIV_ROUND_UP(sizeof(struct mlx5e_tx_umr_wqe) +
142 	    MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param), MLX5_SEND_WQE_DS);
143 	struct mlx5e_tx_umr_wqe *wqe;
144 	int pi;
145 
146 	mtx_lock(&iq->lock);
147 	pi = mlx5e_iq_get_producer_index(iq);
148 	if (pi < 0) {
149 		mtx_unlock(&iq->lock);
150 		return (-ENOMEM);
151 	}
152 	wqe = mlx5_wq_cyc_get_wqe(&iq->wq, pi);
153 
154 	memset(wqe, 0, sizeof(*wqe));
155 
156 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((iq->pc << 8) |
157 	    MLX5_OPCODE_UMR | (MLX5_OPCODE_MOD_UMR_TLS_TIR_STATIC_PARAMS << 24));
158 	wqe->ctrl.qpn_ds = cpu_to_be32((iq->sqn << 8) | ds_cnt);
159 	wqe->ctrl.imm = cpu_to_be32(ptag->tirn << 8);
160 	wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE | MLX5_FENCE_MODE_INITIATOR_SMALL;
161 
162 	/* fill out UMR control segment */
163 	wqe->umr.flags = 0x80;	/* inline data */
164 	wqe->umr.bsf_octowords =
165 	    cpu_to_be16(MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param) / 16);
166 
167 	/* copy in the static crypto parameters */
168 	memcpy(wqe + 1, MLX5_ADDR_OF(sw_tls_rx_cntx, ptag->crypto_params, param),
169 	    MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param));
170 
171 	/* copy data for doorbell */
172 	memcpy(iq->doorbell.d32, &wqe->ctrl, sizeof(iq->doorbell.d32));
173 
174 	iq->data[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
175 	iq->data[pi].callback = &mlx5e_tls_rx_send_static_parameters_cb;
176 	iq->data[pi].arg = ptag;
177 
178 	m_snd_tag_ref(&ptag->tag);
179 
180 	iq->pc += iq->data[pi].num_wqebbs;
181 
182 	mlx5e_iq_notify_hw(iq);
183 
184 	mtx_unlock(&iq->lock);
185 
186 	return (0);	/* success */
187 }
188 
189 static void
190 mlx5e_tls_rx_send_progress_parameters_cb(void *arg)
191 {
192 	struct mlx5e_tls_rx_tag *ptag;
193 
194 	ptag = (struct mlx5e_tls_rx_tag *)arg;
195 
196 	complete(&ptag->progress_complete);
197 }
198 
199 CTASSERT(MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, progress) ==
200     sizeof(((struct mlx5e_tx_psv_wqe *)NULL)->psv));
201 
202 /*
203  * This function resets the state of the TIR context to start
204  * searching for a valid TLS header and is used only when allocating
205  * the TLS RX tag.
206  *
207  * Zero is returned upon success, else some error happened.
208  */
209 static int
210 mlx5e_tls_rx_send_progress_parameters_sync(struct mlx5e_iq *iq,
211     struct mlx5e_tls_rx_tag *ptag)
212 {
213 	const u32 ds_cnt = DIV_ROUND_UP(sizeof(struct mlx5e_tx_psv_wqe),
214 	    MLX5_SEND_WQE_DS);
215 	struct mlx5e_priv *priv;
216 	struct mlx5e_tx_psv_wqe *wqe;
217 	int pi;
218 
219 	mtx_lock(&iq->lock);
220 	pi = mlx5e_iq_get_producer_index(iq);
221 	if (pi < 0) {
222 		mtx_unlock(&iq->lock);
223 		return (-ENOMEM);
224 	}
225 	wqe = mlx5_wq_cyc_get_wqe(&iq->wq, pi);
226 
227 	memset(wqe, 0, sizeof(*wqe));
228 
229 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((iq->pc << 8) |
230 	    MLX5_OPCODE_SET_PSV | (MLX5_OPCODE_MOD_PSV_TLS_TIR_PROGRESS_PARAMS << 24));
231 	wqe->ctrl.qpn_ds = cpu_to_be32((iq->sqn << 8) | ds_cnt);
232 	wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
233 
234 	/* copy in the PSV control segment */
235 	memcpy(&wqe->psv, MLX5_ADDR_OF(sw_tls_rx_cntx, ptag->crypto_params, progress),
236 	    sizeof(wqe->psv));
237 
238 	/* copy data for doorbell */
239 	memcpy(iq->doorbell.d32, &wqe->ctrl, sizeof(iq->doorbell.d32));
240 
241 	iq->data[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
242 	iq->data[pi].callback = &mlx5e_tls_rx_send_progress_parameters_cb;
243 	iq->data[pi].arg = ptag;
244 
245 	iq->pc += iq->data[pi].num_wqebbs;
246 
247 	init_completion(&ptag->progress_complete);
248 
249 	mlx5e_iq_notify_hw(iq);
250 
251 	mtx_unlock(&iq->lock);
252 
253 	while (1) {
254 		if (wait_for_completion_timeout(&ptag->progress_complete, hz) != 0)
255 			break;
256 		priv = container_of(iq, struct mlx5e_channel, iq)->priv;
257 		if (priv->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR ||
258 		    pci_channel_offline(priv->mdev->pdev) != 0)
259 			return (-EWOULDBLOCK);
260 	}
261 
262 	return (0);	/* success */
263 }
264 
265 CTASSERT(MLX5E_TLS_RX_PROGRESS_BUFFER_SIZE >= MLX5_ST_SZ_BYTES(tls_progress_params));
266 CTASSERT(MLX5E_TLS_RX_PROGRESS_BUFFER_SIZE <= PAGE_SIZE);
267 
268 struct mlx5e_get_tls_progress_params_wqe {
269 	struct mlx5_wqe_ctrl_seg ctrl;
270 	struct mlx5_seg_get_psv	 psv;
271 };
272 
273 static void
274 mlx5e_tls_rx_receive_progress_parameters_cb(void *arg)
275 {
276 	struct mlx5e_tls_rx_tag *ptag;
277 	struct mlx5e_iq *iq;
278 	uint32_t tcp_curr_sn_he;
279 	uint32_t tcp_next_sn_he;
280 	uint64_t tls_rcd_num;
281 	void *buffer;
282 
283 	ptag = (struct mlx5e_tls_rx_tag *)arg;
284 	buffer = mlx5e_tls_rx_get_progress_buffer(ptag);
285 
286 	MLX5E_TLS_RX_TAG_LOCK(ptag);
287 
288 	ptag->tcp_resync_pending = 0;
289 
290 	switch (MLX5_GET(tls_progress_params, buffer, record_tracker_state)) {
291 	case MLX5E_TLS_RX_PROGRESS_PARAMS_RECORD_TRACKER_STATE_TRACKING:
292 		break;
293 	default:
294 		goto done;
295 	}
296 
297 	switch (MLX5_GET(tls_progress_params, buffer, auth_state)) {
298 	case MLX5E_TLS_RX_PROGRESS_PARAMS_AUTH_STATE_NO_OFFLOAD:
299 		break;
300 	default:
301 		goto done;
302 	}
303 
304 	tcp_curr_sn_he = MLX5_GET(tls_progress_params, buffer, hw_resync_tcp_sn);
305 
306 	if (mlx5e_tls_rx_snd_tag_find_tcp_sn_and_tls_rcd(ptag, tcp_curr_sn_he,
307 	    &tcp_next_sn_he, &tls_rcd_num)) {
308 
309 		MLX5_SET64(sw_tls_rx_cntx, ptag->crypto_params,
310 		    param.initial_record_number, tls_rcd_num);
311 		MLX5_SET(sw_tls_rx_cntx, ptag->crypto_params,
312 		    param.resync_tcp_sn, tcp_curr_sn_he);
313 
314 		iq = mlx5e_tls_rx_get_iq(
315 		    container_of(ptag->tls_rx, struct mlx5e_priv, tls_rx),
316 		    ptag->flowid, ptag->flowtype);
317 
318 		if (mlx5e_tls_rx_send_static_parameters(iq, ptag) != 0)
319 			MLX5E_TLS_RX_STAT_INC(ptag, rx_error, 1);
320 	}
321 done:
322 	MLX5E_TLS_RX_TAG_UNLOCK(ptag);
323 
324 	m_snd_tag_rele(&ptag->tag);
325 }
326 
327 /*
328  * This function queries the hardware for the current state of the TIR
329  * in question. It is typically called when encrypted data is received
330  * to re-establish hardware decryption of received TLS data.
331  *
332  * Zero is returned upon success, else some error happened.
333  */
334 static int
335 mlx5e_tls_rx_receive_progress_parameters(struct mlx5e_iq *iq, struct mlx5e_tls_rx_tag *ptag)
336 {
337 	struct mlx5e_get_tls_progress_params_wqe *wqe;
338 	const u32 ds_cnt = DIV_ROUND_UP(sizeof(*wqe), MLX5_SEND_WQE_DS);
339 	u64 dma_address;
340 	int pi;
341 
342 	mtx_lock(&iq->lock);
343 	pi = mlx5e_iq_get_producer_index(iq);
344 	if (pi < 0) {
345 		mtx_unlock(&iq->lock);
346 		return (-ENOMEM);
347 	}
348 
349 	mlx5e_iq_load_memory_single(iq, pi,
350 	    mlx5e_tls_rx_get_progress_buffer(ptag),
351 	    MLX5E_TLS_RX_PROGRESS_BUFFER_SIZE,
352 	    &dma_address, BUS_DMASYNC_PREREAD);
353 
354 	wqe = mlx5_wq_cyc_get_wqe(&iq->wq, pi);
355 
356 	memset(wqe, 0, sizeof(*wqe));
357 
358 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((iq->pc << 8) |
359 	    MLX5_OPCODE_GET_PSV | (MLX5_OPCODE_MOD_PSV_TLS_TIR_PROGRESS_PARAMS << 24));
360 	wqe->ctrl.qpn_ds = cpu_to_be32((iq->sqn << 8) | ds_cnt);
361 	wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
362 	wqe->psv.num_psv = 1 << 4;
363 	wqe->psv.l_key = iq->mkey_be;
364 	wqe->psv.psv_index[0] = cpu_to_be32(ptag->tirn);
365 	wqe->psv.va = cpu_to_be64(dma_address);
366 
367 	/* copy data for doorbell */
368 	memcpy(iq->doorbell.d32, &wqe->ctrl, sizeof(iq->doorbell.d32));
369 
370 	iq->data[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
371 	iq->data[pi].callback = &mlx5e_tls_rx_receive_progress_parameters_cb;
372 	iq->data[pi].arg = ptag;
373 
374 	m_snd_tag_ref(&ptag->tag);
375 
376 	iq->pc += iq->data[pi].num_wqebbs;
377 
378 	mlx5e_iq_notify_hw(iq);
379 
380 	mtx_unlock(&iq->lock);
381 
382 	return (0);	/* success */
383 }
384 
385 /*
386  * This is the import function for TLS RX tags.
387  */
388 static int
389 mlx5e_tls_rx_tag_import(void *arg, void **store, int cnt, int domain, int flags)
390 {
391 	struct mlx5e_tls_rx_tag *ptag;
392 	int i;
393 
394 	for (i = 0; i != cnt; i++) {
395 		ptag = malloc_domainset(sizeof(*ptag), M_MLX5E_TLS_RX,
396 		    mlx5_dev_domainset(arg), flags | M_ZERO);
397 		mtx_init(&ptag->mtx, "mlx5-tls-rx-tag-mtx", NULL, MTX_DEF);
398 		INIT_WORK(&ptag->work, mlx5e_tls_rx_work);
399 		store[i] = ptag;
400 	}
401 	return (i);
402 }
403 
404 /*
405  * This is the release function for TLS RX tags.
406  */
407 static void
408 mlx5e_tls_rx_tag_release(void *arg, void **store, int cnt)
409 {
410 	struct mlx5e_tls_rx_tag *ptag;
411 	int i;
412 
413 	for (i = 0; i != cnt; i++) {
414 		ptag = store[i];
415 
416 		flush_work(&ptag->work);
417 		mtx_destroy(&ptag->mtx);
418 		free(ptag, M_MLX5E_TLS_RX);
419 	}
420 }
421 
422 /*
423  * This is a convenience function to free TLS RX tags. It resets some
424  * selected fields, updates the number of resources and returns the
425  * TLS RX tag to the UMA pool of free tags.
426  */
427 static void
428 mlx5e_tls_rx_tag_zfree(struct mlx5e_tls_rx_tag *ptag)
429 {
430 	/* make sure any unhandled taskqueue events are ignored */
431 	ptag->state = MLX5E_TLS_RX_ST_FREED;
432 
433 	/* reset some variables */
434 	ptag->dek_index = 0;
435 	ptag->dek_index_ok = 0;
436 	ptag->tirn = 0;
437 	ptag->flow_rule = NULL;
438 	ptag->tcp_resync_active = 0;
439 	ptag->tcp_resync_pending = 0;
440 
441 	/* avoid leaking keys */
442 	memset(ptag->crypto_params, 0, sizeof(ptag->crypto_params));
443 
444 	/* update number of resources in use */
445 	atomic_add_32(&ptag->tls_rx->num_resources, -1U);
446 
447 	/* return tag to UMA */
448 	uma_zfree(ptag->tls_rx->zone, ptag);
449 }
450 
451 /*
452  * This function enables TLS RX support for the given NIC, if all
453  * needed firmware capabilites are present.
454  */
455 int
456 mlx5e_tls_rx_init(struct mlx5e_priv *priv)
457 {
458 	struct mlx5e_tls_rx *ptls = &priv->tls_rx;
459 	struct sysctl_oid *node;
460 	uint32_t x;
461 
462 	if (MLX5_CAP_GEN(priv->mdev, tls_rx) == 0 ||
463 	    MLX5_CAP_GEN(priv->mdev, log_max_dek) == 0 ||
464 	    MLX5_CAP_FLOWTABLE_NIC_RX(priv->mdev, ft_field_support.outer_ip_version) == 0)
465 		return (0);
466 
467 	ptls->wq = create_singlethread_workqueue("mlx5-tls-rx-wq");
468 	if (ptls->wq == NULL)
469 		return (ENOMEM);
470 
471 	sysctl_ctx_init(&ptls->ctx);
472 
473 	snprintf(ptls->zname, sizeof(ptls->zname),
474 	    "mlx5_%u_tls_rx", device_get_unit(priv->mdev->pdev->dev.bsddev));
475 
476 	ptls->zone = uma_zcache_create(ptls->zname,
477 	    sizeof(struct mlx5e_tls_rx_tag), NULL, NULL, NULL, NULL,
478 	    mlx5e_tls_rx_tag_import, mlx5e_tls_rx_tag_release, priv->mdev,
479 	    UMA_ZONE_UNMANAGED);
480 
481 	/* shared between RX and TX TLS */
482 	ptls->max_resources = 1U << (MLX5_CAP_GEN(priv->mdev, log_max_dek) - 1);
483 
484 	for (x = 0; x != MLX5E_TLS_RX_STATS_NUM; x++)
485 		ptls->stats.arg[x] = counter_u64_alloc(M_WAITOK);
486 
487 	ptls->init = 1;
488 
489 	node = SYSCTL_ADD_NODE(&priv->sysctl_ctx,
490 	    SYSCTL_CHILDREN(priv->sysctl_ifnet), OID_AUTO,
491 	    "tls_rx", CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "Hardware TLS receive offload");
492 	if (node == NULL)
493 		return (0);
494 
495 	mlx5e_create_counter_stats(&ptls->ctx,
496 	    SYSCTL_CHILDREN(node), "stats",
497 	    mlx5e_tls_rx_stats_desc, MLX5E_TLS_RX_STATS_NUM,
498 	    ptls->stats.arg);
499 
500 	return (0);
501 }
502 
503 /*
504  * This function disables TLS RX support for the given NIC.
505  */
506 void
507 mlx5e_tls_rx_cleanup(struct mlx5e_priv *priv)
508 {
509 	struct mlx5e_tls_rx *ptls = &priv->tls_rx;
510 	uint32_t x;
511 
512 	if (ptls->init == 0)
513 		return;
514 
515 	ptls->init = 0;
516 	flush_workqueue(ptls->wq);
517 	sysctl_ctx_free(&ptls->ctx);
518 	uma_zdestroy(ptls->zone);
519 	destroy_workqueue(ptls->wq);
520 
521 	/* check if all resources are freed */
522 	MPASS(priv->tls_rx.num_resources == 0);
523 
524 	for (x = 0; x != MLX5E_TLS_RX_STATS_NUM; x++)
525 		counter_u64_free(ptls->stats.arg[x]);
526 }
527 
528 /*
529  * This function is used to serialize sleeping firmware operations
530  * needed in order to establish and destroy a TLS RX tag.
531  */
532 static void
533 mlx5e_tls_rx_work(struct work_struct *work)
534 {
535 	struct mlx5e_tls_rx_tag *ptag;
536 	struct mlx5e_priv *priv;
537 	int err;
538 
539 	ptag = container_of(work, struct mlx5e_tls_rx_tag, work);
540 	priv = container_of(ptag->tls_rx, struct mlx5e_priv, tls_rx);
541 
542 	switch (ptag->state) {
543 	case MLX5E_TLS_RX_ST_INIT:
544 		/* try to allocate new TIR context */
545 		err = mlx5_tls_open_tir(priv->mdev, priv->tdn,
546 		    priv->channel[mlx5e_tls_rx_get_ch(priv, ptag->flowid, ptag->flowtype)].rqtn,
547 		    &ptag->tirn);
548 		if (err) {
549 			MLX5E_TLS_RX_STAT_INC(ptag, rx_error, 1);
550 			break;
551 		}
552 		MLX5_SET(sw_tls_rx_cntx, ptag->crypto_params, progress.pd, ptag->tirn);
553 
554 		/* try to allocate a DEK context ID */
555 		err = mlx5_encryption_key_create(priv->mdev, priv->pdn,
556 		    MLX5_ADDR_OF(sw_tls_rx_cntx, ptag->crypto_params, key.key_data),
557 		    MLX5_GET(sw_tls_rx_cntx, ptag->crypto_params, key.key_len),
558 		    &ptag->dek_index);
559 		if (err) {
560 			MLX5E_TLS_RX_STAT_INC(ptag, rx_error, 1);
561 			break;
562 		}
563 
564 		MLX5_SET(sw_tls_rx_cntx, ptag->crypto_params, param.dek_index, ptag->dek_index);
565 
566 		ptag->dek_index_ok = 1;
567 
568 		MLX5E_TLS_RX_TAG_LOCK(ptag);
569 		if (ptag->state == MLX5E_TLS_RX_ST_INIT)
570 			ptag->state = MLX5E_TLS_RX_ST_SETUP;
571 		MLX5E_TLS_RX_TAG_UNLOCK(ptag);
572 		break;
573 
574 	case MLX5E_TLS_RX_ST_RELEASE:
575 		/* remove flow rule for incoming traffic, if any */
576 		if (ptag->flow_rule != NULL)
577 			mlx5e_accel_fs_del_inpcb(ptag->flow_rule);
578 
579 		/* try to destroy DEK context by ID */
580 		if (ptag->dek_index_ok)
581 			mlx5_encryption_key_destroy(priv->mdev, ptag->dek_index);
582 
583 		/* try to destroy TIR context by ID */
584 		if (ptag->tirn != 0)
585 			mlx5_tls_close_tir(priv->mdev, ptag->tirn);
586 
587 		/* free tag */
588 		mlx5e_tls_rx_tag_zfree(ptag);
589 		break;
590 
591 	default:
592 		break;
593 	}
594 }
595 
596 /*
597  * This function translates the crypto parameters into the format used
598  * by the firmware and hardware. Currently only AES-128 and AES-256 is
599  * supported for TLS v1.2 and TLS v1.3.
600  *
601  * Returns zero on success, else an error happened.
602  */
603 static int
604 mlx5e_tls_rx_set_params(void *ctx, struct inpcb *inp, const struct tls_session_params *en)
605 {
606 	uint32_t tcp_sn_he;
607 	uint64_t tls_sn_he;
608 
609 	MLX5_SET(sw_tls_rx_cntx, ctx, param.const_2, 2);
610 	if (en->tls_vminor == TLS_MINOR_VER_TWO)
611 		MLX5_SET(sw_tls_rx_cntx, ctx, param.tls_version, 2); /* v1.2 */
612 	else
613 		MLX5_SET(sw_tls_rx_cntx, ctx, param.tls_version, 3); /* v1.3 */
614 	MLX5_SET(sw_tls_rx_cntx, ctx, param.const_1, 1);
615 	MLX5_SET(sw_tls_rx_cntx, ctx, param.encryption_standard, 1); /* TLS */
616 
617 	/* copy the initial vector in place */
618 	switch (en->iv_len) {
619 	case MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param.gcm_iv):
620 	case MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param.gcm_iv) +
621 	     MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, param.implicit_iv):
622 		memcpy(MLX5_ADDR_OF(sw_tls_rx_cntx, ctx, param.gcm_iv),
623 		    en->iv, en->iv_len);
624 		break;
625 	default:
626 		return (EINVAL);
627 	}
628 
629 	if (en->cipher_key_len <= MLX5_FLD_SZ_BYTES(sw_tls_rx_cntx, key.key_data)) {
630 		memcpy(MLX5_ADDR_OF(sw_tls_rx_cntx, ctx, key.key_data),
631 		    en->cipher_key, en->cipher_key_len);
632 		MLX5_SET(sw_tls_rx_cntx, ctx, key.key_len, en->cipher_key_len);
633 	} else {
634 		return (EINVAL);
635 	}
636 
637 	if (__predict_false(inp == NULL ||
638 	    ktls_get_rx_sequence(inp, &tcp_sn_he, &tls_sn_he) != 0))
639 		return (EINVAL);
640 
641 	MLX5_SET64(sw_tls_rx_cntx, ctx, param.initial_record_number, tls_sn_he);
642 	MLX5_SET(sw_tls_rx_cntx, ctx, param.resync_tcp_sn, tcp_sn_he);
643 
644 	return (0);
645 }
646 
647 /* Verify zero default */
648 CTASSERT(MLX5E_TLS_RX_ST_INIT == 0);
649 
650 /*
651  * This functino is responsible for allocating a TLS RX tag. It is a
652  * callback function invoked by the network stack.
653  *
654  * Returns zero on success else an error happened.
655  */
656 int
657 mlx5e_tls_rx_snd_tag_alloc(struct ifnet *ifp,
658     union if_snd_tag_alloc_params *params,
659     struct m_snd_tag **ppmt)
660 {
661 	struct mlx5e_iq *iq;
662 	struct mlx5e_priv *priv;
663 	struct mlx5e_tls_rx_tag *ptag;
664 	struct mlx5_flow_rule *flow_rule;
665 	const struct tls_session_params *en;
666 	uint32_t value;
667 	int error;
668 
669 	priv = ifp->if_softc;
670 
671 	if (unlikely(priv->gone != 0 || priv->tls_rx.init == 0 ||
672 	    params->hdr.flowtype == M_HASHTYPE_NONE))
673 		return (EOPNOTSUPP);
674 
675 	/* allocate new tag from zone, if any */
676 	ptag = uma_zalloc(priv->tls_rx.zone, M_NOWAIT);
677 	if (ptag == NULL)
678 		return (ENOMEM);
679 
680 	/* sanity check default values */
681 	MPASS(ptag->dek_index == 0);
682 	MPASS(ptag->dek_index_ok == 0);
683 
684 	/* setup TLS RX tag */
685 	ptag->tls_rx = &priv->tls_rx;
686 	ptag->flowtype = params->hdr.flowtype;
687 	ptag->flowid = params->hdr.flowid;
688 
689 	value = atomic_fetchadd_32(&priv->tls_rx.num_resources, 1U);
690 
691 	/* check resource limits */
692 	if (value >= priv->tls_rx.max_resources) {
693 		error = ENOMEM;
694 		goto failure;
695 	}
696 
697 	en = &params->tls_rx.tls->params;
698 
699 	/* only TLS v1.2 and v1.3 is currently supported */
700 	if (en->tls_vmajor != TLS_MAJOR_VER_ONE ||
701 	    (en->tls_vminor != TLS_MINOR_VER_TWO
702 #ifdef TLS_MINOR_VER_THREE
703 	     && en->tls_vminor != TLS_MINOR_VER_THREE
704 #endif
705 	     )) {
706 		error = EPROTONOSUPPORT;
707 		goto failure;
708 	}
709 
710 	switch (en->cipher_algorithm) {
711 	case CRYPTO_AES_NIST_GCM_16:
712 		switch (en->cipher_key_len) {
713 		case 128 / 8:
714 			if (en->tls_vminor == TLS_MINOR_VER_TWO) {
715 				if (MLX5_CAP_TLS(priv->mdev, tls_1_2_aes_gcm_128) == 0) {
716 					error = EPROTONOSUPPORT;
717 					goto failure;
718 				}
719 			} else {
720 				if (MLX5_CAP_TLS(priv->mdev, tls_1_3_aes_gcm_128) == 0) {
721 					error = EPROTONOSUPPORT;
722 					goto failure;
723 				}
724 			}
725 			error = mlx5e_tls_rx_set_params(
726 			    ptag->crypto_params, params->tls_rx.inp, en);
727 			if (error)
728 				goto failure;
729 			break;
730 
731 		case 256 / 8:
732 			if (en->tls_vminor == TLS_MINOR_VER_TWO) {
733 				if (MLX5_CAP_TLS(priv->mdev, tls_1_2_aes_gcm_256) == 0) {
734 					error = EPROTONOSUPPORT;
735 					goto failure;
736 				}
737 			} else {
738 				if (MLX5_CAP_TLS(priv->mdev, tls_1_3_aes_gcm_256) == 0) {
739 					error = EPROTONOSUPPORT;
740 					goto failure;
741 				}
742 			}
743 			error = mlx5e_tls_rx_set_params(
744 			    ptag->crypto_params, params->tls_rx.inp, en);
745 			if (error)
746 				goto failure;
747 			break;
748 
749 		default:
750 			error = EINVAL;
751 			goto failure;
752 		}
753 		break;
754 	default:
755 		error = EPROTONOSUPPORT;
756 		goto failure;
757 	}
758 
759 	/* store pointer to mbuf tag */
760 	MPASS(ptag->tag.refcount == 0);
761 	m_snd_tag_init(&ptag->tag, ifp, &mlx5e_tls_rx_snd_tag_sw);
762 	*ppmt = &ptag->tag;
763 
764 	/* reset state */
765 	ptag->state = MLX5E_TLS_RX_ST_INIT;
766 
767 	queue_work(priv->tls_rx.wq, &ptag->work);
768 	flush_work(&ptag->work);
769 
770 	/* check that worker task completed successfully */
771 	MLX5E_TLS_RX_TAG_LOCK(ptag);
772 	if (ptag->state == MLX5E_TLS_RX_ST_SETUP) {
773 		ptag->state = MLX5E_TLS_RX_ST_READY;
774 		error = 0;
775 	} else {
776 		error = ENOMEM;
777 	}
778 	MLX5E_TLS_RX_TAG_UNLOCK(ptag);
779 
780 	if (unlikely(error))
781 		goto cleanup;
782 
783 	iq = mlx5e_tls_rx_get_iq(priv, ptag->flowid, ptag->flowtype);
784 
785 	/* establish connection between DEK and TIR */
786 	if (mlx5e_tls_rx_send_static_parameters(iq, ptag) != 0) {
787 		MLX5E_TLS_RX_STAT_INC(ptag, rx_error, 1);
788 		error = ENOMEM;
789 		goto cleanup;
790 	}
791 
792 	MLX5_SET(sw_tls_rx_cntx, ptag->crypto_params, progress.auth_state,
793 	    MLX5E_TLS_RX_PROGRESS_PARAMS_AUTH_STATE_NO_OFFLOAD);
794 	MLX5_SET(sw_tls_rx_cntx, ptag->crypto_params, progress.record_tracker_state,
795 	    MLX5E_TLS_RX_PROGRESS_PARAMS_RECORD_TRACKER_STATE_START);
796 
797 	/* reset state to all zeros */
798 	if (mlx5e_tls_rx_send_progress_parameters_sync(iq, ptag) != 0) {
799 		MLX5E_TLS_RX_STAT_INC(ptag, rx_error, 1);
800 		error = ENOMEM;
801 		goto cleanup;
802 	}
803 
804 	if (ifp->if_pcp != IFNET_PCP_NONE || params->tls_rx.vlan_id != 0) {
805 		/* create flow rule for TLS RX traffic (tagged) */
806 		flow_rule = mlx5e_accel_fs_add_inpcb(priv, params->tls_rx.inp,
807 		    ptag->tirn, MLX5_FS_DEFAULT_FLOW_TAG, params->tls_rx.vlan_id);
808 	} else {
809 		/* create flow rule for TLS RX traffic (untagged) */
810 		flow_rule = mlx5e_accel_fs_add_inpcb(priv, params->tls_rx.inp,
811 		    ptag->tirn, MLX5_FS_DEFAULT_FLOW_TAG, MLX5E_ACCEL_FS_ADD_INPCB_NO_VLAN);
812 	}
813 
814 	if (IS_ERR_OR_NULL(flow_rule)) {
815 		MLX5E_TLS_RX_STAT_INC(ptag, rx_error, 1);
816 		error = ENOMEM;
817 		goto cleanup;
818 	}
819 
820 	ptag->flow_rule = flow_rule;
821 
822 	return (0);
823 
824 cleanup:
825 	m_snd_tag_rele(&ptag->tag);
826 	return (error);
827 
828 failure:
829 	mlx5e_tls_rx_tag_zfree(ptag);
830 	return (error);
831 }
832 
833 
834 /*
835  * This function adds the TCP sequence number and TLS record number in
836  * host endian format to a small database. When TLS records have the
837  * same length, they are simply accumulated by counting instead of
838  * separated entries in the TLS database. The dimension of the
839  * database is such that it cannot store more than 1GByte of
840  * continuous TCP data to avoid issues with TCP sequence number wrap
841  * around. A record length of zero bytes has special meaning and means
842  * that resync completed and all data in the database can be
843  * discarded. This function is called after the TCP stack has
844  * re-assembled all TCP fragments due to out of order packet reception
845  * and all TCP sequence numbers should be sequential.
846  *
847  * This function returns true if a so-called TLS RX resync operation
848  * is in progress. Else no such operation is in progress.
849  */
850 static bool
851 mlx5e_tls_rx_snd_tag_add_tcp_sequence(struct mlx5e_tls_rx_tag *ptag,
852     uint32_t tcp_sn_he, uint32_t len, uint64_t tls_rcd)
853 {
854 	uint16_t i, j, n;
855 
856 	if (ptag->tcp_resync_active == 0 ||
857 	    ptag->tcp_resync_next != tcp_sn_he ||
858 	    len == 0) {
859 		/* start over again or terminate */
860 		ptag->tcp_resync_active = (len != 0);
861 		ptag->tcp_resync_len[0] = len;
862 		ptag->tcp_resync_num[0] = 1;
863 		ptag->tcp_resync_pc = (len != 0);
864 		ptag->tcp_resync_cc = 0;
865 		ptag->tcp_resync_start = tcp_sn_he;
866 		ptag->rcd_resync_start = tls_rcd;
867 	} else {
868 		i = (ptag->tcp_resync_pc - 1) & (MLX5E_TLS_RX_RESYNC_MAX - 1);
869 		n = ptag->tcp_resync_pc - ptag->tcp_resync_cc;
870 
871 		/* check if same length like last time */
872 		if (ptag->tcp_resync_len[i] == len &&
873 		    ptag->tcp_resync_num[i] != MLX5E_TLS_RX_NUM_MAX) {
874 			/* use existing entry */
875 			ptag->tcp_resync_num[i]++;
876 		} else if (n == MLX5E_TLS_RX_RESYNC_MAX) {
877 			j = ptag->tcp_resync_cc++ & (MLX5E_TLS_RX_RESYNC_MAX - 1);
878 			/* adjust starting TCP sequence number */
879 			ptag->rcd_resync_start += ptag->tcp_resync_num[j];
880 			ptag->tcp_resync_start += ptag->tcp_resync_len[j] * ptag->tcp_resync_num[j];
881 			i = ptag->tcp_resync_pc++ & (MLX5E_TLS_RX_RESYNC_MAX - 1);
882 			/* store new entry */
883 			ptag->tcp_resync_len[i] = len;
884 			ptag->tcp_resync_num[i] = 1;
885 		} else {
886 			i = ptag->tcp_resync_pc++ & (MLX5E_TLS_RX_RESYNC_MAX - 1);
887 			/* add new entry */
888 			ptag->tcp_resync_len[i] = len;
889 			ptag->tcp_resync_num[i] = 1;
890 		}
891 	}
892 
893 	/* store next TCP SN in host endian format */
894 	ptag->tcp_resync_next = tcp_sn_he + len;
895 
896 	return (ptag->tcp_resync_active);
897 }
898 
899 /*
900  * This function checks if the given TCP sequence number points to the
901  * beginning of a valid TLS header.
902  *
903  * Returns true if a match is found. Else false.
904  */
905 static bool
906 mlx5e_tls_rx_snd_tag_find_tcp_sn_and_tls_rcd(struct mlx5e_tls_rx_tag *ptag,
907     uint32_t tcp_sn_he, uint32_t *p_next_tcp_sn_he, uint64_t *p_tls_rcd)
908 {
909 	uint16_t i, j;
910 	uint32_t off = 0;
911 	uint32_t rcd = 0;
912 	uint32_t delta;
913 	uint32_t leap;
914 
915 	for (i = ptag->tcp_resync_cc; i != ptag->tcp_resync_pc; i++) {
916 		delta = tcp_sn_he - off - ptag->tcp_resync_start;
917 
918 		/* check if subtraction went negative */
919 		if ((int32_t)delta < 0)
920 			break;
921 
922 		j = i & (MLX5E_TLS_RX_RESYNC_MAX - 1);
923 		leap = ptag->tcp_resync_len[j] * ptag->tcp_resync_num[j];
924 		if (delta < leap) {
925 			if ((delta % ptag->tcp_resync_len[j]) == 0) {
926 				*p_next_tcp_sn_he = tcp_sn_he +
927 				    ptag->tcp_resync_len[j];
928 				*p_tls_rcd = ptag->rcd_resync_start +
929 				    (uint64_t)rcd +
930 				    (uint64_t)(delta / ptag->tcp_resync_len[j]);
931 				return (true);		/* success */
932 			}
933 			break;	/* invalid offset */
934 		}
935 		rcd += ptag->tcp_resync_num[j];
936 		off += leap;
937 	}
938 	return (false);	/* not found */
939 }
940 
941 /*
942  * This is a callback function from the network stack to keep track of
943  * TLS RX TCP sequence numbers.
944  *
945  * Returns zero on success else an error happened.
946  */
947 static int
948 mlx5e_tls_rx_snd_tag_modify(struct m_snd_tag *pmt, union if_snd_tag_modify_params *params)
949 {
950 	struct mlx5e_tls_rx_tag *ptag;
951 	struct mlx5e_priv *priv;
952 	struct mlx5e_iq *iq;
953 	int err;
954 
955 	ptag = container_of(pmt, struct mlx5e_tls_rx_tag, tag);
956 	priv = container_of(ptag->tls_rx, struct mlx5e_priv, tls_rx);
957 
958 	if (unlikely(priv->gone != 0))
959 		return (ENXIO);
960 
961 	iq = mlx5e_tls_rx_get_iq(priv, ptag->flowid, ptag->flowtype);
962 
963 	MLX5E_TLS_RX_TAG_LOCK(ptag);
964 
965 	if (mlx5e_tls_rx_snd_tag_add_tcp_sequence(ptag,
966 	    params->tls_rx.tls_hdr_tcp_sn,
967 	    params->tls_rx.tls_rec_length,
968 	    params->tls_rx.tls_seq_number) &&
969 	    ptag->tcp_resync_pending == 0) {
970 		err = mlx5e_tls_rx_receive_progress_parameters(iq, ptag);
971 		if (err != 0) {
972 			MLX5E_TLS_RX_STAT_INC(ptag, rx_resync_err, 1);
973 		} else {
974 			ptag->tcp_resync_pending = 1;
975 			MLX5E_TLS_RX_STAT_INC(ptag, rx_resync_ok, 1);
976 		}
977 	} else {
978 		err = 0;
979 	}
980 	MLX5E_TLS_RX_TAG_UNLOCK(ptag);
981 
982 	return (-err);
983 }
984 
985 /*
986  * This function frees a TLS RX tag in a non-blocking way.
987  */
988 static void
989 mlx5e_tls_rx_snd_tag_free(struct m_snd_tag *pmt)
990 {
991 	struct mlx5e_tls_rx_tag *ptag =
992 	    container_of(pmt, struct mlx5e_tls_rx_tag, tag);
993 	struct mlx5e_priv *priv;
994 
995 	MLX5E_TLS_RX_TAG_LOCK(ptag);
996 	ptag->state = MLX5E_TLS_RX_ST_RELEASE;
997 	MLX5E_TLS_RX_TAG_UNLOCK(ptag);
998 
999 	priv = ptag->tag.ifp->if_softc;
1000 	queue_work(priv->tls_rx.wq, &ptag->work);
1001 }
1002 
1003 #else
1004 
1005 int
1006 mlx5e_tls_rx_init(struct mlx5e_priv *priv)
1007 {
1008 
1009 	return (0);
1010 }
1011 
1012 void
1013 mlx5e_tls_rx_cleanup(struct mlx5e_priv *priv)
1014 {
1015 	/* NOP */
1016 }
1017 
1018 #endif
1019