xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c (revision bdd1243df58e60e85101c09001d9812a789b6bc4)
1 /*-
2  * Copyright (c) 2015-2021 Mellanox Technologies. All rights reserved.
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_rss.h"
29 #include "opt_ratelimit.h"
30 
31 #include <dev/mlx5/mlx5_en/en.h>
32 #include <machine/in_cksum.h>
33 
34 static inline int
35 mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq,
36     struct mlx5e_rx_wqe *wqe, u16 ix)
37 {
38 	bus_dma_segment_t segs[MLX5E_MAX_BUSDMA_RX_SEGS];
39 	struct mbuf *mb;
40 	int nsegs;
41 	int err;
42 	struct mbuf *mb_head;
43 	int i;
44 
45 	if (rq->mbuf[ix].mbuf != NULL)
46 		return (0);
47 
48 	mb_head = mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR,
49 	    MLX5E_MAX_RX_BYTES);
50 	if (unlikely(mb == NULL))
51 		return (-ENOMEM);
52 
53 	mb->m_len = MLX5E_MAX_RX_BYTES;
54 	mb->m_pkthdr.len = MLX5E_MAX_RX_BYTES;
55 
56 	for (i = 1; i < rq->nsegs; i++) {
57 		if (mb_head->m_pkthdr.len >= rq->wqe_sz)
58 			break;
59 		mb = mb->m_next = m_getjcl(M_NOWAIT, MT_DATA, 0,
60 		    MLX5E_MAX_RX_BYTES);
61 		if (unlikely(mb == NULL)) {
62 			m_freem(mb_head);
63 			return (-ENOMEM);
64 		}
65 		mb->m_len = MLX5E_MAX_RX_BYTES;
66 		mb_head->m_pkthdr.len += MLX5E_MAX_RX_BYTES;
67 	}
68 	/* rewind to first mbuf in chain */
69 	mb = mb_head;
70 
71 	/* get IP header aligned */
72 	m_adj(mb, MLX5E_NET_IP_ALIGN);
73 
74 	err = -bus_dmamap_load_mbuf_sg(rq->dma_tag, rq->mbuf[ix].dma_map,
75 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
76 	if (err != 0)
77 		goto err_free_mbuf;
78 	if (unlikely(nsegs == 0)) {
79 		bus_dmamap_unload(rq->dma_tag, rq->mbuf[ix].dma_map);
80 		err = -ENOMEM;
81 		goto err_free_mbuf;
82 	}
83 	wqe->data[0].addr = cpu_to_be64(segs[0].ds_addr);
84 	wqe->data[0].byte_count = cpu_to_be32(segs[0].ds_len |
85 	    MLX5_HW_START_PADDING);
86 	for (i = 1; i != nsegs; i++) {
87 		wqe->data[i].addr = cpu_to_be64(segs[i].ds_addr);
88 		wqe->data[i].byte_count = cpu_to_be32(segs[i].ds_len);
89 	}
90 	for (; i < rq->nsegs; i++) {
91 		wqe->data[i].addr = 0;
92 		wqe->data[i].byte_count = 0;
93 	}
94 
95 	rq->mbuf[ix].mbuf = mb;
96 	rq->mbuf[ix].data = mb->m_data;
97 
98 	bus_dmamap_sync(rq->dma_tag, rq->mbuf[ix].dma_map,
99 	    BUS_DMASYNC_PREREAD);
100 	return (0);
101 
102 err_free_mbuf:
103 	m_freem(mb);
104 	return (err);
105 }
106 
107 static void
108 mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
109 {
110 	if (unlikely(rq->enabled == 0))
111 		return;
112 
113 	while (!mlx5_wq_ll_is_full(&rq->wq)) {
114 		struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, rq->wq.head);
115 
116 		if (unlikely(mlx5e_alloc_rx_wqe(rq, wqe, rq->wq.head))) {
117 			callout_reset_curcpu(&rq->watchdog, 1, (void *)&mlx5e_post_rx_wqes, rq);
118 			break;
119 		}
120 		mlx5_wq_ll_push(&rq->wq, be16_to_cpu(wqe->next.next_wqe_index));
121 	}
122 
123 	/* ensure wqes are visible to device before updating doorbell record */
124 	atomic_thread_fence_rel();
125 
126 	mlx5_wq_ll_update_db_record(&rq->wq);
127 }
128 
129 static void
130 mlx5e_lro_update_hdr(struct mbuf *mb, struct mlx5_cqe64 *cqe)
131 {
132 	/* TODO: consider vlans, ip options, ... */
133 	struct ether_header *eh;
134 	uint16_t eh_type;
135 	uint16_t tot_len;
136 	struct ip6_hdr *ip6 = NULL;
137 	struct ip *ip4 = NULL;
138 	struct tcphdr *th;
139 	uint32_t *ts_ptr;
140 	uint8_t l4_hdr_type;
141 	int tcp_ack;
142 
143 	eh = mtod(mb, struct ether_header *);
144 	eh_type = ntohs(eh->ether_type);
145 
146 	l4_hdr_type = get_cqe_l4_hdr_type(cqe);
147 	tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA == l4_hdr_type) ||
148 	    (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
149 
150 	/* TODO: consider vlan */
151 	tot_len = be32_to_cpu(cqe->byte_cnt) - ETHER_HDR_LEN;
152 
153 	switch (eh_type) {
154 	case ETHERTYPE_IP:
155 		ip4 = (struct ip *)(eh + 1);
156 		th = (struct tcphdr *)(ip4 + 1);
157 		break;
158 	case ETHERTYPE_IPV6:
159 		ip6 = (struct ip6_hdr *)(eh + 1);
160 		th = (struct tcphdr *)(ip6 + 1);
161 		break;
162 	default:
163 		return;
164 	}
165 
166 	ts_ptr = (uint32_t *)(th + 1);
167 
168 	if (get_cqe_lro_tcppsh(cqe))
169 		th->th_flags |= TH_PUSH;
170 
171 	if (tcp_ack) {
172 		th->th_flags |= TH_ACK;
173 		th->th_ack = cqe->lro_ack_seq_num;
174 		th->th_win = cqe->lro_tcp_win;
175 
176 		/*
177 		 * FreeBSD handles only 32bit aligned timestamp right after
178 		 * the TCP hdr
179 		 * +--------+--------+--------+--------+
180 		 * |   NOP  |  NOP   |  TSopt |   10   |
181 		 * +--------+--------+--------+--------+
182 		 * |          TSval   timestamp        |
183 		 * +--------+--------+--------+--------+
184 		 * |          TSecr   timestamp        |
185 		 * +--------+--------+--------+--------+
186 		 */
187 		if (get_cqe_lro_timestamp_valid(cqe) &&
188 		    (__predict_true(*ts_ptr) == ntohl(TCPOPT_NOP << 24 |
189 		    TCPOPT_NOP << 16 | TCPOPT_TIMESTAMP << 8 |
190 		    TCPOLEN_TIMESTAMP))) {
191 			/*
192 			 * cqe->timestamp is 64bit long.
193 			 * [0-31] - timestamp.
194 			 * [32-64] - timestamp echo replay.
195 			 */
196 			ts_ptr[1] = *(uint32_t *)&cqe->timestamp;
197 			ts_ptr[2] = *((uint32_t *)&cqe->timestamp + 1);
198 		}
199 	}
200 	if (ip4) {
201 		ip4->ip_ttl = cqe->lro_min_ttl;
202 		ip4->ip_len = cpu_to_be16(tot_len);
203 		ip4->ip_sum = 0;
204 		ip4->ip_sum = in_cksum(mb, ip4->ip_hl << 2);
205 	} else {
206 		ip6->ip6_hlim = cqe->lro_min_ttl;
207 		ip6->ip6_plen = cpu_to_be16(tot_len -
208 		    sizeof(struct ip6_hdr));
209 	}
210 	/* TODO: handle tcp checksum */
211 }
212 
213 static uint64_t
214 mlx5e_mbuf_tstmp(struct mlx5e_priv *priv, uint64_t hw_tstmp)
215 {
216 	struct mlx5e_clbr_point *cp, dcp;
217 	uint64_t tstmp_sec, tstmp_nsec;
218 	uint64_t hw_clocks;
219 	uint64_t rt_cur_to_prev, res_s, res_n, res_s_modulo, res;
220 	uint64_t hw_clk_div;
221 	u_int gen;
222 
223 	do {
224 		cp = &priv->clbr_points[priv->clbr_curr];
225 		gen = atomic_load_acq_int(&cp->clbr_gen);
226 		if (gen == 0)
227 			return (0);
228 		dcp = *cp;
229 		atomic_thread_fence_acq();
230 	} while (gen != dcp.clbr_gen);
231 	/*
232 	 * Our goal here is to have a result that is:
233 	 *
234 	 * (                             (cur_time - prev_time)   )
235 	 * ((hw_tstmp - hw_prev) *  ----------------------------- ) + prev_time
236 	 * (                             (hw_cur - hw_prev)       )
237 	 *
238 	 * With the constraints that we cannot use float and we
239 	 * don't want to overflow the uint64_t numbers we are using.
240 	 *
241 	 * The plan is to take the clocking value of the hw timestamps
242 	 * and split them into seconds and nanosecond equivalent portions.
243 	 * Then we operate on the two portions seperately making sure to
244 	 * bring back the carry over from the seconds when we divide.
245 	 *
246 	 * First up lets get the two divided into separate entities
247 	 * i.e. the seconds. We use the clock frequency for this.
248 	 * Note that priv->cclk was setup with the clock frequency
249 	 * in hz so we are all set to go.
250 	 */
251 	hw_clocks = hw_tstmp - dcp.clbr_hw_prev;
252 	tstmp_sec = hw_clocks / priv->cclk;
253 	tstmp_nsec = hw_clocks % priv->cclk;
254 	/* Now work with them separately */
255 	rt_cur_to_prev = (dcp.base_curr - dcp.base_prev);
256 	res_s = tstmp_sec * rt_cur_to_prev;
257 	res_n = tstmp_nsec * rt_cur_to_prev;
258 	/* Now lets get our divider */
259 	hw_clk_div = dcp.clbr_hw_curr - dcp.clbr_hw_prev;
260 	/* Make sure to save the remainder from the seconds divide */
261 	res_s_modulo = res_s % hw_clk_div;
262 	res_s /= hw_clk_div;
263 	/* scale the remainder to where it should be */
264 	res_s_modulo *= priv->cclk;
265 	/* Now add in the remainder */
266 	res_n += res_s_modulo;
267 	/* Now do the divide */
268 	res_n /= hw_clk_div;
269 	res_s *= priv->cclk;
270 	/* Recombine the two */
271 	res = res_s + res_n;
272 	/* And now add in the base time to get to the real timestamp */
273 	res += dcp.base_prev;
274 	return (res);
275 }
276 
277 static inline void
278 mlx5e_build_rx_mbuf(struct mlx5_cqe64 *cqe,
279     struct mlx5e_rq *rq, struct mbuf *mb,
280     u32 cqe_bcnt)
281 {
282 	if_t ifp = rq->ifp;
283 	struct mlx5e_channel *c;
284 	struct mbuf *mb_head;
285 	int lro_num_seg;	/* HW LRO session aggregated packets counter */
286 	uint64_t tstmp;
287 
288 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
289 	if (lro_num_seg > 1) {
290 		mlx5e_lro_update_hdr(mb, cqe);
291 		rq->stats.lro_packets++;
292 		rq->stats.lro_bytes += cqe_bcnt;
293 	}
294 
295 	mb->m_pkthdr.len = cqe_bcnt;
296 	for (mb_head = mb; mb != NULL; mb = mb->m_next) {
297 		if (mb->m_len > cqe_bcnt)
298 			mb->m_len = cqe_bcnt;
299 		cqe_bcnt -= mb->m_len;
300 		if (likely(cqe_bcnt == 0)) {
301 			if (likely(mb->m_next != NULL)) {
302 				/* trim off empty mbufs */
303 				m_freem(mb->m_next);
304 				mb->m_next = NULL;
305 			}
306 			break;
307 		}
308 	}
309 	/* rewind to first mbuf in chain */
310 	mb = mb_head;
311 
312 	/* check if a Toeplitz hash was computed */
313 	if (cqe->rss_hash_type != 0) {
314 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->rss_hash_result);
315 #ifdef RSS
316 		/* decode the RSS hash type */
317 		switch (cqe->rss_hash_type &
318 		    (CQE_RSS_DST_HTYPE_L4 | CQE_RSS_DST_HTYPE_IP)) {
319 		/* IPv4 */
320 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV4):
321 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV4);
322 			break;
323 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV4):
324 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV4);
325 			break;
326 		case CQE_RSS_DST_HTYPE_IPV4:
327 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV4);
328 			break;
329 		/* IPv6 */
330 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV6):
331 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV6);
332 			break;
333 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV6):
334 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV6);
335 			break;
336 		case CQE_RSS_DST_HTYPE_IPV6:
337 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV6);
338 			break;
339 		default:	/* Other */
340 			M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
341 			break;
342 		}
343 #else
344 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
345 #endif
346 #ifdef M_HASHTYPE_SETINNER
347 		if (cqe_is_tunneled(cqe))
348 			M_HASHTYPE_SETINNER(mb);
349 #endif
350 	} else {
351 		mb->m_pkthdr.flowid = rq->ix;
352 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
353 	}
354 	mb->m_pkthdr.rcvif = ifp;
355 	mb->m_pkthdr.leaf_rcvif = ifp;
356 
357 	if (cqe_is_tunneled(cqe)) {
358 		/*
359 		 * CQE can be tunneled only if TIR is configured to
360 		 * enable parsing of tunneled payload, so no need to
361 		 * check for capabilities.
362 		 */
363 		if (((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK)) ==
364 		    (CQE_L2_OK | CQE_L3_OK))) {
365 			mb->m_pkthdr.csum_flags |=
366 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
367 			    CSUM_IP_CHECKED | CSUM_IP_VALID |
368 			    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
369 			mb->m_pkthdr.csum_data = htons(0xffff);
370 
371 			if (likely((cqe->hds_ip_ext & CQE_L4_OK) == CQE_L4_OK)) {
372 				mb->m_pkthdr.csum_flags |=
373 				    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
374 			}
375 		} else {
376 			rq->stats.csum_none++;
377 		}
378 	} else if (likely((if_getcapenable(ifp) & (IFCAP_RXCSUM |
379 	    IFCAP_RXCSUM_IPV6)) != 0) &&
380 	    ((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK)) ==
381 	    (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK))) {
382 		mb->m_pkthdr.csum_flags =
383 		    CSUM_IP_CHECKED | CSUM_IP_VALID |
384 		    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
385 		mb->m_pkthdr.csum_data = htons(0xffff);
386 	} else {
387 		rq->stats.csum_none++;
388 	}
389 
390 	if (cqe_has_vlan(cqe)) {
391 		mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->vlan_info);
392 		mb->m_flags |= M_VLANTAG;
393 	}
394 
395 	c = container_of(rq, struct mlx5e_channel, rq);
396 	if (c->priv->clbr_done >= 2) {
397 		tstmp = mlx5e_mbuf_tstmp(c->priv, be64_to_cpu(cqe->timestamp));
398 		if ((tstmp & MLX5_CQE_TSTMP_PTP) != 0) {
399 			/*
400 			 * Timestamp was taken on the packet entrance,
401 			 * instead of the cqe generation.
402 			 */
403 			tstmp &= ~MLX5_CQE_TSTMP_PTP;
404 			mb->m_flags |= M_TSTMP_HPREC;
405 		}
406 		if (tstmp != 0) {
407 			mb->m_pkthdr.rcv_tstmp = tstmp;
408 			mb->m_flags |= M_TSTMP;
409 		}
410 	}
411 	switch (get_cqe_tls_offload(cqe)) {
412 	case CQE_TLS_OFFLOAD_DECRYPTED:
413 		/* set proper checksum flag for decrypted packets */
414 		mb->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
415 		rq->stats.decrypted_ok_packets++;
416 		break;
417 	case CQE_TLS_OFFLOAD_ERROR:
418 		rq->stats.decrypted_error_packets++;
419 		break;
420 	default:
421 		break;
422 	}
423 }
424 
425 static inline void
426 mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
427 {
428 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, (cc & cq->wq.sz_m1)),
429 	    sizeof(struct mlx5_cqe64));
430 }
431 
432 static inline void
433 mlx5e_write_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
434 {
435 	memcpy(mlx5_cqwq_get_wqe(&cq->wq, cc & cq->wq.sz_m1),
436 	    data, sizeof(struct mlx5_cqe64));
437 }
438 
439 static inline void
440 mlx5e_decompress_cqe(struct mlx5e_cq *cq, struct mlx5_cqe64 *title,
441     struct mlx5_mini_cqe8 *mini,
442     u16 wqe_counter, int i)
443 {
444 	/*
445 	 * NOTE: The fields which are not set here are copied from the
446 	 * initial and common title. See memcpy() in
447 	 * mlx5e_write_cqe_slot().
448 	 */
449 	title->byte_cnt = mini->byte_cnt;
450 	title->wqe_counter = cpu_to_be16((wqe_counter + i) & cq->wq.sz_m1);
451 	title->rss_hash_result = mini->rx_hash_result;
452 	/*
453 	 * Since we use MLX5_CQE_FORMAT_HASH when creating the RX CQ,
454 	 * the value of the checksum should be ignored.
455 	 */
456 	title->check_sum = 0;
457 	title->op_own = (title->op_own & 0xf0) |
458 	    (((cq->wq.cc + i) >> cq->wq.log_sz) & 1);
459 }
460 
461 #define MLX5E_MINI_ARRAY_SZ 8
462 /* Make sure structs are not packet differently */
463 CTASSERT(sizeof(struct mlx5_cqe64) ==
464     sizeof(struct mlx5_mini_cqe8) * MLX5E_MINI_ARRAY_SZ);
465 static void
466 mlx5e_decompress_cqes(struct mlx5e_cq *cq)
467 {
468 	struct mlx5_mini_cqe8 mini_array[MLX5E_MINI_ARRAY_SZ];
469 	struct mlx5_cqe64 title;
470 	u32 cqe_count;
471 	u32 i = 0;
472 	u16 title_wqe_counter;
473 
474 	mlx5e_read_cqe_slot(cq, cq->wq.cc, &title);
475 	title_wqe_counter = be16_to_cpu(title.wqe_counter);
476 	cqe_count = be32_to_cpu(title.byte_cnt);
477 
478 	/* Make sure we won't overflow */
479 	KASSERT(cqe_count <= cq->wq.sz_m1,
480 	    ("%s: cqe_count %u > cq->wq.sz_m1 %u", __func__,
481 	    cqe_count, cq->wq.sz_m1));
482 
483 	mlx5e_read_cqe_slot(cq, cq->wq.cc + 1, mini_array);
484 	while (true) {
485 		mlx5e_decompress_cqe(cq, &title,
486 		    &mini_array[i % MLX5E_MINI_ARRAY_SZ],
487 		    title_wqe_counter, i);
488 		mlx5e_write_cqe_slot(cq, cq->wq.cc + i, &title);
489 		i++;
490 
491 		if (i == cqe_count)
492 			break;
493 		if (i % MLX5E_MINI_ARRAY_SZ == 0)
494 			mlx5e_read_cqe_slot(cq, cq->wq.cc + i, mini_array);
495 	}
496 }
497 
498 static int
499 mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
500 {
501 	struct pfil_head *pfil;
502 	int i, rv;
503 
504 	CURVNET_SET_QUIET(if_getvnet(rq->ifp));
505 	pfil = rq->channel->priv->pfil;
506 	for (i = 0; i < budget; i++) {
507 		struct mlx5e_rx_wqe *wqe;
508 		struct mlx5_cqe64 *cqe;
509 		struct mbuf *mb;
510 		__be16 wqe_counter_be;
511 		u16 wqe_counter;
512 		u32 byte_cnt, seglen;
513 
514 		cqe = mlx5e_get_cqe(&rq->cq);
515 		if (!cqe)
516 			break;
517 
518 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED)
519 			mlx5e_decompress_cqes(&rq->cq);
520 
521 		mlx5_cqwq_pop(&rq->cq.wq);
522 
523 		wqe_counter_be = cqe->wqe_counter;
524 		wqe_counter = be16_to_cpu(wqe_counter_be);
525 		wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
526 		byte_cnt = be32_to_cpu(cqe->byte_cnt);
527 
528 		bus_dmamap_sync(rq->dma_tag,
529 		    rq->mbuf[wqe_counter].dma_map,
530 		    BUS_DMASYNC_POSTREAD);
531 
532 		if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
533 			mlx5e_dump_err_cqe(&rq->cq, rq->rqn, (const void *)cqe);
534 			rq->stats.wqe_err++;
535 			goto wq_ll_pop;
536 		}
537 		if (pfil != NULL && PFIL_HOOKED_IN(pfil)) {
538 			seglen = MIN(byte_cnt, MLX5E_MAX_RX_BYTES);
539 			rv = pfil_mem_in(rq->channel->priv->pfil,
540 			    rq->mbuf[wqe_counter].data, seglen, rq->ifp, &mb);
541 
542 			switch (rv) {
543 			case PFIL_DROPPED:
544 			case PFIL_CONSUMED:
545 				/*
546 				 * Filter dropped or consumed it. In
547 				 * either case, we can just recycle
548 				 * buffer; there is no more work to do.
549 				 */
550 				rq->stats.packets++;
551 				goto wq_ll_pop;
552 			case PFIL_REALLOCED:
553 				/*
554 				 * Filter copied it; recycle buffer
555 				 * and receive the new mbuf allocated
556 				 * by the Filter
557 				 */
558 				goto rx_common;
559 			default:
560 				/*
561 				 * The Filter said it was OK, so
562 				 * receive like normal.
563 				 */
564 				KASSERT(rv == PFIL_PASS,
565 					("Filter returned %d!\n", rv));
566 			}
567 		}
568 		if ((MHLEN - MLX5E_NET_IP_ALIGN) >= byte_cnt &&
569 		    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
570 			/* set maximum mbuf length */
571 			mb->m_len = MHLEN - MLX5E_NET_IP_ALIGN;
572 			/* get IP header aligned */
573 			mb->m_data += MLX5E_NET_IP_ALIGN;
574 
575 			bcopy(rq->mbuf[wqe_counter].data, mtod(mb, caddr_t),
576 			    byte_cnt);
577 		} else {
578 			mb = rq->mbuf[wqe_counter].mbuf;
579 			rq->mbuf[wqe_counter].mbuf = NULL;	/* safety clear */
580 
581 			bus_dmamap_unload(rq->dma_tag,
582 			    rq->mbuf[wqe_counter].dma_map);
583 		}
584 rx_common:
585 		mlx5e_build_rx_mbuf(cqe, rq, mb, byte_cnt);
586 		rq->stats.bytes += byte_cnt;
587 		rq->stats.packets++;
588 #ifdef NUMA
589 		mb->m_pkthdr.numa_domain = if_getnumadomain(rq->ifp);
590 #endif
591 
592 #if !defined(HAVE_TCP_LRO_RX)
593 		tcp_lro_queue_mbuf(&rq->lro, mb);
594 #else
595 		if (mb->m_pkthdr.csum_flags == 0 ||
596 		    (if_getcapenable(rq->ifp) & IFCAP_LRO) == 0 ||
597 		    rq->lro.lro_cnt == 0 ||
598 		    tcp_lro_rx(&rq->lro, mb, 0) != 0) {
599 			if_input(rq->ifp, mb);
600 		}
601 #endif
602 wq_ll_pop:
603 		mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
604 		    &wqe->next.next_wqe_index);
605 	}
606 	CURVNET_RESTORE();
607 
608 	mlx5_cqwq_update_db_record(&rq->cq.wq);
609 
610 	/* ensure cq space is freed before enabling more cqes */
611 	atomic_thread_fence_rel();
612 	return (i);
613 }
614 
615 void
616 mlx5e_rx_cq_comp(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe __unused)
617 {
618 	struct mlx5e_channel *c = container_of(mcq, struct mlx5e_channel, rq.cq.mcq);
619 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
620 	int i = 0;
621 
622 #ifdef HAVE_PER_CQ_EVENT_PACKET
623 #if (MHLEN < 15)
624 #error "MHLEN is too small"
625 #endif
626 	struct mbuf *mb = m_gethdr(M_NOWAIT, MT_DATA);
627 
628 	if (mb != NULL) {
629 		/* this code is used for debugging purpose only */
630 		mb->m_pkthdr.len = mb->m_len = 15;
631 		memset(mb->m_data, 255, 14);
632 		mb->m_data[14] = rq->ix;
633 		mb->m_pkthdr.rcvif = rq->ifp;
634 		mb->m_pkthdr.leaf_rcvif = rq->ifp;
635 		if_input(rq->ifp, mb);
636 	}
637 #endif
638 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
639 		mtx_lock(&c->sq[j].lock);
640 		c->sq[j].db_inhibit++;
641 		mtx_unlock(&c->sq[j].lock);
642 	}
643 
644 	mtx_lock(&c->iq.lock);
645 	c->iq.db_inhibit++;
646 	mtx_unlock(&c->iq.lock);
647 
648 	mtx_lock(&rq->mtx);
649 
650 	/*
651 	 * Polling the entire CQ without posting new WQEs results in
652 	 * lack of receive WQEs during heavy traffic scenarios.
653 	 */
654 	while (1) {
655 		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
656 		    MLX5E_RX_BUDGET_MAX)
657 			break;
658 		i += MLX5E_RX_BUDGET_MAX;
659 		if (i >= MLX5E_BUDGET_MAX)
660 			break;
661 		mlx5e_post_rx_wqes(rq);
662 	}
663 	mlx5e_post_rx_wqes(rq);
664 	/* check for dynamic interrupt moderation callback */
665 	if (rq->dim.mode != NET_DIM_CQ_PERIOD_MODE_DISABLED)
666 		net_dim(&rq->dim, rq->stats.packets, rq->stats.bytes);
667 	mlx5e_cq_arm(&rq->cq, MLX5_GET_DOORBELL_LOCK(&rq->channel->priv->doorbell_lock));
668 	tcp_lro_flush_all(&rq->lro);
669 	mtx_unlock(&rq->mtx);
670 
671 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
672 		mtx_lock(&c->sq[j].lock);
673 		c->sq[j].db_inhibit--;
674 		/* Update the doorbell record, if any. */
675 		mlx5e_tx_notify_hw(c->sq + j, true);
676 		mtx_unlock(&c->sq[j].lock);
677 	}
678 
679 	mtx_lock(&c->iq.lock);
680 	c->iq.db_inhibit--;
681 	mlx5e_iq_notify_hw(&c->iq);
682 	mtx_unlock(&c->iq.lock);
683 }
684