xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c (revision e0c4386e7e71d93b0edc0c8fa156263fc4a8b0b6)
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 
26 #include "opt_rss.h"
27 #include "opt_ratelimit.h"
28 
29 #include <dev/mlx5/mlx5_en/en.h>
30 #include <machine/in_cksum.h>
31 
32 static inline int
33 mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq,
34     struct mlx5e_rx_wqe *wqe, u16 ix)
35 {
36 	bus_dma_segment_t segs[MLX5E_MAX_BUSDMA_RX_SEGS];
37 	struct mbuf *mb;
38 	int nsegs;
39 	int err;
40 	struct mbuf *mb_head;
41 	int i;
42 
43 	if (rq->mbuf[ix].mbuf != NULL)
44 		return (0);
45 
46 	mb_head = mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR,
47 	    MLX5E_MAX_RX_BYTES);
48 	if (unlikely(mb == NULL))
49 		return (-ENOMEM);
50 
51 	mb->m_len = MLX5E_MAX_RX_BYTES;
52 	mb->m_pkthdr.len = MLX5E_MAX_RX_BYTES;
53 
54 	for (i = 1; i < rq->nsegs; i++) {
55 		if (mb_head->m_pkthdr.len >= rq->wqe_sz)
56 			break;
57 		mb = mb->m_next = m_getjcl(M_NOWAIT, MT_DATA, 0,
58 		    MLX5E_MAX_RX_BYTES);
59 		if (unlikely(mb == NULL)) {
60 			m_freem(mb_head);
61 			return (-ENOMEM);
62 		}
63 		mb->m_len = MLX5E_MAX_RX_BYTES;
64 		mb_head->m_pkthdr.len += MLX5E_MAX_RX_BYTES;
65 	}
66 	/* rewind to first mbuf in chain */
67 	mb = mb_head;
68 
69 	/* get IP header aligned */
70 	m_adj(mb, MLX5E_NET_IP_ALIGN);
71 
72 	err = -bus_dmamap_load_mbuf_sg(rq->dma_tag, rq->mbuf[ix].dma_map,
73 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
74 	if (err != 0)
75 		goto err_free_mbuf;
76 	if (unlikely(nsegs == 0)) {
77 		bus_dmamap_unload(rq->dma_tag, rq->mbuf[ix].dma_map);
78 		err = -ENOMEM;
79 		goto err_free_mbuf;
80 	}
81 	wqe->data[0].addr = cpu_to_be64(segs[0].ds_addr);
82 	wqe->data[0].byte_count = cpu_to_be32(segs[0].ds_len |
83 	    MLX5_HW_START_PADDING);
84 	for (i = 1; i != nsegs; i++) {
85 		wqe->data[i].addr = cpu_to_be64(segs[i].ds_addr);
86 		wqe->data[i].byte_count = cpu_to_be32(segs[i].ds_len);
87 	}
88 	for (; i < rq->nsegs; i++) {
89 		wqe->data[i].addr = 0;
90 		wqe->data[i].byte_count = 0;
91 	}
92 
93 	rq->mbuf[ix].mbuf = mb;
94 	rq->mbuf[ix].data = mb->m_data;
95 
96 	bus_dmamap_sync(rq->dma_tag, rq->mbuf[ix].dma_map,
97 	    BUS_DMASYNC_PREREAD);
98 	return (0);
99 
100 err_free_mbuf:
101 	m_freem(mb);
102 	return (err);
103 }
104 
105 static void
106 mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
107 {
108 	if (unlikely(rq->enabled == 0))
109 		return;
110 
111 	while (!mlx5_wq_ll_is_full(&rq->wq)) {
112 		struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, rq->wq.head);
113 
114 		if (unlikely(mlx5e_alloc_rx_wqe(rq, wqe, rq->wq.head))) {
115 			callout_reset_curcpu(&rq->watchdog, 1, (void *)&mlx5e_post_rx_wqes, rq);
116 			break;
117 		}
118 		mlx5_wq_ll_push(&rq->wq, be16_to_cpu(wqe->next.next_wqe_index));
119 	}
120 
121 	/* ensure wqes are visible to device before updating doorbell record */
122 	atomic_thread_fence_rel();
123 
124 	mlx5_wq_ll_update_db_record(&rq->wq);
125 }
126 
127 static void
128 mlx5e_lro_update_hdr(struct mbuf *mb, struct mlx5_cqe64 *cqe)
129 {
130 	/* TODO: consider vlans, ip options, ... */
131 	struct ether_header *eh;
132 	uint16_t eh_type;
133 	uint16_t tot_len;
134 	struct ip6_hdr *ip6 = NULL;
135 	struct ip *ip4 = NULL;
136 	struct tcphdr *th;
137 	uint32_t *ts_ptr;
138 	uint8_t l4_hdr_type;
139 	int tcp_ack;
140 
141 	eh = mtod(mb, struct ether_header *);
142 	eh_type = ntohs(eh->ether_type);
143 
144 	l4_hdr_type = get_cqe_l4_hdr_type(cqe);
145 	tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA == l4_hdr_type) ||
146 	    (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
147 
148 	/* TODO: consider vlan */
149 	tot_len = be32_to_cpu(cqe->byte_cnt) - ETHER_HDR_LEN;
150 
151 	switch (eh_type) {
152 	case ETHERTYPE_IP:
153 		ip4 = (struct ip *)(eh + 1);
154 		th = (struct tcphdr *)(ip4 + 1);
155 		break;
156 	case ETHERTYPE_IPV6:
157 		ip6 = (struct ip6_hdr *)(eh + 1);
158 		th = (struct tcphdr *)(ip6 + 1);
159 		break;
160 	default:
161 		return;
162 	}
163 
164 	ts_ptr = (uint32_t *)(th + 1);
165 
166 	if (get_cqe_lro_tcppsh(cqe))
167 		th->th_flags |= TH_PUSH;
168 
169 	if (tcp_ack) {
170 		th->th_flags |= TH_ACK;
171 		th->th_ack = cqe->lro_ack_seq_num;
172 		th->th_win = cqe->lro_tcp_win;
173 
174 		/*
175 		 * FreeBSD handles only 32bit aligned timestamp right after
176 		 * the TCP hdr
177 		 * +--------+--------+--------+--------+
178 		 * |   NOP  |  NOP   |  TSopt |   10   |
179 		 * +--------+--------+--------+--------+
180 		 * |          TSval   timestamp        |
181 		 * +--------+--------+--------+--------+
182 		 * |          TSecr   timestamp        |
183 		 * +--------+--------+--------+--------+
184 		 */
185 		if (get_cqe_lro_timestamp_valid(cqe) &&
186 		    (__predict_true(*ts_ptr) == ntohl(TCPOPT_NOP << 24 |
187 		    TCPOPT_NOP << 16 | TCPOPT_TIMESTAMP << 8 |
188 		    TCPOLEN_TIMESTAMP))) {
189 			/*
190 			 * cqe->timestamp is 64bit long.
191 			 * [0-31] - timestamp.
192 			 * [32-64] - timestamp echo replay.
193 			 */
194 			ts_ptr[1] = *(uint32_t *)&cqe->timestamp;
195 			ts_ptr[2] = *((uint32_t *)&cqe->timestamp + 1);
196 		}
197 	}
198 	if (ip4) {
199 		ip4->ip_ttl = cqe->lro_min_ttl;
200 		ip4->ip_len = cpu_to_be16(tot_len);
201 		ip4->ip_sum = 0;
202 		ip4->ip_sum = in_cksum(mb, ip4->ip_hl << 2);
203 	} else {
204 		ip6->ip6_hlim = cqe->lro_min_ttl;
205 		ip6->ip6_plen = cpu_to_be16(tot_len -
206 		    sizeof(struct ip6_hdr));
207 	}
208 	/* TODO: handle tcp checksum */
209 }
210 
211 static uint64_t
212 mlx5e_mbuf_tstmp(struct mlx5e_priv *priv, uint64_t hw_tstmp)
213 {
214 	struct mlx5e_clbr_point *cp, dcp;
215 	uint64_t tstmp_sec, tstmp_nsec;
216 	uint64_t hw_clocks;
217 	uint64_t rt_cur_to_prev, res_s, res_n, res_s_modulo, res;
218 	uint64_t hw_clk_div;
219 	u_int gen;
220 
221 	do {
222 		cp = &priv->clbr_points[priv->clbr_curr];
223 		gen = atomic_load_acq_int(&cp->clbr_gen);
224 		if (gen == 0)
225 			return (0);
226 		dcp = *cp;
227 		atomic_thread_fence_acq();
228 	} while (gen != dcp.clbr_gen);
229 	/*
230 	 * Our goal here is to have a result that is:
231 	 *
232 	 * (                             (cur_time - prev_time)   )
233 	 * ((hw_tstmp - hw_prev) *  ----------------------------- ) + prev_time
234 	 * (                             (hw_cur - hw_prev)       )
235 	 *
236 	 * With the constraints that we cannot use float and we
237 	 * don't want to overflow the uint64_t numbers we are using.
238 	 *
239 	 * The plan is to take the clocking value of the hw timestamps
240 	 * and split them into seconds and nanosecond equivalent portions.
241 	 * Then we operate on the two portions seperately making sure to
242 	 * bring back the carry over from the seconds when we divide.
243 	 *
244 	 * First up lets get the two divided into separate entities
245 	 * i.e. the seconds. We use the clock frequency for this.
246 	 * Note that priv->cclk was setup with the clock frequency
247 	 * in hz so we are all set to go.
248 	 */
249 	hw_clocks = hw_tstmp - dcp.clbr_hw_prev;
250 	tstmp_sec = hw_clocks / priv->cclk;
251 	tstmp_nsec = hw_clocks % priv->cclk;
252 	/* Now work with them separately */
253 	rt_cur_to_prev = (dcp.base_curr - dcp.base_prev);
254 	res_s = tstmp_sec * rt_cur_to_prev;
255 	res_n = tstmp_nsec * rt_cur_to_prev;
256 	/* Now lets get our divider */
257 	hw_clk_div = dcp.clbr_hw_curr - dcp.clbr_hw_prev;
258 	/* Make sure to save the remainder from the seconds divide */
259 	res_s_modulo = res_s % hw_clk_div;
260 	res_s /= hw_clk_div;
261 	/* scale the remainder to where it should be */
262 	res_s_modulo *= priv->cclk;
263 	/* Now add in the remainder */
264 	res_n += res_s_modulo;
265 	/* Now do the divide */
266 	res_n /= hw_clk_div;
267 	res_s *= priv->cclk;
268 	/* Recombine the two */
269 	res = res_s + res_n;
270 	/* And now add in the base time to get to the real timestamp */
271 	res += dcp.base_prev;
272 	return (res);
273 }
274 
275 static inline void
276 mlx5e_build_rx_mbuf(struct mlx5_cqe64 *cqe,
277     struct mlx5e_rq *rq, struct mbuf *mb,
278     u32 cqe_bcnt)
279 {
280 	if_t ifp = rq->ifp;
281 	struct mlx5e_channel *c;
282 	struct mbuf *mb_head;
283 	int lro_num_seg;	/* HW LRO session aggregated packets counter */
284 	uint64_t tstmp;
285 
286 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
287 	if (lro_num_seg > 1) {
288 		mlx5e_lro_update_hdr(mb, cqe);
289 		rq->stats.lro_packets++;
290 		rq->stats.lro_bytes += cqe_bcnt;
291 	}
292 
293 	mb->m_pkthdr.len = cqe_bcnt;
294 	for (mb_head = mb; mb != NULL; mb = mb->m_next) {
295 		if (mb->m_len > cqe_bcnt)
296 			mb->m_len = cqe_bcnt;
297 		cqe_bcnt -= mb->m_len;
298 		if (likely(cqe_bcnt == 0)) {
299 			if (likely(mb->m_next != NULL)) {
300 				/* trim off empty mbufs */
301 				m_freem(mb->m_next);
302 				mb->m_next = NULL;
303 			}
304 			break;
305 		}
306 	}
307 	/* rewind to first mbuf in chain */
308 	mb = mb_head;
309 
310 	/* check if a Toeplitz hash was computed */
311 	if (cqe->rss_hash_type != 0) {
312 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->rss_hash_result);
313 #ifdef RSS
314 		/* decode the RSS hash type */
315 		switch (cqe->rss_hash_type &
316 		    (CQE_RSS_DST_HTYPE_L4 | CQE_RSS_DST_HTYPE_IP)) {
317 		/* IPv4 */
318 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV4):
319 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV4);
320 			break;
321 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV4):
322 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV4);
323 			break;
324 		case CQE_RSS_DST_HTYPE_IPV4:
325 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV4);
326 			break;
327 		/* IPv6 */
328 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV6):
329 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV6);
330 			break;
331 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV6):
332 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV6);
333 			break;
334 		case CQE_RSS_DST_HTYPE_IPV6:
335 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV6);
336 			break;
337 		default:	/* Other */
338 			M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
339 			break;
340 		}
341 #else
342 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
343 #endif
344 #ifdef M_HASHTYPE_SETINNER
345 		if (cqe_is_tunneled(cqe))
346 			M_HASHTYPE_SETINNER(mb);
347 #endif
348 	} else {
349 		mb->m_pkthdr.flowid = rq->ix;
350 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
351 	}
352 	mb->m_pkthdr.rcvif = ifp;
353 	mb->m_pkthdr.leaf_rcvif = ifp;
354 
355 	if (cqe_is_tunneled(cqe)) {
356 		/*
357 		 * CQE can be tunneled only if TIR is configured to
358 		 * enable parsing of tunneled payload, so no need to
359 		 * check for capabilities.
360 		 */
361 		if (((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK)) ==
362 		    (CQE_L2_OK | CQE_L3_OK))) {
363 			mb->m_pkthdr.csum_flags |=
364 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
365 			    CSUM_IP_CHECKED | CSUM_IP_VALID |
366 			    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
367 			mb->m_pkthdr.csum_data = htons(0xffff);
368 
369 			if (likely((cqe->hds_ip_ext & CQE_L4_OK) == CQE_L4_OK)) {
370 				mb->m_pkthdr.csum_flags |=
371 				    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
372 			}
373 		} else {
374 			rq->stats.csum_none++;
375 		}
376 	} else if (likely((if_getcapenable(ifp) & (IFCAP_RXCSUM |
377 	    IFCAP_RXCSUM_IPV6)) != 0) &&
378 	    ((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK)) ==
379 	    (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK))) {
380 		mb->m_pkthdr.csum_flags =
381 		    CSUM_IP_CHECKED | CSUM_IP_VALID |
382 		    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
383 		mb->m_pkthdr.csum_data = htons(0xffff);
384 	} else {
385 		rq->stats.csum_none++;
386 	}
387 
388 	if (cqe_has_vlan(cqe)) {
389 		mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->vlan_info);
390 		mb->m_flags |= M_VLANTAG;
391 	}
392 
393 	c = container_of(rq, struct mlx5e_channel, rq);
394 	if (c->priv->clbr_done >= 2) {
395 		tstmp = mlx5e_mbuf_tstmp(c->priv, be64_to_cpu(cqe->timestamp));
396 		if ((tstmp & MLX5_CQE_TSTMP_PTP) != 0) {
397 			/*
398 			 * Timestamp was taken on the packet entrance,
399 			 * instead of the cqe generation.
400 			 */
401 			tstmp &= ~MLX5_CQE_TSTMP_PTP;
402 			mb->m_flags |= M_TSTMP_HPREC;
403 		}
404 		if (tstmp != 0) {
405 			mb->m_pkthdr.rcv_tstmp = tstmp;
406 			mb->m_flags |= M_TSTMP;
407 		}
408 	}
409 	switch (get_cqe_tls_offload(cqe)) {
410 	case CQE_TLS_OFFLOAD_DECRYPTED:
411 		/* set proper checksum flag for decrypted packets */
412 		mb->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
413 		rq->stats.decrypted_ok_packets++;
414 		break;
415 	case CQE_TLS_OFFLOAD_ERROR:
416 		rq->stats.decrypted_error_packets++;
417 		break;
418 	default:
419 		break;
420 	}
421 }
422 
423 static inline void
424 mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
425 {
426 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, (cc & cq->wq.sz_m1)),
427 	    sizeof(struct mlx5_cqe64));
428 }
429 
430 static inline void
431 mlx5e_write_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
432 {
433 	memcpy(mlx5_cqwq_get_wqe(&cq->wq, cc & cq->wq.sz_m1),
434 	    data, sizeof(struct mlx5_cqe64));
435 }
436 
437 static inline void
438 mlx5e_decompress_cqe(struct mlx5e_cq *cq, struct mlx5_cqe64 *title,
439     struct mlx5_mini_cqe8 *mini,
440     u16 wqe_counter, int i)
441 {
442 	/*
443 	 * NOTE: The fields which are not set here are copied from the
444 	 * initial and common title. See memcpy() in
445 	 * mlx5e_write_cqe_slot().
446 	 */
447 	title->byte_cnt = mini->byte_cnt;
448 	title->wqe_counter = cpu_to_be16((wqe_counter + i) & cq->wq.sz_m1);
449 	title->rss_hash_result = mini->rx_hash_result;
450 	/*
451 	 * Since we use MLX5_CQE_FORMAT_HASH when creating the RX CQ,
452 	 * the value of the checksum should be ignored.
453 	 */
454 	title->check_sum = 0;
455 	title->op_own = (title->op_own & 0xf0) |
456 	    (((cq->wq.cc + i) >> cq->wq.log_sz) & 1);
457 }
458 
459 #define MLX5E_MINI_ARRAY_SZ 8
460 /* Make sure structs are not packet differently */
461 CTASSERT(sizeof(struct mlx5_cqe64) ==
462     sizeof(struct mlx5_mini_cqe8) * MLX5E_MINI_ARRAY_SZ);
463 static void
464 mlx5e_decompress_cqes(struct mlx5e_cq *cq)
465 {
466 	struct mlx5_mini_cqe8 mini_array[MLX5E_MINI_ARRAY_SZ];
467 	struct mlx5_cqe64 title;
468 	u32 cqe_count;
469 	u32 i = 0;
470 	u16 title_wqe_counter;
471 
472 	mlx5e_read_cqe_slot(cq, cq->wq.cc, &title);
473 	title_wqe_counter = be16_to_cpu(title.wqe_counter);
474 	cqe_count = be32_to_cpu(title.byte_cnt);
475 
476 	/* Make sure we won't overflow */
477 	KASSERT(cqe_count <= cq->wq.sz_m1,
478 	    ("%s: cqe_count %u > cq->wq.sz_m1 %u", __func__,
479 	    cqe_count, cq->wq.sz_m1));
480 
481 	mlx5e_read_cqe_slot(cq, cq->wq.cc + 1, mini_array);
482 	while (true) {
483 		mlx5e_decompress_cqe(cq, &title,
484 		    &mini_array[i % MLX5E_MINI_ARRAY_SZ],
485 		    title_wqe_counter, i);
486 		mlx5e_write_cqe_slot(cq, cq->wq.cc + i, &title);
487 		i++;
488 
489 		if (i == cqe_count)
490 			break;
491 		if (i % MLX5E_MINI_ARRAY_SZ == 0)
492 			mlx5e_read_cqe_slot(cq, cq->wq.cc + i, mini_array);
493 	}
494 }
495 
496 static int
497 mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
498 {
499 	struct pfil_head *pfil;
500 	int i, rv;
501 
502 	CURVNET_SET_QUIET(if_getvnet(rq->ifp));
503 	pfil = rq->channel->priv->pfil;
504 	for (i = 0; i < budget; i++) {
505 		struct mlx5e_rx_wqe *wqe;
506 		struct mlx5_cqe64 *cqe;
507 		struct mbuf *mb;
508 		__be16 wqe_counter_be;
509 		u16 wqe_counter;
510 		u32 byte_cnt, seglen;
511 
512 		cqe = mlx5e_get_cqe(&rq->cq);
513 		if (!cqe)
514 			break;
515 
516 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED)
517 			mlx5e_decompress_cqes(&rq->cq);
518 
519 		mlx5_cqwq_pop(&rq->cq.wq);
520 
521 		wqe_counter_be = cqe->wqe_counter;
522 		wqe_counter = be16_to_cpu(wqe_counter_be);
523 		wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
524 		byte_cnt = be32_to_cpu(cqe->byte_cnt);
525 
526 		bus_dmamap_sync(rq->dma_tag,
527 		    rq->mbuf[wqe_counter].dma_map,
528 		    BUS_DMASYNC_POSTREAD);
529 
530 		if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
531 			mlx5e_dump_err_cqe(&rq->cq, rq->rqn, (const void *)cqe);
532 			rq->stats.wqe_err++;
533 			goto wq_ll_pop;
534 		}
535 		if (pfil != NULL && PFIL_HOOKED_IN(pfil)) {
536 			seglen = MIN(byte_cnt, MLX5E_MAX_RX_BYTES);
537 			rv = pfil_mem_in(rq->channel->priv->pfil,
538 			    rq->mbuf[wqe_counter].data, seglen, rq->ifp, &mb);
539 
540 			switch (rv) {
541 			case PFIL_DROPPED:
542 			case PFIL_CONSUMED:
543 				/*
544 				 * Filter dropped or consumed it. In
545 				 * either case, we can just recycle
546 				 * buffer; there is no more work to do.
547 				 */
548 				rq->stats.packets++;
549 				goto wq_ll_pop;
550 			case PFIL_REALLOCED:
551 				/*
552 				 * Filter copied it; recycle buffer
553 				 * and receive the new mbuf allocated
554 				 * by the Filter
555 				 */
556 				goto rx_common;
557 			default:
558 				/*
559 				 * The Filter said it was OK, so
560 				 * receive like normal.
561 				 */
562 				KASSERT(rv == PFIL_PASS,
563 					("Filter returned %d!\n", rv));
564 			}
565 		}
566 		if ((MHLEN - MLX5E_NET_IP_ALIGN) >= byte_cnt &&
567 		    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
568 			/* set maximum mbuf length */
569 			mb->m_len = MHLEN - MLX5E_NET_IP_ALIGN;
570 			/* get IP header aligned */
571 			mb->m_data += MLX5E_NET_IP_ALIGN;
572 
573 			bcopy(rq->mbuf[wqe_counter].data, mtod(mb, caddr_t),
574 			    byte_cnt);
575 		} else {
576 			mb = rq->mbuf[wqe_counter].mbuf;
577 			rq->mbuf[wqe_counter].mbuf = NULL;	/* safety clear */
578 
579 			bus_dmamap_unload(rq->dma_tag,
580 			    rq->mbuf[wqe_counter].dma_map);
581 		}
582 rx_common:
583 		mlx5e_build_rx_mbuf(cqe, rq, mb, byte_cnt);
584 		rq->stats.bytes += byte_cnt;
585 		rq->stats.packets++;
586 #ifdef NUMA
587 		mb->m_pkthdr.numa_domain = if_getnumadomain(rq->ifp);
588 #endif
589 
590 #if !defined(HAVE_TCP_LRO_RX)
591 		tcp_lro_queue_mbuf(&rq->lro, mb);
592 #else
593 		if (mb->m_pkthdr.csum_flags == 0 ||
594 		    (if_getcapenable(rq->ifp) & IFCAP_LRO) == 0 ||
595 		    rq->lro.lro_cnt == 0 ||
596 		    tcp_lro_rx(&rq->lro, mb, 0) != 0) {
597 			if_input(rq->ifp, mb);
598 		}
599 #endif
600 wq_ll_pop:
601 		mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
602 		    &wqe->next.next_wqe_index);
603 	}
604 	CURVNET_RESTORE();
605 
606 	mlx5_cqwq_update_db_record(&rq->cq.wq);
607 
608 	/* ensure cq space is freed before enabling more cqes */
609 	atomic_thread_fence_rel();
610 	return (i);
611 }
612 
613 void
614 mlx5e_rx_cq_comp(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe __unused)
615 {
616 	struct mlx5e_channel *c = container_of(mcq, struct mlx5e_channel, rq.cq.mcq);
617 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
618 	int i = 0;
619 
620 #ifdef HAVE_PER_CQ_EVENT_PACKET
621 #if (MHLEN < 15)
622 #error "MHLEN is too small"
623 #endif
624 	struct mbuf *mb = m_gethdr(M_NOWAIT, MT_DATA);
625 
626 	if (mb != NULL) {
627 		/* this code is used for debugging purpose only */
628 		mb->m_pkthdr.len = mb->m_len = 15;
629 		memset(mb->m_data, 255, 14);
630 		mb->m_data[14] = rq->ix;
631 		mb->m_pkthdr.rcvif = rq->ifp;
632 		mb->m_pkthdr.leaf_rcvif = rq->ifp;
633 		if_input(rq->ifp, mb);
634 	}
635 #endif
636 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
637 		mtx_lock(&c->sq[j].lock);
638 		c->sq[j].db_inhibit++;
639 		mtx_unlock(&c->sq[j].lock);
640 	}
641 
642 	mtx_lock(&c->iq.lock);
643 	c->iq.db_inhibit++;
644 	mtx_unlock(&c->iq.lock);
645 
646 	mtx_lock(&rq->mtx);
647 
648 	/*
649 	 * Polling the entire CQ without posting new WQEs results in
650 	 * lack of receive WQEs during heavy traffic scenarios.
651 	 */
652 	while (1) {
653 		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
654 		    MLX5E_RX_BUDGET_MAX)
655 			break;
656 		i += MLX5E_RX_BUDGET_MAX;
657 		if (i >= MLX5E_BUDGET_MAX)
658 			break;
659 		mlx5e_post_rx_wqes(rq);
660 	}
661 	mlx5e_post_rx_wqes(rq);
662 	/* check for dynamic interrupt moderation callback */
663 	if (rq->dim.mode != NET_DIM_CQ_PERIOD_MODE_DISABLED)
664 		net_dim(&rq->dim, rq->stats.packets, rq->stats.bytes);
665 	mlx5e_cq_arm(&rq->cq, MLX5_GET_DOORBELL_LOCK(&rq->channel->priv->doorbell_lock));
666 	tcp_lro_flush_all(&rq->lro);
667 	mtx_unlock(&rq->mtx);
668 
669 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
670 		mtx_lock(&c->sq[j].lock);
671 		c->sq[j].db_inhibit--;
672 		/* Update the doorbell record, if any. */
673 		mlx5e_tx_notify_hw(c->sq + j, true);
674 		mtx_unlock(&c->sq[j].lock);
675 	}
676 
677 	mtx_lock(&c->iq.lock);
678 	c->iq.db_inhibit--;
679 	mlx5e_iq_notify_hw(&c->iq);
680 	mtx_unlock(&c->iq.lock);
681 }
682