xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c (revision d00f3505efad7c323f74bee63f7d2527daf0f534)
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 #include <dev/mlx5/mlx5_accel/ipsec.h>
32 
33 static inline int
mlx5e_alloc_rx_wqe(struct mlx5e_rq * rq,struct mlx5e_rx_wqe * wqe,u16 ix)34 mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq,
35     struct mlx5e_rx_wqe *wqe, u16 ix)
36 {
37 	bus_dma_segment_t segs[MLX5E_MAX_BUSDMA_RX_SEGS];
38 	struct mbuf *mb;
39 	int nsegs;
40 	int err;
41 	struct mbuf *mb_head;
42 	int i;
43 
44 	if (rq->mbuf[ix].mbuf != NULL)
45 		return (0);
46 
47 	mb_head = mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR,
48 	    MLX5E_MAX_RX_BYTES);
49 	if (unlikely(mb == NULL))
50 		return (-ENOMEM);
51 
52 	mb->m_len = MLX5E_MAX_RX_BYTES;
53 	mb->m_pkthdr.len = MLX5E_MAX_RX_BYTES;
54 
55 	for (i = 1; i < rq->nsegs; i++) {
56 		if (mb_head->m_pkthdr.len >= rq->wqe_sz)
57 			break;
58 		mb = mb->m_next = m_getjcl(M_NOWAIT, MT_DATA, 0,
59 		    MLX5E_MAX_RX_BYTES);
60 		if (unlikely(mb == NULL)) {
61 			m_freem(mb_head);
62 			return (-ENOMEM);
63 		}
64 		mb->m_len = MLX5E_MAX_RX_BYTES;
65 		mb_head->m_pkthdr.len += MLX5E_MAX_RX_BYTES;
66 	}
67 	/* rewind to first mbuf in chain */
68 	mb = mb_head;
69 
70 	/* get IP header aligned */
71 	m_adj(mb, MLX5E_NET_IP_ALIGN);
72 
73 	err = mlx5_accel_ipsec_rx_tag_add(rq->ifp, &rq->mbuf[ix]);
74 	if (err)
75 		goto err_free_mbuf;
76 	err = -bus_dmamap_load_mbuf_sg(rq->dma_tag, rq->mbuf[ix].dma_map,
77 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
78 	if (err != 0)
79 		goto err_free_mbuf;
80 	if (unlikely(nsegs == 0)) {
81 		bus_dmamap_unload(rq->dma_tag, rq->mbuf[ix].dma_map);
82 		err = -ENOMEM;
83 		goto err_free_mbuf;
84 	}
85 	wqe->data[0].addr = cpu_to_be64(segs[0].ds_addr);
86 	wqe->data[0].byte_count = cpu_to_be32(segs[0].ds_len |
87 	    MLX5_HW_START_PADDING);
88 	for (i = 1; i != nsegs; i++) {
89 		wqe->data[i].addr = cpu_to_be64(segs[i].ds_addr);
90 		wqe->data[i].byte_count = cpu_to_be32(segs[i].ds_len);
91 	}
92 	for (; i < rq->nsegs; i++) {
93 		wqe->data[i].addr = 0;
94 		wqe->data[i].byte_count = 0;
95 	}
96 
97 	rq->mbuf[ix].mbuf = mb;
98 	rq->mbuf[ix].data = mb->m_data;
99 
100 	bus_dmamap_sync(rq->dma_tag, rq->mbuf[ix].dma_map,
101 	    BUS_DMASYNC_PREREAD);
102 	return (0);
103 
104 err_free_mbuf:
105 	m_freem(mb);
106 	return (err);
107 }
108 
109 static void
mlx5e_post_rx_wqes(struct mlx5e_rq * rq)110 mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
111 {
112 	if (unlikely(rq->enabled == 0))
113 		return;
114 
115 	while (!mlx5_wq_ll_is_full(&rq->wq)) {
116 		struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, rq->wq.head);
117 
118 		if (unlikely(mlx5e_alloc_rx_wqe(rq, wqe, rq->wq.head))) {
119 			callout_reset_curcpu(&rq->watchdog, 1, (void *)&mlx5e_post_rx_wqes, rq);
120 			break;
121 		}
122 		mlx5_wq_ll_push(&rq->wq, be16_to_cpu(wqe->next.next_wqe_index));
123 	}
124 
125 	/* ensure wqes are visible to device before updating doorbell record */
126 	atomic_thread_fence_rel();
127 
128 	mlx5_wq_ll_update_db_record(&rq->wq);
129 }
130 
131 static void
mlx5e_lro_update_hdr(struct mbuf * mb,struct mlx5_cqe64 * cqe)132 mlx5e_lro_update_hdr(struct mbuf *mb, struct mlx5_cqe64 *cqe)
133 {
134 	/* TODO: consider vlans, ip options, ... */
135 	struct ether_header *eh;
136 	uint16_t eh_type;
137 	uint16_t tot_len;
138 	struct ip6_hdr *ip6 = NULL;
139 	struct ip *ip4 = NULL;
140 	struct tcphdr *th;
141 	uint32_t *ts_ptr;
142 	uint8_t l4_hdr_type;
143 	int tcp_ack;
144 
145 	eh = mtod(mb, struct ether_header *);
146 	eh_type = ntohs(eh->ether_type);
147 
148 	l4_hdr_type = get_cqe_l4_hdr_type(cqe);
149 	tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA == l4_hdr_type) ||
150 	    (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
151 
152 	/* TODO: consider vlan */
153 	tot_len = be32_to_cpu(cqe->byte_cnt) - ETHER_HDR_LEN;
154 
155 	switch (eh_type) {
156 	case ETHERTYPE_IP:
157 		ip4 = (struct ip *)(eh + 1);
158 		th = (struct tcphdr *)(ip4 + 1);
159 		break;
160 	case ETHERTYPE_IPV6:
161 		ip6 = (struct ip6_hdr *)(eh + 1);
162 		th = (struct tcphdr *)(ip6 + 1);
163 		break;
164 	default:
165 		return;
166 	}
167 
168 	ts_ptr = (uint32_t *)(th + 1);
169 
170 	if (get_cqe_lro_tcppsh(cqe))
171 		th->th_flags |= TH_PUSH;
172 
173 	if (tcp_ack) {
174 		th->th_flags |= TH_ACK;
175 		th->th_ack = cqe->lro_ack_seq_num;
176 		th->th_win = cqe->lro_tcp_win;
177 
178 		/*
179 		 * FreeBSD handles only 32bit aligned timestamp right after
180 		 * the TCP hdr
181 		 * +--------+--------+--------+--------+
182 		 * |   NOP  |  NOP   |  TSopt |   10   |
183 		 * +--------+--------+--------+--------+
184 		 * |          TSval   timestamp        |
185 		 * +--------+--------+--------+--------+
186 		 * |          TSecr   timestamp        |
187 		 * +--------+--------+--------+--------+
188 		 */
189 		if (get_cqe_lro_timestamp_valid(cqe) &&
190 		    (__predict_true(*ts_ptr) == ntohl(TCPOPT_NOP << 24 |
191 		    TCPOPT_NOP << 16 | TCPOPT_TIMESTAMP << 8 |
192 		    TCPOLEN_TIMESTAMP))) {
193 			/*
194 			 * cqe->timestamp is 64bit long.
195 			 * [0-31] - timestamp.
196 			 * [32-64] - timestamp echo replay.
197 			 */
198 			ts_ptr[1] = *(uint32_t *)&cqe->timestamp;
199 			ts_ptr[2] = *((uint32_t *)&cqe->timestamp + 1);
200 		}
201 	}
202 	if (ip4) {
203 		ip4->ip_ttl = cqe->lro_min_ttl;
204 		ip4->ip_len = cpu_to_be16(tot_len);
205 		ip4->ip_sum = 0;
206 		ip4->ip_sum = in_cksum(mb, ip4->ip_hl << 2);
207 	} else {
208 		ip6->ip6_hlim = cqe->lro_min_ttl;
209 		ip6->ip6_plen = cpu_to_be16(tot_len -
210 		    sizeof(struct ip6_hdr));
211 	}
212 	/* TODO: handle tcp checksum */
213 }
214 
215 static uint64_t
mlx5e_mbuf_tstmp(struct mlx5e_priv * priv,uint64_t hw_tstmp)216 mlx5e_mbuf_tstmp(struct mlx5e_priv *priv, uint64_t hw_tstmp)
217 {
218 	struct mlx5e_clbr_point *cp, dcp;
219 	uint64_t tstmp_sec, tstmp_nsec;
220 	uint64_t hw_clocks;
221 	uint64_t rt_cur_to_prev, res_s, res_n, res_s_modulo, res;
222 	uint64_t hw_clk_div;
223 	u_int gen;
224 
225 	do {
226 		cp = &priv->clbr_points[priv->clbr_curr];
227 		gen = atomic_load_acq_int(&cp->clbr_gen);
228 		if (gen == 0)
229 			return (0);
230 		dcp = *cp;
231 		atomic_thread_fence_acq();
232 	} while (gen != dcp.clbr_gen);
233 	/*
234 	 * Our goal here is to have a result that is:
235 	 *
236 	 * (                             (cur_time - prev_time)   )
237 	 * ((hw_tstmp - hw_prev) *  ----------------------------- ) + prev_time
238 	 * (                             (hw_cur - hw_prev)       )
239 	 *
240 	 * With the constraints that we cannot use float and we
241 	 * don't want to overflow the uint64_t numbers we are using.
242 	 *
243 	 * The plan is to take the clocking value of the hw timestamps
244 	 * and split them into seconds and nanosecond equivalent portions.
245 	 * Then we operate on the two portions seperately making sure to
246 	 * bring back the carry over from the seconds when we divide.
247 	 *
248 	 * First up lets get the two divided into separate entities
249 	 * i.e. the seconds. We use the clock frequency for this.
250 	 * Note that priv->cclk was setup with the clock frequency
251 	 * in hz so we are all set to go.
252 	 */
253 	hw_clocks = hw_tstmp - dcp.clbr_hw_prev;
254 	tstmp_sec = hw_clocks / priv->cclk;
255 	tstmp_nsec = hw_clocks % priv->cclk;
256 	/* Now work with them separately */
257 	rt_cur_to_prev = (dcp.base_curr - dcp.base_prev);
258 	res_s = tstmp_sec * rt_cur_to_prev;
259 	res_n = tstmp_nsec * rt_cur_to_prev;
260 	/* Now lets get our divider */
261 	hw_clk_div = dcp.clbr_hw_curr - dcp.clbr_hw_prev;
262 	/* Make sure to save the remainder from the seconds divide */
263 	res_s_modulo = res_s % hw_clk_div;
264 	res_s /= hw_clk_div;
265 	/* scale the remainder to where it should be */
266 	res_s_modulo *= priv->cclk;
267 	/* Now add in the remainder */
268 	res_n += res_s_modulo;
269 	/* Now do the divide */
270 	res_n /= hw_clk_div;
271 	res_s *= priv->cclk;
272 	/* Recombine the two */
273 	res = res_s + res_n;
274 	/* And now add in the base time to get to the real timestamp */
275 	res += dcp.base_prev;
276 	return (res);
277 }
278 
279 static inline void
mlx5e_build_rx_mbuf(struct mlx5_cqe64 * cqe,struct mlx5e_rq * rq,struct mbuf * mb,struct mlx5e_rq_mbuf * mr,u32 cqe_bcnt)280 mlx5e_build_rx_mbuf(struct mlx5_cqe64 *cqe, struct mlx5e_rq *rq,
281     struct mbuf *mb, struct mlx5e_rq_mbuf *mr, u32 cqe_bcnt)
282 {
283 	if_t ifp = rq->ifp;
284 	struct mlx5e_channel *c;
285 	struct mbuf *mb_head;
286 	int lro_num_seg;	/* HW LRO session aggregated packets counter */
287 	uint64_t tstmp;
288 
289 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
290 	if (lro_num_seg > 1) {
291 		mlx5e_lro_update_hdr(mb, cqe);
292 		rq->stats.lro_packets++;
293 		rq->stats.lro_bytes += cqe_bcnt;
294 	}
295 
296 	mb->m_pkthdr.len = cqe_bcnt;
297 	for (mb_head = mb; mb != NULL; mb = mb->m_next) {
298 		if (mb->m_len > cqe_bcnt)
299 			mb->m_len = cqe_bcnt;
300 		cqe_bcnt -= mb->m_len;
301 		if (likely(cqe_bcnt == 0)) {
302 			if (likely(mb->m_next != NULL)) {
303 				/* trim off empty mbufs */
304 				m_freem(mb->m_next);
305 				mb->m_next = NULL;
306 			}
307 			break;
308 		}
309 	}
310 	/* rewind to first mbuf in chain */
311 	mb = mb_head;
312 
313 	/* check if a Toeplitz hash was computed */
314 	if (cqe->rss_hash_type != 0) {
315 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->rss_hash_result);
316 #ifdef RSS
317 		/* decode the RSS hash type */
318 		switch (cqe->rss_hash_type &
319 		    (CQE_RSS_DST_HTYPE_L4 | CQE_RSS_DST_HTYPE_IP)) {
320 		/* IPv4 */
321 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV4):
322 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV4);
323 			break;
324 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV4):
325 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV4);
326 			break;
327 		case CQE_RSS_DST_HTYPE_IPV4:
328 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV4);
329 			break;
330 		/* IPv6 */
331 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV6):
332 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV6);
333 			break;
334 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV6):
335 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV6);
336 			break;
337 		case CQE_RSS_DST_HTYPE_IPV6:
338 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV6);
339 			break;
340 		default:	/* Other */
341 			M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
342 			break;
343 		}
344 #else
345 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
346 #endif
347 #ifdef M_HASHTYPE_SETINNER
348 		if (cqe_is_tunneled(cqe))
349 			M_HASHTYPE_SETINNER(mb);
350 #endif
351 	} else {
352 		mb->m_pkthdr.flowid = rq->ix;
353 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
354 	}
355 	mb->m_pkthdr.rcvif = ifp;
356 	mb->m_pkthdr.leaf_rcvif = ifp;
357 
358 	if (cqe_is_tunneled(cqe)) {
359 		/*
360 		 * CQE can be tunneled only if TIR is configured to
361 		 * enable parsing of tunneled payload, so no need to
362 		 * check for capabilities.
363 		 */
364 		if (((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK)) ==
365 		    (CQE_L2_OK | CQE_L3_OK))) {
366 			mb->m_pkthdr.csum_flags |=
367 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
368 			    CSUM_IP_CHECKED | CSUM_IP_VALID |
369 			    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
370 			mb->m_pkthdr.csum_data = htons(0xffff);
371 
372 			if (likely((cqe->hds_ip_ext & CQE_L4_OK) == CQE_L4_OK)) {
373 				mb->m_pkthdr.csum_flags |=
374 				    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
375 			}
376 		} else {
377 			rq->stats.csum_none++;
378 		}
379 	} else if (likely((if_getcapenable(ifp) & (IFCAP_RXCSUM |
380 	    IFCAP_RXCSUM_IPV6)) != 0) &&
381 	    ((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK)) ==
382 	    (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK))) {
383 		mb->m_pkthdr.csum_flags =
384 		    CSUM_IP_CHECKED | CSUM_IP_VALID |
385 		    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
386 		mb->m_pkthdr.csum_data = htons(0xffff);
387 	} else {
388 		rq->stats.csum_none++;
389 	}
390 
391 	if (cqe_has_vlan(cqe)) {
392 		mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->vlan_info);
393 		mb->m_flags |= M_VLANTAG;
394 	}
395 
396 	c = container_of(rq, struct mlx5e_channel, rq);
397 	if (c->priv->clbr_done >= 2) {
398 		tstmp = mlx5e_mbuf_tstmp(c->priv, be64_to_cpu(cqe->timestamp));
399 		if ((tstmp & MLX5_CQE_TSTMP_PTP) != 0) {
400 			/*
401 			 * Timestamp was taken on the packet entrance,
402 			 * instead of the cqe generation.
403 			 */
404 			tstmp &= ~MLX5_CQE_TSTMP_PTP;
405 			mb->m_flags |= M_TSTMP_HPREC;
406 		}
407 		if (tstmp != 0) {
408 			mb->m_pkthdr.rcv_tstmp = tstmp;
409 			mb->m_flags |= M_TSTMP;
410 		}
411 	}
412 	switch (get_cqe_tls_offload(cqe)) {
413 	case CQE_TLS_OFFLOAD_DECRYPTED:
414 		/* set proper checksum flag for decrypted packets */
415 		mb->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
416 		rq->stats.decrypted_ok_packets++;
417 		break;
418 	case CQE_TLS_OFFLOAD_ERROR:
419 		rq->stats.decrypted_error_packets++;
420 		break;
421 	default:
422 		break;
423 	}
424 
425 	mlx5e_accel_ipsec_handle_rx(mb, cqe, mr);
426 }
427 
428 static inline void
mlx5e_read_cqe_slot(struct mlx5e_cq * cq,u32 cc,void * data)429 mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
430 {
431 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, (cc & cq->wq.sz_m1)),
432 	    sizeof(struct mlx5_cqe64));
433 }
434 
435 static inline void
mlx5e_write_cqe_slot(struct mlx5e_cq * cq,u32 cc,void * data)436 mlx5e_write_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
437 {
438 	memcpy(mlx5_cqwq_get_wqe(&cq->wq, cc & cq->wq.sz_m1),
439 	    data, sizeof(struct mlx5_cqe64));
440 }
441 
442 static inline void
mlx5e_decompress_cqe(struct mlx5e_cq * cq,struct mlx5_cqe64 * title,struct mlx5_mini_cqe8 * mini,u16 wqe_counter,int i)443 mlx5e_decompress_cqe(struct mlx5e_cq *cq, struct mlx5_cqe64 *title,
444     struct mlx5_mini_cqe8 *mini,
445     u16 wqe_counter, int i)
446 {
447 	/*
448 	 * NOTE: The fields which are not set here are copied from the
449 	 * initial and common title. See memcpy() in
450 	 * mlx5e_write_cqe_slot().
451 	 */
452 	title->byte_cnt = mini->byte_cnt;
453 	title->wqe_counter = cpu_to_be16((wqe_counter + i) & cq->wq.sz_m1);
454 	title->rss_hash_result = mini->rx_hash_result;
455 	/*
456 	 * Since we use MLX5_CQE_FORMAT_HASH when creating the RX CQ,
457 	 * the value of the checksum should be ignored.
458 	 */
459 	title->check_sum = 0;
460 	title->op_own = (title->op_own & 0xf0) |
461 	    (((cq->wq.cc + i) >> cq->wq.log_sz) & 1);
462 }
463 
464 #define MLX5E_MINI_ARRAY_SZ 8
465 /* Make sure structs are not packet differently */
466 CTASSERT(sizeof(struct mlx5_cqe64) ==
467     sizeof(struct mlx5_mini_cqe8) * MLX5E_MINI_ARRAY_SZ);
468 static void
mlx5e_decompress_cqes(struct mlx5e_cq * cq)469 mlx5e_decompress_cqes(struct mlx5e_cq *cq)
470 {
471 	struct mlx5_mini_cqe8 mini_array[MLX5E_MINI_ARRAY_SZ];
472 	struct mlx5_cqe64 title;
473 	u32 cqe_count;
474 	u32 i = 0;
475 	u16 title_wqe_counter;
476 
477 	mlx5e_read_cqe_slot(cq, cq->wq.cc, &title);
478 	title_wqe_counter = be16_to_cpu(title.wqe_counter);
479 	cqe_count = be32_to_cpu(title.byte_cnt);
480 
481 	/* Make sure we won't overflow */
482 	KASSERT(cqe_count <= cq->wq.sz_m1,
483 	    ("%s: cqe_count %u > cq->wq.sz_m1 %u", __func__,
484 	    cqe_count, cq->wq.sz_m1));
485 
486 	mlx5e_read_cqe_slot(cq, cq->wq.cc + 1, mini_array);
487 	while (true) {
488 		mlx5e_decompress_cqe(cq, &title,
489 		    &mini_array[i % MLX5E_MINI_ARRAY_SZ],
490 		    title_wqe_counter, i);
491 		mlx5e_write_cqe_slot(cq, cq->wq.cc + i, &title);
492 		i++;
493 
494 		if (i == cqe_count)
495 			break;
496 		if (i % MLX5E_MINI_ARRAY_SZ == 0)
497 			mlx5e_read_cqe_slot(cq, cq->wq.cc + i, mini_array);
498 	}
499 }
500 
501 static int
mlx5e_poll_rx_cq(struct mlx5e_rq * rq,int budget)502 mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
503 {
504 	struct pfil_head *pfil;
505 	int i, rv;
506 
507 	CURVNET_SET_QUIET(if_getvnet(rq->ifp));
508 	pfil = rq->channel->priv->pfil;
509 	for (i = 0; i < budget; i++) {
510 		struct mlx5e_rx_wqe *wqe;
511 		struct mlx5_cqe64 *cqe;
512 		struct mbuf *mb;
513 		__be16 wqe_counter_be;
514 		u16 wqe_counter;
515 		u32 byte_cnt, seglen;
516 
517 		cqe = mlx5e_get_cqe(&rq->cq);
518 		if (!cqe)
519 			break;
520 
521 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED)
522 			mlx5e_decompress_cqes(&rq->cq);
523 
524 		mlx5_cqwq_pop(&rq->cq.wq);
525 
526 		wqe_counter_be = cqe->wqe_counter;
527 		wqe_counter = be16_to_cpu(wqe_counter_be);
528 		wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
529 		byte_cnt = be32_to_cpu(cqe->byte_cnt);
530 
531 		bus_dmamap_sync(rq->dma_tag,
532 		    rq->mbuf[wqe_counter].dma_map,
533 		    BUS_DMASYNC_POSTREAD);
534 
535 		if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
536 			mlx5e_dump_err_cqe(&rq->cq, rq->rqn, (const void *)cqe);
537 			rq->stats.wqe_err++;
538 			goto wq_ll_pop;
539 		}
540 		if (pfil != NULL && PFIL_HOOKED_IN(pfil)) {
541 			seglen = MIN(byte_cnt, MLX5E_MAX_RX_BYTES);
542 			rv = pfil_mem_in(rq->channel->priv->pfil,
543 			    rq->mbuf[wqe_counter].data, seglen, rq->ifp, &mb);
544 
545 			switch (rv) {
546 			case PFIL_DROPPED:
547 			case PFIL_CONSUMED:
548 				/*
549 				 * Filter dropped or consumed it. In
550 				 * either case, we can just recycle
551 				 * buffer; there is no more work to do.
552 				 */
553 				rq->stats.packets++;
554 				goto wq_ll_pop;
555 			case PFIL_REALLOCED:
556 				/*
557 				 * Filter copied it; recycle buffer
558 				 * and receive the new mbuf allocated
559 				 * by the Filter
560 				 */
561 				goto rx_common;
562 			default:
563 				/*
564 				 * The Filter said it was OK, so
565 				 * receive like normal.
566 				 */
567 				KASSERT(rv == PFIL_PASS,
568 					("Filter returned %d!\n", rv));
569 			}
570 		}
571 		if (!mlx5e_accel_ipsec_flow(cqe) /* tag is already assigned
572 						    to rq->mbuf */ &&
573 		    MHLEN - MLX5E_NET_IP_ALIGN >= byte_cnt &&
574 		    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
575 			/* set maximum mbuf length */
576 			mb->m_len = MHLEN - MLX5E_NET_IP_ALIGN;
577 			/* get IP header aligned */
578 			mb->m_data += MLX5E_NET_IP_ALIGN;
579 
580 			bcopy(rq->mbuf[wqe_counter].data, mtod(mb, caddr_t),
581 			    byte_cnt);
582 		} else {
583 			mb = rq->mbuf[wqe_counter].mbuf;
584 			rq->mbuf[wqe_counter].mbuf = NULL;	/* safety clear */
585 
586 			bus_dmamap_unload(rq->dma_tag,
587 			    rq->mbuf[wqe_counter].dma_map);
588 		}
589 rx_common:
590 		mlx5e_build_rx_mbuf(cqe, rq, mb, &rq->mbuf[wqe_counter],
591 		    byte_cnt);
592 		rq->stats.bytes += byte_cnt;
593 		rq->stats.packets++;
594 #ifdef NUMA
595 		mb->m_pkthdr.numa_domain = if_getnumadomain(rq->ifp);
596 #endif
597 
598 #if !defined(HAVE_TCP_LRO_RX)
599 		tcp_lro_queue_mbuf(&rq->lro, mb);
600 #else
601 		if (mb->m_pkthdr.csum_flags == 0 ||
602 		    (if_getcapenable(rq->ifp) & IFCAP_LRO) == 0 ||
603 		    rq->lro.lro_cnt == 0 ||
604 		    tcp_lro_rx(&rq->lro, mb, 0) != 0) {
605 			if_input(rq->ifp, mb);
606 		}
607 #endif
608 wq_ll_pop:
609 		mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
610 		    &wqe->next.next_wqe_index);
611 	}
612 	CURVNET_RESTORE();
613 
614 	mlx5_cqwq_update_db_record(&rq->cq.wq);
615 
616 	/* ensure cq space is freed before enabling more cqes */
617 	atomic_thread_fence_rel();
618 	return (i);
619 }
620 
621 void
mlx5e_rx_cq_comp(struct mlx5_core_cq * mcq,struct mlx5_eqe * eqe __unused)622 mlx5e_rx_cq_comp(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe __unused)
623 {
624 	struct mlx5e_channel *c = container_of(mcq, struct mlx5e_channel, rq.cq.mcq);
625 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
626 	int i = 0;
627 
628 #ifdef HAVE_PER_CQ_EVENT_PACKET
629 #if (MHLEN < 15)
630 #error "MHLEN is too small"
631 #endif
632 	struct mbuf *mb = m_gethdr(M_NOWAIT, MT_DATA);
633 
634 	if (mb != NULL) {
635 		/* this code is used for debugging purpose only */
636 		mb->m_pkthdr.len = mb->m_len = 15;
637 		memset(mb->m_data, 255, 14);
638 		mb->m_data[14] = rq->ix;
639 		mb->m_pkthdr.rcvif = rq->ifp;
640 		mb->m_pkthdr.leaf_rcvif = rq->ifp;
641 		if_input(rq->ifp, mb);
642 	}
643 #endif
644 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
645 		mtx_lock(&c->sq[j].lock);
646 		c->sq[j].db_inhibit++;
647 		mtx_unlock(&c->sq[j].lock);
648 	}
649 
650 	mtx_lock(&c->iq.lock);
651 	c->iq.db_inhibit++;
652 	mtx_unlock(&c->iq.lock);
653 
654 	mtx_lock(&rq->mtx);
655 
656 	/*
657 	 * Polling the entire CQ without posting new WQEs results in
658 	 * lack of receive WQEs during heavy traffic scenarios.
659 	 */
660 	while (1) {
661 		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
662 		    MLX5E_RX_BUDGET_MAX)
663 			break;
664 		i += MLX5E_RX_BUDGET_MAX;
665 		if (i >= MLX5E_BUDGET_MAX)
666 			break;
667 		mlx5e_post_rx_wqes(rq);
668 	}
669 	mlx5e_post_rx_wqes(rq);
670 	/* check for dynamic interrupt moderation callback */
671 	if (rq->dim.mode != NET_DIM_CQ_PERIOD_MODE_DISABLED)
672 		net_dim(&rq->dim, rq->stats.packets, rq->stats.bytes);
673 	mlx5e_cq_arm(&rq->cq, MLX5_GET_DOORBELL_LOCK(&rq->channel->priv->doorbell_lock));
674 	tcp_lro_flush_all(&rq->lro);
675 	mtx_unlock(&rq->mtx);
676 
677 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
678 		mtx_lock(&c->sq[j].lock);
679 		c->sq[j].db_inhibit--;
680 		/* Update the doorbell record, if any. */
681 		mlx5e_tx_notify_hw(c->sq + j, true);
682 		mtx_unlock(&c->sq[j].lock);
683 	}
684 
685 	mtx_lock(&c->iq.lock);
686 	c->iq.db_inhibit--;
687 	mlx5e_iq_notify_hw(&c->iq);
688 	mtx_unlock(&c->iq.lock);
689 }
690