xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c (revision fafb1ee7bdc5d8a7d07cd03b2fb0bbb76f7a9d7c)
1 /*-
2  * Copyright (c) 2015 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 "en.h"
29 #include <machine/in_cksum.h>
30 
31 static inline int
32 mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq,
33     struct mlx5e_rx_wqe *wqe, u16 ix)
34 {
35 	bus_dma_segment_t segs[1];
36 	struct mbuf *mb;
37 	int nsegs;
38 	int err;
39 
40 	if (rq->mbuf[ix].mbuf != NULL)
41 		return (0);
42 
43 	mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rq->wqe_sz);
44 	if (unlikely(!mb))
45 		return (-ENOMEM);
46 
47 	/* set initial mbuf length */
48 	mb->m_pkthdr.len = mb->m_len = rq->wqe_sz;
49 
50 	/* get IP header aligned */
51 	m_adj(mb, MLX5E_NET_IP_ALIGN);
52 
53 	err = -bus_dmamap_load_mbuf_sg(rq->dma_tag, rq->mbuf[ix].dma_map,
54 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
55 	if (err != 0)
56 		goto err_free_mbuf;
57 	if (unlikely(nsegs != 1)) {
58 		bus_dmamap_unload(rq->dma_tag, rq->mbuf[ix].dma_map);
59 		err = -ENOMEM;
60 		goto err_free_mbuf;
61 	}
62 	wqe->data.addr = cpu_to_be64(segs[0].ds_addr);
63 
64 	rq->mbuf[ix].mbuf = mb;
65 	rq->mbuf[ix].data = mb->m_data;
66 
67 	bus_dmamap_sync(rq->dma_tag, rq->mbuf[ix].dma_map,
68 	    BUS_DMASYNC_PREREAD);
69 	return (0);
70 
71 err_free_mbuf:
72 	m_freem(mb);
73 	return (err);
74 }
75 
76 static void
77 mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
78 {
79 	if (unlikely(rq->enabled == 0))
80 		return;
81 
82 	while (!mlx5_wq_ll_is_full(&rq->wq)) {
83 		struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, rq->wq.head);
84 
85 		if (unlikely(mlx5e_alloc_rx_wqe(rq, wqe, rq->wq.head)))
86 			break;
87 
88 		mlx5_wq_ll_push(&rq->wq, be16_to_cpu(wqe->next.next_wqe_index));
89 	}
90 
91 	/* ensure wqes are visible to device before updating doorbell record */
92 	wmb();
93 
94 	mlx5_wq_ll_update_db_record(&rq->wq);
95 }
96 
97 static void
98 mlx5e_lro_update_hdr(struct mbuf *mb, struct mlx5_cqe64 *cqe)
99 {
100 	/* TODO: consider vlans, ip options, ... */
101 	struct ether_header *eh;
102 	uint16_t eh_type;
103 	struct ip6_hdr *ip6 = NULL;
104 	struct ip *ip4 = NULL;
105 	struct tcphdr *th;
106 	uint32_t *ts_ptr;
107 
108 	eh = mtod(mb, struct ether_header *);
109 	eh_type = ntohs(eh->ether_type);
110 
111 	u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
112 	int tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA == l4_hdr_type) ||
113 	    (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
114 
115 	/* TODO: consider vlan */
116 	u16 tot_len = be32_to_cpu(cqe->byte_cnt) - ETHER_HDR_LEN;
117 
118 	switch (eh_type) {
119 	case ETHERTYPE_IP:
120 		ip4 = (struct ip *)(eh + 1);
121 		th = (struct tcphdr *)(ip4 + 1);
122 		break;
123 	case ETHERTYPE_IPV6:
124 		ip6 = (struct ip6_hdr *)(eh + 1);
125 		th = (struct tcphdr *)(ip6 + 1);
126 		break;
127 	default:
128 		return;
129 	}
130 
131 	ts_ptr = (uint32_t *)(th + 1);
132 
133 	if (get_cqe_lro_tcppsh(cqe))
134 		th->th_flags |= TH_PUSH;
135 
136 	if (tcp_ack) {
137 		th->th_flags |= TH_ACK;
138 		th->th_ack = cqe->lro_ack_seq_num;
139 		th->th_win = cqe->lro_tcp_win;
140 
141 		/*
142 		 * FreeBSD handles only 32bit aligned timestamp right after
143 		 * the TCP hdr
144 		 * +--------+--------+--------+--------+
145 		 * |   NOP  |  NOP   |  TSopt |   10   |
146 		 * +--------+--------+--------+--------+
147 		 * |          TSval   timestamp        |
148 		 * +--------+--------+--------+--------+
149 		 * |          TSecr   timestamp        |
150 		 * +--------+--------+--------+--------+
151 		 */
152 		if (get_cqe_lro_timestamp_valid(cqe) &&
153 		    (__predict_true(*ts_ptr) == ntohl(TCPOPT_NOP << 24 |
154 		    TCPOPT_NOP << 16 | TCPOPT_TIMESTAMP << 8 |
155 		    TCPOLEN_TIMESTAMP))) {
156 			/*
157 			 * cqe->timestamp is 64bit long.
158 			 * [0-31] - timestamp.
159 			 * [32-64] - timestamp echo replay.
160 			 */
161 			ts_ptr[1] = *(uint32_t *)&cqe->timestamp;
162 			ts_ptr[2] = *((uint32_t *)&cqe->timestamp + 1);
163 		}
164 	}
165 	if (ip4) {
166 		ip4->ip_ttl = cqe->lro_min_ttl;
167 		ip4->ip_len = cpu_to_be16(tot_len);
168 		ip4->ip_sum = 0;
169 		ip4->ip_sum = in_cksum(mb, ip4->ip_hl << 2);
170 	} else {
171 		ip6->ip6_hlim = cqe->lro_min_ttl;
172 		ip6->ip6_plen = cpu_to_be16(tot_len -
173 		    sizeof(struct ip6_hdr));
174 	}
175 	/* TODO: handle tcp checksum */
176 }
177 
178 static inline void
179 mlx5e_build_rx_mbuf(struct mlx5_cqe64 *cqe,
180     struct mlx5e_rq *rq, struct mbuf *mb,
181     u32 cqe_bcnt)
182 {
183 	struct ifnet *ifp = rq->ifp;
184 	int lro_num_seg;	/* HW LRO session aggregated packets counter */
185 
186 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
187 	if (lro_num_seg > 1) {
188 		mlx5e_lro_update_hdr(mb, cqe);
189 		rq->stats.lro_packets++;
190 		rq->stats.lro_bytes += cqe_bcnt;
191 	}
192 
193 	mb->m_pkthdr.len = mb->m_len = cqe_bcnt;
194 	/* check if a Toeplitz hash was computed */
195 	if (cqe->rss_hash_type != 0) {
196 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->rss_hash_result);
197 #ifdef RSS
198 		/* decode the RSS hash type */
199 		switch (cqe->rss_hash_type &
200 		    (CQE_RSS_DST_HTYPE_L4 | CQE_RSS_DST_HTYPE_IP)) {
201 		/* IPv4 */
202 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV4):
203 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV4);
204 			break;
205 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV4):
206 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV4);
207 			break;
208 		case CQE_RSS_DST_HTYPE_IPV4:
209 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV4);
210 			break;
211 		/* IPv6 */
212 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV6):
213 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV6);
214 			break;
215 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV6):
216 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV6);
217 			break;
218 		case CQE_RSS_DST_HTYPE_IPV6:
219 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV6);
220 			break;
221 		default:	/* Other */
222 			M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
223 			break;
224 		}
225 #else
226 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
227 #endif
228 	} else {
229 		mb->m_pkthdr.flowid = rq->ix;
230 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
231 	}
232 	mb->m_pkthdr.rcvif = ifp;
233 
234 	if (likely(ifp->if_capenable & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) &&
235 	    ((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK)) ==
236 	    (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK))) {
237 		mb->m_pkthdr.csum_flags =
238 		    CSUM_IP_CHECKED | CSUM_IP_VALID |
239 		    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
240 		mb->m_pkthdr.csum_data = htons(0xffff);
241 	} else {
242 		rq->stats.csum_none++;
243 	}
244 
245 	if (cqe_has_vlan(cqe)) {
246 		mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->vlan_info);
247 		mb->m_flags |= M_VLANTAG;
248 	}
249 }
250 
251 static inline void
252 mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
253 {
254 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, (cc & cq->wq.sz_m1)),
255 	    sizeof(struct mlx5_cqe64));
256 }
257 
258 static inline void
259 mlx5e_write_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
260 {
261 	memcpy(mlx5_cqwq_get_wqe(&cq->wq, cc & cq->wq.sz_m1),
262 	    data, sizeof(struct mlx5_cqe64));
263 }
264 
265 static inline void
266 mlx5e_decompress_cqe(struct mlx5e_cq *cq, struct mlx5_cqe64 *title,
267     struct mlx5_mini_cqe8 *mini,
268     u16 wqe_counter, int i)
269 {
270 	title->byte_cnt = mini->byte_cnt;
271 	title->wqe_counter = cpu_to_be16((wqe_counter + i) & cq->wq.sz_m1);
272 	title->check_sum = mini->checksum;
273 	title->op_own = (title->op_own & 0xf0) |
274 	    (((cq->wq.cc + i) >> cq->wq.log_sz) & 1);
275 }
276 
277 #define MLX5E_MINI_ARRAY_SZ 8
278 /* Make sure structs are not packet differently */
279 CTASSERT(sizeof(struct mlx5_cqe64) ==
280     sizeof(struct mlx5_mini_cqe8) * MLX5E_MINI_ARRAY_SZ);
281 static void
282 mlx5e_decompress_cqes(struct mlx5e_cq *cq)
283 {
284 	struct mlx5_mini_cqe8 mini_array[MLX5E_MINI_ARRAY_SZ];
285 	struct mlx5_cqe64 title;
286 	u32 cqe_count;
287 	u32 i = 0;
288 	u16 title_wqe_counter;
289 
290 	mlx5e_read_cqe_slot(cq, cq->wq.cc, &title);
291 	title_wqe_counter = be16_to_cpu(title.wqe_counter);
292 	cqe_count = be32_to_cpu(title.byte_cnt);
293 
294 	/* Make sure we won't overflow */
295 	KASSERT(cqe_count <= cq->wq.sz_m1,
296 	    ("%s: cqe_count %u > cq->wq.sz_m1 %u", __func__,
297 	    cqe_count, cq->wq.sz_m1));
298 
299 	mlx5e_read_cqe_slot(cq, cq->wq.cc + 1, mini_array);
300 	while (true) {
301 		mlx5e_decompress_cqe(cq, &title,
302 		    &mini_array[i % MLX5E_MINI_ARRAY_SZ],
303 		    title_wqe_counter, i);
304 		mlx5e_write_cqe_slot(cq, cq->wq.cc + i, &title);
305 		i++;
306 
307 		if (i == cqe_count)
308 			break;
309 		if (i % MLX5E_MINI_ARRAY_SZ == 0)
310 			mlx5e_read_cqe_slot(cq, cq->wq.cc + i, mini_array);
311 	}
312 }
313 
314 static int
315 mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
316 {
317 #ifndef HAVE_TURBO_LRO
318 	struct lro_entry *queued;
319 #endif
320 	int i;
321 
322 	for (i = 0; i < budget; i++) {
323 		struct mlx5e_rx_wqe *wqe;
324 		struct mlx5_cqe64 *cqe;
325 		struct mbuf *mb;
326 		__be16 wqe_counter_be;
327 		u16 wqe_counter;
328 		u32 byte_cnt;
329 
330 		cqe = mlx5e_get_cqe(&rq->cq);
331 		if (!cqe)
332 			break;
333 
334 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED)
335 			mlx5e_decompress_cqes(&rq->cq);
336 
337 		mlx5_cqwq_pop(&rq->cq.wq);
338 
339 		wqe_counter_be = cqe->wqe_counter;
340 		wqe_counter = be16_to_cpu(wqe_counter_be);
341 		wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
342 		byte_cnt = be32_to_cpu(cqe->byte_cnt);
343 
344 		bus_dmamap_sync(rq->dma_tag,
345 		    rq->mbuf[wqe_counter].dma_map,
346 		    BUS_DMASYNC_POSTREAD);
347 
348 		if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
349 			rq->stats.wqe_err++;
350 			goto wq_ll_pop;
351 		}
352 
353 		if (MHLEN >= byte_cnt &&
354 		    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
355 			bcopy(rq->mbuf[wqe_counter].data, mtod(mb, caddr_t),
356 			    byte_cnt);
357 		} else {
358 			mb = rq->mbuf[wqe_counter].mbuf;
359 			rq->mbuf[wqe_counter].mbuf = NULL;	/* safety clear */
360 
361 			bus_dmamap_unload(rq->dma_tag,
362 			    rq->mbuf[wqe_counter].dma_map);
363 		}
364 
365 		mlx5e_build_rx_mbuf(cqe, rq, mb, byte_cnt);
366 		rq->stats.packets++;
367 #ifdef HAVE_TURBO_LRO
368 		if (mb->m_pkthdr.csum_flags == 0 ||
369 		    (rq->ifp->if_capenable & IFCAP_LRO) == 0 ||
370 		    rq->lro.mbuf == NULL) {
371 			/* normal input */
372 			rq->ifp->if_input(rq->ifp, mb);
373 		} else {
374 			tcp_tlro_rx(&rq->lro, mb);
375 		}
376 #else
377 		if (mb->m_pkthdr.csum_flags == 0 ||
378 		    (rq->ifp->if_capenable & IFCAP_LRO) == 0 ||
379 		    rq->lro.lro_cnt == 0 ||
380 		    tcp_lro_rx(&rq->lro, mb, 0) != 0) {
381 			rq->ifp->if_input(rq->ifp, mb);
382 		}
383 #endif
384 wq_ll_pop:
385 		mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
386 		    &wqe->next.next_wqe_index);
387 	}
388 
389 	mlx5_cqwq_update_db_record(&rq->cq.wq);
390 
391 	/* ensure cq space is freed before enabling more cqes */
392 	wmb();
393 #ifndef HAVE_TURBO_LRO
394 	while ((queued = SLIST_FIRST(&rq->lro.lro_active)) != NULL) {
395 		SLIST_REMOVE_HEAD(&rq->lro.lro_active, next);
396 		tcp_lro_flush(&rq->lro, queued);
397 	}
398 #endif
399 	return (i);
400 }
401 
402 void
403 mlx5e_rx_cq_comp(struct mlx5_core_cq *mcq)
404 {
405 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
406 	int i = 0;
407 
408 #ifdef HAVE_PER_CQ_EVENT_PACKET
409 	struct mbuf *mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rq->wqe_sz);
410 
411 	if (mb != NULL) {
412 		/* this code is used for debugging purpose only */
413 		mb->m_pkthdr.len = mb->m_len = 15;
414 		memset(mb->m_data, 255, 14);
415 		mb->m_data[14] = rq->ix;
416 		mb->m_pkthdr.rcvif = rq->ifp;
417 		rq->ifp->if_input(rq->ifp, mb);
418 	}
419 #endif
420 
421 	mtx_lock(&rq->mtx);
422 
423 	/*
424 	 * Polling the entire CQ without posting new WQEs results in
425 	 * lack of receive WQEs during heavy traffic scenarios.
426 	 */
427 	while (1) {
428 		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
429 		    MLX5E_RX_BUDGET_MAX)
430 			break;
431 		i += MLX5E_RX_BUDGET_MAX;
432 		if (i >= MLX5E_BUDGET_MAX)
433 			break;
434 		mlx5e_post_rx_wqes(rq);
435 	}
436 	mlx5e_post_rx_wqes(rq);
437 	mlx5e_cq_arm(&rq->cq);
438 #ifdef HAVE_TURBO_LRO
439 	tcp_tlro_flush(&rq->lro, 1);
440 #endif
441 	mtx_unlock(&rq->mtx);
442 }
443