xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
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 			callout_reset_curcpu(&rq->watchdog, 1, (void *)&mlx5e_post_rx_wqes, rq);
87 			break;
88 		}
89 		mlx5_wq_ll_push(&rq->wq, be16_to_cpu(wqe->next.next_wqe_index));
90 	}
91 
92 	/* ensure wqes are visible to device before updating doorbell record */
93 	wmb();
94 
95 	mlx5_wq_ll_update_db_record(&rq->wq);
96 }
97 
98 static void
99 mlx5e_lro_update_hdr(struct mbuf *mb, struct mlx5_cqe64 *cqe)
100 {
101 	/* TODO: consider vlans, ip options, ... */
102 	struct ether_header *eh;
103 	uint16_t eh_type;
104 	uint16_t tot_len;
105 	struct ip6_hdr *ip6 = NULL;
106 	struct ip *ip4 = NULL;
107 	struct tcphdr *th;
108 	uint32_t *ts_ptr;
109 	uint8_t l4_hdr_type;
110 	int tcp_ack;
111 
112 	eh = mtod(mb, struct ether_header *);
113 	eh_type = ntohs(eh->ether_type);
114 
115 	l4_hdr_type = get_cqe_l4_hdr_type(cqe);
116 	tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA == l4_hdr_type) ||
117 	    (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
118 
119 	/* TODO: consider vlan */
120 	tot_len = be32_to_cpu(cqe->byte_cnt) - ETHER_HDR_LEN;
121 
122 	switch (eh_type) {
123 	case ETHERTYPE_IP:
124 		ip4 = (struct ip *)(eh + 1);
125 		th = (struct tcphdr *)(ip4 + 1);
126 		break;
127 	case ETHERTYPE_IPV6:
128 		ip6 = (struct ip6_hdr *)(eh + 1);
129 		th = (struct tcphdr *)(ip6 + 1);
130 		break;
131 	default:
132 		return;
133 	}
134 
135 	ts_ptr = (uint32_t *)(th + 1);
136 
137 	if (get_cqe_lro_tcppsh(cqe))
138 		th->th_flags |= TH_PUSH;
139 
140 	if (tcp_ack) {
141 		th->th_flags |= TH_ACK;
142 		th->th_ack = cqe->lro_ack_seq_num;
143 		th->th_win = cqe->lro_tcp_win;
144 
145 		/*
146 		 * FreeBSD handles only 32bit aligned timestamp right after
147 		 * the TCP hdr
148 		 * +--------+--------+--------+--------+
149 		 * |   NOP  |  NOP   |  TSopt |   10   |
150 		 * +--------+--------+--------+--------+
151 		 * |          TSval   timestamp        |
152 		 * +--------+--------+--------+--------+
153 		 * |          TSecr   timestamp        |
154 		 * +--------+--------+--------+--------+
155 		 */
156 		if (get_cqe_lro_timestamp_valid(cqe) &&
157 		    (__predict_true(*ts_ptr) == ntohl(TCPOPT_NOP << 24 |
158 		    TCPOPT_NOP << 16 | TCPOPT_TIMESTAMP << 8 |
159 		    TCPOLEN_TIMESTAMP))) {
160 			/*
161 			 * cqe->timestamp is 64bit long.
162 			 * [0-31] - timestamp.
163 			 * [32-64] - timestamp echo replay.
164 			 */
165 			ts_ptr[1] = *(uint32_t *)&cqe->timestamp;
166 			ts_ptr[2] = *((uint32_t *)&cqe->timestamp + 1);
167 		}
168 	}
169 	if (ip4) {
170 		ip4->ip_ttl = cqe->lro_min_ttl;
171 		ip4->ip_len = cpu_to_be16(tot_len);
172 		ip4->ip_sum = 0;
173 		ip4->ip_sum = in_cksum(mb, ip4->ip_hl << 2);
174 	} else {
175 		ip6->ip6_hlim = cqe->lro_min_ttl;
176 		ip6->ip6_plen = cpu_to_be16(tot_len -
177 		    sizeof(struct ip6_hdr));
178 	}
179 	/* TODO: handle tcp checksum */
180 }
181 
182 static inline void
183 mlx5e_build_rx_mbuf(struct mlx5_cqe64 *cqe,
184     struct mlx5e_rq *rq, struct mbuf *mb,
185     u32 cqe_bcnt)
186 {
187 	struct ifnet *ifp = rq->ifp;
188 	int lro_num_seg;	/* HW LRO session aggregated packets counter */
189 
190 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
191 	if (lro_num_seg > 1) {
192 		mlx5e_lro_update_hdr(mb, cqe);
193 		rq->stats.lro_packets++;
194 		rq->stats.lro_bytes += cqe_bcnt;
195 	}
196 
197 	mb->m_pkthdr.len = mb->m_len = cqe_bcnt;
198 	/* check if a Toeplitz hash was computed */
199 	if (cqe->rss_hash_type != 0) {
200 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->rss_hash_result);
201 #ifdef RSS
202 		/* decode the RSS hash type */
203 		switch (cqe->rss_hash_type &
204 		    (CQE_RSS_DST_HTYPE_L4 | CQE_RSS_DST_HTYPE_IP)) {
205 		/* IPv4 */
206 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV4):
207 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV4);
208 			break;
209 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV4):
210 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV4);
211 			break;
212 		case CQE_RSS_DST_HTYPE_IPV4:
213 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV4);
214 			break;
215 		/* IPv6 */
216 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV6):
217 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV6);
218 			break;
219 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV6):
220 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV6);
221 			break;
222 		case CQE_RSS_DST_HTYPE_IPV6:
223 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV6);
224 			break;
225 		default:	/* Other */
226 			M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
227 			break;
228 		}
229 #else
230 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
231 #endif
232 	} else {
233 		mb->m_pkthdr.flowid = rq->ix;
234 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
235 	}
236 	mb->m_pkthdr.rcvif = ifp;
237 
238 	if (likely(ifp->if_capenable & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) &&
239 	    ((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK)) ==
240 	    (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK))) {
241 		mb->m_pkthdr.csum_flags =
242 		    CSUM_IP_CHECKED | CSUM_IP_VALID |
243 		    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
244 		mb->m_pkthdr.csum_data = htons(0xffff);
245 	} else {
246 		rq->stats.csum_none++;
247 	}
248 
249 	if (cqe_has_vlan(cqe)) {
250 		mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->vlan_info);
251 		mb->m_flags |= M_VLANTAG;
252 	}
253 }
254 
255 static inline void
256 mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
257 {
258 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, (cc & cq->wq.sz_m1)),
259 	    sizeof(struct mlx5_cqe64));
260 }
261 
262 static inline void
263 mlx5e_write_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
264 {
265 	memcpy(mlx5_cqwq_get_wqe(&cq->wq, cc & cq->wq.sz_m1),
266 	    data, sizeof(struct mlx5_cqe64));
267 }
268 
269 static inline void
270 mlx5e_decompress_cqe(struct mlx5e_cq *cq, struct mlx5_cqe64 *title,
271     struct mlx5_mini_cqe8 *mini,
272     u16 wqe_counter, int i)
273 {
274 	/*
275 	 * NOTE: The fields which are not set here are copied from the
276 	 * initial and common title. See memcpy() in
277 	 * mlx5e_write_cqe_slot().
278 	 */
279 	title->byte_cnt = mini->byte_cnt;
280 	title->wqe_counter = cpu_to_be16((wqe_counter + i) & cq->wq.sz_m1);
281 	title->check_sum = mini->checksum;
282 	title->op_own = (title->op_own & 0xf0) |
283 	    (((cq->wq.cc + i) >> cq->wq.log_sz) & 1);
284 }
285 
286 #define MLX5E_MINI_ARRAY_SZ 8
287 /* Make sure structs are not packet differently */
288 CTASSERT(sizeof(struct mlx5_cqe64) ==
289     sizeof(struct mlx5_mini_cqe8) * MLX5E_MINI_ARRAY_SZ);
290 static void
291 mlx5e_decompress_cqes(struct mlx5e_cq *cq)
292 {
293 	struct mlx5_mini_cqe8 mini_array[MLX5E_MINI_ARRAY_SZ];
294 	struct mlx5_cqe64 title;
295 	u32 cqe_count;
296 	u32 i = 0;
297 	u16 title_wqe_counter;
298 
299 	mlx5e_read_cqe_slot(cq, cq->wq.cc, &title);
300 	title_wqe_counter = be16_to_cpu(title.wqe_counter);
301 	cqe_count = be32_to_cpu(title.byte_cnt);
302 
303 	/* Make sure we won't overflow */
304 	KASSERT(cqe_count <= cq->wq.sz_m1,
305 	    ("%s: cqe_count %u > cq->wq.sz_m1 %u", __func__,
306 	    cqe_count, cq->wq.sz_m1));
307 
308 	mlx5e_read_cqe_slot(cq, cq->wq.cc + 1, mini_array);
309 	while (true) {
310 		mlx5e_decompress_cqe(cq, &title,
311 		    &mini_array[i % MLX5E_MINI_ARRAY_SZ],
312 		    title_wqe_counter, i);
313 		mlx5e_write_cqe_slot(cq, cq->wq.cc + i, &title);
314 		i++;
315 
316 		if (i == cqe_count)
317 			break;
318 		if (i % MLX5E_MINI_ARRAY_SZ == 0)
319 			mlx5e_read_cqe_slot(cq, cq->wq.cc + i, mini_array);
320 	}
321 }
322 
323 static int
324 mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
325 {
326 	int i;
327 
328 	for (i = 0; i < budget; i++) {
329 		struct mlx5e_rx_wqe *wqe;
330 		struct mlx5_cqe64 *cqe;
331 		struct mbuf *mb;
332 		__be16 wqe_counter_be;
333 		u16 wqe_counter;
334 		u32 byte_cnt;
335 
336 		cqe = mlx5e_get_cqe(&rq->cq);
337 		if (!cqe)
338 			break;
339 
340 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED)
341 			mlx5e_decompress_cqes(&rq->cq);
342 
343 		mlx5_cqwq_pop(&rq->cq.wq);
344 
345 		wqe_counter_be = cqe->wqe_counter;
346 		wqe_counter = be16_to_cpu(wqe_counter_be);
347 		wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
348 		byte_cnt = be32_to_cpu(cqe->byte_cnt);
349 
350 		bus_dmamap_sync(rq->dma_tag,
351 		    rq->mbuf[wqe_counter].dma_map,
352 		    BUS_DMASYNC_POSTREAD);
353 
354 		if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
355 			rq->stats.wqe_err++;
356 			goto wq_ll_pop;
357 		}
358 
359 		if (MHLEN >= byte_cnt &&
360 		    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
361 			bcopy(rq->mbuf[wqe_counter].data, mtod(mb, caddr_t),
362 			    byte_cnt);
363 		} else {
364 			mb = rq->mbuf[wqe_counter].mbuf;
365 			rq->mbuf[wqe_counter].mbuf = NULL;	/* safety clear */
366 
367 			bus_dmamap_unload(rq->dma_tag,
368 			    rq->mbuf[wqe_counter].dma_map);
369 		}
370 
371 		mlx5e_build_rx_mbuf(cqe, rq, mb, byte_cnt);
372 		rq->stats.packets++;
373 
374 #if !defined(HAVE_TCP_LRO_RX)
375 		tcp_lro_queue_mbuf(&rq->lro, mb);
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 	return (i);
394 }
395 
396 void
397 mlx5e_rx_cq_comp(struct mlx5_core_cq *mcq)
398 {
399 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
400 	int i = 0;
401 
402 #ifdef HAVE_PER_CQ_EVENT_PACKET
403 	struct mbuf *mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rq->wqe_sz);
404 
405 	if (mb != NULL) {
406 		/* this code is used for debugging purpose only */
407 		mb->m_pkthdr.len = mb->m_len = 15;
408 		memset(mb->m_data, 255, 14);
409 		mb->m_data[14] = rq->ix;
410 		mb->m_pkthdr.rcvif = rq->ifp;
411 		rq->ifp->if_input(rq->ifp, mb);
412 	}
413 #endif
414 
415 	mtx_lock(&rq->mtx);
416 
417 	/*
418 	 * Polling the entire CQ without posting new WQEs results in
419 	 * lack of receive WQEs during heavy traffic scenarios.
420 	 */
421 	while (1) {
422 		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
423 		    MLX5E_RX_BUDGET_MAX)
424 			break;
425 		i += MLX5E_RX_BUDGET_MAX;
426 		if (i >= MLX5E_BUDGET_MAX)
427 			break;
428 		mlx5e_post_rx_wqes(rq);
429 	}
430 	mlx5e_post_rx_wqes(rq);
431 	mlx5e_cq_arm(&rq->cq);
432 	tcp_lro_flush_all(&rq->lro);
433 	mtx_unlock(&rq->mtx);
434 }
435