xref: /freebsd/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c (revision 63a938566d524836885917d95bd491aa4400b181)
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 	atomic_thread_fence_rel();
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 uint64_t
183 mlx5e_mbuf_tstmp(struct mlx5e_priv *priv, uint64_t hw_tstmp)
184 {
185 	struct mlx5e_clbr_point *cp;
186 	uint64_t a1, a2, res;
187 	u_int gen;
188 
189 	do {
190 		cp = &priv->clbr_points[priv->clbr_curr];
191 		gen = atomic_load_acq_int(&cp->clbr_gen);
192 		a1 = (hw_tstmp - cp->clbr_hw_prev) >> MLX5E_TSTMP_PREC;
193 		a2 = (cp->base_curr - cp->base_prev) >> MLX5E_TSTMP_PREC;
194 		res = (a1 * a2) << MLX5E_TSTMP_PREC;
195 
196 		/*
197 		 * Divisor cannot be zero because calibration callback
198 		 * checks for the condition and disables timestamping
199 		 * if clock halted.
200 		 */
201 		res /= (cp->clbr_hw_curr - cp->clbr_hw_prev) >>
202 		    MLX5E_TSTMP_PREC;
203 
204 		res += cp->base_prev;
205 		atomic_thread_fence_acq();
206 	} while (gen == 0 || gen != cp->clbr_gen);
207 	return (res);
208 }
209 
210 static inline void
211 mlx5e_build_rx_mbuf(struct mlx5_cqe64 *cqe,
212     struct mlx5e_rq *rq, struct mbuf *mb,
213     u32 cqe_bcnt)
214 {
215 	struct ifnet *ifp = rq->ifp;
216 	struct mlx5e_channel *c;
217 	int lro_num_seg;	/* HW LRO session aggregated packets counter */
218 	uint64_t tstmp;
219 
220 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
221 	if (lro_num_seg > 1) {
222 		mlx5e_lro_update_hdr(mb, cqe);
223 		rq->stats.lro_packets++;
224 		rq->stats.lro_bytes += cqe_bcnt;
225 	}
226 
227 	mb->m_pkthdr.len = mb->m_len = cqe_bcnt;
228 	/* check if a Toeplitz hash was computed */
229 	if (cqe->rss_hash_type != 0) {
230 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->rss_hash_result);
231 #ifdef RSS
232 		/* decode the RSS hash type */
233 		switch (cqe->rss_hash_type &
234 		    (CQE_RSS_DST_HTYPE_L4 | CQE_RSS_DST_HTYPE_IP)) {
235 		/* IPv4 */
236 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV4):
237 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV4);
238 			break;
239 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV4):
240 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV4);
241 			break;
242 		case CQE_RSS_DST_HTYPE_IPV4:
243 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV4);
244 			break;
245 		/* IPv6 */
246 		case (CQE_RSS_DST_HTYPE_TCP | CQE_RSS_DST_HTYPE_IPV6):
247 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_TCP_IPV6);
248 			break;
249 		case (CQE_RSS_DST_HTYPE_UDP | CQE_RSS_DST_HTYPE_IPV6):
250 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_UDP_IPV6);
251 			break;
252 		case CQE_RSS_DST_HTYPE_IPV6:
253 			M_HASHTYPE_SET(mb, M_HASHTYPE_RSS_IPV6);
254 			break;
255 		default:	/* Other */
256 			M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
257 			break;
258 		}
259 #else
260 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
261 #endif
262 	} else {
263 		mb->m_pkthdr.flowid = rq->ix;
264 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE);
265 	}
266 	mb->m_pkthdr.rcvif = ifp;
267 
268 	if (likely(ifp->if_capenable & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) &&
269 	    ((cqe->hds_ip_ext & (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK)) ==
270 	    (CQE_L2_OK | CQE_L3_OK | CQE_L4_OK))) {
271 		mb->m_pkthdr.csum_flags =
272 		    CSUM_IP_CHECKED | CSUM_IP_VALID |
273 		    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
274 		mb->m_pkthdr.csum_data = htons(0xffff);
275 	} else {
276 		rq->stats.csum_none++;
277 	}
278 
279 	if (cqe_has_vlan(cqe)) {
280 		mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->vlan_info);
281 		mb->m_flags |= M_VLANTAG;
282 	}
283 
284 	c = container_of(rq, struct mlx5e_channel, rq);
285 	if (c->priv->clbr_done >= 2) {
286 		tstmp = mlx5e_mbuf_tstmp(c->priv, be64_to_cpu(cqe->timestamp));
287 		if ((tstmp & MLX5_CQE_TSTMP_PTP) != 0) {
288 			/*
289 			 * Timestamp was taken on the packet entrance,
290 			 * instead of the cqe generation.
291 			 */
292 			tstmp &= ~MLX5_CQE_TSTMP_PTP;
293 			mb->m_flags |= M_TSTMP_HPREC;
294 		}
295 		mb->m_pkthdr.rcv_tstmp = tstmp;
296 		mb->m_flags |= M_TSTMP;
297 	}
298 }
299 
300 static inline void
301 mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
302 {
303 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, (cc & cq->wq.sz_m1)),
304 	    sizeof(struct mlx5_cqe64));
305 }
306 
307 static inline void
308 mlx5e_write_cqe_slot(struct mlx5e_cq *cq, u32 cc, void *data)
309 {
310 	memcpy(mlx5_cqwq_get_wqe(&cq->wq, cc & cq->wq.sz_m1),
311 	    data, sizeof(struct mlx5_cqe64));
312 }
313 
314 static inline void
315 mlx5e_decompress_cqe(struct mlx5e_cq *cq, struct mlx5_cqe64 *title,
316     struct mlx5_mini_cqe8 *mini,
317     u16 wqe_counter, int i)
318 {
319 	/*
320 	 * NOTE: The fields which are not set here are copied from the
321 	 * initial and common title. See memcpy() in
322 	 * mlx5e_write_cqe_slot().
323 	 */
324 	title->byte_cnt = mini->byte_cnt;
325 	title->wqe_counter = cpu_to_be16((wqe_counter + i) & cq->wq.sz_m1);
326 	title->check_sum = mini->checksum;
327 	title->op_own = (title->op_own & 0xf0) |
328 	    (((cq->wq.cc + i) >> cq->wq.log_sz) & 1);
329 }
330 
331 #define MLX5E_MINI_ARRAY_SZ 8
332 /* Make sure structs are not packet differently */
333 CTASSERT(sizeof(struct mlx5_cqe64) ==
334     sizeof(struct mlx5_mini_cqe8) * MLX5E_MINI_ARRAY_SZ);
335 static void
336 mlx5e_decompress_cqes(struct mlx5e_cq *cq)
337 {
338 	struct mlx5_mini_cqe8 mini_array[MLX5E_MINI_ARRAY_SZ];
339 	struct mlx5_cqe64 title;
340 	u32 cqe_count;
341 	u32 i = 0;
342 	u16 title_wqe_counter;
343 
344 	mlx5e_read_cqe_slot(cq, cq->wq.cc, &title);
345 	title_wqe_counter = be16_to_cpu(title.wqe_counter);
346 	cqe_count = be32_to_cpu(title.byte_cnt);
347 
348 	/* Make sure we won't overflow */
349 	KASSERT(cqe_count <= cq->wq.sz_m1,
350 	    ("%s: cqe_count %u > cq->wq.sz_m1 %u", __func__,
351 	    cqe_count, cq->wq.sz_m1));
352 
353 	mlx5e_read_cqe_slot(cq, cq->wq.cc + 1, mini_array);
354 	while (true) {
355 		mlx5e_decompress_cqe(cq, &title,
356 		    &mini_array[i % MLX5E_MINI_ARRAY_SZ],
357 		    title_wqe_counter, i);
358 		mlx5e_write_cqe_slot(cq, cq->wq.cc + i, &title);
359 		i++;
360 
361 		if (i == cqe_count)
362 			break;
363 		if (i % MLX5E_MINI_ARRAY_SZ == 0)
364 			mlx5e_read_cqe_slot(cq, cq->wq.cc + i, mini_array);
365 	}
366 }
367 
368 static int
369 mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
370 {
371 	int i;
372 
373 	for (i = 0; i < budget; i++) {
374 		struct mlx5e_rx_wqe *wqe;
375 		struct mlx5_cqe64 *cqe;
376 		struct mbuf *mb;
377 		__be16 wqe_counter_be;
378 		u16 wqe_counter;
379 		u32 byte_cnt;
380 
381 		cqe = mlx5e_get_cqe(&rq->cq);
382 		if (!cqe)
383 			break;
384 
385 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED)
386 			mlx5e_decompress_cqes(&rq->cq);
387 
388 		mlx5_cqwq_pop(&rq->cq.wq);
389 
390 		wqe_counter_be = cqe->wqe_counter;
391 		wqe_counter = be16_to_cpu(wqe_counter_be);
392 		wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
393 		byte_cnt = be32_to_cpu(cqe->byte_cnt);
394 
395 		bus_dmamap_sync(rq->dma_tag,
396 		    rq->mbuf[wqe_counter].dma_map,
397 		    BUS_DMASYNC_POSTREAD);
398 
399 		if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
400 			rq->stats.wqe_err++;
401 			goto wq_ll_pop;
402 		}
403 		if ((MHLEN - MLX5E_NET_IP_ALIGN) >= byte_cnt &&
404 		    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
405 			/* get IP header aligned */
406 			mb->m_data += MLX5E_NET_IP_ALIGN;
407 
408 			bcopy(rq->mbuf[wqe_counter].data, mtod(mb, caddr_t),
409 			    byte_cnt);
410 		} else {
411 			mb = rq->mbuf[wqe_counter].mbuf;
412 			rq->mbuf[wqe_counter].mbuf = NULL;	/* safety clear */
413 
414 			bus_dmamap_unload(rq->dma_tag,
415 			    rq->mbuf[wqe_counter].dma_map);
416 		}
417 
418 		mlx5e_build_rx_mbuf(cqe, rq, mb, byte_cnt);
419 		rq->stats.packets++;
420 
421 #if !defined(HAVE_TCP_LRO_RX)
422 		tcp_lro_queue_mbuf(&rq->lro, mb);
423 #else
424 		if (mb->m_pkthdr.csum_flags == 0 ||
425 		    (rq->ifp->if_capenable & IFCAP_LRO) == 0 ||
426 		    rq->lro.lro_cnt == 0 ||
427 		    tcp_lro_rx(&rq->lro, mb, 0) != 0) {
428 			rq->ifp->if_input(rq->ifp, mb);
429 		}
430 #endif
431 wq_ll_pop:
432 		mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
433 		    &wqe->next.next_wqe_index);
434 	}
435 
436 	mlx5_cqwq_update_db_record(&rq->cq.wq);
437 
438 	/* ensure cq space is freed before enabling more cqes */
439 	atomic_thread_fence_rel();
440 	return (i);
441 }
442 
443 void
444 mlx5e_rx_cq_comp(struct mlx5_core_cq *mcq)
445 {
446 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
447 	int i = 0;
448 
449 #ifdef HAVE_PER_CQ_EVENT_PACKET
450 	struct mbuf *mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rq->wqe_sz);
451 
452 	if (mb != NULL) {
453 		/* this code is used for debugging purpose only */
454 		mb->m_pkthdr.len = mb->m_len = 15;
455 		memset(mb->m_data, 255, 14);
456 		mb->m_data[14] = rq->ix;
457 		mb->m_pkthdr.rcvif = rq->ifp;
458 		rq->ifp->if_input(rq->ifp, mb);
459 	}
460 #endif
461 
462 	mtx_lock(&rq->mtx);
463 
464 	/*
465 	 * Polling the entire CQ without posting new WQEs results in
466 	 * lack of receive WQEs during heavy traffic scenarios.
467 	 */
468 	while (1) {
469 		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
470 		    MLX5E_RX_BUDGET_MAX)
471 			break;
472 		i += MLX5E_RX_BUDGET_MAX;
473 		if (i >= MLX5E_BUDGET_MAX)
474 			break;
475 		mlx5e_post_rx_wqes(rq);
476 	}
477 	mlx5e_post_rx_wqes(rq);
478 	mlx5e_cq_arm(&rq->cq, MLX5_GET_DOORBELL_LOCK(&rq->channel->priv->doorbell_lock));
479 	tcp_lro_flush_all(&rq->lro);
480 	mtx_unlock(&rq->mtx);
481 }
482