xref: /freebsd/sys/dev/mlx4/mlx4_en/mlx4_en_rx.c (revision 41059135ce931c0f1014a999ffabc6bc470ce856)
1 /*
2  * Copyright (c) 2007, 2014 Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  */
33 #include "opt_inet.h"
34 #include <dev/mlx4/cq.h>
35 #include <linux/slab.h>
36 #include <dev/mlx4/qp.h>
37 #include <linux/if_ether.h>
38 #include <linux/if_vlan.h>
39 #include <linux/vmalloc.h>
40 #include <dev/mlx4/driver.h>
41 #ifdef CONFIG_NET_RX_BUSY_POLL
42 #include <net/busy_poll.h>
43 #endif
44 
45 #include "en.h"
46 
47 
48 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
49 				 struct mlx4_en_rx_ring *ring,
50 				 int index)
51 {
52 	struct mlx4_en_rx_desc *rx_desc = (struct mlx4_en_rx_desc *)
53 	    (ring->buf + (ring->stride * index));
54 	int possible_frags;
55 	int i;
56 
57 	/* Set size and memtype fields */
58 	rx_desc->data[0].byte_count = cpu_to_be32(priv->rx_mb_size - MLX4_NET_IP_ALIGN);
59 	rx_desc->data[0].lkey = cpu_to_be32(priv->mdev->mr.key);
60 
61 	/*
62 	 * If the number of used fragments does not fill up the ring
63 	 * stride, remaining (unused) fragments must be padded with
64 	 * null address/size and a special memory key:
65 	 */
66 	possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
67 	for (i = 1; i < possible_frags; i++) {
68 		rx_desc->data[i].byte_count = 0;
69 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
70 		rx_desc->data[i].addr = 0;
71 	}
72 }
73 
74 static int
75 mlx4_en_alloc_buf(struct mlx4_en_rx_ring *ring,
76      __be64 *pdma, struct mlx4_en_rx_mbuf *mb_list)
77 {
78 	bus_dma_segment_t segs[1];
79 	bus_dmamap_t map;
80 	struct mbuf *mb;
81 	int nsegs;
82 	int err;
83 
84 	/* try to allocate a new spare mbuf */
85 	if (unlikely(ring->spare.mbuf == NULL)) {
86 		mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, ring->rx_mb_size);
87 		if (unlikely(mb == NULL))
88 			return (-ENOMEM);
89 		/* setup correct length */
90 		mb->m_pkthdr.len = mb->m_len = ring->rx_mb_size;
91 
92 		/* make sure IP header gets aligned */
93 		m_adj(mb, MLX4_NET_IP_ALIGN);
94 
95 		/* load spare mbuf into BUSDMA */
96 		err = -bus_dmamap_load_mbuf_sg(ring->dma_tag, ring->spare.dma_map,
97 		    mb, segs, &nsegs, BUS_DMA_NOWAIT);
98 		if (unlikely(err != 0)) {
99 			m_freem(mb);
100 			return (err);
101 		}
102 
103 		/* store spare info */
104 		ring->spare.mbuf = mb;
105 		ring->spare.paddr_be = cpu_to_be64(segs[0].ds_addr);
106 
107 		bus_dmamap_sync(ring->dma_tag, ring->spare.dma_map,
108 		    BUS_DMASYNC_PREREAD);
109 	}
110 
111 	/* synchronize and unload the current mbuf, if any */
112 	if (likely(mb_list->mbuf != NULL)) {
113 		bus_dmamap_sync(ring->dma_tag, mb_list->dma_map,
114 		    BUS_DMASYNC_POSTREAD);
115 		bus_dmamap_unload(ring->dma_tag, mb_list->dma_map);
116 	}
117 
118 	mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, ring->rx_mb_size);
119 	if (unlikely(mb == NULL))
120 		goto use_spare;
121 
122 	/* setup correct length */
123 	mb->m_pkthdr.len = mb->m_len = ring->rx_mb_size;
124 
125 	/* make sure IP header gets aligned */
126 	m_adj(mb, MLX4_NET_IP_ALIGN);
127 
128 	err = -bus_dmamap_load_mbuf_sg(ring->dma_tag, mb_list->dma_map,
129 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
130 	if (unlikely(err != 0)) {
131 		m_freem(mb);
132 		goto use_spare;
133 	}
134 
135 	*pdma = cpu_to_be64(segs[0].ds_addr);
136 	mb_list->mbuf = mb;
137 
138 	bus_dmamap_sync(ring->dma_tag, mb_list->dma_map, BUS_DMASYNC_PREREAD);
139 	return (0);
140 
141 use_spare:
142 	/* swap DMA maps */
143 	map = mb_list->dma_map;
144 	mb_list->dma_map = ring->spare.dma_map;
145 	ring->spare.dma_map = map;
146 
147 	/* swap MBUFs */
148 	mb_list->mbuf = ring->spare.mbuf;
149 	ring->spare.mbuf = NULL;
150 
151 	/* store physical address */
152 	*pdma = ring->spare.paddr_be;
153 	return (0);
154 }
155 
156 static void
157 mlx4_en_free_buf(struct mlx4_en_rx_ring *ring, struct mlx4_en_rx_mbuf *mb_list)
158 {
159 	bus_dmamap_t map = mb_list->dma_map;
160 	bus_dmamap_sync(ring->dma_tag, map, BUS_DMASYNC_POSTREAD);
161 	bus_dmamap_unload(ring->dma_tag, map);
162 	m_freem(mb_list->mbuf);
163 	mb_list->mbuf = NULL;	/* safety clearing */
164 }
165 
166 static int
167 mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
168     struct mlx4_en_rx_ring *ring, int index)
169 {
170 	struct mlx4_en_rx_desc *rx_desc = (struct mlx4_en_rx_desc *)
171 	    (ring->buf + (index * ring->stride));
172 	struct mlx4_en_rx_mbuf *mb_list = ring->mbuf + index;
173 
174 	mb_list->mbuf = NULL;
175 
176 	if (mlx4_en_alloc_buf(ring, &rx_desc->data[0].addr, mb_list)) {
177 		priv->port_stats.rx_alloc_failed++;
178 		return (-ENOMEM);
179 	}
180 	return (0);
181 }
182 
183 static inline void
184 mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
185 {
186 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
187 }
188 
189 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
190 {
191 	struct mlx4_en_rx_ring *ring;
192 	int ring_ind;
193 	int buf_ind;
194 	int new_size;
195 	int err;
196 
197 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
198 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
199 			ring = priv->rx_ring[ring_ind];
200 
201 			err = mlx4_en_prepare_rx_desc(priv, ring,
202 						      ring->actual_size);
203 			if (err) {
204 				if (ring->actual_size == 0) {
205 					en_err(priv, "Failed to allocate "
206 						     "enough rx buffers\n");
207 					return -ENOMEM;
208 				} else {
209 					new_size =
210 						rounddown_pow_of_two(ring->actual_size);
211 					en_warn(priv, "Only %d buffers allocated "
212 						      "reducing ring size to %d\n",
213 						ring->actual_size, new_size);
214 					goto reduce_rings;
215 				}
216 			}
217 			ring->actual_size++;
218 			ring->prod++;
219 		}
220 	}
221 	return 0;
222 
223 reduce_rings:
224 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
225 		ring = priv->rx_ring[ring_ind];
226 		while (ring->actual_size > new_size) {
227 			ring->actual_size--;
228 			ring->prod--;
229 			mlx4_en_free_buf(ring,
230 			    ring->mbuf + ring->actual_size);
231 		}
232 	}
233 
234 	return 0;
235 }
236 
237 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
238 				struct mlx4_en_rx_ring *ring)
239 {
240 	int index;
241 
242 	en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
243 	       ring->cons, ring->prod);
244 
245 	/* Unmap and free Rx buffers */
246 	BUG_ON((u32) (ring->prod - ring->cons) > ring->actual_size);
247 	while (ring->cons != ring->prod) {
248 		index = ring->cons & ring->size_mask;
249 		en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
250 		mlx4_en_free_buf(ring, ring->mbuf + index);
251 		++ring->cons;
252 	}
253 }
254 
255 void mlx4_en_calc_rx_buf(struct net_device *dev)
256 {
257 	struct mlx4_en_priv *priv = netdev_priv(dev);
258 	int eff_mtu = dev->if_mtu + ETH_HLEN + VLAN_HLEN + ETH_FCS_LEN +
259 	    MLX4_NET_IP_ALIGN;
260 
261 	if (eff_mtu > MJUM16BYTES) {
262 		en_err(priv, "MTU(%d) is too big\n", dev->if_mtu);
263                 eff_mtu = MJUM16BYTES;
264         } else if (eff_mtu > MJUM9BYTES) {
265                 eff_mtu = MJUM16BYTES;
266         } else if (eff_mtu > MJUMPAGESIZE) {
267                 eff_mtu = MJUM9BYTES;
268         } else if (eff_mtu > MCLBYTES) {
269                 eff_mtu = MJUMPAGESIZE;
270         } else {
271                 eff_mtu = MCLBYTES;
272         }
273 
274 	priv->rx_mb_size = eff_mtu;
275 
276 	en_dbg(DRV, priv, "Effective RX MTU: %d bytes\n", eff_mtu);
277 }
278 
279 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
280 			   struct mlx4_en_rx_ring **pring,
281 			   u32 size, int node)
282 {
283 	struct mlx4_en_dev *mdev = priv->mdev;
284 	struct mlx4_en_rx_ring *ring;
285 	int err;
286 	int tmp;
287 	uint32_t x;
288 
289         ring = kzalloc(sizeof(struct mlx4_en_rx_ring), GFP_KERNEL);
290         if (!ring) {
291                 en_err(priv, "Failed to allocate RX ring structure\n");
292                 return -ENOMEM;
293         }
294 
295 	/* Create DMA descriptor TAG */
296 	if ((err = -bus_dma_tag_create(
297 	    bus_get_dma_tag(mdev->pdev->dev.bsddev),
298 	    1,				/* any alignment */
299 	    0,				/* no boundary */
300 	    BUS_SPACE_MAXADDR,		/* lowaddr */
301 	    BUS_SPACE_MAXADDR,		/* highaddr */
302 	    NULL, NULL,			/* filter, filterarg */
303 	    MJUM16BYTES,		/* maxsize */
304 	    1,				/* nsegments */
305 	    MJUM16BYTES,		/* maxsegsize */
306 	    0,				/* flags */
307 	    NULL, NULL,			/* lockfunc, lockfuncarg */
308 	    &ring->dma_tag))) {
309 		en_err(priv, "Failed to create DMA tag\n");
310 		goto err_ring;
311 	}
312 
313 	ring->prod = 0;
314 	ring->cons = 0;
315 	ring->size = size;
316 	ring->size_mask = size - 1;
317 	ring->stride = roundup_pow_of_two(
318 	    sizeof(struct mlx4_en_rx_desc) + DS_SIZE);
319 	ring->log_stride = ffs(ring->stride) - 1;
320 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
321 
322 	tmp = size * sizeof(struct mlx4_en_rx_mbuf);
323 
324         ring->mbuf = kzalloc(tmp, GFP_KERNEL);
325         if (ring->mbuf == NULL) {
326                 err = -ENOMEM;
327                 goto err_dma_tag;
328         }
329 
330 	err = -bus_dmamap_create(ring->dma_tag, 0, &ring->spare.dma_map);
331 	if (err != 0)
332 		goto err_info;
333 
334 	for (x = 0; x != size; x++) {
335 		err = -bus_dmamap_create(ring->dma_tag, 0,
336 		    &ring->mbuf[x].dma_map);
337 		if (err != 0) {
338 			while (x--)
339 				bus_dmamap_destroy(ring->dma_tag,
340 				    ring->mbuf[x].dma_map);
341 			goto err_info;
342 		}
343 	}
344 	en_dbg(DRV, priv, "Allocated MBUF ring at addr:%p size:%d\n",
345 		 ring->mbuf, tmp);
346 
347 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
348 				 ring->buf_size, 2 * PAGE_SIZE);
349 	if (err)
350 		goto err_dma_map;
351 
352 	err = mlx4_en_map_buffer(&ring->wqres.buf);
353 	if (err) {
354 		en_err(priv, "Failed to map RX buffer\n");
355 		goto err_hwq;
356 	}
357 	ring->buf = ring->wqres.buf.direct.buf;
358 	*pring = ring;
359 	return 0;
360 
361 err_hwq:
362 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
363 err_dma_map:
364 	for (x = 0; x != size; x++) {
365 		bus_dmamap_destroy(ring->dma_tag,
366 		    ring->mbuf[x].dma_map);
367 	}
368 	bus_dmamap_destroy(ring->dma_tag, ring->spare.dma_map);
369 err_info:
370 	vfree(ring->mbuf);
371 err_dma_tag:
372 	bus_dma_tag_destroy(ring->dma_tag);
373 err_ring:
374 	kfree(ring);
375 	return (err);
376 }
377 
378 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
379 {
380 	struct mlx4_en_rx_ring *ring;
381 	int i;
382 	int ring_ind;
383 	int err;
384 	int stride = roundup_pow_of_two(
385 	    sizeof(struct mlx4_en_rx_desc) + DS_SIZE);
386 
387 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
388 		ring = priv->rx_ring[ring_ind];
389 
390 		ring->prod = 0;
391 		ring->cons = 0;
392 		ring->actual_size = 0;
393 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
394                 ring->rx_mb_size = priv->rx_mb_size;
395 
396 		ring->stride = stride;
397 		if (ring->stride <= TXBB_SIZE)
398 			ring->buf += TXBB_SIZE;
399 
400 		ring->log_stride = ffs(ring->stride) - 1;
401 		ring->buf_size = ring->size * ring->stride;
402 
403 		memset(ring->buf, 0, ring->buf_size);
404 		mlx4_en_update_rx_prod_db(ring);
405 
406 		/* Initialize all descriptors */
407 		for (i = 0; i < ring->size; i++)
408 			mlx4_en_init_rx_desc(priv, ring, i);
409 
410 #ifdef INET
411 		/* Configure lro mngr */
412 		if (priv->dev->if_capenable & IFCAP_LRO) {
413 			if (tcp_lro_init(&ring->lro))
414 				priv->dev->if_capenable &= ~IFCAP_LRO;
415 			else
416 				ring->lro.ifp = priv->dev;
417 		}
418 #endif
419 	}
420 
421 
422 	err = mlx4_en_fill_rx_buffers(priv);
423 	if (err)
424 		goto err_buffers;
425 
426 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
427 		ring = priv->rx_ring[ring_ind];
428 
429 		ring->size_mask = ring->actual_size - 1;
430 		mlx4_en_update_rx_prod_db(ring);
431 	}
432 
433 	return 0;
434 
435 err_buffers:
436 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
437 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
438 
439 	ring_ind = priv->rx_ring_num - 1;
440 
441 	while (ring_ind >= 0) {
442 		ring = priv->rx_ring[ring_ind];
443 		if (ring->stride <= TXBB_SIZE)
444 			ring->buf -= TXBB_SIZE;
445 		ring_ind--;
446 	}
447 
448 	return err;
449 }
450 
451 
452 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
453 			     struct mlx4_en_rx_ring **pring,
454 			     u32 size, u16 stride)
455 {
456 	struct mlx4_en_dev *mdev = priv->mdev;
457 	struct mlx4_en_rx_ring *ring = *pring;
458 	uint32_t x;
459 
460 	mlx4_en_unmap_buffer(&ring->wqres.buf);
461 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
462 	for (x = 0; x != size; x++)
463 		bus_dmamap_destroy(ring->dma_tag, ring->mbuf[x].dma_map);
464 	/* free spare mbuf, if any */
465 	if (ring->spare.mbuf != NULL) {
466 		bus_dmamap_sync(ring->dma_tag, ring->spare.dma_map,
467 		    BUS_DMASYNC_POSTREAD);
468 		bus_dmamap_unload(ring->dma_tag, ring->spare.dma_map);
469 		m_freem(ring->spare.mbuf);
470 	}
471 	bus_dmamap_destroy(ring->dma_tag, ring->spare.dma_map);
472 	vfree(ring->mbuf);
473 	bus_dma_tag_destroy(ring->dma_tag);
474 	kfree(ring);
475 	*pring = NULL;
476 #ifdef CONFIG_RFS_ACCEL
477 	mlx4_en_cleanup_filters(priv, ring);
478 #endif
479 }
480 
481 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
482 				struct mlx4_en_rx_ring *ring)
483 {
484 #ifdef INET
485 	tcp_lro_free(&ring->lro);
486 #endif
487 	mlx4_en_free_rx_buf(priv, ring);
488 	if (ring->stride <= TXBB_SIZE)
489 		ring->buf -= TXBB_SIZE;
490 }
491 
492 
493 static void validate_loopback(struct mlx4_en_priv *priv, struct mbuf *mb)
494 {
495 	int i;
496 	int offset = ETHER_HDR_LEN;
497 
498 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
499 		if (*(mb->m_data + offset) != (unsigned char) (i & 0xff))
500 			goto out_loopback;
501 	}
502 	/* Loopback found */
503 	priv->loopback_ok = 1;
504 
505 out_loopback:
506 	m_freem(mb);
507 }
508 
509 
510 static inline int invalid_cqe(struct mlx4_en_priv *priv,
511 			      struct mlx4_cqe *cqe)
512 {
513 	/* Drop packet on bad receive or bad checksum */
514 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
515 		     MLX4_CQE_OPCODE_ERROR)) {
516 		en_err(priv, "CQE completed in error - vendor syndrom:%d syndrom:%d\n",
517 		       ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
518 		       ((struct mlx4_err_cqe *)cqe)->syndrome);
519 		return 1;
520 	}
521 	if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
522 		en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
523 		return 1;
524 	}
525 
526 	return 0;
527 }
528 
529 static struct mbuf *
530 mlx4_en_rx_mb(struct mlx4_en_priv *priv, struct mlx4_en_rx_ring *ring,
531     struct mlx4_en_rx_desc *rx_desc, struct mlx4_en_rx_mbuf *mb_list,
532     int length)
533 {
534 	struct mbuf *mb;
535 
536 	/* get mbuf */
537 	mb = mb_list->mbuf;
538 
539 	/* collect used fragment while atomically replacing it */
540 	if (mlx4_en_alloc_buf(ring, &rx_desc->data[0].addr, mb_list))
541 		return (NULL);
542 
543 	/* range check hardware computed value */
544 	if (unlikely(length > mb->m_len))
545 		length = mb->m_len;
546 
547 	/* update total packet length in packet header */
548 	mb->m_len = mb->m_pkthdr.len = length;
549 	return (mb);
550 }
551 
552 /* For cpu arch with cache line of 64B the performance is better when cqe size==64B
553  * To enlarge cqe size from 32B to 64B --> 32B of garbage (i.e. 0xccccccc)
554  * was added in the beginning of each cqe (the real data is in the corresponding 32B).
555  * The following calc ensures that when factor==1, it means we are aligned to 64B
556  * and we get the real cqe data*/
557 #define CQE_FACTOR_INDEX(index, factor) ((index << factor) + factor)
558 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
559 {
560 	struct mlx4_en_priv *priv = netdev_priv(dev);
561 	struct mlx4_cqe *cqe;
562 	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
563 	struct mlx4_en_rx_mbuf *mb_list;
564 	struct mlx4_en_rx_desc *rx_desc;
565 	struct mbuf *mb;
566 	struct mlx4_cq *mcq = &cq->mcq;
567 	struct mlx4_cqe *buf = cq->buf;
568 	int index;
569 	unsigned int length;
570 	int polled = 0;
571 	u32 cons_index = mcq->cons_index;
572 	u32 size_mask = ring->size_mask;
573 	int size = cq->size;
574 	int factor = priv->cqe_factor;
575 
576 	if (!priv->port_up)
577 		return 0;
578 
579 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
580 	 * descriptor offset can be deducted from the CQE index instead of
581 	 * reading 'cqe->index' */
582 	index = cons_index & size_mask;
583 	cqe = &buf[CQE_FACTOR_INDEX(index, factor)];
584 
585 	/* Process all completed CQEs */
586 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
587 		    cons_index & size)) {
588 		mb_list = ring->mbuf + index;
589 		rx_desc = (struct mlx4_en_rx_desc *)
590 		    (ring->buf + (index << ring->log_stride));
591 
592 		/*
593 		 * make sure we read the CQE after we read the ownership bit
594 		 */
595 		rmb();
596 
597 		if (invalid_cqe(priv, cqe)) {
598 			goto next;
599 		}
600 		/*
601 		 * Packet is OK - process it.
602 		 */
603 		length = be32_to_cpu(cqe->byte_cnt);
604 		length -= ring->fcs_del;
605 
606 		mb = mlx4_en_rx_mb(priv, ring, rx_desc, mb_list, length);
607 		if (unlikely(!mb)) {
608 			ring->errors++;
609 			goto next;
610 		}
611 
612 		ring->bytes += length;
613 		ring->packets++;
614 
615 		if (unlikely(priv->validate_loopback)) {
616 			validate_loopback(priv, mb);
617 			goto next;
618 		}
619 
620 		/* forward Toeplitz compatible hash value */
621 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->immed_rss_invalid);
622 		M_HASHTYPE_SET(mb, M_HASHTYPE_OPAQUE_HASH);
623 		mb->m_pkthdr.rcvif = dev;
624 		if (be32_to_cpu(cqe->vlan_my_qpn) &
625 		    MLX4_CQE_VLAN_PRESENT_MASK) {
626 			mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->sl_vid);
627 			mb->m_flags |= M_VLANTAG;
628 		}
629 		if (likely(dev->if_capenable &
630 		    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) &&
631 		    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
632 		    (cqe->checksum == cpu_to_be16(0xffff))) {
633 			priv->port_stats.rx_chksum_good++;
634 			mb->m_pkthdr.csum_flags =
635 			    CSUM_IP_CHECKED | CSUM_IP_VALID |
636 			    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
637 			mb->m_pkthdr.csum_data = htons(0xffff);
638 			/* This packet is eligible for LRO if it is:
639 			 * - DIX Ethernet (type interpretation)
640 			 * - TCP/IP (v4)
641 			 * - without IP options
642 			 * - not an IP fragment
643 			 */
644 #ifdef INET
645 			if (mlx4_en_can_lro(cqe->status) &&
646 					(dev->if_capenable & IFCAP_LRO)) {
647 				if (ring->lro.lro_cnt != 0 &&
648 						tcp_lro_rx(&ring->lro, mb, 0) == 0)
649 					goto next;
650 			}
651 
652 #endif
653 			/* LRO not possible, complete processing here */
654 			INC_PERF_COUNTER(priv->pstats.lro_misses);
655 		} else {
656 			mb->m_pkthdr.csum_flags = 0;
657 			priv->port_stats.rx_chksum_none++;
658 		}
659 
660 		/* Push it up the stack */
661 		dev->if_input(dev, mb);
662 
663 next:
664 		++cons_index;
665 		index = cons_index & size_mask;
666 		cqe = &buf[CQE_FACTOR_INDEX(index, factor)];
667 		if (++polled == budget)
668 			goto out;
669 	}
670 	/* Flush all pending IP reassembly sessions */
671 out:
672 #ifdef INET
673 	tcp_lro_flush_all(&ring->lro);
674 #endif
675 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
676 	mcq->cons_index = cons_index;
677 	mlx4_cq_set_ci(mcq);
678 	wmb(); /* ensure HW sees CQ consumer before we post new buffers */
679 	ring->cons = mcq->cons_index;
680 	ring->prod += polled; /* Polled descriptors were realocated in place */
681 	mlx4_en_update_rx_prod_db(ring);
682 	return polled;
683 
684 }
685 
686 /* Rx CQ polling - called by NAPI */
687 static int mlx4_en_poll_rx_cq(struct mlx4_en_cq *cq, int budget)
688 {
689         struct net_device *dev = cq->dev;
690         int done;
691 
692         done = mlx4_en_process_rx_cq(dev, cq, budget);
693         cq->tot_rx += done;
694 
695         return done;
696 
697 }
698 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
699 {
700 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
701 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
702         int done;
703 
704         // Shoot one within the irq context
705         // Because there is no NAPI in freeBSD
706         done = mlx4_en_poll_rx_cq(cq, MLX4_EN_RX_BUDGET);
707 	if (priv->port_up  && (done == MLX4_EN_RX_BUDGET) ) {
708 		cq->curr_poll_rx_cpu_id = curcpu;
709 		taskqueue_enqueue(cq->tq, &cq->cq_task);
710         }
711 	else {
712 		mlx4_en_arm_cq(priv, cq);
713 	}
714 }
715 
716 void mlx4_en_rx_que(void *context, int pending)
717 {
718         struct mlx4_en_cq *cq;
719 	struct thread *td;
720 
721         cq = context;
722 	td = curthread;
723 
724 	thread_lock(td);
725 	sched_bind(td, cq->curr_poll_rx_cpu_id);
726 	thread_unlock(td);
727 
728         while (mlx4_en_poll_rx_cq(cq, MLX4_EN_RX_BUDGET)
729                         == MLX4_EN_RX_BUDGET);
730         mlx4_en_arm_cq(cq->dev->if_softc, cq);
731 }
732 
733 
734 /* RSS related functions */
735 
736 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
737 				 struct mlx4_en_rx_ring *ring,
738 				 enum mlx4_qp_state *state,
739 				 struct mlx4_qp *qp)
740 {
741 	struct mlx4_en_dev *mdev = priv->mdev;
742 	struct mlx4_qp_context *context;
743 	int err = 0;
744 
745 	context = kmalloc(sizeof *context , GFP_KERNEL);
746 	if (!context) {
747 		en_err(priv, "Failed to allocate qp context\n");
748 		return -ENOMEM;
749 	}
750 
751 	err = mlx4_qp_alloc(mdev->dev, qpn, qp);
752 	if (err) {
753 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
754 		goto out;
755 	}
756 	qp->event = mlx4_en_sqp_event;
757 
758 	memset(context, 0, sizeof *context);
759 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
760 				qpn, ring->cqn, -1, context);
761 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
762 
763 	/* Cancel FCS removal if FW allows */
764 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
765 		context->param3 |= cpu_to_be32(1 << 29);
766 		ring->fcs_del = ETH_FCS_LEN;
767 	} else
768 		ring->fcs_del = 0;
769 
770 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
771 	if (err) {
772 		mlx4_qp_remove(mdev->dev, qp);
773 		mlx4_qp_free(mdev->dev, qp);
774 	}
775 	mlx4_en_update_rx_prod_db(ring);
776 out:
777 	kfree(context);
778 	return err;
779 }
780 
781 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
782 {
783 	int err;
784 	u32 qpn;
785 
786 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn, 0);
787 	if (err) {
788 		en_err(priv, "Failed reserving drop qpn\n");
789 		return err;
790 	}
791 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp);
792 	if (err) {
793 		en_err(priv, "Failed allocating drop qp\n");
794 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
795 		return err;
796 	}
797 
798 	return 0;
799 }
800 
801 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
802 {
803 	u32 qpn;
804 
805 	qpn = priv->drop_qp.qpn;
806 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
807 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
808 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
809 }
810 
811 /* Allocate rx qp's and configure them according to rss map */
812 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
813 {
814 	struct mlx4_en_dev *mdev = priv->mdev;
815 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
816 	struct mlx4_qp_context context;
817 	struct mlx4_rss_context *rss_context;
818 	int rss_rings;
819 	void *ptr;
820 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
821 			MLX4_RSS_TCP_IPV6);
822 	int i;
823 	int err = 0;
824 	int good_qps = 0;
825 	static const u32 rsskey[10] = { 0xD181C62C, 0xF7F4DB5B, 0x1983A2FC,
826 				0x943E1ADB, 0xD9389E6B, 0xD1039C2C, 0xA74499AD,
827 				0x593D56D9, 0xF3253C06, 0x2ADC1FFC};
828 
829 	en_dbg(DRV, priv, "Configuring rss steering\n");
830 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
831 				    priv->rx_ring_num,
832 				    &rss_map->base_qpn, 0);
833 	if (err) {
834 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
835 		return err;
836 	}
837 
838 	for (i = 0; i < priv->rx_ring_num; i++) {
839 		priv->rx_ring[i]->qpn = rss_map->base_qpn + i;
840 		err = mlx4_en_config_rss_qp(priv, priv->rx_ring[i]->qpn,
841 					    priv->rx_ring[i],
842 					    &rss_map->state[i],
843 					    &rss_map->qps[i]);
844 		if (err)
845 			goto rss_err;
846 
847 		++good_qps;
848 	}
849 
850 	/* Configure RSS indirection qp */
851 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
852 	if (err) {
853 		en_err(priv, "Failed to allocate RSS indirection QP\n");
854 		goto rss_err;
855 	}
856 	rss_map->indir_qp.event = mlx4_en_sqp_event;
857 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
858 				priv->rx_ring[0]->cqn, -1, &context);
859 
860 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
861 		rss_rings = priv->rx_ring_num;
862 	else
863 		rss_rings = priv->prof->rss_rings;
864 
865 	ptr = ((u8 *)&context) + offsetof(struct mlx4_qp_context, pri_path) +
866 	    MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
867 	rss_context = ptr;
868 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
869 					    (rss_map->base_qpn));
870 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
871 	if (priv->mdev->profile.udp_rss) {
872 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
873 		rss_context->base_qpn_udp = rss_context->default_qpn;
874 	}
875 	rss_context->flags = rss_mask;
876 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
877 	for (i = 0; i < 10; i++)
878 		rss_context->rss_key[i] = cpu_to_be32(rsskey[i]);
879 
880 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
881 			       &rss_map->indir_qp, &rss_map->indir_state);
882 	if (err)
883 		goto indir_err;
884 
885 	return 0;
886 
887 indir_err:
888 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
889 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
890 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
891 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
892 rss_err:
893 	for (i = 0; i < good_qps; i++) {
894 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
895 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
896 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
897 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
898 	}
899 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
900 	return err;
901 }
902 
903 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
904 {
905 	struct mlx4_en_dev *mdev = priv->mdev;
906 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
907 	int i;
908 
909 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
910 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
911 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
912 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
913 
914 	for (i = 0; i < priv->rx_ring_num; i++) {
915 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
916 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
917 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
918 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
919 	}
920 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
921 }
922 
923